diff --git a/.travis.yml b/.travis.yml index e217c8f5a7..d0e2696f10 100644 --- a/.travis.yml +++ b/.travis.yml @@ -36,10 +36,6 @@ before_install: # protobuf version. - sudo pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt - sudo pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker - - curl https://glide.sh/get | bash - - eval "$(GIMME_GO_VERSION=1.8.3 gimme)" - - go get -u github.com/alecthomas/gometalinter - - gometalinter --install - | function timeout() { perl -e 'alarm shift; exec @ARGV' "$@"; } script: diff --git a/CMakeLists.txt b/CMakeLists.txt index 5739c2a260..4921226ec1 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -27,7 +27,7 @@ if(NOT CMAKE_CROSSCOMPILING) endif(NOT CMAKE_CROSSCOMPILING) find_package(Git REQUIRED) find_package(Threads REQUIRED) -if(NOT ANDROID) +if(NOT ANDROID AND NOT IOS) find_package(Boost QUIET) endif() @@ -64,27 +64,29 @@ if(NOT CMAKE_BUILD_TYPE) FORCE) endif() -if(ANDROID) - if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16") - message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16") - elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21") - # TODO: support glog for Android api 16 ~ 19 in the future - message(WARNING "Using the unofficial git repository instead") +if(ANDROID OR IOS) + if(ANDROID) + if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16") + message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16") + elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21") + # TODO: support glog for Android api 16 ~ 19 in the future + message(WARNING "Using the unofficial git repository instead") + endif() endif() set(WITH_GPU OFF CACHE STRING - "Disable GPU when cross-compiling for Android" FORCE) + "Disable GPU when cross-compiling for Android and iOS" FORCE) set(WITH_AVX OFF CACHE STRING - "Disable AVX when cross-compiling for Android" FORCE) + "Disable AVX when cross-compiling for Android and iOS" FORCE) set(WITH_PYTHON OFF CACHE STRING - "Disable PYTHON when cross-compiling for Android" FORCE) + "Disable PYTHON when cross-compiling for Android and iOS" FORCE) set(WITH_RDMA OFF CACHE STRING - "Disable RDMA when cross-compiling for Android" FORCE) + "Disable RDMA when cross-compiling for Android and iOS" FORCE) set(WITH_MKLDNN OFF CACHE STRING - "Disable MKLDNN when cross-compiling for Android" FORCE) + "Disable MKLDNN when cross-compiling for Android and iOS" FORCE) set(WITH_MKLML OFF CACHE STRING - "Disable MKLML package when cross-compiling for Android" FORCE) -endif(ANDROID) + "Disable MKLML package when cross-compiling for Android and iOS" FORCE) +endif() set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING "A path setting third party libraries download & build directories.") diff --git a/README.md b/README.md index b9793c3eab..db0fbd88b2 100644 --- a/README.md +++ b/README.md @@ -51,19 +51,19 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl - **Connected to Products** In addition, PaddlePaddle is also designed to be easily deployable. At Baidu, - PaddlePaddle has been deployed into products or service with a vast number + PaddlePaddle has been deployed into products and services with a vast number of users, including ad click-through rate (CTR) prediction, large-scale image classification, optical character recognition(OCR), search ranking, computer virus detection, recommendation, etc. It is widely utilized in products at - Baidu and it has achieved a significant impact. We hope you can also exploit - the capability of PaddlePaddle to make a huge impact for your product. + 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. ## Installation It is recommended to check out the [Docker installation guide](http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/docker_install_en.html) before looking into the -[build from source guide](http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/build_from_source_en.html) +[build from source guide](http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/build_from_source_en.html). ## Documentation @@ -72,7 +72,7 @@ We provide [English](http://doc.paddlepaddle.org/develop/doc/) and - [Deep Learning 101](http://book.paddlepaddle.org/index.html) - You might want to start from this online interactive book that can run in Jupyter Notebook. + You might want to start from this online interactive book that can run in a Jupyter Notebook. - [Distributed Training](http://doc.paddlepaddle.org/develop/doc/howto/usage/cluster/cluster_train_en.html) diff --git a/benchmark/paddle/image/provider.py b/benchmark/paddle/image/provider.py index 1ac47212b5..4703944c87 100644 --- a/benchmark/paddle/image/provider.py +++ b/benchmark/paddle/image/provider.py @@ -22,5 +22,5 @@ def initHook(settings, height, width, color, num_class, **kwargs): def process(settings, file_list): for i in xrange(1024): img = np.random.rand(1, settings.data_size).reshape(-1, 1).flatten() - lab = random.randint(0, settings.num_class) + lab = random.randint(0, settings.num_class - 1) yield img.astype('float32'), int(lab) diff --git a/benchmark/paddle/image/run_mkldnn.sh b/benchmark/paddle/image/run_mkldnn.sh new file mode 100755 index 0000000000..e31fec1cd8 --- /dev/null +++ b/benchmark/paddle/image/run_mkldnn.sh @@ -0,0 +1,51 @@ +set -e + +function train() { + unset OMP_NUM_THREADS MKL_NUM_THREADS + export OMP_DYNAMIC="FALSE" + export KMP_AFFINITY="granularity=fine,compact,0,0" + topology=$1 + bs=$2 + use_mkldnn=$3 + if [ $3 == "True" ]; then + thread=1 + log="logs/${topology}-mkldnn-${bs}.log" + elif [ $3 == "False" ]; then + thread=`nproc` + # each trainer_count use only 1 core to avoid conflict + export OMP_NUM_THREADS=1 + export MKL_NUM_THREADS=1 + log="logs/${topology}-${thread}mklml-${bs}.log" + else + echo "Wrong input $3, use True or False." + exit 0 + fi + args="batch_size=${bs}" + 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} +} + +if [ ! -d "train.list" ]; then + echo " " > train.list +fi +if [ ! -d "logs" ]; then + mkdir logs +fi + +#========== mkldnn ==========# +train vgg 64 True +train vgg 128 True +train vgg 256 True + +#========== mklml ===========# +train vgg 64 False +train vgg 128 False +train vgg 256 False diff --git a/benchmark/paddle/image/vgg.py b/benchmark/paddle/image/vgg.py new file mode 100644 index 0000000000..b8429975f5 --- /dev/null +++ b/benchmark/paddle/image/vgg.py @@ -0,0 +1,103 @@ +#!/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) + +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)) + +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.") + +lab = data_layer('label', num_class) +loss = cross_entropy(input=vgg, label=lab) +outputs(loss) diff --git a/cmake/cblas.cmake b/cmake/cblas.cmake index 854066fd1d..8fdc382f0c 100644 --- a/cmake/cblas.cmake +++ b/cmake/cblas.cmake @@ -171,3 +171,10 @@ if (REFERENCE_CBLAS_INCLUDE_DIR AND REFERENCE_CBLAS_LIBRARY) add_definitions(-DPADDLE_USE_REFERENCE_CBLAS) message(STATUS "Found reference-cblas (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})") endif() + +if(IOS_USE_VECLIB_FOR_BLAS AND VECLIB_FOUND) + set(CBLAS_FOUND ON) + set(CBLAS_PROVIDER vecLib) + set(CBLAS_INC_DIR ${VECLIB_INC_DIR}) + add_definitions(-DPADDLE_USE_VECLIB) +endif() diff --git a/cmake/configure.cmake b/cmake/configure.cmake index 51c3b918cc..c1c93e17fd 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -49,11 +49,12 @@ if(NOT WITH_GOLANG) endif(NOT WITH_GOLANG) if(NOT WITH_GPU) - add_definitions(-DPADDLE_ONLY_CPU) add_definitions(-DHPPL_STUB_FUNC) list(APPEND CMAKE_CXX_SOURCE_FILE_EXTENSIONS cu) else() + add_definitions(-DPADDLE_WITH_CUDA) + FIND_PACKAGE(CUDA REQUIRED) if(${CUDA_VERSION_MAJOR} VERSION_LESS 7) diff --git a/cmake/cross_compiling/ios.cmake b/cmake/cross_compiling/ios.cmake new file mode 100644 index 0000000000..0b38943952 --- /dev/null +++ b/cmake/cross_compiling/ios.cmake @@ -0,0 +1,350 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT 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 is a toolchain file for cross-compiling for iOS, and the +# configuration largely refers to public toolchain file: +# https://raw.githubusercontent.com/leetal/ios-cmake/master/ios.toolchain.cmake +# and +# https://github.com/cristeab/ios-cmake +# +# Supports options: +# IOS_PLATFORM = OS (default) or SIMULATOR +# This decides if SDKS will be selected from the iPhoneOS.platform or iPhoneSimulator.platform folders +# OS - the default, used to build for iPhone and iPad physical devices, which have an arm arch. +# SIMULATOR - used to build for the Simulator platforms, which have an x86 arch. +# IOS_ARCH +# The archectures wanted to support, such "arm64", "armv7;arm64" +# IOS_DEPLOYMENT_TARGET +# The minimum iOS deployment version, such as "7.0" +# IOS_ENABLE_BITCODE = ON (default) or OFF +# IOS_USE_VECLIB_FOR_BLAS = OFF (default) or ON +# IOS_DEVELOPER_ROOT = automatic(default) or /path/to/platform/Developer folder +# By default this location is automatcially chosen based on the IOS_PLATFORM value above. +# If set manually, it will override the default location and force the user of a particular Developer Platform +# IOS_SDK_ROOT = automatic(default) or /path/to/platform/Developer/SDKs/SDK folder +# By default this location is automatcially chosen based on the IOS_DEVELOPER_ROOT value. +# In this case it will always be the most up-to-date SDK found in the IOS_DEVELOPER_ROOT path. +# If set manually, this will force the use of a specific SDK version + +# Macros: +# set_xcode_property (TARGET XCODE_PROPERTY XCODE_VALUE) +# A convenience macro for setting xcode specific properties on targets +# example: set_xcode_property (myioslib IPHONEOS_DEPLOYMENT_TARGET "3.1") +# find_host_package (PROGRAM ARGS) +# A macro used to find executable programs on the host system, not within the iOS environment. +# Thanks to the android-cmake project for providing the command + +if(NOT IOS) + return() +endif() + +set(CMAKE_SYSTEM_NAME Darwin) + +# Get the Xcode version being used. +execute_process(COMMAND xcodebuild -version + OUTPUT_VARIABLE XCODE_VERSION + RESULT_VARIABLE XCODE_VERSION_RESULT + ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) +if(NOT ${XCODE_VERSION_RESULT}) + string(REGEX MATCH "Xcode [0-9\\.]+" XCODE_VERSION "${XCODE_VERSION}") + string(REGEX REPLACE "Xcode ([0-9\\.]+)" "\\1" XCODE_VERSION "${XCODE_VERSION}") + message(STATUS "Building with Xcode version: ${XCODE_VERSION}") +else() + message(FATAL_ERROR "Cannot execute xcodebuild, please check whether xcode is installed.") +endif() + +# Required as of cmake 2.8.10 +set(CMAKE_OSX_DEPLOYMENT_TARGET "" CACHE STRING "Force unset of the deployment target for iOS" FORCE) + +# Setup iOS platform unless specified manually with IOS_PLATFORM +if(NOT DEFINED IOS_PLATFORM) + set(IOS_PLATFORM "OS") +endif() +set(IOS_PLATFORM ${IOS_PLATFORM} CACHE STRING "Type of iOS Platform") + +# Set the architecture for iOS +if(NOT DEFINED IOS_ARCH) + if(IOS_PLATFORM STREQUAL "OS") + # FIXME(liuyiqun): support "armv7;armv7s;arm64" future + set(IOS_ARCH "arm64") + elseif(IOS_PLATFORM STREQUAL "SIMULATOR") + set(IOS_ARCH "i386;x86_64") + elseif(IOS_PLATFORM STREQUAL "WATCHOS") + set(IOS_ARCH armv7k) + endif() +endif() +set(CMAKE_OSX_ARCHITECTURES ${IOS_ARCH} CACHE string "Build architecture for iOS") + +# Specify minimum iOS deployment version +if(NOT DEFINED IOS_DEPLOYMENT_TARGET) + set(IOS_DEPLOYMENT_TARGET "7.0") +endif() +set(IOS_DEPLOYMENT_TARGET ${IOS_DEPLOYMENT_TARGET} CACHE STRING "Minimum iOS version") + +# Whether to enable bitcode +if(NOT DEFINED IOS_ENABLE_BITCODE) + set(IOS_ENABLE_BITCODE ON) +endif() +set(IOS_ENABLE_BITCODE ${IOS_ENABLE_BITCODE} CACHE BOOL "Whether to enable bitcode") + +if(NOT DEFINED IOS_USE_VECLIB_FOR_BLAS) + set(IOS_USE_VECLIB_FOR_BLAS OFF) +endif() +set(IOS_USE_VECLIB_FOR_BLAS ${IOS_UES_VECLIB_FOR_BLAS} CACHE BOOL "Whether to use veclib") + +# Check the platform selection and setup for developer root +if(${IOS_PLATFORM} STREQUAL "OS") + set(IOS_PLATFORM_LOCATION "iPhoneOS.platform") + set(XCODE_IOS_PLATFORM iphoneos) + + # This causes the installers to properly locate the output libraries + set(CMAKE_XCODE_EFFECTIVE_PLATFORMS "-iphoneos") +elseif(${IOS_PLATFORM} STREQUAL "SIMULATOR") + set(IOS_PLATFORM_LOCATION "iPhoneSimulator.platform") + set(XCODE_IOS_PLATFORM iphonesimulator) + + # This causes the installers to properly locate the output libraries + set(CMAKE_XCODE_EFFECTIVE_PLATFORMS "-iphonesimulator") +elseif(${IOS_PLATFORM} STREQUAL "WATCHOS") + set(IOS_PLATFORM_LOCATION "WatchOS.platform") + set(XCODE_IOS_PLATFORM watchos) + + # This causes the installers to properly locate the output libraries + set(CMAKE_XCODE_EFFECTIVE_PLATFORMS "-watchos") +else(${IOS_PLATFORM} STREQUAL "OS") + message(FATAL_ERROR "Unsupported IOS_PLATFORM value selected. Please set to\n" + "\t OS, SIMULATOR, or WATCHOS.") +endif() + +# Check iOS developer toolchain +if(NOT DEFINED IOS_DEVELOPER_ROOT) + # Setup iOS developer location + execute_process(COMMAND xcode-select -print-path + OUTPUT_VARIABLE XCODE_DEVELOPER_DIR + RESULT_VARIABLE XCODE_DEVELOPER_DIR_RESULT + ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) + # Xcode 4.3 changed the installation location, choose the most recent one available + if(${XCODE_VERSION} VERSION_LESS "4.3.0") + set(IOS_DEVELOPER_ROOT "/Developer/Platforms/${IOS_PLATFORM_LOCATION}/Developer") + else() + set(IOS_DEVELOPER_ROOT "${XCODE_DEVELOPER_DIR}/Platforms/${IOS_PLATFORM_LOCATION}/Developer") + endif() +endif() +if(EXISTS ${IOS_DEVELOPER_ROOT}) + set(IOS_DEVELOPER_ROOT ${IOS_DEVELOPER_ROOT} CACHE PATH "Location of iOS Platform") +else() + message(FATAL_ERROR "Invalid IOS_DEVELOPER_ROOT: ${IOS_DEVELOPER_ROOT} does not exist.") +endif() + +# Check iOS SDK +if(NOT DEFINED IOS_SDK_ROOT) + # Find and use the most recent iOS sdk + file(GLOB IOS_SDK_LISTS "${IOS_DEVELOPER_ROOT}/SDKs/*") + if(IOS_SDK_LISTS) + list(SORT IOS_SDK_LISTS) + list(REVERSE IOS_SDK_LISTS) + list(GET IOS_SDK_LISTS 0 IOS_SDK_ROOT) + else(IOS_SDK_LISTS) + message(FATAL_ERROR "No iOS SDK's found in default search path ${IOS_DEVELOPER_ROOT}." + " Please manually set IOS_SDK_ROOT or install the iOS SDK.") + endif(IOS_SDK_LISTS) +endif() +if(EXISTS ${IOS_SDK_ROOT}) + set(IOS_SDK_ROOT ${IOS_SDK_ROOT} CACHE PATH "Location of the selected iOS SDK") + message(STATUS "iOS toolchain: ${IOS_SDK_ROOT}") +else() + message(FATAL_ERROR "Invalid IOS_SDK_ROOT: ${IOS_SDK_ROOT} does not exist.") +endif() + +# Set the sysroot default to the most recent SDK +set(CMAKE_OSX_SYSROOT ${IOS_SDK_ROOT} CACHE PATH "Sysroot used for iOS support") + +# Get version of iOS SDK +execute_process(COMMAND xcodebuild -sdk ${CMAKE_OSX_SYSROOT} -version SDKVersion + OUTPUT_VARIABLE IOS_SDK_VERSION + RESULT_VARIABLE IOS_SDK_VERSION_RESULT + ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) +if(${IOS_SDK_VERSION_RESULT}) + string(REGEX MATCH "(([0-9]+)\\.)+([0-9]+)" IOS_SDK_VERSION "${IOS_SDK_ROOT}") +endif() +if(NOT IOS_SDK_VERSION) + message(WARNING "Cannot get SDK's version.") + set(IOS_SDK_VERSION 1) +endif() +set(CMAKE_SYSTEM_VERSION ${IOS_SDK_VERSION}) + +# Find the C & C++ compilers for the specified SDK. +if(NOT CMAKE_C_COMPILER) + # Default to use clang + execute_process(COMMAND xcrun -sdk ${CMAKE_OSX_SYSROOT} -find clang + OUTPUT_VARIABLE IOS_C_COMPILER + RESULT_VARIABLE IOS_C_COMPILER_RESULT + ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) + if(${IOS_C_COMPILER_RESULT}) + get_filename_component(IOS_C_COMPILER clang PROGRAM) + endif() +else(NOT CMAKE_C_COMPILER) + # User can set it in cmake command + get_filename_component(IOS_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM) +endif(NOT CMAKE_C_COMPILER) +if(NOT EXISTS ${IOS_C_COMPILER}) + message(FATAL_ERROR "Cannot find C compiler: ${IOS_C_COMPILER}") +endif() + +if(NOT CMAKE_CXX_COMPILER) + # Default to use clang++ + execute_process(COMMAND xcrun -sdk ${CMAKE_OSX_SYSROOT} -find clang++ + OUTPUT_VARIABLE IOS_CXX_COMPILER + RESULT_VARIABLE IOS_CXX_COMPILER_RESULT + ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) + if(${IOS_CXX_COMPILER_RESULT}) + get_filename_component(IOS_CXX_COMPILER clang++ PROGRAM) + endif() +else(NOT CMAKE_CXX_COMPILER) + # User can set it in cmake command + get_filename_component(IOS_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM) +endif(NOT CMAKE_CXX_COMPILER) +if(NOT EXISTS ${IOS_CXX_COMPILER}) + message(FATAL_ERROR "Cannot find CXX compiler: ${IOS_CXX_COMPILER}") +endif() + +set(CMAKE_C_COMPILER ${IOS_C_COMPILER} CACHE PATH "C compiler" FORCE) +set(CMAKE_CXX_COMPILER ${IOS_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE) + +set(CMAKE_C_OSX_COMPATIBILITY_VERSION_FLAG "-compatibility_version ") +set(CMAKE_C_OSX_CURRENT_VERSION_FLAG "-current_version ") +set(CMAKE_CXX_OSX_COMPATIBILITY_VERSION_FLAG "${CMAKE_C_OSX_COMPATIBILITY_VERSION_FLAG}") +set(CMAKE_CXX_OSX_CURRENT_VERSION_FLAG "${CMAKE_C_OSX_CURRENT_VERSION_FLAG}") + +# Set iOS specific C/C++ flags +if(IOS_PLATFORM STREQUAL "OS") + if(XCODE_VERSION VERSION_LESS "7.0") + set(XCODE_IOS_PLATFORM_VERSION_FLAGS "-mios-version-min=${IOS_DEPLOYMENT_TARGET}") + else() + # Xcode 7.0+ uses flags we can build directly from XCODE_IOS_PLATFORM. + set(XCODE_IOS_PLATFORM_VERSION_FLAGS "-m${XCODE_IOS_PLATFORM}-version-min=${IOS_DEPLOYMENT_TARGET}") + endif() +else() + set(XCODE_IOS_FLATFORM_VERSION_FLAGS "-mios-simulator-version-min=${IOS_DEPLOYMENT_TARGET}") +endif() + +if(IOS_ENABLE_BITCODE) + set(XCODE_IOS_BITCODE_FLAGS "${IOS_COMPILER_FLAGS} -fembed-bitcode") +else() + set(XCODE_IOS_BITCODE_FLAGS "") +endif() + +set(IOS_COMPILER_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} ${XCODE_IOS_BITCODE_FLAGS}") + +# Hidden visibilty is required for cxx on iOS +set(CMAKE_C_FLAGS "${IOS_COMPILER_FLAGS} ${CMAKE_C_FLAGS}" CACHE STRING "C flags") +set(CMAKE_CXX_FLAGS "${IOS_COMPILER_FLAGS} -fvisibility-inlines-hidden ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags") + +set(IOS_LINK_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} -Wl,-search_paths_first") + +if(IOS_USE_VECLIB_FOR_BLAS) + # Find vecLib for iOS + set(VECLIB_SEARCH_DIRS + ${IOS_SDK_ROOT}/System/Library/Frameworks/Accelerate.framework/Versions/Current/Frameworks + ${IOS_SDK_ROOT}/System/Library/Frameworks/Accelerate.framework/Frameworks + ) + find_path(VECLIB_INC_DIR vecLib.h PATHS ${VECLIB_SEARCH_DIRS}/vecLib.framework/Headers) + + include(FindPackageHandleStandardArgs) + find_package_handle_standard_args(vecLib DEFAULT_MSG VECLIB_INC_DIR) + + if(VECLIB_FOUND) + if(VECLIB_INC_DIR MATCHES "^/System/Library/Frameworks/vecLib.framework.*") + set(IOS_LINK_FLAGS ${IOS_LINK_FLAGS} -lcblas "-framework vecLib") + message(STATUS "Found standalone vecLib.framework") + else() + set(IOS_LINK_FLAGS ${IOS_LINK_FLAGS} -lcblas "-framework Accelerate") + message(STATUS "Found vecLib as part of Accelerate.framework") + endif() + + endif() +endif() + +set(CMAKE_C_LINK_FLAGS "${IOS_LINK_FLAGS} ${CMAKE_C_LINK_FLAGS}") +set(CMAKE_CXX_LINK_FLAGS "${IOS_LINK_FLAGS} ${CMAKE_CXX_LINK_FLAGS}") + +set(CMAKE_PLATFORM_HAS_INSTALLNAME 1) +if(NOT IOS_ENABLE_BITCODE) + set(CMAKE_SHARED_LIBRARY_CREATE_C_FLAGS "-dynamiclib -headerpad_max_install_names") + set(CMAKE_SHARED_MODULE_CREATE_C_FLAGS "-bundle -headerpad_max_install_names") +else() + set(CMAKE_SHARED_LIBRARY_CREATE_C_FLAGS "-dynamiclib") + set(CMAKE_SHARED_MODULE_CREATE_C_FLAGS "-bundle") +endif() +set(CMAKE_SHARED_MODULE_LOADER_C_FLAG "-Wl,-bundle_loader,") +set(CMAKE_SHARED_MODULE_LOADER_CXX_FLAG "-Wl,-bundle_loader,") +set(CMAKE_FIND_LIBRARY_SUFFIXES ".dylib" ".so" ".a") + +# hack: if a new cmake (which uses CMAKE_INSTALL_NAME_TOOL) runs on an old build tree +# (where install_name_tool was hardcoded) and where CMAKE_INSTALL_NAME_TOOL isn't in the cache +# and still cmake didn't fail in CMakeFindBinUtils.cmake (because it isn't rerun) +# hardcode CMAKE_INSTALL_NAME_TOOL here to install_name_tool, so it behaves as it did before, Alex +if(NOT DEFINED CMAKE_INSTALL_NAME_TOOL) + find_program(CMAKE_INSTALL_NAME_TOOL install_name_tool) +endif() + +# Set the find root to the iOS developer roots and to user defined paths +set(CMAKE_FIND_ROOT_PATH ${IOS_DEVELOPER_ROOT} ${IOS_SDK_ROOT} ${CMAKE_PREFIX_PATH} + CACHE string "iOS find search path root") + +# default to searching for frameworks first +set(CMAKE_FIND_FRAMEWORK FIRST) + +# set up the default search directories for frameworks +set(CMAKE_SYSTEM_FRAMEWORK_PATH + ${IOS_SDK_ROOT}/System/Library/Frameworks + ${IOS_SDK_ROOT}/System/Library/PrivateFrameworks + ${IOS_SDK_ROOT}/Developer/Library/Frameworks + ) + +# only search the iOS sdks, not the remainder of the host filesystem +set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER) +set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY) +set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY) + +message(STATUS "iOS: Targeting iOS '${CMAKE_SYSTEM_VERSION}', " + "building for '${IOS_PLATFORM}' platform, with architecture '${CMAKE_OSX_ARCHITECTURES}'") +message(STATUS "System CMAKE_C_FLAGS: ${CMAKE_C_FLAGS}") +message(STATUS "System CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}") + +# Used in ExternalProject command +string(REPLACE ";" "\\$" EXTERNAL_IOS_ARCHITECTURES "${CMAKE_OSX_ARCHITECTURES}") +set(EXTERNAL_OPTIONAL_ARGS + -DCMAKE_OSX_SYSROOT=${CMAKE_OSX_SYSROOT} + -DCMAKE_OSX_ARCHITECTURES=${EXTERNAL_IOS_ARCHITECTURES}) + +# This little macro lets you set any XCode specific property +macro(set_xcode_property TARGET XCODE_PROPERTY XCODE_VALUE) + set_property (TARGET ${TARGET} PROPERTY XCODE_ATTRIBUTE_${XCODE_PROPERTY} ${XCODE_VALUE}) +endmacro(set_xcode_property) + +# This macro lets you find executable programs on the host system +macro(find_host_package) + set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER) + set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY NEVER) + set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE NEVER) + set(IOS FALSE) + + find_package(${ARGN}) + + set(IOS TRUE) + set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM ONLY) + set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY) + set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY) +endmacro(find_host_package) diff --git a/cmake/external/gflags.cmake b/cmake/external/gflags.cmake index 01a2f4d5fa..957f8271e4 100644 --- a/cmake/external/gflags.cmake +++ b/cmake/external/gflags.cmake @@ -39,13 +39,14 @@ ExternalProject_Add( PREFIX ${GFLAGS_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} - CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} - CMAKE_ARGS -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} - CMAKE_ARGS -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} - CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${GFLAGS_INSTALL_DIR} - CMAKE_ARGS -DCMAKE_POSITION_INDEPENDENT_CODE=ON - CMAKE_ARGS -DBUILD_TESTING=OFF - CMAKE_ARGS -DCMAKE_BUILD_TYPE=Release + -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} + -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} + -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} + -DCMAKE_INSTALL_PREFIX=${GFLAGS_INSTALL_DIR} + -DCMAKE_POSITION_INDEPENDENT_CODE=ON + -DBUILD_TESTING=OFF + -DCMAKE_BUILD_TYPE=Release + ${EXTERNAL_OPTIONAL_ARGS} CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GFLAGS_INSTALL_DIR} -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON -DCMAKE_BUILD_TYPE:STRING=Release diff --git a/cmake/external/glog.cmake b/cmake/external/glog.cmake index b450a30166..b3fef738cc 100644 --- a/cmake/external/glog.cmake +++ b/cmake/external/glog.cmake @@ -34,16 +34,17 @@ ExternalProject_Add( PREFIX ${GLOG_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} - CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} - CMAKE_ARGS -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} - CMAKE_ARGS -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} - CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${GLOG_INSTALL_DIR} - CMAKE_ARGS -DCMAKE_INSTALL_LIBDIR=${GLOG_INSTALL_DIR}/lib - CMAKE_ARGS -DCMAKE_POSITION_INDEPENDENT_CODE=ON - CMAKE_ARGS -DWITH_GFLAGS=ON - CMAKE_ARGS -Dgflags_DIR=${GFLAGS_INSTALL_DIR}/lib/cmake/gflags - CMAKE_ARGS -DBUILD_TESTING=OFF - CMAKE_ARGS -DCMAKE_BUILD_TYPE=Release + -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} + -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} + -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} + -DCMAKE_INSTALL_PREFIX=${GLOG_INSTALL_DIR} + -DCMAKE_INSTALL_LIBDIR=${GLOG_INSTALL_DIR}/lib + -DCMAKE_POSITION_INDEPENDENT_CODE=ON + -DWITH_GFLAGS=ON + -Dgflags_DIR=${GFLAGS_INSTALL_DIR}/lib/cmake/gflags + -DBUILD_TESTING=OFF + -DCMAKE_BUILD_TYPE=Release + ${EXTERNAL_OPTIONAL_ARGS} CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GLOG_INSTALL_DIR} -DCMAKE_INSTALL_LIBDIR:PATH=${GLOG_INSTALL_DIR}/lib -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON diff --git a/cmake/external/gtest.cmake b/cmake/external/gtest.cmake index e3970073a1..6a2a79b763 100644 --- a/cmake/external/gtest.cmake +++ b/cmake/external/gtest.cmake @@ -48,15 +48,16 @@ IF(WITH_TESTING) PREFIX ${GTEST_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} - CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} - CMAKE_ARGS -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} - CMAKE_ARGS -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} - CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${GTEST_INSTALL_DIR} - CMAKE_ARGS -DCMAKE_POSITION_INDEPENDENT_CODE=ON - CMAKE_ARGS -DBUILD_GMOCK=ON - CMAKE_ARGS -Dgtest_disable_pthreads=ON - CMAKE_ARGS -Dgtest_force_shared_crt=ON - CMAKE_ARGS -DCMAKE_BUILD_TYPE=Release + -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} + -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} + -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} + -DCMAKE_INSTALL_PREFIX=${GTEST_INSTALL_DIR} + -DCMAKE_POSITION_INDEPENDENT_CODE=ON + -DBUILD_GMOCK=ON + -Dgtest_disable_pthreads=ON + -Dgtest_force_shared_crt=ON + -DCMAKE_BUILD_TYPE=Release + ${EXTERNAL_OPTIONAL_ARGS} CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GTEST_INSTALL_DIR} -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON -DCMAKE_BUILD_TYPE:STRING=Release diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index 4fc8d43fc1..143b57a954 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -29,30 +29,41 @@ IF(NOT ${CBLAS_FOUND}) "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}" CACHE FILEPATH "openblas library." FORCE) - IF(APPLE) - SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}") - SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs) - ELSE() - SET(COMMON_ARGS CC=${CMAKE_C_COMPILER} NO_SHARED=1 NO_LAPACK=1 libs) - ENDIF() + SET(OPENBLAS_CC "${CMAKE_C_COMPILER}") IF(CMAKE_CROSSCOMPILING) + SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER}) + GET_FILENAME_COMPONENT(CROSS_SUFFIX ${CMAKE_C_COMPILER} DIRECTORY) + SET(CROSS_SUFFIX ${CROSS_SUFFIX}/) IF(ANDROID) # arm_soft_fp_abi branch of OpenBLAS to support softfp # https://github.com/xianyi/OpenBLAS/tree/arm_soft_fp_abi SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5") IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$") - SET(TARGET "ARMV7") + SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0) ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a") - SET(TARGET "ARMV8") + SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0) + ENDIF() + ELSEIF(IOS) + # FIXME(liuyiqun): support multiple architectures + SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5") + SET(OPENBLAS_CC "${OPENBLAS_CC} ${CMAKE_C_FLAGS} -isysroot ${CMAKE_OSX_SYSROOT}") + IF(CMAKE_OSX_ARCHITECTURES MATCHES "armv7") + SET(OPENBLAS_CC "${OPENBLAS_CC} -arch armv7") + SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0) + ELSEIF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64") + SET(OPENBLAS_CC "${OPENBLAS_CC} -arch arm64") + SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0 CROSS_SUFFIX=${CROSS_SUFFIX}) ENDIF() - SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=${TARGET} ARM_SOFTFP_ABI=1 USE_THREAD=0) ELSEIF(RPI) # use hardfp SET(OPENBLAS_COMMIT "v0.2.20") - SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=ARMV7 USE_THREAD=0) + SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 USE_THREAD=0) ENDIF() ELSE() + IF(APPLE) + SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}") + ENDIF() SET(OPENBLAS_COMMIT "v0.2.20") SET(OPTIONAL_ARGS "") IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^x86(_64)?$") @@ -60,6 +71,8 @@ IF(NOT ${CBLAS_FOUND}) ENDIF() ENDIF() + SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs) + ExternalProject_Add( extern_openblas ${EXTERNAL_PROJECT_LOG_ARGS} diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake index a887be2e2a..7cf7ba85cc 100644 --- a/cmake/external/protobuf.cmake +++ b/cmake/external/protobuf.cmake @@ -173,7 +173,8 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST) "-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}" "-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}" "-Dprotobuf_WITH_ZLIB=ON" - "-DZLIB_ROOT:FILEPATH=${ZLIB_ROOT}") + "-DZLIB_ROOT:FILEPATH=${ZLIB_ROOT}" + ${EXTERNAL_OPTIONAL_ARGS}) SET(OPTIONAL_CACHE_ARGS "-DZLIB_ROOT:STRING=${ZLIB_ROOT}") ENDIF() diff --git a/cmake/external/python.cmake b/cmake/external/python.cmake index 490c87d67e..46c68cce32 100644 --- a/cmake/external/python.cmake +++ b/cmake/external/python.cmake @@ -12,16 +12,17 @@ # See the License for the specific language governing permissions and # limitations under the License. -INCLUDE(ExternalProject) +IF(NOT WITH_PYTHON) + return() +ENDIF() + INCLUDE(python_module) FIND_PACKAGE(PythonInterp 2.7) -IF(WITH_PYTHON) - FIND_PACKAGE(PythonLibs 2.7) - # Fixme: Maybe find a static library. Get SHARED/STATIC by FIND_PACKAGE. - ADD_LIBRARY(python SHARED IMPORTED GLOBAL) - SET_PROPERTY(TARGET python PROPERTY IMPORTED_LOCATION ${PYTHON_LIBRARIES}) -ENDIF(WITH_PYTHON) +FIND_PACKAGE(PythonLibs 2.7) +# Fixme: Maybe find a static library. Get SHARED/STATIC by FIND_PACKAGE. +ADD_LIBRARY(python SHARED IMPORTED GLOBAL) +SET_PROPERTY(TARGET python PROPERTY IMPORTED_LOCATION ${PYTHON_LIBRARIES}) SET(py_env "") IF(PYTHONINTERP_FOUND) @@ -36,9 +37,5 @@ IF(PYTHONINTERP_FOUND) ENDIF() ENDIF(PYTHONINTERP_FOUND) -IF(WITH_PYTHON) - INCLUDE_DIRECTORIES(${PYTHON_INCLUDE_DIR}) - INCLUDE_DIRECTORIES(${PYTHON_NUMPY_INCLUDE_DIR}) -ELSE() - SET(PYTHON_LIBRARIES "") -ENDIF() +INCLUDE_DIRECTORIES(${PYTHON_INCLUDE_DIR}) +INCLUDE_DIRECTORIES(${PYTHON_NUMPY_INCLUDE_DIR}) diff --git a/cmake/external/swig.cmake b/cmake/external/swig.cmake index 744c766ee7..ce088ae7ea 100644 --- a/cmake/external/swig.cmake +++ b/cmake/external/swig.cmake @@ -12,6 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. +IF(NOT WITH_SWIG_PY) + return() +ENDIF() + FIND_PACKAGE(SWIG) IF(NOT SWIG_FOUND) diff --git a/cmake/external/warpctc.cmake b/cmake/external/warpctc.cmake index 2d7daed9bc..bb258c7b55 100644 --- a/cmake/external/warpctc.cmake +++ b/cmake/external/warpctc.cmake @@ -16,25 +16,14 @@ INCLUDE(ExternalProject) SET(WARPCTC_SOURCES_DIR ${THIRD_PARTY_PATH}/warpctc) SET(WARPCTC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/warpctc) -SET(WARPCTC_INCLUDE_DIR "${WARPCTC_INSTALL_DIR}/include" CACHE PATH "Warp-ctc Directory" FORCE) -INCLUDE_DIRECTORIES(${WARPCTC_INCLUDE_DIR}) - -SET(WARPCTC_LIB_DIR "${WARPCTC_INSTALL_DIR}/lib" CACHE PATH "Warp-ctc Library Directory" FORCE) - -IF(WIN32) - SET(WARPCTC_LIBRARIES - "${WARPCTC_INSTALL_DIR}/lib/warpctc.dll" CACHE FILEPATH "Warp-ctc Library" FORCE) -ELSE(WIN32) - IF(APPLE) - SET(_warpctc_SHARED_SUFFIX dylib) - ELSE(APPLE) - SET(_warpctc_SHARED_SUFFIX so) - ENDIF(APPLE) - - SET(WARPCTC_LIBRARIES - "${WARPCTC_INSTALL_DIR}/lib/libwarpctc.${_warpctc_SHARED_SUFFIX}" CACHE FILEPATH "Warp-ctc Library" FORCE) -ENDIF(WIN32) +SET(WARPCTC_INCLUDE_DIR "${WARPCTC_INSTALL_DIR}/include" + CACHE PATH "Warp-ctc Directory" FORCE) +# Used in unit test test_WarpCTCLayer +SET(WARPCTC_LIB_DIR "${WARPCTC_INSTALL_DIR}/lib" + CACHE PATH "Warp-ctc Library Directory" FORCE) +SET(WARPCTC_LIBRARIES "${WARPCTC_INSTALL_DIR}/lib/libwarpctc${CMAKE_SHARED_LIBRARY_SUFFIX}" + CACHE FILEPATH "Warp-ctc Library" FORCE) IF(CMAKE_CXX_COMPILER_ID STREQUAL "Clang" OR CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" ) SET(USE_OMP OFF) @@ -49,22 +38,26 @@ ExternalProject_Add( PREFIX ${WARPCTC_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} - CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} - CMAKE_ARGS -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} - CMAKE_ARGS -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} - CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${WARPCTC_INSTALL_DIR} - CMAKE_ARGS -DWITH_GPU=${WITH_GPU} - CMAKE_ARGS -DWITH_OMP=${USE_OMP} - CMAKE_ARGS -DWITH_TORCH=OFF - CMAKE_ARGS -DCMAKE_DISABLE_FIND_PACKAGE_Torch=ON - CMAKE_ARGS -DBUILD_SHARED=ON - CMAKE_ARGS -DCMAKE_POSITION_INDEPENDENT_CODE=ON - CMAKE_ARGS -DCMAKE_BUILD_TYPE=Release + -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} + -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} + -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} + -DCMAKE_INSTALL_PREFIX=${WARPCTC_INSTALL_DIR} + -DWITH_GPU=${WITH_GPU} + -DWITH_OMP=${USE_OMP} + -DWITH_TORCH=OFF + -DCMAKE_DISABLE_FIND_PACKAGE_Torch=ON + -DBUILD_SHARED=ON + -DCMAKE_POSITION_INDEPENDENT_CODE=ON + -DCMAKE_BUILD_TYPE=Release + ${EXTERNAL_OPTIONAL_ARGS} CMAKE_CACHE_ARGS -DCMAKE_BUILD_TYPE:STRING=Release -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON -DCMAKE_INSTALL_PREFIX:PATH=${WARPCTC_INSTALL_DIR} ) +MESSAGE(STATUS "warp-ctc library: ${WARPCTC_LIBRARIES}") +INCLUDE_DIRECTORIES(${WARPCTC_INCLUDE_DIR}) + ADD_LIBRARY(warpctc STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET warpctc PROPERTY IMPORTED_LOCATION ${WARPCTC_LIBRARIES}) ADD_DEPENDENCIES(warpctc extern_warpctc) diff --git a/cmake/external/zlib.cmake b/cmake/external/zlib.cmake index 5aecab90ca..c496a52b78 100644 --- a/cmake/external/zlib.cmake +++ b/cmake/external/zlib.cmake @@ -34,15 +34,16 @@ ExternalProject_Add( GIT_TAG "v1.2.8" PREFIX ${ZLIB_SOURCES_DIR} UPDATE_COMMAND "" - CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} - CMAKE_ARGS -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} - CMAKE_ARGS -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} - CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${ZLIB_INSTALL_DIR} - CMAKE_ARGS -DBUILD_SHARED_LIBS=OFF - CMAKE_ARGS -DCMAKE_POSITION_INDEPENDENT_CODE=ON - CMAKE_ARGS -DCMAKE_MACOSX_RPATH=ON - CMAKE_ARGS -DCMAKE_BUILD_TYPE=Release + -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} + -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} + -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} + -DCMAKE_INSTALL_PREFIX=${ZLIB_INSTALL_DIR} + -DBUILD_SHARED_LIBS=OFF + -DCMAKE_POSITION_INDEPENDENT_CODE=ON + -DCMAKE_MACOSX_RPATH=ON + -DCMAKE_BUILD_TYPE=Release + ${EXTERNAL_OPTIONAL_ARGS} CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${ZLIB_INSTALL_DIR} -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON -DCMAKE_BUILD_TYPE:STRING=Release diff --git a/cmake/flags.cmake b/cmake/flags.cmake index ff246b2eb4..4593ae6180 100644 --- a/cmake/flags.cmake +++ b/cmake/flags.cmake @@ -128,8 +128,10 @@ set(GPU_COMMON_FLAGS ) if (APPLE) - # On Mac OS X build fat binaries with x86_64 architectures by default. - set (CMAKE_OSX_ARCHITECTURES "x86_64" CACHE STRING "Build architectures for OSX" FORCE) + if(NOT CMAKE_CROSSCOMPILING) + # On Mac OS X build fat binaries with x86_64 architectures by default. + set (CMAKE_OSX_ARCHITECTURES "x86_64" CACHE STRING "Build architectures for OSX" FORCE) + endif() else() set(GPU_COMMON_FLAGS -Wall diff --git a/cmake/generic.cmake b/cmake/generic.cmake index d2aab938d4..ff9868fc4e 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -106,22 +106,22 @@ function(merge_static_libs TARGET_NAME) endforeach() list(REMOVE_DUPLICATES libs_deps) - if(APPLE) # Use OSX's libtool to merge archives - # To produce a library we need at least one source file. - # It is created by add_custom_command below and will helps - # also help to track dependencies. - set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}_dummy.c) + # To produce a library we need at least one source file. + # It is created by add_custom_command below and will helps + # also help to track dependencies. + set(target_SRCS ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}_dummy.c) + if(APPLE) # Use OSX's libtool to merge archives # Make the generated dummy source file depended on all static input # libs. If input lib changes,the source file is touched # which causes the desired effect (relink). - add_custom_command(OUTPUT ${dummyfile} - COMMAND ${CMAKE_COMMAND} -E touch ${dummyfile} + add_custom_command(OUTPUT ${target_SRCS} + COMMAND ${CMAKE_COMMAND} -E touch ${target_SRCS} DEPENDS ${libs}) # Generate dummy staic lib - file(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";") - add_library(${TARGET_NAME} STATIC ${dummyfile}) + file(WRITE ${target_SRCS} "const char *dummy = \"${target_SRCS}\";") + add_library(${TARGET_NAME} STATIC ${target_SRCS}) target_link_libraries(${TARGET_NAME} ${libs_deps}) foreach(lib ${libs}) @@ -130,11 +130,14 @@ function(merge_static_libs TARGET_NAME) endforeach() add_custom_command(TARGET ${TARGET_NAME} POST_BUILD COMMAND rm "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" - COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles}) + COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles} + ) else() # general UNIX: use "ar" to extract objects and re-add to a common lib + set(target_DIR ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}.dir) + foreach(lib ${libs}) - set(objlistfile ${lib}.objlist) # list of objects in the input library - set(objdir ${lib}.objdir) + set(objlistfile ${target_DIR}/${lib}.objlist) # list of objects in the input library + set(objdir ${target_DIR}/${lib}.objdir) add_custom_command(OUTPUT ${objdir} COMMAND ${CMAKE_COMMAND} -E make_directory ${objdir} @@ -142,31 +145,32 @@ function(merge_static_libs TARGET_NAME) add_custom_command(OUTPUT ${objlistfile} COMMAND ${CMAKE_AR} -x "$" - COMMAND ${CMAKE_AR} -t "$" > ../${objlistfile} + COMMAND ${CMAKE_AR} -t "$" > ${objlistfile} DEPENDS ${lib} ${objdir} WORKING_DIRECTORY ${objdir}) - # Empty dummy source file that goes into merged library - set(mergebase ${lib}.mergebase.c) - add_custom_command(OUTPUT ${mergebase} - COMMAND ${CMAKE_COMMAND} -E touch ${mergebase} - DEPENDS ${objlistfile}) - - list(APPEND mergebases "${mergebase}") + list(APPEND target_OBJS "${objlistfile}") endforeach() - add_library(${TARGET_NAME} STATIC ${mergebases}) + # Make the generated dummy source file depended on all static input + # libs. If input lib changes,the source file is touched + # which causes the desired effect (relink). + add_custom_command(OUTPUT ${target_SRCS} + COMMAND ${CMAKE_COMMAND} -E touch ${target_SRCS} + DEPENDS ${libs} ${target_OBJS}) + + # Generate dummy staic lib + file(WRITE ${target_SRCS} "const char *dummy = \"${target_SRCS}\";") + add_library(${TARGET_NAME} STATIC ${target_SRCS}) target_link_libraries(${TARGET_NAME} ${libs_deps}) # Get the file name of the generated library - set(outlibfile "$") + set(target_LIBNAME "$") - foreach(lib ${libs}) - add_custom_command(TARGET ${TARGET_NAME} POST_BUILD - COMMAND ${CMAKE_AR} cr ${outlibfile} *.o - COMMAND ${CMAKE_RANLIB} ${outlibfile} - WORKING_DIRECTORY ${lib}.objdir) - endforeach() + add_custom_command(TARGET ${TARGET_NAME} POST_BUILD + COMMAND ${CMAKE_AR} crs ${target_LIBNAME} `find ${target_DIR} -name '*.o'` + COMMAND ${CMAKE_RANLIB} ${target_LIBNAME} + WORKING_DIRECTORY ${target_DIR}) endif() endfunction(merge_static_libs) @@ -196,7 +200,7 @@ function(cc_library TARGET_NAME) add_style_check_target(${TARGET_NAME} ${cc_library_SRCS} ${cc_library_HEADERS}) else(cc_library_SRCS) - if (cc_library_DEPS) + if(cc_library_DEPS) merge_static_libs(${TARGET_NAME} ${cc_library_DEPS}) else() message(FATAL "Please specify source file or library in cc_library.") @@ -249,7 +253,7 @@ function(nv_library TARGET_NAME) foreach(source_file ${nv_library_SRCS}) string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file}) if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) - list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) + list(APPEND nv_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) endif() endforeach() add_style_check_target(${TARGET_NAME} ${nv_library_SRCS} ${nv_library_HEADERS}) diff --git a/cmake/system.cmake b/cmake/system.cmake index adf5e2c539..396bd1a079 100644 --- a/cmake/system.cmake +++ b/cmake/system.cmake @@ -24,11 +24,10 @@ IF(WIN32) SET(HOST_SYSTEM "win32") ELSE(WIN32) IF(APPLE) - EXEC_PROGRAM (sw_vers ARGS -productVersion OUTPUT_VARIABLE MACOSX_VERSION) - STRING(REGEX MATCH "[0-9]+.[0-9]+" VERSION "${MACOSX_VERSION}") - SET(MACOS_VERSION ${VERSION}) SET(HOST_SYSTEM "macosx") - IF(NOT DEFINED ENV{MACOSX_DEPLOYMENT_TARGET}) + EXEC_PROGRAM(sw_vers ARGS -productVersion OUTPUT_VARIABLE HOST_SYSTEM_VERSION) + STRING(REGEX MATCH "[0-9]+.[0-9]+" MACOS_VERSION "${HOST_SYSTEM_VERSION}") + IF(NOT DEFINED $ENV{MACOSX_DEPLOYMENT_TARGET}) # Set cache variable - end user may change this during ccmake or cmake-gui configure. SET(CMAKE_OSX_DEPLOYMENT_TARGET ${MACOS_VERSION} CACHE STRING "Minimum OS X version to target for deployment (at runtime); newer APIs weak linked. Set to empty string for default value.") @@ -49,6 +48,8 @@ ELSE(WIN32) ELSEIF(LINUX_ISSUE MATCHES "Fedora") SET(HOST_SYSTEM "fedora") ENDIF() + + STRING(REGEX MATCH "(([0-9]+)\\.)+([0-9]+)" HOST_SYSTEM_VERSION "${LINUX_ISSUE}") ENDIF(EXISTS "/etc/issue") IF(EXISTS "/etc/redhat-release") @@ -70,7 +71,7 @@ CMAKE_HOST_SYSTEM_INFORMATION(RESULT CPU_CORES QUERY NUMBER_OF_LOGICAL_CORES) MARK_AS_ADVANCED(HOST_SYSTEM CPU_CORES) -MESSAGE(STATUS "Found Paddle host system: ${HOST_SYSTEM}") +MESSAGE(STATUS "Found Paddle host system: ${HOST_SYSTEM}, version: ${HOST_SYSTEM_VERSION}") MESSAGE(STATUS "Found Paddle host system's CPU: ${CPU_CORES} cores") # configuration for cross-compiling @@ -82,6 +83,9 @@ IF(DEFINED CMAKE_SYSTEM_NAME) ELSEIF(${CMAKE_SYSTEM_NAME} STREQUAL "RPi") SET(RPI TRUE) INCLUDE(cross_compiling/raspberry_pi) + ELSEIF(${CMAKE_SYSTEM_NAME} STREQUAL "iOS") + SET(IOS TRUE) + INCLUDE(cross_compiling/ios) ENDIF() ENDIF() diff --git a/cmake/util.cmake b/cmake/util.cmake index 0da4969d31..d1aee3e170 100644 --- a/cmake/util.cmake +++ b/cmake/util.cmake @@ -25,7 +25,9 @@ function(target_circle_link_libraries TARGET_NAME) endif() endforeach() if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang" OR "${CMAKE_CXX_COMPILER_ID}" STREQUAL "AppleClang") - list(APPEND LIBS "-undefined dynamic_lookup") + if(NOT IOS_ENABLE_BITCODE) + list(APPEND LIBS "-undefined dynamic_lookup") + endif() endif() list(REVERSE libsInArgn) target_link_libraries(${TARGET_NAME} @@ -95,6 +97,10 @@ function(link_paddle_exe TARGET_NAME) target_link_libraries(${TARGET_NAME} log) endif(ANDROID) + if(WITH_MKLDNN AND WITH_MKLML AND MKLDNN_IOMP_DIR) + target_link_libraries(${TARGET_NAME} "-L${MKLDNN_IOMP_DIR} -liomp5 -Wl,--as-needed") + endif() + add_dependencies(${TARGET_NAME} ${external_project_dependencies}) endfunction() diff --git a/doc/api/v1/index_cn.rst b/doc/api/v1/index_cn.rst index 3718cd73a2..cf146dc088 100644 --- a/doc/api/v1/index_cn.rst +++ b/doc/api/v1/index_cn.rst @@ -21,7 +21,7 @@ Model Config API trainer_config_helpers/optimizers.rst trainer_config_helpers/data_sources.rst trainer_config_helpers/layers.rst - trainer_config_helpers/activations.rst + trainer_config_helpers/activations.rst trainer_config_helpers/poolings.rst trainer_config_helpers/networks.rst trainer_config_helpers/evaluators.rst diff --git a/doc/api/v2/config/layer.rst b/doc/api/v2/config/layer.rst index c94627a728..d4e9d53e5c 100644 --- a/doc/api/v2/config/layer.rst +++ b/doc/api/v2/config/layer.rst @@ -345,6 +345,11 @@ clip .. autoclass:: paddle.v2.layer.clip :noindex: +resize +------ +.. autoclass:: paddle.v2.layer.resize + :noindex: + slope_intercept --------------- .. autoclass:: paddle.v2.layer.slope_intercept diff --git a/doc/design/api.md b/doc/design/api.md index 8185d2af0e..e6a4638d91 100644 --- a/doc/design/api.md +++ b/doc/design/api.md @@ -3,7 +3,7 @@ ## Ingredients As our design principle is starting from the essence: how could we -allow users to express and solve their problems at neural networks. +allow users to express and solve their problems as neural networks. Some essential concepts that our API have to provide include: 1. A *topology* is an expression of *layers*. @@ -233,7 +233,7 @@ paddle.dist_train(model, num_parameter_servers=15) ``` -The pseudo code if `paddle.dist_train` is as follows: +The pseudo code of `paddle.dist_train` is as follows: ```python def dist_train(topology, parameters, trainer, reader, ...): diff --git a/doc/design/auto_gradient_check.md b/doc/design/auto_gradient_check.md index 1f4d4ec16f..f9991541bc 100644 --- a/doc/design/auto_gradient_check.md +++ b/doc/design/auto_gradient_check.md @@ -1,17 +1,17 @@ ## Auto Gradient Checker Design ## Backgraound： -- Operator forward computing is easy to check if the result is right because it has a clear definition. **But** backpropagation is a notoriously difficult algorithm to debug and get right: - - 1. you should get the right backpropagation formula according to the forward computation. - - 2. you should implement it right in CPP. - - 3. it's difficult to prepare test data. +- Generally, it is easy to check whether the forward computation of an Operator is correct or not. However, backpropagation is a notoriously difficult algorithm to debug and get right: + 1. you should get the right backpropagation formula according to the forward computation. + 2. you should implement it right in CPP. + 3. it's difficult to prepare test data. -- Auto gradient check gets a numeric gradient by forward Operator and use it as a reference of the backward Operator's result. It has several advantages: - - 1. numeric gradient checker only need forward operator. - - 2. user only need to prepare the input data for forward Operator. +- Auto gradient checking gets a numerical gradient by forward Operator and use it as a reference of the backward Operator's result. It has several advantages: + 1. numerical gradient checker only need forward operator. + 2. user only need to prepare the input data for forward Operator. ## Mathematical Theory -The following two document from stanford has a detailed explanation of how to get numeric gradient and why it's useful. +The following two document from Stanford has a detailed explanation of how to get numerical gradient and why it's useful. - [Gradient checking and advanced optimization(en)](http://deeplearning.stanford.edu/wiki/index.php/Gradient_checking_and_advanced_optimization) - [Gradient checking and advanced optimization(cn)](http://ufldl.stanford.edu/wiki/index.php/%E6%A2%AF%E5%BA%A6%E6%A3%80%E9%AA%8C%E4%B8%8E%E9%AB%98%E7%BA%A7%E4%BC%98%E5%8C%96) @@ -20,7 +20,7 @@ The following two document from stanford has a detailed explanation of how to ge ## Numeric Gradient Implementation ### Python Interface ```python -def get_numeric_gradient(op, +def get_numerical_gradient(op, input_values, output_name, input_to_check, @@ -30,13 +30,13 @@ def get_numeric_gradient(op, Get Numeric Gradient for an operator's input. :param op: C++ operator instance, could be an network - :param input_values: The input variables. Should be an dictionary, key is - variable name. Value is numpy array. + :param input_values: The input variables. Should be an dictionary, whose key is + variable name, and value is numpy array. :param output_name: The final output variable name. - :param input_to_check: The input variable need to get gradient. + :param input_to_check: The input variable with respect to which to compute the gradient. :param delta: The perturbation value for numeric gradient method. The smaller delta is, the more accurate result will get. But if that delta is - too small, it could occur numerical stability problem. + too small, it will suffer from numerical stability problem. :param local_scope: The local scope used for get_numeric_gradient. :return: The gradient array in numpy format. """ @@ -45,28 +45,28 @@ def get_numeric_gradient(op, ### Explaination: - Why need `output_name` - - One Operator may have multiple Output, you can get independent gradient from each Output. So user should set one output to calculate. + - An Operator may have multiple Output, one can get independent gradient from each Output. So caller should specify the name of the output variable. - Why need `input_to_check` - - One operator may have multiple inputs. Gradient Op can calculate the gradient of these Inputs at the same time. But Numeric Gradient needs to calculate them one by one. So `get_numeric_gradient` is designed to calculate the gradient for one input. If you need to compute multiple inputs, you can call `get_numeric_gradient` multiple times. + - One operator may have multiple inputs. Gradient Op can calculate the gradient of these inputs at the same time. But Numeric Gradient needs to calculate them one by one. So `get_numeric_gradient` is designed to calculate the gradient for one input. If you need to compute multiple inputs, you can call `get_numeric_gradient` multiple times. ### Core Algorithm Implementation ```python - # we only compute gradient of one element each time. - # we use a for loop to compute the gradient of every element. + # we only compute gradient of one element a time. + # we use a for loop to compute the gradient of each element. for i in xrange(tensor_size): - # get one input element throw it's index i. + # get one input element by its index i. origin = tensor_to_check.get_float_element(i) - # add delta to it, run op and then get the sum of the result tensor. + # add delta to it, run op and then get the new value of the result tensor. x_pos = origin + delta tensor_to_check.set_float_element(i, x_pos) y_pos = get_output() - # plus delta to this element, run op and get the sum of the result tensor. + # plus delta to this element, run op and get the new value of the result tensor. x_neg = origin - delta tensor_to_check.set_float_element(i, x_neg) y_neg = get_output() @@ -85,15 +85,15 @@ def get_numeric_gradient(op, Each Operator Kernel has three kinds of Gradient: -- 1. Numeric Gradient -- 2. CPU Operator Gradient -- 3. GPU Operator Gradient(if supported) +1. Numerical gradient +2. CPU kernel gradient +3. GPU kernel gradient (if supported) -Numeric Gradient Only relies on forward Operator. So we use Numeric Gradient as the reference value. +The numerical gradient only relies on forward Operator. So we use the numerical gradient as the reference value. And the gradient checking is performed in the following three steps: -- 1. calculate the numeric gradient. -- 2. calculate CPU kernel Gradient with the backward Operator and compare it with the numeric gradient. -- 3. calculate GPU kernel Gradient with the backward Operator and compare it with the numeric gradient.(if support GPU) +1. calculate the numerical gradient +2. calculate CPU kernel gradient with the backward Operator and compare it with the numerical gradient +3. calculate GPU kernel gradient with the backward Operator and compare it with the numeric gradient (if supported) #### Python Interface @@ -110,8 +110,8 @@ Numeric Gradient Only relies on forward Operator. So we use Numeric Gradient as :param forward_op: used to create backward_op :param input_vars: numpy value of input variable. The following computation will use these variables. - :param inputs_to_check: inputs var names that should check gradient. - :param output_name: output name that used to + :param inputs_to_check: the input variable with respect to which to compute the gradient. + :param output_name: The final output variable name. :param max_relative_error: The relative tolerance parameter. :param no_grad_set: used when create backward ops :param only_cpu: only compute and check gradient on cpu kernel. @@ -120,24 +120,24 @@ Numeric Gradient Only relies on forward Operator. So we use Numeric Gradient as ``` ### How to check if two numpy array is close enough? -if `abs_numeric_grad` is nearly zero, then use abs error for numeric_grad, not relative +if `abs_numerical_grad` is nearly zero, then use abs error for numerical_grad ```python -numeric_grad = ... +numerical_grad = ... operator_grad = numpy.array(scope.find_var(grad_var_name(name)).get_tensor()) -abs_numeric_grad = numpy.abs(numeric_grad) -# if abs_numeric_grad is nearly zero, then use abs error for numeric_grad, not relative +abs_numerical_grad = numpy.abs(numerical_grad) +# if abs_numerical_grad is nearly zero, then use abs error for numeric_grad, not relative # error. -abs_numeric_grad[abs_numeric_grad < 1e-3] = 1 +abs_numerical_grad[abs_numerical_grad < 1e-3] = 1 -diff_mat = numpy.abs(abs_numeric_grad - operator_grad) / abs_numeric_grad +diff_mat = numpy.abs(abs_numerical_grad - operator_grad) / abs_numerical_grad max_diff = numpy.max(diff_mat) ``` #### Notes： -1，The Input data for auto gradient checker should be reasonable to avoid numeric problem. +The Input data for auto gradient checker should be reasonable to avoid numerical stability problem. #### Refs: diff --git a/doc/design/block.md b/doc/design/block.md index be88001220..4d5dd4ba95 100644 --- a/doc/design/block.md +++ b/doc/design/block.md @@ -55,17 +55,23 @@ Let us consolidate the discussion by presenting some examples. The following C++ programs shows how blocks are used with the `if-else` structure: ```c++ +namespace pd = paddle; + int x = 10; -int y = 20; -int out; +int y = 1; +int z = 10; bool cond = false; +int o1, o2; if (cond) { int z = x + y; - out = softmax(z); + o1 = z; + o2 = pd::layer::softmax(z); } else { - int z = fc(x); - out = z; + int d = pd::layer::fc(z); + o1 = d; + o2 = d+1; } + ``` An equivalent PaddlePaddle program from the design doc of the [IfElseOp operator](./if_else_op.md) is as follows: @@ -73,57 +79,55 @@ An equivalent PaddlePaddle program from the design doc of the [IfElseOp operator ```python import paddle as pd -x = var(10) -y = var(20) -cond = var(false) -ie = pd.create_ifelseop(inputs=[x], output_num=1) +x = minibatch([10, 20, 30]) # shape=[None, 1] +y = var(1) # shape=[1], value=1 +z = minibatch([10, 20, 30]) # shape=[None, 1] +cond = larger_than(x, 15) # [false, true, true] + +ie = pd.ifelse() with ie.true_block(): - x = ie.inputs(true, 0) - z = operator.add(x, y) - ie.set_output(true, 0, operator.softmax(z)) + d = pd.layer.add_scalar(x, y) + ie.output(d, pd.layer.softmax(d)) with ie.false_block(): - x = ie.inputs(false, 0) - z = layer.fc(x) - ie.set_output(true, 0, operator.softmax(z)) -out = b(cond) + d = pd.layer.fc(z) + ie.output(d, d+1) +o1, o2 = ie(cond) ``` -In both examples, the left branch computes `softmax(x+y)` and the right branch computes `fc(x)`. +In both examples, the left branch computes `x+y` and `softmax(x+y)`, the right branch computes `x+1` and `fc(x)`. A difference is that variables in the C++ program contain scalar values, whereas those in the PaddlePaddle programs are mini-batches of instances. The `ie.input(true, 0)` invocation returns instances in the 0-th input, `x`, that corresponds to true values in `cond` as the local variable `x`, where `ie.input(false, 0)` returns instances corresponding to false values. + ### Blocks with `for` and `RNNOp` The following RNN model from the [RNN design doc](./rnn.md) ```python -x = sequence([10, 20, 30]) -m = var(0) -W = tensor() -U = tensor() - -rnn = create_rnn(inputs=[input]) -with rnn.stepnet() as net: - x = net.set_inputs(0) - h = net.add_memory(init=m) - fc_out = pd.matmul(W, x) - hidden_out = pd.matmul(U, h.pre(n=1)) - sum = pd.add_two(fc_out, hidden_out) - act = pd.sigmoid(sum) - h.update(act) # update memory with act - net.set_outputs(0, act, hidden_out) # two outputs - +x = sequence([10, 20, 30]) # shape=[None, 1] +m = var(0) # shape=[1] +W = var(0.314, param=true) # shape=[1] +U = var(0.375, param=true) # shape=[1] + +rnn = pd.rnn() +with rnn.step(): + h = rnn.memory(init = m) + hh = rnn.previous_memory(h) + a = layer.fc(W, x) + b = layer.fc(U, hh) + s = pd.add(a, b) + act = pd.sigmoid(s) + rnn.update_memory(h, act) + rnn.output(a, b) o1, o2 = rnn() -print o1, o2 ``` - has its equivalent C++ program as follows ```c++ int* x = {10, 20, 30}; -int m = 0; -int W = some_value(); -int U = some_other_value(); +int* m = {0}; +int* W = {0.314}; +int* U = {0.375}; int mem[sizeof(x) / sizeof(x[0]) + 1]; int o1[sizeof(x) / sizeof(x[0]) + 1]; @@ -131,20 +135,16 @@ int o2[sizeof(x) / sizeof(x[0]) + 1]; for (int i = 1; i <= sizeof(x)/sizeof(x[0]); ++i) { int x = x[i-1]; if (i == 1) mem[0] = m; - int fc_out = W * x; - int hidden_out = Y * mem[i-1]; - int sum = fc_out + hidden_out; + int a = W * x; + int b = Y * mem[i-1]; + int s = fc_out + hidden_out; int act = sigmoid(sum); mem[i] = act; o1[i] = act; o2[i] = hidden_out; } - -print_array(o1); -print_array(o2); ``` - ## Compilation and Execution Like TensorFlow programs, a PaddlePaddle program is written in Python. The first part describes a neural network as a protobuf message, and the rest part executes the message for training or inference. @@ -210,11 +210,11 @@ a = pd.Varaible(shape=[20, 20]) b = pd.fc(a, params=["fc.w", "fc.b"]) rnn = pd.create_rnn() -with rnn.stepnet() as net: - x = net.set_inputs(a) +with rnn.stepnet() + x = a.as_step_input() # reuse fc's parameter fc_without_b = pd.get_variable("fc.w") - net.set_outputs(fc_without_b) + rnn.output(fc_without_b) out = rnn() ``` diff --git a/doc/design/functions_operators_layers.md b/doc/design/functions_operators_layers.md index d23ba56b57..984b59f4c6 100644 --- a/doc/design/functions_operators_layers.md +++ b/doc/design/functions_operators_layers.md @@ -53,12 +53,12 @@ Let's explain using an example. Suppose that we are going to compose the FC usi ```python def operator.mul(X1, X2): O = Var() - paddle.cpp.create_operator("mul", input={X1, Y1], output=O) + paddle.cpp.create_operator("mul", input={X1, Y1}, output=O) return O def operator.add(X1, X2): O = Var() - paddle.cpp.create_operator("add", input={X1, X2], output=O) + paddle.cpp.create_operator("add", input={X1, X2}, output=O) return O ``` diff --git a/doc/design/graph.md b/doc/design/graph.md index 51b7f87638..7519a65df8 100644 --- a/doc/design/graph.md +++ b/doc/design/graph.md @@ -56,7 +56,7 @@ For each parameter, like W and b created by `layer.fc`, marked as double circles ## Block and Graph -The word block and graph are interchangable in the desgin of PaddlePaddle. A [Block[(https://github.com/PaddlePaddle/Paddle/pull/3708) is a metaphore of the code and local variables in a pair of curly braces in programming languages, where operators are like statements or instructions. A graph of operators and variables is a representation of the block. +The word block and graph are interchangable in the desgin of PaddlePaddle. A [Block](https://github.com/PaddlePaddle/Paddle/pull/3708) is a metaphore of the code and local variables in a pair of curly braces in programming languages, where operators are like statements or instructions. A graph of operators and variables is a representation of the block. A Block keeps operators in an array `BlockDesc::ops` @@ -67,4 +67,4 @@ message BlockDesc { } ``` -in the order that there appear in user programs, like the Python program at the beginning of this article. We can imagine that in `ops`, we have some forward operators, followed by some gradient operators, and then some optimization operators. +in the order that they appear in user programs, like the Python program at the beginning of this article. We can imagine that in `ops`, we have some forward operators, followed by some gradient operators, and then some optimization operators. diff --git a/doc/design/if_else_op.md b/doc/design/if_else_op.md index 954a19c073..26d140f06d 100644 --- a/doc/design/if_else_op.md +++ b/doc/design/if_else_op.md @@ -1,41 +1,51 @@ -IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has N instances. If cond[i] == True, input instance input[i] will go through true_block() and generate output[i]; otherwise it will produce output from false_bloack(). +# The `IfElse` Operator -```python -import paddle as pd +PaddlePaddle's `IfElse` operator differs from TensorFlow's: -x = var() -y = var() -cond = var() -default_value = var() -b = pd.create_ifelseop(inputs=[x], output_num=1) -with b.true_block(): - x = b.inputs(0) - z = operator.add(x, y) - b.set_output(0, operator.softmax(z)) - -with b.false_block(): - x = b.inputs(0) - z = layer.fc(x) - b.set_output(0, operator.softmax(z)) - -out = b(cond) -``` +- the TensorFlow version takes a scalar boolean value as the condition so that the whole mini-batch goes to either the true or the false branch, whereas +- the PaddlePaddle version takes a vector of boolean value as the condition, and instances corresponding to true values go to the true branch, those corresponding to false values go to the false branch. + +## Example + +The following PaddlePaddle program shows the usage of the IfElse operator: -If only true_block is set in an IfElseOp, a special case is that we can have a default value for false as: ```python import paddle as pd -x = var() -y = var() -cond = var() -default_value = var() -b = pd.create_ifelseop(inputs=[x], output_num=1, default_value) - -with b.true_block(): - x = b.inputs(0) - z = operator.add(x, y) - b.set_output(0, operator.softmax(z)) +x = minibatch([10, 20, 30]) # shape=[None, 1] +y = var(1) # shape=[1], value=1 +z = minibatch([10, 20, 30]) # shape=[None, 1] +cond = larger_than(x, 15) # [false, true, true] + +ie = pd.ifelse() +with ie.true_block(): + d = pd.layer.add(x, y) + ie.output(d, pd.layer.softmax(d)) +with ie.false_block(): + d = pd.layer.fc(z) + ie.output(d, d+1) +o1, o2 = ie(cond) +``` -out = b(cond) +A challenge to implement the `IfElse` operator is to infer those variables to be split, or, say, to identify the variable of the mini-batch or those derived from the mini-batch. + +An equivalent C++ program is as follows: + +```c++ +namespace pd = paddle; + +int x = 10; +int y = 1; +int z = 10; +bool cond = false; +int o1, o2; +if (cond) { + int d = x + y; + o1 = z; + o2 = pd::layer::softmax(z); +} else { + int d = pd::layer::fc(z); + o1 = d; + o2 = d+1; +} ``` -where default_value is a list of vars for `cond` == False. diff --git a/doc/design/parameters_in_cpp.md b/doc/design/parameters_in_cpp.md index b6f99bc7d9..a7ac3f17c4 100644 --- a/doc/design/parameters_in_cpp.md +++ b/doc/design/parameters_in_cpp.md @@ -1,19 +1,19 @@ # Design Doc: The C++ Class `Parameters` -`Parameters` is a concept we designed in Paddle V2 API. `Parameters` is a container of parameters, and make Paddle can shared parameter between topologies. We described usages of `Parameter` in [api.md](./api.md). +`Parameters` is a concept we designed in PaddlePaddle V2 API. `Parameters` is a container of parameters, which makes PaddlePaddle capable of sharing parameter between topologies. We described usages of `Parameter` in [api.md](./api.md). -We used Python to implement Parameters when designing V2 API before. There are several defects for current implementation: +We used Python to implement Parameters when designing V2 API before. There are several defects for the current implementation: * We just use `memcpy` to share Parameters between topologies, but this is very inefficient. -* We did not implement share Parameters while training. We just trigger `memcpy` when start training. +* We did not support sharing Parameters while training. We just trigger `memcpy` when start training. -It is necessary that we implement Parameters in CPP side. However, it could be a code refactoring for Paddle, because Paddle was designed for training only one topology before, i.e., each GradientMachine contains its Parameter as a data member. In current Paddle implementation, there are three concepts associated with `Parameters`: +It is necessary that we implement Parameters in CPP side. However, it could result a code refactoring for PaddlePaddle, because PaddlePaddle was designed for training only one topology before, i.e., each GradientMachine contains its Parameter as a data member. In current PaddlePaddle implementation, there are three concepts associated with `Parameters`: 1. `paddle::Parameter`. A `Parameters` is a container for `paddle::Parameter`. It is evident that we should use `paddle::Parameter` when developing `Parameters`. However, the `Parameter` class contains many functions and does not have a clear interface. It contains `create/store Parameter`, `serialize/deserialize`, `optimize(i.e SGD)`, `randomize/zero`. When we developing `Parameters`, we only use `create/store Parameter` functionality. -We should extract functionalities of Parameter into many classes to clean Paddle CPP implementation. +We should extract functionalities of Parameter into many classes to clean PaddlePaddle CPP implementation. 2. `paddle::GradientMachine` and its sub-classes, e.g., `paddle::MultiGradientMachine`, `paddle::NeuralNetwork`. We should pass `Parameters` to `paddle::GradientMachine` when `forward/backward` to avoid `memcpy` between topologies. @@ -24,7 +24,7 @@ Also, we should handle multi-GPU/CPU training, because `forward` and `backward` So `Parameters` should be used by `paddle::ParameterUpdater`, and `paddle::ParameterUpdater` should optimize `Parameters` (by SGD). -The step by step approach for implementation Parameters in Paddle C++ core is listed below. Each step should be a PR and could be merged into Paddle one by one. +The step by step approach for implementation Parameters in PaddlePaddle C++ core is listed below. Each step should be a PR and could be merged into PaddlePaddle one by one. 1. Clean `paddle::Parameter` interface. Extract the functionalities of `paddle::Parameter` to prepare for the implementation of Parameters. diff --git a/doc/design/program.md b/doc/design/program.md new file mode 100644 index 0000000000..bd2456787c --- /dev/null +++ b/doc/design/program.md @@ -0,0 +1,139 @@ +# Design Doc: PaddlePaddle Programs + +## Compile and Execution + +A PaddlePaddle program consists of two parts -- the first generates a `ProgramDesc` protobuf message that describes the program, and the second runs this message using a C++ class `Executor`. + +A simple example PaddlePaddle program can be found in [graph.md](./graph.md): + +```python +x = layer.data("images") +l = layer.data("label") +y = layer.fc(x) +cost = layer.mse(y, l) +optimize(cost) +train(cost, reader=mnist.train()) +``` + +The first five lines of the following PaddlePaddle program generates, or, compiles, the `ProgramDesc` message. The last line runs it. + +## Programs and Blocks + +The basic structure of a PaddlePaddle program is some nested blocks, as a C++ or Java program. + +- program: some nested blocks +- [block](./block.md): + - some local variable definitions, and + - a sequence of operators + +The concept of block comes from usual programs. For example, the following C++ program has three blocks: + +```c++ +int main() { // block 0 + int i = 0; + if (i < 10) { // block 1 + for (int j = 0; j < 10; j++) { // block 2 + } + } + return 0; +} +``` + +The following PaddlePaddle program has three blocks: + +```python +import paddle as pd // block 0 + +x = minibatch([10, 20, 30]) # shape=[None, 1] +y = var(1) # shape=[1], value=1 +z = minibatch([10, 20, 30]) # shape=[None, 1] +cond = larger_than(x, 15) # [false, true, true] + +ie = pd.ifelse() +with ie.true_block(): // block 1 + d = pd.layer.add_scalar(x, y) + ie.output(d, pd.layer.softmax(d)) +with ie.false_block(): // block 2 + d = pd.layer.fc(z) + ie.output(d, d+1) +o1, o2 = ie(cond) +``` + +## `BlockDesc` and `ProgramDesc` + +All protobuf messages are defined in `framework.proto`. + +`BlockDesc` is straight-forward -- it includes local variable definitions, `vars`, and a sequence of operators, `ops`. + +```protobuf +message BlockDesc { + required int32 parent = 1; + repeated VarDesc vars = 2; + repeated OpDesc ops = 3; +} +``` + +The parent ID indicates the parent block so that operators in a block can refer to variables defined locally and also those defined in their ancestor blocks. + +All hierarchical blocks in a program are flattened and stored in an array. The block ID is the index of the block in this array. + +```protobuf +message ProgramDesc { + repeated BlockDesc blocks = 1; +} +``` + + +### Global Block + +The global block is the first one in the above array. + +## Operators that Use Blocks + +In the above example, the operator `IfElseOp` has two blocks -- the true branch and the false branch. + +The definition of `OpDesc` shows that an operator could have some attributes: + +```protobuf +message OpDesc { + AttrDesc attrs = 1; + ... +} +``` + +and an attribute could be of type block, which is, in fact, a block ID as described above: + +``` +message AttrDesc { + required string name = 1; + + enum AttrType { + INT = 1, + STRING = 2, + ... + BLOCK = ... + } + required AttrType type = 2; + + optional int32 block = 10; // when type == BLOCK + ... +} +``` + +## InferShape + +With this design, the InferShape function should take the following parameters: + +```c++ +void InferShape(int current_block, + int current_operator, + ProgramDesc* program // might change VarDesc values. + ) { + ... +} +``` + +where + +- `current_block` indices into `ProgramDesc::blocks`, +- `current_operator` indices into `BlockDesc::ops`. diff --git a/doc/design/python_api.md b/doc/design/python_api.md new file mode 100644 index 0000000000..6213da65c8 --- /dev/null +++ b/doc/design/python_api.md @@ -0,0 +1,216 @@ +# Design Doc: Python API + +Due to the refactorization of the PaddlePaddle core, we need Python classes to construct corresponding protobuf messages that describe a DL program. + +| Python classes | Protobuf messages | +| --- | --- | +| Program | ProgramDesc | +| Block | BlockDesc | +| Operator | OpDesc | +| Variable | VarDesc | + +Please be aware that these Python classes need to maintain some construction-time information, which are not part of the protobuf messages. + +## Core Concepts + +### Program + +A `ProgramDesc` describes a [DL program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md), which is composed of an array of `BlockDesc`s. The `BlockDesc`s in a `ProgramDesc` can have a tree-like hierarchical structure. However, the `ProgramDesc` onlys stores a flattened array of `BlockDesc`s. A `BlockDesc` refers to its parent block by its index in the array. For example, operators in the step block of an RNN operator need to be able to access variables in its ancestor blocks. + +Whenever we create a block, we need to set its parent block to the current block, hence the Python class `Program` needs to maintain a data member `current_block`. + +```python +class Program(objects): + def __init__(self): + self.proto = core.NewProgram() # a C++ ProgramDesc pointer. + self.blocks = vector() + self.blocks.append(Block(self, -1)) # the global block + self.current_block = 0 # initialized to the global block + + def global_block(): + return self.blocks[0] + + def current_block(): + return self.get_block(self.current_block) + + def rollback(): + self.current_block = self.current_block().parent_idx + + def create_block(): + new_block_idx = len(self.block) + self.blocks.append(Block(self, self.current_block)) + self.current_block = new_block_idx + return current_block() +``` + +`Program` is an accessor to the protobuf message `ProgramDesc`, which is created in C++ space, because the InferShape function is in C++, which manipulates `VarDesc` messages, which are in turn members of `BlockDesc`, which is a member of `ProgramDesc`. + +`Program` creates the first block as the global block in its constructor. All parameters and their initializer operators are in the global block. + +### Block + +A [Block](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md) includes + +1. a map from variable names to an instance of the Python `Variable` class, and +1. a list of `Operator` instances. + +```python +class Block(objects): + def __init__(self, program, parent_idx): + self.proto = core.NewBlock(program.proto) + self.program = program + self.vars = map() + self.ops = vector() + self.parent_idx = parent_idx + + def create_var(self, ...): + return Variable(self, ...) + + def _create_global_var(self, ...): + program.global_block().create_var(...) + + def create_parameter(self, name, ...): + # Parameter is a subclass of variable. See Parameter section for details. + self.vars[name] = Parameter(self._create_global_var(...), ...) + return self.vars[name] + + def append_operator(self, ...): + self.ops.append(Operator(self, ...)) + + def prepend_operator(self, ...): # Parameter's ctor prepands initialize operators. + self.ops.prepend(Operator(self, ...)) +``` + +`create_parameter` is necessary because parameters are global variables, defined in the global block, but can be created in some sub-blocks. For example, an FC layer in the step block of an RNN operator. + +`prepend_operator` is necessary because the constructor of `Parameter` needs to create the initialize (or load) operator of the parameter, and would like to put it in the *preamble* of the global block. + +### Operator + +The `Operator` class fills in the `OpDesc` message and calls the C++ function `InferShape` to infer the output shapes from the input shapes. + +```python +class Operator(object): + def __init__(self, + block, # Block + type, # string + inputs, # dict + outputs,# dict + attrs # dict + ): + self.proto = core.NewOpDesc(block.proto, type, inputs, outputs, attrs) + core.infer_shape(self.proto, inputs, outputs) + + def type(self): + return self.proto.type() +``` + +`Operator` creates the `OpDesc` message in C++ space, so that it can call the `InferShape` function, which is in C++. + +### Variable + +Operators take Variables as its inputs and outputs. + +```python +class Variable(object): + def __init__(self, + block=None, # Block + name=None, # string + shape, # tuple + dtype="float32", # string + lod_level=None # int + ): + if name is None: + name = unique_name_generator() + self.name = name + self.block = block + self.proto = core.NewVarDesc(block.proto, name, shape, lod_level) + self.writer = None +``` + +Please be aware of `self.writer`, that tracks operator who creates the variable. It possible that there are more than one operators who write a variable, but in Python space, each write to a variable is represented by a Variable class. This is guaranteed by the fact that **`core.NewVarDesc` must NOT create a new `VarDesc` message if its name already exists in the specified block**. + +### Parameter + +A parameter is a global variable with an initializer (or load) operator. + +```python +class Parameter(Variable): + def __init__(self, + block=None, # Block + name=None, # string + shape, # tuple + dtype="float32", # string + lod_level=None # int + trainable, # bool + initialize_op_attrs, + optimize_op_attrs): + super(Parameter, self).__init__(block, name, shape, dtype, lod_level) + self.trainable = trainable + self.optimize_op_attrs = optimize_op_attrs + block.prepend(Operator(block, # Block + initialize_op_attrs['type'], # string + None, # no inputs + self, # output is the parameter + initialize_op_attrs) +``` + +When users create a parameter, they can call + +```python +program.create_parameter( + ..., + init_attr={ + type: "uniform_random", + min: -1.0, + max: 1.0, + }) +) +``` + +In above example, `init_attr.type` names an initialize operator. It can also name the load operator + +```python +init_attr={ + type: "load", + filename: "something.numpy", +} +``` + +`optimize_op_attrs` is not in the `VarDesc` message, but kept in the Python instance, as it will be used in the Python space when creating the optimize operator's `OpDesc`, and will be in the `OpDesc` message. + +## Layer Functions + +A layer is a Python function that creates some operators and variables. Layers simplify the work of application programmers. + +### Data Layer + +```python +def data_layer(name, type, column_name): + block = the_current_program.glolal_block() + var = block.create_global_var( + name=name, + shape=[None] + type.dims(), + dtype=type.dtype) + block.prepend_operator(block, + type="Feed", + inputs = None, + outputs = [var], + {column_name: column_name}) + return var +``` + +The input to the feed operator is a special variable in the global scope, which is the output of [Python readers](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/reader/README.md). + +### FC Layer + +```python +def fc_layer(input, size, ...): + block = program.current_block() + w = block.create_parameter(...) + b = block.create_parameter(...) + out = block.create_var() + op = block.append_operator("FC", X=input, W=w, b=b, out=out) + out.writer = op + return out +``` diff --git a/doc/design/reader/README.md b/doc/design/reader/README.md index f21f7af520..320dccec3d 100644 --- a/doc/design/reader/README.md +++ b/doc/design/reader/README.md @@ -52,7 +52,7 @@ Here are valid outputs: # a mini batch of three data items, each data item is a list (single column). [([1,1,1],), ([2,2,2],), -([3,3,3],), +([3,3,3],)] ``` Please note that each item inside the list must be a tuple, below is an invalid output: diff --git a/doc/design/refactor/distributed_architecture.md b/doc/design/refactor/distributed_architecture.md new file mode 100644 index 0000000000..ac7e98ccf1 --- /dev/null +++ b/doc/design/refactor/distributed_architecture.md @@ -0,0 +1,222 @@ +# Design Doc: Distributed Training Architecture + +## Abstract + +PaddlePaddle v0.10.0 uses the "trainer-parameter server" +architecture. We run multiple replicated instances of trainers (runs +the same code written by the user) and parameter servers for +distributed training. This architecture served us well, but has some +limitations: + +1. Need to write special code to handle tasks which should only be run + by a single trainer. E.g., initializing model and saving model. + +2. Model parallelism is hard: need to write if-else branches conditioned + on the trainer ID to partition model onto each trainer, and manually + write the inter-model-shard communication code. + +3. The user can not directly specify the parameter update rule: need + to modify the parameter server C++ code and compile a new + binary. This adds complication for researchers: A lot of extra + effort is required. Besides, the training job submission program + may not allow running arbitrary binaries. + +This design doc discusses PaddlePaddle's new distributed training +architecture that addresses the above limitations. + +## Analysis + +We will assume the user writes the trainer program by Python, the same +analysis holds if the trainer program is written in C++. + +### Limitation 1 + +If we look at the Python code that the user writes, there are two +kinds of functionalities: + +- The training logic such as load / save model and print log. +- The neural network definition such as the definition of the data + layer, the fully connected layer, the cost function and the + optimizer. + +When we training with PaddlePaddle v0.10.0 distributedly, multiple +replicated Python instances are running on different nodes: both the +training logic and the neural network computation is replicated. + +The tasks that should only run once all belong to the training logic, +if we only replicate the neural network computation, but do **not** +replicate the training logic, the limitation could be solved. + +### Limitation 2 + +Model parallelism means running a single model on multiple nodes by +partitioning the model onto different nodes and managing the +inter-model-shard communications. + +PaddlePaddle should be able to modify the nerual network computation +definition to support model parallelism automatically. However, the +computation is only specified in Python code, and PaddlePaddle can not +modify Python code. + +Just like compiler uses a intermediate representation (IR) so that +programmer does not need to manually optimize their code in most of +the cases - the compiler will optimize the IR: + + + +We can have our own IR too: PaddlePaddle can support model parallel by +converting the IR so the user no longer need to manually do it in +Python: + + + +The IR for PaddlePaddle after refactor is called `Block`, it specifies +the computation dependency graph and the variables used in the +computation. + +### Limitation 3 + +The user can not directly specify the parameter update rule for the +parameter server because the parameter server does not use the same +computation definition as the trainer. Instead, the update rule is +baked in the parameter server. The user can not specify the update +rule in the same way of specifying the trainer computation. + +This could be fixed by making the parameter server run the same +computation definition as the trainer. For a detailed explanation, +please +see +[Design Doc: Operation Graph Based Parameter Server](./dist_train.md) + +## Distributed Training Architecture + +The new distributed training architecture can address the above +limitations. Below is the illustration: + + + +The architecture includes major components: *PaddlePaddle Python*, +*PaddlePaddle converter* and *PaddlePaddle runtime*: + +### PaddlePaddle Python + +PaddlePaddle Python is the Python library that user's Python trainer +invoke to build the neural network topology, start training, etc. + +```Python +paddle.init() +input = paddle.op.recordIO("/home/data/mnist.recordio") # file stored on the cluster +img, label = input[0], input[1] +hidden = paddle.layer.fc(input=img, size=200, act=paddle.activation.Tanh()) +prediction = paddle.layer.fc(input=img, size=10, act=paddle.activation.Softmax()) +cost = paddle.layer.classification_cost(input=prediction, label=label) +optimizer = paddle.optimizer.SGD(cost, learning_rate=0.01) +session = paddle.session.NewRemote(num_trainer=3, num_ps=2, GPU_per_trainer=1) +for i in range(1000): + _, cost_val = session.eval(targets=[cost, optimizer]) + print cost_val +``` + +The code above is a typical Python trainer code, the neural network +topology is built using helper functions such as +`paddle.layer.fc`. The training is done by calling `session.eval` +iteratively. + +#### session.eval + +As shown in the graph, `session.eval` sends the IR and the evaluation +inputs/targets to the PaddlePaddle cluster for evaluation. The +targets can be any variable in the computation graph. When the target +is the `optimizer` variable, the neural network will be optimized +once. When the target is the `cost` variable, `session.eval` returns +the cost value. + +The Python `session` is a wrapper of the C++ `Session` class. For more +information about `Session`, please +see [Design Doc: Session](./session.md). + +### PaddlePaddle Converter + +PaddlePaddle converter automatically converts the IR in the request +(IR and evaluation inputs/targets) from PaddlePaddle Python to new +partitioned IRs and dispatch the new IRs and evaluation inputs/targets +to different PaddlePaddle runtimes. Below are the steps: + +1. Add `feed` OP that feeds the eval inputs, and `fetch` OP that + fetches the eval targets to the IR. + +1. Extract a new computation (sub)graph with `feed` and `fetch` OP as + the boundary. The runtime does not need to run the OP that is not + dependent by the `fetch` OP. + +1. Optimizes the computation graph. + +1. Place the OPs in the graph onto different devices on different + PaddlePaddle runtime according to a placement algorithm and device + constraint specified by the user. + +1. Partition the graph according to runtime boundaries and add `send` / + `recv` OP pair on the runtime boundaries. + +1. Dispatch the partitioned graph to different PaddlePaddle runtimes. + +1. PaddlePaddle runtimes with the `fetch` OP reports evaluation + results back to the converter, the convert reports the evaluation + results back to the PaddlePaddle Python. + +The output IRs will be cached to optimize the conversion latency. + + +#### Placement Algorithm + +Our first implementation will only support "trainer-parameter server" +placement: the parameters, initializers, and optimizers are placed on +the PaddlePaddle runtimes with the parameter server role. And +everything else will be placed on the PaddlePaddle runtimes with the +trainer role. This has the same functionality of our +"trainer-parameter server" architecture of PaddlePaddle v0.10.0, but +is more general and flexible. + +In the future, we will implement the general placement algorithm, +which makes placements according to the input IR, and a model of +device computation time and device communication time. Model +parallelism requires the general placement algorithm. + + +### PaddlePaddle Runtime + +The PaddlePaddle runtime owns multiple devices (e.g., CPUs, GPUs) and +runs the IR. The runtime does not need to do OP placement since it's +already done by the converter. + + +### Local Training Architecture + +The local training architecture will be the same as the distributed +training architecture, the differences are everything runs locally, +and there is just one PaddlePaddle runtime: + + + + +### Training Data + +In PaddlePaddle v0.10.0, training data is typically read +with [data reader](../reader/README.md) from Python. This approach is +no longer efficient when training distributedly since the Python +process no longer runs on the same node with the trainer processes, +the Python reader will need to read from the distributed filesystem +(assuming it has the access) and send to the trainers, doubling the +network traffic. + +When doing distributed training, the user can still use Python data +reader: the training data are sent with `session.eval`. However should +be used for debugging purpose only. The users are encouraged to use +the read data OPs. + + +## References: + +[1] [TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems](https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45166.pdf) + +[2] [TensorFlow: A System for Large-Scale Machine Learning](https://www.usenix.org/system/files/conference/osdi16/osdi16-abadi.pdf) diff --git a/doc/design/ops/dist_train.md b/doc/design/refactor/parameter_server.md similarity index 100% rename from doc/design/ops/dist_train.md rename to doc/design/refactor/parameter_server.md diff --git a/doc/design/refactor/session.md b/doc/design/refactor/session.md new file mode 100644 index 0000000000..1d9a26683c --- /dev/null +++ b/doc/design/refactor/session.md @@ -0,0 +1,180 @@ +# Design Doc: Session + +## Abstract + +The *session* object encapsulates the environment in which the +computation graph is executed. + +We will have the *local* session and *remote* session, they offer the +same [interface](#interface). The local session encapsulates the local +runtime environment and the remote session encapsulates the cluster +runtime environment. + +The local runtime environment contains: + +1. computation devices (i.e., CPU, GPU) handles, and +1. the [scope](../scope.md) which holds all variables. + +The remote runtime environment contains: + +1. computation devices (i.e., CPU and GPU on node 0, 1) in a cluster, + and +1. the distributed [scope](../scope.md) in a cluster which holds all + variables. + +The user can create a remote session on Paddle Cloud and evaluate the +computation graph with it. In this way, the user can control the +remote computation resource in a cluster from his local computer. + + +## Background + +The current design has an implicit global session in which +`paddle.eval()` is executed. The pain point is: + +Since the user is not able to explicitly switch between runtime +environments, the user cannot run a topology in two independent +environments. + +For example, in reinforcement learning, the user may want to have a +stale model for inference and a fresh model for training, and only +replace the stale model with the fresh model periodically. + +Furthermore, we have no concept that encapsulates a remote environment +that executes a computation graph. + +We need the session object to address above issues. + + +## Session + +A session is an object that owns the runtime environment. All +computations are executed through `session.eval()`. + + +### Interface + +```python +eval( + targets, + feed_dict=None, +) +``` + +Evaluates the target Operations or Variables in `targets`. + +- *targets*: the evaluation targets. Can be a single Operation or + Variable, or a list with the Operations or Variables as + elements. The value returned by `eval()` has the same shape as the + `target` argument. + + The PaddlePaddle program is represented by + the [ProgramDesc](../design/program.md), `eval()` will infer the + ProgramDesc from the given targets and run the PaddlePaddle + program. Please + see + [this graph](./distributed_architecture.md#local-training-architecture) for + the detailed illustration for the local session + and + [this graph](./distributed_architecture.md#distributed-training-architecture) for + the detailed illustration for the remote session. + +- *feed_dict*: a dictionary that contains the tensors which override + the edges of the computation graph. + + feed_dict not only can provide the input data, it can override any + OP's input as well: + + ```python + a = pd.constant(2.0, name="a") + b = pd.variable(name="b") + c = pd.mul(a,b) + sess.eval(targets=c, feed_dict={"b":3.0}) # returns 6.0 + ``` + +```python +close() +``` + +Closes the session and releases the scope that the session owns. + + +### Create a Local Session + +```python +session( + devices=None +) +``` + +Creates a new session. One session owns one global scope, so creating +multiple sessions will create different scopes. + +- *devices*: a single `string` or a list of `string` of device names, + the corresponding devices will be the computation devices for + `eval()`. If not specified, all available devices (e.g., all GPUs) + will be used. The user doesn't need to specify the CPU device since + it will be always used. Multiple sessions can use the same device. + + +#### Example + +```Python +a = paddle.constant(1.0) +b = paddle.constant(2.0) +c = a + b +sess = paddle.session(devices=["gpu:0", "gpu:1", "fpga:0"]) +sess.eval(c) +sess.close() +``` + +### Create a Remote Session + +```python +create_cloud_job( + name, + num_trainer, + mem_per_trainer, + gpu_per_trainer, + cpu_per_trainer, + num_ps, + mem_per_ps, + cpu_per_ps, +) +``` + +Creates a Paddle Cloud job. Fails if the job name exists. + +```python +get_cloud_job( + name +) +``` + +Gets a Paddle Cloud job. + +```python +remote_session( + job +) +``` + +- *job*: the Paddle Cloud job. + +#### Example + +```Python +reader = paddle.reader.recordio("/pfs/home/peter/mnist-train-*") # data stored on Paddle Cloud +image = reader.column(0) +label = reader.column(1) +fc1 = paddle.op.fc(image, size=256, act="sigmoid") +fc2 = paddle.op.fc(fc1, size=10, act="softmax") +cost = paddle.op.cross_entropy(fc2, label) +opt = paddle.optimizer.sgd(cost) + +job = paddle.create_cloud_job("test", 3, "1G", 1, 1, 2, "1G", 1) +sess = paddle.remote_ession(job) +for i in range(1000): + sess.eval(opt) +sess.close() +``` diff --git a/doc/design/refactor/src/compiler.graffle b/doc/design/refactor/src/compiler.graffle new file mode 100644 index 0000000000..8cc678fea3 Binary files /dev/null and b/doc/design/refactor/src/compiler.graffle differ diff --git a/doc/design/refactor/src/compiler.png b/doc/design/refactor/src/compiler.png new file mode 100644 index 0000000000..65d34f841a Binary files /dev/null and b/doc/design/refactor/src/compiler.png differ diff --git a/doc/design/ops/src/dist-graph.graffle b/doc/design/refactor/src/dist-graph.graffle similarity index 100% rename from doc/design/ops/src/dist-graph.graffle rename to doc/design/refactor/src/dist-graph.graffle diff --git a/doc/design/ops/src/dist-graph.png b/doc/design/refactor/src/dist-graph.png similarity index 100% rename from doc/design/ops/src/dist-graph.png rename to doc/design/refactor/src/dist-graph.png diff --git a/doc/design/refactor/src/distributed_architecture.graffle b/doc/design/refactor/src/distributed_architecture.graffle new file mode 100644 index 0000000000..f8496e5732 Binary files /dev/null and b/doc/design/refactor/src/distributed_architecture.graffle differ diff --git a/doc/design/refactor/src/distributed_architecture.png b/doc/design/refactor/src/distributed_architecture.png new file mode 100644 index 0000000000..410c4510c6 Binary files /dev/null and b/doc/design/refactor/src/distributed_architecture.png differ diff --git a/doc/design/ops/src/local-graph.graffle b/doc/design/refactor/src/local-graph.graffle similarity index 100% rename from doc/design/ops/src/local-graph.graffle rename to doc/design/refactor/src/local-graph.graffle diff --git a/doc/design/ops/src/local-graph.png b/doc/design/refactor/src/local-graph.png similarity index 100% rename from doc/design/ops/src/local-graph.png rename to doc/design/refactor/src/local-graph.png diff --git a/doc/design/refactor/src/local_architecture.graffle b/doc/design/refactor/src/local_architecture.graffle new file mode 100644 index 0000000000..cc7783c453 Binary files /dev/null and b/doc/design/refactor/src/local_architecture.graffle differ diff --git a/doc/design/refactor/src/local_architecture.png b/doc/design/refactor/src/local_architecture.png new file mode 100644 index 0000000000..4b999538b7 Binary files /dev/null and b/doc/design/refactor/src/local_architecture.png differ diff --git a/doc/design/refactor/src/paddle-compile.graffle b/doc/design/refactor/src/paddle-compile.graffle new file mode 100644 index 0000000000..a6348cc3db Binary files /dev/null and b/doc/design/refactor/src/paddle-compile.graffle differ diff --git a/doc/design/refactor/src/paddle-compile.png b/doc/design/refactor/src/paddle-compile.png new file mode 100644 index 0000000000..e0f13d551a Binary files /dev/null and b/doc/design/refactor/src/paddle-compile.png differ diff --git a/doc/design/refactorization.md b/doc/design/refactorization.md index e105861e92..629422e774 100644 --- a/doc/design/refactorization.md +++ b/doc/design/refactorization.md @@ -1,40 +1,40 @@ # Design Doc: Refactorization Overview -The goal of refactorizaiton include: +The goals of refactoring include: -1. Make it easy for external contributors to write new elementory computaiton operations. -1. Make the codebase clean and readable. -1. Introduce a new design of computation representation -- a computation graph of operators and variables. -1. The graph representation helps implementing auto-scalable and auto fault recoverable distributed computing. +1. Making it easy for external contributors to write new elementary computation operations. +1. Making the codebase clean and readable. +1. Designing a new computation representation -- a computation graph of operators and variables. +1. Implementing auto-scalability and auto fault recoverable distributed computing with the help of computation graphs. ## Computation Graphs -1. PaddlePaddle represent the computation, training and inference of DL models, by computation graphs. +1. PaddlePaddle represents the computation, training and inference of Deep Learning models, by computation graphs. - 1. Please dig into [computation graphs](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/graph.md) for a solid example. + 1. Please refer to [computation graphs](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/graph.md) for a concrete example. -1. Users write Python programs to describe the graphs and run it (locally or remotely). +1. Users write Python programs to describe the graphs and run them (locally or remotely). -1. A graph is composed of *variabels* and *operators*. +1. A graph is composed of *variables* and *operators*. -1. The description of graphs must be able to be serialized/deserialized, so it +1. The description of graphs must be capable of being serialized/deserialized, so that: - 1. could to be sent to the cloud for distributed execution, and - 1. be sent to clients for mobile or enterprise deployment. + 1. It can to be sent to the cloud for distributed execution, and + 1. It can be sent to clients for mobile or enterprise deployment. -1. The Python program do +1. The Python program does the following steps - 1. *compilation*: runs a Python program to generate a protobuf message representation of the graph and send it to + 1. *compilation*: run a Python program to generate a protobuf message representation of the graph and send it to 1. the C++ library `libpaddle.so` for local execution, 1. the master process of a distributed training job for training, or 1. the server process of a Kubernetes serving job for distributed serving. - 1. *execution*: according to the protobuf message, constructs instances of class `Variable` and `OperatorBase`, and run them. + 1. *execution*: execute the graph by constructing instances of class [`Variable`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h#L24) and [`OperatorBase`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L70), according to the protobuf message. -## Description and Realization +## Description and Realization of Computation Graph -At compile time, the Python program generates protobuf message representation of the graph, or the description of the graph. +At compile time, the Python program generates a protobuf message representation of the graph, or the description of the graph. -At runtime, the C++ program realizes the graph and run it. +At runtime, the C++ program realizes the graph and runs it. | | Representation (protobuf messages) | Realization (C++ class objects) | |---|---|---| @@ -42,30 +42,31 @@ At runtime, the C++ program realizes the graph and run it. |Operation|[OpDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L35)|[Operator](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L64)| |Block|BlockDesc|Block| -The word *graph* is exchangable with *block* in this document. A graph represent computation steps and local variables as a C++/Java program block, or a pair of { and }. +The word *graph* is interchangeable with *block* in this document. A graph represents computation steps and local variables similar to a C++/Java program block, or a pair of parentheses(`{` and `}`). ## Compilation and Execution -1. Run an applicaton Python program to describe the graph. In particular, +1. Run an application Python program to describe the graph. In particular, the Python application program does the following: - 1. create VarDesc to represent local/intermediate variables, - 1. create operators and set attributes, - 1. validate attribute values, - 1. inference the type and the shape of variables, - 1. plan for memory-reuse for variables, - 1. generate backward and optimization part of the Graph. - 1. possiblly split the graph for distributed training. + 1. Create `VarDesc` to represent local/intermediate variables, + 1. Create operators and set attributes, + 1. Validate attribute values, + 1. Infer the type and the shape of variables, + 1. Plan memory-reuse for variables, + 1. Generate the backward graph + 1. Optimize the computation graph. + 1. Potentially, split the graph for distributed training. -1. The invocation of `train` or `infer` in the application Python program: +1. The invocation of `train` or [`infer`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/inference.py#L108) methods in the application Python program does the following: - 1. create a new Scope instance in the [scope hierarchy](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md) for each run of a block, + 1. Create a new Scope instance in the [scope hierarchy](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md) for each run of a block, 1. realize local variables defined in the BlockDesc message in the new scope, 1. a scope is similar to the stack frame in programming languages, - 1. create an instance of class `Block`, in which, + 1. Create an instance of class `Block`, in which, 1. realize operators in the BlockDesc message, - 1. run the Block by calling + 1. Run the Block by calling 1. `Block::Eval(vector* targets)` for forward and backward computations, or 1. `Block::Eval(vector* targets)` for optimization. @@ -76,14 +77,14 @@ The word *graph* is exchangable with *block* in this document. A graph represen Compile Time -> IR -> Runtime ``` -### Benefit +### Benefits of IR - Optimization ```text Compile Time -> IR -> Optimized IR -> Runtime ``` -- Send automatically partitioned IR to different nodes. - - Automatic data parallel +- Automatically send partitioned IR to different nodes. + - Automatic Data Parallelism ```text Compile Time |-> Single GPU IR @@ -92,7 +93,7 @@ Compile Time -> IR -> Runtime |-> Node-1 (runs trainer-IR-1) |-> Node-2 (runs pserver-IR) ``` - - Automatic model parallel (planned for future) + - Automatic Model Parallelism (planned for future) --- @@ -105,10 +106,10 @@ Compile Time -> IR -> Runtime # Operator ![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/dd598e8f1976f5759f58af5e5ef94738a6b2e661/op.dot) -* `Operator` is the fundamental building block as the user interface. - * Operator stores input/output variable name, and attributes. - * The `InferShape` interface is used to infer output variable shapes by its input shapes. - * Use `Run` to compute `input variables` to `output variables`. +* `Operator` is the fundamental building block of the user interface. + * Operator stores input/output variable names, and attributes. + * The `InferShape` interface is used to infer the shape of the output variable shapes based on the shapes of the input variables. + * Use `Run` to compute the `output` variables from the `input` variables. --- @@ -126,30 +127,29 @@ Compile Time -> IR -> Runtime # Why separate Kernel and Operator * Separate GPU and CPU code. - * Make Paddle can run without GPU. -* Make one operator (which is user interface) can contain many implementations. - * Same mul op, different FP16, FP32 Kernel. different MKL, eigen kernel. + * Make Paddle capable of running without GPU. +* Make one operator (which is a user interface) and create many implementations. + * For example, same multiplication op can have different implementations kernels such as FP16 kernel, FP32 kernel, MKL, eigen kernel. --- # Libraries for Kernel development * `Eigen::Tensor` contains basic math and element-wise functions. * Note that `Eigen::Tensor` has broadcast implementation. - * Limit number of `tensor.device(dev) = ` in your code. -* `thrust::tranform` and `std::transform`. - * `thrust` has the same API as C++ standard library. Using `transform` can quickly implement a customized elementwise kernel. - * `thrust` has more complex API, like `scan`, `reduce`, `reduce_by_key`. + * Limit the number of `tensor.device(dev) = ` in your code. +* `thrust::transform` and `std::transform`. + * `thrust` has the same API as C++ standard library. Using `transform`, one can quickly implement customized element-wise kernels. + * `thrust` also has more complex APIs, like `scan`, `reduce`, `reduce_by_key`. * Hand-writing `GPUKernel` and `CPU` code - * Do not write `.h`. CPU Kernel should be in `.cc`. CPU kernel should be in `.cu`. (`GCC` cannot compile GPU code.) + * Do not write in header (`.h`) files. CPU Kernel should be in cpp source (`.cc`) and GPU kernels should be in cuda (`.cu`) files. (GCC cannot compile GPU code.) --- -# Operator Register +# Operator Registration -## Why register is necessary? +## Why is registration necessary? We need a method to build mappings between Op type names and Op classes. -## How to do the register? - -Maintain a map, whose key is the type name and value is corresponding Op constructor. +## How is registration implemented? +Maintaining a map, whose key is the type name and the value is the corresponding Op constructor. --- # The Registry Map @@ -169,7 +169,7 @@ Maintain a map, whose key is the type name and value is corresponding Op constru # Related Concepts ### Op_Maker -It's constructor takes `proto` and `checker`. They are compeleted during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)) +It's constructor takes `proto` and `checker`. They are completed during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)) ### Register Macros ```cpp @@ -177,34 +177,34 @@ REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, grad_op_class) REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) ``` -### `USE` Macros -make sure the registration process is executed and linked. +### USE Macros +Make sure the registration process is executed and linked. --- -# Register Process -1. Write Op class, as well as its gradient Op class if there is. -2. Write Op maker class. In the constructor, describe its inputs, outputs, and attributes. -3. Invoke macro `REGISTER_OP`. The macro will - 1. call maker class to complete `proto` and `checker` - 2. with the completed `proto` and `checker`, build a new key-value pair in the `OpInfoMap` +# Registration Process +1. Write an Op class and its gradient Op class, if required. +2. Write an Op maker class. In the constructor of this class, describe the inputs, outputs and attributes of the operator. +3. Invoke the macro `REGISTER_OP`. This macro will + 1. Call maker class to complete the `proto` and the `checker` + 2. Using the completed `proto` and `checker`, it will add a new key-value pair to the `OpInfoMap` -4. Invoke `USE` macro in where the Op is used to make sure it is linked. +4. Invoke the `USE` macro in which the Op is used, to make sure that it is linked. --- # Backward Module (1/2) ### Create Backward Operator -- Mapping from forwarding Op to backward Op +- Mapping from forward Op to backward Op ![backward](https://gist.githubusercontent.com/dzhwinter/a6fbd4623ee76c459f7f94591fd1abf0/raw/61026ab6e518e66bde66a889bc42557a1fccff33/backward.png) --- # Backward Module (2/2) ### Build Backward Network -- **Input** graph of forwarding operators -- **Output** graph of backward operators -- **corner case in construction** - - shared variable => insert `Add` operator - - no gradient => insert `fill_zero_grad` operator - - recursive netOp => call `Backward` recursively +- **Input**: graph of forward operators +- **Output**: graph of backward operators +- **Corner cases in construction** + - Shared Variables => insert an `Add` operator to combine gradients + - No Gradient => insert a `fill_zero_grad` operator + - Recursive NetOp => call `Backward` recursively - RNN Op => recursively call `Backward` on stepnet @@ -213,41 +213,41 @@ make sure the registration process is executed and linked. * `Tensor` is an n-dimension array with type. * Only dims and data pointers are stored in `Tensor`. - * All operators on `Tensor` is written in `Operator` or global functions. - * variable length Tensor design [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) -* `Variable` is the inputs and outputs of an operator. Not just `Tensor`. - * step_scopes in RNN is a variable and not a tensor. -* `Scope` is where variables store at. - * map - * `Scope` has a hierarchical structure. The local scope can get variable from its parent scope. + * All operations on `Tensor` are written in `Operator` or global functions. + * Variable length Tensor design [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) +* `Variable` instances are the inputs and the outputs of an operator. Not just `Tensor`. + * `step_scopes` in RNN is a variable and not a tensor. +* `Scope` is where variables are stores. + * map + * `Scope` has a hierarchical structure. The local scope can get variables from its parent scope. --- # Block (in design) -## the difference with original RNNOp -- as an operator is more intuitive than `RNNOp`, -- offers new interface `Eval(targets)` to deduce the minimal block to `Run`, -- fits the compile-time/ runtime separation design. - - during the compilation, `SymbolTable` stores `VarDesc`s and `OpDesc`s and serialize to a `BlockDesc` - - when graph executes, a Block with `BlockDesc` passed in creates `Op` and `Var` then `Run` +## the difference between original RNNOp and Block +- As an operator is more intuitive than `RNNOp`, +- Offers a new interface `Eval(targets)` to deduce the minimal block to `Run`, +- Fits the compile-time/ runtime separation design paradigm. + - During the compilation, `SymbolTable` stores `VarDesc`s and `OpDesc`s and serialize to a `BlockDesc` + - When graph executes, a Block with `BlockDesc` is passed. It then creates `Op` and `Var` instances and then invokes `Run`. --- # Milestone -- take Paddle/books as the main line, the requirement of the models motivates framework refactoring, -- model migration - - framework development gives **priority support** to model migration, for example, +- Take Paddle/books as the main line, the requirement of the models motivates framework refactoring, +- Model migration + - Framework development gives **priority support** to model migration, for example, - the MNIST demo needs a Python interface, - the RNN models require the framework to support `LoDTensor`. - - determine some timelines, - - heavily-relied Ops need to be migrated first, - - different models can be migrated parallelly. -- improve the framework at the same time -- accept imperfection, concentrated on solving the specific problem at the right price. + - Determine some timelines, + - Frequently used Ops need to be migrated first, + - Different models can be migrated in parallel. +- Improve the framework at the same time +- Accept imperfection, concentrate on solving the specific problem at the right price. --- # Control the migration quality -- compare the performance of migrated models with old ones. -- follow google C style -- build the automatic workflow of generating Python/C++ documentations - - the documentation of layers and ops should be written inside the code - - take the documentation quality into account when doing PR - - preview the documentations, read and improve them from users' perspective +- Compare the performance of migrated models with old ones. +- Follow the google C++ style +- Build the automatic workflow of generating Python/C++ documentations. + - The documentation of layers and ops should be written inside the code. + - Take the documentation quality into account when submitting pull requests. + - Preview the documentations, read and improve them from a user's perspective. diff --git a/doc/design/register_grad_op.md b/doc/design/register_grad_op.md new file mode 100644 index 0000000000..3cf8a59446 --- /dev/null +++ b/doc/design/register_grad_op.md @@ -0,0 +1,90 @@ +# Design Doc: Gradient Operators Registration + + +## The Problem Posed + +In our current operator registration mechanism, for each operator, the programmer should register a *gradient operator creator* function, which takes a C++ operator instance, and returns the corresponding gradient instance. + +However, as we decided to separate the *compilation* and *execution* of DL models, we need to reshape the creator to take a protobuf `OpDesc` message, and returns a corresponding message. + +More than that, the new registration mechanism need to support the fact that an operators' gradient computation might be a composition of operators. + +## Current Implementation + +OpInfos store in a association map which key is the operator type. The `grad_op_type` indicate associated gradient operator type. Operator can create gradient operator by `OpInfo::creator_` of gradient. The pseudo code is + +```cpp +struct OpInfo { + std::function creator_; + std::string grad_op_type_; + ... +}; + +map OpInfoMap; + +OperatorBase* CreateGradientOperator(const OperatorBase& op) { + return OpInfoMap.at(op.Type()).creator_(...); +} +``` + +## Proposed Solution + +The mapping relationship between an operator and its gradient operators is a function. The interface of that function is: + +```cpp +// (OpDesc) --> vector +std::function(const OpDescBind&)>; +``` + +The function takes an `OpDescBind` of the forward operator and returns one or many gradient operator descriptions. `OpDescBind` is a C++ wrapper for protobuf message `OpDesc` to manipulate `OpDesc` fast. + +The `GradOpDescMaker` will be registered in `OpInfo`, to replace `grad_op_type_` field. The `OpInfo` should be + +```cpp +struct OpInfo { + std::function>(const OpDescBind&)> grad_op_maker_; + ... +}; +``` + +The `grad_op_maker_ ` is `nullptr` if the operator does not have associated gradient operators. + +We propose a base class called `GradOpDescMakerBase` to let operator developers generate `Gradient Operators` easily. The public interface of that class is + +```cpp +class GradOpDescMakerBase { +public: + GradOpDescMakerBase(const OpDescBind& ); + virtual std::vector> operator()()const = 0; +}; +``` + +We can convert `GradOpDescMakerBase` to `std::function>(const OpDescBind&)>` by + +```cpp +using GradOpMaker = ...; +std::function(const OpDescBind&)> func; +func = [] (const OpDescBind& fwd_op) { + GradOpMaker maker(fwd_op); + return maker(); +}; +``` + +We can write many helper functions since the `GradOpDescMakerBase` is a class now. The basic helper functions get the variables of `Input`, `Output`, `InputGradient` and `OutputGradient` in the forwarding operator. + +We should chagne register macros at the same time. In the current solution, there is no difference between forwarding operators and backward operators. So `REGISTER_OP` just register one operator. If the `REGISTER_OPERATOR ` contains `OpProtoAndCheckerMaker` and `GradOpDescMaker`, we just list them in the same macro. It can be done by a macro contains `__VA_ARGS__`. + +The user interface should be + +```cpp +vector MinusOpGradMaker(OpDesc) {...} +REGISTER_OPERATOR(minus, MinusOp, MinusOpProtoAndCheckerMaker, SumOpGradMaker); +// Developers can still manually implement gradient operator. +REGISTER_OPERATOR(minus_grad, MinusGradOp); +``` + +The interface of current `REGISTER_OP` macro could not be changed. In `REGISTER_OP`, it will invoke `REGISTER_OPERATOR` two times and generate GradOpDescMaker inside. + +```cpp +REGISTER_OP(minus, MinusOp, MinusOpProtoAndCheckerMaker, minus_grad, MinusGradOp); +``` diff --git a/doc/design/releasing_process.md b/doc/design/releasing_process.md index 0c10e78280..62ff8f3229 100644 --- a/doc/design/releasing_process.md +++ b/doc/design/releasing_process.md @@ -1,8 +1,8 @@ -# Paddle发行规范 +# PaddlePaddle发行规范 -Paddle使用git-flow branching model做分支管理，使用[Semantic Versioning](http://semver.org/)标准表示Paddle版本号。 +PaddlePaddle使用git-flow branching model做分支管理，使用[Semantic Versioning](http://semver.org/)标准表示PaddlePaddle版本号。 -Paddle每次发新的版本，遵循以下流程: +PaddlePaddle每次发新的版本，遵循以下流程: 1. 从`develop`分支派生出新的分支，分支名为`release/版本号`。例如，`release/0.10.0` 2. 将新分支的版本打上tag，tag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`，第二个为`0.10.0rc2`，依次类推。 @@ -27,14 +27,14 @@ Paddle每次发新的版本，遵循以下流程: 需要注意的是: -* `release/版本号`分支一旦建立，一般不允许再从`develop`分支合入`release/版本号`。这样保证`release/版本号`分支功能的封闭，方便测试人员测试Paddle的行为。 +* `release/版本号`分支一旦建立，一般不允许再从`develop`分支合入`release/版本号`。这样保证`release/版本号`分支功能的封闭，方便测试人员测试PaddlePaddle的行为。 * 在`release/版本号`分支存在的时候，如果有bugfix的行为，需要将bugfix的分支同时merge到`master`, `develop`和`release/版本号`这三个分支。 -# Paddle 分支规范 +# PaddlePaddle 分支规范 -Paddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，并适应github的特性做了一些区别。 +PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，并适应github的特性做了一些区别。 -* Paddle的主版本库遵循[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范。其中: +* PaddlePaddle的主版本库遵循[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范。其中: * `master`分支为稳定(stable branch)版本分支。每一个`master`分支的版本都是经过单元测试和回归测试的版本。 * `develop`分支为开发(develop branch)版本分支。每一个`develop`分支的版本都经过单元测试，但并没有经过回归测试。 * `release/版本号`分支为每一次Release时建立的临时分支。在这个阶段的代码正在经历回归测试。 @@ -42,18 +42,18 @@ Paddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branch * 其他用户的fork版本库并不需要严格遵守[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，但所有fork的版本库的所有分支都相当于特性分支。 * 建议，开发者fork的版本库使用`develop`分支同步主版本库的`develop`分支 * 建议，开发者fork的版本库中，再基于`develop`版本fork出自己的功能分支。 - * 当功能分支开发完毕后，向Paddle的主版本库提交`Pull Reuqest`，进而进行代码评审。 + * 当功能分支开发完毕后，向PaddlePaddle的主版本库提交`Pull Reuqest`，进而进行代码评审。 * 在评审过程中，开发者修改自己的代码，可以继续在自己的功能分支提交代码。 * BugFix分支也是在开发者自己的fork版本库维护，与功能分支不同的是，BugFix分支需要分别给主版本库的`master`、`develop`与可能有的`release/版本号`分支，同时提起`Pull Request`。 -# Paddle回归测试列表 +# PaddlePaddle回归测试列表 -本列表说明Paddle发版之前需要测试的功能点。 +本列表说明PaddlePaddle发版之前需要测试的功能点。 -## Paddle Book中所有章节 +## PaddlePaddle Book中所有章节 -Paddle每次发版本首先要保证Paddle Book中所有章节功能的正确性。功能的正确性包括验证Paddle目前的`paddle_trainer`训练和纯使用`Python`训练模型正确性。 +PaddlePaddle每次发版本首先要保证PaddlePaddle Book中所有章节功能的正确性。功能的正确性包括验证PaddlePaddle目前的`paddle_trainer`训练和纯使用`Python`训练模型正确性。 | | 新手入门章节 | 识别数字 | 图像分类 | 词向量 | 情感分析 | 语意角色标注 | 机器翻译 | 个性化推荐 | | --- | --- | --- | --- | --- | --- | --- | --- | --- | diff --git a/doc/design/scope.md b/doc/design/scope.md index c9e0be716b..b1f9bb4378 100644 --- a/doc/design/scope.md +++ b/doc/design/scope.md @@ -17,7 +17,7 @@ Scope is an association of a name to variable. All variables belong to `Scope`. 1. Scope only contains a map of a name to variable. - All parameters, data, states in a Net should be variables and stored inside a scope. Each op should get inputs and outputs to do computation from a scope, such as data buffer, state(momentum) etc. + All parameters, data, states in a Net should be variables and stored inside a scope. Each op should get inputs and outputs to do computation from a scope, such as data buffer, state (momentum) etc. 1. Variable can only be created by Scope and a variable can only be got from Scope. User cannot create or get a variable outside a scope. This is a constraints of our framework, and will keep our framework simple and clear. @@ -32,7 +32,7 @@ Scope is an association of a name to variable. All variables belong to `Scope`. 1. Scope should destruct all Variables inside it when itself is destructed. User can never store `Variable` pointer somewhere else. - Because Variable can only be got from Scope. When destroying Scope, we also need to destroy all the Variables in it. If user store `Variable` pointer to private data member or some global variable, the pointer will be a invalid pointer when associated `Scope` is destroyed. + Because Variable can only be got from Scope. When destroying Scope, we also need to destroy all the Variables in it. If user store `Variable` pointer to private data member or some global variable, the pointer will be an invalid pointer when associated `Scope` is destroyed. ```cpp class Scope { @@ -50,7 +50,7 @@ class Scope { Just like [scope](https://en.wikipedia.org/wiki/Scope_(computer_science)) in programming languages, `Scope` in the neural network can also be a local scope. There are two attributes about local scope. -1. We can create local variables in a local scope. When that local scope are destroyed, all local variables should also be destroyed. +1. We can create local variables in a local scope. When that local scope is destroyed, all local variables should also be destroyed. 2. Variables in a parent scope can be retrieved from local scopes of that parent scope, i.e., when user get a variable from a scope, it will try to search this variable in current scope. If there is no such variable in the local scope, `scope` will keep searching from its parent, until the variable is found or there is no parent. ```cpp @@ -121,4 +121,4 @@ Also, as the parent scope is a `shared_ptr`, we can only `Create()` a scope shar ## Orthogonal interface -`FindVar` will return `nullptr` when `name` is not found. It can be used as `Contains` method. `NewVar` will return a `Error` when there is a name conflict locally. Combine `FindVar` and `NewVar`, we can implement `NewVar` easily. +`FindVar` will return `nullptr` when `name` is not found. It can be used as `Contains` method. `NewVar` will return an `Error` when there is a name conflict locally. Combine `FindVar` and `NewVar`, we can implement `NewVar` easily. diff --git a/doc/design/simple_op_design.md b/doc/design/simple_op_design.md index fded4a6861..c7aeed7f9b 100644 --- a/doc/design/simple_op_design.md +++ b/doc/design/simple_op_design.md @@ -6,9 +6,9 @@ The Interaction between Python and C++ can be simplified as two steps: 1. C++ tells Python how many Ops there are, and what parameter do users need to offer to initialize a new Op. Python then builds API for each Op at compile time. -2. Users invoke APIs built by Python and provide necessary parameters. These parameters will be sent to C++ fo finish Op construction task. +2. Users invoke APIs built by Python and provide necessary parameters. These parameters will be sent to C++ for finishing the Op construction task. -### Message form C++ to Python +### Message from C++ to Python We define a Protobuf message class `OpProto` to hold message needed in the first step. What should an `OpProto` contain? This question is equivalent to “What message do we need to offer, to build a Python API which is legal and user oriented and can use to describe a whole Op.” @@ -193,7 +193,7 @@ def fc_layer(input, size, with_bias, activation): elif: # ... return act_output; -``` +``` ### Low Leval API diff --git a/doc/design/tensor_array.md b/doc/design/tensor_array.md new file mode 100644 index 0000000000..8378e97bf7 --- /dev/null +++ b/doc/design/tensor_array.md @@ -0,0 +1,271 @@ +# Design for TensorArray +This design doc presents the necessity of a new C++ class `TensorArray`. +In addition to the very simple C++ implementation + +```c++ +class TensorArray { + public: + explicit TensorArray(const LoDTensor&); + explicit TensorArray(size_t size); + + private: + vector values_; +}; +``` + +We also need to expose it to PaddlePaddle's Python API, +because users would want to use it with our very flexible operators `WhileLoop`. +An example for a RNN based on dynamic operators is + +```python +input = pd.data(...) +num_steps = Var(12) + +TensorArray states(size=num_steps) +TensorArray step_inputs(unstack_from=input) +TensorArray step_outputs(size=num_steps) + +W = Tensor(...) +U = Tensor(...) +default_state = some_op() + +step = Var(1) + +wloop = paddle.create_whileloop(loop_vars=[step]) +with wloop.frame(): + wloop.break_if(pd.equal(step, num_steps) + pre_state = states.read(step-1, default_state) + step_input = step_inputs.read(step) + state = pd.sigmoid(pd.matmul(U, pre_state) + pd.matmul(W, step_input)) + states.write(step, state) + step_outputs.write(step, state) # output state + step.update(state+1) + +output = step_outputs.stack() +``` + +## Background +Steps are one of the core concepts of RNN. In each time step of RNN, there should be several input segments, states, and output segments; all these components act like arrays, for example, call `states[step_id]` will get the state in `step_id`th time step. + +An RNN can be implemented with the following pseudocode + +```c++ +Array states; +Array input_segments; +Array output_segments; +Parameter W, U; + +step = 1 +seq_len = 12 +while_loop { + if (step == seq_len) break; + states[step] = sigmoid(W * states[step-1] + U * input_segments[step]); + output_segments[step] = states[step] // take state as output + step++; +} +``` +According to the [RNN roadmap](https://github.com/PaddlePaddle/Paddle/issues/4561), there are several different RNNs that PaddlePaddle will eventually support. + +Currently, the basic RNN implementation supported by PaddlePaddle is the `recurrent_op` which takes tensors as input and splits them into `input_segments`. + + +Since a tensor cannot store variable-length sequences directly, PaddlePaddle implements the tensor with level of details (`LoDTensor` for short). +Segmenting the `LoDTensor` is much more complicated than splitting a tensor, that makes it necessary to refactor the `recurrent_op` with `LoDTensor` segmenting support. + +As the next step in RNN support, `dynamic_recurrent_op` should be introduced to handle inputs with variable-length sequences. + +The implementation is similar to `recurrent_op`. +The key difference is the way **the original input `LoDTensors` and outupts are split to get the `input_segments` and the `output_segments`.** + + +Though it can't be built over `recurrent_op` or `dynamic_recurrent_op` directly, +the logic behind splitting a tensor or a LoD tensor into `input_segments` remains the same. + +## Why `TensorArray` +The logic behind splitting the inputs to segments, states and outputs is similar and can be shared in a seperate module. + +The array of `states`, `input_segments` and `output_segments` would be exposed to users when writing a dynamic RNN model similar to the above pseudo codes. + +So there should be an array-like container, which can store the segments of a tensor or LoD tensor. + +**This container can store an array of tensors and provides several methods to split a tensor or a LoD tensor** . +This is where the notion of `TensorArray` comes from. + +## Introduce TensorArray to uniform all the three RNNs +TensorArray as a new concept is borrowed from TensorFlow, +it is meant to be used with dynamic iteration primitives such as `while_loop` and `map_fn`. + +This concept can be used to support our new design of dynamic operations, and help to refactor some existing variant-sentence-related layers, +such as `recurrent_op`, `RecurrentGradientMachine`. + +In [our design for dynamic RNN](https://github.com/PaddlePaddle/Paddle/pull/4401), +`TensorArray` is used to segment inputs and store states in all time steps. +By providing some methods similar to a C++ array, +the definition of some state-based dynamic models such as RNN can be more natural and highly flexible. + +## Dynamic-operations on TensorArray + +`TensorArray` will be used directly when defining dynamic models, so some operators listed below should be implemented + +```python +# several helper operators for TensorArray +def tensor_array_stack(ta, tensor): + ''' + get a tensor array `ta`, return a packed `tensor`. + ''' + pass + +def tensor_array_unstack(tensor, ta): + ''' + get a `tensor`, unstack it and get a tensor array `ta`. + ''' + pass + +def tensor_array_write(ta, index, tensor, data_shared): + ''' + get a `tensor` and a scalar tensor `index`, write `tensor` into index-th + value of the tensor array `ta`. + `data_shared` is an attribute that specifies whether to copy or reference the tensors. + ''' + pass + +def tensor_array_read(ta, index, tensor): + ''' + get a tensor array `ta`, a scalar tensor `index`, read the index-th value of + `ta` and return as the `tensor`. + ''' + pass + +def tensor_array_size(ta, tensor): + ''' + get a tensor array `ta`, return the size of `ta` and return as the scalar `tensor`. + ''' + pass +``` + +It is trivial for users to use so many low-level operators, so some helper methods should be proposed in python wrapper to make `TensorArray` easier to use, +for example + +```python +class TensorArray: + def __init__(self, name): + self.name = name + self.desc = TensorArrayDesc() + + def stack(self, name=None): + ''' + Pack the values in a `TensorArray` into a tensor with rank one higher + than each tensor in `values`. + `stack` can be used to split tensor into time steps for RNN or whileloop. + + @name: str + the name of the variable to output. + ''' + tensor = NewVar(name) + tensor_array_stack(self.name, tensor) + return tensor + + def unstack(self, input): + ''' + Unpacks the given dimension of a rank-`R` tensor into rank-`(R-1)` tensors. + `unstack` can be used to concatenate all the time steps for RNN or whileloop. + + @input: str + the name of input tensor + ''' + tensor_array_unstack(tensor, self.name) + + def write(self, index, value, data_shared=True): + ''' + Write value into index of the TensorArray. + If `data_shared` is set to True, than the index-th value in TensorArray will + be shared with the tensor passed in. + + @index: str + name of a scalar tensor + @value: str + name of a tensor + @data_shared: bool + ''' + tensor_array_write(self.name, index, value, data_shared) + + def read(self, index, output): + ''' + Read the value at location `index` in the `TensorArray`. + + @index: str + name of a scalar tensor + @output: + name of a output variable + ''' + tensor_array_read(self.name, index, output) + + + def size(self, output): + ''' + Return the number of values. + + @output: str + name of a scalar tensor + ''' + tensor_array_size(self.name, output) +``` + +## LoDTensor-related Supports +The `RecurrentGradientMachine` in Paddle serves as a flexible RNN layer; it takes varience-length sequences as input, and output sequences too. + +Since each step of RNN can only take a tensor-represented batch of data as input, +some preprocess should be taken on the inputs such as sorting the sentences by their length in descending order and cut each word and pack to new batches. + +Such cut-like operations can be embedded into `TensorArray` as general methods called `unpack` and `pack`, +these two operations are similar to `stack` and `unstack` except that they operate on variable-length sequences formated as a LoD tensor rather than a tensor. + +Some definitions are like + +```python +def unpack(level): + ''' + Split LodTensor in some `level` and generate batches, if set `sort_by_length`, + will sort by length. + + Returns: + - a new `TensorArray`, whose values are LodTensors and represents batches + of data. + - an int32 Tensor, which stores the map from the new batch's indices to + original LoDTensor + ''' + pass + +def pack(level, indices_map): + ''' + Recover the original LoD-arranged LoDTensor with the values in a `TensorArray` + and `level` and `indices_map`. + ''' + pass +``` + +With these two methods, a varience-length sentence supported RNN can be implemented like + +```c++ +// input is the varient-length data +LodTensor sentence_input(xxx); +TensorArray ta; +Tensor indice_map; +Tensor boot_state = xxx; // to initialize rnn's first state +TensorArray::unpack(input, 1/*level*/, true/*sort_by_length*/, &ta, &indice_map); +TessorArray step_outputs; +TensorArray states; + +for (int step = 0; step = ta.size(); step++) { + auto state = states.read(step); + // rnnstep is a function which acts like a step of RNN + auto step_input = ta.read(step); + auto step_output = rnnstep(step_input, state); + step_outputs.write(step_output, true/*data_shared*/); +} + +// rnn_output is the final output of an rnn +LoDTensor rnn_output = ta.pack(ta, indice_map); +``` +the code above shows that by embedding the LoDTensor-related preprocess operations into `TensorArray`, +the implementation of a RNN that supports varient-length sentences is far more concise than `RecurrentGradientMachine` because the latter mixes all the codes together, hard to read and extend. diff --git a/doc/design/var_desc.md b/doc/design/var_desc.md index 86a95c10d5..bfbbdd0578 100644 --- a/doc/design/var_desc.md +++ b/doc/design/var_desc.md @@ -1,7 +1,7 @@ ## Background PaddlePaddle divides the description of neural network computation graph into two stages: compile time and runtime. -PaddlePaddle use proto message to describe compile time graph for +PaddlePaddle use proto message to describe compile time graph because 1. Computation graph should be able to be saved to a file. 1. In distributed training, the graph will be serialized and send to multiple workers. diff --git a/doc/faq/build_and_install/index_cn.rst b/doc/faq/build_and_install/index_cn.rst new file mode 100644 index 0000000000..f1677e216f --- /dev/null +++ b/doc/faq/build_and_install/index_cn.rst @@ -0,0 +1,111 @@ +################### +编译安装与单元测试 +################### + +.. contents:: + +1. 运行Docker GPU镜像出现 "CUDA driver version is insufficient" +---------------------------------------------------------------- + +用户在使用PaddlePaddle GPU的Docker镜像的时候，常常出现 `Cuda Error: CUDA driver version is insufficient for CUDA runtime version`, 原因在于没有把机器上CUDA相关的驱动和库映射到容器内部。 +具体的解决方法是： + +.. code-block:: bash + + $ export CUDA_SO="$(\ls usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')" + $ export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}') + $ docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddlepaddle:latest-gpu + +更多关于Docker的安装与使用, 请参考 `PaddlePaddle Docker 文档 `_ 。 + + +2. CMake源码编译, 找到的PythonLibs和PythonInterp版本不一致 +---------------------------------------------------------------- + +这是目前CMake寻找Python的逻辑存在缺陷，如果系统安装了多个Python版本，CMake找到的Python库和Python解释器版本可能有不一致现象，导致编译PaddlePaddle失败。正确的解决方法是， +用户强制指定特定的Python版本，具体操作如下： + + .. code-block:: bash + + cmake .. -DPYTHON_EXECUTABLE= -DPYTHON_LIBRARY= -DPYTHON_INCLUDE_DIR= + +用户需要指定本机上Python的路径：````, ````, ```` + +3. CMake源码编译，Paddle版本号为0.0.0 +-------------------------------------- + +如果运行 :code:`paddle version`, 出现 :code:`PaddlePaddle 0.0.0`；或者运行 :code:`cmake ..`，出现 + +.. code-block:: bash + + CMake Warning at cmake/version.cmake:20 (message): + Cannot add paddle version from git tag + +那么用户需要拉取所有的远程分支到本机，命令为 :code:`git fetch upstream`，然后重新cmake即可。 + +4. paddlepaddle\*.whl is not a supported wheel on this platform. +------------------------------------------------------------------------ + +出现这个问题的主要原因是，没有找到和当前系统匹配的paddlepaddle安装包。最新的paddlepaddle python安装包支持Linux x86_64和MacOS 10.12操作系统，并安装了python 2.7和pip 9.0.1。 + +更新 :code:`pip` 包的方法是\: + +.. code-block:: bash + + pip install --upgrade pip + +如果还不行，可以执行 :code:`python -c "import pip; print(pip.pep425tags.get_supported())"` 获取当前系统支持的python包的后缀， +并对比是否和正在安装的后缀一致。 + +如果系统支持的是 :code:`linux_x86_64` 而安装包是 :code:`manylinux1_x86_64` ，需要升级pip版本到最新； +如果系统支持 :code:`manylinux1_x86_64` 而安装包（本地）是 :code:`linux_x86_64` ，可以重命名这个whl包为 :code:`manylinux1_x86_64` 再安装。 + +5. 编译安装后执行 import paddle.v2 as paddle 报ImportError: No module named v2 +------------------------------------------------------------------------------------------ +先查看一下是否曾经安装过paddle v1版本，有的话需要先卸载： + +pip uninstall py_paddle paddle + +然后安装paddle的python环境, 在build目录下执行 + +pip install python/dist/paddle*.whl && pip install ../paddle/dist/py_paddle*.whl + +6. 遇到“非法指令”或者是“illegal instruction” +-------------------------------------------- + +PaddlePaddle使用avx SIMD指令提高cpu执行效率，因此错误的使用二进制发行版可能会导致这种错误，请选择正确的版本。 + +7. python相关的单元测试都过不了 +-------------------------------- + +如果出现以下python相关的单元测试都过不了的情况： + +.. code-block:: bash + + 24 - test_PyDataProvider (Failed) + 26 - test_RecurrentGradientMachine (Failed) + 27 - test_NetworkCompare (Failed) + 28 - test_PyDataProvider2 (Failed) + 32 - test_Prediction (Failed) + 33 - test_Compare (Failed) + 34 - test_Trainer (Failed) + 35 - test_TrainerOnePass (Failed) + 36 - test_CompareTwoNets (Failed) + 37 - test_CompareTwoOpts (Failed) + 38 - test_CompareSparse (Failed) + 39 - test_recurrent_machine_generation (Failed) + 40 - test_PyDataProviderWrapper (Failed) + 41 - test_config_parser (Failed) + 42 - test_swig_api (Failed) + 43 - layers_test (Failed) + +并且查询PaddlePaddle单元测试的日志，提示： + +.. code-block:: bash + + paddle package is already in your PYTHONPATH. But unittest need a clean environment. + Please uninstall paddle package before start unittest. Try to 'pip uninstall paddle'. + +解决办法是： + +* 卸载PaddlePaddle包 :code:`pip uninstall paddle`, 清理掉老旧的PaddlePaddle安装包，使得单元测试有一个干净的环境。如果PaddlePaddle包已经在python的site-packages里面，单元测试会引用site-packages里面的python包，而不是源码目录里 :code:`/python` 目录下的python包。同时，即便设置 :code:`PYTHONPATH` 到 :code:`/python` 也没用，因为python的搜索路径是优先已经安装的python包。 diff --git a/doc/faq/cluster/index_cn.rst b/doc/faq/cluster/index_cn.rst new file mode 100644 index 0000000000..e59c1e1a54 --- /dev/null +++ b/doc/faq/cluster/index_cn.rst @@ -0,0 +1,17 @@ +############### +集群训练与预测 +############### + +.. contents:: + +1. 集群多节点训练，日志中保存均为网络通信类错误 +------------------------------------------------ + +集群多节点训练，日志报错为网络通信类错误，比如 :code:`Connection reset by peer` 等。 +此类报错通常是由于某一个节点的错误导致这个节点的训练进程退出，从而引发其他节点无法连接导致，可以参考下面的步骤排查： + +* 从 :code:`train.log` ， :code:`server.log` 找到最早报错的地方，查看是否是其他错误引发的报错（比如FPE，内存不足，磁盘空间不足等）。 + +* 如果发现最早的报错就是网络通信的问题，很有可能是非独占方式执行导致的端口冲突，可以联系OP，看当前MPI集群是否支持resource=full参数提交，如果支持增加此参数提交，并更换job 端口。 + +* 如果当前MPI集群并不支持任务独占模式，可以联系OP是否可以更换集群或升级当前集群。 diff --git a/doc/faq/index_cn.rst b/doc/faq/index_cn.rst index 138efb566e..9929767cac 100644 --- a/doc/faq/index_cn.rst +++ b/doc/faq/index_cn.rst @@ -1,323 +1,11 @@ -#################### FAQ -#################### +==== -.. contents:: +.. toctree:: + :maxdepth: 1 -1. 如何减少内存占用 ---------------------------------- - -神经网络的训练本身是一个非常消耗内存和显存的工作，经常会消耗数10GB的内存和数GB的显存。 -PaddlePaddle的内存占用主要分为如下几个方面\: - -* DataProvider缓冲池内存（只针对内存） -* 神经元激活内存（针对内存和显存） -* 参数内存 （针对内存和显存） -* 其他内存杂项 - -其中，其他内存杂项是指PaddlePaddle本身所用的一些内存，包括字符串分配，临时变量等等，暂不考虑在内。 - -减少DataProvider缓冲池内存 -++++++++++++++++++++++++++ - -PyDataProvider使用的是异步加载，同时在内存里直接随即选取数据来做Shuffle。即 - -.. graphviz:: - - digraph { - rankdir=LR; - 数据文件 -> 内存池 -> PaddlePaddle训练 - } - -所以，减小这个内存池即可减小内存占用，同时也可以加速开始训练前数据载入的过程。但是，这 -个内存池实际上决定了shuffle的粒度。所以，如果将这个内存池减小，又要保证数据是随机的， -那么最好将数据文件在每次读取之前做一次shuffle。可能的代码为 - -.. literalinclude:: src/reduce_min_pool_size.py - -这样做可以极大的减少内存占用，并且可能会加速训练过程，详细文档参考 :ref:`api_pydataprovider2` 。 - -神经元激活内存 -++++++++++++++ - -神经网络在训练的时候，会对每一个激活暂存一些数据，如神经元激活值等。 -在反向传递的时候，这些数据会被用来更新参数。这些数据使用的内存主要和两个参数有关系， -一是batch size，另一个是每条序列(Sequence)长度。所以，其实也是和每个mini-batch中包含 -的时间步信息成正比。 - -所以做法可以有两种： - -* 减小batch size。 即在网络配置中 :code:`settings(batch_size=1000)` 设置成一个小一些的值。但是batch size本身是神经网络的超参数，减小batch size可能会对训练结果产生影响。 -* 减小序列的长度，或者直接扔掉非常长的序列。比如，一个数据集大部分序列长度是100-200, - 但是突然有一个10000长的序列，就很容易导致内存超限，特别是在LSTM等RNN中。 - -参数内存 -++++++++ - -PaddlePaddle支持非常多的优化算法(Optimizer)，不同的优化算法需要使用不同大小的内存。 -例如使用 :code:`adadelta` 算法，则需要使用等于权重参数规模大约5倍的内存。举例，如果参数保存下来的模型目录 -文件为 :code:`100M`， 那么该优化算法至少需要 :code:`500M` 的内存。 - -可以考虑使用一些优化算法，例如 :code:`momentum`。 - -2. 如何加速PaddlePaddle的训练速度 ---------------------------------- - -加速PaddlePaddle训练可以考虑从以下几个方面\： - -* 减少数据载入的耗时 -* 加速训练速度 -* 利用分布式训练驾驭更多的计算资源 - -减少数据载入的耗时 -++++++++++++++++++ - -使用\ :code:`pydataprovider`\ 时，可以减少缓存池的大小，同时设置内存缓存功能，即可以极大的加速数据载入流程。 -:code:`DataProvider` 缓存池的减小，和之前减小通过减小缓存池来减小内存占用的原理一致。 - -.. literalinclude:: src/reduce_min_pool_size.py - -同时 :code:`@provider` 接口有一个 :code:`cache` 参数来控制缓存方法，将其设置成 :code:`CacheType.CACHE_PASS_IN_MEM` 的话，会将第一个 :code:`pass` (过完所有训练数据即为一个pass)生成的数据缓存在内存里，在之后的 :code:`pass` 中，不会再从 :code:`python` 端读取数据，而是直接从内存的缓存里读取数据。这也会极大减少数据读入的耗时。 - - -加速训练速度 -++++++++++++ - -PaddlePaddle支持Sparse的训练，sparse训练需要训练特征是 :code:`sparse_binary_vector` 、 :code:`sparse_vector` 、或者 :code:`integer_value` 的任一一种。同时，与这个训练数据交互的Layer，需要将其Parameter设置成 sparse 更新模式，即设置 :code:`sparse_update=True` - -这里使用简单的 :code:`word2vec` 训练语言模型距离，具体使用方法为\: - -使用一个词前两个词和后两个词，来预测这个中间的词。这个任务的DataProvider为\: - -.. literalinclude:: src/word2vec_dataprovider.py - -这个任务的配置为\: - -.. literalinclude:: src/word2vec_config.py - - -利用更多的计算资源 -++++++++++++++++++ - -利用更多的计算资源可以分为一下几个方式来进行\: - -* 单机CPU训练 - - * 使用多线程训练。设置命令行参数 :code:`trainer_count`。 - -* 单机GPU训练 - - * 使用显卡训练。设置命令行参数 :code:`use_gpu`。 - * 使用多块显卡训练。设置命令行参数 :code:`use_gpu` 和 :code:`trainer_count` 。 - -* 多机训练 - - * 请参考 :ref:`cluster_train` 。 - - -3. 遇到“非法指令”或者是“illegal instruction” --------------------------------------------- - -PaddlePaddle使用avx SIMD指令提高cpu执行效率，因此错误的使用二进制发行版可能会导致这种错误，请选择正确的版本。 - -4. 如何选择SGD算法的学习率 --------------------------- - -在采用sgd/async_sgd进行训练时，一个重要的问题是选择正确的learning_rate。如果learning_rate太大，那么训练有可能不收敛，如果learning_rate太小，那么收敛可能很慢，导致训练时间过长。 - -通常做法是从一个比较大的learning_rate开始试，如果不收敛，那减少学习率10倍继续试验，直到训练收敛为止。那么如何判断训练不收敛呢？可以估计出如果模型采用不变的输出最小的cost0是多少。 - -如果训练过程的的cost明显高于这个常数输出的cost，那么我们可以判断为训练不收敛。举一个例子，假如我们是三分类问题，采用multi-class-cross-entropy作为cost，数据中0,1,2三类的比例为 :code:`0.2, 0.5, 0.3` , 那么常数输出所能达到的最小cost是 :code:`-(0.2*log(0.2)+0.5*log(0.5)+0.3*log(0.3))=1.03` 。如果训练一个pass（或者更早）后，cost还大于这个数，那么可以认为训练不收敛，应该降低学习率。 - - -5. 如何初始化参数 ------------------ - -默认情况下，PaddlePaddle使用均值0，标准差为 :math:`\frac{1}{\sqrt{d}}` 来初始化参数。其中 :math:`d` 为参数矩阵的宽度。这种初始化方式在一般情况下不会产生很差的结果。如果用户想要自定义初始化方式，PaddlePaddle目前提供两种参数初始化的方式\: - -* 高斯分布。将 :code:`param_attr` 设置成 :code:`param_attr=ParamAttr(initial_mean=0.0, initial_std=1.0)` -* 均匀分布。将 :code:`param_attr` 设置成 :code:`param_attr=ParamAttr(initial_max=1.0, initial_min=-1.0)` - -比如设置一个全连接层的参数初始化方式和bias初始化方式，可以使用如下代码。 - -.. code-block:: python - - hidden = fc_layer(input=ipt, param_attr=ParamAttr(initial_max=1.0, initial_min=-1.0), - bias_attr=ParamAttr(initial_mean=1.0, initial_std=0.0)) - -上述代码将bias全部初始化为1.0, 同时将参数初始化为 :code:`[1.0, -1.0]` 的均匀分布。 - -6. 如何共享参数 ---------------- - -PaddlePaddle的参数使用名字 :code:`name` 作为参数的ID，相同名字的参数，会共享参数。设置参数的名字，可以使用 :code:`ParamAttr(name="YOUR_PARAM_NAME")` 来设置。更方便的设置方式，是使得要共享的参数使用同样的 :code:`ParamAttr` 对象。 - -简单的全连接网络，参数共享的配置示例为\: - -.. literalinclude:: ../../python/paddle/trainer_config_helpers/tests/configs/shared_fc.py - -这里 :code:`hidden_a` 和 :code:`hidden_b` 使用了同样的parameter和bias。并且softmax层的两个输入也使用了同样的参数 :code:`softmax_param`。 - -7. \*-cp27mu-linux_x86_64.whl is not a supported wheel on this platform. ------------------------------------------------------------------------- - -出现这个问题的主要原因是，系统编译wheel包的时候，使用的 :code:`wheel` 包是最新的， -而系统中的 :code:`pip` 包比较老。具体的解决方法是，更新 :code:`pip` 包并重新编译PaddlePaddle。 -更新 :code:`pip` 包的方法是\: - -.. code-block:: bash - - pip install --upgrade pip - -8. python相关的单元测试都过不了 --------------------------------- - -如果出现以下python相关的单元测试都过不了的情况： - -.. code-block:: bash - - 24 - test_PyDataProvider (Failed) - 26 - test_RecurrentGradientMachine (Failed) - 27 - test_NetworkCompare (Failed) - 28 - test_PyDataProvider2 (Failed) - 32 - test_Prediction (Failed) - 33 - test_Compare (Failed) - 34 - test_Trainer (Failed) - 35 - test_TrainerOnePass (Failed) - 36 - test_CompareTwoNets (Failed) - 37 - test_CompareTwoOpts (Failed) - 38 - test_CompareSparse (Failed) - 39 - test_recurrent_machine_generation (Failed) - 40 - test_PyDataProviderWrapper (Failed) - 41 - test_config_parser (Failed) - 42 - test_swig_api (Failed) - 43 - layers_test (Failed) - -并且查询PaddlePaddle单元测试的日志，提示： - -.. code-block:: bash - - paddle package is already in your PYTHONPATH. But unittest need a clean environment. - Please uninstall paddle package before start unittest. Try to 'pip uninstall paddle'. - -解决办法是： - -* 卸载PaddlePaddle包 :code:`pip uninstall paddle`, 清理掉老旧的PaddlePaddle安装包，使得单元测试有一个干净的环境。如果PaddlePaddle包已经在python的site-packages里面，单元测试会引用site-packages里面的python包，而不是源码目录里 :code:`/python` 目录下的python包。同时，即便设置 :code:`PYTHONPATH` 到 :code:`/python` 也没用，因为python的搜索路径是优先已经安装的python包。 - - -9. 运行Docker GPU镜像出现 "CUDA driver version is insufficient" ----------------------------------------------------------------- - -用户在使用PaddlePaddle GPU的Docker镜像的时候，常常出现 `Cuda Error: CUDA driver version is insufficient for CUDA runtime version`, 原因在于没有把机器上CUDA相关的驱动和库映射到容器内部。 -具体的解决方法是： - -.. code-block:: bash - - $ export CUDA_SO="$(\ls usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')" - $ export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}') - $ docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddlepaddle:latest-gpu - -更多关于Docker的安装与使用, 请参考 `PaddlePaddle Docker 文档 `_ 。 - - -10. CMake源码编译, 找到的PythonLibs和PythonInterp版本不一致 ----------------------------------------------------------------- - -这是目前CMake寻找Python的逻辑存在缺陷，如果系统安装了多个Python版本，CMake找到的Python库和Python解释器版本可能有不一致现象，导致编译PaddlePaddle失败。正确的解决方法是， -用户强制指定特定的Python版本，具体操作如下： - - .. code-block:: bash - - cmake .. -DPYTHON_EXECUTABLE= -DPYTHON_LIBRARY= -DPYTHON_INCLUDE_DIR= - -用户需要指定本机上Python的路径：````, ````, ```` - -11. CMake源码编译，Paddle版本号为0.0.0 --------------------------------------- - -如果运行 :code:`paddle version`, 出现 :code:`PaddlePaddle 0.0.0`；或者运行 :code:`cmake ..`，出现 - -.. code-block:: bash - - CMake Warning at cmake/version.cmake:20 (message): - Cannot add paddle version from git tag - -那么用户需要拉取所有的远程分支到本机，命令为 :code:`git fetch upstream`，然后重新cmake即可。 - -12. A protocol message was rejected because it was too big ----------------------------------------------------------- - -如果在训练NLP相关模型时，出现以下错误： - -.. code-block:: bash - - [libprotobuf ERROR google/protobuf/io/coded_stream.cc:171] A protocol message was rejected because it was too big (more than 67108864 bytes). To increase the limit (or to disable these warnings), see CodedInputStream::SetTotalBytesLimit() in google/protobuf/io/coded_stream.h. - F1205 14:59:50.295174 14703 TrainerConfigHelper.cpp:59] Check failed: m->conf.ParseFromString(configProtoStr) - -可能的原因是：传给dataprovider的某一个args过大，一般是由于直接传递大字典导致的。错误的define_py_data_sources2类似： - -.. code-block:: python - - src_dict = dict() - for line_count, line in enumerate(open(src_dict_path, "r")): - src_dict[line.strip()] = line_count - - define_py_data_sources2( - train_list, - test_list, - module="dataprovider", - obj="process", - args={"src_dict": src_dict}) - -解决方案是：将字典的地址作为args传给dataprovider，然后在dataprovider里面根据该地址加载字典。即define_py_data_sources2应改为： - -.. code-block:: python - - define_py_data_sources2( - train_list, - test_list, - module="dataprovider", - obj="process", - args={"src_dict_path": src_dict_path}) - -完整源码可参考 `seqToseq `_ 示例。 - -13. 如何指定GPU设备 -------------------- - -例如机器上有4块GPU，编号从0开始，指定使用2、3号GPU： - -* 方式1：通过 `CUDA_VISIBLE_DEVICES `_ 环境变量来指定特定的GPU。 - -.. code-block:: bash - - env CUDA_VISIBLE_DEVICES=2,3 paddle train --use_gpu=true --trainer_count=2 - -* 方式2：通过命令行参数 ``--gpu_id`` 指定。 - -.. code-block:: bash - - paddle train --use_gpu=true --trainer_count=2 --gpu_id=2 - - -14. 训练过程中出现 :code:`Floating point exception`, 训练因此退出怎么办? ------------------------------------------------------------------------- - -Paddle二进制在运行时捕获了浮点数异常，只要出现浮点数异常(即训练过程中出现NaN或者Inf)，立刻退出。浮点异常通常的原因是浮点数溢出、除零等问题。 -主要原因包括两个方面: - -* 训练过程中参数或者训练过程中的梯度尺度过大，导致参数累加，乘除等时候，导致了浮点数溢出。 -* 模型一直不收敛，发散到了一个数值特别大的地方。 -* 训练数据有问题，导致参数收敛到了一些奇异的情况。或者输入数据尺度过大，有些特征的取值达到数百万，这时进行矩阵乘法运算就可能导致浮点数溢出。 - -主要的解决办法是减小学习律或者对数据进行归一化处理。 - -15. 编译安装后执行 import paddle.v2 as paddle 报ImportError: No module named v2 ------------------------------------------------------------------------- -先查看一下是否曾经安装过paddle v1版本，有的话需要先卸载： - -pip uninstall py_paddle paddle - -然后安装paddle的python环境, 在build目录下执行 - -pip install python/dist/paddle*.whl && pip install ../paddle/dist/py_paddle*.whl + build_and_install/index_cn.rst + model/index_cn.rst + parameter/index_cn.rst + local/index_cn.rst + cluster/index_cn.rst diff --git a/doc/faq/local/index_cn.rst b/doc/faq/local/index_cn.rst new file mode 100644 index 0000000000..75c4ba028e --- /dev/null +++ b/doc/faq/local/index_cn.rst @@ -0,0 +1,213 @@ +############### +本地训练与预测 +############### + +.. contents:: + +1. 如何减少内存占用 +------------------- + +神经网络的训练本身是一个非常消耗内存和显存的工作，经常会消耗数10GB的内存和数GB的显存。 +PaddlePaddle的内存占用主要分为如下几个方面\: + +* DataProvider缓冲池内存（只针对内存） +* 神经元激活内存（针对内存和显存） +* 参数内存 （针对内存和显存） +* 其他内存杂项 + +其中，其他内存杂项是指PaddlePaddle本身所用的一些内存，包括字符串分配，临时变量等等，暂不考虑在内。 + +减少DataProvider缓冲池内存 +++++++++++++++++++++++++++ + +PyDataProvider使用的是异步加载，同时在内存里直接随即选取数据来做Shuffle。即 + +.. graphviz:: + + digraph { + rankdir=LR; + 数据文件 -> 内存池 -> PaddlePaddle训练 + } + +所以，减小这个内存池即可减小内存占用，同时也可以加速开始训练前数据载入的过程。但是，这 +个内存池实际上决定了shuffle的粒度。所以，如果将这个内存池减小，又要保证数据是随机的， +那么最好将数据文件在每次读取之前做一次shuffle。可能的代码为 + +.. literalinclude:: src/reduce_min_pool_size.py + +这样做可以极大的减少内存占用，并且可能会加速训练过程，详细文档参考 :ref:`api_pydataprovider2` 。 + +神经元激活内存 +++++++++++++++ + +神经网络在训练的时候，会对每一个激活暂存一些数据，如神经元激活值等。 +在反向传递的时候，这些数据会被用来更新参数。这些数据使用的内存主要和两个参数有关系， +一是batch size，另一个是每条序列(Sequence)长度。所以，其实也是和每个mini-batch中包含 +的时间步信息成正比。 + +所以做法可以有两种： + +* 减小batch size。 即在网络配置中 :code:`settings(batch_size=1000)` 设置成一个小一些的值。但是batch size本身是神经网络的超参数，减小batch size可能会对训练结果产生影响。 +* 减小序列的长度，或者直接扔掉非常长的序列。比如，一个数据集大部分序列长度是100-200, + 但是突然有一个10000长的序列，就很容易导致内存超限，特别是在LSTM等RNN中。 + +参数内存 +++++++++ + +PaddlePaddle支持非常多的优化算法(Optimizer)，不同的优化算法需要使用不同大小的内存。 +例如使用 :code:`adadelta` 算法，则需要使用等于权重参数规模大约5倍的内存。举例，如果参数保存下来的模型目录 +文件为 :code:`100M`， 那么该优化算法至少需要 :code:`500M` 的内存。 + +可以考虑使用一些优化算法，例如 :code:`momentum`。 + +2. 如何加速训练速度 +------------------- + +加速PaddlePaddle训练可以考虑从以下几个方面\： + +* 减少数据载入的耗时 +* 加速训练速度 +* 利用分布式训练驾驭更多的计算资源 + +减少数据载入的耗时 +++++++++++++++++++ + +使用\ :code:`pydataprovider`\ 时，可以减少缓存池的大小，同时设置内存缓存功能，即可以极大的加速数据载入流程。 +:code:`DataProvider` 缓存池的减小，和之前减小通过减小缓存池来减小内存占用的原理一致。 + +.. literalinclude:: src/reduce_min_pool_size.py + +同时 :code:`@provider` 接口有一个 :code:`cache` 参数来控制缓存方法，将其设置成 :code:`CacheType.CACHE_PASS_IN_MEM` 的话，会将第一个 :code:`pass` (过完所有训练数据即为一个pass)生成的数据缓存在内存里，在之后的 :code:`pass` 中，不会再从 :code:`python` 端读取数据，而是直接从内存的缓存里读取数据。这也会极大减少数据读入的耗时。 + + +加速训练速度 +++++++++++++ + +PaddlePaddle支持Sparse的训练，sparse训练需要训练特征是 :code:`sparse_binary_vector` 、 :code:`sparse_vector` 、或者 :code:`integer_value` 的任一一种。同时，与这个训练数据交互的Layer，需要将其Parameter设置成 sparse 更新模式，即设置 :code:`sparse_update=True` + +这里使用简单的 :code:`word2vec` 训练语言模型距离，具体使用方法为\: + +使用一个词前两个词和后两个词，来预测这个中间的词。这个任务的DataProvider为\: + +.. literalinclude:: src/word2vec_dataprovider.py + +这个任务的配置为\: + +.. literalinclude:: src/word2vec_config.py + + +利用更多的计算资源 +++++++++++++++++++ + +利用更多的计算资源可以分为一下几个方式来进行\: + +* 单机CPU训练 + + * 使用多线程训练。设置命令行参数 :code:`trainer_count`。 + +* 单机GPU训练 + + * 使用显卡训练。设置命令行参数 :code:`use_gpu`。 + * 使用多块显卡训练。设置命令行参数 :code:`use_gpu` 和 :code:`trainer_count` 。 + +* 多机训练 + + * 请参考 :ref:`cluster_train` 。 + +3. 如何指定GPU设备 +------------------ + +例如机器上有4块GPU，编号从0开始，指定使用2、3号GPU： + +* 方式1：通过 `CUDA_VISIBLE_DEVICES `_ 环境变量来指定特定的GPU。 + +.. code-block:: bash + + env CUDA_VISIBLE_DEVICES=2,3 paddle train --use_gpu=true --trainer_count=2 + +* 方式2：通过命令行参数 ``--gpu_id`` 指定。 + +.. code-block:: bash + + paddle train --use_gpu=true --trainer_count=2 --gpu_id=2 + + +4. 训练过程中出现 :code:`Floating point exception`, 训练因此退出怎么办? +------------------------------------------------------------------------ + +Paddle二进制在运行时捕获了浮点数异常，只要出现浮点数异常(即训练过程中出现NaN或者Inf)，立刻退出。浮点异常通常的原因是浮点数溢出、除零等问题。 +主要原因包括两个方面: + +* 训练过程中参数或者训练过程中的梯度尺度过大，导致参数累加，乘除等时候，导致了浮点数溢出。 +* 模型一直不收敛，发散到了一个数值特别大的地方。 +* 训练数据有问题，导致参数收敛到了一些奇异的情况。或者输入数据尺度过大，有些特征的取值达到数百万，这时进行矩阵乘法运算就可能导致浮点数溢出。 + +这里有两种有效的解决方法： + +1. 设置 :code:`gradient_clipping_threshold` 参数，示例代码如下： + +.. code-block:: python + +optimizer = paddle.optimizer.RMSProp( + learning_rate=1e-3, + gradient_clipping_threshold=10.0, + regularization=paddle.optimizer.L2Regularization(rate=8e-4)) + +具体可以参考 `nmt_without_attention `_ 示例。 + +2. 设置 :code:`error_clipping_threshold` 参数，示例代码如下： + +.. code-block:: python + +decoder_inputs = paddle.layer.fc( + act=paddle.activation.Linear(), + size=decoder_size * 3, + bias_attr=False, + input=[context, current_word], + layer_attr=paddle.attr.ExtraLayerAttribute( + error_clipping_threshold=100.0)) + +完整代码可以参考示例 `machine translation `_ 。 + +两种方法的区别： + +1. 两者都是对梯度的截断，但截断时机不同，前者在 :code:`optimzier` 更新网络参数时应用；后者在激活函数反向计算时被调用； +2. 截断对象不同：前者截断可学习参数的梯度，后者截断回传给前层的梯度; + +除此之外，还可以通过减小学习律或者对数据进行归一化处理来解决这类问题。 + +5. 如何调用 infer 接口输出多个layer的预测结果 +----------------------------------------------- + +* 将需要输出的层作为 :code:`paddle.inference.Inference()` 接口的 :code:`output_layer` 参数输入，代码如下： + +.. code-block:: python + + inferer = paddle.inference.Inference(output_layer=[layer1, layer2], parameters=parameters) + +* 指定要输出的字段进行输出。以输出 :code:`value` 字段为例，代码如下： + +.. code-block:: python + + out = inferer.infer(input=data_batch, field=["value"]) + +需要注意的是： + +* 如果指定了2个layer作为输出层，实际上需要的输出结果是两个矩阵； +* 假设第一个layer的输出A是一个 N1 * M1 的矩阵，第二个 Layer 的输出B是一个 N2 * M2 的矩阵； +* paddle.v2 默认会将A和B 横向拼接，当N1 和 N2 大小不一样时，会报如下的错误： + +.. code-block:: python + + ValueError: all the input array dimensions except for the concatenation axis must match exactly + +多个层的输出矩阵的高度不一致导致拼接失败，这种情况常常发生在： + +* 同时输出序列层和非序列层； +* 多个输出层处理多个不同长度的序列; + +此时可以在调用infer接口时通过设置 :code:`flatten_result=False` , 跳过“拼接”步骤，来解决上面的问题。这时，infer接口的返回值是一个python list: + +* list 中元素的个数等于网络中输出层的个数； +* list 中每个元素是一个layer的输出结果矩阵，类型是numpy的ndarray； +* 每一个layer输出矩阵的高度，在非序列输入时：等于样本数；序列输入时等于：输入序列中元素的总数；宽度等于配置中layer的size； diff --git a/doc/faq/src/reduce_min_pool_size.py b/doc/faq/local/src/reduce_min_pool_size.py similarity index 100% rename from doc/faq/src/reduce_min_pool_size.py rename to doc/faq/local/src/reduce_min_pool_size.py diff --git a/doc/faq/src/word2vec_config.py b/doc/faq/local/src/word2vec_config.py similarity index 100% rename from doc/faq/src/word2vec_config.py rename to doc/faq/local/src/word2vec_config.py diff --git a/doc/faq/src/word2vec_dataprovider.py b/doc/faq/local/src/word2vec_dataprovider.py similarity index 100% rename from doc/faq/src/word2vec_dataprovider.py rename to doc/faq/local/src/word2vec_dataprovider.py diff --git a/doc/faq/model/index_cn.rst b/doc/faq/model/index_cn.rst new file mode 100644 index 0000000000..b47bbe05bd --- /dev/null +++ b/doc/faq/model/index_cn.rst @@ -0,0 +1,69 @@ +######### +模型配置 +######### + +.. contents:: + +1. 出现 :code:`Duplicated layer name` 错误怎么办 +-------------------------------------------------- + +出现该错误的原因一般是用户对不同layer的参数 :code:`name` 设置了相同的取值。遇到该错误时，先找出参数 :code:`name` 取值相同的layer，然后将这些layer的参数 :code:`name` 设置为不同的值。 + +2. :code:`paddle.layer.memory` 的参数 :code:`name` 如何使用 +------------------------------------------------------------- + +* :code:`paddle.layer.memory` 用于获取特定layer上一时间步的输出，该layer是通过参数 :code:`name` 指定，即，:code:`paddle.layer.memory` 会关联参数 :code:`name` 取值相同的layer，并将该layer上一时间步的输出作为自身当前时间步的输出。 + +* PaddlePaddle的所有layer都有唯一的name，用户通过参数 :code:`name` 设定，当用户没有显式设定时，PaddlePaddle会自动设定。而 :code:`paddle.layer.memory` 不是真正的layer，其name由参数 :code:`memory_name` 设定，当用户没有显式设定时，PaddlePaddle会自动设定。:code:`paddle.layer.memory` 的参数 :code:`name` 用于指定其要关联的layer，需要用户显式设定。 + +3. 两种使用 drop_out 的方法有何区别 +------------------------------------ + +* 在PaddlePaddle中使用dropout有两种方式 + + * 在相应layer的 :code:`layer_atter` 设置 :code:`drop_rate`，以 :code:`paddle.layer.fc` 为例，代码如下： + + .. code-block:: python + + fc = paddle.layer.fc(input=input, layer_attr=paddle.attr.ExtraLayerAttribute(drop_rate=0.5)) + + * 使用 :code:`paddle.layer.dropout`，以 :code:`paddle.layer.fc` 为例，代码如下： + + .. code-block:: python + + fc = paddle.layer.fc(input=input) + drop_fc = paddle.layer.dropout(input=fc, dropout_rate=0.5) + +* :code:`paddle.layer.dropout` 实际上使用了 :code:`paddle.layer.add_to`，并在该layer里采用第一种方式设置 :code:`drop_rate` 来使用dropout的。这种方式对内存消耗较大。 + +* PaddlePaddle在激活函数里实现dropout，而不是在layer里实现。 + +* :code:`paddle.layer.lstmemory`、:code:`paddle.layer.grumemory`、:code:`paddle.layer.recurrent` 不是通过一般的方式来实现对输出的激活，所以不能采用第一种方式在这几个layer里设置 :code:`drop_rate` 来使用dropout。若要对这几个layer使用dropout，可采用第二种方式，即使用 :code:`paddle.layer.dropout`。 + +4. 不同的 recurrent layer 的区别 +---------------------------------- +以LSTM为例，在PaddlePaddle中包含以下 recurrent layer： + +* :code:`paddle.layer.lstmemory` +* :code:`paddle.networks.simple_lstm` +* :code:`paddle.networks.lstmemory_group` +* :code:`paddle.networks.bidirectional_lstm` + +按照具体实现方式可以归纳为2类： + +1. 由 recurrent_group 实现的 recurrent layer： + + * 用户在使用这一类recurrent layer时，可以访问由recurrent unit在一个时间步内计算得到的中间值（例如：hidden states, memory cells等）； + * 上述的 :code:`paddle.networks.lstmemory_group` 是这一类的 recurrent layer ； + +2. 将recurrent layer作为一个整体来实现： + + * 用户在使用这一类recurrent layer，只能访问它们的输出值； + * 上述的 :code:`paddle.networks.lstmemory_group` 、 :code:`paddle.networks.simple_lstm` 和 :code:`paddle.networks.bidirectional_lstm` 属于这一类的实现； + +将recurrent layer作为一个整体来实现， 能够针对CPU和GPU的计算做更多优化， 所以相比于recurrent group的实现方式， 第二类 recurrent layer 计算效率更高。 在实际应用中，如果用户不需要访问LSTM的中间变量，而只需要获得recurrent layer计算的输出，我们建议使用第二类实现。 + +此外，关于LSTM, PaddlePaddle中还包含 :code:`paddle.networks.lstmemory_unit` 这一计算单元： + + * 不同于上述介绍的recurrent layer , :code:`paddle.networks.lstmemory_unit` 定义了LSTM单元在一个时间步内的计算过程，它并不是一个完整的recurrent layer，也不能接收序列数据作为输入； + * :code:`paddle.networks.lstmemory_unit` 只能在recurrent_group中作为step function使用； diff --git a/doc/faq/parameter/index_cn.rst b/doc/faq/parameter/index_cn.rst new file mode 100644 index 0000000000..c721b62318 --- /dev/null +++ b/doc/faq/parameter/index_cn.rst @@ -0,0 +1,201 @@ +######### +参数设置 +######### + +.. contents:: + +1. 如何选择SGD算法的学习率 +-------------------------- + +在采用sgd/async_sgd进行训练时，一个重要的问题是选择正确的learning_rate。如果learning_rate太大，那么训练有可能不收敛，如果learning_rate太小，那么收敛可能很慢，导致训练时间过长。 + +通常做法是从一个比较大的learning_rate开始试，如果不收敛，那减少学习率10倍继续试验，直到训练收敛为止。那么如何判断训练不收敛呢？可以估计出如果模型采用不变的输出最小的cost0是多少。 + +如果训练过程的的cost明显高于这个常数输出的cost，那么我们可以判断为训练不收敛。举一个例子，假如我们是三分类问题，采用multi-class-cross-entropy作为cost，数据中0,1,2三类的比例为 :code:`0.2, 0.5, 0.3` , 那么常数输出所能达到的最小cost是 :code:`-(0.2*log(0.2)+0.5*log(0.5)+0.3*log(0.3))=1.03` 。如果训练一个pass（或者更早）后，cost还大于这个数，那么可以认为训练不收敛，应该降低学习率。 + +2. 如何设置学习率退火（learning rate annealing） +------------------------------------------------ + +在相应的优化算法里设置learning_rate_schedule及相关参数，以使用Adam算法为例，代码如下： + +.. code-block:: python + + optimizer = paddle.optimizer.Adam( + learning_rate=1e-3, + learning_rate_decay_a=0.5, + learning_rate_decay_b=0.75, + learning_rate_schedule="poly",) + +PaddlePaddle目前支持8种learning_rate_schedule，这8种learning_rate_schedule及其对应学习率计算方式如下： + +* "constant" + + lr = learning_rate + +* "poly" + + lr = learning_rate * pow(1 + learning_rate_decay_a * num_samples_processed, -learning_rate_decay_b) + + 其中，num_samples_processed为已训练样本数，下同。 + +* "caffe_poly" + + lr = learning_rate * pow(1.0 - num_samples_processed / learning_rate_decay_a, learning_rate_decay_b) + +* "exp" + + lr = learning_rate * pow(learning_rate_decay_a, num_samples_processed / learning_rate_decay_b) + +* "discexp" + + lr = learning_rate * pow(learning_rate_decay_a, floor(num_samples_processed / learning_rate_decay_b)) + +* "linear" + + lr = max(learning_rate - learning_rate_decay_a * num_samples_processed, learning_rate_decay_b) + +* "manual" + + 这是一种按已训练样本数分段取值的学习率退火方法。使用该learning_rate_schedule时，用户通过参数 :code:`learning_rate_args` 设置学习率衰减因子分段函数，当前的学习率为所设置 :code:`learning_rate` 与当前的衰减因子的乘积。以使用Adam算法为例，代码如下： + + .. code-block:: python + + optimizer = paddle.optimizer.Adam( + learning_rate=1e-3, + learning_rate_schedule="manual", + learning_rate_args="1000:1.0,2000:0.9,3000:0.8",) + + 在该示例中，当已训练样本数小于等于1000时，学习率为 :code:`1e-3 * 1.0`；当已训练样本数大于1000小于等于2000时，学习率为 :code:`1e-3 * 0.9`；当已训练样本数大于2000时，学习率为 :code:`1e-3 * 0.8`。 + +* "pass_manual" + + 这是一种按已训练pass数分段取值的学习率退火方法。使用该learning_rate_schedule时，用户通过参数 :code:`learning_rate_args` 设置学习率衰减因子分段函数，当前的学习率为所设置 :code:`learning_rate` 与当前的衰减因子的乘积。以使用Adam算法为例，代码如下： + + .. code-block:: python + + optimizer = paddle.optimizer.Adam( + learning_rate=1e-3, + learning_rate_schedule="manual", + learning_rate_args="1:1.0,2:0.9,3:0.8",) + + 在该示例中，当已训练pass数小于等于1时，学习率为 :code:`1e-3 * 1.0`；当已训练pass数大于1小于等于2时，学习率为 :code:`1e-3 * 0.9`；当已训练pass数大于2时，学习率为 :code:`1e-3 * 0.8`。 + +3. 如何初始化参数 +----------------- + +默认情况下，PaddlePaddle使用均值0，标准差为 :math:`\frac{1}{\sqrt{d}}` 来初始化参数。其中 :math:`d` 为参数矩阵的宽度。这种初始化方式在一般情况下不会产生很差的结果。如果用户想要自定义初始化方式，PaddlePaddle目前提供两种参数初始化的方式\: + +* 高斯分布。将 :code:`param_attr` 设置成 :code:`param_attr=ParamAttr(initial_mean=0.0, initial_std=1.0)` +* 均匀分布。将 :code:`param_attr` 设置成 :code:`param_attr=ParamAttr(initial_max=1.0, initial_min=-1.0)` + +比如设置一个全连接层的参数初始化方式和bias初始化方式，可以使用如下代码。 + +.. code-block:: python + + hidden = fc_layer(input=ipt, param_attr=ParamAttr(initial_max=1.0, initial_min=-1.0), + bias_attr=ParamAttr(initial_mean=1.0, initial_std=0.0)) + +上述代码将bias全部初始化为1.0, 同时将参数初始化为 :code:`[1.0, -1.0]` 的均匀分布。 + +4. 如何共享参数 +--------------- + +PaddlePaddle的参数使用名字 :code:`name` 作为参数的ID，相同名字的参数，会共享参数。设置参数的名字，可以使用 :code:`ParamAttr(name="YOUR_PARAM_NAME")` 来设置。更方便的设置方式，是使得要共享的参数使用同样的 :code:`ParamAttr` 对象。 + +简单的全连接网络，参数共享的配置示例为\: + +.. literalinclude:: ../../python/paddle/trainer_config_helpers/tests/configs/shared_fc.py + +这里 :code:`hidden_a` 和 :code:`hidden_b` 使用了同样的parameter和bias。并且softmax层的两个输入也使用了同样的参数 :code:`softmax_param`。 + +5. 如何加载预训练参数 +------------------------ + +* 对加载预训练参数的层，设置其参数属性 :code:`is_static=True`，使该层的参数在训练过程中保持不变。以embedding层为例，代码如下： + +.. code-block:: python + + emb_para = paddle.attr.Param(name='emb', is_static=True) + paddle.layer.embedding(size=word_dim, input=x, param_attr=emb_para) + + +* 从模型文件将预训练参数载入 :code:`numpy.array`，在创建parameters后，使用 :code:`parameters.set()` 加载预训练参数。PaddlePaddle保存的模型参数文件前16字节为头信息，用户将参数载入 :code:`numpy.array` 时须从第17字节开始。以embedding层为例，代码如下： + +.. code-block:: python + + def load_parameter(file_name, h, w): + with open(file_name, 'rb') as f: + f.read(16) # skip header. + return np.fromfile(f, dtype=np.float32).reshape(h, w) + + parameters = paddle.parameters.create(my_cost) + parameters.set('emb', load_parameter(emb_param_file, 30000, 256)) + +6. 存储的参数格式是什么，如何和明文进行相互转化 +-------------------------------------------------- + +PaddlePaddle保存的模型参数文件内容由16字节头信息和网络参数两部分组成。头信息中，1~4字节表示PaddlePaddle版本信息，请直接填充0；5~8字节表示每个参数占用的字节数，当保存的网络参数为float类型时为4，double类型时为8；9~16字节表示保存的参数总个数。 + +将PaddlePaddle保存的模型参数还原回明文时，可以使用相应数据类型的 :code:`numpy.array` 加载具体网络参数，此时可以跳过PaddlePaddle模型参数文件的头信息。若在PaddlePaddle编译时，未指定按照double精度编译，默认情况下按照float精度计算，保存的参数也是float类型。这时在使用 :code:`numpy.array` 时，一般设置 :code:`dtype=float32` 。示例如下： + +.. code-block:: python + + def read_parameter(fname, width): + s = open(fname).read() + # skip header + vec = np.fromstring(s[16:], dtype=np.float32) + # width is the size of the corresponding layer + np.savetxt(fname + ".csv", vec.reshape(width, -1), + fmt="%.6f", delimiter=",") + + +将明文参数转化为PaddlePaddle可加载的模型参数时，首先构造头信息，再写入网络参数。下面的代码将随机生成的矩阵转化为可以被PaddlePaddle加载的模型参数。 + +.. code-block:: python + + def gen_rand_param(param_file, width, height, need_trans): + np.random.seed() + header = struct.pack("iil", 0, 4, height * width) + param = np.float32(np.random.rand(height, width)) + with open(param_file, "w") as fparam: + fparam.write(header + param.tostring()) + +7. A protocol message was rejected because it was too big +------------------------------------------------------------ + +如果在训练NLP相关模型时，出现以下错误： + +.. code-block:: bash + + [libprotobuf ERROR google/protobuf/io/coded_stream.cc:171] A protocol message was rejected because it was too big (more than 67108864 bytes). To increase the limit (or to disable these warnings), see CodedInputStream::SetTotalBytesLimit() in google/protobuf/io/coded_stream.h. + F1205 14:59:50.295174 14703 TrainerConfigHelper.cpp:59] Check failed: m->conf.ParseFromString(configProtoStr) + +可能的原因是：传给dataprovider的某一个args过大，一般是由于直接传递大字典导致的。错误的define_py_data_sources2类似： + +.. code-block:: python + + src_dict = dict() + for line_count, line in enumerate(open(src_dict_path, "r")): + src_dict[line.strip()] = line_count + + define_py_data_sources2( + train_list, + test_list, + module="dataprovider", + obj="process", + args={"src_dict": src_dict}) + +解决方案是：将字典的地址作为args传给dataprovider，然后在dataprovider里面根据该地址加载字典。即define_py_data_sources2应改为： + +.. code-block:: python + + define_py_data_sources2( + train_list, + test_list, + module="dataprovider", + obj="process", + args={"src_dict_path": src_dict_path}) + +完整源码可参考 `seqToseq `_ 示例。 + + diff --git a/doc/getstarted/build_and_install/docker_install_cn.rst b/doc/getstarted/build_and_install/docker_install_cn.rst index 84e3317774..30b144d849 100644 --- a/doc/getstarted/build_and_install/docker_install_cn.rst +++ b/doc/getstarted/build_and_install/docker_install_cn.rst @@ -20,7 +20,7 @@ Docker使用入门 docker pull paddlepaddle/paddle:0.10.0 - 来下载Docker镜像，paddlepaddle/paddle是从官方镜像源Dockerhub.com下载的，推荐国内用户使用ocker.paddlepaddle.org/paddle下载。 + 来下载Docker镜像，paddlepaddle/paddle是从官方镜像源Dockerhub.com下载的，推荐国内用户使用docker.paddlepaddle.org/paddle下载。 - *容器*： 如果说一个Docker镜像就是一个程序，那容器就是这个程序运行时产生的“进程”。 实际上，一个容器就是一个操作系统的进程，但是是运行在独立的进程空间，文件系统以及网络之上。 diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md index c6570b89ae..c823d7e9fc 100644 --- a/doc/howto/dev/new_op_cn.md +++ b/doc/howto/dev/new_op_cn.md @@ -54,9 +54,9 @@ class MulOpMaker : public framework::OpProtoAndCheckerMaker { public: MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The first input of mul op"); - AddInput("Y", "The second input of mul op"); - AddOutput("Out", "The output of mul op"); + AddInput("X", "(Tensor), 2D tensor of size (M x K)"); + AddInput("Y", "(Tensor), 2D tensor of size (K x N)"); + AddOutput("Out", "(Tensor), 2D tensor of size (M x N)"); AddComment(R"DOC( Two Element Mul Operator. The equation is: Out = X * Y @@ -72,7 +72,7 @@ The equation is: Out = X * Y 构造函数里通过`AddInput`添加输入参数，通过`AddOutput`添加输出参数，通过`AddComment`添加Op的注释。这些函数会将对应内容添加到`OpProto`中。 -上面的代码在`MulOp`中添加两个输入`X`和`Y`，添加了一个输出`Out`，并解释了各自含义，命名请遵守命名规范。 +上面的代码在`MulOp`中添加两个输入`X`和`Y`，添加了一个输出`Out`，并解释了各自含义，命名请遵守[命名规范](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/name_convention.md)。 再以[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)为例： @@ -206,7 +206,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs, - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`。 - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。 - - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。 + - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulGradKernel`类。 - 在 `.cu`文件中注册GPU Kernel。 @@ -285,41 +285,27 @@ class TestMulGradOp(GradientChecker): 'Y': np.random.random((84, 100)).astype("float32") } - def test_cpu_gpu_compare(self): - self.compare_grad(self.op, self.inputs) - - def test_normal(self): + def test_check_grad_normal(self): # mul op will enlarge the relative error - self.check_grad( - self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5) + self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5) - def test_ignore_x(self): + def test_check_grad_ingore_x(self): self.check_grad( - self.op, - self.inputs, ["Y"], - "Out", - max_relative_error=0.5, - no_grad_set={"X"}) + ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X")) - def test_ignore_y(self): + def test_check_grad_ingore_y(self): self.check_grad( - self.op, - self.inputs, ["X"], - "Out", - max_relative_error=0.5, - no_grad_set={"Y"}) + ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) ``` 下面解释代码中一些关键的地方: - 调用`create_op("mul")`创建反向Op对应的前向Op。 -- 调用`compare_grad`函数对比CPU、GPU计算结果。 -- `test_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。 - - 第一个参数`self.op` : 前向Op。 - - 第二个参数`self.inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。 - - 第三个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。 - - 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out` -- `test_ignore_x`和`test_ignore_y`分支用来测试只需要计算一个输入梯度的情况。 +- `test_check_grad_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。 + - 第一个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。 + - 第二个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`。 + - 第三个参数`max_relative_error`：指定检测梯度时能容忍的最大错误值。 +- `test_check_grad_ingore_x`和`test_check_grad_ingore_y`分支用来测试只需要计算一个输入梯度的情况。 ### 编译和执行单元测试 diff --git a/doc/howto/dev/new_op_en.md b/doc/howto/dev/new_op_en.md new file mode 100644 index 0000000000..1e88e1f5b4 --- /dev/null +++ b/doc/howto/dev/new_op_en.md @@ -0,0 +1,342 @@ +# How to write a new operator + + - [Background](#background) + - [Implementing C++ Types](#implementing-c++-types) + - [Defining ProtoMaker](#defining-protoMaker) + - [Defining Operator](#defining-operator) + - [Registering Operator](#registering-operator) + - [Compilation](#compilation) + - [Python Binding](#python-binding) + - [Unit Tests](#unit-tests) + - [Testing Forward Operators](#testing-forward-operators) + - [Testing Backward Operators](#testing-backward-operators) + - [Compiling and Running](#compiling-and-running) + - [Remarks](#remarks) +## Background + +Here are the base types needed. For details, please refer to the design docs. + +- `framework::OperatorBase`: Operator (Op)base class. +- `framework::OpKernel`: Base class for Op computation. +- `framework::OperatorWithKernel`: Inherited from OperatorBase, describing an operator with computation. +- `class OpProtoAndCheckerMaker`: Describes an Operator's input, output, attributes and description, mainly used to interface with Python API. + +An operator can be differentiated by whether in has kernel methods. An operator with kernel inherits from `OperatorWithKernel` while the ones without inherit from `OperatorBase`. This tutorial focuses on implementing operators with kernels. In short, an operator includes the following information: + + + Information | Where is it defined +-------------- | :---------------------- +OpProtoMake definition | `.cc`files, Backward Op does not need an OpProtoMake interface. +Op definition | `.cc` files +Kernel implementation | The kernel methods shared between CPU and GPU are defined in `.h` files. CPU-specific kernels live in `.cc` files, while GPU-specific kernels are implemented in `.cu`files. +Registering the Op | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the GPU implementation. + + +New Operator implementations are added to the list [paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators), with file names in the format `*_op.h` (if applicable), `*_op.cc`, `*_op.cu` (if applicable).** The system will use the naming scheme to automatically build operators and their corresponding Python extensions. ** + + +Let's take matrix multiplication operator, [MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc), as an example to introduce the writing of an Operator with Kernel. + + +## Implementing C++ Types + + +### 1. Defining Class ProtoMaker + +Matrix Multiplication can be written as $Out = X * Y$, meaning that the operation consists of two inputs and pne output. + +First, define `ProtoMaker` to describe the Operator's input, output, and additional comments: + +```cpp +class MulOpMaker : public framework::OpProtoAndCheckerMaker { + public: + MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "(Tensor), 2D tensor of size (M x K)"); + AddInput("Y", "(Tensor), 2D tensor of size (K x N)"); + AddOutput("Out", "(Tensor), 2D tensor of size (M x N)"); + AddComment(R"DOC( +Two Element Mul Operator. +The equation is: Out = X * Y +)DOC"); + } +}; +``` + +[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L43)is inherited from`framework::OpProtoAndCheckerMaker`, consisting of 2 variables in the constructor： + + - `framework::OpProto` stores Operator input and variable attribute, used for generating Python API interfaces. + - `framework::OpAttrChecker` is used to validate variable attributes. + +The constructor utilizes `AddInput`, `AddOutput`, and `AddComment`, so that the corresponding information will be added to `OpProto`. + +The code above adds two inputs `X` and `Y` to `MulOp`, an output `Out`, and their corresponding descriptions, in accordance to Paddle's [naming convention](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/name_convention.md). + + +An additional example [`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37) is implemented as follows: + +```cpp +template +class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The input tensor of scale operator.").NotInGradient(); + AddOutput("Out", "The output tensor of scale operator.").NotInGradient(); + AddComment(R"DOC(Scale operator +The equation is: Out = scale*X +)DOC"); + AddAttr("scale", "scale of scale operator.").SetDefault(1.0); + } +}; +``` + +There are two changes in this example: + +- `AddInput("X","...").NotInGradient()` expresses that input `X` is not involved in `ScaleOp`'s corresponding computation. If an input to an operator is not participating in back-propagation, please explicitly set `.NotInGradient()`. + +- `AddAttr("scale", "...").SetDefault(1.0);` adds `scale`constant as an attribute, and sets the default value to 1.0. + + +### 2. Defining Operator + +The following code defines the interface for MulOp: + +```cpp +class MulOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + auto dim0 = ctx.Input("X")->dims(); + auto dim1 = ctx.Input("Y")->dims(); + PADDLE_ENFORCE_EQ(dim0.size(), 2, + "input X(%s) should be a tensor with 2 dims, a matrix", + ctx.op_.Input("X")); + PADDLE_ENFORCE_EQ(dim1.size(), 2, + "input Y(%s) should be a tensor with 2 dims, a matrix", + ctx.op_.Input("Y")); + PADDLE_ENFORCE_EQ( + dim0[1], dim1[0], + "First matrix's width must be equal with second matrix's height."); + ctx.Output("Out")->Resize({dim0[0], dim1[1]}); + } +}; +``` + +[`MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L22) is inherited from `OperatorWithKernel`. Its `public` member + +```cpp +using framework::OperatorWithKernel::OperatorWithKernel; +``` + +expresses an operator constructor using base class `OperatorWithKernel`, alternatively written as + +```cpp +MulOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} +``` + +`InferShape` interface needs to be re-written.`InferShape` is a constant method and cannot modify Op's member variables, its constant member `const framework::InferShapeContext &ctx` can be used to extract input, output, and attributes. It functions to + + - 1). validate and error out early: it checks input data dimensions and types. + - 2). configures the tensor shape in the output. + +Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, which also include the registration methods introduced later. + +### 3. Defining OpKernel + +`MulKernel` inherits `framework::OpKernel`, which includes the following templates: + +- `typename Place` denotes device type. When different devices, namely the CPU and the GPU, share the same kernel, this template needs to be added. If they don't share kernels, this must not be added. An example of a non-sharing kernel is [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43). + +- `typename T` denotes data type, such as `float` or `double`. + +`MulKernel` types need to rewrite the interface for `Compute`. +- `Compute` takes one input variable `const framework::ExecutionContext& context`. +- Compared with `InferShapeContext`, `ExecutionContext` includes device types, and can similarly extract input, output, and attribute variables. +- `Compute` implements the computation logics of an `OpKernel`. + +`MulKernel`'s implementation of `Compute` is as follows: + + ```cpp + template + class MulKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* X = context.Input("X"); + auto* Y = context.Input("Y"); + auto* Z = context.Output("Out"); + Z->mutable_data(context.GetPlace()); + auto* device_context = + const_cast(context.device_context_); + math::matmul(*X, false, *Y, false, 1, Z, 0, device_context); + } + }; + ``` + +Note that **different devices (CPU, GPU)share an Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions that support both devices.** + +`MulOp`'s CPU and GPU share the same `Kernel`. A non-sharing `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43). + +To ease the writing of `OpKernel` compute, and for reusing code cross-device, [`Eigen-unsupported Tensor`](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md?fileviewer=file-view-default) module is used to implement `Compute` interface. To learn about how the Eigen library is used in PaddlePaddle, please see [usage document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md). + + +This concludes the forward implementation of an operator. Next its operation and kernel need to be registered in a `.cc` file. + +The definition of its corresponding backward operator, if applicable, is similar to that of an forward operator. **Note that a backward operator does not include a `ProtoMaker`**. + +### 4. Registering Operator + +- In `.cc` files, register forward and backward operator classes and the CPU kernel. + + ```cpp + namespace ops = paddle::operators; + REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad); + REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); + REGISTER_OP_CPU_KERNEL(mul_grad, + ops::MulGradKernel); + ``` + + In that code block, + + - `REGISTER_OP` registers the `ops::MulOp` class, type named `mul`, its type `ProtoMaker` is `ops::MulOpMaker`, registering `ops::MulOpGrad` as `mul_grad`. + - `REGISTER_OP_WITHOUT_GRADIENT` registers an operator without gradient. + - `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulGradKernel`. + + +- Registering GPU Kernel in `.cu` files + - Note that if GPU Kernel is implemented using the `Eigen unsupported` module, then on top of `.cu`, a macro definition `#define EIGEN_USE_GPU` is needed, such as + + ```cpp + // if use Eigen unsupported module before include head files + #define EIGEN_USE_GPU + + namespace ops = paddle::operators; + REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel); + REGISTER_OP_GPU_KERNEL(mul_grad, + ops::MulGradKernel); + ``` + +### 5. Compilation + +Run the following commands to compile. + +``` +make mul_op +``` + +## Python Binding + +The system will automatically bind to Python and link it to a generated library. + +## Unit Tests + +Unit tests for an operator include + +1. comparing a forward operator's implementations on different devices, + +2. comparing a backward operator's implementation on different devices, and + +3. a scaling test for the backward operator. + +Here, we introduce the [unit tests for `MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py). + +### Testing Forward Operators + +A forward operator unit test inherits `unittest.TestCase` and defines metaclass `__metaclass__ = OpTestMeta`. More concrete tests are performed in `OpTestMeta`. Testing a forward operator requires the following: + +1. Defining input, output and relevant attributes in `setUp` method. + +2. Generating random input data. + +3. Implementing the same computation logic in a Python script: + + ```python + import unittest + import numpy as np + from gradient_checker import GradientChecker, create_op + from op_test_util import OpTestMeta + + class TestMulOp(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "mul" + self.inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'Y': np.random.random((84, 100)).astype("float32") + } + self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} + ``` +Get its output, and compare it with the forward operator's own output. + +The code above first loads required packages. In addition, we have + +- `self.type = "mul" ` defines the type that is identical to what the operator's registered type. +- `self.inputs` defines input, with type `numpy.array` and initializes it. +- `self.outputs` defines output and completes the same operator computation in the Python script, and returns its result from the Python script. + +### Testing Backward Operators + +A backward operator unit test inherits `GradientChecker`, which inherits `unittest.TestCase`. As a result, **a backward operator unit test needs to be have the prefix `test_`**. + +```python +class TestMulGradOp(GradientChecker): + def setUp(self): + self.op = create_op("mul") + self.inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'Y': np.random.random((84, 100)).astype("float32") + } + + def test_check_grad_normal(self): + # mul op will enlarge the relative error + self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5) + + def test_check_grad_ingore_x(self): + self.check_grad( + ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X")) + + def test_check_grad_ingore_y(self): + self.check_grad( + ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) +``` + +Some key points in the code above include: + +- `create_op("mul")` creates the backward operator's corresponding forward operator. +- `test_normal` calls `check_grad` to validate scaling tests' correctness and stability through numeric methods. + - The first variable `["X", "Y"]` appoints `X` and `Y` to be scale tested. + - The second variable `"Out"` points to the network's final output target `Out`. + - The third variable `max_relative_error` points to the maximum relative tolerance error during scaling tests. +- `test_check_grad_ingore_x` and `test_check_grad_ingore_y`branches test the cases where there is only one scaling input. + +### Compiling and Running + + +Any new unit testing file of the format `test_*.py` added to the director `python/paddle/v2/framework/tests` is automatically added to the project to compile. + +Note that **unlike the compile test for Ops, running unit tests requires compiling the entire project** and requires compiling with flag `WITH_TESTING` on i.e. `cmake paddle_dir -DWITH_TESTING=ON`. + +After successfully compiling the project, run the following command to run unit tests: + +```bash +make test ARGS="-R test_mul_op -V" +``` + +Or, + +```bash +ctest -R test_mul_op +``` + +## Remarks + +- Every `*_op.h` (if applicable), `*_op.cc`, and `*_op.cu` (if applicable) must be created for a unique Op. Compiling will fail if multiple operators are included per file. +- The type with which an operator is registered needs to be identical to the Op's name. Registering `REGISTER_OP(B, ...)` in `A_op.cc` will cause unit testing failures. +- If the operator does not implement a GPU kernel, please refrain from creating an empty `*_op.cu` file, or else unit tests will fail. +- If multiple operators rely on some shared methods, a file NOT named `*_op.*` can be created to store them, such as `gather.h`. diff --git a/doc/howto/dev/use_eigen_en.md b/doc/howto/dev/use_eigen_en.md new file mode 100644 index 0000000000..e169106e12 --- /dev/null +++ b/doc/howto/dev/use_eigen_en.md @@ -0,0 +1,146 @@ +## How to use Eigen in Paddle + +Essentially, a neural network is a compute graph. T data needed for the computation is stored in `Tensor`s and its computation procedure is described by `Operator`s. An `Operator` calls the `Compute` interface in its corresponding `OpKernel` and operates on the `Tensor`. + + +### Eigen Tensor Module + +The Eigen Tensor module supports powerful element-wise computation. In addition, a piece of code written using it can be run on both the CPU and the GPU. + +Note that Eigen Tensor is still being actively developed, so its tests are not completely covered and its documentation may be sparse. + +For details on Eigen Tensor module, please see [doc 1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md) and [doc 2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md). + + +### paddle::framework::Tensor + +Paddle Tensor's is defined in the framework directory with the following interface: + +```cpp +class Tensor { + public: + /*! Return a pointer to mutable memory block. */ + template + inline T* data(); + + /** + * @brief Return a pointer to mutable memory block. + * @note If not exist, then allocation. + */ + template + inline T* mutable_data(platform::Place place); + + /** + * @brief Return a pointer to mutable memory block. + * + * @param[in] dims The dimensions of the memory block. + * @param[in] place The place of the memory block. + * + * @note If not exist, then allocation. + */ + template + inline T* mutable_data(DDim dims, platform::Place place); + + /*! Resize the dimensions of the memory block. */ + inline Tensor& Resize(const DDim& dims); + + /*! Return the dimensions of the memory block. */ + inline const DDim& dims() const; + + private: + /*! holds the memory block if allocated. */ + std::shared_ptr holder_; + + /*! points to dimensions of memory block. */ + DDim dim_; +}; +``` + +`Placeholder` is used to delay memory allocation; that is, we can first define a tensor, using `Resize` to configure its shape, and then call `mutuable_data` to allocate the actual memory. + +```cpp +paddle::framework::Tensor t; +paddle::platform::CPUPlace place; +// set size first +t.Resize({2, 3}); +// allocate memory on CPU later +t.mutable_data(place); +``` + +### paddle::framework::Tensor Usage +`AddOp` demonstrates Tensor's usage. + +- InferShape + +When computing a neural network's compute graph, first call every `Operator`'s `InferShape` method, and use `Resize` to configure the size of the output tensor. + +```cpp +void InferShape(const framework::InferShapeContext &ctx) const override { + PADDLE_ENFORCE_EQ(ctx.Input("X")->dims(), + ctx.Input("Y")->dims(), + "Two input of Add Op's dimension must be same."); + ctx.Output("Out")->Resize(ctx.Input("X")->dims()); +} +``` + + +- Run + +```cpp +void Compute(const framework::ExecutionContext& context) const override { + auto* input0 = context.Input("X"); + auto* input1 = context.Input("Y"); + auto* output = context.Output("Out"); + + output->mutable_data(context.GetPlace()); + + auto x = EigenVector::Flatten(*input0); + auto y = EigenVector::Flatten(*input1); + auto z = EigenVector::Flatten(*output); + + auto place = context.GetEigenDevice(); + + z.device(place) = x + y; +} +``` + + +### paddle::framework::Tensor到EigenTensor的转换 + +As shown above, in actual computation, we need to transform the input and output `Tensor`s into formats Eigen supports. We show some functions in [eigen.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen.h) to implement the transformation from `paddle::framework::Tensor`to `EigenTensor/EigenMatrix/EigenVector/EigenScalar`. + +Using EigenTensor as an example: + +```cpp +Tensor t; +float* p = t.mutable_data(make_ddim({1, 2, 3}), platform::CPUPlace()); +for (int i = 0; i < 1 * 2 * 3; i++) { + p[i] = static_cast(i); +} + +EigenTensor::Type et = EigenTensor::From(t); +``` + +`From` is an interfacing method provided by the EigenTensor template, which implements the transformation from a `paddle::framework::Tensor` object to an EigenTensor. Since `rank` is a template parameter, it needs to be explicitly specified at the time of the transformation. + +In Eigen, tensors with different ranks are different types, with `Vector` bring a rank-1 instance. Note that `EigenVector::From` uses a transformation from an 1-dimensional Paddle tensor to a 1-dimensional Eigen tensor while `EigenVector::Flatten` reshapes a paddle tensor and flattens it into a 1-dimensional Eigen tensor. Both resulting tensors are still typed EigenVector. + +For more transformations, see the [unit tests](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen_test.cc) in the `eigen_test.cc` file. + + + +### Implementing Computation + +While computing, the device interface is needed from the EigenTensors on the left hand side of the assignments. Note that the computation between EigenTensors only changes the data originally inthe Tensor and does not change all the shape information associated with the Tensor. + +```cpp +auto x = EigenVector::Flatten(*input0); +auto y = EigenVector::Flatten(*input1); +auto z = EigenVector::Flatten(*output); +auto place = context.GetEigenDevice(); +z.device(place) = x + y; +``` + +In this code segment, input0/input1/output can be Tensors of arbitrary dimension. We are calling Flatten from EigenVector, transforming a tensor of any dimension into a 1-dimensional EigenVector. After completing computation, input0/input1/output will retain the same shape information, and they can be resized using the `Resize` interface. + +Because the Eigen Tensor module is under-documented, please refer to `OpKernel`'s computation code in TensorFlow's [kernel module documentation](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/core/kernels). diff --git a/doc/survey/cluster_bootstrapping_tools.md b/doc/survey/cluster_bootstrapping_tools.md new file mode 100644 index 0000000000..1cd9962700 --- /dev/null +++ b/doc/survey/cluster_bootstrapping_tools.md @@ -0,0 +1,71 @@ +# Cluster bootstrapping tool survey +## Abstract +In order to bring up a cluster from bare metal machine to a fully functional kubernetes cluster for Paddlepaddle to run, we need to utilize some tools. Here we are going to compare [Sextant](https://github.com/k8sp/sextant) and [Tectonic installer](https://github.com/coreos/tectonic-installer) + +## Basic assumptions +Here are some basic assumptions before we move on to details +1. You are an administrator of a bare metal machine cluster, which means: + * you have full control to each of the machines. + * you have full control to the network which machines are connected to. +2. Machines can be booted from network with PEX or iPXE +3. You understand the [general procedure to bring up a cluster](#appendix-general-procedure-to-bring-up-a-cluster) + +if your cluster is able to mark above items with checkmarks, then keep reading. + +## Comparing Sextant and Tectonic installer +### Sextant +Sextant is an end2end solution to bring up a bare metal cluster to a fully functional k8s cluster, it integrates DHCP, name service, PEX, cloud-config-service, docker registry services altogether. + +#### Pros +1. End2End: basically all admin need to do is to config the cluster.yaml and power on the cluster. +2. Offline cluster configuration: Sextant has 2 phases during working with it, config time and deploy time. when admin is configuring, it requires admin's machine has internet connectivity, which will download some images, etc. But in deploy time, it's completely OK to go offline since all dependencies are ready during config time. +3. docker registry integrated. +4. GPU machine took care of. + +### Cons +1. k8s API server is not deployed with high availability in considering by default. +2. No grouping support. +3. No API interface, a one-off service. + + +### Tectonic installer +First of all, Tectonic is not free, it requires coreos.com account as a step of installation, and free user can only create less than 10 nodes. + +Tectonic is a suite of software which wraps around k8s and providing more utility regarding dev ops, ie, +Tectonic installer as it's named, it installs Tectonic to a bare metal cluster which means it's not totally an equivalent of Sextant. At the "booting a cluster" part, it mostly utilizes [Matchbox](https://github.com/coreos/matchbox), which is a general cluster bootstrapper. + +Matchbox's Approach is similar to Sexstant. + +### Pros +1. supports grouping machines. +2. supports running provisioning service in rtk. (not a big deal though). +3. supports http/gRPC API interface. +4. supports multi-template. + +### Cons +1. Not an e2e solution to bring up a cluster, need a lot of extra work and other software. +2. [Not fully supporting](https://github.com/coreos/matchbox/issues/550) centOS deployment yet. + +## Conclusion +Sextant is a better solution overall for paddle cloud deploying to a bare metal cluster. It would be great if Sextant can also 1) deploy k8s api server with high availability by default; 2) not designed as a one-off service. + + + +## Appendix: General procedure to bring up a cluster +It's physically impossible for a cluster admin to manually install OS and applications into cluster nodes one by one, here is what an admin would do in cloud industry: +1. setup a bootstrap machine with static IP in the cluster, which has following services: + * DHCP: assigns ip address for rest of the nodes. + * name service: to map node name to a IP + * PXE related services: the booting related info will be delivered to newly booted machines as their IP is assigned via DHCP service, PXE service will provide further booting and installing info and image with TFTP and http protocol. + * cluster config service: this is for providing cluster node with OS config via http + * optional docker registry: a built-in docker registry makes the whole cluster independent from connecting internet, and speeds up software distribution. +2. New node powers on, it will + * broadcast the request for an IP address + * DHCP server assigns the IP address, and deliver the PXE booting related info to the node. + * cluster node will request config files with booting info delivered with DHCP via the TFTP service, and in most of the cases, the config file will point to a http service for the booting image. + * Since PXE is configured with initrd, it will utilize the cloud config service and do further installations like coreOS or K8s installations. + * then restart the node. + +For further understanding, following 2 links from Matchbox are some good readings: +* [Machine lifecycle](https://github.com/coreos/matchbox/blob/master/Documentation/machine-lifecycle.md) +* [PXE booting](https://github.com/coreos/matchbox/blob/master/Documentation/network-booting.md) diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index ec866b2907..b435de80a2 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -19,7 +19,7 @@ if(Boost_FOUND) endif() if(WITH_C_API) - add_subdirectory(capi) + add_subdirectory(capi) endif() if(WITH_SWIG_PY) diff --git a/paddle/api/Util.cpp b/paddle/api/Util.cpp index d369df5d4e..11bd05c09d 100644 --- a/paddle/api/Util.cpp +++ b/paddle/api/Util.cpp @@ -47,7 +47,7 @@ bool isUsingGpu() { return FLAGS_use_gpu; } void setUseGpu(bool useGpu) { FLAGS_use_gpu = useGpu; } bool isGpuVersion() { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA return false; #else return true; diff --git a/paddle/capi/CMakeLists.txt b/paddle/capi/CMakeLists.txt index 3af111eb57..b9bbe58951 100644 --- a/paddle/capi/CMakeLists.txt +++ b/paddle/capi/CMakeLists.txt @@ -28,53 +28,47 @@ add_style_check_target(paddle_capi ${CAPI_SOURCES} ${CAPI_HEADER} add_dependencies(paddle_capi paddle_proto) - # combine all paddle static libraries together, into libpaddle_capi_whole.a # user should use PaddleCAPI as -lpaddle_capi_whole -set(capi_whole_library libpaddle_capi_whole.a) -add_custom_target(paddle_capi_whole ALL - COMMAND mkdir -p o_files/capi && cd o_files/capi/ && ar -x $ - COMMAND mkdir -p o_files/utils && cd o_files/utils/ && ar -x $ - COMMAND mkdir -p o_files/parameter && cd o_files/parameter/ && ar -x $ - COMMAND mkdir -p o_files/math && cd o_files/math/ && ar -x $ - COMMAND mkdir -p o_files/cuda && cd o_files/cuda/ && ar -x $ - COMMAND mkdir -p o_files/function && cd o_files/function/ && ar -x $ - COMMAND mkdir -p o_files/gserver && cd o_files/gserver/ && ar -x $ - COMMAND mkdir -p o_files/proto && cd o_files/proto/ && ar -x $ - COMMAND mkdir -p o_files/network && cd o_files/network/ && ar -x $ - COMMAND mkdir -p o_files/pserver && cd o_files/pserver/ && ar -x $ - COMMAND ar crs ${capi_whole_library} `find ./o_files -name '*.o'` - COMMAND rm -rf o_files - WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR} - DEPENDS paddle_capi paddle_utils paddle_parameter paddle_math - paddle_cuda paddle_function paddle_gserver - paddle_proto paddle_pserver paddle_network - ) -set_target_properties(paddle_capi_whole - PROPERTIES IMPORTED_LOCATION ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library}) +set(PADDLE_CAPI_INFER_LIBS + paddle_utils + paddle_parameter + paddle_math + paddle_cuda + paddle_function + paddle_gserver + paddle_proto + paddle_pserver + paddle_network) + +cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_CAPI_INFER_LIBS}) -set(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map") -# TODO: merge mkl into paddle_capi_shared -add_library(paddle_capi_shared SHARED ${CAPI_SOURCES}) -set_target_properties(paddle_capi_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}") -target_include_directories(paddle_capi_shared PUBLIC ${CMAKE_CURRENT_BINARY_DIR}) -link_paddle_exe(paddle_capi_shared) +# No shared library for iOS +if(NOT IOS) + set(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map") + # TODO: merge mkl into paddle_capi_shared + add_library(paddle_capi_shared SHARED ${CAPI_SOURCES}) + set_target_properties(paddle_capi_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}") + target_include_directories(paddle_capi_shared PUBLIC ${CMAKE_CURRENT_BINARY_DIR}) + link_paddle_exe(paddle_capi_shared) +endif() # install library & headers. install(FILES ${CAPI_HEADERS} DESTINATION include/paddle) install(FILES ${CMAKE_CURRENT_BINARY_DIR}/config.h DESTINATION include/paddle) if(ANDROID) + install(TARGETS paddle_capi_whole paddle_capi_shared + ARCHIVE DESTINATION lib/${ANDROID_ABI} + LIBRARY DESTINATION lib/${ANDROID_ABI}) execute_process( COMMAND ${GIT_EXECUTABLE} log --pretty=oneline -1 + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR} OUTPUT_VARIABLE GIT_COMMITS_LIST RESULT_VARIABLE GIT_COMMITS_LIST_RESULT ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) if(${GIT_COMMITS_LIST_RESULT}) set(GIT_COMMITS_LIST "No commits.") endif() - install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library} - DESTINATION lib/${ANDROID_ABI}) - install(TARGETS paddle_capi_shared DESTINATION lib/${ANDROID_ABI}) install(CODE "FILE(WRITE ${CMAKE_INSTALL_PREFIX}/lib/${ANDROID_ABI}/BUILD.txt \"Compiler:\n\" \"\\t${CMAKE_C_COMPILER}\\n\" @@ -88,8 +82,10 @@ if(ANDROID) )" ) else(ANDROID) - install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library} DESTINATION lib) - install(TARGETS paddle_capi_shared DESTINATION lib) + install(TARGETS paddle_capi_whole ARCHIVE DESTINATION lib) + if(NOT IOS) + install(TARGETS paddle_capi_shared DESTINATION lib) + endif() endif(ANDROID) # this variable used for unittest diff --git a/paddle/capi/Matrix.cpp b/paddle/capi/Matrix.cpp index d898ebe261..4547afaf1d 100644 --- a/paddle/capi/Matrix.cpp +++ b/paddle/capi/Matrix.cpp @@ -46,7 +46,7 @@ paddle_error paddle_matrix_set_row(paddle_matrix mat, if (rowID >= ptr->mat->getHeight()) return kPD_OUT_OF_RANGE; paddle::real* buf = ptr->mat->getRowBuf(rowID); size_t width = ptr->mat->getWidth(); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA hl_memcpy(buf, rowArray, sizeof(paddle::real) * width); #else std::copy(rowArray, rowArray + width, buf); diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt index 3371962c63..3e0e0f5903 100644 --- a/paddle/framework/CMakeLists.txt +++ b/paddle/framework/CMakeLists.txt @@ -19,14 +19,15 @@ cc_test(scope_test SRCS scope_test.cc DEPS scope) proto_library(framework_proto SRCS framework.proto) cc_library(attribute SRCS attribute.cc DEPS framework_proto) -cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto) -cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope) +cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS attribute) +cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute) +cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker) +cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto proto_desc) +cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope proto_desc) cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry) -cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS operator) -cc_library(op_registry SRCS op_registry.cc DEPS grad_op_builder) +cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator) cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry) -cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry add_op) py_proto_compile(framework_py_proto SRCS framework.proto) # Generate an empty __init__.py to make framework_py_proto as a valid python module. @@ -40,3 +41,6 @@ add_custom_command(TARGET framework_py_proto POST_BUILD cc_library(backward SRCS backward.cc DEPS net_op) cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context) + +cc_library(tensor_array SRCS tensor_array.cc DEPS lod_tensor) +cc_test(tensor_array_test SRCS tensor_array_test.cc DEPS tensor_array place) diff --git a/paddle/framework/attribute.cc b/paddle/framework/attribute.cc index 27132eaa0b..d6a2975aaa 100644 --- a/paddle/framework/attribute.cc +++ b/paddle/framework/attribute.cc @@ -19,74 +19,62 @@ limitations under the License. */ namespace paddle { namespace framework { -template <> -AttrType AttrTypeID() { - return INT; -} -template <> -AttrType AttrTypeID() { - return FLOAT; -} -template <> -AttrType AttrTypeID() { - return STRING; -} -template <> -AttrType AttrTypeID>() { - return INTS; -} -template <> -AttrType AttrTypeID>() { - return FLOATS; -} -template <> -AttrType AttrTypeID>() { - return STRINGS; -} -template <> -AttrType AttrTypeID>>() { - return INT_PAIRS; +static ProgramDesc* g_program_desc = nullptr; + +ProgramDesc& GetProgramDesc() { + if (g_program_desc == nullptr) { + g_program_desc = new ProgramDesc(); + auto root_block = g_program_desc->mutable_blocks()->Add(); + root_block->set_idx(0); + root_block->set_parent_idx(-1); + } + return *g_program_desc; } Attribute GetAttrValue(const OpDesc::Attr& attr_desc) { switch (attr_desc.type()) { - case paddle::framework::AttrType::INT: { + case framework::AttrType::BOOLEAN: { + return attr_desc.b(); + } + case framework::AttrType::INT: { return attr_desc.i(); } - case paddle::framework::AttrType::FLOAT: { + case framework::AttrType::FLOAT: { return attr_desc.f(); } - case paddle::framework::AttrType::STRING: { + case framework::AttrType::STRING: { return attr_desc.s(); } - case paddle::framework::AttrType::INTS: { + case framework::AttrType::BOOLEANS: { + std::vector val(attr_desc.bools_size()); + for (int i = 0; i < attr_desc.bools_size(); ++i) { + val[i] = attr_desc.bools(i); + } + return val; + } + case framework::AttrType::INTS: { std::vector val(attr_desc.ints_size()); for (int i = 0; i < attr_desc.ints_size(); ++i) { val[i] = attr_desc.ints(i); } return val; } - case paddle::framework::AttrType::FLOATS: { + case framework::AttrType::FLOATS: { std::vector val(attr_desc.floats_size()); for (int i = 0; i < attr_desc.floats_size(); ++i) { val[i] = attr_desc.floats(i); } return val; } - case paddle::framework::AttrType::STRINGS: { + case framework::AttrType::STRINGS: { std::vector val(attr_desc.strings_size()); for (int i = 0; i < attr_desc.strings_size(); ++i) { val[i] = attr_desc.strings(i); } return val; } - case paddle::framework::AttrType::INT_PAIRS: { - std::vector> val(attr_desc.int_pairs_size()); - for (int i = 0; i < attr_desc.int_pairs_size(); ++i) { - val[i].first = attr_desc.int_pairs(i).first(); - val[i].second = attr_desc.int_pairs(i).second(); - } - return val; + case framework::AttrType::BLOCK: { + return GetProgramDesc().mutable_blocks(attr_desc.block_idx()); } } PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !"); diff --git a/paddle/framework/attribute.h b/paddle/framework/attribute.h index 2b788a76ca..d13530e340 100644 --- a/paddle/framework/attribute.h +++ b/paddle/framework/attribute.h @@ -21,24 +21,37 @@ limitations under the License. */ #include #include "paddle/framework/framework.pb.h" +#include "paddle/framework/type_defs.h" #include "paddle/platform/enforce.h" -#include "paddle/platform/variant.h" namespace paddle { namespace framework { -typedef boost::variant, - std::vector, std::vector, - std::vector>> - Attribute; - -typedef std::unordered_map AttributeMap; +ProgramDesc& GetProgramDesc(); template -AttrType AttrTypeID(); +inline AttrType AttrTypeID() { + Attribute tmp = T(); + return static_cast(tmp.which() - 1); +} Attribute GetAttrValue(const OpDesc::Attr& attr_desc); +class AttrReader { + public: + explicit AttrReader(const AttributeMap& attrs) : attrs_(attrs) {} + + template + inline const T& Get(const std::string& name) const { + PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap", + name); + return boost::get(attrs_.at(name)); + } + + private: + const AttributeMap& attrs_; +}; + // check whether a value(attribute) fit a certain limit template class GreaterThanChecker { diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc index c5d4662215..c970e01dd1 100644 --- a/paddle/framework/backward.cc +++ b/paddle/framework/backward.cc @@ -13,10 +13,13 @@ limitations under the License. */ #include "paddle/framework/backward.h" +#include "paddle/operators/net_op.h" +#include #include #include +#include "paddle/framework/block_desc.h" #include "paddle/framework/op_registry.h" #include "paddle/operators/net_op.h" #include "paddle/operators/recurrent_op.h" @@ -24,6 +27,35 @@ namespace paddle { namespace framework { +static inline std::unique_ptr CreateGradOp( + const OperatorBase& op) { + OpDescBind op_desc; + op_desc.SetInputMap(op.Inputs()); + op_desc.SetOutputMap(op.Outputs()); + op_desc.SetType(op.Type()); + op_desc.SetAttrMap(op.Attrs()); + auto& info = OpInfoMap::Instance().Get(op.Type()); + auto grad_descs = info.GradOpMaker()(op_desc); + std::vector> grad_ops; + grad_ops.reserve(grad_descs.size()); + std::transform(grad_descs.begin(), grad_descs.end(), + std::back_inserter(grad_ops), + [](const std::unique_ptr& grad_desc) { + return OpRegistry::CreateOp(*grad_desc); + }); + PADDLE_ENFORCE(!grad_ops.empty()); + if (grad_ops.size() == 1) { + return std::move(grad_ops[0]); + } else { + auto net_op = new operators::NetOp(); + for (auto& grad_op : grad_ops) { + net_op->AppendOp(std::move(grad_op)); + } + net_op->CompleteAddOp(); + return std::unique_ptr(net_op); + } +} + template static void ForEachVarName(const Map& names, T callback) { for (auto& name : names) { @@ -141,9 +173,26 @@ static std::unique_ptr BackwardRecursive( net->ops_[op_offset]->Rename(name, dup_outputs.back()); } // collect all the offset to append `add` op for each alias - insert_position.push_back( - {dup_op.back(), OpRegistry::CreateOp("add", {{"X", {dup_outputs}}}, - {{"Out", {name}}}, {})}); + // + // one variable is shared between multiple operators. + // insert add operator one by one, then add it to output + for (size_t output_idx = 0; output_idx < dup_outputs.size() - 1; + ++output_idx) { + auto insert_add_x = dup_outputs[output_idx]; + auto insert_add_y = dup_outputs[output_idx + 1]; + auto insert_add_out = name + "@SHARED@" + std::to_string(output_idx); + // first add op inserted + if (output_idx == dup_outputs.size() - 2) { + insert_add_out = name; + } + if (output_idx != 0) { + insert_add_y = name + "@SHARED@" + std::to_string(output_idx - 1); + } + insert_position.push_back( + {dup_op.back(), + OpRegistry::CreateOp("sum", {{"X", {insert_add_x, insert_add_y}}}, + {{"Out", {insert_add_out}}}, {})}); + } } // make sure the inserted `add` ops follow the BFS order. @@ -154,7 +203,7 @@ static std::unique_ptr BackwardRecursive( net->InsertOp(pos.first + 1, std::move(pos.second)); } } else { - std::unique_ptr grad_op(OpRegistry::CreateGradOp(forwardOp)); + std::unique_ptr grad_op(CreateGradOp(forwardOp)); ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op]( const std::string& grad_input) { @@ -166,9 +215,8 @@ static std::unique_ptr BackwardRecursive( // If part of input gradient of that operator is not calculated, fill // zero variables to that input gradient. - net->AppendOp(OpRegistry::CreateOp("fill_zeros_like", - {{"Src", {prefix}}}, - {{"Dst", {grad_input}}}, {})); + net->AppendOp(OpRegistry::CreateOp("fill_zeros_like", {{"X", {prefix}}}, + {{"Y", {grad_input}}}, {})); } return false; }); @@ -183,7 +231,8 @@ static std::unique_ptr BackwardRecursive( // process recurrent gradient op as a special operator. if (forwardOp.Type() == "recurrent") { - // NOTE clean up cycle call somewhere (RNN's stepnet constains itself), or + // NOTE clean up cycle call somewhere (RNN's stepnet constains itself), + // or // this will result in infinite loop. const auto& rnnop = *static_cast(&forwardOp); @@ -223,5 +272,145 @@ std::unique_ptr Backward( return BackwardRecursive(forwardOp, no_grad_names, uid); } +// ==================================== // + +static bool AllGradInSet(const std::vector& names, + const std::unordered_set& set) { + for (const std::string& name : names) { + if (!set.count(GradVarName(name))) { + return false; + } + } + return true; +} + +std::vector> MakeOpGrad( + const std::unique_ptr& op_desc, + std::unordered_set& no_grad_vars) { + std::vector> grad_op_descs; + // All input gradients of forwarding operator do not need to calculat. + const std::vector& inputs = op_desc->InputArgumentNames(); + if (AllGradInSet(inputs, no_grad_vars)) { + return grad_op_descs; // empty vector + } + // All output gradients of forwarding operator do not need to calculate. + const std::vector& outputs = op_desc->OutputArgumentNames(); + if (AllGradInSet(outputs, no_grad_vars)) { + for (const std::string& name : inputs) { + no_grad_vars.insert(GradVarName(name)); + } + return grad_op_descs; // empty vector + } + + grad_op_descs = OpRegistry::CreateGradOpDescs(*op_desc); + + std::list> pending_fill_zeros_ops; + for (auto& desc : grad_op_descs) { + for (const std::string& in_name : desc->InputArgumentNames()) { + if (no_grad_vars.count(in_name)) { + std::string prefix = in_name.substr( + 0, in_name.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1); + std::string new_name = prefix + kZeroVarSuffix; + desc->Rename(in_name, new_name); + std::unique_ptr fill_zeros_op(new OpDescBind( + "fill_zeros_like", {{"X", {prefix}}}, {{"Y", {new_name}}}, {})); + pending_fill_zeros_ops.push_back(std::move(fill_zeros_op)); + } + } + for (const std::string& out_name : desc->OutputArgumentNames()) { + if (no_grad_vars.count(out_name)) { + desc->Rename(out_name, kEmptyVarName); + } + } + } + + for (auto& p : pending_fill_zeros_ops) { + grad_op_descs.insert(grad_op_descs.begin(), std::move(p)); + } + return grad_op_descs; +} + +std::vector> MakeBlockBackward( + ProgramDescBind& program_desc, int block_idx, + std::unordered_set& no_grad_vars) { + BlockDescBind* cur_block = program_desc.Block(block_idx); + std::deque>& op_descs = cur_block->ops_; + std::unordered_map> dup_out_ops; + size_t grad_desc_idx = 0; + std::vector> backward_descs; + for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) { + std::vector> op_grads = + MakeOpGrad(*it, no_grad_vars); + + if ((*it)->Type() == "recurrent") { + PADDLE_ENFORCE_EQ( + op_grads.size(), size_t(1), + "rnn_op's gradient process should contain only one op."); + int step_block_idx = (*it)->GetBlockAttr("stop_block"); + auto backward_block_op_descs = + MakeBlockBackward(program_desc, step_block_idx, no_grad_vars); + BlockDescBind* backward_block = program_desc.AppendBlock(*cur_block); + for (auto& ptr : backward_block_op_descs) { + backward_block->ops_.push_back(std::move(ptr)); + } + op_grads[0]->SetBlockAttr("step_block", *backward_block); + } + + for (const auto& desc : op_grads) { + for (const std::string& out_name : desc->OutputArgumentNames()) { + dup_out_ops[out_name].emplace_back(grad_desc_idx); + } + ++grad_desc_idx; + } + std::transform( + op_grads.begin(), op_grads.end(), std::back_inserter(backward_descs), + [](std::unique_ptr& ptr) { return std::move(ptr); }); + } + // Check whether some variables are written more than once + std::list>> pending_sum_ops; + for (const auto& dup : dup_out_ops) { + const std::string& out_name = dup.first; + const std::vector dup_op = dup.second; + if (out_name != kEmptyVarName && dup_op.size() > 1) { + std::vector sum_op_inputs; + for (size_t i = 0; i < dup_op.size(); ++i) { + std::string new_name = out_name + "@RENAME@" + std::to_string(i); + backward_descs[dup_op[i]]->Rename(out_name, new_name); + sum_op_inputs.emplace_back(new_name); + } + std::unique_ptr sum_op(new OpDescBind( + "sum", {{"X", sum_op_inputs}}, {{"Out", {out_name}}}, {})); + pending_sum_ops.push_back({dup_op.back(), std::move(sum_op)}); + } + } + pending_sum_ops.sort( + [](const std::pair>& a, + const std::pair>& b) { + return a.first > b.first; + }); + for (auto& p : pending_sum_ops) { + backward_descs.insert(backward_descs.begin() + p.first + 1, + std::move(p.second)); + } + return backward_descs; +} + +void AppendBackward(ProgramDescBind& program_desc, + const std::unordered_set& no_grad_vars) { + std::unordered_set no_grad_var_names; + no_grad_var_names.reserve(no_grad_vars.size() + 1); + no_grad_var_names.insert(std::string(kEmptyVarName) + kGradVarSuffix); + for (auto& name : no_grad_vars) { + no_grad_var_names.insert(GradVarName(name)); + } + const int root_block_idx = 0; + auto backward_op_descs = + MakeBlockBackward(program_desc, root_block_idx, no_grad_var_names); + auto& forw_op_descs = program_desc.Block(root_block_idx)->ops_; + for (auto& ptr : backward_op_descs) { + forw_op_descs.push_back(std::move(ptr)); + } +} + } // namespace framework } // namespace paddle diff --git a/paddle/framework/backward.h b/paddle/framework/backward.h index 1ecf69881b..7ffe4c2810 100644 --- a/paddle/framework/backward.h +++ b/paddle/framework/backward.h @@ -13,8 +13,11 @@ limitations under the License. */ #pragma once + #include -#include "operator.h" +#include "paddle/framework/operator.h" +#include "paddle/framework/program_desc.h" + namespace paddle { namespace framework { @@ -23,5 +26,9 @@ namespace framework { extern std::unique_ptr Backward( const OperatorBase& forwardOp, const std::unordered_set& no_grad_vars); + +void AppendBackward(ProgramDescBind& program_desc, + const std::unordered_set& no_grad_vars); + } // namespace framework } // namespace paddle diff --git a/paddle/framework/backward.md b/paddle/framework/backward.md index 0a6d762bc8..ac60be5724 100644 --- a/paddle/framework/backward.md +++ b/paddle/framework/backward.md @@ -2,7 +2,7 @@ ## Motivation -In Neural Network, many model is solved by the the backpropagation algorithm(known as BP) at present. Technically it caculates the gradient of the loss function, then distributed back through the networks. Follows the chain rule, so we need a module chains the gradient operators/expressions together with to construct the backward pass. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass. +In Neural Network, most models are solved by the backpropagation algorithm(known as **BP**) at present. Technically, BP calculates the gradient of the loss function, then propagates it back through the networks following the chain rule. Hence we need a module that chains the gradient operators/expressions together to construct the backward pass. Every forward network needs a backward network to construct the full computation graph. The operator/expression's backward pass will be generated with respect to the forward pass. ## Implementation @@ -24,9 +24,9 @@ A backward network is built up with several backward operators. Backward operato | **Operator::inputs_** | Inputs | Inputs, Outputs, OutputGradients | | **Operator::outputs_** | Outputs | InputGradients | - In most cases, there is a one-to-one correspondence between the forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced. + In most cases, there is a one-to-one relation between the forward and backward operators. These relations are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and to make operators pluggable, the registry mechanism is introduced. -For example, we have got a `mul_op`, and we can register its information and corresponding backward operator by the following macro: +For example, we have `mul_op`, and we can register its information and corresponding backward operator by the following macro: ```cpp REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad); @@ -48,7 +48,7 @@ The function `BuildGradOp` will sequentially execute following processes: 1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`. -2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing. +2. Build two maps named `inputs` and `outputs` to temporarily store backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing. 3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`. @@ -56,11 +56,11 @@ The function `BuildGradOp` will sequentially execute following processes: ### Backward Network Building -A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and append them together one by one. There is some corner case need to process specially. +A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and appending them together one by one. There are some corner cases that need special processing. 1. Op - When the input forward network is an Op, return its gradient Operator Immediately. If all of its outputs are in no gradient set, then return a special `NOP`. + When the input forward network is an Op, return its gradient Operator immediately. If all of its outputs are in no gradient set, then return a special `NOP`. 2. NetOp @@ -68,33 +68,33 @@ A backward network is a series of backward operators. The main idea of building 3. RnnOp - RnnOp is a nested stepnet operator. Backward module need to recusively call `Backward` for every stepnet. + RnnOp is a nested stepnet operator. Backward module needs to recusively call `Backward` for every stepnet. 4. Sharing Variables - **sharing variables**. As illustrated in the pictures, two operator's share the same variable name of W@GRAD, which will overwrite their sharing input variable. + As illustrated in the figure 1 and figure 2, two operators share the same variable name **W@GRAD**, which will overwrite their shared input variable.


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

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


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

-​ Because our framework finds variables accord to their names, we need to rename the output links. We add a suffix of number to represent its position in clockwise. +​ Because the framework finds variables according to their names, we need to rename the output links. We add an integer suffix to represent its position in the clockwise direction. -5. Part of Gradient is Zero. +5. Part of the Gradient is Zero. - In the whole graph, there is some case of that one operator's gradient is not needed, but its input's gradient is a dependency link of other operator, we need to fill a same shape gradient matrix in the position. In our implement, we insert a special `fillZeroLike` operator. + In the whole graph, there is some case of that one operator's gradient is not needed, but its input's gradient is a dependency link of other operator, we need to fill a same shape gradient matrix in the position. In our implementation, we insert a special `fillZeroLike` operator. Follow these rules above, then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it. diff --git a/paddle/framework/backward_test.cc b/paddle/framework/backward_test.cc index ad8003420d..30225a4a99 100644 --- a/paddle/framework/backward_test.cc +++ b/paddle/framework/backward_test.cc @@ -15,30 +15,42 @@ #include "paddle/framework/backward.h" #include +#include "paddle/framework/block_desc.h" +#include "paddle/framework/op_desc.h" #include "paddle/framework/op_registry.h" #include "paddle/operators/net_op.h" namespace paddle { namespace framework { -using OperatorBase = framework::OperatorBase; -using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker; -using OpProto = framework::OpProto; -using OpAttrChecker = framework::OpAttrChecker; -using Scope = framework::Scope; using DeviceContext = platform::DeviceContext; class RowWiseAddOpMaker : public OpProtoAndCheckerMaker { public: RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "Input X of Add").NotInGradient(); - AddInput("b", "Bias of Add").NotInGradient(); - AddOutput("Out", "Out of Add").NotInGradient(); + AddInput("X", "Input X of Add"); + AddInput("b", "Bias of Add"); + AddOutput("Out", "Out of Add"); AddComment("Add Op"); } }; +class RowWiseAddGradMaker : public SingleGradOpDescMaker { + public: + using SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto grad_op = new OpDescBind(); + grad_op->SetInput(GradVarName("Out"), OutputGrad("Out")); + grad_op->SetOutput(GradVarName("X"), InputGrad("X")); + grad_op->SetOutput(GradVarName("b"), InputGrad("b")); + grad_op->SetType("rowwise_add_grad"); + return std::unique_ptr(grad_op); + } +}; + class MulOpMaker : public OpProtoAndCheckerMaker { public: MulOpMaker(OpProto *proto, OpAttrChecker *op_checker) @@ -127,48 +139,52 @@ class FillZeroOpMaker : public OpProtoAndCheckerMaker { public: FillZeroOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("Src", "x"); - AddOutput("Dst", "out"); + AddInput("X", "x"); + AddOutput("Y", "out"); + AddComment(""); + } +}; + +class SumOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SumOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "the input tensors of sum operator.").AsDuplicable(); + AddOutput("Out", "the output tensor of sum operator."); AddComment(""); } }; -class AddOpMaker : public OpProtoAndCheckerMaker { +class MultInOutOpMaker : public OpProtoAndCheckerMaker { public: - AddOpMaker(OpProto *proto, OpAttrChecker *op_checker) + MultInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "x").AsDuplicable(); - AddOutput("Out", "out"); + AddInput("X", "x"); + AddInput("H", "h"); + AddOutput("Y", "y"); + AddOutput("Z", "z"); AddComment(""); } }; + } // namespace framework } // namespace paddle namespace f = paddle::framework; namespace ops = paddle::operators; using EnforceNotMet = paddle::platform::EnforceNotMet; -REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad, - f::NOP); +REGISTER_OPERATOR(rowwise_add, f::NOP, f::RowWiseAddOpMaker, + f::RowWiseAddGradMaker); +REGISTER_OPERATOR(rowwise_add_grad, f::NOP); REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP); REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP); REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker); REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker); -REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP); +REGISTER_OP(sum, f::NOP, f::SumOpMaker, sum_grad, f::NOP); REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker); REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad, f::NOP); - -TEST(Backward, simple_op_grad) { - auto fwd = f::OpRegistry::CreateOp( - "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {}); - ASSERT_NE(fwd, nullptr); - auto gop = f::OpRegistry::CreateGradOp(*fwd); - ASSERT_EQ(1UL, gop->Inputs().size()); - ASSERT_EQ("rowwise_add_grad", gop->Type()); - ASSERT_EQ(f::GradVarName("x"), gop->Output(f::GradVarName("X"))); - ASSERT_EQ(f::GradVarName("b"), gop->Output(f::GradVarName("b"))); -} +REGISTER_OP(mult_in_out, f::NOP, f::MultInOutOpMaker, mult_in_out_grad, f::NOP); TEST(Backward, simple_op_not_need_grad) { auto fwd = f::OpRegistry::CreateOp( @@ -283,18 +299,7 @@ TEST(Backward, net_shared_weight) { ASSERT_TRUE(bwd->IsNetOp()); auto bwd_net = static_cast(bwd.get()); ASSERT_EQ(3UL, bwd_net->ops_.size()); - ASSERT_EQ("add", bwd_net->ops_[2]->Type()); -} - -TEST(Backward, op_register_grad_not_for_network) { - auto fwd = - f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}}, - {{"mul_result", {"mul_out"}}, - {"add_result", {"add_out"}}, - {"Out", {"out1"}}}, - {{"temporary_index", std::vector{0, 1}}}); - - ASSERT_THROW(f::OpRegistry::CreateGradOp(*fwd), EnforceNotMet); + ASSERT_EQ("sum", bwd_net->ops_[2]->Type()); } TEST(Backward, op_all_input_are_not_need) { @@ -325,10 +330,10 @@ TEST(Backward, op_part_of_output_are_not_need) { auto &fill_zero = *net->ops_[0]; ASSERT_EQ("fill_zeros_like", fill_zero.Type()); - ASSERT_EQ(1UL, fill_zero.Inputs("Src").size()); - ASSERT_EQ("Z", fill_zero.Input("Src")); - ASSERT_EQ(1UL, fill_zero.Outputs("Dst").size()); - ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Dst")); + ASSERT_EQ(1UL, fill_zero.Inputs("X").size()); + ASSERT_EQ("Z", fill_zero.Input("X")); + ASSERT_EQ(1UL, fill_zero.Outputs("Y").size()); + ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Y")); auto &d_many_out = *net->ops_[1]; ASSERT_EQ("many_output_op_grad", d_many_out.Type()); @@ -399,3 +404,293 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) { EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL); EXPECT_EQ(bwd_net->ops_[2]->Outputs(all).size(), 0UL); } + +// =================================== // + +f::ProgramDesc *GetNewProgramDesc() { + auto *program_desc = new f::ProgramDesc(); + auto *root_block = program_desc->add_blocks(); + root_block->set_idx(0); + root_block->set_parent_idx(-1); + return program_desc; +} + +TEST(Backward, simple_single_op) { + f::ProgramDesc *program_desc = GetNewProgramDesc(); + f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc); + f::BlockDescBind *block = program.Block(0); + f::OpDescBind *op = block->AppendOp(); + op->SetType("rowwise_add"); + op->SetInput("X", {"x"}); + op->SetInput("b", {"b"}); + op->SetOutput("Out", {"out"}); + + AppendBackward(program, {}); + + ASSERT_EQ(block->AllOps().size(), 2UL); + f::OpDescBind *grad_op = block->AllOps()[1]; + EXPECT_EQ(grad_op->Type(), "rowwise_add_grad"); + ASSERT_EQ(grad_op->InputNames().size(), 1UL); + ASSERT_EQ(grad_op->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out")})); + EXPECT_EQ(grad_op->Output(f::GradVarName("X")), + std::vector({f::GradVarName("x")})); + EXPECT_EQ(grad_op->Output(f::GradVarName("b")), + std::vector({f::GradVarName("b")})); +} + +TEST(Backward, simple_mult_op) { + f::ProgramDesc *program_desc = GetNewProgramDesc(); + f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc); + f::BlockDescBind *block = program.Block(0); + f::OpDescBind *op1 = block->AppendOp(); + op1->SetType("rowwise_add"); + op1->SetInput("X", {"x1"}); + op1->SetInput("b", {"b1"}); + op1->SetOutput("Out", {"out1"}); + + f::OpDescBind *op2 = block->AppendOp(); + op2->SetType("mul"); + op2->SetInput("X", {"out1"}); + op2->SetInput("Y", {"y2"}); + op2->SetOutput("Out", {"out2"}); + + f::OpDescBind *op3 = block->AppendOp(); + op3->SetType("rowwise_add"); + op3->SetInput("X", {"out2"}); + op3->SetInput("b", {"b3"}); + op3->SetOutput("Out", {"out3"}); + + AppendBackward(program, {}); + + ASSERT_EQ(block->AllOps().size(), 6UL); + f::OpDescBind *grad_op1 = block->AllOps()[5]; + EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad"); + ASSERT_EQ(grad_op1->InputNames().size(), 1UL); + ASSERT_EQ(grad_op1->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("X")), + std::vector({f::GradVarName("x1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("b")), + std::vector({f::GradVarName("b1")})); + + f::OpDescBind *grad_op2 = block->AllOps()[4]; + EXPECT_EQ(grad_op2->Type(), "mul_grad"); + ASSERT_EQ(grad_op2->InputNames().size(), 4UL); + ASSERT_EQ(grad_op2->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op2->Input("X"), std::vector({"out1"})); + EXPECT_EQ(grad_op2->Input("Y"), std::vector({"y2"})); + EXPECT_EQ(grad_op2->Input("Out"), std::vector({"out2"})); + EXPECT_EQ(grad_op2->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out2")})); + EXPECT_EQ(grad_op2->Output(f::GradVarName("X")), + std::vector({f::GradVarName("out1")})); + EXPECT_EQ(grad_op2->Output(f::GradVarName("Y")), + std::vector({f::GradVarName("y2")})); + + f::OpDescBind *grad_op3 = block->AllOps()[3]; + EXPECT_EQ(grad_op3->Type(), "rowwise_add_grad"); + ASSERT_EQ(grad_op3->InputNames().size(), 1UL); + ASSERT_EQ(grad_op3->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op3->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out3")})); + EXPECT_EQ(grad_op3->Output(f::GradVarName("X")), + std::vector({f::GradVarName("out2")})); + EXPECT_EQ(grad_op3->Output(f::GradVarName("b")), + std::vector({f::GradVarName("b3")})); +} + +TEST(Backward, intermedia_var_no_grad) { + f::ProgramDesc *program_desc = GetNewProgramDesc(); + f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc); + f::BlockDescBind *block = program.Block(0); + f::OpDescBind *op1 = block->AppendOp(); + op1->SetType("rowwise_add"); + op1->SetInput("X", {"x1"}); + op1->SetInput("b", {"b1"}); + op1->SetOutput("Out", {"out1"}); + + f::OpDescBind *op2 = block->AppendOp(); + op2->SetType("mul"); + op2->SetInput("X", {"x2"}); + op2->SetInput("Y", {"y2"}); + op2->SetOutput("Out", {"out2"}); + + f::OpDescBind *op3 = block->AppendOp(); + op3->SetType("rowwise_add"); + op3->SetInput("X", {"out2"}); + op3->SetInput("b", {"b3"}); + op3->SetOutput("Out", {"out3"}); + + f::OpDescBind *op4 = block->AppendOp(); + op4->SetType("mul"); + op4->SetInput("X", {"out1"}); + op4->SetInput("Y", {"out3"}); + op4->SetOutput("Out", {"out4"}); + + AppendBackward(program, {"out3"}); + + ASSERT_EQ(block->AllOps().size(), 6UL); + f::OpDescBind *grad_op1 = block->AllOps()[5]; + EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad"); + ASSERT_EQ(grad_op1->InputNames().size(), 1UL); + ASSERT_EQ(grad_op1->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("X")), + std::vector({f::GradVarName("x1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("b")), + std::vector({f::GradVarName("b1")})); + + f::OpDescBind *grad_op4 = block->AllOps()[4]; + EXPECT_EQ(grad_op4->Type(), "mul_grad"); + ASSERT_EQ(grad_op4->InputNames().size(), 4UL); + ASSERT_EQ(grad_op4->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op4->Input("X"), std::vector({"out1"})); + EXPECT_EQ(grad_op4->Input("Y"), std::vector({"out3"})); + EXPECT_EQ(grad_op4->Input("Out"), std::vector({"out4"})); + EXPECT_EQ(grad_op4->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out4")})); + EXPECT_EQ(grad_op4->Output(f::GradVarName("X")), + std::vector({f::GradVarName("out1")})); + EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")), + std::vector({f::kEmptyVarName})); +} + +TEST(Backward, var_no_grad) { + f::ProgramDesc *program_desc = GetNewProgramDesc(); + f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc); + f::BlockDescBind *block = program.Block(0); + f::OpDescBind *op1 = block->AppendOp(); + op1->SetType("mult_in_out"); + op1->SetInput("X", {"x1"}); + op1->SetInput("H", {"h1"}); + op1->SetOutput("Y", {"y1"}); + op1->SetOutput("Z", {"z1"}); + + f::OpDescBind *op2 = block->AppendOp(); + op2->SetType("mult_in_out"); + op2->SetInput("X", {"y1"}); + op2->SetInput("H", {"z1"}); + op2->SetOutput("Y", {"y2"}); + op2->SetOutput("Z", {"z2"}); + + AppendBackward(program, {"z1"}); + + ASSERT_EQ(block->AllOps().size(), 5UL); + f::OpDescBind *grad_op2 = block->AllOps()[2]; + ASSERT_EQ(grad_op2->Type(), "mult_in_out_grad"); + ASSERT_EQ(grad_op2->InputNames().size(), 6UL); + ASSERT_EQ(grad_op2->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op2->Input("X"), std::vector({"y1"})); + EXPECT_EQ(grad_op2->Input("H"), std::vector({"z1"})); + EXPECT_EQ(grad_op2->Input("Y"), std::vector({"y2"})); + EXPECT_EQ(grad_op2->Input("Z"), std::vector({"z2"})); + EXPECT_EQ(grad_op2->Input(f::GradVarName("Y")), + std::vector({f::GradVarName("y2")})); + EXPECT_EQ(grad_op2->Input(f::GradVarName("Z")), + std::vector({f::GradVarName("z2")})); + EXPECT_EQ(grad_op2->Output(f::GradVarName("X")), + std::vector({f::GradVarName("y1")})); + EXPECT_EQ(grad_op2->Output(f::GradVarName("H")), + std::vector({f::kEmptyVarName})); + + f::OpDescBind *fill_zero_op = block->AllOps()[3]; + ASSERT_EQ(fill_zero_op->Type(), "fill_zeros_like"); + ASSERT_EQ(fill_zero_op->InputNames().size(), 1UL); + ASSERT_EQ(fill_zero_op->OutputNames().size(), 1UL); + EXPECT_EQ(fill_zero_op->Input("X"), std::vector({"z1"})); + EXPECT_EQ(fill_zero_op->Output("Y"), + std::vector({std::string("z1") + f::kZeroVarSuffix})); + + f::OpDescBind *grad_op1 = block->AllOps()[4]; + ASSERT_EQ(grad_op1->Type(), "mult_in_out_grad"); + ASSERT_EQ(grad_op1->InputNames().size(), 6UL); + ASSERT_EQ(grad_op1->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op1->Input("X"), std::vector({"x1"})); + EXPECT_EQ(grad_op1->Input("H"), std::vector({"h1"})); + EXPECT_EQ(grad_op1->Input("Y"), std::vector({"y1"})); + EXPECT_EQ(grad_op1->Input("Z"), std::vector({"z1"})); + EXPECT_EQ(grad_op1->Input(f::GradVarName("Y")), + std::vector({f::GradVarName("y1")})); + EXPECT_EQ(grad_op1->Input(f::GradVarName("Z")), + std::vector({std::string("z1") + f::kZeroVarSuffix})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("X")), + std::vector({f::GradVarName("x1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("H")), + std::vector({f::GradVarName("h1")})); +} + +TEST(Backward, shared_var) { + f::ProgramDesc *program_desc = GetNewProgramDesc(); + f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc); + f::BlockDescBind *block = program.Block(0); + f::OpDescBind *op1 = block->AppendOp(); + op1->SetType("rowwise_add"); + op1->SetInput("X", {"x1"}); + op1->SetInput("b", {"b1"}); + op1->SetOutput("Out", {"out1"}); + + f::OpDescBind *op2 = block->AppendOp(); + op2->SetType("mul"); + op2->SetInput("X", {"out1"}); + op2->SetInput("Y", {"y2"}); + op2->SetOutput("Out", {"out2"}); + + f::OpDescBind *op3 = block->AppendOp(); + op3->SetType("rowwise_add"); + op3->SetInput("X", {"out1"}); + op3->SetInput("b", {"b3"}); + op3->SetOutput("Out", {"out3"}); + + AppendBackward(program, {}); + + ASSERT_EQ(block->AllOps().size(), 7UL); + f::OpDescBind *grad_op3 = block->AllOps()[3]; + ASSERT_EQ(grad_op3->Type(), "rowwise_add_grad"); + ASSERT_EQ(grad_op3->InputNames().size(), 1UL); + ASSERT_EQ(grad_op3->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op3->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out3")})); + EXPECT_EQ(grad_op3->Output(f::GradVarName("X")), + std::vector({f::GradVarName("out1") + "@RENAME@0"})); + EXPECT_EQ(grad_op3->Output(f::GradVarName("b")), + std::vector({f::GradVarName("b3")})); + + f::OpDescBind *grad_op4 = block->AllOps()[4]; + ASSERT_EQ(grad_op4->Type(), "mul_grad"); + ASSERT_EQ(grad_op4->InputNames().size(), 4UL); + ASSERT_EQ(grad_op4->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op4->Input("X"), std::vector({"out1"})); + EXPECT_EQ(grad_op4->Input("Y"), std::vector({"y2"})); + EXPECT_EQ(grad_op4->Input("Out"), std::vector({"out2"})); + EXPECT_EQ(grad_op4->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out2")})); + EXPECT_EQ(grad_op4->Output(f::GradVarName("X")), + std::vector({f::GradVarName("out1") + "@RENAME@1"})); + EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")), + std::vector({f::GradVarName("y2")})); + + f::OpDescBind *sum_op = block->AllOps()[5]; + ASSERT_EQ(sum_op->Type(), "sum"); + ASSERT_EQ(sum_op->InputNames().size(), 1UL); + ASSERT_EQ(sum_op->OutputNames().size(), 1UL); + EXPECT_EQ(sum_op->Input("X"), + std::vector({f::GradVarName("out1") + "@RENAME@0", + f::GradVarName("out1") + "@RENAME@1"})); + EXPECT_EQ(sum_op->Output("Out"), + std::vector({f::GradVarName("out1")})); + + f::OpDescBind *grad_op1 = block->AllOps()[6]; + ASSERT_EQ(grad_op1->Type(), "rowwise_add_grad"); + ASSERT_EQ(grad_op1->InputNames().size(), 1UL); + ASSERT_EQ(grad_op1->OutputNames().size(), 2UL); + EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")), + std::vector({f::GradVarName("out1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("X")), + std::vector({f::GradVarName("x1")})); + EXPECT_EQ(grad_op1->Output(f::GradVarName("b")), + std::vector({f::GradVarName("b1")})); +} \ No newline at end of file diff --git a/paddle/framework/block_desc.cc b/paddle/framework/block_desc.cc new file mode 100644 index 0000000000..01f50e1393 --- /dev/null +++ b/paddle/framework/block_desc.cc @@ -0,0 +1,93 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/block_desc.h" +#include "paddle/framework/program_desc.h" + +namespace paddle { +namespace framework { + +VarDescBind *BlockDescBind::NewVar(const std::string &name) { + need_update_ = true; + auto it = vars_.find(name); + PADDLE_ENFORCE(it == vars_.end(), "Duplicated variable %s", name); + auto var = new VarDescBind(name); + vars_[name].reset(var); + return var; +} + +VarDescBind *BlockDescBind::Var(const std::string &name) const { + auto it = vars_.find(name); + PADDLE_ENFORCE(it != vars_.end(), + "Can not find variable %s in current block.", name); + return it->second.get(); +} + +bool BlockDescBind::HasVar(const std::string &name) const { + return vars_.find(name) != vars_.end(); +} + +std::vector BlockDescBind::AllVars() const { + std::vector res; + for (const auto &p : vars_) { + res.push_back(p.second.get()); + } + return res; +} + +OpDescBind *BlockDescBind::AppendOp() { + need_update_ = true; + ops_.emplace_back(new OpDescBind()); + return ops_.back().get(); +} + +OpDescBind *BlockDescBind::PrependOp() { + need_update_ = true; + ops_.emplace_front(new OpDescBind()); + return ops_.front().get(); +} + +std::vector BlockDescBind::AllOps() const { + std::vector res; + for (const auto &op : ops_) { + res.push_back(op.get()); + } + return res; +} + +void BlockDescBind::Sync() { + if (need_update_) { + auto &op_field = *this->desc_->mutable_ops(); + op_field.Clear(); + op_field.Reserve(static_cast(ops_.size())); + for (auto &op_desc : ops_) { + op_field.AddAllocated(op_desc->Proto()); + } + need_update_ = false; + } +} + +BlockDescBind *BlockDescBind::ParentBlock() const { + if (this->desc_->parent_idx() == -1) { + return nullptr; + } + return prog_->Block(static_cast(this->desc_->parent_idx())); +} + +void OpDescBind::SetBlockAttr(const std::string &name, BlockDescBind &block) { + BlockDesc *desc = block.RawPtr(); + this->attrs_[name] = desc; +} +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/block_desc.h b/paddle/framework/block_desc.h new file mode 100644 index 0000000000..2de270f60e --- /dev/null +++ b/paddle/framework/block_desc.h @@ -0,0 +1,81 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include +#include "paddle/framework/op_desc.h" +#include "paddle/framework/var_desc.h" +#include "paddle/platform/macros.h" + +namespace paddle { +namespace framework { + +class ProgramDescBind; + +// Each Protobuf Message, we provide a XXXBind class. In that class, we optimize +// read/write speed. Only when we want the protobuf message, the local changes +// will be synchronized (by `Sync` method). + +class BlockDescBind { + public: + friend std::vector> MakeBlockBackward( + ProgramDescBind &program_desc, int block_idx, + std::unordered_set &no_grad_vars); + + friend void AppendBackward( + ProgramDescBind &program_desc, + const std::unordered_set &no_grad_vars); + + BlockDescBind(ProgramDescBind *prog, BlockDesc *desc) + : prog_(prog), desc_(desc), need_update_(false) {} + + int32_t ID() const { return desc_->idx(); } + + int32_t Parent() const { return desc_->parent_idx(); } + + VarDescBind *NewVar(const std::string &name_bytes); + + VarDescBind *Var(const std::string &name_bytes) const; + + bool HasVar(const std::string &var_name) const; + + std::vector AllVars() const; + + BlockDescBind *ParentBlock() const; + + OpDescBind *AppendOp(); + + OpDescBind *PrependOp(); + + std::vector AllOps() const; + + void Sync(); + + BlockDesc *RawPtr() { return desc_; } + + private: + ProgramDescBind *prog_; // not_own + BlockDesc *desc_; // not_own + bool need_update_; + + std::deque> ops_; + std::unordered_map> vars_; + + DISABLE_COPY_AND_ASSIGN(BlockDescBind); +}; +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/data_type.h b/paddle/framework/data_type.h new file mode 100644 index 0000000000..55e3931f87 --- /dev/null +++ b/paddle/framework/data_type.h @@ -0,0 +1,36 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include +#include "paddle/framework/framework.pb.h" + +namespace paddle { +namespace framework { + +inline DataType ToDataType(std::type_index type) { + if (typeid(float).hash_code() == type.hash_code()) { + return DataType::FP32; + } else if (typeid(double).hash_code() == type.hash_code()) { + return DataType::FP64; + } else if (typeid(int).hash_code() == type.hash_code()) { + return DataType::INT32; + } else { + PADDLE_THROW("Not supported"); + return static_cast(-1); + } +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/ddim.cc b/paddle/framework/ddim.cc index fc3d508553..a335786753 100644 --- a/paddle/framework/ddim.cc +++ b/paddle/framework/ddim.cc @@ -292,5 +292,13 @@ DDim flatten_to_2d(const DDim& src, int num_col_dims) { DDim flatten_to_1d(const DDim& src) { return make_ddim({product(src)}); } +DDim stride(const DDim& ddim) { + std::vector strides(ddim.size()); + strides[ddim.size() - 1] = 1; + for (int i = ddim.size() - 2; i >= 0; --i) { + strides[i] = strides[i + 1] * ddim[i + 1]; + } + return framework::make_ddim(strides); +} } // namespace framework } // namespace paddle diff --git a/paddle/framework/ddim.h b/paddle/framework/ddim.h index ca29e7e8c7..4a871bb0a9 100644 --- a/paddle/framework/ddim.h +++ b/paddle/framework/ddim.h @@ -121,6 +121,7 @@ DDim flatten_to_2d(const DDim& src, int num_col_dims); DDim flatten_to_1d(const DDim& src); +DDim stride(const DDim& ddim); } // namespace framework } // namespace paddle diff --git a/paddle/framework/details/op_registry.h b/paddle/framework/details/op_registry.h new file mode 100644 index 0000000000..daa474e8c5 --- /dev/null +++ b/paddle/framework/details/op_registry.h @@ -0,0 +1,109 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/grad_op_desc_maker.h" +#include "paddle/framework/op_info.h" +#include "paddle/framework/op_proto_maker.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace framework { +namespace details { + +enum OpInfoFillType { + kOperator = 0, + kOpProtoAndCheckerMaker = 1, + kGradOpDescMaker = 2 +}; + +template +struct OpInfoFillTypeID { + static constexpr OpInfoFillType ID() { + return std::is_base_of::value + ? kOperator + : (std::is_base_of::value + ? kOpProtoAndCheckerMaker + : (std::is_base_of::value + ? kGradOpDescMaker + : static_cast(-1))); + } +}; + +template ::ID()> +struct OpInfoFiller; + +template +class OperatorRegistrarRecursive; + +template +class OperatorRegistrarRecursive { + public: + using T = typename std::tuple_element>::type; + OperatorRegistrarRecursive(const char* op_type, OpInfo* info) { + OpInfoFiller fill; + fill(op_type, info); + constexpr auto size = sizeof...(ARGS); + OperatorRegistrarRecursive reg(op_type, + info); + (void)(reg); + } +}; + +template +class OperatorRegistrarRecursive { + public: + OperatorRegistrarRecursive(const char* op_type, OpInfo* info) {} +}; + +template +struct OpInfoFiller { + void operator()(const char* op_type, OpInfo* info) const { + info->creator_ = [](const std::string& type, const VariableNameMap& inputs, + const VariableNameMap& outputs, + const AttributeMap& attrs) { + return new T(type, inputs, outputs, attrs); + }; + } +}; + +template +struct OpInfoFiller { + void operator()(const char* op_type, OpInfo* info) const { + info->proto_ = new OpProto; + info->checker_ = new OpAttrChecker(); + auto maker = T(info->proto_, info->checker_); + maker.Validate(); + info->proto_->set_type(op_type); + PADDLE_ENFORCE( + info->proto_->IsInitialized(), + "Fail to initialize %s's OpProto, because %s is not initialized", + op_type, info->proto_->InitializationErrorString()); + } +}; + +template +struct OpInfoFiller { + void operator()(const char* op_type, OpInfo* info) const { + info->grad_op_maker_ = [](const OpDescBind& fwd_op) { + T maker(fwd_op); + return maker(); + }; + } +}; +} // namespace details + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/framework.proto b/paddle/framework/framework.proto index dfcb5fb621..ac2827e547 100644 --- a/paddle/framework/framework.proto +++ b/paddle/framework/framework.proto @@ -22,14 +22,11 @@ enum AttrType { INTS = 3; FLOATS = 4; STRINGS = 5; - INT_PAIRS = 6; + BOOLEAN = 6; + BOOLEANS = 7; + BLOCK = 8; } -message IntPair { - required int32 first = 1; - required int32 second = 2; -}; - // OpDesc describes an instance of a C++ framework::OperatorBase // derived class type. message OpDesc { @@ -43,7 +40,9 @@ message OpDesc { repeated int32 ints = 6; repeated float floats = 7; repeated string strings = 8; - repeated IntPair int_pairs = 9; + optional bool b = 10; + repeated bool bools = 11; + optional int32 block_idx = 12; }; message Var { @@ -67,7 +66,6 @@ message OpProto { optional bool duplicable = 3 [ default = false ]; optional bool intermediate = 4 [ default = false ]; - optional bool not_in_gradient = 5 [ default = false ]; } // AttrProto describes the C++ type Attribute. @@ -100,11 +98,24 @@ enum DataType { message LoDTensorDesc { required DataType data_type = 1; - repeated int32 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480] + repeated int64 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480] optional int32 lod_level = 3 [ default = 0 ]; } message VarDesc { required string name = 1; optional LoDTensorDesc lod_tensor = 2; + optional bool persistable = 3 [ default = false ]; } + +message BlockDesc { + required int32 idx = 1; + required int32 parent_idx = 2; + repeated VarDesc vars = 3; + repeated OpDesc ops = 4; +} + +// Please refer to +// https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md +// for more details. +message ProgramDesc { repeated BlockDesc blocks = 1; } diff --git a/paddle/framework/grad_op_builder.cc b/paddle/framework/grad_op_builder.cc deleted file mode 100644 index b02a599a80..0000000000 --- a/paddle/framework/grad_op_builder.cc +++ /dev/null @@ -1,58 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOpArgType::OUT WARRANTIES OR CONDITIONS OF ANY KOpArgType::IND, either -express or implied. See the License for the specific language governing -permissions and limitations under the License. */ - -#include "paddle/framework/grad_op_builder.h" -#include "paddle/framework/op_registry.h" - -namespace paddle { -namespace framework { -enum class OpArgType { IN, OUT }; - -static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type, - bool is_grad, VariableNameMap* vars) { - const auto& src_inout = - src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs(); - auto& dst_inout = *vars; - auto& proto = OpInfoMap::Instance().Get(src_op->Type()).Proto(); - const auto& src_arg_list = - src_type == OpArgType::IN ? proto.inputs() : proto.outputs(); - for (const auto& arg : src_arg_list) { - if (arg.not_in_gradient() && !is_grad) continue; - const std::string src_name = arg.name(); - std::string dst_name = is_grad ? GradVarName(src_name) : src_name; - dst_inout[dst_name].reserve(src_inout.at(src_name).size()); - for (auto& var_name : src_inout.at(src_name)) { - std::string s = is_grad ? GradVarName(var_name) : var_name; - dst_inout[dst_name].emplace_back(s); - } - } -} - -OperatorBase* BuildGradOp(const OperatorBase* op) { - auto& info = OpInfoMap::Instance().Get(op->Type()); - PADDLE_ENFORCE(info.HasGradientOp()); - - VariableNameMap inputs; - VariableNameMap outputs; - TransOpArg(op, OpArgType::IN, false, &inputs); // I - TransOpArg(op, OpArgType::OUT, false, &inputs); // O - TransOpArg(op, OpArgType::OUT, true, &inputs); // OG - TransOpArg(op, OpArgType::IN, true, &outputs); // IG - - auto& grad_info = OpInfoMap::Instance().Get(info.grad_op_type_); - return grad_info.Creator()(info.grad_op_type_, inputs, outputs, op->Attrs()); -} - -} // namespace framework -} // namespace paddle diff --git a/paddle/framework/grad_op_builder_test.cc b/paddle/framework/grad_op_builder_test.cc deleted file mode 100644 index 9e3ca563c6..0000000000 --- a/paddle/framework/grad_op_builder_test.cc +++ /dev/null @@ -1,122 +0,0 @@ -#include "paddle/framework/grad_op_builder.h" -#include -#include "paddle/framework/op_registry.h" -#include "paddle/framework/operator.h" - -USE_OP(add); - -namespace paddle { -namespace framework { - -class MutiInOutOpMaker : public OpProtoAndCheckerMaker { - public: - MutiInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("In1", "a single input"); - AddInput("In2_mult", "a multiple input").AsDuplicable(); - AddInput("In3", "another single input"); - AddOutput("Out1", "a single output"); - AddOutput("Out2_mult", "a multiple output").AsDuplicable(); - AddComment("test op with multiple inputs and outputs"); - } -}; - -class IOIgnoredOpMaker : public OpProtoAndCheckerMaker { - public: - IOIgnoredOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("In1", "a single input"); - AddInput("In2_mult", "a multiple input").AsDuplicable().NotInGradient(); - AddInput("In3_mult", "another multiple input").AsDuplicable(); - AddOutput("Out1_mult", "a multiple output").AsDuplicable(); - AddOutput("Out2", "a single output").NotInGradient(); - AddComment("op with inputs and outputs ignored in gradient calculating"); - } -}; - -} // namespace framework -} // namespace paddle - -namespace f = paddle::framework; - -TEST(GradOpBuilder, AddTwo) { - std::shared_ptr add_op(f::OpRegistry::CreateOp( - "add", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {})); - std::shared_ptr grad_add_op = - f::OpRegistry::CreateGradOp(*add_op); - EXPECT_EQ(grad_add_op->Inputs().size(), 4UL); - EXPECT_EQ(grad_add_op->Outputs().size(), 2UL); - EXPECT_EQ(grad_add_op->Input("X"), "x"); - EXPECT_EQ(grad_add_op->Input("Y"), "y"); - EXPECT_EQ(grad_add_op->Input("Out"), "out"); - EXPECT_EQ(grad_add_op->Input(f::GradVarName("Out")), f::GradVarName("out")); - EXPECT_EQ(grad_add_op->Output(f::GradVarName("X")), f::GradVarName("x")); - EXPECT_EQ(grad_add_op->Output(f::GradVarName("Y")), f::GradVarName("y")); -} - -REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, mult_io_grad, f::NOP); -REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker, io_ignored_grad, f::NOP); - -TEST(GradOpBuilder, MutiInOut) { - std::shared_ptr test_op(f::OpRegistry::CreateOp( - "mult_io", {{"In1", {"in1"}}, - {"In2_mult", {"in2_1", "in2_2", "in2_3"}}, - {"In3", {"in3"}}}, - {{"Out1", {"out1"}}, {"Out2_mult", {"out2_1", "out2_2"}}}, {})); - std::shared_ptr grad_test_op = - f::OpRegistry::CreateGradOp(*test_op); - - ASSERT_EQ(grad_test_op->Inputs().size(), 3UL + 2UL + 2UL); - EXPECT_EQ(grad_test_op->Input("In1"), "in1"); - EXPECT_EQ(grad_test_op->Inputs("In2_mult"), - std::vector({"in2_1", "in2_2", "in2_3"})); - EXPECT_EQ(grad_test_op->Input("In3"), "in3"); - EXPECT_EQ(grad_test_op->Input("Out1"), "out1"); - EXPECT_EQ(grad_test_op->Inputs("Out2_mult"), - std::vector({"out2_1", "out2_2"})); - EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out1")), - f::GradVarName("out1")); - EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out2_mult")), - std::vector( - {f::GradVarName("out2_1"), f::GradVarName("out2_2")})); - - ASSERT_EQ(grad_test_op->Outputs().size(), 3UL); - EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1")); - EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")), - std::vector({f::GradVarName("in2_1"), - f::GradVarName("in2_2"), - f::GradVarName("in2_3")})); - EXPECT_EQ(grad_test_op->Output(f::GradVarName("In3")), f::GradVarName("in3")); -} - -TEST(GradOpBuilder, IOIgnoredInGradient) { - std::shared_ptr test_op(f::OpRegistry::CreateOp( - "io_ignored", {{"In1", {"in1"}}, - {"In2_mult", {"in2_1", "in2_2"}}, - {"In3_mult", {"in3_1", "in3_2"}}}, - {{"Out1_mult", {"out1_1", "out1_2"}}, {"Out2", {"out2"}}}, {})); - std::shared_ptr grad_test_op = - f::OpRegistry::CreateGradOp(*test_op); - - // 'In2' and 'Out2' are ignored in gradient calculating - ASSERT_EQ(grad_test_op->Inputs().size(), 2UL + 1UL + 2UL); - EXPECT_EQ(grad_test_op->Input("In1"), "in1"); - EXPECT_EQ(grad_test_op->Inputs("In3_mult"), - std::vector({"in3_1", "in3_2"})); - EXPECT_EQ(grad_test_op->Inputs("Out1_mult"), - std::vector({"out1_1", "out1_2"})); - EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out1_mult")), - std::vector( - {f::GradVarName("out1_1"), f::GradVarName("out1_2")})); - EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out2")), - f::GradVarName("out2")); - - ASSERT_EQ(grad_test_op->Outputs().size(), 3UL); - EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1")); - EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")), - std::vector( - {f::GradVarName("in2_1"), f::GradVarName("in2_2")})); - EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In3_mult")), - std::vector( - {f::GradVarName("in3_1"), f::GradVarName("in3_2")})); -} diff --git a/paddle/framework/grad_op_desc_maker.h b/paddle/framework/grad_op_desc_maker.h new file mode 100644 index 0000000000..e9ae6e2206 --- /dev/null +++ b/paddle/framework/grad_op_desc_maker.h @@ -0,0 +1,124 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/op_desc.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace framework { + +class GradOpDescMakerBase { + public: + explicit GradOpDescMakerBase(const OpDescBind& fwd_op) : fwd_op_(fwd_op) {} + + virtual ~GradOpDescMakerBase() = default; + virtual std::vector> operator()() const = 0; + + protected: + static std::vector ToGradNames( + const std::vector& var_names) { + std::vector ret_val; + ret_val.reserve(var_names.size()); + std::transform(var_names.begin(), var_names.end(), + std::back_inserter(ret_val), GradVarName); + return ret_val; + } + + std::vector InputGrad(const std::string& name) const { + return ToGradNames(fwd_op_.Input(name)); + } + + std::vector OutputGrad(const std::string& name) const { + return ToGradNames(fwd_op_.Output(name)); + } + + std::vector InputNames() const { + return this->fwd_op_.InputNames(); + } + + std::vector OutputNames() const { + return this->fwd_op_.OutputNames(); + } + + std::vector Input(const std::string& name) const { + return fwd_op_.Input(name); + } + + std::vector Output(const std::string& name) const { + return fwd_op_.Output(name); + } + + const std::unordered_map& Attrs() const { + return fwd_op_.GetAttrMap(); + } + + const Attribute& GetAttr(const std::string& name) const { + auto& map = fwd_op_.GetAttrMap(); + auto it = map.find(name); + PADDLE_ENFORCE(it != map.end(), "Cannot find attribute %s", name); + return it->second; + } + + std::string ForwardOpType() const { return this->fwd_op_.Type(); } + + private: + const OpDescBind& fwd_op_; +}; + +class SingleGradOpDescMaker : public GradOpDescMakerBase { + public: + using GradOpDescMakerBase::GradOpDescMakerBase; + + std::vector> operator()() const { + std::vector> retv; + retv.emplace_back(this->Apply()); + return retv; + } + + protected: + virtual std::unique_ptr Apply() const = 0; +}; + +class DefaultGradOpDescMaker : public SingleGradOpDescMaker { + public: + using SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + virtual std::unique_ptr Apply() const { + auto* grad = new OpDescBind(); + grad->SetType(this->GradOpType()); + + for (auto& input_param : this->InputNames()) { + grad->SetInput(input_param, this->Input(input_param)); + grad->SetOutput(GradVarName(input_param), this->InputGrad(input_param)); + } + + for (auto& output_param : this->OutputNames()) { + grad->SetInput(output_param, this->Output(output_param)); + grad->SetInput(GradVarName(output_param), this->OutputGrad(output_param)); + } + + grad->SetAttrMap(this->Attrs()); + + return std::unique_ptr(grad); + } + + virtual std::string GradOpType() const { + return this->ForwardOpType() + "_grad"; + } +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc index 908a1f2fd0..5b7badf89c 100644 --- a/paddle/framework/lod_tensor.cc +++ b/paddle/framework/lod_tensor.cc @@ -72,20 +72,32 @@ bool operator==(const LoD& a, const LoD& b) { return true; } -void LoDTensor::SliceLevels(size_t level_begin, size_t level_end) { +size_t LoDTensor::NumElements(size_t level, size_t idx) const { + PADDLE_ENFORCE_LT(level, NumLevels()); + PADDLE_ENFORCE_LT(idx, NumElements(level)); + // the last level of LoD, just return number of records in Tensor + if (level == NumLevels() - 1) { + return lod_[level][idx + 1] - lod_[level][idx]; + } + // high level of LoD, and there is another lower level, return number of + // lower-level elements + auto tmp = SliceInLevel(lod_, level, idx, idx + 1); + PADDLE_ENFORCE_GE(tmp.size(), 2); + // there is a 0 as a placeholder stored in LoD, so the number of elements + // equals lod.size() - 1 + return tmp[1].size() - 1; +} + +void LoDTensor::ShrinkLevels(size_t level_begin, size_t level_end) { auto new_lod = framework::SliceLevels(lod_, level_begin, level_end); lod_ = new_lod; } -void LoDTensor::SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) { - PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level, - NumLevels()); - PADDLE_ENFORCE(elem_begin < NumElements(level), - "element begin [%d] out of range [%d]", elem_begin, - NumElements(level)); - PADDLE_ENFORCE(elem_end < NumElements(level) + 1, - "element end [%d] out of range [%d]", elem_end, - NumElements(level)); +void LoDTensor::ShrinkInLevel(size_t level, size_t elem_begin, + size_t elem_end) { + PADDLE_ENFORCE_LT(level, NumLevels()); + PADDLE_ENFORCE_LT(elem_begin, NumElements(level)); + PADDLE_ENFORCE_LT(elem_end, NumElements(level) + 1); auto new_lod = framework::SliceInLevel(lod_, level, elem_begin, elem_end); lod_ = new_lod; diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h index fac5cd20aa..4db36ee766 100644 --- a/paddle/framework/lod_tensor.h +++ b/paddle/framework/lod_tensor.h @@ -15,7 +15,7 @@ #pragma once #include -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include #include #include @@ -29,7 +29,7 @@ namespace paddle { namespace framework { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA template using Vector = std::vector; #else @@ -38,6 +38,18 @@ using Vector = thrust::host_vector< T, thrust::system::cuda::experimental::pinned_allocator>; #endif +/* + * 3-level LoD stores + * + * 0 10 20 + * 0 5 10 15 20 + * 0 2 5 7 10 12 15 20 + * + * - in a level, each element indicates offset in the underlying Tensor + * - the first element should be 0 and that indicates that this sequence start + * from 0 + * - each sequence's begin and end(no-inclusive) is level[id, id+1] + */ using LoD = std::vector>; LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end); @@ -65,11 +77,8 @@ class LoDTensor : public Tensor { * Get a element from LoD. */ size_t lod_element(size_t level, size_t elem) const { - PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level, - NumLevels()); - PADDLE_ENFORCE(elem < NumElements(level), - "element begin [%d] out of range [%d]", elem, - NumElements(level)); + PADDLE_ENFORCE_LT(level, NumLevels()); + PADDLE_ENFORCE_LT(elem, NumElements(level)); return (lod_)[level][elem]; } @@ -82,22 +91,33 @@ class LoDTensor : public Tensor { * Number of elements in a level. */ size_t NumElements(size_t level = 0) const { - PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level, - NumLevels()); + PADDLE_ENFORCE_LT(level, NumLevels()); // the last offset is the end of last element return (lod_)[level].size() - 1; } /* - * Slice of levels[level_begin:level_end] + * Number of lower-level elements. + * For example, a 2-level lod-tensor + * + * 0-th level | | + * 1-th level || ||| + * + * NumElements(0, 0) get 2 + * NumElements(0, 1) get 3 + */ + size_t NumElements(size_t level, size_t idx) const; + + /* + * Shrink levels[level_begin:level_end] */ - void SliceLevels(size_t level_begin, size_t level_end); + void ShrinkLevels(size_t level_begin, size_t level_end); /* - * Slice of elements of a level, [elem_begin: elem_end] + * Shrink elements of a level, [elem_begin: elem_end] * @note: low performance in slice lod_. */ - void SliceInLevel(size_t level, size_t elem_begin, size_t elem_end); + void ShrinkInLevel(size_t level, size_t elem_begin, size_t elem_end); private: LoD lod_; diff --git a/paddle/framework/lod_tensor.md b/paddle/framework/lod_tensor.md index 769b61f175..d147f1c425 100644 --- a/paddle/framework/lod_tensor.md +++ b/paddle/framework/lod_tensor.md @@ -1,146 +1,175 @@ # Design Doc: LoD (Level-of-Detail) Tensor -PaddlePaddle's RNN doesn't require that all instances have the same length. To do so, we introduce an extension to Tensor, namely, LoD Tensor. +Like other deep learning systems, PaddlePaddle supports training models from sequence data. Also, like other systems, PaddlePaddle represent a mini-batch of sequences as a Tensor. What is different is that PaddlePaddle doesn't require all sequences in a mini-batch to be of the same length. Thus no need for padding zeros. -## Challenge of Variable-length Inputs +| | TensorFlow | PaddlePaddle | +|-----------------------|------------|--------------| +| RNN | Support | Support | +| recursive RNN | Support | Support | +| padding zeros | Must | No need | +| blob data type | Tensor | LoDTensor | -People usually represent a mini-batch by a Tensor. For example, a mini-batch of 32 images, each of size 32x32, is a 10x32x32 Tensor. So a transformation, T, of all images can be a matrix multiplication of the 32x32xO-dimensional tensor T and the 10x32x32 Tensor. +PaddlePaddle achieves this flexibility by passing through a new data type, *LoD Tensor*, which is a Tensor attached with segmentation index known as *LoD*, between operators. The LoD index doesn't only segment a tensor, but also recursively segments sub-sequences. This document presents the design of LoD and LoDTensor. -Another example is that each mini-batch contains 32 sentences, where each word is a D-dimensional one-hot vector. If all sentences have the same length L, we can represent this mini-batch by a 32xLxD tensor. However, in most cases, sentences have variable lengths, and we will need an index data structure to record these variable lengths. -## LoD as a Solution +## The Challenge: Variable-length Sequences -### Mini-Batch of variable-length sentenses +Most deep learning systems represent a mini-batch as a Tensor. For example, a mini-batch of 10 images, each of size 32x32, is a 10x32x32 Tensor. Another example is that each mini-batch contains N sentences, where each word is a D-dimensional one-hot vector. Suppose that all sentences have the same length L, we can represent this mini-batch by a NxLxD tensor. -Let's imagine a mini-batch of 3 variable lengths sentences, containing 3, 1, and 2 words respectively. We can represent it by a (3+1+2)xD tensor plus some index information: +Both examples show that the elements of sequences are usually of the same size. In the first example, all images are 32x32, and in the second one, all words are D-dimensional vectors. It doesn't make sense to allow variable-sized images, as that would require transformations like convolution to handle variable-sized Tensors. + +The real challenge is that in most cases, sentences have variable lengths, and we will need an index data structure to segment the tensor into sequences. Also, sequences might consist of sub-sequences. + + +## A Solution: The LoD Index + +To understand our solution, it is best to look at some examples. + +### A Mini-Batch of Sentences + +Let's imagine a mini-batch of 3 variable lengths sentences composed of 3, 1, and 2 words, respectively. We can represent the mini-batch by a (3+1+2)xD tensor plus some index information: ``` - 3 3 1 2 ||| | || ``` -Each `|` represents a D-dimensional word vectors. The number 3 on top indicate 3 sentences, and numbers 3, 1, and 2 on the second level represent the number of words in each sentence. +where each `|` represents a D-dimensional word vector. The numbers, 3, 1, and 2, form a 1-level LoD. + +### Recursive Sequences + +Let check another example of a 2-level LoD Tensor. Consider a mini-batch of three articles with 3, 1, and 2 sentences, and each sentence consists of a variable number of words: + +``` +3 1 2 +3 2 4 1 2 3 +||| || |||| | || ||| +``` -### Mini-Batch of variable-length videos +### A Mini-Batch of Videos -This approach generalizes to the case where elements are not words, but higher dimensional objects, like images. Suppose that a mini-batch contains videos of the same frame size 640x480. If a mini-batch contains 3 videos of 3, 1, and 2 frames respectively. The underlying tensor is of size (3+1+2)x640x480. The index information illustrates as: +LoD tensors generalize to the case where elements are higher dimensional objects, like images. Suppose that a mini-batch contains videos of the same frame size 640x480. Here is a mini-batch of 3 videos with 3, 1, and 2 frames, respectively. ``` - 3 3 1 2 口口口 口 口口 ``` -where each `口` represents an image. +The underlying tensor is of size (3+1+2)x640x480, and each `口` represents a 640x480 image. -### Mini-Batch of fixed-size images +### A Mini-Batch of Images -Let's get back to a typical example, image classification, where each mini-batch has M fixed-sized images. The LoD Tensor representation is +In traditional cases like a mini-batch with N fixed-sized images, the LoD Tensor representation is as ``` - M 1 1 1 1 1 口口口口 ... 口 ``` -The many 1's on the second level seem duplicated. For this particular case of 2 levels and the second level always have length 1, we can ignore the LoD index. - -### Design and summarization +In this case, we don't lose any information by ignoring the many 1's in the index and simply considering this LoD Tensor as a usual Tensor: -In summary, as long as that the essential elements (words or images) have the same size, we can represent mini-batches by a LoD Tensor: +``` +口口口口 ... 口 +``` -- The underlying tensor has size LxD1xD2x..., where D1xD2... is the size of the essential elements, and -- the first dimension size L has an additon property -- a LoD index as a nested vector: +### Model Parameters - ```c++ - typedef std::vector > LoD; - ``` +A model parameter is just a usual Tensor, which, just like the above example, is a **0-level LoD Tensor**. -- The LoD index can is not necessary when there are only two levels and all elements of the second level have length 1. -## Slicing of LoD Tensor +## The LoD Tensor -Consider that we have a network with three levels of RNN: the top level one handles articles, the second level one handles sentences, and the basic level one handles words. This network requires that mini-batches represented by 4 level LoD Tensor, for example, +Let us revisit above example of the 2-level LoD Tensor ``` - 3 3 1 2 3 2 4 1 2 3 ||| || |||| | || ||| ``` -To allow each level of RNN to handle its input, we define **the slicing of a LoD Tensor is defined as getting the j-th sequence on level i, or the -slice** +It is indeed a tree, where leaves are elementary sequences identified by **branches**. + +For example, the third sentence in above example is identified by branch <0,2>, where 0 indicates the first article with length 3, and 2 indicates the third sentence in this article with length 4. + +### The LoD Index -For example, the <2,1>-slice of above slice is +We can save the LoD index in the above example ``` -2 -|| +3 1 2 +3 2 4 1 2 3 ``` -and the <1,2>-slice of above example is +in a not-full 2D matrix: +```c++ +typedef std::vector > LoD; ``` -2 -2 3 -|| ||| -``` -Let's go on slicing this slice. Its <1,1>-slice is +where + +- `LoD.size()` is the number of levels, or the maximum length of branches, +- `LoD[i][j]` is the length of the j-th segment at the i-th level. + +## The Offset Representation + +To quickly access elementary sequences, we adopt an offset representation -- instead of saving the lengths, we save the beginning and ending elements of sequences. + +In the above example, we accumulate the length of elementary sequences: ``` -3 -||| +3 2 4 1 2 3 ``` -### The Slicing Algorithm +into offsets -The algorithm, with over-simplified data structure, is defined as +``` +0 3 5 9 10 12 15 + = = = = = = + 3 2+3 4+5 1+9 2+10 3+12 +``` -```c++ -typedef vector > LoD; +so we know that the first sentence is from word 0 to word 3, and the second sentence from work 3 to word 5. -struct LoDTensor { - LoD lod_; - float* tensor_; -}; +Similarly, the lengths in the top level LoD -LoDTensor Slice(const LoDTensor& lodt, int level, int sequence); +``` +3 1 2 ``` -Let us revisit the example above +are transformed into offsets of elements/words as follows: ``` - 3 -3 1 2 -3 2 4 1 2 3 -||| || |||| | || ||| +0 9 10 15 + = = = + 3+2+4 1+9 2+3+10 ``` -Suppose that we want to retrieve the <1,2>-slice +so we can tell that the first article is from word 0 to word 9, and the second article is from word 9 to word 10. + +The complete offset representation is as follows: ``` -2 -2 3 -|| ||| +0 9 10 15 +0 3 5 9 10 12 15 + ||| || |||| | || ||| ``` -we will need to find out the starting position of this slice by summing over all leaf nodes in `LoD` to the left of the slice, i.e., 3 + 2 + 4 + 1 = 10. +## Slicing of LoD Tensors + +When we use the above 2-level LoD Tensor as the input to a nested-RNN, we need to retrieve certain sequences. Here we define the sequence identified by branch as the **-slice**. -To avoid the traversal of the LoD tree at slcing time, we can do it at the construction time -- instead of saving the lengths of the next level in the LoD tree, we can save the starting offset of the next level. For example, above LoD Tensor can be transformed into +For example, the <2>-slice of above example is ``` - 0 -0 9 10 -0 3 5 9 10 12 -||| || |||| | || ||| +10 15 +10 12 15 + || ||| ``` -We don't really need the 0 on top, so the LoD Tensor could be +and the <2,0>-slice of above slice is ``` -0 9 10 -0 3 5 9 10 12 -||| || |||| | || ||| +10 12 + || ``` diff --git a/paddle/framework/lod_tensor_test.cc b/paddle/framework/lod_tensor_test.cc index 7915326b27..44f09f584f 100644 --- a/paddle/framework/lod_tensor_test.cc +++ b/paddle/framework/lod_tensor_test.cc @@ -56,19 +56,25 @@ TEST_F(LoDTensorTester, NumElements) { ASSERT_EQ(lod_tensor_.NumElements(2), 8UL); } -TEST_F(LoDTensorTester, SliceLevels) { +TEST_F(LoDTensorTester, NumElements2) { + ASSERT_EQ(lod_tensor_.NumElements(0, 0), 2UL); + ASSERT_EQ(lod_tensor_.NumElements(0, 1), 2UL); + ASSERT_EQ(lod_tensor_.NumElements(1, 1), 2UL); +} + +TEST_F(LoDTensorTester, ShrinkLevels) { // slice 1 level for (size_t level = 0; level < 3UL; ++level) { LoDTensor new_lod_tensor = lod_tensor_; - new_lod_tensor.SliceLevels(level, level + 1); + new_lod_tensor.ShrinkLevels(level, level + 1); ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL); ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level)); ASSERT_EQ(new_lod_tensor.data(), lod_tensor_.data()); } - // slice 2 level + // shrink 2 level for (size_t level = 0; level < 2UL; ++level) { LoDTensor new_lod_tensor = lod_tensor_; - new_lod_tensor.SliceLevels(level, level + 2); + new_lod_tensor.ShrinkLevels(level, level + 2); ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL); ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level)); ASSERT_EQ(new_lod_tensor.NumElements(1), @@ -77,10 +83,10 @@ TEST_F(LoDTensorTester, SliceLevels) { } } -TEST_F(LoDTensorTester, SliceInLevel) { +TEST_F(LoDTensorTester, ShrinkInLevel) { size_t level = 0; LoDTensor new_lod_tensor = lod_tensor_; - new_lod_tensor.SliceInLevel(level, 0, 2); + new_lod_tensor.ShrinkInLevel(level, 0, 2); EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL); EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL); EXPECT_EQ(new_lod_tensor.NumElements(1), 4UL); @@ -89,7 +95,7 @@ TEST_F(LoDTensorTester, SliceInLevel) { level = 1; new_lod_tensor = lod_tensor_; - new_lod_tensor.SliceInLevel(level, 0, 2); + new_lod_tensor.ShrinkInLevel(level, 0, 2); ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL); ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL); ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL); diff --git a/paddle/framework/lod_tensor_test.cu b/paddle/framework/lod_tensor_test.cu index 97e69cdb2e..647d07536d 100644 --- a/paddle/framework/lod_tensor_test.cu +++ b/paddle/framework/lod_tensor_test.cu @@ -36,8 +36,8 @@ TEST(LoDTensor, LoDInGPU) { lod_tensor.mutable_data(place); lod_tensor.set_lod(src_lod); - CHECK_EQ(lod_tensor.lod_element(0, 2), 4); - CHECK_EQ(lod_tensor.lod_element(0, 4), 8); + CHECK_EQ(lod_tensor.lod_element(0, 2), 4UL); + CHECK_EQ(lod_tensor.lod_element(0, 4), 8UL); auto lod = lod_tensor.lod(); diff --git a/paddle/framework/op_desc.cc b/paddle/framework/op_desc.cc new file mode 100644 index 0000000000..02aa74a842 --- /dev/null +++ b/paddle/framework/op_desc.cc @@ -0,0 +1,188 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/op_desc.h" +#include "paddle/framework/block_desc.h" + +namespace paddle { +namespace framework { + +OpDescBind::OpDescBind(const std::string &type, const VariableNameMap &inputs, + const VariableNameMap &outputs, + const AttributeMap &attrs) { + op_desc_.set_type(type); + inputs_ = inputs; + outputs_ = outputs; + attrs_ = attrs; +} + +OpDesc *OpDescBind::Proto() { + Sync(); + return &op_desc_; +} + +const std::vector &OpDescBind::Input( + const std::string &name) const { + auto it = inputs_.find(name); + PADDLE_ENFORCE(it != inputs_.end(), "Input %s cannot be found in Op %s", name, + Type()); + return it->second; +} + +std::vector OpDescBind::InputArgumentNames() const { + std::vector retv; + for (auto &ipt : this->inputs_) { + retv.insert(retv.end(), ipt.second.begin(), ipt.second.end()); + } + return retv; +} + +void OpDescBind::SetInput(const std::string ¶m_name, + const std::vector &args) { + need_update_ = true; + inputs_[param_name] = args; +} + +const std::vector &OpDescBind::Output( + const std::string &name) const { + auto it = outputs_.find(name); + PADDLE_ENFORCE(it != outputs_.end(), "Output %s cannot be found in Op %s", + name, Type()); + return it->second; +} + +std::vector OpDescBind::OutputArgumentNames() const { + std::vector retv; + for (auto &ipt : this->outputs_) { + retv.insert(retv.end(), ipt.second.begin(), ipt.second.end()); + } + return retv; +} + +void OpDescBind::SetOutput(const std::string ¶m_name, + const std::vector &args) { + need_update_ = true; + this->outputs_[param_name] = args; +} + +AttrType OpDescBind::GetAttrType(const std::string &name) const { + auto it = attrs_.find(name); + PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); + return static_cast(it->second.which() - 1); +} + +std::vector OpDescBind::AttrNames() const { + std::vector retv; + retv.reserve(attrs_.size()); + for (auto &attr : attrs_) { + retv.push_back(attr.first); + } + return retv; +} + +void OpDescBind::SetAttr(const std::string &name, const Attribute &v) { + this->attrs_[name] = v; + need_update_ = true; +} + +void OpDescBind::SetAttrMap( + const std::unordered_map &attr_map) { + attrs_ = attr_map; + need_update_ = true; +} + +Attribute OpDescBind::GetAttr(const std::string &name) const { + auto it = attrs_.find(name); + PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); + return it->second; +} + +int OpDescBind::GetBlockAttr(const std::string &name) const { + auto it = attrs_.find(name); + PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name); + return boost::get(it->second)->idx(); +} + +const std::unordered_map &OpDescBind::GetAttrMap() + const { + return attrs_; +} + +void OpDescBind::Rename(const std::string &old_name, + const std::string &new_name) { + for (auto &input : inputs_) { + std::replace(input.second.begin(), input.second.end(), old_name, new_name); + } + for (auto &output : outputs_) { + std::replace(output.second.begin(), output.second.end(), old_name, + new_name); + } + need_update_ = true; +} + +struct SetAttrDescVisitor : public boost::static_visitor { + explicit SetAttrDescVisitor(OpDesc::Attr *attr) : attr_(attr) {} + mutable OpDesc::Attr *attr_; + void operator()(int v) const { attr_->set_i(v); } + void operator()(float v) const { attr_->set_f(v); } + void operator()(const std::string &v) const { attr_->set_s(v); } + void operator()(bool b) const { attr_->set_b(b); } + + void operator()(const std::vector &v) const { + VectorToRepeated(v, attr_->mutable_ints()); + } + void operator()(const std::vector &v) const { + VectorToRepeated(v, attr_->mutable_floats()); + } + void operator()(const std::vector &v) const { + VectorToRepeated(v, attr_->mutable_strings()); + } + void operator()(const std::vector &v) const { + VectorToRepeated(v, attr_->mutable_bools()); + } + void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->idx()); } + void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); } +}; + +void OpDescBind::Sync() { + if (need_update_) { + this->op_desc_.mutable_inputs()->Clear(); + for (auto &ipt : inputs_) { + auto *input = op_desc_.add_inputs(); + input->set_parameter(ipt.first); + VectorToRepeated(ipt.second, input->mutable_arguments()); + } + + this->op_desc_.mutable_outputs()->Clear(); + for (auto &opt : outputs_) { + auto *output = op_desc_.add_outputs(); + output->set_parameter(opt.first); + VectorToRepeated(opt.second, output->mutable_arguments()); + } + + this->op_desc_.mutable_attrs()->Clear(); + for (auto &attr : attrs_) { + auto *attr_desc = op_desc_.add_attrs(); + attr_desc->set_name(attr.first); + attr_desc->set_type( + static_cast(attr.second.which() - 1)); + SetAttrDescVisitor visitor(attr_desc); + boost::apply_visitor(visitor, attr.second); + } + + need_update_ = false; + } +} +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/op_desc.h b/paddle/framework/op_desc.h new file mode 100644 index 0000000000..b39808dad1 --- /dev/null +++ b/paddle/framework/op_desc.h @@ -0,0 +1,121 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "paddle/framework/attribute.h" +#include "paddle/framework/type_defs.h" +#include "paddle/framework/var_desc.h" + +namespace paddle { +namespace framework { + +class BlockDescBind; + +class OpDescBind { + public: + OpDescBind() {} + + OpDescBind(const std::string &type, const VariableNameMap &inputs, + const VariableNameMap &outputs, const AttributeMap &attrs); + + OpDesc *Proto(); + + std::string Type() const { return op_desc_.type(); } + + void SetType(const std::string &type) { op_desc_.set_type(type); } + + const std::vector &Input(const std::string &name) const; + + std::vector InputArgumentNames() const; + + void SetInput(const std::string ¶m_name, + const std::vector &args); + + const std::vector &Output(const std::string &name) const; + + std::vector OutputArgumentNames() const; + + void SetOutput(const std::string ¶m_name, + const std::vector &args); + + std::string DebugString() { return this->Proto()->DebugString(); } + + bool HasAttr(const std::string &name) const { + return attrs_.find(name) != attrs_.end(); + } + + AttrType GetAttrType(const std::string &name) const; + + std::vector AttrNames() const; + + void SetAttr(const std::string &name, const Attribute &v); + + void SetBlockAttr(const std::string &name, BlockDescBind &block); + + Attribute GetAttr(const std::string &name) const; + + int GetBlockAttr(const std::string &name) const; + + void Rename(const std::string &old_name, const std::string &new_name); + + // Only be used in C++ + const AttributeMap &GetAttrMap() const; + + // Only be used in C++ + void SetAttrMap(const AttributeMap &attr_map); + + std::vector InputNames() const { return MapKeys(inputs_); } + std::vector OutputNames() const { return MapKeys(outputs_); } + + void SetInputMap(const VariableNameMap &input) { + this->inputs_ = input; + this->need_update_ = true; + } + + void SetOutputMap(const VariableNameMap &output) { + this->outputs_ = output; + this->need_update_ = true; + } + + void Sync(); + + const VariableNameMap &Inputs() const { return inputs_; } + + const VariableNameMap &Outputs() const { return outputs_; } + + private: + template + static std::vector MapKeys(const MapType &map) { + std::vector ret_val; + ret_val.reserve(map.size()); + std::transform( + map.begin(), map.end(), std::back_inserter(ret_val), + [](const typename MapType::value_type &pair) { return pair.first; }); + return ret_val; + } + + OpDesc op_desc_; + VariableNameMap inputs_; + VariableNameMap outputs_; + AttributeMap attrs_; + + // need_update_ indicate there some local changes not be synchronized. If + // local changes should be synchronized, need_update_ should be set to true. + bool need_update_{false}; +}; +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/op_info.h b/paddle/framework/op_info.h index b98d8f23a1..c504f69e30 100644 --- a/paddle/framework/op_info.h +++ b/paddle/framework/op_info.h @@ -19,21 +19,18 @@ #include #include "paddle/framework/attribute.h" +#include "paddle/framework/op_desc.h" +#include "paddle/framework/type_defs.h" +#include "paddle/platform/macros.h" namespace paddle { namespace framework { -class OperatorBase; -using VariableNameMap = std::map>; - -using OpCreator = std::function; struct OpInfo { OpCreator creator_; - std::string grad_op_type_; - OpProto* proto_; - OpAttrChecker* checker_; + GradOpMakerFN grad_op_maker_; + OpProto* proto_{nullptr}; + OpAttrChecker* checker_{nullptr}; bool HasOpProtoAndChecker() const { return proto_ != nullptr && checker_ != nullptr; @@ -46,30 +43,25 @@ struct OpInfo { return *proto_; } - const OpAttrChecker& Checker() const { - PADDLE_ENFORCE_NOT_NULL(checker_, - "Operator Checker has not been registered"); - return *checker_; - } - const OpCreator& Creator() const { PADDLE_ENFORCE_NOT_NULL(creator_, "Operator Creator has not been registered"); return creator_; } - bool HasGradientOp() const { return !grad_op_type_.empty(); } + const GradOpMakerFN& GradOpMaker() const { + PADDLE_ENFORCE_NOT_NULL(grad_op_maker_, + "Operator GradOpMaker has not been registered."); + return grad_op_maker_; + } + + const OpAttrChecker* Checker() const { return checker_; } }; class OpInfoMap { public: static OpInfoMap& Instance(); - OpInfoMap(const OpInfoMap& o) = delete; - OpInfoMap(OpInfoMap&& o) = delete; - OpInfoMap& operator=(const OpInfoMap& o) = delete; - OpInfoMap& operator=(OpInfoMap&& o) = delete; - bool Has(const std::string& op_type) const { return map_.find(op_type) != map_.end(); } @@ -105,6 +97,8 @@ class OpInfoMap { private: OpInfoMap() = default; std::unordered_map map_; + + DISABLE_COPY_AND_ASSIGN(OpInfoMap); }; } // namespace framework diff --git a/paddle/framework/op_proto_maker.cc b/paddle/framework/op_proto_maker.cc new file mode 100644 index 0000000000..151d61d5b1 --- /dev/null +++ b/paddle/framework/op_proto_maker.cc @@ -0,0 +1,58 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/op_proto_maker.h" + +namespace paddle { +namespace framework { + +void OpProtoAndCheckerMaker::Validate() { + validated_ = true; + CheckNoDuplicatedInOutAttrs(); +} + +OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddInput( + const std::string& name, const std::string& comment) { + auto* input = proto_->add_inputs(); + input->set_name(name); + input->set_comment(comment); + return OpProtoAndCheckerMaker::VariableBuilder{input}; +} + +OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddOutput( + const std::string& name, const std::string& comment) { + auto* output = proto_->add_outputs(); + output->set_name(name); + output->set_comment(comment); + return OpProtoAndCheckerMaker::VariableBuilder{output}; +} + +void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() { + std::unordered_set names; + auto checker = [&](const std::string& name) { + PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name); + names.insert(name); + }; + for (auto& attr : proto_->attrs()) { + checker(attr.name()); + } + for (auto& input : proto_->inputs()) { + checker(input.name()); + } + for (auto& output : proto_->outputs()) { + checker(output.name()); + } +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/op_proto_maker.h b/paddle/framework/op_proto_maker.h new file mode 100644 index 0000000000..a134befd90 --- /dev/null +++ b/paddle/framework/op_proto_maker.h @@ -0,0 +1,83 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/framework/attribute.h" +#include "paddle/framework/framework.pb.h" + +namespace paddle { +namespace framework { + +// this class not only make proto but also init attribute checkers. +class OpProtoAndCheckerMaker { + public: + OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) + : proto_(proto), op_checker_(op_checker) {} + + virtual ~OpProtoAndCheckerMaker() { + PADDLE_ENFORCE(validated_, "should call Validate after build"); + } + + void Validate(); + + protected: + struct VariableBuilder { + OpProto::Var* var_; + + VariableBuilder& AsDuplicable() { + var_->set_duplicable(true); + return *this; + } + + VariableBuilder& AsIntermediate() { + var_->set_intermediate(true); + return *this; + } + }; + + VariableBuilder AddInput(const std::string& name, const std::string& comment); + + VariableBuilder AddOutput(const std::string& name, + const std::string& comment); + + template + TypedAttrChecker& AddAttr(const std::string& name, + const std::string& comment, + bool generated = false) { + auto* attr = proto_->add_attrs(); + attr->set_name(name); + attr->set_comment(comment); + attr->set_generated(generated); + attr->set_type(AttrTypeID()); + return op_checker_->AddAttrChecker(name); + } + + void AddComment(const std::string& comment) { proto_->set_comment(comment); } + + private: + void CheckNoDuplicatedInOutAttrs(); + + OpProto* proto_; + OpAttrChecker* op_checker_; + bool validated_{false}; +}; + +class NOPMaker : public OpProtoAndCheckerMaker { + public: + NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) {} +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/op_proto_maker_test.cc b/paddle/framework/op_proto_maker_test.cc new file mode 100644 index 0000000000..988a14cf4d --- /dev/null +++ b/paddle/framework/op_proto_maker_test.cc @@ -0,0 +1,51 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/op_proto_maker.h" + +#include "gtest/gtest.h" + +class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { + public: + TestAttrProtoMaker(paddle::framework::OpProto* proto, + paddle::framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddAttr("scale", "scale of test op"); + AddAttr("scale", "scale of test op"); + } +}; + +TEST(ProtoMaker, DuplicatedAttr) { + paddle::framework::OpProto op_proto; + paddle::framework::OpAttrChecker op_checker; + auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker); + ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); +} + +class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { + public: + TestInOutProtoMaker(paddle::framework::OpProto* proto, + paddle::framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("input", "input of test op"); + AddInput("input", "input of test op"); + } +}; + +TEST(ProtoMaker, DuplicatedInOut) { + paddle::framework::OpProto op_proto; + paddle::framework::OpAttrChecker op_checker; + auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker); + ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); +} diff --git a/paddle/framework/op_registry.cc b/paddle/framework/op_registry.cc index b0e85dd49f..66043f6e04 100644 --- a/paddle/framework/op_registry.cc +++ b/paddle/framework/op_registry.cc @@ -23,7 +23,9 @@ std::unique_ptr OpRegistry::CreateOp( const std::string& type, const VariableNameMap& inputs, const VariableNameMap& outputs, AttributeMap attrs) { auto& info = OpInfoMap::Instance().Get(type); - info.Checker().Check(attrs); + if (info.Checker() != nullptr) { + info.Checker()->Check(attrs); + } auto op = info.Creator()(type, inputs, outputs, attrs); return std::unique_ptr(op); } @@ -52,9 +54,15 @@ std::unique_ptr OpRegistry::CreateOp(const OpDesc& op_desc) { return CreateOp(op_desc.type(), inputs, outputs, attrs); } -std::unique_ptr OpRegistry::CreateGradOp(const OperatorBase& op) { - PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops"); - return std::unique_ptr(BuildGradOp(&op)); +std::unique_ptr OpRegistry::CreateOp(const OpDescBind& op_desc) { + return CreateOp(op_desc.Type(), op_desc.Inputs(), op_desc.Outputs(), + op_desc.GetAttrMap()); +} + +std::vector> OpRegistry::CreateGradOpDescs( + const OpDescBind& op_desc) { + auto& info = OpInfoMap::Instance().Get(op_desc.Type()); + return info.grad_op_maker_(op_desc); } } // namespace framework diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h index 572dff860a..cce3605fd4 100644 --- a/paddle/framework/op_registry.h +++ b/paddle/framework/op_registry.h @@ -21,48 +21,54 @@ limitations under the License. */ #include #include #include "paddle/framework/attribute.h" +#include "paddle/framework/details/op_registry.h" #include "paddle/framework/framework.pb.h" -#include "paddle/framework/grad_op_builder.h" -#include "paddle/framework/op_info.h" +#include "paddle/framework/grad_op_desc_maker.h" +#include "paddle/framework/op_desc.h" #include "paddle/framework/operator.h" #include "paddle/framework/scope.h" namespace paddle { namespace framework { +class Registrar { + public: + // In our design, various kinds of classes, e.g., operators and kernels, + // have their corresponding registry and registrar. The action of + // registration is in the constructor of a global registrar variable, which, + // however, are not used in the code that calls package framework, and would + // be removed from the generated binary file by the linker. To avoid such + // removal, we add Touch to all registrar classes and make USE_OP macros to + // call this method. So, as long as the callee code calls USE_OP, the global + // registrar variable won't be removed by the linker. + void Touch() {} +}; + +template +struct OperatorRegistrar : public Registrar { + explicit OperatorRegistrar(const char* op_type) : op_type(op_type) { + PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type), + "'%s' is registered more than once.", op_type); + static_assert(sizeof...(ARGS) != 0, + "OperatorRegistrar should be invoked at least by OpClass"); + details::OperatorRegistrarRecursive<0, false, ARGS...>(op_type, &info); + OpInfoMap::Instance().Insert(op_type, info); + } + + const char* op_type; + + OpInfo info; +}; class OpRegistry { public: template static void RegisterOp(const std::string& op_type, const std::string& grad_op_type) { - PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type), - "'%s' is registered more than once.", op_type); - OpInfo op_info; - op_info.creator_ = []( - const std::string& type, const VariableNameMap& inputs, - const VariableNameMap& outputs, const AttributeMap& attrs) { - return new OpType(type, inputs, outputs, attrs); - }; - op_info.grad_op_type_ = grad_op_type; - if (std::type_index(typeid(ProtoMakerType)) != - std::type_index(typeid(NOPMaker))) { - op_info.proto_ = new OpProto; - op_info.checker_ = new OpAttrChecker; - auto maker = ProtoMakerType(op_info.proto_, op_info.checker_); - maker.Validate(); - op_info.proto_->set_type(op_type); - PADDLE_ENFORCE( - op_info.proto_->IsInitialized(), - "Fail to initialize %s's OpProto, because %s is not initialized", - op_type, op_info.proto_->InitializationErrorString()); - } else { - op_info.proto_ = nullptr; - op_info.checker_ = nullptr; - } - OpInfoMap::Instance().Insert(op_type, op_info); + OperatorRegistrar reg(op_type.c_str()); + reg.info.grad_op_type_ = grad_op_type; // register gradient op if (!grad_op_type.empty()) { - RegisterOp(grad_op_type, ""); + OperatorRegistrar grad_reg(grad_op_type.c_str()); } } @@ -73,20 +79,10 @@ class OpRegistry { static std::unique_ptr CreateOp(const OpDesc& op_desc); - static std::unique_ptr CreateGradOp(const OperatorBase& op); -}; + static std::vector> CreateGradOpDescs( + const OpDescBind& op_desc); -class Registrar { - public: - // In our design, various kinds of classes, e.g., operators and kernels, - // have their corresponding registry and registrar. The action of - // registration is in the constructor of a global registrar variable, which, - // however, are not used in the code that calls package framework, and would - // be removed from the generated binary file by the linker. To avoid such - // removal, we add Touch to all registrar classes and make USE_OP macros to - // call this method. So, as long as the callee code calls USE_OP, the global - // registrar variable won't be removed by the linker. - void Touch() {} + static std::unique_ptr CreateOp(const OpDescBind& op_desc); }; template @@ -99,13 +95,39 @@ class OpRegistrar : public Registrar { } }; -template +template +struct OpKernelRegistrarFunctor; + +template +struct OpKernelRegistrarFunctor { + using KERNEL_TYPE = + typename std::tuple_element>::type; + + void operator()(const char* op_type) const { + using T = typename KERNEL_TYPE::ELEMENT_TYPE; + OperatorWithKernel::OpKernelKey key(ToDataType(std::type_index(typeid(T))), + PlaceType()); + OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KERNEL_TYPE); + + constexpr auto size = std::tuple_size>::value; + OpKernelRegistrarFunctor + func; + func(op_type); + } +}; + +template +struct OpKernelRegistrarFunctor { + void operator()(const char* op_type) const {} +}; + +// User can register many kernel in one place. The data type could be different. +template class OpKernelRegistrar : public Registrar { public: explicit OpKernelRegistrar(const char* op_type) { - OperatorWithKernel::OpKernelKey key; - key.place_ = PlaceType(); - OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KernelType); + OpKernelRegistrarFunctor func; + func(op_type); } }; @@ -118,33 +140,41 @@ class OpKernelRegistrar : public Registrar { __test_global_namespace_##uniq_name##__>::value, \ msg) +#define REGISTER_OPERATOR(op_type, op_class, ...) \ + STATIC_ASSERT_GLOBAL_NAMESPACE( \ + __reg_op__##op_type, \ + "REGISTER_OPERATOR must be called in global namespace"); \ + class _OpClass_##op_type##_ : public op_class { \ + public: \ + DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_); \ + DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class); \ + }; \ + static ::paddle::framework::OperatorRegistrar<_OpClass_##op_type##_, \ + ##__VA_ARGS__> \ + __op_registrar_##op_type##__(#op_type); \ + int TouchOpRegistrar_##op_type() { \ + __op_registrar_##op_type##__.Touch(); \ + return 0; \ + } + /** * Macro to register Operator. */ -#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, \ - grad_op_class) \ - STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __reg_op__##op_type, "REGISTER_OP must be called in global namespace"); \ - class _OpClass_##op_type##_ : public op_class { \ - public: \ - DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_); \ - DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class); \ - }; \ - class _OpGradClass_##op_type##_ : public grad_op_class { \ - public: \ - DEFINE_OP_CLONE_METHOD(_OpGradClass_##op_type##_); \ - DEFINE_OP_CONSTRUCTOR(_OpGradClass_##op_type##_, grad_op_class); \ - }; \ - static ::paddle::framework::OpRegistrar< \ - _OpClass_##op_type##_, op_maker_class, _OpGradClass_##op_type##_> \ - __op_registrar_##op_type##__(#op_type, #grad_op_type); \ - int TouchOpRegistrar_##op_type() { \ - __op_registrar_##op_type##__.Touch(); \ - return 0; \ - } +#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, \ + grad_op_class) \ + REGISTER_OPERATOR(grad_op_type, grad_op_class); \ + class _GradOpDescMaker_##grad_op_type##_ \ + : public ::paddle::framework::DefaultGradOpDescMaker { \ + using ::paddle::framework::DefaultGradOpDescMaker::DefaultGradOpDescMaker; \ + \ + protected: \ + virtual std::string GradOpType() const { return #grad_op_type; } \ + }; \ + REGISTER_OPERATOR(op_type, op_class, _GradOpDescMaker_##grad_op_type##_, \ + op_maker_class); #define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \ - REGISTER_OP(op_type, op_class, op_maker_class, , ::paddle::framework::NOP) + REGISTER_OPERATOR(op_type, op_class, op_maker_class) /** * Macro to register OperatorKernel. @@ -191,7 +221,7 @@ class OpKernelRegistrar : public Registrar { // TODO(fengjiayi): The following macros // seems ugly, do we have better method? -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA #define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU) #else #define USE_OP_KERNEL(op_type) \ diff --git a/paddle/framework/op_registry_test.cc b/paddle/framework/op_registry_test.cc index e00c6e8d90..b860fe6cac 100644 --- a/paddle/framework/op_registry_test.cc +++ b/paddle/framework/op_registry_test.cc @@ -10,7 +10,6 @@ class CosineOp : public OperatorBase { using OperatorBase::OperatorBase; void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override {} - void InferShape(const Scope& scope) const override {} }; class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { @@ -29,7 +28,6 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { class MyTestOp : public OperatorBase { public: using OperatorBase::OperatorBase; - void InferShape(const Scope& scope) const override {} void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override {} }; @@ -174,4 +172,15 @@ TEST(OpRegistry, CustomChecker) { op->Run(scope, dev_ctx); int test_attr = op->Attr("test_attr"); ASSERT_EQ(test_attr, 4); -} \ No newline at end of file +} + +class CosineOpComplete : public paddle::framework::CosineOp { + public: + DEFINE_OP_CONSTRUCTOR(CosineOpComplete, paddle::framework::CosineOp); + DEFINE_OP_CLONE_METHOD(CosineOpComplete); +}; + +TEST(OperatorRegistrar, Test) { + using namespace paddle::framework; + OperatorRegistrar reg("cos"); +} diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc index f8a64a7866..2ca838f838 100644 --- a/paddle/framework/operator.cc +++ b/paddle/framework/operator.cc @@ -14,7 +14,7 @@ limitations under the License. */ #include "paddle/framework/operator.h" #include -#include "paddle/framework/op_registry.h" +#include namespace paddle { namespace framework { @@ -22,17 +22,35 @@ namespace framework { template <> Eigen::DefaultDevice& ExecutionContext::GetEigenDevice< platform::CPUPlace, Eigen::DefaultDevice>() const { - return *device_context_.get_eigen_device(); + return *device_context_.GetEigenDevice(); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA template <> Eigen::GpuDevice& ExecutionContext::GetEigenDevice() const { - return *device_context_.get_eigen_device(); + return *device_context_.GetEigenDevice(); } #endif +const Tensor* GetTensorFromVar(const Variable* var) { + if (var->IsType()) { + return &var->Get(); + } + PADDLE_ENFORCE(var->IsType(), + "The Input must be LoDTensor or Tensor."); + return &var->Get(); +} + +Tensor* GetTensorFromVar(Variable* var) { + if (var->IsType()) { + return var->GetMutable(); + } + PADDLE_ENFORCE(var->IsType(), + "The Input must be LoDTensor or Tensor."); + return var->GetMutable(); +} + std::string OperatorBase::Input(const std::string& name) const { auto& ins = Inputs(name); PADDLE_ENFORCE_LE(ins.size(), 1UL, @@ -60,8 +78,8 @@ std::string OperatorBase::Output(const std::string& name) const { const std::vector& OperatorBase::Outputs( const std::string& name) const { auto it = outputs_.find(name); - PADDLE_ENFORCE(it != outputs_.end(), "Op %s does not have output %s", type_, - name); + PADDLE_ENFORCE(it != outputs_.end(), "Op %s does not have output called %s", + type_, name); return it->second; } @@ -207,63 +225,31 @@ const std::vector InferShapeContext::MultiInput( } template <> -Tensor* ExecutionContext::Output(const std::string& name) const { - auto* var = OutputVar(name); - return var == nullptr ? nullptr : const_cast(GetTensorFromVar(var)); +Tensor* InferShapeContext::Output(const std::string& name) const { + auto var = OutputVar(name); + return var == nullptr ? nullptr : var->GetMutable(); } template <> -std::vector ExecutionContext::MultiOutput( +std::vector InferShapeContext::MultiOutput( const std::string& name) const { auto names = op().Outputs(name); std::vector res; res.reserve(names.size()); std::transform(names.begin(), names.end(), std::back_inserter(res), [&](const std::string& sub_name) { - auto var = scope().FindVar(sub_name); - return var == nullptr - ? nullptr - : const_cast(GetTensorFromVar(var)); + auto var = scope_.FindVar(sub_name); + return var == nullptr ? nullptr + : var->GetMutable(); }); return res; } -void OpProtoAndCheckerMaker::Validate() { - validated_ = true; - CheckNoDuplicatedInOutAttrs(); -} - -OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddInput( - const std::string& name, const std::string& comment) { - auto* input = proto_->add_inputs(); - input->set_name(name); - input->set_comment(comment); - return OpProtoAndCheckerMaker::VariableBuilder{input}; -} - -OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddOutput( - const std::string& name, const std::string& comment) { - auto* output = proto_->add_outputs(); - output->set_name(name); - output->set_comment(comment); - return OpProtoAndCheckerMaker::VariableBuilder{output}; -} - -void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() { - std::unordered_set names; - auto checker = [&](const std::string& name) { - PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name); - names.insert(name); - }; - for (auto& attr : proto_->attrs()) { - checker(attr.name()); - } - for (auto& input : proto_->inputs()) { - checker(input.name()); - } - for (auto& output : proto_->outputs()) { - checker(output.name()); - } +std::ostream& operator<<(std::ostream& os, + const OperatorWithKernel::OpKernelKey& kernel_key) { + os << "place[" << kernel_key.place_ << "]:data_type[" << kernel_key.data_type_ + << "]"; + return os; } } // namespace framework diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h index b7c9c39402..d7bc9c9ffb 100644 --- a/paddle/framework/operator.h +++ b/paddle/framework/operator.h @@ -15,15 +15,19 @@ limitations under the License. */ #pragma once #include +#include #include #include #include #include "op_info.h" #include "paddle/framework/attribute.h" +#include "paddle/framework/block_desc.h" +#include "paddle/framework/data_type.h" #include "paddle/framework/framework.pb.h" #include "paddle/framework/lod_tensor.h" #include "paddle/framework/scope.h" +#include "paddle/framework/shape_inference.h" #include "paddle/framework/tensor.h" #include "paddle/platform/device_context.h" #include "paddle/platform/place.h" @@ -56,6 +60,9 @@ class OperatorBase; class InferShapeContext; class ExecutionContext; +extern const Tensor* GetTensorFromVar(const Variable* var); +extern Tensor* GetTensorFromVar(Variable* var); + /** * OperatorBase has the basic element that Net will call to do computation. * Only CreateOperator from OpRegistry will new Operator directly. User @@ -78,10 +85,6 @@ class OperatorBase { virtual std::string DebugString() const; - /// InferShape infer the size of Variables used by this Operator with - /// information inside scope - virtual void InferShape(const Scope& scope) const = 0; - /// Net will call this function to Run an op. virtual void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const = 0; @@ -159,7 +162,6 @@ class OperatorBase { class NOP : public OperatorBase { public: using OperatorBase::OperatorBase; - void InferShape(const Scope& scope) const override {} void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override {} std::unique_ptr Clone() const override { @@ -167,71 +169,6 @@ class NOP : public OperatorBase { } }; -// this class not only make proto but also init attribute checkers. -class OpProtoAndCheckerMaker { - public: - OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : proto_(proto), op_checker_(op_checker) {} - - ~OpProtoAndCheckerMaker() { - PADDLE_ENFORCE(validated_, "should call Validate after build"); - } - - void Validate(); - - protected: - struct VariableBuilder { - OpProto::Var* var_; - - VariableBuilder& AsDuplicable() { - var_->set_duplicable(true); - return *this; - } - - VariableBuilder& AsIntermediate() { - var_->set_intermediate(true); - return *this; - } - - VariableBuilder& NotInGradient() { - var_->set_not_in_gradient(true); - return *this; - } - }; - - VariableBuilder AddInput(const std::string& name, const std::string& comment); - - VariableBuilder AddOutput(const std::string& name, - const std::string& comment); - - template - TypedAttrChecker& AddAttr(const std::string& name, - const std::string& comment, - bool generated = false) { - auto* attr = proto_->add_attrs(); - attr->set_name(name); - attr->set_comment(comment); - attr->set_generated(generated); - attr->set_type(AttrTypeID()); - return op_checker_->AddAttrChecker(name); - } - - void AddComment(const std::string& comment) { proto_->set_comment(comment); } - - private: - void CheckNoDuplicatedInOutAttrs(); - - OpProto* proto_; - OpAttrChecker* op_checker_; - bool validated_{false}; -}; - -class NOPMaker : public OpProtoAndCheckerMaker { - public: - NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) {} -}; - class InferShapeContext { public: InferShapeContext(const OperatorBase& op, const Scope& scope) @@ -277,9 +214,9 @@ class InferShapeContext { return res; } - std::vector MultiOutputVar(const std::string& name) const { + std::vector MultiOutputVar(const std::string& name) const { auto names = op_.Outputs(name); - std::vector res; + std::vector res; res.reserve(names.size()); std::transform(names.begin(), names.end(), std::back_inserter(res), [this](const std::string& name) { @@ -327,13 +264,18 @@ class InferShapeContext { return res; } - const Tensor* GetTensorFromVar(const Variable* var) const { - if (var->IsType()) { - return &var->Get(); - } - PADDLE_ENFORCE(var->IsType(), - "The Input(%s) must be LoDTensor or Tensor."); - return &var->Get(); + void ShareLoD(const std::string& in, const std::string& out, size_t i = 0, + size_t j = 0) const { + PADDLE_ENFORCE_LT(i, InputSize(in)); + PADDLE_ENFORCE_LT(j, OutputSize(out)); + auto* in_var = MultiInputVar(in)[i]; + auto* out_var = MultiOutputVar(out)[j]; + if (!in_var->IsType()) return; + PADDLE_ENFORCE(out_var->IsType(), + "The %d-th output of Output(%s) must be LoDTensor.", j, out); + auto in_tensor = in_var->Get(); + auto* out_tensor = out_var->GetMutable(); + out_tensor->set_lod(in_tensor.lod()); } private: @@ -348,20 +290,12 @@ template <> const std::vector InferShapeContext::MultiInput( const std::string& name) const; -template -struct EigenDeviceConverter; - template <> -struct EigenDeviceConverter { - using EigenDeviceType = Eigen::DefaultDevice; -}; +Tensor* InferShapeContext::Output(const std::string& name) const; -#ifndef PADDLE_ONLY_CPU template <> -struct EigenDeviceConverter { - using EigenDeviceType = Eigen::GpuDevice; -}; -#endif +std::vector InferShapeContext::MultiOutput( + const std::string& name) const; class ExecutionContext : public InferShapeContext { public: @@ -370,8 +304,8 @@ class ExecutionContext : public InferShapeContext { : InferShapeContext(op, scope), device_context_(device_context) {} template ::EigenDeviceType> + typename DeviceType = typename platform::EigenDeviceConverter< + PlaceType>::EigenDeviceType> DeviceType& GetEigenDevice() const; platform::Place GetPlace() const { return device_context_.GetPlace(); } @@ -380,39 +314,207 @@ class ExecutionContext : public InferShapeContext { return device_context_; } - // redefine Output function, - // use Variable::Get instead of Variable::GetMutable - template - T* Output(const std::string& name) const { - auto var = OutputVar(name); - return var == nullptr ? nullptr : const_cast(&var->Get()); + private: + const platform::DeviceContext& device_context_; +}; + +class CompileTimeInferShapeContext : public InferShapeContextBase { + public: + CompileTimeInferShapeContext(const OpDescBind& op, const BlockDescBind& block) + : op_(op), block_(block) {} + + bool HasInput(const std::string& name) const override { + const std::vector& input_names = op_.Input(name); + auto length = input_names.size(); + PADDLE_ENFORCE_EQ(length, 1UL, + "Input(%s) should have only one value, " + "but it have %d now", + name, length); + return block_.HasVar(input_names[0]); } - // redefine MultiOutput function. - // use Variable::Get instead of Variable::GetMutable - template - std::vector MultiOutput(const std::string& name) const { - auto names = op().Outputs(name); - std::vector res; - res.reserve(names.size()); - std::transform( - names.begin(), names.end(), std::back_inserter(res), - [&](const std::string& sub_name) { return Output(sub_name); }); - return res; + bool HasOutput(const std::string& name) const override { + const std::vector& output_names = op_.Output(name); + auto length = output_names.size(); + PADDLE_ENFORCE_EQ(length, 1UL, + "Output(%s) should have only one value, " + "but it have %d now", + name, length); + return block_.HasVar(output_names[0]); + } + + bool HasInputs(const std::string& name) const override { + const std::vector& input_names = op_.Input(name); + PADDLE_ENFORCE(!input_names.empty(), "Inputs(%s) length is 0", name); + for (auto& input : input_names) { + if (!block_.HasVar(input)) return false; + } + return true; + } + + bool HasOutputs(const std::string& name) const override { + const std::vector& output_names = op_.Output(name); + PADDLE_ENFORCE(!output_names.empty(), "Inputs(%s) length is 0", name); + for (auto& output : output_names) { + if (!block_.HasVar(output)) return false; + } + return true; + } + + DDim GetInputDim(const std::string& name) const override { + std::vector ddims = GetInputsDim(name); + auto length = ddims.size(); + PADDLE_ENFORCE_EQ(length, 1UL, + "Input(%s) should have 1 value, " + "but it has %d now", + name, length); + return ddims[0]; + } + + void SetInputDim(const std::string& name, const DDim& dim) override { + SetInputsDim(name, {dim}); + } + + DDim GetOutputDim(const std::string& name) const override { + std::vector ddims = GetOutputsDim(name); + auto length = ddims.size(); + PADDLE_ENFORCE_EQ(length, 1UL, + "Output(%s) should have 1 value, " + "but it has %d now", + name, length); + return ddims[0]; + } + + void SetOutputDim(const std::string& name, const DDim& dim) override { + SetOutputsDim(name, {dim}); + } + + AttrReader Attrs() const override { return AttrReader(op_.GetAttrMap()); } + + const std::vector& Inputs( + const std::string& name) const override { + return op_.Input(name); + } + + const std::vector& Outputs( + const std::string& name) const override { + return op_.Output(name); } private: - const platform::DeviceContext& device_context_; + DDim GetDim(const std::string& name) const override { + return framework::make_ddim(block_.Var(name)->Shape()); + } + + void SetDim(const std::string& name, const DDim& dim) override { + block_.Var(name)->SetShape(framework::vectorize(dim)); + } + + const OpDescBind& op_; + const BlockDescBind& block_; }; -template <> -Tensor* ExecutionContext::Output(const std::string& name) const; +class RuntimeInferShapeContext : public InferShapeContextBase { + public: + RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope) + : op_(op), scope_(scope) {} -template <> -std::vector ExecutionContext::MultiOutput( - const std::string& name) const; + bool HasInput(const std::string& name) const override { + auto ipt = op_.Input(name); + auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt); + return var != nullptr; + } + + bool HasOutput(const std::string& name) const override { + auto ipt = op_.Output(name); + auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt); + return var != nullptr; + } + + bool HasInputs(const std::string& name) const override { + auto inputs = op_.Inputs(name); + if (inputs.empty()) { + return false; + } + for (auto& input : inputs) { + if (scope_.FindVar(input) == nullptr) { + return false; + } + } + return true; + } + + bool HasOutputs(const std::string& name) const override { + auto outputs = op_.Outputs(name); + if (outputs.empty()) { + return false; + } + for (auto& output : outputs) { + if (scope_.FindVar(output) == nullptr) { + return false; + } + } + return true; + } + + DDim GetInputDim(const std::string& name) const override { + return GetDim(op_.Input(name)); + } + + void SetInputDim(const std::string& name, const DDim& dim) override { + SetDim(op_.Input(name), dim); + } + + DDim GetOutputDim(const std::string& name) const override { + return GetDim(op_.Output(name)); + } + + void SetOutputDim(const std::string& name, const DDim& dim) override { + SetDim(op_.Output(name), dim); + } + + AttrReader Attrs() const override { return AttrReader(op_.Attrs()); } + + const std::vector& Inputs( + const std::string& name) const override { + return op_.Inputs(name); + } -class OpKernel { + const std::vector& Outputs( + const std::string& name) const override { + return op_.Outputs(name); + } + + private: + template + Tensor* GetTensor(const std::string& name) const { + Tensor* t = nullptr; + auto* var = scope_.FindVar(name); + if (!var->IsType() && !var->IsType()) { + if (Allocate) { + t = var->GetMutable(); + } else { + PADDLE_THROW("Variable(%s) should be tensor", name); + } + } else { + t = GetTensorFromVar(scope_.FindVar(name)); + } + return t; + } + + DDim GetDim(const std::string& name) const override { + return GetTensor(name)->dims(); + } + + void SetDim(const std::string& name, const DDim& dim) override { + GetTensor(name)->Resize(dim); + } + + const OperatorBase& op_; + const Scope& scope_; +}; + +class OpKernelBase { public: /** * ExecutionContext is the only parameter of Kernel Run function. @@ -423,46 +525,77 @@ class OpKernel { virtual void Compute(const ExecutionContext& context) const = 0; - virtual ~OpKernel() {} + virtual ~OpKernelBase() = default; +}; + +template +class OpKernel : public OpKernelBase { + public: + using ELEMENT_TYPE = T; }; class OperatorWithKernel : public OperatorBase { public: struct OpKernelKey { platform::Place place_; + DataType data_type_; - OpKernelKey() = default; - explicit OpKernelKey(const platform::DeviceContext& dev_ctx) { - place_ = dev_ctx.GetPlace(); - } + OpKernelKey(DataType data_type, platform::Place place) + : place_(place), data_type_(data_type) {} + + OpKernelKey(DataType data_type, const platform::DeviceContext& dev_ctx) + : place_(dev_ctx.GetPlace()), data_type_(data_type) {} bool operator==(const OpKernelKey& o) const { - return platform::places_are_same_class(place_, o.place_); + return platform::places_are_same_class(place_, o.place_) && + data_type_ == o.data_type_; } }; struct OpKernelHash { - std::hash hash_; + std::hash hash_; size_t operator()(const OpKernelKey& key) const { - return hash_(platform::is_gpu_place(key.place_)); + int place = key.place_.which(); + int data_type = static_cast(key.data_type_); + int pre_hash = data_type << NUM_PLACE_TYPE_LIMIT_IN_BIT | + (place & ((1 << NUM_PLACE_TYPE_LIMIT_IN_BIT) - 1)); + return hash_(pre_hash); } }; using OpKernelMap = - std::unordered_map, OpKernelHash>; + std::unordered_map, + OpKernelHash>; OperatorWithKernel(const std::string& type, const VariableNameMap& inputs, const VariableNameMap& outputs, const AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs) {} - void InferShape(const Scope& scope) const override { - InferShape(InferShapeContext(*this, scope)); - } - void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const final { - auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx)); - opKernel->Compute(ExecutionContext(*this, scope, dev_ctx)); + RuntimeInferShapeContext infer_shape_ctx(*this, scope); + this->InferShape(&infer_shape_ctx); + + ExecutionContext ctx(*this, scope, dev_ctx); + + // check if op[type] has kernel registered. + auto& all_op_kernels = AllOpKernels(); + auto kernels_iter = all_op_kernels.find(type_); + if (kernels_iter == all_op_kernels.end()) { + PADDLE_THROW("op[%s] has no kernel", type_); + } + + // check if op[type] have kernel for kernel_key + OpKernelMap& kernels = kernels_iter->second; + auto kernel_key = OpKernelKey(IndicateDataType(ctx), dev_ctx); + auto kernel_iter = kernels.find(kernel_key); + + if (kernel_iter == kernels.end()) { + PADDLE_THROW("op[%s] has no kernel with kernel_key[%s]", type_, + kernel_key); + } + + kernel_iter->second->Compute(ctx); } static std::unordered_map& @@ -472,14 +605,47 @@ class OperatorWithKernel : public OperatorBase { } bool SupportGPU() const override { - OperatorWithKernel::OpKernelKey key; - key.place_ = platform::GPUPlace(); - return OperatorWithKernel::AllOpKernels().at(type_).count(key) != 0; + auto& op_kernels = OperatorWithKernel::AllOpKernels().at(type_); + return std::any_of(op_kernels.begin(), op_kernels.end(), + [](OpKernelMap::const_reference kern_pair) { + return platform::is_gpu_place(kern_pair.first.place_); + }); } + virtual void InferShape(InferShapeContextBase* ctx) const = 0; + protected: - virtual void InferShape(const InferShapeContext& ctx) const = 0; + // indicate kernel DataType by input data. Defaultly all input data must be + // same. + virtual DataType IndicateDataType(const ExecutionContext& ctx) const { + auto& scope = ctx.scope(); + int data_type = -1; + for (auto& input : this->inputs_) { + for (auto& ipt_name : input.second) { + auto* var = scope.FindVar(ipt_name); + if (var != nullptr) { + const Tensor* t = nullptr; + if (var->IsType()) { + t = &var->Get(); + } else if (var->IsType()) { + t = &var->Get(); + } + if (t != nullptr) { + int tmp = static_cast(ToDataType(t->type())); + PADDLE_ENFORCE(tmp == data_type || data_type == -1, + "DataType of Paddle Op must be same."); + data_type = tmp; + } + } + } + } + PADDLE_ENFORCE(data_type != -1, "DataType should be indicated by input"); + return static_cast(data_type); + } }; +std::ostream& operator<<(std::ostream& os, + const OperatorWithKernel::OpKernelKey& kernel_key); + } // namespace framework } // namespace paddle diff --git a/paddle/framework/operator_test.cc b/paddle/framework/operator_test.cc index 20bbb11896..a0c17b41f2 100644 --- a/paddle/framework/operator_test.cc +++ b/paddle/framework/operator_test.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include "paddle/framework/operator.h" #include "gtest/gtest.h" +#include "paddle/framework/op_info.h" #include "paddle/framework/op_registry.h" namespace paddle { @@ -26,7 +27,6 @@ class OpWithoutKernelTest : public OperatorBase { OpWithoutKernelTest(const std::string& type, const VariableNameMap& inputs, const VariableNameMap& outputs, const AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs), x(1) {} - void InferShape(const Scope& scope) const override {} void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override { ++op_run_num; @@ -86,7 +86,6 @@ TEST(OperatorBase, all) { auto op = paddle::framework::OpRegistry::CreateOp(op_desc); scope.NewVar("OUT1"); ASSERT_EQ(paddle::framework::op_run_num, 0); - op->InferShape(scope); op->Run(scope, device_context); ASSERT_EQ(paddle::framework::op_run_num, 1); } @@ -114,11 +113,14 @@ class OpWithKernelTest : public OperatorWithKernel { using OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override {} + void InferShape(framework::InferShapeContextBase* ctx) const override {} + DataType IndicateDataType(const ExecutionContext& ctx) const override { + return DataType::FP32; + } }; template -class CPUKernelTest : public OpKernel { +class CPUKernelTest : public OpKernel { public: void Compute(const ExecutionContext& ctx) const { std::cout << "this is cpu kernel" << std::endl; @@ -145,7 +147,7 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker } }; -class CPUKernalMultiInputsTest : public OpKernel { +class CPUKernalMultiInputsTest : public OpKernel { public: void Compute(const ExecutionContext& ctx) const { auto xs = ctx.op().Inputs("xs"); @@ -254,7 +256,6 @@ class OperatorClone : public paddle::framework::OperatorBase { const paddle::framework::VariableNameMap& outputs, const paddle::framework::AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs) {} - void InferShape(const paddle::framework::Scope& scope) const override {} void Run(const paddle::framework::Scope& scope, const paddle::platform::DeviceContext& dev_ctx) const override {} }; @@ -264,37 +265,3 @@ TEST(Operator, Clone) { auto b = a.Clone(); ASSERT_EQ(a.Type(), b->Type()); } - -class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { - public: - TestAttrProtoMaker(paddle::framework::OpProto* proto, - paddle::framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddAttr("scale", "scale of test op"); - AddAttr("scale", "scale of test op"); - } -}; - -TEST(ProtoMaker, DuplicatedAttr) { - paddle::framework::OpProto op_proto; - paddle::framework::OpAttrChecker op_checker; - auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker); - ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); -} - -class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { - public: - TestInOutProtoMaker(paddle::framework::OpProto* proto, - paddle::framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("input", "input of test op"); - AddInput("input", "input of test op"); - } -}; - -TEST(ProtoMaker, DuplicatedInOut) { - paddle::framework::OpProto op_proto; - paddle::framework::OpAttrChecker op_checker; - auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker); - ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); -} \ No newline at end of file diff --git a/paddle/framework/program_desc.cc b/paddle/framework/program_desc.cc new file mode 100644 index 0000000000..e89f9a46d5 --- /dev/null +++ b/paddle/framework/program_desc.cc @@ -0,0 +1,60 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/program_desc.h" +#include "paddle/framework/block_desc.h" + +namespace paddle { +namespace framework { + +using ProgDescMap = + std::unordered_map>; +static ProgDescMap *g_bind_map = nullptr; + +ProgramDescBind &ProgramDescBind::Instance(ProgramDesc *prog) { + if (g_bind_map == nullptr) { + g_bind_map = new ProgDescMap(); + } + auto &map = *g_bind_map; + auto &ptr = map[prog]; + + if (ptr == nullptr) { + ptr.reset(new ProgramDescBind(prog)); + } + return *ptr; +} + +BlockDescBind *ProgramDescBind::AppendBlock(const BlockDescBind &parent) { + auto *b = prog_->add_blocks(); + b->set_parent_idx(parent.ID()); + b->set_idx(prog_->blocks_size() - 1); + blocks_.emplace_back(new BlockDescBind(this, b)); + return blocks_.back().get(); +} + +ProgramDesc *ProgramDescBind::Proto() { + for (auto &block : blocks_) { + block->Sync(); + } + return prog_; +} + +ProgramDescBind::ProgramDescBind(ProgramDesc *prog) { + prog_ = prog; + for (auto &block : *prog->mutable_blocks()) { + blocks_.emplace_back(new BlockDescBind(this, &block)); + } +} +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/program_desc.h b/paddle/framework/program_desc.h new file mode 100644 index 0000000000..9b34a06aef --- /dev/null +++ b/paddle/framework/program_desc.h @@ -0,0 +1,51 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/framework/framework.pb.h" +#include "paddle/platform/macros.h" + +namespace paddle { +namespace framework { + +class BlockDescBind; + +class ProgramDescBind { + public: + static ProgramDescBind &Instance(ProgramDesc *prog); + + BlockDescBind *AppendBlock(const BlockDescBind &parent); + + BlockDescBind *Block(size_t idx) { return blocks_[idx].get(); } + + std::string DebugString() { return Proto()->DebugString(); } + + size_t Size() const { return blocks_.size(); } + + ProgramDesc *Proto(); + + private: + explicit ProgramDescBind(ProgramDesc *prog); + + // Not owned + ProgramDesc *prog_; + + std::vector> blocks_; + + DISABLE_COPY_AND_ASSIGN(ProgramDescBind); +}; +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/scope.h b/paddle/framework/scope.h index 2ba3f8ed35..7047f0d55e 100644 --- a/paddle/framework/scope.h +++ b/paddle/framework/scope.h @@ -19,6 +19,7 @@ limitations under the License. */ #include #include "paddle/framework/variable.h" +#include "paddle/platform/macros.h" namespace paddle { namespace framework { @@ -38,11 +39,6 @@ class Scope { Scope() {} ~Scope(); - // Disable Copy, Assign, Move. - Scope(const Scope& other) = delete; - Scope& operator=(const Scope& other) = delete; - Scope(Scope&& other) = delete; - /// Create a sub-scope. Returns a reference other than a pointer so /// to prevent from manual deletion. /// Mark it to const because that new kid scope cannot change parent scope. @@ -58,6 +54,8 @@ class Scope { /// nullptr if cannot find. Variable* FindVar(const std::string& name) const; + const Scope& parent() const { return *parent_; } + /// Find the scope or an ancestor scope that contains the given variable. const Scope* FindScope(const Variable* var) const; @@ -71,6 +69,8 @@ class Scope { std::unordered_map vars_; mutable std::list kids_; Scope const* parent_{nullptr}; + + DISABLE_COPY_AND_ASSIGN(Scope); }; } // namespace framework diff --git a/paddle/framework/shape_inference.h b/paddle/framework/shape_inference.h new file mode 100644 index 0000000000..74e0371e32 --- /dev/null +++ b/paddle/framework/shape_inference.h @@ -0,0 +1,89 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/framework/ddim.h" + +namespace paddle { +namespace framework { + +// TODO(longfei): Once after both CompileTimeInferShapeContext and +// RuntimeInferShapeContext get merged, we can rename InferShapeContextBase into +// InferShapeContext so to replace the current InferShapeContext. +class InferShapeContextBase { + public: + virtual ~InferShapeContextBase() {} + virtual bool HasInput(const std::string &name) const = 0; + virtual bool HasOutput(const std::string &name) const = 0; + + virtual bool HasInputs(const std::string &name) const = 0; + virtual bool HasOutputs(const std::string &name) const = 0; + + virtual framework::DDim GetInputDim(const std::string &name) const = 0; + std::vector GetInputsDim(const std::string &name) const { + const std::vector &names = Inputs(name); + return GetDims(names); + } + virtual void SetInputDim(const std::string &name, + const framework::DDim &dim) = 0; + void SetInputsDim(const std::string &name, + const std::vector &dims) { + auto &names = Inputs(name); + SetDims(names, dims); + } + virtual framework::DDim GetOutputDim(const std::string &name) const = 0; + std::vector GetOutputsDim(const std::string &name) const { + const std::vector &names = Outputs(name); + return GetDims(names); + } + virtual void SetOutputDim(const std::string &name, const DDim &dim) = 0; + void SetOutputsDim(const std::string &name, + const std::vector &dims) { + auto &names = Outputs(name); + SetDims(names, dims); + } + virtual AttrReader Attrs() const = 0; + virtual const std::vector &Inputs( + const std::string &name) const = 0; + virtual const std::vector &Outputs( + const std::string &name) const = 0; + // TODO(qiao) implement this function + void ShareLoD(const std::string &in, const std::string &out, size_t i = 0, + size_t j = 0) const {} + + protected: + virtual framework::DDim GetDim(const std::string &name) const = 0; + virtual void SetDim(const std::string &name, const framework::DDim &dim) = 0; + std::vector GetDims( + const std::vector &names) const { + std::vector ret; + ret.reserve(names.size()); + std::transform( + names.begin(), names.end(), std::back_inserter(ret), + [this](const std::string &name) { return this->GetDim(name); }); + return ret; + } + void SetDims(const std::vector &names, + const std::vector &dims) { + size_t length = names.size(); + PADDLE_ENFORCE_EQ(length, dims.size()); + for (size_t i = 0; i < length; ++i) { + SetDim(names[i], dims[i]); + } + } +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/tensor.h b/paddle/framework/tensor.h index 4b5a2ae523..80a3f0a393 100644 --- a/paddle/framework/tensor.h +++ b/paddle/framework/tensor.h @@ -30,16 +30,9 @@ limitations under the License. */ namespace paddle { namespace framework { -namespace details { -template -struct CastToPyBufferImpl; -} class Tensor { public: - template - friend struct details::CastToPyBufferImpl; - template friend struct EigenTensor; @@ -116,6 +109,8 @@ class Tensor { return holder_->place(); } + std::type_index type() const { return holder_->type(); } + private: template inline void check_memory_size() const; @@ -165,12 +160,6 @@ class Tensor { /*! points to dimensions of memory block. */ DDim dims_; - /** - * A cache of the number of elements in a tensor. - * Would be 0 for an uninitialized tensor. - */ - int64_t numel_; - /** * @brief A PlaceHolder may be shared by more than one tensor. * diff --git a/paddle/framework/tensor_array.cc b/paddle/framework/tensor_array.cc new file mode 100644 index 0000000000..2728bce1c1 --- /dev/null +++ b/paddle/framework/tensor_array.cc @@ -0,0 +1,283 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + + + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/framework/tensor_array.h" + +#include +#include +#include + +namespace paddle { +namespace framework { + +namespace detail { + +/* + * Offer an iterator over the length-sorted lod-tensor's top level. The top + * level of a lod-tensor stores batch-size of sequences, each top-level sequence + * may contains several lower-level sequences, sort top-level lod by the numbers + * of lower-level sequences in descending order, so that during RNN's running, + * the batch-size will keep decreasing, the short sentences will end at the tail + * of each batch. + * + * Let's take a simple lod-tensor for example + * + * |(0) |(1) top-level has two instances + * ||| ||||| lower-level + * + * sort by lower-level's length + * + * |(1) |(0) + * ||||| ||| + * + * when RNN runs, it get 5 batches (equals the number of elements the longest + * sequence has) + * + * ||||| + * ||| + * + * the first three batches has two elements, the last two elements just has 1 + * element each. + */ +struct DynamicBatchUnpacker { + using value_type = float; + + DynamicBatchUnpacker(const LoDTensor& source, size_t level, + bool descend = true) + : source(&source), level(level) { + BuildLengthSortedMeta(descend); + } + + LoDTensor GetBatch(size_t index); + + std::vector meta; + + LoDTensor const* source; + size_t level; + + protected: + void BuildLengthSortedMeta(bool descend); +}; + +LoDTensor PackDynamicBatch(const std::vector& source, + const std::vector& meta, const LoD& lod, + size_t level); + +} // namespace detail + +const LoDTensor& TensorArray::Read(size_t index) const { + PADDLE_ENFORCE_LE(index, MAX_SIZE, "index[%d] too large", index); + if (index >= size()) { + values_.resize(index + 1); + } + return values_[index]; +} + +void TensorArray::Write(size_t index, const LoDTensor& value) { + PADDLE_ENFORCE_LE(index, MAX_SIZE, "index[%d] too large", index); + + if (index >= size()) { + values_.resize(index + 1); + } + + values_[index].Resize(value.dims()); + values_[index].mutable_data(platform::CPUPlace()); + values_[index].CopyFrom(value, platform::CPUPlace()); +} + +void TensorArray::WriteShared(size_t index, const LoDTensor& value) { + PADDLE_ENFORCE_LE(index, MAX_SIZE, "index[%d] too large", index); + if (index >= size()) { + values_.resize(index + 1); + } + + values_[index].ShareDataWith(value); +} + +LoDTensor TensorArray::Pack(size_t level, const std::vector& meta, + const LoD& lod) const { + return detail::PackDynamicBatch(values_, meta, lod, level); +} + +std::vector TensorArray::Unpack(const LoDTensor& source, int level, + bool length_desend) { + detail::DynamicBatchUnpacker unpacker(source, level, + length_desend /*descend*/); + + // find max length of all the sequences + size_t max_length = 0; + for (const auto& seq : unpacker.meta) { + max_length = std::max(max_length, seq.end - seq.begin); + } + + // write batches to values + for (size_t batch_id = 0; batch_id < max_length; batch_id++) { + Write(batch_id, unpacker.GetBatch(batch_id)); + } + + return unpacker.meta; +} + +LoDTensor TensorArray::Stack() const { + LoDTensor result; + if (size() == 0) return result; + + const auto& first_dims = values_.front().dims(); + // check all the values have the same shape + // TODO(superjom) check the same dtypes + for (size_t idx = 1; idx < size(); idx++) { + const auto& value_dims = values_[idx].dims(); + PADDLE_ENFORCE_EQ(first_dims, value_dims); + } + + // copy + auto result_dims = vectorize(first_dims); + result_dims.insert(result_dims.begin(), size()); + result.Resize(make_ddim(result_dims)); + result.mutable_data(platform::CPUPlace()); + + for (size_t idx = 0; idx < size(); idx++) { + result.Slice(idx, idx + 1) + .CopyFrom(Read(idx), platform::CPUPlace()); + } + return result; +} + +void TensorArray::Unstack(const LoDTensor& source) const { + Unstack(source, false /*data_shared*/); +} + +void TensorArray::UnstackShared(const LoDTensor& source) const { + Unstack(source, true /*data_shared*/); +} + +void TensorArray::Unstack(const LoDTensor& source, bool data_shared) const { + size_t first_dim = source.dims()[0]; + DDim value_dims = slice_ddim(source.dims(), 1, source.dims().size()); + PADDLE_ENFORCE_GT(first_dim, 0, + "source should have some data to be unstacked"); + + values_.resize(first_dim); + + for (size_t elem = 0; elem < first_dim; elem++) { + // create a new value + auto& value = values_[elem]; + if (data_shared) { + // share memory + value.ShareDataWith(source.Slice(elem, elem + 1)); + } else { + // copy + value.Resize(value_dims); + value.CopyFrom(source.Slice(elem, elem + 1), + platform::CPUPlace()); + } + } +} + +size_t TensorArray::size() const { return values_.size(); } + +namespace detail { + +void DynamicBatchUnpacker::BuildLengthSortedMeta(bool descend) { + PADDLE_ENFORCE(meta.empty(), "duplicate build meta"); + // collect meta for each sequence in some level + auto lod = SliceLevels(source->lod(), level, level + 1)[0]; + + for (size_t seq_id = 0; seq_id < lod.size() - 1; seq_id++) { + DySeqMeta seq_meta({lod[seq_id], lod[seq_id + 1], seq_id}); + meta.push_back(seq_meta); + } + + PADDLE_ENFORCE_GT(meta.size(), 0, "meta is empty"); + + // sort by length + sort(meta.begin(), meta.end(), + [descend](const DySeqMeta& a, const DySeqMeta& b) { + bool a_ge_b = (a.end - a.begin) > (b.end - b.begin); + return descend ? a_ge_b : !a_ge_b; + }); +} + +LoDTensor DynamicBatchUnpacker::GetBatch(size_t index) { + PADDLE_ENFORCE(!meta.empty(), "should build meta first"); + LoDTensor result; + + // collect indice need to copy to the batch + std::vector indice; + for (const auto& seq : meta) { + size_t id = seq.begin + index; + if (id >= seq.end) break; + indice.push_back(id); + } + PADDLE_ENFORCE(!indice.empty(), "invalid batch at %d", index); + + // copy the indice of records in LoDTensor + auto record_dims = slice_ddim(source->dims(), 1, source->dims().size()); + auto record_dims_vec = vectorize(record_dims); + record_dims_vec.insert(record_dims_vec.begin(), indice.size()); + result.Resize(make_ddim(record_dims_vec)); + result.mutable_data(platform::CPUPlace()); + + for (size_t i = 0; i < indice.size(); i++) { + auto index = indice[i]; + auto target = result.Slice(i, i + 1); + auto source_ = source->Slice(index, index + 1); + + target.CopyFrom(source_, platform::CPUPlace()); + } + + return result; +} + +// TODO(supejom) to cache lod if reasonable +LoDTensor PackDynamicBatch(const std::vector& source, + const std::vector& meta, const LoD& lod, + size_t level) { + PADDLE_ENFORCE(!source.empty()); + PADDLE_ENFORCE(!meta.empty()); + PADDLE_ENFORCE(!lod.empty()); + + LoDTensor result; + + // init result space + auto record_dims = slice_ddim(source[0].dims(), 1, source[0].dims().size()); + auto record_dims_vec = vectorize(record_dims); + auto height = lod[level].back(); + record_dims_vec.insert(record_dims_vec.begin(), height); + result.Resize(make_ddim(record_dims_vec)); + result.mutable_data(platform::CPUPlace()); + + for (size_t batch_id = 0; batch_id < source.size(); batch_id++) { + for (size_t seq_id = 0; seq_id < meta.size(); seq_id++) { + const auto& seq_meta = meta[seq_id]; + // source is source[batch_id][seq_id] + // target is result[index] + auto index = seq_meta.begin + batch_id; + if (index >= seq_meta.end) break; + auto source_ = source[batch_id].Slice(seq_id, seq_id + 1); + auto target = result.Slice(index, index + 1); + target.CopyFrom(source_, platform::CPUPlace()); + } + } + + result.set_lod(lod); + return result; +} + +} // namespace detail + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/tensor_array.h b/paddle/framework/tensor_array.h new file mode 100644 index 0000000000..94a14c2df4 --- /dev/null +++ b/paddle/framework/tensor_array.h @@ -0,0 +1,113 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include + +#include "paddle/framework/lod_tensor.h" + +namespace paddle { +namespace framework { + +/* + * DyBatchSeqPosition stores indices of the basic element in tensor. It is used + * after lod-tensor's re-assembling, its info can be used to recover the order + * in original lod-tensor. + */ +struct DySeqMeta { + DySeqMeta(size_t begin, size_t end, size_t ori_idx) + : begin(begin), end(end), ori_idx(ori_idx) {} + + size_t begin; + size_t end; // not included + size_t ori_idx; +}; + +/* + * TensorArray is a C-array-like array of tensors, it is meant to be used with + * dynamic iteration primitives such as while_loop. It is used to segment inputs + * and store states in all time steps. + * + * By providing some methods similar to a C++ array, the difinition of some + * state-based dynamic models such as RNN cound be more natural and highly + * flexible. + */ +class TensorArray { + public: + using value_type = float; + + // max number of values allowed to store. + const size_t MAX_SIZE{100000}; + + /* + * Read the value at location `index` in the `TensorArray`. + */ + const LoDTensor &Read(size_t index) const; + + /* + * Write value into the index of the TensorArray. + */ + void Write(size_t index, const LoDTensor &value); + + /* + * Write value into the index of the TensorArray, with memory shared. + */ + void WriteShared(size_t index, const LoDTensor &value); + + /* + * Recover the original LoD-arranged LoDTensor with the `values`, `level` and + * `indice_map`. + */ + LoDTensor Pack(size_t level, const std::vector &meta, + const LoD &lod) const; + + /* + * Split LoDTensor in some `level` and write the generated batches to + * `values`, if set `desend`, will sort by length in descending order else in + * ascending order. + */ + std::vector Unpack(const LoDTensor &source, int level, + bool length_desend); + + /* + * Pack the values into a tensor with rank one higher than each tensor in + * values. + */ + LoDTensor Stack() const; + + /* + * Unpacks the given division of a rank-`R` tensor into rank-`(R-1)` tensors. + */ + void Unstack(const LoDTensor &source) const; + + /* + * Unpacks the given division of a rank-`R` tensor into rank-`(R-1)` tensors, + * with memory of tensors shared. + */ + void UnstackShared(const LoDTensor &source) const; + + /* + * Return the number of values. + */ + size_t size() const; + + protected: + void Unstack(const LoDTensor &source, bool data_shared) const; + + private: + mutable std::vector values_; +}; // class TensorArray + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/tensor_array_test.cc b/paddle/framework/tensor_array_test.cc new file mode 100644 index 0000000000..d9f52509cd --- /dev/null +++ b/paddle/framework/tensor_array_test.cc @@ -0,0 +1,130 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/framework/tensor_array.h" + +#include + +namespace paddle { +namespace framework { + +class TensorArrayTester : public ::testing::Test { + protected: + void SetUp() override { + LoDTensor source; + source.Resize(make_ddim({batch_size, dim})); + int* data = source.mutable_data(platform::CPUPlace()); + for (int i = 0; i < 16 * 32; i++) { + data[i] = i; + } + ta.Unstack(source); + } + + TensorArray ta; + const int batch_size = 16; + const int dim = 32; +}; + +TEST_F(TensorArrayTester, Read) { + for (int i = 0; i < batch_size; i++) { + const auto& tensor = ta.Read(i); + ASSERT_EQ(tensor.dims()[0], 1); + ASSERT_EQ(tensor.dims()[1], dim); + } +} + +TEST_F(TensorArrayTester, Write) { + LoDTensor source; + source.Resize(make_ddim({1, dim})); + for (int i = 0; i < dim; i++) { + *(source.mutable_data(platform::CPUPlace()) + i) = i; + } + + ta.Write(2, source); + + const auto& tensor = ta.Read(2); + for (int i = 0; i < dim; i++) { + EXPECT_EQ(*(tensor.data() + i), *(source.data() + i)); + } +} + +TEST_F(TensorArrayTester, WriteShared) { + LoDTensor source; + source.Resize(make_ddim({1, dim})); + for (int i = 0; i < dim; i++) { + *(source.mutable_data(platform::CPUPlace()) + i) = i; + } + + ta.WriteShared(2, source); + + const auto& tensor = ta.Read(2); + for (int i = 0; i < dim; i++) { + EXPECT_EQ(*(tensor.data() + i), *(source.data() + i)); + } + + EXPECT_EQ(source.data(), tensor.data()); +} + +class TensorArrayPackTester : public ::testing::Test { + protected: + virtual void SetUp() override { + lod.push_back(std::vector{0, 2, 9, 13}); + + source.set_lod(lod); + source.Resize(make_ddim({13, 128})); + source.mutable_data(platform::CPUPlace()); + + // content of each setence: 0 1 2 3 4 + const auto& level = lod.front(); + for (size_t i = 0; i < level.size() - 1; i++) { + size_t begin = level[i]; + size_t end = level[i + 1]; + for (size_t j = begin; j < end; j++) { + auto record = source.Slice(j, j + 1); + for (int dim = 0; dim < 128; dim++) { + record.mutable_data(platform::CPUPlace())[dim] = j - begin; + } + } + } + + // unpack + meta = ta.Unpack(source, 0, true); + } + + LoD lod; + TensorArray ta; + LoDTensor source; + std::vector meta; +}; + +TEST_F(TensorArrayPackTester, Unpack) { + ASSERT_EQ(ta.size(), 7UL); + + const auto& t0 = ta.Read(0); + const auto& t1 = ta.Read(1); + + ASSERT_EQ(t0.data()[0], int(0)); + ASSERT_EQ(t1.data()[0], int(1)); +} + +TEST_F(TensorArrayPackTester, Pack) { + LoDTensor packed = ta.Pack(0, meta, lod); +} + +TEST_F(TensorArrayTester, size) { + ASSERT_EQ(ta.size(), static_cast(batch_size)); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/tensor_impl.h b/paddle/framework/tensor_impl.h index ed166935f7..379eac94f9 100644 --- a/paddle/framework/tensor_impl.h +++ b/paddle/framework/tensor_impl.h @@ -65,7 +65,7 @@ inline T* Tensor::mutable_data(platform::Place place) { holder_.reset(new PlaceholderImpl( boost::get(place), size)); } else if (platform::is_gpu_place(place)) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA PADDLE_THROW("'GPUPlace' is not supported in CPU only device."); } #else @@ -103,7 +103,7 @@ inline void Tensor::CopyFrom(const Tensor& src, memory::Copy(boost::get(dst_place), dst_ptr, boost::get(src_place), src_ptr, size); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA else if (platform::is_gpu_place(src_place) && platform::is_cpu_place(dst_place)) { memory::Copy(boost::get(dst_place), dst_ptr, @@ -130,26 +130,29 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const { PADDLE_ENFORCE_LE(end_idx, dims_[0], "Slice end index is out of bound."); PADDLE_ENFORCE_LT(begin_idx, end_idx, "Begin index must be less than end index."); - PADDLE_ENFORCE_NE(dims_[0], 1, "Can not slice a tensor with dims_[0] = 1."); - size_t base = numel() / dims_[0]; - Tensor dst; - dst.holder_ = holder_; - DDim dst_dims = dims_; - dst_dims[0] = end_idx - begin_idx; - dst.Resize(dst_dims); - dst.offset_ = offset_ + begin_idx * base * sizeof(T); - return dst; + + if (dims_[0] == 1) { + return *this; + } else { + size_t base = numel() / dims_[0]; + Tensor dst; + dst.holder_ = holder_; + DDim dst_dims = dims_; + dst_dims[0] = end_idx - begin_idx; + dst.Resize(dst_dims); + dst.offset_ = offset_ + begin_idx * base * sizeof(T); + return dst; + } } inline Tensor& Tensor::Resize(const DDim& dims) { dims_ = dims; - numel_ = product(dims_); return *this; } inline const DDim& Tensor::dims() const { return dims_; } -inline int64_t Tensor::numel() const { return numel_; } +inline int64_t Tensor::numel() const { return product(dims_); } template inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) { diff --git a/paddle/framework/tensor_test.cc b/paddle/framework/tensor_test.cc index e2ec738de3..58cf0fc3cb 100644 --- a/paddle/framework/tensor_test.cc +++ b/paddle/framework/tensor_test.cc @@ -74,7 +74,7 @@ TEST(Tensor, MutableData) { EXPECT_EQ(p1, p2); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA { Tensor src_tensor; float* p1 = nullptr; @@ -126,7 +126,7 @@ TEST(Tensor, ShareDataWith) { ASSERT_EQ(src_tensor.data(), dst_tensor.data()); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA { Tensor src_tensor; Tensor dst_tensor; @@ -163,7 +163,7 @@ TEST(Tensor, Slice) { EXPECT_EQ(src_data_address + 3 * 4 * 1 * sizeof(int), slice_data_address); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA { Tensor src_tensor; src_tensor.mutable_data(make_ddim({6, 9}), GPUPlace()); @@ -218,7 +218,7 @@ TEST(Tensor, CopyFrom) { EXPECT_EQ(dst_ptr[i], slice_ptr[i]); } } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA { Tensor src_tensor; Tensor gpu_tensor; diff --git a/paddle/framework/type_defs.h b/paddle/framework/type_defs.h new file mode 100644 index 0000000000..a5b9472213 --- /dev/null +++ b/paddle/framework/type_defs.h @@ -0,0 +1,42 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include +#include +#include "paddle/platform/variant.h" + +namespace paddle { +namespace framework { +class OperatorBase; +class OpDescBind; +using VariableNameMap = std::map>; + +// The order should be as same as framework.proto +using Attribute = + boost::variant, + std::vector, std::vector, bool, + std::vector, BlockDesc*>; + +using AttributeMap = std::unordered_map; + +using OpCreator = std::function; + +using GradOpMakerFN = + std::function>(const OpDescBind&)>; + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/var_desc.cc b/paddle/framework/var_desc.cc new file mode 100644 index 0000000000..13b9c5f3cd --- /dev/null +++ b/paddle/framework/var_desc.cc @@ -0,0 +1,36 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/var_desc.h" + +namespace paddle { +namespace framework { + +void VarDescBind::SetShape(const std::vector &dims) { + VectorToRepeated(dims, desc_.mutable_lod_tensor()->mutable_dims()); +} + +void VarDescBind::SetDataType(DataType data_type) { + desc_.mutable_lod_tensor()->set_data_type(data_type); +} + +std::vector VarDescBind::Shape() const { + return RepeatedToVector(desc_.lod_tensor().dims()); +} + +DataType VarDescBind::GetDataType() const { + return desc_.lod_tensor().data_type(); +} +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/var_desc.h b/paddle/framework/var_desc.h new file mode 100644 index 0000000000..4763bf09d0 --- /dev/null +++ b/paddle/framework/var_desc.h @@ -0,0 +1,73 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/framework/framework.pb.h" + +namespace paddle { +namespace framework { + +// convert between std::vector and protobuf repeated. +template +inline std::vector RepeatedToVector( + const google::protobuf::RepeatedField &repeated_field) { + std::vector ret; + ret.reserve(repeated_field.size()); + std::copy(repeated_field.begin(), repeated_field.end(), + std::back_inserter(ret)); + return ret; +} + +template +inline void VectorToRepeated(const std::vector &vec, + RepeatedField *repeated_field) { + repeated_field->Reserve(vec.size()); + for (const auto &elem : vec) { + *repeated_field->Add() = elem; + } +} + +// Specialize vector. +template +inline void VectorToRepeated(const std::vector &vec, + RepeatedField *repeated_field) { + repeated_field->Reserve(vec.size()); + for (auto elem : vec) { + *repeated_field->Add() = elem; + } +} + +class VarDescBind { + public: + explicit VarDescBind(const std::string &name) { desc_.set_name(name); } + + VarDesc *Proto() { return &desc_; } + + std::string Name() const { return desc_.name(); } + + void SetShape(const std::vector &dims); + + void SetDataType(DataType data_type); + + std::vector Shape() const; + + DataType GetDataType() const; + + private: + VarDesc desc_; +}; +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/variable.md b/paddle/framework/variable.md index f44d5ea46e..442ef6b718 100644 --- a/paddle/framework/variable.md +++ b/paddle/framework/variable.md @@ -7,7 +7,7 @@ Variable is also known as *blob* in MxNet and Caffe2. It is the input and outpu For the flexibility of a DL system, a variable should be able to contain any typed value -- a tensor in most cases, but could also be some integer IDs or a scope of other variables in the case of RNN. -To use the minimum amount of memory, we'd like that a variable to allocate memory when it has to, or, lazy memory allocation. Let's take the following example: +To use the minimum amount of memory, we would like that a variable allocates memory only when it has to, or, lazy memory allocation. Let's take the following example: ```cpp Variable vr, v1, v2; @@ -38,7 +38,7 @@ This syntax for lazy memory allocation when we call `Randomize` and `Mult`, thos To make memory allocation lazy, we cannot assume that we know the type held by a variable at definition time. In other words, `class Variable` cannot be a template `template class Variable`. -Because we don't know the type `T`, we cannot save a `T*` as `Variable's` data member. Instead, we save an interface object `Placeholder`, who can return the pointer to the saved object via `Placeholder::Ptr()` as `void*`. +Because we don't know the type `T`, we cannot save a `T*` as `Variable's` data member. Instead, we save an interface object `Placeholder`, which can return the pointer to the saved object via `Placeholder::Ptr()` as `void*`. But anyway, Variable needs to know `T` so could it `delete(ptr)` and so could `Variable::Get` checks the expected type and the saved object's type. @@ -49,4 +49,4 @@ Because `PlaceholderImpl` knows `T`, it can save and return `typeid(T)` for the ## Conclusion -The technique type hiding utilizes C++ class templates, interface and derivation, and C++ RTTI (typeid). This combination saves us from definition something like `caffe2::TypeMata`, which takes hundreds of lines of C++ code. +The technique type hiding utilizes C++ class templates, interface and derivation, and C++ RTTI (typeid). This combination saves us from defining something like `caffe2::TypeMeta`, which takes hundreds of lines of C++ code. diff --git a/paddle/function/BlockExpandOp.cpp b/paddle/function/BlockExpandOp.cpp index a89b6bba45..bd0fe119ce 100644 --- a/paddle/function/BlockExpandOp.cpp +++ b/paddle/function/BlockExpandOp.cpp @@ -194,7 +194,7 @@ public: REGISTER_TYPED_FUNC(BlockExpand, CPU, BlockExpandForward); REGISTER_TYPED_FUNC(BlockExpandGrad, CPU, BlockExpandBackward); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(BlockExpand, GPU, BlockExpandForward); REGISTER_TYPED_FUNC(BlockExpandGrad, GPU, BlockExpandBackward); #endif diff --git a/paddle/function/ContextProjectionOp.cpp b/paddle/function/ContextProjectionOp.cpp index b87750b742..23916c0f4b 100644 --- a/paddle/function/ContextProjectionOp.cpp +++ b/paddle/function/ContextProjectionOp.cpp @@ -395,7 +395,7 @@ REGISTER_TYPED_FUNC(ContextProjectionForward, REGISTER_TYPED_FUNC(ContextProjectionBackward, CPU, ContextProjectionBackwardFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(ContextProjectionForward, GPU, ContextProjectionForwardFunc); diff --git a/paddle/function/CosSimOp.cpp b/paddle/function/CosSimOp.cpp index 7ece7b2dfe..2e5c281f37 100644 --- a/paddle/function/CosSimOp.cpp +++ b/paddle/function/CosSimOp.cpp @@ -233,7 +233,7 @@ private: REGISTER_TYPED_FUNC(CosSimForward, CPU, CosSimForwardFunc); REGISTER_TYPED_FUNC(CosSimBackward, CPU, CosSimBackwardFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(CosSimForward, GPU, CosSimForwardFunc); REGISTER_TYPED_FUNC(CosSimBackward, GPU, CosSimBackwardFunc); #endif diff --git a/paddle/function/CropOp.cpp b/paddle/function/CropOp.cpp index f12ee43e3d..46f98f12c1 100644 --- a/paddle/function/CropOp.cpp +++ b/paddle/function/CropOp.cpp @@ -169,7 +169,7 @@ private: REGISTER_TYPED_FUNC(Crop, CPU, CropFunc); REGISTER_TYPED_FUNC(CropGrad, CPU, CropGradFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(Crop, GPU, CropFunc); REGISTER_TYPED_FUNC(CropGrad, GPU, CropGradFunc); #endif diff --git a/paddle/function/CrossMapNormalOp.cpp b/paddle/function/CrossMapNormalOp.cpp index ef878bfbba..9e88669d37 100644 --- a/paddle/function/CrossMapNormalOp.cpp +++ b/paddle/function/CrossMapNormalOp.cpp @@ -336,7 +336,7 @@ private: REGISTER_TYPED_FUNC(CrossMapNormal, CPU, CrossMapNormalFunc); REGISTER_TYPED_FUNC(CrossMapNormalGrad, CPU, CrossMapNormalGradFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(CrossMapNormal, GPU, CrossMapNormalFunc); REGISTER_TYPED_FUNC(CrossMapNormalGrad, GPU, CrossMapNormalGradFunc); #endif diff --git a/paddle/function/DepthwiseConvOp.cpp b/paddle/function/DepthwiseConvOp.cpp index 2f3112fe65..9863e3ae1d 100644 --- a/paddle/function/DepthwiseConvOp.cpp +++ b/paddle/function/DepthwiseConvOp.cpp @@ -292,7 +292,7 @@ REGISTER_TYPED_FUNC(DepthwiseConvGradInput, REGISTER_TYPED_FUNC(DepthwiseConvGradFilter, CPU, DepthwiseConvGradFilterFunction); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(DepthwiseConv, GPU, DepthwiseConvFunction); REGISTER_TYPED_FUNC(DepthwiseConvGradInput, GPU, diff --git a/paddle/function/DepthwiseConvOpTest.cpp b/paddle/function/DepthwiseConvOpTest.cpp index d8e8c889d5..b1a90da7db 100644 --- a/paddle/function/DepthwiseConvOpTest.cpp +++ b/paddle/function/DepthwiseConvOpTest.cpp @@ -17,7 +17,7 @@ limitations under the License. */ namespace paddle { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(DepthwiseConv, Forward) { DepthwiseConvolution( "GemmConv-CPU", "DepthwiseConv-GPU", forward); diff --git a/paddle/function/GemmConvOp.cpp b/paddle/function/GemmConvOp.cpp index f8cf4ebea8..bdb56ddac3 100644 --- a/paddle/function/GemmConvOp.cpp +++ b/paddle/function/GemmConvOp.cpp @@ -340,7 +340,7 @@ public: REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvFunction); REGISTER_TYPED_FUNC(GemmConvGradInput, CPU, GemmConvGradInputFunction); REGISTER_TYPED_FUNC(GemmConvGradFilter, CPU, GemmConvGradFilterFunction); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(GemmConv, GPU, GemmConvFunction); REGISTER_TYPED_FUNC(GemmConvGradInput, GPU, GemmConvGradInputFunction); REGISTER_TYPED_FUNC(GemmConvGradFilter, GPU, GemmConvGradFilterFunction); diff --git a/paddle/function/GemmConvOpTest.cpp b/paddle/function/GemmConvOpTest.cpp index 5283d79a5a..b5b5e1f35b 100644 --- a/paddle/function/GemmConvOpTest.cpp +++ b/paddle/function/GemmConvOpTest.cpp @@ -24,7 +24,7 @@ TEST(GemmConv, NaiveConv) { "NaiveConv-CPU", "GemmConv-CPU", forward); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(GemmConv, Forward) { Convolution( "GemmConv-CPU", "GemmConv-GPU", forward); diff --git a/paddle/function/Im2ColTest.cpp b/paddle/function/Im2ColTest.cpp index acc88a553a..a0a01a5fc7 100644 --- a/paddle/function/Im2ColTest.cpp +++ b/paddle/function/Im2ColTest.cpp @@ -116,7 +116,7 @@ void TestIm2ColFunctor() { TEST(Im2ColFunctor, CPU) { TestIm2ColFunctor(); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(Im2ColFunctor, GPU) { TestIm2ColFunctor(); } diff --git a/paddle/function/MulOp.cpp b/paddle/function/MulOp.cpp index 25e41edad5..704a8c4132 100644 --- a/paddle/function/MulOp.cpp +++ b/paddle/function/MulOp.cpp @@ -341,7 +341,7 @@ private: }; REGISTER_TYPED_FUNC(MulOp, CPU, MulFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(MulOp, GPU, MulFunc); #endif } // namespace paddle diff --git a/paddle/function/PadOp.cpp b/paddle/function/PadOp.cpp index adba7c92ec..eed2f2e308 100644 --- a/paddle/function/PadOp.cpp +++ b/paddle/function/PadOp.cpp @@ -207,7 +207,7 @@ private: REGISTER_TYPED_FUNC(Pad, CPU, PadFunc); REGISTER_TYPED_FUNC(PadGrad, CPU, PadGradFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(Pad, GPU, PadFunc); REGISTER_TYPED_FUNC(PadGrad, GPU, PadGradFunc); #endif diff --git a/paddle/function/RowConvOp.cpp b/paddle/function/RowConvOp.cpp index b6501e8f4d..7c802d6627 100644 --- a/paddle/function/RowConvOp.cpp +++ b/paddle/function/RowConvOp.cpp @@ -217,7 +217,7 @@ public: REGISTER_TYPED_FUNC(RowConv, CPU, RowConvFunc); REGISTER_TYPED_FUNC(RowConvGrad, CPU, RowConvGradFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(RowConv, GPU, RowConvFunc); REGISTER_TYPED_FUNC(RowConvGrad, GPU, RowConvGradFunc); #endif diff --git a/paddle/function/SwitchOp.cpp b/paddle/function/SwitchOp.cpp index 01e252a8dc..597723a2dd 100644 --- a/paddle/function/SwitchOp.cpp +++ b/paddle/function/SwitchOp.cpp @@ -132,7 +132,7 @@ public: REGISTER_TYPED_FUNC(NCHW2NHWC, CPU, NCHW2NHWCFunc); REGISTER_TYPED_FUNC(NHWC2NCHW, CPU, NHWC2NCHWFunc); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA REGISTER_TYPED_FUNC(NCHW2NHWC, GPU, NCHW2NHWCFunc); REGISTER_TYPED_FUNC(NHWC2NCHW, GPU, NHWC2NCHWFunc); #endif diff --git a/paddle/function/neon/NeonDepthwiseConv.cpp b/paddle/function/neon/NeonDepthwiseConv.cpp index 18126152ea..38aa667061 100644 --- a/paddle/function/neon/NeonDepthwiseConv.cpp +++ b/paddle/function/neon/NeonDepthwiseConv.cpp @@ -52,7 +52,7 @@ public: int outputHeight = output[2]; int outputWidth = output[3]; int filterMultiplier = outputChannels / groups_; - CHECK_EQ(inputChannels, groups_); + CHECK_EQ(static_cast(inputChannels), groups_); // only support strideH() == strideW() and filterHeight == filterWidth. CHECK_EQ(strideH(), strideW()); diff --git a/paddle/function/neon/NeonDepthwiseConv.h b/paddle/function/neon/NeonDepthwiseConv.h index 33722d3cac..98a86d278f 100644 --- a/paddle/function/neon/NeonDepthwiseConv.h +++ b/paddle/function/neon/NeonDepthwiseConv.h @@ -18,7 +18,6 @@ limitations under the License. */ #include "neon_util.h" namespace paddle { - namespace neon { #if defined(__ARM_NEON__) || defined(__ARM_NEON) @@ -26,17 +25,20 @@ namespace neon { template struct DepthwiseConvKernel {}; -inline float32_t conv3x3(float32x4_t r0, - float32x4_t r1, - float32x4_t r2, +inline float32_t conv3x3(const float* r0, + const float* r1, + const float* r2, float32x4_t k0, float32x4_t k1, float32x4_t k2) { - float32x4_t tmp; - tmp = vmulq_f32(r0, k0); - tmp = vmlaq_f32(tmp, r1, k1); - tmp = vmlaq_f32(tmp, r2, k2); - return vaddvq_f32(tmp); + float32_t tmp[12]; + vst1q_f32(&(tmp[0]), k0); + vst1q_f32(&(tmp[4]), k1); + vst1q_f32(&(tmp[8]), k2); + float32_t sum0 = r0[0] * tmp[0] + r0[1] * tmp[1] + r0[2] * tmp[2]; + float32_t sum1 = r1[0] * tmp[4] + r1[1] * tmp[5] + r1[2] * tmp[6]; + float32_t sum2 = r2[0] * tmp[8] + r2[1] * tmp[9] + r2[2] * tmp[10]; + return sum0 + sum1 + sum2; } inline float32_t conv4x4(float32x4_t r0, @@ -136,10 +138,7 @@ struct DepthwiseConvKernel<3, 1> { } for (int r = 0; r < remain; r++) { - float32x4_t i0 = vld1q_f32(r0); - float32x4_t i1 = vld1q_f32(r1); - float32x4_t i2 = vld1q_f32(r2); - *outputData = conv3x3(i0, i1, i2, k[0], k[1], k[2]); + *outputData = conv3x3(r0, r1, r2, k[0], k[1], k[2]); r0++; r1++; r2++; @@ -243,10 +242,7 @@ struct DepthwiseConvKernel<3, 2> { } for (int r = 0; r < remain; r++) { - float32x4_t i0 = vld1q_f32(r0); - float32x4_t i1 = vld1q_f32(r1); - float32x4_t i2 = vld1q_f32(r2); - *outputData = conv3x3(i0, i1, i2, k[0], k[1], k[2]); + *outputData = conv3x3(r0, r1, r2, k[0], k[1], k[2]); r0 += 2; r1 += 2; r2 += 2; diff --git a/paddle/gserver/activations/ActivationFunction.cpp b/paddle/gserver/activations/ActivationFunction.cpp index 78e958e06f..8b7b2e9b65 100644 --- a/paddle/gserver/activations/ActivationFunction.cpp +++ b/paddle/gserver/activations/ActivationFunction.cpp @@ -22,9 +22,12 @@ limitations under the License. */ #include #include "paddle/parameter/Argument.h" #include "paddle/utils/ClassRegistrar.h" - #include "paddle/utils/Logging.h" +#ifdef PADDLE_USE_MKLDNN +#include "MKLDNNActivation.h" +#endif + namespace paddle { static ClassRegistrar gActivationRegistrar; @@ -456,6 +459,12 @@ Error __must_check backward(Argument& act) { END_DEFINE_ACTIVATION(log) ActivationFunction* ActivationFunction::create(const std::string& type) { +#ifdef PADDLE_USE_MKLDNN + if (!type.empty() && type.compare(0, 7, "mkldnn_") == 0) { + return MKLDNNActivation::create(type); + } +#endif + return gActivationRegistrar.createByType(type); } diff --git a/paddle/gserver/activations/MKLDNNActivation.cpp b/paddle/gserver/activations/MKLDNNActivation.cpp new file mode 100644 index 0000000000..18c5638100 --- /dev/null +++ b/paddle/gserver/activations/MKLDNNActivation.cpp @@ -0,0 +1,249 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MKLDNNActivation.h" +#include "mkldnn.hpp" +#include "paddle/utils/ClassRegistrar.h" + +namespace paddle { + +static ClassRegistrar gMKLDNNActivationRegistrar; +/** + * @def MKLDNN_ACTIVATION_CLASS_NAME + * @note MKLDNN_ACTIVATION_CLASS_NAME(relu) relu_; + * means mkldnn_reluActivation relu_; + */ +#define MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE) mkldnn_##ACT_TYPE##Activation + +/** + * @def BEGIN_MKLDNN_ACTIVATION + */ +#define BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ + class MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE) : public BASE_CLASS { +/** + * @def END_MKLDNN_ACTIVATION + */ +#define END_MKLDNN_ACTIVATION(ACT_TYPE) \ +private: \ + static const std::string name; \ + \ +public: \ + const std::string& getName() const { return name; } \ + } \ + ; \ + const std::string MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::name = \ + "mkldnn_" #ACT_TYPE; \ + static InitFunction __reg_activation__mkldnn_##ACT_TYPE([] { \ + gMKLDNNActivationRegistrar \ + .registerClass( \ + "mkldnn_" #ACT_TYPE); \ + }); + +/** + * @def DEFINE_MKLDNN_ACTIVATION + */ +#define DEFINE_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ + BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ + END_MKLDNN_ACTIVATION(ACT_TYPE) + +/** + * @def DEFINE_MKLDNN_ELTWISE_ACTIVATION + */ +#define DEFINE_MKLDNN_ELTWISE_ACTIVATION( \ + ACT_TYPE, BASE_CLASS, ALPHA, BWD_ALPHA) \ + BEGIN_MKLDNN_ACTIVATION(ACT_TYPE, BASE_CLASS) \ +private: \ + static const float alpha; \ + static const float bwdAlpha; \ + \ +public: \ + float getAlpha() const { return alpha; } \ + float getBwdAlpha() const { return bwdAlpha; } \ + END_MKLDNN_ACTIVATION(ACT_TYPE) \ + const float MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::alpha = ALPHA; \ + const float MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::bwdAlpha = BWD_ALPHA; + +/** + * @brief MKLDNN Relu Activation. + * Actually mkldnn_relu is Leaky Relu. + * f(x) = x (x >= 0) + * f(x) = negative_slope * x (x < 0) + * @note the negative_slope should be -0.f in forward + */ +DEFINE_MKLDNN_ELTWISE_ACTIVATION(relu, MKLDNNEltwiseActivation, -0.f, 0.f) + +/** + * @brief MKLDNN Tanh Activation. + */ +DEFINE_MKLDNN_ELTWISE_ACTIVATION(tanh, MKLDNNEltwiseActivation, 0.f, 0.f) + +/** + * @brief MKLDNN ELU(Exponential Linear Unit) Activation. + * f(x) = x (x >= 0) + * f(x) = negative_slope * (exp(x) - 1) (x < 0) + */ +DEFINE_MKLDNN_ELTWISE_ACTIVATION(elu, MKLDNNEltwiseActivation, 0.f, 0.f) + +mkldnn::algorithm MKLDNNEltwiseActivation::getAlgo(std::string type) const { + const std::map algoMap = { + {"relu", algorithm::eltwise_relu}, + {"tanh", algorithm::eltwise_tanh}, + {"elu", algorithm::eltwise_elu}}; + type.erase(0, 7); // remove mkldnn_ + algorithm algo = (algorithm)0; + mapGet(type, algoMap, &algo); + return algo; +} + +void MKLDNNEltwiseActivation::resetFwd(Argument& act) { + if (cnt_ == act.value->getElementCnt()) { + return; + } + MKLDNNActivation::resetFwd(act); + // note: alpha represents the NegativeSlope when used in relu. + float alpha = getAlpha(); + float beta = getBeta(); + algorithm algo = getAlgo(this->getName()); + auto fwdDesc = eltwise_fwd::desc(mkldnn::prop_kind::forward_training, + algo, + val_->getMemoryDesc(), + alpha, + beta); + fwdPD_.reset(new eltwise_fwd::primitive_desc(fwdDesc, *engine_)); + // use inplace for forward but save input value before submit + inVal_ = val_; + copyInVal_ = nullptr; + if (act.grad && algo == algorithm::eltwise_tanh) { + // tanh need save src input for backward + inVal_ = MKLDNNMatrix::create(nullptr, val_->getPrimitiveDesc()); + copyInVal_ = std::make_shared(*val_, *inVal_); + CHECK(copyInVal_) << "should not be emptry"; + pipelineFwd_.push_back(*copyInVal_); + } + fwd_.reset(new eltwise_fwd(*fwdPD_, *val_, *val_)); + pipelineFwd_.push_back(*fwd_); + needResetBwd_ = true; +} + +void MKLDNNEltwiseActivation::resetBwd(Argument& act) { + if (!needResetBwd_) { + return; + } + VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward"; + needResetBwd_ = false; + algorithm algo = getAlgo(this->getName()); + float alpha = getBwdAlpha(); + float beta = getBeta(); + grad_ = MKLDNNMatrix::create(act.grad, val_->getPrimitiveDesc()); + auto eng = CPUEngine::Instance().getEngine(); + auto bwdDesc = eltwise_bwd::desc( + algo, grad_->getMemoryDesc(), val_->getMemoryDesc(), alpha, beta); + auto bwdPD = eltwise_bwd::primitive_desc(bwdDesc, eng, *fwdPD_); + CHECK(inVal_); + bwd_.reset(new eltwise_bwd(bwdPD, *inVal_, *grad_, *grad_)); + pipelineBwd_.clear(); + pipelineBwd_.push_back(*bwd_); +} + +/** + * @brief MKLDNN Softmax Activation + */ +DEFINE_MKLDNN_ACTIVATION(softmax, MKLDNNSoftmaxActivation) + +void MKLDNNSoftmaxActivation::resetFwd(Argument& act) { + if (cnt_ == act.value->getElementCnt()) { + return; + } + MKLDNNActivation::resetFwd(act); + int axis = 1; + auto fwdDesc = softmax_fwd::desc( + mkldnn::prop_kind::forward_scoring, val_->getMemoryDesc(), axis); + auto fwdPD = softmax_fwd::primitive_desc(fwdDesc, *engine_); + fwd_.reset(new softmax_fwd(fwdPD, *val_, *val_)); + pipelineFwd_.push_back(*fwd_); +} + +Error __must_check MKLDNNSoftmaxActivation::forward(Argument& act) { + resetFwd(act); + stream_->submit(pipelineFwd_); + real* v = act.value->getData(); + real threshold = exp(-64); +#pragma omp parallel for + for (size_t i = 0; i < act.value->getElementCnt(); ++i) { + v[i] = v[i] < threshold ? threshold : v[i]; + } + return Error(); +} + +Error __must_check MKLDNNSoftmaxActivation::backward(Argument& act) { + MatrixPtr outputV = act.value; + MatrixPtr outputG = act.grad; + Matrix::resizeOrCreate(sftMaxDot_, + outputG->getHeight(), + outputG->getWidth(), + /* trans */ false, + /* useGpu */ false); + Matrix::resizeOrCreate(sftMaxSum_, + outputG->getHeight(), + 1, + /* trans */ false, + /* useGpu */ false); + sftMaxDot_->dotMul(*outputG, *outputV); + sftMaxSum_->colMerge(*sftMaxDot_); + act.grad->softmaxDerivative(*act.value, *sftMaxSum_); + return Error(); +} + +ActivationFunction* MKLDNNActivation::create(const std::string& type) { + return gMKLDNNActivationRegistrar.createByType(type); +} + +std::vector MKLDNNActivation::getAllRegisteredTypes() { + std::vector types; + gMKLDNNActivationRegistrar.forEachType( + [&](const std::string& type) { types.push_back(type); }); + return types; +} + +void MKLDNNActivation::resetFwd(Argument& act) { + VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward"; + cnt_ = act.value->getElementCnt(); + pipelineFwd_.clear(); + stream_.reset(new MKLDNNStream()); + engine_.reset(new mkldnn::engine(mkldnn::engine::cpu, 0)); + val_ = std::dynamic_pointer_cast(act.value); + if (val_ == nullptr) { + int bs = act.getBatchSize(); + int ih = act.getFrameHeight() > 0 ? act.getFrameHeight() : 1; + int iw = act.getFrameWidth() > 0 ? act.getFrameWidth() : 1; + int ic = cnt_ / bs / ih / iw; + CHECK_EQ(cnt_, (size_t)bs * ic * ih * iw); + val_ = MKLDNNMatrix::create( + act.value, {bs, ic, ih, iw}, mkldnn::memory::format::nchw, *engine_); + CHECK(val_); + val_->downSpatial(); + } +} + +Error __must_check MKLDNNActivation::forward(Argument& act) { + resetFwd(act); + stream_->submit(pipelineFwd_); + return Error(); +} +Error __must_check MKLDNNActivation::backward(Argument& act) { + resetBwd(act); + stream_->submit(pipelineBwd_); + return Error(); +} +} // namespace paddle diff --git a/paddle/gserver/activations/MKLDNNActivation.h b/paddle/gserver/activations/MKLDNNActivation.h new file mode 100644 index 0000000000..dd16421fd6 --- /dev/null +++ b/paddle/gserver/activations/MKLDNNActivation.h @@ -0,0 +1,119 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "ActivationFunction.h" +#include "mkldnn.hpp" +#include "paddle/gserver/layers/MKLDNNBase.h" +#include "paddle/math/MKLDNNMatrix.h" +#include "paddle/parameter/Argument.h" + +namespace paddle { + +/** + * @brief Base class of MKLDNN Activation. + * Common activation function are provieded, + * including mkldnn_relu, mkldnn_elu, mkldnn_tanh, mkldnn_softmax + */ +class MKLDNNActivation : public ActivationFunction { +protected: + // input value element count + size_t cnt_; + // should not merge the resetBwd into resetFwd, + // because the grad data would be changing before backward. + bool needResetBwd_; + // mkldnn matrix, primitive, stream and pipeline + MKLDNNMatrixPtr val_; + MKLDNNMatrixPtr grad_; + std::shared_ptr engine_; + std::shared_ptr stream_; + std::shared_ptr fwd_; + std::shared_ptr bwd_; + std::vector pipelineFwd_; + std::vector pipelineBwd_; + +public: + MKLDNNActivation() : cnt_(0), needResetBwd_(true) {} + ~MKLDNNActivation() {} + static ActivationFunction* create(const std::string& type); + static std::vector getAllRegisteredTypes(); + virtual const std::string& getName() const = 0; + /** + * reset the forward primitives + */ + virtual void resetFwd(Argument& act); + /** + * reset the backward primitives, + * can not merge this functions into resetFwd as the grad data + * would be changing before backward. + */ + virtual void resetBwd(Argument& act) {} + virtual Error __must_check forward(Argument& act); + virtual Error __must_check backward(Argument& act); +}; + +/** + * @brief Base class of MKLDNN Eltwise Activation, + * includes mkldnn_relu, mkldnn_elu and mkldnn_tanh. + */ +class MKLDNNEltwiseActivation : public MKLDNNActivation { + typedef mkldnn::eltwise_forward eltwise_fwd; + typedef mkldnn::eltwise_backward eltwise_bwd; + typedef mkldnn::algorithm algorithm; + +protected: + // save the forward primitive desc, which can be used backward + std::shared_ptr fwdPD_; + // eltwise_bwd need src input value + MKLDNNMatrixPtr inVal_; + // use for copy data + std::shared_ptr copyInVal_; + +public: + MKLDNNEltwiseActivation() {} + ~MKLDNNEltwiseActivation() {} + virtual const std::string& getName() const = 0; + + // in common, the alpha of forward and backward should be equal. + // but for relu, to avoid negative value, they should be opposite + virtual float getAlpha() const = 0; + virtual float getBwdAlpha() const = 0; + virtual float getBeta() const { return 0.f; } + virtual algorithm getAlgo(std::string type) const; + void resetFwd(Argument& act) override; + void resetBwd(Argument& act) override; +}; + +/** + * @brief Base class of MKLDNN softmax Activation, + * only have mkldnn forward, use cpu implement for backward. + */ +class MKLDNNSoftmaxActivation : public MKLDNNActivation { + typedef mkldnn::softmax_forward softmax_fwd; + +private: + // for backward + MatrixPtr sftMaxSum_; + MatrixPtr sftMaxDot_; + +public: + MKLDNNSoftmaxActivation() {} + ~MKLDNNSoftmaxActivation() {} + virtual const std::string& getName() const = 0; + void resetFwd(Argument& act) override; + Error __must_check forward(Argument& act) override; + Error __must_check backward(Argument& act) override; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/BatchNormBaseLayer.cpp b/paddle/gserver/layers/BatchNormBaseLayer.cpp index f7a80e23e1..bc7d1c83a4 100644 --- a/paddle/gserver/layers/BatchNormBaseLayer.cpp +++ b/paddle/gserver/layers/BatchNormBaseLayer.cpp @@ -16,7 +16,7 @@ limitations under the License. */ #include "BatchNormalizationLayer.h" #include "Layer.h" #include "paddle/utils/Stat.h" -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include "CudnnBatchNormLayer.h" #endif diff --git a/paddle/gserver/layers/BatchNormalizationLayer.cpp b/paddle/gserver/layers/BatchNormalizationLayer.cpp index 412762d384..dacff25e59 100644 --- a/paddle/gserver/layers/BatchNormalizationLayer.cpp +++ b/paddle/gserver/layers/BatchNormalizationLayer.cpp @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/utils/Stat.h" -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include "hl_batch_transpose.h" #endif #include "BatchNormalizationLayer.h" @@ -90,7 +90,7 @@ void BatchNormalizationLayer::expandMat(const MatrixPtr& in, MatrixPtr& out) { size_t batchSize = in->getHeight(); CHECK_EQ(out->getHeight(), batchSize * imgPixels_); if (useGpu_) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA LOG(FATAL) << "paddle is compiled only for cpu"; #else batchTranspose( @@ -127,7 +127,7 @@ void BatchNormalizationLayer::shrinkMat(const MatrixPtr& in, MatrixPtr& out) { } CHECK_EQ(in->getHeight(), static_cast(batchSize * imgPixels_)); if (useGpu_) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA LOG(FATAL) << "paddle is compiled only for cpu"; #else batchTranspose( diff --git a/paddle/gserver/layers/Layer.cpp b/paddle/gserver/layers/Layer.cpp index 2bc20eee6c..e95f42c863 100644 --- a/paddle/gserver/layers/Layer.cpp +++ b/paddle/gserver/layers/Layer.cpp @@ -14,26 +14,12 @@ limitations under the License. */ #include "paddle/utils/Util.h" +#include "CostLayer.h" +#include "ValidationLayer.h" #include "paddle/math/SparseMatrix.h" #include "paddle/utils/Error.h" #include "paddle/utils/Logging.h" -#include "AddtoLayer.h" -#include "CRFLayer.h" -#include "CosSimLayer.h" -#include "CostLayer.h" -#include "DataLayer.h" -#include "ExpandConvLayer.h" -#include "FullyConnectedLayer.h" -#include "HierarchicalSigmoidLayer.h" -#include "MaxLayer.h" -#include "MixedLayer.h" -#include "NormLayer.h" -#include "PoolLayer.h" -#include "TensorLayer.h" -#include "TransLayer.h" -#include "ValidationLayer.h" - DEFINE_bool(log_error_clipping, false, "enable log error clipping or not"); namespace paddle { @@ -109,6 +95,10 @@ ClassRegistrar Layer::registrar_; LayerPtr Layer::create(const LayerConfig& config) { std::string type = config.type(); + // NOTE: As following types have illegal character '-', + // they can not use REGISTER_LAYER to registrar. + // Besides, to fit with old training models, + // they can not use '_' instead. if (type == "multi-class-cross-entropy") return LayerPtr(new MultiClassCrossEntropy(config)); else if (type == "rank-cost") @@ -117,8 +107,6 @@ LayerPtr Layer::create(const LayerConfig& config) { return LayerPtr(new AucValidation(config)); else if (type == "pnpair-validation") return LayerPtr(new PnpairValidation(config)); - // NOTE: stop adding "if" statements here. - // Instead, use REGISTER_LAYER to add more layer types return LayerPtr(registrar_.createByType(config.type(), config)); } diff --git a/paddle/gserver/layers/MKLDNNConvLayer.cpp b/paddle/gserver/layers/MKLDNNConvLayer.cpp index 9088744bee..0d6742e909 100644 --- a/paddle/gserver/layers/MKLDNNConvLayer.cpp +++ b/paddle/gserver/layers/MKLDNNConvLayer.cpp @@ -28,7 +28,7 @@ bool MKLDNNConvLayer::init(const LayerMap& layerMap, if (!MKLDNNLayer::init(layerMap, parameterMap)) { return false; } - CHECK_EQ(inputLayers_.size(), 1) << "Only support one input layer yet"; + CHECK_EQ(inputLayers_.size(), 1UL) << "Only support one input layer yet"; CHECK_EQ(inputLayers_.size(), parameters_.size()); CHECK(config_.shared_biases()) << "Only support shared biases yet"; @@ -64,7 +64,7 @@ bool MKLDNNConvLayer::init(const LayerMap& layerMap, // create biases if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, oc_, biasParameter_)); + biases_ = std::unique_ptr(new Weight(1, oc_, biasParameter_, 0)); } return true; } @@ -251,22 +251,31 @@ void MKLDNNConvLayer::resetInValue( // create buffer and reorder if input value do not match cpuInVal_ = nullptr; cvtInVal_ = nullptr; - if (inputIsOnlyMKLDNN()) { - MKLDNNMatrixPtr dnnIn = std::dynamic_pointer_cast(inMat); - CHECK(dnnIn) << "Input should be MKLDNNMatrix"; - if (dnnIn->getPrimitiveDesc() != in->getPrimitiveDesc()) { - CHECK_EQ(dnnIn->getFormat(), format::nc); + + MKLDNNMatrixPtr dnnIn = std::dynamic_pointer_cast(inMat); + CHECK_EQ(inputIsOnlyMKLDNN(), dnnIn != nullptr); + if (dnnIn != nullptr && dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc()) { + in = dnnIn; + return; + } + if (dnnIn) { + if (dnnIn->getFormat() == format::nc) { CHECK(ih_ == 1 && iw_ == 1) << "when input is nc format"; // create a new one with nchw format and same data memory::dims inDims = memory::dims{bs_, ic_, 1, 1}; dnnIn = MKLDNNMatrix::create(inMat, inDims, format::nchw, engine_); - CHECK(dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc()); } - in = dnnIn; + if (dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc()) { + in = dnnIn; + return; + } + cpuInVal_ = dnnIn; + in = MKLDNNMatrix::create(nullptr, pd->src_primitive_desc()); + cvtInVal_ = MKLDNNMatrix::createReorder(cpuInVal_, in); + CHECK(cvtInVal_) << "should not be emptry"; } else { - const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE); memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_}; - cpuInVal_ = MKLDNNMatrix::create(cpuIn, inDims, format::nchw, engine_); + cpuInVal_ = MKLDNNMatrix::create(inMat, inDims, format::nchw, engine_); if (cpuInVal_->getPrimitiveDesc() != in->getPrimitiveDesc()) { // create new mkldnn matrix in = MKLDNNMatrix::create(nullptr, pd->src_primitive_desc()); @@ -294,12 +303,9 @@ void MKLDNNConvLayer::resetOutValue( std::shared_ptr& pd, MKLDNNMatrixPtr& out) { out = MKLDNNMatrix::create(output_.value, pd->dst_primitive_desc()); - // change original output value from cpu matrix to mkldnn matrix - output_.value = std::dynamic_pointer_cast(out); - // create reorder if output value has cpu device and pd do not match cpuOutVal_ = nullptr; - cpuOutVal_ = nullptr; + cvtOutVal_ = nullptr; if (!outputIsOnlyMKLDNN()) { const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value; memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_}; @@ -452,13 +458,14 @@ void MKLDNNConvLayer::resetOutGrad( cvtOutGrad_ = nullptr; if (!outputIsOnlyMKLDNN()) { const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad; + outMat->setData(cpuOut->getData()); // same PrimitiveDesc with cpuInVal_ CHECK(cpuOutVal_); cpuOutGrad_ = MKLDNNMatrix::create(cpuOut, cpuOutVal_->getPrimitiveDesc()); if (cpuOutGrad_->getPrimitiveDesc() == out->getPrimitiveDesc()) { - outMat->setData(cpuOut->getData()); out = cpuOutGrad_; } else { + out = MKLDNNMatrix::create(nullptr, wgtPD->diff_dst_primitive_desc()); cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out); CHECK(cvtOutGrad_); } @@ -537,7 +544,7 @@ void MKLDNNConvLayer::resetWgtValBwdData( } else { wgtValBwdData_ = wgtVal_; } - VLOG(MKLDNN_FMTS) << "weight value format for backward data" + VLOG(MKLDNN_FMTS) << "weight value format for backward data: " << wgtValBwdData_->getFormat(); } diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp index f60e221a6e..e829456d6a 100644 --- a/paddle/gserver/layers/MKLDNNFcLayer.cpp +++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp @@ -28,7 +28,7 @@ bool MKLDNNFcLayer::init(const LayerMap& layerMap, return false; } - CHECK_EQ(inputLayers_.size(), 1) << "Only support one input layer yet"; + CHECK_EQ(inputLayers_.size(), 1UL) << "Only support one input layer yet"; CHECK_EQ(inputLayers_.size(), parameters_.size()); CHECK(!parameters_[0]->isSparse()) << "Do not support sparse yet"; @@ -49,7 +49,7 @@ bool MKLDNNFcLayer::init(const LayerMap& layerMap, // create biases if (biasParameter_.get() != NULL) { - biases_ = std::unique_ptr(new Weight(1, oc_, biasParameter_)); + biases_ = std::unique_ptr(new Weight(1, oc_, biasParameter_, 0)); } return true; } @@ -161,9 +161,16 @@ void MKLDNNFcLayer::resetInValue(MKLDNNMatrixPtr& in) { void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias) { + format wgtFmt = format::oihw; + if (inVal_->getFormat() == format::nChw8c) { + wgtFmt = format::oIhw8i; + } else if (inVal_->getFormat() == format::nChw16c) { + wgtFmt = format::oIhw16i; + } wgt = MKLDNNMatrix::create( - weight_->getW(), {oc_, ic_, ih_, iw_}, format::oihw, engine_); + weight_->getW(), {oc_, ic_, ih_, iw_}, wgtFmt, engine_); wgt->downSpatial(); + VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat(); bias = (biases_ && biases_->getW()) ? MKLDNNMatrix::create(biases_->getW(), {oc_}, format::x, engine_) @@ -172,12 +179,10 @@ void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt, void MKLDNNFcLayer::resetOutValue(MKLDNNMatrixPtr& out) { out = MKLDNNMatrix::create(output_.value, {bs_, oc_}, format::nc, engine_); - // change original output value to mkldnn output value - output_.value = std::dynamic_pointer_cast(out); if (!outputIsOnlyMKLDNN()) { // fc cpu output value do not need create convert // just share point - getOutput(CPU_DEVICE).value->setData(output_.value->getData()); + getOutput(CPU_DEVICE).value->setData(out->getData()); } } @@ -234,6 +239,7 @@ void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in, void MKLDNNFcLayer::resetOutGrad(MKLDNNMatrixPtr& out) { // TODO(TJ): merge outgrad int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE; + output_.grad->setData(getOutput(device).grad->getData()); // for MKLDNN device: // can not directly cast outputgrad to mkldnnmatrix, // since each layer can not write the inputgrad to mkldnn inputgrad. diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h index 169679c829..c09fd89462 100644 --- a/paddle/gserver/layers/MKLDNNLayer.h +++ b/paddle/gserver/layers/MKLDNNLayer.h @@ -115,10 +115,15 @@ public: copySeqInfoToOutputs(); size_t elemenCnt = inputLayers_[0]->getOutput().value->getElementCnt(); if (inputElemenCnt_ != elemenCnt) { + VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward"; // reset when input total sizes changed, not only the batchsize inputElemenCnt_ = elemenCnt; reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_); resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_); + if (outVal_) { + // change original output value to mkldnn output value + output_.value = std::dynamic_pointer_cast(outVal_); + } convertWeightsFromPaddle(); needResetBwd_ = true; } @@ -137,18 +142,17 @@ public: } void backward(const UpdateCallback& callback) override { - /* Do derivation */ { + if (needResetBwd_) { + VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward"; + resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_); + needResetBwd_ = false; + } + { REGISTER_TIMER_INFO("BpActTimer", getName().c_str()); backwardActivation(); } - { REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str()); - if (needResetBwd_) { - resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_); - needResetBwd_ = false; - } - stream_->submit(pipelineBwd_); } diff --git a/paddle/gserver/layers/MKLDNNPoolLayer.cpp b/paddle/gserver/layers/MKLDNNPoolLayer.cpp index 48b2f5a4cb..b62dfb7c54 100644 --- a/paddle/gserver/layers/MKLDNNPoolLayer.cpp +++ b/paddle/gserver/layers/MKLDNNPoolLayer.cpp @@ -134,7 +134,6 @@ void MKLDNNPoolLayer::resetOutValue(MKLDNNMatrixPtr& out) { memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_}; out = MKLDNNMatrix::create( output_.value, outDims, inVal_->getFormat(), engine_); - output_.value = std::dynamic_pointer_cast(out); // create reorder if output value has cpu device and pd do not match cpuOutVal_ = nullptr; diff --git a/paddle/gserver/layers/PoolLayer.cpp b/paddle/gserver/layers/PoolLayer.cpp index 96d5c54acc..7b932d5a76 100644 --- a/paddle/gserver/layers/PoolLayer.cpp +++ b/paddle/gserver/layers/PoolLayer.cpp @@ -15,7 +15,7 @@ limitations under the License. */ #include "PoolLayer.h" #include "PoolProjectionLayer.h" #include "paddle/utils/Logging.h" -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include "CudnnPoolLayer.h" #endif namespace paddle { @@ -53,7 +53,7 @@ Layer* PoolLayer::create(const LayerConfig& config) { const std::string& pool = config.inputs(0).pool_conf().pool_type(); if (pool == "max-projection" || pool == "avg-projection") { return new PoolProjectionLayer(config); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA } else if (CudnnPoolLayer::typeCheck(pool)) { return new CudnnPoolLayer(config); #endif diff --git a/paddle/gserver/layers/SequenceSliceLayer.cpp b/paddle/gserver/layers/SequenceSliceLayer.cpp index d3a83fad27..ce68ca4494 100644 --- a/paddle/gserver/layers/SequenceSliceLayer.cpp +++ b/paddle/gserver/layers/SequenceSliceLayer.cpp @@ -73,9 +73,10 @@ void SequenceSliceLayer::checkInputs() { CHECK(inputSeq.hasSeq()) << "The first input of sequence slice layer " << "must be a sequence."; const MatrixPtr indices1 = getInputValue(1); - CHECK_EQ(static_cast(indices1->getHeight()), - inputSeq.hasSubseq() ? inputSeq.getNumSubSequences() - : inputSeq.getNumSequences()) + CHECK_EQ( + indices1->getHeight(), + static_cast(inputSeq.hasSubseq() ? inputSeq.getNumSubSequences() + : inputSeq.getNumSequences())) << "Height of the second input should be equal to number of sequence " << "in the first input."; if (inputLayers_.size() == 3) { @@ -151,7 +152,7 @@ void SequenceSliceLayer::calSelectedRows(const MatrixPtr starts, if (ends) endPos = inputSeqInfoVec_[i][j] + ends->getElement(rowIdx, k); int seqLen = endPos - begPos + 1; - CHECK_GT(seqLen, 0U); + CHECK_GT(seqLen, 0); for (int m = begPos; m <= endPos; ++m) selectedRows_.push_back(m); hasSubseq ? outSubSeqStartPos_.push_back(outSubSeqStartPos_.back() + seqLen) diff --git a/paddle/gserver/tests/LayerGradUtil.cpp b/paddle/gserver/tests/LayerGradUtil.cpp index a38880e14c..cd957c7c0b 100644 --- a/paddle/gserver/tests/LayerGradUtil.cpp +++ b/paddle/gserver/tests/LayerGradUtil.cpp @@ -674,7 +674,7 @@ void testLayerGradKernel(TestConfig testConf, bool useGpu, bool useWeight, float epsilon) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) return; #endif FLAGS_use_gpu = useGpu; diff --git a/paddle/gserver/tests/MKLDNNTester.cpp b/paddle/gserver/tests/MKLDNNTester.cpp index 2f48e5b2d3..f59618be9d 100644 --- a/paddle/gserver/tests/MKLDNNTester.cpp +++ b/paddle/gserver/tests/MKLDNNTester.cpp @@ -64,15 +64,17 @@ void MKLDNNTester::reset(const TestConfig& dnn, configs_[i], &(layerMaps_[i]), &(parameters_[i]), &(testLayers_[i])); } refLayer_ = testLayers_[REF]; - dnnLayer_ = std::dynamic_pointer_cast(testLayers_[DNN]); - CHECK(dnnLayer_); - // for comparison with Paddle reference results, - // need manually add cpu device output for test - dnnLayer_->addOutputArgument(CPU_DEVICE); + dnnLayer_ = testLayers_[DNN]; EXPECT_EQ(dataLayers_[DNN].size(), dataLayers_[REF].size()); EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size()); - setInputImgSize(); + + // for comparison with Paddle reference results, + // need manually add cpu device output for test + MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast(dnnLayer_); + if (dnnLayer) { + dnnLayer->addOutputArgument(CPU_DEVICE); + } } void MKLDNNTester::setInputImgSize() { @@ -122,7 +124,7 @@ void MKLDNNTester::randomTopDiffs() { void MKLDNNTester::checkForward() { VLOG(MKLDNN_ALL) << "Check Forward"; printTopDatas(); - double delta = compareMatrix(dnnLayer_->getOutput(-1).value, + double delta = compareMatrix(dnnLayer_->getOutput(CPU_DEVICE).value, refLayer_->getOutputValue()); EXPECT_LE(fabs(delta), eps_); } @@ -155,7 +157,10 @@ void MKLDNNTester::checkBackwardWgts() { vector dnnWgts; // used to temply save mkldnn weights saveWgt(parameters_[DNN], dnnWgts); - dnnLayer_->convertWeightsToPaddle(); + MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast(dnnLayer_); + if (dnnLayer) { + dnnLayer->convertWeightsToPaddle(); + } for (size_t i = 0; i < parameters_[DNN].size(); ++i) { const VectorPtr& dnn = parameters_[DNN][i]->getBuf(PARAMETER_VALUE); const VectorPtr& ref = parameters_[REF][i]->getBuf(PARAMETER_VALUE); @@ -322,6 +327,10 @@ void MKLDNNTester::runOnce() { // and clearTopDatas(REF) should be coverd by ref layers clearBotDiffs(REF); clearWgtDiffs(REF); + // it is necessary to clear bottom diffs when only activation is dnn type + if (configs_[DNN].layerConfig.active_type().compare(0, 7, "mkldnn_") == 0) { + clearBotDiffs(DNN); + } } void MKLDNNTester::run(const TestConfig& dnn, @@ -333,8 +342,19 @@ void MKLDNNTester::run(const TestConfig& dnn, float epsilon, bool log, int level) { - VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: " << dnn.layerConfig.type() - << " vs " << ref.layerConfig.type(); + CHECK(dnn.layerConfig.type().compare(0, 7, "mkldnn_") == 0 || + dnn.layerConfig.active_type().compare(0, 7, "mkldnn_") == 0) + << "should be MKLDNN layer or MKLDNN activation"; + if (dnn.layerConfig.type() == ref.layerConfig.type()) { + VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: " + << dnn.layerConfig.active_type() << " vs " + << ref.layerConfig.active_type(); + } else { + VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: " + << dnn.layerConfig.type() << " vs " + << ref.layerConfig.type(); + } + ih_ = inputImgH; iw_ = inputImgW; iter_ = iter; diff --git a/paddle/gserver/tests/MKLDNNTester.h b/paddle/gserver/tests/MKLDNNTester.h index 5ac885638c..171d176ee7 100644 --- a/paddle/gserver/tests/MKLDNNTester.h +++ b/paddle/gserver/tests/MKLDNNTester.h @@ -41,8 +41,7 @@ protected: vector layerMaps_; vector> parameters_; vector testLayers_; - LayerPtr refLayer_; - MKLDNNLayerPtr dnnLayer_; + LayerPtr refLayer_, dnnLayer_; /// run some iterations, all the result should pass size_t iter_; diff --git a/paddle/gserver/tests/test_BatchNorm.cpp b/paddle/gserver/tests/test_BatchNorm.cpp index 659eefa31b..050fde9d0a 100644 --- a/paddle/gserver/tests/test_BatchNorm.cpp +++ b/paddle/gserver/tests/test_BatchNorm.cpp @@ -119,7 +119,7 @@ TEST(Layer, batchNorm) { CHECK_EQ(static_cast(convLayer->getOutputValue()->getWidth()), 576); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA void batchNormInference(int n, int c, int h, int w) { MatrixPtr input = std::make_shared(n, c * h * w); MatrixPtr cudnnOut = std::make_shared(n, c * h * w); diff --git a/paddle/gserver/tests/test_ConvUnify.cpp b/paddle/gserver/tests/test_ConvUnify.cpp index e7325e0cc3..ffcc47e2a8 100644 --- a/paddle/gserver/tests/test_ConvUnify.cpp +++ b/paddle/gserver/tests/test_ConvUnify.cpp @@ -117,7 +117,7 @@ MatrixPtr doOneConvTest(size_t imgSize, } TEST(Layer, convParaUnified) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA MatrixPtr input, resultCpu, resultGpu; /// TEST1 for conv /// diff --git a/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp b/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp index 538d18cdc3..c922237d33 100644 --- a/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp +++ b/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp @@ -228,7 +228,7 @@ void genGroundTruth(vector& beamExpansions, curBeam.groundTruth[j] = *(start + n); curBeam.inBeam[j] = 1; } else { - CHECK_LE(curBeam.rowIdxInBeam[j] + 1, + CHECK_LE((size_t)curBeam.rowIdxInBeam[j] + 1, curBeam.subSeqStartPos.size() - 1); int start = curBeam.subSeqStartPos[curBeam.rowIdxInBeam[j]]; int end = curBeam.subSeqStartPos[curBeam.rowIdxInBeam[j] + 1]; diff --git a/paddle/gserver/tests/test_DetectionOutput.cpp b/paddle/gserver/tests/test_DetectionOutput.cpp index af43dc51fa..dc39c97a87 100644 --- a/paddle/gserver/tests/test_DetectionOutput.cpp +++ b/paddle/gserver/tests/test_DetectionOutput.cpp @@ -150,7 +150,7 @@ TEST(Layer, detectionOutputLayerFwd) { useGpu, result2); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA // GPU case 1. useGpu = true; inputLoc = Matrix::create(1, 16, false, useGpu); diff --git a/paddle/gserver/tests/test_Evaluator.cpp b/paddle/gserver/tests/test_Evaluator.cpp index 93996392d2..62a131171f 100644 --- a/paddle/gserver/tests/test_Evaluator.cpp +++ b/paddle/gserver/tests/test_Evaluator.cpp @@ -51,7 +51,7 @@ void testEvaluator(TestConfig testConf, string testEvaluatorName, size_t batchSize, bool useGpu) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) return; #endif FLAGS_use_gpu = useGpu; diff --git a/paddle/gserver/tests/test_KmaxSeqScore.cpp b/paddle/gserver/tests/test_KmaxSeqScore.cpp index 308abe6816..6386259882 100644 --- a/paddle/gserver/tests/test_KmaxSeqScore.cpp +++ b/paddle/gserver/tests/test_KmaxSeqScore.cpp @@ -97,7 +97,7 @@ TEST(Layer, kmaxSeqScoreLayer) { Matrix::create(subSeqStartPosition.back(), 1, false, false); std::vector mode = {false}; -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA mode.push_back(true); #endif diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp index 090bde7b20..90a3352898 100644 --- a/paddle/gserver/tests/test_LayerGrad.cpp +++ b/paddle/gserver/tests/test_LayerGrad.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include #endif #include @@ -258,7 +258,7 @@ void testProjectionConv(size_t groups, bool isDeconv) { true); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(Projection, conv) { /// test ConvProjection testProjectionConv(1, false); @@ -422,7 +422,7 @@ TEST(Layer, depthwiseConvLayer) { // 'depthwise_conv' is a sepecial case of 'exconv' whose // groups size equals to the input channels size. testDepthwiseConvLayer("exconv", /* useGpu= */ false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testDepthwiseConvLayer("exconv", /* useGpu= */ true); #endif } @@ -480,7 +480,7 @@ void testConvLayer(const string& type, bool trans, bool useGpu) { TEST(Layer, convLayer) { testConvLayer("exconv", /* trans= */ false, /* useGpu= */ false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testConvLayer("exconv", /* trans= */ false, /* useGpu= */ true); testConvLayer("cudnn_conv", /* trans= */ false, /* useGpu= */ true); #endif @@ -525,7 +525,7 @@ TEST(Layer, convTransLayer) { for (auto useGpu : {false, true}) { testConvTransLayer("exconvt", /* trans= */ false, /* useGpu= */ useGpu); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testConvTransLayer("cudnn_convt", /* trans= */ false, /* useGpu= */ true); #endif } @@ -638,7 +638,7 @@ TEST(Layer, SelectiveFullyConnectedLayer) { /* trans= */ false, /* useGup= */ false, false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testLayerGrad(config, "selective_fc", 100, @@ -1210,7 +1210,7 @@ void testPoolLayer(const string& poolType, bool trans, bool useGpu) { testLayerGrad(config, "pool", 100, trans, useGpu); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA void testPoolLayer2(const string& poolType, bool trans, bool useGpu) { TestConfig config; config.inputDefs.push_back({INPUT_DATA, "layer_0", 3200, 0}); @@ -1236,7 +1236,7 @@ TEST(Layer, PoolLayer) { testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ false); testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ true); testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ true); testPoolLayer("cudnn-max-pool", /* trans= */ false, /* useGpu= */ true); @@ -1309,7 +1309,7 @@ void testPool3DLayer(const string& poolType, bool trans, bool useGpu) { TEST(Layer, Pool3DLayer) { testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ false); testPool3DLayer("max", /* trans= */ false, /* useGpu= */ false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ true); testPool3DLayer("max", /* trans= */ false, /* useGpu= */ true); #endif @@ -1695,7 +1695,7 @@ void testBatchNormLayer(const string& type, bool trans, bool useGpu) { TEST(Layer, BatchNormalizationLayer) { testBatchNormLayer("batch_norm", false, false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testBatchNormLayer("batch_norm", false, true); if (hl_get_cudnn_lib_version() >= int(4000)) { testBatchNormLayer("cudnn_batch_norm", false, true); @@ -1744,7 +1744,7 @@ void testBatchNorm3DLayer(const string& type, bool trans, bool useGpu) { TEST(Layer, testBatchNorm3DLayer) { testBatchNorm3DLayer("batch_norm", false, false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testBatchNorm3DLayer("batch_norm", false, true); if (hl_get_cudnn_lib_version() >= int(4000)) { testBatchNorm3DLayer("cudnn_batch_norm", false, true); @@ -2262,7 +2262,7 @@ void test3DConvLayer(const string& type, bool trans, bool useGpu) { TEST(Layer, test3DConvLayer) { test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ true); #endif } @@ -2339,7 +2339,7 @@ void test3DDeConvLayer(const string& type, bool trans, bool useGpu) { TEST(Layer, test3DDeConvLayer) { test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ true); #endif } diff --git a/paddle/gserver/tests/test_MKLDNN.cpp b/paddle/gserver/tests/test_MKLDNN.cpp index b593f65fe4..a70b2f17f4 100644 --- a/paddle/gserver/tests/test_MKLDNN.cpp +++ b/paddle/gserver/tests/test_MKLDNN.cpp @@ -17,6 +17,7 @@ limitations under the License. */ #include #include "MKLDNNTester.h" #include "ModelConfig.pb.h" +#include "paddle/gserver/activations/MKLDNNActivation.h" #include "paddle/math/MathUtils.h" using namespace paddle; // NOLINT @@ -25,17 +26,26 @@ DECLARE_bool(thread_local_rand_use_global_seed); DECLARE_bool(use_gpu); DECLARE_bool(use_mkldnn); -struct testFCDesc { +#define RUN_MKLDNN_TEST(DNN_CONFIG, REF_CONFIG, DESC) \ + MKLDNNTester tester; \ + for (auto bs : {DESC.bs, 1}) { \ + tester.run(DNN_CONFIG, REF_CONFIG, bs, DESC.ih, DESC.iw); \ + } + +#define RUN_MKLDNN_TEST_LAYER(DNN_CONFIG, REF_TYPE, DESC) \ + TestConfig ref = DNN_CONFIG; \ + ref.layerConfig.set_type(REF_TYPE); \ + RUN_MKLDNN_TEST(DNN_CONFIG, ref, DESC) + +struct testFcDesc { int bs; int ic; int oc; int ih, iw; // oh == ow == 1 }; -void testFcLayer(const testFCDesc& pm) { - const std::string compareTypes[] = {"mkldnn_fc", "fc"}; - TestConfig cfg; - cfg.layerConfig.set_type(compareTypes[0]); +static void getMKLDNNFcConfig(TestConfig& cfg, const testFcDesc& pm) { + cfg.layerConfig.set_type("mkldnn_fc"); cfg.layerConfig.set_size(pm.oc); cfg.inputDefs.push_back( {INPUT_DATA, @@ -43,25 +53,25 @@ void testFcLayer(const testFCDesc& pm) { /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw), /* size of weight= */ size_t(pm.oc * pm.ic * pm.ih * pm.iw)}); cfg.layerConfig.add_inputs(); +} - MKLDNNTester tester; +void testFcLayer(const testFcDesc& pm) { + TestConfig dnnConfig; + getMKLDNNFcConfig(dnnConfig, pm); for (auto biasSize : {pm.oc, 0}) { - cfg.biasSize = biasSize; - TestConfig ref = cfg; - ref.layerConfig.set_type(compareTypes[1]); - for (auto bs : {pm.bs, 1}) { - tester.run(cfg, ref, bs, pm.ih, pm.iw); - } + dnnConfig.biasSize = biasSize; + RUN_MKLDNN_TEST_LAYER(dnnConfig, "fc", pm) } } TEST(MKLDNNLayer, FcLayer) { - testFcLayer({/*bs*/ 2, /*ic*/ 2, /*oc*/ 3, /*ih*/ 1, /*iw*/ 1}); - testFcLayer({/*bs*/ 3, /*ic*/ 7, /*oc*/ 19, /*ih*/ 1, /*iw*/ 1}); - testFcLayer({/*bs*/ 8, /*ic*/ 16, /*oc*/ 32, /*ih*/ 13, /*iw*/ 13}); - testFcLayer({/*bs*/ 4, /*ic*/ 12, /*oc*/ 18, /*ih*/ 13, /*iw*/ 11}); - testFcLayer({/*bs*/ 2, /*ic*/ 64, /*oc*/ 32, /*ih*/ 16, /*iw*/ 16}); - testFcLayer({/*bs*/ 15, /*ic*/ 3, /*oc*/ 6, /*ih*/ 16, /*iw*/ 16}); + /* bs, ic, ih, iw, oc */ + testFcLayer({2, 2, 1, 1, 3}); + testFcLayer({3, 7, 1, 1, 19}); + testFcLayer({8, 16, 13, 13, 32}); + testFcLayer({4, 12, 13, 13, 18}); + testFcLayer({2, 64, 16, 16, 32}); + testFcLayer({15, 3, 16, 16, 6}); } struct testConvDesc { @@ -74,13 +84,10 @@ struct testConvDesc { int dh, dw; }; -void testConvLayer(const testConvDesc& pm) { - const std::string compareTypes[] = {"mkldnn_conv", "exconv"}; - TestConfig cfg; - cfg.layerConfig.set_type(compareTypes[0]); +static void getMKLDNNConvConfig(TestConfig& cfg, const testConvDesc& pm) { + cfg.layerConfig.set_type("mkldnn_conv"); cfg.layerConfig.set_num_filters(pm.oc); cfg.layerConfig.set_size(pm.oc * pm.oh * pm.ow); - // cfg.layerConfig.set_partial_sum(1); // TODO: check it cfg.layerConfig.set_shared_biases(true); cfg.inputDefs.push_back( {INPUT_DATA, @@ -114,15 +121,14 @@ void testConvLayer(const testConvDesc& pm) { int oh = outputSize(pm.ih, fh, pm.ph, pm.sh, true); CHECK_EQ(ow, pm.ow) << "output size check failed"; CHECK_EQ(oh, pm.oh) << "output size check failed"; +} - MKLDNNTester tester; +void testConvLayer(const testConvDesc& pm) { + TestConfig dnnConfig; + getMKLDNNConvConfig(dnnConfig, pm); for (auto biasSize : {pm.oc, 0}) { - cfg.biasSize = biasSize; - TestConfig ref = cfg; - ref.layerConfig.set_type(compareTypes[1]); - for (auto bs : {pm.bs, 1}) { - tester.run(cfg, ref, bs, pm.ih, pm.iw); - } + dnnConfig.biasSize = biasSize; + RUN_MKLDNN_TEST_LAYER(dnnConfig, "exconv", pm) } } @@ -142,7 +148,7 @@ TEST(MKLDNNLayer, ConvLayer) { } struct testPoolDesc { - int bs, ch; // input channel and output channel are the same + int bs, ic; // input channel and output channel are the same int ih, iw; int oh, ow; int fh, fw; @@ -150,20 +156,18 @@ struct testPoolDesc { int sh, sw; }; -void testPoolLayer(const testPoolDesc& pm) { - const std::string compareTypes[] = {"mkldnn_pool", "pool"}; - TestConfig cfg; - cfg.layerConfig.set_type(compareTypes[0]); - cfg.layerConfig.set_size(pm.ch * pm.oh * pm.ow); +static void getMKLDNNPoolConfig(TestConfig& cfg, const testPoolDesc& pm) { + cfg.layerConfig.set_type("mkldnn_pool"); + cfg.layerConfig.set_size(pm.ic * pm.oh * pm.ow); cfg.inputDefs.push_back( {INPUT_DATA, "layer_0", - /* size of input layer= */ size_t(pm.ch * pm.ih * pm.iw), + /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw), 0}); LayerInputConfig* input = cfg.layerConfig.add_inputs(); PoolConfig* pool = input->mutable_pool_conf(); - // pool->set_pool_type(poolType); - pool->set_channels(pm.ch); + pool->set_pool_type("avg-projection"); + pool->set_channels(pm.ic); pool->set_img_size(pm.iw); pool->set_img_size_y(pm.ih); pool->set_output_x(pm.ow); @@ -179,20 +183,21 @@ void testPoolLayer(const testPoolDesc& pm) { int ow = outputSize(pm.iw, pm.fw, pm.pw, pm.sw, false); CHECK_EQ(ow, pm.ow) << "output size check failed"; CHECK_EQ(oh, pm.oh) << "output size check failed"; +} - MKLDNNTester tester; +void testPoolLayer(const testPoolDesc& pm) { + TestConfig dnnConfig; + getMKLDNNPoolConfig(dnnConfig, pm); + LayerInputConfig* input = dnnConfig.layerConfig.mutable_inputs(0); + PoolConfig* pool = input->mutable_pool_conf(); for (auto type : {"max-projection", "avg-projection"}) { pool->set_pool_type(type); - TestConfig ref = cfg; - ref.layerConfig.set_type(compareTypes[1]); - for (auto bs : {pm.bs, 1}) { - tester.run(cfg, ref, bs, pm.ih, pm.iw); - } + RUN_MKLDNN_TEST_LAYER(dnnConfig, "pool", pm) } } -TEST(MkldnnLayer, PoolLayer) { - /* bs, ch, ih, iw, oh, ow, fh, fw, ph, pw, sh, sw*/ +TEST(MKLDNNLayer, PoolLayer) { + /* bs, ch, ih, iw, oh, ow, fh, fw, ph, pw, sh, sw */ testPoolLayer({2, 1, 4, 4, 2, 2, 3, 3, 0, 0, 2, 2}); testPoolLayer({10, 8, 16, 16, 8, 8, 2, 2, 0, 0, 2, 2}); testPoolLayer({4, 2, 5, 5, 3, 3, 3, 3, 1, 1, 2, 2}); @@ -203,6 +208,42 @@ TEST(MkldnnLayer, PoolLayer) { testPoolLayer({2, 8, 56, 56, 29, 29, 3, 3, 1, 1, 2, 2}); } +struct testActDesc { + int bs, ic, ih, iw; +}; + +static void getAddtoConfig(TestConfig& cfg, const testActDesc& pm) { + cfg.biasSize = 0; + cfg.layerConfig.set_type("addto"); + size_t layerSize = pm.ic * pm.ih * pm.iw; + cfg.layerConfig.set_size(layerSize); + cfg.inputDefs.push_back({INPUT_DATA, "layer_0", layerSize, 0}); + cfg.layerConfig.add_inputs(); +} + +void testActivation(std::string actType, const testActDesc& pm) { + // TODO(TJ): remove me when paddle support elu activation + if (actType == "mkldnn_elu") { + return; + } + const std::string compareTypes[] = {actType, actType.erase(0, 7)}; + TestConfig cfg; + getAddtoConfig(cfg, pm); + TestConfig ref = cfg; + cfg.layerConfig.set_active_type(compareTypes[0]); + ref.layerConfig.set_active_type(compareTypes[1]); + RUN_MKLDNN_TEST(cfg, ref, pm) +} + +TEST(MKLDNNActivation, Activations) { + auto types = MKLDNNActivation::getAllRegisteredTypes(); + for (auto type : types) { + /* bs, c, h, w*/ + testActivation(type, {16, 64, 32, 32}); + testActivation(type, {2, 8, 1, 1}); + } +} + // TODO(TJ): add branch test int main(int argc, char** argv) { diff --git a/paddle/gserver/tests/test_NetworkCompare.cpp b/paddle/gserver/tests/test_NetworkCompare.cpp index d36f72360f..2b92211936 100644 --- a/paddle/gserver/tests/test_NetworkCompare.cpp +++ b/paddle/gserver/tests/test_NetworkCompare.cpp @@ -243,7 +243,7 @@ TEST(Compare, concat_slice) { compareNetwork(config_file_a, config_file_b); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(Compare, img_pool) { std::string config_file_a = "./gserver/tests/img_pool_a.conf"; std::string config_file_b = "./gserver/tests/img_pool_b.conf"; diff --git a/paddle/gserver/tests/test_PriorBox.cpp b/paddle/gserver/tests/test_PriorBox.cpp index ae0e3bc3d2..8dc5568784 100644 --- a/paddle/gserver/tests/test_PriorBox.cpp +++ b/paddle/gserver/tests/test_PriorBox.cpp @@ -151,7 +151,7 @@ TEST(Layer, priorBoxLayerFwd) { useGpu, result); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA // reset the input parameters variance[1] = 0.1; variance[3] = 0.2; diff --git a/paddle/gserver/tests/test_ProtoDataProvider.cpp b/paddle/gserver/tests/test_ProtoDataProvider.cpp index e11bf402c2..af6472619d 100644 --- a/paddle/gserver/tests/test_ProtoDataProvider.cpp +++ b/paddle/gserver/tests/test_ProtoDataProvider.cpp @@ -485,7 +485,7 @@ TEST(ProtoDataProvider, test) { // Currently in async mode, useGpu is not supported continue; } -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) { continue; } @@ -525,7 +525,7 @@ TEST(ProtoDataProvider, constant_slots) { for (int numConstantSlots : {1, 2}) { for (int useGpu : numTwoArray) { for (int dataCompression : numTwoArray) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) { continue; } @@ -708,7 +708,7 @@ TEST(ProtoSequenceDataProvider, test) { // Currently in async mode, useGpu is not supported continue; } -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) { continue; } diff --git a/paddle/gserver/tests/test_PyDataProvider.cpp b/paddle/gserver/tests/test_PyDataProvider.cpp index db883543c3..fe54799259 100644 --- a/paddle/gserver/tests/test_PyDataProvider.cpp +++ b/paddle/gserver/tests/test_PyDataProvider.cpp @@ -37,7 +37,7 @@ TEST(PyDataProvider, py_fill_slots) { config.clear_files(); std::string dataFile = "gserver/tests/pyDataProvider/pyDataProviderList"; config.set_files(dataFile); -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA bool useGpu = false; #else bool useGpu = true; @@ -71,7 +71,7 @@ TEST(PyDataProvider, py_fill_nest_slots) { std::string dataFile = "gserver/tests/pyDataProvider/pyDataProviderList"; config.set_files(dataFile); EXPECT_EQ(config.IsInitialized(), true); -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA bool useGpu = false; #else bool useGpu = true; diff --git a/paddle/gserver/tests/test_SelectiveFCLayer.cpp b/paddle/gserver/tests/test_SelectiveFCLayer.cpp index ab23d00a2c..4c87fe1bba 100644 --- a/paddle/gserver/tests/test_SelectiveFCLayer.cpp +++ b/paddle/gserver/tests/test_SelectiveFCLayer.cpp @@ -321,7 +321,7 @@ TEST(Layer, SelectiveFcLayer_train_dense_mul) { "filelist=gserver/tests/SelectiveFcTest/dense_mul_list"; for (auto useGpu : {false, true}) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) { break; } @@ -388,7 +388,7 @@ void testSelectiveFcLayerTrainSparseMul(const LayerConfig& config, outMatSelfc->getWidth(), outMatSelfc->getElementCnt())); cpuOutMatSelfc->copyFrom(*outMatSelfc, HPPL_STREAM_DEFAULT); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA if (useGpu) { hl_stream_synchronize(HPPL_STREAM_DEFAULT); } @@ -418,7 +418,7 @@ void testSelectiveFcLayerTrainSparseMul(const LayerConfig& config, MatrixPtr cpuOutMatFc( new CpuMatrix(outMatFc->getHeight(), outMatFc->getWidth())); cpuOutMatFc->copyFrom(*outMatFc, HPPL_STREAM_DEFAULT); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA if (useGpu) { hl_stream_synchronize(HPPL_STREAM_DEFAULT); } @@ -443,7 +443,7 @@ TEST(Layer, SelectiveFcLayer_train_sparse_mul) { selLayerConfig.set_size(fcLayerWidth); testSelectiveFcLayerTrainSparseMul(selLayerConfig, false); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testSelectiveFcLayerTrainSparseMul(selLayerConfig, true); #endif } diff --git a/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp b/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp index e1d4ae1617..3366002ca1 100644 --- a/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp +++ b/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp @@ -195,7 +195,7 @@ TEST(Layer, SeqSliceLayer) { vector> ends; std::vector mode = {false}; -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA mode.push_back(true); #endif genSeqInfo(seqStartPos, subSeqStartPos); diff --git a/paddle/gserver/tests/test_WarpCTCLayer.cpp b/paddle/gserver/tests/test_WarpCTCLayer.cpp index 55427e2f12..da82946006 100644 --- a/paddle/gserver/tests/test_WarpCTCLayer.cpp +++ b/paddle/gserver/tests/test_WarpCTCLayer.cpp @@ -199,7 +199,7 @@ TEST(Layer, WarpCTCLayer) { for (auto batchSize : {1, 10, 32}) { for (auto normByTimes : {false, true}) { for (auto useGpu : {false, true}) { -#ifdef PADDLE_ONLY_CPU +#ifndef PADDLE_WITH_CUDA if (useGpu) continue; #endif LOG(INFO) << "layerSize=" << layerSize << " batchSize=" << batchSize diff --git a/paddle/math/MathFunctions.h b/paddle/math/MathFunctions.h index e8ea6e37ac..8193aa4adf 100644 --- a/paddle/math/MathFunctions.h +++ b/paddle/math/MathFunctions.h @@ -26,7 +26,7 @@ limitations under the License. */ #include #endif -#ifdef PADDLE_USE_ATLAS +#if defined(PADDLE_USE_ATLAS) || defined(PADDLE_USE_VECLIB) extern "C" { #include #include diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp index 0023b4d0f5..c3e34d5309 100644 --- a/paddle/math/Matrix.cpp +++ b/paddle/math/Matrix.cpp @@ -670,7 +670,7 @@ void GpuMatrix::leftMul(Matrix& a, real scaleAB, real scaleT) { } void GpuMatrix::selectRows(Matrix& table, IVector& ids) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA CHECK(dynamic_cast(&table)); CHECK(table.useGpu()); CHECK(ids.useGpu()); @@ -694,7 +694,7 @@ void GpuMatrix::selectRows(Matrix& table, IVector& ids) { } void GpuMatrix::addToRows(Matrix& table, IVector& ids) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA CHECK(dynamic_cast(&table)); CHECK(table.useGpu()); CHECK(ids.useGpu()); @@ -741,7 +741,7 @@ void GpuMatrix::rowMax(Matrix& max) { } void GpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal"; size_t numSamples = getHeight(); size_t beam = maxVal.getWidth(); diff --git a/paddle/math/RowBuffer.h b/paddle/math/RowBuffer.h index dbb829c4e2..9ef5b89680 100644 --- a/paddle/math/RowBuffer.h +++ b/paddle/math/RowBuffer.h @@ -99,7 +99,11 @@ public: /** * @brief clear local buffer. It only affect auto-growth buffer. */ - inline void clear() { rowStore_.clear(); } + inline void clear() { + // swap an empty vector to it to free the memory. + std::vector> empty; + rowStore_.swap(empty); + } /** * @brief get current number of rows. diff --git a/paddle/math/SparseMatrix.cpp b/paddle/math/SparseMatrix.cpp index 6370c77386..284b68d590 100644 --- a/paddle/math/SparseMatrix.cpp +++ b/paddle/math/SparseMatrix.cpp @@ -836,7 +836,7 @@ void GpuSparseMatrix::zeroMem() { } void GpuSparseMatrix::rowMax(IVector& maxIds, Matrix& maxVal) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal"; size_t numSamples = getHeight(); size_t beam = maxVal.getWidth(); diff --git a/paddle/math/Vector.cpp b/paddle/math/Vector.cpp index eb87ee9bb7..ff72672e3a 100644 --- a/paddle/math/Vector.cpp +++ b/paddle/math/Vector.cpp @@ -172,7 +172,7 @@ void GpuVectorT::isEqualTo(const VectorT& b, const T& value) { template void GpuVectorT::selectFrom(const VectorT& src, const VectorT& ids) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA hl_vector_select_from(this->getData(), this->getSize(), src.getData(), @@ -850,7 +850,7 @@ CpuGpuVectorT::CpuGpuVectorT(CpuGpuVectorT& src, size_t size) : sync_(nullptr) { CHECK_LE(offset + size, static_cast(src.getSize())); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA SyncedFlag* flag = src.getSync(); if (*flag == DATA_AT_CPU) { src.copyToGpu(); // will set synchronous data between CPU and GPU @@ -861,7 +861,7 @@ CpuGpuVectorT::CpuGpuVectorT(CpuGpuVectorT& src, auto cMemHandle = (src.getVector(false))->getMemoryHandle(); cpuVectorT_ = std::make_shared>( size, std::dynamic_pointer_cast(cMemHandle), offset); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA auto gMemHandle = (src.getVector(true))->getMemoryHandle(); gpuVectorT_ = std::make_shared>( size, std::dynamic_pointer_cast(gMemHandle), offset); diff --git a/paddle/math/tests/test_Allocator.cpp b/paddle/math/tests/test_Allocator.cpp index 1ca70ea84c..1fecf659e5 100644 --- a/paddle/math/tests/test_Allocator.cpp +++ b/paddle/math/tests/test_Allocator.cpp @@ -68,7 +68,7 @@ void testPoolAllocator() { TEST(Allocator, Pool) { testPoolAllocator(); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testPoolAllocator(); #endif } @@ -92,7 +92,7 @@ TEST(MemoryHandle, Cpu) { EXPECT_EQ(ptr1, ptr2); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(MemoryHandle, Gpu) { int numGpu = hl_get_device_count(); diff --git a/paddle/math/tests/test_BaseMatrix.cpp b/paddle/math/tests/test_BaseMatrix.cpp index 22ce39701f..1766257860 100644 --- a/paddle/math/tests/test_BaseMatrix.cpp +++ b/paddle/math/tests/test_BaseMatrix.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA /** * This test file use autotest::AutoCompare and cmpWithoutArg to compares the * implementation of CPU and GPU member function in diff --git a/paddle/math/tests/test_CpuGpuVector.cpp b/paddle/math/tests/test_CpuGpuVector.cpp index 58bc43a38b..c72f89c824 100644 --- a/paddle/math/tests/test_CpuGpuVector.cpp +++ b/paddle/math/tests/test_CpuGpuVector.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include #include "paddle/math/Vector.h" diff --git a/paddle/math/tests/test_ExecViaCpu.cpp b/paddle/math/tests/test_ExecViaCpu.cpp index 04c856453d..25e0ba11de 100644 --- a/paddle/math/tests/test_ExecViaCpu.cpp +++ b/paddle/math/tests/test_ExecViaCpu.cpp @@ -94,7 +94,7 @@ void testWrapper(F&& f) { } } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(ExecViaCpu, test1) { testWrapper(f); testWrapper(&f); diff --git a/paddle/math/tests/test_GpuProfiler.cpp b/paddle/math/tests/test_GpuProfiler.cpp index e6b5dba446..9402bd3ec4 100644 --- a/paddle/math/tests/test_GpuProfiler.cpp +++ b/paddle/math/tests/test_GpuProfiler.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include #include "paddle/math/Matrix.h" diff --git a/paddle/math/tests/test_Matrix.cpp b/paddle/math/tests/test_Matrix.cpp index 1c21da5b76..2f99fa3581 100644 --- a/paddle/math/tests/test_Matrix.cpp +++ b/paddle/math/tests/test_Matrix.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA /** * This test file use autotest::AutoCompare and cmpWithArg to compares the * implementation of CPU and GPU member function in Matrix.cpp. diff --git a/paddle/math/tests/test_SparseMatrix.cpp b/paddle/math/tests/test_SparseMatrix.cpp index c0572dfdbf..8abbe8d82e 100644 --- a/paddle/math/tests/test_SparseMatrix.cpp +++ b/paddle/math/tests/test_SparseMatrix.cpp @@ -47,7 +47,7 @@ struct MatrixPara { SparseFormat format; }; -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA void test_sparse_matrix_mul(MatrixPara paraA, MatrixPara paraB, MatrixPara paraC) { @@ -452,7 +452,7 @@ TEST(Matrix, SparseMatrixCSRFormatTrimFrom) { matB->trimFrom(*mat); checkSMatrixEqual2(matA, matB); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA GpuSparseMatrixPtr matC = std::make_shared( height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSR, true); matC->trimFrom(*mat); @@ -546,7 +546,7 @@ TEST(Matrix, SparseMatrixCSCFormatTrimFrom) { matB->trimFrom(*mat); checkSMatrixEqual2(matA, matB); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA GpuSparseMatrixPtr matC = std::make_shared( height, trimedWidth, height, FLOAT_VALUE, SPARSE_CSC, true); matC->trimFrom(*mat); diff --git a/paddle/math/tests/test_Tensor.cu b/paddle/math/tests/test_Tensor.cu index 31b693afa8..d03698dee2 100644 --- a/paddle/math/tests/test_Tensor.cu +++ b/paddle/math/tests/test_Tensor.cu @@ -270,7 +270,7 @@ TEST(Unary, BaseOp) { TestUnaryVectorT testCpuIVector( testUnaryBaseOpInt); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestUnaryMatrix testGpuMatrix(testUnaryBaseOp); TestUnaryVectorT testGpuVector(testUnaryBaseOp); TestUnaryVectorT testGpuIVector( @@ -317,7 +317,7 @@ void testUnayrMathOp(Tensor& A1, Tensor& A2) { TEST(Unary, MathOp) { TestUnaryMatrix testCpu(testUnayrMathOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestUnaryMatrix testGpu(testUnayrMathOp); #endif } @@ -374,7 +374,7 @@ void testUnayrCompareOp(Tensor& A1, Tensor& A2) { TEST(Unary, CompareOp) { TestUnaryMatrix testCpu(testUnayrCompareOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestUnaryMatrix testGpu(testUnayrCompareOp); #endif } @@ -536,7 +536,7 @@ void testBinaryBaseOp(Tensor& A1, Tensor& A2, Tensor& B) { TEST(Binary, BaseOp) { TestBinaryMatrix testCpu(testBinaryBaseOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestBinaryMatrix testGpu(testBinaryBaseOp); #endif } @@ -710,7 +710,7 @@ void testBinaryMathOp(Tensor& A1, Tensor& A2, Tensor& B) { TEST(Binary, MathOp) { TestBinaryMatrix testCpu(testBinaryMathOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestBinaryMatrix testGpu(testBinaryMathOp); #endif } @@ -810,7 +810,7 @@ void testBinaryCompareOp(Tensor& A1, Tensor& A2, Tensor& B) { TEST(Binary, CompareOp) { TestBinaryMatrix testCpu(testBinaryCompareOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestBinaryMatrix testGpu(testBinaryCompareOp); #endif } @@ -955,7 +955,7 @@ void testTernaryBaseOp(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { TEST(Ternary, BaseOp) { TestTernaryMatrix testCpu(testTernaryBaseOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestTernaryMatrix testGpu(testTernaryBaseOp); #endif } @@ -1058,7 +1058,7 @@ void testTernaryCompareOp(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { TEST(Ternary, CompareOp) { TestTernaryMatrix testCpu(testTernaryCompareOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestTernaryMatrix testGpu(testTernaryCompareOp); #endif } @@ -1086,7 +1086,7 @@ void testQuaternaryAdd( TEST(Quaternary, BaseOp) { TestQuaternaryMatrix testCpu(testQuaternaryAdd); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestQuaternaryMatrix testGpu(testQuaternaryAdd); #endif } @@ -1156,7 +1156,7 @@ void testQuaternaryCompareOp( TEST(Quaternary, CompareOp) { TestQuaternaryMatrix testCpu(testQuaternaryCompareOp); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TestQuaternaryMatrix testGpu(testQuaternaryCompareOp); #endif } diff --git a/paddle/math/tests/test_TrainingAlgorithm.cpp b/paddle/math/tests/test_TrainingAlgorithm.cpp index 4a88844b43..5ae0aa036f 100644 --- a/paddle/math/tests/test_TrainingAlgorithm.cpp +++ b/paddle/math/tests/test_TrainingAlgorithm.cpp @@ -91,7 +91,7 @@ int VectorCheckErr(const VectorPtr& vector1, const VectorPtr& vector2) { typedef std::function testMatrixFunc; void testCase(testMatrixFunc matrixFunc) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA for (auto useGpu : {false, true}) { #else for (auto useGpu : {false}) { diff --git a/paddle/math/tests/test_batchTranspose.cpp b/paddle/math/tests/test_batchTranspose.cpp index 4eb9837909..b70a619764 100644 --- a/paddle/math/tests/test_batchTranspose.cpp +++ b/paddle/math/tests/test_batchTranspose.cpp @@ -17,7 +17,7 @@ limitations under the License. */ using namespace paddle; // NOLINT -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(MatrixBatchTransTest, test_batch_matrix_transpose) { const int nx = 100; const int ny = 50; diff --git a/paddle/math/tests/test_lazyAssign.cu b/paddle/math/tests/test_lazyAssign.cu index 92afab4ff7..04f23cff55 100644 --- a/paddle/math/tests/test_lazyAssign.cu +++ b/paddle/math/tests/test_lazyAssign.cu @@ -72,7 +72,7 @@ void testLazyAssign(int height, int width) { TEST(lazyAssign, CPU) { testMatrixCase(testLazyAssign); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TEST(lazyAssign, GPU) { testMatrixCase(testLazyAssign); } #endif @@ -142,6 +142,6 @@ void testSgdUpdate(int height, int width) { TEST(sgdUpdate, CPU) { testMatrixCase(testSgdUpdate); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_GPU TEST(sgdUpdate, GPU) { testMatrixCase(testSgdUpdate); } #endif diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp index 061fb22e3f..7e5a1db44a 100644 --- a/paddle/math/tests/test_matrixCompare.cpp +++ b/paddle/math/tests/test_matrixCompare.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA /// This unittest checks GpuMatrix/CpuMatrix get same result, so disable when /// only cpu version. diff --git a/paddle/math/tests/test_perturbation.cpp b/paddle/math/tests/test_perturbation.cpp index 60ebae0153..c7c07c817a 100644 --- a/paddle/math/tests/test_perturbation.cpp +++ b/paddle/math/tests/test_perturbation.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include #include diff --git a/paddle/math/tests/test_sparseMatrixCompare.cpp b/paddle/math/tests/test_sparseMatrixCompare.cpp index a9185a4b24..2b2a391b9d 100644 --- a/paddle/math/tests/test_sparseMatrixCompare.cpp +++ b/paddle/math/tests/test_sparseMatrixCompare.cpp @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA /// This unittest checks GpuSparseMatrix/CpuSparseMatrix get same result, // so disable when /// only cpu version. diff --git a/paddle/memory/.clang-format b/paddle/memory/.clang-format deleted file mode 100644 index 29282dc87e..0000000000 --- a/paddle/memory/.clang-format +++ /dev/null @@ -1,5 +0,0 @@ ---- -Language: Cpp -BasedOnStyle: Google -Standard: Cpp11 -... diff --git a/paddle/memory/.clang-format b/paddle/memory/.clang-format new file mode 120000 index 0000000000..7d28cb3924 --- /dev/null +++ b/paddle/memory/.clang-format @@ -0,0 +1 @@ +../framework/.clang-format \ No newline at end of file diff --git a/paddle/memory/detail/buddy_allocator.cc b/paddle/memory/detail/buddy_allocator.cc index bb44970109..fdc5ed19dc 100644 --- a/paddle/memory/detail/buddy_allocator.cc +++ b/paddle/memory/detail/buddy_allocator.cc @@ -175,7 +175,7 @@ void* BuddyAllocator::SystemAlloc(size_t size) { } BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA if (system_allocator_->UseGpu()) { if ((total_used_ + total_free_) == 0) { // Compute the maximum allocation size for the first allocation. diff --git a/paddle/memory/detail/system_allocator.cc b/paddle/memory/detail/system_allocator.cc index a270bd5958..6c9a46dd09 100644 --- a/paddle/memory/detail/system_allocator.cc +++ b/paddle/memory/detail/system_allocator.cc @@ -62,7 +62,7 @@ void CPUAllocator::Free(void* p, size_t size, size_t index) { bool CPUAllocator::UseGpu() const { return false; } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA void* GPUAllocator::Alloc(size_t& index, size_t size) { // CUDA documentation doesn't explain if cudaMalloc returns nullptr diff --git a/paddle/memory/detail/system_allocator.h b/paddle/memory/detail/system_allocator.h index 82ba322e05..ee9b012f91 100644 --- a/paddle/memory/detail/system_allocator.h +++ b/paddle/memory/detail/system_allocator.h @@ -40,7 +40,7 @@ class CPUAllocator : public SystemAllocator { virtual bool UseGpu() const; }; -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA class GPUAllocator : public SystemAllocator { public: virtual void* Alloc(size_t& index, size_t size); diff --git a/paddle/memory/detail/system_allocator_test.cc b/paddle/memory/detail/system_allocator_test.cc index ba44e06ddb..cd563844e7 100644 --- a/paddle/memory/detail/system_allocator_test.cc +++ b/paddle/memory/detail/system_allocator_test.cc @@ -56,7 +56,7 @@ TEST(CPUAllocator, LockMem) { TestAllocator(a, 0); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(GPUAllocator, Alloc) { paddle::memory::detail::GPUAllocator a; TestAllocator(a, 2048); diff --git a/paddle/memory/memcpy.cc b/paddle/memory/memcpy.cc index 19ec9ba9b2..790420a8ab 100644 --- a/paddle/memory/memcpy.cc +++ b/paddle/memory/memcpy.cc @@ -26,7 +26,7 @@ void Copy(platform::CPUPlace, void* dst, std::memcpy(dst, src, num); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA template <> void Copy(platform::CPUPlace dst_place, void* dst, @@ -80,6 +80,15 @@ void Copy(platform::GPUPlace dst_place, platform::GpuMemcpySync(dst, src, num, cudaMemcpyHostToDevice); } +template <> +void Copy(platform::GPUPlace dst_place, + void* dst, + platform::GPUPlace src_place, + const void* src, size_t num) { + platform::SetDeviceId(dst_place.device); + platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToDevice); +} + #endif // PADDLE_ONLY_CPU } // namespace memory diff --git a/paddle/memory/memcpy.h b/paddle/memory/memcpy.h index 2b9c0eada6..0bccee58c3 100644 --- a/paddle/memory/memcpy.h +++ b/paddle/memory/memcpy.h @@ -33,7 +33,7 @@ namespace memory { template void Copy(DstPlace, void* dst, SrcPlace, const void* src, size_t num); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA /** * \brief Copy memory from one place to another place. diff --git a/paddle/memory/memory.cc b/paddle/memory/memory.cc index 29bc26f9d3..30ce8a82e1 100644 --- a/paddle/memory/memory.cc +++ b/paddle/memory/memory.cc @@ -62,7 +62,7 @@ size_t Used(platform::CPUPlace place) { return GetCPUBuddyAllocator()->Used(); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { using BuddyAllocVec = std::vector; @@ -77,7 +77,7 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { // GPU buddy allocator initialization std::call_once(gpu_allocator_flag, [&]() { - int gpu_num = platform::GetDeviceCount(); + int gpu_num = platform::GetCUDADeviceCount(); allocators.reserve(gpu_num); for (int gpu = 0; gpu < gpu_num; gpu++) { platform::SetDeviceId(gpu); diff --git a/paddle/memory/memory_test.cc b/paddle/memory/memory_test.cc index 53cc63a098..0d402038a0 100644 --- a/paddle/memory/memory_test.cc +++ b/paddle/memory/memory_test.cc @@ -80,7 +80,7 @@ TEST(BuddyAllocator, CPUMultAlloc) { } } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA size_t align(size_t size, paddle::platform::GPUPlace place) { size += sizeof(paddle::memory::detail::Metadata); diff --git a/paddle/operators/.clang-format b/paddle/operators/.clang-format deleted file mode 100644 index 47b8a85206..0000000000 --- a/paddle/operators/.clang-format +++ /dev/null @@ -1,5 +0,0 @@ ---- -Language: Cpp -BasedOnStyle: Google -Standard: Cpp11 -... diff --git a/paddle/operators/.clang-format b/paddle/operators/.clang-format new file mode 120000 index 0000000000..7d28cb3924 --- /dev/null +++ b/paddle/operators/.clang-format @@ -0,0 +1 @@ +../framework/.clang-format \ No newline at end of file diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt index e3e934bccc..0fa1fca2bc 100644 --- a/paddle/operators/CMakeLists.txt +++ b/paddle/operators/CMakeLists.txt @@ -55,6 +55,26 @@ function(op_library TARGET) set(pybind_flag 1) endif() + if ("${TARGET}" STREQUAL "pool_op") + set(pybind_flag 1) + # It's enough to just adding one operator to pybind + file(APPEND ${pybind_file} "USE_OP(pool2d);\n") + endif() + + # activation_op contains several operators + if ("${TARGET}" STREQUAL "activation_op") + set(pybind_flag 1) + # It's enough to just adding one operator to pybind + file(APPEND ${pybind_file} "USE_OP(sigmoid);\n") + endif() + + # reduce_op contains several operators + if ("${TARGET}" STREQUAL "reduce_op") + set(pybind_flag 1) + # It's enough to just adding one operator to pybind + file(APPEND ${pybind_file} "USE_OP(reduce_sum);\n") + endif() + # pybind USE_NO_KERNEL_OP file(READ ${TARGET}.cc TARGET_CONTENT) string(REGEX MATCH "OperatorWithKernel" regex_result "${TARGET_CONTENT}") @@ -81,10 +101,18 @@ add_subdirectory(math) set(DEPS_OPS recurrent_op - cond_op) + cond_op + cross_entropy_op + softmax_with_cross_entropy_op + sum_op) + + op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc DEPS framework_proto tensor net_op) op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op) +op_library(cross_entropy_op DEPS cross_entropy) +op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax) +op_library(sum_op DEPS net_op) list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS}) foreach(src ${GENERAL_OPS}) @@ -96,3 +124,4 @@ set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library") cc_test(gather_test SRCS gather_test.cc DEPS tensor) cc_test(net_op_test SRCS net_op_test.cc DEPS net_op) cc_test(scatter_test SRCS scatter_test.cc DEPS tensor) +cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memory) diff --git a/paddle/operators/accuracy_op.cc b/paddle/operators/accuracy_op.cc index 0c813748b2..82010bfb53 100644 --- a/paddle/operators/accuracy_op.cc +++ b/paddle/operators/accuracy_op.cc @@ -22,24 +22,23 @@ class AccuracyOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL( - ctx.InputVar("Inference"), - "Input(Inference) of AccuracyOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"), - "Input(Label) of AccuracyOp should not be null."); - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Accuracy"), - "Output(Accuracy) of AccuracyOp should not be null."); + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Inference"), + "Input(Inference) of AccuracyOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), + "Input(Label) of AccuracyOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Accuracy"), + "Output(Accuracy) of AccuracyOp should not be null."); - auto *inference = ctx.Input("Inference"); - auto *label = ctx.Input("Label"); + auto inference_dim = ctx->GetInputDim("Inference"); + auto label_dim = ctx->GetInputDim("Label"); - PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label must be a vector"); - PADDLE_ENFORCE_EQ(inference->dims()[0], label->dims()[0], + PADDLE_ENFORCE_EQ(label_dim.size(), 1, "label must be a vector"); + PADDLE_ENFORCE_EQ(inference_dim[0], label_dim[0], "inference size must be the same as label size"); - ctx.Output("Accuracy")->Resize({1}); + ctx->SetOutputDim("Accuracy", {1}); + ctx->ShareLoD("Inference", /*->*/ "Accuracy"); } }; @@ -54,11 +53,15 @@ class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker { // TODO(typhoonzero): AddInput("Weight", ... AddOutput("Accuracy", "The accuracy of current batch"); - AddComment( - R"DOC(Accuracy. It will print accuracy rate for classification. + AddComment(R"DOC( +Accuracy. It will print accuracy rate for classification. The accuracy is: .. math:: -accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})DOC"); +accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples}) + +Both the input `Inference` and `Label` can carry the LoD (Level of Details) +information, or not. But the output only shares the LoD with input `Inference`. +)DOC"); } }; diff --git a/paddle/operators/accuracy_op.cu b/paddle/operators/accuracy_op.cu index 0a6a0fd15c..0ca9ef941d 100644 --- a/paddle/operators/accuracy_op.cu +++ b/paddle/operators/accuracy_op.cu @@ -47,7 +47,7 @@ __global__ void AccuracyCudaKernel(const int N, const int D, const int* Xdata, } template -class AccuracyOpCUDAKernel : public framework::OpKernel { +class AccuracyOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), @@ -69,8 +69,12 @@ class AccuracyOpCUDAKernel : public framework::OpKernel { return; } - AccuracyCudaKernel<<<1, PADDLE_CUDA_NUM_THREADS>>>( - num_samples, infer_width, inference_data, label_data, accuracy_data); + AccuracyCudaKernel<<< + 1, PADDLE_CUDA_NUM_THREADS, 0, + reinterpret_cast( + ctx.device_context()) + .stream()>>>(num_samples, infer_width, inference_data, label_data, + accuracy_data); } }; diff --git a/paddle/operators/accuracy_op.h b/paddle/operators/accuracy_op.h index fe704efe1c..12c6b9aac8 100644 --- a/paddle/operators/accuracy_op.h +++ b/paddle/operators/accuracy_op.h @@ -35,7 +35,7 @@ template ; template -class AccuracyKernel : public framework::OpKernel { +class AccuracyKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* inference = ctx.Input("Inference"); diff --git a/paddle/operators/activation_op.cc b/paddle/operators/activation_op.cc new file mode 100644 index 0000000000..66e9d2c401 --- /dev/null +++ b/paddle/operators/activation_op.cc @@ -0,0 +1,295 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/activation_op.h" + +namespace paddle { +namespace operators { + +class ActivationOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase *ctx) const override { + ctx->SetOutputDim("Y", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Y"); + } +}; + +class ActivationOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase *ctx) const override { + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Y")); + } +}; + +class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SigmoidOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Sigmoid operator"); + AddOutput("Y", "Output of Sigmoid operator"); + AddComment("Sigmoid activation operator, sigmoid = 1 / (1 + exp(-x))"); + } +}; + +class ExpOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ExpOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Exp operator"); + AddOutput("Y", "Output of Exp operator"); + AddComment("Exp activation operator, exp(x) = e^x"); + } +}; + +class ReluOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ReluOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Relu operator"); + AddOutput("Y", "Output of Relu operator"); + AddComment("Relu activation operator, relu(x) = max(x, 0)"); + } +}; + +template +class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker { + public: + LeakyReluOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of LeakyRelu operator"); + AddOutput("Y", "Output of LeakyRelu operator"); + AddComment( + "LeakyRelu activation operator, " + "leaky_relu = max(x, alpha * x)"); + AddAttr("alpha", "The small negative slope") + .SetDefault(static_cast(0.02f)); + } +}; + +class TanhOpMaker : public framework::OpProtoAndCheckerMaker { + public: + TanhOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Tanh operator"); + AddOutput("Y", "Output of Tanh operator"); + AddComment( + "Tanh activation operator, tanh = (exp(x) - exp(-x)) / (exp(x) + " + "exp(-x))"); + } +}; + +class TanhShrinkOpMaker : public framework::OpProtoAndCheckerMaker { + public: + TanhShrinkOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of TanhShrink operator"); + AddOutput("Y", "Output of TanhShrink operator"); + AddComment("TanhShrink activation operator, tanhshrink(x) = x - tanh(x)"); + } +}; + +class SqrtOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SqrtOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Sqrt operator"); + AddOutput("Y", "Output of Sqrt operator"); + AddComment("Sqrt activation operator, sqrt(x) = x^(1/2)"); + } +}; + +class AbsOpMaker : public framework::OpProtoAndCheckerMaker { + public: + AbsOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Abs operator"); + AddOutput("Y", "Output of Abs operator"); + AddComment("Abs activation operator, abs(x) = |x|"); + } +}; + +class ReciprocalOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ReciprocalOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Reciprocal operator"); + AddOutput("Y", "Output of Reciprocal operator"); + AddComment("Reciprocal activation operator, reciprocal(x) = 1 / x"); + } +}; + +class LogOpMaker : public framework::OpProtoAndCheckerMaker { + public: + LogOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Log operator"); + AddOutput("Y", "Output of Log operator"); + AddComment("Log activation operator, log(x) = natural logarithm of x"); + } +}; + +class SquareOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SquareOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Square operator"); + AddOutput("Y", "Output of Square operator"); + AddComment("Square activation operator, square(x) = x^2"); + } +}; + +class SoftsignOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SoftsignOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Softsign operator"); + AddOutput("Y", "Output of Softsign operator"); + AddComment("Softsign activation operator, softsign(x) = x / (1 + |x|)"); + } +}; + +template +class BReluOpMaker : public framework::OpProtoAndCheckerMaker { + public: + BReluOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of BRelu operator"); + AddOutput("Y", "Output of BRelu operator"); + AddComment("BRelu activation operator, brelu = max(min(x, t_min), t_max)"); + AddAttr("t_min", "The min marginal value of BRelu") + .SetDefault(static_cast(0)); + AddAttr("t_max", "The max marginal value of BRelu") + .SetDefault(static_cast(24)); + } +}; + +template +class SoftReluOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SoftReluOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of SoftRelu operator"); + AddOutput("Y", "Output of SoftRelu operator"); + AddComment( + "SoftRelu activation operator, soft_relu = log(1 + exp(max(min(x, " + "threshold), threshold)))"); + AddAttr("threshold", "The threshold value of SoftRelu") + .SetDefault(static_cast(40)); + } +}; + +template +class PowOpMaker : public framework::OpProtoAndCheckerMaker { + public: + PowOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Pow operator"); + AddOutput("Y", "Output of Pow operator"); + AddComment("Pow activation operator, pow(x, factor) = x^factor"); + AddAttr("factor", "The exponential factor of Pow") + .SetDefault(static_cast(1)); + } +}; + +template +class STanhOpMaker : public framework::OpProtoAndCheckerMaker { + public: + STanhOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of STanh operator"); + AddOutput("Y", "Output of STanh operator"); + AddComment("STanh activation operator, stanh = b * tanh(a * x)"); + AddAttr("scale_a", "The scale parameter of a for the input") + .SetDefault(static_cast(2 / 3)); + AddAttr("scale_b", "The scale parameter of b for the input") + .SetDefault(static_cast(1.7159)); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP(sigmoid, ops::ActivationOp, ops::SigmoidOpMaker, sigmoid_grad, + ops::ActivationOpGrad); + +REGISTER_OP(exp, ops::ActivationOp, ops::ExpOpMaker, exp_grad, + ops::ActivationOpGrad); + +REGISTER_OP(relu, ops::ActivationOp, ops::ReluOpMaker, relu_grad, + ops::ActivationOpGrad); + +REGISTER_OP(tanh, ops::ActivationOp, ops::TanhOpMaker, tanh_grad, + ops::ActivationOpGrad); + +REGISTER_OP(tanh_shrink, ops::ActivationOp, ops::TanhShrinkOpMaker, + tanh_shrink_grad, ops::ActivationOpGrad); + +REGISTER_OP(sqrt, ops::ActivationOp, ops::SqrtOpMaker, sqrt_grad, + ops::ActivationOpGrad); + +REGISTER_OP(abs, ops::ActivationOp, ops::AbsOpMaker, abs_grad, + ops::ActivationOpGrad); + +REGISTER_OP(reciprocal, ops::ActivationOp, ops::ReciprocalOpMaker, + reciprocal_grad, ops::ActivationOpGrad); + +REGISTER_OP(log, ops::ActivationOp, ops::LogOpMaker, log_grad, + ops::ActivationOpGrad); + +REGISTER_OP(square, ops::ActivationOp, ops::SquareOpMaker, square_grad, + ops::ActivationOpGrad); + +REGISTER_OP(softsign, ops::ActivationOp, ops::SoftsignOpMaker, softsign_grad, + ops::ActivationOpGrad); + +REGISTER_OP(brelu, ops::ActivationOp, ops::BReluOpMaker, brelu_grad, + ops::ActivationOpGrad); + +REGISTER_OP(leaky_relu, ops::ActivationOp, ops::LeakyReluOpMaker, + leaky_relu_grad, ops::ActivationOpGrad); + +REGISTER_OP(soft_relu, ops::ActivationOp, ops::SoftReluOpMaker, + soft_relu_grad, ops::ActivationOpGrad); + +REGISTER_OP(pow, ops::ActivationOp, ops::PowOpMaker, pow_grad, + ops::ActivationOpGrad); + +REGISTER_OP(stanh, ops::ActivationOp, ops::STanhOpMaker, stanh_grad, + ops::ActivationOpGrad); + +#define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor) \ + REGISTER_OP_CPU_KERNEL( \ + act_type, \ + paddle::operators::ActivationKernel>); \ + REGISTER_OP_CPU_KERNEL(act_type##_grad, \ + paddle::operators::ActivationGradKernel< \ + paddle::platform::CPUPlace, \ + paddle::operators::grad_functor>); + +FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL); diff --git a/paddle/operators/activation_op.cu b/paddle/operators/activation_op.cu new file mode 100644 index 0000000000..93e9f1c694 --- /dev/null +++ b/paddle/operators/activation_op.cu @@ -0,0 +1,28 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/activation_op.h" + +#define REGISTER_ACTIVATION_GPU_KERNEL(act_type, functor, grad_functor) \ + REGISTER_OP_GPU_KERNEL( \ + act_type, \ + paddle::operators::ActivationKernel>); \ + REGISTER_OP_GPU_KERNEL(act_type##_grad, \ + paddle::operators::ActivationGradKernel< \ + paddle::platform::GPUPlace, \ + paddle::operators::grad_functor>); + +FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_GPU_KERNEL); diff --git a/paddle/operators/activation_op.h b/paddle/operators/activation_op.h new file mode 100644 index 0000000000..2450601742 --- /dev/null +++ b/paddle/operators/activation_op.h @@ -0,0 +1,429 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class ActivationKernel + : public framework::OpKernel { + public: + using T = typename Functor::ELEMENT_TYPE; + + void Compute(const framework::ExecutionContext& context) const override { + auto* X = context.Input("X"); + auto* Y = context.Output("Y"); + Y->mutable_data(context.GetPlace()); + + auto x = framework::EigenVector::Flatten(*X); + auto y = framework::EigenVector::Flatten(*Y); + auto place = context.GetEigenDevice(); + Functor functor; + + auto attrs = functor.GetAttrs(); + for (auto& attr : attrs) { + *attr.second = context.Attr(attr.first); + } + functor(place, x, y); + } +}; + +template +class ActivationGradKernel + : public framework::OpKernel { + public: + using T = typename Functor::ELEMENT_TYPE; + void Compute(const framework::ExecutionContext& context) const override { + auto* X = context.Input("X"); + auto* Y = context.Input("Y"); + auto* dY = context.Input(framework::GradVarName("Y")); + auto* dX = context.Output(framework::GradVarName("X")); + dX->mutable_data(context.GetPlace()); + + auto dy = framework::EigenVector::Flatten(*dY); + auto x = framework::EigenVector::Flatten(*X); + auto y = framework::EigenVector::Flatten(*Y); + auto dx = framework::EigenVector::Flatten(*dX); + auto place = context.GetEigenDevice(); + Functor functor; + auto attrs = functor.GetAttrs(); + for (auto& attr : attrs) { + *attr.second = context.Attr(attr.first); + } + functor(place, x, y, dy, dx); + } +}; + +template +struct BaseActivationFunctor { + using ELEMENT_TYPE = T; + + using AttrPair = std::vector>; + + AttrPair GetAttrs() { return AttrPair(); } +}; + +// sigmoid(x) = 1 / (1 + exp(-x)) +template +struct SigmoidFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) const { + y.device(d) = static_cast(1) / (static_cast(1) + (-x).exp()); + } +}; + +template +struct SigmoidGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * y * (static_cast(1) - y); + } +}; + +// exp(x) = e^x +template +struct ExpFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.exp(); + } +}; + +template +struct ExpGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * y; + } +}; + +// relu(x) = max(x, 0) +template +struct ReluFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.cwiseMax(static_cast(0)); + } +}; + +template +struct ReluGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * (x > static_cast(0)).template cast(); + } +}; + +// tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x)) +template +struct TanhFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.tanh(); + } +}; + +template +struct TanhGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * (static_cast(1) - y * y); + } +}; + +// tanhshrink(x) = x - tanh(x) +// where tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x)) +template +struct TanhShrinkFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x - x.tanh(); + } +}; + +template +struct TanhShrinkGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * (x.tanh() * x.tanh()); + } +}; + +// sqrt(x) = x^(1/2) +template +struct SqrtFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.sqrt(); + } +}; + +template +struct SqrtGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + const Y y_conj = Eigen::numext::conj(y); + dx.device(d) = static_cast(0.5) * dy / y_conj; + } +}; + +// abs(x) = |x| +template +struct AbsFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.abs(); + } +}; + +template +struct AbsGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * x.sign(); + } +}; + +// reciprocal(x) = 1 / x +template +struct ReciprocalFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) const { + y.device(d) = static_cast(1) / x; + } +}; + +template +struct ReciprocalGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * static_cast(-1) * y * y; + } +}; + +// log(x) = natural logarithm of x +template +struct LogFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.log(); + } +}; + +template +struct LogGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * (static_cast(1) / x); + } +}; + +// square(x) = x^2 +template +struct SquareFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.square(); + } +}; + +template +struct SquareGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * static_cast(2) * x; + } +}; + +template +struct BReluFunctor : public BaseActivationFunctor { + float t_min; + float t_max; + + // NOTE: Explicit hides the `BaseActivationFunctor::GetAttrs` + // not polymorphism for speed. + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"t_min", &t_min}, {"t_max", &t_max}}; + } + + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.cwiseMax(t_min).cwiseMin(t_max); + } +}; + +template +struct BReluGradFunctor : public BaseActivationFunctor { + float t_min; + float t_max; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"t_min", &t_min}, {"t_max", &t_max}}; + } + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * ((x > t_min) * (x < t_max)).template cast(); + } +}; + +// softsign(x) = x / (1 + |x|) +template +struct SoftsignFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y) { + y.device(d) = x / (static_cast(1) + x.abs()); + } +}; + +// d(softsign(x))/dx = 1 / (1 + |x|)^2 +// Taken from https://en.wikipedia.org/wiki/Activation_function +template +struct SoftsignGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Y y, dY dy, dX dx) { + dx.device(d) = + dy * (static_cast(1) / (static_cast(1) + x.abs()).square()); + } +}; + +template +struct SoftReluFunctor : public BaseActivationFunctor { + float threshold; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"threshold", &threshold}}; + } + + template + void operator()(Device d, X x, Y y) const { + auto temp = x.cwiseMax(-threshold).cwiseMin(threshold); + y.device(d) = (static_cast(1) + temp.exp()).log(); + } +}; + +template +struct SoftReluGradFunctor : public BaseActivationFunctor { + float threshold; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"threshold", &threshold}}; + } + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + auto temp = ((x > -threshold) * (x < threshold)).template cast().eval(); + dx.device(d) = dy * (static_cast(1) - (-y).exp()) * temp; + } +}; + +template +struct LeakyReluFunctor : public BaseActivationFunctor { + float alpha; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"alpha", &alpha}}; + } + + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.cwiseMax(alpha * x); + } +}; + +template +struct LeakyReluGradFunctor : public BaseActivationFunctor { + float alpha; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"alpha", &alpha}}; + } + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + auto temp1 = alpha * (x < static_cast(0)).template cast().eval(); + auto temp2 = (x >= static_cast(0)).template cast().eval(); + dx.device(d) = dy * (temp1 + temp2).template cast(); + } +}; + +template +struct PowFunctor : public BaseActivationFunctor { + float factor; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"factor", &factor}}; + } + template + void operator()(Device d, X x, Y y) const { + y.device(d) = x.pow(factor); + } +}; + +template +struct PowGradFunctor : public BaseActivationFunctor { + float factor; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"factor", &factor}}; + } + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + dx.device(d) = dy * factor * x.pow(factor - static_cast(1)); + } +}; + +template +struct STanhFunctor : public BaseActivationFunctor { + float scale_a; + float scale_b; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"scale_a", &scale_a}, {"scale_b", &scale_b}}; + } + + template + void operator()(Device d, X x, Y y) const { + y.device(d) = scale_b * (scale_a * x).tanh(); + } +}; + +template +struct STanhGradFunctor : public BaseActivationFunctor { + float scale_a; + float scale_b; + typename BaseActivationFunctor::AttrPair GetAttrs() { + return {{"scale_a", &scale_a}, {"scale_b", &scale_b}}; + } + + template + void operator()(Device d, X x, Y y, dY dy, dX dx) const { + auto temp = (scale_a * x).tanh() * (scale_a * x).tanh(); + dx.device(d) = dy * scale_a * scale_b * (static_cast(1) - temp); + } +}; + +} // namespace operators +} // namespace paddle + +#define FOR_EACH_KERNEL_FUNCTOR(__macro) \ + __macro(sigmoid, SigmoidFunctor, SigmoidGradFunctor); \ + __macro(exp, ExpFunctor, ExpGradFunctor); \ + __macro(relu, ReluFunctor, ReluGradFunctor); \ + __macro(tanh, TanhFunctor, TanhGradFunctor); \ + __macro(sqrt, SqrtFunctor, SqrtGradFunctor); \ + __macro(abs, AbsFunctor, AbsGradFunctor); \ + __macro(reciprocal, ReciprocalFunctor, ReciprocalGradFunctor); \ + __macro(log, LogFunctor, LogGradFunctor); \ + __macro(square, SquareFunctor, SquareGradFunctor); \ + __macro(brelu, BReluFunctor, BReluGradFunctor); \ + __macro(soft_relu, SoftReluFunctor, SoftReluGradFunctor); \ + __macro(pow, PowFunctor, PowGradFunctor); \ + __macro(stanh, STanhFunctor, STanhGradFunctor); \ + __macro(softsign, SoftsignFunctor, SoftsignGradFunctor); \ + __macro(leaky_relu, LeakyReluFunctor, LeakyReluGradFunctor); \ + __macro(tanh_shrink, TanhShrinkFunctor, TanhShrinkGradFunctor) diff --git a/paddle/operators/adadelta_op.cc b/paddle/operators/adadelta_op.cc new file mode 100644 index 0000000000..bd8c93b4a1 --- /dev/null +++ b/paddle/operators/adadelta_op.cc @@ -0,0 +1,115 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/adadelta_op.h" + +namespace paddle { +namespace operators { + +class AdadeltaOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Param"), + "Input(Param) of AdadeltaOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grad"), + "Input(Grad) of AdadeltaOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("AvgSquaredGrad"), + "Input(AvgSquaredGrad) of AdadeltaOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("AvgSquaredUpdate"), + "Input(AvgSquaredUpdate) of AdadeltaOp should not be null."); + + PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), + "Output(ParamOut) of AdadeltaOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("AvgSquaredGradOut"), + "Output(AvgSquaredGradOut) of AdadeltaOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("AvgSquaredUpdateOut"), + "Output(AvgSquaredUpdateOut) of AdadeltaOp should not be null."); + + auto param_dim = ctx->GetInputDim("Param"); + PADDLE_ENFORCE_EQ( + param_dim, ctx->GetInputDim("Grad"), + "param and grad input of AdadeltaOp should have same dimension"); + PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("AvgSquaredGrad"), + "Param and AvgSquaredGrad input of AdadeltaOp " + "should have same dimension"); + PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("AvgSquaredUpdate"), + "Param and AvgSquaredUpdate input of AdadeltaOp " + "should have same dimension"); + + ctx->SetOutputDim("ParamOut", param_dim); + ctx->SetOutputDim("AvgSquaredGradOut", param_dim); + ctx->SetOutputDim("AvgSquaredUpdateOut", param_dim); + } +}; + +class AdadeltaOpMaker : public framework::OpProtoAndCheckerMaker { + public: + AdadeltaOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Param", "(Tensor) Input parameter"); + AddInput("Grad", "(Tensor) Input gradient"); + AddInput("AvgSquaredGrad", + "(Tensor) Input expectation of squared gradient"); + AddInput("AvgSquaredUpdate", + "(Tensor) Input expectation of squared parameter updates"); + + AddOutput("ParamOut", "(Tensor) Output parameter"); + AddOutput("AvgSquaredGradOut", + "(Tensor) Output expectation of squared gradient"); + AddOutput("AvgSquaredUpdateOut", + "(Tensor) Output expectation of squared parameter updates"); + + AddAttr("rho", + "(float, default 0.95) Exponential decay rate " + "for squared gradients.") + .SetDefault(0.95f); + AddAttr("epsilon", + "(float, default 1.0e-6) Constant for " + "numerical stability") + .SetDefault(1.0e-6f); + AddComment(R"DOC( +Adadelta Updates Operator. + +This implements the Adadelta optimizer[1]. Adadelta is a per-dimension +adaptive learning rate method for gradient descent. + +Adadelta updates: + +avg_squared_grad_out = rho * avg_squared_grad + (1 - rho) * grad * grad +param_update = - sqrt((avg_squared_update + epsilon) / + (avg_squared_grad_out + epsilon)) * grad +avg_squared_update_out = rho * avg_squared_update + (1 - rho) * param_update**2 +param_out = param + param_update + +References: + [1] ADADELTA: An Adaptive Learning Rate Method + https://arxiv.org/abs/1212.5701 + +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(adadelta, ops::AdadeltaOp, ops::AdadeltaOpMaker); +REGISTER_OP_CPU_KERNEL( + adadelta, ops::AdadeltaOpKernel); diff --git a/paddle/operators/add_op.cu b/paddle/operators/adadelta_op.cu similarity index 80% rename from paddle/operators/add_op.cu rename to paddle/operators/adadelta_op.cu index d9c6d20a6c..3af1c8c8e9 100644 --- a/paddle/operators/add_op.cu +++ b/paddle/operators/adadelta_op.cu @@ -12,7 +12,9 @@ See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/operators/add_op.h" +#define EIGEN_USE_GPU +#include "paddle/operators/adadelta_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(add, ops::AddKernel); +REGISTER_OP_GPU_KERNEL( + adadelta, ops::AdadeltaOpKernel); diff --git a/paddle/operators/adadelta_op.h b/paddle/operators/adadelta_op.h new file mode 100644 index 0000000000..d29e15c435 --- /dev/null +++ b/paddle/operators/adadelta_op.h @@ -0,0 +1,69 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class AdadeltaOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto param_out_tensor = ctx.Output("ParamOut"); + auto avg_squared_grad_out_tensor = + ctx.Output("AvgSquaredGradOut"); + auto avg_squared_update_out_tensor = + ctx.Output("AvgSquaredUpdateOut"); + + param_out_tensor->mutable_data(ctx.GetPlace()); + avg_squared_grad_out_tensor->mutable_data(ctx.GetPlace()); + avg_squared_update_out_tensor->mutable_data(ctx.GetPlace()); + + float rho = ctx.Attr("rho"); + float epsilon = ctx.Attr("epsilon"); + + auto param = framework::EigenVector::Flatten( + *ctx.Input("Param")); + auto grad = framework::EigenVector::Flatten( + *ctx.Input("Grad")); + // Squared gradient accumulator + auto avg_squared_grad = framework::EigenVector::Flatten( + *ctx.Input("AvgSquaredGrad")); + // Squared updates accumulator + auto avg_squared_update = framework::EigenVector::Flatten( + *ctx.Input("AvgSquaredUpdate")); + auto param_out = framework::EigenVector::Flatten(*param_out_tensor); + auto avg_squared_grad_out = + framework::EigenVector::Flatten(*avg_squared_grad_out_tensor); + auto avg_squared_update_out = + framework::EigenVector::Flatten(*avg_squared_update_out_tensor); + auto place = ctx.GetEigenDevice(); + + avg_squared_grad_out.device(place) = + rho * avg_squared_grad + (1 - rho) * grad.square(); + auto update = + -((avg_squared_update + epsilon) / (avg_squared_grad_out + epsilon)) + .sqrt() * + grad; + avg_squared_update_out.device(place) = + rho * avg_squared_update + (1 - rho) * update.square(); + param_out.device(place) = param + update; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/adagrad_op.cc b/paddle/operators/adagrad_op.cc new file mode 100644 index 0000000000..ea2ff3c503 --- /dev/null +++ b/paddle/operators/adagrad_op.cc @@ -0,0 +1,93 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/adagrad_op.h" + +namespace paddle { +namespace operators { + +class AdagradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Param"), + "Input(Param) of AdagradOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grad"), + "Input(Grad) of AdagradOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Moment"), + "Input(Moment) of AdagradOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("LearningRate"), + "Input(LearningRate) of AdagradOp should not be null."); + + PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), + "Output(ParamOut) of AdagradOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("MomentOut"), + "Output(MomentOut) of AdagradOp should not be null."); + + auto lr_dims = ctx->GetInputDim("LearningRate"); + PADDLE_ENFORCE_EQ(framework::product(lr_dims), 1, + "LearningRate should have one element"); + auto param_dims = ctx->GetInputDim("Param"); + PADDLE_ENFORCE_EQ( + param_dims, ctx->GetInputDim("Grad"), + "Param and Grad input of AdagradOp should have the same dimension."); + PADDLE_ENFORCE_EQ( + param_dims, ctx->GetInputDim("Moment"), + "Param and Moment input of AdagradOp should have the same dimension."); + + ctx->SetOutputDim("ParamOut", param_dims); + ctx->SetOutputDim("MomentOut", param_dims); + } +}; + +class AdagradOpMaker : public framework::OpProtoAndCheckerMaker { + public: + AdagradOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Param", "(Tensor) Input parameter"); + AddInput("Grad", "(Tensor) Input gradient"); + AddInput("Moment", "(Tensor) Second moment"); + AddInput("LearningRate", "(Tensor) Learning rate"); + + AddOutput("ParamOut", "(Tensor) Output parameter"); + AddOutput("MomentOut", "(Tensor) Output second moment"); + + AddAttr("epsilon", + "(float, default 1.0e-6) " + "Constant for numerical stability") + .SetDefault(1.0e-6f); + AddComment(R"DOC( + +Adaptive Gradient Algorithm (Adagrad). + +moment_out = moment + grad * grad +param_out = param - learning_rate * grad / (sqrt(moment_out) + epsilon) + +The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) +does not have the epsilon attribute. It is added here for numerical stability +by avoiding division by zero. + +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(adagrad, ops::AdagradOp, ops::AdagradOpMaker); +REGISTER_OP_CPU_KERNEL(adagrad, + ops::AdagradOpKernel); diff --git a/paddle/operators/adagrad_op.cu b/paddle/operators/adagrad_op.cu new file mode 100644 index 0000000000..a5b7951121 --- /dev/null +++ b/paddle/operators/adagrad_op.cu @@ -0,0 +1,20 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/adagrad_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(adagrad, + ops::AdagradOpKernel); diff --git a/paddle/operators/adagrad_op.h b/paddle/operators/adagrad_op.h new file mode 100644 index 0000000000..c5d8f751d3 --- /dev/null +++ b/paddle/operators/adagrad_op.h @@ -0,0 +1,55 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class AdagradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto param_out_tensor = ctx.Output("ParamOut"); + auto moment_out_tensor = ctx.Output("MomentOut"); + + param_out_tensor->mutable_data(ctx.GetPlace()); + moment_out_tensor->mutable_data(ctx.GetPlace()); + + float epsilon = ctx.Attr("epsilon"); + + auto param = framework::EigenVector::Flatten( + *ctx.Input("Param")); + auto grad = framework::EigenVector::Flatten( + *ctx.Input("Grad")); + auto moment = framework::EigenVector::Flatten( + *ctx.Input("Moment")); + auto lr = framework::EigenVector::Flatten( + *ctx.Input("LearningRate")); + + auto param_out = framework::EigenVector::Flatten(*param_out_tensor); + auto moment_out = framework::EigenVector::Flatten(*moment_out_tensor); + auto place = ctx.GetEigenDevice(); + + moment_out.device(place) = moment + grad * grad; + Eigen::DSizes m_dsize(moment_out_tensor->numel()); + param_out.device(place) = + param - lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc deleted file mode 100644 index e83c1efeaf..0000000000 --- a/paddle/operators/add_op.cc +++ /dev/null @@ -1,70 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/operators/add_op.h" - -namespace paddle { -namespace operators { - -class AddOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of AddOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), - "Input(Y) of AddOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of AddOp should not be null."); - - PADDLE_ENFORCE_EQ(ctx.Input("X")->dims(), - ctx.Input("Y")->dims(), - "Two input of Add Op's dimension must be same."); - ctx.Output("Out")->Resize( - ctx.Input("X")->dims()); - } -}; - -class AddOpMaker : public framework::OpProtoAndCheckerMaker { - public: - AddOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The first input of add op"); - AddInput("Y", "The second input of add op"); - AddOutput("Out", "The output of add op"); - AddComment(R"DOC( -Two Element Add Operator. - -The equation is: Out = X + Y -)DOC"); - } -}; - -class AddOpGrad : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext &ctx) const override {} -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; -REGISTER_OP(add, ops::AddOp, ops::AddOpMaker, add_grad, ops::AddOpGrad); - -REGISTER_OP_CPU_KERNEL(add, ops::AddKernel); diff --git a/paddle/operators/add_op.h b/paddle/operators/add_op.h deleted file mode 100644 index a7307b6818..0000000000 --- a/paddle/operators/add_op.h +++ /dev/null @@ -1,48 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include "paddle/framework/eigen.h" -#include "paddle/framework/op_registry.h" - -namespace paddle { -namespace operators { - -using Tensor = framework::Tensor; -template -using EigenVector = framework::EigenVector; - -template -class AddKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* input0 = context.Input("X"); - auto* input1 = context.Input("Y"); - auto* output = context.Output("Out"); - - output->mutable_data(context.GetPlace()); - - auto X = EigenVector::Flatten(*input0); - auto Y = EigenVector::Flatten(*input1); - auto Z = EigenVector::Flatten(*output); - - auto place = context.GetEigenDevice(); - - Z.device(place) = X + Y; - } -}; - -} // namespace operators -} // namespace paddle diff --git a/paddle/operators/clip_op.cc b/paddle/operators/clip_op.cc new file mode 100644 index 0000000000..b3dd060fd7 --- /dev/null +++ b/paddle/operators/clip_op.cc @@ -0,0 +1,84 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/clip_op.h" + +namespace paddle { +namespace operators { + +class ClipOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ClipOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ClipOp should not be null."); + auto x_dims = ctx->GetInputDim("X"); + auto max = Attr("max"); + auto min = Attr("min"); + PADDLE_ENFORCE_LT(min, max, "max should be greater than min."); + ctx->SetOutputDim("Out", x_dims); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +template +class ClipOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ClipOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "(Tensor)The input of clip op." + "The number of dimensions must be between [1, 9]."); + AddOutput("Out", "(Tensor)The output of clip op with shape as input(X)"); + AddAttr( + "min", "(float)Minimum value, under which element is replaced by min."); + AddAttr( + "max", "(float)Maximum value, above which element is replaced by max"); + AddComment(R"DOC( +Clip operator limits the given input within an interval. The interval is +specified with arguments 'min' and 'max'. +)DOC"); + } +}; + +class ClipOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto x_dims = ctx->GetInputDim("X"); + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(clip, ops::ClipOp, ops::ClipOpMaker, clip_grad, + ops::ClipOpGrad); +REGISTER_OP_CPU_KERNEL(clip, + ops::ClipKernel); +REGISTER_OP_CPU_KERNEL(clip_grad, + ops::ClipGradKernel); diff --git a/paddle/operators/clip_op.cu b/paddle/operators/clip_op.cu new file mode 100644 index 0000000000..ca9701298f --- /dev/null +++ b/paddle/operators/clip_op.cu @@ -0,0 +1,21 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/clip_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(clip, + ops::ClipKernel); +REGISTER_OP_GPU_KERNEL(clip_grad, + ops::ClipGradKernel); diff --git a/paddle/operators/clip_op.h b/paddle/operators/clip_op.h new file mode 100644 index 0000000000..ac702e9935 --- /dev/null +++ b/paddle/operators/clip_op.h @@ -0,0 +1,97 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/platform/transform.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; +using platform::Transform; + +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 +class ClipKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto max = context.Attr("max"); + auto min = context.Attr("min"); + auto* x = context.Input("X"); + auto* out = context.Output("Out"); + T* out_data = out->mutable_data(context.GetPlace()); + const T* x_data = x->data(); + int64_t numel = x->numel(); + Transform trans; + trans(context.device_context(), x_data, x_data + numel, out_data, + ClipFunctor(min, max)); + } +}; + +template +class ClipGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto max = context.Attr("max"); + auto min = context.Attr("min"); + auto* d_out = context.Input(framework::GradVarName("Out")); + auto* d_x = context.Output(framework::GradVarName("X")); + if (d_x != nullptr) { + auto* x = context.Input("X"); + int64_t numel = d_out->numel(); + auto* d_x_data = d_x->mutable_data(context.GetPlace()); + const T* d_out_data = d_out->data(); + const T* x_data = x->data(); + Transform trans; + trans(context.device_context(), d_out_data, d_out_data + numel, x_data, + d_x_data, ClipGradFunctor(min, max)); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/concat_op.cc b/paddle/operators/concat_op.cc index 223bb0ffe6..1ffa02c8f9 100644 --- a/paddle/operators/concat_op.cc +++ b/paddle/operators/concat_op.cc @@ -24,31 +24,32 @@ class ConcatOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of ConcatOp should not be null."); + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE_GE(ctx->Inputs("X").size(), 1UL, + "Inputs(X) of ConcatOp should be empty.") + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ConcatOp should not be null."); - auto ins = ctx.MultiInput("X"); - auto *out = ctx.Output("Out"); - size_t axis = static_cast(ctx.Attr("axis")); - size_t n = ins.size(); + auto ins = ctx->GetInputsDim("X"); + size_t axis = static_cast(ctx->Attrs().Get("axis")); + const size_t n = ins.size(); PADDLE_ENFORCE_GT(n, 1, "Input tensors count should > 1."); - auto out_dims = ins[0]->dims(); + auto out_dims = ins[0]; size_t in_zero_dims_size = out_dims.size(); for (size_t i = 1; i < n; i++) { for (size_t j = 0; j < in_zero_dims_size; j++) { if (j == axis) { - out_dims[axis] += ins[i]->dims()[j]; + out_dims[axis] += ins[i][j]; continue; } - PADDLE_ENFORCE_EQ(out_dims[j], ins[i]->dims()[j], + PADDLE_ENFORCE_EQ(out_dims[j], ins[i][j], "Input tensors should have the same " "elements except the specify axis.") } } - out->Resize(out_dims); + ctx->SetOutputDim("Out", out_dims); } }; @@ -73,10 +74,27 @@ class ConcatOpMaker : public framework::OpProtoAndCheckerMaker { } }; +class ConcatOpGrad : public framework::OperatorWithKernel { + public: + ConcatOpGrad(const std::string &type, + const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + protected: + void InferShape(framework::InferShapeContextBase *ctx) const override { + ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X")); + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_WITHOUT_GRADIENT(concat, ops::ConcatOp, ops::ConcatOpMaker) +REGISTER_OP(concat, ops::ConcatOp, ops::ConcatOpMaker, concat_grad, + ops::ConcatOpGrad) REGISTER_OP_CPU_KERNEL(concat, ops::ConcatKernel) +REGISTER_OP_CPU_KERNEL(concat_grad, + ops::ConcatGradKernel) diff --git a/paddle/operators/concat_op.cu b/paddle/operators/concat_op.cu new file mode 100644 index 0000000000..ede832ddcd --- /dev/null +++ b/paddle/operators/concat_op.cu @@ -0,0 +1,20 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/concat_op.h" +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(concat, + ops::ConcatKernel); +REGISTER_OP_GPU_KERNEL( + concat_grad, ops::ConcatGradKernel); diff --git a/paddle/operators/concat_op.h b/paddle/operators/concat_op.h index f977054fdf..c113f19fb5 100644 --- a/paddle/operators/concat_op.h +++ b/paddle/operators/concat_op.h @@ -16,46 +16,51 @@ limitations under the License. */ #include #include "paddle/framework/op_registry.h" +#include "paddle/operators/strided_memcpy.h" namespace paddle { namespace operators { template -class ConcatKernel : public framework::OpKernel { +class ConcatKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto ins = ctx.MultiInput("X"); auto* out = ctx.Output("Out"); int64_t axis = static_cast(ctx.Attr("axis")); - size_t n = ins.size(); - size_t output_axis_dim = 0; - size_t before = 1, after = 1; - for (size_t i = 0; i < n; i++) { - output_axis_dim += ins[i]->dims()[axis]; - } - auto& input_zero = ins[0]; - for (int64_t i = 0; i < input_zero->dims().size(); i++) { - if (i == axis) { - continue; - } - if (i < axis) { - before *= input_zero->dims()[i]; - } else { - after *= input_zero->dims()[i]; - } - } + const size_t n = ins.size(); size_t output_offset = 0; + out->mutable_data(ctx.GetPlace()); + auto out_stride = framework::stride(out->dims()); for (size_t i = 0; i < n; i++) { auto& in = ins[i]; auto axis_dim = in->dims()[axis]; - for (size_t j = 0; j < before; j++) { - size_t len = axis_dim * after * sizeof(T); - const T* src = in->data() + axis_dim * after * j; - T* out_data = out->mutable_data(platform::CPUPlace()); - T* dest = out_data + output_offset + output_axis_dim * after * j; - memcpy(dest, src, len); - } - output_offset += axis_dim * after; + auto in_stride = framework::stride(in->dims()); + StridedMemcpy(ctx.device_context(), in->data(), in_stride, + in->dims(), out_stride, out->data() + output_offset); + output_offset += axis_dim * in_stride[axis]; + } + } +}; + +template +class ConcatGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const { + auto* in = ctx.Input(framework::GradVarName("Out")); + auto outs = ctx.MultiOutput(framework::GradVarName("X")); + int64_t axis = static_cast(ctx.Attr("axis")); + const size_t n = outs.size(); + size_t input_offset = 0; + auto in_stride = framework::stride(in->dims()); + for (size_t i = 0; i < n; i++) { + auto& out = outs[i]; + out->mutable_data(ctx.GetPlace()); + size_t axis_dim = out->dims()[axis]; + auto out_stride = framework::stride(out->dims()); + StridedMemcpy(ctx.device_context(), in->data() + input_offset, + in_stride, out->dims(), out_stride, out->data()); + input_offset += axis_dim * in_stride[axis]; } } }; diff --git a/paddle/operators/cond_op.cc b/paddle/operators/cond_op.cc index 8262a7a5c8..2737104a20 100644 --- a/paddle/operators/cond_op.cc +++ b/paddle/operators/cond_op.cc @@ -14,12 +14,7 @@ limitations under the License. */ #include "paddle/operators/cond_op.h" -#include -#include - -#include "paddle/framework/op_registry.h" #include "paddle/operators/gather.h" -#include "paddle/operators/net_op.h" #include "paddle/operators/scatter.h" namespace paddle { @@ -31,175 +26,183 @@ using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; using DDim = framework::DDim; -void CondOp::CreateScope(const Scope& scope) const { +framework::Scope& CondOp::AddSubScope(const Scope& scope) const { auto sub_scopes_var = scope.FindVar("SubScopes"); PADDLE_ENFORCE_NOT_NULL(sub_scopes_var, "Output(SubScopes) of CondOp should not be null."); auto sub_scopes = sub_scopes_var->GetMutable>(); auto& sub_scope = scope.NewScope(); sub_scopes->push_back(&sub_scope); + return sub_scope; } -void CondOp::CreateIndexTensor(const Scope& scope) const { +std::vector& CondOp::GetSubScopes( + const framework::Scope& scope) const { + auto sub_scopes_var = scope.FindVar("SubScopes"); + PADDLE_ENFORCE_NOT_NULL(sub_scopes_var, + "Output(SubScopes) of CondOp should not be null."); + return *sub_scopes_var->GetMutable>(); +} + +LoDTensor& CondOp::AddIndexTensor(const Scope& scope) const { auto index_tensors_var = scope.FindVar("IndexTensors"); PADDLE_ENFORCE_NOT_NULL(index_tensors_var, "Output(IndexTensors) of CondOp should not be null."); auto& index_tensors = *index_tensors_var->GetMutable>(); index_tensors.push_back(LoDTensor()); + return index_tensors.back(); } -void CondOp::InferShape(const Scope& scope) const { - auto sub_scopes_var = scope.FindVar("SubScopes"); - PADDLE_ENFORCE_NOT_NULL(sub_scopes_var, - "Output(SubScopes) of CondOp should not be null."); - auto& sub_scopes = *sub_scopes_var->GetMutable>(); - - for (int i = 0; i < 2; ++i) { - // Create two sub scopes for true and false branches - // sub_scopes[0] for the true branch and sub_scopes[1] for the false - // branch - CreateScope(scope); - - // Create two tensors for true and false indices - // index_tensors[0] for the true branch and index_tensors[1] for the false - // branch - CreateIndexTensor(scope); - - PADDLE_ENFORCE(!Inputs("Xs").empty(), - "Inputs(Xs) of CondOp can't be empty."); - for (auto& input : Inputs("Xs")) { - // Create a new tensor in sub-scope for input-type tensor - Variable* v = sub_scopes[i]->NewVar(input); - LoDTensor* sub_input = v->GetMutable(); - sub_input->Resize(scope.FindVar(input)->GetMutable()->dims()); - } - - for (auto& output : (*sub_net_op_[i]).Outputs()) { - for (auto& var_name : output.second) { - sub_scopes[i]->NewVar(var_name); - } - } - - // each net calls InferShape - sub_net_op_[i]->InferShape(*sub_scopes[i]); - } - - for (auto& output : Outputs("Outs")) { - LoDTensor* tensor_t_out = - sub_scopes[0]->FindVar(output)->GetMutable(); - PADDLE_ENFORCE_NOT_NULL(tensor_t_out, "True output should not be NULL"); - LoDTensor* tensor_f_out = - sub_scopes[1]->FindVar(output)->GetMutable(); - PADDLE_ENFORCE_NOT_NULL(tensor_f_out, "False output should not be NULL"); - - auto* tensor_out_var = scope.FindVar(output); - PADDLE_ENFORCE_NOT_NULL(tensor_out_var, "Output not found"); - LoDTensor* tensor_out = tensor_out_var->GetMutable(); - PADDLE_ENFORCE_NOT_NULL(tensor_t_out, - "True output tensor should not be NULL"); - - // check output size should be same - PADDLE_ENFORCE_EQ(tensor_t_out->dims(), tensor_f_out->dims(), - "Outputs not of the same shape"); - tensor_out->Resize(tensor_t_out->dims()); - // tensor_out->mutable_data(tensor_out->dims(), - // platform::CPUPlace()); - tensor_out->mutable_data(platform::CPUPlace()); - } -} - -void CondOp::Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const { - auto* sub_scopes_var = scope.FindVar("SubScopes"); - PADDLE_ENFORCE_NOT_NULL(sub_scopes_var, - "Output(SubScopes) of CondOp should not be null."); - auto sub_scopes = sub_scopes_var->Get>(); +std::vector& CondOp::GetIndexTensors( + const framework::Scope& scope) const { auto* index_tensors_var = scope.FindVar("IndexTensors"); PADDLE_ENFORCE_NOT_NULL(index_tensors_var, "Output(IndexTensors) of CondOp should not be null."); - auto index_tensors = index_tensors_var->Get>(); + return *index_tensors_var->GetMutable>(); +} - std::string cond_name = Input("Cond"); - Variable* cond_var = scope.FindVar(cond_name); +void CondOp::PrepareDataForSubnet( + const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const { + PADDLE_ENFORCE(!Inputs("Xs").empty(), "Inputs(Xs) of CondOp can't be empty."); + + for (int i = 0; i < BRANCH_NUM; ++i) { + // Create two sub scopes for true and false branches + // sub_scopes[0] for the true branch + // sub_scopes[1] for the false branch + AddSubScope(scope); + // Create two tensors for true and false indices: + // index_tensors[0] for the true branch + // index_tensors[1] for the false branch + AddIndexTensor(scope); + } + + Variable* cond_var = scope.FindVar(Input("Cond")); PADDLE_ENFORCE_NOT_NULL(cond_var, "Input(Cond) of CondOp should not be null."); const LoDTensor* cond = cond_var->GetMutable(); - // Step 1: get the true/false index at runtime - // index_[0]: vector, contains all index for cond[i] == true - // index_[1]: vector, contains all index for cond[i] == false - for (int i = 0; i < 2; ++i) index_[i].clear(); + // get the true/false index at runtime according to cond tensor + // index_vectors[0]: vector, contains all index for cond[i] == true + // index_vectors[1]: vector, contains all index for cond[i] == false + std::vector> index_vectors; + index_vectors.resize(BRANCH_NUM); const int* cond_data = cond->data(); for (int i = 0; i < cond->dims()[0]; ++i) { if (cond_data[i]) - index_[0].push_back(i); + index_vectors[TRUE_BRANCH].push_back(i); else - index_[1].push_back(i); + index_vectors[FALSE_BRANCH].push_back(i); } - // put index_[0] and index_[1] into two tensors: - // index_tensor_[0] and index_tensor_[1] - DDim dim = paddle::framework::make_ddim({0}); - for (int i = 0; i < 2; ++i) { - dim[0] = index_[i].size(); - int* tmp_ptr = + // put index_vectors[0] and index_vectors[1] into two tensors: + // index_tensors[0] and index_tensors[1] + std::vector& index_tensors = GetIndexTensors(scope); + std::vector& sub_scopes = GetSubScopes(scope); + + for (int i = 0; i < BRANCH_NUM; ++i) { + DDim dim = {static_cast(index_vectors[i].size())}; + int* index_tensor_data_ptr = index_tensors[i].mutable_data(dim, platform::CPUPlace()); - index_tensors[i].Resize(dim); - memcpy(tmp_ptr, index_[i].data(), dim[0] * sizeof(int)); + memcpy(index_tensor_data_ptr, index_vectors[i].data(), + dim[0] * sizeof(int)); } - // Step 2: collect data by calling gather - for (int i = 0; i < 2; ++i) { - // i= 0/i for True and False branches respectively - for (auto& input : Inputs("Xs")) { - // find Tensor - Variable* v = scope.FindVar(input); - PADDLE_ENFORCE_NOT_NULL(v); - LoDTensor* tensor_parent = v->GetMutable(); + // create input in subscopes according to index_vectors + for (auto& input : Inputs("Xs")) { + Variable* var_parent = scope.FindVar(input); + PADDLE_ENFORCE_NOT_NULL(var_parent); + const auto* tensor_parent = &var_parent->Get(); - v = sub_scopes[i]->FindVar(input); - PADDLE_ENFORCE_NOT_NULL(v); - LoDTensor* tensor_child = v->GetMutable(); + for (int i = 0; i < BRANCH_NUM; ++i) { + Variable* var_child = sub_scopes[i]->FindVar(input); + PADDLE_ENFORCE_NOT_NULL(var_child); + auto* tensor_child = var_child->GetMutable(); // Resize child - DDim dim = tensor_child->dims(); - dim[0] = index_[i].size(); - tensor_child->Resize(dim); + DDim dim = tensor_parent->dims(); + dim[0] = index_tensors[i].dims()[0]; tensor_child->mutable_data(dim, platform::CPUPlace()); - Gather(dev_ctx.GetPlace(), tensor_parent, &index_tensors[i], - tensor_child); + CPUGather(dev_ctx, *tensor_parent, index_tensors[i], tensor_child); } } - // Step 3: run - for (int i = 0; i < 2; ++i) { - sub_net_op_[i]->Run(*sub_scopes[i], dev_ctx); + // create output_tensors in subscope for sub_net + for (int i = 0; i < BRANCH_NUM; ++i) { + for (auto& output : (*sub_net_op_[i]).Outputs()) { + for (auto& var_name : output.second) { + sub_scopes[i]->NewVar(var_name); + } + } } +} - // Step 4: merge output results +void CondOp::MergeDataFromSubnet(const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const { + std::vector& sub_scopes = GetSubScopes(scope); + const std::vector& index_tensors = + GetIndexTensors(scope); + + // Infer the output dim, out_dim[0] = true_dim[0] + false_dim[0] PADDLE_ENFORCE(!Outputs("Outs").empty(), "Outputs(Outs) of CondOp can't be empty."); - for (int i = 0; i < 2; ++i) { - // i= 0/i for True and False branches respectively - for (auto& output : Outputs("Outs")) { - // find Tensor - Variable* v = scope.FindVar(output); - PADDLE_ENFORCE_NOT_NULL(v); - LoDTensor* tensor_parent = v->GetMutable(); - - v = sub_scopes[i]->FindVar(output); - PADDLE_ENFORCE_NOT_NULL(v); - LoDTensor* tensor_child = v->GetMutable(); - - ScatterUpdate(dev_ctx.GetPlace(), tensor_child, &index_tensors[i], + for (auto& output : Outputs("Outs")) { + const LoDTensor* tensor_t_out = + &sub_scopes[TRUE_BRANCH]->FindVar(output)->Get(); + PADDLE_ENFORCE_NOT_NULL(tensor_t_out, "True output should not be NULL"); + const LoDTensor* tensor_f_out = + &sub_scopes[FALSE_BRANCH]->FindVar(output)->Get(); + PADDLE_ENFORCE_NOT_NULL(tensor_f_out, "False output should not be NULL"); + + auto* var_out = scope.FindVar(output); + PADDLE_ENFORCE_NOT_NULL(var_out, "Output not found"); + LoDTensor* tensor_out = var_out->GetMutable(); + PADDLE_ENFORCE_NOT_NULL(tensor_t_out, + "True output tensor should not be NULL"); + + DDim true_dim = tensor_t_out->dims(); + DDim false_dim = tensor_f_out->dims(); + true_dim[0] = 0; + false_dim[0] = 0; + PADDLE_ENFORCE_EQ(true_dim, false_dim, + "Outputs not of the same shape except the first dim"); + + DDim out_dim = tensor_t_out->dims(); + out_dim[0] = tensor_t_out->dims()[0] + tensor_f_out->dims()[0]; + tensor_out->Resize(out_dim); + tensor_out->mutable_data(platform::CPUPlace()); + } + + // merge output results: + // output_tensor = true_output_tensor + false_output_tensor + for (auto& output : Outputs("Outs")) { + Variable* var_parent = scope.FindVar(output); + PADDLE_ENFORCE_NOT_NULL(var_parent); + auto* tensor_parent = var_parent->GetMutable(); + + for (int i = 0; i < BRANCH_NUM; ++i) { + Variable* var_child = sub_scopes[i]->FindVar(output); + PADDLE_ENFORCE_NOT_NULL(var_child); + auto* tensor_child = &var_child->Get(); + ScatterAssign(dev_ctx, *tensor_child, index_tensors[i], tensor_parent); } } } +void CondOp::Run(const Scope& scope, + const platform::DeviceContext& dev_ctx) const { + PrepareDataForSubnet(scope, dev_ctx); + std::vector& sub_scopes = GetSubScopes(scope); + for (int i = 0; i < BRANCH_NUM; ++i) { + sub_net_op_[i]->Run(*sub_scopes[i], dev_ctx); + } + MergeDataFromSubnet(scope, dev_ctx); +} + class CondOpProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker { public: CondOpProtoAndCheckerMaker(framework::OpProto* proto, @@ -215,7 +218,7 @@ class CondOpProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( Sample dependent Cond Operator: Given Cond[i] as a 1/0 vector to indicate true/false -The equation is: +The equation is: Out[i] = subnet_t[i], if Cond[i] == true Out[i] = subnet_t[i], if Cond[i] == false )DOC"); diff --git a/paddle/operators/cond_op.h b/paddle/operators/cond_op.h index b09e32331e..93121fb31b 100644 --- a/paddle/operators/cond_op.h +++ b/paddle/operators/cond_op.h @@ -40,8 +40,7 @@ class CondOp : public framework::OperatorBase { const framework::VariableNameMap& outputs, const framework::AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs) { - index_.resize(2); - sub_net_op_.resize(2); + sub_net_op_.resize(BRANCH_NUM); } CondOp(const CondOp& o) @@ -51,40 +50,44 @@ class CondOp : public framework::OperatorBase { PADDLE_THROW("Not implemented"); } - void CreateScope(const framework::Scope& scope) const; + framework::Scope& AddSubScope(const framework::Scope& scope) const; + std::vector& GetSubScopes( + const framework::Scope& scope) const; - void CreateIndexTensor(const framework::Scope& scope) const; + framework::LoDTensor& AddIndexTensor(const framework::Scope& scope) const; + std::vector& GetIndexTensors( + const framework::Scope& scope) const; - /* - * InferShape must be called before Run. - */ - void InferShape(const framework::Scope& scope) const override; + void PrepareDataForSubnet(const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const; + void MergeDataFromSubnet(const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const; /* * Set True Block */ void set_truenet(std::unique_ptr&& net) { - sub_net_op_[0] = std::move(net); + sub_net_op_[TRUE_BRANCH] = std::move(net); } /* * Set False Block */ void set_falsenet(std::unique_ptr&& net) { - sub_net_op_[1] = std::move(net); + sub_net_op_[FALSE_BRANCH] = std::move(net); } void Run(const framework::Scope& scope, const platform::DeviceContext& dev_ctx) const override; private: + const int TRUE_BRANCH = 0; + const int FALSE_BRANCH = 1; + const int BRANCH_NUM = 2; + // sub_net_op_[0]: subnet_t // sub_net_op_[1]: subnet_f std::vector> sub_net_op_; - - // index_[0]: True_index; - // index_[1]: False_index; - mutable std::vector> index_; }; } // namespace operators diff --git a/paddle/operators/conv2d_op.cc b/paddle/operators/conv2d_op.cc new file mode 100644 index 0000000000..5cc82944bb --- /dev/null +++ b/paddle/operators/conv2d_op.cc @@ -0,0 +1,131 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/gemm_conv2d_op.h" + +namespace paddle { +namespace operators { + +int outputSize(int input_size, int filter_size, int padding, int stride) { + int output_size = (input_size - filter_size + 2 * padding) / stride + 1; + return output_size; +} + +class Conv2DOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input(Input) of Conv2DOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Filter"), + "Input(Filter) of Conv2DOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Output"), + "Output(Output) of Conv2DOp should not be null."); + + auto in_dims = ctx->GetInputDim("Input"); + auto filter_dims = ctx->GetInputDim("Filter"); + std::vector strides = ctx->Attrs().Get>("strides"); + std::vector paddings = ctx->Attrs().Get>("paddings"); + int groups = ctx->Attrs().Get("groups"); + int input_channels = in_dims[1]; + int output_channels = filter_dims[0]; + + PADDLE_ENFORCE_EQ(in_dims.size(), 4, "Conv2DOp input should be 4-D."); + PADDLE_ENFORCE_EQ(filter_dims.size(), 4, "Conv2DOp filter should be 4-D."); + PADDLE_ENFORCE_EQ(input_channels, filter_dims[1] * groups, + "The number of input channels should be equal to filter " + "channels * groups."); + PADDLE_ENFORCE_EQ( + output_channels % groups, 0, + "The number of output channels should be divided by groups."); + + auto output_height = + outputSize(in_dims[2], filter_dims[2], paddings[0], strides[0]); + auto output_width = + outputSize(in_dims[3], filter_dims[3], paddings[1], strides[1]); + ctx->SetOutputDim( + "Output", {in_dims[0], filter_dims[0], output_height, output_width}); + } +}; + +class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker { + public: + Conv2DOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput( + "Input", + "The input tensor of convolution operator. " + "The format of input tensor is NCHW. Where N is batch size, C is the " + "number of channels, H and W is the height and width of image."); + AddInput( + "Filter", + "The filter tensor of convolution operator." + "The format of the filter tensor is MCHW, where M is the number of " + "output image channels, C is the number of input image channels, " + "H and W is height and width of filter. " + "If the groups attribute is greater than 1, C equal the number of " + "input image channels divided by the groups."); + AddOutput("Output", + "The output tensor of convolution operator." + "The format of output tensor is also NCHW."); + AddAttr>("strides", "strides of convolution operator.") + .SetDefault({1, 1}); + AddAttr>("paddings", "paddings of convolution operator.") + .SetDefault({0, 0}); + AddAttr( + "groups", + "group size of convolution operator. " + "Refer to grouped convolution in Alex Krizhevsky's paper: " + "when group=2, the first half of the filters are only connected to the " + "first half of the input channels, and the second half only connected " + "to the second half.") + .SetDefault(1); + AddComment(R"DOC( +The convolution operation calculates the output based on the input, filter +and strides, paddings, groups parameters. The size of each dimension of the +parameters is checked in the infer-shape. +)DOC"); + } +}; + +class Conv2DOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + auto in_dims = ctx->GetInputDim("Input"); + auto filter_dims = ctx->GetInputDim("Filter"); + if (ctx->HasOutput(framework::GradVarName("Input"))) { + ctx->SetOutputDim(framework::GradVarName("Input"), in_dims); + } + if (ctx->HasOutput(framework::GradVarName("Filter"))) { + ctx->SetOutputDim(framework::GradVarName("Filter"), filter_dims); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(conv2d, ops::Conv2DOp, ops::Conv2DOpMaker, conv2d_grad, + ops::Conv2DOpGrad); + +REGISTER_OP_CPU_KERNEL( + conv2d, ops::GemmConv2DKernel); +REGISTER_OP_CPU_KERNEL( + conv2d_grad, ops::GemmConvGrad2DKernel); diff --git a/paddle/operators/conv2d_op.cu b/paddle/operators/conv2d_op.cu new file mode 100644 index 0000000000..5df818ba04 --- /dev/null +++ b/paddle/operators/conv2d_op.cu @@ -0,0 +1,22 @@ +/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/gemm_conv2d_op.h" + +namespace ops = paddle::operators; + +REGISTER_OP_GPU_KERNEL( + conv2d, ops::GemmConv2DKernel); +REGISTER_OP_GPU_KERNEL( + conv2d_grad, ops::GemmConvGrad2DKernel); diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc index 72c4464936..040546f1a6 100644 --- a/paddle/operators/cos_sim_op.cc +++ b/paddle/operators/cos_sim_op.cc @@ -24,22 +24,22 @@ class CosSimOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { + void InferShape(framework::InferShapeContextBase* ctx) const override { // notnull check - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of CosSimOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), - "Input(Y) of CosSimOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of CosSimOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("XNorm"), - "Output(XNorm) of CosSimOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("YNorm"), - "Output(YNorm) of CosSimOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of CosSimOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), + "Input(Y) of CosSimOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of CosSimOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("XNorm"), + "Output(XNorm) of CosSimOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("YNorm"), + "Output(YNorm) of CosSimOp should not be null."); // shape check - auto x_dims = ctx.Input("X")->dims(); - auto y_dims = ctx.Input("Y")->dims(); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(), "Ranks of Input(X) and Input(Y) must be equal."); @@ -54,15 +54,16 @@ class CosSimOp : public framework::OperatorWithKernel { " just 1 (which will be broadcasted to match Input(X))."); // resize tensor - ctx.Output("Out")->Resize({x_dims[0], 1}); - ctx.Output("XNorm")->Resize({x_dims[0], 1}); - ctx.Output("YNorm")->Resize({y_dims[0], 1}); + ctx->SetOutputDim("Out", {x_dims[0], 1}); + ctx->SetOutputDim("XNorm", {x_dims[0], 1}); + ctx->SetOutputDim("YNorm", {y_dims[0], 1}); + ctx->ShareLoD("X", /*->*/ "Out"); } }; class CosSimOpMaker : public framework::OpProtoAndCheckerMaker { public: - CosSimOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + CosSimOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The 1st input of cos_sim op."); AddInput("Y", "The 2nd input of cos_sim op."); @@ -81,10 +82,13 @@ Cosine Similarity Operator. The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y)). -Input(X) and Input(Y) must have the same shape, except that the 1st dimension -of Input(Y) could be just 1 (different from Input(X)), which will be -broadcasted to match the shape of Input(X) before computing their cosine +The input `X` and `Y` must have the same shape, except that the 1st dimension +of input `Y` could be just 1 (different from input `X`), which will be +broadcasted to match the shape of input `X` before computing their cosine similarity. + +Both the input `X` and `Y` can carry the LoD (Level of Details) information, +or not. But the output only shares the LoD with input `X`. )DOC"); } }; @@ -94,27 +98,23 @@ class CosSimOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { + void InferShape(framework::InferShapeContextBase* ctx) const override { // notnull check - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("XNorm"), - "Input(XNorm) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("YNorm"), - "Input(YNorm) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Out"), - "Input(Out) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("XNorm"), "Input(XNorm) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("YNorm"), "Input(YNorm) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("Out"), "Input(Out) must not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) must not be null."); // shape check - auto x_dims = ctx.Input("X")->dims(); - auto y_dims = ctx.Input("Y")->dims(); - auto xnorm_dims = ctx.Input("XNorm")->dims(); - auto ynorm_dims = ctx.Input("YNorm")->dims(); - auto out_dims = ctx.Input("Out")->dims(); - auto out_grad_dims = - ctx.Input(framework::GradVarName("Out"))->dims(); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + auto xnorm_dims = ctx->GetInputDim("XNorm"); + auto ynorm_dims = ctx->GetInputDim("YNorm"); + auto out_dims = ctx->GetInputDim("Out"); + auto out_grad_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), "Ranks of Input(X) and Input(Y) must be equal."); @@ -139,12 +139,14 @@ class CosSimOpGrad : public framework::OperatorWithKernel { "Shape of Input(Out@Grad) must be [X.Dim(0), 1]."); // resize tensor - auto *x_grad = - ctx.Output(framework::GradVarName("X")); - auto *y_grad = - ctx.Output(framework::GradVarName("Y")); - if (x_grad) x_grad->Resize(x_dims); - if (y_grad) y_grad->Resize(y_dims); + auto x_grad_name = framework::GradVarName("X"); + auto y_grad_name = framework::GradVarName("Y"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); + } + if (ctx->HasOutput(y_grad_name)) { + ctx->SetOutputDim(y_grad_name, y_dims); + } } }; diff --git a/paddle/operators/cos_sim_op.h b/paddle/operators/cos_sim_op.h index bcf6f758ca..68c56f531f 100644 --- a/paddle/operators/cos_sim_op.h +++ b/paddle/operators/cos_sim_op.h @@ -28,7 +28,7 @@ template ; template -class CosSimKernel : public framework::OpKernel { +class CosSimKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { // get Tensor @@ -67,7 +67,7 @@ class CosSimKernel : public framework::OpKernel { }; template -class CosSimGradKernel : public framework::OpKernel { +class CosSimGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { // get Tensor diff --git a/paddle/operators/crop_op.cc b/paddle/operators/crop_op.cc new file mode 100644 index 0000000000..9b2305e90e --- /dev/null +++ b/paddle/operators/crop_op.cc @@ -0,0 +1,137 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/crop_op.h" +#include + +namespace paddle { +namespace operators { + +using framework::Tensor; + +class CropOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of CropOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of CropOp should not be null."); + auto x_dim = ctx->GetInputDim("X"); + if (!ctx->HasInput("Y")) { + auto shape = ctx->Attrs().Get>("shape"); + PADDLE_ENFORCE_EQ( + int64_t(shape.size()), x_dim.size(), + "Shape size should be equal to dimention size of input tensor."); + std::vector tensor_shape(shape.size()); + for (size_t i = 0; i < shape.size(); ++i) { + tensor_shape[i] = static_cast(shape[i]); + } + ctx->SetOutputDim("Out", framework::make_ddim(tensor_shape)); + } else { + auto y_dim = ctx->GetInputDim("Y"); + PADDLE_ENFORCE_EQ(framework::arity(x_dim), framework::arity(y_dim), + "Tensor rank of both CropOp's " + "inputs must be same."); + ctx->SetOutputDim("Out", y_dim); + } + } +}; + +class CropOpMaker : public framework::OpProtoAndCheckerMaker { + public: + CropOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "The input of pad op. " + "The input should be a k-D tensor(k > 0 and k < 7)"); + AddInput("Y", + "The input used as reference for cropping" + " with the same dimension as X. "); + AddOutput("Out", + "The output of crop op " + "with the same dimension as X."); + AddAttr>("offsets", + "A list describing offsets to be cropped." + "The size of offsets list should be as same as " + "dimension size of input X."); + AddAttr>("shape", + "A list describing the shape of output." + "The size of shape list should be as same as " + "dimension size of input X.") + .SetDefault(std::vector()); + AddComment(R"DOC( +Crop Operator. +Crop input into output, as specified by offsets and shape. + +There are two ways to set shape: +1. referenc input: crop input X as shape as reference input. + The dimension of reference input should + be as same as input X. +2. shape list: crop input X by shape described by a list. + The size of shape list should be as same as + dimension size of input X. + +The input should be a k-D tensor(k > 0 and k < 7). As an example: + +Given: + + X = [[0, 1, 2, 0, 0] + [0, 3, 4, 0, 0] + [0, 0, 0, 0, 0]] + +and + + offsets = [0, 1] + +and + + shape = [2, 2] + +then we get + + Out = [[1, 2], + [3, 4]] + +)DOC"); + } +}; + +class CropOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto x_dims = ctx->GetInputDim("X"); + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(crop, ops::CropOp, ops::CropOpMaker, crop_grad, ops::CropOpGrad); +REGISTER_OP_CPU_KERNEL(crop, ops::CropKernel); +REGISTER_OP_CPU_KERNEL(crop_grad, + ops::CropGradKernel); diff --git a/paddle/operators/crop_op.cu b/paddle/operators/crop_op.cu new file mode 100644 index 0000000000..f8ee18a1d6 --- /dev/null +++ b/paddle/operators/crop_op.cu @@ -0,0 +1,21 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/crop_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(crop, ops::CropKernel); +REGISTER_OP_GPU_KERNEL(crop_grad, + ops::CropGradKernel); diff --git a/paddle/operators/crop_op.h b/paddle/operators/crop_op.h new file mode 100644 index 0000000000..2e72583d68 --- /dev/null +++ b/paddle/operators/crop_op.h @@ -0,0 +1,104 @@ +/* Copyright (c) 2016 CropdleCropdle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/strided_memcpy.h" + +namespace paddle { +namespace operators { // Internal + +template +using EigenTensor = framework::EigenTensor; +using framework::Tensor; + +template +class CropKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* x = context.Input("X"); + auto* out = context.Output("Out"); + const T* x_data = x->data(); + T* out_data = out->mutable_data(context.GetPlace()); + auto x_stride = framework::stride(x->dims()); + auto out_stride = framework::stride(out->dims()); + auto offsets = context.Attr>("offsets"); + PADDLE_ENFORCE_EQ( + x->dims().size(), static_cast(offsets.size()), + "Offsets size should be equal to dimension size of input tensor."); + int64_t offset = 0; + for (size_t i = 0; i < offsets.size(); ++i) { + offset += (x_stride[i] * offsets[i]); + } + StridedMemcpy(context.device_context(), x_data + offset, x_stride, + out->dims(), out_stride, out_data); + } +}; + +template +void CropGradFunction(const framework::ExecutionContext& context) { + auto* d_x = context.Output(framework::GradVarName("X")); + if (d_x != nullptr) { + auto* d_out = context.Input(framework::GradVarName("Out")); + d_x->mutable_data(context.GetPlace()); + auto offsets = context.Attr>("offsets"); + Eigen::array, D> paddings; + for (size_t i = 0; i < D; ++i) { + paddings[i].first = offsets[i]; + paddings[i].second = d_x->dims()[i] - d_out->dims()[i] - offsets[i]; + } + auto d_x_tensor = EigenTensor::From(*d_x); + auto d_out_tensor = EigenTensor::From(*d_out); + d_x_tensor.device(context.GetEigenDevice()) = + d_out_tensor.pad(paddings, 0); + } +} + +template +class CropGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + size_t rank = + context.Input(framework::GradVarName("Out"))->dims().size(); + switch (rank) { + case 1: + CropGradFunction(context); + break; + case 2: + CropGradFunction(context); + break; + case 3: + CropGradFunction(context); + break; + case 4: + CropGradFunction(context); + break; + case 5: + CropGradFunction(context); + break; + case 6: + CropGradFunction(context); + break; + default: + PADDLE_THROW( + "CropOp only support tensors with no more than 6 dimensions."); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/cross_entropy_op.cc b/paddle/operators/cross_entropy_op.cc index 953367eb8b..4b67887f36 100644 --- a/paddle/operators/cross_entropy_op.cc +++ b/paddle/operators/cross_entropy_op.cc @@ -17,41 +17,41 @@ limitations under the License. */ namespace paddle { namespace operators { -using framework::LoDTensor; - class CrossEntropyOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"), - "Input(Label) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"), "Output(Y) must not be null."); - - auto x = ctx.Input("X"); - auto label = ctx.Input("Label"); - PADDLE_ENFORCE_EQ(x->dims().size(), 2, "Input(X)'s rank must be 2."); - PADDLE_ENFORCE_EQ(label->dims().size(), 2, - "Input(Label)'s rank must be 2."); - // TODO(xinghai-sun): remove this check after swtiching to bool - PADDLE_ENFORCE(ctx.Attr("soft_label") == 0 || - ctx.Attr("soft_label") == 1); - PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0], - "The 1st dimension of Input(X) and Input(Label) must " + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null."); + + auto x_dims = ctx->GetInputDim("X"); + auto label_dims = ctx->GetInputDim("Label"); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); + PADDLE_ENFORCE_EQ(label_dims.size(), 2, "Input(Label)'s rank should be 2."); + PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0], + "The 1st dimension of Input(X) and Input(Label) should " "be equal."); - if (ctx.Attr("soft_label") == 1) { - PADDLE_ENFORCE_EQ(x->dims()[1], label->dims()[1], - "If Attr(soft_label) == 1, The 2nd dimension of " - "Input(X) and Input(Label) must be equal."); + if (ctx->Attrs().Get("softLabel")) { + PADDLE_ENFORCE_EQ(x_dims[1], label_dims[1], + "If Attr(softLabel) == true, the 2nd dimension of " + "Input(X) and Input(Label) should be equal."); } else { - PADDLE_ENFORCE_EQ(label->dims()[1], 1, - "If Attr(soft_label) == 0, The 2nd dimension of " - "Input(Label) must be 1."); + PADDLE_ENFORCE_EQ(label_dims[1], 1, + "If Attr(softLabel) == false, the 2nd dimension of " + "Input(Label) should be 1."); } - ctx.Output("Y")->Resize({x->dims()[0], 1}); + ctx->SetOutputDim("Y", {x_dims[0], 1}); + ctx->ShareLoD("X", /*->*/ "Y"); + } + + // CrossEntropy's data type just determined by "X" + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("X")->type()); } }; @@ -60,68 +60,85 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"), - "Input(Label) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Y")), - "Input(Y@GRAD) must not be null."); - - auto x = ctx.Input("X"); - auto label = ctx.Input("Label"); - auto dy = ctx.Input(framework::GradVarName("Y")); - PADDLE_ENFORCE_EQ(x->dims().size(), 2, "Input(X)'s rank must be 2."); - PADDLE_ENFORCE_EQ(dy->dims().size(), 2, "Input(Y@Grad)'s rank must be 2."); - PADDLE_ENFORCE_EQ(label->dims().size(), 2, - "Input(Label)'s rank must be 2."); - // TODO(xinghai-sun): remove this check after swtiching to bool - PADDLE_ENFORCE(ctx.Attr("soft_label") == 0 || - ctx.Attr("soft_label") == 1); - PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0], - "The 1st dimension of Input(X) and Input(Label) must " + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")), + "Input(Y@GRAD) shoudl be not null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Output(X@GRAD) should be not null."); + + auto x_dims = ctx->GetInputDim("X"); + auto label_dims = ctx->GetInputDim("Label"); + auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y")); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); + PADDLE_ENFORCE_EQ(dy_dims.size(), 2, "Input(Y@Grad)'s rank should be 2."); + PADDLE_ENFORCE_EQ(label_dims.size(), 2, "Input(Label)'s rank should be 2."); + PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0], + "The 1st dimension of Input(X) and Input(Label) should " "be equal."); - PADDLE_ENFORCE_EQ(x->dims()[0], dy->dims()[0], - "The 1st dimension of Input(X) and Input(Y@Grad) must " + PADDLE_ENFORCE_EQ(x_dims[0], dy_dims[0], + "The 1st dimension of Input(X) and Input(Y@Grad) should " "be equal."); - PADDLE_ENFORCE_EQ(dy->dims()[1], 1, - "The 2nd dimension of Input(Y@Grad) must be 1."); - if (ctx.Attr("soft_label") == 1) { - PADDLE_ENFORCE_EQ(x->dims()[1], label->dims()[1], - "If Attr(soft_label) == 1, The 2nd dimension of " - "Input(X) and Input(Label) must be equal."); + PADDLE_ENFORCE_EQ(dy_dims[1], 1, + "The 2nd dimension of Input(Y@Grad) should be 1."); + if (ctx->Attrs().Get("softLabel")) { + PADDLE_ENFORCE_EQ(x_dims[1], label_dims[1], + "When Attr(softLabel) == true, the 2nd dimension of " + "Input(X) and Input(Label) should be equal."); } else { - PADDLE_ENFORCE_EQ(label->dims()[1], 1, - "If Attr(soft_label) == 0, The 2nd dimension of " - "Input(Label) must be 1."); + PADDLE_ENFORCE_EQ(label_dims[1], 1, + "When Attr(softLabel) == false, the 2nd dimension of " + "Input(Label) should be 1."); } + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + } - auto dx = ctx.Output(framework::GradVarName("X")); - dx->Resize(x->dims()); + // CrossEntropy's data type just determined by "X" + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("X")->type()); } }; class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker { public: - CrossEntropyOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + CrossEntropyOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The first input of CrossEntropyOp"); - AddInput("Label", "The second input of CrossEntropyOp"); - AddOutput("Y", "The output of CrossEntropyOp"); - AddAttr("soft_label", "Is soft label. Default zero.").SetDefault(0); - + AddInput("X", + "(Tensor, default Tensor), a 2-D tensor with shape N x D, " + "where N is the batch size and D is the number of classes. " + "This input is a probability computed by the previous operator, " + "which is almost always the result of a softmax operator."); + AddInput( + "Label", + "(Tensor, default Tensor), the ground truth which is " + "a 2-D tensor. " + "When softLabel is set to false, `Label` is a Tensor with shape " + "[N x 1]. " + "When softLabel is set to true, `Label` is a Tensor " + "with shape [N x K]."); + AddOutput("Y", + "(Tensor, default Tensor), a 2-D tensor " + "with shape [N x 1]. The cross entropy loss."); + AddAttr( + "softLabel", + "(bool, default false), a flag to indicate whether to interpretate " + "the given labels as soft labels.") + .SetDefault(false); AddComment(R"DOC( CrossEntropy Operator. It supports both standard cross-entropy and soft-label cross-entropy loss computation. 1) One-hot cross-entropy: - soft_label = 0, Label[i, 0] indicates the class index for sample i: + softLabel = false, Label[i, 0] indicates the class index for sample i: Y[i] = -log(X[i, Label[i]]) 2) Soft-label cross-entropy: - soft_label = 1, Label[i, j] indicates the soft label of class j + softLabel = true, Label[i, j] indicates the soft label of class j for sample i: Y[i] = \sum_j{-Label[i, j] * log(X[i, j])} @@ -133,6 +150,9 @@ computation. As a special case of 2), when each row of Input(Label) has only one non-zero element (equals 1), soft-label cross-entropy degenerates to a one-hot cross-entropy with one-hot label representation. + +Both the input `X` and `Label` can carry the LoD (Level of Details) information, +or not. But the output only shares the LoD with input `X`. )DOC"); } }; diff --git a/paddle/operators/cross_entropy_op.cu b/paddle/operators/cross_entropy_op.cu index ab6ad0e062..5e2024e0ea 100644 --- a/paddle/operators/cross_entropy_op.cu +++ b/paddle/operators/cross_entropy_op.cu @@ -12,47 +12,12 @@ See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/framework/op_registry.h" #include "paddle/operators/cross_entropy_op.h" -#include "paddle/platform/assert.h" -#include "paddle/platform/hostdevice.h" namespace paddle { namespace operators { -template -__global__ void CrossEntropyKernel(T* Y, const T* X, const int* label, - const int N, const int D) { - // TOOD(qingqing) define CUDA_1D_KERNEL_LOOP macro in a common file. - // CUDA_1D_KERNEL_LOOP(i, N) { - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; - i += blockDim.x * gridDim.x) { - PADDLE_ASSERT(label[i] >= 0 && label[i] < D); - Y[i] = -tolerable_value(log(X[i * D + label[i]])); - } -} - -template -__global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, - const int N, const int D) { - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; - i += blockDim.x * gridDim.x) { - T sum = static_cast(0); - for (int j = 0; j < D; j++) { - sum += label[i * D + j] * tolerable_value(log(X[i * D + j])); - } - Y[i] = -sum; - } -} - -// TODO(qingqing): make zero setting an common function. -template -__global__ void zero(T* X, const int N) { - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; - i += blockDim.x * gridDim.x) { - X[i] = 0.0; - } -} +namespace { template __global__ void CrossEntropyGradientKernel(T* dX, const T* dY, const T* X, @@ -71,80 +36,69 @@ template __global__ void SoftCrossEntropyGradientKernel(T* dX, const T* dY, const T* X, const T* label, const int N, const int D) { - // TOOD(qingqing): optimize for this kernel - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; - i += blockDim.x * gridDim.x) { - for (int j = 0; j < D; ++j) { - int idx = i * D + j; - dX[idx] = -label[idx] * dY[i] / X[idx]; - } + int ids = blockIdx.x * blockDim.x + threadIdx.x; + if (ids < N * D) { + int row_ids = ids / D; + dX[ids] = -label[ids] * dY[row_ids] / X[ids]; } } +} // namespace template -class CrossEntropyOpCUDAKernel : public framework::OpKernel { +class CrossEntropyOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); - - auto x = ctx.Input("X"); - auto y = ctx.Output("Y"); - auto label = ctx.Input("Label"); - - auto* x_data = x->data(); + "This kernel only runs on GPU device."); + const Tensor* x = ctx.Input("X"); + const Tensor* label = ctx.Input("Label"); + Tensor* y = ctx.Output("Y"); y->mutable_data(ctx.GetPlace()); - auto* y_data = y->data(); - int n = x->dims()[0]; - int d = x->dims()[1]; - int block = 512; - int grid = (n + block - 1) / block; - // TODO(qingqing) launch kernel on specified stream - // base on ExecutionContext. - if (ctx.Attr("soft_label") == 1) { - auto* label_data = ctx.Input("Label")->data(); - SoftCrossEntropyKernel<<>>(y_data, x_data, label_data, n, - d); - } else { - auto* label_data = ctx.Input("Label")->data(); - CrossEntropyKernel<<>>(y_data, x_data, label_data, n, d); - } + math::CrossEntropyFunctor()( + ctx.device_context(), y, x, label, ctx.Attr("softLabel")); } }; template -class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel { +class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "It must use GPUPlace."); + "This kernel only runs on GPU device."); + + const Tensor* x = ctx.Input("X"); + const Tensor* label = ctx.Input("Label"); + Tensor* dx = ctx.Output(framework::GradVarName("X")); + dx->mutable_data(ctx.GetPlace()); - auto x = ctx.Input("X"); - auto dx = ctx.Output(framework::GradVarName("X")); - auto dy = ctx.Input(framework::GradVarName("Y")); - auto label = ctx.Input("Label"); + const T* dy_data = + ctx.Input(framework::GradVarName("Y"))->data(); + T* dx_data = dx->mutable_data(ctx.GetPlace()); + const T* x_data = x->data(); - auto* dx_data = dx->mutable_data(ctx.GetPlace()); - auto* dy_data = dy->data(); - auto* x_data = x->data(); + int batch_size = x->dims()[0]; + int class_num = x->dims()[1]; - int n = x->dims()[0]; - int d = x->dims()[1]; int block = 512; - int grid = (n * d + block - 1) / block; - zero<<>>(dx_data, n * d); - grid = (n + block - 1) / block; - // TODO(qingqing): launch kernel on specified stream - // base on ExecutionContext. - if (ctx.Attr("soft_label") == 1) { + int grid = (batch_size * class_num + block - 1) / block; + + if (ctx.Attr("softLabel")) { auto* label_data = label->data(); - SoftCrossEntropyGradientKernel<<>>( - dx_data, dy_data, x_data, label_data, n, d); + SoftCrossEntropyGradientKernel<<< + grid, block, 0, reinterpret_cast( + ctx.device_context()) + .stream()>>>(dx_data, dy_data, x_data, label_data, + batch_size, class_num); } else { + math::SetConstant(ctx.device_context(), dx, 0); auto* label_data = label->data(); - CrossEntropyGradientKernel<<>>(dx_data, dy_data, x_data, - label_data, n, d); + grid = (batch_size + block - 1) / block; + CrossEntropyGradientKernel<<< + grid, block, 0, reinterpret_cast( + ctx.device_context()) + .stream()>>>(dx_data, dy_data, x_data, label_data, + batch_size, class_num); } } }; diff --git a/paddle/operators/cross_entropy_op.h b/paddle/operators/cross_entropy_op.h index 1b4b23ac20..d2d321aa7e 100644 --- a/paddle/operators/cross_entropy_op.h +++ b/paddle/operators/cross_entropy_op.h @@ -13,97 +13,65 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include "paddle/framework/eigen.h" #include "paddle/framework/op_registry.h" -#include "paddle/platform/hostdevice.h" +#include "paddle/operators/math/cross_entropy.h" +#include "paddle/operators/math/math_function.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; +template +using EigenMatrix = framework::EigenMatrix; template -HOSTDEVICE T tolerable_value(const T x) { - PADDLE_ASSERT(std::is_floating_point::value); - const T kApproInf = 1e20; - if (x == INFINITY) { - return kApproInf; - } - if (x == -INFINITY) { - return -kApproInf; - } - return x; -} - -template -class CrossEntropyOpKernel : public framework::OpKernel { +class CrossEntropyOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), - "It must use CPUPlace."); - - auto x = ctx.Input("X"); - auto y = ctx.Output("Y"); - - auto* x_data = x->data(); + "This kernel only runs on CPU."); + const Tensor* x = ctx.Input("X"); + const Tensor* labels = ctx.Input("Label"); + Tensor* y = ctx.Output("Y"); y->mutable_data(ctx.GetPlace()); - auto* y_data = y->data(); - int batch_size = x->dims()[0]; - int class_num = x->dims()[1]; - - if (ctx.Attr("soft_label") == 1) { - auto* label_data = ctx.Input("Label")->data(); - int index = 0; - for (int i = 0; i < batch_size; ++i) { - T sum = static_cast(0); - for (int j = 0; j < class_num; ++j) { - sum += label_data[index] * tolerable_value(std::log(x_data[index])); - y_data[i] = -sum; - index++; - } - } - } else { - auto* label_data = ctx.Input("Label")->data(); - for (int i = 0; i < batch_size; ++i) { - int index = i * class_num + label_data[i]; - y_data[i] = -tolerable_value(std::log(x_data[index])); - } - } + math::CrossEntropyFunctor()( + ctx.device_context(), y, x, labels, ctx.Attr("softLabel")); } }; template -class CrossEntropyGradientOpKernel : public framework::OpKernel { +class CrossEntropyGradientOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), - "It must use CPUPlace."); - - auto x = ctx.Input("X"); - auto dx = ctx.Output(framework::GradVarName("X")); - auto dy = ctx.Input(framework::GradVarName("Y")); - auto label = ctx.Input("Label"); - - auto* dx_data = dx->mutable_data(ctx.GetPlace()); - auto* dy_data = dy->data(); - auto* x_data = x->data(); + "This kernel only runs on CPU."); + const Tensor* x = ctx.Input("X"); + const Tensor* dy = ctx.Input(framework::GradVarName("Y")); + const Tensor* label = ctx.Input("Label"); + Tensor* dx = ctx.Output(framework::GradVarName("X")); + T* dx_data = dx->mutable_data(ctx.GetPlace()); - int batch_size = x->dims()[0]; int class_num = x->dims()[1]; - - // TODO(qingqing): make zero setting an common function. - if (ctx.Attr("soft_label") == 1) { - auto* label_data = ctx.Input("Label")->data(); - int index = 0; - for (int i = 0; i < batch_size; ++i) { - for (int j = 0; j < class_num; ++j) { - dx_data[index] = -label_data[index] * dy_data[i] / x_data[index]; - index++; - } - } + if (ctx.Attr("softLabel")) { + auto x_mat = EigenMatrix::From(*x); + auto dy_mat = EigenMatrix::From(*dy); + auto lbl_mat = EigenMatrix::From(*label); + auto dx_mat = EigenMatrix::From(*dx); + + dx_mat.device(ctx.GetEigenDevice()) = + -(lbl_mat * dy_mat.broadcast(Eigen::DSizes(1, class_num)) / + x_mat); } else { - auto* label_data = label->data(); - memset(dx_data, 0, sizeof(T) * batch_size * class_num); + int batch_size = x->dims()[0]; + const T* dy_data = dy->data(); + const T* x_data = x->data(); + const int* label_data = label->data(); + + math::SetConstant(ctx.device_context(), dx, 0); + for (int i = 0; i < batch_size; ++i) { PADDLE_ASSERT(label_data[i] >= 0 || label_data[i] < class_num); int index = i * class_num + label_data[i]; diff --git a/paddle/operators/detail/strided_memcpy.h b/paddle/operators/detail/strided_memcpy.h new file mode 100644 index 0000000000..068c82f399 --- /dev/null +++ b/paddle/operators/detail/strided_memcpy.h @@ -0,0 +1,93 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/ddim.h" +#include "paddle/memory/memcpy.h" +#include "paddle/platform/device_context.h" + +namespace paddle { +namespace operators { +namespace detail { + +template +struct StridedMemcpyFunctor; + +template +struct StridedMemcpyFunctor { + void operator()(const platform::DeviceContext& dev_ctx, const T* src, + framework::Dim<1> src_stride, framework::Dim<1> dst_dim, + framework::Dim<1> dst_stride, T* dst) const { + auto place = dev_ctx.GetPlace(); + if (platform::is_cpu_place(place)) { + auto& cpu_place = boost::get(place); + memory::Copy(cpu_place, dst, cpu_place, src, sizeof(T) * dst_dim.head); + } else { +#ifdef PADDLE_WITH_CUDA + auto& gpu_place = boost::get(place); + auto& cuda_ctx = + reinterpret_cast(dev_ctx); + memory::Copy(gpu_place, dst, gpu_place, src, sizeof(T) * dst_dim.head, + cuda_ctx.stream()); +#else + PADDLE_THROW("Paddle is not compiled with GPU"); +#endif + } + } +}; + +template +struct StridedMemcpyFunctor { + void operator()(const platform::DeviceContext& dev_ctx, const T* src, + framework::Dim src_stride, framework::Dim dst_dim, + framework::Dim dst_stride, T* dst) const { + for (int64_t i = 0; i < dst_dim.head; ++i) { + StridedMemcpyFunctor func; + func(dev_ctx, src, src_stride.tail, dst_dim.tail, dst_stride.tail, dst); + src += src_stride.head; + dst += dst_stride.head; + } + } +}; + +template +struct StridedCopyDimVisitor : public boost::static_visitor { + StridedCopyDimVisitor(const platform::DeviceContext& dev_ctx, const T* src, + const framework::DDim& src_stride, + const framework::DDim& dst_stride, T* dst) + : dev_ctx_(dev_ctx), + src_(src), + src_stride_(src_stride), + dst_stride_(dst_stride), + dst_(dst) {} + + template + void operator()(Dim dst_dim) const { + Dim src_stride = boost::get(src_stride_); + Dim dst_stride = boost::get(dst_stride_); + constexpr int dim = Dim::dimensions; + StridedMemcpyFunctor functor; + functor(dev_ctx_, src_, src_stride, dst_dim, dst_stride, dst_); + } + + const platform::DeviceContext& dev_ctx_; + const T* src_; + const framework::DDim& src_stride_; + const framework::DDim& dst_stride_; + T* dst_; +}; + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/dropout_op.cc b/paddle/operators/dropout_op.cc index b111b9fccb..a669b5cf00 100644 --- a/paddle/operators/dropout_op.cc +++ b/paddle/operators/dropout_op.cc @@ -18,39 +18,35 @@ namespace paddle { namespace operators { using framework::Tensor; -using framework::LoDTensor; class DropoutOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_GE(ctx.Attr("dropout_prob"), 0); - PADDLE_ENFORCE_LE(ctx.Attr("dropout_prob"), 1); - // TODO(xinghai-sun): remove this check after swtiching to bool - PADDLE_ENFORCE(ctx.Attr("is_training") == 0 || - ctx.Attr("is_training") == 1); - - auto dims = ctx.Input("X")->dims(); - ctx.Output("Out")->Resize(dims); - if (ctx.Attr("is_training") == 1) { - ctx.Output("Mask")->Resize(dims); + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); + PADDLE_ENFORCE_GE(ctx->Attrs().Get("dropout_prob"), 0); + PADDLE_ENFORCE_LE(ctx->Attrs().Get("dropout_prob"), 1); + + auto x_dims = ctx->GetInputDim("X"); + ctx->SetOutputDim("Out", x_dims); + if (ctx->Attrs().Get("is_training") == 1) { + ctx->SetOutputDim("Mask", x_dims); } + ctx->ShareLoD("X", /*->*/ "Out"); } }; template class DropoutOpMaker : public framework::OpProtoAndCheckerMaker { public: - DropoutOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + DropoutOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddAttr("dropout_prob", "Probability of setting units to zero.") .SetDefault(.5f); - // TODO(xinghai-sun): use bool for is_training after bool is supported. - AddAttr("is_training", "Whether in training phase.").SetDefault(1); + AddAttr("is_training", "Whether in training phase.").SetDefault(true); AddAttr("seed", "Dropout random seed.").SetDefault(0); AddInput("X", "The input of dropout op."); AddOutput("Out", "The output of dropout op."); @@ -59,7 +55,7 @@ class DropoutOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( Dropout Operator. -"Dropout" refers to randomly dropping out units in a nerual network. It is a +'Dropout' refers to randomly dropping out units in a nerual network. It is a regularization technique for reducing overfitting by preventing neuron co-adaption during training. The dropout operator randomly set (according to the given dropout probability) the outputs of some units to zero, while others @@ -74,30 +70,26 @@ class DropoutOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.Attr("is_training"), 1, + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE_EQ(ctx->Attrs().Get("is_training"), 1, "GradOp is only callable when is_training is true"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Mask"), "Mask must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) must not be null."); - - PADDLE_ENFORCE_GE(ctx.Attr("dropout_prob"), 0); - PADDLE_ENFORCE_LE(ctx.Attr("dropout_prob"), 1); - // TODO(xinghai-sun): remove this check after swtiching to bool - PADDLE_ENFORCE(ctx.Attr("is_training") == 0 || - ctx.Attr("is_training") == 1); - auto x_dims = ctx.Input("X")->dims(); - auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("Mask"), "Mask must not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) must not be null."); + + PADDLE_ENFORCE_GE(ctx->Attrs().Get("dropout_prob"), 0); + PADDLE_ENFORCE_LE(ctx->Attrs().Get("dropout_prob"), 1); + auto x_dims = ctx->GetInputDim("X"); + auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_EQ(x_dims, out_dims, "Dimensions of Input(X) and Out@Grad must be the same."); - auto mask_dims = ctx.Input("Mask")->dims(); + auto mask_dims = ctx->GetInputDim("Mask"); PADDLE_ENFORCE_EQ(x_dims, mask_dims, "Dimensions of Input(X) and Mask must be the same."); - auto *x_grad = ctx.Output(framework::GradVarName("X")); - x_grad->Resize(x_dims); + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); } }; diff --git a/paddle/operators/dropout_op.cu b/paddle/operators/dropout_op.cu index 186237fb23..30c769000f 100644 --- a/paddle/operators/dropout_op.cu +++ b/paddle/operators/dropout_op.cu @@ -47,7 +47,7 @@ struct MaskGenerator { // Use std::random and thrust::random(thrust is a std library in CUDA) to // implement uniform random. template -class GPUDropoutKernel : public framework::OpKernel { +class GPUDropoutKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* x = context.Input("X"); @@ -59,7 +59,7 @@ class GPUDropoutKernel : public framework::OpKernel { auto Y = EigenMatrix::Reshape(*y, 1); auto place = context.GetEigenDevice(); - if (context.Attr("is_training") == 1) { + if (context.Attr("is_training")) { auto* mask = context.Output("Mask"); auto* mask_data = mask->mutable_data(context.GetPlace()); int size = framework::product(mask->dims()); diff --git a/paddle/operators/dropout_op.h b/paddle/operators/dropout_op.h index 82eafee0e0..745525fe81 100644 --- a/paddle/operators/dropout_op.h +++ b/paddle/operators/dropout_op.h @@ -26,7 +26,7 @@ template ; template -class CPUDropoutKernel : public framework::OpKernel { +class CPUDropoutKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* x = context.Input("X"); @@ -35,7 +35,7 @@ class CPUDropoutKernel : public framework::OpKernel { auto* y_data = y->mutable_data(context.GetPlace()); AttrType dropout_prob = context.Attr("dropout_prob"); - if (context.Attr("is_training") == 1) { + if (context.Attr("is_training")) { auto* mask = context.Output("Mask"); auto* mask_data = mask->mutable_data(context.GetPlace()); int seed = context.Attr("seed"); @@ -62,11 +62,11 @@ class CPUDropoutKernel : public framework::OpKernel { }; template -class DropoutGradKernel : public framework::OpKernel { +class DropoutGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - PADDLE_ENFORCE_EQ(context.Attr("is_training"), 1, - "GradOp is only callable when is_training is true"); + PADDLE_ENFORCE(context.Attr("is_training"), + "GradOp is only callable when is_training is true"); auto* grad_x = context.Output(framework::GradVarName("X")); auto* grad_y = context.Input(framework::GradVarName("Out")); diff --git a/paddle/operators/elementwise_add_op.cc b/paddle/operators/elementwise_add_op.cc new file mode 100644 index 0000000000..d9bc80c869 --- /dev/null +++ b/paddle/operators/elementwise_add_op.cc @@ -0,0 +1,40 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/elementwise_add_op.h" +#include "paddle/operators/elementwise_op.h" + +namespace paddle { +namespace operators { +class ElementwiseAddOpMaker : public ElementwiseOpMaker { + public: + ElementwiseAddOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : ElementwiseOpMaker(proto, op_checker) { + SetComment("add", "Out = X + Y"); + AddComment(comment_); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(elementwise_add, ops::ElementwiseOp, ops::ElementwiseAddOpMaker, + elementwise_add_grad, ops::ElementwiseOpGrad); +REGISTER_OP_CPU_KERNEL( + elementwise_add, + ops::ElementwiseAddKernel); +REGISTER_OP_CPU_KERNEL( + elementwise_add_grad, + ops::ElementwiseAddGradKernel); diff --git a/paddle/operators/sequence_avg_pool_op.cu b/paddle/operators/elementwise_add_op.cu similarity index 74% rename from paddle/operators/sequence_avg_pool_op.cu rename to paddle/operators/elementwise_add_op.cu index bc9d1611fc..85d063a76b 100644 --- a/paddle/operators/sequence_avg_pool_op.cu +++ b/paddle/operators/elementwise_add_op.cu @@ -13,13 +13,13 @@ limitations under the License. */ #define EIGEN_USE_GPU - -#include "paddle/operators/sequence_avg_pool_op.h" +#include "paddle/operators/elementwise_add_op.h" namespace ops = paddle::operators; + REGISTER_OP_GPU_KERNEL( - sequence_avg_pool, - ops::SequenceAvgPoolKernel); + elementwise_add, + ops::ElementwiseAddKernel); REGISTER_OP_GPU_KERNEL( - sequence_avg_pool_grad, - ops::SequenceAvgPoolGradKernel); + elementwise_add_grad, + ops::ElementwiseAddGradKernel); diff --git a/paddle/operators/elementwise_add_op.h b/paddle/operators/elementwise_add_op.h new file mode 100644 index 0000000000..f04fe3ec60 --- /dev/null +++ b/paddle/operators/elementwise_add_op.h @@ -0,0 +1,115 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/operators/elementwise_op_function.h" + +namespace paddle { +namespace operators { + +template +class ElementwiseAddKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + ElementwiseCompute(ctx); + } +}; + +template +struct ElementwiseAddGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) { + auto dz_e = framework::EigenVector::Flatten(*dz); + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e; + } + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = dz_e; + } + } +}; + +template +struct ElementwiseAddOneGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) { + auto dz_e = framework::EigenVector::Flatten(*dz); + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e; + } + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = dz_e.sum(); + } + } +}; + +template +struct ElementwiseAddBroadCastGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) { + auto dz_e = framework::EigenVector::Flatten(*dz); + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e; + } + + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = dz_e.reshape(Eigen::DSizes(pre, n)) + .sum(Eigen::array{{0}}); + } + } +}; + +template +struct ElementwiseAddBroadCast2GradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n, + Post post) { + auto dz_e = framework::EigenVector::Flatten(*dz); + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e; + } + + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = dz_e.reshape(Eigen::DSizes(pre, n, post)) + .sum(Eigen::array{{0, 2}}); + } + } +}; + +template +class ElementwiseAddGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + ElementwiseGradCompute, + ElementwiseAddOneGradFunctor, + ElementwiseAddBroadCastGradFunctor, + ElementwiseAddBroadCast2GradFunctor>(ctx); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/elementwise_div_op.cc b/paddle/operators/elementwise_div_op.cc new file mode 100644 index 0000000000..3f56344d00 --- /dev/null +++ b/paddle/operators/elementwise_div_op.cc @@ -0,0 +1,41 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/elementwise_div_op.h" +#include "paddle/operators/elementwise_op.h" + +namespace paddle { +namespace operators { +class ElementwiseDivOpMaker : public ElementwiseOpMaker { + public: + ElementwiseDivOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : ElementwiseOpMaker(proto, op_checker) { + SetComment("Div", "Out = X / Y"); + AddComment(comment_); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(elementwise_div, ops::ElementwiseOp, ops::ElementwiseDivOpMaker, + elementwise_div_grad, ops::ElementwiseOpGrad); +REGISTER_OP_CPU_KERNEL( + elementwise_div, + ops::ElementwiseDivKernel); +REGISTER_OP_CPU_KERNEL( + elementwise_div_grad, + ops::ElementwiseDivGradKernel); diff --git a/paddle/operators/elementwise_div_op.cu b/paddle/operators/elementwise_div_op.cu new file mode 100644 index 0000000000..b96aa31748 --- /dev/null +++ b/paddle/operators/elementwise_div_op.cu @@ -0,0 +1,25 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/elementwise_div_op.h" + +namespace ops = paddle::operators; + +REGISTER_OP_GPU_KERNEL( + elementwise_div, + ops::ElementwiseDivKernel); +REGISTER_OP_GPU_KERNEL( + elementwise_div_grad, + ops::ElementwiseDivGradKernel); diff --git a/paddle/operators/elementwise_div_op.h b/paddle/operators/elementwise_div_op.h new file mode 100644 index 0000000000..8946ff3d25 --- /dev/null +++ b/paddle/operators/elementwise_div_op.h @@ -0,0 +1,117 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/operators/elementwise_op_function.h" + +namespace paddle { +namespace operators { + +template +class ElementwiseDivKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + ElementwiseCompute(ctx); + } +}; + +template +struct ElementwiseDivGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) { + auto y_e = framework::EigenVector::Flatten(*y); + auto z_e = framework::EigenVector::Flatten(*z); + auto dz_e = framework::EigenVector::Flatten(*dz); + + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e / y_e; + } + + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = -1.0 * dz_e * z_e / y_e; + } + } +}; + +template +struct ElementwiseDivBroadCastGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) { + auto x_e = framework::EigenVector::Flatten(*x); + auto y_e = framework::EigenVector::Flatten(*y); + auto dz_e = framework::EigenVector::Flatten(*dz); + + auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n)) + .broadcast(Eigen::DSizes(pre, 1)) + .reshape(Eigen::DSizes(x_e.size())); + + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e / y_e_bcast; + } + + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = (-1.0 * (x_e * dz_e) / (y_e_bcast * y_e_bcast)) + .reshape(Eigen::DSizes(pre, n)) + .sum(Eigen::array{{0}}); + } + } +}; + +template +struct ElementwiseDivBroadCast2GradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n, + Post post) { + auto x_e = framework::EigenVector::Flatten(*x); + auto y_e = framework::EigenVector::Flatten(*y); + auto dz_e = framework::EigenVector::Flatten(*dz); + + auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) + .broadcast(Eigen::DSizes(pre, 1, post)) + .reshape(Eigen::DSizes(x_e.size())); + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e / y_e_bcast; + } + + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = (-1.0 * (x_e * dz_e) / (y_e_bcast * y_e_bcast)) + .reshape(Eigen::DSizes(pre, n, post)) + .sum(Eigen::array{{0, 2}}); + } + } +}; + +template +class ElementwiseDivGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + ElementwiseGradCompute, + ElementwiseDivGradFunctor, + ElementwiseDivBroadCastGradFunctor, + ElementwiseDivBroadCast2GradFunctor>(ctx); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/elementwise_mul_op.cc b/paddle/operators/elementwise_mul_op.cc index ee6e975b44..da7765aa6a 100644 --- a/paddle/operators/elementwise_mul_op.cc +++ b/paddle/operators/elementwise_mul_op.cc @@ -13,105 +13,32 @@ limitations under the License. */ #include "paddle/operators/elementwise_mul_op.h" +#include "paddle/operators/elementwise_op.h" namespace paddle { namespace operators { -using Tensor = framework::Tensor; - -class ElementWiseMulOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of ElementWiseMulOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), - "Input(Y) of ElementWiseMulOp should not be null."); - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Out"), - "Output(Out) of ElementWiseMulOp should not be null."); - - auto x_dim = ctx.Input("X")->dims(); - auto y_dim = ctx.Input("Y")->dims(); - PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(), - "Rank of first input must >= rank of second input.") - ctx.Output("Out")->Resize(x_dim); - } -}; - -class ElementWiseMulOpMaker : public framework::OpProtoAndCheckerMaker { +class ElementwiseMulOpMaker : public ElementwiseOpMaker { public: - ElementWiseMulOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The first input of elementwise mul op"); - AddInput("Y", "The second input of elementwise mul op"); - AddAttr("axis", - R"DOC( -When shape(Y) does not equal shape(X),Y will be broadcasted -to match the shape of X and axis should be dimension index Y in X - )DOC") - .SetDefault(-1) - .EqualGreaterThan(-1); - - AddOutput("Out", "The output of elementwise mul op"); - AddComment(R"DOC( -Limited elementwise multiple operator.The equation is: Out = X ⊙ Y. -1. The shape of Y should be same with X or -2. Y's shape is a subset of X. - Y will be broadcasted to match the shape of X and axis should be dimension index Y in X. - example: - shape(X) = (2, 3, 4, 5), shape(Y) = (,) - shape(X) = (2, 3, 4, 5), shape(Y) = (5,) - shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5) - shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1 - shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0 -)DOC"); + ElementwiseMulOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : ElementwiseOpMaker(proto, op_checker) { + SetComment("Mul", "Out = X ⊙ Y"); + AddComment(comment_); } }; -class ElementWiseMulOpGrad : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); - - auto x_dims = ctx.Input("X")->dims(); - auto y_dims = ctx.Input("Y")->dims(); - auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); - auto *x_grad = - ctx.Output(framework::GradVarName("X")); - auto *y_grad = - ctx.Output(framework::GradVarName("Y")); - - PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), - "Rank of first input must >= rank of second input.") - - if (x_grad) { - x_grad->Resize(x_dims); - } - - if (y_grad) { - y_grad->Resize(y_dims); - } - } -}; } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(elementwise_mul, ops::ElementWiseMulOp, ops::ElementWiseMulOpMaker, - elementwise_mul_grad, ops::ElementWiseMulOpGrad); +REGISTER_OP(elementwise_mul, ops::ElementwiseOp, ops::ElementwiseMulOpMaker, + elementwise_mul_grad, ops::ElementwiseOpGrad); REGISTER_OP_CPU_KERNEL( elementwise_mul, - ops::ElementWiseMulKernel); + ops::ElementwiseMulKernel, + ops::ElementwiseMulKernel); REGISTER_OP_CPU_KERNEL( elementwise_mul_grad, - ops::ElementWiseMulGradKernel); + ops::ElementwiseMulGradKernel, + ops::ElementwiseMulGradKernel); diff --git a/paddle/operators/elementwise_mul_op.cu b/paddle/operators/elementwise_mul_op.cu index 56f2087c22..056f081d3e 100644 --- a/paddle/operators/elementwise_mul_op.cu +++ b/paddle/operators/elementwise_mul_op.cu @@ -19,7 +19,9 @@ namespace ops = paddle::operators; REGISTER_OP_GPU_KERNEL( elementwise_mul, - ops::ElementWiseMulKernel); + ops::ElementwiseMulKernel, + ops::ElementwiseMulKernel); REGISTER_OP_GPU_KERNEL( elementwise_mul_grad, - ops::ElementWiseMulGradKernel); + ops::ElementwiseMulGradKernel, + ops::ElementwiseMulGradKernel); diff --git a/paddle/operators/elementwise_mul_op.h b/paddle/operators/elementwise_mul_op.h index 6d58da580b..4469b07eaa 100644 --- a/paddle/operators/elementwise_mul_op.h +++ b/paddle/operators/elementwise_mul_op.h @@ -13,171 +13,104 @@ limitations under the License. */ #pragma once -#include "paddle/framework/eigen.h" -#include "paddle/framework/op_registry.h" +#include "paddle/operators/elementwise_op_function.h" namespace paddle { namespace operators { -/* - * Out = X ⊙ Y - * 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1 - * pre=2, n=3*4, post=5 - * 2. shape(X) = (2, 3, 4, 5), shape(Y) = (4,5) - * pre=2*3, n=4*5, post=1 - */ - -inline void get_mid_dims(const framework::DDim& x_dims, - const framework::DDim& y_dims, const int axis, - int& pre, int& n, int& post) { - pre = 1; - n = 1; - post = 1; - for (int i = 0; i < axis; ++i) { - pre *= x_dims[i]; - } - - for (int i = 0; i < y_dims.size(); ++i) { - PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i], - "Broadcast dimension mismatch."); - n *= y_dims[i]; - } - - for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) { - post *= x_dims[i]; - } -} template -class ElementWiseMulKernel : public framework::OpKernel { +class ElementwiseMulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - using Tensor = framework::Tensor; - - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* z = ctx.Output("Out"); - z->mutable_data(ctx.GetPlace()); + ElementwiseCompute(ctx); + } +}; +template +struct ElementwiseMulGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) { auto x_e = framework::EigenVector::Flatten(*x); auto y_e = framework::EigenVector::Flatten(*y); - auto z_e = framework::EigenVector::Flatten(*z); + auto dz_e = framework::EigenVector::Flatten(*dz); - auto x_dims = x->dims(); - auto y_dims = y->dims(); - PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), - "Rank of first input must >= rank of second input.") - - if (x_dims == y_dims || product(y_dims) == 1) { - z_e.device(ctx.GetEigenDevice()) = x_e * y_e; - return; + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e * y_e; } - int axis = ctx.Attr("axis"); - axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); - PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(), - "Axis should be in range [0, x_dims)"); - - int pre, n, post; - get_mid_dims(x_dims, y_dims, axis, pre, n, post); - if (post == 1) { - auto y_bcast = y_e.reshape(Eigen::DSizes(1, n)) - .broadcast(Eigen::DSizes(pre, 1)) - .reshape(Eigen::DSizes(x_e.size())); - z_e.device(ctx.GetEigenDevice()) = x_e * y_bcast; - return; - } else { - auto y_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) - .broadcast(Eigen::DSizes(pre, 1, post)) - .reshape(Eigen::DSizes(x_e.size())); - z_e.device(ctx.GetEigenDevice()) = x_e * y_bcast; - return; + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = x_e * dz_e; } } }; -template -class ElementWiseMulGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - using Tensor = framework::Tensor; - - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* dout = ctx.Input(framework::GradVarName("Out")); - +template +struct ElementwiseMulBroadCastGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) { auto x_e = framework::EigenVector::Flatten(*x); auto y_e = framework::EigenVector::Flatten(*y); - auto dout_e = framework::EigenVector::Flatten(*dout); + auto dz_e = framework::EigenVector::Flatten(*dz); - auto x_dims = x->dims(); - auto y_dims = y->dims(); + auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n)) + .broadcast(Eigen::DSizes(pre, 1)) + .reshape(Eigen::DSizes(x_e.size())); - auto* dx = ctx.Output(framework::GradVarName("X")); - auto* dy = ctx.Output(framework::GradVarName("Y")); if (dx) { - dx->mutable_data(ctx.GetPlace()); + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e * y_e_bcast; } + if (dy) { - dy->mutable_data(ctx.GetPlace()); + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = (x_e * dz_e) + .reshape(Eigen::DSizes(pre, n)) + .sum(Eigen::array{{0}}); } + } +}; - if (x_dims == y_dims || product(y_dims) == 1) { - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(ctx.GetEigenDevice()) = dout_e * y_e; - } - - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(ctx.GetEigenDevice()) = x_e * dout_e; - } - return; +template +struct ElementwiseMulBroadCast2GradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n, + Post post) { + auto x_e = framework::EigenVector::Flatten(*x); + auto y_e = framework::EigenVector::Flatten(*y); + auto dz_e = framework::EigenVector::Flatten(*dz); + + auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) + .broadcast(Eigen::DSizes(pre, 1, post)) + .reshape(Eigen::DSizes(x_e.size())); + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e * y_e_bcast; } - int axis = ctx.Attr("axis"); - axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); - - int pre, n, post; - get_mid_dims(x_dims, y_dims, axis, pre, n, post); - - // TODO(gongweibao): wrap reshape to a function. - if (post == 1) { - auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n)) - .broadcast(Eigen::DSizes(pre, 1)) - .reshape(Eigen::DSizes(x_e.size())); - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(ctx.GetEigenDevice()) = dout_e * y_e_bcast; - } - - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(ctx.GetEigenDevice()) = - (x_e * dout_e) - .reshape(Eigen::DSizes(pre, n)) - .sum(Eigen::array{{0}}); - } - return; - } else { - auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) - .broadcast(Eigen::DSizes(pre, 1, post)) - .reshape(Eigen::DSizes(x_e.size())); - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(ctx.GetEigenDevice()) = dout_e * y_e_bcast; - } - - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(ctx.GetEigenDevice()) = - (x_e * dout_e) - .reshape(Eigen::DSizes(pre, n, post)) - .sum(Eigen::array{{0, 2}}); - } - return; + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = (x_e * dz_e) + .reshape(Eigen::DSizes(pre, n, post)) + .sum(Eigen::array{{0, 2}}); } } }; +template +class ElementwiseMulGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + ElementwiseGradCompute, + ElementwiseMulGradFunctor, + ElementwiseMulBroadCastGradFunctor, + ElementwiseMulBroadCast2GradFunctor>(ctx); + } +}; + } // namespace operators } // namespace paddle diff --git a/paddle/operators/elementwise_op.h b/paddle/operators/elementwise_op.h new file mode 100644 index 0000000000..3082f37422 --- /dev/null +++ b/paddle/operators/elementwise_op.h @@ -0,0 +1,133 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/op_registry.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace operators { + +class ElementwiseOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + using Tensor = framework::Tensor; + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of elementwise op should not be null"); + PADDLE_ENFORCE(ctx->HasInput("Y"), + "Input(Y) of elementwise op should not be null"); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of elementwise op should not be null."); + + auto x_dim = ctx->GetInputDim("X"); + auto y_dim = ctx->GetInputDim("Y"); + PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(), + "Rank of first input must >= rank of second input.") + ctx->SetOutputDim("Out", x_dim); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ElementwiseOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", R"DOC( +The first input of elementwise op, it's a tensor of any dimensions. +)DOC"); + AddInput("Y", R"DOC( +The sencond input of elementwise op, it's a tensor and it's dimensions +must be small or equal to X's dimensions. +)DOC"); + AddAttr("axis", + R"DOC( +When the shape(Y) does not equal the shape(X),Y will be broadcasted +to match the shape of X and axis should be dimension index Y in X + )DOC") + .SetDefault(-1) + .EqualGreaterThan(-1); + + AddOutput("Out", "The output of elementwise op"); + comment_ = R"DOC( +Limited elementwise {name} operator.The equation is: Out = {equation}. +1. The shape of Y should be same with X or +2. Y's shape is a subset of X. + Y will be broadcasted to match the shape of X and axis should be dimension index Y in X. + + example: + shape(X) = (2, 3, 4, 5), shape(Y) = (,) + shape(X) = (2, 3, 4, 5), shape(Y) = (5,) + shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5) + shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1 + shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0 + +Both the input X and Y can carry the LoD (Level of Details) information, +or not. But the output only shares the LoD with input X. +)DOC"; + AddComment(comment_); + } + + protected: + std::string comment_; + + void Replace(std::string& src, std::string from, std::string to) { + std::size_t len_from = std::strlen(from.c_str()); + std::size_t len_to = std::strlen(to.c_str()); + for (std::size_t pos = src.find(from); pos != std::string::npos; + pos = src.find(from, pos + len_to)) { + src.replace(pos, len_from, to); + } + } + + void SetComment(std::string name, std::string equation) { + Replace(comment_, "{name}", name); + Replace(comment_, "{equation}", equation); + } +}; + +class ElementwiseOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + using Tensor = framework::Tensor; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); + + PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), + "Rank of first input must >= rank of second input.") + + auto x_grad_name = framework::GradVarName("X"); + auto y_grad_name = framework::GradVarName("Y"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); + } + if (ctx->HasOutput(y_grad_name)) { + ctx->SetOutputDim(y_grad_name, y_dims); + } + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/elementwise_op_function.h b/paddle/operators/elementwise_op_function.h new file mode 100644 index 0000000000..3eb97f60b5 --- /dev/null +++ b/paddle/operators/elementwise_op_function.h @@ -0,0 +1,200 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/framework/operator.h" + +#include "paddle/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +/* + * Out = X ⊙ Y + * If Y's shape does not match X' shape, they will be reshaped. + * For example: + * 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1 + * pre=2, n=3*4, post=5 + * x.shape(2, 12, 5) * y.shape(1,12,1).broadcast(2,12,5) + * 2. shape(X) = (2, 3, 4, 5), shape(Y) = (4,5) + * pre=2*3, n=4*5, post=1 + * x.shape(2, 3, 20) * y.shape(1,1,20).broadcast(2,3,20) + */ +inline void get_mid_dims(const framework::DDim& x_dims, + const framework::DDim& y_dims, const int axis, + int& pre, int& n, int& post) { + pre = 1; + n = 1; + post = 1; + for (int i = 0; i < axis; ++i) { + pre *= x_dims[i]; + } + + for (int i = 0; i < y_dims.size(); ++i) { + PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i], + "Broadcast dimension mismatch."); + n *= y_dims[i]; + } + + for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) { + post *= x_dims[i]; + } +} + +#define EIGEN_FUNCTOR(name, eigen_op) \ + struct Eigen##name##Functor { \ + template \ + inline void Run(const framework::Tensor* x, const framework::Tensor* y, \ + framework::Tensor* z, \ + const framework::ExecutionContext& ctx) { \ + auto x_e = framework::EigenVector::Flatten(*x); \ + auto y_e = framework::EigenVector::Flatten(*y); \ + auto z_e = framework::EigenVector::Flatten(*z); \ + z_e.device(ctx.GetEigenDevice()) = eigen_op(x_e, y_e); \ + } \ + template \ + inline void RunBroadCast(const framework::Tensor* x, \ + const framework::Tensor* y, framework::Tensor* z, \ + const framework::ExecutionContext& ctx, int pre, \ + int n) { \ + auto x_e = framework::EigenVector::Flatten(*x); \ + auto y_e = framework::EigenVector::Flatten(*y); \ + auto z_e = framework::EigenVector::Flatten(*z); \ + auto y_bcast = y_e.reshape(Eigen::DSizes(1, n)) \ + .broadcast(Eigen::DSizes(pre, 1)) \ + .reshape(Eigen::DSizes(x_e.size())); \ + z_e.device(ctx.GetEigenDevice()) = eigen_op(x_e, y_bcast); \ + } \ + template \ + inline void RunBroadCast2(const framework::Tensor* x, \ + const framework::Tensor* y, \ + framework::Tensor* z, \ + const framework::ExecutionContext& ctx, int pre, \ + int n, int post) { \ + auto x_e = framework::EigenVector::Flatten(*x); \ + auto y_e = framework::EigenVector::Flatten(*y); \ + auto z_e = framework::EigenVector::Flatten(*z); \ + auto y_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) \ + .broadcast(Eigen::DSizes(pre, 1, post)) \ + .reshape(Eigen::DSizes(x_e.size())); \ + z_e.device(ctx.GetEigenDevice()) = eigen_op(x_e, y_bcast); \ + } \ + } + +template +void ElementwiseCompute(const framework::ExecutionContext& ctx) { + using Tensor = framework::Tensor; + + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* z = ctx.Output("Out"); + z->mutable_data(ctx.GetPlace()); + + auto x_dims = x->dims(); + auto y_dims = y->dims(); + PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), + "Rank of first input must >= rank of second input.") + + if (x_dims == y_dims || product(y_dims) == 1) { + functor f; + f.template Run(x, y, z, ctx); + return; + } + + int axis = ctx.Attr("axis"); + axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); + PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(), + "Axis should be in range [0, x_dims)"); + + int pre, n, post; + get_mid_dims(x_dims, y_dims, axis, pre, n, post); + if (post == 1) { + functor f; + f.template RunBroadCast(x, y, z, ctx, pre, n); + return; + } else { + functor f; + f.template RunBroadCast2(x, y, z, ctx, pre, n, post); + return; + } +} + +#define EIGEN_ADD(x, y) ((x) + (y)) +EIGEN_FUNCTOR(Add, EIGEN_ADD); + +#define EIGEN_SUB(x, y) ((x) - (y)) +EIGEN_FUNCTOR(Sub, EIGEN_SUB); + +#define EIGEN_MUL(x, y) ((x) * (y)) +EIGEN_FUNCTOR(Mul, EIGEN_MUL); + +#define EIGEN_DIV(x, y) ((x) / (y)) +EIGEN_FUNCTOR(Div, EIGEN_DIV); + +template +void ElementwiseGradCompute(const framework::ExecutionContext& ctx) { + using Tensor = framework::Tensor; + + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* out = ctx.Input("Out"); + auto* dout = ctx.Input(framework::GradVarName("Out")); + + auto place = ctx.GetEigenDevice(); + + auto x_dims = x->dims(); + auto y_dims = y->dims(); + + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dy = ctx.Output(framework::GradVarName("Y")); + if (dx) { + dx->mutable_data(ctx.GetPlace()); + } + if (dy) { + dy->mutable_data(ctx.GetPlace()); + } + + if (x_dims == y_dims) { + functor f; + f(place, x, y, out, dx, dy, dout); + return; + } + + if (product(y_dims) == 1) { + functor1 f; + f(place, x, y, out, dx, dy, dout); + return; + } + + int axis = ctx.Attr("axis"); + axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); + + int pre, n, post; + get_mid_dims(x_dims, y_dims, axis, pre, n, post); + + if (post == 1) { + broadcastfunctor f; + f(place, x, y, out, dx, dy, dout, pre, n); + return; + } else { + broadcast2functor f; + f(place, x, y, out, dx, dy, dout, pre, n, post); + return; + } +} +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/elementwise_sub_op.cc b/paddle/operators/elementwise_sub_op.cc new file mode 100644 index 0000000000..3e4f98fdb3 --- /dev/null +++ b/paddle/operators/elementwise_sub_op.cc @@ -0,0 +1,40 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/elementwise_sub_op.h" +#include "paddle/operators/elementwise_op.h" + +namespace paddle { +namespace operators { +class ElementwiseSubOpMaker : public ElementwiseOpMaker { + public: + ElementwiseSubOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : ElementwiseOpMaker(proto, op_checker) { + SetComment("Sub", "Out = X - Y"); + AddComment(comment_); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(elementwise_sub, ops::ElementwiseOp, ops::ElementwiseSubOpMaker, + elementwise_sub_grad, ops::ElementwiseOpGrad); +REGISTER_OP_CPU_KERNEL( + elementwise_sub, + ops::ElementwiseSubKernel); +REGISTER_OP_CPU_KERNEL( + elementwise_sub_grad, + ops::ElementwiseSubGradKernel); diff --git a/paddle/operators/elementwise_sub_op.cu b/paddle/operators/elementwise_sub_op.cu new file mode 100644 index 0000000000..0efb92fce9 --- /dev/null +++ b/paddle/operators/elementwise_sub_op.cu @@ -0,0 +1,25 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/elementwise_sub_op.h" + +namespace ops = paddle::operators; + +REGISTER_OP_GPU_KERNEL( + elementwise_sub, + ops::ElementwiseSubKernel); +REGISTER_OP_GPU_KERNEL( + elementwise_sub_grad, + ops::ElementwiseSubGradKernel); diff --git a/paddle/operators/elementwise_sub_op.h b/paddle/operators/elementwise_sub_op.h new file mode 100644 index 0000000000..3f40c1c5bc --- /dev/null +++ b/paddle/operators/elementwise_sub_op.h @@ -0,0 +1,116 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/operators/elementwise_op_function.h" + +namespace paddle { +namespace operators { + +template +class ElementwiseSubKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + ElementwiseCompute(ctx); + } +}; + +template +struct ElementwiseSubGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) { + auto dz_e = framework::EigenVector::Flatten(*dz); + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e; + } + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = (-1.0) * dz_e; + } + } +}; + +template +struct ElementwiseSubOneGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) { + auto dz_e = framework::EigenVector::Flatten(*dz); + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e; + } + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = (-1.0) * dz_e.sum(); + } + } +}; + +template +struct ElementwiseSubBroadCastGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) { + auto dz_e = framework::EigenVector::Flatten(*dz); + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e; + } + + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = (-1.0) * + dz_e.reshape(Eigen::DSizes(pre, n)) + .sum(Eigen::array{{0}}); + } + } +}; + +template +struct ElementwiseSubBroadCast2GradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n, + Post post) { + auto dz_e = framework::EigenVector::Flatten(*dz); + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = dz_e; + } + + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = (-1.0) * + dz_e.reshape(Eigen::DSizes(pre, n, post)) + .sum(Eigen::array{{0, 2}}); + } + } +}; + +template +class ElementwiseSubGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + ElementwiseGradCompute, + ElementwiseSubOneGradFunctor, + ElementwiseSubBroadCastGradFunctor, + ElementwiseSubBroadCast2GradFunctor>(ctx); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/fc_op.cc b/paddle/operators/fc_op.cc index e5d0f3c372..7c422c81fc 100644 --- a/paddle/operators/fc_op.cc +++ b/paddle/operators/fc_op.cc @@ -100,7 +100,7 @@ class FCOp : public NetOp { add_out = Output("AddOut"); AppendOp(framework::OpRegistry::CreateOp( - "rowwise_add", {{"X", {sum_out}}, {"b", {Input("B")}}}, + "elementwise_add", {{"X", {sum_out}}, {"Y", {Input("B")}}}, {{"Out", {add_out}}}, {})); } else { if (Output("AddOut") != framework::kEmptyVarName) { @@ -186,6 +186,9 @@ W_i is a 2-D matrix of size (K x N), where N means the number of neurons in the fully connected layer. B is a 1-D vector of size N. Thus, the output Out is a 2-D matrix of size (M x N). Activation type can be set to `identity` (default), `sigmoid` or `softmax`. + +All the inputs can carry the LoD (Level of Details) information, +or not. But the output only shares the LoD with first input (`X[0]`). )DOC"); } }; diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc index ba7857cc65..e164de6584 100644 --- a/paddle/operators/fill_zeros_like_op.cc +++ b/paddle/operators/fill_zeros_like_op.cc @@ -22,16 +22,13 @@ class FillZerosLikeOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL( - ctx.InputVar("Src"), - "Input(Src) of FillZerosLikeOp should not be null."); - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Dst"), - "Output(Dst) of FillZerosLikeOp should not be null."); - - ctx.Output("Dst")->Resize( - ctx.Input("Src")->dims()); + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of FillZerosLikeOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Y"), + "Output(Y) of FillZerosLikeOp should not be null."); + ctx->SetOutputDim("Y", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Y"); } }; @@ -40,8 +37,8 @@ class FillZerosLikeOpMaker : public framework::OpProtoAndCheckerMaker { FillZerosLikeOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : framework::OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("Src", "The input of fill-zeros-like op."); - AddOutput("Dst", "The varibale will be filled up with zeros."); + AddInput("X", "The input of fill-zeros-like op."); + AddOutput("Y", "The varibale will be filled up with zeros."); AddComment(R"DOC( Fill up a vriable with zeros. diff --git a/paddle/operators/fill_zeros_like_op.h b/paddle/operators/fill_zeros_like_op.h index 969998ce2e..cdf56a723b 100644 --- a/paddle/operators/fill_zeros_like_op.h +++ b/paddle/operators/fill_zeros_like_op.h @@ -20,10 +20,10 @@ namespace paddle { namespace operators { template -class FillZerosLikeKernel : public framework::OpKernel { +class FillZerosLikeKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* output = context.Output("Dst"); + auto* output = context.Output("Y"); output->mutable_data(context.GetPlace()); auto t = framework::EigenVector::Flatten(*output); t.device(context.GetEigenDevice()) = t.constant(static_cast(0)); diff --git a/paddle/operators/gather.cu.h b/paddle/operators/gather.cu.h new file mode 100644 index 0000000000..8d04ecd284 --- /dev/null +++ b/paddle/operators/gather.cu.h @@ -0,0 +1,79 @@ +/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; +using platform::Place; + +#define CUDA_1D_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \ + i += blockDim.x * gridDim.x) + +template +__global__ void GatherCUDAKernel(const T* params, const int* indices, T* output, + size_t index_size, size_t slice_size) { + CUDA_1D_KERNEL_LOOP(i, index_size * slice_size) { + int indices_i = i / slice_size; + int slice_i = i - indices_i * slice_size; // offset inside the slice + int gather_i = indices[indices_i]; + int params_i = gather_i * slice_size + slice_i; + *(output + i) = *(params + params_i); + } +} + +/** + * A thin wrapper on gpu tensor + * Return a new tensor from source tensor, gathered according to index + * input[src]: type-T source Tensor + * input[index]: type-int index Tensor (1-D) + * return: output tensor + */ +template +void GPUGather(const platform::DeviceContext& ctx, const Tensor& src, + const Tensor& index, Tensor* output) { + // PADDLE_ENFORCE(platform::is_gpu_place(place)); + // check index of shape 1-D + PADDLE_ENFORCE(index.dims().size() == 1); + int index_size = index.dims()[0]; + + auto src_dims = src.dims(); + framework::DDim output_dims(src_dims); + output_dims[0] = index_size; + + // slice size + int slice_size = 1; + for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i]; + + const T* p_src = src.data(); + const int* p_index = index.data(); + T* p_output = output->data(); + + int block = 512; + int n = slice_size * index_size; + int grid = (n + block - 1) / block; + + GatherCUDAKernel<<< + grid, block, 0, + reinterpret_cast(ctx).stream()>>>( + p_src, p_index, p_output, index_size, slice_size); +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/gather.h b/paddle/operators/gather.h index 92fb51ec17..052db49cb3 100644 --- a/paddle/operators/gather.h +++ b/paddle/operators/gather.h @@ -24,49 +24,40 @@ limitations under the License. */ namespace paddle { namespace operators { -// Implementation of CPU copy -template -void CPUGather(const T* src, const int* indices, const int slice_size, - const int index_size, T* output) { - const size_t slice_bytes = slice_size * sizeof(T); - - for (int i = 0; i < index_size; ++i) { - int index_ = indices[i]; - memcpy(output + i * slice_size, src + index_ * slice_size, slice_bytes); - } -} - -// Implementation of GPU copy: -template -void GPUGather(const T* src, const int* index, const int slice_size, - const int index_size, T* output); +using framework::Tensor; /** + * A thin wrapper for gathering on cpu tensor * Return a new tensor from source tensor, gathered according to index * input[src]: type-T source Tensor * input[index]: type-int index Tensor (1-D) * return: output tensor */ template -void Gather(const platform::Place& place, const paddle::framework::Tensor* src, - const paddle::framework::Tensor* index, - paddle::framework::Tensor* output) { +void CPUGather(const platform::DeviceContext& ctx, const Tensor& src, + const Tensor& index, Tensor* output) { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace())); // check index of shape 1-D - PADDLE_ENFORCE(index->dims().size() == 1); - int index_size = index->dims()[0]; + PADDLE_ENFORCE(index.dims().size() == 1); + int index_size = index.dims()[0]; - auto src_dims = src->dims(); + auto src_dims = src.dims(); framework::DDim output_dims(src_dims); output_dims[0] = index_size; + const T* p_src = src.data(); + const int* p_index = index.data(); + T* p_output = output->data(); + // slice size int slice_size = 1; for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i]; - // Gathering - if (platform::is_cpu_place(place)) { - CPUGather(src->data(), index->data(), slice_size, index_size, - output->data()); + const size_t slice_bytes = slice_size * sizeof(T); + + for (int i = 0; i < index_size; ++i) { + int index_ = p_index[i]; + memcpy(p_output + i * slice_size, p_src + index_ * slice_size, slice_bytes); } } diff --git a/paddle/operators/gather_op.cc b/paddle/operators/gather_op.cc index d445b61c16..fe305337cb 100644 --- a/paddle/operators/gather_op.cc +++ b/paddle/operators/gather_op.cc @@ -23,19 +23,26 @@ class GatherOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of GatherOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"), - "Input(Index) of GatherOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of GatherOp should not be null."); + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of GatherOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Index"), + "Input(Index) of GatherOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of GatherOp should not be null."); - int batch_size = ctx.Input("Index")->dims()[0]; + auto index_dims = ctx->GetInputDim("Index"); + PADDLE_ENFORCE(index_dims.size() == 1); + int batch_size = ctx->GetInputDim("Index")[0]; PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0"); - framework::DDim output_dims(ctx.Input("X")->dims()); + framework::DDim output_dims(ctx->GetInputDim("X")); output_dims[0] = batch_size; - ctx.Output("Out")->Resize(output_dims); + ctx->SetOutputDim("Out", output_dims); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("X")->type()); } }; @@ -44,23 +51,25 @@ class GatherGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - auto X_grad = ctx.Output(framework::GradVarName("X")); - auto X = ctx.Input("X"); + void InferShape(framework::InferShapeContextBase* ctx) const override { + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } - X_grad->Resize(X->dims()); + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("X")->type()); } }; class GatherOpMaker : public framework::OpProtoAndCheckerMaker { public: - GatherOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + GatherOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The source input of gather op"); AddInput("Index", "The index input of gather op"); AddOutput("Out", "The output of add op"); AddComment(R"DOC( -Gather Operator by selecting from the first axis, +Gather Operator by selecting from the first axis, Out = X[Index] )DOC"); @@ -72,8 +81,5 @@ Out = X[Index] namespace ops = paddle::operators; REGISTER_OP(gather, ops::GatherOp, ops::GatherOpMaker, gather_grad, ops::GatherGradOp); -REGISTER_OP_CPU_KERNEL(gather, - ops::GatherOpKernel); -REGISTER_OP_CPU_KERNEL( - gather_grad, - ops::GatherGradientOpKernel); +REGISTER_OP_CPU_KERNEL(gather, ops::GatherOpKernel); +REGISTER_OP_CPU_KERNEL(gather_grad, ops::GatherGradientOpKernel); diff --git a/paddle/operators/gather_op.cu b/paddle/operators/gather_op.cu new file mode 100644 index 0000000000..92219d6a43 --- /dev/null +++ b/paddle/operators/gather_op.cu @@ -0,0 +1,64 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "gather.cu.h" +#include "paddle/framework/eigen.h" +#include "paddle/operators/gather_op.h" +#include "scatter.cu.h" + +namespace paddle { +namespace operators { + +template +class GatherOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "This kernel only runs on GPU device."); + auto *x = ctx.Input("X"); + auto *index = ctx.Input("Index"); + auto *output = ctx.Output("Out"); + + output->mutable_data(ctx.GetPlace()); + + GPUGather(ctx.device_context(), *x, *index, output); + } +}; + +template +class GatherGradOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "This kernel only runs on GPU device."); + 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.GetEigenDevice(); + dxt.device(place) = dxt.constant(static_cast(0)); + + GPUScatterAssign(ctx.device_context(), *dO, *Index, dX); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(gather, ops::GatherOpCUDAKernel); +REGISTER_OP_GPU_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel); diff --git a/paddle/operators/gather_op.h b/paddle/operators/gather_op.h index 381854f301..8276ed0d3d 100644 --- a/paddle/operators/gather_op.h +++ b/paddle/operators/gather_op.h @@ -23,29 +23,40 @@ namespace operators { using Tensor = framework::Tensor; -template -class GatherOpKernel : public framework::OpKernel { +template +class GatherOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { - auto *X = ctx.Input("X"); - auto *Index = ctx.Input("Index"); - auto *Y = ctx.Output("Out"); + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "This kernel only runs on CPU."); + + auto *x = ctx.Input("X"); + auto *index = ctx.Input("Index"); + auto *output = ctx.Output("Out"); + + output->mutable_data(ctx.GetPlace()); - Y->mutable_data(ctx.GetPlace()); - Gather(ctx.GetPlace(), X, Index, Y); + CPUGather(ctx.device_context(), *x, *index, output); } }; -template -class GatherGradientOpKernel : public framework::OpKernel { +template +class GatherGradientOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "This kernel only runs on CPU."); + auto *Index = ctx.Input("Index"); auto *dX = ctx.Output(framework::GradVarName("X")); auto *dO = ctx.Input(framework::GradVarName("Out")); dX->mutable_data(ctx.GetPlace()); - ScatterUpdate(ctx.GetPlace(), dO, Index, dX); + auto dxt = framework::EigenVector::Flatten(*dX); + auto place = ctx.GetEigenDevice(); + dxt.device(place) = dxt.constant(static_cast(0)); + + ScatterAssign(ctx.device_context(), *dO, *Index, dX); } }; diff --git a/paddle/operators/gather_test.cc b/paddle/operators/gather_test.cc index 0ae1e99452..cbd86b8796 100644 --- a/paddle/operators/gather_test.cc +++ b/paddle/operators/gather_test.cc @@ -41,7 +41,9 @@ TEST(Gather, GatherData) { int* p_output = output->mutable_data(make_ddim({2, 4}), CPUPlace()); - Gather(CPUPlace(), src, index, output); + auto* cpu_place = new paddle::platform::CPUPlace(); + paddle::platform::CPUDeviceContext ctx(*cpu_place); + CPUGather(ctx, *src, *index, output); for (int i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4); for (int i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4); diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc index c0e161bbc0..5cd2c7d2c0 100644 --- a/paddle/operators/gaussian_random_op.cc +++ b/paddle/operators/gaussian_random_op.cc @@ -16,7 +16,7 @@ namespace paddle { namespace operators { template -class CPUGaussianRandomKernel : public framework::OpKernel { +class CPUGaussianRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { float mean = context.Attr("mean"); @@ -43,13 +43,10 @@ class GaussianRandomOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Out"), - "Output(Out) of GaussianRandomOp should not be null."); - - auto* tensor = ctx.Output("Out"); - auto dims = Attr>("dims"); + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of GaussianRandomOp should not be null."); + auto dims = ctx->Attrs().Get>("dims"); std::vector temp; temp.reserve(dims.size()); for (auto dim : dims) { @@ -57,7 +54,12 @@ class GaussianRandomOp : public framework::OperatorWithKernel { } PADDLE_ENFORCE(dims.size() > 0UL, "dims can be one int or array. dims must be set."); - tensor->Resize(framework::make_ddim(temp)); + ctx->SetOutputDim("Out", framework::make_ddim(temp)); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return static_cast(Attr("data_type")); } }; @@ -79,6 +81,8 @@ Use to initialize tensor with gaussian random generator. "Random seed of generator." "0 means use system wide seed") .SetDefault(0); + AddAttr("data_type", "output data type") + .SetDefault(framework::DataType::FP32); } }; diff --git a/paddle/operators/gaussian_random_op.cu b/paddle/operators/gaussian_random_op.cu index 2d63b30499..315560bf1b 100644 --- a/paddle/operators/gaussian_random_op.cu +++ b/paddle/operators/gaussian_random_op.cu @@ -37,7 +37,7 @@ struct GaussianGenerator { }; template -class GPUGaussianRandomKernel : public framework::OpKernel { +class GPUGaussianRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* tensor = context.Output("Out"); diff --git a/paddle/operators/gemm_conv2d_op.h b/paddle/operators/gemm_conv2d_op.h new file mode 100644 index 0000000000..323e3f7c3b --- /dev/null +++ b/paddle/operators/gemm_conv2d_op.h @@ -0,0 +1,226 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/math/im2col.h" +#include "paddle/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class GemmConv2DKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* input = context.Input("Input"); + // The filter will be reshaped in the calculations, + // so here use an assignment operation, + // that avoids modifying the variable in the Scope. + Tensor filter = *context.Input("Filter"); + Tensor* output = context.Output("Output"); + output->mutable_data(context.GetPlace()); + + std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); + int groups = context.Attr("groups"); + + int batch_size = input->dims()[0]; + int input_channels = input->dims()[1]; + int filter_height = filter.dims()[filter.dims().size() - 2]; + int filter_width = filter.dims()[filter.dims().size() - 1]; + int output_channels = output->dims()[1]; + int output_height = output->dims()[2]; + int output_width = output->dims()[3]; + + paddle::operators::math::Im2ColFunctor< + paddle::operators::math::ColFormat::kCFO, Place, T> + im2col; + // use col_shape in the im2col calculation + framework::DDim col_shape = {input_channels / groups, filter_height, + filter_width, output_height, output_width}; + // use col_matrix_shape in the gemm calculation + framework::DDim col_matrix_shape = { + input_channels / groups * filter_height * filter_width, + output_height * output_width}; + Tensor col; + col.mutable_data(col_shape, context.GetPlace()); + // col_matrix shares the same piece of data with col, + // but will be reshaped into a two-dimensional matrix shape + // to call the matrix multiplication interface. + Tensor col_matrix = col; + col_matrix.Resize(col_matrix_shape); + + framework::DDim input_shape = {input->dims()[1], input->dims()[2], + input->dims()[3]}; + framework::DDim filter_matrix_shape = {filter.dims()[0], + filter.numel() / filter.dims()[0]}; + filter.Resize(filter_matrix_shape); + + framework::DDim output_matrix_shape = {output_channels, + output_height * output_width}; + + // convolution operator: im2col + gemm + int in_step = input_channels / groups; + int out_step = output_channels / groups; + for (int i = 0; i < batch_size; i++) { + Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape); + Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape); + for (int g = 0; g < groups; g++) { + // im2col + Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step); + im2col(context.device_context(), in_slice, col, strides[0], strides[1], + paddings[0], paddings[1]); + + // gemm + Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step); + Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step); + math::matmul(context.device_context(), filter_slice, false, + col_matrix, false, T(1.0), &out_slice, T(0.0)); + } + } + } +}; + +template +class GemmConvGrad2DKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* input = context.Input("Input"); + const Tensor* output_grad = + context.Input(framework::GradVarName("Output")); + Tensor* input_grad = + context.Output(framework::GradVarName("Input")); + Tensor* filter_grad = + context.Output(framework::GradVarName("Filter")); + + // The filter and filter_grad will be reshaped in the calculations, + // so here use an assignment operation, + // that avoids modifying the variable in the Scope. + Tensor filter = *context.Input("Filter"); + + std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); + int groups = context.Attr("groups"); + + int batch_size = input->dims()[0]; + int input_channels = input->dims()[1]; + int filter_height = filter.dims()[filter.dims().size() - 2]; + int filter_width = filter.dims()[filter.dims().size() - 1]; + int output_channels = output_grad->dims()[1]; + int output_height = output_grad->dims()[2]; + int output_width = output_grad->dims()[3]; + + paddle::operators::math::Col2ImFunctor< + paddle::operators::math::ColFormat::kCFO, Place, T> + col2im; + paddle::operators::math::Im2ColFunctor< + paddle::operators::math::ColFormat::kCFO, Place, T> + im2col; + // use col_shape in the im2col and col2im calculation + framework::DDim col_shape = {input_channels / groups, filter_height, + filter_width, output_height, output_width}; + // use col_matrix_shape in the gemm calculation + framework::DDim col_matrix_shape = { + input_channels / groups * filter_height * filter_width, + output_height * output_width}; + Tensor col; + col.mutable_data(col_shape, context.GetPlace()); + // col_matrix shares the same piece of data with col, + // but will be reshaped into a two-dimensional matrix shape + // to call the matrix multiplication interface. + Tensor col_matrix = col; + col_matrix.Resize(col_matrix_shape); + + framework::DDim input_shape = {input->dims()[1], input->dims()[2], + input->dims()[3]}; + framework::DDim output_matrix_shape = { + output_grad->dims()[1], + output_grad->dims()[2] * output_grad->dims()[3]}; + + framework::DDim filter_matrix_shape = {filter.dims()[0], + filter.numel() / filter.dims()[0]}; + filter.Resize(filter_matrix_shape); + + // convolution backward input operator: gemm + col2im + // convolution backward weight operator: im2col + gemm + int in_step = input_channels / groups; + int out_step = output_channels / groups; + + if (input_grad) { + input_grad->mutable_data(context.GetPlace()); + auto t = framework::EigenVector::Flatten(*input_grad); + t.device(context.GetEigenDevice()) = t.constant(static_cast(0)); + + for (int i = 0; i < batch_size; i++) { + Tensor out_grad_batch = + output_grad->Slice(i, i + 1).Resize(output_matrix_shape); + Tensor in_grad_batch = + input_grad->Slice(i, i + 1).Resize(input_shape); + for (int g = 0; g < groups; g++) { + // gemm + Tensor out_grad_slice = + out_grad_batch.Slice(g * out_step, (g + 1) * out_step); + Tensor filter_slice = + filter.Slice(g * out_step, (g + 1) * out_step); + math::matmul(context.device_context(), filter_slice, true, + out_grad_slice, false, T(1.0), &col_matrix, + T(0.0)); + + // col2im + Tensor in_grad_slice = + in_grad_batch.Slice(g * in_step, (g + 1) * in_step); + col2im(context.device_context(), in_grad_slice, col, strides[0], + strides[1], paddings[0], paddings[1]); + } + } + } + + if (filter_grad) { + filter_grad->mutable_data(context.GetPlace()); + Tensor filter_grad_ = *filter_grad; + filter_grad_.Resize(filter_matrix_shape); + auto t = framework::EigenVector::Flatten(filter_grad_); + t.device(context.GetEigenDevice()) = t.constant(static_cast(0)); + + for (int i = 0; i < batch_size; i++) { + Tensor out_grad_batch = + output_grad->Slice(i, i + 1).Resize(output_matrix_shape); + Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape); + for (int g = 0; g < groups; g++) { + // im2col + Tensor out_grad_slice = + out_grad_batch.Slice(g * out_step, (g + 1) * out_step); + Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step); + im2col(context.device_context(), in_slice, col, strides[0], + strides[1], paddings[0], paddings[1]); + + // gemm + Tensor filter_grad_slice = + filter_grad_.Slice(g * out_step, (g + 1) * out_step); + math::matmul(context.device_context(), out_grad_slice, + false, col_matrix, true, T(1.0), + &filter_grad_slice, T(1.0)); + } + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc index 07f6dfabca..929008fbcb 100644 --- a/paddle/operators/lookup_table_op.cc +++ b/paddle/operators/lookup_table_op.cc @@ -22,26 +22,31 @@ class LookupTableOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("W"), - "Input(W) of LookupTableOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Ids"), - "Input(Ids) of LookupTableOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of LookupTableOp should not be null."); - - auto table_t = ctx.Input("W"); - auto ids_t = ctx.Input("Ids"); - auto output_t = ctx.Output("Out"); - - output_t->Resize({ids_t->dims()[0], table_t->dims()[1]}); + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("W"), + "Input(W) of LookupTableOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Ids"), + "Input(Ids) of LookupTableOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of LookupTableOp should not be null."); + + auto table_dims = ctx->GetInputDim("W"); + auto ids_dims = ctx->GetInputDim("Ids"); + + ctx->SetOutputDim("Out", {ids_dims[0], table_dims[1]}); + ctx->ShareLoD("Ids", /*->*/ "Out"); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("W")->type()); } }; class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { public: - LookupTableOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + LookupTableOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("W", "An input represents embedding tensors," @@ -50,9 +55,13 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { "An input with type int32 or int64" "contains the ids to be looked up in W."); AddOutput("Out", "The lookup results, which have the same type with W."); - AddComment( - "This operator is used to perform lookups on the parameter W," - "then concatenated into a dense tensor."); + AddComment(R"DOC( +This operator is used to perform lookups on the parameter W, +then concatenated into a dense tensor. + +The input `Ids` can carry the LoD (Level of Details) information, +or not. And the output only shares the LoD with input `Ids`. +)DOC"); } }; @@ -61,11 +70,14 @@ class LookupTableOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &context) const override { - auto table = context.Input("W"); - auto d_table = - context.Output(framework::GradVarName("W")); - d_table->Resize(table->dims()); + void InferShape(framework::InferShapeContextBase* ctx) const override { + auto table_dims = ctx->GetInputDim("W"); + ctx->SetOutputDim(framework::GradVarName("W"), table_dims); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("W")->type()); } }; diff --git a/paddle/operators/lookup_table_op.cu b/paddle/operators/lookup_table_op.cu index 7083440467..c3808fa9a8 100644 --- a/paddle/operators/lookup_table_op.cu +++ b/paddle/operators/lookup_table_op.cu @@ -61,7 +61,7 @@ __global__ void LookupTableGrad(T* table, const T* output, const int32_t* ids, } template -class LookupTableCUDAKernel : public framework::OpKernel { +class LookupTableCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto table_t = context.Input("W"); @@ -77,12 +77,15 @@ class LookupTableCUDAKernel : public framework::OpKernel { dim3 threads(128, 8); dim3 grids(8, 1); - LookupTable<<>>(output, table, ids, N, K, D); + LookupTable<<< + grids, threads, 0, reinterpret_cast( + context.device_context()) + .stream()>>>(output, table, ids, N, K, D); } }; template -class LookupTableGradCUDAKernel : public framework::OpKernel { +class LookupTableGradCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto ids_t = context.Input("Ids"); @@ -102,8 +105,10 @@ class LookupTableGradCUDAKernel : public framework::OpKernel { dim3 threads(128, 8); dim3 grids(8, 1); - LookupTableGrad<<>>(d_table, d_output, ids, N, - K, D); + LookupTableGrad<<< + grids, threads, 0, reinterpret_cast( + context.device_context()) + .stream()>>>(d_table, d_output, ids, N, K, D); } }; diff --git a/paddle/operators/lookup_table_op.h b/paddle/operators/lookup_table_op.h index a1298906dd..dfead2fc5b 100644 --- a/paddle/operators/lookup_table_op.h +++ b/paddle/operators/lookup_table_op.h @@ -23,7 +23,7 @@ namespace operators { using Tensor = framework::Tensor; template -class LookupTableKernel : public framework::OpKernel { +class LookupTableKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto table_t = context.Input("W"); // float tensor @@ -44,7 +44,7 @@ class LookupTableKernel : public framework::OpKernel { }; template -class LookupTableGradKernel : public framework::OpKernel { +class LookupTableGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto ids_t = context.Input("Ids"); diff --git a/paddle/operators/lstm_unit_op.cc b/paddle/operators/lstm_unit_op.cc new file mode 100644 index 0000000000..dad56731de --- /dev/null +++ b/paddle/operators/lstm_unit_op.cc @@ -0,0 +1,102 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/lstm_unit_op.h" + +namespace paddle { +namespace operators { + +class LstmUnitOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasInput("C_prev"), + "Input(C_prev) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("C"), + "Output(C) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("H"), + "Output(H) of LSTM should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + auto c_prev_dims = ctx->GetInputDim("C_prev"); + + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank must be 2."); + PADDLE_ENFORCE(x_dims[0] == c_prev_dims[0], + "Batch size of inputs and states must be equal"); + PADDLE_ENFORCE(x_dims[1] == c_prev_dims[1] * 4, + "Dimension of FC should equal to prev state * 4"); + + int b_size = c_prev_dims[0]; // batch size + int s_dim = c_prev_dims[1]; // state dim + ctx->SetOutputDim("C", {b_size, s_dim}); + ctx->SetOutputDim("H", {b_size, s_dim}); + } +}; + +class LstmUnitOpMaker : public framework::OpProtoAndCheckerMaker { + public: + LstmUnitOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "FC input before the non-linear activation."); + AddInput( + "C_prev", + "The cell state tensor of last time-step in the Lstm Unit operator."); + AddOutput("C", "The cell tensor of Lstm Unit operator."); + AddOutput("H", "The hidden state tensor of Lstm Unit operator."); + + AddComment(R"DOC(Lstm-Unit Operator + +Equation: + i, f, o, j = split(X) + C = C_prev * sigm(f + forget_bias) + sigm(i) * tanh(j) + H = C * sigm(o) + +)DOC"); + AddAttr("forget_bias", "The forget bias of Lstm Unit.") + .SetDefault(0.0); + } +}; + +class LstmUnitGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("C")), + "Input(C@GRAD) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("H")), + "Input(H@GRAD) should not be null"); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + ctx->SetOutputDim(framework::GradVarName("C_prev"), + ctx->GetInputDim("C_prev")); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(lstm_unit, ops::LstmUnitOp, ops::LstmUnitOpMaker, lstm_unit_grad, + ops::LstmUnitGradOp); +REGISTER_OP_CPU_KERNEL(lstm_unit, + ops::LstmUnitKernel, + ops::LstmUnitKernel); +REGISTER_OP_CPU_KERNEL( + lstm_unit_grad, ops::LstmUnitGradKernel, + ops::LstmUnitGradKernel); diff --git a/paddle/operators/lstm_unit_op.cu b/paddle/operators/lstm_unit_op.cu new file mode 100644 index 0000000000..49ea550b6f --- /dev/null +++ b/paddle/operators/lstm_unit_op.cu @@ -0,0 +1,175 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/framework/op_registry.h" +#include "paddle/operators/cross_entropy_op.h" +#include "paddle/platform/assert.h" +#include "paddle/platform/hostdevice.h" + +namespace paddle { +namespace operators { + +#define CUDA_1D_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \ + i += blockDim.x * gridDim.x) + +template +__device__ Dtype cuda_sigmoid(const Dtype x) { + return Dtype(1) / (Dtype(1) + exp(-x)); +} + +template +__device__ Dtype cuda_tanh(const Dtype x) { + return Dtype(1 - exp(-2. * x)) / (Dtype(1) + exp(-2. * x)); +} + +template +__global__ void LSTMUnitKernel(const int nthreads, const int dim, + const T* C_prev, const T* X, T* C, T* H, + const T forget_bias) { + CUDA_1D_KERNEL_LOOP(index, nthreads) { + const int n = index / dim; + const int d = index % dim; + + const T* X_offset = X + 4 * dim * n; + const T i = cuda_sigmoid(X_offset[d]); + const T f = cuda_sigmoid(X_offset[1 * dim + d] + forget_bias); + const T o = cuda_sigmoid(X_offset[2 * dim + d]); + const T g = cuda_tanh(X_offset[3 * dim + d]); + const T c_prev = C_prev[index]; + const T c = f * c_prev + i * g; + C[index] = c; + const T tanh_c = cuda_tanh(c); + H[index] = o * tanh_c; + } +} + +template +__global__ void LSTMUnitGradientKernel(const int nthreads, const int dim, + const T* C_prev, const T* X, const T* C, + const T* H, const T* C_diff, + const T* H_diff, T* C_prev_diff, + T* X_diff, const T forget_bias) { + CUDA_1D_KERNEL_LOOP(index, nthreads) { + const int n = index / dim; + const int d = index % dim; + const T* X_offset = X + 4 * dim * n; + T* c_prev_diff = C_prev_diff + index; + T* X_diff_offset = X_diff + 4 * dim * n; + T* i_diff = X_diff_offset + d; + T* f_diff = X_diff_offset + 1 * dim + d; + T* o_diff = X_diff_offset + 2 * dim + d; + T* g_diff = X_diff_offset + 3 * dim + d; + + const T i = cuda_sigmoid(X_offset[d]); + const T f = cuda_sigmoid(X_offset[1 * dim + d] + forget_bias); + const T o = cuda_sigmoid(X_offset[2 * dim + d]); + const T g = cuda_tanh(X_offset[3 * dim + d]); + const T c_prev = C_prev[index]; + const T c = C[index]; + const T tanh_c = cuda_tanh(c); + const T c_term_diff = + C_diff[index] + H_diff[index] * o * (1 - tanh_c * tanh_c); + *c_prev_diff = c_term_diff * f; + *i_diff = c_term_diff * g * i * (1 - i); + *f_diff = c_term_diff * c_prev * f * (1 - f); + *o_diff = H_diff[index] * tanh_c * o * (1 - o); + *g_diff = c_term_diff * i * (1 - g * g); + } +} + +template +class LstmUnitOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use GPUPlace."); + + auto* x_tensor = ctx.Input("X"); + auto* c_prev_tensor = ctx.Input("C_prev"); + auto* c_tensor = ctx.Output("C"); + auto* h_tensor = ctx.Output("H"); + + auto forget_bias = static_cast(ctx.Attr("forget_bias")); + + int b_size = c_tensor->dims()[0]; + int D = c_tensor->dims()[1]; + + const T* X = x_tensor->data(); + const T* C_prev = c_prev_tensor->data(); + + T* C = c_tensor->mutable_data(ctx.GetPlace()); + T* H = h_tensor->mutable_data(ctx.GetPlace()); + + int block = 512; + int n = b_size * D; + int grid = (n + block - 1) / block; + + LSTMUnitKernel<<>>(n, D, C_prev, X, C, H, forget_bias); + } +}; + +template +class LstmUnitGradOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use GPUPlace."); + + auto x_tensor = ctx.Input("X"); + auto c_prev_tensor = ctx.Input("C_prev"); + auto c_tensor = ctx.Input("C"); + auto h_tensor = ctx.Input("H"); + + auto hdiff_tensor = ctx.Input(framework::GradVarName("H")); + auto cdiff_tensor = ctx.Input(framework::GradVarName("C")); + + auto xdiff_tensor = ctx.Output(framework::GradVarName("X")); + auto c_prev_diff_tensor = + ctx.Output(framework::GradVarName("C_prev")); + + auto* X = x_tensor->data(); + auto* C_prev = c_prev_tensor->data(); + auto* C = c_tensor->data(); + auto* H = h_tensor->data(); + + auto* H_diff = hdiff_tensor->data(); + auto* C_diff = cdiff_tensor->data(); + + auto* C_prev_diff = c_prev_diff_tensor->mutable_data(ctx.GetPlace()); + auto* X_diff = xdiff_tensor->mutable_data(ctx.GetPlace()); + + int N = c_tensor->dims()[0]; + int D = c_tensor->dims()[1]; + + auto forget_bias = static_cast(ctx.Attr("forget_bias")); + + int block = 512; + int n = N * D; + int grid = (n + block - 1) / block; + + LSTMUnitGradientKernel<<>>(n, D, C_prev, X, C, H, C_diff, + H_diff, C_prev_diff, X_diff, + forget_bias); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(lstm_unit, ops::LstmUnitOpCUDAKernel, + ops::LstmUnitOpCUDAKernel); +REGISTER_OP_GPU_KERNEL(lstm_unit_grad, ops::LstmUnitGradOpCUDAKernel, + ops::LstmUnitGradOpCUDAKernel); diff --git a/paddle/operators/lstm_unit_op.h b/paddle/operators/lstm_unit_op.h new file mode 100644 index 0000000000..a0ff498c1d --- /dev/null +++ b/paddle/operators/lstm_unit_op.h @@ -0,0 +1,148 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT 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 "glog/logging.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using framework::LoDTensor; +using framework::Tensor; + +template +inline T sigmoid(T x) { + return 1. / (1. + exp(-x)); +} + +template +inline T tanh(T x) { + return 2. * sigmoid(2. * x) - 1.; +} + +template +class LstmUnitKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + + auto* x_tensor = ctx.Input("X"); + auto* c_prev_tensor = ctx.Input("C_prev"); + auto* c_tensor = ctx.Output("C"); + auto* h_tensor = ctx.Output("H"); + + auto forget_bias = static_cast(ctx.Attr("forget_bias")); + + int b_size = c_tensor->dims()[0]; + int D = c_tensor->dims()[1]; + + T* C = c_tensor->mutable_data(ctx.GetPlace()); + T* H = h_tensor->mutable_data(ctx.GetPlace()); + + const T* X = x_tensor->data(); + const T* C_prev = c_prev_tensor->data(); + + for (int n = 0; n < b_size; ++n) { + for (int d = 0; d < D; ++d) { + const T i = sigmoid(X[d]); + const T f = sigmoid(X[1 * D + d] + forget_bias); + const T o = sigmoid(X[2 * D + d]); + const T g = tanh(X[3 * D + d]); + const T c_prev = C_prev[d]; + const T c = f * c_prev + i * g; + C[d] = c; + const T tanh_c = tanh(c); + H[d] = o * tanh_c; + } + C_prev += D; + X += 4 * D; + C += D; + H += D; + } + } +}; + +template +class LstmUnitGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + + auto x_tensor = ctx.Input("X"); + auto c_prev_tensor = ctx.Input("C_prev"); + auto c_tensor = ctx.Input("C"); + auto h_tensor = ctx.Input("H"); + + auto hdiff_tensor = ctx.Input(framework::GradVarName("H")); + auto cdiff_tensor = ctx.Input(framework::GradVarName("C")); + + auto xdiff_tensor = ctx.Output(framework::GradVarName("X")); + auto c_prev_diff_tensor = + ctx.Output(framework::GradVarName("C_prev")); + + auto* X = x_tensor->data(); + auto* C_prev = c_prev_tensor->data(); + auto* C = c_tensor->data(); + auto* H = h_tensor->data(); + + auto* H_diff = hdiff_tensor->data(); + auto* C_diff = cdiff_tensor->data(); + + auto* C_prev_diff = c_prev_diff_tensor->mutable_data(ctx.GetPlace()); + auto* X_diff = xdiff_tensor->mutable_data(ctx.GetPlace()); + + int N = c_tensor->dims()[0]; + int D = c_tensor->dims()[1]; + + auto forget_bias = static_cast(ctx.Attr("forget_bias")); + + for (int n = 0; n < N; ++n) { + for (int d = 0; d < D; ++d) { + T* c_prev_diff = C_prev_diff + d; + T* i_diff = X_diff + d; + T* f_diff = X_diff + 1 * D + d; + T* o_diff = X_diff + 2 * D + d; + T* g_diff = X_diff + 3 * D + d; + + const T i = sigmoid(X[d]); + const T f = sigmoid(X[1 * D + d] + forget_bias); + const T o = sigmoid(X[2 * D + d]); + const T g = tanh(X[3 * D + d]); + const T c_prev = C_prev[d]; + const T c = C[d]; + const T tanh_c = tanh(c); + const T c_term_diff = C_diff[d] + H_diff[d] * o * (1 - tanh_c * tanh_c); + *c_prev_diff = c_term_diff * f; + *i_diff = c_term_diff * g * i * (1 - i); + *f_diff = c_term_diff * c_prev * f * (1 - f); + *o_diff = H_diff[d] * tanh_c * o * (1 - o); + *g_diff = c_term_diff * i * (1 - g * g); + } + C_prev += D; + X += 4 * D; + C += D; + H += D; + C_diff += D; + H_diff += D; + X_diff += 4 * D; + C_prev_diff += D; + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/CMakeLists.txt b/paddle/operators/math/CMakeLists.txt index f8333f34f7..a0ceb029e3 100644 --- a/paddle/operators/math/CMakeLists.txt +++ b/paddle/operators/math/CMakeLists.txt @@ -1,10 +1,13 @@ - if(WITH_GPU) - nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc - im2col.cu DEPS cblas device_context) + nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc im2col.cu pooling.cc pooling.cu DEPS cblas device_context operator) + nv_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor) + nv_library(softmax SRCS softmax.cc softmax.cu DEPS operator) + nv_library(cross_entropy SRCS cross_entropy.cc cross_entropy.cu DEPS operator) else() - cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context) + cc_library(math_function SRCS math_function.cc im2col.cc pooling.cc DEPS cblas device_context operator) + cc_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor) + cc_library(softmax SRCS softmax.cc DEPS operator) + cc_library(cross_entropy SRCS cross_entropy.cc DEPS operator) endif() -nv_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor) cc_test(im2col_test SRCS im2col_test.cc DEPS math_function tensor) diff --git a/paddle/operators/math/cross_entropy.cc b/paddle/operators/math/cross_entropy.cc new file mode 100644 index 0000000000..150a65f275 --- /dev/null +++ b/paddle/operators/math/cross_entropy.cc @@ -0,0 +1,59 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/math/cross_entropy.h" + +namespace paddle { +namespace operators { +namespace math { + +using Tensor = framework::Tensor; +template +using EigenMatrix = framework::EigenMatrix; + +template +class CrossEntropyFunctor { + public: + void operator()(const platform::DeviceContext& ctx, framework::Tensor* out, + const framework::Tensor* prob, + const framework::Tensor* labels, const bool softLabel) { + const int batch_size = prob->dims()[0]; + if (softLabel) { + auto in = EigenMatrix::From(*prob); + auto lbl = EigenMatrix::From(*labels); + auto loss = EigenMatrix::From(*out); + + loss.device(*ctx.GetEigenDevice()) = + -((lbl * in.log().unaryExpr(math::TolerableValue())) + .sum(Eigen::DSizes(1)) + .reshape(Eigen::DSizes(batch_size, 1))); + } else { + const int class_num = prob->dims()[1]; + const T* prob_data = prob->data(); + T* loss_data = out->data(); + + const int* label_data = labels->data(); + for (int i = 0; i < batch_size; ++i) { + int index = i * class_num + label_data[i]; + loss_data[i] = -math::TolerableValue()(std::log(prob_data[index])); + } + } + } +}; + +template class CrossEntropyFunctor; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/cross_entropy.cu b/paddle/operators/math/cross_entropy.cu new file mode 100644 index 0000000000..367190e6b0 --- /dev/null +++ b/paddle/operators/math/cross_entropy.cu @@ -0,0 +1,109 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/math/cross_entropy.h" + +namespace paddle { +namespace operators { +namespace math { + +namespace { +template +__global__ void CrossEntropyKernel(T* Y, const T* X, const int* label, + const int N, const int D) { + // TOOD(qingqing) define CUDA_1D_KERNEL_LOOP macro in a common file. + // CUDA_1D_KERNEL_LOOP(i, N) { + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; + i += blockDim.x * gridDim.x) { + PADDLE_ASSERT(label[i] >= 0 && label[i] < D); + Y[i] = -math::TolerableValue()(log(X[i * D + label[i]])); + } +} + +template +__device__ __forceinline__ T sum_single_warp(T val) { + val += __shfl_down(val, 16); + val += __shfl_down(val, 8); + val += __shfl_down(val, 4); + val += __shfl_down(val, 2); + val += __shfl_down(val, 1); + return val; +} + +template +__global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, + const int class_num) { + int tid = threadIdx.x; + extern __shared__ T d_sum[]; + d_sum[tid] = 0; + + int cur_idx = tid; + int next_idx = blockIdx.x * class_num + tid; + while (cur_idx < class_num) { + d_sum[tid] += + math::TolerableValue()(std::log(X[next_idx])) * label[next_idx]; + next_idx += blockDim.x; + cur_idx += blockDim.x; + } + __syncthreads(); + + for (unsigned int stride = blockDim.x >> 1; stride >= 32; stride >>= 1) { + if (tid < stride) d_sum[tid] += d_sum[tid + stride]; + __syncthreads(); + } + + T val = d_sum[tid]; + val = sum_single_warp(val); + if (tid == 0) Y[blockIdx.x] = -val; +} +} // namespace + +using Tensor = framework::Tensor; + +template +class CrossEntropyFunctor { + public: + void operator()(const platform::DeviceContext& ctx, framework::Tensor* out, + const framework::Tensor* prob, + const framework::Tensor* labels, bool softLabel) { + const T* prob_data = prob->data(); + T* loss_data = out->mutable_data(ctx.GetPlace()); + + int batch_size = prob->dims()[0]; + int class_num = prob->dims()[1]; + + if (softLabel) { + const T* label_data = labels->data(); + int block = class_num > 512 ? 512 : pow(2, int(std::log2(class_num))); + + SoftCrossEntropyKernel<<< + batch_size, block, block * sizeof(T), + reinterpret_cast(ctx).stream()>>>( + loss_data, prob_data, label_data, class_num); + } else { + const int* label_data = labels->data(); + int block = 512; + int grid = (batch_size + block - 1) / block; + CrossEntropyKernel<<< + grid, block, 0, + reinterpret_cast(ctx).stream()>>>( + loss_data, prob_data, label_data, batch_size, class_num); + } + } +}; + +template class CrossEntropyFunctor; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/cross_entropy.h b/paddle/operators/math/cross_entropy.h new file mode 100644 index 0000000000..0ab6827ffa --- /dev/null +++ b/paddle/operators/math/cross_entropy.h @@ -0,0 +1,46 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/operator.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/hostdevice.h" + +namespace paddle { +namespace operators { +namespace math { + +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; + } +}; + +template +class CrossEntropyFunctor { + public: + void operator()(const platform::DeviceContext& context, + framework::Tensor* out, const framework::Tensor* prob, + const framework::Tensor* labels, const bool softLabel); +}; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/im2col.cc b/paddle/operators/math/im2col.cc index 5727c1cab1..c08a3380f0 100644 --- a/paddle/operators/math/im2col.cc +++ b/paddle/operators/math/im2col.cc @@ -27,9 +27,10 @@ template class Im2ColFunctor { public: - void operator()(const framework::Tensor& im, framework::Tensor& col, + void operator()(const platform::DeviceContext& context, + const framework::Tensor& im, framework::Tensor& col, int stride_height, int stride_width, int padding_height, - int padding_width, platform::DeviceContext* context) { + int padding_width) { PADDLE_ENFORCE(im.dims().size() == 3); PADDLE_ENFORCE(col.dims().size() == 5); @@ -79,9 +80,9 @@ template class Col2ImFunctor { public: - void operator()(framework::Tensor& im, const framework::Tensor& col, - int stride_height, int stride_width, int padding_height, - int padding_width, platform::DeviceContext* context) { + void operator()(const platform::DeviceContext& context, framework::Tensor& im, + const framework::Tensor& col, int stride_height, + int stride_width, int padding_height, int padding_width) { PADDLE_ENFORCE(im.dims().size() == 3); PADDLE_ENFORCE(col.dims().size() == 5); int input_channels = im.dims()[0]; @@ -137,9 +138,10 @@ template class Im2ColFunctor { public: - void operator()(const framework::Tensor& im, framework::Tensor& col, + void operator()(const platform::DeviceContext& context, + const framework::Tensor& im, framework::Tensor& col, int stride_height, int stride_width, int padding_height, - int padding_width, platform::DeviceContext* context) { + int padding_width) { PADDLE_ENFORCE(im.dims().size() == 3); PADDLE_ENFORCE(col.dims().size() == 5); int input_channels = im.dims()[0]; @@ -197,9 +199,9 @@ template class Col2ImFunctor { public: - void operator()(framework::Tensor& im, const framework::Tensor& col, - int stride_height, int stride_width, int padding_height, - int padding_width, platform::DeviceContext* context) { + void operator()(const platform::DeviceContext& context, framework::Tensor& im, + const framework::Tensor& col, int stride_height, + int stride_width, int padding_height, int padding_width) { PADDLE_ENFORCE(im.dims().size() == 3); PADDLE_ENFORCE(col.dims().size() == 5); int input_channels = im.dims()[0]; diff --git a/paddle/operators/math/im2col.cu b/paddle/operators/math/im2col.cu index 9bff7bee3c..01f60bfe70 100644 --- a/paddle/operators/math/im2col.cu +++ b/paddle/operators/math/im2col.cu @@ -64,9 +64,10 @@ template class Im2ColFunctor { public: - void operator()(const framework::Tensor& im, framework::Tensor& col, + void operator()(const platform::DeviceContext& context, + const framework::Tensor& im, framework::Tensor& col, int stride_height, int stride_width, int padding_height, - int padding_width, platform::DeviceContext* context) { + int padding_width) { PADDLE_ENFORCE(im.dims().size() == 3); PADDLE_ENFORCE(col.dims().size() == 5); @@ -84,9 +85,9 @@ class Im2ColFunctor<<< - grid, threads, 0, - reinterpret_cast(context)->stream()>>>( + im2col<<(context) + .stream()>>>( im.data(), num_outputs, input_height, input_width, filter_height, filter_width, stride_height, stride_width, padding_height, padding_width, output_height, output_width, col.data()); @@ -149,9 +150,9 @@ template class Col2ImFunctor { public: - void operator()(framework::Tensor& im, const framework::Tensor& col, - int stride_height, int stride_width, int padding_height, - int padding_width, platform::DeviceContext* context) { + void operator()(const platform::DeviceContext& context, framework::Tensor& im, + const framework::Tensor& col, int stride_height, + int stride_width, int padding_height, int padding_width) { PADDLE_ENFORCE(im.dims().size() == 3); PADDLE_ENFORCE(col.dims().size() == 5); @@ -174,9 +175,9 @@ class Col2ImFunctor<<< - grid, threads, 0, - reinterpret_cast(context)->stream()>>>( + col2im<<(context) + .stream()>>>( num_kernels, col.data(), input_height + 2 * padding_height, input_width + 2 * padding_width, input_channels, filter_height, filter_width, stride_height, stride_width, padding_height, @@ -235,9 +236,10 @@ template class Im2ColFunctor { public: - void operator()(const framework::Tensor& im, framework::Tensor& col, + void operator()(const platform::DeviceContext& context, + const framework::Tensor& im, framework::Tensor& col, int stride_height, int stride_width, int padding_height, - int padding_width, platform::DeviceContext* context) { + int padding_width) { PADDLE_ENFORCE(im.dims().size() == 3); PADDLE_ENFORCE(col.dims().size() == 5); int input_channels = im.dims()[0]; @@ -268,9 +270,9 @@ class Im2ColFunctor<<< - grid, threads, 0, - reinterpret_cast(context)->stream()>>>( + im2colOCF<<(context) + .stream()>>>( im.data(), col.data(), input_channels, input_height, input_width, filter_height, filter_width, stride_height, stride_width, padding_height, padding_width, output_height, output_width); @@ -318,9 +320,9 @@ template class Col2ImFunctor { public: - void operator()(framework::Tensor& im, const framework::Tensor& col, - int stride_height, int stride_width, int padding_height, - int padding_width, platform::DeviceContext* context) { + void operator()(const platform::DeviceContext& context, framework::Tensor& im, + const framework::Tensor& col, int stride_height, + int stride_width, int padding_height, int padding_width) { PADDLE_ENFORCE(im.dims().size() == 3); PADDLE_ENFORCE(col.dims().size() == 5); int input_channels = im.dims()[0]; @@ -351,9 +353,9 @@ class Col2ImFunctor<<< - grid, threads, 0, - reinterpret_cast(context)->stream()>>>( + col2imOCF<<(context) + .stream()>>>( im.data(), col.data(), input_channels, input_height, input_width, filter_height, filter_width, stride_height, stride_width, padding_height, padding_width, output_height, output_width); diff --git a/paddle/operators/math/im2col.h b/paddle/operators/math/im2col.h index 8958c5457c..7b717e1603 100644 --- a/paddle/operators/math/im2col.h +++ b/paddle/operators/math/im2col.h @@ -72,17 +72,18 @@ enum class ColFormat { kCFO = 0, kOCF = 1 }; template class Im2ColFunctor { public: - void operator()(const framework::Tensor& im, framework::Tensor& col, + void operator()(const platform::DeviceContext& context, + const framework::Tensor& im, framework::Tensor& col, int stride_height, int stride_width, int padding_height, - int padding_width, platform::DeviceContext* context); + int padding_width); }; template class Col2ImFunctor { public: - void operator()(framework::Tensor& im, const framework::Tensor& col, - int stride_height, int stride_width, int padding_height, - int padding_width, platform::DeviceContext* context); + void operator()(const platform::DeviceContext& context, framework::Tensor& im, + const framework::Tensor& col, int stride_height, + int stride_width, int padding_height, int padding_width); }; } // namespace math diff --git a/paddle/operators/math/im2col_test.cc b/paddle/operators/math/im2col_test.cc index 4f380388b1..40bdbfe733 100644 --- a/paddle/operators/math/im2col_test.cc +++ b/paddle/operators/math/im2col_test.cc @@ -71,15 +71,15 @@ void testIm2col() { context = new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace()); } else { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA context = new paddle::platform::CUDADeviceContext(paddle::platform::GPUPlace()); #else PADDLE_THROW("no GPU support"); #endif // PADDLE_ONLY_CPU } - im2col(input, output_cfo, stride, stride, padding, padding, context); - im2col_ocf(input, output_ocf, stride, stride, padding, padding, context); + im2col(*context, input, output_cfo, stride, stride, padding, padding); + im2col_ocf(*context, input, output_ocf, stride, stride, padding, padding); float* out_cfo_ptr; if (paddle::platform::is_cpu_place(*place)) { @@ -116,7 +116,7 @@ void testIm2col() { TEST(math, im2col) { testIm2col(); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA testIm2col(); #endif } diff --git a/paddle/operators/math/math_function.cc b/paddle/operators/math/math_function.cc index def4b01da0..ba653afa2c 100644 --- a/paddle/operators/math/math_function.cc +++ b/paddle/operators/math/math_function.cc @@ -48,6 +48,32 @@ void gemm(const platform::DeviceContext& context, beta, C, ldc); } +template <> +void gemm(const platform::DeviceContext& context, + const bool transA, const bool transB, + const int M, const int N, const int K, + const float alpha, const float* A, + const int lda, const float* B, + const int ldb, const float beta, float* C, + const int ldc) { + cblas_sgemm(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans, + transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A, + lda, B, ldb, beta, C, ldc); +} + +template <> +void gemm(const platform::DeviceContext& context, + const bool transA, const bool transB, + const int M, const int N, const int K, + const double alpha, const double* A, + const int lda, const double* B, + const int ldb, const double beta, + double* C, const int ldc) { + cblas_dgemm(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans, + transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A, + lda, B, ldb, beta, C, ldc); +} + template <> void matmul( const platform::DeviceContext& context, const framework::Tensor& matrix_a, diff --git a/paddle/operators/math/math_function.cu b/paddle/operators/math/math_function.cu index 71563b77b4..649f1f352c 100644 --- a/paddle/operators/math/math_function.cu +++ b/paddle/operators/math/math_function.cu @@ -63,6 +63,42 @@ void gemm(const platform::DeviceContext& context, cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N)); } +template <> +void gemm(const platform::DeviceContext& context, + const bool transA, const bool transB, + const int M, const int N, const int K, + const float alpha, const float* A, + const int lda, const float* B, + const int ldb, const float beta, float* C, + const int ldc) { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T; + PADDLE_ENFORCE(platform::dynload::cublasSgemm( + reinterpret_cast(context) + .cublas_handle(), + cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, ldc)); +} + +template <> +void gemm(const platform::DeviceContext& context, + const bool transA, const bool transB, + const int M, const int N, const int K, + const double alpha, const double* A, + const int lda, const double* B, + const int ldb, const double beta, + double* C, const int ldc) { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T; + PADDLE_ENFORCE(platform::dynload::cublasDgemm( + reinterpret_cast(context) + .cublas_handle(), + cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, ldc)); +} + template <> void matmul( const platform::DeviceContext& context, const framework::Tensor& matrix_a, diff --git a/paddle/operators/math/math_function.h b/paddle/operators/math/math_function.h index d8518e77fa..473eff4d19 100644 --- a/paddle/operators/math/math_function.h +++ b/paddle/operators/math/math_function.h @@ -52,6 +52,7 @@ int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda, #include +#include "paddle/framework/eigen.h" #include "paddle/framework/tensor.h" #include "paddle/platform/device_context.h" #include "paddle/platform/enforce.h" @@ -70,6 +71,13 @@ void gemm(const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA, const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, const T alpha, const T* A, const T* B, const T beta, T* C); +// gemm wrapper with stride args for matrix uncontinuous in memory +template +void gemm(const platform::DeviceContext& context, const bool transA, + const bool transB, const int M, const int N, const int K, + const T alpha, const T* A, const int lda, const T* B, const int ldb, + const T beta, T* C, const int ldc); + // matrix multiply with continuous memory template void matmul(const platform::DeviceContext& context, @@ -77,6 +85,13 @@ void matmul(const platform::DeviceContext& context, const framework::Tensor& matrix_b, bool trans_b, T alpha, framework::Tensor* matrix_out, T beta); +template +void SetConstant(const platform::DeviceContext& context, + framework::Tensor* tensor, T num) { + auto t = framework::EigenVector::Flatten(*tensor); + t.device(*context.GetEigenDevice()) = t.constant(static_cast(num)); +} + } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/operators/math/math_function_test.cc b/paddle/operators/math/math_function_test.cc index 7e339457f7..9945ba101d 100644 --- a/paddle/operators/math/math_function_test.cc +++ b/paddle/operators/math/math_function_test.cc @@ -1,7 +1,7 @@ #include "paddle/operators/math/math_function.h" #include "gtest/gtest.h" -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA TEST(math_function, notrans_mul_trans) { paddle::framework::Tensor input1; paddle::framework::Tensor input1_gpu; @@ -72,4 +72,195 @@ TEST(math_function, trans_mul_notrans) { EXPECT_EQ(out_ptr[8], 29); delete gpu_place; } + +TEST(math_function, gemm_notrans_cublas) { + paddle::framework::Tensor input1; + paddle::framework::Tensor input2; + paddle::framework::Tensor input3; + paddle::framework::Tensor input1_gpu; + paddle::framework::Tensor input2_gpu; + paddle::framework::Tensor input3_gpu; + + int m = 2; + int n = 3; + int k = 3; + auto* cpu_place = new paddle::platform::CPUPlace(); + float* input1_ptr = input1.mutable_data({2, 3}, *cpu_place); + float arr1[6] = {0, 1, 2, 3, 4, 5}; + memcpy(input1_ptr, arr1, 6 * sizeof(float)); + float* input2_ptr = input2.mutable_data({3, 4}, *cpu_place); + float arr2[12] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}; + memcpy(input2_ptr, arr2, 12 * sizeof(float)); + float* input3_ptr = input3.mutable_data({2, 4}, *cpu_place); + float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7}; + memcpy(input3_ptr, arr3, 8 * sizeof(float)); + + auto* gpu_place = new paddle::platform::GPUPlace(0); + paddle::platform::CUDADeviceContext context(*gpu_place); + + input1_gpu.CopyFrom(input1, *gpu_place); + input2_gpu.CopyFrom(input2, *gpu_place); + input3_gpu.CopyFrom(input3, *gpu_place); + float* a = input1_gpu.data(); + float* b = input2_gpu.data(); + float* c = input3_gpu.mutable_data(*gpu_place); + + paddle::operators::math::gemm( + context, false, false, m, n, k, 1, a, 3, b + 1, 4, 1, c + 1, 4); + + input3.CopyFrom(input3_gpu, *cpu_place); + + // numpy code: + // a = np.arange(6).reshape(2, 3) + // b = np.arange(12).reshape(3, 4)[:, 1:] + // c = np.arange(8).reshape(2, 4)[:, 1:] + // out = np.arange(8).reshape(2, 4) + // out[:, 1:] = np.dot(a, b) + c + EXPECT_EQ(input3_ptr[0], 0); + EXPECT_EQ(input3_ptr[1], 24); + EXPECT_EQ(input3_ptr[2], 28); + EXPECT_EQ(input3_ptr[3], 32); + EXPECT_EQ(input3_ptr[4], 4); + EXPECT_EQ(input3_ptr[5], 73); + EXPECT_EQ(input3_ptr[6], 86); + EXPECT_EQ(input3_ptr[7], 99); + delete gpu_place; +} + +TEST(math_function, gemm_trans_cublas) { + paddle::framework::Tensor input1; + paddle::framework::Tensor input2; + paddle::framework::Tensor input3; + paddle::framework::Tensor input1_gpu; + paddle::framework::Tensor input2_gpu; + paddle::framework::Tensor input3_gpu; + + int m = 2; + int n = 3; + int k = 3; + auto* cpu_place = new paddle::platform::CPUPlace(); + float* input1_ptr = input1.mutable_data({2, 3}, *cpu_place); + float arr1[6] = {0, 1, 2, 3, 4, 5}; + memcpy(input1_ptr, arr1, 6 * sizeof(float)); + float* input2_ptr = input2.mutable_data({4, 3}, *cpu_place); + float arr2[12] = {0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11}; + memcpy(input2_ptr, arr2, 12 * sizeof(float)); + float* input3_ptr = input3.mutable_data({2, 4}, *cpu_place); + float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7}; + memcpy(input3_ptr, arr3, 8 * sizeof(float)); + + auto* gpu_place = new paddle::platform::GPUPlace(0); + paddle::platform::CUDADeviceContext context(*gpu_place); + + input1_gpu.CopyFrom(input1, *gpu_place); + input2_gpu.CopyFrom(input2, *gpu_place); + input3_gpu.CopyFrom(input3, *gpu_place); + float* a = input1_gpu.data(); + float* b = input2_gpu.data(); + float* c = input3_gpu.mutable_data(*gpu_place); + + paddle::operators::math::gemm( + context, false, true, m, n, k, 1, a, 3, b + 3, 3, 1, c + 1, 4); + + input3.CopyFrom(input3_gpu, *cpu_place); + + EXPECT_EQ(input3_ptr[0], 0); + EXPECT_EQ(input3_ptr[1], 24); + EXPECT_EQ(input3_ptr[2], 28); + EXPECT_EQ(input3_ptr[3], 32); + EXPECT_EQ(input3_ptr[4], 4); + EXPECT_EQ(input3_ptr[5], 73); + EXPECT_EQ(input3_ptr[6], 86); + EXPECT_EQ(input3_ptr[7], 99); + delete gpu_place; +} #endif + +TEST(math_function, gemm_notrans_cblas) { + paddle::framework::Tensor input1; + paddle::framework::Tensor input2; + paddle::framework::Tensor input3; + + int m = 2; + int n = 3; + int k = 3; + auto* cpu_place = new paddle::platform::CPUPlace(); + float* input1_ptr = input1.mutable_data({2, 3}, *cpu_place); + float arr1[6] = {0, 1, 2, 3, 4, 5}; + memcpy(input1_ptr, arr1, 6 * sizeof(float)); + float* input2_ptr = input2.mutable_data({3, 4}, *cpu_place); + float arr2[12] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}; + memcpy(input2_ptr, arr2, 12 * sizeof(float)); + float* input3_ptr = input3.mutable_data({2, 4}, *cpu_place); + float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7}; + memcpy(input3_ptr, arr3, 8 * sizeof(float)); + + paddle::platform::CPUDeviceContext context(*cpu_place); + paddle::operators::math::gemm( + context, false, false, m, n, k, 1, input1_ptr, 3, input2_ptr + 1, 4, 1, + input3_ptr + 1, 4); + + EXPECT_EQ(input3_ptr[0], 0); + EXPECT_EQ(input3_ptr[1], 24); + EXPECT_EQ(input3_ptr[2], 28); + EXPECT_EQ(input3_ptr[3], 32); + EXPECT_EQ(input3_ptr[4], 4); + EXPECT_EQ(input3_ptr[5], 73); + EXPECT_EQ(input3_ptr[6], 86); + EXPECT_EQ(input3_ptr[7], 99); +} + +TEST(math_function, gemm_trans_clbas) { + paddle::framework::Tensor input1; + paddle::framework::Tensor input2; + paddle::framework::Tensor input3; + + int m = 2; + int n = 3; + int k = 3; + auto* cpu_place = new paddle::platform::CPUPlace(); + float* input1_ptr = input1.mutable_data({2, 3}, *cpu_place); + float arr1[6] = {0, 1, 2, 3, 4, 5}; + memcpy(input1_ptr, arr1, 6 * sizeof(float)); + float* input2_ptr = input2.mutable_data({4, 3}, *cpu_place); + float arr2[12] = {0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11}; + memcpy(input2_ptr, arr2, 12 * sizeof(float)); + float* input3_ptr = input3.mutable_data({2, 4}, *cpu_place); + float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7}; + memcpy(input3_ptr, arr3, 8 * sizeof(float)); + + paddle::platform::CPUDeviceContext context(*cpu_place); + paddle::operators::math::gemm( + context, false, true, m, n, k, 1, input1_ptr, 3, input2_ptr + 3, 3, 1, + input3_ptr + 1, 4); + + EXPECT_EQ(input3_ptr[0], 0); + EXPECT_EQ(input3_ptr[1], 24); + EXPECT_EQ(input3_ptr[2], 28); + EXPECT_EQ(input3_ptr[3], 32); + EXPECT_EQ(input3_ptr[4], 4); + EXPECT_EQ(input3_ptr[5], 73); + EXPECT_EQ(input3_ptr[6], 86); + EXPECT_EQ(input3_ptr[7], 99); +} + +TEST(math_function, zero) { + paddle::framework::Tensor tensor; + auto* cpu_place = new paddle::platform::CPUPlace(); + float* t = tensor.mutable_data({2, 2}, *cpu_place); + paddle::platform::CPUDeviceContext context(*cpu_place); + paddle::operators::math::SetConstant( + context, &tensor, 0); + EXPECT_EQ(t[0], 0); + EXPECT_EQ(t[1], 0); + EXPECT_EQ(t[2], 0); + EXPECT_EQ(t[3], 0); + + paddle::operators::math::SetConstant( + context, &tensor, 1); + + EXPECT_EQ(t[0], 1); + EXPECT_EQ(t[1], 1); + EXPECT_EQ(t[2], 1); + EXPECT_EQ(t[3], 1); +} diff --git a/paddle/operators/math/pooling.cc b/paddle/operators/math/pooling.cc new file mode 100644 index 0000000000..3b706529d8 --- /dev/null +++ b/paddle/operators/math/pooling.cc @@ -0,0 +1,463 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/math/pooling.h" + +namespace paddle { +namespace operators { +namespace math { + +template +class Pool2dFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_process) { + const int batch_size = input.dims()[0]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + + const int input_stride = input_height * input_width; + const int output_stride = output_height * output_width; + + const T* input_data = input.data(); + T* output_data = output.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + + T ele = pool_process.initial(); + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + pool_process.compute(ele, input_data[h * input_width + w]); + } + } + int pool_size = (hend - hstart) * (wend - wstart); + pool_process.finalize(ele, (static_cast(pool_size))); + output_data[ph * output_width + pw] = ele; + } + } + input_data += input_stride; + output_data += output_stride; + } + } + } +}; + +template +class Pool2dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings, + PoolProcess pool_grad_process) { + const int batch_size = input.dims()[0]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + const int input_stride = input_height * input_width; + const int output_stride = output_height * output_width; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + int pool_size = (hend - hstart) * (wend - wstart); + float scale = 1.0 / pool_size; + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + pool_grad_process.compute( + input_data[h * input_width + w], + output_data[ph * output_width + pw], + output_grad_data[ph * output_width + pw], + input_grad_data[h * input_width + w], + static_cast(scale)); + } + } + } + } + input_data += input_stride; + output_data += output_stride; + input_grad_data += input_stride; + output_grad_data += output_stride; + } + } + } +}; + +template +class MaxPool2dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input.dims()[0]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + const int input_stride = input_height * input_width; + const int output_stride = output_height * output_width; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + + bool stop = false; + for (int h = hstart; h < hend && !stop; ++h) { + for (int w = wstart; w < wend && !stop; ++w) { + int input_idx = h * input_width + w; + int output_idx = ph * output_width + pw; + if (input_data[input_idx] == output_data[output_idx]) { + input_grad_data[input_idx] += output_grad_data[output_idx]; + stop = true; + } + } + } + } + } + input_data += input_stride; + output_data += output_stride; + input_grad_data += input_stride; + output_grad_data += output_stride; + } + } + } +}; + +template class MaxPool2dGradFunctor; +// template class MaxPool2dGradFunctor; + +template class Pool2dFunctor, float>; +template class Pool2dFunctor, float>; +template class Pool2dGradFunctor< + platform::CPUPlace, paddle::operators::math::MaxPoolGrad, float>; +template class Pool2dGradFunctor< + platform::CPUPlace, paddle::operators::math::AvgPoolGrad, float>; +template class Pool2dFunctor, double>; +template class Pool2dFunctor, double>; +template class Pool2dGradFunctor< + platform::CPUPlace, paddle::operators::math::MaxPoolGrad, double>; +template class Pool2dGradFunctor< + platform::CPUPlace, paddle::operators::math::AvgPoolGrad, double>; + +template +class Pool3dFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_process) { + const int batch_size = input.dims()[0]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + + const int input_stride = input_depth * input_height * input_width; + const int output_stride = output_depth * output_height * output_width; + + const T* input_data = input.data(); + T* output_data = output.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int pd = 0; pd < output_depth; ++pd) { + int dstart = pd * stride_depth - padding_depth; + int dend = std::min(dstart + ksize_depth, input_depth); + dstart = std::max(dstart, 0); + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + int output_idx = (pd * output_height + ph) * output_width + pw; + T ele = pool_process.initial(); + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + pool_process.compute( + ele, + input_data[(d * input_height + h) * input_width + w]); + } + } + } + int pool_size = + (dend - dstart) * (hend - hstart) * (wend - wstart); + pool_process.finalize(ele, static_cast(pool_size)); + output_data[output_idx] = ele; + } + } + } + input_data += input_stride; + output_data += output_stride; + } + } + } +}; + +template +class Pool3dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings, + PoolProcess pool_grad_process) { + const int batch_size = input.dims()[0]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + const int input_stride = input_depth * input_height * input_width; + const int output_stride = output_depth * output_height * output_width; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int pd = 0; pd < output_depth; ++pd) { + int dstart = pd * stride_depth - padding_depth; + int dend = std::min(dstart + ksize_depth, input_depth); + dstart = std::max(dstart, 0); + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + + int pool_size = + (dend - dstart) * (hend - hstart) * (wend - wstart); + float scale = 1.0 / pool_size; + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + int input_idx = (d * input_height + h) * input_width + w; + int output_idx = + (pd * output_height + ph) * output_width + pw; + pool_grad_process.compute( + input_data[input_idx], output_data[output_idx], + output_grad_data[output_idx], + input_grad_data[input_idx], static_cast(scale)); + } + } + } + } + } + } + input_data += input_stride; + output_data += output_stride; + input_grad_data += input_stride; + output_grad_data += output_stride; + } + } + } +}; + +template +class MaxPool3dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input.dims()[0]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + const int input_stride = input_depth * input_height * input_width; + const int output_stride = output_depth * output_height * output_width; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + for (int i = 0; i < batch_size; i++) { + for (int c = 0; c < output_channels; ++c) { + for (int pd = 0; pd < output_depth; ++pd) { + int dstart = pd * stride_depth - padding_depth; + int dend = std::min(dstart + ksize_depth, input_depth); + dstart = std::max(dstart, 0); + for (int ph = 0; ph < output_height; ++ph) { + int hstart = ph * stride_height - padding_height; + int hend = std::min(hstart + ksize_height, input_height); + hstart = std::max(hstart, 0); + for (int pw = 0; pw < output_width; ++pw) { + int wstart = pw * stride_width - padding_width; + int wend = std::min(wstart + ksize_width, input_width); + wstart = std::max(wstart, 0); + bool stop = false; + for (int d = dstart; d < dend && !stop; ++d) { + for (int h = hstart; h < hend && !stop; ++h) { + for (int w = wstart; w < wend && !stop; ++w) { + int input_idx = (d * input_height + h) * input_width + w; + int output_idx = + (pd * output_height + ph) * output_width + pw; + + if (input_data[input_idx] == output_data[output_idx]) { + input_grad_data[input_idx] += + output_grad_data[output_idx]; + stop = true; + } + } + } + } + } + } + } + input_data += input_stride; + output_data += output_stride; + input_grad_data += input_stride; + output_grad_data += output_stride; + } + } + } +}; + +template class MaxPool3dGradFunctor; +// template class MaxPool3dGradFunctor; + +template class Pool3dFunctor, float>; +template class Pool3dFunctor, float>; +template class Pool3dGradFunctor< + platform::CPUPlace, paddle::operators::math::MaxPoolGrad, float>; +template class Pool3dGradFunctor< + platform::CPUPlace, paddle::operators::math::AvgPoolGrad, float>; +template class Pool3dFunctor, double>; +template class Pool3dFunctor, double>; +template class Pool3dGradFunctor< + platform::CPUPlace, paddle::operators::math::MaxPoolGrad, double>; +template class Pool3dGradFunctor< + platform::CPUPlace, paddle::operators::math::AvgPoolGrad, double>; +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/pooling.cu b/paddle/operators/math/pooling.cu new file mode 100644 index 0000000000..8aeccd1f8e --- /dev/null +++ b/paddle/operators/math/pooling.cu @@ -0,0 +1,635 @@ +/* Copyright (c) 2016 paddlepaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/math/pooling.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { +namespace math { + +template +__global__ void KernelPool2D(const int nthreads, const T* input_data, + T* output_data, const int channels, + const int input_height, const int input_width, + const int output_height, const int output_width, + const int ksize_height, const int ksize_width, + const int stride_height, const int stride_width, + const int padding_height, const int padding_width, + PoolProcess pool_process) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int pw = index % output_width; + int ph = (index / output_width) % output_height; + int c = (index / output_width / output_height) % channels; + int batch_idx = index / output_width / output_height / channels; + + int hstart = ph * stride_height - padding_height; + int hend = min(hstart + ksize_height, input_height); + hstart = max(hstart, 0); + + int wstart = pw * stride_width - padding_width; + int wend = min(wstart + ksize_width, input_width); + wstart = max(wstart, 0); + + input_data += (batch_idx * channels + c) * input_height * input_width; + T ele = pool_process.initial(); + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + pool_process.compute(ele, input_data[h * input_width + w]); + } + } + int pool_size = (hend - hstart) * (wend - wstart); + pool_process.finalize(ele, (static_cast(pool_size))); + output_data[index] = ele; + } +} + +template +__global__ void KernelPool2DGrad( + const int nthreads, const T* input_data, const T* output_data, + const T* output_grad, T* input_grad, const int channels, + const int input_height, const int input_width, const int output_height, + const int output_width, const int ksize_height, const int ksize_width, + const int stride_height, const int stride_width, const int padding_height, + const int padding_width, PoolProcess pool_process) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int offsetW = index % input_width + padding_width; + int offsetH = (index / input_width) % input_height + padding_height; + int offsetC = (index / input_width / input_height) % channels; + int batch_idx = index / input_width / input_height / channels; + + int phstart = (offsetH < ksize_height) + ? 0 + : (offsetH - ksize_height) / stride_height + 1; + int pwstart = (offsetW < ksize_width) + ? 0 + : (offsetW - ksize_width) / stride_width + 1; + int phend = min(offsetH / stride_height + 1, output_height); + int pwend = min(offsetW / stride_width + 1, output_width); + T gradient = 0; + T input = input_data[index]; + int output_idx = + (batch_idx * channels + offsetC) * output_height * output_width; + output_data += output_idx; + output_grad += output_idx; + for (int ph = phstart; ph < phend; ++ph) { + for (int pw = pwstart; pw < pwend; ++pw) { + int hstart = ph * stride_height - padding_height; + int wstart = pw * stride_width - padding_width; + int hend = min(hstart + ksize_height, input_height); + int wend = min(wstart + ksize_width, input_width); + hstart = max(hstart, 0); + wstart = max(wstart, 0); + int pool_size = (hend - hstart) * (wend - wstart); + int output_sub_idx = ph * output_width + pw; + pool_process.compute(input, output_data[output_sub_idx], + output_grad[output_sub_idx], gradient, + static_cast(1.0 / pool_size)); + } + } + input_grad[index] = gradient; + } +} + +template +__global__ void KernelMaxPool2DGrad( + const int nthreads, const T* input_data, const T* output_data, + const T* output_grad, T* input_grad, const int channels, + const int input_height, const int input_width, const int output_height, + const int output_width, const int ksize_height, const int ksize_width, + const int stride_height, const int stride_width, const int padding_height, + const int padding_width) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int pw = index % output_width; + int ph = (index / output_width) % output_height; + int c = (index / output_width / output_height) % channels; + int batch_idx = index / output_width / output_height / channels; + + int hstart = ph * stride_height - padding_height; + int hend = min(hstart + ksize_height, input_height); + hstart = max(hstart, 0); + + int wstart = pw * stride_width - padding_width; + int wend = min(wstart + ksize_width, input_width); + wstart = max(wstart, 0); + + input_data += (batch_idx * channels + c) * input_height * input_width; + input_grad += (batch_idx * channels + c) * input_height * input_width; + + T ele = output_data[index]; + int maxIndex = -1; + bool stop = false; + for (int h = hstart; h < hend && !stop; ++h) { + for (int w = wstart; w < wend && !stop; ++w) { + if (ele == input_data[h * input_width + w]) { + maxIndex = h * input_width + w; + stop = true; + } + } + } + + if (maxIndex != -1) { + // atomic add + atomicAdd(input_grad + maxIndex, output_grad[index]); + } + } +} + +template +class Pool2dFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_process) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + + const T* input_data = input.data(); + T* output_data = output.mutable_data(context.GetPlace()); + + int nthreads = batch_size * output_channels * output_height * output_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelPool2D< + PoolProcess, + T><<(context) + .stream()>>>(nthreads, input_data, output_data, input_channels, + input_height, input_width, output_height, + output_width, ksize_height, ksize_width, + stride_height, stride_width, padding_height, + padding_width, pool_process); + } +}; + +template +class Pool2dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings, + PoolProcess pool_process) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + int nthreads = batch_size * input_channels * input_height * input_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelPool2DGrad< + PoolProcess, + T><<(context) + .stream()>>>( + nthreads, input_data, output_data, output_grad_data, input_grad_data, + input_channels, input_height, input_width, output_height, output_width, + ksize_height, ksize_width, stride_height, stride_width, padding_height, + padding_width, pool_process); + } +}; + +template +class MaxPool2dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_height = input.dims()[2]; + const int input_width = input.dims()[3]; + const int output_channels = output.dims()[1]; + const int output_height = output.dims()[2]; + const int output_width = output.dims()[3]; + const int ksize_height = ksize[0]; + const int ksize_width = ksize[1]; + const int stride_height = strides[0]; + const int stride_width = strides[1]; + const int padding_height = paddings[0]; + const int padding_width = paddings[1]; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + int nthreads = batch_size * output_channels * output_height * output_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelMaxPool2DGrad< + T><<(context) + .stream()>>>( + nthreads, input_data, output_data, output_grad_data, input_grad_data, + input_channels, input_height, input_width, output_height, output_width, + ksize_height, ksize_width, stride_height, stride_width, padding_height, + padding_width); + } +}; + +template class MaxPool2dGradFunctor; +// template class MaxPool2dGradFunctor; // The +// 64-bit floating-point version of atomicAdd() is only supported by devices of +// compute capability 6.x and higher. + +template class Pool2dFunctor, float>; +template class Pool2dFunctor, float>; +template class Pool2dGradFunctor< + platform::GPUPlace, paddle::operators::math::MaxPoolGrad, float>; +template class Pool2dGradFunctor< + platform::GPUPlace, paddle::operators::math::AvgPoolGrad, float>; +template class Pool2dFunctor, double>; +template class Pool2dFunctor, double>; +template class Pool2dGradFunctor< + platform::GPUPlace, paddle::operators::math::MaxPoolGrad, double>; +template class Pool2dGradFunctor< + platform::GPUPlace, paddle::operators::math::AvgPoolGrad, double>; + +template +__global__ void KernelPool3D( + const int nthreads, const T* input_data, T* output_data, const int channels, + const int input_depth, const int input_height, const int input_width, + const int output_depth, const int output_height, const int output_width, + const int ksize_depth, const int ksize_height, const int ksize_width, + const int stride_depth, const int stride_height, const int stride_width, + const int padding_depth, const int padding_height, const int padding_width, + PoolProcess pool_process) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int pw = index % output_width; + int ph = (index / output_width) % output_height; + int pd = (index / output_width / output_height) % output_depth; + int c = (index / output_width / output_height / output_depth) % channels; + int batch_idx = + index / output_width / output_height / output_depth / channels; + int dstart = pd * stride_depth - padding_depth; + int hstart = ph * stride_height - padding_height; + int wstart = pw * stride_width - padding_width; + int dend = min(dstart + ksize_depth, input_depth); + int hend = min(hstart + ksize_height, input_height); + int wend = min(wstart + ksize_width, input_width); + dstart = max(dstart, 0); + hstart = max(hstart, 0); + wstart = max(wstart, 0); + T ele = pool_process.initial(); + input_data += + (batch_idx * channels + c) * input_depth * input_height * input_width; + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + pool_process.compute( + ele, input_data[(d * input_height + h) * input_width + w]); + } + } + } + int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); + pool_process.finalize(ele, static_cast(pool_size)); + output_data[index] = ele; + } +} + +template +__global__ void KernelPool3DGrad( + const int nthreads, const T* input_data, const T* output_data, + const T* output_grad, T* input_grad, const int channels, + const int input_depth, const int input_height, const int input_width, + const int output_depth, const int output_height, const int output_width, + const int ksize_depth, const int ksize_height, const int ksize_width, + const int stride_depth, const int stride_height, const int stride_width, + const int padding_depth, const int padding_height, const int padding_width, + PoolProcess pool_process) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int offsetW = index % input_width + padding_width; + int offsetH = (index / input_width) % input_height + padding_height; + int offsetD = + (index / input_width / input_height) % input_depth + padding_depth; + int offsetC = (index / input_width / input_height / input_depth) % channels; + int batch_idx = index / input_width / input_height / input_depth / channels; + + int pdstart = (offsetD < ksize_depth) + ? 0 + : (offsetD - ksize_depth) / stride_depth + 1; + int phstart = (offsetH < ksize_height) + ? 0 + : (offsetH - ksize_height) / stride_height + 1; + int pwstart = (offsetW < ksize_width) + ? 0 + : (offsetW - ksize_width) / stride_width + 1; + int pdend = min((offsetD) / stride_depth + 1, output_depth); + int phend = min((offsetH) / stride_height + 1, output_height); + int pwend = min((offsetW) / stride_width + 1, output_width); + + T gradient = 0; + T input = input_data[index]; + int output_idx = (batch_idx * channels + offsetC) * output_depth * + output_height * output_width; + output_data += output_idx; + output_grad += output_idx; + + for (int pd = pdstart; pd < pdend; ++pd) { + for (int ph = phstart; ph < phend; ++ph) { + for (int pw = pwstart; pw < pwend; ++pw) { + // figure out the pooling size + int dstart = pd * stride_depth - padding_depth; + int hstart = ph * stride_height - padding_height; + int wstart = pw * stride_width - padding_width; + int dend = min(dstart + ksize_depth, input_depth); + int hend = min(hstart + ksize_height, input_height); + int wend = min(wstart + ksize_width, input_width); + dstart = max(dstart, 0); + hstart = max(hstart, 0); + wstart = max(wstart, 0); + int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); + int output_sub_idx = (pd * output_height + ph) * output_width + pw; + pool_process.compute(input, output_data[output_sub_idx], + output_grad[output_sub_idx], gradient, + static_cast(1.0 / pool_size)); + } + } + } + input_grad[index] = gradient; + } +} + +template +__global__ void KernelMaxPool3DGrad( + const int nthreads, const T* input_data, const T* output_data, + const T* output_grad, T* input_grad, const int channels, + const int input_depth, const int input_height, const int input_width, + const int output_depth, const int output_height, const int output_width, + const int ksize_depth, const int ksize_height, const int ksize_width, + const int stride_depth, const int stride_height, const int stride_width, + const int padding_depth, const int padding_height, + const int padding_width) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; + index += blockDim.x * gridDim.x) { + int pw = index % output_width; + int ph = (index / output_width) % output_height; + int pd = (index / output_width / output_height) % output_depth; + int c = (index / output_width / output_height / output_depth) % channels; + int batch_idx = + index / output_width / output_height / output_depth / channels; + int dstart = pd * stride_depth - padding_depth; + int hstart = ph * stride_height - padding_height; + int wstart = pw * stride_width - padding_width; + int dend = min(dstart + ksize_depth, input_depth); + int hend = min(hstart + ksize_height, input_height); + int wend = min(wstart + ksize_width, input_width); + dstart = max(dstart, 0); + hstart = max(hstart, 0); + wstart = max(wstart, 0); + T ele = output_data[index]; + bool stop = false; + int maxIdx = -1; + input_data += + (batch_idx * channels + c) * input_depth * input_height * input_width; + input_grad += + (batch_idx * channels + c) * input_depth * input_height * input_width; + + for (int d = dstart; d < dend && !stop; ++d) { + for (int h = hstart; h < hend && !stop; ++h) { + for (int w = wstart; w < wend && !stop; ++w) { + if (ele == input_data[(d * input_height + h) * input_width + w]) { + stop = true; + maxIdx = (d * input_height + h) * input_width + w; + } + } + } + } + if (maxIdx != -1) { + // atomic add + atomicAdd(input_grad + maxIdx, output_grad[index]); + } + } +} + +template +class Pool3dFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_process) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + + const T* input_data = input.data(); + T* output_data = output.mutable_data(context.GetPlace()); + + int nthreads = batch_size * output_channels * output_depth * output_height * + output_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelPool3D< + PoolProcess, + T><<(context) + .stream()>>>( + nthreads, input_data, output_data, input_channels, input_depth, + input_height, input_width, output_depth, output_height, output_width, + ksize_depth, ksize_height, ksize_width, stride_depth, stride_height, + stride_width, padding_depth, padding_height, padding_width, + pool_process); + } +}; + +template +class Pool3dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings, + PoolProcess pool_process) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + int nthreads = + batch_size * input_channels * input_depth * input_height * input_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelPool3DGrad< + PoolProcess, + T><<(context) + .stream()>>>( + nthreads, input_data, output_data, output_grad_data, input_grad_data, + input_channels, input_depth, input_height, input_width, output_depth, + output_height, output_width, ksize_depth, ksize_height, ksize_width, + stride_depth, stride_height, stride_width, padding_depth, + padding_height, padding_width, pool_process); + } +}; + +template +class MaxPool3dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings) { + const int batch_size = input.dims()[0]; + const int input_channels = input.dims()[1]; + const int input_depth = input.dims()[2]; + const int input_height = input.dims()[3]; + const int input_width = input.dims()[4]; + const int output_channels = output.dims()[1]; + const int output_depth = output.dims()[2]; + const int output_height = output.dims()[3]; + const int output_width = output.dims()[4]; + const int ksize_depth = ksize[0]; + const int ksize_height = ksize[1]; + const int ksize_width = ksize[2]; + const int stride_depth = strides[0]; + const int stride_height = strides[1]; + const int stride_width = strides[2]; + const int padding_depth = paddings[0]; + const int padding_height = paddings[1]; + const int padding_width = paddings[2]; + + const T* input_data = input.data(); + const T* output_data = output.data(); + const T* output_grad_data = output_grad.data(); + T* input_grad_data = input_grad.mutable_data(context.GetPlace()); + + int nthreads = batch_size * output_channels * output_depth * output_height * + output_width; + int blocks = (nthreads + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KernelMaxPool3DGrad< + T><<(context) + .stream()>>>( + nthreads, input_data, output_data, output_grad_data, input_grad_data, + input_channels, input_depth, input_height, input_width, output_depth, + output_height, output_width, ksize_depth, ksize_height, ksize_width, + stride_depth, stride_height, stride_width, padding_depth, + padding_height, padding_width); + } +}; + +template class MaxPool3dGradFunctor; +// template class MaxPool3dGradFunctor; // The +// 64-bit floating-point version of atomicAdd() is only supported by devices of +// compute capability 6.x and higher. + +template class Pool3dFunctor, float>; +template class Pool3dFunctor, float>; +template class Pool3dGradFunctor< + platform::GPUPlace, paddle::operators::math::MaxPoolGrad, float>; +template class Pool3dGradFunctor< + platform::GPUPlace, paddle::operators::math::AvgPoolGrad, float>; +template class Pool3dFunctor, double>; +template class Pool3dFunctor, double>; +template class Pool3dGradFunctor< + platform::GPUPlace, paddle::operators::math::MaxPoolGrad, double>; +template class Pool3dGradFunctor< + platform::GPUPlace, paddle::operators::math::AvgPoolGrad, double>; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/pooling.h b/paddle/operators/math/pooling.h new file mode 100644 index 0000000000..d214c68923 --- /dev/null +++ b/paddle/operators/math/pooling.h @@ -0,0 +1,122 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/device_context.h" +#include "paddle/platform/hostdevice.h" + +namespace paddle { +namespace operators { +namespace math { +////////////////////// +#define FLT_MAX __FLT_MAX__ // + +template +class MaxPool { + public: + DEVICE inline T initial() { return static_cast(-FLT_MAX); } + DEVICE inline void compute(T& y, const T& x) { y = y > x ? y : x; } + DEVICE inline void finalize(T& y, const T& poo_size) {} +}; + +template +class AvgPool { + public: + DEVICE inline T initial() { return static_cast(0); } + DEVICE inline void compute(T& y, const T& x) { y += x; } + DEVICE inline void finalize(T& y, const T& poo_size) { y /= poo_size; } +}; +template +class MaxPoolGrad { + public: + DEVICE inline void compute(const T& x, const T& y, const T& dy, T& dx, + T scale) { + dx += dy * (x == y); + } +}; + +template +class AvgPoolGrad { + public: + DEVICE inline void compute(const T& x, const T& y, const T& dy, T& dx, + T scale) { + dx += (scale * dy); + } +}; + +template +class Pool2dFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_compute); +}; + +template +class Pool2dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings, + PoolProcess pool_compute); +}; + +template +class MaxPool2dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings); +}; + +template +class Pool3dFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& output, + std::vector& ksize, std::vector& strides, + std::vector& paddings, PoolProcess pool_compute); +}; + +template +class Pool3dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings, + PoolProcess pool_compute); +}; + +template +class MaxPool3dGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor& input_grad, + const framework::Tensor& output, + const framework::Tensor& output_grad, std::vector& ksize, + std::vector& strides, std::vector& paddings); +}; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/framework/grad_op_builder.h b/paddle/operators/math/softmax.cc similarity index 71% rename from paddle/framework/grad_op_builder.h rename to paddle/operators/math/softmax.cc index 998f8ebbb5..0ba8197ab8 100644 --- a/paddle/framework/grad_op_builder.h +++ b/paddle/operators/math/softmax.cc @@ -12,14 +12,15 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#pragma once - -#include "paddle/framework/operator.h" +#include "paddle/operators/math/softmax.h" namespace paddle { -namespace framework { +namespace operators { +namespace math { -OperatorBase* BuildGradOp(const OperatorBase* op); +template class SoftmaxFunctor; +template class SoftmaxGradFunctor; -} // namespace framework +} // namespace math +} // namespace operators } // namespace paddle diff --git a/paddle/operators/math/softmax.cu b/paddle/operators/math/softmax.cu new file mode 100644 index 0000000000..99f988d51e --- /dev/null +++ b/paddle/operators/math/softmax.cu @@ -0,0 +1,28 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#define EIGEN_USE_GPU + +#include "paddle/operators/math/softmax.h" + +namespace paddle { +namespace operators { +namespace math { + +template class SoftmaxFunctor; +template class SoftmaxGradFunctor; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/softmax.h b/paddle/operators/math/softmax.h new file mode 100644 index 0000000000..b7f627eee7 --- /dev/null +++ b/paddle/operators/math/softmax.h @@ -0,0 +1,104 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/operator.h" +#include "paddle/framework/tensor.h" + +namespace paddle { +namespace operators { +namespace math { + +template +using EigenMatrix = framework::EigenMatrix; + +template +struct ValueClip { + HOSTDEVICE T operator()(const T& x) const { + const T kThreshold = -64.; + return x < kThreshold ? kThreshold : x; + } +}; + +template +class SoftmaxFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor* X, framework::Tensor* Y) { + auto logits = EigenMatrix::From(*X); + auto softmax = EigenMatrix::From(*Y); + + const int kBatchDim = 0; + const int kClassDim = 1; + + const int batch_size = logits.dimension(kBatchDim); + const int num_classes = logits.dimension(kClassDim); + + Eigen::DSizes along_class(kClassDim); + Eigen::DSizes batch_by_one(batch_size, 1); + Eigen::DSizes one_by_class(1, num_classes); + + auto shifted_logits = (logits - + logits.maximum(along_class) + .eval() + .reshape(batch_by_one) + .broadcast(one_by_class)) + .unaryExpr(ValueClip()); + + softmax.device(*context.GetEigenDevice()) = shifted_logits.exp(); + softmax.device(*context.GetEigenDevice()) = + (softmax * + softmax.sum(along_class) + .inverse() + .eval() + .reshape(batch_by_one) + .broadcast(one_by_class)); + } +}; + +template +class SoftmaxGradFunctor { + public: + void operator()(const platform::DeviceContext& context, + const framework::Tensor* y, const framework::Tensor* y_grad, + framework::Tensor* x_grad) { + auto softmax = EigenMatrix::From(*y); + auto softmax_grad = EigenMatrix::From(*y_grad); + auto logits_grad = EigenMatrix::From(*x_grad); + + const int kBatchDim = 0; + const int kClassDim = 1; + + const int batch_size = softmax.dimension(kBatchDim); + const int num_classes = softmax.dimension(kClassDim); + + Eigen::DSizes along_class(kClassDim); + Eigen::DSizes batch_by_one(batch_size, 1); + Eigen::DSizes one_by_class(1, num_classes); + + auto dot = (softmax * softmax_grad) + .sum(along_class) + .eval() + .reshape(batch_by_one) + .broadcast(one_by_class); + logits_grad.device(*context.GetEigenDevice()) = + (softmax_grad - dot) * softmax; + } +}; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc index 7d7eeb59a2..2332c9546b 100644 --- a/paddle/operators/mean_op.cc +++ b/paddle/operators/mean_op.cc @@ -22,22 +22,23 @@ class MeanOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of MeanOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of MeanOp should not be null."); - ctx.Output("Out")->Resize({1}); + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of MeanOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of MeanOp should not be null."); + ctx->SetOutputDim("Out", {1}); } }; class MeanOpMaker : public framework::OpProtoAndCheckerMaker { public: - MeanOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + MeanOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of mean op"); - AddOutput("Out", "The output of mean op").NotInGradient(); - AddComment("Mean Operator"); + AddOutput("Out", "The output of mean op"); + AddComment(R"DOC( Mean Operator +)DOC"); } }; @@ -46,9 +47,23 @@ class MeanGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - ctx.Output(framework::GradVarName("X")) - ->Resize(ctx.Input("X")->dims()); + void InferShape(framework::InferShapeContextBase* ctx) const override { + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } +}; + +class MeanGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* grad_op = new framework::OpDescBind(); + grad_op->SetType("mean_grad"); + grad_op->SetInput("X", Input("X")); + grad_op->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); + grad_op->SetOutput(framework::GradVarName("X"), InputGrad("X")); + return std::unique_ptr(grad_op); } }; @@ -56,7 +71,8 @@ class MeanGradOp : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker, mean_grad, ops::MeanGradOp); +REGISTER_OPERATOR(mean, ops::MeanOp, ops::MeanOpMaker, ops::MeanGradMaker); +REGISTER_OPERATOR(mean_grad, ops::MeanGradOp); REGISTER_OP_CPU_KERNEL(mean, ops::MeanKernel); REGISTER_OP_CPU_KERNEL(mean_grad, diff --git a/paddle/operators/mean_op.h b/paddle/operators/mean_op.h index ce31e178d8..c99286a5b9 100644 --- a/paddle/operators/mean_op.h +++ b/paddle/operators/mean_op.h @@ -28,7 +28,7 @@ template ; template -class MeanKernel : public framework::OpKernel { +class MeanKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* input = context.Input("X"); @@ -45,7 +45,7 @@ class MeanKernel : public framework::OpKernel { }; template -class MeanGradKernel : public framework::OpKernel { +class MeanGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto OG = context.Input(framework::GradVarName("Out")); diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc index a97bbecdca..7057dcbd6e 100644 --- a/paddle/operators/minus_op.cc +++ b/paddle/operators/minus_op.cc @@ -26,21 +26,22 @@ class MinusOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of MinusOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), - "Input(Y) of MinusOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of MinusOp should not be null."); + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of MinusOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), + "Input(Y) of MinusOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of MinusOp should not be null."); - auto *left_tensor = ctx.Input("X"); - auto *right_tensor = ctx.Input("Y"); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); PADDLE_ENFORCE_EQ( - left_tensor->numel(), right_tensor->numel(), + x_dims, y_dims, "Minus operator must take two tensor with same num of elements"); - ctx.Output("Out")->Resize(left_tensor->dims()); + ctx->SetOutputDim("Out", x_dims); + ctx->ShareLoD("X", /*->*/ "Out"); } }; @@ -48,36 +49,50 @@ class MinusOpMaker : public framework::OpProtoAndCheckerMaker { public: MinusOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The left tensor of minus operator.").NotInGradient(); - AddInput("Y", "The right tensor of minus operator.").NotInGradient(); - AddOutput("Out", "The output tensor of minus operator.").NotInGradient(); + AddInput("X", "The left tensor of minus operator."); + AddInput("Y", "The right tensor of minus operator."); + AddOutput("Out", "The output tensor of minus operator."); AddComment(R"DOC(Minus Operator -Equation: Out = X - Y +Equation: + + Out = X - Y + +Both the input `X` and `Y` can carry the LoD (Level of Details) information, +or not. But the output only shares the LoD with input `X`. )DOC"); } }; -template -class MinusGradOp : public NetOp { + +class MinusGradMaker : public framework::GradOpDescMakerBase { public: - MinusGradOp(const std::string &type, const framework::VariableNameMap &inputs, - const framework::VariableNameMap &outputs, - const framework::AttributeMap &attrs) - : NetOp(type, inputs, outputs, attrs) { - auto out_grad = Input(framework::GradVarName("Out")); - auto x_grad = Output(framework::GradVarName("X")); - auto y_grad = Output(framework::GradVarName("Y")); - - // x_grad = out_grad - AppendOp(framework::OpRegistry::CreateOp("identity", {{"X", {out_grad}}}, - {{"Y", {x_grad}}}, {})); - - framework::AttributeMap scale_attr; - scale_attr["scale"] = static_cast(-1); - AppendOp(framework::OpRegistry::CreateOp("scale", {{"X", {out_grad}}}, - {{"Out", {y_grad}}}, scale_attr)); - CompleteAddOp(false); + using framework::GradOpDescMakerBase::GradOpDescMakerBase; + + std::vector> operator()() + const override { + std::vector> ops; + auto x_g = InputGrad("X"); + if (!x_g.empty()) { + auto *x_g_op = new framework::OpDescBind(); + x_g_op->SetType("scale"); + x_g_op->SetInput("X", OutputGrad("Out")); + x_g_op->SetOutput("Out", x_g); + x_g_op->SetAttr("scale", 1.0f); + ops.emplace_back(x_g_op); + } + + auto y_g = InputGrad("Y"); + if (!y_g.empty()) { + auto *y_g_op = new framework::OpDescBind(); + y_g_op->SetType("scale"); + y_g_op->SetInput("X", OutputGrad("Out")); + y_g_op->SetOutput("Out", y_g); + y_g_op->SetAttr("scale", -1.0f); + ops.emplace_back(y_g_op); + } + + return ops; } }; @@ -85,7 +100,6 @@ class MinusGradOp : public NetOp { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, minus_grad, - ops::MinusGradOp); +REGISTER_OPERATOR(minus, ops::MinusOp, ops::MinusOpMaker, ops::MinusGradMaker); REGISTER_OP_CPU_KERNEL(minus, ops::MinusKernel); diff --git a/paddle/operators/minus_op.h b/paddle/operators/minus_op.h index 6310a4fd51..bd9a2790aa 100644 --- a/paddle/operators/minus_op.h +++ b/paddle/operators/minus_op.h @@ -20,7 +20,7 @@ namespace paddle { namespace operators { template -class MinusKernel : public framework::OpKernel { +class MinusKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* left_tensor = context.Input("X"); diff --git a/paddle/operators/modified_huber_loss_op.cc b/paddle/operators/modified_huber_loss_op.cc new file mode 100644 index 0000000000..84212a2b3b --- /dev/null +++ b/paddle/operators/modified_huber_loss_op.cc @@ -0,0 +1,114 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/modified_huber_loss_op.h" + +namespace paddle { +namespace operators { + +class ModifiedHuberLossOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "X must be initialized."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Y must be initialized."); + + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + + PADDLE_ENFORCE_EQ(x_dims, y_dims, "The shape of X and Y must be the same."); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "The tensor rank of X must be 2."); + PADDLE_ENFORCE_EQ(x_dims[1], 1, "The 2nd dimension of X must be 1."); + + ctx->SetOutputDim("IntermediateVal", x_dims); + ctx->SetOutputDim("Out", {x_dims[0], 1}); + } +}; + +class ModifiedHuberLossOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ModifiedHuberLossOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "The input tensor of modified huber loss op." + "X is 2-D tensor with shape [batch_size, 1]."); + AddInput("Y", + "The target labels of modified huber loss op." + "The shape of Y is same as X. Values of Y must be 0 or 1."); + AddOutput("IntermediateVal", + "Variable to save intermediate result which will be reused in " + "backward processing.") + .AsIntermediate(); + AddOutput("Out", "Classification loss for X."); + AddComment(R"DOC( +Modified huber loss is used in binary classification problem. The shape of +input X and target Y are both [N, 1] and so is the shape of output loss. +Since target Y is not differentiable, cacluating gradient for Y is illegal. +The formulation of modified huber loss is: + +L(y, f(x)) = max(0, 1 - yf(x))^2 for yf(x) >= -1, + -4yf(x) otherwise. + +Make sure the values of target label Y are in {0, 1} here. The operator will +scale values of Y to {-1, +1} when computing losses and gradients. +)DOC"); + } +}; + +class ModifiedHuberLossGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "X must be initialized."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Y must be initialized."); + PADDLE_ENFORCE(ctx->HasInput("IntermediateVal"), + "Intermediate value must not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@Grad) must not be null."); + + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + auto intermediate_dims = ctx->GetInputDim("IntermediateVal"); + auto out_grad_dims = ctx->GetInputDim(framework::GradVarName("Out")); + + PADDLE_ENFORCE_EQ( + intermediate_dims, x_dims, + "The shape of X and intermediate value must be the same."); + PADDLE_ENFORCE_EQ(out_grad_dims, x_dims, + "The shape of Input(Out@Grad) and X must be the same."); + + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(modified_huber_loss, ops::ModifiedHuberLossOp, + ops::ModifiedHuberLossOpMaker, modified_huber_loss_grad, + ops::ModifiedHuberLossGradOp); + +REGISTER_OP_CPU_KERNEL( + modified_huber_loss, + ops::ModifiedHuberLossKernel); +REGISTER_OP_CPU_KERNEL(modified_huber_loss_grad, + ops::ModifiedHuberLossGradCPUKernel); diff --git a/paddle/operators/modified_huber_loss_op.cu b/paddle/operators/modified_huber_loss_op.cu new file mode 100644 index 0000000000..8854e166cd --- /dev/null +++ b/paddle/operators/modified_huber_loss_op.cu @@ -0,0 +1,78 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include +#include +#include +#include +#include "paddle/framework/op_registry.h" +#include "paddle/operators/modified_huber_loss_op.h" +#include "paddle/platform/hostdevice.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +struct ModifiedHuberLossBackward { + template + HOSTDEVICE void operator()(Tuple t) const { + auto inter_val = thrust::get<1>(t); + auto y_val = thrust::get<2>(t); + auto out_grad = thrust::get<3>(t); + if (inter_val < -1) { + thrust::get<0>(t) = -4 * (2 * y_val - 1) * out_grad; + } else if (inter_val < 1) { + thrust::get<0>(t) = -2 * (1 - inter_val) * (2 * y_val - 1) * out_grad; + } else { + thrust::get<0>(t) = 0; + } + } +}; + +template +class ModifiedHuberLossGradGPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* in0 = context.Input("Y"); + auto* in1 = context.Input("IntermediateVal"); + auto* in2 = context.Input(framework::GradVarName("Out")); + auto* out0 = context.Output(framework::GradVarName("X")); + + if (out0) { + auto counts = framework::product(in1->dims()); + auto y_ptr = thrust::device_pointer_cast(in0->data()); + auto inter_val_ptr = thrust::device_pointer_cast(in1->data()); + auto out_grad_ptr = thrust::device_pointer_cast(in2->data()); + thrust::device_ptr x_grad_ptr( + out0->mutable_data(context.GetPlace())); + + auto iter_begin = thrust::make_zip_iterator( + thrust::make_tuple(x_grad_ptr, inter_val_ptr, y_ptr, out_grad_ptr)); + + auto iter_end = thrust::make_zip_iterator( + thrust::make_tuple(x_grad_ptr + counts, inter_val_ptr + counts, + y_ptr + counts, out_grad_ptr + counts)); + + thrust::for_each(iter_begin, iter_end, ModifiedHuberLossBackward()); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL( + modified_huber_loss, + ops::ModifiedHuberLossKernel); +REGISTER_OP_GPU_KERNEL(modified_huber_loss_grad, + ops::ModifiedHuberLossGradGPUKernel); diff --git a/paddle/operators/modified_huber_loss_op.h b/paddle/operators/modified_huber_loss_op.h new file mode 100644 index 0000000000..aba75efad9 --- /dev/null +++ b/paddle/operators/modified_huber_loss_op.h @@ -0,0 +1,105 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/platform/hostdevice.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenVector = framework::EigenVector; + +template +struct CheckLabelValue { + HOSTDEVICE T operator()(const T& val) const { + PADDLE_ASSERT(val == static_cast(0) || val == static_cast(1)); + } +}; + +template +struct ModifiedHuberLossForward { + HOSTDEVICE T operator()(const T& val) const { + if (val < -1) { + return -4 * val; + } else if (val < 1) { + return (1 - val) * (1 - val); + } else { + return static_cast(0); + } + } +}; + +template +class ModifiedHuberLossKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* in0 = context.Input("X"); + auto* in1 = context.Input("Y"); + auto* out0 = context.Output("IntermediateVal"); + auto* out1 = context.Output("Out"); + + out0->mutable_data(context.GetPlace()); + out1->mutable_data(context.GetPlace()); + auto place = context.GetEigenDevice(); + + auto x = EigenVector::Flatten(*in0); + auto y = EigenVector::Flatten(*in1); + // make sure value's of Y in {0, 1} + y.unaryExpr(CheckLabelValue()); + auto inter_val = EigenVector::Flatten(*out0); + // scale y to {-1, +1} and compute x * y + inter_val.device(place) = x * (2 * y - static_cast(1)); + auto loss = EigenVector::Flatten(*out1); + loss.device(place) = inter_val.unaryExpr(ModifiedHuberLossForward()); + } +}; + +// CPU backward kernel +template +class ModifiedHuberLossGradCPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* in0 = context.Input("Y"); + auto* in1 = context.Input("IntermediateVal"); + auto* in2 = context.Input(framework::GradVarName("Out")); + auto* out0 = context.Output(framework::GradVarName("X")); + + if (out0) { + const T* y_ptr = in0->data(); + const T* inter_val_ptr = in1->data(); + const T* out_grad_ptr = in2->data(); + size_t counts = static_cast(framework::product(in1->dims())); + T* x_grad_ptr = out0->mutable_data(context.GetPlace()); + for (size_t i = 0; i < counts; ++i) { + if (inter_val_ptr[i] < -1) { + x_grad_ptr[i] = -4 * (2 * y_ptr[i] - 1) * out_grad_ptr[i]; + } else if (inter_val_ptr[i] < 1) { + x_grad_ptr[i] = -2 * (1 - inter_val_ptr[i]) * (2 * y_ptr[i] - 1) * + out_grad_ptr[i]; + } else { + x_grad_ptr[i] = 0; + } + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc index b6d320b415..3c8fe04d2e 100644 --- a/paddle/operators/mul_op.cc +++ b/paddle/operators/mul_op.cc @@ -1,16 +1,16 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 + http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ #include "paddle/operators/mul_op.h" @@ -18,34 +18,31 @@ namespace paddle { namespace operators { using framework::Tensor; -using framework::LoDTensor; class MulOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of MulOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), - "Input(Y) of MulOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of MulOp should not be null."); - - auto x_dims = ctx.Input("X")->dims(); - auto y_dims = ctx.Input("Y")->dims(); - int x_num_col_dims = Attr("x_num_col_dims"); - int y_num_col_dims = Attr("y_num_col_dims"); - - PADDLE_ENFORCE(x_dims.size() > x_num_col_dims, - "The rank of input tensor X(%s) should be larger than " - "`mul_op`'s `x_num_col_dims`.", - ctx.op().Input("X")); - PADDLE_ENFORCE(y_dims.size() > y_num_col_dims, - "The rank of input tensor Y(%s) should be larger than " - "`mul_op`'s `y_num_col_dims`.", - ctx.op().Input("Y")); + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of MulOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) of MulOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of MulOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + int x_num_col_dims = ctx->Attrs().Get("x_num_col_dims"); + int y_num_col_dims = ctx->Attrs().Get("y_num_col_dims"); + + PADDLE_ENFORCE_GT( + x_dims.size(), x_num_col_dims, + "The input tensor X's rank of MulOp should be larger than " + "x_num_col_dims."); + PADDLE_ENFORCE_GT( + y_dims.size(), y_num_col_dims, + "The input tensor Y's rank of MulOp should be larger than " + "y_num_col_dims."); auto x_mat_dims = framework::flatten_to_2d(x_dims, x_num_col_dims); auto y_mat_dims = framework::flatten_to_2d(y_dims, y_num_col_dims); @@ -53,23 +50,23 @@ class MulOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_EQ( x_mat_dims[1], y_mat_dims[0], "First matrix's width must be equal with second matrix's height."); - ctx.Output("Out")->Resize( - {x_mat_dims[0], y_mat_dims[1]}); + ctx->SetOutputDim("Out", {x_mat_dims[0], y_mat_dims[1]}); + ctx->ShareLoD("X", /*->*/ "Out"); } }; class MulOpMaker : public framework::OpProtoAndCheckerMaker { public: - MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + MulOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The first input of mul op"); AddInput("Y", "The second input of mul op"); AddOutput("Out", "The output of mul op"); AddAttr( "x_num_col_dims", - R"DOC(mul_op can take tensors with more than two dimensions as input `X`, - in that case, tensors will be reshaped to a matrix. The matrix's first - dimension(column length) will be the product of tensor's last + R"DOC(mul_op can take tensors with more than two dimensions as input `X`, + in that case, tensors will be reshaped to a matrix. The matrix's first + dimension(column length) will be the product of tensor's last `num_col_dims` dimensions, and the matrix's second dimension(row length) will be the product of tensor's first `rank - num_col_dims` dimensions. )DOC") @@ -83,9 +80,14 @@ class MulOpMaker : public framework::OpProtoAndCheckerMaker { .SetDefault(1) .EqualGreaterThan(1); AddComment(R"DOC( -Two Element Mul Operator. +Mul operator is used to perform matrix multiplication for input X and Y. -The equation is: Out = X * Y +The equation is: + + Out = X * Y + +Both the input `X` and `Y` can carry the LoD (Level of Details) information, +or not. But the output only shares the LoD with input `X`. )DOC"); } }; @@ -95,18 +97,14 @@ class MulOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); - auto x_dims = ctx.Input("X")->dims(); - auto y_dims = ctx.Input("Y")->dims(); - auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); - auto *x_grad = - ctx.Output(framework::GradVarName("X")); - auto *y_grad = - ctx.Output(framework::GradVarName("Y")); + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); auto x_mat_dims = framework::flatten_to_2d(x_dims, Attr("x_num_col_dims")); @@ -122,8 +120,15 @@ class MulOpGrad : public framework::OperatorWithKernel { "The second dimension of Out@GRAD must equal to the second " "dimension of the second operand."); - if (x_grad) x_grad->Resize(x_dims); - if (y_grad) y_grad->Resize(y_dims); + auto x_grad_name = framework::GradVarName("X"); + auto y_grad_name = framework::GradVarName("Y"); + + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); + } + if (ctx->HasOutput(y_grad_name)) { + ctx->SetOutputDim(y_grad_name, y_dims); + } } }; diff --git a/paddle/operators/mul_op.h b/paddle/operators/mul_op.h index ac7136a769..684b1ea0c0 100644 --- a/paddle/operators/mul_op.h +++ b/paddle/operators/mul_op.h @@ -28,7 +28,7 @@ template ; template -class MulKernel : public framework::OpKernel { +class MulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { const Tensor* x = context.Input("X"); @@ -52,7 +52,7 @@ class MulKernel : public framework::OpKernel { }; template -class MulGradKernel : public framework::OpKernel { +class MulGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { int x_num_col_dims = ctx.template Attr("x_num_col_dims"); diff --git a/paddle/operators/multiplex_op.cc b/paddle/operators/multiplex_op.cc new file mode 100644 index 0000000000..a069127a19 --- /dev/null +++ b/paddle/operators/multiplex_op.cc @@ -0,0 +1,124 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/multiplex_op.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +class MultiplexOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Ids"), "Input(Ids) shouldn't be null."); + PADDLE_ENFORCE(!ctx->Inputs("X").empty(), + "MultiInput(X) shouldn't be empty."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) shouldn't be null."); + auto ids_dim = ctx->GetInputDim("Ids"); + PADDLE_ENFORCE( + ids_dim.size() == 2 && ids_dim[1] == 1, + "The index tensor must be a vector with size batchSize x 1."); + + auto ins_dims = ctx->GetInputsDim("X"); + auto num_ins = ins_dims.size(); + PADDLE_ENFORCE(num_ins > 1, + "multiplex operator should have more than " + "one candidate input tensors."); + + auto in_dim = ins_dims[0]; + PADDLE_ENFORCE(in_dim.size() >= 2, + "The rank of candidate tensors must be not less than 2."); + for (size_t i = 1; i < num_ins; i++) { + auto dim = ins_dims[i]; + PADDLE_ENFORCE(in_dim == dim, + "All the candidate tensors must have the same size."); + } + ctx->SetOutputDim("Out", in_dim); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.MultiInput("X")[0]->type()); + } +}; + +class MultiplexOpMaker : public framework::OpProtoAndCheckerMaker { + public: + MultiplexOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Ids", "The index tensor of multiplex operator."); + AddInput("X", "The candidate tensors of multiplex operator.") + .AsDuplicable(); + AddOutput("Out", "The output tensor of multiplex operator."); + AddComment(R"DOC(Multiplex operator + +Multiplex multiple tensors according to the index provided by the index tensor. + +Ids: the index tensor. +X[0 : N - 1]: the candidate tensors for output (N >= 2). +For each index i from 0 to batchSize - 1, the output is the i-th row of the +the (Ids[i])-th tensor. + +For i-th row of the output tensor: + +y[i] = x_{k}[i] + +where y is the output tensor. `x_{k}` is the k-th input tensor +and `k = Ids[i]`. +)DOC"); + } +}; + +class MultiplexGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(!ctx->Inputs("X").empty(), "Input(X) should not be null."); + PADDLE_ENFORCE(!ctx->Outputs(framework::GradVarName("X")).empty(), + "Output(X@Grad) should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null."); + std::vector d_ins; + auto ins = ctx->GetInputsDim("X"); + // No need to compute gradient for Input(Ids) + for (size_t i = 0; i < ins.size(); i++) { + d_ins.push_back(ins[i]); + } + ctx->SetOutputsDim(framework::GradVarName("X"), d_ins); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.MultiInput("X")[0]->type()); + } +}; + +} // namespace operators +} // namespace paddle +namespace ops = paddle::operators; + +REGISTER_OP(multiplex, ops::MultiplexOp, ops::MultiplexOpMaker, multiplex_grad, + ops::MultiplexGradOp); +REGISTER_OP_CPU_KERNEL( + multiplex, ops::MultiplexCPUKernel); +REGISTER_OP_CPU_KERNEL( + multiplex_grad, + ops::MultiplexGradCPUKernel); diff --git a/paddle/operators/multiplex_op.cu b/paddle/operators/multiplex_op.cu new file mode 100644 index 0000000000..72b1f96eaf --- /dev/null +++ b/paddle/operators/multiplex_op.cu @@ -0,0 +1,98 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/framework/op_registry.h" +#include "paddle/operators/multiplex_op.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class MultiplexGPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const { + auto ins = ctx.MultiInput("X"); + auto* ids = ctx.Input("Ids"); + auto* out = ctx.Output("Out"); + out->mutable_data(ctx.GetPlace()); + + auto rows = ins[0]->dims()[0]; + auto cols = ins[0]->numel() / rows; + // copy index to cpu + Tensor index_t_cpu; + index_t_cpu.CopyFrom(*ids, platform::CPUPlace()); + auto* index = index_t_cpu.data(); + auto stream = reinterpret_cast( + ctx.device_context()) + .stream(); + Place place = boost::get(ctx.GetPlace()); + for (auto i = 0; i < rows; i++) { + int32_t k = index[i]; + PADDLE_ENFORCE_GE(k, 0, "index must be nonnegative."); + PADDLE_ENFORCE_LT((size_t)k, ins.size(), + "index exceeds the number of candidate tensors."); + memory::Copy(place, out->data() + i * cols, place, + ins[k]->data() + i * cols, cols * sizeof(T), stream); + } + } +}; + +template +class MultiplexGradGPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const { + auto* d_out = ctx.Input(framework::GradVarName("Out")); + auto ins = ctx.MultiInput("X"); + auto* ids = ctx.Input("Ids"); + auto d_ins = ctx.MultiOutput(framework::GradVarName("X")); + for (size_t i = 0; i < d_ins.size(); i++) { + if (d_ins[i]) { + d_ins[i]->mutable_data(ctx.GetPlace()); + auto t = framework::EigenVector::Flatten(*d_ins[i]); + t.device(ctx.GetEigenDevice()) = t.constant(static_cast(0)); + } + } + + auto rows = ins[0]->dims()[0]; + auto cols = ins[0]->numel() / rows; + // copy index to cpu + Tensor index_t_cpu; + index_t_cpu.CopyFrom(*ids, platform::CPUPlace()); + auto* index = index_t_cpu.data(); + + auto stream = reinterpret_cast( + ctx.device_context()) + .stream(); + Place place = boost::get(ctx.GetPlace()); + for (auto i = 0; i < rows; i++) { + size_t k = static_cast(index[i]); + if (d_ins[k]) { + memory::Copy(place, d_ins[k]->data() + i * cols, place, + d_out->data() + i * cols, cols * sizeof(T), stream); + } + } + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_GPU_KERNEL( + multiplex, ops::MultiplexGPUKernel); +REGISTER_OP_GPU_KERNEL( + multiplex_grad, + ops::MultiplexGradGPUKernel); diff --git a/paddle/operators/multiplex_op.h b/paddle/operators/multiplex_op.h new file mode 100644 index 0000000000..ab3cafaa32 --- /dev/null +++ b/paddle/operators/multiplex_op.h @@ -0,0 +1,81 @@ + +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/memory/memcpy.h" + +namespace paddle { +namespace operators { + +template +class MultiplexCPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const { + auto ins = ctx.MultiInput("X"); + auto ids = ctx.Input("Ids"); + auto* out = ctx.Output("Out"); + + out->mutable_data(ctx.GetPlace()); + + auto rows = ins[0]->dims()[0]; + auto cols = ins[0]->numel() / rows; + auto index = ids->data(); + Place place = boost::get(ctx.GetPlace()); + for (auto i = 0; i < rows; i++) { + int32_t k = index[i]; + PADDLE_ENFORCE_GE(k, 0, "index must be nonnegative."); + PADDLE_ENFORCE_LT(static_cast(k), ins.size(), + "index exceeds the number of candidate tensors."); + memory::Copy(place, out->data() + i * cols, place, + ins[k]->data() + i * cols, cols * sizeof(T)); + } + } +}; + +template +class MultiplexGradCPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const { + auto* d_out = ctx.Input(framework::GradVarName("Out")); + auto* ids = ctx.Input("Ids"); + auto ins = ctx.MultiInput("X"); + auto d_ins = + ctx.MultiOutput(framework::GradVarName("X")); + for (size_t i = 0; i < d_ins.size(); i++) { + if (d_ins[i]) { + d_ins[i]->mutable_data(ctx.GetPlace()); + auto t = framework::EigenVector::Flatten(*d_ins[i]); + t.device(ctx.GetEigenDevice()) = t.constant(static_cast(0)); + } + } + + auto rows = ins[0]->dims()[0]; + auto cols = ins[0]->numel() / rows; + auto* index = ids->data(); + Place place = boost::get(ctx.GetPlace()); + for (auto i = 0; i < rows; i++) { + size_t k = static_cast(index[i]); + if (d_ins[k]) { + memory::Copy(place, d_ins[k]->data() + i * cols, place, + d_out->data() + i * cols, cols * sizeof(T)); + } + } + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/net_op.h b/paddle/operators/net_op.h index fcd8134b2c..2388b094d2 100644 --- a/paddle/operators/net_op.h +++ b/paddle/operators/net_op.h @@ -53,16 +53,6 @@ class NetOp : public framework::OperatorBase { this->CompleteAddOp(); } - /** - * Infer all the operators' input and output variables' shapes, will be called - * before every mini-batch - */ - void InferShape(const framework::Scope& scope) const override { - for (auto& op : ops_) { - op->InferShape(scope); - } - } - /** * @brief Run the network. * diff --git a/paddle/operators/net_op_test.cc b/paddle/operators/net_op_test.cc index f2e98ee7a1..63bebd5b44 100644 --- a/paddle/operators/net_op_test.cc +++ b/paddle/operators/net_op_test.cc @@ -7,14 +7,12 @@ namespace operators { using Scope = framework::Scope; using DeviceContext = platform::DeviceContext; -static int infer_shape_cnt = 0; static int run_cnt = 0; class TestOp : public framework::OperatorBase { public: using framework::OperatorBase::OperatorBase; DEFINE_OP_CLONE_METHOD(TestOp); - void InferShape(const Scope& scope) const override { ++infer_shape_cnt; } void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override { ++run_cnt; diff --git a/paddle/operators/pad_op.cc b/paddle/operators/pad_op.cc index a0b1c6b631..15aa05f266 100644 --- a/paddle/operators/pad_op.cc +++ b/paddle/operators/pad_op.cc @@ -24,14 +24,13 @@ class PadOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of PadOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of PadOp should not be null."); - - auto x_dim = ctx.Input("X")->dims(); - auto paddings = Attr>("paddings"); + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of PadOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of PadOp should not be null."); + + auto x_dim = ctx->GetInputDim("X"); + auto paddings = ctx->Attrs().Get>("paddings"); PADDLE_ENFORCE_EQ(x_dim.size() * 2, int64_t(paddings.size()), "Size of paddings should be equal to 2 * dimension size " "of input tensor."); @@ -39,22 +38,25 @@ class PadOp : public framework::OperatorWithKernel { for (int i = 0; i < x_dim.size(); ++i) { out_dims[i] = x_dim[i] + paddings[i * 2] + paddings[i * 2 + 1]; } - ctx.Output("Out")->Resize( - framework::make_ddim(out_dims)); + ctx->SetOutputDim("Out", framework::make_ddim(out_dims)); + if (out_dims[0] == x_dim[0]) { + // Only pass LoD when the first dimension is equal between + // output and input. + ctx->ShareLoD("X", /*->*/ "Out"); + } } }; class PadOpMaker : public framework::OpProtoAndCheckerMaker { public: - PadOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + PadOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of pad op. " "The input should be a k-D tensor(k > 0 and k < 7)"); AddOutput("Out", "The output of pad op." - "A tensor with the same shape as X.") - .NotInGradient(); + "A tensor with the same shape as X."); AddComment(R"DOC( Pad input into output, as specified by paddings and pad_value. The input should be a k-D tensor(k > 0 and k < 7). As an example: @@ -63,15 +65,15 @@ Given: X = [[1, 2], [3, 4]] -and +and paddings = [0, 1, 1, 2] and - -pad_value = 0 -then we get +pad_value = 0 + +then we get Out = [[0, 1, 2, 0, 0] [0, 3, 4, 0, 0] @@ -96,23 +98,41 @@ class PadOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); - auto x_dims = ctx.Input("X")->dims(); - auto *x_g = ctx.Output(framework::GradVarName("X")); - if (x_g != nullptr) { - x_g->Resize(x_dims); + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto x_dims = ctx->GetInputDim("X"); + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); } } }; +class PadOpGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* bind = new framework::OpDescBind(); + bind->SetInput("X", Input("X")); + bind->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); + bind->SetOutput(framework::GradVarName("X"), InputGrad("X")); + bind->SetAttrMap(Attrs()); + bind->SetType("pad_grad"); + return std::unique_ptr(bind); + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(pad, ops::PadOp, ops::PadOpMaker, pad_grad, ops::PadOpGrad); + +REGISTER_OPERATOR(pad, ops::PadOp, ops::PadOpMaker, ops::PadOpGradMaker); +REGISTER_OPERATOR(pad_grad, ops::PadOpGrad); REGISTER_OP_CPU_KERNEL(pad, ops::PadKernel); REGISTER_OP_CPU_KERNEL(pad_grad, ops::PadGradKernel); diff --git a/paddle/operators/pad_op.h b/paddle/operators/pad_op.h index 2cc3b945ae..9534dbf545 100644 --- a/paddle/operators/pad_op.h +++ b/paddle/operators/pad_op.h @@ -47,7 +47,7 @@ void PadFunction(const framework::ExecutionContext& context) { } template -class PadKernel : public framework::OpKernel { +class PadKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { int rank = context.Input("X")->dims().size(); @@ -97,7 +97,7 @@ void PadGradFunction(const framework::ExecutionContext& context) { } template -class PadGradKernel : public framework::OpKernel { +class PadGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { size_t rank = diff --git a/paddle/operators/pool_op.cc b/paddle/operators/pool_op.cc new file mode 100644 index 0000000000..c29f51f056 --- /dev/null +++ b/paddle/operators/pool_op.cc @@ -0,0 +1,195 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/pool_op.h" + +namespace paddle { +namespace operators { + +int OutputSizePool(int input_size, int filter_size, int padding, int stride) { + int output_size = (input_size - filter_size + 2 * padding) / stride + 1; + return output_size; +} + +class PoolOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "X(Input) of Pooling should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Out(Output) of Pooling should not be null."); + + auto in_x_dims = ctx->GetInputDim("X"); + + std::string pooling_type = ctx->Attrs().Get("poolingType"); + std::vector ksize = ctx->Attrs().Get>("ksize"); + std::vector strides = ctx->Attrs().Get>("strides"); + std::vector paddings = ctx->Attrs().Get>("paddings"); + + PADDLE_ENFORCE(pooling_type == "max" || pooling_type == "avg", + "pooling_type should be 'max' or 'avg'"); + PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5, + "Pooling intput should be 4-D or 5-D"); + + if (ctx->Attrs().Get("globalPooling")) { + ksize.resize(static_cast(in_x_dims.size()) - 2); + for (size_t i = 0; i < ksize.size(); ++i) + ksize[i] = static_cast(in_x_dims[i + 2]); + } + + PADDLE_ENFORCE(in_x_dims.size() - ksize.size() == 2U, + "Input size and Pooling size should be consistent."); + PADDLE_ENFORCE(ksize.size() == 2 || ksize.size() == 3, + "Pooling size should be 2 elements. or 3 elements."); + PADDLE_ENFORCE_EQ(ksize.size(), strides.size(), + "strides size and pooling size should be the same."); + PADDLE_ENFORCE_EQ(ksize.size(), paddings.size(), + "paddings size and pooling size should be the same."); + + std::vector output_shape({in_x_dims[0], in_x_dims[1]}); + for (size_t i = 0; i < ksize.size(); ++i) { + output_shape.push_back( + OutputSizePool(in_x_dims[i + 2], ksize[i], paddings[i], strides[i])); + } + ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); + } +}; + +class PoolOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "X(Input) of Pooling should not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Input@Grad of Pooling should not be null."); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } +}; + +class Pool2dOpMaker : public framework::OpProtoAndCheckerMaker { + public: + Pool2dOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput( + "X", + "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 and W is the height and width of feature."); + AddOutput("Out", + "The output tensor of pooling operator." + "The format of output tensor is also NCHW."); + + AddAttr("poolingType", + "PoolingType of pooling operator." + "Str constant equal to 'max' or 'avg'.") + .InEnum({"max", "avg"}); + AddAttr>( + "ksize", + "Pooling size(depth, height, width) of pooling operator." + "If globalPooling = true, ksize is ignored and need not be " + "specified."); // TODO(Add checker) + AddAttr( + "globalPooling", + "Whether to use the globalPooling." + "Bool constant equal to false or true." + "Default false." + "If globalPooling = true, ksize is ignored and need not be specified.") + .SetDefault(false); + AddAttr>("strides", + "Strides(height, width) of pooling operator." + "Default {1,1}") + .SetDefault({1, 1}); // TODO(Add checker) + AddAttr>("paddings", + "Paddings(height, width) of pooling operator." + "Default {0,0}.") + .SetDefault({0, 0}); // TODO(Add checker) + AddComment(R"DOC( +The pooling2d operation calculates the output based on +the input, poolingType and ksize, strides, paddings parameters. +)DOC"); + } +}; + +class Pool3dOpMaker : public framework::OpProtoAndCheckerMaker { + public: + Pool3dOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "The input tensor of pooling operator. " + "The format of input tensor is NCDHW. Where N is batch size, C is " + "the " + "number of channels, D, H and W is the depth, height and width of " + "feature."); + AddOutput("Out", + "The output tensor of pooling operator." + "The format of output tensor is also NCDHW."); + + AddAttr("poolingType", + "PoolingType of pooling operator." + "str constant equal to 'max' or 'avg'.") + .InEnum({"max", "avg"}); + AddAttr>( + "ksize", + "Pooling size(depth, height, width) of pooling operator." + "If globalPooling = true, ksize is ignored and need not be " + "specified."); // TODO(Add checker) + AddAttr( + "globalPooling", + "Whether to use the globalPooling." + "Bool constant equal to false or true." + "Default false." + "If globalPooling = true, ksize is ignored and need not be specified.") + .SetDefault(false); + AddAttr>( + "strides", + "Strides(depth, height, width) of pooling operator." + "Default {1,1,1}.") + .SetDefault({1, 1, 1}); // TODO(Add checker) + AddAttr>( + "paddings", + "Paddings(depth, height, width) of pooling operator." + "Default {0,0,0}.") + .SetDefault({0, 0, 0}); // TODO(Add checker) + AddComment(R"DOC( +The pooling3d operation calculates the output based on +the input, poolingType and ksize, strides, paddings parameters. +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP(pool2d, ops::PoolOp, ops::Pool2dOpMaker, pool2d_grad, + ops::PoolOpGrad); + +REGISTER_OP_CPU_KERNEL(pool2d, + ops::PoolKernel); +REGISTER_OP_CPU_KERNEL(pool2d_grad, + ops::PoolGradKernel) + +REGISTER_OP(pool3d, ops::PoolOp, ops::Pool3dOpMaker, pool3d_grad, + ops::PoolOpGrad); + +REGISTER_OP_CPU_KERNEL(pool3d, + ops::PoolKernel); +REGISTER_OP_CPU_KERNEL(pool3d_grad, + ops::PoolGradKernel); diff --git a/paddle/operators/pool_op.cu b/paddle/operators/pool_op.cu new file mode 100644 index 0000000000..0e3b80868f --- /dev/null +++ b/paddle/operators/pool_op.cu @@ -0,0 +1,27 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/pool_op.h" + +namespace ops = paddle::operators; + +REGISTER_OP_GPU_KERNEL(pool2d, + ops::PoolKernel); +REGISTER_OP_GPU_KERNEL(pool2d_grad, + ops::PoolGradKernel); + +REGISTER_OP_GPU_KERNEL(pool3d, + ops::PoolKernel); +REGISTER_OP_GPU_KERNEL(pool3d_grad, + ops::PoolGradKernel); diff --git a/paddle/operators/pool_op.h b/paddle/operators/pool_op.h new file mode 100644 index 0000000000..c2bc358def --- /dev/null +++ b/paddle/operators/pool_op.h @@ -0,0 +1,147 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/math/math_function.h" +#include "paddle/operators/math/pooling.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class PoolKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* in_x = context.Input("X"); + Tensor* out = context.Output("Out"); + + std::string pooling_type = context.Attr("poolingType"); + std::vector ksize = context.Attr>("ksize"); + std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); + if (context.Attr("globalPooling")) { + for (size_t i = 0; i < ksize.size(); ++i) { + ksize[i] = static_cast(in_x->dims()[i + 2]); + } + } + + switch (ksize.size()) { + case 2: { + if (pooling_type == "max") { + paddle::operators::math::Pool2dFunctor< + Place, paddle::operators::math::MaxPool, T> + pool2d_forward; + paddle::operators::math::MaxPool pool_process; + pool2d_forward(context.device_context(), *in_x, *out, ksize, strides, + paddings, pool_process); + + } else if (pooling_type == "avg") { + paddle::operators::math::Pool2dFunctor< + Place, paddle::operators::math::AvgPool, T> + pool2d_forward; + paddle::operators::math::AvgPool pool_process; + pool2d_forward(context.device_context(), *in_x, *out, ksize, strides, + paddings, pool_process); + } + } break; + case 3: { + if (pooling_type == "max") { + paddle::operators::math::Pool3dFunctor< + Place, paddle::operators::math::MaxPool, T> + pool3d_forward; + paddle::operators::math::MaxPool pool_process; + pool3d_forward(context.device_context(), *in_x, *out, ksize, strides, + paddings, pool_process); + } else if (pooling_type == "avg") { + paddle::operators::math::Pool3dFunctor< + Place, paddle::operators::math::AvgPool, T> + pool3d_forward; + paddle::operators::math::AvgPool pool_process; + pool3d_forward(context.device_context(), *in_x, *out, ksize, strides, + paddings, pool_process); + } + } break; + } + } +}; + +template +class PoolGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* in_x = context.Input("X"); + const Tensor* out = context.Input("Out"); + const Tensor* out_grad = + context.Input(framework::GradVarName("Out")); + Tensor* in_x_grad = context.Output(framework::GradVarName("X")); + + std::string pooling_type = context.Attr("poolingType"); + std::vector ksize = context.Attr>("ksize"); + std::vector strides = context.Attr>("strides"); + std::vector paddings = context.Attr>("paddings"); + + if (context.Attr("globalPooling")) { + for (size_t i = 0; i < ksize.size(); ++i) + ksize[i] = static_cast(in_x->dims()[i + 2]); + } + + if (in_x_grad) { + in_x_grad->mutable_data(context.GetPlace()); + auto temp = framework::EigenVector::Flatten(*in_x_grad); + temp.device(context.GetEigenDevice()) = + temp.constant(static_cast(0)); + + switch (ksize.size()) { + case 2: { + if (pooling_type == "max") { + paddle::operators::math::MaxPool2dGradFunctor + pool2d_backward; + pool2d_backward(context.device_context(), *in_x, *in_x_grad, *out, + *out_grad, ksize, strides, paddings); + } else if (pooling_type == "avg") { + paddle::operators::math::Pool2dGradFunctor< + Place, paddle::operators::math::AvgPoolGrad, T> + pool2d_backward; + paddle::operators::math::AvgPoolGrad pool_process; + pool2d_backward(context.device_context(), *in_x, *in_x_grad, *out, + *out_grad, ksize, strides, paddings, pool_process); + } + } break; + case 3: { + if (pooling_type == "max") { + paddle::operators::math::MaxPool3dGradFunctor + pool3d_backward; + pool3d_backward(context.device_context(), *in_x, *in_x_grad, *out, + *out_grad, ksize, strides, paddings); + } else if (pooling_type == "avg") { + paddle::operators::math::Pool3dGradFunctor< + Place, paddle::operators::math::AvgPoolGrad, T> + pool3d_backward; + paddle::operators::math::AvgPoolGrad pool_process; + pool3d_backward(context.device_context(), *in_x, *in_x_grad, *out, + *out_grad, ksize, strides, paddings, pool_process); + } + } break; + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/prelu_op.cc b/paddle/operators/prelu_op.cc index 7ae80b2968..1692464f28 100644 --- a/paddle/operators/prelu_op.cc +++ b/paddle/operators/prelu_op.cc @@ -26,18 +26,14 @@ class PReluOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); - auto *in = ctx.Input("X"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Alpha"), - "Input(Alpha) should not be null"); - auto *alpha = ctx.Input("Alpha"); - PADDLE_ENFORCE(alpha->numel() == 1, "Size of weight Alpha must be one."); - - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) should not be null"); - auto *out = ctx.Output("Out"); - out->Resize(in->dims()); + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput("Alpha"), "Input(Alpha) should not be null"); + PADDLE_ENFORCE(product(ctx->GetInputDim("Alpha")) == 1, + "Size of weight Alpha must be one."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null"); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Out"); } }; @@ -55,6 +51,8 @@ The equation is: f(x) = alpha * x , for x < 0 f(x) = x , for x >= 0 +The input `X` can carry the LoD (Level of Details) information, +or not. And the output shares the LoD with input `X`. )DOC"); } }; @@ -65,19 +63,13 @@ class PReluGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); - auto *dx = ctx.Output(framework::GradVarName("X")); - auto *x = ctx.Input("X"); - - auto *dalpha = - ctx.Output(framework::GradVarName("Alpha")); - auto *alpha = ctx.Input("Alpha"); - - dx->Resize(x->dims()); - dalpha->Resize(alpha->dims()); + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + ctx->SetOutputDim(framework::GradVarName("Alpha"), + ctx->GetInputDim("Alpha")); } }; diff --git a/paddle/operators/prelu_op.h b/paddle/operators/prelu_op.h index 63031c25cc..5ad31c2203 100644 --- a/paddle/operators/prelu_op.h +++ b/paddle/operators/prelu_op.h @@ -40,7 +40,7 @@ class PReluFunctor { }; template -class PReluKernel : public framework::OpKernel { +class PReluKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* x = context.Input("X"); @@ -54,8 +54,9 @@ class PReluKernel : public framework::OpKernel { int numel = x->numel(); - Transform(context.device_context(), x_ptr, x_ptr + numel, o_ptr, - PReluFunctor(alpha_ptr)); + Transform trans; + trans(context.device_context(), x_ptr, x_ptr + numel, o_ptr, + PReluFunctor(alpha_ptr)); } }; @@ -76,7 +77,7 @@ class PReluGradFunctor { }; template -class PReluGradKernel : public framework::OpKernel { +class PReluGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* dx = context.Output(framework::GradVarName("X")); @@ -91,10 +92,11 @@ class PReluGradKernel : public framework::OpKernel { const T* out_ptr = out->data(); int numel = dx->numel(); - Transform(context.device_context(), out_ptr, out_ptr + numel, dout_ptr, - dx_ptr, PReluGradFunctor(alpha_ptr)); + Transform trans; + trans(context.device_context(), out_ptr, out_ptr + numel, dout_ptr, dx_ptr, + PReluGradFunctor(alpha_ptr)); - // TODO (Zhuoyuan): add dalpha upgrade when GPU kernels ready + // TODO(Zhuoyuan): add dalpha upgrade when GPU kernels ready } }; diff --git a/paddle/operators/rank_loss_op.cc b/paddle/operators/rank_loss_op.cc new file mode 100644 index 0000000000..1ba22006f2 --- /dev/null +++ b/paddle/operators/rank_loss_op.cc @@ -0,0 +1,122 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/rank_loss_op.h" + +namespace paddle { +namespace operators { + +class RankLossOp : public framework::OperatorWithKernel { + public: + RankLossOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + protected: + void InferShape(framework::InferShapeContextBase *ctx) const override { + // input check + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) shouldn't be null"); + PADDLE_ENFORCE(ctx->HasInput("Left"), "Input(Left) shouldn't be null"); + PADDLE_ENFORCE(ctx->HasInput("Right"), "Input(Right) shouldn't be null"); + + auto label_dims = ctx->GetInputDim("Label"); + auto left_dims = ctx->GetInputDim("Left"); + auto right_dims = ctx->GetInputDim("Right"); + + PADDLE_ENFORCE((label_dims == left_dims) && (left_dims == right_dims), + "All inputs must have the same size"); + PADDLE_ENFORCE((label_dims.size() == 2) && (label_dims[1] == 1), + "All inputs must be row vector with size batch_size x 1."); + ctx->SetOutputDim("Out", label_dims); + } +}; + +class RankLossOpMaker : public framework::OpProtoAndCheckerMaker { + public: + RankLossOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Label", + "The label indicating A ranked higher than B or not, row vector."); + AddInput("Left", "The output of RankNet for doc A, vector."); + AddInput("Right", "The output of RankNet for doc B, vetor"); + AddOutput("Out", "The output loss of RankLoss operator, vector."); + AddComment(R"DOC(RankLoss operator + +Rank loss operator for RankNet[1]. RankNet is a pairwise ranking model with +one training sample consisting of a pair of doc A and B, and the label P +indicating that A is ranked higher than B or not: + +P = {0, 1} or {0, 0.5, 1}, where 0.5 means no information about the rank of +the input pair. + +The RankLoss operator contains three inputs: Left (o_i), Right (o_j) and Label +(P_{i,j}), which represent the output of RankNet for two docs and the label +respectively, and yields the rank loss C_{i,j} by following the expression + +\f[ + C_{i,j} = -\tilde{P_{ij}} * o_{i,j} + log(1 + e^{o_{i,j}}) \\ + o_{i,j} = o_i - o_j \\ + \tilde{P_{i,j}} = \left \{0, 0.5, 1 \right \} \ or \ \left \{0, 1 \right \} +\f] + +The operator can take inputs of one sample or in batch. + +[1]. Chris Burges, Tal Shaked, Erin Renshaw, et al. Learning to + Rank using Gradient Descent. + http://icml.cc/2015/wp-content/uploads/2015/06/icml_ranking.pdf +)DOC"); + } +}; + +class RankLossGradOp : public framework::OperatorWithKernel { + public: + RankLossGradOp(const std::string &type, + const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + protected: + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) shouldn't be null."); + PADDLE_ENFORCE(ctx->HasInput("Left"), "Input(Left) shouldn't be null."); + PADDLE_ENFORCE(ctx->HasInput("Right"), "Input(Right) shouldn't be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) shouldn't be null."); + auto dims = ctx->GetInputDim("Left"); + auto left_grad_name = framework::GradVarName("Left"); + auto right_grad_name = framework::GradVarName("Right"); + + if (ctx->HasOutput(left_grad_name)) { + ctx->SetOutputDim(left_grad_name, dims); + } + + if (ctx->HasOutput(right_grad_name)) { + ctx->SetOutputDim(right_grad_name, dims); + } + } +}; + +} // namespace operators +} // namespace paddle +namespace ops = paddle::operators; + +REGISTER_OP(rank_loss, ops::RankLossOp, ops::RankLossOpMaker, rank_loss_grad, + ops::RankLossGradOp); +REGISTER_OP_CPU_KERNEL(rank_loss, + ops::RankLossKernel); +REGISTER_OP_CPU_KERNEL( + rank_loss_grad, ops::RankLossGradKernel); diff --git a/paddle/operators/rank_loss_op.cu b/paddle/operators/rank_loss_op.cu new file mode 100644 index 0000000000..779588ff36 --- /dev/null +++ b/paddle/operators/rank_loss_op.cu @@ -0,0 +1,22 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/rank_loss_op.h" + +REGISTER_OP_GPU_KERNEL( + rank_loss, + paddle::operators::RankLossKernel); +REGISTER_OP_GPU_KERNEL( + rank_loss_grad, + paddle::operators::RankLossGradKernel); diff --git a/paddle/operators/rank_loss_op.h b/paddle/operators/rank_loss_op.h new file mode 100644 index 0000000000..f184d6efcb --- /dev/null +++ b/paddle/operators/rank_loss_op.h @@ -0,0 +1,80 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class RankLossKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const { + auto* out_t = ctx.Output("Out"); + auto* label_t = ctx.Input("Label"); + auto* left_t = ctx.Input("Left"); + auto* right_t = ctx.Input("Right"); + out_t->mutable_data(ctx.GetPlace()); + + auto out = framework::EigenVector::Flatten(*out_t); + auto label = framework::EigenVector::Flatten(*label_t); + auto left = framework::EigenVector::Flatten(*left_t); + auto right = framework::EigenVector::Flatten(*right_t); + + auto& dev = ctx.GetEigenDevice(); + out.device(dev) = + (1. + (left - right).exp()).log() - label * (left - right); + } +}; + +template +class RankLossGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const { + auto* d_left_t = + ctx.Output(framework::GradVarName("Left")); + auto* d_right_t = + ctx.Output(framework::GradVarName("Right")); + + auto* d_out_t = ctx.Input(framework::GradVarName("Out")); + auto* label_t = ctx.Input("Label"); + auto* left_t = ctx.Input("Left"); + auto* right_t = ctx.Input("Right"); + + auto& dev = ctx.GetEigenDevice(); + auto d_out = framework::EigenVector::Flatten(*d_out_t); + auto label = framework::EigenVector::Flatten(*label_t); + auto left = framework::EigenVector::Flatten(*left_t); + auto right = framework::EigenVector::Flatten(*right_t); + + // compute d_left + if (d_left_t) { + d_left_t->mutable_data(ctx.GetPlace()); + auto d_left = framework::EigenVector::Flatten(*d_left_t); + d_left.device(dev) = d_out * (1. / (1. + (right - left).exp()) - label); + } + // compute d_right + if (d_right_t) { + d_right_t->mutable_data(ctx.GetPlace()); + auto d_right = framework::EigenVector::Flatten(*d_right_t); + d_right.device(dev) = + -d_out * (1.0 / (1. + (right - left).exp()) - label); + } + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/recurrent_op.cc b/paddle/operators/recurrent_op.cc index d3413d7cb9..04c4c24951 100644 --- a/paddle/operators/recurrent_op.cc +++ b/paddle/operators/recurrent_op.cc @@ -28,60 +28,41 @@ using Variable = framework::Variable; using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; -void RecurrentAlgorithm::InferShape(const Scope& scope) const { - seq_len_ = scope.FindVar((arg_->inlinks[0]).external) - ->GetMutable() - ->dims()[0]; - CreateScopes(scope); - auto step_scopes = GetStepScopes(scope); - rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, - true /*infer_shape_mode*/); - InitMemories(step_scopes[0], true /*infer_shape_mode*/); - - for (size_t i = 0; i < seq_len_; i++) { - if (i > 0) { - rnn::LinkMemories(step_scopes, arg_->memories, i, -1, - true /*infer_shape_mode*/); - } - (*stepnet_)->InferShape(*step_scopes[i]); - } - rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, - true /*infer_shape_mode*/); -} - void RecurrentAlgorithm::Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const { - auto step_scopes = GetStepScopes(scope); - rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, - false /*infer_shape_mode*/); - InitMemories(step_scopes[0], false /*infer_shape_mode*/); + auto* input0 = scope.FindVar(arg_->inlinks[0]); + PADDLE_ENFORCE_NOT_NULL(input0); + size_t seq_len = input0->GetMutable()->dims()[0]; + PADDLE_ENFORCE_GT(seq_len, 0); + + CreateScopes(scope, seq_len); + auto& step_scopes = GetStepScopes(scope); + rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len); + InitMemories(step_scopes[0]); - for (size_t step_id = 0; step_id < seq_len_; step_id++) { - // create output alias variables + for (size_t step_id = 0; step_id < seq_len; step_id++) { if (step_id > 0) { - rnn::LinkMemories(step_scopes, arg_->memories, step_id, -1, - false /*infer_shape_mode*/); + rnn::LinkMemories(step_scopes, arg_->memories, step_id, -1); } (*stepnet_)->Run(*step_scopes[step_id], dev_ctx); } - rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, - false /*infer_shape_mode*/); + rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len); } -void RecurrentAlgorithm::CreateScopes(const Scope& scope) const { +void RecurrentAlgorithm::CreateScopes(const Scope& scope, + size_t seq_len) const { // TODO(superjom) Only two scopes are needed for inference, this case will be // supported later. - auto step_scopes_var = scope.FindVar(arg_->step_scopes); + auto* step_scopes_var = scope.FindVar(arg_->step_scopes); PADDLE_ENFORCE(step_scopes_var != nullptr, ""); - auto step_scopes = step_scopes_var->GetMutable>(); + auto* step_scopes = step_scopes_var->GetMutable>(); // Now all variables in scope must be created outside of op. PADDLE_ENFORCE_NOT_NULL(stepnet_); PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "stepnet_ op has no outputs"); - PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "net_op has no outputs"); - if (seq_len_ > step_scopes->size()) { - for (size_t i = step_scopes->size(); i < seq_len_; ++i) { + if (seq_len > step_scopes->size()) { + for (size_t i = step_scopes->size(); i < seq_len; ++i) { auto& step_scope = scope.NewScope(); // create step net's temp inputs @@ -104,8 +85,7 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const { } } -void RecurrentAlgorithm::InitMemories(Scope* step_scope, - bool infer_shape_mode) const { +void RecurrentAlgorithm::InitMemories(Scope* step_scope) const { for (auto& attr : arg_->memories) { auto* pre_mem = step_scope->NewVar(attr.pre_var)->GetMutable(); PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr, @@ -113,24 +93,19 @@ void RecurrentAlgorithm::InitMemories(Scope* step_scope, attr.boot_var); auto* boot_mem = step_scope->FindVar(attr.boot_var)->GetMutable(); - if (infer_shape_mode) { - pre_mem->Resize(boot_mem->dims()); - PADDLE_ENFORCE_EQ(pre_mem->dims().size(), 2); - } else { - pre_mem->ShareDataWith(*boot_mem); - } + pre_mem->Resize(boot_mem->dims()); + PADDLE_ENFORCE_EQ(pre_mem->dims().size(), 2); + pre_mem->ShareDataWith(*boot_mem); } } const rnn::ArgumentName RecurrentOp::kArgName{ - "step_net", "step_scopes", "inlinks", - "outlinks", "inlink_alias", "outlink_alias", + "step_net", "step_scopes", "inlinks", "outlinks", "memories", "pre_memories", "boot_memories"}; const rnn::ArgumentName RecurrentGradientOp::kArgName{ - "step_net", "step_scopes", "outlink@grad", - "inlink@grad", "inlink_alias", "outlink_alias", - "memories", "pre_memories", "boot_memories@grad"}; + "step_net", "step_scopes@GRAD", "outlinks@GRAD", "inlinks@GRAD", + "memories", "pre_memories", "boot_memories@GRAD"}; RecurrentOp::RecurrentOp(const std::string& type, const framework::VariableNameMap& inputs, @@ -160,8 +135,6 @@ class RecurrentAlgorithmProtoAndCheckerMaker AddOutput(name.step_scopes, "step scopes"); // Attributes stored in AttributeMap - AddAttr>(name.inlink_alias, "alias of inlinks"); - AddAttr>(name.outlink_alias, "alias of outlinks"); AddAttr>(name.pre_memories, "names of pre-memories"); AddAttr>(name.memories, "names of memories"); @@ -172,23 +145,23 @@ class RecurrentAlgorithmProtoAndCheckerMaker void RecurrentGradientAlgorithm::Run( const Scope& scope, const platform::DeviceContext& dev_ctx) const { - auto step_scopes = GetStepScopes(scope); - rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, - false /*infer_shape_mode*/); - for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) { - if (static_cast(step_id) != seq_len_ - 1) { - rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1, - false /*infer_shape_mode*/); + auto* input0 = scope.FindVar(arg_->inlinks[0]); + PADDLE_ENFORCE_NOT_NULL(input0); + size_t seq_len = input0->GetMutable()->dims()[0]; + auto& step_scopes = GetStepScopes(scope); + rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len); + for (int step_id = seq_len - 1; step_id >= 0; --step_id) { + if (step_id != seq_len - 1) { + rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1); } (*stepnet_)->Run(*step_scopes[step_id], dev_ctx); } - LinkBootMemoryGradients(step_scopes[0], false); - rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, - false /*infer_shape_mode*/); + rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len); + LinkBootMemoryGradients(step_scopes[0]); } void RecurrentGradientAlgorithm::LinkBootMemoryGradients( - Scope* step_scope, bool infer_shape_mode) const { + Scope* step_scope) const { for (auto& attr : arg_->memories) { PADDLE_ENFORCE(step_scope->FindVar(attr.var) != nullptr, "memory variable [%s] does not exists", attr.var); @@ -197,31 +170,9 @@ void RecurrentGradientAlgorithm::LinkBootMemoryGradients( auto* mem_grad = step_scope->NewVar(attr.var)->GetMutable(); auto* boot_mem_grad = step_scope->NewVar(attr.boot_var)->GetMutable(); - if (infer_shape_mode) { - boot_mem_grad->Resize(mem_grad->dims()); - } else { - boot_mem_grad->ShareDataWith(*mem_grad); - } - } -} - -void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const { - seq_len_ = scope.FindVar((arg_->inlinks[0]).external) - ->GetMutable() - ->dims()[0]; - auto step_scopes = GetStepScopes(scope); - rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, - true /*infer_shape_mode*/); - for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) { - if (static_cast(step_id) != seq_len_ - 1) { - rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1, - true /*infer_shape_mode*/); - } - (*stepnet_)->InferShape(*step_scopes[step_id]); + boot_mem_grad->Resize(mem_grad->dims()); + boot_mem_grad->ShareDataWith(*mem_grad); } - rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, - true /*infer_shape_mode*/); - LinkBootMemoryGradients(step_scopes[0], true /*infer_shape_mode*/); } RecurrentGradientOp::RecurrentGradientOp( @@ -229,13 +180,13 @@ RecurrentGradientOp::RecurrentGradientOp( const framework::VariableNameMap& outputs, const framework::AttributeMap& attrs) : OperatorBase(type, inputs, outputs, attrs) { - rnn::InitArgument(kArgName, &arg_, *this); + rnn::InitArgument(kArgName, &arg_, *this, true /*is grad*/); alg_.Init(&arg_, &stepnet_); } } // namespace operators } // namespace paddle -REGISTER_OP_WITHOUT_GRADIENT( - recurrent, paddle::operators::RecurrentOp, - paddle::operators::RecurrentAlgorithmProtoAndCheckerMaker); +REGISTER_OP(recurrent, paddle::operators::RecurrentOp, + paddle::operators::RecurrentAlgorithmProtoAndCheckerMaker, + recurrent_grad, paddle::operators::RecurrentGradientOp); diff --git a/paddle/operators/recurrent_op.h b/paddle/operators/recurrent_op.h index 1033d657a3..253d7e3284 100644 --- a/paddle/operators/recurrent_op.h +++ b/paddle/operators/recurrent_op.h @@ -22,7 +22,7 @@ namespace paddle { namespace operators { // The sequence format in RecurrentOp is Tensor now. -// TODO(Yan Chunwei): +// TODO(Superjom) // 1. No-padding computing for sequences with indifinite length in one batch. // 2. Hierarchical RNN for sequence with sub-sequence. // 3. Internal Memory. @@ -41,11 +41,6 @@ class RecurrentAlgorithm { stepnet_ = stepnet; } - /** - * InferShape must be called before Run. - */ - void InferShape(const framework::Scope& scope) const; - protected: /* * The step scopes will be stored in the father scope as a variable. @@ -53,7 +48,7 @@ class RecurrentAlgorithm { * NOTE the scopes are reused in both the forward and backward, so just * create once and expand its size if more steps need. */ - void CreateScopes(const framework::Scope& scope) const; + void CreateScopes(const framework::Scope& scope, size_t seq_len) const; const std::vector& GetStepScopes( const framework::Scope& scope) const { @@ -61,12 +56,11 @@ class RecurrentAlgorithm { ->GetMutable>(); } - void InitMemories(framework::Scope* step_scopes, bool infer_shape_mode) const; + void InitMemories(framework::Scope* step_scopes) const; private: std::unique_ptr* stepnet_; rnn::Argument* arg_; - mutable size_t seq_len_; }; class RecurrentGradientAlgorithm { @@ -91,13 +85,7 @@ class RecurrentGradientAlgorithm { void Run(const framework::Scope& scope, const platform::DeviceContext& dev_ctx) const; - void LinkBootMemoryGradients(framework::Scope* step_scopes, - bool infer_shape_mode) const; - - /** - * InferShape must be called before Run. - */ - void InferShape(const framework::Scope& scope) const; + void LinkBootMemoryGradients(framework::Scope* step_scopes) const; protected: inline const std::vector& GetStepScopes( @@ -108,7 +96,6 @@ class RecurrentGradientAlgorithm { private: rnn::Argument* arg_; - mutable size_t seq_len_; std::unique_ptr* stepnet_; }; @@ -124,12 +111,6 @@ class RecurrentOp : public framework::OperatorBase { // TODO(yuyang18): Implement copy ctor well. PADDLE_THROW("Not implemented"); } - /** - * InferShape must be called before Run. - */ - void InferShape(const framework::Scope& scope) const override { - alg_.InferShape(scope); - } void Run(const framework::Scope& scope, const platform::DeviceContext& dev_ctx) const override { @@ -139,6 +120,7 @@ class RecurrentOp : public framework::OperatorBase { void set_stepnet(std::unique_ptr net) { stepnet_ = std::move(net); } + const OperatorBase& stepnet() const { return *stepnet_; } static const rnn::ArgumentName kArgName; @@ -163,13 +145,6 @@ class RecurrentGradientOp : public framework::OperatorBase { PADDLE_THROW("Not Implemented"); } - /** - * InferShape must be called before Run. - */ - void InferShape(const framework::Scope& scope) const override { - alg_.InferShape(scope); - } - void Run(const framework::Scope& scope, const platform::DeviceContext& dev_ctx) const override { alg_.Run(scope, dev_ctx); @@ -177,6 +152,9 @@ class RecurrentGradientOp : public framework::OperatorBase { static const rnn::ArgumentName kArgName; + /* + * set a stepnet that is created according to a RecurrentOp's stepnet. + */ void set_stepnet(std::unique_ptr net) { stepnet_ = std::move(net); } diff --git a/paddle/operators/reduce_op.cc b/paddle/operators/reduce_op.cc new file mode 100644 index 0000000000..3ef443d1c7 --- /dev/null +++ b/paddle/operators/reduce_op.cc @@ -0,0 +1,203 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/reduce_op.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; + +class ReduceOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ReduceOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ReduceOp should not be null."); + auto x_dims = ctx->GetInputDim("X"); + auto x_rank = x_dims.size(); + PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); + int dim = ctx->Attrs().Get("dim"); + if (dim < 0) dim = x_rank + dim; + PADDLE_ENFORCE_LT( + dim, x_rank, + "The dim should be in the range [-rank(input), rank(input))."); + bool keep_dim = ctx->Attrs().Get("keep_dim"); + auto dims_vector = vectorize(x_dims); + if (keep_dim || x_rank == 1) { + dims_vector[dim] = 1; + } else { + dims_vector.erase(dims_vector.begin() + dim); + } + auto out_dims = framework::make_ddim(dims_vector); + ctx->SetOutputDim("Out", out_dims); + if (dim != 0) { + // Only pass LoD when not reducing on the first dim. + ctx->ShareLoD("X", /*->*/ "Out"); + } + } +}; + +class ReduceGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null."); + auto x_dims = ctx->GetInputDim("X"); + auto x_rank = x_dims.size(); + PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); + int dim = ctx->Attrs().Get("dim"); + if (dim < 0) dim = x_rank + dim; + PADDLE_ENFORCE_LT( + dim, x_rank, + "The dim should be in the range [-rank(input), rank(input))."); + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); + } + } +}; + +class ReduceOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ReduceOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput( + "X", + "(Tensor) The input tensor. Tensors with rank at most 6 are supported"); + AddOutput("Out", "(Tensor) The result tensor."); + AddAttr( + "dim", + "(int, default 1) The dimension to reduce. " + "Must be in the range [-rank(input), rank(input)). " + "If `dim < 0`, the dim to reduce is `rank + dim`. " + "Noting that reducing on the first dim will make the LoD info lost.") + .SetDefault(0); + AddAttr("keep_dim", + "(bool, default false) " + "If true, retain the reduced dimension with length 1.") + .SetDefault(false); + comment_ = R"DOC( +{ReduceOP} operator computes the {reduce} of input tensor along the given dimension. +The result tensor has 1 fewer dimension than the input unless `keep_dim` is true. +)DOC"; + AddComment(comment_); + } + + protected: + std::string comment_; + + void Replace(std::string &src, std::string from, std::string to) { + std::size_t len_from = std::strlen(from.c_str()); + std::size_t len_to = std::strlen(to.c_str()); + for (std::size_t pos = src.find(from); pos != std::string::npos; + pos = src.find(from, pos + len_to)) { + src.replace(pos, len_from, to); + } + } + + void SetComment(std::string name, std::string op) { + Replace(comment_, "{ReduceOP}", name); + Replace(comment_, "{reduce}", op); + } +}; + +class ReduceSumOpMaker : public ReduceOpMaker { + public: + ReduceSumOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : ReduceOpMaker(proto, op_checker) { + SetComment("ReduceSum", "sum"); + AddComment(comment_); + } +}; + +class ReduceMeanOpMaker : public ReduceOpMaker { + public: + ReduceMeanOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : ReduceOpMaker(proto, op_checker) { + SetComment("ReduceMean", "mean"); + AddComment(comment_); + } +}; + +class ReduceMaxOpMaker : public ReduceOpMaker { + public: + ReduceMaxOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : ReduceOpMaker(proto, op_checker) { + SetComment("ReduceMax", "max"); + AddComment(comment_); + } +}; + +class ReduceMinOpMaker : public ReduceOpMaker { + public: + ReduceMinOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : ReduceOpMaker(proto, op_checker) { + SetComment("ReduceMin", "min"); + AddComment(comment_); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP(reduce_sum, ops::ReduceOp, ops::ReduceSumOpMaker, reduce_sum_grad, + ops::ReduceGradOp); +REGISTER_OP_CPU_KERNEL( + reduce_sum, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL(reduce_sum_grad, + ops::ReduceGradKernel); + +REGISTER_OP(reduce_mean, ops::ReduceOp, ops::ReduceMeanOpMaker, + reduce_mean_grad, ops::ReduceGradOp); +REGISTER_OP_CPU_KERNEL( + reduce_mean, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL(reduce_mean_grad, + ops::ReduceGradKernel); + +REGISTER_OP(reduce_max, ops::ReduceOp, ops::ReduceMaxOpMaker, reduce_max_grad, + ops::ReduceGradOp); +REGISTER_OP_CPU_KERNEL( + reduce_max, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL(reduce_max_grad, + ops::ReduceGradKernel); + +REGISTER_OP(reduce_min, ops::ReduceOp, ops::ReduceMaxOpMaker, reduce_min_grad, + ops::ReduceGradOp); +REGISTER_OP_CPU_KERNEL( + reduce_min, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL(reduce_min_grad, + ops::ReduceGradKernel); diff --git a/paddle/operators/reduce_op.cu b/paddle/operators/reduce_op.cu new file mode 100644 index 0000000000..595127b858 --- /dev/null +++ b/paddle/operators/reduce_op.cu @@ -0,0 +1,46 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/reduce_op.h" + +namespace ops = paddle::operators; + +REGISTER_OP_GPU_KERNEL( + reduce_sum, + ops::ReduceKernel); +REGISTER_OP_GPU_KERNEL(reduce_sum_grad, + ops::ReduceGradKernel); + +REGISTER_OP_GPU_KERNEL( + reduce_mean, + ops::ReduceKernel); +REGISTER_OP_GPU_KERNEL(reduce_mean_grad, + ops::ReduceGradKernel); + +REGISTER_OP_GPU_KERNEL( + reduce_max, + ops::ReduceKernel); +REGISTER_OP_GPU_KERNEL(reduce_max_grad, + ops::ReduceGradKernel); + +REGISTER_OP_GPU_KERNEL( + reduce_min, + ops::ReduceKernel); +REGISTER_OP_GPU_KERNEL(reduce_min_grad, + ops::ReduceGradKernel); diff --git a/paddle/operators/reduce_op.h b/paddle/operators/reduce_op.h new file mode 100644 index 0000000000..ba3f3db81d --- /dev/null +++ b/paddle/operators/reduce_op.h @@ -0,0 +1,200 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using DDim = framework::DDim; +template +using EigenTensor = framework::EigenTensor; + +struct SumFunctor { + template + void operator()(const Place& place, X& x, Y& y, const Dim& dim) { + y.device(place) = x.sum(dim); + } +}; + +struct SumGradFunctor { + template + void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy, + const Dim& dim, int size) { + dx.device(place) = dy.broadcast(dim); + } +}; + +struct MeanFunctor { + template + void operator()(const Place& place, X& x, Y& y, const Dim& dim) { + y.device(place) = x.mean(dim); + } +}; + +struct MeanGradFunctor { + template + void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy, + const Dim& dim, int size) { + dx.device(place) = dy.broadcast(dim) / dx.constant(size); + } +}; + +struct MaxFunctor { + template + void operator()(const Place& place, X& x, Y& y, const Dim& dim) { + y.device(place) = x.maximum(dim); + } +}; + +struct MinFunctor { + template + void operator()(const Place& place, X& x, Y& y, const Dim& dim) { + y.device(place) = x.minimum(dim); + } +}; + +struct MaxOrMinGradFunctor { + template + void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy, + const Dim& dim, int size) { + auto equals = x == y.broadcast(dim); + auto ones = dx.constant(1); + auto zeros = dx.constant(0); + // If there are multiple minimum or maximum elements, the subgradient of + // each is the set [0, 1], and we pass gradient to all of them here. + dx.device(place) = dy.broadcast(dim) * equals.select(ones, zeros); + } +}; + +template +class ReduceKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + int rank = context.Input("X")->dims().size(); + switch (rank) { + case 1: + ReduceCompute<1>(context); + break; + case 2: + ReduceCompute<2>(context); + break; + case 3: + ReduceCompute<3>(context); + break; + case 4: + ReduceCompute<4>(context); + break; + case 5: + ReduceCompute<5>(context); + break; + case 6: + ReduceCompute<6>(context); + break; + } + } + + private: + template + void ReduceCompute(const framework::ExecutionContext& context) const { + auto* input = context.Input("X"); + auto* output = context.Output("Out"); + output->mutable_data(context.GetPlace()); + + auto x = EigenTensor::From(*input); + auto x_rank = static_cast(x.dimensions().size()); + int dim = static_cast(context.Attr("dim")); + if (dim < 0) dim = x_rank + dim; + auto reduce_dim = Eigen::array({{dim}}); + // construct the squeezed output tensor + bool keep_dim = context.Attr("keep_dim"); + DDim dims = output->dims(); + auto dims_vector = vectorize(dims); + if (keep_dim && x_rank > 1) { + dims_vector.erase(dims_vector.begin() + dim); + dims = framework::make_ddim(dims_vector); + } + auto out = EigenTensor < T, D == 1 ? 1 : (D - 1) > ::From(*output, dims); + auto& place = context.GetEigenDevice(); + Functor functor; + functor(place, x, out, reduce_dim); + } +}; + +template +class ReduceGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + int rank = context.Input("X")->dims().size(); + switch (rank) { + case 1: + ReduceGradCompute<1>(context); + break; + case 2: + ReduceGradCompute<2>(context); + break; + case 3: + ReduceGradCompute<3>(context); + break; + case 4: + ReduceGradCompute<4>(context); + break; + case 5: + ReduceGradCompute<5>(context); + break; + case 6: + ReduceGradCompute<6>(context); + break; + } + } + + private: + template + void ReduceGradCompute(const framework::ExecutionContext& context) const { + auto* input0 = context.Input("X"); + auto* input1 = context.Input("Out"); + auto* input2 = context.Input(framework::GradVarName("Out")); + auto* output = context.Output(framework::GradVarName("X")); + + output->mutable_data(context.GetPlace()); + auto x = EigenTensor::From(*input0); + auto x_grad = EigenTensor::From(*output); + auto x_rank = static_cast(x.dimensions().size()); + int dim = static_cast(context.Attr("dim")); + if (dim < 0) dim = x_rank + dim; + DDim dims = input0->dims(); + dims[dim] = 1; + auto x_reduce = EigenTensor::From(*input1, dims); + auto x_reduce_grad = EigenTensor::From(*input2, dims); + + Eigen::array braodcast_dim; + for (size_t i = 0; i < D; ++i) braodcast_dim[i] = 1; + braodcast_dim[dim] = input0->dims()[dim]; + auto& place = context.GetEigenDevice(); + Functor functor; + functor(place, x, x_reduce, x_grad, x_reduce_grad, braodcast_dim, + braodcast_dim[dim]); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/reshape_op.cc b/paddle/operators/reshape_op.cc index 0d05e34414..a3c3fa2716 100644 --- a/paddle/operators/reshape_op.cc +++ b/paddle/operators/reshape_op.cc @@ -26,14 +26,14 @@ class ReshapeOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} protected: - void InferShape(const framework::InferShapeContext &ctx) const override { + void InferShape(framework::InferShapeContextBase *ctx) const override { // input check - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of ReshapeOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of ReshapeOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ReshapeOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ReshapeOp should not be null."); - auto shape = ctx.Attr>("shape"); + auto shape = ctx->Attrs().Get>("shape"); PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty."); for (auto dim : shape) { PADDLE_ENFORCE(dim > 0, "Each dimension of shape must be positive."); @@ -41,8 +41,8 @@ class ReshapeOp : public framework::OperatorWithKernel { // capacity check int64_t capacity = std::accumulate(shape.begin(), shape.end(), 1, std::multiplies()); - auto *in = ctx.Input("X"); - int64_t in_size = framework::product(in->dims()); + auto x_dims = ctx->GetInputDim("X"); + int64_t in_size = framework::product(x_dims); PADDLE_ENFORCE_EQ(capacity, in_size, "The size of Input(X) mismatches with Attr(shape)."); // resize output @@ -50,7 +50,12 @@ class ReshapeOp : public framework::OperatorWithKernel { std::transform(shape.begin(), shape.end(), shape_int64.begin(), [](int a) { return static_cast(a); }); auto out_dims = framework::make_ddim(shape_int64); - ctx.Output("Out")->Resize(out_dims); + ctx->SetOutputDim("Out", out_dims); + if (shape[0] == x_dims[0]) { + // Only pass LoD when the first dimension is equal between + // output and input. + ctx->ShareLoD("X", /*->*/ "Out"); + } } }; @@ -71,7 +76,7 @@ Given a 2-D tensor X with 2 rows and 2 columns [[1, 2], [3, 4]] -with target shape = [1, 4], the reshape operator will transform +with target shape = [1, 4], the reshape operator will transform the tensor X into a 1-D tensor: [1, 2, 3, 4] @@ -89,13 +94,11 @@ class ReshapeGradOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) shouldn't be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) shouldn't be null."); - auto dims = ctx.Input("X")->dims(); - auto *d_in = ctx.Output(framework::GradVarName("X")); - d_in->Resize(dims); + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) shouldn't be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) shouldn't be null."); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); } }; diff --git a/paddle/operators/reshape_op.h b/paddle/operators/reshape_op.h index 873acf3078..628dfe4c0f 100644 --- a/paddle/operators/reshape_op.h +++ b/paddle/operators/reshape_op.h @@ -21,7 +21,7 @@ namespace paddle { namespace operators { template -class ReshapeKernel : public framework::OpKernel { +class ReshapeKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { auto* out = ctx.Output("Out"); @@ -39,7 +39,7 @@ class ReshapeKernel : public framework::OpKernel { }; template -class ReshapeGradKernel : public framework::OpKernel { +class ReshapeGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { auto* d_out = ctx.Input(framework::GradVarName("Out")); diff --git a/paddle/operators/rmsprop_op.cc b/paddle/operators/rmsprop_op.cc new file mode 100644 index 0000000000..8f61c7fdda --- /dev/null +++ b/paddle/operators/rmsprop_op.cc @@ -0,0 +1,120 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/rmsprop_op.h" + +namespace paddle { +namespace operators { + +class RmspropOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Param"), + "Input(Param) of RmspropOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("MeanSquare"), + "Input(MeanSquare) of RmspropOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("LearningRate"), + "Input(LearningRate) of RmspropOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grad"), + "Input(Grad) of RmspropOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Moment"), + "Input(Moment) of RmspropOp should not be null."); + + PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), + "Output(param_out) of RmspropOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("MomentOut"), + "Output(Momentum_out) of RmspropOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("MeanSquareOut"), + "Output(MeanSquareOut) of RmspropOp should not be null."); + + auto param_dim = ctx->GetInputDim("Param"); + PADDLE_ENFORCE_EQ( + param_dim, ctx->GetInputDim("Grad"), + "Param and grad input of RmspropOp should have the same dimension."); + PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("Moment"), + "Param and Momentum input of RmspropOp " + "should have the same dimension."); + PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("MeanSquare"), + "Param and Momentum input of RmspropOp " + "should have the same dimension."); + + auto lr_dim = ctx->GetInputDim("LearningRate"); + PADDLE_ENFORCE_EQ(framework::product(lr_dim), 1, + "Learning Rate should be a scalar."); + + ctx->SetOutputDim("ParamOut", param_dim); + ctx->SetOutputDim("MomentOut", param_dim); + ctx->SetOutputDim("MeanSquareOut", param_dim); + } +}; + +class RmspropOpMaker : public framework::OpProtoAndCheckerMaker { + public: + RmspropOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Param", + "(Tensor, default Tensor) " + "Input parameter value that has to be updated"); + AddInput("MeanSquare", + "(Tensor, default Tensor)" + " The mean square value that gets updated"); + AddInput("LearningRate", + "(Tensor, default Tensor) " + "The learning rate should be a tensor of size 1"); + AddInput("Grad", + "(Tensor, default Tensor) " + "Input gradient of the parameter"); + AddInput("Moment", + "(Tensor, default Tensor) The moment that gets updated"); + + AddOutput("ParamOut", "(Tensor) Output updated parameter value"); + AddOutput("MomentOut", "(Tensor) Output updated moment"); + AddOutput("MeanSquareOut", "(Tensor) Output Mean squared updated value"); + + AddAttr("epsilon", + "(float, default 1e-10) Constant " + "for numerical stability.") + .SetDefault(1.0e-10f); + AddAttr("decay", + "(float, default 0.9) " + "Discounting factor for coming gradient.") + .SetDefault(0.9f); + AddAttr("momentum", "(float, default 0.0) Constant value") + .SetDefault(0.0f); + AddComment(R"DOC( + +RMSprop + +MeanSquareOut = decay * MeanSquare + (1 - decay) * Grad * Grad +MomentOut = momentum * Moment + + LearningRate * Grad / sqrt(MeanSquareOut + epsilon) +ParamOut = Param - MomentOut + +The original slides that proposed RMSprop: Slide 29 of +http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf) + +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(rmsprop, ops::RmspropOp, ops::RmspropOpMaker); +REGISTER_OP_CPU_KERNEL(rmsprop, + ops::RmspropOpKernel); diff --git a/paddle/operators/rmsprop_op.cu b/paddle/operators/rmsprop_op.cu new file mode 100644 index 0000000000..52634a5481 --- /dev/null +++ b/paddle/operators/rmsprop_op.cu @@ -0,0 +1,20 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/rmsprop_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(rmsprop, + ops::RmspropOpKernel); diff --git a/paddle/operators/rmsprop_op.h b/paddle/operators/rmsprop_op.h new file mode 100644 index 0000000000..7bf2129010 --- /dev/null +++ b/paddle/operators/rmsprop_op.h @@ -0,0 +1,67 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenVector = framework::EigenVector; + +template +class RmspropOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* param_out = ctx.Output("ParamOut"); + auto* moment_out = ctx.Output("MomentOut"); + auto* mean_square_out = ctx.Output("MeanSquareOut"); + + auto grad = ctx.Input("Grad"); + + param_out->mutable_data(ctx.GetPlace()); + moment_out->mutable_data(ctx.GetPlace()); + mean_square_out->mutable_data(ctx.GetPlace()); + + float epsilon = ctx.Attr("epsilon"); + float rho = ctx.Attr("decay"); + float momentum = ctx.Attr("momentum"); + + auto p = EigenVector::Flatten(*ctx.Input("Param")); + auto ms = EigenVector::Flatten(*ctx.Input("MeanSquare")); + auto lr = EigenVector::Flatten(*ctx.Input("LearningRate")); + auto g = EigenVector::Flatten(*grad); + auto mom = EigenVector::Flatten(*ctx.Input("Moment")); + + auto p_out = EigenVector::Flatten(*param_out); + auto mom_out = EigenVector::Flatten(*moment_out); + auto ms_out = EigenVector::Flatten(*mean_square_out); + auto place = ctx.GetEigenDevice(); + + Eigen::DSizes grad_dsize(grad->numel()); + + ms_out.device(place) = rho * ms + (1 - rho) * g * g; + mom_out.device(place) = + momentum * mom + + lr.broadcast(grad_dsize) * g / (ms_out + epsilon).sqrt(); + p_out.device(place) = p - mom_out; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/rnn/recurrent_op_utils.cc b/paddle/operators/rnn/recurrent_op_utils.cc index 6c082cb182..ef317a71f1 100644 --- a/paddle/operators/rnn/recurrent_op_utils.cc +++ b/paddle/operators/rnn/recurrent_op_utils.cc @@ -24,69 +24,62 @@ using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; void SegmentInputs(const std::vector& step_scopes, - const std::vector& inlinks, const size_t seq_len, - bool infer_shape_mode) { + const std::vector& inlinks, + const size_t seq_len) { PADDLE_ENFORCE(!inlinks.empty(), "no in links are provided."); for (size_t i = 0; i < inlinks.size(); ++i) { - auto input_var = step_scopes[0]->FindVar(inlinks[i].external); - PADDLE_ENFORCE(input_var != nullptr, "input link [%s] is not in scope.", - inlinks[i].external); + // global inputs + auto input_var = step_scopes[0]->parent().FindVar(inlinks[i]); + PADDLE_ENFORCE_NOT_NULL(input_var, "input link [%s] is not in scope.", + inlinks[i]); LoDTensor* input = input_var->GetMutable(); f::DDim dims = input->dims(); - PADDLE_ENFORCE(static_cast(dims[0]) == seq_len, - "all the inlinks must have same length"); + PADDLE_ENFORCE_EQ(static_cast(dims[0]), seq_len, + "all the inlinks be the same length"); f::DDim step_dims = slice_ddim(dims, 1, dims.size()); for (size_t j = 0; j < seq_len; j++) { Tensor* step_input = - step_scopes[j]->NewVar(inlinks[i].internal)->GetMutable(); - if (!infer_shape_mode) { - // The input of operators of each step is Tensor here. - // Maybe need to modify Slice function. - *step_input = input->Slice(j, j + 1); - } + step_scopes[j]->NewVar(inlinks[i])->GetMutable(); + // The input of operators of each step is Tensor here. + // Maybe need to modify Slice function. + *step_input = input->Slice(j, j + 1); step_input->Resize(step_dims); } } } void ConcatOutputs(const std::vector& step_scopes, - const std::vector& outlinks, const size_t seq_len, - bool infer_shape_mode) { + const std::vector& outlinks, + const size_t seq_len) { for (size_t i = 0; i < outlinks.size(); i++) { - auto output_var = step_scopes[0]->FindVar(outlinks[i].external); - PADDLE_ENFORCE(output_var != nullptr, "output link [%s] is not in scope.", - outlinks[i].external); + auto* output_var = step_scopes[0]->parent().FindVar(outlinks[i]); + PADDLE_ENFORCE_NOT_NULL(output_var, "output link [%s] is not in scope.", + outlinks[i]); LoDTensor* output = output_var->GetMutable(); - if (infer_shape_mode) { - auto step_scope_var = step_scopes[0]->FindVar(outlinks[i].internal); - PADDLE_ENFORCE(step_scope_var != nullptr, "%s not in scope", - outlinks[i].internal); - f::DDim step_dims = - step_scope_var->template GetMutable()->dims(); - std::vector dims_vec = vectorize(step_dims); - dims_vec.insert(dims_vec.begin(), seq_len); - output->Resize(f::make_ddim(dims_vec)); - } else { - output->mutable_data(platform::CPUPlace()); - for (size_t j = 0; j < seq_len; j++) { - LoDTensor* step_output = step_scopes[j] - ->FindVar(outlinks[i].internal) - ->GetMutable(); - // TODO(luotao02) data type and platform::DeviceContext() should set - // correctly - (output->Slice(j, j + 1)) - .CopyFrom(*step_output, platform::CPUPlace()); - } + auto* step_scope_var = step_scopes[0]->FindVar(outlinks[i]); + PADDLE_ENFORCE_NOT_NULL(step_scope_var, "%s not in scope", outlinks[i]); + f::DDim step_dims = + step_scope_var->template GetMutable()->dims(); + std::vector dims_vec = vectorize(step_dims); + dims_vec.insert(dims_vec.begin(), seq_len); + output->Resize(f::make_ddim(dims_vec)); + output->mutable_data(platform::CPUPlace()); + for (size_t j = 0; j < seq_len; j++) { + LoDTensor* step_output = + step_scopes[j]->FindVar(outlinks[i])->GetMutable(); + // TODO(luotao02) data type and platform::DeviceContext() should set + // correctly + (output->Slice(j, j + 1)) + .CopyFrom(*step_output, platform::CPUPlace()); } } } void LinkMemories(const std::vector& scopes, const std::vector& memories, - const size_t step_id, const int offset, - bool infer_shape_mode) { + const size_t step_id, const int offset) { PADDLE_ENFORCE_LT(step_id, scopes.size(), "step [%d] is out of range of step scopes' size [%d]", step_id, scopes.size()); @@ -96,52 +89,28 @@ void LinkMemories(const std::vector& scopes, step_id + offset, scopes.size(), "offset [%d] is out of range, it must be less than (%d - %d)", offset, scopes.size(), step_id); - auto scope = scopes[step_id]; - auto linked_scope = scopes[step_id + offset]; + auto* scope = scopes[step_id]; + auto* linked_scope = scopes[step_id + offset]; for (auto& attr : memories) { - auto mem = scope->FindVar(attr.pre_var)->GetMutable(); - auto linked_mem = linked_scope->FindVar(attr.var)->GetMutable(); - if (infer_shape_mode) { - mem->Resize(linked_mem->dims()); - } else { - mem->ShareDataWith(*linked_mem); - } + auto* mem = scope->FindVar(attr.pre_var)->GetMutable(); + auto* linked_mem = linked_scope->FindVar(attr.var)->GetMutable(); + mem->Resize(linked_mem->dims()); + mem->ShareDataWith(*linked_mem); } } void InitArgument(const ArgumentName& name, Argument* arg, - const framework::OperatorBase& op) { - arg->step_scopes = op.Output(name.step_scopes); - - auto inlinks = op.Inputs(name.inlinks); - auto inlink_alias = op.Attr>(name.inlink_alias); - PADDLE_ENFORCE(inlinks.size() == inlink_alias.size(), - "the size of inlinks and inlink_alias don't match:%d,%d", - inlinks.size(), inlink_alias.size()); - for (size_t i = 0; i < inlinks.size(); ++i) { - rnn::Link link; - link.external = inlinks[i]; - link.internal = inlink_alias[i]; - (arg->inlinks).push_back(link); - } - - auto outlinks = op.Outputs(name.outlinks); - auto outlink_alias = op.Attr>(name.outlink_alias); - PADDLE_ENFORCE(outlinks.size() == outlink_alias.size(), - "the size of outlinks and outlink_alias don't match:%d,%d", - outlinks.size(), outlink_alias.size()); - for (size_t i = 0; i < outlinks.size(); ++i) { - rnn::Link link; - link.external = outlinks[i]; - link.internal = outlink_alias[i]; - (arg->outlinks).push_back(link); - } - - auto boot_memories = op.Inputs(name.boot_memories); - + const framework::OperatorBase& op, bool is_grad) { + arg->step_scopes = + is_grad ? op.Input(name.step_scopes) : op.Output(name.step_scopes); + arg->inlinks = op.Inputs(name.inlinks); + arg->outlinks = op.Outputs(name.outlinks); + + auto& boot_memories = + is_grad ? op.Outputs(name.boot_memories) : op.Inputs(name.boot_memories); // attributes - auto memories = op.Attr>(name.memories); - auto pre_memories = op.Attr>(name.pre_memories); + auto& memories = op.Attr>(name.memories); + auto& pre_memories = op.Attr>(name.pre_memories); PADDLE_ENFORCE(memories.size() == boot_memories.size(), "the size of memories, boot_memories don't match:%d,%d", diff --git a/paddle/operators/rnn/recurrent_op_utils.h b/paddle/operators/rnn/recurrent_op_utils.h index 17941c503c..fd17b9b889 100644 --- a/paddle/operators/rnn/recurrent_op_utils.h +++ b/paddle/operators/rnn/recurrent_op_utils.h @@ -41,18 +41,11 @@ struct MemoryAttr { std::string boot_var; }; -struct Link { - // input or output links name. - std::string internal; - // alias to avoid duplicate keys in scopes. - std::string external; -}; - struct Argument { std::string step_net; std::string step_scopes; - std::vector inlinks; - std::vector outlinks; + std::vector inlinks; + std::vector outlinks; std::vector memories; }; @@ -61,8 +54,6 @@ struct ArgumentName { std::string step_scopes; std::string inlinks; std::string outlinks; - std::string inlink_alias; // the alias of inlinks in step net. - std::string outlink_alias; // the alias of outlinks in step net. std::string memories; // the memory name std::string pre_memories; // the previous memory name std::string boot_memories; // the boot memory name @@ -72,22 +63,22 @@ struct ArgumentName { * Prepare inputs for each step net. */ void SegmentInputs(const std::vector& step_scopes, - const std::vector& inlinks, const size_t seq_len, - bool infer_shape_mode); + const std::vector& inlinks, + const size_t seq_len); /** * Process outputs of step nets and merge to variables. */ void ConcatOutputs(const std::vector& step_scopes, - const std::vector& outlinks, const size_t seq_len, - bool infer_shape_mode); + const std::vector& outlinks, + const size_t seq_len); void LinkMemories(const std::vector& step_scopes, const std::vector& memories, const size_t step_id, - const int offset, bool infer_shape_mode); + const int offset); void InitArgument(const ArgumentName& name, Argument* arg, - const framework::OperatorBase& op); + const framework::OperatorBase& op, bool is_grad = false); } // namespace rnn } // namespace operators diff --git a/paddle/operators/rowwise_add_op.cc b/paddle/operators/rowwise_add_op.cc deleted file mode 100644 index 2a3fd3be94..0000000000 --- a/paddle/operators/rowwise_add_op.cc +++ /dev/null @@ -1,103 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ - -#include "paddle/operators/rowwise_add_op.h" - -namespace paddle { -namespace operators { - -using framework::Tensor; - -class RowwiseAddOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of RowwiseAddOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("b"), - "Input(b) of RowwiseAddOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of RowwiseAddOp should not be null."); - - auto x_dims = ctx.Input("X")->dims(); - auto b_dims = ctx.Input("b")->dims(); - PADDLE_ENFORCE_GT( - x_dims.size(), b_dims.size(), - "The rank of input `X` must be larger than the one of input `b`."); - - int num_col_dims = x_dims.size() - b_dims.size(); - - PADDLE_ENFORCE_EQ( - framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims, - "The width of two operands must be same"); - PADDLE_ENFORCE_EQ(ctx.OutputSize("Out"), 1, "The output size must be 1"); - ctx.Output("Out")->Resize(x_dims); - } -}; - -class RowwiseAddOpMaker : public framework::OpProtoAndCheckerMaker { - public: - RowwiseAddOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The left input of row-wise add op, must be matrix"); - AddInput("b", "The right input of row-wise add op, must be vector"); - AddOutput("Out", "The output of row-wise add op"); - AddComment(R"DOC(Row-wise Add operator - -for i in xrange(X.shape[0]): - Out = X[i] + b -)DOC"); - } -}; -class RowwiseAddGradOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "X should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("b"), "b should not be null"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null"); - auto x_dims = ctx.Input("X")->dims(); - auto b_dims = ctx.Input("b")->dims(); - PADDLE_ENFORCE_GT( - x_dims.size(), b_dims.size(), - "The rank of input `X` must be larger than the one of input `b`."); - - int num_col_dims = x_dims.size() - b_dims.size(); - PADDLE_ENFORCE_EQ( - framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims, - "The width of two operands must be same"); - auto *dx = ctx.Output(framework::GradVarName("X")); - auto *db = ctx.Output(framework::GradVarName("b")); - if (dx) dx->Resize(x_dims); - if (db) db->Resize(b_dims); - } -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; -REGISTER_OP(rowwise_add, ops::RowwiseAddOp, ops::RowwiseAddOpMaker, - rowwise_add_grad, ops::RowwiseAddGradOp); -REGISTER_OP_CPU_KERNEL( - rowwise_add, ops::RowwiseAddKernel); -REGISTER_OP_CPU_KERNEL( - rowwise_add_grad, - ops::RowwiseAddGradKernel); diff --git a/paddle/operators/rowwise_add_op.h b/paddle/operators/rowwise_add_op.h deleted file mode 100644 index 35774b9409..0000000000 --- a/paddle/operators/rowwise_add_op.h +++ /dev/null @@ -1,80 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include "paddle/framework/eigen.h" -#include "paddle/framework/op_registry.h" - -namespace paddle { -namespace operators { - -using Tensor = framework::Tensor; -template -using EigenVector = framework::EigenVector; -template -using EigenMatrix = framework::EigenMatrix; - -template -class RowwiseAddKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto out = context.Output("Out"); - out->mutable_data(context.GetPlace()); - int num_col_dims = context.Input("X")->dims().size() - - context.Input("b")->dims().size(); - auto input = - EigenMatrix::Reshape(*context.Input("X"), num_col_dims); - auto bias = EigenVector::Flatten(*context.Input("b")); - auto output = EigenMatrix::Reshape(*out, num_col_dims); - - const int bias_size = bias.dimension(0); - const int rest_size = input.size() / bias_size; - Eigen::DSizes one_d(input.size()); - Eigen::DSizes bcast(rest_size); - output.reshape(one_d).device(context.GetEigenDevice()) = - input.reshape(one_d) + bias.broadcast(bcast).reshape(one_d); - } -}; - -template -class RowwiseAddGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* dout = context.Input(framework::GradVarName("Out")); - auto* dx = context.Output(framework::GradVarName("X")); - auto* db = context.Output(framework::GradVarName("b")); - int num_col_dims = context.Input("X")->dims().size() - - context.Input("b")->dims().size(); - - auto out_grad = EigenMatrix::Reshape(*dout, num_col_dims); - auto place = context.GetEigenDevice(); - - if (dx) { - dx->mutable_data(context.GetPlace()); - EigenMatrix::Reshape(*dx, num_col_dims).device(place) = out_grad; - } - - if (db) { - db->mutable_data(context.GetPlace()); - // https://eigen.tuxfamily.org/dox/unsupported/TensorBase_8h_source.html - // colwise add - Eigen::array dims{{0}}; /* dimension to reduce */ - EigenVector::Flatten(*db).device(place) = out_grad.sum(dims); - } - } -}; -} // namespace operators -} // namespace paddle diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc index d1f42e8662..e225aecc27 100644 --- a/paddle/operators/scale_op.cc +++ b/paddle/operators/scale_op.cc @@ -26,15 +26,13 @@ class ScaleOp : public framework::OperatorWithKernel { : OperatorWithKernel(type, inputs, outputs, attrs) {} protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of ScaleOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of ScaleOp should not be null."); - - auto *in = ctx.Input("X"); - auto *out = ctx.Output("Out"); - out->Resize(in->dims()); + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ScaleOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ScaleOp should not be null."); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Out"); } }; @@ -43,8 +41,8 @@ class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { public: ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The input tensor of scale operator.").NotInGradient(); - AddOutput("Out", "The output tensor of scale operator.").NotInGradient(); + AddInput("X", "The input tensor of scale operator."); + AddOutput("Out", "The output tensor of scale operator."); AddComment(R"DOC(Scale operator The equation is: Out = scale*X @@ -54,21 +52,18 @@ The equation is: Out = scale*X } }; -// The operator to calculate gradients of a scale operator is just the scale -// operator itself. -// Grad(Out=scale(X)) => Grad(X) = scale(Grad(Out)) -template -class ScaleGradOp : public NetOp { +class ScaleGradMaker : public framework::SingleGradOpDescMaker { public: - ScaleGradOp(const std::string &type, const framework::VariableNameMap &inputs, - const framework::VariableNameMap &outputs, - const framework::AttributeMap &attrs) - : NetOp(type, inputs, outputs, attrs) { - AppendOp(framework::OpRegistry::CreateOp( - "scale", {{"X", {Input(framework::GradVarName("Out"))}}}, - {{"Out", {Output(framework::GradVarName("X"))}}}, - {{"scale", Attr("scale")}})); - CompleteAddOp(false); + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDescBind(); + grad_op->SetType("scale"); + grad_op->SetInput("X", OutputGrad("Out")); + grad_op->SetOutput("Out", InputGrad("X")); + grad_op->SetAttr("scale", GetAttr("scale")); + return std::unique_ptr(grad_op); } }; @@ -77,7 +72,7 @@ class ScaleGradOp : public NetOp { namespace ops = paddle::operators; -REGISTER_OP(scale, ops::ScaleOp, ops::ScaleOpMaker, scale_grad, - ops::ScaleGradOp); +REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker, + ops::ScaleGradMaker); REGISTER_OP_CPU_KERNEL(scale, ops::ScaleKernel); diff --git a/paddle/operators/scale_op.h b/paddle/operators/scale_op.h index 02fbdc52bb..dc6bc76899 100644 --- a/paddle/operators/scale_op.h +++ b/paddle/operators/scale_op.h @@ -20,7 +20,7 @@ namespace paddle { namespace operators { template -class ScaleKernel : public framework::OpKernel { +class ScaleKernel : public framework::OpKernel { public: virtual void Compute(const framework::ExecutionContext& context) const { auto* tensor = context.Output("Out"); diff --git a/paddle/operators/scatter.cu.h b/paddle/operators/scatter.cu.h new file mode 100644 index 0000000000..d95436be4f --- /dev/null +++ b/paddle/operators/scatter.cu.h @@ -0,0 +1,80 @@ +/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +#define CUDA_1D_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \ + i += blockDim.x * gridDim.x) + +template +__global__ void ScatterCUDAKernel(const T* params, const int* indices, + T* output, size_t index_size, + size_t slice_size) { + CUDA_1D_KERNEL_LOOP(i, index_size * slice_size) { + int indices_i = i / slice_size; + int slice_i = i - indices_i * slice_size; // offset inside the slice + int scatter_i = indices[indices_i]; + int out_i = scatter_i * slice_size + slice_i; + *(output + out_i) = *(params + i); + } +} + +/** + * A thin wrapper on gpu tensor + * Return a new updated tensor from source tensor, scatter-assigned according to + * index + * input[src]: type-T source Tensor + * input[index]: type-int index Tensor (1-D) + * return: output tensor + */ +template +void GPUScatterAssign(const platform::DeviceContext& ctx, const Tensor& src, + const Tensor& index, Tensor* output) { + // PADDLE_ENFORCE(platform::is_gpu_place(place)); + // check index of shape 1-D + PADDLE_ENFORCE(index.dims().size() == 1); + int index_size = index.dims()[0]; + + auto src_dims = src.dims(); + framework::DDim output_dims(src_dims); + output_dims[0] = index_size; + + // slice size + int slice_size = 1; + for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i]; + + const T* p_src = src.data(); + const int* p_index = index.data(); + T* p_output = output->data(); + + int block = 512; + int n = slice_size * index_size; + int grid = (n + block - 1) / block; + + ScatterCUDAKernel<<< + grid, block, 0, + reinterpret_cast(ctx).stream()>>>( + p_src, p_index, p_output, index_size, slice_size); +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/scatter.h b/paddle/operators/scatter.h index 6b542675c2..c1fb844ebd 100644 --- a/paddle/operators/scatter.h +++ b/paddle/operators/scatter.h @@ -24,67 +24,42 @@ namespace paddle { namespace operators { using Tensor = framework::Tensor; -template -using EigenVector = framework::EigenVector; - -// Implementation of CPU copy -template -void CPUScatterUpdate(const paddle::framework::Tensor* src, const int* index, - const size_t index_size, - paddle::framework::Tensor* output) { - paddle::framework::DDim output_dims = output->dims(); - - for (size_t i = 0; i < index_size; ++i) { - int index_ = index[i]; - - paddle::framework::Tensor src_ = *src; - paddle::framework::Tensor output_ = *output; - if (index_size > 1) src_ = src->Slice(i, i + 1); - if (output_dims[0] > 1) output_ = output->Slice(index_, index_ + 1); - - auto X = EigenVector::Flatten(src_); - auto Y = EigenVector::Flatten(output_); - - Y = X + Y; - } -} - -// Implementation of GPU scatter: -template -void GPUScatterUpdate(const T* src, const int* index, const int slice_size, - const int index_size, T* output); /** * Return a updated tensor from source tensor, scattered according to index: - * dst[i] += src[index[i]] + * dst[i] = src[index[i]] * input[src]: type-T source Tensor * input[index]: type-int index Tensor (1-D) * return: output tensor */ template -void ScatterUpdate(const platform::Place& place, - const paddle::framework::Tensor* src, - const paddle::framework::Tensor* index, - paddle::framework::Tensor* output) { +void ScatterAssign(const platform::DeviceContext& ctx, const Tensor& src, + const Tensor& index, Tensor* output) { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace())); // check index of shape 1-D - PADDLE_ENFORCE(index->dims().size() == 1); - int index_size = index->dims()[0]; + PADDLE_ENFORCE(index.dims().size() == 1); + int index_size = index.dims()[0]; - auto src_dims = src->dims(); + auto src_dims = src.dims(); auto dst_dims = output->dims(); + const T* p_src = src.data(); + const int* p_index = index.data(); + T* p_output = output->data(); + // check src shape and dst shape should match for (int i = 1; i < src_dims.size(); i++) PADDLE_ENFORCE(src_dims[i] == dst_dims[i]); // slice size size_t slice_size = 1; - for (int i = 0; i < src_dims.size(); ++i) slice_size *= src_dims[i]; + for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i]; + + const size_t slice_bytes = slice_size * sizeof(T); - if (platform::is_cpu_place(place)) { - CPUScatterUpdate(src, index->data(), index_size, output); - } else { + for (int i = 0; i < index_size; ++i) { + int index_ = p_index[i]; + memcpy(p_output + index_ * slice_size, p_src + i * slice_size, slice_bytes); } } diff --git a/paddle/operators/scatter_op.cc b/paddle/operators/scatter_op.cc index 8820262732..d15ba15153 100644 --- a/paddle/operators/scatter_op.cc +++ b/paddle/operators/scatter_op.cc @@ -23,29 +23,35 @@ class ScatterOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Ref"), - "Input(Ref) of ScatterOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"), - "Input(Index) of ScatterOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Updates"), - "Input(Updates) of ScatterOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of ScatterOp should not be null."); - - PADDLE_ENFORCE_EQ(ctx.Input("Index")->dims().size(), 1, + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Ref"), + "Input(Ref) of ScatterOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Index"), + "Input(Index) of ScatterOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Updates"), + "Input(Updates) of ScatterOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ScatterOp should not be null."); + + auto updates_dims = ctx->GetInputDim("Updates"); + auto ref_dims = ctx->GetInputDim("Ref"); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Index").size(), 1, "Update Index should be 1-D."); - PADDLE_ENFORCE_EQ(ctx.Input("Ref")->dims().size(), - ctx.Input("Updates")->dims().size(), + PADDLE_ENFORCE_EQ(ref_dims.size(), updates_dims.size(), "Reference and Updates should have the same shape size"); - PADDLE_ENFORCE_EQ(ctx.Input("Updates")->dims()[0], - ctx.Input("Index")->dims()[0], + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Updates")[0], + ctx->GetInputDim("Index")[0], "Updates and Index should have same batch-size."); - framework::DDim data_dim(ctx.Input("Updates")->dims()); - for (int i = 1; i < data_dim.size(); ++i) - PADDLE_ENFORCE_EQ(data_dim[i], ctx.Input("Updates")->dims()[i]); - ctx.Output("Out")->Resize( - ctx.Input("Ref")->dims()); + framework::DDim data_dim(updates_dims); + for (int i = 1; i < data_dim.size(); ++i) { + PADDLE_ENFORCE_EQ(data_dim[i], updates_dims[i]); + } + ctx->SetOutputDim("Out", ref_dims); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("Ref")->type()); } }; @@ -54,23 +60,22 @@ class ScatterGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - auto *dUpdates = - ctx.Output(framework::GradVarName("Updates")); - auto *Updates = ctx.Input("Updates"); - auto *dRef = - ctx.Output(framework::GradVarName("Ref")); - auto *Ref = ctx.Input("Ref"); - - dRef->Resize(Ref->dims()); - dUpdates->Resize(Updates->dims()); + void InferShape(framework::InferShapeContextBase* ctx) const override { + ctx->SetOutputDim(framework::GradVarName("Updates"), + ctx->GetInputDim("Updates")); + ctx->SetOutputDim(framework::GradVarName("Ref"), ctx->GetInputDim("Ref")); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("Ref")->type()); } }; class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { public: - ScatterOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + ScatterOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("Ref", "The source input of scatter op"); AddInput("Index", @@ -78,21 +83,19 @@ class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Updates", "The updated value of updates op"); AddOutput("Out", "The output of add op"); AddComment(R"DOC( -Scatter Operator by selecting from the first axis, +Scatter Operator by selecting from the first axis, Out = Ref Out[Index] = Ref[Index] + Updates )DOC"); } }; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OP(scatter, ops::ScatterOp, ops::ScatterOpMaker, scatter_grad, ops::ScatterGradOp); -REGISTER_OP_CPU_KERNEL(scatter, - ops::ScatterOpKernel); -REGISTER_OP_CPU_KERNEL( - scatter_grad, - ops::ScatterGradientOpKernel); +REGISTER_OP_CPU_KERNEL(scatter, ops::ScatterOpKernel); +REGISTER_OP_CPU_KERNEL(scatter_grad, ops::ScatterGradientOpKernel); diff --git a/paddle/operators/scatter_op.cu b/paddle/operators/scatter_op.cu new file mode 100644 index 0000000000..06f4d75944 --- /dev/null +++ b/paddle/operators/scatter_op.cu @@ -0,0 +1,63 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "gather.cu.h" +#include "paddle/operators/gather_op.h" +#include "scatter.cu.h" + +namespace paddle { +namespace operators { + +template +class ScatterOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "This kernel only runs on GPU device."); + auto *Ref = ctx.Input("Ref"); + auto *Index = ctx.Input("Index"); + auto *Updates = ctx.Input("Updates"); + auto *Out = ctx.Output("Out"); + + Out->ShareDataWith(*Ref); + + GPUScatterAssign(ctx.device_context(), *Updates, *Index, Out); + } +}; + +template +class ScatterGradOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "This kernel only runs on GPU device."); + auto *dRef = ctx.Output(framework::GradVarName("Ref")); + auto *dUpdates = ctx.Output(framework::GradVarName("Updates")); + auto *Index = ctx.Input("Index"); + auto *dOut = ctx.Input(framework::GradVarName("Out")); + + // In place gradient: dRef = dO + dRef->ShareDataWith(*dOut); + dUpdates->mutable_data(ctx.GetPlace()); + // Gradient by Gather: dUpdates = dO[Index] + GPUGather(ctx.device_context(), *dOut, *Index, dUpdates); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(scatter, ops::ScatterOpCUDAKernel); +REGISTER_OP_GPU_KERNEL(scatter_grad, ops::ScatterGradOpCUDAKernel); diff --git a/paddle/operators/scatter_op.h b/paddle/operators/scatter_op.h index e9595638a8..6101219006 100644 --- a/paddle/operators/scatter_op.h +++ b/paddle/operators/scatter_op.h @@ -23,10 +23,12 @@ namespace operators { using Tensor = framework::Tensor; -template -class ScatterOpKernel : public framework::OpKernel { +template +class ScatterOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "This kernel only runs on CPU."); auto *Ref = ctx.Input("Ref"); auto *Index = ctx.Input("Index"); auto *Updates = ctx.Input("Updates"); @@ -35,14 +37,16 @@ class ScatterOpKernel : public framework::OpKernel { // In place output: Out = Ref, Out[Index] += Updates Out->ShareDataWith(*Ref); // Apply ScatterUpdate: Out[index] += Updates[:] - ScatterUpdate(ctx.GetPlace(), Updates, Index, Out); + ScatterAssign(ctx.device_context(), *Updates, *Index, Out); } }; -template -class ScatterGradientOpKernel : public framework::OpKernel { +template +class ScatterGradientOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "This kernel only runs on CPU."); auto *dRef = ctx.Output(framework::GradVarName("Ref")); auto *dUpdates = ctx.Output(framework::GradVarName("Updates")); auto *Index = ctx.Input("Index"); @@ -52,7 +56,7 @@ class ScatterGradientOpKernel : public framework::OpKernel { dRef->ShareDataWith(*dOut); dUpdates->mutable_data(ctx.GetPlace()); // Gradient by Gather: dUpdates += dO[Index] - Gather(ctx.GetPlace(), dOut, Index, dUpdates); + CPUGather(ctx.device_context(), *dOut, *Index, dUpdates); } }; diff --git a/paddle/operators/scatter_test.cc b/paddle/operators/scatter_test.cc index 26fdaff146..00dbdacbfe 100644 --- a/paddle/operators/scatter_test.cc +++ b/paddle/operators/scatter_test.cc @@ -40,7 +40,9 @@ TEST(scatter, ScatterUpdate) { float* p_output = output->mutable_data(make_ddim({4, 4}), CPUPlace()); - ScatterUpdate(CPUPlace(), src, index, output); + auto* cpu_place = new paddle::platform::CPUPlace(); + paddle::platform::CPUDeviceContext ctx(*cpu_place); + ScatterAssign(ctx, *src, *index, output); for (size_t i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], float(0)); for (size_t i = 0; i < 4; ++i) EXPECT_EQ(output->data()[i], float(0)); diff --git a/paddle/operators/sequence_avg_pool_op.cc b/paddle/operators/sequence_avg_pool_op.cc deleted file mode 100644 index 9815b8f3a8..0000000000 --- a/paddle/operators/sequence_avg_pool_op.cc +++ /dev/null @@ -1,95 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/operators/sequence_avg_pool_op.h" - -namespace paddle { -namespace operators { - -class SequenceAvgPoolOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL( - ctx.InputVar("X"), "Input(X) of SequenceAvgPoolOp should not be null."); - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Out"), - "Output(Out) of SequenceAvgPoolOp should not be null."); - - auto* x = ctx.Input("X"); - auto dims = x->dims(); - auto lod = x->lod(); - PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now."); - PADDLE_ENFORCE_GE( - dims[0], - /*batch size = */ static_cast(lod[0].size() - 1), - "The first dimension of Input(X) must be large than batch size."); - dims[0] = lod[0].size() - 1; - ctx.Output("Out")->Resize({dims}); - } -}; - -class SequenceAvgPoolOpMaker : public framework::OpProtoAndCheckerMaker { - public: - SequenceAvgPoolOpMaker(framework::OpProto* proto, - framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "Input of SequenceAvgPoolOp."); - AddOutput("Out", "The output of SequenceAvgPoolOp."); - AddComment(R"DOC( - SequenceAvgPoolOp averages features of all time-steps of each instance. - More detailed comments will be added later. - )DOC"); - } -}; - -class SequenceAvgPoolGradOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Gradient of Out should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "The input X should not be null."); - auto og_dims = - ctx.Input(framework::GradVarName("Out"))->dims(); - auto x_dims = ctx.Input("X")->dims(); - PADDLE_ENFORCE_EQ(og_dims.size(), x_dims.size(), - "The rank of output grad must equal to Input(X)."); - for (int64_t i = 1; i < og_dims.size(); ++i) { - PADDLE_ENFORCE_EQ(og_dims[i], x_dims[i], "The dimension mismatch."); - } - auto* x_grad = - ctx.Output(framework::GradVarName("X")); - x_grad->Resize(x_dims); - } -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; -REGISTER_OP(sequence_avg_pool, ops::SequenceAvgPoolOp, - ops::SequenceAvgPoolOpMaker, sequence_avg_pool_grad, - ops::SequenceAvgPoolGradOp); -REGISTER_OP_CPU_KERNEL( - sequence_avg_pool, - ops::SequenceAvgPoolKernel); -REGISTER_OP_CPU_KERNEL( - sequence_avg_pool_grad, - ops::SequenceAvgPoolGradKernel); diff --git a/paddle/operators/sequence_pool_op.cc b/paddle/operators/sequence_pool_op.cc new file mode 100644 index 0000000000..bc4af2f704 --- /dev/null +++ b/paddle/operators/sequence_pool_op.cc @@ -0,0 +1,102 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/sequence_pool_op.h" + +namespace paddle { +namespace operators { + +class SequencePoolOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequencePoolOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SequencePoolOp should not be null."); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + } +}; + +class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SequencePoolOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "A float LoDTensor, the variable-length input of SequencePoolOp"); + AddOutput( + "Out", + "A float LoDTensor, the variable-length output of SequencePoolOp."); + AddAttr( + "strategy", + "(int, default AVERAGE) the pooling strategy of SequencePoolOp.") + .SetDefault(AVERAGE) + .InEnum({AVERAGE, SUM, SQRT, MAX, LAST, FIRST}); + AddComment(R"DOC( + SequencePoolOp pools features of all time-steps of each instance. + + For a mini-batch of 3 variable lengths sentences, containing 2, 3, and 2 time-steps: + + Assume X is a [7,M,N] float LoDTensor, and X->lod()[0] = [0, 2, 5, 7]. + Besides, for the sake of simplicity, we assume M=1 and N=1, + and the value of X = [[1, 3], [2, 4, 6], [5, 1]]. + + Thus, Out is a [3,1,1] float LoDTensor, but Out->lod() is nullptr. + And for different strategy, the value of Out is as follows: + + - AVERAGE: [2, 4, 3], where 2=(1+3)/2, 4=(2+4+6)/3, 3=(5+1)/2 + - SUM: [4, 12, 6], where 4=1+3, 12=2+4+6, 6=5+1 + - SQRT: [2.82, 6.93, 4.24], where 2.82=(1+3)/sqrt(2), + 6.93=(2+4+6)/sqrt(3), 4.24=(5+1)/sqrt(2) + - MAX: [3, 6, 5], where 3=max(1,3), 6=max(2,4,6), 5=max(5,1) + - LAST: [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1) + - FIRST: [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1) + )DOC"); + } +}; + +class SequencePoolGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Gradient of Out should not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), "The input X should not be null."); + auto og_dims = ctx->GetInputDim(framework::GradVarName("Out")); + auto x_dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE_EQ(og_dims.size(), x_dims.size(), + "The rank of output grad must equal to Input(X)."); + for (int64_t i = 1; i < og_dims.size(); ++i) { + PADDLE_ENFORCE_EQ(og_dims[i], x_dims[i], "The dimension mismatch."); + } + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(sequence_pool, ops::SequencePoolOp, ops::SequencePoolOpMaker, + sequence_pool_grad, ops::SequencePoolGradOp); +REGISTER_OP_CPU_KERNEL( + sequence_pool, ops::SequencePoolKernel); +REGISTER_OP_CPU_KERNEL( + sequence_pool_grad, + ops::SequencePoolGradKernel); diff --git a/paddle/operators/rowwise_add_op.cu b/paddle/operators/sequence_pool_op.cu similarity index 76% rename from paddle/operators/rowwise_add_op.cu rename to paddle/operators/sequence_pool_op.cu index 4a57f64c89..66850772d5 100644 --- a/paddle/operators/rowwise_add_op.cu +++ b/paddle/operators/sequence_pool_op.cu @@ -13,11 +13,12 @@ limitations under the License. */ #define EIGEN_USE_GPU -#include "paddle/operators/rowwise_add_op.h" + +#include "paddle/operators/sequence_pool_op.h" namespace ops = paddle::operators; REGISTER_OP_GPU_KERNEL( - rowwise_add, ops::RowwiseAddKernel); + sequence_pool, ops::SequencePoolKernel); REGISTER_OP_GPU_KERNEL( - rowwise_add_grad, - ops::RowwiseAddGradKernel); + sequence_pool_grad, + ops::SequencePoolGradKernel); diff --git a/paddle/operators/sequence_avg_pool_op.h b/paddle/operators/sequence_pool_op.h similarity index 56% rename from paddle/operators/sequence_avg_pool_op.h rename to paddle/operators/sequence_pool_op.h index ebe0956344..752d714125 100644 --- a/paddle/operators/sequence_avg_pool_op.h +++ b/paddle/operators/sequence_pool_op.h @@ -28,54 +28,96 @@ template using EigenMatrix = framework::EigenMatrix; +enum SeqPoolType { + AVERAGE = 0, + SUM = 1, + SQRT = 2, // square_root_n + MAX = 3, + LAST = 4, + FIRST = 5 +}; + template -class SequenceAvgPoolKernel : public framework::OpKernel { +class SequencePoolKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in = context.Input("X"); auto* out = context.Output("Out"); + int strategy = context.Attr("strategy"); auto dims = in->dims(); auto lod = in->lod(); int64_t w = in->numel() / dims[0]; + // InferShape by lod + PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now."); + PADDLE_ENFORCE_GE( + dims[0], + /*batch size = */ static_cast(lod[0].size() - 1), + "The first dimension of Input(X) must be large than batch size."); + dims[0] = lod[0].size() - 1; + out->Resize({dims}); + + auto lod_level_0 = lod[0]; + out->mutable_data(context.GetPlace()); auto place = context.GetEigenDevice(); - for (int i = 0; i < static_cast(lod[0].size()) - 1; ++i) { - Tensor in_t = in->Slice(static_cast(lod[0][i]), - static_cast(lod[0][i + 1])); + for (int i = 0; i < static_cast(lod_level_0.size()) - 1; ++i) { + Tensor in_t = in->Slice(static_cast(lod_level_0[i]), + static_cast(lod_level_0[i + 1])); Tensor out_t = out->Slice(i, i + 1); - int64_t h = static_cast(lod[0][i + 1] - lod[0][i]); + int64_t h = static_cast(lod_level_0[i + 1] - lod_level_0[i]); auto in_e = EigenMatrix::From(in_t, framework::make_ddim({h, w})); auto out_e = EigenVector::Flatten(out_t); - out_e.device(place) = in_e.mean(Eigen::array({{0}})); + + switch (strategy) { + case AVERAGE: + out_e.device(place) = in_e.mean(Eigen::array({{0}})); + break; + case SUM: + out_e.device(place) = in_e.sum(Eigen::array({{0}})); + break; + default: + PADDLE_THROW("unsupported pooling strategy"); + } } } }; template -class SequenceAvgPoolGradKernel : public framework::OpKernel { +class SequencePoolGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in = context.Input("X"); auto* out_g = context.Input(framework::GradVarName("Out")); auto* in_g = context.Output(framework::GradVarName("X")); + int strategy = context.Attr("strategy"); auto dims = in->dims(); - auto lod = in->lod(); + auto lod = in->lod()[0]; int64_t w = in->numel() / dims[0]; in_g->mutable_data(context.GetPlace()); auto place = context.GetEigenDevice(); - for (int i = 0; i < static_cast(lod[0].size()) - 1; ++i) { - auto in_g_t = in_g->Slice(static_cast(lod[0][i]), - static_cast(lod[0][i + 1])); + for (int i = 0; i < static_cast(lod.size()) - 1; ++i) { + auto in_g_t = in_g->Slice(static_cast(lod[i]), + static_cast(lod[i + 1])); auto out_g_t = out_g->Slice(i, i + 1); - int64_t h = static_cast(lod[0][i + 1] - lod[0][i]); + int64_t h = static_cast(lod[i + 1] - lod[i]); auto in_g_e = EigenMatrix::From(in_g_t, {h, w}); auto out_g_e = EigenMatrix::From(out_g_t, {1, w}); Eigen::DSizes bcast(h, 1); - in_g_e.device(place) = (out_g_e / static_cast(h)).broadcast(bcast); + + switch (strategy) { + case AVERAGE: + in_g_e.device(place) = (out_g_e / static_cast(h)).broadcast(bcast); + break; + case SUM: + in_g_e.device(place) = (out_g_e).broadcast(bcast); + break; + default: + PADDLE_THROW("unsupported pooling strategy"); + } } } }; diff --git a/paddle/operators/sequence_softmax_op.cc b/paddle/operators/sequence_softmax_op.cc new file mode 100644 index 0000000000..621779ab61 --- /dev/null +++ b/paddle/operators/sequence_softmax_op.cc @@ -0,0 +1,103 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/sequence_softmax_op.h" + +namespace paddle { +namespace operators { + +class SequenceSoftmaxOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequenceSoftmaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SequenceSoftmaxOp should not be null."); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +class SequenceSoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SequenceSoftmaxOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "(LoDTensor) 1-D or 2-D input LoDTensor with the 2-nd dimension " + "of length 1."); + AddOutput("Out", + "(LoDTensor) 1-D or 2-D output LoDTensor with the 2-nd dimension " + "of length 1."); + AddComment(R"DOC( +SequenceSoftmaxOp computes softmax activation among all time-steps for each +sequence. The dimension of each time-step should be 1. Thus, the shape of +input Tensor can be either [N, 1] or [N], where N is the sum of all sequences' +lengths. + +Equation: + for i-th sequence in a mini-batch: + Out(X[lod[i]:lod[i+1]], :) = + exp(X[lod[i]:lod[i+1], :]) / sum(exp(X[lod[i]:lod[i+1], :])) + +For example, for a mini-batch of 3 sequences with variable-length, +each containing 2, 3, 2 time-steps, the lod of which is [0, 2, 5, 7], +then softmax will be computed among X[0:2, :], X[2:5, :], X[5:7, :] +and N turns out to be 7. +)DOC"); + } +}; + +class SequenceSoftmaxGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Out"), + "Input(Out) of SequenceSoftmaxGradOp should not be null."); + PADDLE_ENFORCE( + ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) of SequenceSoftmaxGradOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequenceSoftmaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Output(X@GRAD) of SequenceSoftmaxOp should not be null."); + + PADDLE_ENFORCE_EQ( + ctx->GetInputDim("Out"), + ctx->GetInputDim(framework::GradVarName("Out")), + "Input(Out) and Input(Out@GRAD) of SequenceSoftmaxGradOp should be of " + "the same shape."); + + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(sequence_softmax, ops::SequenceSoftmaxOp, + ops::SequenceSoftmaxOpMaker, sequence_softmax_grad, + ops::SequenceSoftmaxGradOp); +REGISTER_OP_CPU_KERNEL( + sequence_softmax, + ops::SequenceSoftmaxKernel); +REGISTER_OP_CPU_KERNEL( + sequence_softmax_grad, + ops::SequenceSoftmaxGradKernel); diff --git a/paddle/operators/sequence_softmax_op.cu b/paddle/operators/sequence_softmax_op.cu new file mode 100644 index 0000000000..f2a1e3d5e3 --- /dev/null +++ b/paddle/operators/sequence_softmax_op.cu @@ -0,0 +1,25 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#define EIGEN_USE_GPU + +#include "paddle/operators/sequence_softmax_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL( + sequence_softmax, + ops::SequenceSoftmaxKernel) +REGISTER_OP_GPU_KERNEL( + sequence_softmax_grad, + ops::SequenceSoftmaxGradKernel); diff --git a/paddle/operators/sequence_softmax_op.h b/paddle/operators/sequence_softmax_op.h new file mode 100644 index 0000000000..96d87c404d --- /dev/null +++ b/paddle/operators/sequence_softmax_op.h @@ -0,0 +1,94 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/math/softmax.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; + +template +class SequenceSoftmaxKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* out = ctx.Output("Out"); + + auto lod = x->lod(); + auto dims = x->dims(); + + const size_t level = lod.size() - 1; + PADDLE_ENFORCE_EQ(dims[0], static_cast(lod[level].back()), + "The first dimension of Input(X) should be equal to the " + "sum of all sequences' lengths."); + PADDLE_ENFORCE_EQ(dims[0], x->numel(), + "The width of each timestep in Input(X) of " + "SequenceSoftmaxOp should be 1."); + + out->mutable_data(ctx.GetPlace()); + for (int i = 0; i < static_cast(lod[level].size()) - 1; ++i) { + int start_pos = static_cast(lod[level][i]); + int end_pos = static_cast(lod[level][i + 1]); + Tensor x_i = x->Slice(start_pos, end_pos); + Tensor out_i = out->Slice(start_pos, end_pos); + + // Reshape from (end_pos - start_pos) x 1UL to 1UL x (end_pos - start_pos) + framework::DDim dims_i = framework::make_ddim({1UL, end_pos - start_pos}); + x_i.Resize(dims_i); + out_i.Resize(dims_i); + math::SoftmaxFunctor()(ctx.device_context(), &x_i, &out_i); + } + } +}; + +template +class SequenceSoftmaxGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* out = ctx.Input("Out"); + auto* out_grad = ctx.Input(framework::GradVarName("Out")); + auto* x = ctx.Input("X"); + auto* x_grad = ctx.Output(framework::GradVarName("X")); + + auto lod = x->lod(); + const size_t level = lod.size() - 1; + + x_grad->mutable_data(ctx.GetPlace()); + for (int i = 0; i < static_cast(lod[level].size()) - 1; ++i) { + int start_pos = static_cast(lod[level][i]); + int end_pos = static_cast(lod[level][i + 1]); + + Tensor out_i = out->Slice(start_pos, end_pos); + Tensor out_grad_i = out_grad->Slice(start_pos, end_pos); + Tensor x_grad_i = x_grad->Slice(start_pos, end_pos); + + // Reshape from (end_pos - start_pos) x 1UL to 1UL x (end_pos - start_pos) + framework::DDim dims_i = framework::make_ddim({1UL, end_pos - start_pos}); + out_i.Resize(dims_i); + out_grad_i.Resize(dims_i); + x_grad_i.Resize(dims_i); + math::SoftmaxGradFunctor()(ctx.device_context(), &out_i, + &out_grad_i, &x_grad_i); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc index 1232e64c7f..31d491f130 100644 --- a/paddle/operators/sgd_op.cc +++ b/paddle/operators/sgd_op.cc @@ -22,19 +22,23 @@ class SGDOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("param"), - "Input(param) of SGDOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("grad"), - "Input(grad) of SGDOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("param_out"), - "Output(param_out) of SGDOp should not be null."); + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Param"), + "Input(Param) of SGDOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grad"), + "Input(Grad) of SGDOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("LearningRate"), + "Input(LearningRate) of SGDOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), + "Output(ParamOut) of SGDOp should not be null."); - PADDLE_ENFORCE_EQ(ctx.Input("param")->dims(), - ctx.Input("grad")->dims(), + auto lr_dims = ctx->GetInputDim("LearningRate"); + PADDLE_ENFORCE_EQ(framework::product(lr_dims), 1, + "Learning rate should have 1 element"); + auto param_dim = ctx->GetInputDim("Param"); + PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("Grad"), "Two input of SGD Op's dimension must be same."); - ctx.Output("param_out") - ->Resize(ctx.Input("param")->dims()); + ctx->SetOutputDim("ParamOut", param_dim); } }; @@ -42,10 +46,10 @@ class SGDOpMaker : public framework::OpProtoAndCheckerMaker { public: SGDOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("param", "input parameter"); - AddInput("grad", "input gradient"); - AddOutput("param_out", "output parameter"); - AddAttr("learning_rate", "learning rate of sgd"); + AddInput("Param", "Input parameter"); + AddInput("LearningRate", "Learning rate of SGD"); + AddInput("Grad", "Input gradient"); + AddOutput("ParamOut", "output parameter"); AddComment(R"DOC( Simplest sgd algorithm. diff --git a/paddle/operators/sgd_op.h b/paddle/operators/sgd_op.h index f8888f9c36..26f4012f25 100644 --- a/paddle/operators/sgd_op.h +++ b/paddle/operators/sgd_op.h @@ -19,28 +19,25 @@ limitations under the License. */ namespace paddle { namespace operators { -using Tensor = framework::Tensor; -template -using EigenVector = framework::EigenVector; - template -class SGDOpKernel : public framework::OpKernel { +class SGDOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - auto param = ctx.Input("param"); - auto grad = ctx.Input("grad"); - auto param_out = ctx.Output("param_out"); - float lr = ctx.Attr("learning_rate"); + auto param = ctx.Input("Param"); + auto grad = ctx.Input("Grad"); + auto param_out = ctx.Output("ParamOut"); + auto learning_rate = ctx.Input("LearningRate"); param_out->mutable_data(ctx.GetPlace()); - auto p = EigenVector::Flatten(*param); - auto g = EigenVector::Flatten(*grad); - auto o = EigenVector::Flatten(*param_out); + auto p = framework::EigenVector::Flatten(*param); + auto g = framework::EigenVector::Flatten(*grad); + auto o = framework::EigenVector::Flatten(*param_out); + auto lr = framework::EigenVector::Flatten(*learning_rate); auto place = ctx.GetEigenDevice(); - o.device(place) = p - lr * g; + Eigen::DSizes grad_dsize(grad->numel()); + o.device(place) = p - lr.broadcast(grad_dsize) * g; } }; diff --git a/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc new file mode 100644 index 0000000000..ede458e011 --- /dev/null +++ b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cc @@ -0,0 +1,150 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/sigmoid_cross_entropy_with_logits_op.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; + +class SigmoidCrossEntropyWithLogitsOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input(Labels) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should be not null."); + + auto x_dims = ctx->GetInputDim("X"); + auto labels_dims = ctx->GetInputDim("Labels"); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); + PADDLE_ENFORCE_EQ(labels_dims.size(), 2, + "Input(Labels)'s rank should be 2."); + PADDLE_ENFORCE_EQ(x_dims[0], labels_dims[0], + "The 1st dimension of Input(X) and Input(Labels) should " + "be equal."); + PADDLE_ENFORCE_EQ(x_dims[1], labels_dims[1], + "The 2nd dimension of Input(X) and Input(Labels) should " + "be equal."); + + ctx->SetOutputDim("Out", x_dims); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +class SigmoidCrossEntropyWithLogitsGradOp + : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input(Labels) should be not null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) shoudl be not null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), + "Output(X@GRAD) should be not null."); + + auto x_dims = ctx->GetInputDim("X"); + auto labels_dims = ctx->GetInputDim("Labels"); + auto dout_dims = ctx->GetInputDim(framework::GradVarName("Out")); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2."); + PADDLE_ENFORCE_EQ(labels_dims.size(), 2, + "Input(Labels)'s rank should be 2."); + PADDLE_ENFORCE_EQ(dout_dims.size(), 2, + "Input(Out@Grad)'s rank should be 2."); + PADDLE_ENFORCE_EQ(x_dims[0], labels_dims[0], + "The 1st dimension of Input(X) and Input(Labels) should " + "be equal."); + PADDLE_ENFORCE_EQ(x_dims[1], labels_dims[1], + "The 2nd dimension of Input(X) and Input(Labels) should " + "be equal."); + PADDLE_ENFORCE_EQ(x_dims[0], dout_dims[0], + "The 1st dimension of Input(X) and Input(Out@Grad) " + "should be equal."); + PADDLE_ENFORCE_EQ(x_dims[1], dout_dims[1], + "The 2nd dimension of Input(X) and Input(Out@Grad) " + "should be equal."); + + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + } +}; + +class SigmoidCrossEntropyWithLogitsOpMaker + : public framework::OpProtoAndCheckerMaker { + public: + SigmoidCrossEntropyWithLogitsOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "(Tensor, default Tensor), a 2-D tensor with shape N x D, " + "where N is the batch size and D is the number of classes. " + "This input is a tensor of logits computed by the previous " + " operator. Logits are unscaled log probabilities given as " + "log(p/(1-p))."); + AddInput("Labels", + "(Tensor, default Tensor), a 2-D tensor of the same type " + "and shape as X. This input is a tensor of probabalistic labels " + "for each logit"); + AddOutput("Out", + "(Tensor, default Tensor), a 2-D tensor with shape N x D " + " of elementwise logistic losses."); + AddComment(R"DOC( +SigmoidCrossEntropyWithLogits Operator. + +This measures the elementwise probability error in discrete classification tasks +in which each class is independent. This can be thought of as predicting labels +for a data-point that are not mutually exclusive. For example, a news article +can be about politics, technology or sports at the same time or none of these. + +The logistic loss is given as follows: + + loss = -Labels * log(sigmoid(X)) - (1 - Labels) * log(1 - sigmoid(X)) + +We know that sigmoid(X) = (1 / (1 + exp(-X))). By substituting this we get + + loss = X - X * Labels + log(1 + exp(-X)) + +For stability and to prevent overflow of exp(-X) when X < 0, +we can reformulate the loss as follows: + + loss = max(X, 0) - X * Labels + log(1 + exp(-abs(X))) + +Both the input `X` and `Labels` can carry the LoD (Level of Details) information. +However the output only shares the LoD with input `X`. +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(sigmoid_cross_entropy_with_logits, + ops::SigmoidCrossEntropyWithLogitsOp, + ops::SigmoidCrossEntropyWithLogitsOpMaker, + sigmoid_cross_entropy_with_logits_grad, + ops::SigmoidCrossEntropyWithLogitsGradOp); +REGISTER_OP_CPU_KERNEL(sigmoid_cross_entropy_with_logits, + ops::SigmoidCrossEntropyWithLogitsKernel< + paddle::platform::CPUPlace, float>); +REGISTER_OP_CPU_KERNEL(sigmoid_cross_entropy_with_logits_grad, + ops::SigmoidCrossEntropyWithLogitsGradKernel< + paddle::platform::CPUPlace, float>); diff --git a/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu new file mode 100644 index 0000000000..32a39956a1 --- /dev/null +++ b/paddle/operators/sigmoid_cross_entropy_with_logits_op.cu @@ -0,0 +1,24 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/sigmoid_cross_entropy_with_logits_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(sigmoid_cross_entropy_with_logits, + ops::SigmoidCrossEntropyWithLogitsKernel< + paddle::platform::GPUPlace, float>); +REGISTER_OP_GPU_KERNEL(sigmoid_cross_entropy_with_logits_grad, + ops::SigmoidCrossEntropyWithLogitsGradKernel< + paddle::platform::GPUPlace, float>); diff --git a/paddle/operators/sigmoid_cross_entropy_with_logits_op.h b/paddle/operators/sigmoid_cross_entropy_with_logits_op.h new file mode 100644 index 0000000000..41c619f181 --- /dev/null +++ b/paddle/operators/sigmoid_cross_entropy_with_logits_op.h @@ -0,0 +1,75 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +// Out = max(X, 0) - X * Labels + log(1 + exp(-abs(X))) +template +class SigmoidCrossEntropyWithLogitsKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + const framework::Tensor *X = context.Input("X"); + const framework::Tensor *Labels = + context.Input("Labels"); + framework::Tensor *Out = context.Output("Out"); + Out->mutable_data(context.GetPlace()); + + auto x = framework::EigenVector::Flatten(*X); + auto labels = framework::EigenVector::Flatten(*Labels); + auto out = framework::EigenVector::Flatten(*Out); + auto place = context.GetEigenDevice(); + + // term1 = max(x, 0) + auto term1 = x.cwiseMax(static_cast(0)); + // term2 = x * labels + auto term2 = x * labels; + // term3 = log(1 + exp(-abs(x))) + auto term3 = (static_cast(1) + (-(x.abs())).exp()).log(); + + out.device(place) = term1 - term2 + term3; + } +}; + +// dX = sigmoid(X) - labels +template +class SigmoidCrossEntropyWithLogitsGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + const framework::Tensor *X = context.Input("X"); + const framework::Tensor *Labels = + context.Input("Labels"); + const framework::Tensor *dOut = + context.Input(framework::GradVarName("Out")); + framework::Tensor *dX = + context.Output(framework::GradVarName("X")); + dX->mutable_data(context.GetPlace()); + + auto x = framework::EigenVector::Flatten(*X); + auto labels = framework::EigenVector::Flatten(*Labels); + auto dout = framework::EigenVector::Flatten(*dOut); + auto dx = framework::EigenVector::Flatten(*dX); + auto place = context.GetEigenDevice(); + + auto sigmoid_x = static_cast(1) / (static_cast(1) + (-x).exp()); + dx.device(place) = dout * (sigmoid_x - labels); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/sigmoid_op.cc b/paddle/operators/sigmoid_op.cc deleted file mode 100644 index 992b19965e..0000000000 --- a/paddle/operators/sigmoid_op.cc +++ /dev/null @@ -1,67 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ - -#include "paddle/operators/sigmoid_op.h" - -namespace paddle { -namespace operators { - -class SigmoidOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of SigmoidOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"), - "Output(Y) of SigmoidOp should not be null."); - - ctx.Output("Y")->Resize( - ctx.Input("X")->dims()); - } -}; - -class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker { - public: - SigmoidOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "sigmoid input"); - AddOutput("Y", "sigmoid output"); - AddComment("Sigmoid function"); - } -}; - -class SigmoidOpGrad : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - ctx.Output(framework::GradVarName("X")) - ->Resize(ctx.Input("Y")->dims()); - } -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; -REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker, sigmoid_grad, - ops::SigmoidOpGrad); -REGISTER_OP_CPU_KERNEL(sigmoid, - ops::SigmoidKernel); -REGISTER_OP_CPU_KERNEL( - sigmoid_grad, ops::SigmoidGradKernel); diff --git a/paddle/operators/sigmoid_op.h b/paddle/operators/sigmoid_op.h deleted file mode 100644 index b01a9b3f23..0000000000 --- a/paddle/operators/sigmoid_op.h +++ /dev/null @@ -1,62 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ - -#pragma once -#include "paddle/framework/eigen.h" -#include "paddle/framework/op_registry.h" - -namespace paddle { -namespace operators { - -using Tensor = framework::Tensor; -template -using EigenVector = framework::EigenVector; - -template -class SigmoidKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto input = context.Input("X"); - auto output = context.Output("Y"); - output->mutable_data(context.GetPlace()); - - // The clipping is used in Paddle's raw implenmention - auto X = EigenVector::Flatten(*input); - auto Y = EigenVector::Flatten(*output); - auto place = context.GetEigenDevice(); - - Y.device(place) = 1. / (1. + (-X).exp()); - } -}; - -template -class SigmoidGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto Y_t = context.Input("Y"); - auto dY_t = context.Input(framework::GradVarName("Y")); - auto dX_t = context.Output(framework::GradVarName("X")); - - dX_t->mutable_data(context.GetPlace()); - - auto dX = EigenVector::Flatten(*dX_t); - auto Y = EigenVector::Flatten(*Y_t); - auto dY = EigenVector::Flatten(*dY_t); - dX.device(context.GetEigenDevice()) = dY * Y * (1. - Y); - } -}; - -} // namespace operators -} // namespace paddle diff --git a/paddle/operators/smooth_l1_loss_op.cc b/paddle/operators/smooth_l1_loss_op.cc new file mode 100644 index 0000000000..2d197e3b1b --- /dev/null +++ b/paddle/operators/smooth_l1_loss_op.cc @@ -0,0 +1,131 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/smooth_l1_loss_op.h" + +namespace paddle { +namespace operators { + +class SmoothL1LossOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "X must be initialized."); + PADDLE_ENFORCE(ctx->HasInput("Y"), "Y must be initialized."); + + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); + PADDLE_ENFORCE_EQ(x_dims, y_dims, "The shape of X and Y must be the same."); + PADDLE_ENFORCE_GE(x_dims.size(), 2, + "The tensor rank of X must be at least 2."); + if (ctx->HasInput("InsideWeight")) { + PADDLE_ENFORCE(ctx->HasInput("OutsideWeight"), + "If weights are provided, must specify both " + "inside and outside weights."); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("InsideWeight"), x_dims, + "The shape of InsideWeight must be same as X."); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("OutsideWeight"), x_dims, + "The shape of OutsideWeight must be same as X."); + } + + ctx->SetOutputDim("Diff", x_dims); + // loss is a two-rank tensor + ctx->SetOutputDim("Out", {x_dims[0], 1}); + } +}; + +template +class SmoothL1LossOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SmoothL1LossOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "The input tensor of smooth l1 loss op." + "The rank should be greater or equal to 2 with shape " + "[batch_size, value_dim1, value_dim2, ..., value_dimN]"); + AddInput("Y", + "The target tensor of smooth l1 loss op " + "with the same shape as X."); + AddInput("InsideWeight", + "Optional input tensor of smooth l1 loss op with the same shape " + "as X. If provided, the result of (X - Y) will be multiplied " + "by this tensor element by element."); + AddInput("OutsideWeight", + "Optinal input of smooth l1 loss op with the same shape as X." + "If provided, the output smooth l1 loss will be multiplied by " + "this tensor element by element."); + AddOutput("Diff", "Intermediate variable to cache InsideWeight*(X-Y).") + .AsIntermediate(); + AddOutput("Out", "Smooth l1 loss."); + AddAttr("sigma", + "Hyper parameter of smooth l1 loss op." + "A float scalar with default value 3.0.") + .SetDefault(3.0); + AddComment(R"DOC( +Compute smooth l1 loss for input and target. The operator take the 1st +dimension of input as batch size. For each instance, it will compute +smooth l1 loss element by element first and sum all losses to one value. +So the output shape is [batch_size, 1]. + +The equation is: +loss = 0.5 * (sigma * (x-y))^2 if abs(x - y) < 1 / sigma^2 + abs(x - y) - 0.5 / sigma^2 otherwise + +)DOC"); + } +}; + +class SmoothL1LossGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + auto in_dims = ctx->GetInputDim("X"); + auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); + + PADDLE_ENFORCE_GE(out_dims.size(), 2, + "The tensor rank of Input(Out@Grad) should be 2."); + PADDLE_ENFORCE_EQ(out_dims[0], in_dims[0], + "The 1st dimension of Input(Out@Grad) must be " + "same as input."); + PADDLE_ENFORCE_EQ(out_dims[1], 1, + "The 2nd dimension of Input(Out@Grad) must be 1."); + + auto x_grad_name = framework::GradVarName("X"); + auto y_grad_name = framework::GradVarName("Y"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, in_dims); + } + if (ctx->HasOutput(y_grad_name)) { + ctx->SetOutputDim(y_grad_name, in_dims); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(smooth_l1_loss, ops::SmoothL1LossOp, + ops::SmoothL1LossOpMaker, smooth_l1_loss_grad, + ops::SmoothL1LossGradOp); +REGISTER_OP_CPU_KERNEL( + smooth_l1_loss, ops::SmoothL1LossKernel); +REGISTER_OP_CPU_KERNEL( + smooth_l1_loss_grad, + ops::SmoothL1LossGradKernel); diff --git a/paddle/operators/sigmoid_op.cu b/paddle/operators/smooth_l1_loss_op.cu similarity index 73% rename from paddle/operators/sigmoid_op.cu rename to paddle/operators/smooth_l1_loss_op.cu index 1a50dfe14a..1c3172f438 100644 --- a/paddle/operators/sigmoid_op.cu +++ b/paddle/operators/smooth_l1_loss_op.cu @@ -13,11 +13,12 @@ limitations under the License. */ #define EIGEN_USE_GPU -#include "paddle/operators/sigmoid_op.h" -namespace ops = paddle::operators; +#include "paddle/operators/smooth_l1_loss_op.h" -REGISTER_OP_GPU_KERNEL(sigmoid, - ops::SigmoidKernel); +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL( + smooth_l1_loss, ops::SmoothL1LossKernel); REGISTER_OP_GPU_KERNEL( - sigmoid_grad, ops::SigmoidGradKernel); + smooth_l1_loss_grad, + ops::SmoothL1LossGradKernel); diff --git a/paddle/operators/smooth_l1_loss_op.h b/paddle/operators/smooth_l1_loss_op.h new file mode 100644 index 0000000000..39d0070b6c --- /dev/null +++ b/paddle/operators/smooth_l1_loss_op.h @@ -0,0 +1,182 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/platform/hostdevice.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenVector = framework::EigenVector; +template +using EigenMatrix = framework::EigenMatrix; + +template +struct SmoothL1LossForward { + HOSTDEVICE SmoothL1LossForward(const T& sigma2) : sigma2(sigma2) {} + + HOSTDEVICE T operator()(const T& val) const { + T abs_val = std::abs(val); + if (abs_val < 1.0 / sigma2) { + return 0.5 * val * val * sigma2; + } else { + return abs_val - 0.5 / sigma2; + } + } + + T sigma2; +}; + +template +class SmoothL1LossKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* in0 = context.Input("X"); + auto* in1 = context.Input("Y"); + auto* in2 = context.Input("InsideWeight"); + auto* in3 = context.Input("OutsideWeight"); + auto* out0 = context.Output("Diff"); + auto* out1 = context.Output("Out"); + + out0->mutable_data(context.GetPlace()); + out1->mutable_data(context.GetPlace()); + auto place = context.GetEigenDevice(); + + auto sigma = static_cast(context.Attr("sigma")); + T sigma2 = sigma * sigma; + bool has_weight = (in2 != nullptr) && (in3 != nullptr); + + auto x = EigenVector::Flatten(*in0); + auto y = EigenVector::Flatten(*in1); + auto diff = EigenVector::Flatten(*out0); + + diff.device(place) = x - y; + // multiply inside weight + if (has_weight) { + auto inside_weight = EigenVector::Flatten(*in2); + // cache diff, reused in bp + diff.device(place) = diff * inside_weight; + } + + auto in_counts = in0->numel(); + Tensor ptensor_errors; + ptensor_errors.mutable_data({static_cast(in_counts)}, + context.GetPlace()); + auto errors = EigenVector::Flatten(ptensor_errors); + // apply smooth l1 forward + errors.device(place) = diff.unaryExpr(SmoothL1LossForward(sigma2)); + + // multiply outside weight + if (has_weight) { + auto outside_weight = EigenVector::Flatten(*in3); + errors.device(place) = errors * outside_weight; + } + auto loss = EigenVector::Flatten(*out1); + // first dimension of 'X' is the number of samples + auto mat_dims = + framework::make_ddim({static_cast(in0->dims()[0]), + static_cast(in_counts / in0->dims()[0])}); + auto errors_mat_view = EigenMatrix::From(ptensor_errors, mat_dims); + loss.device(place) = errors_mat_view.sum(Eigen::array({{1}})); + } +}; + +template +struct SmoothL1LossBackward { + HOSTDEVICE SmoothL1LossBackward(const T& sigma2) : sigma2(sigma2) {} + + HOSTDEVICE T operator()(const T& val) const { + T abs_val = std::abs(val); + if (abs_val < 1.0 / sigma2) { + return sigma2 * val; + } else { + return (0 < val) - (val < 0); + } + } + + T sigma2; +}; + +template +class SmoothL1LossGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* in0 = context.Input("InsideWeight"); + auto* in1 = context.Input("OutsideWeight"); + auto* in2 = context.Input("Diff"); + auto* og = context.Input(framework::GradVarName("Out")); + auto sigma = static_cast(context.Attr("sigma")); + T sigma2 = sigma * sigma; + bool has_weight = (in0 != nullptr) && (in1 != nullptr); + + auto place = context.GetEigenDevice(); + + auto in_dims = in2->dims(); + auto counts = in2->numel(); + auto cols = counts / in_dims[0]; + auto mat_dims = framework::make_ddim( + {static_cast(in_dims[0]), static_cast(cols)}); + + Tensor ptensor_diff; + ptensor_diff.mutable_data({static_cast(counts)}, + context.GetPlace()); + auto diff = EigenVector::Flatten(ptensor_diff); + // apply smooth l1 backwoard + diff.device(place) = EigenVector::Flatten(*in2).unaryExpr( + SmoothL1LossBackward(sigma2)); + + // compute weights + Tensor ptensor_weights; + ptensor_weights.mutable_data(mat_dims, context.GetPlace()); + auto weights = EigenMatrix::From(ptensor_weights); + // initialize to 1.0 + weights.device(place) = weights.constant(static_cast(1.0)); + if (has_weight) { + auto inside_weight = EigenMatrix::From(*in0, mat_dims); + auto outside_weight = EigenMatrix::From(*in1, mat_dims); + weights.device(place) = inside_weight * outside_weight; + } + + // compute gradients + auto out_grad = EigenMatrix::From(*og); + auto diff_mat_view = EigenMatrix::From(ptensor_diff, mat_dims); + auto gradients = out_grad.broadcast( + Eigen::array({{1, static_cast(cols)}})) * + weights * diff_mat_view; + + auto* out0 = context.Output(framework::GradVarName("X")); + auto* out1 = context.Output(framework::GradVarName("Y")); + + if (out0) { + out0->mutable_data(context.GetPlace()); + auto x_grad = EigenMatrix::From(*out0, mat_dims); + x_grad.device(place) = gradients; + } + + if (out1) { + out1->mutable_data(context.GetPlace()); + auto y_grad = EigenMatrix::From(*out1, mat_dims); + y_grad.device(place) = -1 * gradients; + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc index c67eb028c8..e353afee3e 100644 --- a/paddle/operators/softmax_op.cc +++ b/paddle/operators/softmax_op.cc @@ -22,23 +22,23 @@ class SoftmaxOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of SoftmaxOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"), - "Output(Y) of SoftmaxOp should not be null."); - - PADDLE_ENFORCE(ctx.Input("X")->dims().size() == 2UL, + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SoftmaxOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Y"), + "Output(Y) of SoftmaxOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE(x_dims.size() == 2UL, "The input of softmax op must be a matrix."); - ctx.Output("Y")->Resize( - ctx.Input("X")->dims()); + ctx->SetOutputDim("Y", x_dims); } }; class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftmaxOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) + SoftmaxOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input tensor of softmax. " @@ -69,16 +69,15 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should be not null."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Y")), - "Input(Y@GRAD) should be not null."); - PADDLE_ENFORCE_EQ(ctx.Input("Y")->dims(), - ctx.Input(framework::GradVarName("Y"))->dims(), + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should be not null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")), + "Input(Y@GRAD) should be not null."); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Y"), + ctx->GetInputDim(framework::GradVarName("Y")), "Input(Y) and its gradients should have a same shape."); - ctx.Output(framework::GradVarName("X")) - ->Resize(ctx.Input("X")->dims()); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); } }; diff --git a/paddle/operators/softmax_op.h b/paddle/operators/softmax_op.h index 8a3a5ab927..2c08853f4f 100644 --- a/paddle/operators/softmax_op.h +++ b/paddle/operators/softmax_op.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include "paddle/framework/eigen.h" #include "paddle/framework/op_registry.h" +#include "paddle/operators/math/softmax.h" namespace paddle { namespace operators { @@ -25,73 +26,31 @@ template ; template -class SoftmaxKernel : public framework::OpKernel { +class SoftmaxKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto X = context.Input("X"); - auto Y = context.Output("Y"); - Y->mutable_data(context.GetPlace()); - - auto logits = EigenMatrix::From(*X); - auto softmax = EigenMatrix::From(*Y); - - const int kBatchDim = 0; - const int kClassDim = 1; - - const int batch_size = logits.dimension(kBatchDim); - const int num_classes = logits.dimension(kClassDim); - - Eigen::DSizes along_class(kClassDim); - Eigen::DSizes batch_by_one(batch_size, 1); - Eigen::DSizes one_by_class(1, num_classes); + auto* X = context.Input("X"); + auto* Y = context.Output("Y"); - auto shifted_logits = (logits - - logits.maximum(along_class) - .eval() - .reshape(batch_by_one) - .broadcast(one_by_class)); - - softmax.device(context.GetEigenDevice()) = shifted_logits.exp(); + // allocate memory on device. + Y->mutable_data(context.GetPlace()); - softmax.device(context.GetEigenDevice()) = - (softmax * - softmax.sum(along_class) - .inverse() - .eval() - .reshape(batch_by_one) - .broadcast(one_by_class)); + math::SoftmaxFunctor()(context.device_context(), X, Y); } }; template -class SoftmaxGradKernel : public framework::OpKernel { +class SoftmaxGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - std::shared_ptr scale_ = std::make_shared(); + auto* Y = context.Input("Y"); + auto* dY = context.Input(framework::GradVarName("Y")); + auto* dX = context.Output(framework::GradVarName("X")); - auto Y = context.Input("Y"); - auto dY = context.Input(framework::GradVarName("Y")); - auto dX = context.Output(framework::GradVarName("X")); + // allocate memory on device. dX->mutable_data(context.GetPlace()); - const int batch_size = Y->dims()[0]; - const int class_num = Y->dims()[1]; - - Eigen::DSizes along_class(1); - Eigen::DSizes batch_by_one(batch_size, 1); - Eigen::DSizes one_by_class(1, class_num); - - auto Y_eigen = EigenMatrix::From(*Y); - auto dY_eigen = EigenMatrix::From(*dY); - auto dX_eigen = EigenMatrix::From(*dX); - auto place = context.GetEigenDevice(); - - auto dot = (Y_eigen * dY_eigen) - .sum(along_class) - .eval() - .reshape(batch_by_one) - .broadcast(one_by_class); - dX_eigen.device(place) = (dY_eigen - dot) * Y_eigen; + math::SoftmaxGradFunctor()(context.device_context(), Y, dY, dX); } }; diff --git a/paddle/operators/softmax_with_cross_entropy_op.cc b/paddle/operators/softmax_with_cross_entropy_op.cc new file mode 100644 index 0000000000..42c1ba6fdf --- /dev/null +++ b/paddle/operators/softmax_with_cross_entropy_op.cc @@ -0,0 +1,197 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/softmax_with_cross_entropy_op.h" +#include +#include + +namespace paddle { +namespace operators { + +class SoftmaxWithCrossEntropyOpMaker + : public framework::OpProtoAndCheckerMaker { + public: + SoftmaxWithCrossEntropyOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + 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("Label", + "(Tensor, default: Tensor), The ground truth which is a 2-D " + "tensor. " + "If softLable is set to 0, Label is a Tensor with shape [N x " + "1]. " + "If softLable is set to 1, Label is a Tensor " + "with shape [N x K]."); + AddOutput( + "Softmax", + "(Tensor, default: Tensor), A 2-D tensor with shape [N x K]. " + "The outputs value of softmax activation by given the input batch, " + "which will be used in backward calculation.") + .AsIntermediate(); + AddOutput("Loss", + "(Tensor, default: Tensor), A 2-D tensor. The cross " + "entropy loss with shape [N x 1]."); + AddAttr( + "softLabel", + "(bool, default: false), A flag to indicate whether to interpretate " + "the given labels as soft labels.") + .SetDefault(false); + AddComment(R"DOC( +Cross entropy loss with softmax are used as the output layer extensively. This +operator computes the softmax normalized values for each row of the input +tensor, after which cross-entropy loss is then computed. This provides a more +numerically stable gradient. + +Because this operators performs a softmax on logits internally, it expects +unscaled logits. Please do not call this op with the output of softmax operator, +which will produce incorrect results. + +This operators expects mutually exclusive hard labels, each sample in a batch +is in exactly one class with probabilities 1. Each sample in the batch with one +and only one label. + +Equation: + +1) hard label (one-hot label) + +Loss_j = -\text{Logit}_{Label_j} + \log\left(\sum_{i=0}^{K}\exp(\text{Logit}_i)\right), j = 1, ..., K + +2) soft label (a distribution over all classes) + +Loss_j = -\sum_{i=0}^{K}\text{Label}_i\left(\text{Logit}_i-\log\left(\sum_{i=0}^{K}\exp(\text{Logit}_i)\right)\right), j = 1,...,K + +)DOC"); + } +}; + +class SoftmaxWithCrossEntropyOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Logits"), + "Input(Logits) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); + + PADDLE_ENFORCE(ctx->HasOutput("Softmax"), + "Output(Softmax) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("Loss"), "Output(Loss) should be not null."); + + auto logits_dims = ctx->GetInputDim("Logits"); + auto labels_dims = ctx->GetInputDim("Label"); + PADDLE_ENFORCE_EQ( + logits_dims.size(), 2UL, + "The input 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."); + + if (ctx->Attrs().Get("softLabel")) { + PADDLE_ENFORCE_EQ(logits_dims[1], labels_dims[1], + "If Attr(softLabel) == true, the 2nd dimension of " + "Input(X) and Input(Label) should be equal."); + } else { + PADDLE_ENFORCE_EQ(labels_dims[1], 1UL, + "If Attr(softLabel) == false, the 2nd dimension of " + "Input(Label) should be 1."); + } + + ctx->SetOutputDim("Softmax", logits_dims); + ctx->SetOutputDim("Loss", {logits_dims[0], 1}); + + ctx->ShareLoD("Logits", /*->*/ "Softmax"); + ctx->ShareLoD("Logits", /*->*/ "Loss"); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType(ctx.Input("Logits")->type()); + } +}; + +class SoftmaxWithCrossEntropyOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Loss")), + "Input(Loss@Grad) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Softmax"), + "Input(Softmax) should be not null."); + PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Logits")), + "Output(Logits@Grad) should be not null."); + + auto softmax_dims = ctx->GetInputDim("Softmax"); + auto labels_dims = ctx->GetInputDim("Label"); + PADDLE_ENFORCE_EQ(labels_dims.size(), 2UL, + "The labels should be a 2-D tensor."); + + if (ctx->Attrs().Get("softLabel")) { + PADDLE_ENFORCE_EQ(softmax_dims[1], labels_dims[1], + "When Attr(softLabel) == true, the 2nd dimension of " + "Input(X) and Input(Label) should be equal."); + } else { + PADDLE_ENFORCE_EQ(labels_dims[1], 1UL, + "When Attr(softLabel) == false, the 2nd dimension of " + "Input(Label) should be 1."); + } + + ctx->SetOutputDim(framework::GradVarName("Logits"), + ctx->GetInputDim("Softmax")); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return framework::ToDataType( + ctx.Input(framework::GradVarName("Loss"))->type()); + } +}; + +class SoftmaxGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* grad_op = new framework::OpDescBind(); + grad_op->SetType("softmax_with_cross_entropy_grad"); + grad_op->SetInput("Label", Input("Label")); + grad_op->SetInput("Softmax", Output("Softmax")); + grad_op->SetInput("Loss", Output("Loss")); + grad_op->SetInput(framework::GradVarName("Softmax"), OutputGrad("Softmax")); + grad_op->SetInput(framework::GradVarName("Loss"), OutputGrad("Loss")); + grad_op->SetOutput(framework::GradVarName("Logits"), InputGrad("Logits")); + grad_op->SetAttrMap(Attrs()); + return std::unique_ptr(grad_op); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(softmax_with_cross_entropy, ops::SoftmaxWithCrossEntropyOp, + ops::SoftmaxWithCrossEntropyOpMaker, ops::SoftmaxGradMaker); +REGISTER_OPERATOR(softmax_with_cross_entropy_grad, + ops::SoftmaxWithCrossEntropyOpGrad); +REGISTER_OP_CPU_KERNEL(softmax_with_cross_entropy, + ops::SoftmaxWithCrossEntropyKernel); +REGISTER_OP_CPU_KERNEL(softmax_with_cross_entropy_grad, + ops::SoftmaxWithCrossEntropyGradKernel); diff --git a/paddle/operators/softmax_with_cross_entropy_op.cu b/paddle/operators/softmax_with_cross_entropy_op.cu new file mode 100644 index 0000000000..2bc53ecf87 --- /dev/null +++ b/paddle/operators/softmax_with_cross_entropy_op.cu @@ -0,0 +1,121 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU + +#include "paddle/operators/softmax_with_cross_entropy_op.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +namespace { +template +__global__ void CrossEntropyGrad(T* out_grad, const T* in_grad, + const int* labels, const int batch_size, + const int class_num) { + int tid = blockIdx.x * blockDim.x + threadIdx.x; + int sample_idx = tid / class_num; + + if (tid < batch_size * class_num) out_grad[tid] *= in_grad[sample_idx]; + __syncthreads(); + + if (tid < batch_size) { + PADDLE_ASSERT(labels[sample_idx] >= 0 && labels[sample_idx] < class_num); + out_grad[tid * class_num + labels[tid]] -= 1.; + } +} + +template +__global__ void SoftCrossEntropyGradientKernel(T* logit_grad, + const T* loss_grad, + const T* labels, + const int batch_size, + const int class_num) { + int ids = blockIdx.x * blockDim.x + threadIdx.x; + if (ids < batch_size * class_num) { + int row_ids = ids / class_num; + logit_grad[ids] = logit_grad[ids] * loss_grad[row_ids] - labels[ids]; + } +} +} // namespace + +template +class SoftmaxWithCrossEntropyCUDAKernel : 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."); + const Tensor* logits = context.Input("Logits"); + const Tensor* labels = context.Input("Label"); + Tensor* softmax = context.Output("Softmax"); + + Tensor* loss = context.Output("Loss"); + softmax->mutable_data(context.GetPlace()); + loss->mutable_data(context.GetPlace()); + + math::SoftmaxFunctor()(context.device_context(), + logits, softmax); + math::CrossEntropyFunctor()( + context.device_context(), loss, softmax, labels, + context.Attr("softLabel")); + } +}; + +template +class SoftmaxWithCrossEntropyGradCUDAKernel : 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."); + const Tensor* labels = context.Input("Label"); + const T* loss_grad_data = + context.Input(framework::GradVarName("Loss"))->data(); + Tensor* logit_grad = + context.Output(framework::GradVarName("Logits")); + logit_grad->ShareDataWith(*context.Input("Softmax")); + T* logit_grad_data = logit_grad->data(); + + const int batch_size = logit_grad->dims()[0]; + const int class_num = logit_grad->dims()[1]; + int block = 512; + int grid = (batch_size * class_num + block - 1) / block; + + if (context.Attr("softLabel")) { + const T* label_data = labels->data(); + SoftCrossEntropyGradientKernel<<< + grid, block, 0, reinterpret_cast( + context.device_context()) + .stream()>>>(logit_grad_data, loss_grad_data, + label_data, batch_size, class_num); + } else { + const int* label_data = labels->data(); + CrossEntropyGrad<<< + grid, block, 0, reinterpret_cast( + context.device_context()) + .stream()>>>(logit_grad_data, loss_grad_data, + label_data, batch_size, class_num); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(softmax_with_cross_entropy, + ops::SoftmaxWithCrossEntropyCUDAKernel); +REGISTER_OP_GPU_KERNEL(softmax_with_cross_entropy_grad, + ops::SoftmaxWithCrossEntropyGradCUDAKernel); diff --git a/paddle/operators/softmax_with_cross_entropy_op.h b/paddle/operators/softmax_with_cross_entropy_op.h new file mode 100644 index 0000000000..cffd422f18 --- /dev/null +++ b/paddle/operators/softmax_with_cross_entropy_op.h @@ -0,0 +1,88 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/math/cross_entropy.h" +#include "paddle/operators/math/softmax.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenMatrix = framework::EigenMatrix; + +template +class SoftmaxWithCrossEntropyKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + PADDLE_ENFORCE(platform::is_cpu_place(context.GetPlace()), + "This kernel only runs on CPU."); + const Tensor* logits = context.Input("Logits"); + const Tensor* labels = context.Input("Label"); + Tensor* softmax = context.Output("Softmax"); + Tensor* loss = context.Output("Loss"); + + softmax->mutable_data(context.GetPlace()); + loss->mutable_data(context.GetPlace()); + + math::SoftmaxFunctor()(context.device_context(), + logits, softmax); + math::CrossEntropyFunctor()( + context.device_context(), loss, softmax, labels, + context.Attr("softLabel")); + } +}; + +template +class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* out_grad = + context.Input(framework::GradVarName("Loss")); + const Tensor* labels = context.Input("Label"); + Tensor* logit_grad = + context.Output(framework::GradVarName("Logits")); + logit_grad->ShareDataWith(*context.Input("Softmax")); + + const int class_num = logit_grad->dims()[1]; + if (context.Attr("softLabel")) { + auto out_grad_mat = EigenMatrix::From(*out_grad); + auto logit_grad_mat = EigenMatrix::From(*logit_grad); + auto lbl_mat = EigenMatrix::From(*labels); + + logit_grad_mat.device(context.GetEigenDevice()) = + logit_grad_mat * + out_grad_mat.broadcast(Eigen::DSizes(1, class_num)) - + lbl_mat; + } else { + const int batch_size = logit_grad->dims()[0]; + const int* label_data = labels->data(); + const T* out_grad_data = out_grad->data(); + T* logit_grad_data = logit_grad->data(); + + for (int i = 0; i < batch_size; ++i) { + int index = i * class_num + label_data[i]; + logit_grad_data[index] = + (out_grad_data[i] * logit_grad_data[index] - 1.); + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/split_op.cc b/paddle/operators/split_op.cc index 61296f5c81..5f4b5539af 100644 --- a/paddle/operators/split_op.cc +++ b/paddle/operators/split_op.cc @@ -24,40 +24,43 @@ class SplitOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - // infershape - auto *in = ctx.Input("X"); - auto outs = ctx.MultiOutput("Out"); - size_t axis = static_cast(ctx.Attr("axis")); - size_t num = static_cast(ctx.Attr("num")); - std::vector sections = - static_cast>(ctx.Attr>("sections")); - const size_t n = outs.size(); + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SplitOp should not be null."); + PADDLE_ENFORCE_GE(ctx->Outputs("Out").size(), 1UL, + "Outputs(Out) of SplitOp should not be empty."); + auto in_dims = ctx->GetInputDim("X"); + auto outs_names = ctx->Outputs("Out"); + size_t axis = static_cast(ctx->Attrs().Get("axis")); + size_t num = static_cast(ctx->Attrs().Get("num")); + std::vector sections = static_cast>( + ctx->Attrs().Get>("sections")); + const size_t outs_number = outs_names.size(); + std::vector outs_dims; + outs_dims.reserve(outs_number); if (num > 0) { - int64_t in_axis_dim = in->dims()[axis]; + int64_t in_axis_dim = in_dims[axis]; PADDLE_ENFORCE_EQ(in_axis_dim % num, 0, "tensor split does not result" " in an equal division"); size_t out_axis_dim = in_axis_dim / num; - for (size_t i = 0; i < n; ++i) { - auto dim = in->dims(); + for (size_t i = 0; i < outs_number; ++i) { + auto dim = in_dims; dim[axis] = out_axis_dim; - outs[i]->Resize(dim); + outs_dims.push_back(dim); } } else if (sections.size() > 0) { - PADDLE_ENFORCE_EQ(sections.size(), n, + PADDLE_ENFORCE_EQ(sections.size(), outs_number, "tensor split sections size" "should be equal to output size."); - for (size_t i = 0; i < n; ++i) { - auto dim = in->dims(); + for (size_t i = 0; i < outs_number; ++i) { + auto dim = in_dims; dim[axis] = sections[i]; - outs[i]->Resize(dim); + outs_dims.push_back(dim); } - } else { - PADDLE_ENFORCE_NOT_NULL(nullptr, "split operator should", - " specify indices or sections."); } + ctx->SetOutputsDim("Out", outs_dims); } }; @@ -115,4 +118,4 @@ USE_CPU_ONLY_OP(concat); REGISTER_OP(split, ops::SplitOp, ops::SplitOpMaker, split_grad, ops::SplitOpGrad); REGISTER_OP_CPU_KERNEL(split, - ops::SplitKernel); + ops::SplitOpKernel); diff --git a/paddle/operators/split_op.cu b/paddle/operators/split_op.cu new file mode 100644 index 0000000000..93d1fc3c44 --- /dev/null +++ b/paddle/operators/split_op.cu @@ -0,0 +1,18 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/split_op.h" +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(split, + ops::SplitOpKernel); diff --git a/paddle/operators/split_op.h b/paddle/operators/split_op.h index 860690ee89..fa26e5f677 100644 --- a/paddle/operators/split_op.h +++ b/paddle/operators/split_op.h @@ -16,44 +16,29 @@ limitations under the License. */ #include #include "paddle/framework/op_registry.h" +#include "paddle/operators/strided_memcpy.h" namespace paddle { namespace operators { template -class SplitKernel : public framework::OpKernel { +class SplitOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* in = ctx.Input("X"); auto outs = ctx.MultiOutput("Out"); + auto in_stride = framework::stride(in->dims()); int64_t axis = static_cast(ctx.Attr("axis")); - size_t before = 1, after = 1; const size_t n = outs.size(); - size_t input_axis_dim = in->dims()[axis]; - - for (int64_t i = 0; i < in->dims().size(); ++i) { - if (i == axis) { - continue; - } - if (i < axis) { - before *= in->dims()[i]; - } else { - after *= in->dims()[i]; - } - } size_t input_offset = 0; for (size_t i = 0; i < n; i++) { auto& out = outs[i]; + out->mutable_data(ctx.GetPlace()); size_t axis_dim = out->dims()[axis]; - for (size_t j = 0; j < before; j++) { - size_t len = axis_dim * after * sizeof(T); - T* dest = - out->mutable_data(platform::CPUPlace()) + axis_dim * after * j; - const T* src = - in->data() + input_offset + input_axis_dim * after * j; - memcpy(dest, src, len); - } - input_offset += axis_dim * after; + auto out_stride = framework::stride(out->dims()); + StridedMemcpy(ctx.device_context(), in->data() + input_offset, + in_stride, out->dims(), out_stride, out->data()); + input_offset += axis_dim * in_stride[axis]; } } }; diff --git a/paddle/operators/squared_l2_distance_op.cc b/paddle/operators/squared_l2_distance_op.cc index 39f4305877..5a0cb59600 100644 --- a/paddle/operators/squared_l2_distance_op.cc +++ b/paddle/operators/squared_l2_distance_op.cc @@ -22,24 +22,19 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL( - ctx.InputVar("X"), - "Input(X) of SquaredL2DistanceOp should not be null."); - PADDLE_ENFORCE_NOT_NULL( - ctx.InputVar("Y"), - "Input(Y) of SquaredL2DistanceOp should not be null."); - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("sub_result"), + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SquaredL2DistanceOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Y"), + "Input(Y) of SquaredL2DistanceOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("sub_result"), "Output(sub_result) of SquaredL2DistanceOp should not be null."); - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Out"), - "Output(Out) of SquaredL2DistanceOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SquaredL2DistanceOp should not be null."); - auto* x = ctx.Input("X"); - auto x_dims = x->dims(); - auto* y = ctx.Input("Y"); - auto y_dims = y->dims(); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); PADDLE_ENFORCE_EQ(framework::arity(x_dims), framework::arity(y_dims), "Tensor rank of both SquaredL2DistanceOp's " @@ -47,16 +42,16 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel { int rank = framework::arity(x_dims); PADDLE_ENFORCE_GE(rank, 2, "Tensor rank should be at least equal to 2."); - PADDLE_ENFORCE_EQ(x->numel() / x_dims[0], y->numel() / y_dims[0], + PADDLE_ENFORCE_EQ(product(x_dims) / x_dims[0], product(y_dims) / y_dims[0], "Product of dimensions expcet the first dimension of " "input and target must be equal."); PADDLE_ENFORCE(y_dims[0] == 1 || y_dims[0] == x_dims[0], "First dimension of target must be equal to input " "or to 1."); - ctx.Output("sub_result") - ->Resize({x_dims[0], x->numel() / x_dims[0]}); - ctx.Output("Out")->Resize({x_dims[0], 1}); + ctx->SetOutputDim("sub_result", {x_dims[0], product(x_dims) / x_dims[0]}); + ctx->SetOutputDim("Out", {x_dims[0], 1}); + ctx->ShareLoD("X", /*->*/ "Out"); } }; @@ -79,6 +74,9 @@ class SquaredL2DistanceOpMaker : public framework::OpProtoAndCheckerMaker { input or to 1. If the first dimension of target is 1, SquaredL2DistanceOp will broadcast target's first dimension to input's first dimension. You can decide whether calculate the gradient of input and target. + + Both the input X and Y can carry the LoD (Level of Details) information, + or not. But the output only shares the LoD with input X. )DOC"); } }; @@ -88,24 +86,22 @@ class SquaredL2DistanceGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), - "Gradient of Out should not be null"); - auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); - auto x_dims = ctx.Input("X")->dims(); - auto y_dims = ctx.Input("Y")->dims(); + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Gradient of Out should not be null"); + auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); + auto x_dims = ctx->GetInputDim("X"); + auto y_dims = ctx->GetInputDim("Y"); PADDLE_ENFORCE_EQ(out_dims[0], x_dims[0], "First dimension of output gradient and " "input value must be equal."); PADDLE_ENFORCE_EQ(out_dims[1], 1, "Second dimension of output gradient " "must be 1."); - auto* x_grad = - ctx.Output(framework::GradVarName("X")); - auto* y_grad = - ctx.Output(framework::GradVarName("Y")); - if (x_grad) x_grad->Resize(x_dims); - if (y_grad) y_grad->Resize(y_dims); + auto x_grad_name = framework::GradVarName("X"); + auto y_grad_name = framework::GradVarName("Y"); + if (ctx->HasOutput(x_grad_name)) ctx->SetOutputDim(x_grad_name, x_dims); + if (ctx->HasOutput(y_grad_name)) ctx->SetOutputDim(y_grad_name, y_dims); } }; diff --git a/paddle/operators/squared_l2_distance_op.h b/paddle/operators/squared_l2_distance_op.h index 097ac04fc0..259ef40296 100644 --- a/paddle/operators/squared_l2_distance_op.h +++ b/paddle/operators/squared_l2_distance_op.h @@ -28,7 +28,7 @@ template ; template -class SquaredL2DistanceKernel : public framework::OpKernel { +class SquaredL2DistanceKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in0 = context.Input("X"); @@ -68,7 +68,7 @@ class SquaredL2DistanceKernel : public framework::OpKernel { }; template -class SquaredL2DistanceGradKernel : public framework::OpKernel { +class SquaredL2DistanceGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in0 = context.Input("sub_result"); diff --git a/paddle/operators/strided_memcpy.h b/paddle/operators/strided_memcpy.h new file mode 100644 index 0000000000..c9dd805184 --- /dev/null +++ b/paddle/operators/strided_memcpy.h @@ -0,0 +1,45 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/operators/detail/strided_memcpy.h" + +namespace paddle { +namespace operators { + +// Strided memory copy from src to dst. +// +// The src and dst should be both on dev_ctx.GetPlace(), otherwise, there will +// be a segment fault. +// +// The stride of an array (also referred to as increment, pitch or step size) is +// the number of locations in memory between beginnings of successive array +// elements +// +// For example, for tensor like [1, 3, 300, 300]. If there is no padding, the +// stride is [270000, 90000, 300, 1]. +// +// NOTE: When use GPU, the memcpy is async. To sync memcpy, please invoke +// `dev_ctx.Wait()`. +template +inline void StridedMemcpy(const platform::DeviceContext& dev_ctx, const T* src, + const framework::DDim& src_stride, + const framework::DDim& dst_dim, + const framework::DDim& dst_stride, T* dst) { + using namespace detail; + StridedCopyDimVisitor func(dev_ctx, src, src_stride, dst_stride, dst); + boost::apply_visitor(func, dst_dim); +} +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/strided_memcpy_test.cc b/paddle/operators/strided_memcpy_test.cc new file mode 100644 index 0000000000..68f064eaee --- /dev/null +++ b/paddle/operators/strided_memcpy_test.cc @@ -0,0 +1,160 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/strided_memcpy.h" +#include "gtest/gtest.h" +#include "paddle/memory/memory.h" + +namespace paddle { +namespace operators { + +TEST(StridedMemcpy, CPUCrop) { + // clang-format off + int src[] = { + 0, 1, 2, 0, 0, + 0, 3, 4, 0, 0, + 0, 0, 0, 0, 0, + }; + // clang-format on + + framework::DDim src_stride({5, 1}); + + int dst[4]; + framework::DDim dst_dim({2, 2}); + framework::DDim dst_stride({2, 1}); + + platform::CPUDeviceContext ctx; + StridedMemcpy(ctx, src + 1, src_stride, dst_dim, dst_stride, dst); + + ASSERT_EQ(1, dst[0]); + ASSERT_EQ(2, dst[1]); + ASSERT_EQ(3, dst[2]); + ASSERT_EQ(4, dst[3]); +} + +TEST(StridedMemcpy, CPUConcat) { + // clang-format off + int src[] = { + 1, 2, + 3, 4 + }; + // clang-format on + + int dst[8]; + + framework::DDim src_stride({2, 1}); + framework::DDim dst_dim({2, 2}); + framework::DDim dst_stride({4, 1}); + platform::CPUDeviceContext ctx; + + StridedMemcpy(ctx, src, src_stride, dst_dim, dst_stride, dst); + StridedMemcpy(ctx, src, src_stride, dst_dim, dst_stride, dst + 2); + + // clang-format off + int expect_dst[] = { + 1, 2, 1, 2, + 3, 4, 3, 4 + }; + // clang-format on + for (size_t i = 0; i < sizeof(expect_dst) / sizeof(int); ++i) { + ASSERT_EQ(expect_dst[i], dst[i]); + } +} + +#ifdef PADDLE_WITH_CUDA +TEST(StridedMemcpy, GPUCrop) { + // clang-format off + int src[] = { + 0, 1, 2, 0, 0, + 0, 3, 4, 0, 0, + 0, 0, 0, 0, 0, + }; + // clang-format on + + platform::GPUPlace gpu0(0); + platform::CPUPlace cpu; + + int* gpu_src = reinterpret_cast(memory::Alloc(gpu0, sizeof(src))); + memory::Copy(gpu0, gpu_src, cpu, src, sizeof(src)); + + framework::DDim src_stride({5, 1}); + + int dst[4]; + int* gpu_dst = reinterpret_cast(memory::Alloc(gpu0, sizeof(dst))); + + framework::DDim dst_dim({2, 2}); + framework::DDim dst_stride({2, 1}); + + platform::CUDADeviceContext ctx(gpu0); + StridedMemcpy(ctx, gpu_src + 1, src_stride, dst_dim, dst_stride, + gpu_dst); + + memory::Copy(cpu, dst, gpu0, gpu_dst, sizeof(dst), ctx.stream()); + ctx.Wait(); + + ASSERT_EQ(1, dst[0]); + ASSERT_EQ(2, dst[1]); + ASSERT_EQ(3, dst[2]); + ASSERT_EQ(4, dst[3]); + + memory::Free(gpu0, gpu_dst); + memory::Free(gpu0, gpu_src); +} + +TEST(StridedMemcpy, GPUConcat) { + // clang-format off + int src[] = { + 1, 2, + 3, 4 + }; + // clang-format on + + platform::GPUPlace gpu0(0); + platform::CPUPlace cpu; + + int* gpu_src = reinterpret_cast(memory::Alloc(gpu0, sizeof(src))); + memory::Copy(gpu0, gpu_src, cpu, src, sizeof(src)); + + int dst[8]; + int* gpu_dst = reinterpret_cast(memory::Alloc(gpu0, sizeof(dst))); + + framework::DDim src_stride({2, 1}); + framework::DDim dst_dim({2, 2}); + framework::DDim dst_stride({4, 1}); + platform::CUDADeviceContext ctx(gpu0); + + StridedMemcpy(ctx, gpu_src, src_stride, dst_dim, dst_stride, gpu_dst); + StridedMemcpy(ctx, gpu_src, src_stride, dst_dim, dst_stride, + gpu_dst + 2); + + memory::Copy(cpu, dst, gpu0, gpu_dst, sizeof(dst), ctx.stream()); + ctx.Wait(); + + // clang-format off + int expect_dst[] = { + 1, 2, 1, 2, + 3, 4, 3, 4 + }; + // clang-format on + for (size_t i = 0; i < sizeof(expect_dst) / sizeof(int); ++i) { + ASSERT_EQ(expect_dst[i], dst[i]); + } + + memory::Free(gpu0, gpu_dst); + memory::Free(gpu0, gpu_src); +} + +#endif +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc index 41e05c27f9..c701ee8dde 100644 --- a/paddle/operators/sum_op.cc +++ b/paddle/operators/sum_op.cc @@ -11,6 +11,7 @@ limitations under the License. */ #include "paddle/operators/sum_op.h" #include +#include "paddle/operators/net_op.h" namespace paddle { namespace operators { @@ -21,51 +22,60 @@ class SumOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(!ctx.MultiInputVar("X").empty(), - "Input(X) of SumOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of SumOp should not be null."); - - auto ins = ctx.MultiInput("X"); - auto *out = ctx.Output("Out"); - int N = ins.size(); - - auto in_dim = ins[0]->dims(); + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInputs("X"), "Inputs(X) should not be null"); + auto x_dims = ctx->GetInputsDim("X"); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SumOp should not be null."); + size_t N = x_dims.size(); PADDLE_ENFORCE_GT(N, 1, "Input tensors count should > 1."); - for (int i = 1; i < N; i++) { - auto dim = ins[i]->dims(); + + auto in_dim = x_dims[0]; + for (size_t i = 1; i < N; i++) { + auto dim = x_dims[i]; PADDLE_ENFORCE(in_dim == dim, "Input tensors must have same shape"); } - out->Resize(in_dim); + ctx->SetOutputDim("Out", in_dim); + ctx->ShareLoD("X", /*->*/ "Out"); } }; class SumOpMaker : public framework::OpProtoAndCheckerMaker { public: - SumOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + SumOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "the input tensors of sum operator.").AsDuplicable(); AddOutput("Out", "the output tensor of sum operator."); AddComment(R"DOC( - Sum the input tensors. - )DOC"); +Sum the input tensors. + +All the inputs can carry the LoD (Level of Details) information, +or not. But the output only shares the LoD with the first input. +)DOC"); } }; -class SumGradOp : public framework::OperatorWithKernel { +class SumGradMaker : public framework::GradOpDescMakerBase { public: - using framework::OperatorWithKernel::OperatorWithKernel; + using framework::GradOpDescMakerBase::GradOpDescMakerBase; - protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - auto outputs = - ctx.MultiOutput(framework::GradVarName("X")); - auto dims = ctx.Input(framework::GradVarName("Out"))->dims(); - for (auto output : outputs) { - output->Resize(dims); - } + std::vector> operator()() + const override { + auto x_grads = InputGrad("X"); + std::vector> grad_ops; + grad_ops.reserve(x_grads.size()); + auto og = OutputGrad("Out"); + std::transform(x_grads.begin(), x_grads.end(), std::back_inserter(grad_ops), + [&og](const std::string& x_grad) { + auto* grad_op = new framework::OpDescBind(); + grad_op->SetType("scale"); + grad_op->SetInput("X", og); + grad_op->SetOutput("Out", {x_grad}); + grad_op->SetAttr("scale", 1.0f); + return std::unique_ptr(grad_op); + }); + return grad_ops; } }; @@ -73,7 +83,6 @@ class SumGradOp : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(sum, ops::SumOp, ops::SumOpMaker, sum_grad, ops::SumGradOp); + +REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker); REGISTER_OP_CPU_KERNEL(sum, ops::SumKernel); -REGISTER_OP_CPU_KERNEL(sum_grad, - ops::SumGradKernel); diff --git a/paddle/operators/sum_op.cu b/paddle/operators/sum_op.cu index a465cf3659..b1896d3cd8 100644 --- a/paddle/operators/sum_op.cu +++ b/paddle/operators/sum_op.cu @@ -14,5 +14,3 @@ limitations under the License. */ namespace ops = paddle::operators; REGISTER_OP_GPU_KERNEL(sum, ops::SumKernel); -REGISTER_OP_GPU_KERNEL(sum_grad, - ops::SumGradKernel); diff --git a/paddle/operators/sum_op.h b/paddle/operators/sum_op.h index 0b1e9ebaa3..91e5da8b40 100644 --- a/paddle/operators/sum_op.h +++ b/paddle/operators/sum_op.h @@ -22,7 +22,7 @@ template ; template -class SumKernel : public framework::OpKernel { +class SumKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto ins = context.MultiInput("X"); @@ -42,24 +42,5 @@ class SumKernel : public framework::OpKernel { } }; -template -class SumGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* input = context.Input(framework::GradVarName("Out")); - auto outs = context.MultiOutput(framework::GradVarName("X")); - for (auto out : outs) { - out->mutable_data(context.GetPlace()); - } - - auto place = context.GetEigenDevice(); - auto in = EigenVector::Flatten(*input); - for (auto out : outs) { - auto result = EigenVector::Flatten(*out); - result.device(place) = in; - } - } -}; - } // namespace operators } // namespace paddle diff --git a/paddle/operators/top_k_op.cc b/paddle/operators/top_k_op.cc index 169b815fef..5f22bf1df8 100644 --- a/paddle/operators/top_k_op.cc +++ b/paddle/operators/top_k_op.cc @@ -22,26 +22,26 @@ class TopkOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), - "Input(X) of TopkOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"), - "Output(Out) of TopkOp should not be null."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Indices"), - "Output(Indices) of TopkOp should not be null."); + void InferShape(framework::InferShapeContextBase *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of TopkOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of TopkOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Indices"), + "Output(Indices) of TopkOp should not be null."); - auto *input = ctx.Input("X"); - const int k = static_cast(ctx.Attr("k")); + auto input_dims = ctx->GetInputDim("X"); + const int k = static_cast(ctx->Attrs().Get("k")); PADDLE_ENFORCE_GE(k, 1, "k must >= 1"); - PADDLE_ENFORCE_GE(input->dims().size(), 1, "input must have >= 1d shape"); - PADDLE_ENFORCE_GE(input->dims()[input->dims().size() - 1], k, + PADDLE_ENFORCE_GE(input_dims.size(), 1, "input must have >= 1d shape"); + PADDLE_ENFORCE_GE(input_dims[input_dims.size() - 1], k, "input must have >= k columns"); - framework::DDim dims = input->dims(); + framework::DDim dims = input_dims; dims[dims.size() - 1] = k; - ctx.Output("Out")->Resize(dims); - ctx.Output("Indices")->Resize(dims); + ctx->SetOutputDim("Out", dims); + ctx->SetOutputDim("Indices", dims); } }; diff --git a/paddle/operators/top_k_op.cu b/paddle/operators/top_k_op.cu index afe4d149c5..7be6932f1e 100644 --- a/paddle/operators/top_k_op.cu +++ b/paddle/operators/top_k_op.cu @@ -279,7 +279,7 @@ __global__ void KeMatrixTopK(T* output, int output_stride, int* indices, } template -class TopkOpCUDAKernel : public framework::OpKernel { +class TopkOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), @@ -301,14 +301,16 @@ class TopkOpCUDAKernel : public framework::OpKernel { // NOTE: pass lds and dim same to input width. // NOTE: old matrix implementation of stride is different to eigen. - // TODO(typhoonzero): launch kernel on specified stream. // TODO(typhoonzero): refine this kernel. dim3 threads(256, 1); dim3 grid(input_height, 1); - KeMatrixTopK<<>>( - output_data, output->dims()[1], indices_data, input_data, input_width, - input_width, int(k)); + KeMatrixTopK<<< + grid, threads, 0, reinterpret_cast( + ctx.device_context()) + .stream()>>>(output_data, output->dims()[1], + indices_data, input_data, + input_width, input_width, int(k)); } }; diff --git a/paddle/operators/top_k_op.h b/paddle/operators/top_k_op.h index ef66acc1d5..4b248faa12 100644 --- a/paddle/operators/top_k_op.h +++ b/paddle/operators/top_k_op.h @@ -28,7 +28,7 @@ template ; template -class TopkKernel : public framework::OpKernel { +class TopkKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { // Get the top k elements of each row of input tensor diff --git a/paddle/operators/transpose_op.cc b/paddle/operators/transpose_op.cc new file mode 100644 index 0000000000..0672f9342d --- /dev/null +++ b/paddle/operators/transpose_op.cc @@ -0,0 +1,118 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/transpose_op.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; + +class TransposeOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null"); + auto x_dims = ctx->GetInputDim("X"); + std::vector axis = ctx->Attrs().Get>("axis"); + size_t x_rank = x_dims.size(); + size_t axis_size = axis.size(); + + PADDLE_ENFORCE_EQ(x_rank, axis_size, + "the input tensor's rank(%d) " + "should be equal to the axis's size(%d)", + x_rank, axis_size); + + std::vector count(axis_size, 0); + for (size_t i = 0; i < axis_size; i++) { + PADDLE_ENFORCE( + axis[i] < static_cast(axis_size) && ++count[axis[i]] == 1, + "Each element of Attribute axis should be a unique value " + "range from 0 to (dims - 1), " + "where the dims is the axis's size"); + } + + framework::DDim out_dims(x_dims); + for (size_t i = 0; i < axis_size; i++) { + out_dims[i] = x_dims[axis[i]]; + } + ctx->SetOutputDim("Out", out_dims); + } +}; + +class TransposeOpMaker : public framework::OpProtoAndCheckerMaker { + public: + TransposeOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput( + "X", + "(Tensor)The input tensor, tensors with rank at most 6 are supported"); + AddOutput("Out", "(Tensor)The output tensor"); + AddAttr>( + "axis", + "(vector)a list of values, and the size of the list should be " + "the same with the input tensor rank, the tensor will " + "permute the axes according the the values given"); + AddComment(R"DOC( +The Tensor will be permuted according to the axis values given. +The op is very much like the numpy.transpose function in python +For example: + >> input = numpy.arange(6).reshape((2,3)) + >> input + array([[0, 1, 2], + [3, 4, 5]]) + >> axis = [1, 0] + >> output = input.transpose(axis) + >> output + array([[0, 3], + [1, 4], + [2, 5]]) +So, given a input tensor of shape(N, C, H, W) and the axis is {0, 2, 3, 1}, +the output tensor shape will be (N, H, W, C) +)DOC"); + } +}; + +class TransposeOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto x_dims = ctx->GetInputDim("X"); + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), x_dims); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(transpose, ops::TransposeOp, ops::TransposeOpMaker, transpose_grad, + ops::TransposeOpGrad); +REGISTER_OP_CPU_KERNEL(transpose, + ops::TransposeKernel); +REGISTER_OP_CPU_KERNEL( + transpose_grad, + ops::TransposeGradKernel); diff --git a/paddle/operators/transpose_op.cu b/paddle/operators/transpose_op.cu new file mode 100644 index 0000000000..af3f581462 --- /dev/null +++ b/paddle/operators/transpose_op.cu @@ -0,0 +1,22 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/transpose_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(transpose, + ops::TransposeKernel); +REGISTER_OP_GPU_KERNEL( + transpose_grad, + ops::TransposeGradKernel); diff --git a/paddle/operators/transpose_op.h b/paddle/operators/transpose_op.h new file mode 100644 index 0000000000..aaa3f47ab5 --- /dev/null +++ b/paddle/operators/transpose_op.h @@ -0,0 +1,128 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +void EigenTranspose(const framework::ExecutionContext& context, + const framework::Tensor& in, framework::Tensor& out, + std::vector axis) { + Eigen::array permute; + for (int i = 0; i < Rank; i++) { + permute[i] = axis[i]; + } + auto in_dim = in.dims(); + auto out_dim = out.dims(); + + auto eigen_in = framework::EigenTensor::From(in); + auto eigen_out = framework::EigenTensor::From(out); + auto& dev = context.GetEigenDevice(); + eigen_out.device(dev) = eigen_in.shuffle(permute); +} + +template +class TransposeKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* x = context.Input("X"); + auto* out = context.Output("Out"); + out->mutable_data(context.GetPlace()); + + std::vector axis = context.Attr>("axis"); + int ndims = axis.size(); + switch (ndims) { + case 1: + EigenTranspose(context, *x, *out, axis); + break; + case 2: + EigenTranspose(context, *x, *out, axis); + break; + case 3: + EigenTranspose(context, *x, *out, axis); + break; + case 4: + EigenTranspose(context, *x, *out, axis); + break; + case 5: + EigenTranspose(context, *x, *out, axis); + break; + case 6: + EigenTranspose(context, *x, *out, axis); + break; + default: + PADDLE_THROW("Tensors with rank at most 6 are supported"); + } + } +}; + +template +class TransposeGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* out_grad = + context.Input(framework::GradVarName("Out")); + auto* x_grad = + context.Output(framework::GradVarName("X")); + if (x_grad) { + x_grad->mutable_data(context.GetPlace()); + + std::vector axis = context.Attr>("axis"); + std::vector reversed_axis(axis); + + for (size_t i = 0; i < axis.size(); i++) { + reversed_axis[axis[i]] = i; + } + + int ndims = axis.size(); + + switch (ndims) { + case 1: + EigenTranspose(context, *out_grad, *x_grad, + reversed_axis); + break; + case 2: + EigenTranspose(context, *out_grad, *x_grad, + reversed_axis); + break; + case 3: + EigenTranspose(context, *out_grad, *x_grad, + reversed_axis); + break; + case 4: + EigenTranspose(context, *out_grad, *x_grad, + reversed_axis); + break; + case 5: + EigenTranspose(context, *out_grad, *x_grad, + reversed_axis); + break; + case 6: + EigenTranspose(context, *out_grad, *x_grad, + reversed_axis); + break; + default: + PADDLE_THROW("Tensors with rank at most 6 are supported"); + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc index 184bcbc29c..97b1d0bed4 100644 --- a/paddle/operators/uniform_random_op.cc +++ b/paddle/operators/uniform_random_op.cc @@ -21,20 +21,20 @@ namespace operators { // Use std::random and thrust::random(thrust is a std library in CUDA) to // implement uniform random. template -class CPUUniformRandomKernel : public framework::OpKernel { +class CPUUniformRandomKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& context) const override { - auto* tensor = context.Output("Out"); - T* data = tensor->mutable_data(context.GetPlace()); - unsigned int seed = static_cast(context.Attr("seed")); + void Compute(const framework::ExecutionContext& ctx) const override { + auto* tensor = ctx.Output("Out"); + T* data = tensor->mutable_data(ctx.GetPlace()); + unsigned int seed = static_cast(ctx.Attr("seed")); std::minstd_rand engine; if (seed == 0) { seed = std::random_device()(); } engine.seed(seed); std::uniform_real_distribution dist( - static_cast(context.Attr("min")), - static_cast(context.Attr("max"))); + static_cast(ctx.Attr("min")), + static_cast(ctx.Attr("max"))); int64_t size = tensor->numel(); for (int64_t i = 0; i < size; ++i) { data[i] = dist(engine); @@ -47,21 +47,25 @@ class UniformRandomOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE_NOT_NULL( - ctx.OutputVar("Out"), - "Output(Out) of UniformRandomOp should not be null."); + void InferShape(framework::InferShapeContextBase* ctx) const override { + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of UniformRandomOp should not be null."); - PADDLE_ENFORCE(Attr("min") < Attr("max"), - "uniform_random's min must less then max"); - auto* tensor = ctx.Output("Out"); + PADDLE_ENFORCE( + ctx->Attrs().Get("min") < ctx->Attrs().Get("max"), + "uniform_random's min must less then max"); auto dims = Attr>("dims"); std::vector temp; temp.reserve(dims.size()); for (auto dim : dims) { temp.push_back(static_cast(dim)); } - tensor->Resize(framework::make_ddim(temp)); + ctx->SetOutputDim("Out", framework::make_ddim(temp)); + } + + framework::DataType IndicateDataType( + const framework::ExecutionContext& ctx) const override { + return static_cast(Attr("data_type")); } }; @@ -81,6 +85,8 @@ Used to initialize tensor with uniform random generator. "Random seed of uniform random. " "0 means generate a seed by system") .SetDefault(0); + AddAttr("data_type", "output tensor data type") + .SetDefault(framework::DataType::FP32); } }; } // namespace operators diff --git a/paddle/operators/uniform_random_op.cu b/paddle/operators/uniform_random_op.cu index 6614b53b3f..5612ce9eb1 100644 --- a/paddle/operators/uniform_random_op.cu +++ b/paddle/operators/uniform_random_op.cu @@ -40,7 +40,7 @@ struct UniformGenerator { // Use std::random and thrust::random(thrust is a std library in CUDA) to // implement uniform random. template -class GPUUniformRandomKernel : public framework::OpKernel { +class GPUUniformRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* tensor = context.Output("Out"); diff --git a/paddle/parameter/FirstOrderOptimizer.h b/paddle/parameter/FirstOrderOptimizer.h index caa78acd98..895e8d6a63 100644 --- a/paddle/parameter/FirstOrderOptimizer.h +++ b/paddle/parameter/FirstOrderOptimizer.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include "ParameterOptimizer.h" +#include "ParameterUpdateFunctions.h" #include "Regularizer.h" namespace paddle { @@ -37,6 +38,15 @@ public: real torch_learningRate = optConfig_.learning_method() == "torch_momentum" ? 1.0 - paraConfig.momentum() : 1.0; +#ifdef PADDLE_USE_MKLDNN + sgdUpdate(learningRate_ * paraConfig.learning_rate() * + (firstTime_ ? 1.0 : torch_learningRate), + paraConfig.momentum(), + applyDecay_ ? paraConfig.decay_rate() : 0, + vecs[PARAMETER_VALUE].get(), + vecs[PARAMETER_GRADIENT].get(), + vecs[PARAMETER_MOMENTUM].get()); +#else vecs[PARAMETER_VALUE]->sgdUpdate( *vecs[PARAMETER_GRADIENT], *vecs[PARAMETER_MOMENTUM], @@ -44,6 +54,7 @@ public: (firstTime_ ? 1.0 : torch_learningRate), paraConfig.momentum(), applyDecay_ ? paraConfig.decay_rate() : 0); +#endif } virtual void finishBatch() { firstTime_ = false; } }; diff --git a/paddle/parameter/ParameterUpdateFunctions.cpp b/paddle/parameter/ParameterUpdateFunctions.cpp index c8af7105c7..8b3be062b6 100644 --- a/paddle/parameter/ParameterUpdateFunctions.cpp +++ b/paddle/parameter/ParameterUpdateFunctions.cpp @@ -30,6 +30,9 @@ void sgdUpdateCpu(real learningRate, const real* grad, real* momentumVec) { decayRate *= learningRate; +#ifdef PADDLE_USE_MKLDNN +#pragma omp parallel for +#endif for (size_t i = 0; i < size; ++i) { momentumVec[i] = momentum * momentumVec[i] - learningRate * grad[i] - decayRate * value[i]; diff --git a/paddle/platform/device_context.cc b/paddle/platform/device_context.cc index 93b472b41c..a9b6b79903 100644 --- a/paddle/platform/device_context.cc +++ b/paddle/platform/device_context.cc @@ -16,8 +16,8 @@ namespace paddle { namespace platform { template <> -Eigen::DefaultDevice* DeviceContext::get_eigen_device() - const { +Eigen::DefaultDevice* DeviceContext::GetEigenDevice< + platform::CPUPlace, Eigen::DefaultDevice>() const { return reinterpret_cast(this)->eigen_device(); } @@ -35,7 +35,13 @@ Eigen::DefaultDevice* CPUDeviceContext::eigen_device() const { Place CPUDeviceContext::GetPlace() const { return CPUPlace(); } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA + +template <> +Eigen::GpuDevice* +DeviceContext::GetEigenDevice() const { + return reinterpret_cast(this)->eigen_device(); +} class EigenCudaStreamDevice : public Eigen::StreamInterface { public: @@ -90,11 +96,6 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface { mutable unsigned int* semaphore_; }; -template <> -Eigen::GpuDevice* DeviceContext::get_eigen_device() const { - return reinterpret_cast(this)->eigen_device(); -} - CUDADeviceContext::CUDADeviceContext(GPUPlace place) : place_(place) { SetDeviceId(place_.device); PADDLE_ENFORCE(cudaStreamCreate(&stream_)); diff --git a/paddle/platform/device_context.h b/paddle/platform/device_context.h index a106592e45..ef5f19214d 100644 --- a/paddle/platform/device_context.h +++ b/paddle/platform/device_context.h @@ -14,7 +14,7 @@ limitations under the License. */ #include "paddle/platform/enforce.h" #include "paddle/platform/place.h" -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include "paddle/platform/dynload/cublas.h" #include "paddle/platform/dynload/cudnn.h" #include "paddle/platform/gpu_info.h" @@ -27,20 +27,31 @@ limitations under the License. */ namespace paddle { namespace platform { +template +struct EigenDeviceConverter; + +template <> +struct EigenDeviceConverter { + using EigenDeviceType = Eigen::DefaultDevice; +}; + class DeviceContext { public: virtual ~DeviceContext() {} virtual Place GetPlace() const = 0; - template - DeviceType* get_eigen_device() const; + template ::EigenDeviceType> + DeviceType* GetEigenDevice() const; + + virtual void Wait() const {} }; class CPUDeviceContext : public DeviceContext { public: CPUDeviceContext(); explicit CPUDeviceContext(CPUPlace place); - virtual ~CPUDeviceContext() {} Eigen::DefaultDevice* eigen_device() const; @@ -50,7 +61,12 @@ class CPUDeviceContext : public DeviceContext { std::unique_ptr eigen_device_; }; -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA +template <> +struct EigenDeviceConverter { + using EigenDeviceType = Eigen::GpuDevice; +}; + class EigenCudaStreamDevice; class CUDADeviceContext : public DeviceContext { @@ -59,7 +75,7 @@ class CUDADeviceContext : public DeviceContext { virtual ~CUDADeviceContext(); /*! \brief Wait for all operations completion in the stream. */ - void Wait() const; + void Wait() const override; /*! \brief Return place in the device context. */ Place GetPlace() const override; diff --git a/paddle/platform/device_context_test.cc b/paddle/platform/device_context_test.cc index 5883a55272..8bf5174c4a 100644 --- a/paddle/platform/device_context_test.cc +++ b/paddle/platform/device_context_test.cc @@ -20,11 +20,11 @@ TEST(Device, Init) { using paddle::platform::CUDADeviceContext; using paddle::platform::GPUPlace; - int count = paddle::platform::GetDeviceCount(); + int count = paddle::platform::GetCUDADeviceCount(); for (int i = 0; i < count; i++) { DeviceContext* device_context = new CUDADeviceContext(GPUPlace(i)); Eigen::GpuDevice* gpu_device = - device_context->template get_eigen_device(); + device_context->template GetEigenDevice(); ASSERT_NE(nullptr, gpu_device); delete device_context; } @@ -34,7 +34,7 @@ TEST(Device, CUDADeviceContext) { using paddle::platform::CUDADeviceContext; using paddle::platform::GPUPlace; - int count = paddle::platform::GetDeviceCount(); + int count = paddle::platform::GetCUDADeviceCount(); for (int i = 0; i < count; i++) { CUDADeviceContext* device_context = new CUDADeviceContext(GPUPlace(i)); Eigen::GpuDevice* gpu_device = device_context->eigen_device(); diff --git a/paddle/platform/enforce.h b/paddle/platform/enforce.h index df5f71ed76..15d8446cd8 100644 --- a/paddle/platform/enforce.h +++ b/paddle/platform/enforce.h @@ -29,7 +29,7 @@ limitations under the License. */ #include // for __cxa_demangle #endif -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include "paddle/platform/dynload/cublas.h" #include "paddle/platform/dynload/cudnn.h" @@ -107,13 +107,13 @@ struct EnforceNotMet : public std::exception { template inline typename std::enable_if::type throw_on_error( - int stat, const Args&... args) { + bool stat, const Args&... args) { if (UNLIKELY(!(stat))) { throw std::runtime_error(string::Sprintf(args...)); } } -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA template inline typename std::enable_if::type throw_on_error( @@ -185,7 +185,7 @@ inline void throw_on_error(T e) { std::make_exception_ptr( \ std::runtime_error(paddle::string::Sprintf(__VA_ARGS__))), \ __FILE__, __LINE__); \ - } while (0) + } while (false) #define PADDLE_ENFORCE(...) \ do { \ @@ -195,7 +195,7 @@ inline void throw_on_error(T e) { throw ::paddle::platform::EnforceNotMet(std::current_exception(), \ __FILE__, __LINE__); \ } \ - } while (0) + } while (false) /* * Some enforce helpers here, usage: diff --git a/paddle/platform/enforce_test.cc b/paddle/platform/enforce_test.cc index 80bdee3d9d..8206a055ea 100644 --- a/paddle/platform/enforce_test.cc +++ b/paddle/platform/enforce_test.cc @@ -213,4 +213,4 @@ 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); -} \ No newline at end of file +} diff --git a/paddle/platform/gpu_info.cc b/paddle/platform/gpu_info.cc index be381a4e26..70ad611d5d 100644 --- a/paddle/platform/gpu_info.cc +++ b/paddle/platform/gpu_info.cc @@ -26,11 +26,11 @@ DEFINE_double(fraction_of_gpu_memory_to_use, 0.95, namespace paddle { namespace platform { -int GetDeviceCount() { +int GetCUDADeviceCount() { int count; PADDLE_ENFORCE( cudaGetDeviceCount(&count), - "cudaGetDeviceCount failed in paddle::platform::GetDeviceCount"); + "cudaGetDeviceCount failed in paddle::platform::GetCUDADeviceCount"); return count; } diff --git a/paddle/platform/gpu_info.h b/paddle/platform/gpu_info.h index ed2420b874..fb33db07bd 100644 --- a/paddle/platform/gpu_info.h +++ b/paddle/platform/gpu_info.h @@ -14,7 +14,7 @@ limitations under the License. */ #pragma once -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA #include #include @@ -28,7 +28,7 @@ const std::string kEnvFractionGpuMemoryToUse = "PADDLE_FRACTION_GPU_MEMORY_TO_USE"; //! Get the total number of GPU devices in system. -int GetDeviceCount(); +int GetCUDADeviceCount(); //! Get the current GPU device id in system. int GetCurrentDeviceId(); @@ -36,7 +36,7 @@ int GetCurrentDeviceId(); //! Set the GPU device id for next execution. void SetDeviceId(int device_id); -//！Get the memory usage of current GPU device. +//! Get the memory usage of current GPU device. void GpuMemoryUsage(size_t &available, size_t &total); //! Get the maximum allocation size of current GPU device. diff --git a/paddle/platform/hostdevice.h b/paddle/platform/hostdevice.h index e7de86b7b2..eb2df291cc 100644 --- a/paddle/platform/hostdevice.h +++ b/paddle/platform/hostdevice.h @@ -2,8 +2,10 @@ #ifdef __CUDACC__ #define HOSTDEVICE __host__ __device__ +#define DEVICE __device__ #define HOST __host__ #else #define HOSTDEVICE +#define DEVICE #define HOST #endif diff --git a/paddle/platform/macros.h b/paddle/platform/macros.h index 4a04a38c0c..feae7bdd77 100644 --- a/paddle/platform/macros.h +++ b/paddle/platform/macros.h @@ -16,8 +16,10 @@ limitations under the License. */ // Disable the copy and assignment operator for a class. #ifndef DISABLE_COPY_AND_ASSIGN -#define DISABLE_COPY_AND_ASSIGN(classname) \ - private: \ - classname(const classname&) = delete; \ - classname& operator=(const classname&) = delete +#define DISABLE_COPY_AND_ASSIGN(classname) \ + private: \ + classname(const classname&) = delete; \ + classname(const classname&&) = delete; \ + classname& operator=(const classname&) = delete; \ + classname& operator=(const classname&&) = delete #endif diff --git a/paddle/platform/place.cc b/paddle/platform/place.cc index b31515e1f0..856e54df89 100644 --- a/paddle/platform/place.cc +++ b/paddle/platform/place.cc @@ -47,7 +47,7 @@ bool is_cpu_place(const Place &p) { } bool places_are_same_class(const Place &p1, const Place &p2) { - return is_gpu_place(p1) == is_gpu_place(p2); + return p1.which() == p2.which(); } std::ostream &operator<<(std::ostream &os, const Place &p) { diff --git a/paddle/platform/place.h b/paddle/platform/place.h index 1117476bb3..0efc693234 100644 --- a/paddle/platform/place.h +++ b/paddle/platform/place.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include + #include "paddle/platform/variant.h" namespace paddle { @@ -46,8 +47,18 @@ struct IsGPUPlace : public boost::static_visitor { bool operator()(const GPUPlace &gpu) const { return true; } }; +// Define the max number of Place in bit length. i.e., the max number of places +// should be less equal than 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT) +#define NUM_PLACE_TYPE_LIMIT_IN_BIT 4 + typedef boost::variant Place; +// static check number of place types is less equal than +// 2^(NUM_PLACE_TYPE_LIMIT_IN_BIT) +BOOST_MPL_ASSERT((boost::mpl::less_equal< + Place::types::size, + boost::mpl::long_<1 << NUM_PLACE_TYPE_LIMIT_IN_BIT>>)); + void set_place(const Place &); const Place &get_place(); diff --git a/paddle/platform/transform.h b/paddle/platform/transform.h index 8eaab047fd..f196868c72 100644 --- a/paddle/platform/transform.h +++ b/paddle/platform/transform.h @@ -29,45 +29,71 @@ namespace paddle { namespace platform { + // Transform on host or device. It provides the same API in std library. -template -void Transform(const DeviceContext& context, InputIter first, InputIter last, - OutputIter result, UnaryOperation op) { - auto place = context.GetPlace(); - if (is_cpu_place(place)) { +template +struct Transform { + template + void operator()(const DeviceContext& context, InputIter first, InputIter last, + OutputIter result, UnaryOperation op); + + template + void operator()(const DeviceContext& context, InputIter1 first1, + InputIter1 last1, InputIter2 first2, OutputIter result, + BinaryOperation op); +}; + +template <> +struct Transform { + template + void operator()(const DeviceContext& context, InputIter first, InputIter last, + OutputIter result, UnaryOperation op) { + auto place = context.GetPlace(); + PADDLE_ENFORCE(is_cpu_place(place), "It must use CPU place."); std::transform(first, last, result, op); - } else { -#ifdef __NVCC__ - auto& ctx = reinterpret_cast(context); - using namespace details; - thrust::transform(thrust::cuda::par.on(ctx.stream()), DevPtrCast(first), - DevPtrCast(last), DevPtrCast(result), op); -#else - PADDLE_THROW("Do not invoke `Transform` in .cc file"); -#endif } -} -template -void Transform(const DeviceContext& context, InputIter1 first1, - InputIter1 last1, InputIter2 first2, OutputIter result, - BinaryOperation op) { - auto place = context.GetPlace(); - if (is_cpu_place(place)) { + template + void operator()(const DeviceContext& context, InputIter1 first1, + InputIter1 last1, InputIter2 first2, OutputIter result, + BinaryOperation op) { + auto place = context.GetPlace(); + PADDLE_ENFORCE(is_cpu_place(place), "It must use CPU place."); std::transform(first1, last1, first2, result, op); - } else { + } +}; + #ifdef __NVCC__ +template <> +struct Transform { + template + void operator()(const DeviceContext& context, InputIter first, InputIter last, + OutputIter result, UnaryOperation op) { + auto place = context.GetPlace(); + PADDLE_ENFORCE(is_gpu_place(place), "It must use GPU place."); auto& ctx = reinterpret_cast(context); - using namespace details; - thrust::transform(thrust::cuda::par.on(ctx.stream()), DevPtrCast(first1), - DevPtrCast(last1), DevPtrCast(first2), DevPtrCast(result), + thrust::transform(thrust::cuda::par.on(ctx.stream()), + details::DevPtrCast(first), details::DevPtrCast(last), + details::DevPtrCast(result), op); + } + + template + void operator()(const DeviceContext& context, InputIter1 first1, + InputIter1 last1, InputIter2 first2, OutputIter result, + BinaryOperation op) { + auto place = context.GetPlace(); + PADDLE_ENFORCE(is_gpu_place(place), "It must use GPU place."); + auto& ctx = reinterpret_cast(context); + thrust::transform(thrust::cuda::par.on(ctx.stream()), + details::DevPtrCast(first1), details::DevPtrCast(last1), + details::DevPtrCast(first2), details::DevPtrCast(result), op); -#else - PADDLE_THROW("Do not invoke `Transform` in .cc file"); -#endif } }; +#endif } // namespace platform } // namespace paddle diff --git a/paddle/platform/transform_test.cu b/paddle/platform/transform_test.cu index b8a6200bb0..c76cab80e4 100644 --- a/paddle/platform/transform_test.cu +++ b/paddle/platform/transform_test.cu @@ -15,6 +15,7 @@ #include #include "paddle/memory/memcpy.h" #include "paddle/memory/memory.h" +#include "paddle/platform/hostdevice.h" #include "paddle/platform/transform.h" template @@ -38,7 +39,8 @@ TEST(Transform, CPUUnary) { using namespace paddle::platform; CPUDeviceContext ctx; float buf[4] = {0.1, 0.2, 0.3, 0.4}; - Transform(ctx, buf, buf + 4, buf, Scale(10)); + Transform trans; + trans(ctx, buf, buf + 4, buf, Scale(10)); for (int i = 0; i < 4; ++i) { ASSERT_NEAR(buf[i], static_cast(i + 1), 1e-5); } @@ -52,7 +54,8 @@ TEST(Transform, GPUUnary) { float cpu_buf[4] = {0.1, 0.2, 0.3, 0.4}; float* gpu_buf = static_cast(Alloc(gpu0, sizeof(float) * 4)); Copy(gpu0, gpu_buf, CPUPlace(), cpu_buf, sizeof(cpu_buf)); - Transform(ctx, gpu_buf, gpu_buf + 4, gpu_buf, Scale(10)); + Transform trans; + trans(ctx, gpu_buf, gpu_buf + 4, gpu_buf, Scale(10)); ctx.Wait(); Copy(CPUPlace(), cpu_buf, gpu0, gpu_buf, sizeof(cpu_buf)); Free(gpu0, gpu_buf); @@ -65,7 +68,9 @@ TEST(Transform, CPUBinary) { using namespace paddle::platform; using namespace paddle::memory; int buf[4] = {1, 2, 3, 4}; - Transform(CPUDeviceContext(), buf, buf + 4, buf, buf, Multiply()); + Transform trans; + CPUDeviceContext ctx; + trans(ctx, buf, buf + 4, buf, buf, Multiply()); for (int i = 0; i < 4; ++i) { ASSERT_EQ((i + 1) * (i + 1), buf[i]); } @@ -79,11 +84,12 @@ TEST(Transform, GPUBinary) { CUDADeviceContext ctx(gpu0); int* gpu_buf = static_cast(Alloc(gpu0, sizeof(buf))); Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf)); - Transform(ctx, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply()); + Transform trans; + trans(ctx, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply()); ctx.Wait(); Copy(CPUPlace(), buf, gpu0, gpu_buf, sizeof(buf)); Free(gpu0, gpu_buf); for (int i = 0; i < 4; ++i) { ASSERT_EQ((i + 1) * (i + 1), buf[i]); } -} \ No newline at end of file +} diff --git a/paddle/platform/variant.h b/paddle/platform/variant.h index c2257af1b5..619897ca19 100644 --- a/paddle/platform/variant.h +++ b/paddle/platform/variant.h @@ -16,7 +16,7 @@ #include -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA // Because boost's variadic templates has bug on nvcc, boost will disable // variadic template support when GPU enabled on nvcc. @@ -29,4 +29,6 @@ #endif #endif +#include +#include #include diff --git a/paddle/pserver/CMakeLists.txt b/paddle/pserver/CMakeLists.txt index 2245c7d88c..ccfc0e7602 100644 --- a/paddle/pserver/CMakeLists.txt +++ b/paddle/pserver/CMakeLists.txt @@ -45,14 +45,18 @@ add_dependencies(paddle_pserver paddle_proto ${external_project_dependencies}) set(PSERVER_MAIN_SOURCES ParameterServer2Main.cpp) -add_executable(paddle_pserver_main - ${PSERVER_MAIN_SOURCES}) -link_paddle_exe(paddle_pserver_main) if(WITH_TESTING) add_subdirectory(test) endif() -install(TARGETS paddle_pserver_main - RUNTIME DESTINATION opt/paddle/bin - PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ - GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ) -set_target_properties(paddle_pserver_main PROPERTIES INSTALL_RPATH_USE_LINK_PATH TRUE) + +if(NOT WITH_C_API) + add_executable(paddle_pserver_main ${PSERVER_MAIN_SOURCES}) + link_paddle_exe(paddle_pserver_main) + + install(TARGETS paddle_pserver_main + RUNTIME DESTINATION opt/paddle/bin + PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ + GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ) + + set_target_properties(paddle_pserver_main PROPERTIES INSTALL_RPATH_USE_LINK_PATH TRUE) +endif() diff --git a/paddle/pserver/test/SocketTest.cpp b/paddle/pserver/test/SocketTest.cpp index 6f6c9e596c..b43461d61b 100644 --- a/paddle/pserver/test/SocketTest.cpp +++ b/paddle/pserver/test/SocketTest.cpp @@ -215,7 +215,7 @@ int main(int argc, char** argv) { uint64_t dataSize = FLAGS_dim * sizeof(real); -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA GpuVector gpuParam(FLAGS_dim); GpuVector gpuGrad(FLAGS_dim); #else diff --git a/paddle/pserver/test/test_ProtoServer.cpp b/paddle/pserver/test/test_ProtoServer.cpp index 04236fda2f..ad8ffed9c1 100644 --- a/paddle/pserver/test/test_ProtoServer.cpp +++ b/paddle/pserver/test/test_ProtoServer.cpp @@ -99,7 +99,7 @@ TEST(ProtoServer, regular) { } TEST(ProtoServer, extended) { -#ifndef PADDLE_ONLY_CPU +#ifdef PADDLE_WITH_CUDA ProtoClient* client; if (FLAGS_rdma_tcp == "rdma") client = new ProtoClient(FLAGS_server_addr, FLAGS_port, F_RDMA); diff --git a/paddle/pybind/.clang-format b/paddle/pybind/.clang-format new file mode 120000 index 0000000000..7d28cb3924 --- /dev/null +++ b/paddle/pybind/.clang-format @@ -0,0 +1 @@ +../framework/.clang-format \ No newline at end of file diff --git a/paddle/pybind/CMakeLists.txt b/paddle/pybind/CMakeLists.txt index 4f05406c7f..97364f2db9 100644 --- a/paddle/pybind/CMakeLists.txt +++ b/paddle/pybind/CMakeLists.txt @@ -1,6 +1,6 @@ if(WITH_PYTHON) cc_library(paddle_pybind SHARED - SRCS pybind.cc - DEPS pybind python backward + SRCS pybind.cc exception.cc protobuf.cc + DEPS pybind python backward proto_desc tensor_array ${GLOB_OP_LIB}) endif(WITH_PYTHON) diff --git a/paddle/pybind/exception.cc b/paddle/pybind/exception.cc new file mode 100644 index 0000000000..ff79b12ee4 --- /dev/null +++ b/paddle/pybind/exception.cc @@ -0,0 +1,34 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/pybind/exception.h" + +namespace paddle { +namespace pybind { + +void BindException(pybind11::module& m) { + static pybind11::exception exc(m, "EnforceNotMet"); + pybind11::register_exception_translator([](std::exception_ptr p) { + try { + if (p) std::rethrow_exception(p); + } catch (const platform::EnforceNotMet& e) { + exc(e.what()); + } + }); + + m.def("__unittest_throw_exception__", [] { PADDLE_THROW("test exception"); }); +} + +} // namespace pybind +} // namespace paddle diff --git a/paddle/pybind/exception.h b/paddle/pybind/exception.h new file mode 100644 index 0000000000..70beac1460 --- /dev/null +++ b/paddle/pybind/exception.h @@ -0,0 +1,24 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include +#include "paddle/platform/enforce.h" +#include "pybind11/pybind11.h" +namespace paddle { +namespace pybind { + +extern void BindException(pybind11::module& m); +} // namespace pybind +} // namespace paddle diff --git a/paddle/pybind/protobuf.cc b/paddle/pybind/protobuf.cc new file mode 100644 index 0000000000..218821b35b --- /dev/null +++ b/paddle/pybind/protobuf.cc @@ -0,0 +1,206 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/pybind/protobuf.h" +#include +#include +#include "paddle/framework/block_desc.h" +#include "paddle/framework/op_desc.h" +#include "paddle/framework/program_desc.h" +#include "paddle/framework/var_desc.h" + +// Cast boost::variant for PyBind. +// Copy from +// https://github.com/pybind/pybind11/issues/576#issuecomment-269563199 +namespace pybind11 { +namespace detail { + +// Can be replaced by a generic lambda in C++14 +struct variant_caster_visitor : public boost::static_visitor { + return_value_policy policy; + handle parent; + + variant_caster_visitor(return_value_policy policy, handle parent) + : policy(policy), parent(parent) {} + + template + handle operator()(T const &src) const { + return make_caster::cast(src, policy, parent); + } +}; + +template +struct variant_caster; + +template