diff --git a/.clang_format.hook b/.clang_format.hook new file mode 100755 index 0000000000000000000000000000000000000000..1d928216867c0ba3897d71542fea44debf8d72a0 --- /dev/null +++ b/.clang_format.hook @@ -0,0 +1,15 @@ +#!/bin/bash +set -e + +readonly VERSION="3.8" + +version=$(clang-format -version) + +if ! [[ $version == *"$VERSION"* ]]; then + echo "clang-format version check failed." + echo "a version contains '$VERSION' is needed, but get '$version'" + echo "you can install the right version, and make an soft-link to '\$PATH' env" + exit -1 +fi + +clang-format $@ diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index bb8c88787d37faf9ce4d7d856a307c11f1085d98..83fe9af768964003130d02b7d913ad1c2102dd1d 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -19,10 +19,10 @@ - id: end-of-file-fixer - repo: local hooks: - - id: clang-format + - id: clang-format-with-version-check name: clang-format description: Format files with ClangFormat. - entry: clang-format -i + entry: bash ./.clang_format.hook -i language: system files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$ - repo: https://github.com/PaddlePaddle/pre-commit-golang diff --git a/.travis.yml b/.travis.yml index b4b83fcdbc84ce0fb0c91c816ebc3c964acfa590..e217c8f5a740ef5ab7315656ed7839ffa219c805 100644 --- a/.travis.yml +++ b/.travis.yml @@ -4,7 +4,6 @@ cache: - $HOME/.ccache - $HOME/.cache/pip - $TRAVIS_BUILD_DIR/build/third_party - - $TRAVIS_BUILD_DIR/build_android/third_party sudo: required dist: trusty os: @@ -12,7 +11,6 @@ os: env: - JOB=build_doc - JOB=check_style - - JOB=build_android addons: apt: packages: @@ -23,7 +21,6 @@ addons: - python - python-pip - python2.7-dev - - python-numpy - python-wheel - libboost-dev - curl @@ -37,8 +34,8 @@ before_install: - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python # protobuf version. - - pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt - - pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker + - 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 diff --git a/CMakeLists.txt b/CMakeLists.txt index c75b83e50cf9cef8290c37f88b38cdc3d77df39c..08237cd850ae20c515a39c8783a18deaac431626 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -36,8 +36,8 @@ include(simd) ################################ Configurations ####################################### option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_FOUND}) option(WITH_AVX "Compile PaddlePaddle with AVX intrinsics" ${AVX_FOUND}) -option(WITH_MKLDNN "Compile PaddlePaddle with mkl-dnn support." OFF) -option(WITH_MKLML "Compile PaddlePaddle with mklml package." OFF) +option(WITH_MKLDNN "Compile PaddlePaddle with mkl-dnn support." ${AVX_FOUND}) +option(WITH_MKLML "Compile PaddlePaddle with mklml package." ${AVX_FOUND}) option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON) option(WITH_TESTING "Compile PaddlePaddle with unit testing" ON) option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON) @@ -55,6 +55,7 @@ option(WITH_C_API "Compile PaddlePaddle with C-API(Prediction)" OFF) option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF) option(GLIDE_INSTALL "Download and install go dependencies " ON) option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF) +option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) # CMAKE_BUILD_TYPE if(NOT CMAKE_BUILD_TYPE) @@ -64,8 +65,8 @@ if(NOT CMAKE_BUILD_TYPE) endif() if(ANDROID) - if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21") - message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 21") + if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16") + message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16") endif() set(WITH_GPU OFF CACHE STRING @@ -137,9 +138,9 @@ set(EXTERNAL_LIBS ) if(WITH_GPU) - list(APPEND EXTERNAL_LIB ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY}) + list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY}) if(NOT WITH_DSO) - list(APPEND EXTERNAL_LIB ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY}) + list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY}) endif(NOT WITH_DSO) endif(WITH_GPU) diff --git a/Dockerfile b/Dockerfile index 41b6729124228cec16be35d9b26da8042824b0b0..136db772cc6a24b8084120fa6bab666bc1eda78e 100644 --- a/Dockerfile +++ b/Dockerfile @@ -10,13 +10,11 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub ARG WITH_GPU ARG WITH_AVX ARG WITH_DOC -ARG WITH_STYLE_CHECK ENV WOBOQ OFF -ENV WITH_GPU=${WITH_GPU:-OFF} +ENV WITH_GPU=${WITH_GPU:-ON} ENV WITH_AVX=${WITH_AVX:-ON} ENV WITH_DOC=${WITH_DOC:-OFF} -ENV WITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} ENV HOME /root # Add bash enhancements @@ -34,9 +32,6 @@ RUN apt-get update && \ net-tools && \ apt-get clean -y -# paddle is using numpy.flip, which is introduced since 1.12.0 -RUN pip --no-cache-dir install 'numpy>=1.12.0' - # Install Go and glide RUN wget -qO- https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \ tar -xz -C /usr/local && \ @@ -58,33 +53,22 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8 # FIXME: due to temporary ipykernel dependency issue, specify ipykernel jupyter # version util jupyter fixes this issue. RUN pip install --upgrade pip && \ - pip install -U 'protobuf==3.1.0' && \ - pip install -U wheel pillow BeautifulSoup && \ + pip install -U wheel && \ pip install -U docopt PyYAML sphinx && \ - pip install -U sphinx-rtd-theme==0.1.9 recommonmark && \ - pip install pre-commit 'requests==2.9.2' 'ipython==5.3.0' && \ + pip install -U sphinx-rtd-theme==0.1.9 recommonmark + +RUN pip install pre-commit 'ipython==5.3.0' && \ pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \ - pip install opencv-python rarfile 'scipy>=0.19.0' 'nltk>=3.2.2' + pip install opencv-python + +COPY ./python/requirements.txt /root/ +RUN pip install -r /root/requirements.txt # To fix https://github.com/PaddlePaddle/Paddle/issues/1954, we use # the solution in https://urllib3.readthedocs.io/en/latest/user-guide.html#ssl-py2 RUN apt-get install -y libssl-dev libffi-dev RUN pip install certifi urllib3[secure] -# TODO(qijun) The template library Eigen doesn't work well with GCC 5 -# coming with the default Docker image, so we switch to use GCC 4.8 -# by default. And I will check Eigen library later. - -RUN ln -sf gcc-4.8 /usr/bin/gcc && \ - ln -sf gcc-ar-4.8 /usr/bin/gcc-ar && \ - ln -sf gcc-nm-4.8 /usr/bin/gcc-nm && \ - ln -sf gcc-ranlib-4.8 /usr/bin/gcc-ranlib && \ - ln -sf gcc-4.8 /usr/bin/x86_64-linux-gnu-gcc && \ - ln -sf gcc-ar-4.8 /usr/bin/x86_64-linux-gnu-gcc-ar && \ - ln -sf gcc-nm-4.8 /usr/bin/x86_64-linux-gnu-gcc-nm && \ - ln -sf gcc-ranlib-4.8 /usr/bin/x86_64-linux-gnu-gcc-ranlib && \ - ln -sf g++-4.8 /usr/bin/g++ && \ - ln -sf g++-4.8 /usr/bin/x86_64-linux-gnu-g++ # Install woboq_codebrowser to /woboq RUN git clone https://github.com/woboq/woboq_codebrowser /woboq && \ diff --git a/Dockerfile.android b/Dockerfile.android index c0fa58c384f9ebcae60477ffce49ea4ffa929db9..452aa1574550627c2cce6375e12e154a9763254d 100644 --- a/Dockerfile.android +++ b/Dockerfile.android @@ -4,9 +4,15 @@ MAINTAINER PaddlePaddle Authors ARG UBUNTU_MIRROR RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi' +# ENV variables +ARG ANDROID_ABI + +ENV ANDROID_ABI=${ANDROID_ABI:-"armeabi-v7a"} + ENV HOME=/root \ ANDROID_NDK_HOME=/opt/android-ndk-linux \ - ANDROID_STANDALONE_TOOLCHAIN=/opt/android-toolchain-gcc + ANDROID_ARM_STANDALONE_TOOLCHAIN=/opt/arm-toolchain \ + ANDROID_ARM64_STANDALONE_TOOLCHAIN=/opt/arm64-toolchain RUN apt-get update && \ apt-get install -y \ @@ -15,12 +21,11 @@ RUN apt-get update && \ apt-get clean -y # Install Go and glide -RUN wget -O go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz && \ - tar -C /usr/local -xzf go.tgz && \ +RUN wget -qO- go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \ + tar -xz -C /usr/local && \ mkdir /root/gopath && \ mkdir /root/gopath/bin && \ - mkdir /root/gopath/src && \ - rm go.tgz + mkdir /root/gopath/src ENV GOROOT=/usr/local/go GOPATH=/root/gopath # should not be in the same line with GOROOT definition, otherwise docker build could not find GOROOT. ENV PATH=${PATH}:${GOROOT}/bin:${GOPATH}/bin @@ -42,7 +47,8 @@ RUN mkdir /opt/android-ndk-tmp && \ wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \ unzip -q android-ndk-r14b-linux-x86_64.zip && \ mv android-ndk-r14b ${ANDROID_NDK_HOME} && \ - ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-21 --install-dir=${ANDROID_STANDALONE_TOOLCHAIN} && \ + ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-23 --install-dir=${ANDROID_ARM_STANDALONE_TOOLCHAIN} && \ + ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm64 --platform=android-23 --install-dir=${ANDROID_ARM64_STANDALONE_TOOLCHAIN} && \ rm -rf /opt/android-ndk-tmp && \ rm -rf ${ANDROID_NDK_HOME} diff --git a/cmake/configure.cmake b/cmake/configure.cmake index 209f9078a637ac581d90212a48216eb388c477ed..51c3b918cc4ef4cf6c8052ccc14028a872309fcf 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -28,6 +28,10 @@ if(NOT WITH_TIMER) add_definitions(-DPADDLE_DISABLE_TIMER) endif(NOT WITH_TIMER) +if(USE_EIGEN_FOR_BLAS) + add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS) +endif(USE_EIGEN_FOR_BLAS) + if(NOT WITH_PROFILER) add_definitions(-DPADDLE_DISABLE_PROFILER) endif(NOT WITH_PROFILER) diff --git a/cmake/cross_compiling/android.cmake b/cmake/cross_compiling/android.cmake index 5e3e437a8da9624df35a5c754fe77be73f20361d..84219cfa5587f5b765b2e8f35180797d7053169f 100644 --- a/cmake/cross_compiling/android.cmake +++ b/cmake/cross_compiling/android.cmake @@ -20,6 +20,7 @@ # The supported variables are listed belows: # # ANDROID_STANDALONE_TOOLCHAIN +# ANDROID_TOOLCHAIN # ANDROID_ABI # ANDROID_NATIVE_API_LEVEL # ANDROID_ARM_MODE @@ -57,6 +58,10 @@ IF(NOT DEFINED CMAKE_SYSTEM_VERSION AND ANDROID_NATIVE_API_LEVEL) ENDIF() ENDIF() +IF(NOT DEFINED ANDROID_TOOLCHAIN) + SET(ANDROID_TOOLCHAIN clang) +ENDIF() + IF(NOT DEFINED ANDROID_ABI) SET(ANDROID_ABI "armeabi-v7a") ENDIF() @@ -82,6 +87,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") "${CMAKE_VERSION}), when cross-compiling for Android.") IF(ANDROID_STANDALONE_TOOLCHAIN) + # Use standalone toolchain SET(CMAKE_SYSROOT "${ANDROID_STANDALONE_TOOLCHAIN}/sysroot") IF(NOT CMAKE_SYSTEM_VERSION) @@ -96,26 +102,44 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") ENDIF() # Toolchain - SET(ANDROID_TOOLCHAIN "gcc") SET(ANDROID_TOOLCHAIN_ROOT ${ANDROID_STANDALONE_TOOLCHAIN}) - IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$") - SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi) - IF(ANDROID_ABI STREQUAL "armeabi") - SET(CMAKE_SYSTEM_PROCESSOR armv5te) - ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a") - SET(CMAKE_SYSTEM_PROCESSOR armv7-a) - ENDIF() - ENDIF() - IF(ANDROID_ABI STREQUAL "arm64-v8a") - SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android) - SET(CMAKE_SYSTEM_PROCESSOR aarch64) + ELSE(ANDROID_NDK) + # TODO: use android ndk + ENDIF() + + IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$") + SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi) + IF(ANDROID_ABI STREQUAL "armeabi") + SET(CMAKE_SYSTEM_PROCESSOR armv5te) + SET(ANDROID_CLANG_TRIPLE armv5te-none-linux-androideabi) + ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a") + SET(CMAKE_SYSTEM_PROCESSOR armv7-a) + SET(ANDROID_CLANG_TRIPLE armv7-none-linux-androideabi) ENDIF() - SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-") + ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a") + SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android) + SET(CMAKE_SYSTEM_PROCESSOR aarch64) + SET(ANDROID_CLANG_TRIPLE aarch64-none-linux-android) + ELSE() + MESSAGE(FATAL_ERROR "Invalid Android ABI: ${ANDROID_ABI}.") + ENDIF() + SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-") + + IF(ANDROID_TOOLCHAIN STREQUAL clang) + SET(ANDROID_C_COMPILER_NAME clang) + SET(ANDROID_CXX_COMPILER_NAME clang++) + SET(CMAKE_C_COMPILER_TARGET ${ANDROID_CLANG_TRIPLE}) + SET(CMAKE_CXX_COMPILER_TARGET ${ANDROID_CLANG_TRIPLE}) + ELSEIF(ANDROID_TOOLCHAIN STREQUAL gcc) + SET(ANDROID_C_COMPILER_NAME gcc) + SET(ANDROID_CXX_COMPILER_NAME g++) + ELSE() + MESSAGE(FATAL_ERROR "Invalid Android toolchain: ${ANDROID_TOOLCHAIN}") ENDIF() # C compiler IF(NOT CMAKE_C_COMPILER) - SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}gcc") + SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_C_COMPILER_NAME}") ELSE() GET_FILENAME_COMPONENT(ANDROID_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM) ENDIF() @@ -125,7 +149,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") # CXX compiler IF(NOT CMAKE_CXX_COMPILER) - SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}g++") + SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_CXX_COMPILER_NAME}") ELSE() GET_FILENAME_COMPONENT(ANDROID_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM) ENDIF() @@ -137,7 +161,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") SET(CMAKE_CXX_COMPILER ${ANDROID_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE) # Toolchain and ABI specific flags. - SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections -finline-limit=64") + SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections") SET(ANDROID_LINKER_FLAGS "-Wl,--gc-sections") IF(ANDROID_ABI STREQUAL "armeabi") @@ -145,8 +169,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") -march=armv5te -mtune=xscale -msoft-float) - ENDIF() - IF(ANDROID_ABI STREQUAL "armeabi-v7a") + ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a") LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv7-a -mfloat-abi=softfp) @@ -156,6 +179,8 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") LIST(APPEND ANDROID_COMPILER_FLAGS -mfpu=vfpv3-d16) ENDIF() LIST(APPEND ANDROID_LINKER_FLAGS -Wl,--fix-cortex-a8) + ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a") + LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a) ENDIF() IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$") @@ -164,10 +189,18 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") ELSE() LIST(APPEND ANDROID_COMPILER_FLAGS -mthumb) ENDIF() + IF(ANDROID_TOOLCHAIN STREQUAL clang) + # Disable integrated-as for better compatibility. + LIST(APPEND ANDROID_COMPILER_FLAGS -fno-integrated-as) + ENDIF() ENDIF() - IF(ANDROID_ABI STREQUAL "arm64-v8a") - LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a) + IF(ANDROID_TOOLCHAIN STREQUAL clang) + # CMake automatically forwards all compiler flags to the linker, + # and clang doesn't like having -Wa flags being used for linking. + # To prevent CMake from doing this would require meddling with + # the CMAKE__COMPILE_OBJECT rules, which would get quite messy. + LIST(APPEND ANDROID_LINKER_FLAGS -Qunused-arguments) ENDIF() STRING(REPLACE ";" " " ANDROID_COMPILER_FLAGS "${ANDROID_COMPILER_FLAGS}") diff --git a/cmake/cudnn.cmake b/cmake/cudnn.cmake index 69f40df51680a104c47d9335c070c570dcaff59a..2c84061ff572de4687b4d496f8ded6deee8d1011 100644 --- a/cmake/cudnn.cmake +++ b/cmake/cudnn.cmake @@ -2,7 +2,7 @@ if(NOT WITH_GPU) return() endif() -set(CUDNN_ROOT "" CACHE PATH "CUDNN ROOT") +set(CUDNN_ROOT "/usr" CACHE PATH "CUDNN ROOT") find_path(CUDNN_INCLUDE_DIR cudnn.h PATHS ${CUDNN_ROOT} ${CUDNN_ROOT}/include $ENV{CUDNN_ROOT} $ENV{CUDNN_ROOT}/include ${CUDA_TOOLKIT_INCLUDE} diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake index 25c6b4ef52d3f8ebff1572ae8d348be7c577c08c..9686df00219001769d074ee815d9cc8db0258496 100644 --- a/cmake/external/mkldnn.cmake +++ b/cmake/external/mkldnn.cmake @@ -51,7 +51,7 @@ ExternalProject_Add( ${EXTERNAL_PROJECT_LOG_ARGS} DEPENDS ${MKLDNN_DEPENDS} GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git" - GIT_TAG "v0.9" + GIT_TAG "v0.10" PREFIX ${MKLDNN_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR} diff --git a/cmake/external/mklml.cmake b/cmake/external/mklml.cmake index e9fd3d4bedc983ae7c544cf289dc841cf22f9de4..74f3279831357c21038df133df0f5a432a6dfd20 100644 --- a/cmake/external/mklml.cmake +++ b/cmake/external/mklml.cmake @@ -28,7 +28,7 @@ INCLUDE(ExternalProject) SET(MKLML_PROJECT "extern_mklml") SET(MKLML_VER "mklml_lnx_2018.0.20170720") -SET(MKLML_URL "https://github.com/01org/mkl-dnn/releases/download/v0.9/${MKLML_VER}.tgz") +SET(MKLML_URL "https://github.com/01org/mkl-dnn/releases/download/v0.10/${MKLML_VER}.tgz") SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml") SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}") SET(MKLML_DST_DIR "mklml") @@ -54,7 +54,8 @@ ExternalProject_Add( ${EXTERNAL_PROJECT_LOG_ARGS} PREFIX ${MKLML_SOURCE_DIR} DOWNLOAD_DIR ${MKLML_DOWNLOAD_DIR} - DOWNLOAD_COMMAND wget --no-check-certificate -qO- ${MKLML_URL} | tar xz -C ${MKLML_DOWNLOAD_DIR} + DOWNLOAD_COMMAND wget --no-check-certificate ${MKLML_URL} -c -q -O ${MKLML_VER}.tgz + && tar zxf ${MKLML_VER}.tgz DOWNLOAD_NO_PROGRESS 1 UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLML_INSTALL_ROOT} diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index db09232c0e69016bf18c1d981e4620e9e804ff7c..f9e05af59fed7a8ad049390eda2c94d8577db1e7 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -12,6 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. +IF(USE_EIGEN_FOR_BLAS) + return() +ENDIF(USE_EIGEN_FOR_BLAS) + INCLUDE(cblas) IF(NOT ${CBLAS_FOUND}) @@ -25,7 +29,12 @@ IF(NOT ${CBLAS_FOUND}) "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}" CACHE FILEPATH "openblas library." FORCE) - SET(COMMON_ARGS CC=${CMAKE_C_COMPILER} NO_SHARED=1 NO_LAPACK=1 libs) + IF(APPLE) + SET(OPENBLAS_CC "${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() IF(CMAKE_CROSSCOMPILING) IF(ANDROID) @@ -40,11 +49,11 @@ IF(NOT ${CBLAS_FOUND}) SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=${TARGET} ARM_SOFTFP_ABI=1 USE_THREAD=0) ELSEIF(RPI) # use hardfp - SET(OPENBLAS_COMMIT "v0.2.19") + SET(OPENBLAS_COMMIT "v0.2.20") SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=ARMV7 USE_THREAD=0) ENDIF() ELSE() - SET(OPENBLAS_COMMIT "v0.2.19") + SET(OPENBLAS_COMMIT "v0.2.20") SET(OPTIONAL_ARGS "") IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^x86(_64)?$") SET(OPTIONAL_ARGS DYNAMIC_ARCH=1 NUM_THREADS=64) @@ -73,10 +82,18 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR}) # linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas) SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c) FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";") -ADD_LIBRARY(cblas STATIC ${dummyfile}) +IF(${CBLAS_PROVIDER} MATCHES MKL) + ADD_LIBRARY(cblas SHARED ${dummyfile}) +ELSE() + ADD_LIBRARY(cblas STATIC ${dummyfile}) +ENDIF() TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES}) IF(NOT ${CBLAS_FOUND}) ADD_DEPENDENCIES(cblas extern_openblas) LIST(APPEND external_project_dependencies cblas) +ELSE() + IF("${CBLAS_PROVIDER}" STREQUAL "MKLML") + ADD_DEPENDENCIES(cblas mklml) + ENDIF() ENDIF(NOT ${CBLAS_FOUND}) diff --git a/cmake/flags.cmake b/cmake/flags.cmake index b27eb71550b68b5c27e47bf067ae0df329bbd628..ff246b2eb4ed97dd14d45763569b661cefd203c8 100644 --- a/cmake/flags.cmake +++ b/cmake/flags.cmake @@ -9,13 +9,6 @@ function(CheckCompilerCXX11Flag) if(${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 4.8) message(FATAL_ERROR "Unsupported GCC version. GCC >= 4.8 required.") endif() - if(NOT ANDROID) - # TODO(qijun) gcc 4.9 or later versions raise SEGV due to the optimization problem. - # Use Debug mode instead for now. - if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 4.9 OR CMAKE_CXX_COMPILER_VERSION VERSION_EQUAL 4.9) - set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "" FORCE) - endif() - endif() elseif(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang") # cmake >= 3.0 compiler id "AppleClang" on Mac OS X, otherwise "Clang" # Apple Clang is a different compiler than upstream Clang which havs different version numbers. @@ -160,7 +153,7 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF) # Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc. # So, don't set these flags here. -LIST(APPEND CUDA_NVCC_FLAGS -std=c++11 --default-stream per-thread) +LIST(APPEND CUDA_NVCC_FLAGS -std=c++11) LIST(APPEND CUDA_NVCC_FLAGS --use_fast_math) if(CMAKE_BUILD_TYPE STREQUAL "Debug") diff --git a/doc/about/index_cn.md b/doc/about/index_cn.md deleted file mode 100644 index 3bf030004d4de8c6f3cb773c6e78c09f40878c5f..0000000000000000000000000000000000000000 --- a/doc/about/index_cn.md +++ /dev/null @@ -1,11 +0,0 @@ -关于PaddlePaddle -================ - -PaddlePaddle是一个最早由百度科学家和工程师共同研发的并行分布式深度学习平台，兼备易用性、高效性、灵活性和可扩展性，目前已被百度内部多个产品线广泛使用。 -PaddlePaddle目前已经开放源码, 但是远未完善，我们希望能在这个基础上不断的改进、扩展和延伸。 -同时我们希望广大开发者积极提供反馈和贡献源代码，建立一个活跃的开源社区。 - -致谢 --------- - -在此，特别感谢PaddlePaddle的[所有贡献者](https://github.com/PaddlePaddle/Paddle/graphs/contributors)。 diff --git a/doc/about/index_en.rst b/doc/about/index_en.rst deleted file mode 100644 index 065c430cdea802ed3c9f487cd00255b85a5598a5..0000000000000000000000000000000000000000 --- a/doc/about/index_en.rst +++ /dev/null @@ -1,14 +0,0 @@ -ABOUT -======= - -PaddlPaddle is an easy-to-use, efficient, flexible and scalable deep learning platform, -which is originally developed by Baidu scientists and engineers for the purpose of applying deep learning to many products at Baidu. - -PaddlePaddle is now open source but far from complete, which is intended to be built upon, improved, scaled, and extended. -We hope to build an active open source community both by providing feedback and by actively contributing to the source code. - - -Credits --------- - -We owe many thanks to `all contributors and developers `_ of PaddlePaddle! diff --git a/doc/api/v2/config/layer.rst b/doc/api/v2/config/layer.rst index cb330ea5e1b914587a725c9b90a33053f3fbbc3d..c94627a72806fa2eca77c79da24f7f3ca18f0259 100644 --- a/doc/api/v2/config/layer.rst +++ b/doc/api/v2/config/layer.rst @@ -257,6 +257,11 @@ seq_concat .. autoclass:: paddle.v2.layer.seq_concat :noindex: +seq_slice +--------- +.. autoclass:: paddle.v2.layer.seq_slice + :noindex: + kmax_sequence_score ------------------- .. autoclass:: paddle.v2.layer.kmax_sequence_score @@ -362,6 +367,11 @@ trans .. autoclass:: paddle.v2.layer.trans :noindex: +scale_shift +----------- +.. autoclass:: paddle.v2.layer.scale_shift + :noindex: + Sampling Layers =============== @@ -409,9 +419,14 @@ multi_binary_label_cross_entropy_cost .. autoclass:: paddle.v2.layer.multi_binary_label_cross_entropy_cost :noindex: -huber_cost ----------- -.. autoclass:: paddle.v2.layer.huber_cost +huber_regression_cost +------------------------- +.. autoclass:: paddle.v2.layer.huber_regression_cost + :noindex: + +huber_classification_cost +------------------------- +.. autoclass:: paddle.v2.layer.huber_classification_cost :noindex: lambda_cost @@ -419,9 +434,9 @@ lambda_cost .. autoclass:: paddle.v2.layer.lambda_cost :noindex: -mse_cost +square_error_cost -------- -.. autoclass:: paddle.v2.layer.mse_cost +.. autoclass:: paddle.v2.layer.square_error_cost :noindex: rank_cost diff --git a/doc/design/cluster_train/README.md b/doc/design/cluster_train/README.md index 74961f80050c6b2723889b51416a2e8048174b00..177a5f5d54bd924fab34795219ce1f7b270c8e25 100644 --- a/doc/design/cluster_train/README.md +++ b/doc/design/cluster_train/README.md @@ -54,17 +54,18 @@ The life cycle of a single task is illustrated below: 1. When a new pass of training starts, all tasks will be placed in the todo queue. -1. The master server will dispatch few tasks to each trainer at a time, puts them in the pending queue and waits for completion. -1. The trainer will work on its tasks and tell the master server once a task is completed. The master server will dispatch a new task to that trainer. -1. If a task timeout. the master server will move it back to the todo queue. The timeout count will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, so it will be discarded. +1. Upon trainer requests for new task, the master server will dispatch a task from todo queue to it, put the task in the pending queue and wait for completion. +1. The trainer will work on its task and tell the master server once the task is completed and ask for new task. The master server will dispatch a new task to that trainer. +1. If a task fails for any reason in trainer, or takes longer than a specific period of time, the master server will move the task back to the todo queue. The timeout count for that task will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, then it will be discarded. 1. The master server will move completed task to the done queue. When the todo queue is empty, the master server will start a new pass by moving all tasks in the done queue to todo queue and reset the timeout counter of all tasks to zero. ### Trainer Process The trainer process will: -- Receive tasks from the master. -- Work on the tasks: calculate and upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers. +- Request tasks from the master. +- Work on the tasks +- Upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers. ### Parameter Server Process @@ -119,8 +120,8 @@ When the master is started by the Kubernetes, it executes the following steps at 1. Grabs a unique *master* lock in etcd, which prevents concurrent master instantiations. 1. Recovers the task queues from etcd if they already exist, otherwise, the master will create them. -1. Watches the trainer prefix keys `/trainer/` on etcd to find the live trainers. -1. Starts dispatching the tasks to the trainers, and updates task queue using an etcd transaction to ensure lock is held during the update. +1. Write its ip address to */master/addr* so that trainers can discover it. +1. Listens to trainers' request of task, dispatch one upon request, and updates task queue using an etcd transaction to ensure lock is held during the update. When the master server process is dead for any reason, Kubernetes will restart it. It will be online again with all states recovered from etcd in few minutes. @@ -128,13 +129,11 @@ When the master server process is dead for any reason, Kubernetes will restart i When the trainer is started by the Kubernetes, it executes the following steps at startup: -1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count. -1. Generates a unique ID, and sets key `/trainer/` with its contact address as value. The key will be deleted when the lease expires, so the master will be aware of the trainer being online and offline. -1. Waits for tasks from the master to start training. +1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count */ps_desired*. +1. Finds and watches */master/addr* to get master's address. +1. Requests for tasks from the master to start training. -If trainer's etcd lease expires, it will try set key `/trainer/` again so that the master server can discover the trainer again. - -When a trainer fails, Kuberentes would try to restart it. The recovered trainer would fetch tasks from the TODO queue and go on training. +When a trainer fails, Kuberentes would try to restart it. The recovered trainer would fetch tasks from master and go on training. ### Parameter Server Process diff --git a/doc/design/cluster_train/large_model_dist_train.md b/doc/design/cluster_train/large_model_dist_train.md new file mode 100644 index 0000000000000000000000000000000000000000..0c4b5bc24c854b7062d509249bea9c50d42bd5f1 --- /dev/null +++ b/doc/design/cluster_train/large_model_dist_train.md @@ -0,0 +1,101 @@ +# Alalysis of large model distributed training in Paddle + +***NOTE: This is only some note for how we implemeted this scheme in V1, not a new design.*** + +## What is it + +We often encounter cases that the embedding layer parameters(sparse) are so large that we can not store it in the trainer's memory when training. So we need to put them to several servers, and fetch them row by row instead of fetch all of the parameters. + +## How to use + +Specify command-line argument like `--loadsave_parameters_in_pserver=true --ports_num_for_sparse=1 --use_old_updater=1` when starting the paddle trainer. And also add something like `--ports_num_for_sparse=1 --pserver_num_threads=5` when starting pserver processes. + +Accrodingly, configure your embedding layers like: + +```python +SPARSE_REMOTE=True + +w1 = data_layer(name="w1", size=dict_size) +emb1 = embedding_layer(input=w1, size=32, param_attr=ParameterAttribute(sparse_update=SPARSE_REMOTE)) +w2 = data_layer(name="w2", size=dict_size) +emb2 = embedding_layer(input=w2, size=32, param_attr=ParameterAttribute(sparse_update=SPARSE_REMOTE)) +... +``` + +## Implementation details + +```c++ +enum MatType { + MAT_NORMAL, + MAT_NORMAL_SHARED, + MAT_VALUE_SHARED, + MAT_SPARSE_ROW_IDS, + MAT_SPARSE_ROW_AUTO_GROW, + MAT_CACHE_ROW, + MAT_SPARSE_ROW, + MAT_SPARSE_ROW_PREFETCH, + MAT_SPARSE_ROW_PREFETCH_FULL_SIZE, +}; +``` + +`MAT_SPARSE_ROW_PREFETCH` is what we use when configured to fetch only row of matrix when training. + +In `trainer_internal.cpp:L93 trainOneBatch`: + +```c++ + if (config_->getOptConfig().use_sparse_remote_updater()) { + REGISTER_TIMER("prefetch"); + gradientMachine_->prefetch(inArgs); + parameterUpdater_->getParametersRemote(); + } +``` + +When doing actual network forward and backward, at the beginning of each batch, the trainer will try to download one row of data from pserver. + +In `trainer/RemoteParameterUpdater.cpp`: `parameterUpdater_->getParametersRemote();`: + +```c++ +if (fullSize) { + ... +} else { +getParams = [&] { + parameterClient_->getParameterSparse( + /* recvParameterType= */ PARAMETER_VALUE, sendBackParameterType); +}; +applyL1 = [](Parameter& para, real decayRate) { + para.getMat(PARAMETER_VALUE)->applyL1(/*lr=*/1.0f, decayRate); +}; +} +``` + +Calling `parameterClient_->getParameterSparse` will do remote call to pserver's `getParameterSparse`: + +```c++ +void ParameterServer2::getParameterSparse(const SendParameterRequest& request, + std::vector& inputBuffers, + SendParameterResponse* response, + std::vector* outputBuffers) { + (void)inputBuffers; + auto& buffer = *readWriteBuffer_; + size_t numReals = 0; + for (const auto& block : request.blocks()) { + numReals += getParameterConfig(block).dims(1); + } + buffer.resize(numReals); + + VLOG(3) << "pserver: getParameterSparse, numReals=" << numReals; + + ReadLockGuard guard(parameterMutex_); + size_t offset = 0; + for (const auto& block : request.blocks()) { + size_t width = getParameterConfig(block).dims(1); + Buffer buf = {buffer.data() + offset, width}; + int type = request.send_back_parameter_type(); + sendBackParameterSparse(block, type, response, &buf, width, outputBuffers); + offset += width; + } +} +``` + +`getParameterConfig(block).dims(1)` returns the width of the current "parameter block"(a shard of parameter object), +then `getParameterSparse` remote call returns only one row of data to the client. diff --git a/doc/design/cluster_train/src/paddle-etcd.graffle b/doc/design/cluster_train/src/paddle-etcd.graffle index 56681ae5bbe11849116d621b066a6317e003e4ca..f973dc9b9dbf72e9bc31e2d32822916cd281f8d9 100644 Binary files a/doc/design/cluster_train/src/paddle-etcd.graffle and b/doc/design/cluster_train/src/paddle-etcd.graffle differ diff --git a/doc/design/cluster_train/src/paddle-etcd.png b/doc/design/cluster_train/src/paddle-etcd.png index 4f9c9762b3a8c089dd5e9b2c07cb9dfc78296a21..57981ceb4b94f0f7d6dfa63f3d28c0402bf9cc31 100644 Binary files a/doc/design/cluster_train/src/paddle-etcd.png and b/doc/design/cluster_train/src/paddle-etcd.png differ diff --git a/doc/design/functions_operators_layers.md b/doc/design/functions_operators_layers.md new file mode 100644 index 0000000000000000000000000000000000000000..d23ba56b5773a36d448a99e4abdebc1475ed789c --- /dev/null +++ b/doc/design/functions_operators_layers.md @@ -0,0 +1,100 @@ +# Design Doc: Functions, Operators, and Layers + +In a DL system, we can compose one or more fine grained operators into a coarse grained one. For example, the FC layer can be composed of a multiplication operator and an add operator. + +Historically, some fine grained operations are known as operators, and some coarse level ones are known as layers. But we need a well-defined separation. + +In general, operators are those very fine grained operations, e.g., mul and add. In the implementation, we can write them as C++ functions: + +```c++ +template T add(T x, T y) { return x + y; } +template T mul(T x, T y) { return x * y; } +``` + +Then we can wrap them into operators which are C++ classes and can be created from Python bindings by name. A C macro can do this. For example, the following macro invocation + +```c++ +#define MAKE_FUNCTION_OPERATOR(mul); +``` + +generates + +```c++ +template class mulOp : public OperatorBase {...}; +REGISTER_OP(mulOp, "mul"); +``` + +so that in Python we can create operator mul by: + +```python +X1 = Var() +X2 = Var() +Y = Var() +paddle.cpp.create_operator("mul", input=[X1, X2], output=Y) +``` + +Also, at the same time, we can compose a coarse level C++ operator class by composing functions `mul` and `add`: + +```c++ +template +class FCOp : public OperatorBase { + public: + void Run(...) { + add(mul(Input("X"), Input("W")), Input("b"); + } +}; +REGISTER_OP(FCOp, "fc"); +``` + +We need to support such composition in Python as well. To do so, we need a higher level Python wrapping of operator creation than `paddle.cpp.create_operator`. This higher level operator API should be compatible with the layer API. + +Let's explain using an example. Suppose that we are going to compose the FC using mul and add in Python, we'd like to have Python functions `mul` and `add` defined in module `operator`: + +```python +def operator.mul(X1, X2): + O = Var() + 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) + return O +``` + +Above code snippets are automatically generated. Given them, users can define + +```python +def layer.fc(X): + W = Var() + b = Var() + return operator.add(operator.mul(X, W), b) +``` + +If we don't have `operator.mul` and `operator.add`, the definiton of `layer.fc` would be complicated: + +```python +def layer.fc(X): + W = Var() + b = Var() + O1 = Var() + paddle.cpp.create_operator("mul", input=[X, W], output=O1) + O2 = Var() + paddle.cpp.create_operator("add", input=[O1, b], output=O2) + return O2 +``` + +We'd like to have Python bindings to operators in package `paddle.operator`, and Python compositions of operators in package `paddle.layer`. So we have the following concepts in above illustrative example: + + +| C++ functions/functors | mul | add | | | +|------------------------|--------------|--------------|-------------|----------| +| C++ operator class | mulOp | addOp | FCOp | | +| Python binding | operator.mul | operator.add | operator.fc | | +| Python function | | | | layer.fc | + + +This is how we differentiate layer and operators in PaddlePaddle: + +- those defined in C++ and have a lightweighted Python wrapper in module `operators` are operators; whereas +- those who don't have C++ implementations but a Python implementation that compose C++ operators are known as layers. diff --git a/doc/design/graph.md b/doc/design/graph.md new file mode 100644 index 0000000000000000000000000000000000000000..51b7f87638f8ddff752328a562fe0dd0fe56cfd1 --- /dev/null +++ b/doc/design/graph.md @@ -0,0 +1,70 @@ +# Design Doc: Computations as a Graph + +A primary goal of the refactorization of PaddlePaddle is a more flexible representation of deep learning computation, in particular, a graph of operators and variables, instead of sequences of layers as before. + +This document explains that the construction of a graph as three steps: + +- construct the forward part +- construct the backward part +- construct the optimization part + +## The Construction of a Graph + +Let us take the problem of image classification as a simple example. The application program that trains the model looks like: + +```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()) +``` + +### Forward Part + +The first four lines of above program build the forward part of the graph. + +![](images/graph_construction_example_forward_only.png) + +In particular, the first line `x = layer.data("images")` creates variable x and a Feed operator that copies a column from the minibatch to x. `y = layer.fc(x)` creates not only the FC operator and output variable y, but also two parameters, W and b, and the initialization operators. + +Initialization operators are kind of "run-once" operators -- the `Run` method increments a class data member counter so to run at most once. By doing so, a parameter wouldn't be initialized repeatedly, say, in every minibatch. + +In this example, all operators are created as `OpDesc` protobuf messages, and all variables are `VarDesc`. These protobuf messages are saved in a `BlockDesc` protobuf message. + +### Backward Part + +The fifth line `optimize(cost)` calls two functions, `ConstructBackwardGraph` and `ConstructOptimizationGraph`. + +`ConstructBackwardGraph` traverses the forward graph in the `BlockDesc` protobuf message and builds the backward part. + +![](images/graph_construction_example_forward_backward.png) + +According to the chain rule of gradient computation, `ConstructBackwardGraph` would + +1. create a gradient operator G for each operator F, +1. make all inputs, outputs, and outputs' gradient of F as inputs of G, +1. create gradients for all inputs of F, except for those who don't have gradients, like x and l, and +1. make all these gradients as outputs of G. + +### Optimization Part + +For each parameter, like W and b created by `layer.fc`, marked as double circles in above graphs, `ConstructOptimizationGraph` creates an optimization operator to apply its gradient. Here results in the complete graph: + +![](images/graph_construction_example_all.png) + +## 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. + +A Block keeps operators in an array `BlockDesc::ops` + +```protobuf +message BlockDesc { + repeated OpDesc ops = 1; + repeated VarDesc vars = 2; +} +``` + +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. diff --git a/doc/design/if_else_op.md b/doc/design/if_else_op.md new file mode 100644 index 0000000000000000000000000000000000000000..7370c2a24fa644a64e738f202bac9b9209642e08 --- /dev/null +++ b/doc/design/if_else_op.md @@ -0,0 +1,59 @@ +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 M (M<=N) instances, each corresponds to a true element in `cond`. + +```python +import paddle as pd + +x = var() +y = var() +cond = var() + +b = pd.create_ifop(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)) + +out = b(cond) +``` + +If we want the output still has N instances, we can use IfElseOp with a default value, whose minibatch size must be N: + +```python +import paddle as pd + +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) +``` + +If only true_block is set in an IfElseOp, 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)) + +out = b(cond) +``` +where default_value is a list of vars for `cond` == False. diff --git a/doc/design/images/graph_construction_example.bash b/doc/design/images/graph_construction_example.bash new file mode 100755 index 0000000000000000000000000000000000000000..35e6997abd17588e17a82d448918fc1b3bd7220e --- /dev/null +++ b/doc/design/images/graph_construction_example.bash @@ -0,0 +1,11 @@ +cat ./graph_construction_example.dot | \ + sed 's/color=red/color=red, style=invis/g' | \ + sed 's/color=green/color=green, style=invis/g' | \ + dot -Tpng > graph_construction_example_forward_only.png + +cat ./graph_construction_example.dot | \ + sed 's/color=green/color=green, style=invis/g' | \ + dot -Tpng > graph_construction_example_forward_backward.png + +cat ./graph_construction_example.dot | \ + dot -Tpng > graph_construction_example_all.png diff --git a/doc/design/images/graph_construction_example.dot b/doc/design/images/graph_construction_example.dot new file mode 100644 index 0000000000000000000000000000000000000000..8d1b673abf6b78c851676fa379dc850c4818f0e5 --- /dev/null +++ b/doc/design/images/graph_construction_example.dot @@ -0,0 +1,69 @@ +digraph ImageClassificationGraph { + ///////// The forward part ///////// + FeedX [label="Feed", color=blue, shape=box]; + FeedY [label="Feed", color=blue, shape=box]; + InitW [label="Init", color=blue, shape=diamond]; + Initb [label="Init", color=blue, shape=diamond]; + FC [label="FC", color=blue, shape=box]; + MSE [label="MSE", color=blue, shape=box]; + + x [label="x", color=blue, shape=oval]; + l [label="l", color=blue, shape=oval]; + y [label="y", color=blue, shape=oval]; + W [label="W", color=blue, shape=doublecircle]; + b [label="b", color=blue, shape=doublecircle]; + cost [label="cost", color=blue, shape=oval]; + + FeedX -> x -> FC -> y -> MSE -> cost [color=blue]; + FeedY -> l [color=blue]; + InitW -> W [color=blue]; + Initb -> b [color=blue]; + W -> FC [color=blue]; + b -> FC [color=blue]; + l -> MSE [color=blue]; + + ////////// The backward part ///////// + MSE_Grad [label="MSE_grad", color=red, shape=box]; + FC_Grad [label="FC_grad", color=red, shape=box]; + + d_cost [label="d cost", color=red, shape=oval]; + d_y [label="d y", color=red, shape=oval]; + d_b [label="d b", color=red, shape=oval]; + d_W [label="d W", color=red, shape=oval]; + + cost -> MSE_Grad [color=red]; + d_cost -> MSE_Grad [color=red]; + x -> MSE_Grad [color=red]; + l -> MSE_Grad [color=red]; + y -> MSE_Grad -> d_y [color=red]; + + x -> FC_Grad [color=red]; + y -> FC_Grad [color=red]; + d_y -> FC_Grad [color=red]; + W -> FC_Grad -> d_W [color=red]; + b -> FC_Grad -> d_b [color=red]; + + ////////// The optimizaiton part ////////// + + OPT_W [label="SGD", color=green, shape=box]; + OPT_b [label="SGD", color=green, shape=box]; + + W -> OPT_W [color=green]; + b -> OPT_b [color=green]; + d_W -> OPT_W -> W [color=green]; + d_b -> OPT_b -> b [color=green]; + + ////////// Groupings ////////// + + subgraph clusterMSE { + style=invis; + MSE; + MSE_Grad; + } + + subgraph clusterFC { + style=invis; + FC; + FC_Grad; + } +} diff --git a/doc/design/images/graph_construction_example_all.png b/doc/design/images/graph_construction_example_all.png new file mode 100644 index 0000000000000000000000000000000000000000..181187503472d15779b87284105841168b3945c4 Binary files /dev/null and b/doc/design/images/graph_construction_example_all.png differ diff --git a/doc/design/images/graph_construction_example_forward_backward.png b/doc/design/images/graph_construction_example_forward_backward.png new file mode 100644 index 0000000000000000000000000000000000000000..3049a9315fd616464dec54e33064cb75598ca536 Binary files /dev/null and b/doc/design/images/graph_construction_example_forward_backward.png differ diff --git a/doc/design/images/graph_construction_example_forward_only.png b/doc/design/images/graph_construction_example_forward_only.png new file mode 100644 index 0000000000000000000000000000000000000000..25d19088cbf0b5f68cf734f2ff21eba8af4a2860 Binary files /dev/null and b/doc/design/images/graph_construction_example_forward_only.png differ diff --git a/doc/design/mkldnn/README.MD b/doc/design/mkldnn/README.MD index e956994431fbb43438c56dcd96ad8313cf516090..fe8da907d9d45a2164031430ac5b7a3d5523967a 100644 --- a/doc/design/mkldnn/README.MD +++ b/doc/design/mkldnn/README.MD @@ -101,6 +101,7 @@ if use_mkldnn 5. 在**Argument**里添加两个`MkldnnMatrixPtr`，取名为`mkldnnValue`和`mkldnnGrad`，用于存放`MkldnnLayer`会用到的memory buffer。 并且添加函数cvt(会修改为一个更加合适的函数名)，用于处理"CPU device"和"MKL-DNN device"之间memory的相互转化。 6. 在父类`Layer`中的`getOutput`函数中添加一段逻辑，用于判断`deviceId`，并针对device在MKL-DNN和CPU之间不统一的情况，做一个前期转换。 也就是调用`Argument`的cvt函数把output统一到需要的device上。 7. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。 +8. 关于MKLDNN参数的保存。由于MKLDNN参数的格式与PaddlePaddle原有的格式存在不一样的情况，所以需要在保存参数时同时保存该格式信息。目前准备扩展[Header](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/parameter/Parameter.h#L247)里面的`int32_t version`。这个值不管是在v1还是在v2里面，一直保存的是0，所以可以充分利用这个信息，定义一个枚举处理所有MKLDNN的参数格式，从而`MKLDNNLayer`就可以从输入的参数中获取需要的格式信息。 ## References diff --git a/doc/design/ops/dist_train.md b/doc/design/ops/dist_train.md new file mode 100644 index 0000000000000000000000000000000000000000..fa3c5d7990213cf2b0d236e66e592dd2699da876 --- /dev/null +++ b/doc/design/ops/dist_train.md @@ -0,0 +1,106 @@ +# Design Doc: Operation Graph Based Parameter Server + +## Abstract + +We propose an approach to implement the parameter server. In this +approach, there is no fundamental difference between the trainer and +the parameter server: they both run subgraphs, but subgraphs of +different purposes. + +## Background + +The previous implementations of the parameter server does not run a +subgraph. parameter initialization, optimizer computation, network +communication and checkpointing are implemented twice on both the +trainer and the parameter server. + +It would be great if we can write code once and use them on both the +trainer and the parameter server: reduces code duplication and +improves extensibility. Given that after the current refactor, we are +representing everything as a computing graph on the +trainer. Representing everything as a computing graph on the parameter +server becomes a natural extension. + +## Design + +### Graph Converter + +The *graph converter* converts the user-defined operation (OP) graph +into subgraphs to be scheduled on different nodes with the following +steps: + +1. OP placement: the OPs will be placed on different nodes according + to heuristic that minimizes estimated total computation + time. Currently we will use a simple heuristic that puts parameter + varable on parameter server workers and everything else on trainer + workers. + +1. Add communication OPs to enable the communication between nodes. + +We will need these OPs: *Send*, *Recv*, *Enqueue*, *Dequeue*. + +Below is an example of converting the user defined graph to the +subgraphs for the trainer and the parameter server: + + + +After converting: + + + +1. The parameter variable W and it's optimizer subgraph are placed on the parameter server. +1. Operators are added to the subgraphs. + - *Send* sends data to the connected *Recv* operator. The + scheduler on the receive node will only schedule *Recv* operator + to run when the *Send* operator has ran (the *Send* OP will mark + the *Recv* OP runnable automatically). + - *Enueue* enqueues the input variable, it can block until space + become available in the queue. + - *Dequeue* outputs configurable numbers of tensors from the + queue. It will block until the queue have the required number of + tensors. + + +### Benefits + +- Model parallelism become easier to implement: it's an extension to + the trainer - parameter server approach. we already have the + communication OPs, but need to extend the graph converter's + placement functionality. + +- User-defined optimizer is easier to add - user can now express it as + a subgraph. + +- No more duplication logic inside the trainer and the parameter + server mentioned in the background section. + +### Challenges + +- It might be hard for the graph converter to cut a general graph + (without any hint for which subgraph is the optimizer). We may need + to label which subgraph inside the OP graph is the optimizer. + +- It's important to balance the parameter shards of on multiple + parameter server. If a single parameter is very big (some + word-embedding, fully connected, softmax layer), we need to + automatically partition the single parameter onto different + parameter servers when possible (only element-wise optimizer depends + on the parameter variable). + +### Discussion + +- In the "Aync SGD" figure, the "W" variable on the parameter server + could be read and wrote concurrently, what is our locking strategy? + E.g., each variable have a lock cpp method to be invoked by every + OP, or, have a lock OP. + +- Can the Enqueue OP be implemented under our current tensor design + (puts the input tensor into the queue tensor)? + +- *Dequeue* OP will have variable numbers of output (depends on the + `min_count` attribute), does our current design support it? (similar + question for the *Add* OP) + + +### References: +[1] [TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems](https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45166.pdf) diff --git a/doc/design/ops/src/dist-graph.graffle b/doc/design/ops/src/dist-graph.graffle new file mode 100644 index 0000000000000000000000000000000000000000..941399c6ced8d5f65b6c595522b770c88259df4b Binary files /dev/null and b/doc/design/ops/src/dist-graph.graffle differ diff --git a/doc/design/ops/src/dist-graph.png b/doc/design/ops/src/dist-graph.png new file mode 100644 index 0000000000000000000000000000000000000000..3546b09f1c2ee3e4f60f519d5e47f823f08051a7 Binary files /dev/null and b/doc/design/ops/src/dist-graph.png differ diff --git a/doc/design/ops/src/local-graph.graffle b/doc/design/ops/src/local-graph.graffle new file mode 100644 index 0000000000000000000000000000000000000000..19e509bd9af3c1e9a3f5e0f16ddd281457a339c5 Binary files /dev/null and b/doc/design/ops/src/local-graph.graffle differ diff --git a/doc/design/ops/src/local-graph.png b/doc/design/ops/src/local-graph.png new file mode 100644 index 0000000000000000000000000000000000000000..ada51200f793a9bb18911e7d63cfdb3244b967d7 Binary files /dev/null and b/doc/design/ops/src/local-graph.png differ diff --git a/doc/design/simple_op_design.md b/doc/design/simple_op_design.md index 5e07c29c56d21728599195d420d3222213d77e7c..fded4a68612396a262121a5a886a8ae573dfa662 100644 --- a/doc/design/simple_op_design.md +++ b/doc/design/simple_op_design.md @@ -147,7 +147,7 @@ class CosineOp { struct CosineOpProtoMaker : public OpProtoMaker { CosineOpProtoMaker(OpProto* proto) : OpProtoMaker(proto) { AddInput("input", "input of cosine op"); - AddAttr("scale", "scale of cosine op", float).Default(1.0).LargerThan(0.0); + AddAttr("scale", "scale of cosine op", float).Default(1.0).GreaterThan(0.0); AddType("cos"); AddComment("This is cos op"); } diff --git a/doc/design/var_desc.md b/doc/design/var_desc.md new file mode 100644 index 0000000000000000000000000000000000000000..86a95c10d5729704f86c285c9fe92db0cf2158be --- /dev/null +++ b/doc/design/var_desc.md @@ -0,0 +1,124 @@ +## 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 + +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. + +The computation graph is constructed by Data Node and Operation Node. The concept to represent them is in the table below. + +| |compile time|runtime| +|---|---|---| +|Data|VarDesc(proto)|Variable(cpp)| +|Operation|OpDesc(proto)|Operator(cpp)| + + +## Definition of VarDesc + +A VarDesc should have a name and value, in PaddlePaddle, the value will always be a tensor. Since we use LoDTensor most of the time. We add a LoDTesnorDesc to represent it. + +```proto +message VarDesc { + required string name = 1; + optional LoDTensorDesc lod_tensor = 2; +} +``` + +## Definition of LodTensorDesc + +```proto +enum DataType { + BOOL = 0; + INT16 = 1; + INT32 = 2; + INT64 = 3; + FP16 = 4; + FP32 = 5; + FP64 = 6; +} + +message LoDTensorDesc { + required DataType data_type = 1; + repeated int32 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480] + optional int32 lod_level = 3 [default=0]; +} +``` + +## Definition of Variable in Python + +In Python API, layer will take Variable as Input, and return Variable as Output. There should be a class `Variable` in python to help create and manage Variable. + +```python +image = Variable(dims=[-1, 640, 480]) +# fc1 and fc2 are both Variable +fc1 = layer.fc(input=image, output_size=10) +fc2 = layer.fc(input=fc1, output_size=20) +``` +### what should class `Variable` Have +1. `name`.a name of string type is used to mark the value of the Variable. +1. `initializer`. Since our Tensor does not have value. we will always use some Operator to fullfill it when run. So we should have a initialize method to help add the init operator. +1. `operator`. Variable should record which operator produce itself. The reaon is: + - we use pd.eval(targets=[var1, var2]) to run the related ops to get the value of var1 and var2. var.op is used to trace the dependency of the current variable. + +In PaddlePaddle, we use Block to describe Computation Graph, so in the code we will use Block but not Graph. + +```python +import VarDesc +import LoDTensorDesc +import framework + +def AddInitialOperator(variable, initializer): + # add an initialize Operator to block to init this Variable + +class Variable(object): + def __init__(self, name, dims, type, initializer): + self._block = get_default_block() + self._name = name + self.op = None + + tensor_desc = LoDTensorDesc(data_type=type, dims=dims) + _var_desc = VarDesc(name=name, lod_tensor=tensor_desc) + self._var = framework.CreateVar(_var_desc) + self._block.add_var(self) + + # add initial op according to initializer + if initializer is not None: + AddInitialOperator(self, initializer) + + def dims(self): + return self._var.dims() + + def data_type(self): + return self._var.data_type() + + def to_proto(self): + pass +``` + +Then we can use this Variable to create a fc layer in Python. + +```python +import paddle as pd + +def flatten_size(X, num_flatten_dims): + prod = 1 # of last num_flatten_dims + for i in xrange(num_flatten_dims): + prod = prod * X.dims[-i-1] + return prod + +def layer.fc(X, output_size, num_flatten_dims): + W = Variable(pd.random_uniform(), type=FP32, dims=[flatten_size(X, num_flatten_dims), output_size]) + b = Variable(pd.random_uniform(), type=FP32, dims=[output_size]) + out = Variable(type=FP32) + y = operator.fc(X, W, b, output=out) # fc will put fc op input into out + pd.InferShape(y) + return out + +x = Variable(dims=[-1, 640, 480]) +y = layer.fc(x, output_size=100) +z = layer.fc(y, output_size=200) + +paddle.eval(targets=[z], ...) +print(z) +``` diff --git a/doc/getstarted/basic_usage/index_cn.rst b/doc/getstarted/basic_usage/index_cn.rst index 428f58830e0b10c024f31238b7404c6df193eecd..b473944fc7fb89d3e0a0b330933f2226734bb5bd 100644 --- a/doc/getstarted/basic_usage/index_cn.rst +++ b/doc/getstarted/basic_usage/index_cn.rst @@ -55,7 +55,7 @@ PaddlePaddle是源于百度的一个深度学习平台。这份简短的介绍 # 线性计算网络层: ȳ = wx + b ȳ = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b')) # 计算误差函数，即 ȳ 和真实 y 之间的距离 - cost = mse_cost(input= ȳ, label=y) + cost = square_error_cost(input= ȳ, label=y) outputs(cost) @@ -69,7 +69,7 @@ PaddlePaddle是源于百度的一个深度学习平台。这份简短的介绍 - **数据层**：数据层 `data_layer` 是神经网络的入口，它读入数据并将它们传输到接下来的网络层。这里数据层有两个，分别对应于变量 `x` 和 `y`。 - **全连接层**：全连接层 `fc_layer` 是基础的计算单元，这里利用它建模变量之间的线性关系。计算单元是神经网络的核心，PaddlePaddle支持大量的计算单元和任意深度的网络连接，从而可以拟合任意的函数来学习复杂的数据关系。 - - **回归误差代价层**：回归误差代价层 `mse_cost` 是众多误差代价函数层的一种，它们在训练过程作为网络的出口，用来计算模型的误差，是模型参数优化的目标函数。 + - **回归误差代价层**：回归误差代价层 `square_error_cost` 是众多误差代价函数层的一种，它们在训练过程作为网络的出口，用来计算模型的误差，是模型参数优化的目标函数。 定义了网络结构并保存为 `trainer_config.py` 之后，运行以下训练命令： diff --git a/doc/getstarted/basic_usage/index_en.rst b/doc/getstarted/basic_usage/index_en.rst index 6775da20c2f51000f305b095d40abd27b8fa6c0e..2cc438ebbe0f97345d25354b93b4ebbd43502415 100644 --- a/doc/getstarted/basic_usage/index_en.rst +++ b/doc/getstarted/basic_usage/index_en.rst @@ -49,7 +49,7 @@ To recover this relationship between ``X`` and ``Y``, we use a neural network wi x = data_layer(name='x', size=1) y = data_layer(name='y', size=1) y_predict = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b')) - cost = mse_cost(input=y_predict, label=y) + cost = square_error_cost(input=y_predict, label=y) outputs(cost) Some of the most fundamental usages of PaddlePaddle are demonstrated: diff --git a/doc/getstarted/build_and_install/build_from_source_en.md b/doc/getstarted/build_and_install/build_from_source_en.md index c0608ede8e57b224dae4b3d510d704a8b0918b53..2f1461489495618718d5abaeab9cbeda9b93700f 100644 --- a/doc/getstarted/build_and_install/build_from_source_en.md +++ b/doc/getstarted/build_and_install/build_from_source_en.md @@ -68,7 +68,7 @@ As a simple example, consider the following: 1. **BLAS Dependencies(optional)** - CMake will search BLAS libraries from system. If not found, OpenBLAS will be downloaded, built and installed automatically. + CMake will search BLAS libraries from the system. If not found, OpenBLAS will be downloaded, built and installed automatically. To utilize preinstalled BLAS， you can simply specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`. ```bash @@ -131,9 +131,9 @@ As a simple example, consider the following: To build GPU version, you will need the following installed: 1. a CUDA-capable GPU - 2. A supported version of Linux with a gcc compiler and toolchain + 2. A supported version of Linux with a GCC compiler and toolchain 3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads) - 4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn) + 4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn) The CUDA development environment relies on tight integration with the host development environment, including the host compiler and C runtime libraries, and is therefore only supported on @@ -172,6 +172,7 @@ export PATH=/bin:$PATH # install PaddlePaddle Python modules. sudo pip install /opt/paddle/share/wheels/*.whl ``` + ## Build on Centos 7 ### Install Dependencies @@ -192,9 +193,9 @@ sudo pip install /opt/paddle/share/wheels/*.whl To build GPU version, you will need the following installed: 1. a CUDA-capable GPU - 2. A supported version of Linux with a gcc compiler and toolchain + 2. A supported version of Linux with a GCC compiler and toolchain 3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads) - 4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn) + 4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn) The CUDA development environment relies on tight integration with the host development environment, including the host compiler and C runtime libraries, and is therefore only supported on @@ -222,7 +223,7 @@ mkdir build && cd build ``` Finally, you can build and install PaddlePaddle: - + ```bash # you can add build option here, such as: cmake3 .. -DCMAKE_INSTALL_PREFIX= diff --git a/doc/getstarted/build_and_install/index_cn.rst b/doc/getstarted/build_and_install/index_cn.rst index a24df6c518fad84a48061ecb34ee46cb312a4995..dd9923697ab85825557aa89a08870bece7c76673 100644 --- a/doc/getstarted/build_and_install/index_cn.rst +++ b/doc/getstarted/build_and_install/index_cn.rst @@ -6,14 +6,12 @@ 安装流程 ++++++++ -PaddlePaddle提供数个预编译的二进制来进行安装，包括Docker镜像，ubuntu的deb安装包等。我们推荐使用Docker镜像来部署环境，同时欢迎贡献更多的安装包。 +PaddlePaddle提供Docker镜像来部署环境。 .. toctree:: :maxdepth: 1 docker_install_cn.rst - ubuntu_install_cn.rst - 编译流程 diff --git a/doc/getstarted/build_and_install/index_en.rst b/doc/getstarted/build_and_install/index_en.rst index 1bfd4f75c0b9b82d61d28a30f03181f7be159f24..8a53588e0439df8f4d5fd529b7a20262c67d4e58 100644 --- a/doc/getstarted/build_and_install/index_en.rst +++ b/doc/getstarted/build_and_install/index_en.rst @@ -8,14 +8,13 @@ Install PaddlePaddle :maxdepth: 1 docker_install_en.rst - ubuntu_install_en.rst Build from Source ----------------- .. warning:: - Please use :code:`deb` package or :code:`docker` image to install paddle. The building guide is used for hacking or contributing PaddlePaddle source code. + Please use :code:`docker` image to install paddle. The building guide is used for hacking or contributing PaddlePaddle source code. .. toctree:: :maxdepth: 1 diff --git a/doc/getstarted/build_and_install/ubuntu_install_cn.rst b/doc/getstarted/build_and_install/ubuntu_install_cn.rst deleted file mode 100644 index 9e39ccb00f5d5655c30148900a3d76a22aacfc01..0000000000000000000000000000000000000000 --- a/doc/getstarted/build_and_install/ubuntu_install_cn.rst +++ /dev/null @@ -1,71 +0,0 @@ -Ubuntu部署PaddlePaddle -=================================== - -PaddlePaddle提供了ubuntu 14.04 deb安装包。 - -安装 ------- - -安装包的下载地址是\: https://github.com/PaddlePaddle/Paddle/releases - -它包含四个版本\: - -* cpu版本: 支持主流x86处理器平台, 使用了avx指令集。 - -* cpu-noavx版本：支持主流x86处理器平台，没有使用avx指令集。 - -* gpu版本：支持主流x86处理器平台，支持nvidia cuda平台，使用了avx指令集。 - -* gpu-noavx版本：支持主流x86处理器平台，支持nvidia cuda平台，没有使用avx指令集。 - -下载完相关安装包后，执行: - -.. code-block:: shell - - sudo apt-get install gdebi - gdebi paddle-*-cpu.deb - -或者: - -.. code-block:: shell - - dpkg -i paddle-*-cpu.deb - apt-get install -f - - -在 :code:`dpkg -i` 的时候如果报一些依赖未找到的错误是正常的， -在 :code:`apt-get install -f` 里会继续安装 PaddlePaddle。 - -安装完成后，可以使用命令 :code:`paddle version` 查看安装后的paddle 版本: - -.. code-block:: shell - - PaddlePaddle 0.8.0b1, compiled with - with_avx: ON - with_gpu: OFF - with_double: OFF - with_python: ON - with_rdma: OFF - with_timer: OFF - with_predict_sdk: - - -可能遇到的问题 --------------- - -libcudart.so/libcudnn.so找不到 -++++++++++++++++++++++++++++++ - -安装完成后，运行 :code:`paddle train` 报错\: - -.. code-block:: shell - - 0831 12:36:04.151525 1085 hl_dso_loader.cc:70] Check failed: nullptr != *dso_handle For Gpu version of PaddlePaddle, it couldn't find CUDA library: libcudart.so Please make sure you already specify its path.Note: for training data on Cpu using Gpu version of PaddlePaddle,you must specify libcudart.so via LD_LIBRARY_PATH. - -原因是未设置cuda运行时环境变量。 如果使用GPU版本的PaddlePaddle，请安装CUDA 7.5 和CUDNN 5到本地环境中，并设置： - -.. code-block:: shell - - export LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda/lib:$LD_LIBRARY_PATH - export PATH=/usr/local/cuda/bin:$PATH - diff --git a/doc/getstarted/build_and_install/ubuntu_install_en.rst b/doc/getstarted/build_and_install/ubuntu_install_en.rst deleted file mode 100644 index ea8042085bf458be96e71017d229d88ad867695b..0000000000000000000000000000000000000000 --- a/doc/getstarted/build_and_install/ubuntu_install_en.rst +++ /dev/null @@ -1,25 +0,0 @@ -Debian Package installation guide -================================= - -PaddlePaddle supports :code:`deb` pacakge. The installation of this :code:`deb` package is tested in ubuntu 14.04, but it should be support other debian based linux, too. - -There are four versions of debian package, :code:`cpu`, :code:`gpu`, :code:`cpu-noavx`, :code:`gpu-noavx`. And :code:`noavx` version is used to support CPU which does not contain :code:`AVX` instructions. The download url of :code:`deb` package is \: https://github.com/baidu/Paddle/releases/ - - -After downloading PaddlePaddle deb packages, you can use :code:`gdebi` install. - -.. code-block:: bash - - gdebi paddle-*.deb - -If :code:`gdebi` is not installed, you can use :code:`sudo apt-get install gdebi` to install it. - -Or you can use following commands to install PaddlePaddle. - -.. code-block:: bash - - dpkg -i paddle-*.deb - apt-get install -f - -And if you use GPU version deb package, you need to install CUDA toolkit and cuDNN, and set related environment variables(such as LD_LIBRARY_PATH) first. It is normal when `dpkg -i` get errors. `apt-get install -f` will continue install paddle, and install dependences. - diff --git a/doc/getstarted/concepts/src/train.py b/doc/getstarted/concepts/src/train.py index 7e604f23de38543a00f305d508af0791193f78ba..8aceb23406a476f08639cc6223cdf730b728a705 100644 --- a/doc/getstarted/concepts/src/train.py +++ b/doc/getstarted/concepts/src/train.py @@ -8,7 +8,7 @@ paddle.init(use_gpu=False) x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(2)) y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear()) y = paddle.layer.data(name='y', type=paddle.data_type.dense_vector(1)) -cost = paddle.layer.mse_cost(input=y_predict, label=y) +cost = paddle.layer.square_error_cost(input=y_predict, label=y) # create parameters parameters = paddle.parameters.create(cost) diff --git a/doc/getstarted/concepts/use_concepts_cn.rst b/doc/getstarted/concepts/use_concepts_cn.rst index f15b11bd780402a3ec1755900e8c648f5d2a7bc5..c243083794bb3c4659242de99b3b2715af9d7c24 100644 --- a/doc/getstarted/concepts/use_concepts_cn.rst +++ b/doc/getstarted/concepts/use_concepts_cn.rst @@ -81,9 +81,9 @@ PaddlePaddle支持不同类型的输入数据，主要包括四种类型，和 .. code-block:: bash y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear()) - cost = paddle.layer.mse_cost(input=y_predict, label=y) + cost = paddle.layer.square_error_cost(input=y_predict, label=y) -其中，x与y为之前描述的输入层；而y_predict是接收x作为输入，接上一个全连接层；cost接收y_predict与y作为输入，接上均方误差层。 +其中，x与y为之前描述的输入层；而y_predict是接收x作为输入，接上一个全连接层；cost接收y_predict与y作为输入，接上平方误差层。 最后一层cost中记录了神经网络的所有拓扑结构，通过组合不同的layer，我们即可完成神经网络的搭建。 @@ -147,4 +147,4 @@ PaddlePaddle支持不同类型的输入数据，主要包括四种类型，和 .. literalinclude:: src/train.py :linenos: -有关线性回归的实际应用，可以参考PaddlePaddle book的 `第一章节 `_。 \ No newline at end of file +有关线性回归的实际应用，可以参考PaddlePaddle book的 `第一章节 `_。 diff --git a/doc/howto/dev/build_cn.md b/doc/howto/dev/build_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..0b911f7b7509da4a147c65954acb7e7c38f489da --- /dev/null +++ b/doc/howto/dev/build_cn.md @@ -0,0 +1,124 @@ +# 编译PaddlePaddle和运行单元测试 + +## 需要的软硬件 + +为了开发PaddlePaddle，我们需要 + +1. 一台电脑，可以装的是 Linux, BSD, Windows 或者 MacOS 操作系统，以及 +1. Docker。 + +不需要依赖其他任何软件了。即便是 Python 和 GCC 都不需要，因为我们会把所有编译工具都安装进一个 Docker image 里。 + +## 总体流程 + +1. 获取源码 + + ```bash + git clone https://github.com/paddlepaddle/paddle + ``` + +2. 安装开发工具到 Docker image 里 + + ```bash + cd paddle; docker build -t paddle:dev . + ``` + + 请注意这个命令结尾处的 `.`；它表示 `docker build` 应该读取当前目录下的 [`Dockerfile`文件](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile)，按照其内容创建一个名为 `paddle:dev` 的 Docker image，并且把各种开发工具安装进去。 + +3. 编译 + + 以下命令启动一个 Docker container 来执行 `paddle:dev` 这个 Docker image，同时把当前目录（源码树根目录）映射为 container 里的 `/paddle` 目录，并且运行 `Dockerfile` 描述的默认入口程序 [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `cmake` 和 `make` 来编译 `/paddle` 里的源码，结果输出到 `/paddle/build`，也就是本地的源码树根目录里的 `build` 子目录。 + + ```bash + docker run --rm -v $PWD:/paddle paddle:dev + ``` + + 上述命令编译出一个 CUDA-enabled 版本。如果我们只需要编译一个只支持 CPU 的版本，可以用 + + ```bash + docker run --rm -e WITH_GPU=OFF -v $PWD:/paddle paddle:dev + ``` + +4. 运行单元测试 + + 用本机的第一个 GPU 来运行包括 GPU 单元测试在内的所有单元测试： + + ```bash + NV_GPU=0 nvidia-docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest" + ``` + + 如果编译的时候我们用了 `WITH_GPU=OFF` 选项，那么编译过程只会产生 CPU-based 单元测试，那么我们也就不需要 nvidia-docker 来运行单元测试了。我们只需要： + + ```bash + docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest" + ``` + + 有时候我们只想运行一个特定的单元测试，比如 `memory_test`，我们可以 + + ```bash + nvidia-docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest -V -R memory_test" + ``` + +5. 清理 + + 有时候我们会希望清理掉已经下载的第三方依赖以及已经编译的二进制文件。此时只需要： + + ```bash + rm -rf build + ``` + +## 为什么要 Docker 呀？ + +- 什么是 Docker? + + 如果您没有听说 Docker，可以把它想象为一个类似 virtualenv 的系统，但是虚拟的不仅仅是 Python 的运行环境。 + +- Docker 还是虚拟机？ + + 有人用虚拟机来类比 Docker。需要强调的是：Docker 不会虚拟任何硬件，Docker container 里运行的编译工具实际上都是在本机的 CPU 和操作系统上直接运行的，性能和把编译工具安装在本机运行一样。 + +- 为什么用 Docker? + + 把工具和配置都安装在一个 Docker image 里可以标准化编译环境。这样如果遇到问题，其他人可以复现问题以便帮助。 + + 另外，对于习惯使用Windows和MacOS的开发者来说，使用Docker就不用配置交叉编译环境了。 + +- 我可以选择不用Docker吗？ + + 当然可以。大家可以用把开发工具安装进入 Docker image 一样的方式，把这些工具安装到本机。这篇文档介绍基于 Docker 的开发流程，是因为这个流程比其他方法都更简便。 + +- 学习 Docker 有多难？ + + 理解 Docker 并不难，大概花十分钟看一下[这篇文章](https://zhuanlan.zhihu.com/p/19902938)。这可以帮您省掉花一小时安装和配置各种开发工具，以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。 + +- 我可以用 IDE 吗？ + + 当然可以，因为源码就在本机上。IDE 默认调用 make 之类的程序来编译源码，我们只需要配置 IDE 来调用 Docker 命令编译源码即可。 + + 很多 PaddlePaddle 开发者使用 Emacs。他们在自己的 `~/.emacs` 配置文件里加两行 + + ```emacs + (global-set-key "\C-cc" 'compile) + (setq compile-command + "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev") + ``` + + 就可以按 `Ctrl-C` 和 `c` 键来启动编译了。 + +- 可以并行编译吗？ + + 是的。我们的 Docker image 运行一个 [Bash 脚本](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。 + +## 可能碰到的问题 + +- Docker 需要 sudo + + 如果用自己的电脑开发，自然也就有管理员权限（sudo）了。如果用公用的电脑开发，需要请管理员安装和配置好 Docker。此外，PaddlePaddle 项目在努力开始支持其他不需要 sudo 的集装箱技术，比如 rkt。 + +- 在 Windows/MacOS 上编译很慢 + + Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存，以保证编译高效。具体做法请参考[这个issue](https://github.com/PaddlePaddle/Paddle/issues/627)。 + +- 磁盘不够 + + 本文中的例子里，`docker run` 命令里都用了 `--rm` 参数，这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 `docker ps -a` 命令看到停止后但是没有删除的 containers。`docker build` 命令有时候会产生一些中间结果，是没有名字的 images，也会占用磁盘。可以参考[这篇文章](https://zaiste.net/posts/removing_docker_containers/)来清理这些内容。 diff --git a/doc/howto/dev/build_en.md b/doc/howto/dev/build_en.md new file mode 100644 index 0000000000000000000000000000000000000000..d0048e3714a5861a503736879d6c0870e5906c95 --- /dev/null +++ b/doc/howto/dev/build_en.md @@ -0,0 +1,124 @@ +# Build PaddlePaddle from Source Code and Run Unit Test + +## What Developers Need + +To contribute to PaddlePaddle, you need + +1. A computer -- Linux, BSD, Windows, MacOS, and +1. Docker. + +Nothing else. Not even Python and GCC, because you can install all build tools into a Docker image. We run all the tools by running this image. + +## General Process + +1. Retrieve source code. + + ```bash + git clone https://github.com/paddlepaddle/paddle + ``` + +2. Install build tools into a Docker image. + + ```bash + cd paddle; docker build -t paddle:dev . + ``` + + Please be aware of the `.` at the end of the command, which refers to the [`./Dockerfile` file](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile). `docker build` follows instructions in this file to create a Docker image named `paddle:dev`, and installs building tools into it. + +3. Build from source. + + This following command starts a Docker container that executes the Docker image `paddle:dev`, mapping the current directory to `/paddle/` in the container, and runs the default entry-point [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh) as specified in the Dockefile. `build.sh` invokes `cmake` and `make` to build PaddlePaddle source code, which had been mapped to `/paddle`, and writes outputs to `/paddle/build`, which maps to `build` in the current source directory on the computer. + + ```bash + docker run -v $PWD:/paddle paddle:dev + ``` + + Above command builds a CUDA-enabled version. If we want to build a CPU-only version, we can type + + ```bash + docker run -e WITH_GPU=OFF -v $PWD:/paddle paddle:dev + ``` + +4. Run unit tests. + + To run all unit tests using the first GPU of a node: + + ```bash + NV_GPU=0 nvidia-docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest" + ``` + + If we used `WITH_GPU=OFF` at build time, it generates only CPU-based unit tests, and we don't need nvidia-docker to run them. We can just run + + ```bash + docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest" + ``` + + Sometimes we want to run a specific unit test, say `memory_test`, we can run + + ```bash + nvidia-docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest -V -R memory_test" + ``` + +5. Clean Build. + + Sometimes, we might want to clean all thirt-party dependents and built binaries. To do so, just + + ```bash + rm -rf build + ``` + +## Docker, Or Not? + +- What is Docker? + + If you haven't heard of it, consider it something like Python's virtualenv. + +- Docker or virtual machine? + + Some people compare Docker with VMs, but Docker doesn't virtualize any hardware nor running a guest OS, which means there is no compromise on the performance. + +- Why Docker? + + Using a Docker image of build tools standardizes the building environment, which makes it easier for others to reproduce your problems and to help. + + Also, some build tools don't run on Windows or Mac or BSD, but Docker runs almost everywhere, so developers can use whatever computer they want. + +- Can I choose not to use Docker? + + Sure, you don't have to install build tools into a Docker image; instead, you can install them in your local computer. This document exists because Docker would make the development way easier. + +- How difficult is it to learn Docker? + + It takes you ten minutes to read [an introductory article](https://docs.docker.com/get-started) and saves you more than one hour to install all required build tools, configure them, especially when new versions of PaddlePaddle require some new tools. Not even to mention the time saved when other people trying to reproduce the issue you have. + +- Can I use my favorite IDE? + + Yes, of course. The source code resides on your local computer, and you can edit it using whatever editor you like. + + Many PaddlePaddle developers are using Emacs. They add the following few lines into their `~/.emacs` configure file: + + ```emacs + (global-set-key "\C-cc" 'compile) + (setq compile-command + "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev") + ``` + + so they could type `Ctrl-C` and `c` to build PaddlePaddle from source. + +- Does Docker do parallel building? + + Our building Docker image runs a [Bash script](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh), which calls `make -j$(nproc)` to starts as many processes as the number of your CPU cores. + +## Some Gotchas + +- Docker requires sudo + + An owner of a computer has the administrative privilege, a.k.a., sudo, and Docker requires this privilege to work properly. If you use a shared computer for development, please ask the administrator to install and configure Docker. We will do our best to support rkt, another container technology that doesn't require sudo. + +- Docker on Windows/MacOS builds slowly + + On Windows and MacOS, Docker containers run in a Linux VM. You might want to give this VM some more memory and CPUs so to make the building efficient. Please refer to [this issue](https://github.com/PaddlePaddle/Paddle/issues/627) for details. + +- Not enough disk space + + Examples in this article uses option `--rm` with the `docker run` command. This option ensures that stopped containers do not exist on hard disks. We can use `docker ps -a` to list all containers, including stopped. Sometimes `docker build` generates some intermediate dangling images, which also take disk space. To clean them, please refer to [this article](https://zaiste.net/posts/removing_docker_containers/). diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..58665e9f2b6299ec3959ed6858ab01d459f64dd8 --- /dev/null +++ b/doc/howto/dev/new_op_cn.md @@ -0,0 +1,373 @@ +# 如何写新的Operator + + - [概念简介](#概念简介) + - [实现C++类](#实现C++类) + - [定义ProtoMaker类](#定义ProtoMaker类) + - [定义Operator类](#定义Operator类) + - [定义OpKernel类](#定义OpKernel类) + - [注册Operator](#注册Operator) + - [编译](#编译) + - [绑定Python](#绑定Python) + - [实现单元测试](#实现单元测试) + - [前向Operator单测](#前向Operator单测) + - [反向Operator单测](#反向Operator单测) + - [编译和执行](#编译和执行) + + +## 概念简介 + +简单介绍需要用到基类，详细介绍请参考设计文档。 + +- `framework::OperatorBase`: Operator(简写，Op)基类。 +- `framework::OpKernel`: Op计算函数的基类，称作Kernel。 +- `framework::OperatorWithKernel`：继承自OperatorBase，Op有计算函数，称作有Kernel。 +- `class OpProtoAndCheckerMaker`：描述该Op的输入、输出、属性、注释,主要用于Python API接口生成 + +依据是否包含kernel，可以将Op分为两种：包含Kernel的Op和不包含kernel的Op，前者Op的定义继承自`OperatorBase`，后者继承自`OperatorWithKernel`。本教程主要介绍带Kernel的Op如何写，简单总结Op需要包含的内容如下： + + + 内容 | 定义位置 +-------------- | :---------------------- +OpProtoMake定义 | `.cc`文件，Backward Op不需要定义OpProtoMake +Op定义 | `.cc`文件 +Kernel实现 | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，GPU 实现在`.cu`文件中。 +注册Op | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，GPU实现在`.cu`文件中 + + +实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。 + + +下面以矩阵乘操作，即[MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc)为例来介绍如何写带Kernel的Operator。 + + +## 实现C++类 + + +### 1. 定义ProtoMaker类 + +矩阵乘法的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。 + +首先定义`ProtoMaker`来描述该Op的输入、输出，并添加注释： + +```cpp +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"); + 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)继承自`framework::OpProtoAndCheckerMaker`，构造函数含有2个参数： + + - `framework::OpProto` ： 前者存储Op的输入输出和参数属性，将用于Python API接口的生成。 + - `framework::OpAttrChecker` ：后者用于检查参数属性的合法性。 + +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。 + + +再以[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)为例： + +```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); + } +}; +``` + +这个例子有两处不同： + +- `AddInput("X","...").NotInGradient()` : 表示`X`这个输入不参与`ScaleOp`对应的梯度Op计算之中，如果Op的某个输入不参与反向梯度的计算，请显示地调用`.NotInGradient()`进行设置。 + +- `AddAttr("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。 + + +### 2. 定义Operator类 + +下面的点实现了MulOp的定义： + +```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)继承自`OperatorWithKernel`。`public`成员： + +```cpp +using framework::OperatorWithKernel::OperatorWithKernel; +``` + +这句表示使用基类`OperatorWithKernel`的构造函数，也可写成： + +```cpp +MulOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} +``` + +还需要重写`InferShape`接口。`InferShape`为const函数，不能修改Op的成员变量，参数为`const framework::InferShapeContext &ctx`，通过该参数可获取到输入输出以及属性。它的功能是： + + - 1). 做检查， 尽早报错：检查输入数据维度、类型等是否合法。 + - 2). 设置输出Tensor的形状。 + +通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和下面将要介绍的注册函数一起放在`.cc`中 + +### 3. 定义OpKernel类 + +`MulKernel`继承自`framework::OpKernel`，带有下面两个模板参数: + +- `typename Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 + +- `typename T` : 表示数据类型，如`float`, `double`等。 + +需要为`MulKernel`类重写`Compute`接口。 +- `Compute`接受一个输入参数：`const framework::ExecutionContext& context`。 +- 与`InferShapeContext`相比，`ExecutionContext`增加了设备类型，同样可获取到输入输出和属性参数。 +- `Compute`函数里实现`OpKernel`的具体计算逻辑。 + +下面是 `MulKernel` `Compute`的实现： + + ```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); + } + }; + ``` + +需要注意：**不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。** + +`MulOp`的CPU、GPU实现共享同一个`Kernel`。`OpKernel`不共享的例子可以参考：[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 + +为了使`OpKernel`的计算过程书写更加简单，并且CPU、GPU的代码可以复用，我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。关于在PaddlePaddle中如何使用Eigen库，请参考[使用文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md)。 + + +到此，前向Op实现完成。接下来，需要在`.cc`文件中注册该op和kernel。 +反向Op类的定义，反向OpKernel的定义与前向Op类似，这里不再赘述。**但需注意反向Op没有`ProtoMaker`**。 + +### 4. 注册Operator + +- 在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。 + + ```cpp + namespace ops = paddle::operators; + REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad); + REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); + REGISTER_OP_CPU_KERNEL(mul_grad, + ops::MulGradKernel); + ``` + + 在上面的代码中： + + - `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。 + + +- 在 `.cu`文件中注册GPU Kernel。 + - 请注意，如果GPU Kernel的实现基于Eigen unsupported模块，那么在 `.cu`的开始请加上宏定义 `#define EIGEN_USE_GPU`，代码示例如下： + + ```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. 编译 + +- 简单**无特殊依赖**的OP无需修改CMakeList.txt文件。[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt) 会自动将 `paddle/operators` 目录下新增的 `*_op.cc` 文件加入编译。 +- 较为复杂、**有额外依赖** 的operator仍需要修改[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)。如，`mul_op` 依赖 `math_function`，需要在`CMakeLists.txt`中添加如下内容： + + ``` + op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function) + + ``` + +- 运行下面命令可以进行编译： + + ``` + make mul_op + ``` + +## 绑定Python + +- 绑定Python + + 在 [`paddle/pybind/pybind.cc +`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc) 使用`USE_OP`告知编译器需要链接的Op，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。 + + ``` + USE_OP(mul); + ``` + 如果只实现了CPU版本，则使用`USE_CPU_ONLY_OP`: + + ``` + USE_CPU_ONLY_OP(gather); + ``` + + 如果OP不带Kernel，则使用`USE_NO_KENREL_OP`: + + ``` + USE_NO_KENREL_OP(recurrent); + ``` + + + - 生成库 + + 无需修改 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件，`paddle/operators` 目录下新增的 `*_op.cc` 文件会被自动添加链接到生成的lib库中。 + +## 实现单元测试 + +单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。 + +### 前向Operator单元测试 + +前向Op单元测试继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`。各项更加具体的单元测试在`OpTestMeta`里完成。测试前向Operator，需要： + +1. 在`setUp`函数定义输入、输出，以及相关的属性参数。 +2. 生成随机的输入数据。 +3. 在Python脚本中实现与前向operator相同的计算逻辑，得到输出值，与operator前向计算的输出进行对比。 + + + ```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'])} + ``` + +上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释： + +- `self.type = "mul" ` : 定义类型，与operator注册时注册的类型一致。 +- `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。 +- `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。 + + +### 反向Operator单元测试 + +反向Op单元测试继承自`GradientChecker`，而`GradientChecker`继承自`unittest.TestCase`，因此，**反向单元测试函数需要以`test_`开头**。 + +```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_cpu_gpu_compare(self): + self.compare_grad(self.op, self.inputs) + + def test_normal(self): + # mul op will enlarge the relative error + self.check_grad( + self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5) + + def test_ignore_x(self): + self.check_grad( + self.op, + self.inputs, ["Y"], + "Out", + max_relative_error=0.5, + no_grad_set={"X"}) + + def test_ignore_y(self): + self.check_grad( + self.op, + self.inputs, ["X"], + "Out", + max_relative_error=0.5, + no_grad_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。 + + +### 编译和执行单元测试 + +单元测试编写完成之后，在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)中添加以下内容，将单元测试加入工程： + +``` +py_test(test_mul_op SRCS test_mul_op.py) +``` + +请注意，**不同于Op的编译测试，运行单元测试测时需要编译整个工程**，并且编译时需要打开`WITH_TESTING`, 即`cmake paddle_dir -DWITH_TESTING=ON`。编译成功后，执行下面的命令来运行单元测试： + +```bash +make test ARGS="-R test_mul_op -V" +``` + +或者: + +```bash +ctest -R test_mul_op +``` diff --git a/doc/howto/dev/use_eigen_cn.md b/doc/howto/dev/use_eigen_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..1367323b71277984834d9d4f0d9bea0f69478479 --- /dev/null +++ b/doc/howto/dev/use_eigen_cn.md @@ -0,0 +1,146 @@ +## 在Paddle中如何使用Eigen + +神经网络本质上是一个计算图，计算需要的数据存放在`Tensor`中，而计算过程是由`Operartor`来描述的。在执行时，`Operator`调用对应`OpKernel`中的`Compute`接口，实现对`Tensor`的操作。 + + +### Eigen Tensor模块 + +Eigen Tensor模块对element-wise计算提供了强大的支持，并且书写一份代码，可以同时在CPU、GPU执行。但Eigen Tensor是一个正在开发中的模块，因此可能测试不够完备，文档较少。 + +关于Eigen Tensor模块的详细介绍请参考[文档1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md) 和[文档2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md) + + +### paddle::framework::Tensor + +Paddle Tensor定义在framework目录下，其主要接口如下： + +```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`的作用是延迟分配内存，即我们可以先定义一个Tensor，然后使用Resize接口设置Tensor的大小，最后再调用mutable_data接口分配实际的内存。 + +```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使用样例 +下面以AddOp为例说明Tensor的使用过程： + +- InferShape + +在运行神经网络计算图时，我们先调用每个`Operator`的`InferShape`接口，根据输入Tensor的大小来设置输出Tensor的大小，`Resize`接口会被调用。 + +```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 + +`Operator`的`Run`接口最终会调用对应`OpKernel`的`Compute`接口，在这时真正的分配内存，`mutable_data`接口会被调用。 + +```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 + +如上一小节所示，在具体的计算中，我们需要先把输入Tensor和输出Tensor转换为Eigen支持的格式。我们在[eigen.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen.h)中提供了一些全局函数用来实现paddle::framework::Tensor到EigenTensor/EigenMatrix/EigenVector/EigenScalar的转换。 + +以EigenTensor为例，做一个介绍 + +```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是EigenTensor模板提供的一个接口，可以实现从paddle::framework::Tensor到对EigenTensor的转换。由于Tensor的rank是模板参数，因此在转换时需要显示的指定。 + +在Eigen中，不同rank的Tensor是不同类型，Vector是rank为1的Tensor。需要额外注意的是，EigenVector::From方法是把paddle中的一维Tensor转为Eigen的一维Tensor，在这里用EigenVector来表示；而EigenVector::Flatten方法是把paddle中的一个Tensor进行reshape操作，压扁成为Eigen的一维Tensor，类型仍然为EigenVector。 + +更多的转换方法请参考eigen_test.cc中的[单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen_test.cc)。 + + + +### 实现计算 + +当需要完成计算时，我们需要等式左边的EigenTensor调用device接口。在这里需要注意的是，这里的EigenTensor之间的运算只是改变了原有Tensor中的数据，而不会改变原有Tensor的shape信息。 + +```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; +``` + +在这段代码中，input0/input1/output可以是任意维度的Tensor。我们调用了EigenVector的Flatten接口，把任意维度的Tensor转为了一维的EigenVector。而在计算结束之后，input0/input1/output的原有shape信息不变。如果想改变原有Tensor的shape信息，可以调用Resize接口进行改变。 + +由于Eigen Tensor模块的文档较少，我们可以参考TensorFlow的[kernels](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/core/kernels)模块下的相关`OpKernel`的计算代码。 diff --git a/doc/howto/index_cn.rst b/doc/howto/index_cn.rst index 26449a6365843b526b3ac3111b337d2f17524c9d..0608aa30968b0e8474eef330e4d2cc63c9def97d 100644 --- a/doc/howto/index_cn.rst +++ b/doc/howto/index_cn.rst @@ -19,6 +19,7 @@ .. toctree:: :maxdepth: 1 + dev/build_cn.rst dev/write_docs_cn.rst dev/contribute_to_paddle_cn.md diff --git a/doc/howto/index_en.rst b/doc/howto/index_en.rst index 1fbfcd260b912078f00ed5b720ed607db725c4e2..1b6034be4edffd2cbc822018b733b9a3836ea84a 100644 --- a/doc/howto/index_en.rst +++ b/doc/howto/index_en.rst @@ -18,6 +18,7 @@ Development .. toctree:: :maxdepth: 1 + dev/build_en.rst dev/new_layer_en.rst dev/contribute_to_paddle_en.md diff --git a/doc/howto/usage/k8s/k8s_distributed_cn.md b/doc/howto/usage/k8s/k8s_distributed_cn.md index 3121b3f59df650c0a22d0bd305a6f793b202d30e..a9bebf09558b06993119803458977abedbbfbdd0 100644 --- a/doc/howto/usage/k8s/k8s_distributed_cn.md +++ b/doc/howto/usage/k8s/k8s_distributed_cn.md @@ -213,7 +213,7 @@ I1116 09:10:17.123440 50 Util.cpp:130] Calling runInitFunctions I1116 09:10:17.123764 50 Util.cpp:143] Call runInitFunctions done. [WARNING 2016-11-16 09:10:17,227 default_decorators.py:40] please use keyword arguments in paddle config. [INFO 2016-11-16 09:10:17,239 networks.py:1282] The input order is [movie_id, title, genres, user_id, gender, age, occupation, rating] -[INFO 2016-11-16 09:10:17,239 networks.py:1289] The output order is [__mse_cost_0__] +[INFO 2016-11-16 09:10:17,239 networks.py:1289] The output order is [__square_error_cost_0__] I1116 09:10:17.392917 50 Trainer.cpp:170] trainer mode: Normal I1116 09:10:17.613910 50 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process I1116 09:10:17.680917 50 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process diff --git a/doc/index_en.rst b/doc/index_en.rst index 168c7667c61da677905585d6c4b5037ce80b3765..64684b8b9b27e245c6b32ea28809d3bbce22fab9 100644 --- a/doc/index_en.rst +++ b/doc/index_en.rst @@ -7,4 +7,3 @@ PaddlePaddle Documentation getstarted/index_en.rst howto/index_en.rst api/index_en.rst - about/index_en.rst diff --git a/go/master/client.go b/go/master/client.go index 62801b9b7fe85fe27147b12160f48d988623d547..f04cf50ce3cf765a79cbe555d3edb68f3dbb911e 100644 --- a/go/master/client.go +++ b/go/master/client.go @@ -63,13 +63,24 @@ func WithAddr(addr string) func(c *Client) error { // WithEtcd sets the client to use etcd for master discovery. func WithEtcd(endpoints []string, timeout time.Duration) func(*Client) error { return func(c *Client) error { - cli, err := clientv3.New(clientv3.Config{ - Endpoints: endpoints, - DialTimeout: timeout, - }) - if err != nil { + var cli *clientv3.Client + f := func() error { + var err error + cli, err = clientv3.New(clientv3.Config{ + Endpoints: endpoints, + DialTimeout: timeout, + }) return err } + for { + err := f() + if err != nil { + log.Warningln(err) + } else { + break + } + time.Sleep(time.Second) + } ch := make(chan string, 1) a, err := GetKey(cli, DefaultAddrPath, timeout) @@ -101,9 +112,6 @@ func NewClient(opts ...func(*Client) error) (*Client, error) { } } c.ch = make(chan record, c.bufSize) - // FIXME: connection is created asyncrosly in monitorMaster go routine, - // ensure the connection is ready for use before calling c.addClient. - time.Sleep(time.Second) return c, nil } diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index cf61a243e9df2fd4a580e41f07cb0a22dcc72083..ec866b2907d4623e8a94a249bc9af624071ade97 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -15,6 +15,7 @@ if(Boost_FOUND) add_subdirectory(platform) add_subdirectory(framework) add_subdirectory(operators) + add_subdirectory(pybind) endif() if(WITH_C_API) diff --git a/paddle/capi/CMakeLists.txt b/paddle/capi/CMakeLists.txt index 11022d17541476c97a2b29be8eb8fecce7e39435..dde99ab3400be4e61bfe119fc272270518acf070 100644 --- a/paddle/capi/CMakeLists.txt +++ b/paddle/capi/CMakeLists.txt @@ -53,7 +53,10 @@ add_custom_target(paddle_capi_whole ALL set_target_properties(paddle_capi_whole PROPERTIES IMPORTED_LOCATION ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library}) +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) diff --git a/paddle/capi/export.map b/paddle/capi/export.map new file mode 100644 index 0000000000000000000000000000000000000000..8d673f675dd5511f554bff9519a8c078e11868bd --- /dev/null +++ b/paddle/capi/export.map @@ -0,0 +1,6 @@ +{ + global: + paddle_*; + local: + *; +}; diff --git a/paddle/capi/export.sym b/paddle/capi/export.sym new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/paddle/capi/gradient_machine.cpp b/paddle/capi/gradient_machine.cpp index b3287552db87d25edbf6e7f3d5e68121df49e9d6..629449bbd497a7444144c533ad079b3ae6b51438 100644 --- a/paddle/capi/gradient_machine.cpp +++ b/paddle/capi/gradient_machine.cpp @@ -146,3 +146,19 @@ paddle_error paddle_gradient_machine_randomize_param( m->machine->randParameters(); return kPD_NO_ERROR; } + +paddle_error paddle_gradient_machine_get_layer_output( + paddle_gradient_machine machine, + const char* layerName, + paddle_arguments args) { + auto m = cast(machine); + auto out = paddle::capi::cast(args); + if (m == nullptr || layerName == nullptr || out == nullptr || + m->machine == nullptr) { + return kPD_NULLPTR; + } + + auto layerOutput = m->machine->getLayerOutput(layerName); + out->args.push_back(layerOutput); + return kPD_NO_ERROR; +} diff --git a/paddle/capi/gradient_machine.h b/paddle/capi/gradient_machine.h index c613ade5b24efbbf52f21c7ee86dd3189981c5ef..28eeb23e3bbdd4cc22a25c14170bf56c294f8cd7 100644 --- a/paddle/capi/gradient_machine.h +++ b/paddle/capi/gradient_machine.h @@ -39,7 +39,11 @@ PD_API paddle_error paddle_gradient_machine_create_for_inference( /** * @brief Create a gradient machine used for model inference, using config with * parameters which is generated by `paddle merge_model`. - * @param [out] machine that used for model inference. + * Example: + * paddle merge_model \ + * --model_dir="pass-00000" \ + * --model_file="merged_model.paddle" + * @param [out] machine that used for model inference * @param [in] mergedModel * @param [in] size * @return paddle_error @@ -97,6 +101,18 @@ paddle_gradient_machine_randomize_param(paddle_gradient_machine machine); PD_API paddle_error paddle_gradient_machine_destroy(paddle_gradient_machine machine); +/** + * @brief Get the output of the layer named `layerName`. + * @param [in] gradient machine that have run a inference + * @param [in] layerName name of specified layer + * @param [out] args output of the specified layer + * @return paddle_error + */ +PD_API paddle_error +paddle_gradient_machine_get_layer_output(paddle_gradient_machine machine, + const char* layerName, + paddle_arguments args); + #ifdef __cplusplus } #endif diff --git a/paddle/cuda/include/hl_cnn.h b/paddle/cuda/include/hl_cnn.h index 9f84db72da24b0e678520b077f9cba7ffc2d589a..6b56d9ec8d3daae96aaaa04ed79cb637331e2281 100644 --- a/paddle/cuda/include/hl_cnn.h +++ b/paddle/cuda/include/hl_cnn.h @@ -173,6 +173,96 @@ extern void hl_avgpool_backward(const int frameCnt, real* backGrad, const int outStride); +extern void hl_maxpool3D_forward(const int frameCnt, + const real* inputData, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int paddingD, + const int paddingH, + const int paddingW, + real* tgtData, + real* maxPoolIdxData, + const int tgtStride); + +extern void hl_maxpool3D_backward(const int frameCnt, + const real* outGrad, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int paddingD, + const int paddingH, + const int paddingW, + real scaleA, + real scaleB, + real* targetGrad, + real* maxPoolIdxData, + const int outStride); + +extern void hl_avgpool3D_forward(const int frameCnt, + const real* inputData, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int paddingD, + const int paddingH, + const int paddingW, + real* tgtData, + const int tgtStride); + +extern void hl_avgpool3D_backward(const int frameCnt, + const real* outGrad, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + int paddingD, + int paddingH, + int paddingW, + real scaleA, + real scaleB, + real* backGrad, + const int outStride); + /** * @brief Bilinear interpolation forward. * @@ -275,4 +365,4 @@ extern void hl_maxout_backward(real* inGrad, size_t featLen, size_t groups); -#endif /* HL_CNN_H_ */ +#endif // HL_CNN_H_ diff --git a/paddle/cuda/include/hl_cpu_gru.cuh b/paddle/cuda/include/hl_cpu_gru.cuh index c0a37ced2a72a1ab410025e2aa45313c23f1349a..e4f6bf42c61694e9826a127c9628730cfd43ada7 100644 --- a/paddle/cuda/include/hl_cpu_gru.cuh +++ b/paddle/cuda/include/hl_cpu_gru.cuh @@ -18,14 +18,6 @@ limitations under the License. */ #ifndef __NVCC__ -#include "paddle/math/MathFunctions.h" - -#ifndef PADDLE_TYPE_DOUBLE -#define CBLAS_GEMM paddle::gemm -#else -#define CBLAS_GEMM paddle::gemm -#endif - template void hl_naive_gru_forward_reset_output(OpResetOutput opResetOutput, real *gateValue, @@ -210,51 +202,6 @@ inline void forward_final_output(OpFinalOutput opFinalOutput, } } -template -void hl_cpu_gru_forward(OpResetOutput opResetOutput, - OpFinalOutput opFinalOutput, - hl_gru_value value, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate) { - if (value.prevOutValue) { - CBLAS_GEMM(CblasNoTrans, - CblasNoTrans, - batchSize, - 2 * frameSize, - frameSize, - 1, - value.prevOutValue, - frameSize, - value.gateWeight, - frameSize * 2, - 1, - value.gateValue, - frameSize * 3); - } - - forward_reset_output(opResetOutput, value, frameSize, batchSize, active_gate); - - if (value.prevOutValue) { - CBLAS_GEMM(CblasNoTrans, - CblasNoTrans, - batchSize, - frameSize, - frameSize, - 1, - value.resetOutputValue, - frameSize, - value.stateWeight, - frameSize, - 1, - value.gateValue + frameSize * 2, - frameSize * 3); - } - - forward_final_output(opFinalOutput, value, frameSize, batchSize, active_node); -} - template void hl_naive_gru_backward_state_grad(OpStateGrad opStateGrad, real *gateValue, @@ -525,86 +472,6 @@ inline void backward_reset_grad(OpResetGrad opResetGrad, } } -template -void hl_cpu_gru_backward(OpStateGrad opStateGrad, - OpResetGrad opResetGrad, - hl_gru_value value, - hl_gru_grad grad, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate) { - backward_state_grad(opStateGrad, value, grad, - frameSize, batchSize, active_node); - - if (value.prevOutValue && grad.prevOutGrad) { - CBLAS_GEMM(CblasNoTrans, - CblasTrans, - batchSize, - frameSize, - frameSize, - 1, - grad.gateGrad + frameSize * 2, - frameSize * 3, - value.stateWeight, - frameSize, - 0, - grad.resetOutputGrad, - frameSize); - - if (grad.stateWeightGrad) { - CBLAS_GEMM(CblasTrans, - CblasNoTrans, - frameSize, - frameSize, - batchSize, - 1, - value.resetOutputValue, - frameSize, - grad.gateGrad + frameSize * 2, - frameSize * 3, - 1, - grad.stateWeightGrad, - frameSize); - } - } - - backward_reset_grad(opResetGrad, value, grad, - frameSize, batchSize, active_gate); - - if (grad.prevOutGrad && value.prevOutValue) { - CBLAS_GEMM(CblasNoTrans, - CblasTrans, - batchSize, - frameSize, - frameSize * 2, - 1, - grad.gateGrad, - frameSize * 3, - value.gateWeight, - frameSize * 2, - 1, - grad.prevOutGrad, - frameSize); - - if (grad.gateWeightGrad) { - CBLAS_GEMM(CblasTrans, - CblasNoTrans, - frameSize, - frameSize * 2, - batchSize, - 1, - value.prevOutValue, - frameSize, - grad.gateGrad, - frameSize * 3, - 1, - grad.gateWeightGrad, - frameSize * 2); - } - } -} - #endif #endif // HL_CPU_GRU_CUH_ diff --git a/paddle/cuda/include/hl_cuda_cudnn.h b/paddle/cuda/include/hl_cuda_cudnn.h index db18e4912b63ec18dcfff3ef3aaf0c7947e0af18..3f68c62de6d9b3aaadc9180d86159089dc728ea9 100644 --- a/paddle/cuda/include/hl_cuda_cudnn.h +++ b/paddle/cuda/include/hl_cuda_cudnn.h @@ -214,7 +214,8 @@ extern void hl_conv_workspace(hl_tensor_descriptor input, int* convBwdDataAlgo, size_t* bwdDataLimitBytes, int* convBwdFilterAlgo, - size_t* bwdFilterLimitBytes); + size_t* bwdFilterLimitBytes, + bool useDilation); /** * @brief destroy filter descriptor. @@ -242,7 +243,9 @@ extern void hl_create_convolution_descriptor(hl_convolution_descriptor* conv, int padding_height, int padding_width, int stride_height, - int stride_width); + int stride_width, + int dilation_h = 1, + int dilation_w = 1); /** * @brief reset convolution descriptor. @@ -262,7 +265,9 @@ extern void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, int padding_height, int padding_width, int stride_height, - int stride_width); + int stride_width, + int dilation_h = 1, + int dilation_w = 1); /** * @brief destroy convolution descriptor. diff --git a/paddle/cuda/include/hl_matrix.h b/paddle/cuda/include/hl_matrix.h index eb454c59c1e58cf2b4817b4cb3230b9d75e320ac..c7f25109972195fb56b9e96c4b68d952363e6338 100644 --- a/paddle/cuda/include/hl_matrix.h +++ b/paddle/cuda/include/hl_matrix.h @@ -224,4 +224,80 @@ extern void hl_matrix_collect_shared_bias(real* B_d, extern void hl_matrix_rotate( real* mat, real* matRot, int dimM, int dimN, bool clockWise); +/** + * @brief Matrix vol2Col: Convert 3D volume into col matrix + * + * @param[in] matSrc input matrix. + * @param[in] channel channel of matSrc. + * @param[in] depth depth of matSrc. + * @param[in] height height of matSrc. + * @param[in] width width of matSrc. + * @param[in] filterD depth of filter. + * @param[in] filterH height of filter. + * @param[in] filterW width of filter. + * @param[in] strideD stride in the depth. + * @param[in] strideH stride in the height. + * @param[in] strideW stride in the width. + * @param[in] paddingD padding in the depth. + * @param[in] paddingH padding in the height. + * @param[in] paddingW padding in the width. + * @param[out] dataDst output matrix. + * + */ +extern void hl_matrix_vol2Col(const real* dataSrc, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real* dataDst); + +/** + * @brief Matrix col2Vol: Convert col matrix into 3D volume + * + * @param[out] matDst output matrix. + * @param[in] channel channel of matDst. + * @param[in] depth depth of matDst. + * @param[in] height height of matDst. + * @param[in] width width of matDst. + * @param[in] filterD depth of filter. + * @param[in] filterH height of filter. + * @param[in] filterW width of filter. + * @param[in] strideD stride in the depth. + * @param[in] strideH stride in the height. + * @param[in] strideW stride in the width. + * @param[in] paddingD padding in the depth. + * @param[in] paddingH padding in the height. + * @param[in] paddingW padding in the width. + * @param[in] matSrc input matrix. + * @param[in] beta input + * @param[in] alpha input + * + */ +extern void hl_matrix_col2Vol(real* dataDst, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + const real* dataSrc, + real alpha, + real beta); + #endif /* HL_MATRIX_H_ */ diff --git a/paddle/cuda/include/stub/hl_cnn_stub.h b/paddle/cuda/include/stub/hl_cnn_stub.h index 2bbb9fa8dfd5eeac9d55aa67a28ebfbffa2acd46..a76dbf0b6578de0606702ad1af227fbf6e1cd62e 100644 --- a/paddle/cuda/include/stub/hl_cnn_stub.h +++ b/paddle/cuda/include/stub/hl_cnn_stub.h @@ -87,6 +87,96 @@ inline void hl_avgpool_backward(const int frameCnt, real* backGrad, const int outStride) {} +inline void hl_maxpool3D_forward(const int frameCnt, + const real* inputData, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int paddingD, + const int paddingH, + const int paddingW, + real* tgtData, + real* maxPoolIdxData, + const int tgtStride) {} + +inline void hl_maxpool3D_backward(const int frameCnt, + const real* outGrad, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int paddingD, + const int paddingH, + const int paddingW, + real scaleA, + real scaleB, + real* targetGrad, + real* maxPoolIdxData, + const int outStride) {} + +inline void hl_avgpool3D_forward(const int frameCnt, + const real* inputData, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int paddingD, + const int paddingH, + const int paddingW, + real* tgtData, + const int tgtStride) {} + +inline void hl_avgpool3D_backward(const int frameCnt, + const real* outGrad, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int paddingD, + const int paddingH, + const int paddingW, + real scaleA, + real scaleB, + real* backGrad, + const int outStride) {} + inline void hl_bilinear_forward(const real* inData, const size_t inImgH, const size_t inImgW, diff --git a/paddle/cuda/include/stub/hl_cuda_cudnn_stub.h b/paddle/cuda/include/stub/hl_cuda_cudnn_stub.h index abd0d6b09901a7cd124c245e359f9d38f52bda26..3afcc6fa85a4a6a03697663719b6ab685897b68b 100644 --- a/paddle/cuda/include/stub/hl_cuda_cudnn_stub.h +++ b/paddle/cuda/include/stub/hl_cuda_cudnn_stub.h @@ -78,7 +78,9 @@ inline void hl_create_convolution_descriptor(hl_convolution_descriptor* conv, int padding_height, int padding_width, int stride_height, - int stride_width) {} + int stride_width, + int dilation_h, + int dilation_w) {} inline void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, hl_tensor_descriptor image, @@ -86,7 +88,9 @@ inline void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, int padding_height, int padding_width, int stride_height, - int stride_width) {} + int stride_width, + int dilation_h, + int dilation_w) {} inline void hl_destroy_convolution_descriptor(hl_convolution_descriptor conv) {} @@ -99,7 +103,8 @@ inline void hl_conv_workspace(hl_tensor_descriptor input, int* convBwdDataAlgo, size_t* bwdDataLimitBytes, int* convBwdFilterAlgo, - size_t* bwdFilterLimitBytes) {} + size_t* bwdFilterLimitBytes, + bool useDilation) {} inline void hl_convolution_forward(hl_tensor_descriptor input, real* input_data, diff --git a/paddle/cuda/include/stub/hl_matrix_stub.h b/paddle/cuda/include/stub/hl_matrix_stub.h index 127cb7e27983e8ff2c1ff6ef5108b5f8c5bd6ca5..6ac332945c8f09fef23f35680ba5bb1d9ba9f4fd 100644 --- a/paddle/cuda/include/stub/hl_matrix_stub.h +++ b/paddle/cuda/include/stub/hl_matrix_stub.h @@ -99,4 +99,38 @@ inline void hl_matrix_collect_shared_bias(real* B_d, inline void hl_matrix_rotate( real* mat, real* matRot, int dimM, int dimN, bool clockWise) {} +inline void hl_matrix_vol2Col(const real* dataSrc, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real* dataDst) {} + +inline void hl_matrix_col2Vol(real* dataDst, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + const real* dataSrc, + real alpha, + real beta) {} + #endif // HL_MATRIX_STUB_H_ diff --git a/paddle/cuda/src/hl_cuda_cnn.cu b/paddle/cuda/src/hl_cuda_cnn.cu index aac19b1ea566ad69f1f7374e393676c8debd9883..9ba3d142617537c0160f6dccb86ddca43ada15a5 100644 --- a/paddle/cuda/src/hl_cuda_cnn.cu +++ b/paddle/cuda/src/hl_cuda_cnn.cu @@ -353,6 +353,433 @@ void hl_avgpool_backward(const int frameCnt, CHECK_SYNC("hl_avgpool_backward failed"); } +__global__ void KeMaxPool3DForward(const int nthreads, + const real* inputData, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int ksizeD, + const int ksizeH, + const int ksizeW, + const int strideD, + const int strideH, + const int strideW, + const int padD, + const int padH, + const int padW, + real* tgtData, + real* maxPoolIdxData, + const int tgtStride) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < (nthreads); + index += blockDim.x * gridDim.x) { + int pw = index % pooledW; + int ph = (index / pooledW) % pooledH; + int pd = (index / pooledW / pooledH) % pooledD; + int c = (index / pooledW / pooledH / pooledD) % channels; + int frameNum = index / pooledW / pooledH / pooledD / channels; + int dstart = pd * strideD - padD; + int hstart = ph * strideH - padH; + int wstart = pw * strideW - padW; + int dend = min(dstart + ksizeD, depth); + int hend = min(hstart + ksizeH, height); + int wend = min(wstart + ksizeW, width); + dstart = max(dstart, 0); + hstart = max(hstart, 0); + wstart = max(wstart, 0); + real maxval = -FLT_MAX; + int maxIdx = -1; + inputData += (frameNum * channels + c) * depth * height * width; + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + if (maxval < inputData[(d * height + h) * width + w]) { + maxval = inputData[(d * height + h) * width + w]; + maxIdx = (d * height + h) * width + w; + } + } + } + } + int tgtIndex = + index % (pooledW * pooledH * pooledD * channels) + frameNum * tgtStride; + tgtData[tgtIndex] = maxval; + maxPoolIdxData[tgtIndex] = maxIdx; + } +} + +void hl_maxpool3D_forward(const int frameCnt, + const real* inputData, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int padD, + const int padH, + const int padW, + real* tgtData, + real* maxPoolIdxData, + const int tgtStride) { + int num_kernels = pooledD * pooledH * pooledW * channels * frameCnt; + int blocks = (num_kernels + 1024 - 1) / 1024; + dim3 threads(1024, 1); + dim3 grid(blocks, 1); + + KeMaxPool3DForward<<>>(num_kernels, + inputData, + channels, + depth, + height, + width, + pooledD, + pooledH, + pooledW, + sizeZ, + sizeY, + sizeX, + strideD, + strideH, + strideW, + padD, + padH, + padW, + tgtData, + maxPoolIdxData, + tgtStride); + CHECK_SYNC("hl_maxpool3D_forward failed"); +} + +__global__ void KeMaxPool3DBackward(const int nthreads, + const real* outGrad, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int padD, + const int padH, + const int padW, + real scaleA, + real scaleB, + real* targetGrad, + real* maxPoolIdxData, + const int outStride) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < (nthreads); + index += blockDim.x * gridDim.x) { + int offsetW = index % width; + int offsetH = (index / width) % height; + int offsetD = (index / width / height) % depth; + int offsetC = (index / width / height / depth) % channels; + int frameNum = index / width / height / depth / channels; + + int pdstart = + (offsetD + padD < sizeZ) ? 0 : (offsetD + padD - sizeZ) / strideD + 1; + int phstart = + (offsetH + padH < sizeY) ? 0 : (offsetH + padH - sizeY) / strideH + 1; + int pwstart = + (offsetW + padW < sizeX) ? 0 : (offsetW + padW - sizeX) / strideW + 1; + int pdend = min((offsetD + padD) / strideD + 1, pooledD); + int phend = min((offsetH + padH) / strideH + 1, pooledH); + int pwend = min((offsetW + padW) / strideW + 1, pooledW); + + real gradient = 0; + outGrad += ((frameNum * channels + offsetC) * pooledD * pooledH * pooledW); + maxPoolIdxData += + ((frameNum * channels + offsetC) * pooledD * pooledH * pooledW); + for (int pd = pdstart; pd < pdend; ++pd) { + for (int ph = phstart; ph < phend; ++ph) { + for (int pw = pwstart; pw < pwend; ++pw) { + if (((offsetD * height + offsetH) * width + offsetW) == + maxPoolIdxData[(pd * pooledH + ph) * pooledW + pw]) + gradient += outGrad[(pd * pooledH + ph) * pooledW + pw]; + } + } + } + targetGrad[index] = scaleA * gradient + scaleB * targetGrad[index]; + } +} + +void hl_maxpool3D_backward(const int frameCnt, + const real* outGrad, + const int channels, + const int depth, + const int height, + const int width, + const int outputD, + const int outputH, + const int outputW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int paddingD, + const int paddingH, + const int paddingW, + real scaleA, + real scaleB, + real* targetGrad, + real* maxPoolIdxData, + const int outStride) { + int num_kernels = depth * height * width * channels * frameCnt; + int blocks = (num_kernels + 1024 - 1) / 1024; + + KeMaxPool3DBackward<<>>(num_kernels, + outGrad, + channels, + depth, + height, + width, + outputD, + outputH, + outputW, + sizeZ, + sizeY, + sizeX, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + scaleA, + scaleB, + targetGrad, + maxPoolIdxData, + outStride); + CHECK_SYNC("hl_maxpool3D_backward"); +} + +__global__ void KeAvgPool3DForward(const int nthreads, + const real* inputData, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int padD, + const int padH, + const int padW, + real* tgtData, + const int tgtStride) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < (nthreads); + index += blockDim.x * gridDim.x) { + int pw = index % pooledW; + int ph = (index / pooledW) % pooledH; + int pd = (index / pooledW / pooledH) % pooledD; + int c = (index / pooledW / pooledH / pooledD) % channels; + int frameNum = index / pooledW / pooledH / pooledD / channels; + int dstart = pd * strideD - padD; + int hstart = ph * strideH - padH; + int wstart = pw * strideW - padW; + int dend = min(dstart + sizeZ, depth + padD); + int hend = min(hstart + sizeY, height + padH); + int wend = min(wstart + sizeX, width + padW); + int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); + dstart = max(dstart, 0); + hstart = max(hstart, 0); + wstart = max(wstart, 0); + dend = min(dend, depth); + hend = min(hend, height); + wend = min(wend, width); + + real aveval = 0; + inputData += (frameNum * channels + c) * depth * height * width; + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + aveval += inputData[(d * height + h) * width + w]; + } + } + } + int tgtIndex = + index % (pooledW * pooledH * pooledD * channels) + frameNum * tgtStride; + tgtData[tgtIndex] = aveval / pool_size; + } +} + +void hl_avgpool3D_forward(const int frameCnt, + const real* inputData, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int paddingD, + const int paddingH, + const int paddingW, + real* tgtData, + const int tgtStride) { + int num_kernels = pooledD * pooledH * pooledW * channels * frameCnt; + int blocks = (num_kernels + 1024 - 1) / 1024; + KeAvgPool3DForward<<>>(num_kernels, + inputData, + channels, + depth, + height, + width, + pooledD, + pooledH, + pooledW, + sizeZ, + sizeY, + sizeX, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + tgtData, + tgtStride); + CHECK_SYNC("hl_avgpool3D_forward failed"); +} + +__global__ void KeAvgPool3DBackward(const int nthreads, + const real* outGrad, + const int channels, + const int depth, + const int height, + const int width, + const int pooledD, + const int pooledH, + const int pooledW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + const int padD, + const int padH, + const int padW, + real scaleA, + real scaleB, + real* tgtGrad, + const int outStride) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < (nthreads); + index += blockDim.x * gridDim.x) { + int offsetW = index % width + padW; + int offsetH = (index / width) % height + padH; + int offsetD = (index / width / height) % depth + padD; + int offsetC = (index / width / height / depth) % channels; + int frameNum = index / width / height / depth / channels; + + int pdstart = (offsetD < sizeZ) ? 0 : (offsetD - sizeZ) / strideD + 1; + int phstart = (offsetH < sizeY) ? 0 : (offsetH - sizeY) / strideH + 1; + int pwstart = (offsetW < sizeX) ? 0 : (offsetW - sizeX) / strideW + 1; + int pdend = min(offsetD / strideD + 1, pooledD); + int phend = min(offsetH / strideH + 1, pooledH); + int pwend = min(offsetW / strideW + 1, pooledW); + + real gradient = 0; + outGrad += (frameNum * channels + offsetC) * pooledD * pooledH * pooledW; + + 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 * strideD - padD; + int hstart = ph * strideH - padH; + int wstart = pw * strideW - padW; + int dend = min(dstart + sizeZ, depth + padD); + int hend = min(hstart + sizeY, height + padH); + int wend = min(wstart + sizeX, width + padW); + int poolsize = (dend - dstart) * (hend - hstart) * (wend - wstart); + gradient += outGrad[(pd * pooledH + ph) * pooledW + pw] / poolsize; + } + } + } + tgtGrad[index] = scaleA * gradient + scaleB * tgtGrad[index]; + } +} + +void hl_avgpool3D_backward(const int frameCnt, + const real* outGrad, + const int channels, + const int depth, + const int height, + const int width, + const int outputD, + const int outputH, + const int outputW, + const int sizeZ, + const int sizeY, + const int sizeX, + const int strideD, + const int strideH, + const int strideW, + int paddingD, + int paddingH, + int paddingW, + real scaleA, + real scaleB, + real* backGrad, + const int outStride) { + int num_kernels = depth * height * width * channels * frameCnt; + int blocks = (num_kernels + 1024 - 1) / 1024; + + KeAvgPool3DBackward<<>>(num_kernels, + outGrad, + channels, + depth, + height, + width, + outputD, + outputH, + outputW, + sizeZ, + sizeY, + sizeX, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + scaleA, + scaleB, + backGrad, + outStride); + CHECK_SYNC("hl_avgpool3D_backward failed"); +} + __global__ void KeBilinearInterpFw(const real* in, const size_t inImgH, const size_t inImgW, diff --git a/paddle/cuda/src/hl_cuda_cudnn.cc b/paddle/cuda/src/hl_cuda_cudnn.cc index 78642a17443b0b4d81defaa46579332ef20c71a1..f38ef692558b908ed65d2c84821bbb7c3b439742 100644 --- a/paddle/cuda/src/hl_cuda_cudnn.cc +++ b/paddle/cuda/src/hl_cuda_cudnn.cc @@ -201,7 +201,8 @@ void hl_conv_workspace(hl_tensor_descriptor input, int* convBwdDataAlgo, size_t* bwdDataLimitBytes, int* convBwdFilterAlgo, - size_t* bwdFilterLimitBytes) { + size_t* bwdFilterLimitBytes, + bool useDilation) { #if CUDNN_VERSION >= 4000 CHECK_NOTNULL(input); @@ -213,21 +214,60 @@ void hl_conv_workspace(hl_tensor_descriptor input, size_t memoryLimitBytes = (1LL << 20) * FLAGS_cudnn_conv_workspace_limit_in_mb; + // For dilation + int algo = 0; + // cudnn convolution forward configuration cudnnTensorDescriptor_t fwd_src_desc = GET_TENSOR_DESCRIPTOR(input); cudnnTensorDescriptor_t fwd_dest_desc = GET_TENSOR_DESCRIPTOR(output); cudnnFilterDescriptor_t fwd_filter_desc = GET_FILTER_DESCRIPTOR(filter); cudnnConvolutionDescriptor_t fwd_conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv); + // cudnn convolution backward data configuration + cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter); + cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output); + cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input); + cudnnConvolutionDescriptor_t bwd_data_conv_desc = + GET_CONVOLUTION_DESCRIPTOR(conv); + // cudnn convolution backward filter configuration + cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input); + cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output); + cudnnConvolutionDescriptor_t bwd_filter_conv_desc = + GET_CONVOLUTION_DESCRIPTOR(conv); + cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter); - CHECK_CUDNN(dynload::cudnnGetConvolutionForwardAlgorithm( - t_resource.cudnn_handle, - fwd_src_desc, - fwd_filter_desc, - fwd_conv_desc, - fwd_dest_desc, - CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, - memoryLimitBytes, - reinterpret_cast(convFwdAlgo))); + if (useDilation) { + convFwdAlgo = &algo; + convBwdDataAlgo = &algo; + convBwdFilterAlgo = &algo; + } else { + CHECK_CUDNN(dynload::cudnnGetConvolutionForwardAlgorithm( + t_resource.cudnn_handle, + fwd_src_desc, + fwd_filter_desc, + fwd_conv_desc, + fwd_dest_desc, + CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, + memoryLimitBytes, + reinterpret_cast(convFwdAlgo))); + CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataAlgorithm( + t_resource.cudnn_handle, + bwd_data_filter_desc, + bwd_data_diff_desc, + bwd_data_conv_desc, + bwd_data_grad_desc, + CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT, + memoryLimitBytes, + reinterpret_cast(convBwdDataAlgo))); + CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterAlgorithm( + t_resource.cudnn_handle, + bwd_filter_src_desc, + bwd_filter_diff_desc, + bwd_filter_conv_desc, + bwd_filter_grad_desc, + CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT, + memoryLimitBytes, + reinterpret_cast(convBwdFilterAlgo))); + } CHECK_CUDNN(dynload::cudnnGetConvolutionForwardWorkspaceSize( t_resource.cudnn_handle, @@ -238,23 +278,6 @@ void hl_conv_workspace(hl_tensor_descriptor input, static_cast(*convFwdAlgo), fwdLimitBytes)); - // cudnn convolution backward data configuration - cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter); - cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output); - cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input); - cudnnConvolutionDescriptor_t bwd_data_conv_desc = - GET_CONVOLUTION_DESCRIPTOR(conv); - - CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataAlgorithm( - t_resource.cudnn_handle, - bwd_data_filter_desc, - bwd_data_diff_desc, - bwd_data_conv_desc, - bwd_data_grad_desc, - CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT, - memoryLimitBytes, - reinterpret_cast(convBwdDataAlgo))); - CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataWorkspaceSize( t_resource.cudnn_handle, bwd_data_filter_desc, @@ -264,23 +287,6 @@ void hl_conv_workspace(hl_tensor_descriptor input, static_cast(*convBwdDataAlgo), bwdDataLimitBytes)); - // cudnn convolution backward filter configuration - cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input); - cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output); - cudnnConvolutionDescriptor_t bwd_filter_conv_desc = - GET_CONVOLUTION_DESCRIPTOR(conv); - cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter); - - CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterAlgorithm( - t_resource.cudnn_handle, - bwd_filter_src_desc, - bwd_filter_diff_desc, - bwd_filter_conv_desc, - bwd_filter_grad_desc, - CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT, - memoryLimitBytes, - reinterpret_cast(convBwdFilterAlgo))); - CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize( t_resource.cudnn_handle, bwd_filter_src_desc, @@ -603,7 +609,9 @@ void hl_create_convolution_descriptor(hl_convolution_descriptor* conv, int padding_height, int padding_width, int stride_height, - int stride_width) { + int stride_width, + int dilation_h, + int dilation_w) { CHECK_NOTNULL(conv); cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)malloc( @@ -625,18 +633,24 @@ void hl_create_convolution_descriptor(hl_convolution_descriptor* conv, padding_width, stride_height, stride_width, - 1, - 1, + dilation_h, + dilation_w, mode, data_type)); #else + if (dilation_h > 1 || dilation_w > 1) { + LOG(FATAL) + << "Current cuDNN version does't support for dilation convolution. " + << "The dilation convolution requires cuDNN >= v6.0."; + } + CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(hl_conv->desc, padding_height, padding_width, stride_height, stride_width, - 1, - 1, + dilation_h, + dilation_w, mode)); #endif @@ -659,7 +673,9 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, int padding_height, int padding_width, int stride_height, - int stride_width) { + int stride_width, + int dilation_h, + int dilation_w) { CHECK_NOTNULL(conv); CHECK_NOTNULL(image); CHECK_NOTNULL(filter); @@ -678,8 +694,8 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, padding_width, stride_height, stride_width, - 1, - 1, + dilation_h, + dilation_w, mode, data_type)); #else @@ -688,8 +704,8 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, padding_width, stride_height, stride_width, - 1, - 1, + dilation_h, + dilation_w, mode)); #endif diff --git a/paddle/cuda/src/hl_cuda_matrix.cu b/paddle/cuda/src/hl_cuda_matrix.cu index 39272456c394adc0509e60cf5972df832f7b3424..b41a3a1e06db7b2566acef19ce430645f79d486d 100644 --- a/paddle/cuda/src/hl_cuda_matrix.cu +++ b/paddle/cuda/src/hl_cuda_matrix.cu @@ -592,3 +592,204 @@ void hl_matrix_rotate( mat, matRot, dimM, dimN, clockWise); CHECK_SYNC("hl_matrix_rotate failed"); } + +__global__ void keMatrixVol2Col(int num_kernels, + const real* dataSrc, + real* dataDst, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + int depth_col, + int height_col, + int width_col) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels; + index += blockDim.x * gridDim.x) { + int w_out = index % width_col; + int h_out = (index / width_col) % height_col; + int d_out = (index / width_col / height_col) % depth_col; + int channel_in = index / width_col / height_col / depth_col; + int channel_out = channel_in * filterD * filterH * filterW; + int w_in = w_out * strideW - paddingW; + int h_in = h_out * strideH - paddingH; + int d_in = d_out * strideD - paddingD; + + dataDst += + ((channel_out * depth_col + d_out) * height_col + h_out) * width_col + + w_out; + dataSrc += ((channel_in * depth + d_in) * height + h_in) * width + w_in; + for (int k = 0; k < filterD; ++k) { + for (int i = 0; i < filterH; ++i) { + for (int j = 0; j < filterW; ++j) { + int d = d_in + k; + int h = h_in + i; + int w = w_in + j; + *dataDst = (d >= 0 && d < depth && h >= 0 && h < height && w >= 0 && + w < width) + ? dataSrc[(k * height + i) * width + j] + : 0; + dataDst += depth_col * height_col * width_col; + } + } + } + } +} + +void hl_matrix_vol2Col(const real* dataSrc, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real* dataDst) { + int depth_col = (depth + 2 * paddingD - filterD) / strideD + 1; + int height_col = (height + 2 * paddingH - filterH) / strideH + 1; + int width_col = (width + 2 * paddingW - filterW) / strideW + 1; + int num_kernels = channels * depth_col * height_col * width_col; + + const int threads = 512; + const int blocks = DIVUP(num_kernels, threads); + + keMatrixVol2Col<<>>(num_kernels, + dataSrc, + dataDst, + depth, + height, + width, + filterD, + filterH, + filterW, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + depth_col, + height_col, + width_col); + CHECK_SYNC("hl_matrix_vol2Col failed"); +} + +__global__ void keMatrixCol2Vol(int num_kernels, + real* dataDst, + const real* dataSrc, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + int depth_col, + int height_col, + int width_col, + real alpha, + real beta) { + for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels; + index += blockDim.x * gridDim.x) { + real srcVal = 0; + real dstVal = dataDst[index]; + int w = index % width + paddingW; + int h = (index / width) % height + paddingH; + int d = (index / width / height) % depth + paddingD; + int c = index / width / height / depth; + // compute the start and end of the output + int w_col_start = (w < filterW) ? 0 : (w - filterW) / strideW + 1; + int w_col_end = min(w / strideW + 1, width_col); + int h_col_start = (h < filterH) ? 0 : (h - filterH) / strideH + 1; + int h_col_end = min(h / strideH + 1, height_col); + int d_col_start = (d < filterD) ? 0 : (d - filterD) / strideD + 1; + int d_col_end = min(d / strideD + 1, depth_col); + + int offset = (c * filterD * filterW * filterH + d * filterW * filterH + + h * filterW + w) * + depth_col * height_col * width_col; + + int coeff_d_col = + (1 - strideD * filterW * filterH * depth_col) * height_col * width_col; + int coeff_h_col = + (1 - strideH * filterW * depth_col * height_col) * width_col; + int coeff_w_col = (1 - strideW * depth_col * height_col * width_col); + + for (int d_col = d_col_start; d_col < d_col_end; ++d_col) { + for (int h_col = h_col_start; h_col < h_col_end; ++h_col) { + for (int w_col = w_col_start; w_col < w_col_end; ++w_col) { + srcVal += dataSrc[offset + d_col * coeff_d_col + h_col * coeff_h_col + + w_col * coeff_w_col]; + } + } + } + dataDst[index] = alpha * srcVal + beta * dstVal; + } +} + +void hl_matrix_col2Vol(real* dataDst, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + const real* dataSrc, + real alpha, + real beta) { + int depth_col = (depth + 2 * paddingD - filterD) / strideD + 1; + int height_col = (height + 2 * paddingH - filterH) / strideH + 1; + int width_col = (width + 2 * paddingW - filterW) / strideW + 1; + int num_kernels = channels * depth * height * width; + + const int threads = 512; + const int blocks = DIVUP(num_kernels, threads); + + keMatrixCol2Vol<<>>(num_kernels, + dataDst, + dataSrc, + depth, + height, + width, + filterD, + filterH, + filterW, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + depth_col, + height_col, + width_col, + alpha, + beta); + + CHECK_SYNC("hl_matrix_col2Vol failed"); +} diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt index 9024ed2fd427d7d21d7899a3b8df61d86f08a2cd..c0838d9b759110fd706577386d2c81bda6876223 100644 --- a/paddle/framework/CMakeLists.txt +++ b/paddle/framework/CMakeLists.txt @@ -15,23 +15,19 @@ cc_test(variable_test SRCS variable_test.cc) cc_library(scope SRCS scope.cc) cc_test(scope_test SRCS scope_test.cc DEPS scope) -proto_library(attribute_proto SRCS attribute.proto) -proto_library(op_proto SRCS op_proto.proto DEPS attribute_proto) -proto_library(op_desc SRCS op_desc.proto DEPS attribute_proto) -cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf) -cc_test(op_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf) +proto_library(framework_proto SRCS framework.proto) -cc_library(attribute SRCS attribute.cc DEPS op_desc op_proto) - -cc_library(operator SRCS operator.cc DEPS op_desc device_context tensor scope attribute) +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_test(operator_test SRCS operator_test.cc DEPS operator op_registry) -cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS op_proto operator) -cc_library(op_registry SRCS op_registry.cc DEPS op_desc grad_op_builder) +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_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 attribute.proto op_proto.proto op_desc.proto) +py_proto_compile(framework_py_proto SRCS framework.proto) # Generate an empty __init__.py to make framework_py_proto as a valid python module. add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py) add_dependencies(framework_py_proto framework_py_proto_init) @@ -42,22 +38,4 @@ add_custom_command(TARGET framework_py_proto POST_BUILD WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) cc_library(backward SRCS backward.cc DEPS net_op) -cc_test(backward_test SRCS backward_test.cc DEPS backward) - -if(WITH_PYTHON) -cc_library(paddle_pybind SHARED - SRCS pybind.cc - DEPS pybind python backward - sgd_op - add_op - mul_op - rowwise_add_op - sigmoid_op - softmax_op - mean_op - cross_entropy_op - recurrent_op - uniform_random_op - gaussian_random_op - fill_zeros_like_op) -endif(WITH_PYTHON) +cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context) diff --git a/paddle/framework/attribute.cc b/paddle/framework/attribute.cc index 4c5790693b7e48396e945d09f4fdc72b86aa5978..27132eaa0b3b0666fc042faf052dac2e169ba9e7 100644 --- a/paddle/framework/attribute.cc +++ b/paddle/framework/attribute.cc @@ -43,8 +43,12 @@ template <> AttrType AttrTypeID>() { return STRINGS; } +template <> +AttrType AttrTypeID>>() { + return INT_PAIRS; +} -Attribute GetAttrValue(const AttrDesc& attr_desc) { +Attribute GetAttrValue(const OpDesc::Attr& attr_desc) { switch (attr_desc.type()) { case paddle::framework::AttrType::INT: { return attr_desc.i(); @@ -76,6 +80,14 @@ Attribute GetAttrValue(const AttrDesc& attr_desc) { } 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; + } } PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !"); return boost::blank(); diff --git a/paddle/framework/attribute.h b/paddle/framework/attribute.h index 49a62bedb6aadab5ff05d8aa7dda42fe983314a0..2b788a76cafe198abb9aed8ba842e37cc6ff73a6 100644 --- a/paddle/framework/attribute.h +++ b/paddle/framework/attribute.h @@ -20,8 +20,7 @@ limitations under the License. */ #include #include -#include "paddle/framework/attribute.pb.h" -#include "paddle/framework/op_desc.pb.h" +#include "paddle/framework/framework.pb.h" #include "paddle/platform/enforce.h" #include "paddle/platform/variant.h" @@ -29,7 +28,8 @@ namespace paddle { namespace framework { typedef boost::variant, - std::vector, std::vector> + std::vector, std::vector, + std::vector>> Attribute; typedef std::unordered_map AttributeMap; @@ -37,15 +37,27 @@ typedef std::unordered_map AttributeMap; template AttrType AttrTypeID(); -Attribute GetAttrValue(const AttrDesc& attr_desc); +Attribute GetAttrValue(const OpDesc::Attr& attr_desc); // check whether a value(attribute) fit a certain limit template -class LargerThanChecker { +class GreaterThanChecker { public: - explicit LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {} + explicit GreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {} void operator()(T& value) const { - PADDLE_ENFORCE(value > lower_bound_, "larger_than check fail"); + PADDLE_ENFORCE(value > lower_bound_, "larger_than check fails."); + } + + private: + T lower_bound_; +}; + +template +class EqualGreaterThanChecker { + public: + explicit EqualGreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {} + void operator()(T& value) const { + PADDLE_ENFORCE_GE(value, lower_bound_, "equal_larger_than check fails."); } private: @@ -110,8 +122,13 @@ class TypedAttrChecker { return *this; } - TypedAttrChecker& LargerThan(const T& lower_bound) { - value_checkers_.push_back(LargerThanChecker(lower_bound)); + TypedAttrChecker& GreaterThan(const T& lower_bound) { + value_checkers_.push_back(GreaterThanChecker(lower_bound)); + return *this; + } + + TypedAttrChecker& EqualGreaterThan(const T& lower_bound) { + value_checkers_.push_back(EqualGreaterThanChecker(lower_bound)); return *this; } diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc index 437a44a8aafa650d654a1a77c60613abe07679fe..c5d46622156c56acb98fb77e7db5ee7bca8c937a 100644 --- a/paddle/framework/backward.cc +++ b/paddle/framework/backward.cc @@ -15,31 +15,44 @@ #include "paddle/framework/backward.h" #include +#include + #include "paddle/framework/op_registry.h" #include "paddle/operators/net_op.h" +#include "paddle/operators/recurrent_op.h" namespace paddle { namespace framework { -static bool AllInSet(const std::vector& names, - const std::string& suffix, - const std::unordered_set& set) { +template +static void ForEachVarName(const Map& names, T callback) { for (auto& name : names) { - if (set.find(name + suffix) == set.end()) { - return false; + for (auto& n : name.second) { + if (callback(n)) return; } } - return true; } -static std::shared_ptr NOP() { - auto net_op = std::make_shared(); - net_op->type_ = "@NOP@"; +// return whether all the names + suffixes in the set +static bool AllInSet( + const std::map>& names, + const std::string& suffix, const std::unordered_set& set) { + bool all_in_set = true; + ForEachVarName(names, [&all_in_set, &set, &suffix](const std::string& n) { + all_in_set = set.find(n + suffix) != set.end(); + return !all_in_set; + }); + return all_in_set; +} + +static std::unique_ptr NOP() { + auto net_op = new operators::NetOp(); + net_op->SetType("@NOP@"); net_op->CompleteAddOp(); - return net_op; + return std::unique_ptr(net_op); } -// Get backward operator from a forward operator, recursively implementation. +// Get backward operator from a forward operator, a recursive implementation. // // no_grad_names the gradient variable names without gradient calculating. // @@ -47,126 +60,159 @@ static std::shared_ptr NOP() { // BackwardRecursive. use `uid = uniq_id++;` to get the unique index, and // pass `uniq_id` through recursive calling. // -// returns The backward operator. For simple situation, it is a simple -// operator. For complex situation, it is a NetOp. +// returns The backward operator. In a simple situation, it may be a simple +// operator, in a complex situation, it maybe a NetOp. // // See Backward.h for details -static std::shared_ptr BackwardRecursive( - const OperatorBase& forwardOp, - std::unordered_set& no_grad_names, size_t& uniq_id); -std::shared_ptr BackwardRecursive( +static std::unique_ptr BackwardRecursive( const OperatorBase& forwardOp, std::unordered_set& no_grad_names, size_t& uniq_id) { // If all input gradients of forwarding operator do not need to calculate, // just return an NOP. Not return null ptr because NOP does not take // too much time for calculation, but it is useful for simplifying logic. - if (AllInSet(forwardOp.inputs_, kGradVarSuffix, no_grad_names)) { + if (AllInSet(forwardOp.Inputs() /*names*/, kGradVarSuffix /*suffix*/, + no_grad_names /*set*/)) { return NOP(); } // All output gradients of forwarding operator do not need to calculate. // Then all input gradients cannot be computed at all, and we put them into // `no_grad_names` set. Return an NOP. - if (AllInSet(forwardOp.outputs_, kGradVarSuffix, no_grad_names)) { - for (auto& name : forwardOp.inputs_) { - // Mark all input is not need - no_grad_names.insert(name + kGradVarSuffix); - } + if (AllInSet(forwardOp.Outputs() /*names*/, kGradVarSuffix /*suffix*/, + no_grad_names /*set*/)) { + ForEachVarName(forwardOp.Inputs(), + [&no_grad_names](const std::string& name) -> bool { + no_grad_names.insert(GradVarName(name)); + return false; + }); return NOP(); } // Returned gradient network - auto net = std::make_shared(); + auto net = std::unique_ptr(new operators::NetOp()); if (forwardOp.IsNetOp()) { // Because forwardOp is a net op, it can static_cast. auto& forwardNet = static_cast(forwardOp); // Map from output gradient variable name to operator's indices in - // backward net. That operator generates that variable. + // backward net's ops_. That operator generates that variable. std::unordered_map> dup_output_ops; size_t local_op_id = 0; - // reversely travel forwardNet + // reversely travel forwardNet and collect all duplicate outputs. for (auto it = forwardNet.ops_.rbegin(); it != forwardNet.ops_.rend(); ++it, ++local_op_id) { - auto fwd = *it; + auto& fwd = *it; auto bwd = BackwardRecursive(*fwd, no_grad_names, uniq_id); - net->AddOp(bwd); - for (auto& out : bwd->outputs_) { - dup_output_ops[out].emplace_back(local_op_id); - } + ForEachVarName(bwd->Outputs(), + [&dup_output_ops, local_op_id](const std::string& out) { + dup_output_ops[out].emplace_back(local_op_id); + return false; + }); + net->AppendOp(std::move(bwd)); } // Get unique ID for this method. auto uid = uniq_id++; // TODO(dzh): more comment - using Pos = std::pair>; + // multiple operators which have the same output (y for example) may + // overwrite the same y variable when backward, special operations are token + // to handle this case. For each duplicate output, rename it to an alias + // (original name with a offset), append an `add` op for its operator, + // and finally sum all the alias variable to the final output variable y. + using Pos = std::pair>; std::list insert_position; for (auto& dup_output_op : dup_output_ops) { const std::string& name = dup_output_op.first; + // duplicate @Empty@ don't need to be added + if (name == kEmptyVarName) continue; + auto& dup_op = dup_output_op.second; + // no duplicate output if (dup_op.size() == 1) continue; - std::vector dup_outputs; + // process the duplicate outputs + std::vector dup_outputs; for (size_t i = 0; i < dup_op.size(); ++i) { + // rename each duplicate output to an alias auto op_offset = dup_op[i]; dup_outputs.push_back(name + "@RENAME@" + std::to_string(uid) + "@" + std::to_string(i)); 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", {dup_outputs}, {name}, - {{"input_format", - std::vector{0, static_cast(dup_outputs.size())}}})}); + {dup_op.back(), OpRegistry::CreateOp("add", {{"X", {dup_outputs}}}, + {{"Out", {name}}}, {})}); } + // make sure the inserted `add` ops follow the BFS order. insert_position.sort( [](const Pos& l, const Pos& r) { return l.first > r.first; }); for (auto& pos : insert_position) { - net->InsertOp(pos.first + 1, pos.second); + net->InsertOp(pos.first + 1, std::move(pos.second)); } - } else { - std::shared_ptr grad_op = OpRegistry::CreateGradOp(forwardOp); - for (std::string& grad_input : grad_op->inputs_) { + std::unique_ptr grad_op(OpRegistry::CreateGradOp(forwardOp)); + + ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op]( + const std::string& grad_input) { if (no_grad_names.count(grad_input)) { // +1 for \0 std::string prefix = grad_input.substr( 0, grad_input.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1); - grad_input = prefix + kZeroVarSuffix; + grad_op->Rename(grad_input, prefix + kZeroVarSuffix); // If part of input gradient of that operator is not calculated, fill // zero variables to that input gradient. - net->AddOp(OpRegistry::CreateOp("fill_zeros_like", {prefix}, - {grad_input}, {})); - } - } - - for (std::string& grad_output : grad_op->outputs_) { - if (no_grad_names.count(grad_output)) { - grad_output = kEmptyVarName; + net->AppendOp(OpRegistry::CreateOp("fill_zeros_like", + {{"Src", {prefix}}}, + {{"Dst", {grad_input}}}, {})); } + return false; + }); + + ForEachVarName(grad_op->Outputs(), + [&no_grad_names, &grad_op](const std::string& grad_output) { + if (no_grad_names.count(grad_output)) { + grad_op->Rename(grad_output, kEmptyVarName); + } + return false; + }); + + // process recurrent gradient op as a special operator. + if (forwardOp.Type() == "recurrent") { + // NOTE clean up cycle call somewhere (RNN's stepnet constains itself), or + // this will result in infinite loop. + const auto& rnnop = + *static_cast(&forwardOp); + auto rnn_grad_op = + static_cast(grad_op.get()); + const auto& stepnet_op = + *static_cast(&rnnop.stepnet()); + // create stepnet's gradient op + rnn_grad_op->set_stepnet( + BackwardRecursive(stepnet_op, no_grad_names, uniq_id)); } if (net->ops_.empty()) { // Current no aux op is added to network return grad_op; } - net->AddOp(grad_op); + net->AppendOp(std::move(grad_op)); } - net->type_ = "@GENERATED_BACKWARD@"; + net->SetType("@GENERATED_BACKWARD@"); net->CompleteAddOp(); - return net; + return std::unique_ptr( + static_cast(net.release())); } // See header for comments -std::shared_ptr Backward( +std::unique_ptr Backward( const OperatorBase& forwardOp, const std::unordered_set& no_grad_vars) { std::unordered_set no_grad_names; - no_grad_names.reserve(no_grad_vars.size()); + no_grad_names.reserve(no_grad_vars.size() + 1); no_grad_names.insert(std::string(kEmptyVarName) + kGradVarSuffix); diff --git a/paddle/framework/backward.h b/paddle/framework/backward.h index c181919dc165cf0b49362f85e22ceb4131bbd387..1ecf69881b3126c2904920b9f4b77bfcccc9cf86 100644 --- a/paddle/framework/backward.h +++ b/paddle/framework/backward.h @@ -20,7 +20,7 @@ namespace framework { // Create the backward operator from a forward operator. // TODO(yuyang18): Add more API reference comment. -extern std::shared_ptr Backward( +extern std::unique_ptr Backward( const OperatorBase& forwardOp, const std::unordered_set& no_grad_vars); } // namespace framework diff --git a/paddle/framework/backward.md b/paddle/framework/backward.md index 74c001b06a9e7b2279abf998604f2acf1b1168e4..c762811dfc190b255e0a3389885a081ce8315caf 100644 --- a/paddle/framework/backward.md +++ b/paddle/framework/backward.md @@ -1,38 +1,82 @@ -## Operator/expression 's Backward +# Operator/expression 's Backward -### Motivation +## Motivation -In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation lineage, the operator/ expression's Backward feature will generate the backward pass respect to forward pass. +In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the gradient operators/expressions together with the chain rule. 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. -### Implement : gradient operator registry +## Backward Operator Registry -| | forward operator | backward operator | -| ---------------------- | ---------------- | -------------------------------- | -| **Operator::inputs_** | Inputs | Inputs, Outputs, OutputGradients | -| **Operator::outputs_** | Outputs | InputGradients | +A backward network is built up with several backward operators. Backward operators take forward operators' inputs outputs, and output gradients and then calculate its input gradients. -Inputs/Outputs means the input/output of the operator, InputGradients/OutputGradients is the gradient respect to forward opeartor. Forward operator and Backward operator are isomorphic, save their corresponding needs into member attribute. +| | forward operator | backward operator +| ---------------------- | ---------------- |------------------------- | +| **Operator::inputs_** | Inputs | Inputs, Outputs, OutputGradients | +| **Operator::outputs_** | Outputs | InputGradients | -We use a global hash map record the gradient operators available, follow the philosophy of minimum core, make operator pluggable unit. Each gradient is an operator and it needs to regist itself. + 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. -grad_op_builder(fengjiayi) +For example, we have got a `mul_op`, and we can register its information and corresponding backward operator by the following macro: -### Implement : Backward network +```cpp +REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad); +``` -given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`,`InputGradients`. +`mul` is the operator's type. `MulOp` and `MulOpMaker` are the operator class and the operator maker class respectively. -1. bla bla bla (yuyang) +`mul_grad` is the type of backward operator, and `MulOpGrad` is its class name. + +## Backward Opeartor Creating + +Given a certain forward operator, we can get its corresponding backward operator by calling: + +```cpp +OperatorBase* bwd_op = BuildGradOp(const OperatorBase* fwd_op); +``` + +The function `BuildGradOp` will sequentially execute following processes: + +1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`. + +2. Build two maps named `inputs` and `outputs` to 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. + +3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`. + +4. Building backward operator with `inputs`, `outputs` and forward operator's attributes. + +## Backward Network Building + +A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and put them together. + +In our design, the network itself is also a kind of operator. So the operators contained by a big network may be some small network. + +given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`, `InputGradients`. + +1. Op + + when the input forward network is an Op, return its gradient Operator Immediately. 2. NetOp - when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively and ensure them done. During the process, we need to collect the `OutputGradients` name. + when the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp. + + **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwrite their shared input variable. + +

+
+ + 1. Shared variable in operators. + +

+ + Share 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 replace the overwrite links. + +

+
- We share variable in the same scope, as a result, duplicate operator `OutputGradients` will overwirte then duplicate variable. + 2. Replace shared variable's gradient with `Add` operator. - ![./images/duplicate_op]() +

- Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator instead. -![./images/duplicate_op2]() -​ Then collect the sub graph OutputGradients/InputGradients as the NetOp's and return it. +​ 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 da3b9c8bed7cd123f2f8ef982a5f0e23abcc0ec7..ad8003420dc14538d0dae9a1cb19d6459b154576 100644 --- a/paddle/framework/backward_test.cc +++ b/paddle/framework/backward_test.cc @@ -28,21 +28,13 @@ using OpAttrChecker = framework::OpAttrChecker; using Scope = framework::Scope; using DeviceContext = platform::DeviceContext; -class EmptyOp : public OperatorBase { - public: - DEFINE_OPERATOR_CTOR(EmptyOp, OperatorBase) - - void InferShape(const Scope &scope) const override {} - void Run(const Scope &scope, const DeviceContext &dev_ctx) const override {} -}; - class RowWiseAddOpMaker : public OpProtoAndCheckerMaker { public: RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "Input X of Add").IgnoreGradient(); - AddInput("b", "Bias of Add").IgnoreGradient(); - AddOutput("Out", "Out of Add").IgnoreGradient(); + AddInput("X", "Input X of Add").NotInGradient(); + AddInput("b", "Bias of Add").NotInGradient(); + AddOutput("Out", "Out of Add").NotInGradient(); AddComment("Add Op"); } }; @@ -51,8 +43,8 @@ class MulOpMaker : public OpProtoAndCheckerMaker { public: MulOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("A", "A"); - AddInput("B", "B"); + AddInput("X", "A"); + AddInput("Y", "B"); AddOutput("Out", "Out"); AddComment("Mul"); } @@ -63,7 +55,7 @@ class SigmoidOpMaker : public OpProtoAndCheckerMaker { SigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "X"); - AddOutput("Y", "Y"); + AddOutput("Out", "Y"); AddComment("Sigmoid"); } }; @@ -73,21 +65,25 @@ class NoGradOpMaker : public OpProtoAndCheckerMaker { NoGradOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "X input"); - AddOutput("Y", "Y output"); + AddOutput("Out", "Y output"); AddComment("NoGradOp, same input output. no Grad"); } }; class FcOp : public operators::NetOp { public: - void Init() override { - AddOp(OpRegistry::CreateOp("mul", {Input("X"), Input("W")}, - {Output("mul_result")}, {})); - auto b_name = Input("b"); + FcOp(const std::string &type, const VariableNameMap &inputs, + const VariableNameMap &outputs, const AttributeMap &attrs) + : NetOp(type, inputs, outputs, attrs) { + AppendOp(OpRegistry::CreateOp("mul", + {{"X", {Input("X")}}, {"Y", {Input("W")}}}, + {{"Out", {Output("mul_result")}}}, {})); + auto input_b = Inputs("b"); std::string before_act = "mul_result"; - if (b_name != kEmptyVarName) { - AddOp(OpRegistry::CreateOp("rowwise_add", {Output("mul_result"), b_name}, - {Output("add_result")}, {})); + if (input_b.size() != 0) { + AppendOp(OpRegistry::CreateOp( + "rowwise_add", {{"X", {Output("mul_result")}}, {"b", {input_b[0]}}}, + {{"Out", {Output("add_result")}}}, {})); before_act = "add_result"; } else { auto out_varname = Output("add_result"); @@ -96,8 +92,8 @@ class FcOp : public operators::NetOp { } } - AddOp(OpRegistry::CreateOp("sigmoid", {Output(before_act)}, {Output("Out")}, - {})); + AppendOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}}, + {{"Out", {Output("Out")}}}, {})); CompleteAddOp(false); } }; @@ -109,8 +105,8 @@ class FcOpMaker : public OpProtoAndCheckerMaker { AddInput("X", "x"); AddInput("W", "w"); AddInput("b", "b"); - AddOutput("mul_result", "").SetTemporary(); - AddOutput("add_result", "").SetTemporary(); + AddOutput("mul_result", "").AsIntermediate(); + AddOutput("add_result", "").AsIntermediate(); AddOutput("Out", ""); AddComment(""); } @@ -131,8 +127,8 @@ class FillZeroOpMaker : public OpProtoAndCheckerMaker { public: FillZeroOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("x", "x"); - AddOutput("out", "out"); + AddInput("Src", "x"); + AddOutput("Dst", "out"); AddComment(""); } }; @@ -141,8 +137,8 @@ class AddOpMaker : public OpProtoAndCheckerMaker { public: AddOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "x").SetMultiple(); - AddOutput("Y", "y"); + AddInput("X", "x").AsDuplicable(); + AddOutput("Out", "out"); AddComment(""); } }; @@ -152,51 +148,48 @@ class AddOpMaker : public OpProtoAndCheckerMaker { namespace f = paddle::framework; namespace ops = paddle::operators; using EnforceNotMet = paddle::platform::EnforceNotMet; -REGISTER_OP(rowwise_add, f::EmptyOp, f::RowWiseAddOpMaker); -REGISTER_GRADIENT_OP(rowwise_add, rowwise_add_grad, f::EmptyOp); -REGISTER_OP(mul, f::EmptyOp, f::MulOpMaker); -REGISTER_GRADIENT_OP(mul, mul_grad, f::EmptyOp); -REGISTER_OP(sigmoid, f::EmptyOp, f::SigmoidOpMaker); -REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, f::EmptyOp); -REGISTER_OP(nograd, f::EmptyOp, f::NoGradOpMaker); -REGISTER_OP(fill_zeros_like, f::EmptyOp, f::FillZeroOpMaker); -REGISTER_OP(add, f::EmptyOp, f::AddOpMaker); -REGISTER_GRADIENT_OP(add, add_grad, f::EmptyOp); -REGISTER_OP(fc, f::FcOp, f::FcOpMaker); -REGISTER_OP(many_output_op, f::EmptyOp, f::ManyOutputOpMaker); -REGISTER_GRADIENT_OP(many_output_op, many_output_op_grad, f::EmptyOp); +REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, 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_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", "b"}, {"Out"}, {}); + 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(4UL, gop->inputs_.size()); - ASSERT_EQ(f::kEmptyVarName, gop->inputs_[0]); - ASSERT_EQ("rowwise_add_grad", gop->type_); - ASSERT_EQ(f::GradVarName("X"), gop->outputs_[0]); - ASSERT_EQ(f::GradVarName("b"), gop->outputs_[1]); - - ASSERT_EQ(f::GradVarName("X"), gop->Output(f::GradVarName("X"))); + 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"))); } TEST(Backward, simple_op_not_need_grad) { - auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {}); + auto fwd = f::OpRegistry::CreateOp( + "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {}); ASSERT_NE(fwd, nullptr); - auto gop = f::Backward(*fwd, {"X"}); - ASSERT_EQ(std::find(gop->outputs_.begin(), gop->outputs_.end(), - f::GradVarName("X")), - gop->outputs_.end()); + auto gop = f::Backward(*fwd, {"x"}); + ASSERT_EQ(gop->Output(f::GradVarName("X")), f::kEmptyVarName); - auto no_input_gop = f::Backward(*fwd, {"X", "b"}); + auto no_input_gop = f::Backward(*fwd, {"x", "b"}); ASSERT_NE(no_input_gop, nullptr); ASSERT_TRUE(no_input_gop->IsNetOp()); - ASSERT_EQ(0UL, - std::static_pointer_cast(no_input_gop)->ops_.size()); + ASSERT_EQ(0UL, static_cast(no_input_gop.get())->ops_.size()); } TEST(Backward, net_fc_backward_normal) { - std::shared_ptr fwd = f::OpRegistry::CreateOp( - "fc", {"X", "w", "b"}, {"mul_result", "add_result", "out"}, {}); + std::shared_ptr fwd = + f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}}, + {{"mul_result", {"mul_res"}}, + {"add_result", {"add_re"}}, + {"Out", {"out"}}}, + {}); ASSERT_NE(fwd, nullptr); std::shared_ptr gop = f::Backward(*fwd, {}); ASSERT_TRUE(gop->IsNetOp()); @@ -207,19 +200,22 @@ TEST(Backward, net_fc_backward_normal) { ASSERT_EQ(3UL, net->ops_.size()); f::OperatorBase &d_sigmoid = *net->ops_[0]; - ASSERT_EQ("sigmoid_grad", d_sigmoid.type_); + ASSERT_EQ("sigmoid_grad", d_sigmoid.Type()); f::OperatorBase &d_add = *net->ops_[1]; - ASSERT_EQ("rowwise_add_grad", d_add.type_); + ASSERT_EQ("rowwise_add_grad", d_add.Type()); f::OperatorBase &d_mul = *net->ops_[2]; - ASSERT_EQ("mul_grad", d_mul.type_); + ASSERT_EQ("mul_grad", d_mul.Type()); } TEST(Backward, net_fc_backward_not_have_b) { std::shared_ptr fwd = - f::OpRegistry::CreateOp("fc", {"X", "w", f::kEmptyVarName}, - {"mul_result", "add_result", "tmp"}, {}); + f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {}}}, + {{"mul_result", {"mul_res"}}, + {"add_result", {"add_res"}}, + {"Out", {"tmp"}}}, + {}); ASSERT_NE(fwd, nullptr); std::shared_ptr gop = f::Backward(*fwd, {}); ASSERT_TRUE(gop->IsNetOp()); @@ -230,96 +226,113 @@ TEST(Backward, net_fc_backward_not_have_b) { ASSERT_EQ(2UL, net->ops_.size()); f::OperatorBase &d_sigmoid = *net->ops_[0]; - ASSERT_EQ("sigmoid_grad", d_sigmoid.type_); + ASSERT_EQ("sigmoid_grad", d_sigmoid.Type()); f::OperatorBase &d_mul = *net->ops_[1]; - ASSERT_EQ("mul_grad", d_mul.type_); + ASSERT_EQ("mul_grad", d_mul.Type()); } TEST(Backward, net_input_of_network_not_need_grad) { ops::NetOp net; - net.AddOp(f::OpRegistry::CreateOp("fc", {"X", "W1", "b1"}, - {"mul_tmp_0", "add_tmp_0", "hidden0"}, {})); - net.AddOp(f::OpRegistry::CreateOp("fc", {"hidden0", "W2", "b2"}, - {"mul_tmp_1", "add_tmp_1", "hidden1"}, {})); + net.AppendOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"x"}}, {"W", {"W1"}}, {"b", {"b1"}}}, + {{"mul_result", {"mul_tmp_0"}}, + {"add_result", {"add_tmp_0"}}, + {"Out", {"hidden0"}}}, + {})); + net.AppendOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"hidden0"}}, {"W", {"W2"}}, {"b", {"b2"}}}, + {{"mul_result", {"mul_tmp_1"}}, + {"add_result", {"add_tmp_1"}}, + {"Out", {"hidden1"}}}, + {})); net.CompleteAddOp(); - auto bwd = Backward(net, {"X"}); // X@GRAD is not need. + auto bwd = Backward(net, {"x"}); // x@GRAD is not need. ASSERT_TRUE(bwd->IsNetOp()); auto bwd_net = static_cast(bwd.get()); - std::unordered_set all_output = std::unordered_set( - bwd_net->outputs_.begin(), bwd_net->outputs_.end()); - all_output.erase(f::kEmptyVarName); + auto output_vars = bwd_net->OutputVars(true); + std::unordered_set all_outputs = + std::unordered_set(output_vars.begin(), output_vars.end()); + all_outputs.erase(f::kEmptyVarName); for (auto &out : {"W1", "b1", "hidden0", "W2", "b2"}) { - ASSERT_NE(all_output.find(f::GradVarName(out)), all_output.end()); + ASSERT_NE(all_outputs.find(f::GradVarName(out)), all_outputs.end()); } // Not Generated X - ASSERT_EQ(all_output.find(f::GradVarName("X")), all_output.end()); + ASSERT_EQ(all_outputs.find(f::GradVarName("X")), all_outputs.end()); ASSERT_EQ(2UL, bwd_net->ops_.size()); ASSERT_TRUE(bwd_net->ops_[1]->IsNetOp()); auto first_fc_grad = static_cast(bwd_net->ops_[1].get()); ASSERT_EQ(3UL, first_fc_grad->ops_.size()); ASSERT_EQ(f::kEmptyVarName, - first_fc_grad->ops_[2]->Output(f::GradVarName("A"))); + first_fc_grad->ops_[2]->Output(f::GradVarName("X"))); } TEST(Backward, net_shared_weight) { ops::NetOp net; - net.AddOp(f::OpRegistry::CreateOp("mul", {"X", "W"}, {"Out"}, {})); - net.AddOp(f::OpRegistry::CreateOp("mul", {"Out", "W"}, {"FinalOut"}, {})); + net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}}, + {{"Out", {"out"}}}, {})); + net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}}, + {{"Out", {"FinalOut"}}}, {})); net.CompleteAddOp(); auto bwd = f::Backward(net, {}); 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_); + ASSERT_EQ("add", bwd_net->ops_[2]->Type()); } TEST(Backward, op_register_grad_not_for_network) { - auto fwd = f::OpRegistry::CreateOp( - "fc", {"X", "W", "b"}, {"mul_out", "add_out", "out1"}, - {{"temporary_index", std::vector{0, 1}}}); + 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); } TEST(Backward, op_all_input_are_not_need) { - auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {}); - auto backward = f::Backward(*fwd, {"X", "b"}); + auto fwd = f::OpRegistry::CreateOp( + "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {}); + auto backward = f::Backward(*fwd, {"x", "b"}); ASSERT_TRUE(backward->IsNetOp()); auto net = static_cast(backward.get()); ASSERT_TRUE(net->ops_.empty()); } TEST(Backward, op_all_output_are_not_need) { - auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {}); - auto backward = f::Backward(*fwd, {"Out"}); + auto fwd = f::OpRegistry::CreateOp( + "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {}); + auto backward = f::Backward(*fwd, {"out"}); ASSERT_TRUE(backward->IsNetOp()); auto net = static_cast(backward.get()); ASSERT_TRUE(net->ops_.empty()); } TEST(Backward, op_part_of_output_are_not_need) { - auto fwd = f::OpRegistry::CreateOp("many_output_op", {"X"}, {"Y", "Z"}, {}); + auto fwd = f::OpRegistry::CreateOp("many_output_op", {{"x", {"X"}}}, + {{"y", {"Y"}}, {"z", {"Z"}}}, {}); auto backward = f::Backward(*fwd, {"Z"}); ASSERT_TRUE(backward->IsNetOp()); auto net = static_cast(backward.get()); ASSERT_EQ(net->ops_.size(), 2UL); auto &fill_zero = *net->ops_[0]; - ASSERT_EQ("fill_zeros_like", fill_zero.type_); - ASSERT_EQ(1UL, fill_zero.inputs_.size()); - ASSERT_EQ("Z", fill_zero.inputs_[0]); - ASSERT_EQ(1UL, fill_zero.outputs_.size()); - ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.outputs_[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")); auto &d_many_out = *net->ops_[1]; - ASSERT_EQ("many_output_op_grad", d_many_out.type_); - ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.inputs_.size()); // I/O/OG + ASSERT_EQ("many_output_op_grad", d_many_out.Type()); + ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.Inputs().size()); // I/O/OG ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, d_many_out.Input(f::GradVarName("z"))); ASSERT_EQ(f::GradVarName("Y"), d_many_out.Input(f::GradVarName("y"))); @@ -327,44 +340,62 @@ TEST(Backward, op_part_of_output_are_not_need) { } TEST(Backward, op_part_of_input_are_not_need) { - auto fwd = f::OpRegistry::CreateOp("mul", {"a", "b"}, {"out"}, {}); + auto fwd = f::OpRegistry::CreateOp("mul", {{"X", {"a"}}, {"Y", {"b"}}}, + {{"Out", {"out"}}}, {}); auto backward = f::Backward(*fwd, {"a"}); auto &grad_mul = *backward; - ASSERT_EQ(grad_mul.type_, "mul_grad"); - ASSERT_EQ(grad_mul.inputs_.size(), 2UL + 1UL + 1UL); - ASSERT_EQ(grad_mul.outputs_.size(), 2UL); - ASSERT_EQ(grad_mul.Output(f::GradVarName("A")), f::kEmptyVarName); - ASSERT_EQ(grad_mul.Output(f::GradVarName("B")), f::GradVarName("b")); + ASSERT_EQ(grad_mul.Type(), "mul_grad"); + ASSERT_EQ(grad_mul.Inputs().size(), 2UL + 1UL + 1UL); + ASSERT_EQ(grad_mul.Outputs().size(), 2UL); + ASSERT_EQ(grad_mul.Output(f::GradVarName("X")), f::kEmptyVarName); + ASSERT_EQ(grad_mul.Output(f::GradVarName("Y")), f::GradVarName("b")); ASSERT_EQ(grad_mul.Input(f::GradVarName("Out")), f::GradVarName("out")); - ASSERT_EQ(grad_mul.Input("A"), "a"); - ASSERT_EQ(grad_mul.Input("B"), "b"); + ASSERT_EQ(grad_mul.Input("X"), "a"); + ASSERT_EQ(grad_mul.Input("Y"), "b"); ASSERT_EQ(grad_mul.Input("Out"), "out"); } TEST(Backward, linear_net_intermediate_variable_has_no_grad) { ops::NetOp net; - net.AddOp(f::OpRegistry::CreateOp("fc", {"x1", "w1", "b1"}, - {"mul_out1", "add_out1", "out1"}, {})); - net.AddOp(f::OpRegistry::CreateOp("fc", {"out1", "w2", "b2"}, - {"mul_out2", "tmp_out2", "out2"}, {})); - net.AddOp(f::OpRegistry::CreateOp("fc", {"out2", "w3", "b3"}, - {"mul_out3", "tmp_out3", "out3"}, {})); + net.AppendOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"x1"}}, {"W", {"w1"}}, {"b", {"b1"}}}, + {{"mul_result", {"mul_out1"}}, + {"add_result", {"add_out1"}}, + {"Out", {"out1"}}}, + {})); + net.AppendOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"out1"}}, {"W", {"w2"}}, {"b", {"b2"}}}, + {{"mul_result", {"mul_out2"}}, + {"add_result", {"tmp_out2"}}, + {"Out", {"out2"}}}, + {})); + net.AppendOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"out2"}}, {"W", {"w3"}}, {"b", {"b3"}}}, + {{"mul_result", {"mul_out3"}}, + {"add_result", {"tmp_out3"}}, + {"Out", {"out3"}}}, + {})); net.CompleteAddOp(); + auto backward = f::Backward(net, {"mul_out2", "tmp_out2", "out2"}); ASSERT_TRUE(backward->IsNetOp()); auto bwd_net = static_cast(backward.get()); ASSERT_EQ(bwd_net->ops_.size(), 3UL); auto &grad_fc = *bwd_net->ops_[0]; - EXPECT_EQ(grad_fc.inputs_.size(), - 3UL /* external input number */ + + const char *all = paddle::operators::NetOp::kAll; + EXPECT_EQ(grad_fc.Inputs(all).size(), + 2UL /* external input number */ + 1UL /* external output number*/ + 1UL /* number of gradient of external output*/ + 2U /* internal variable number*/); - EXPECT_EQ(grad_fc.outputs_.size(), 2UL /* input number of mul*/ - + 2UL /* input number of rowwise_add */ - + 1UL /* input number of sigmod */); - EXPECT_EQ(bwd_net->ops_[1]->inputs_.size(), 0UL); - EXPECT_EQ(bwd_net->ops_[1]->outputs_.size(), 0UL); - EXPECT_EQ(bwd_net->ops_[2]->inputs_.size(), 0UL); - EXPECT_EQ(bwd_net->ops_[2]->outputs_.size(), 0UL); + EXPECT_EQ(grad_fc.Outputs(all).size(), + 2UL /* input number of mul*/ + + 2UL /* input number of rowwise_add + */ + + 1UL /* input number of sigmod */); + EXPECT_EQ(bwd_net->ops_[1]->Inputs(all).size(), 0UL); + EXPECT_EQ(bwd_net->ops_[1]->Outputs(all).size(), 0UL); + EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL); + EXPECT_EQ(bwd_net->ops_[2]->Outputs(all).size(), 0UL); } diff --git a/paddle/framework/ddim.cc b/paddle/framework/ddim.cc index 545c1dcc2a1682839d90194002fdbb748d85e808..fc3d508553c0e966978b28d58127bdbff10d45f1 100644 --- a/paddle/framework/ddim.cc +++ b/paddle/framework/ddim.cc @@ -21,16 +21,16 @@ namespace framework { /// @cond HIDDEN template -Dim make_dim(const int* d) { +Dim make_dim(const int64_t* d) { return Dim(*d, make_dim(d + 1)); } template <> -Dim<1> make_dim<1>(const int* d) { +Dim<1> make_dim<1>(const int64_t* d) { return Dim<1>(*d); } -void make_ddim(DDim& ddim, const int* dims, int n) { +void make_ddim(DDim& ddim, const int64_t* dims, int n) { switch (n) { case 1: ddim = make_dim<1>(dims); @@ -67,13 +67,13 @@ void make_ddim(DDim& ddim, const int* dims, int n) { /// @endcond -DDim make_ddim(std::initializer_list dims) { +DDim make_ddim(std::initializer_list dims) { DDim result(make_dim(0)); make_ddim(result, dims.begin(), dims.size()); return result; } -DDim make_ddim(const std::vector& dims) { +DDim make_ddim(const std::vector& dims) { DDim result(make_dim(0)); make_ddim(result, &dims[0], dims.size()); return result; @@ -81,12 +81,12 @@ DDim make_ddim(const std::vector& dims) { /// @cond HIDDEN // XXX For some reason, putting this in an anonymous namespace causes errors -class DynamicMutableIndexer : public boost::static_visitor { +class DynamicMutableIndexer : public boost::static_visitor { public: explicit DynamicMutableIndexer(int idx) : idx_(idx) {} template - int& operator()(Dim& dim) const { + int64_t& operator()(Dim& dim) const { return dim[idx_]; } @@ -94,12 +94,12 @@ class DynamicMutableIndexer : public boost::static_visitor { int idx_; }; -class DynamicConstIndexer : public boost::static_visitor { +class DynamicConstIndexer : public boost::static_visitor { public: explicit DynamicConstIndexer(int idx) : idx_(idx) {} template - int operator()(const Dim& dim) const { + int64_t operator()(const Dim& dim) const { return dim[idx_]; } @@ -109,22 +109,22 @@ class DynamicConstIndexer : public boost::static_visitor { /// @endcond -int& DDim::operator[](int idx) { +int64_t& DDim::operator[](int idx) { return boost::apply_visitor(DynamicMutableIndexer(idx), var); } -int DDim::operator[](int idx) const { +int64_t DDim::operator[](int idx) const { return boost::apply_visitor(DynamicConstIndexer(idx), var); } -ssize_t DDim::size() const { return arity(*this); } +int64_t DDim::size() const { return arity(*this); } bool DDim::operator==(DDim d) const { if (var.which() != d.getVar().which()) { return false; } else { - std::vector v1 = vectorize(*this); - std::vector v2 = vectorize(d); + std::vector v1 = vectorize(*this); + std::vector v2 = vectorize(d); for (unsigned int i = 0; i < v1.size(); i++) { if (v1[i] != v2[i]) { @@ -139,10 +139,10 @@ bool DDim::operator==(DDim d) const { bool DDim::operator!=(DDim d) const { return !(*this == d); } DDim DDim::operator+(DDim d) const { - std::vector v1 = vectorize(*this); - std::vector v2 = vectorize(d); + std::vector v1 = vectorize(*this); + std::vector v2 = vectorize(d); - std::vector v3; + std::vector v3; assert(v1.size() == v2.size()); @@ -154,10 +154,10 @@ DDim DDim::operator+(DDim d) const { } DDim DDim::operator*(DDim d) const { - std::vector v1 = vectorize(*this); - std::vector v2 = vectorize(d); + std::vector v1 = vectorize(*this); + std::vector v2 = vectorize(d); - std::vector v3; + std::vector v3; assert(v1.size() == v2.size()); @@ -168,15 +168,15 @@ DDim DDim::operator*(DDim d) const { return make_ddim(v3); } -int get(const DDim& ddim, int idx) { return ddim[idx]; } +int64_t get(const DDim& ddim, int idx) { return ddim[idx]; } void set(DDim& ddim, int idx, int value) { ddim[idx] = value; } /// @cond HIDDEN struct VectorizeVisitor : public boost::static_visitor<> { - std::vector& vector; + std::vector& vector; - explicit VectorizeVisitor(std::vector& v) : vector(v) {} + explicit VectorizeVisitor(std::vector& v) : vector(v) {} template void operator()(const T& t) { @@ -188,31 +188,31 @@ struct VectorizeVisitor : public boost::static_visitor<> { }; /// @endcond -std::vector vectorize(const DDim& ddim) { - std::vector result; +std::vector vectorize(const DDim& ddim) { + std::vector result; VectorizeVisitor visitor(result); boost::apply_visitor(visitor, ddim); return result; } -struct ProductVisitor : public boost::static_visitor { +struct ProductVisitor : public boost::static_visitor { template - ssize_t operator()(const Dim& dim) { + int64_t operator()(const Dim& dim) { return product(dim); } }; -ssize_t product(const DDim& ddim) { +int64_t product(const DDim& ddim) { ProductVisitor visitor; return boost::apply_visitor(visitor, ddim); } struct SliceVectorizeVisitor : public boost::static_visitor<> { - std::vector& vector; + std::vector& vector; int begin; int end; - SliceVectorizeVisitor(std::vector& v, int b, int e) + SliceVectorizeVisitor(std::vector& v, int b, int e) : vector(v), begin(b), end(e) { PADDLE_ENFORCE(begin < end, "Begin index must be less than end index in ddim slice."); @@ -240,7 +240,7 @@ struct SliceVectorizeVisitor : public boost::static_visitor<> { }; DDim slice_ddim(const DDim& dim, int begin, int end) { - std::vector vec; + std::vector vec; vec.reserve(end - begin); SliceVectorizeVisitor visitor(vec, begin, end); boost::apply_visitor(visitor, dim); @@ -280,9 +280,17 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) { return os; } -DDim::DDim(std::initializer_list init_list) { +DDim::DDim(std::initializer_list init_list) { *this = make_ddim(init_list); } +DDim flatten_to_2d(const DDim& src, int num_col_dims) { + int rank = src.size(); + return make_ddim({product(slice_ddim(src, 0, num_col_dims)), + product(slice_ddim(src, num_col_dims, rank))}); +} + +DDim flatten_to_1d(const DDim& src) { return make_ddim({product(src)}); } + } // namespace framework } // namespace paddle diff --git a/paddle/framework/ddim.h b/paddle/framework/ddim.h index 95f294b62737be5c3eac39303148ac35da29fe7d..ca29e7e8c7776de6adf3e3b0e8f11f0d4d8487c3 100644 --- a/paddle/framework/ddim.h +++ b/paddle/framework/ddim.h @@ -40,7 +40,7 @@ struct DDim { template explicit DDim(const Dim& in) : var(in) {} - /*implicit*/ DDim(std::initializer_list init_list); + /*implicit*/ DDim(std::initializer_list init_list); template DDim& operator=(const Dim& in) { @@ -48,8 +48,8 @@ struct DDim { return *this; } - int& operator[](int idx); - int operator[](int idx) const; + int64_t& operator[](int idx); + int64_t operator[](int idx) const; template typename Visitor::result_type apply_visitor(Visitor& visitor) { @@ -71,15 +71,15 @@ struct DDim { DDim operator*(DDim d) const; - ssize_t size() const; + int64_t size() const; }; /** - * \brief Make a DDim from std::vector + * \brief Make a DDim from std::vector * * \param dims An vector of ints. Must be sized between [1, 9] */ -DDim make_ddim(const std::vector& dims); +DDim make_ddim(const std::vector& dims); /** * \brief Make a DDim from an initializer list @@ -87,14 +87,14 @@ DDim make_ddim(const std::vector& dims); * \param dims An initializer list of ints. Must be sized between [1, 9] * */ -DDim make_ddim(std::initializer_list dims); +DDim make_ddim(std::initializer_list dims); -int get(const DDim& dim, int idx); +int64_t get(const DDim& dim, int idx); void set(DDim& dim, int idx, int val); -std::vector vectorize(const DDim& ddim); +std::vector vectorize(const DDim& ddim); -ssize_t product(const DDim& ddim); +int64_t product(const DDim& ddim); /** * \brief Slice a ddim @@ -115,6 +115,12 @@ int arity(const DDim& ddim); std::ostream& operator<<(std::ostream&, const DDim&); +// Reshape a tensor to a matrix. The matrix's first dimension(column length) +// will be the product of tensor's first `num_col_dims` dimensions. +DDim flatten_to_2d(const DDim& src, int num_col_dims); + +DDim flatten_to_1d(const DDim& src); + } // namespace framework } // namespace paddle diff --git a/paddle/framework/ddim_test.cc b/paddle/framework/ddim_test.cc index 9d18a2972ce62139430b240b4599854b14290a32..756232b1b56a49d2c91cc2cac950ca508c54fb3f 100644 --- a/paddle/framework/ddim_test.cc +++ b/paddle/framework/ddim_test.cc @@ -12,7 +12,7 @@ TEST(DDim, Equality) { EXPECT_EQ(ddim[2], 5); // construct a DDim from a vector - std::vector vec({9, 1, 5}); + std::vector vec({9, 1, 5}); paddle::framework::DDim vddim = paddle::framework::make_ddim(vec); EXPECT_EQ(ddim[0], 9); EXPECT_EQ(ddim[1], 1); @@ -25,7 +25,7 @@ TEST(DDim, Equality) { EXPECT_EQ(paddle::framework::get(ddim, 0), 6); // vectorize a DDim - std::vector res_vec = paddle::framework::vectorize(vddim); + std::vector res_vec = paddle::framework::vectorize(vddim); EXPECT_EQ(res_vec[0], 9); EXPECT_EQ(res_vec[1], 1); EXPECT_EQ(res_vec[2], 5); diff --git a/paddle/framework/dim.h b/paddle/framework/dim.h index 883fdc55eb929ebc51e8ae05938e9d07374406ce..04d4b0e604e6f73ad94e0ca79d6b69f663bd4076 100644 --- a/paddle/framework/dim.h +++ b/paddle/framework/dim.h @@ -17,13 +17,13 @@ struct Dim { static constexpr int dimensions = i; template - HOSTDEVICE Dim(int _head, Args... _tail) : head(_head), tail(_tail...) { + HOSTDEVICE Dim(int64_t _head, Args... _tail) : head(_head), tail(_tail...) { static_assert(sizeof...(_tail) == i - 1, "Dim initialized with the wrong number of parameters"); } HOSTDEVICE - Dim(int _head, const Dim& _tail) : head(_head), tail(_tail) {} + Dim(int64_t _head, const Dim& _tail) : head(_head), tail(_tail) {} HOSTDEVICE Dim() : head(0), tail() {} @@ -31,12 +31,12 @@ struct Dim { /** Construct a Dim from a linear index and size. Uses Fortran order * indexing. */ HOSTDEVICE - Dim(int idx, const Dim& size) + Dim(int64_t idx, const Dim& size) : head(idx % size.head), tail(idx / size.head, size.tail) {} /** Construct a Dim with each dimension set to the given index */ HOSTDEVICE - Dim(int idx) : head(idx), tail(idx) {} + Dim(int64_t idx) : head(idx), tail(idx) {} HOSTDEVICE bool operator==(const Dim& o) const { @@ -47,13 +47,13 @@ struct Dim { bool operator!=(const Dim& o) const { return !(*this == o); } HOSTDEVICE - int& operator[](int idx); + int64_t& operator[](int idx); HOSTDEVICE - int operator[](int idx) const; + int64_t operator[](int idx) const; HOST std::string to_string() const; - int head; + int64_t head; Dim tail; }; @@ -63,7 +63,7 @@ struct Dim<1> { static constexpr int dimensions = 1; HOSTDEVICE - Dim(int _head) : head(_head) {} + Dim(int64_t _head) : head(_head) {} HOSTDEVICE Dim() : head(0) {} @@ -86,11 +86,11 @@ struct Dim<1> { bool operator!=(const Dim<1>& o) const { return !(*this == o); } HOSTDEVICE - int& operator[](int idx); + int64_t& operator[](int idx); HOSTDEVICE - int operator[](int idx) const; + int64_t operator[](int idx) const; - int head; + int64_t head; }; namespace { @@ -100,12 +100,12 @@ template struct DimGetter { // Return a copy if Dim is const template - HOSTDEVICE static int impl(const D& d) { + HOSTDEVICE static int64_t impl(const D& d) { return DimGetter::impl(d.tail); } // Return a reference if Dim is mutable template - HOSTDEVICE static int& impl(D& d) { + HOSTDEVICE static int64_t& impl(D& d) { return DimGetter::impl(d.tail); } }; @@ -115,18 +115,18 @@ template <> struct DimGetter<0> { // Return a copy if Dim is const template - HOSTDEVICE static int impl(const D& d) { + HOSTDEVICE static int64_t impl(const D& d) { return d.head; } // Return a reference if Dim is mutable template - HOSTDEVICE static int& impl(D& d) { + HOSTDEVICE static int64_t& impl(D& d) { return d.head; } }; template -HOSTDEVICE int& indexer(Dim& dim, int idx) { +HOSTDEVICE int64_t& indexer(Dim& dim, int idx) { #ifndef __CUDA_ARCH__ if (idx < 0) { throw std::invalid_argument("Tried to access a negative dimension"); @@ -141,7 +141,7 @@ HOSTDEVICE int& indexer(Dim& dim, int idx) { } template <> -HOSTDEVICE int& indexer<1>(Dim<1>& dim, int idx) { +HOSTDEVICE int64_t& indexer<1>(Dim<1>& dim, int idx) { #ifndef __CUDA_ARCH__ if (idx != 0) { throw std::invalid_argument("Invalid index"); @@ -153,7 +153,7 @@ HOSTDEVICE int& indexer<1>(Dim<1>& dim, int idx) { } template -HOSTDEVICE int indexer(const Dim& dim, int idx) { +HOSTDEVICE int64_t indexer(const Dim& dim, int idx) { #ifndef __CUDA_ARCH__ if (idx < 0) { throw std::invalid_argument("Tried to access a negative dimension"); @@ -168,7 +168,7 @@ HOSTDEVICE int indexer(const Dim& dim, int idx) { } template <> -HOSTDEVICE int indexer<1>(const Dim<1>& dim, int idx) { +HOSTDEVICE int64_t indexer<1>(const Dim<1>& dim, int idx) { #ifndef __CUDA_ARCH__ if (idx != 0) { throw std::invalid_argument("Invalid index"); @@ -182,73 +182,76 @@ HOSTDEVICE int indexer<1>(const Dim<1>& dim, int idx) { } // namespace // Static access to constant Dim template -HOSTDEVICE int get(const Dim& d) { +HOSTDEVICE int64_t get(const Dim& d) { return DimGetter::impl(d); } // Static access to mutable Dim template -HOSTDEVICE int& get(Dim& d) { +HOSTDEVICE int64_t& get(Dim& d) { return DimGetter::impl(d); } // Dynamic access to constant Dim template -HOSTDEVICE int Dim::operator[](int i) const { +HOSTDEVICE int64_t Dim::operator[](int i) const { return indexer(*this, i); } // Dynamic access to mutable Dim template -HOSTDEVICE int& Dim::operator[](int i) { +HOSTDEVICE int64_t& Dim::operator[](int i) { return indexer(*this, i); } // Dynamic access to constant Dim -inline HOSTDEVICE int Dim<1>::operator[](int i) const { +inline HOSTDEVICE int64_t Dim<1>::operator[](int i) const { return indexer(*this, i); } // Dynamic access to mutable Dim -inline HOSTDEVICE int& Dim<1>::operator[](int i) { return indexer(*this, i); } +inline HOSTDEVICE int64_t& Dim<1>::operator[](int i) { + return indexer(*this, i); +} // Dynamic access to constant Dim // without std::enable_if will try to instantiate this on get<0>(d) template -HOSTDEVICE typename std::enable_if<(l > 0), int>::type get(const Dim& d, - int i) { +HOSTDEVICE typename std::enable_if<(l > 0), int64_t>::type get(const Dim& d, + int i) { return d[i]; } // Dynamic access to mutable Dim template -HOSTDEVICE typename std::enable_if<(l > 0), int&>::type get(Dim& d, int i) { +HOSTDEVICE typename std::enable_if<(l > 0), int64_t&>::type get(Dim& d, + int i) { return d[i]; } // Dot product of two dims template -HOSTDEVICE int linearize(const Dim& a, const Dim& b) { +HOSTDEVICE int64_t linearize(const Dim& a, const Dim& b) { return a.head * b.head + linearize(a.tail, b.tail); } // Base case dot product of two Dims // Notice it is inline because it is no longer a template template <> -HOSTDEVICE inline int linearize(const Dim<1>& a, const Dim<1>& b) { +HOSTDEVICE inline int64_t linearize(const Dim<1>& a, const Dim<1>& b) { return a.head * b.head; } // Product of a Dim template -HOSTDEVICE int product(const Dim& a, int prod = 1) { +HOSTDEVICE int64_t product(const Dim& a, int prod = 1) { return prod * a.head * product(a.tail); } // Base case product of a Dim // Notice it is inline because it is no longer a template template <> -HOSTDEVICE inline int product(const Dim<1>& a, int prod) { +HOSTDEVICE inline int64_t product(const Dim<1>& a, int prod) { return prod * a.head; } diff --git a/paddle/framework/dim_test.cu b/paddle/framework/dim_test.cu index 3898d0a447aa502813b3cb5e86c29eebb814ff84..0a6a87669c900de6cb507dd48f0cfc871defe279 100644 --- a/paddle/framework/dim_test.cu +++ b/paddle/framework/dim_test.cu @@ -8,7 +8,7 @@ __global__ void test(paddle::framework::Dim<2>* o) { o[0] = paddle::framework::make_dim(5, 6); } -__global__ void dyn_idx_gpu(int* o) { +__global__ void dyn_idx_gpu(int64_t* o) { auto d = paddle::framework::make_dim(5, 6); o[0] = d[1]; } @@ -47,9 +47,9 @@ TEST(Dim, Equality) { EXPECT_EQ(b[1], 11); // dynamic access on GPU - thrust::device_vector r(1); + thrust::device_vector r(1); dyn_idx_gpu<<<1, 1>>>(thrust::raw_pointer_cast(r.data())); - int res = r[0]; + int64_t res = r[0]; EXPECT_EQ(res, 6); // ex_prefix_mul diff --git a/paddle/framework/eigen.h b/paddle/framework/eigen.h index a4667cc51fadfc020d3211b7a82356db386fced1..54bbeafcabdeeb1e2c1017c156b3512c83dada3a 100644 --- a/paddle/framework/eigen.h +++ b/paddle/framework/eigen.h @@ -28,7 +28,7 @@ struct EigenDim { static Type From(const DDim& dims) { PADDLE_ENFORCE(arity(dims) == D, "D must match arity(DDim)"); Type ret; - for (int d = 0; d < arity(dims); d++) { + for (int64_t d = 0; d < arity(dims); d++) { ret[d] = dims[d]; } return ret; @@ -63,20 +63,35 @@ struct EigenTensor { template -struct EigenMatrix : public EigenTensor {}; +struct EigenMatrix : public EigenTensor { + static typename EigenMatrix::Type Reshape(Tensor& tensor, int num_col_dims) { + int rank = tensor.dims_.size(); + PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank, + "`num_col_dims` must be between (0, rank_of_tensor)."); + return EigenMatrix::From(tensor, + flatten_to_2d(tensor.dims(), num_col_dims)); + } + + static typename EigenMatrix::ConstType Reshape(const Tensor& tensor, + int num_col_dims) { + int rank = tensor.dims_.size(); + PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank, + "`num_col_dims` must be between (0, rank_of_tensor)."); + return EigenMatrix::From(tensor, + flatten_to_2d(tensor.dims(), num_col_dims)); + } +}; template struct EigenVector : public EigenTensor { // Flatten reshapes a Tensor into an EigenVector. static typename EigenVector::Type Flatten(Tensor& tensor) { - return EigenVector::From( - tensor, make_ddim({static_cast(product(tensor.dims_))})); + return EigenVector::From(tensor, {product(tensor.dims_)}); } static typename EigenVector::ConstType Flatten(const Tensor& tensor) { - return EigenVector::From( - tensor, make_ddim({static_cast(product(tensor.dims_))})); + return EigenVector::From(tensor, {product(tensor.dims_)}); } }; diff --git a/paddle/framework/eigen_test.cc b/paddle/framework/eigen_test.cc index dc1957691b1a202826e10e84c21ac8874df9e378..bc4a2db32cfba66bef2c444e1f822e0d2a57b91e 100644 --- a/paddle/framework/eigen_test.cc +++ b/paddle/framework/eigen_test.cc @@ -108,5 +108,24 @@ TEST(Eigen, Matrix) { } } +TEST(Eigen, MatrixReshape) { + Tensor t; + float* p = t.mutable_data({2, 3, 6, 4}, platform::CPUPlace()); + for (int i = 0; i < 2 * 3 * 6 * 4; ++i) { + p[i] = static_cast(i); + } + + EigenMatrix::Type em = EigenMatrix::Reshape(t, 2); + + ASSERT_EQ(2 * 3, em.dimension(0)); + ASSERT_EQ(6 * 4, em.dimension(1)); + + for (int i = 0; i < 2 * 3; i++) { + for (int j = 0; j < 6 * 4; j++) { + ASSERT_NEAR(i * 6 * 4 + j, em(i, j), 1e-6f); + } + } +} + } // namespace framework } // namespace paddle diff --git a/paddle/framework/framework.proto b/paddle/framework/framework.proto new file mode 100644 index 0000000000000000000000000000000000000000..dfcb5fb6210a08f35193b83e3b5f7cee92f618d7 --- /dev/null +++ b/paddle/framework/framework.proto @@ -0,0 +1,110 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +syntax = "proto2"; +package paddle.framework; + +enum AttrType { + INT = 0; + FLOAT = 1; + STRING = 2; + INTS = 3; + FLOATS = 4; + STRINGS = 5; + INT_PAIRS = 6; +} + +message IntPair { + required int32 first = 1; + required int32 second = 2; +}; + +// OpDesc describes an instance of a C++ framework::OperatorBase +// derived class type. +message OpDesc { + + message Attr { + required string name = 1; + required AttrType type = 2; + optional int32 i = 3; + optional float f = 4; + optional string s = 5; + repeated int32 ints = 6; + repeated float floats = 7; + repeated string strings = 8; + repeated IntPair int_pairs = 9; + }; + + message Var { + required string parameter = 1; + repeated string arguments = 2; + }; + + required string type = 3; + repeated Var inputs = 1; + repeated Var outputs = 2; + repeated Attr attrs = 4; +}; + +// OpProto describes a C++ framework::OperatorBase derived class. +message OpProto { + + // VarProto describes the C++ type framework::Variable. + message Var { + required string name = 1; + required string comment = 2; + + 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. + message Attr { + required string name = 1; + required AttrType type = 2; + required string comment = 3; + // If that attribute is generated, it means the Paddle third + // language binding has responsibility to fill that + // attribute. End-User should not set that attribute. + optional bool generated = 4 [ default = false ]; + } + + required string type = 1; + repeated Var inputs = 2; + repeated Var outputs = 3; + repeated Attr attrs = 4; + required string comment = 5; +} + +enum DataType { + BOOL = 0; + INT16 = 1; + INT32 = 2; + INT64 = 3; + FP16 = 4; + FP32 = 5; + FP64 = 6; +} + +message LoDTensorDesc { + required DataType data_type = 1; + repeated int32 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; +} diff --git a/paddle/framework/grad_op_builder.cc b/paddle/framework/grad_op_builder.cc index 8bd2bc590272256fed79f4ab38ad52b470e87012..b02a599a800668b22e7fe39a10fa6dc132e305bd 100644 --- a/paddle/framework/grad_op_builder.cc +++ b/paddle/framework/grad_op_builder.cc @@ -13,105 +13,45 @@ 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_proto.pb.h" #include "paddle/framework/op_registry.h" namespace paddle { namespace framework { - -typedef std::vector Ints; - enum class OpArgType { IN, OUT }; -const Ints* AttrFormat(const AttributeMap& attrs, const std::string& key) { - return (attrs.count(key) > 0) ? &boost::get(attrs.at(key)) : nullptr; -} - -Ints* AttrFormat(AttributeMap& attrs, const std::string& key) { - return (attrs.count(key) > 0) ? &boost::get(attrs.at(key)) : nullptr; -} - -static void TransOpArg(const OperatorBase* src_op, - std::vector& grad_inputs, - std::vector& grad_outputs, - AttributeMap& grad_attrs, - std::unordered_map& grad_idxs, - const std::string& src_type, const std::string& dst_type, - int& idx, bool is_grad) { - const std::vector& src_inout = - (src_type == "input_format") ? src_op->inputs_ : src_op->outputs_; - - const std::vector* src_format = AttrFormat(src_op->Attrs(), src_type); - - std::vector& dst_inout = - (dst_type == "input_format") ? grad_inputs : grad_outputs; - - std::vector* dst_format = AttrFormat(grad_attrs, dst_type); - - const OpProto& proto = OpRegistry::protos().at(src_op->type_); - +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 == "input_format") ? proto.inputs() : proto.outputs(); - + src_type == OpArgType::IN ? proto.inputs() : proto.outputs(); for (const auto& arg : src_arg_list) { - std::string src_name = arg.name(); - std::string dst_name = is_grad ? src_name + kGradVarSuffix : src_name; - grad_idxs[dst_name] = idx++; - int src_arg_idx = src_op->in_out_idxs_->at(src_name); - int src_begin = - src_format == nullptr ? src_arg_idx : src_format->at(src_arg_idx); - int src_end = src_format == nullptr ? src_arg_idx + 1 - : src_format->at(src_arg_idx + 1); - for (int i = src_begin; i < src_end; ++i) { - std::string s = - is_grad ? src_inout[i] + kGradVarSuffix - : (arg.ignore_gradient() ? kEmptyVarName : src_inout[i]); - dst_inout.emplace_back(s); - } - if (dst_format != nullptr) { - dst_format->push_back(dst_inout.size()); + 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) { - const std::string& grad_op_type = OpRegistry::grad_ops().at(op->Type()); - - AttributeMap grad_attrs(op->Attrs()); - grad_attrs.erase("input_format"); - grad_attrs.erase("output_format"); - if (op->Attrs().count("input_format") > 0) { - grad_attrs["output_format"] = std::vector({0}); - } - if (op->Attrs().count("input_format") > 0 || - op->Attrs().count("output_format") > 0) { - grad_attrs["input_format"] = std::vector({0}); - } - - std::vector grad_inputs, grad_outputs; - - using VarIndexMap = std::unordered_map; - VarIndexMap* grad_idxs = new VarIndexMap; - int in_idx = 0; - int out_idx = 0; - TransOpArg(op, grad_inputs, grad_outputs, grad_attrs, *grad_idxs, - "input_format", "input_format", in_idx, false); // I - TransOpArg(op, grad_inputs, grad_outputs, grad_attrs, *grad_idxs, - "output_format", "input_format", in_idx, false); // G - TransOpArg(op, grad_inputs, grad_outputs, grad_attrs, *grad_idxs, - "output_format", "input_format", in_idx, true); // OG - TransOpArg(op, grad_inputs, grad_outputs, grad_attrs, *grad_idxs, - "input_format", "output_format", out_idx, true); // IG - - OperatorBase* grad_op = OpRegistry::op_creators().at(grad_op_type)(); - - grad_op->type_ = grad_op_type; - grad_op->inputs_ = grad_inputs; - grad_op->outputs_ = grad_outputs; - grad_op->attrs_ = grad_attrs; - grad_op->in_out_idxs_.reset(grad_idxs); - - return grad_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 diff --git a/paddle/framework/grad_op_builder_test.cc b/paddle/framework/grad_op_builder_test.cc index 19e552b7458c966d473bdee99515a2beee1f6089..9e3ca563c6765637f8471d142d32cec447f0b977 100644 --- a/paddle/framework/grad_op_builder_test.cc +++ b/paddle/framework/grad_op_builder_test.cc @@ -3,29 +3,20 @@ #include "paddle/framework/op_registry.h" #include "paddle/framework/operator.h" -USE_OP(add_two); +USE_OP(add); namespace paddle { namespace framework { -class NOP : public OperatorBase { - public: - DEFINE_OPERATOR_CTOR(NOP, OperatorBase) - - void InferShape(const Scope &scope) const override {} - void Run(const Scope &scope, - const platform::DeviceContext &dev_ctx) const override {} -}; - 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").SetMultiple(); + AddInput("In2_mult", "a multiple input").AsDuplicable(); AddInput("In3", "another single input"); AddOutput("Out1", "a single output"); - AddOutput("Out2_mult", "a multiple output").SetMultiple(); + AddOutput("Out2_mult", "a multiple output").AsDuplicable(); AddComment("test op with multiple inputs and outputs"); } }; @@ -35,10 +26,10 @@ class IOIgnoredOpMaker : public OpProtoAndCheckerMaker { IOIgnoredOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("In1", "a single input"); - AddInput("In2_mult", "a multiple input").SetMultiple().IgnoreGradient(); - AddInput("In3_mult", "another multiple input").SetMultiple(); - AddOutput("Out1_mult", "a multiple output").SetMultiple(); - AddOutput("Out2", "a single output").IgnoreGradient(); + 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"); } }; @@ -49,35 +40,33 @@ class IOIgnoredOpMaker : public OpProtoAndCheckerMaker { namespace f = paddle::framework; TEST(GradOpBuilder, AddTwo) { - std::shared_ptr add_op( - f::OpRegistry::CreateOp("add_two", {"x", "y"}, {"out"}, {})); + 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(static_cast(grad_add_op->inputs_.size()), 4); - EXPECT_EQ(static_cast(grad_add_op->outputs_.size()), 2); + 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("Out@GRAD"), "out@GRAD"); - EXPECT_EQ(grad_add_op->Output("X@GRAD"), "x@GRAD"); - EXPECT_EQ(grad_add_op->Output("Y@GRAD"), "y@GRAD"); + 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); -REGISTER_GRADIENT_OP(mult_io, mult_io_grad, f::NOP); -REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker); -REGISTER_GRADIENT_OP(io_ignored, io_ignored_grad, f::NOP); +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) { - f::AttributeMap attrs{{"input_format", std::vector{0, 1, 4, 5}}, - {"output_format", std::vector{0, 1, 3}}}; std::shared_ptr test_op(f::OpRegistry::CreateOp( - "mult_io", {"in1", "in2_1", "in2_2", "in2_3", "in3"}, - {"out1", "out2_1", "out2_2"}, attrs)); + "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(), 5UL + 3UL + 3UL); + 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"})); @@ -91,7 +80,7 @@ TEST(GradOpBuilder, MutiInOut) { std::vector( {f::GradVarName("out2_1"), f::GradVarName("out2_2")})); - ASSERT_EQ(grad_test_op->outputs_.size(), 5UL); + 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"), @@ -101,31 +90,28 @@ TEST(GradOpBuilder, MutiInOut) { } TEST(GradOpBuilder, IOIgnoredInGradient) { - f::AttributeMap attrs{{"input_format", std::vector{0, 1, 3, 5}}, - {"output_format", std::vector{0, 2, 3}}}; std::shared_ptr test_op(f::OpRegistry::CreateOp( - "io_ignored", {"in1", "in2_1", "in2_2", "in3_1", "in3_2"}, - {"out1_1", "out1_2", "out2"}, attrs)); + "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(), 5UL + 3UL + 3UL); + ASSERT_EQ(grad_test_op->Inputs().size(), 2UL + 1UL + 2UL); EXPECT_EQ(grad_test_op->Input("In1"), "in1"); - EXPECT_EQ(grad_test_op->Inputs("In2_mult"), - std::vector({f::kEmptyVarName, f::kEmptyVarName})); 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->Input("Out2"), f::kEmptyVarName); 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(), 5UL); + 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( diff --git a/paddle/framework/images/duplicate_op2.graffle b/paddle/framework/images/duplicate_op2.graffle index 2b658085d6a55d368c320051ba7f94ec2900f13c..ede3bca30ae17d5af52505fd94dc2f79b23b57e0 100644 Binary files a/paddle/framework/images/duplicate_op2.graffle and b/paddle/framework/images/duplicate_op2.graffle differ diff --git a/paddle/framework/images/duplicate_op2.png b/paddle/framework/images/duplicate_op2.png index c5588015d1450fd8c1bda3580680d884494868bb..4e872dc2caf3b0cbd0d5176f11a14801b538dc86 100644 Binary files a/paddle/framework/images/duplicate_op2.png and b/paddle/framework/images/duplicate_op2.png differ diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc index 2b178907747b3911292b070b65160a24c120b726..908a1f2fd0abe0aa4016c72dbcbc18dcc144232c 100644 --- a/paddle/framework/lod_tensor.cc +++ b/paddle/framework/lod_tensor.cc @@ -19,25 +19,24 @@ namespace paddle { namespace framework { -LODTensor::LOD LODTensor::LOD::SliceLevels(size_t level_begin, - size_t level_end) const { - LOD new_lod; +LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) { + LoD new_lod; new_lod.reserve(level_end - level_begin); for (size_t i = level_begin; i < level_end; i++) { - new_lod.emplace_back(at(i)); + new_lod.emplace_back(in.at(i)); } return new_lod; } -LODTensor::LOD LODTensor::LOD::SliceInLevel(size_t level, size_t elem_begin, - size_t elem_end) const { +LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin, + size_t elem_end) { // slice the lod. - LOD new_lod; - new_lod.reserve(size() - level); - auto start = this->at(level)[elem_begin]; - auto end = this->at(level)[elem_end]; + LoD new_lod; + new_lod.reserve(in.size() - level); + auto start = in.at(level)[elem_begin]; + auto end = in.at(level)[elem_end]; - for (auto it = this->begin() + level; it != this->end(); it++) { + for (auto it = in.begin() + level; it != in.end(); it++) { auto it_begin = std::find(it->begin(), it->end(), start); auto it_end = std::find(it_begin, it->end(), end); PADDLE_ENFORCE(it_begin != it->end(), "error in parsing lod info"); @@ -47,13 +46,13 @@ LODTensor::LOD LODTensor::LOD::SliceInLevel(size_t level, size_t elem_begin, std::transform(new_lod.back().begin(), new_lod.back().end(), new_lod.back().begin(), [start](int v) { return v - start; }); - PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LOD"); + PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LoD"); } - PADDLE_ENFORCE_LE(new_lod.size(), this->size()); + PADDLE_ENFORCE_LE(new_lod.size(), in.size()); return new_lod; } -bool operator==(const LODTensor::LOD& a, const LODTensor::LOD& b) { +bool operator==(const LoD& a, const LoD& b) { if (a.size() != b.size()) { return false; } @@ -70,9 +69,27 @@ bool operator==(const LODTensor::LOD& a, const LODTensor::LOD& b) { } } } - return true; } +void LoDTensor::SliceLevels(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)); + + auto new_lod = framework::SliceInLevel(lod_, level, elem_begin, elem_end); + lod_ = new_lod; +} + } // namespace framework } // namespace paddle diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h index 0290ec09b446acc5d0ca8f604a730e84a5bf844c..154068fef69bc96edbd85b731fe8091b3b1ff823 100644 --- a/paddle/framework/lod_tensor.h +++ b/paddle/framework/lod_tensor.h @@ -15,7 +15,7 @@ #pragma once #include -#if (!PADDLE_ONLY_CPU) +#ifndef PADDLE_ONLY_CPU #include #include #endif @@ -27,36 +27,42 @@ namespace paddle { namespace framework { +#ifdef PADDLE_ONLY_CPU +template +using Vector = std::vector; +#else +template +using Vector = thrust::host_vector; +#endif + +using LoD = std::vector>; + +LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end); + +LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin, + size_t elem_end); + +bool operator==(const LoD& a, const LoD& b); + /* - * LODTensor (Level of details Tensor) + * LoDTensor (Level of details Tensor) * see https://en.wikipedia.org/wiki/Level_of_details for reference. */ -class LODTensor : public Tensor { +class LoDTensor { public: -// Level save offsets of each unit. -#ifdef PADDLE_ONLY_CPU - template - using Vector = std::vector; -#else - template - using Vector = thrust::host_vector; -#endif - // LoD stores offsets of each level of units, the largest units level first, - // then the smaller units level. Each Level stores the offsets of units in - // Tesor. - class LOD : public std::vector> { - public: - LOD SliceLevels(size_t level_begin, size_t level_end) const; - LOD SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) const; - }; + LoDTensor() {} + LoDTensor(const LoD& lod, Tensor* t) : lod_(lod), tensor_(t) {} + + void set_lod(const LoD& lod) { lod_ = lod; } - LODTensor() {} - explicit LODTensor(const LOD &lod) : lod_(lod) {} + void set_tensor(Tensor* tensor) { tensor_ = tensor; } - virtual Tensor *Clone() const { return new LODTensor(lod_); } + Tensor& tensor() { return *tensor_; } + + LoD lod() { return lod_; } /* - * Get a element from LOD. + * 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, @@ -68,7 +74,7 @@ class LODTensor : public Tensor { } /* - * Number of LODTensor's levels, each level has units of data, for example, + * Number of LoDTensor's levels, each level has units of data, for example, * in the sentence's view, article, paragraph, sentence are 3 levels. */ size_t NumLevels() const { return lod_.size(); } @@ -79,71 +85,23 @@ class LODTensor : public Tensor { PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level, NumLevels()); // the last offset is the end of last element - return lod_[level].size() - 1; + return (lod_)[level].size() - 1; } /* - * Slice of levels[level_begin:level_end], with tensor shared. + * Slice of levels[level_begin:level_end] */ - template - LODTensor SliceLevels(size_t level_begin, size_t level_end) const; + void SliceLevels(size_t level_begin, size_t level_end); /* - * Slice of elements of a level, [elem_begin: elem_end], with tensor shared. + * Slice of elements of a level, [elem_begin: elem_end] * @note: low performance in slice lod_. */ - template - LODTensor SliceInLevel(size_t level, size_t elem_begin, - size_t elem_end) const; - - /* - * Copy other's lod_'s content, free to mutate. - */ - void CopyLOD(const LODTensor &other) { lod_ = other.lod_; } - /* - * Determine whether LODTensor has a valid LOD info. - */ - const LOD &lod() const { return lod_; } - LOD *mutable_lod() { return &lod_; } - - virtual ~LODTensor() {} + void SliceInLevel(size_t level, size_t elem_begin, size_t elem_end); private: - LOD lod_; + LoD lod_; + Tensor* tensor_; // not owned }; - -bool operator==(const LODTensor::LOD &a, const LODTensor::LOD &b); - -template -LODTensor LODTensor::SliceLevels(size_t level_begin, size_t level_end) const { - auto new_lod = lod_.SliceLevels(level_begin, level_end); - // slice levels just need to update LOD info, each level will contains the - // whole tensor_, so no need to modify tensor_. - LODTensor new_tensor(new_lod); - new_tensor.ShareDataWith(*this); - return new_tensor; -} - -template -LODTensor LODTensor::SliceInLevel(size_t level, size_t elem_begin, - size_t elem_end) const { - 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)); - - auto new_lod = lod_.SliceInLevel(level, elem_begin, elem_end); - - // slice elements just need to update LOD info, because offsets are not - // changed, so the original tensor_ can be reused. - LODTensor new_tensor(new_lod); - new_tensor.ShareDataWith(*this); - return new_tensor; -} - } // namespace framework } // namespace paddle diff --git a/paddle/framework/lod_tensor.md b/paddle/framework/lod_tensor.md new file mode 100644 index 0000000000000000000000000000000000000000..769b61f175a2f462258c1242d027c04c0abd12a9 --- /dev/null +++ b/paddle/framework/lod_tensor.md @@ -0,0 +1,146 @@ +# 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. + +## Challenge of Variable-length Inputs + +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. + +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 + +### Mini-Batch of variable-length sentenses + +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: + +``` + 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. + +### Mini-Batch of variable-length 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: + +``` + 3 +3 1 2 +口口口 口 口口 +``` + +where each `口` represents an image. + +### Mini-Batch of fixed-size 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 + +``` + 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 summary, as long as that the essential elements (words or images) have the same size, we can represent mini-batches by a LoD Tensor: + +- 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: + + ```c++ + typedef std::vector > LoD; + ``` + +- 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 + +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, + +``` + 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** + +For example, the <2,1>-slice of above slice is + +``` +2 +|| +``` + +and the <1,2>-slice of above example is + +``` +2 +2 3 +|| ||| +``` + +Let's go on slicing this slice. Its <1,1>-slice is + +``` +3 +||| +``` + +### The Slicing Algorithm + +The algorithm, with over-simplified data structure, is defined as + +```c++ +typedef vector > LoD; + +struct LoDTensor { + LoD lod_; + float* tensor_; +}; + +LoDTensor Slice(const LoDTensor& lodt, int level, int sequence); +``` + +Let us revisit the example above + +``` + 3 +3 1 2 +3 2 4 1 2 3 +||| || |||| | || ||| +``` + +Suppose that we want to retrieve the <1,2>-slice + +``` +2 +2 3 +|| ||| +``` + +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. + +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 + +``` + 0 +0 9 10 +0 3 5 9 10 12 +||| || |||| | || ||| +``` + +We don't really need the 0 on top, so the LoD Tensor could be + +``` +0 9 10 +0 3 5 9 10 12 +||| || |||| | || ||| +``` diff --git a/paddle/framework/lod_tensor_test.cc b/paddle/framework/lod_tensor_test.cc index 2881136ced6ef957a192e303e529b9b2867b3dda..1da8553134f377f7a4fbe8008d12fe8d4a0e47f4 100644 --- a/paddle/framework/lod_tensor_test.cc +++ b/paddle/framework/lod_tensor_test.cc @@ -21,16 +21,15 @@ namespace paddle { namespace framework { -class LODTensorTester : public ::testing::Test { +class LoDTensorTester : public ::testing::Test { public: virtual void SetUp() override { - lod_tensor.reset(new LODTensor); // tensor's batch_size: 30 // 3 levels // 0 10 20 // 0 5 10 15 20 // 0 2 5 7 10 12 15 20 - LODTensor::LOD lod; + LoD lod; lod.push_back(std::vector{0, 10, 20}); lod.push_back(std::vector{0, 5, 10, 15, 20}); lod.push_back(std::vector{0, 2, 5, 7, 10, 12, 15, 17, 20}); @@ -41,75 +40,65 @@ class LODTensorTester : public ::testing::Test { // malloc memory tensor.mutable_data(place); - lod_tensor.reset(new LODTensor(lod)); - lod_tensor->Resize({20 /*batch size*/, 128 /*dim*/}); - - lod_tensor->ShareDataWith(tensor); - // lod_tensor->ShareDataWith(tensor); + lod_tensor.set_lod(lod); + lod_tensor.set_tensor(&tensor); } protected: - std::unique_ptr lod_tensor; platform::CPUPlace place; Tensor tensor; + LoDTensor lod_tensor; }; -TEST_F(LODTensorTester, NumLevels) { ASSERT_EQ(lod_tensor->NumLevels(), 3UL); } +TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); } -TEST_F(LODTensorTester, NumElements) { - ASSERT_EQ(lod_tensor->NumElements(0), 2UL); - ASSERT_EQ(lod_tensor->NumElements(1), 4UL); - ASSERT_EQ(lod_tensor->NumElements(2), 8UL); +TEST_F(LoDTensorTester, NumElements) { + ASSERT_EQ(lod_tensor.NumElements(0), 2UL); + ASSERT_EQ(lod_tensor.NumElements(1), 4UL); + ASSERT_EQ(lod_tensor.NumElements(2), 8UL); } -TEST_F(LODTensorTester, SliceLevels) { +TEST_F(LoDTensorTester, SliceLevels) { // slice 1 level for (size_t level = 0; level < 3UL; ++level) { - auto new_lod_tensor = lod_tensor->SliceLevels(level, level + 1); + LoDTensor new_lod_tensor = lod_tensor; + new_lod_tensor.SliceLevels(level, level + 1); ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL); - ASSERT_EQ(new_lod_tensor.NumElements(0UL), lod_tensor->NumElements(level)); - // ASSERT_EQ(new_lod_tensor, *lod_tensor); + ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level)); + ASSERT_EQ(new_lod_tensor.tensor().data(), + lod_tensor.tensor().data()); } // slice 2 level for (size_t level = 0; level < 2UL; ++level) { - auto new_lod_tensor = lod_tensor->SliceLevels(level, level + 2); + LoDTensor new_lod_tensor = lod_tensor; + new_lod_tensor.SliceLevels(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), - lod_tensor->NumElements(level + 1)); - ASSERT_EQ(new_lod_tensor.data(), lod_tensor->data()); + ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level)); + ASSERT_EQ(new_lod_tensor.NumElements(1), lod_tensor.NumElements(level + 1)); + ASSERT_EQ(new_lod_tensor.tensor().data(), + lod_tensor.tensor().data()); } } -TEST_F(LODTensorTester, SliceInLevel) { +TEST_F(LoDTensorTester, SliceInLevel) { size_t level = 0; - auto new_lod_tensor = lod_tensor->SliceInLevel(level, 0, 2); + LoDTensor new_lod_tensor = lod_tensor; + new_lod_tensor.SliceInLevel(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); EXPECT_EQ(new_lod_tensor.NumElements(2), 8UL); - ASSERT_EQ(new_lod_tensor.data(), lod_tensor->data()); + ASSERT_EQ(new_lod_tensor.tensor().data(), + lod_tensor.tensor().data()); level = 1; - new_lod_tensor = lod_tensor->SliceInLevel(level, 0, 2); + new_lod_tensor = lod_tensor; + new_lod_tensor.SliceInLevel(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); - ASSERT_EQ(new_lod_tensor.data(), lod_tensor->data()); -} - -TEST_F(LODTensorTester, ShareLOD) { - LODTensor new_lod_tensor; - new_lod_tensor.CopyLOD(*lod_tensor); - ASSERT_EQ(new_lod_tensor.lod(), lod_tensor->lod()); -} - -TEST_F(LODTensorTester, CopyLOD) { - LODTensor new_lod_tensor; - new_lod_tensor.CopyLOD(*lod_tensor); - bool equals = std::equal(lod_tensor->lod().begin(), lod_tensor->lod().end(), - new_lod_tensor.lod().begin()); - ASSERT_TRUE(equals); + ASSERT_EQ(new_lod_tensor.tensor().data(), + lod_tensor.tensor().data()); } } // namespace framework diff --git a/paddle/framework/op_desc.proto b/paddle/framework/op_desc.proto deleted file mode 100644 index d95ba26f88ae181f991440e0df30c80f80a7eb2a..0000000000000000000000000000000000000000 --- a/paddle/framework/op_desc.proto +++ /dev/null @@ -1,56 +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. */ - -syntax = "proto2"; -package paddle.framework; - -import "attribute.proto"; - -// AttrDesc is used to describe Attributes of an Operator. It contain's -// name, type, and value of Attribute. -// -// e.g, for scale=3.0: name=scala, type=AttrType.FLOAT, value=3.0 -message AttrDesc { - required string name = 1; - required AttrType type = 2; - optional int32 i = 3; - optional float f = 4; - optional string s = 5; - repeated int32 ints = 6; - repeated float floats = 7; - repeated string strings = 8; -}; - -// Protocol Message to describe an Operator. -// -// In PaddlePaddle, Operator is used to do a certain computation such -// as "add", "sub", "cosine", etc. -// (1) Operator needs to know the input and output variable names. -// (2) Some ops may have special attributes such as "scale" in "CosineOp". -// -// 3rd-party language can build this proto message and call -// AddOp(const OpDesc& op_desc) of Paddle core to create an Operator. -message OpDesc { - // input names of this Operator. - repeated string inputs = 1; - - // output names of this Operator. - repeated string outputs = 2; - - // type of this Operator, such as "add", "sub", "fc". - required string type = 3; - - // Attributes of this Operator. e.g., scale=3.0 in cosine op. - repeated AttrDesc attrs = 4; -}; \ No newline at end of file diff --git a/paddle/framework/op_desc_test.cc b/paddle/framework/op_desc_test.cc deleted file mode 100644 index d0c52523b64725ee11c281b086f9ffed6a09e787..0000000000000000000000000000000000000000 --- a/paddle/framework/op_desc_test.cc +++ /dev/null @@ -1,35 +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 -#include - -TEST(OpDesc, Create) { - paddle::framework::OpDesc op_desc; - op_desc.set_type("add"); - op_desc.add_inputs("X"); - op_desc.add_inputs("Y"); - op_desc.add_outputs("Z"); - - auto attr = op_desc.mutable_attrs()->Add(); - attr->set_type(paddle::framework::AttrType::FLOAT); - attr->set_f(3.14); - - // required field name is not set, so IsInitialized should be false. - ASSERT_FALSE(op_desc.IsInitialized()); - - attr->set_name("add"); - // after all required fields are set, IsInitialized should be true now. - ASSERT_TRUE(op_desc.IsInitialized()); -} \ No newline at end of file diff --git a/paddle/framework/op_info.cc b/paddle/framework/op_info.cc new file mode 100644 index 0000000000000000000000000000000000000000..81ba29797c5f478e5d6a91236f3e8de1e6b43e49 --- /dev/null +++ b/paddle/framework/op_info.cc @@ -0,0 +1,29 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/framework/op_info.h" + +namespace paddle { +namespace framework { + +static OpInfoMap* g_op_info_map = nullptr; + +OpInfoMap& OpInfoMap::Instance() { + if (g_op_info_map == nullptr) { + g_op_info_map = new OpInfoMap(); + } + return *g_op_info_map; +} +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/op_info.h b/paddle/framework/op_info.h new file mode 100644 index 0000000000000000000000000000000000000000..b98d8f23a14cf6fbe787953ad16b5c9ab99222ad --- /dev/null +++ b/paddle/framework/op_info.h @@ -0,0 +1,111 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include +#include +#include +#include + +#include "paddle/framework/attribute.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_; + + bool HasOpProtoAndChecker() const { + return proto_ != nullptr && checker_ != nullptr; + } + + const OpProto& Proto() const { + PADDLE_ENFORCE_NOT_NULL(proto_, "Operator Proto has not been registered"); + PADDLE_ENFORCE(proto_->IsInitialized(), + "Operator Proto must be initialized in op info"); + 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(); } +}; + +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(); + } + + void Insert(const std::string& type, const OpInfo& info) { + PADDLE_ENFORCE(!Has(type), "Operator %s has been registered", type); + map_.insert({type, info}); + } + + const OpInfo& Get(const std::string& type) const { + auto op_info_ptr = GetNullable(type); + PADDLE_ENFORCE_NOT_NULL(op_info_ptr, "Operator %s has not been registered", + type); + return *op_info_ptr; + } + + const OpInfo* GetNullable(const std::string& type) const { + auto it = map_.find(type); + if (it == map_.end()) { + return nullptr; + } else { + return &it->second; + } + } + + template + void IterAllInfo(Callback callback) { + for (auto& it : map_) { + callback(it.first, it.second); + } + } + + private: + OpInfoMap() = default; + std::unordered_map map_; +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/op_proto.proto b/paddle/framework/op_proto.proto deleted file mode 100644 index 52292162874b9ca207fb0d3917df41ade096b143..0000000000000000000000000000000000000000 --- a/paddle/framework/op_proto.proto +++ /dev/null @@ -1,116 +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. */ - -// Protocol Message for 3rd-party language binding. -// -// Paddle Python package will use `OpProto` to generate op creation methods. -// The op creation methods take user's input and generate `OpDesc` proto -// message, -// then pass `OpDesc` to C++ side and create Op pointer. -// -syntax = "proto2"; -package paddle.framework; - -import "attribute.proto"; - -// Attribute protocol message for 3rd-party language binding. -// It will store the Op support what attribute and what type. -message AttrProto { - // Supported attribute name. e.g. `scale` for cosine op. - required string name = 1; - - // Supported attribute type. - required AttrType type = 2; - - // Supported attribute comments. It helps 3rd-party language generate - // doc-string. - required string comment = 3; - - // If that attribute is generated, it means the Paddle third language - // binding has responsibility to fill that attribute. End-User should - // not set that attribute. - optional bool generated = 4 [ default = false ]; -} - -// Input or output message for 3rd-party language binding. -// It contains parameter name and its comments. -message VarProto { - // Input or output name in that op creation function. - // e.g. `cos(a, b, output, ...)`, "a", "b", "output" are names. - required string name = 1; - - // The comment for that input. It helps 3rd-party language generate - // doc-string. - required string comment = 2; - - // Is that input/output could be a list or not. - // If so, that Op should write a attributed named `input_format` or - // `output_format`. - // - // e.g. - // If the op is a fc op, the inputs are `X`, `W`, `b`. The `X` and `W` - // could be multiple, so the multiple of `X` and `W` is True, and OpDesc - // will hold a attribute of them. - // - // The Op desc of same fc could be - // { - // "type": "fc", - // "input": ["X1", "X2", "W1", "W2", "b"], - // "output": "fc.out", - // "attrs" : { - // "input_format": [0, 2, 4, 5] - // } - // } - // - optional bool multiple = 3 [ default = false ]; - - // It marks that output is a temporary output. That output is not used by - // user, but used by other op internally as input. If other op is not use - // that output, it could be optimized early. - // - // Attribute temporary_index will be set in OpDesc if there is some - // outputs are temporary. - // - // output = [ "xxx.out1", "xxx.tmp", "xxx.out2"], - // attrs = { - // "temporary_index": [1] - // } - optional bool temporary = 4 [ default = false ]; - - // The gradient of operator can be ignored immediately - // e.g. operator AddOp, y = x1 + x2, the gradient of dy/dx1, dy/dx2 - // can be ignored for the future optimized on graph. - optional bool ignore_gradient = 6; -} - -// Op protocol message for 3rd-party language binding. -// It contains all information for generating op creation method. -message OpProto { - // The input information to generate op creation method. - repeated VarProto inputs = 1; - - // The output information to generate op creation method. - repeated VarProto outputs = 2; - - // The attribute information to generate op creation method. - repeated AttrProto attrs = 3; - - // The comments for that Op. It helps 3rd-party language generate - // doc-string. The whole documentation of that Op is generated by comment, - // inputs, outputs, attrs together. - required string comment = 4; - - // The type of that Op. - required string type = 5; -} diff --git a/paddle/framework/op_proto_test.cc b/paddle/framework/op_proto_test.cc deleted file mode 100644 index 9c054bde44e77571330cbc59074705f0cfc1cfb6..0000000000000000000000000000000000000000 --- a/paddle/framework/op_proto_test.cc +++ /dev/null @@ -1,31 +0,0 @@ -#include -#include - -TEST(TestOpProto, ALL) { - paddle::framework::OpProto proto; - { - auto ipt = proto.mutable_inputs()->Add(); - *ipt->mutable_name() = "a"; - *ipt->mutable_comment() = "the one input of cosine op"; - } - { - auto ipt = proto.mutable_inputs()->Add(); - *ipt->mutable_name() = "b"; - *ipt->mutable_comment() = "the other input of cosine op"; - } - { - auto opt = proto.mutable_outputs()->Add(); - *opt->mutable_name() = "output"; - *opt->mutable_comment() = "the output of cosine op"; - } - { - auto attr = proto.mutable_attrs()->Add(); - *attr->mutable_name() = "scale"; - attr->set_type(paddle::framework::AttrType::FLOAT); - *attr->mutable_comment() = "the scale attribute of cosine op"; - } - proto.set_type("cos"); - *proto.mutable_comment() = "cosine op, output = scale * cos(a, b)"; - - ASSERT_TRUE(proto.IsInitialized()); -} \ No newline at end of file diff --git a/paddle/framework/op_registry.cc b/paddle/framework/op_registry.cc index 1caa02a2a1d046778f875d04eeaef957be741302..b0e85dd49f97da4a7f889fde0b5f060954947be8 100644 --- a/paddle/framework/op_registry.cc +++ b/paddle/framework/op_registry.cc @@ -17,5 +17,45 @@ limitations under the License. */ #include namespace paddle { -namespace framework {} // namespace framework +namespace framework { + +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); + auto op = info.Creator()(type, inputs, outputs, attrs); + return std::unique_ptr(op); +} + +static VariableNameMap ConvertOpDescVarsToVarNameMap( + const google::protobuf::RepeatedPtrField& op_desc_vars) { + VariableNameMap ret_val; + for (auto& var : op_desc_vars) { + auto& var_names = ret_val[var.parameter()]; + auto& var_names_in_proto = var.arguments(); + var_names.reserve(static_cast(var_names_in_proto.size())); + std::copy(var_names_in_proto.begin(), var_names_in_proto.end(), + std::back_inserter(var_names)); + } + return ret_val; +} + +std::unique_ptr OpRegistry::CreateOp(const OpDesc& op_desc) { + VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs()); + VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs()); + AttributeMap attrs; + for (auto& attr : op_desc.attrs()) { + attrs[attr.name()] = GetAttrValue(attr); + } + + 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)); +} + +} // namespace framework } // namespace paddle diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h index cb9164eec1788c2c19176115e8687bed49d8c0b6..572dff860a306bb03ba9e6702fec85e4a2ea1b54 100644 --- a/paddle/framework/op_registry.h +++ b/paddle/framework/op_registry.h @@ -17,320 +17,85 @@ limitations under the License. */ #include #include #include +#include #include #include #include "paddle/framework/attribute.h" +#include "paddle/framework/framework.pb.h" #include "paddle/framework/grad_op_builder.h" -#include "paddle/framework/op_desc.pb.h" +#include "paddle/framework/op_info.h" +#include "paddle/framework/operator.h" #include "paddle/framework/scope.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) {} - - ~OpProtoAndCheckerMaker() { - PADDLE_ENFORCE(validated_, "should call Validate after build"); - } - - void Validate() { - validated_ = true; - CheckNoDuplicatedInOutAttrs(); - } - - protected: - struct VariableBuilder { - VarProto* var_; - std::function on_multiple_; - std::function on_temporary_; - - VariableBuilder& SetMultiple() { - var_->set_multiple(true); - on_multiple_(); - return *this; - } - - VariableBuilder& SetTemporary() { - PADDLE_ENFORCE(bool(on_temporary_), "Cannot set temporary"); - var_->set_temporary(true); - on_temporary_(); - return *this; - } - - VariableBuilder& IgnoreGradient() { - var_->set_ignore_gradient(true); - return *this; - } - }; - - VariableBuilder AddInput(const std::string& name, - const std::string& comment) { - VarProto* input = proto_->add_inputs(); - input->set_name(name); - input->set_comment(comment); - return VariableBuilder{input, [=] { this->SetHasMultipleInput(); }, - nullptr}; - } - - VariableBuilder AddOutput(const std::string& name, - const std::string& comment) { - VarProto* output = proto_->add_outputs(); - output->set_name(name); - output->set_comment(comment); - return VariableBuilder{output, [=] { this->SetHasMultipleOutput(); }, - [=] { this->SetHasTemporaryOutput(); }}; - } - - template - TypedAttrChecker& AddAttr(const std::string& name, - const std::string& comment, - bool generated = false) { - AttrProto* 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 SetHasMultiple(const std::string& in_out, bool* flag) { - if (!*flag) { - AddAttr>(in_out + "_format", - "The multiple index of " + in_out + - "\n" - R"DOC( -This attribute is used by Paddle core framework. Paddle's Op support each input -or output could be a list of variable. This attribute is used to show how that -list organized. - -e.g. - input = ["a", "b", "c", "d", "e", "f"] - input_format = [0, 4, 5, 6] - -means - The number of all input variables this op is six, and they are segmented into - three inputs. - - The first input is input[0:4], second is input[4:5], third is input[5:6]. -)DOC", - /*generated*/ true); - *flag = true; - } - } - - void SetHasMultipleInput() { SetHasMultiple("input", &has_multiple_input_); } - void SetHasMultipleOutput() { - SetHasMultiple("output", &has_multiple_output_); - } - - void SetHasTemporaryOutput() { - if (!has_temporary_output_) { - AddAttr>("temporary_index", - R"DOC(The temporary index of output. - -Not all output of Paddle Op is used by user. For faster computation, each op -could output some its internal state to other op, other op could take that -output to make compute faster. - -Add a mark to which output is temporary is helpful for future optimization. -)DOC", - /*generated*/ true) - .SetDefault(std::vector()); - has_temporary_output_ = true; - } - } - - void 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()); - } - } - - OpProto* proto_; - OpAttrChecker* op_checker_; - bool validated_{false}; - bool has_multiple_input_{false}; - bool has_multiple_output_{false}; - bool has_temporary_output_{false}; -}; - class OpRegistry { - using OpCreator = std::function; - using VarIndexMap = std::unordered_map; - using VarNameList = std::vector; - public: - template - static void RegisterOp(const std::string& op_type) { - op_creators()[op_type] = [] { return new OpType; }; - OpAttrChecker& op_checker = op_checkers()[op_type]; - OpProto& op_proto = protos()[op_type]; - auto maker = ProtoMakerType(&op_proto, &op_checker); - maker.Validate(); - op_proto.set_type(op_type); - PADDLE_ENFORCE( - op_proto.IsInitialized(), - "Fail to initialize %s's OpProto, because %s is not initialized", - op_type, op_proto.InitializationErrorString()); - - VarIndexMaps()[op_type].reset(new VarIndexMap()); - auto& varmap = *VarIndexMaps()[op_type]; - int idx = 0; - for (auto& var : op_proto.inputs()) { - varmap[var.name()] = idx++; - } - idx = 0; - for (auto& var : op_proto.outputs()) { - varmap[var.name()] = idx++; - } - } - - template - static void RegisterGradOp(const std::string& op_type, - const std::string& grad_op_type) { - op_creators()[grad_op_type] = [] { return new GradOpType; }; - grad_ops()[op_type] = grad_op_type; - } - - static std::shared_ptr CreateOp(const std::string& type, - const VarNameList& inputs, - const VarNameList& outputs, - const AttributeMap& attrs) { - auto op_create_it = op_creators().find(type); - PADDLE_ENFORCE(op_create_it != op_creators().end(), - "Operator %s cannot be found.", type); - - auto op = op_create_it->second(); - op->type_ = type; - op->inputs_ = inputs; - op->outputs_ = outputs; - - op->attrs_ = attrs; - op_checkers().at(type).Check(op->attrs_); - - GenerateTempVariableName(op); - - { - auto var_index_it = VarIndexMaps().find(type); - if (var_index_it != VarIndexMaps().end()) { - op->in_out_idxs_ = var_index_it->second; - } + 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; } - - op->Init(); - return std::shared_ptr(op); - } - - static std::shared_ptr CreateOp(const OpDesc& op_desc) { - std::vector inputs; - inputs.reserve((size_t)op_desc.inputs_size()); - std::copy(op_desc.inputs().begin(), op_desc.inputs().end(), - std::back_inserter(inputs)); - - std::vector outputs; - outputs.reserve((size_t)op_desc.outputs_size()); - std::copy(op_desc.outputs().begin(), op_desc.outputs().end(), - std::back_inserter(outputs)); - - AttributeMap attrs; - for (auto& attr : op_desc.attrs()) { - attrs[attr.name()] = GetAttrValue(attr); + OpInfoMap::Instance().Insert(op_type, op_info); + // register gradient op + if (!grad_op_type.empty()) { + RegisterOp(grad_op_type, ""); } - - return CreateOp(op_desc.type(), inputs, outputs, attrs); - } - - static std::shared_ptr CreateGradOp(const OperatorBase& op) { - PADDLE_ENFORCE(!op.IsNetOp(), - "Use framework::Backward to get backward ops"); - std::shared_ptr grad_op(BuildGradOp(&op)); - grad_op->Init(); - return grad_op; - } - - static std::unordered_map& protos() { - static std::unordered_map protos_; - return protos_; - } - - static std::unordered_map& grad_ops() { - static std::unordered_map grad_ops_; - return grad_ops_; } - static std::unordered_map>& - VarIndexMaps() { - static std::unordered_map> maps_; - return maps_; - } - - static std::unordered_map& op_creators() { - static std::unordered_map op_creators_; - return op_creators_; - } + static std::unique_ptr CreateOp(const std::string& type, + const VariableNameMap& inputs, + const VariableNameMap& outputs, + AttributeMap attrs); - private: - static std::unordered_map& op_checkers() { - static std::unordered_map op_checkers_; - return op_checkers_; - } + static std::unique_ptr CreateOp(const OpDesc& op_desc); - static void GenerateTempVariableName(OperatorBase* op) { - static std::atomic gUniqId(0UL); - for (auto& outname : op->outputs_) { - if (outname == kTempVarName) { - outname += op->type_; - outname += "@"; - outname += std::to_string(gUniqId.fetch_add(1)); - } - } - } + static std::unique_ptr CreateGradOp(const OperatorBase& op); }; 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 + // 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 +template class OpRegistrar : public Registrar { public: - explicit OpRegistrar(const char* op_type) { - OpRegistry::RegisterOp(op_type); - } -}; - -template -class GradOpRegistrar : public Registrar { - public: - GradOpRegistrar(const char* op_type, const char* grad_op_type) { - OpRegistry::RegisterGradOp(op_type, grad_op_type); + explicit OpRegistrar(const char* op_type) { OpRegistrar(op_type, ""); } + OpRegistrar(const char* op_type, const char* grad_op_type) { + OpRegistry::RegisterOp(op_type, + grad_op_type); } }; @@ -356,30 +121,30 @@ class OpKernelRegistrar : public Registrar { /** * Macro to register Operator. */ -#define REGISTER_OP(op_type, op_class, op_maker_class) \ +#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"); \ - static ::paddle::framework::OpRegistrar \ - __op_registrar_##op_type##__(#op_type); \ + 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; \ } -/** - * Macro to register Gradient Operator. - */ -#define REGISTER_GRADIENT_OP(op_type, grad_op_type, grad_op_class) \ - STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __reg_gradient_op__##op_type##_##grad_op_type, \ - "REGISTER_GRADIENT_OP must be called in global namespace"); \ - static ::paddle::framework::GradOpRegistrar \ - __op_gradient_registrar_##op_type##_##grad_op_type##__(#op_type, \ - #grad_op_type); \ - int TouchOpGradientRegistrar_##op_type() { \ - __op_gradient_registrar_##op_type##_##grad_op_type##__.Touch(); \ - return 0; \ - } +#define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \ + REGISTER_OP(op_type, op_class, op_maker_class, , ::paddle::framework::NOP) /** * Macro to register OperatorKernel. @@ -395,14 +160,6 @@ class OpKernelRegistrar : public Registrar { return 0; \ } -/** - * Macro to Forbid user register Gradient Operator. - */ -#define NO_GRADIENT(op_type) \ - STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __reg_gradient_op__##op_type##_##op_type##_grad, \ - "NO_GRADIENT must be called in global namespace") - #define REGISTER_OP_GPU_KERNEL(op_type, ...) \ REGISTER_OP_KERNEL(op_type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__) @@ -410,7 +167,8 @@ class OpKernelRegistrar : public Registrar { REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__) /** - * Macro to mark what Operator and Kernel we will use and tell the compiler to + * Macro to mark what Operator and Kernel + * we will use and tell the compiler to * link them into target. */ #define USE_OP_ITSELF(op_type) \ @@ -421,23 +179,6 @@ class OpKernelRegistrar : public Registrar { static int use_op_itself_##op_type##_ __attribute__((unused)) = \ TouchOpRegistrar_##op_type() -// TODO(fengjiayi): Most ops' gradient op have not been compeleted. So we use -// `NO_GRAD` to disable micro USE_OP_GRADIENT(op_type). Otherwise the code can't -// be compiled. `NO_GRAD` should be removed after all gradient ops are -// compeleted. -#define NO_GRAD -#ifndef NO_GRAD -#define USE_OP_GRADIENT(op_type) \ - STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __use_op_gradient_##op_type, \ - "USE_OP_GRADIENT must be called in global namespace"); \ - extern int TouchOpGradientRegistrar_##op_type(); \ - static int use_op_gradient_##op_type##_ __attribute__((unused)) = \ - TouchOpGradientRegistrar_##op_type() -#else -#define USE_OP_GRADIENT(op_type) -#endif - #define USE_OP_DEVICE_KERNEL(op_type, DEVICE_TYPE) \ STATIC_ASSERT_GLOBAL_NAMESPACE( \ __use_op_kernel_##op_type##_##DEVICE_TYPE##__, \ @@ -447,7 +188,8 @@ class OpKernelRegistrar : public Registrar { __attribute__((unused)) = \ TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE() -// TODO(fengjiayi): The following macros seems ugly, do we have better method? +// TODO(fengjiayi): The following macros +// seems ugly, do we have better method? #ifdef PADDLE_ONLY_CPU #define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU) @@ -457,18 +199,15 @@ class OpKernelRegistrar : public Registrar { USE_OP_DEVICE_KERNEL(op_type, GPU) #endif -#define USE_NO_GRAD_OP(op_type) \ - USE_OP_ITSELF(op_type); \ - USE_OP_KERNEL(op_type) +#define USE_NO_KERNEL_OP(op_type) USE_OP_ITSELF(op_type); -#define USE_CPU_OP(op_type) \ - USE_OP_ITSELF(op_type); \ - USE_OP_DEVICE_KERNEL(op_type, CPU); \ - USE_OP_GRADIENT(op_type) +#define USE_CPU_ONLY_OP(op_type) \ + USE_OP_ITSELF(op_type); \ + USE_OP_DEVICE_KERNEL(op_type, CPU); -#define USE_OP(op_type) \ - USE_NO_GRAD_OP(op_type); \ - USE_OP_GRADIENT(op_type) +#define USE_OP(op_type) \ + USE_OP_ITSELF(op_type); \ + USE_OP_KERNEL(op_type) } // namespace framework } // namespace paddle diff --git a/paddle/framework/op_registry_test.cc b/paddle/framework/op_registry_test.cc index e64126c7093a8eebc219afa4979d941ddc1afc97..e00c6e8d904508ec9985537fc703c7c61a14e0de 100644 --- a/paddle/framework/op_registry_test.cc +++ b/paddle/framework/op_registry_test.cc @@ -7,8 +7,7 @@ namespace paddle { namespace framework { class CosineOp : public OperatorBase { public: - DEFINE_OPERATOR_CTOR(CosineOp, OperatorBase) - + using OperatorBase::OperatorBase; void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override {} void InferShape(const Scope& scope) const override {} @@ -22,15 +21,14 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { AddOutput("output", "output of cosine op"); AddAttr("scale", "scale of cosine op") .SetDefault(1.0) - .LargerThan(0.0); + .GreaterThan(0.0); AddComment("This is cos op"); } }; class MyTestOp : public OperatorBase { public: - DEFINE_OPERATOR_CTOR(MyTestOp, OperatorBase) - + using OperatorBase::OperatorBase; void InferShape(const Scope& scope) const override {} void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override {} @@ -40,8 +38,8 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: MyTestOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("input", "input of cosine op").SetMultiple(); - AddOutput("output", "output of cosine op").SetTemporary(); + AddInput("input", "input of cosine op").AsDuplicable(); + AddOutput("output", "output of cosine op").AsIntermediate(); auto my_checker = [](int i) { PADDLE_ENFORCE(i % 2 == 0, "'test_attr' must be even!"); }; @@ -53,16 +51,24 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { } // namespace framework } // namespace paddle -REGISTER_OP(cos_sim, paddle::framework::CosineOp, - paddle::framework::CosineOpProtoAndCheckerMaker); -REGISTER_OP(my_test_op, paddle::framework::MyTestOp, - paddle::framework::MyTestOpProtoAndCheckerMaker); +static void BuildVar(const std::string& param_name, + std::initializer_list arguments, + paddle::framework::OpDesc::Var* var) { + var->set_parameter(param_name); + for (auto& arg_name : arguments) { + var->add_arguments(arg_name); + } +} +REGISTER_OP_WITHOUT_GRADIENT(cos_sim, paddle::framework::CosineOp, + paddle::framework::CosineOpProtoAndCheckerMaker); +REGISTER_OP_WITHOUT_GRADIENT(my_test_op, paddle::framework::MyTestOp, + paddle::framework::MyTestOpProtoAndCheckerMaker); TEST(OpRegistry, CreateOp) { paddle::framework::OpDesc op_desc; op_desc.set_type("cos_sim"); - op_desc.add_inputs("aa"); - op_desc.add_outputs("bb"); + BuildVar("input", {"aa"}, op_desc.add_inputs()); + BuildVar("output", {"bb"}, op_desc.add_outputs()); float scale = 3.3; auto attr = op_desc.mutable_attrs()->Add(); @@ -70,20 +76,19 @@ TEST(OpRegistry, CreateOp) { attr->set_type(paddle::framework::AttrType::FLOAT); attr->set_f(scale); - std::shared_ptr op = - paddle::framework::OpRegistry::CreateOp(op_desc); + auto op = paddle::framework::OpRegistry::CreateOp(op_desc); paddle::framework::Scope scope; paddle::platform::CPUDeviceContext dev_ctx; op->Run(scope, dev_ctx); - float scale_get = op->GetAttr("scale"); + float scale_get = op->Attr("scale"); ASSERT_EQ(scale_get, scale); } TEST(OpRegistry, IllegalAttr) { paddle::framework::OpDesc op_desc; op_desc.set_type("cos_sim"); - op_desc.add_inputs("aa"); - op_desc.add_outputs("bb"); + BuildVar("input", {"aa"}, op_desc.add_inputs()); + BuildVar("output", {"bb"}, op_desc.add_outputs()); auto attr = op_desc.mutable_attrs()->Add(); attr->set_name("scale"); @@ -107,33 +112,23 @@ TEST(OpRegistry, IllegalAttr) { TEST(OpRegistry, DefaultValue) { paddle::framework::OpDesc op_desc; op_desc.set_type("cos_sim"); - op_desc.add_inputs("aa"); - op_desc.add_outputs("bb"); + BuildVar("input", {"aa"}, op_desc.add_inputs()); + BuildVar("output", {"bb"}, op_desc.add_outputs()); ASSERT_TRUE(op_desc.IsInitialized()); - std::shared_ptr op = - paddle::framework::OpRegistry::CreateOp(op_desc); + auto op = paddle::framework::OpRegistry::CreateOp(op_desc); paddle::framework::Scope scope; paddle::platform::CPUDeviceContext dev_ctx; op->Run(scope, dev_ctx); - ASSERT_EQ(op->GetAttr("scale"), 1.0); -} - -static void SetInputFormat(paddle::framework::OpDesc* desc) { - auto attr = desc->add_attrs(); - attr->set_name("input_format"); - attr->set_type(paddle::framework::INTS); - attr->mutable_ints()->Add(0); - attr->mutable_ints()->Add(1); + ASSERT_EQ(op->Attr("scale"), 1.0); } TEST(OpRegistry, CustomChecker) { paddle::framework::OpDesc op_desc; op_desc.set_type("my_test_op"); - op_desc.add_inputs("ii"); - op_desc.add_outputs("oo"); - SetInputFormat(&op_desc); + BuildVar("input", {"ii"}, op_desc.add_inputs()); + BuildVar("output", {"oo"}, op_desc.add_outputs()); // attr 'test_attr' is not set bool caught = false; @@ -173,43 +168,10 @@ TEST(OpRegistry, CustomChecker) { attr->set_name("test_attr"); attr->set_type(paddle::framework::AttrType::INT); attr->set_i(4); - SetInputFormat(&op_desc); auto op = paddle::framework::OpRegistry::CreateOp(op_desc); paddle::platform::CPUDeviceContext dev_ctx; paddle::framework::Scope scope; op->Run(scope, dev_ctx); - int test_attr = op->GetAttr("test_attr"); + int test_attr = op->Attr("test_attr"); ASSERT_EQ(test_attr, 4); -} - -class TestAttrProtoMaker : public pd::OpProtoAndCheckerMaker { - public: - TestAttrProtoMaker(pd::OpProto* proto, pd::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddAttr("scale", "scale of test op"); - AddAttr("scale", "scale of test op"); - } -}; - -TEST(ProtoMaker, DuplicatedAttr) { - pd::OpProto op_proto; - pd::OpAttrChecker op_checker; - auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker); - ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); -} - -class TestInOutProtoMaker : public pd::OpProtoAndCheckerMaker { - public: - TestInOutProtoMaker(pd::OpProto* proto, pd::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("input", "input of test op"); - AddInput("input", "input of test op"); - } -}; - -TEST(ProtoMaker, DuplicatedInOut) { - pd::OpProto op_proto; - pd::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/operator.cc b/paddle/framework/operator.cc index d9a013b883abdec4422806f90e36da7410a4fa0c..e1e122091f7759b1a68f1f982bc2a35e8241f9f0 100644 --- a/paddle/framework/operator.cc +++ b/paddle/framework/operator.cc @@ -12,9 +12,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include - #include "paddle/framework/operator.h" +#include +#include "paddle/framework/op_registry.h" namespace paddle { namespace framework { @@ -33,84 +33,195 @@ ExecutionContext::GetEigenDevice() const { } #endif -const std::string& OperatorBase::Input(const std::string& name) const { - PADDLE_ENFORCE_NOT_NULL(in_out_idxs_, - "Input Output Indices could not be nullptr"); - auto it = in_out_idxs_->find(name); - PADDLE_ENFORCE(it != in_out_idxs_->end(), "no key [%s] in in_out_idxs_", - name); - if (attrs_.count("input_format") == 0) { - return inputs_.at((size_t)it->second); - } else { - const auto& input_format = GetAttr>("input_format"); - int idx = input_format[it->second]; - return inputs_.at((size_t)idx); - } +std::string OperatorBase::Input(const std::string& name) const { + auto& ins = Inputs(name); + PADDLE_ENFORCE_LE(ins.size(), 1UL, + "Op %s input %s should contain only one variable", type_, + name); + return ins.empty() ? kEmptyVarName : ins[0]; } -std::vector OperatorBase::Inputs(const std::string& name) const { - PADDLE_ENFORCE_NOT_NULL(in_out_idxs_, "IO Idx could not be nullptr"); - auto input_format = GetAttr>("input_format"); - auto offset = in_out_idxs_->at(name); - PADDLE_ENFORCE(input_format.at(static_cast(offset) + 1) <= - static_cast(inputs_.size()), - "Input Out Of Range"); - - return std::vector{ - inputs_.begin() + input_format.at(offset), - inputs_.begin() + input_format.at(offset + 1)}; +const std::vector& OperatorBase::Inputs( + const std::string& name) const { + auto it = inputs_.find(name); + PADDLE_ENFORCE(it != inputs_.end(), "Op %s do not have input %s", type_, + name); + return it->second; } -const std::string& OperatorBase::Output(const std::string& name) const { - PADDLE_ENFORCE_NOT_NULL(in_out_idxs_, "InOut Indice could not be nullptr"); - auto it = in_out_idxs_->find(name); - PADDLE_ENFORCE(it != in_out_idxs_->end(), "no key [%s] in in_out_idxs_", - name); - if (attrs_.count("output_format") == 0) { - return outputs_.at((size_t)it->second); - } else { - const auto& output_format = GetAttr>("output_format"); - int idx = output_format[it->second]; - return outputs_.at((size_t)idx); - } +std::string OperatorBase::Output(const std::string& name) const { + auto& outs = Outputs(name); + PADDLE_ENFORCE_LE(outs.size(), 1UL, + "Op %s output %s should contain only one variable", type_, + name); + return outs.empty() ? kEmptyVarName : outs[0]; } -std::vector OperatorBase::Outputs(const std::string& name) const { - PADDLE_ENFORCE_NOT_NULL(in_out_idxs_, "InOut Indice could not be nullptr"); - auto output_format = GetAttr>("output_format"); - auto offset = in_out_idxs_->at(name); - PADDLE_ENFORCE(output_format.at(static_cast(offset) + 1) <= - static_cast(outputs_.size()), - "Output Out of Range"); - return std::vector{ - outputs_.begin() + output_format.at(offset), - outputs_.begin() + output_format.at(offset + 1)}; +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); + return it->second; } std::string OperatorBase::DebugString() const { std::stringstream ss; - ss << "Op(" << type_ << "), inputs:("; - for (size_t i = 0; i < inputs_.size(); ++i) { - ss << inputs_[i]; - if (i != inputs_.size() - 1) { + ss << "Op(" << type_ << "), inputs:{"; + for (auto it = inputs_.begin(); it != inputs_.end();) { + auto& input = *it; + ss << input.first << "["; + for (size_t i = 0; i < input.second.size(); ++i) { + ss << input.second[i]; + if (i != input.second.size() - 1) { + ss << ", "; + } + } + ss << "]"; + ++it; + if (it != inputs_.end()) { ss << ", "; } } - ss << "), outputs:("; - for (size_t i = 0; i < outputs_.size(); ++i) { - ss << outputs_[i]; - if (i != outputs_.size() - 1) { + ss << "}, outputs:{"; + for (auto it = outputs_.begin(); it != outputs_.end();) { + auto& output = *it; + ss << output.first << "["; + for (size_t i = 0; i < output.second.size(); ++i) { + ss << output.second[i]; + if (i != output.second.size() - 1) { + ss << ", "; + } + } + ss << "]"; + ++it; + if (it != outputs_.end()) { ss << ", "; } } - ss << ")."; + ss << "}."; return ss.str(); } void OperatorBase::Rename(const std::string& old_name, const std::string& new_name) { - std::replace(inputs_.begin(), inputs_.end(), old_name, new_name); - std::replace(outputs_.begin(), outputs_.end(), old_name, 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); + } +} + +OperatorBase::OperatorBase(const std::string& type, + const VariableNameMap& inputs, + const VariableNameMap& outputs, + const AttributeMap& attrs) + : type_(type), inputs_(inputs), outputs_(outputs), attrs_(attrs) { + GenerateTemporaryNames(); + CheckAllInputOutputSet(); +} + +std::vector OperatorBase::InputVars() const { + std::vector ret_val; + for (auto& o : outputs_) { + ret_val.reserve(ret_val.size() + o.second.size()); + ret_val.insert(ret_val.end(), o.second.begin(), o.second.end()); + } + return ret_val; +} + +std::vector OperatorBase::OutputVars(bool has_intermediate) const { + std::vector ret_val; + if (has_intermediate) { + // push all outputs into ret_val + for (auto& o : outputs_) { + ret_val.reserve(ret_val.size() + o.second.size()); + ret_val.insert(ret_val.end(), o.second.begin(), o.second.end()); + } + return ret_val; + } + auto& info = OpInfoMap::Instance().Get(Type()); + + // get all OpProto::Var for outputs + for (auto& o : info.Proto().outputs()) { + // ignore all intermediate output + if (o.intermediate()) continue; + auto out = outputs_.find(o.name()); + if (out != outputs_.end()) { + ret_val.reserve(ret_val.size() + out->second.size()); + ret_val.insert(ret_val.end(), out->second.begin(), out->second.end()); + } + } + return ret_val; +} + +void OperatorBase::CheckAllInputOutputSet() const { + auto& info_map = OpInfoMap::Instance(); + auto* op_info = info_map.GetNullable(Type()); + if (op_info == nullptr || op_info->proto_ == nullptr) return; + + for (auto& in : op_info->Proto().inputs()) { + PADDLE_ENFORCE(inputs_.find(in.name()) != inputs_.end(), + "Type %s's input %s is not set", Type(), in.name()); + } + + for (auto& out : op_info->Proto().outputs()) { + PADDLE_ENFORCE(outputs_.find(out.name()) != outputs_.end(), + "Type %s's output %s is not set", Type(), out.name()); + } +} + +void OperatorBase::GenerateTemporaryNames() { + static std::atomic gUniqId(0UL); + for (auto& output : outputs_) { + for (auto& output_name : output.second) { + if (output_name == kTempVarName) { + output_name += type_; + output_name += "@"; + output_name += std::to_string(gUniqId.fetch_add(1)); + } + } + } +} + +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 diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h index 68e7fedcd6102435a3c30326aa91043b8abecb9e..4600b06009bcef7d0774d25b816aac4733f30795 100644 --- a/paddle/framework/operator.h +++ b/paddle/framework/operator.h @@ -19,9 +19,9 @@ limitations under the License. */ #include #include +#include "op_info.h" #include "paddle/framework/attribute.h" -#include "paddle/framework/op_desc.pb.h" -#include "paddle/framework/op_proto.pb.h" +#include "paddle/framework/framework.pb.h" #include "paddle/framework/scope.h" #include "paddle/framework/tensor.h" #include "paddle/platform/device_context.h" @@ -63,21 +63,13 @@ class ExecutionContext; */ class OperatorBase { public: - OperatorBase() {} // TODO(yi): This constructor is to be removed. - OperatorBase(const std::string& type, const std::vector& inputs, - const std::vector& outputs, - const AttributeMap& attrs, - std::unordered_map* in_out_idxs) - : type_(type), - inputs_(inputs), - outputs_(outputs), - attrs_(attrs), - in_out_idxs_(in_out_idxs) {} + OperatorBase(const std::string& type, const VariableNameMap& inputs, + const VariableNameMap& outputs, const AttributeMap& attrs); virtual ~OperatorBase() {} template - inline const T& GetAttr(const std::string& name) const { + inline const T& Attr(const std::string& name) const { PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap", name); return boost::get(attrs_.at(name)); @@ -85,10 +77,6 @@ class OperatorBase { virtual std::string DebugString() const; - /// Init will be called after CreateOperator, you can put some initialization - /// logic here. - virtual void Init() {} - /// InferShape infer the size of Variables used by this Operator with /// information inside scope virtual void InferShape(const Scope& scope) const = 0; @@ -104,39 +92,143 @@ class OperatorBase { /// rename inputs outputs name void Rename(const std::string& old_name, const std::string& new_name); + const VariableNameMap& Inputs() const { return inputs_; } + const VariableNameMap& Outputs() const { return outputs_; } + //! Get a input with argument's name described in `op_proto` - const std::string& Input(const std::string& name) const; + std::string Input(const std::string& name) const; //! Get a input which has multiple variables. - //! TODO add a vector_view to prevent memory copy. - std::vector Inputs(const std::string& name) const; + const std::vector& Inputs(const std::string& name) const; + + std::vector InputVars() const; //! Get a output with argument's name described in `op_proto` - const std::string& Output(const std::string& name) const; + std::string Output(const std::string& name) const; //! Get an output which has multiple variables. //! TODO add a vector_view to prevent memory copy. - std::vector Outputs(const std::string& name) const; + const std::vector& Outputs(const std::string& name) const; - const std::string Type() const { return type_; } - const std::vector Inputs() const { return inputs_; } - const std::vector Outputs() const { return outputs_; } + virtual std::vector OutputVars(bool has_intermediate) const; + + const std::string& Type() const { return type_; } + void SetType(const std::string& type) { type_ = type; } const AttributeMap& Attrs() const { return attrs_; } - const std::unordered_map* InOutIdx() const { - return in_out_idxs_.get(); - } - public: + // Return a new operator instance, which is as same as this. + // Use unique_ptr to prevent caller forget to delete this pointer. + virtual std::unique_ptr Clone() const = 0; + + protected: std::string type_; // NOTE: in case of OpGrad, inputs_ contains: - // I (Inputs) + // I (Inputs)opear // O (Outputs) // OG (Output Gradients) - std::vector inputs_; + VariableNameMap inputs_; + // NOTE: in case of OpGrad, outputs_ contains // IG (Inputs Gradients) - std::vector outputs_; + VariableNameMap outputs_; AttributeMap attrs_; - // store the arguments' offset described in op_desc. - std::shared_ptr> in_out_idxs_; + + private: + void GenerateTemporaryNames(); + void CheckAllInputOutputSet() const; +}; + +// Macro for define a clone method. +// If you are writing an kernel operator, `Clone` will be defined when you +// register it. i.e. `Clone` method is not needed to define by yourself. +#define DEFINE_OP_CLONE_METHOD(cls) \ + std::unique_ptr Clone() const final { \ + return std::unique_ptr(new cls(*this)); \ + } + +// Macro for define a default constructor for Operator. +// You can also use +// using PARENT_CLASS::PARENT_CLASS; +// to use parent's constructor. +#define DEFINE_OP_CONSTRUCTOR(cls, parent_cls) \ + cls(const std::string& type, \ + const ::paddle::framework::VariableNameMap& inputs, \ + const ::paddle::framework::VariableNameMap& outputs, \ + const paddle::framework::AttributeMap& attrs) \ + : parent_cls(type, inputs, outputs, attrs) {} + +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 { + return std::unique_ptr(new NOP(*this)); + } +}; + +// 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 { @@ -144,24 +236,31 @@ class InferShapeContext { InferShapeContext(const OperatorBase& op, const Scope& scope) : op_(op), scope_(scope) {} - size_t InputSize() const { return op_.inputs_.size(); } + const OperatorBase& op() const { return op_; } - size_t OutputSize() const { return op_.outputs_.size(); } + const Scope& scope() const { return scope_; } - const Variable* InputVar(const size_t index) const { - return scope_.FindVar(op_.inputs_.at(index)); + template + inline const T& Attr(const std::string& name) const { + return op_.Attr(name); + } + + size_t InputSize(const std::string& name) const { + return op_.Inputs(name).size(); } - Variable* OutputVar(const size_t index) const { - return scope_.FindVar(op_.outputs_.at(index)); + size_t OutputSize(const std::string& name) const { + return op_.Outputs(name).size(); } const Variable* InputVar(const std::string& name) const { - return scope_.FindVar(op_.Input(name)); + auto ipt = op_.Input(name); + return ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt); } Variable* OutputVar(const std::string& name) const { - return scope_.FindVar(op_.Output(name)); + auto opt = op_.Output(name); + return opt == kEmptyVarName ? nullptr : scope_.FindVar(opt); } const std::vector MultiInputVar( @@ -169,9 +268,11 @@ class InferShapeContext { auto names = op_.Inputs(name); std::vector res; res.reserve(names.size()); - std::transform( - names.begin(), names.end(), std::back_inserter(res), - [this](const std::string& name) { return scope_.FindVar(name); }); + std::transform(names.begin(), names.end(), std::back_inserter(res), + [this](const std::string& name) { + return name == kEmptyVarName ? nullptr + : scope_.FindVar(name); + }); return res; } @@ -179,42 +280,24 @@ class InferShapeContext { auto names = op_.Outputs(name); std::vector res; res.reserve(names.size()); - std::transform( - names.begin(), names.end(), std::back_inserter(res), - [this](const std::string& name) { return scope_.FindVar(name); }); + std::transform(names.begin(), names.end(), std::back_inserter(res), + [this](const std::string& name) { + return name == kEmptyVarName ? nullptr + : scope_.FindVar(name); + }); return res; } - template - const T* Input(const size_t index) const { - auto var = InputVar(index); - PADDLE_ENFORCE_NOT_NULL(var, "Input(%d) should not be nullptr", index); - return &var->Get(); - } - - template - T* Output(const size_t index) const { - auto var = OutputVar(index); - PADDLE_ENFORCE_NOT_NULL( - var, - "Output(%d) not be nullptr, which means variable [%s] does not " - "exist in scope", - index, op_.outputs_[index]); - return var->GetMutable(); - } - template const T* Input(const std::string& name) const { - auto var = InputVar(name); - PADDLE_ENFORCE_NOT_NULL(var, "Input(%s) should not be nullptr", name); - return &var->Get(); + auto* var = InputVar(name); + return var == nullptr ? nullptr : &var->Get(); } template T* Output(const std::string& name) const { auto var = OutputVar(name); - PADDLE_ENFORCE_NOT_NULL(var, "Output(%s) should not be nullptr", name); - return var->GetMutable(); + return var == nullptr ? nullptr : var->GetMutable(); } template @@ -225,30 +308,25 @@ class InferShapeContext { std::transform(names.begin(), names.end(), std::back_inserter(res), [&](const std::string& sub_name) { auto var = scope_.FindVar(sub_name); - PADDLE_ENFORCE_NOT_NULL( - var, "MultiInput(%s:%s) should not be nullptr", name, - sub_name); - return &var->Get(); + return var == nullptr ? nullptr : &var->Get(); }); return res; } template - std::vector MultiOutput(const std::string& name) const { + std::vector MultiOutput(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), [&](const std::string& sub_name) { auto var = scope_.FindVar(sub_name); - PADDLE_ENFORCE_NOT_NULL( - var, "MultiOutput(%s:%s) should not be nullptr", name, - sub_name); - return var->GetMutable(); + return var == nullptr ? nullptr : var->GetMutable(); }); return res; } + private: const OperatorBase& op_; const Scope& scope_; }; @@ -281,6 +359,10 @@ class ExecutionContext : public InferShapeContext { platform::Place GetPlace() const { return device_context_->GetPlace(); } + const platform::DeviceContext* device_context() const { + return device_context_; + } + const platform::DeviceContext* device_context_; }; @@ -300,14 +382,6 @@ class OpKernel { class OperatorWithKernel : public OperatorBase { public: - OperatorWithKernel() {} // TODO(yi): This constructor is to be removed. - OperatorWithKernel(const std::string& type, - const std::vector& inputs, - const std::vector& outputs, - const AttributeMap& attrs, - std::unordered_map* in_out_idxs) - : OperatorBase(type, inputs, outputs, attrs, in_out_idxs) {} - struct OpKernelKey { platform::Place place_; @@ -331,6 +405,10 @@ class OperatorWithKernel : public OperatorBase { using OpKernelMap = 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)); } @@ -357,15 +435,5 @@ class OperatorWithKernel : public OperatorBase { virtual void InferShape(const InferShapeContext& ctx) const = 0; }; -#define DEFINE_OPERATOR_CTOR(Class, ParentClass) \ - public: \ - Class() { /* TODO(yi): This constructor is to be removed. */ \ - } \ - Class(const std::string& type, const std::vector& inputs, \ - const std::vector& outputs, \ - const ::paddle::framework::AttributeMap& attrs, \ - std::unordered_map* in_out_idxs) \ - : ParentClass(type, inputs, outputs, attrs, in_out_idxs) {} - } // namespace framework } // namespace paddle diff --git a/paddle/framework/operator_test.cc b/paddle/framework/operator_test.cc index 7dbd5b14ab6ec89ae9940a3d12ec9d2b169153ad..20bbb11896a4c6f11079669f0b25773f6460594d 100644 --- a/paddle/framework/operator_test.cc +++ b/paddle/framework/operator_test.cc @@ -23,22 +23,22 @@ static int op_run_num = 0; class OpWithoutKernelTest : public OperatorBase { public: - DEFINE_OPERATOR_CTOR(OpWithoutKernelTest, OperatorBase) - - void Init() override { x = 1; } + 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++; - ASSERT_EQ((int)inputs_.size(), 1); - ASSERT_EQ((int)outputs_.size(), 1); - ASSERT_EQ(scope.FindVar(inputs_[0]), nullptr); + ++op_run_num; + ASSERT_EQ(static_cast(inputs_.size()), 1); + ASSERT_EQ(static_cast(outputs_.size()), 1); + ASSERT_EQ(scope.FindVar(inputs_.at("input")[0]), nullptr); ASSERT_EQ(x, 1); - ASSERT_NE(scope.FindVar(outputs_[0]), nullptr); + ASSERT_NE(scope.FindVar(outputs_.at("output")[0]), nullptr); } public: - float x = 0; + int x{0}; }; class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { @@ -56,14 +56,25 @@ class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { } // namespace framework } // namespace paddle -REGISTER_OP(test_operator, paddle::framework::OpWithoutKernelTest, - paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker); +static void BuildVar(const std::string& param_name, + std::initializer_list arguments, + paddle::framework::OpDesc::Var* var) { + var->set_parameter(param_name); + for (auto& arg_name : arguments) { + *var->mutable_arguments()->Add() = arg_name; + } +} + +REGISTER_OP_WITHOUT_GRADIENT( + test_operator, paddle::framework::OpWithoutKernelTest, + paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker); TEST(OperatorBase, all) { paddle::framework::OpDesc op_desc; op_desc.set_type("test_operator"); - *op_desc.mutable_inputs()->Add() = "IN1"; - *op_desc.mutable_outputs()->Add() = "OUT1"; + BuildVar("input", {"IN1"}, op_desc.add_inputs()); + BuildVar("output", {"OUT1"}, op_desc.add_outputs()); + auto attr = op_desc.mutable_attrs()->Add(); attr->set_name("scale"); attr->set_type(paddle::framework::AttrType::FLOAT); @@ -91,7 +102,7 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { AddOutput("y", "output of test op"); AddAttr("scale", "scale of cosine op") .SetDefault(1.0) - .LargerThan(0.0); + .GreaterThan(0.0); AddComment("This is test op"); } }; @@ -100,7 +111,8 @@ static int cpu_kernel_run_num = 0; class OpWithKernelTest : public OperatorWithKernel { public: - DEFINE_OPERATOR_CTOR(OpWithKernelTest, OperatorWithKernel) + using OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext& ctx) const override {} }; @@ -110,31 +122,11 @@ class CPUKernelTest : public OpKernel { public: void Compute(const ExecutionContext& ctx) const { std::cout << "this is cpu kernel" << std::endl; - std::cout << ctx.op_.DebugString() << std::endl; + std::cout << ctx.op().DebugString() << std::endl; cpu_kernel_run_num++; - ASSERT_EQ(ctx.op_.Input("x"), "IN1"); - ASSERT_EQ(ctx.op_.Output("y"), "OUT1"); - } -}; - -// multiple inputs test -class OperatorMultiInputsTest : public OperatorBase { - public: - DEFINE_OPERATOR_CTOR(OperatorMultiInputsTest, OperatorBase) - - void Init() override { x = 1; } - void InferShape(const Scope& scope) const override {} - void Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const override { - ASSERT_EQ(scope.FindVar(inputs_[0]), nullptr); - ASSERT_EQ(x, 1); - ASSERT_NE(scope.FindVar(outputs_[0]), nullptr); - ASSERT_EQ(Input("x"), "IN1"); - ASSERT_EQ(Input("y"), "OUT1"); + ASSERT_EQ(ctx.op().Input("x"), "IN1"); + ASSERT_EQ(ctx.op().Output("y"), "OUT1"); } - - public: - float x = 0; }; class OpKernelTestMultiInputsProtoAndCheckerMaker @@ -143,12 +135,12 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker OpKernelTestMultiInputsProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("xs", "inputs of test op").SetMultiple(); + AddInput("xs", "inputs of test op").AsDuplicable(); AddInput("k", "input of test op"); - AddOutput("ys", "outputs of test op").SetMultiple(); + AddOutput("ys", "outputs of test op").AsDuplicable(); AddAttr("scale", "scale of cosine op") .SetDefault(1.0) - .LargerThan(0.0); + .GreaterThan(0.0); AddComment("This is test op"); } }; @@ -156,7 +148,7 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker class CPUKernalMultiInputsTest : public OpKernel { public: void Compute(const ExecutionContext& ctx) const { - auto xs = ctx.op_.Inputs("xs"); + auto xs = ctx.op().Inputs("xs"); ASSERT_EQ(xs.size(), 3UL); ASSERT_EQ(xs[0], "x0"); ASSERT_EQ(xs[1], "x1"); @@ -180,10 +172,10 @@ class CPUKernalMultiInputsTest : public OpKernel { auto outTensor0 = ctx.MultiOutput("ys"); ASSERT_EQ(outTensor0.size(), 2U); - auto k = ctx.op_.Input("k"); + auto k = ctx.op().Input("k"); ASSERT_EQ(k, "k0"); - auto ys = ctx.op_.Outputs("ys"); + auto ys = ctx.op().Outputs("ys"); ASSERT_EQ(ys.size(), 2UL); ASSERT_EQ(ys[0], "y0"); ASSERT_EQ(ys[1], "y1"); @@ -193,8 +185,9 @@ class CPUKernalMultiInputsTest : public OpKernel { } // namespace framework } // namespace paddle -REGISTER_OP(op_with_kernel, paddle::framework::OpWithKernelTest, - paddle::framework::OpKernelTestProtoAndCheckerMaker); +REGISTER_OP_WITHOUT_GRADIENT( + op_with_kernel, paddle::framework::OpWithKernelTest, + paddle::framework::OpKernelTestProtoAndCheckerMaker); REGISTER_OP_CPU_KERNEL(op_with_kernel, paddle::framework::CPUKernelTest); @@ -202,8 +195,9 @@ REGISTER_OP_CPU_KERNEL(op_with_kernel, TEST(OpKernel, all) { paddle::framework::OpDesc op_desc; op_desc.set_type("op_with_kernel"); - *op_desc.mutable_inputs()->Add() = "IN1"; - *op_desc.mutable_outputs()->Add() = "OUT1"; + BuildVar("x", {"IN1"}, op_desc.add_inputs()); + BuildVar("y", {"OUT1"}, op_desc.add_outputs()); + auto attr = op_desc.mutable_attrs()->Add(); attr->set_name("scale"); attr->set_type(paddle::framework::AttrType::FLOAT); @@ -218,8 +212,9 @@ TEST(OpKernel, all) { ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1); } -REGISTER_OP(op_multi_inputs_with_kernel, paddle::framework::OpWithKernelTest, - paddle::framework::OpKernelTestMultiInputsProtoAndCheckerMaker); +REGISTER_OP_WITHOUT_GRADIENT( + op_multi_inputs_with_kernel, paddle::framework::OpWithKernelTest, + paddle::framework::OpKernelTestMultiInputsProtoAndCheckerMaker); REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel, paddle::framework::CPUKernalMultiInputsTest); @@ -229,32 +224,15 @@ TEST(OpKernel, multi_inputs) { OpDesc op_desc; op_desc.set_type("op_multi_inputs_with_kernel"); - *op_desc.mutable_inputs()->Add() = "x0"; - *op_desc.mutable_inputs()->Add() = "x1"; - *op_desc.mutable_inputs()->Add() = "x2"; - *op_desc.mutable_inputs()->Add() = "k0"; - *op_desc.mutable_outputs()->Add() = "y0"; - *op_desc.mutable_outputs()->Add() = "y1"; + BuildVar("xs", {"x0", "x1", "x2"}, op_desc.add_inputs()); + BuildVar("k", {"k0"}, op_desc.add_inputs()); + BuildVar("ys", {"y0", "y1"}, op_desc.add_outputs()); + auto attr = op_desc.mutable_attrs()->Add(); attr->set_name("scale"); attr->set_type(paddle::framework::AttrType::FLOAT); attr->set_f(3.14); - auto attr0 = op_desc.mutable_attrs()->Add(); - attr0->set_name("input_format"); - attr0->set_type(paddle::framework::AttrType::INTS); - auto input_format = attr0->mutable_ints(); - input_format->Add(0); // x0 - input_format->Add(3); // k - input_format->Add(4); // end - - auto attr1 = op_desc.mutable_attrs()->Add(); - attr1->set_name("output_format"); - attr1->set_type(paddle::framework::AttrType::INTS); - auto output_format = attr1->mutable_ints(); - output_format->Add(0); // y0 - output_format->Add(2); // y1 - paddle::platform::CPUDeviceContext cpu_device_context; paddle::framework::Scope scope; scope.NewVar("x0")->GetMutable(); @@ -267,3 +245,56 @@ TEST(OpKernel, multi_inputs) { auto op = paddle::framework::OpRegistry::CreateOp(op_desc); op->Run(scope, cpu_device_context); } + +class OperatorClone : public paddle::framework::OperatorBase { + public: + DEFINE_OP_CLONE_METHOD(OperatorClone); + OperatorClone(const std::string& type, + const paddle::framework::VariableNameMap& inputs, + 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 {} +}; + +TEST(Operator, Clone) { + OperatorClone a("ABC", {}, {}, {}); + 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/tensor.h b/paddle/framework/tensor.h index cd1b4de426a49fa66dbbf8cf7d09990ac8d21227..ce938b21437195fed8c1adad4329fd139f3f96ab 100644 --- a/paddle/framework/tensor.h +++ b/paddle/framework/tensor.h @@ -43,6 +43,9 @@ class Tensor { template friend struct EigenTensor; + template + friend struct EigenMatrix; + template friend struct EigenVector; @@ -105,6 +108,11 @@ class Tensor { template inline Tensor Slice(const int& begin_idx, const int& end_idx) const; + platform::Place place() const { + PADDLE_ENFORCE_NOT_NULL(holder_, "Tensor get place() must contains holder"); + return holder_->place(); + } + private: template inline void check_memory_size() const; diff --git a/paddle/framework/tensor_impl.h b/paddle/framework/tensor_impl.h index 7d7263b899afb7a2128548f264065a8013b6f0c9..637f04ae0037bd402d855b8bcde8087bfe8328d1 100644 --- a/paddle/framework/tensor_impl.h +++ b/paddle/framework/tensor_impl.h @@ -58,7 +58,7 @@ inline T* Tensor::mutable_data(platform::Place place) { "Tensor's numel must be larger than zero to call " "Tensor::mutable_data. Call Tensor::set_dim first."); /* some versions of boost::variant don't have operator!= */ - size_t size = product(dims_) * sizeof(T); + int64_t size = product(dims_) * sizeof(T); if (holder_ == nullptr || !(holder_->place() == place) || holder_->size() < size + offset_) { if (platform::is_cpu_place(place)) { @@ -117,6 +117,8 @@ inline void Tensor::CopyFrom(const Tensor& src, memory::Copy(boost::get(dst_place), dst_ptr, boost::get(src_place), src_ptr, size, 0); } + PADDLE_ENFORCE(cudaStreamSynchronize(0), + "cudaStreamSynchronize failed in Tensor CopyFrom"); #endif } @@ -129,7 +131,7 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const { 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."); - int base = product(dims_) / dims_[0]; + size_t base = product(dims_) / dims_[0]; Tensor dst; dst.holder_ = holder_; DDim dst_dims = dims_; @@ -146,5 +148,13 @@ inline Tensor& Tensor::Resize(const DDim& dims) { inline const DDim& Tensor::dims() const { return dims_; } +template +inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) { + Tensor res; + res.ShareDataWith(src); + res.Resize(flatten_to_2d(src.dims(), num_col_dims)); + return res; +} + } // namespace framework } // namespace paddle diff --git a/paddle/framework/tensor_test.cc b/paddle/framework/tensor_test.cc index 7db38d5caeebccf710334e854faf785ef0f64063..55302ea47120f420e952b26830c8ea4cbcce6435 100644 --- a/paddle/framework/tensor_test.cc +++ b/paddle/framework/tensor_test.cc @@ -262,3 +262,16 @@ TEST(Tensor, CopyFrom) { } #endif } + +TEST(Tensor, ReshapeToMatrix) { + using namespace paddle::framework; + using namespace paddle::platform; + Tensor src; + int* src_ptr = src.mutable_data({2, 3, 4, 9}, CPUPlace()); + for (int i = 0; i < 2 * 3 * 4 * 9; ++i) { + src_ptr[i] = i; + } + Tensor res = ReshapeToMatrix(src, 2); + ASSERT_EQ(res.dims()[0], 2 * 3); + ASSERT_EQ(res.dims()[1], 4 * 9); +} \ No newline at end of file diff --git a/paddle/function/CMakeLists.txt b/paddle/function/CMakeLists.txt index 7dfb6f61c50959f7269725a00dbc4f9c27474bdf..4fd72d64a90ae6f16dd1499ceb7fba6e40fe4cea 100644 --- a/paddle/function/CMakeLists.txt +++ b/paddle/function/CMakeLists.txt @@ -4,6 +4,10 @@ file(GLOB cpp_files . *Op.cpp) list(APPEND h_files Function.h) list(APPEND cpp_files Function.cpp) list(APPEND cpp_files BufferArg.cpp) +list(APPEND cpp_files GemmFunctor.cpp) +if(USE_EIGEN_FOR_BLAS) + list(APPEND cpp_files EigenGemm.cpp) +endif(USE_EIGEN_FOR_BLAS) if(WITH_GPU) file(GLOB cu_files . *OpGpu.cu) @@ -17,6 +21,8 @@ if(USE_NNPACK) endif() endif() +list(APPEND cpp_files neon/NeonDepthwiseConv.cpp) + add_library(paddle_function STATIC ${cpp_files} ${cu_objs}) add_dependencies(paddle_function ${external_project_dependencies}) add_dependencies(paddle_function paddle_proto) @@ -38,11 +44,12 @@ if(WITH_GPU) add_simple_unittest(RowConvOpTest) add_simple_unittest(BlockExpandOpTest) add_simple_unittest(CropOpTest) - add_simple_unittest(DepthwiseConvOpTest) + add_simple_unittest(SwitchOpTest) endif() add_simple_unittest(Im2ColTest) add_simple_unittest(GemmConvOpTest) +add_simple_unittest(DepthwiseConvOpTest) endif() add_style_check_target(paddle_function ${h_files}) diff --git a/paddle/function/DepthwiseConvOp.cpp b/paddle/function/DepthwiseConvOp.cpp index 490e8d546cbd460217abe95f6291b13fa207faa9..2f3112fe657cd381891dc53c7179e7520911e8c9 100644 --- a/paddle/function/DepthwiseConvOp.cpp +++ b/paddle/function/DepthwiseConvOp.cpp @@ -14,7 +14,6 @@ limitations under the License. */ #include "DepthwiseConvOp.h" #include "ConvOp.h" -#include "GemmFunctor.h" namespace paddle { diff --git a/paddle/function/DepthwiseConvOpGpu.cu b/paddle/function/DepthwiseConvOpGpu.cu index 33463805cbd4746c05548028e0bc4a0e2a90453e..2d722dfcfca0f328edeecf185ea37b8512b91907 100644 --- a/paddle/function/DepthwiseConvOpGpu.cu +++ b/paddle/function/DepthwiseConvOpGpu.cu @@ -13,7 +13,6 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "DepthwiseConvOp.h" -#include "GemmFunctor.h" #include "paddle/math/BaseMatrix.h" namespace paddle { diff --git a/paddle/function/DepthwiseConvOpTest.cpp b/paddle/function/DepthwiseConvOpTest.cpp index f44ae0c342e9536366e2b537694cee81fcb1a6ed..d8e8c889d5c23bf9b2b5fd0b0393395883188fd8 100644 --- a/paddle/function/DepthwiseConvOpTest.cpp +++ b/paddle/function/DepthwiseConvOpTest.cpp @@ -34,4 +34,13 @@ TEST(DepthwiseConv, BackwardFilter) { } #endif +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + +TEST(DepthwiseConv, Forward) { + DepthwiseConvolution( + "GemmConv-CPU", "NeonDepthwiseConv-CPU", forward); +} + +#endif + } // namespace paddle diff --git a/paddle/function/EigenGemm.cpp b/paddle/function/EigenGemm.cpp new file mode 100644 index 0000000000000000000000000000000000000000..b3e666e860d29d89650d48a23cf44917035a02d7 --- /dev/null +++ b/paddle/function/EigenGemm.cpp @@ -0,0 +1,91 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "unsupported/Eigen/CXX11/Tensor" + +namespace paddle { + +template +struct EigenBlasGemm { + typedef Eigen::TensorMap, + Eigen::Aligned> + Matrix; + + static void compute(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) { + Eigen::array sizeA; + if (transA) { + sizeA[0] = K; + sizeA[1] = M; + CHECK_EQ(M, lda); + } else { + sizeA[0] = M; + sizeA[1] = K; + CHECK_EQ(K, lda); + } + Eigen::array sizeB; + if (transB) { + sizeB[0] = N; + sizeB[1] = K; + CHECK_EQ(K, ldb); + } else { + sizeB[0] = K; + sizeB[1] = N; + CHECK_EQ(N, ldb); + } + Eigen::array sizeC; + sizeC[0] = M; + sizeC[1] = N; + CHECK_EQ(N, ldc); + + const Matrix a(const_cast(A), sizeA); + const Matrix b(const_cast(B), sizeB); + Matrix c(C, sizeC); + + typedef typename Eigen::Tensor::DimensionPair DimPair; + Eigen::array dims; + dims[0] = DimPair(1, 0); + dims[0].first = transA ? 0 : 1; + dims[0].second = transB ? 1 : 0; + + Eigen::DefaultDevice device; + if (alpha == T(1) && beta == T(0)) { + c.device(device) = a.contract(b, dims); + } else if (alpha == T(1) && beta == T(1)) { + c.device(device) += a.contract(b, dims); + } else { + c.device(device) = alpha * a.contract(b, dims) + beta * c; + } + } +}; + +#ifdef PADDLE_TYPE_DOUBLE +template struct EigenBlasGemm; +#else +template struct EigenBlasGemm; +#endif + +} // namespace paddle diff --git a/paddle/function/GemmConvOp.cpp b/paddle/function/GemmConvOp.cpp index 0ada4d70a0c7d13f9b5fb1a42eac07fc4c775a87..f8cf4ebea8d724f0291b981647622b63e3d84495 100644 --- a/paddle/function/GemmConvOp.cpp +++ b/paddle/function/GemmConvOp.cpp @@ -85,7 +85,6 @@ public: } Im2ColFunctor im2col; - GemmFunctor gemm; size_t inputOffset = imShape.getElements(); size_t outputOffset = (outputChannels / groups_) * outputHeight * outputWidth; @@ -108,19 +107,19 @@ public: int M = outputChannels / groups_; int N = outputHeight * outputWidth; int K = inputChannels / groups_ * filterHeight * filterWidth; - gemm(CblasNoTrans, - CblasNoTrans, - M, - N, - K, - 1.0f, - filterData + g * filterOffset, - K, - colData, - N, - beta, - outputData + g * outputOffset, - N); + BlasGemm::compute(false, + false, + M, + N, + K, + 1.0f, + filterData + g * filterOffset, + K, + colData, + N, + beta, + outputData + g * outputOffset, + N); } inputData += inputChannels * inputHeight * inputWidth; outputData += outputChannels * outputHeight * outputWidth; @@ -188,8 +187,6 @@ public: } Col2ImFunctor col2im; - GemmFunctor gemm; - size_t inputOffset = imShape.getElements(); size_t outputOffset = (outputChannels / groups_) * outputHeight * outputWidth; @@ -205,19 +202,19 @@ public: colData = inputGrad + g * inputOffset; scale = 1.0f; } - gemm(CblasTrans, - CblasNoTrans, - M, - N, - K, - 1.0f, - filterData + g * filterOffset, - M, - outputGrad + g * outputOffset, - N, - scale, - colData, - N); + BlasGemm::compute(true, + false, + M, + N, + K, + 1.0f, + filterData + g * filterOffset, + M, + outputGrad + g * outputOffset, + N, + scale, + colData, + N); if (needIm2col) { col2im(inputGrad + g * inputOffset, imShape, @@ -299,7 +296,6 @@ public: } Im2ColFunctor im2col; - GemmFunctor gemm; size_t inputOffset = imShape.getElements(); size_t outputOffset = (outputChannels / groups_) * outputHeight * outputWidth; @@ -321,19 +317,19 @@ public: int M = outputChannels / groups_; int K = outputHeight * outputWidth; int N = inputChannels / groups_ * filterHeight * filterWidth; - gemm(CblasNoTrans, - CblasTrans, - M, - N, - K, - 1.0f, - outputGrad + g * outputOffset, - K, - colData, - K, - i == 0 ? beta : 1.0f, - filterGrad + g * filterOffset, - N); + BlasGemm::compute(false, + true, + M, + N, + K, + 1.0f, + outputGrad + g * outputOffset, + K, + colData, + K, + i == 0 ? beta : 1.0f, + filterGrad + g * filterOffset, + N); } inputData += inputChannels * inputHeight * inputWidth; outputGrad += outputChannels * outputHeight * outputWidth; diff --git a/paddle/function/GemmFunctor.cpp b/paddle/function/GemmFunctor.cpp new file mode 100644 index 0000000000000000000000000000000000000000..9e25ee58a12490a1454436b3fe4a89176478d5c0 --- /dev/null +++ b/paddle/function/GemmFunctor.cpp @@ -0,0 +1,90 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "GemmFunctor.h" +#include "paddle/math/MathFunctions.h" + +namespace paddle { + +template +struct BlasGemm { + static void compute(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) { +#ifdef PADDLE_USE_EIGEN_FOR_BLAS + EigenBlasGemm::compute( + transA, transB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc); +#else + gemm(transA == false ? CblasNoTrans : CblasTrans, + transB == false ? CblasNoTrans : CblasTrans, + M, + N, + K, + alpha, + A, + lda, + B, + ldb, + beta, + C, + ldc); +#endif + } +}; + +template +struct BlasGemm { + static void compute(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) { + hl_matrix_mul((T*)A, + transA == false ? HPPL_OP_N : HPPL_OP_T, + (T*)B, + transB == false ? HPPL_OP_N : HPPL_OP_T, + C, + M, + N, + K, + alpha, + beta, + lda, + ldb, + ldc); + } +}; + +template struct BlasGemm; +template struct BlasGemm; + +} // namespace paddle diff --git a/paddle/function/GemmFunctor.h b/paddle/function/GemmFunctor.h index d5db5cf5e7a855d89b262fe8cf42aa2c55f419f1..0809953b4eb17c25eadcce7f474a3dab0469bba1 100644 --- a/paddle/function/GemmFunctor.h +++ b/paddle/function/GemmFunctor.h @@ -14,7 +14,7 @@ limitations under the License. */ #pragma once -#include "paddle/math/MathFunctions.h" +#include "TensorType.h" namespace paddle { @@ -24,73 +24,42 @@ namespace paddle { // of MatMulFunction, we need to consider the reconstruction of hl_matrix_mul // interface. template -class GemmFunctor { -public: - void operator()(const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE 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); +struct BlasGemm { + static void compute(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); }; +// TODO(hedaoyuan): Since the definition of the real type in the Paddle +// conflicts with the Eigen library, so compile the Eigen code can not +// include the Paddle header file. And need an EigenBlasGemm template class +// that does not contain the DeviceType parameter. +// I will fix this problem and merge BlasGemm and EigenBlasGemm into one. template -class GemmFunctor { -public: - void operator()(const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE 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) { - gemm(transA, TransB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc); - } -}; - -template -class GemmFunctor { -public: - void operator()(const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE 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) { - hl_matrix_mul((T*)A, - transA == CblasNoTrans ? HPPL_OP_N : HPPL_OP_T, - (T*)B, - TransB == CblasNoTrans ? HPPL_OP_N : HPPL_OP_T, - C, - M, - N, - K, - alpha, - beta, - lda, - ldb, - ldc); - } +struct EigenBlasGemm { + static void compute(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); }; } // namespace paddle diff --git a/paddle/function/GruFunctor.h b/paddle/function/GruFunctor.h new file mode 100644 index 0000000000000000000000000000000000000000..9f6392198ea360502f313cbe15dfae46ece69758 --- /dev/null +++ b/paddle/function/GruFunctor.h @@ -0,0 +1,159 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT 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 "GemmFunctor.h" +#include "hl_cpu_gru.cuh" + +namespace paddle { + +template +struct GruFunctor { + template + static void compute(OpResetOutput opResetOutput, + OpFinalOutput opFinalOutput, + hl_gru_value value, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate) { +#ifndef __NVCC__ + if (value.prevOutValue) { + BlasGemm::compute(false, + false, + batchSize, + 2 * frameSize, + frameSize, + 1, + value.prevOutValue, + frameSize, + value.gateWeight, + frameSize * 2, + 1, + value.gateValue, + frameSize * 3); + } + + forward_reset_output( + opResetOutput, value, frameSize, batchSize, active_gate); + + if (value.prevOutValue) { + BlasGemm::compute(false, + false, + batchSize, + frameSize, + frameSize, + 1, + value.resetOutputValue, + frameSize, + value.stateWeight, + frameSize, + 1, + value.gateValue + frameSize * 2, + frameSize * 3); + } + + forward_final_output( + opFinalOutput, value, frameSize, batchSize, active_node); +#endif + } +}; + +template +struct GruGradFunctor { + template + static void compute(OpStateGrad opStateGrad, + OpResetGrad opResetGrad, + hl_gru_value value, + hl_gru_grad grad, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate) { +#ifndef __NVCC__ + backward_state_grad( + opStateGrad, value, grad, frameSize, batchSize, active_node); + + if (value.prevOutValue && grad.prevOutGrad) { + BlasGemm::compute(false, + true, + batchSize, + frameSize, + frameSize, + 1, + grad.gateGrad + frameSize * 2, + frameSize * 3, + value.stateWeight, + frameSize, + 0, + grad.resetOutputGrad, + frameSize); + + if (grad.stateWeightGrad) { + BlasGemm::compute(true, + false, + frameSize, + frameSize, + batchSize, + 1, + value.resetOutputValue, + frameSize, + grad.gateGrad + frameSize * 2, + frameSize * 3, + 1, + grad.stateWeightGrad, + frameSize); + } + } + + backward_reset_grad( + opResetGrad, value, grad, frameSize, batchSize, active_gate); + + if (grad.prevOutGrad && value.prevOutValue) { + BlasGemm::compute(false, + true, + batchSize, + frameSize, + frameSize * 2, + 1, + grad.gateGrad, + frameSize * 3, + value.gateWeight, + frameSize * 2, + 1, + grad.prevOutGrad, + frameSize); + + if (grad.gateWeightGrad) { + BlasGemm::compute(true, + false, + frameSize, + frameSize * 2, + batchSize, + 1, + value.prevOutValue, + frameSize, + grad.gateGrad, + frameSize * 3, + 1, + grad.gateWeightGrad, + frameSize * 2); + } + } +#endif + } +}; + +} // namespace paddle diff --git a/paddle/function/Im2Col.h b/paddle/function/Im2Col.h index 48e2e32f9256fb49c67ba25e9b5a47d72499758b..1e0cff436ff60d5a029e89657d00af2b0bf8b454 100644 --- a/paddle/function/Im2Col.h +++ b/paddle/function/Im2Col.h @@ -16,6 +16,7 @@ limitations under the License. */ #include "TensorShape.h" #include "TensorType.h" +#include "neon/neon_util.h" namespace paddle { diff --git a/paddle/function/MulOp.cpp b/paddle/function/MulOp.cpp index 91b4b8ed91b6055babcfbab8f7adb2c55e2747d0..25e41edad54bec0f76a3de4799fab14241407272 100644 --- a/paddle/function/MulOp.cpp +++ b/paddle/function/MulOp.cpp @@ -13,18 +13,10 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "MulOp.h" -/// todo(tianbing), delete it -#include -#include "paddle/math/MathFunctions.h" +#include "GemmFunctor.h" #include "paddle/math/SIMDFunctions.h" #include "paddle/utils/ThreadLocal.h" -#ifndef PADDLE_TYPE_DOUBLE -#define GEMM paddle::gemm -#else -#define GEMM paddle::gemm -#endif - namespace { inline void vecAddTo(real* a, const real* b, real scaleB, size_t len) { for (unsigned int i = 0; i < len; ++i) { @@ -114,19 +106,20 @@ void MulOp(CpuMatrix& out, real scaleT, bool aTrans, bool bTrans) { - GEMM(aTrans ? CblasTrans : CblasNoTrans, - bTrans ? CblasTrans : CblasNoTrans, - out.getHeight(), - out.getWidth(), - !aTrans ? a.getWidth() : a.getHeight(), - scaleAB, - a.getData(), - a.getStride(), - b.getData(), - b.getStride(), - scaleT, - out.getData(), - out.getStride()); + BlasGemm::compute( + aTrans, + bTrans, + out.getHeight(), + out.getWidth(), + !aTrans ? a.getWidth() : a.getHeight(), + scaleAB, + a.getData(), + a.getStride(), + b.getData(), + b.getStride(), + scaleT, + out.getData(), + out.getStride()); } /// dense matrix (+)= sparse matrix * dense matrix diff --git a/paddle/function/SwitchOp.cpp b/paddle/function/SwitchOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..01e252a8dc0cd5fa1e964efa01d04cf282b3dfe7 --- /dev/null +++ b/paddle/function/SwitchOp.cpp @@ -0,0 +1,140 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "SwitchOp.h" +#include "paddle/math/Vector.h" + +namespace paddle { + +template <> +void NCHW2NHWC(real* outputs, + const real* inputs, + const int num, + const int inC, + const int inH, + const int inW, + const int argType) { + for (int n = 0; n < num; ++n) { + for (int c = 0; c < inC; ++c) { + for (int h = 0; h < inH; ++h) { + for (int w = 0; w < inW; ++w) { + if (argType == ADD_TO) { + outputs[((n * inH + h) * inW + w) * inC + c] += *(inputs++); + } else { + outputs[((n * inH + h) * inW + w) * inC + c] = *(inputs++); + } + } + } + } + } +} + +template <> +void NHWC2NCHW(real* outputs, + const real* inputs, + const int num, + const int inH, + const int inW, + const int inC, + const int argType) { + for (int n = 0; n < num; ++n) { + for (int h = 0; h < inH; ++h) { + for (int w = 0; w < inW; ++w) { + for (int c = 0; c < inC; ++c) { + if (argType == ADD_TO) { + outputs[((n * inC + c) * inH + h) * inW + w] += *(inputs++); + } else { + outputs[((n * inC + c) * inH + h) * inW + w] = *(inputs++); + } + } + } + } + } +} + +/** + * \brief Switch dimension order of image input. + * The input and output is a 4D tensor. Switch order + * 'batch_size,channels, height, width' to + * order 'batch_size, height, width, channels'. + * + * Argument in this Function: + * \param inputs input data with order 'batch_size,channels, height, width'. + * \param outputs output data with order 'batch_size, height, width, channels'. + */ +template +class NCHW2NHWCFunc : public FunctionBase { +public: + void init(const FuncConfig& config) override {} + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + + size_t num = inputs[0].shape()[0]; + size_t inC = inputs[0].shape()[1]; + size_t inH = inputs[0].shape()[2]; + size_t inW = inputs[0].shape()[3]; + NCHW2NHWC(outputs[0].data(), + inputs[0].data(), + num, + inC, + inH, + inW, + outputs[0].getArgType()); + } +}; + +/** + * \brief Switch dimension order of image input. + * The input and output is a 4D tensor. Switch order + * 'batch_size, height, width, channels' to + * order 'batch_size, channels, height, width'. + * + * Argument in this Function: + * \param inputs input data with order 'batch_size, height, width, channels'. + * \param outputs output data with order 'batch_size, channels, height, width'. + */ +template +class NHWC2NCHWFunc : public FunctionBase { +public: + void init(const FuncConfig& config) override {} + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + + size_t num = inputs[0].shape()[0]; + size_t inH = inputs[0].shape()[1]; + size_t inW = inputs[0].shape()[2]; + size_t inC = inputs[0].shape()[3]; + + NHWC2NCHW(outputs[0].data(), + inputs[0].data(), + num, + inH, + inW, + inC, + outputs[0].getArgType()); + } +}; + +REGISTER_TYPED_FUNC(NCHW2NHWC, CPU, NCHW2NHWCFunc); +REGISTER_TYPED_FUNC(NHWC2NCHW, CPU, NHWC2NCHWFunc); +#ifndef PADDLE_ONLY_CPU +REGISTER_TYPED_FUNC(NCHW2NHWC, GPU, NCHW2NHWCFunc); +REGISTER_TYPED_FUNC(NHWC2NCHW, GPU, NHWC2NCHWFunc); +#endif + +} // namespace paddle diff --git a/paddle/function/SwitchOp.h b/paddle/function/SwitchOp.h new file mode 100644 index 0000000000000000000000000000000000000000..e4c1c3ac922f88c3e5424b5943082810aabfacdb --- /dev/null +++ b/paddle/function/SwitchOp.h @@ -0,0 +1,66 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Function.h" + +namespace paddle { + +/** + * \brief This funtion switch dimension order of image input. + * The input and output is a 4D tensor. Switch order 'batch_size, + *channels, height, width' to + * order 'batch_size, height, width, channels'. + * + * \param[out] outputs save results. + * \param[in] inputs input data. + * \param[in] num batch size of input data. + * \param[in] inC channel number of input data. + * \param[in] inH height of input data. + * \param[in] inH with of input data. + * \param[in] argType type of output argument. + */ +template +void NCHW2NHWC(real* outputs, + const real* inputs, + const int num, + const int inC, + const int inH, + const int inW, + const int argtype); + +/** + * \brief This funtion switch dimension order of image input. + * The input and output is a 4D tensor. Switch order 'batch_size, + *height, width, channels' to + * order 'batch_size, channels, height, width'. + * + * \param[out] inGrad gradients of previous layer. + * \param[in] outGrad output gradients. + * \param[in] num batch size of input data. + * \param[in] inH height of input data. + * \param[in] inW with of input data. + * \param[in] inC channel number of input data. + * \param[in] argType type of output argument. + */ +template +void NHWC2NCHW(real* inGrad, + const real* outGrad, + const int num, + const int inH, + const int inW, + const int inC, + const int argType); +} // namespace paddle diff --git a/paddle/function/SwitchOpGpu.cu b/paddle/function/SwitchOpGpu.cu new file mode 100644 index 0000000000000000000000000000000000000000..45390a56c3f776ec18a65a6ba2f7149a7a6ef6c3 --- /dev/null +++ b/paddle/function/SwitchOpGpu.cu @@ -0,0 +1,98 @@ +/* Copyright (c) 2016 Paddle + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "SwitchOp.h" +#include "hl_base.h" + +namespace paddle { + +__global__ void KeNCHW2NHWC(real* outputs, + const real* inputs, + int inC, + int inH, + int inW, + int nthreads, + int argType) { + const int idx = threadIdx.x + blockIdx.x * blockDim.x; + if (idx < nthreads) { + const int w = idx % inW; + const int h = (idx / inW) % inH; + const int c = (idx / inW / inH) % inC; + const int n = idx / inW / inH / inC; + + const int off = ((n * inH + h) * inW + w) * inC + c; + if (argType == ADD_TO) { + outputs[off] += inputs[idx]; + } else { + outputs[off] = inputs[idx]; + } + } +} + +template <> +void NCHW2NHWC(real* outputs, + const real* inputs, + const int num, + const int inC, + const int inH, + const int inW, + const int argType) { + size_t nth = num * inC * inH * inW; + int blockSize = 1024; + int gridSize = (nth + 1024 - 1) / 1024; + KeNCHW2NHWC<<>>( + outputs, inputs, inC, inH, inW, nth, argType); + CHECK_SYNC("NCHW2NHWC"); +} + +__global__ void KeNHWC2NCHW(real* outputs, + const real* inputs, + int inH, + int inW, + int inC, + int nthreads, + int argType) { + const int idx = threadIdx.x + blockIdx.x * blockDim.x; + if (idx < nthreads) { + const int c = idx % inC; + const int w = (idx / inC) % inW; + const int h = (idx / inC / inW) % inH; + const int n = idx / inW / inH / inC; + + const int off = ((n * inC + c) * inH + h) * inW + w; + if (argType == ADD_TO) { + outputs[off] += inputs[idx]; + } else { + outputs[off] = inputs[idx]; + } + } +} + +template <> +void NHWC2NCHW(real* outputs, + const real* inputs, + const int num, + const int inH, + const int inW, + const int inC, + const int argType) { + int nth = num * inC * inH * inW; + int blockSize = 1024; + int gridSize = (nth + 1024 - 1) / 1024; + KeNHWC2NCHW<<>>( + outputs, inputs, inH, inW, inC, nth, argType); + CHECK_SYNC("NHWC2NCHW"); +} + +} // namespace paddle diff --git a/paddle/function/SwitchOpTest.cpp b/paddle/function/SwitchOpTest.cpp new file mode 100644 index 0000000000000000000000000000000000000000..03b0dd66ddcbab713969ed747601ecb1b2eb7955 --- /dev/null +++ b/paddle/function/SwitchOpTest.cpp @@ -0,0 +1,44 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "FunctionTest.h" + +namespace paddle { + +TEST(Pad, real) { + for (size_t numSamples : {1, 4, 8, 16}) { + for (size_t channels : {1, 4, 8, 16}) { + for (size_t imgSizeH : {1, 4, 8, 16}) { + for (size_t imgSizeW : {1, 4, 8, 16}) { + VLOG(3) << " numSamples=" << numSamples << " channels=" << channels + << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW; + for (bool test_grad : {true, false}) { + CpuGpuFuncCompare compare(test_grad ? "NHWC2NCHW" : "NCHW2NHWC", + FuncConfig()); + TensorShape inDims{numSamples, channels, imgSizeH, imgSizeW}; + TensorShape outDims{numSamples, imgSizeH, imgSizeW, channels}; + compare.addInputs( + BufferArg(VALUE_TYPE_FLOAT, test_grad ? outDims : inDims)); + compare.addOutputs(BufferArg( + VALUE_TYPE_FLOAT, test_grad ? inDims : outDims, ASSIGN_TO)); + compare.run(); + } + } + } + } + } +} + +} // namespace paddle diff --git a/paddle/function/neon/NeonDepthwiseConv.cpp b/paddle/function/neon/NeonDepthwiseConv.cpp new file mode 100644 index 0000000000000000000000000000000000000000..18126152ea0b4ebfe4ec5c8084479787814ed173 --- /dev/null +++ b/paddle/function/neon/NeonDepthwiseConv.cpp @@ -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 "NeonDepthwiseConv.h" +#include "paddle/function/ConvOp.h" + +namespace paddle { + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + +template +class NeonDepthwiseConvFunction : public ConvFunctionBase { +public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + check(inputs, outputs); + + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + + int batchSize = input[0]; + int inputChannels = input[1]; + int inputHeight = input[2]; + int inputWidth = input[3]; + int filterHeight = getFilterHeight(filter); + int filterWidth = getFilterWidth(filter); + int outputChannels = output[1]; + int outputHeight = output[2]; + int outputWidth = output[3]; + int filterMultiplier = outputChannels / groups_; + CHECK_EQ(inputChannels, groups_); + + // only support strideH() == strideW() and filterHeight == filterWidth. + CHECK_EQ(strideH(), strideW()); + CHECK_EQ(filterHeight, filterWidth); + + float* inputData = inputs[0].data(); + float* filterData = inputs[1].data(); + float* outputData = outputs[0].data(); + + // padding the input + float* inputPadding = inputData; + int padInputHeight = inputHeight + 2 * paddingH(); + int padInputWidth = inputWidth + 2 * paddingW(); + if (paddingH() > 0 || paddingW() > 0) { + int newSize = batchSize * inputChannels * padInputHeight * padInputWidth; + resizeBuffer(newSize); + inputPadding = reinterpret_cast(memory_->getBuf()); + neon::Padding::run(inputData, + inputPadding, + batchSize * inputChannels, + inputHeight, + inputWidth, + padInputHeight, + padInputWidth); + } + + std::function + DepthWiseConv; + + if (filterWidth == 3 && strideW() == 1) { + DepthWiseConv = neon::DepthwiseConvKernel<3, 1>::run; + } else if (filterWidth == 3 && strideW() == 2) { + DepthWiseConv = neon::DepthwiseConvKernel<3, 2>::run; + } else if (filterWidth == 4 && strideW() == 1) { + DepthWiseConv = neon::DepthwiseConvKernel<4, 1>::run; + } else if (filterWidth == 4 && strideW() == 2) { + DepthWiseConv = neon::DepthwiseConvKernel<4, 2>::run; + } else { + LOG(FATAL) << "Not supported"; + } + + for (int i = 0; i < batchSize; i++) { + DepthWiseConv(inputPadding, + filterData, + padInputHeight, + padInputWidth, + outputChannels, + outputHeight, + outputWidth, + filterMultiplier, + outputData); + inputPadding += inputChannels * padInputHeight * padInputWidth; + outputData += outputChannels * outputHeight * outputWidth; + } + } +}; + +#ifndef PADDLE_TYPE_DOUBLE +REGISTER_TYPED_FUNC(NeonDepthwiseConv, CPU, NeonDepthwiseConvFunction); +#endif + +#endif + +} // namespace paddle diff --git a/paddle/function/neon/NeonDepthwiseConv.h b/paddle/function/neon/NeonDepthwiseConv.h new file mode 100644 index 0000000000000000000000000000000000000000..aefeea78badbca3d0d09e292e4e1e148618f8ac6 --- /dev/null +++ b/paddle/function/neon/NeonDepthwiseConv.h @@ -0,0 +1,631 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT 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 "neon_util.h" + +namespace paddle { + +namespace neon { + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + +template +struct DepthwiseConvKernel {}; + +inline float32_t conv3x3(float32x4_t r0, + float32x4_t r1, + float32x4_t 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); +} + +inline float32_t conv4x4(float32x4_t r0, + float32x4_t r1, + float32x4_t r2, + float32x4_t r3, + float32x4_t k0, + float32x4_t k1, + float32x4_t k2, + float32x4_t k3) { + float32x4_t tmp; + tmp = vmulq_f32(r0, k0); + tmp = vmlaq_f32(tmp, r1, k1); + tmp = vmlaq_f32(tmp, r2, k2); + tmp = vmlaq_f32(tmp, r3, k3); + return vaddvq_f32(tmp); +} + +/** + * Each step calculates four elements of the output. + * First step: + * R0[0, 1, 2, 3...] * K[0][0] + * R0[1, 2, 3, 4...] * K[0][1] + * R0[2, 3, 4, 5...] * K[0][2] + * R1[0, 1, 2, 3...] * K[1][0] + * R1[1, 2, 3, 4...] * K[1][1] + * R1[2, 3, 4, 5...] * K[1][2] + * R2[0, 1, 2, 3...] * K[2][0] + * R2[1, 2, 3, 4...] * K[2][1] + * + R2[2, 3, 4, 5...] * K[2][2] + * ------------------------------ + * Output[0, 1, 2, 3] + */ +template <> +struct DepthwiseConvKernel<3, 1> { + static void run(const float* inputData, + const float* filterData, + int inputHeight, + int inputWidth, + int outputChannels, + int outputHeight, + int outputWidth, + int filterMultiplier, + float* outputData) { + const int steps = outputWidth >> 2; + const int remain = outputWidth & 3; + for (int c = 0; c < outputChannels; c++, filterData += 9) { + // Load the filters + float32x4_t k[3]; + k[0] = vld1q_f32(filterData); + k[1] = vld1q_f32(filterData + 3); + k[2] = vld1q_f32(filterData + 6); + k[0] = vsetq_lane_f32(0.f, k[0], 3); + k[1] = vsetq_lane_f32(0.f, k[1], 3); + k[2] = vsetq_lane_f32(0.f, k[2], 3); + + const float* r0 = + inputData + (c / filterMultiplier) * (inputHeight * inputWidth); + const float* r1 = r0 + inputWidth; + const float* r2 = r0 + inputWidth * 2; + float32x4_t input[3][3]; + for (int h = 0; h < outputHeight; h++) { + for (int s = 0; s < steps; s++) { + // Load the inputs + float32x4_t tmp; + input[0][0] = vld1q_f32(r0); + tmp = vld1q_f32(r0 + 4); + input[0][1] = vextq_f32(input[0][0], tmp, 1); + input[0][2] = vextq_f32(input[0][0], tmp, 2); + input[1][0] = vld1q_f32(r1); + tmp = vld1q_f32(r1 + 4); + input[1][1] = vextq_f32(input[1][0], tmp, 1); + input[1][2] = vextq_f32(input[1][0], tmp, 2); + input[2][0] = vld1q_f32(r2); + tmp = vld1q_f32(r2 + 4); + input[2][1] = vextq_f32(input[2][0], tmp, 1); + input[2][2] = vextq_f32(input[2][0], tmp, 2); + + float32x4_t tmp1 = vdupq_n_f32(0.f); + float32x4_t tmp2 = vdupq_n_f32(0.f); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][0], k[1], 0); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][1], k[1], 1); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][2], k[1], 2); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); + tmp1 = vaddq_f32(tmp1, tmp2); + + vst1q_f32(outputData, tmp1); + r0 += 4; + r1 += 4; + r2 += 4; + outputData += 4; + } + + 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]); + r0++; + r1++; + r2++; + outputData++; + } + + r0 += 2; + r1 += 2; + r2 += 2; + } + } + } +}; + +/** + * Each step calculates four elements of the output. + * First step: + * R0[0, 2, 4, 6...] * K[0][0] + * R0[1, 3, 5, 7...] * K[0][1] + * R0[2, 4, 6, 8...] * K[0][2] + * R1[0, 2, 4, 6...] * K[1][0] + * R1[1, 3, 5, 7...] * K[1][1] + * R1[2, 4, 6, 8...] * K[1][2] + * R2[0, 2, 4, 6...] * K[2][0] + * R2[1, 3, 5, 7...] * K[2][1] + * R2[2, 4, 6, 8...] * K[2][2] + * ------------------------------ + * Output[0, 1, 2, 3] + */ +template <> +struct DepthwiseConvKernel<3, 2> { + static void run(const float* inputData, + const float* filterData, + int inputHeight, + int inputWidth, + int outputChannels, + int outputHeight, + int outputWidth, + int filterMultiplier, + float* outputData) { + const int steps = outputWidth >> 2; + const int remain = outputWidth & 3; + for (int c = 0; c < outputChannels; c++, filterData += 9) { + // Load the filters + float32x4_t k[3]; + k[0] = vld1q_f32(filterData); + k[1] = vld1q_f32(filterData + 3); + k[2] = vld1q_f32(filterData + 6); + k[0] = vsetq_lane_f32(0.f, k[0], 3); + k[1] = vsetq_lane_f32(0.f, k[1], 3); + k[2] = vsetq_lane_f32(0.f, k[2], 3); + + const float* start = + inputData + (c / filterMultiplier) * (inputHeight * inputWidth); + float32x4_t input[3][3]; + for (int h = 0; h < outputHeight; h++) { + const float* r0 = start + 2 * h * inputWidth; + const float* r1 = start + (2 * h + 1) * inputWidth; + const float* r2 = start + (2 * h + 2) * inputWidth; + for (int s = 0; s < steps; s++) { + // Load the inputs + float32x4_t data1; + float32x4x2_t data2; + + data2 = vld2q_f32(r0); + input[0][0] = data2.val[0]; + input[0][1] = data2.val[1]; + data1 = vld1q_f32(r0 + 8); + input[0][2] = vextq_f32(data2.val[0], data1, 1); + + data2 = vld2q_f32(r1); + input[1][0] = data2.val[0]; + input[1][1] = data2.val[1]; + data1 = vld1q_f32(r1 + 8); + input[1][2] = vextq_f32(data2.val[0], data1, 1); + + data2 = vld2q_f32(r2); + input[2][0] = data2.val[0]; + input[2][1] = data2.val[1]; + data1 = vld1q_f32(r2 + 8); + input[2][2] = vextq_f32(data2.val[0], data1, 1); + + float32x4_t tmp1 = vdupq_n_f32(0.f); + float32x4_t tmp2 = vdupq_n_f32(0.f); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][0], k[1], 0); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][1], k[1], 1); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][2], k[1], 2); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); + tmp1 = vaddq_f32(tmp1, tmp2); + + vst1q_f32(outputData, tmp1); + r0 += 8; + r1 += 8; + r2 += 8; + outputData += 4; + } + + 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]); + r0 += 2; + r1 += 2; + r2 += 2; + outputData++; + } + } + } + } +}; + +/** + * Each step calculates four elements of the output. + */ +template <> +struct DepthwiseConvKernel<4, 1> { + static void run(const float* inputData, + const float* filterData, + int inputHeight, + int inputWidth, + int outputChannels, + int outputHeight, + int outputWidth, + int filterMultiplier, + float* outputData) { + const int steps = outputWidth >> 2; + const int remain = outputWidth & 3; + for (int c = 0; c < outputChannels; c++, filterData += 16) { + // Load the filters + float32x4_t k[4]; + k[0] = vld1q_f32(filterData); + k[1] = vld1q_f32(filterData + 4); + k[2] = vld1q_f32(filterData + 8); + k[3] = vld1q_f32(filterData + 12); + + const float* r0 = + inputData + (c / filterMultiplier) * (inputHeight * inputWidth); + const float* r1 = r0 + inputWidth; + const float* r2 = r0 + inputWidth * 2; + const float* r3 = r0 + inputWidth * 3; + float32x4_t input[4][4]; + for (int h = 0; h < outputHeight; h++) { + for (int s = 0; s < steps; s++) { + // Load the inputs + float32x4_t tmp; + input[0][0] = vld1q_f32(r0); + tmp = vld1q_f32(r0 + 4); + input[0][1] = vextq_f32(input[0][0], tmp, 1); + input[0][2] = vextq_f32(input[0][0], tmp, 2); + input[0][3] = vextq_f32(input[0][0], tmp, 3); + + input[1][0] = vld1q_f32(r1); + tmp = vld1q_f32(r1 + 4); + input[1][1] = vextq_f32(input[1][0], tmp, 1); + input[1][2] = vextq_f32(input[1][0], tmp, 2); + input[1][3] = vextq_f32(input[1][0], tmp, 3); + + input[2][0] = vld1q_f32(r2); + tmp = vld1q_f32(r2 + 4); + input[2][1] = vextq_f32(input[2][0], tmp, 1); + input[2][2] = vextq_f32(input[2][0], tmp, 2); + input[2][3] = vextq_f32(input[2][0], tmp, 3); + + input[3][0] = vld1q_f32(r3); + tmp = vld1q_f32(r3 + 4); + input[3][1] = vextq_f32(input[3][0], tmp, 1); + input[3][2] = vextq_f32(input[3][0], tmp, 2); + input[3][3] = vextq_f32(input[3][0], tmp, 3); + + float32x4_t tmp1 = vdupq_n_f32(0.f); + float32x4_t tmp2 = vdupq_n_f32(0.f); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3); + tmp1 = vaddq_f32(tmp1, tmp2); + + vst1q_f32(outputData, tmp1); + r0 += 4; + r1 += 4; + r2 += 4; + r3 += 4; + outputData += 4; + } + + 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); + float32x4_t i3 = vld1q_f32(r3); + *outputData = conv4x4(i0, i1, i2, i3, k[0], k[1], k[2], k[3]); + r0++; + r1++; + r2++; + r3++; + outputData++; + } + + r0 += 3; + r1 += 3; + r2 += 3; + r3 += 3; + } + } + } +}; + +/** + * Each step calculates four elements of the output. + */ +template <> +struct DepthwiseConvKernel<4, 2> { + static void run(const float* inputData, + const float* filterData, + int inputHeight, + int inputWidth, + int outputChannels, + int outputHeight, + int outputWidth, + int filterMultiplier, + float* outputData) { + const int steps = outputWidth >> 2; + const int remain = outputWidth & 3; + for (int c = 0; c < outputChannels; c++, filterData += 16) { + // Load the filters + float32x4_t k[4]; + k[0] = vld1q_f32(filterData); + k[1] = vld1q_f32(filterData + 4); + k[2] = vld1q_f32(filterData + 8); + k[3] = vld1q_f32(filterData + 12); + + const float* start = + inputData + (c / filterMultiplier) * (inputHeight * inputWidth); + float32x4_t input[4][4]; + for (int h = 0; h < outputHeight; h++) { + const float* r0 = start + 2 * h * inputWidth; + const float* r1 = start + (2 * h + 1) * inputWidth; + const float* r2 = start + (2 * h + 2) * inputWidth; + const float* r3 = start + (2 * h + 3) * inputWidth; + for (int s = 0; s < steps; s++) { + // Load the inputs + float32x4x2_t data1; + float32x4x2_t data2; + + data1 = vld2q_f32(r0); + data2 = vld2q_f32(r0 + 8); + input[0][0] = data1.val[0]; + input[0][1] = data1.val[1]; + input[0][2] = vextq_f32(data1.val[0], data2.val[0], 1); + input[0][3] = vextq_f32(data1.val[1], data2.val[1], 1); + + data1 = vld2q_f32(r1); + data2 = vld2q_f32(r1 + 8); + input[1][0] = data1.val[0]; + input[1][1] = data1.val[1]; + input[1][2] = vextq_f32(data1.val[0], data2.val[0], 1); + input[1][3] = vextq_f32(data1.val[1], data2.val[1], 1); + + data1 = vld2q_f32(r2); + data2 = vld2q_f32(r2 + 8); + input[2][0] = data1.val[0]; + input[2][1] = data1.val[1]; + input[2][2] = vextq_f32(data1.val[0], data2.val[0], 1); + input[2][3] = vextq_f32(data1.val[1], data2.val[1], 1); + + data1 = vld2q_f32(r3); + data2 = vld2q_f32(r3 + 8); + input[3][0] = data1.val[0]; + input[3][1] = data1.val[1]; + input[3][2] = vextq_f32(data1.val[0], data2.val[0], 1); + input[3][3] = vextq_f32(data1.val[1], data2.val[1], 1); + + float32x4_t tmp1 = vdupq_n_f32(0.f); + float32x4_t tmp2 = vdupq_n_f32(0.f); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3); + tmp1 = vaddq_f32(tmp1, tmp2); + + vst1q_f32(outputData, tmp1); + r0 += 8; + r1 += 8; + r2 += 8; + r3 += 8; + outputData += 4; + } + + 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); + float32x4_t i3 = vld1q_f32(r3); + *outputData = conv4x4(i0, i1, i2, i3, k[0], k[1], k[2], k[3]); + r0 += 2; + r1 += 2; + r2 += 2; + r3 += 2; + outputData++; + } + } + } + } +}; + +template +struct Padding { + static void run(const T* input, + T* inputPadding, + int channels, + int inputHeight, + int inputWidth, + int padInputHeight, + int padInputWidth) { + const int paddingHeight = (padInputHeight - inputHeight) / 2; + const int paddingWidth = (padInputWidth - inputWidth) / 2; + for (int c = 0; c < channels; c++) { + if (paddingHeight > 0) { + memset(inputPadding, 0, padInputWidth * paddingHeight * sizeof(T)); + inputPadding += padInputWidth * paddingHeight; + } + + for (int i = 0; i < inputHeight; i++) { + // padding head + for (int j = 0; j < paddingWidth; j++) { + *inputPadding++ = T(0); + } + + memcpy(inputPadding, input, inputWidth * sizeof(T)); + inputPadding += inputWidth; + input += inputWidth; + + // padding tail + for (int j = 0; j < paddingWidth; j++) { + *inputPadding++ = T(0); + } + } + + if (paddingHeight > 0) { + memset(inputPadding, 0, padInputWidth * paddingHeight * sizeof(T)); + inputPadding += padInputWidth * paddingHeight; + } + } + } +}; + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) +template <> +struct Padding { + static void run(const float* input, + float* inputPadding, + int channels, + int inputHeight, + int inputWidth, + int padInputHeight, + int padInputWidth) { + const int paddingHeight = (padInputHeight - inputHeight) / 2; + const int paddingWidth = (padInputWidth - inputWidth) / 2; + for (int c = 0; c < channels; c++) { + if (paddingHeight > 0) { + memset(inputPadding, 0, padInputWidth * paddingHeight * sizeof(float)); + inputPadding += padInputWidth * paddingHeight; + } + + for (int i = 0; i < inputHeight; i++) { + // padding head + for (int j = 0; j < paddingWidth; j++) { + *inputPadding++ = float(0); + } + + int step = inputWidth >> 2; + int remain = inputWidth & 3; + for (int s = 0; s < step; s++) { + float32x4_t s0 = vld1q_f32(input); + vst1q_f32(inputPadding, s0); + input += 4; + inputPadding += 4; + } + for (int r = 0; r < remain; r++) { + *inputPadding++ = *input++; + } + + // padding tail + for (int j = 0; j < paddingWidth; j++) { + *inputPadding++ = float(0); + } + } + + if (paddingHeight > 0) { + memset(inputPadding, 0, padInputWidth * paddingHeight * sizeof(float)); + inputPadding += padInputWidth * paddingHeight; + } + } + } +}; + +// for stride is 2 +struct StridePadding { + static void run(const float* input, + float* inputPadding, + int channels, + int inputHeight, + int inputWidth, + int padInputHeight, + int padInputWidth) { + const int paddingHeight = (padInputHeight - (inputHeight * 2 - 1)) / 2; + const int paddingWidth = (padInputWidth - (inputWidth * 2 - 1)) / 2; + for (int c = 0; c < channels; c++) { + if (paddingHeight > 0) { + memset(inputPadding, 0, padInputWidth * paddingHeight * sizeof(float)); + inputPadding += padInputWidth * paddingHeight; + } + + for (int i = 0; i < inputHeight; i++) { + // padding head + for (int j = 0; j < paddingWidth; j++) { + *inputPadding++ = float(0); + } + + int step = inputWidth >> 2; + int remain = inputWidth & 3; + float32x4_t s1 = vdupq_n_f32(0.f); + for (int s = 0; s < step; s++) { + float32x4_t s0 = vld1q_f32(input); + float32x4x2_t v = {s0, s1}; + vst2q_f32(inputPadding, v); + input += 4; + inputPadding += 8; + } + for (int r = 0; r < remain; r++) { + *inputPadding++ = *input++; + *inputPadding++ = float(0); + } + inputPadding--; + + // padding tail + for (int j = 0; j < paddingWidth; j++) { + *inputPadding++ = float(0); + } + if (i != inputHeight - 1) { + memset(inputPadding, 0, padInputWidth * sizeof(float)); + inputPadding += padInputWidth; + } + } + + if (paddingHeight > 0) { + memset(inputPadding, 0, padInputWidth * paddingHeight * sizeof(float)); + inputPadding += padInputWidth * paddingHeight; + } + } + } +}; + +#endif + +#endif + +} // namespace neon +} // namespace paddle diff --git a/paddle/function/neon/NeonDepthwiseConvTranspose.cpp b/paddle/function/neon/NeonDepthwiseConvTranspose.cpp new file mode 100644 index 0000000000000000000000000000000000000000..49ca4bc8a0947ba329bd991e9f7d001623901a67 --- /dev/null +++ b/paddle/function/neon/NeonDepthwiseConvTranspose.cpp @@ -0,0 +1,136 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "NeonDepthwiseConv.h" +#include "paddle/function/ConvOp.h" + +namespace paddle { + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + +template +class NeonDepthwiseConvTransposeFunction : public ConvFunctionBase { +public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + check(inputs, outputs); + + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + + int batchSize = input[0]; + int inputChannels = input[1]; + int inputHeight = input[2]; + int inputWidth = input[3]; + int filterHeight = getFilterHeight(filter); + int filterWidth = getFilterWidth(filter); + int outputChannels = output[1]; + int outputHeight = output[2]; + int outputWidth = output[3]; + int filterMultiplier = outputChannels / groups_; + CHECK_EQ(inputChannels, groups_); + + // only support strideH() == strideW() and filterHeight == filterWidth. + CHECK_EQ(strideH(), strideW()); + CHECK_EQ(paddingH(), paddingW()); + CHECK_EQ(filterHeight, filterWidth); + + float* inputData = inputs[0].data(); + float* filterData = inputs[1].data(); + float* outputData = outputs[0].data(); + + // padding the input, input -> inputPadding + float* inputPadding = inputData; + int padInputHeight = + (inputHeight - 1) * strideH() + 2 * filterHeight - 1 - 2 * paddingH(); + int padInputWidth = + (inputWidth - 1) * strideW() + 2 * filterWidth - 1 - 2 * paddingW(); + + if (padInputHeight > inputHeight || padInputWidth > inputWidth) { + int newSize = batchSize * inputChannels * padInputHeight * padInputWidth; + resizeBuffer(newSize); + inputPadding = reinterpret_cast(memory_->getBuf()); + if (strideH() == 1) { + neon::Padding::run(inputData, + inputPadding, + batchSize * inputChannels, + inputHeight, + inputWidth, + padInputHeight, + padInputWidth); + } else if (strideH() == 2) { + neon::StridePadding::run(inputData, + inputPadding, + batchSize * inputChannels, + inputHeight, + inputWidth, + padInputHeight, + padInputWidth); + } else { + LOG(FATAL) << "Not supported"; + } + } + + std::function + DepthWiseConv; + + if (filterWidth == 3) { + DepthWiseConv = neon::DepthwiseConvKernel<3, 1>::run; + } else if (filterWidth == 4) { + DepthWiseConv = neon::DepthwiseConvKernel<4, 1>::run; + } else { + LOG(FATAL) << "Not supported"; + } + + for (int i = 0; i < batchSize; i++) { + DepthWiseConv(inputPadding, + filterData, + padInputHeight, + padInputWidth, + outputChannels, + outputHeight, + outputWidth, + filterMultiplier, + outputData); + inputPadding += inputChannels * padInputHeight * padInputWidth; + outputData += outputChannels * outputHeight * outputWidth; + } + } +}; + +#ifndef PADDLE_TYPE_DOUBLE + +REGISTER_TYPED_FUNC(NeonDepthwiseConvTranspose, + CPU, + NeonDepthwiseConvTransposeFunction); + +#endif + +#endif + +} // namespace paddle diff --git a/paddle/function/neon/neon_util.h b/paddle/function/neon/neon_util.h new file mode 100644 index 0000000000000000000000000000000000000000..e2db0450675084345ad55559d8988c5375801cc9 --- /dev/null +++ b/paddle/function/neon/neon_util.h @@ -0,0 +1,43 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + +#include + +namespace paddle { + +namespace neon { + +inline float32x4_t vld1q_f32_aligned(const float* p) { + return vld1q_f32( + (const float*)__builtin_assume_aligned(p, sizeof(float32x4_t))); +} + +#ifndef __aarch64__ +inline float32_t vaddvq_f32(float32x4_t a) { + float32x2_t v = vadd_f32(vget_high_f32(a), vget_low_f32(a)); + return vget_lane_f32(vpadd_f32(v, v), 0); +} + +#define vmlaq_laneq_f32(a, b, v, lane) \ + vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane)) +#endif + +} // namespace neon +} // namespace paddle + +#endif diff --git a/paddle/gserver/evaluators/CTCErrorEvaluator.cpp b/paddle/gserver/evaluators/CTCErrorEvaluator.cpp index 132119015f967c6e8d055792de8afe8450df5ec6..92087fa32b1e48b50fbf447ec6f3c43e2a510220 100644 --- a/paddle/gserver/evaluators/CTCErrorEvaluator.cpp +++ b/paddle/gserver/evaluators/CTCErrorEvaluator.cpp @@ -14,18 +14,20 @@ limitations under the License. */ #include "Evaluator.h" #include "paddle/gserver/gradientmachines/NeuralNetwork.h" +#include "paddle/utils/StringUtil.h" namespace paddle { /** * calculate sequence-to-sequence edit distance */ -class CTCErrorEvaluator : public NotGetableEvaluator { +class CTCErrorEvaluator : public Evaluator { private: MatrixPtr outActivations_; int numTimes_, numClasses_, numSequences_, blank_; real deletions_, insertions_, substitutions_; int seqClassficationError_; + mutable std::unordered_map evalResults_; std::vector path2String(const std::vector& path) { std::vector str; @@ -183,6 +185,18 @@ private: return stringAlignment(gtStr, recogStr); } + void storeLocalValues() const { + evalResults_["error"] = numSequences_ ? totalScore_ / numSequences_ : 0; + evalResults_["deletion_error"] = + numSequences_ ? deletions_ / numSequences_ : 0; + evalResults_["insertion_error"] = + numSequences_ ? insertions_ / numSequences_ : 0; + evalResults_["substitution_error"] = + numSequences_ ? substitutions_ / numSequences_ : 0; + evalResults_["sequence_error"] = + (real)seqClassficationError_ / numSequences_; + } + public: CTCErrorEvaluator() : numTimes_(0), @@ -245,16 +259,12 @@ public: } virtual void printStats(std::ostream& os) const { - os << config_.name() << "=" - << (numSequences_ ? totalScore_ / numSequences_ : 0); - os << " deletions error" - << "=" << (numSequences_ ? deletions_ / numSequences_ : 0); - os << " insertions error" - << "=" << (numSequences_ ? insertions_ / numSequences_ : 0); - os << " substitutions error" - << "=" << (numSequences_ ? substitutions_ / numSequences_ : 0); - os << " sequences error" - << "=" << (real)seqClassficationError_ / numSequences_; + storeLocalValues(); + os << config_.name() << " error = " << evalResults_["error"]; + os << " deletions error = " << evalResults_["deletion_error"]; + os << " insertions error = " << evalResults_["insertion_error"]; + os << " substitution error = " << evalResults_["substitution_error"]; + os << " sequence error = " << evalResults_["sequence_error"]; } virtual void distributeEval(ParameterClient2* client) { @@ -272,6 +282,37 @@ public: seqClassficationError_ = (int)buf[4]; numSequences_ = (int)buf[5]; } + + void getNames(std::vector* names) { + storeLocalValues(); + names->reserve(names->size() + evalResults_.size()); + for (auto it = evalResults_.begin(); it != evalResults_.end(); ++it) { + names->push_back(config_.name() + "." + it->first); + } + } + + real getValue(const std::string& name, Error* err) const { + storeLocalValues(); + + std::vector buffers; + paddle::str::split(name, '.', &buffers); + auto it = evalResults_.find(buffers[buffers.size() - 1]); + + if (it == evalResults_.end()) { + *err = Error("Evaluator does not have the key %s", name.c_str()); + return 0.0f; + } + + return it->second; + } + + std::string getType(const std::string& name, Error* err) const { + this->getValue(name, err); + if (!err->isOK()) { + return ""; + } + return "ctc_edit_distance"; + } }; REGISTER_EVALUATOR(ctc_edit_distance, CTCErrorEvaluator); diff --git a/paddle/gserver/evaluators/ChunkEvaluator.cpp b/paddle/gserver/evaluators/ChunkEvaluator.cpp index 1658282f3a5f79b128ce8685e92fd5cf9db2e41a..a2ab15eedee4aaa7b47504d50e25300359f18173 100644 --- a/paddle/gserver/evaluators/ChunkEvaluator.cpp +++ b/paddle/gserver/evaluators/ChunkEvaluator.cpp @@ -268,7 +268,13 @@ public: } // get type of evaluator - std::string getTypeImpl() const { return "chunk"; } + std::string getType(const std::string& name, Error* err) const { + this->getValue(name, err); + if (!err->isOK()) { + return ""; + } + return "chunk"; + } private: void storeLocalValues() const { diff --git a/paddle/gserver/evaluators/Evaluator.h b/paddle/gserver/evaluators/Evaluator.h index b114500e2b7c1e460a02c78b99b5f1a8fb63b8c3..90203553e0a5fe8cc8183274f374da178bae30d0 100644 --- a/paddle/gserver/evaluators/Evaluator.h +++ b/paddle/gserver/evaluators/Evaluator.h @@ -211,6 +211,7 @@ public: *err = Error("Not implemented"); return .0f; } + std::string getType(const std::string& name, Error* err) const { *err = Error("Not implemented"); return ""; @@ -331,6 +332,7 @@ private: protected: std::string getTypeImpl() const; }; + /** * @brief precision, recall and f1 score Evaluator * \f[ @@ -358,6 +360,12 @@ public: virtual void distributeEval(ParameterClient2* client); + void getNames(std::vector* names); + + real getValue(const std::string& name, Error* err) const; + + std::string getType(const std::string& name, Error* err) const; + struct StatsInfo { /// numbers of true positives double TP; @@ -428,11 +436,6 @@ private: mutable std::unordered_map values_; void storeLocalValues() const; - // Evaluator interface -public: - void getNames(std::vector* names); - real getValue(const std::string& name, Error* err) const; - std::string getType(const std::string& name, Error* err) const; }; /* diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.cpp b/paddle/gserver/gradientmachines/NeuralNetwork.cpp index cfa80a89365af5111746eec9599d16e37532a9f7..26cff3e67710b2f38d93572c5d58849aa94a5135 100644 --- a/paddle/gserver/gradientmachines/NeuralNetwork.cpp +++ b/paddle/gserver/gradientmachines/NeuralNetwork.cpp @@ -202,7 +202,7 @@ void NeuralNetwork::prefetch(const std::vector& inArgs) { auto mat = dynamic_cast( para->getMat(PARAMETER_VALUE).get()); para->clearGradient(); - mat->clearIndices(); + if (mat) mat->clearIndices(); } } } diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp index f98bf95064fa539b990309dfe0bff10c1e99d096..9f29b97466910f1daf88e3ca86f92d10661462c5 100644 --- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp +++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp @@ -184,7 +184,7 @@ public: } void backward(const UpdateCallback& callback) override { - if (biases_) { + if (biases_ && biases_->getWGrad()) { backwardActivation(); biases_->getWGrad()->collectBias(*getOutputGrad(), 1); biases_->getParameterPtr()->incUpdate(callback); @@ -1012,11 +1012,6 @@ void RecurrentGradientMachine::generateSequence() { /* width */ resultNum, false, /* useGpu */ false); - Matrix::resizeOrCreate(generator_.outArg.value, - /* height */ maxGenWordCount, - /* width */ 1, - false, - /* useGpu */ false); } ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions, numSequences + 1, @@ -1026,7 +1021,7 @@ void RecurrentGradientMachine::generateSequence() { } else { oneWaySearch(numSequences); } - if (dataArgsSize_) createDataOutlink(batchMachineIdVec_); + if (dataArgsSize_) createDataOutlink(); size_t size = generator_.ids.size(); generator_.outArg.ids->resize(size); @@ -1106,6 +1101,7 @@ void RecurrentGradientMachine::oneWaySearch(size_t batchSize) { } batchMachineIdVec_.clear(); + batchMachineStartPos_.clear(); int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false); starts[0] = 0; generator_.ids.clear(); @@ -1312,13 +1308,20 @@ void RecurrentGradientMachine::fillGenOutputs() { finalPaths_[i].resize(minFinalPathsSize); } - batchMachineIdVec_.clear(); generator_.ids.clear(); int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false); starts[0] = 0; if (numResults > 1) { - real* probs = generator_.outArg.in->getData(); + int idsProbSaveSize = 0; + for (auto inSeq : finalPaths_) { + for (auto path : inSeq) idsProbSaveSize += path.ids.size(); + idsProbSaveSize += inSeq.size(); + } + Matrix::resizeOrCreate( + generator_.outArg.value, idsProbSaveSize, 1, false, false); real* idsProb = generator_.outArg.value->getData(); + + real* probs = generator_.outArg.in->getData(); size_t curPos = 0; for (size_t i = 0; i < finalPaths_.size(); ++i) { for (size_t j = 0; j < finalPaths_[i].size(); ++j) { @@ -1333,24 +1336,16 @@ void RecurrentGradientMachine::fillGenOutputs() { curPos += genLen; idsProb[curPos++] = -1.0; probs[i * numResults + j] = path.logProb; - - if (!j && dataArgsSize_) { - // in beam search, here only reserved the top 1 generated result - // for out_links that are not the generated word indices. - batchMachineIdVec_.insert(batchMachineIdVec_.end(), - path.machineIdVec.begin(), - path.machineIdVec.end()); - } } starts[i + 1] = generator_.ids.size(); } } else { for (size_t i = 0; i < finalPaths_.size(); ++i) { CHECK(!finalPaths_[i].empty()); - generator_.ids.insert(generator_.ids.begin(), - finalPaths_[i][0].ids.begin(), - finalPaths_[i][0].ids.end()); - starts[i + 1] = starts[i] + finalPaths_[i][0].ids.size(); + Path& path = finalPaths_[i][0]; + generator_.ids.insert( + generator_.ids.end(), path.ids.begin(), path.ids.end()); + starts[i + 1] = starts[i] + path.ids.size(); } } } @@ -1364,25 +1359,76 @@ void RecurrentGradientMachine::copyDataOutlinkFrame(size_t machineCur) { } } -void RecurrentGradientMachine::createDataOutlink( - std::vector& machineIdVec) { - size_t seqNum = - getBeamSize() > 1UL ? finalPaths_.size() : finalPaths_[0].size(); - std::vector starts(seqNum + 1, 0); - for (size_t i = 0; i < seqNum; ++i) { - size_t seqLen = getBeamSize() > 1UL ? finalPaths_[i][0].ids.size() - : finalPaths_[0][i].ids.size(); - starts[i + 1] = starts[i] + seqLen; +void RecurrentGradientMachine::createDataOutlinkSelRowsInfo( + bool isSeq, std::vector& outArgs) { + batchMachineIdVec_.clear(); + + size_t seqIdx = 0; + for (size_t i = 0; i < finalPaths_.size(); ++i) { + for (size_t j = 0; j < finalPaths_[i].size(); ++j) { + std::vector& machineIdVec = finalPaths_[i][j].machineIdVec; + if (isSeq) { + for (size_t i = 0; i < machineIdVec.size(); ++i) { + size_t rowId = machineIdVec[i]; + int* seqPos = + outArgs[i].sequenceStartPositions->getMutableData(false); + batchMachineIdVec_.push_back(seqPos[rowId]); + } + } else { + batchMachineIdVec_.insert( + batchMachineIdVec_.end(), machineIdVec.begin(), machineIdVec.end()); + } + seqIdx++; + } + } +} + +void RecurrentGradientMachine::createDataOutlinkCopySizeInfo( + bool isSeq, std::vector& outArgs, std::vector& copySize) { + size_t totalSeqNum = std::accumulate( + finalPaths_.begin(), + finalPaths_.end(), + 0UL, + [](size_t a, const std::vector& b) { return a + b.size(); }); + copySize.resize(totalSeqNum, 1); + + batchMachineStartPos_.resize(totalSeqNum + 1, 0); + if (isSeq) { + ICpuGpuVectorPtr inputSeqStartPos = outArgs[0].sequenceStartPositions; + CHECK_EQ(static_cast(inputSeqStartPos->getSize() - 1), + getBeamSize() > 1 ? finalPaths_.size() : finalPaths_[0].size()); + int* starts = inputSeqStartPos->getMutableData(false); + int seqId = 0; + for (size_t i = 0; i < finalPaths_.size(); ++i) { + for (size_t j = 0; j < finalPaths_[i].size(); ++j) { + copySize[seqId] = getBeamSize() > 1 ? starts[i + 1] - starts[i] + : starts[j + 1] - starts[j]; + batchMachineStartPos_[seqId + 1] = + batchMachineStartPos_[seqId] + finalPaths_[i][j].ids.size(); + seqId++; + } + } + } else { + for (size_t i = 0; i < finalPaths_[0].size(); ++i) + batchMachineStartPos_[i + 1] = + batchMachineStartPos_[i] + finalPaths_[0][i].ids.size(); } +} +void RecurrentGradientMachine::createDataOutlink() { for (size_t i = 0; i < dataArgsSize_; i++) { + bool isSeq = dataArgsFrame_[i][0].hasSeq(); + std::vector copySize; + createDataOutlinkCopySizeInfo(isSeq, dataArgsFrame_[i], copySize); + createDataOutlinkSelRowsInfo(isSeq, dataArgsFrame_[i]); + dataArgs_[i].concat(dataArgsFrame_[i], - machineIdVec, - starts, + batchMachineIdVec_, + batchMachineStartPos_, + copySize, useGpu_, HPPL_STREAM_1, PASS_TEST); - auto dataAgent = dynamic_cast(outFrameLines_[i + 1].agentLayer.get()); CHECK_NOTNULL(dataAgent); diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h index fb3fc5877ac96323e891f800db80af83b6809831..c16fae6d1770e616fdcfabd440624c9be9753c91 100644 --- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h +++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h @@ -190,7 +190,7 @@ public: std::vector ids; /** - * @brief idsProb, log probability of each generated words. + * @brief idsProb, log probability of each generated word. */ std::vector idsProb; @@ -472,15 +472,43 @@ private: void copyDataOutlinkFrame(size_t machineCur); /* - * @brief In generation, if the layer group has more than 1 outlink, outlinks - * except the first one are data outlinks. This function creates the data - * outlinks. - * @note In beam search, only one generated sequence with the hightest log - * probabilites are retained. - * @param machineIdVec : select a row of output matrix in each frame - * that the generation process expanded. + * @brief In generation, if the layer group has more than 1 outlink, outlink + * except the first one is a data outlink. In RecurrentLayerGroup, each time + * step is a separate Network, outputs of a layer inside the + * RecurrentLayerGroup are stored in separate Arguments. If one layer is + * specified as an outlink of RecurrentLayerGroup. This function will + * collect outputs in each time step of each generated sequence which are + * dispersed in separate Arguments to form a new single Argument as output of + * RecurrentLayerGroup. */ - void createDataOutlink(std::vector& machineIdVec); + void createDataOutlink(); + + /* + * @brief decide to select how many rows from the Matrix stored the forward + * pass results from a start position. + * + * @param isSeq: a flag indicating whetehr the layer to be output of the + * RecurrentGradientMachine is a sequence or not + * @param outArgs: all of the the returned Arguments of the forward pass + * during the generation process. + * @param copySize: the returned result, number of rows to select from the + * Matrix stored the forward pass results from a start position. + */ + void createDataOutlinkCopySizeInfo(bool isSeq, + std::vector& outArgs, + std::vector& copySize); + + /* + * @brief decide index of the start row for each time step of a generated + * sequence in Matrix stored the entire beam search batch's forward pass + * results. + * + * @param isSeq: a flag indicating whether the layer to be output of the + * RecurrentGradientMachine is a sequence or not + * @param outArgs: all of the returned Arguments of the forward pass + * during the generation process. + */ + void createDataOutlinkSelRowsInfo(bool isSeq, std::vector& outArgs); /* * @brief used in beam search, connect previous frame to form recurrent link @@ -543,6 +571,7 @@ private: std::vector topIds_; std::vector seqIds_; std::vector batchMachineIdVec_; + std::vector batchMachineStartPos_; std::vector> finalPaths_; std::vector minFinalPathLogProb_; BeamSearchControlCallbacks* beamSearchCtrlCallbacks_; diff --git a/paddle/gserver/layers/BatchNormBaseLayer.cpp b/paddle/gserver/layers/BatchNormBaseLayer.cpp index 1ceaaaa206ee3cbc5421238574c7f310011ccaa5..f7a80e23e1bd49549bec57b360587adc6b423794 100644 --- a/paddle/gserver/layers/BatchNormBaseLayer.cpp +++ b/paddle/gserver/layers/BatchNormBaseLayer.cpp @@ -62,14 +62,18 @@ void BatchNormBaseLayer::calFeatureMapSize() { const ImageConfig& conf = config_.inputs(0).image_conf(); imageH_ = inputLayers_[0]->getOutput().getFrameHeight(); imageW_ = inputLayers_[0]->getOutput().getFrameWidth(); + imageD_ = inputLayers_[0]->getOutput().getFrameDepth(); + + if (0 == imageD_) imageD_ = conf.img_size_z(); if (imageH_ == 0 && imageW_ == 0) { imageH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); imageW_ = conf.img_size(); } else { getOutput().setFrameHeight(imageH_); getOutput().setFrameWidth(imageW_); + getOutput().setFrameDepth(imageD_); } - imgPixels_ = imageH_ * imageW_; + imgPixels_ = imageH_ * imageW_ * imageD_; } } // namespace paddle diff --git a/paddle/gserver/layers/BatchNormBaseLayer.h b/paddle/gserver/layers/BatchNormBaseLayer.h index 230bafc31d96bbd49481a7ed135be6888688627e..e721d2d267a31cae46407673b8b1281e87055608 100644 --- a/paddle/gserver/layers/BatchNormBaseLayer.h +++ b/paddle/gserver/layers/BatchNormBaseLayer.h @@ -80,6 +80,7 @@ protected: /// Height or width of input image feature. /// Both of them are 1 if the input is fully-connected layer. + int imageD_; int imageH_; int imageW_; /// Height * Width. diff --git a/paddle/gserver/layers/Conv3DLayer.cpp b/paddle/gserver/layers/Conv3DLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..9deda2de989a55d34510560c49b213ea1a52fd07 --- /dev/null +++ b/paddle/gserver/layers/Conv3DLayer.cpp @@ -0,0 +1,253 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Conv3DLayer.h" +#include "paddle/utils/Logging.h" +#include "paddle/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(conv3d, Conv3DLayer); + +bool Conv3DLayer::init(const LayerMap &layerMap, + const ParameterMap ¶meterMap) { + if (!ConvBaseLayer::init(layerMap, parameterMap)) return false; + int index = 0; + for (auto &inputConfig : config_.inputs()) { + const ConvConfig &conf = inputConfig.conv_conf(); + M_.push_back(numFilters_ / conf.groups()); + K_.push_back(filterPixels_[index] * filterChannels_[index]); + + // create a new weight + size_t height, width; + width = filterPixels_[index] * filterChannels_[index]; + height = numFilters_; + CHECK_EQ(parameters_[index]->getSize(), width * height); + Weight *w = new Weight(height, width, parameters_[index]); + weights_.emplace_back(w); + ++index; + } + if (biasParameter_.get()) { + if (sharedBiases_) { + CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); + biases_ = + std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); + } else { + biases_ = + std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); + } + } + return true; +} + +size_t Conv3DLayer::getSize() { + CHECK_NE(inputLayers_.size(), 0UL); + outputH_.clear(); + outputW_.clear(); + outputD_.clear(); + N_.clear(); + size_t layerSize = 0; + for (size_t i = 0; i < inputLayers_.size(); ++i) { + outputW_.push_back(outputSize( + imgSizeW_[i], filterSize_[i], padding_[i], stride_[i], true)); + outputH_.push_back(outputSize( + imgSizeH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true)); + outputD_.push_back(outputSize( + imgSizeD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true)); + + N_.push_back(outputD_[i] * outputH_[i] * outputW_[i]); + CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize); + layerSize += N_[i] * numFilters_; + } + getOutput().setFrameHeight(outputH_[0]); + getOutput().setFrameWidth(outputW_[0]); + getOutput().setFrameDepth(outputD_[0]); + return layerSize; +} + +void Conv3DLayer::forward(PassType passType) { + Layer::forward(passType); + + int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); + int outWidth = getSize(); + resetOutput(batchSize, outWidth); + + REGISTER_TIMER_INFO("FwdConv3D", getName().c_str()); + for (size_t i = 0; i != inputLayers_.size(); ++i) { + const MatrixPtr &inMat = getInputValue(i); + const MatrixPtr &outMat = getOutputValue(); + int M = M_[i]; + int N = N_[i]; + int K = K_[i]; + Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); + MatrixPtr wMat = weights_[i]->getW(); + for (int n = 0; n < batchSize; ++n) { + colBuf_->vol2Col(inMat->getData() + n * inMat->getStride(), + channels_[i], + imgSizeD_[i], + imgSizeH_[i], + imgSizeW_[i], + filterSizeZ_[i], + filterSizeY_[i], + filterSize_[i], + strideZ_[i], + strideY_[i], + stride_[i], + paddingZ_[i], + paddingY_[i], + padding_[i]); + + real *outData = outMat->getData() + n * outMat->getStride(); + MatrixPtr outMatSub = + Matrix::create(outData, groups_[i] * M, N, false, useGpu_); + for (int g = 0; g < groups_[i]; g++) { + MatrixPtr wMatSub = wMat->subMatrix(g * M, M); + MatrixPtr in = colBuf_->subMatrix(g * K, K); + MatrixPtr out = outMatSub->subMatrix(g * M, M); + out->mul(*wMatSub, *in, 1.0, 1.0); + } + } + } + if (nullptr != this->biasParameter_) { + this->addBias(); + } + forwardActivation(); +} + +void Conv3DLayer::backward(const UpdateCallback &callback) { + backwardActivation(); + + if (biases_ && biases_->getWGrad()) { + bpropBiases(); + biases_->getParameterPtr()->incUpdate(callback); + } + + REGISTER_TIMER_INFO("BwdConv3D", getName().c_str()); + for (size_t i = 0; i != inputLayers_.size(); ++i) { + if (weights_[i]->getWGrad()) { + bpropWeights(i); + } + if (getInputGrad(i)) { + bpropData(i); + } + weights_[i]->getParameterPtr()->incUpdate(callback); + } +} + +void Conv3DLayer::bpropWeights(int i) { + int M = M_[i]; + int N = N_[i]; + int K = K_[i]; + const MatrixPtr &inMat = getInputValue(i); + Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); + MatrixPtr wGradMat = weights_[i]->getWGrad(); + int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); + for (int n = 0; n < batchSize; ++n) { + colBuf_->vol2Col(inMat->getData() + n * inMat->getStride(), + channels_[i], + imgSizeD_[i], + imgSizeH_[i], + imgSizeW_[i], + filterSizeZ_[i], + filterSizeY_[i], + filterSize_[i], + strideZ_[i], + strideY_[i], + stride_[i], + paddingZ_[i], + paddingY_[i], + padding_[i]); + + real *outGradData = + getOutputGrad()->getData() + n * getOutputGrad()->getStride(); + MatrixPtr outGradSub = + Matrix::create(outGradData, groups_[i] * M, N, false, useGpu_); + for (int g = 0; g < groups_[i]; ++g) { + MatrixPtr inMatSub = colBuf_->subMatrix(g * K, K); + MatrixPtr outG = outGradSub->subMatrix(g * M, M); + MatrixPtr wGradSub = wGradMat->subMatrix(g * M, M); + wGradSub->mul(*outG, *(inMatSub->getTranspose()), 1.0, 1.0); + } + } +} + +void Conv3DLayer::bpropData(int i) { + int M = M_[i]; + int N = N_[i]; + int K = K_[i]; + Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); + MatrixPtr wMat = weights_[i]->getW(); + int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); + for (int n = 0; n < batchSize; ++n) { + real *outGradData = + getOutputGrad()->getData() + n * getOutputGrad()->getStride(); + real *preGradData = + getInputGrad(i)->getData() + n * getInputGrad(i)->getStride(); + MatrixPtr outGradSub = + Matrix::create(outGradData, M * groups_[i], N, false, useGpu_); + for (int g = 0; g < groups_[i]; ++g) { + MatrixPtr wMatSub = wMat->subMatrix(g * M, M); + MatrixPtr outG = outGradSub->subMatrix(g * M, M); + MatrixPtr inGradMatSub = colBuf_->subMatrix(g * K, K); + inGradMatSub->mul(*(wMatSub->getTranspose()), *outG, 1.0, 0.0); + } + colBuf_->col2Vol(preGradData, + channels_[i], + imgSizeD_[i], + imgSizeH_[i], + imgSizeW_[i], + filterSizeZ_[i], + filterSizeY_[i], + filterSize_[i], + strideZ_[i], + strideY_[i], + stride_[i], + paddingZ_[i], + paddingY_[i], + padding_[i], + 1.0, + 1.0); + } +} + +void Conv3DLayer::bpropBiases() { + MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(), + 1, + biases_->getWGrad()->getElementCnt(), + false, + useGpu_); + MatrixPtr outGradMat = getOutputGrad(); + + if (this->sharedBiases_) { + biases->collectSharedBias(*outGradMat, 1.0f); + } else { + biases->collectBias(*outGradMat, 1.0f); + } +} + +void Conv3DLayer::addBias() { + MatrixPtr outMat = getOutputValue(); + MatrixPtr bias = Matrix::create(biases_->getW()->getData(), + 1, + biases_->getW()->getElementCnt(), + false, + useGpu_); + if (this->sharedBiases_) { + outMat->addSharedBias(*(bias), 1.0f); + } else { + outMat->addBias(*(bias), 1.0f); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/Conv3DLayer.h b/paddle/gserver/layers/Conv3DLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..b622508d0ce1b0938c44f5c7f1371a34c86b2c1d --- /dev/null +++ b/paddle/gserver/layers/Conv3DLayer.h @@ -0,0 +1,51 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT 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 "ConvBaseLayer.h" +#include "paddle/math/MathUtils.h" +#include "paddle/math/Matrix.h" + +namespace paddle { + +/** + * @brief A subclass of convolution layer. + * This layer expands input and use matrix multiplication to + * calculate convolution operation. + */ +class Conv3DLayer : public ConvBaseLayer { +public: + explicit Conv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {} + ~Conv3DLayer() {} + + bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); + + void forward(PassType passType); + void addBias(); + void backward(const UpdateCallback& callback); + void bpropBiases(); + void bpropData(int i); + void bpropWeights(int i); + size_t getSize(); + +protected: + // Figure out the dimensions for individual gemms. + IntV M_; /// numFilters_ / filter_group_; + IntV N_; /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_ + IntV K_; /// outputD_ * outputH_ * outputW_ + MatrixPtr colBuf_; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/ConvBaseLayer.cpp b/paddle/gserver/layers/ConvBaseLayer.cpp index e161d89c38a290000a2cbdb2905e56901ae4c144..b848ab6bdd44f8fe81cbbf63b35a321599fd93fe 100644 --- a/paddle/gserver/layers/ConvBaseLayer.cpp +++ b/paddle/gserver/layers/ConvBaseLayer.cpp @@ -32,11 +32,12 @@ bool ConvBaseLayer::init(const LayerMap& layerMap, const ConvConfig& conf = inputConfig.conv_conf(); padding_.push_back(conf.padding()); stride_.push_back(conf.stride()); + dilation_.push_back(conf.dilation()); filterSize_.push_back(conf.filter_size()); paddingY_.push_back(conf.padding_y()); strideY_.push_back(conf.stride_y()); + dilationY_.push_back(conf.dilation_y()); filterSizeY_.push_back(conf.filter_size_y()); - filterPixels_.push_back(filterSize_.back() * filterSizeY_.back()); channels_.push_back(conf.channels()); imgSizeH_.push_back(conf.has_img_size_y() ? conf.img_size_y() : conf.img_size()); @@ -45,31 +46,20 @@ bool ConvBaseLayer::init(const LayerMap& layerMap, filterChannels_.push_back(conf.filter_channels()); outputH_.push_back(conf.has_output_y() ? conf.output_y() : conf.output_x()); outputW_.push_back(conf.output_x()); + + paddingZ_.push_back(conf.padding_z()); + strideZ_.push_back(conf.stride_z()); + filterSizeZ_.push_back(conf.filter_size_z()); + imgSizeD_.push_back(conf.img_size_z()); + outputD_.push_back(conf.output_z()); + filterPixels_.push_back(filterSize_.back() * filterSizeY_.back() * + filterSizeZ_.back()); } CHECK(inputLayers_.size() == parameters_.size()); - for (size_t i = 0; i < inputLayers_.size(); i++) { - size_t height, width; - height = filterPixels_[i] * filterChannels_[i]; - width = (!isDeconv_) ? numFilters_ : channels_[i]; - - // create a new weight - CHECK_EQ(parameters_[i]->getSize(), width * height); - Weight* w = new Weight(height, width, parameters_[i]); - weights_.emplace_back(w); - } - /* initialize the biases_ */ - if (biasParameter_.get()) { - if (sharedBiases_) { - CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); - biases_ = - std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); - } else { - biases_ = - std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); - } - } + // create new weights_ in derived class + // create new biases_ in derived class // default caffe model caffeMode_ = true; @@ -89,7 +79,11 @@ size_t ConvBaseLayer::calOutputSize() { size_t layerSize = 0; auto setLayerSize = [&](IntV& inH, IntV& inW, IntV& outH, IntV& outW) { + size_t filterSizeY; + size_t filterSize; for (size_t i = 0; i < inputLayers_.size(); i++) { + filterSizeY = (filterSizeY_[i] - 1) * dilationY_[i] + 1; + filterSize = (filterSize_[i] - 1) * dilation_[i] + 1; inH.push_back(inputLayers_[i]->getOutput().getFrameHeight()); inW.push_back(inputLayers_[i]->getOutput().getFrameWidth()); const ConvConfig& conf = config_.inputs(i).conv_conf(); @@ -98,17 +92,17 @@ size_t ConvBaseLayer::calOutputSize() { inH[i] = conf.has_output_y() ? conf.output_y() : conf.output_x(); if (inW[i] == 0) inW[i] = conf.output_x(); outH.push_back(imageSize( - inH[i], filterSizeY_[i], paddingY_[i], strideY_[i], caffeMode_)); - outW.push_back(imageSize( - inW[i], filterSize_[i], padding_[i], stride_[i], caffeMode_)); + inH[i], filterSizeY, paddingY_[i], strideY_[i], caffeMode_)); + outW.push_back( + imageSize(inW[i], filterSize, padding_[i], stride_[i], caffeMode_)); } else { if (inH[i] == 0) inH[i] = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); if (inW[i] == 0) inW[i] = conf.img_size(); outH.push_back(outputSize( - inH[i], filterSizeY_[i], paddingY_[i], strideY_[i], caffeMode_)); + inH[i], filterSizeY, paddingY_[i], strideY_[i], caffeMode_)); outW.push_back(outputSize( - inW[i], filterSize_[i], padding_[i], stride_[i], caffeMode_)); + inW[i], filterSize, padding_[i], stride_[i], caffeMode_)); } CHECK_EQ(outH[i], outH[0]); CHECK_EQ(outW[i], outW[0]); diff --git a/paddle/gserver/layers/ConvBaseLayer.h b/paddle/gserver/layers/ConvBaseLayer.h index e9d15d94f806a5d2e6f11cbbfc29e291dfe8538f..ccd170d9d85f573dff7340c26b2038c17a548471 100644 --- a/paddle/gserver/layers/ConvBaseLayer.h +++ b/paddle/gserver/layers/ConvBaseLayer.h @@ -40,6 +40,10 @@ protected: IntV stride_; /// The y dimension of the stride. IntV strideY_; + /// The x dimension of the dilation. + IntV dilation_; + /// The y dimension of the dilation. + IntV dilationY_; /// The x dimension of a filter kernel. IntV filterSize_; /// The y dimension of a filter kernel. @@ -58,6 +62,13 @@ protected: IntV outputH_; /// The spatial dimensions of output feature map width. IntV outputW_; + + IntV outputD_; + IntV imgSizeD_; + IntV filterSizeZ_; + IntV strideZ_; + IntV paddingZ_; + /// Group size, 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, diff --git a/paddle/gserver/layers/ConvBaseOperator.cpp b/paddle/gserver/layers/ConvBaseOperator.cpp index 5c231986292d2cd26ee30ccc122142fccd5b4949..5469c41c87468001232f7bae0d5b6bf26693b9e0 100644 --- a/paddle/gserver/layers/ConvBaseOperator.cpp +++ b/paddle/gserver/layers/ConvBaseOperator.cpp @@ -59,7 +59,8 @@ void ConvBaseOperator::allocConvWorkSpace() { &bwdDataAlgo_, &bwdDataLimitBytes_, &bwdFilterAlgo_, - &bwdFilterLimitBytes_); + &bwdFilterLimitBytes_, + /*useDilation*/ false); size_t maxWorkSpace = 0; maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_); diff --git a/paddle/gserver/layers/ConvBaseProjection.cpp b/paddle/gserver/layers/ConvBaseProjection.cpp index eb6b0445c95a9e9a7acd5d693ecdb11a263f41fd..08f36c516cfdadd42e9333c1c5a7a247df1f263e 100644 --- a/paddle/gserver/layers/ConvBaseProjection.cpp +++ b/paddle/gserver/layers/ConvBaseProjection.cpp @@ -41,6 +41,11 @@ void ConvBaseProjection::getConvParams() { strideH_ = conf.stride_y(); strideW_ = conf.stride(); + dilationH_ = conf.dilation_y(); + dilationW_ = conf.dilation(); + CHECK_GT(dilationH_, 0); + CHECK_GT(dilationW_, 0); + filterH_ = conf.filter_size_y(); filterW_ = conf.filter_size(); @@ -77,7 +82,9 @@ void ConvBaseProjection::initCudnn() { paddingH_, paddingW_, strideH_, - strideW_); + strideW_, + dilationH_, + dilationW_); // initialize all to default algorithms fwdAlgo_ = 0; @@ -131,7 +138,9 @@ void ConvBaseProjection::reshapeTensorDesc(int batchSize) { paddingH_, paddingW_, strideH_, - strideW_); + strideW_, + dilationH_, + dilationW_); } void ConvBaseProjection::reshape(int batchSize) { @@ -140,6 +149,10 @@ void ConvBaseProjection::reshape(int batchSize) { CHECK_EQ(calInputSize(), in_->value->getWidth()); reshapeTensorDesc(batchSize); + bool useDilation = false; + if (dilationH_ > 1 || dilationW_ > 1) { + useDilation = true; + } hl_conv_workspace(imageDesc_, outputDesc_, filterDesc_, @@ -149,7 +162,8 @@ void ConvBaseProjection::reshape(int batchSize) { &bwdDataAlgo_, &bwdDataLimitBytes_, &bwdFilterAlgo_, - &bwdFilterLimitBytes_); + &bwdFilterLimitBytes_, + useDilation); size_t maxWorkSpace = 0; maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_); diff --git a/paddle/gserver/layers/ConvBaseProjection.h b/paddle/gserver/layers/ConvBaseProjection.h index e9d9f8f1b2937b3a3b7323c43ef5608ffc5f82ca..ebdb57845bb36ac607b1e4c8e02f9d20b6e82a36 100644 --- a/paddle/gserver/layers/ConvBaseProjection.h +++ b/paddle/gserver/layers/ConvBaseProjection.h @@ -63,6 +63,7 @@ protected: int configChannels_, configNumFilters_; int paddingH_, paddingW_; int strideH_, strideW_; + int dilationH_, dilationW_; int filterH_, filterW_; /// One group offset of input data. int inputOffset_; diff --git a/paddle/gserver/layers/ConvProjection.cpp b/paddle/gserver/layers/ConvProjection.cpp index 5b7ecc5560c1e7431305b34a331fe1fbc96c6b06..6f0106b713d93494ba9baa5c7afa0a6b1f167262 100644 --- a/paddle/gserver/layers/ConvProjection.cpp +++ b/paddle/gserver/layers/ConvProjection.cpp @@ -25,12 +25,12 @@ size_t ConvProjection::calOutputSize() { if (imageH_ == 0) imageH_ = configImgH_; if (imageW_ == 0) imageW_ = configImgW_; outputH_ = outputSize(imageH_, - filterH_, + (filterH_ - 1) * dilationH_ + 1, paddingH_, strideH_, /* caffeMode */ true); outputW_ = outputSize(imageW_, - filterW_, + (filterW_ - 1) * dilationW_ + 1, paddingW_, strideW_, /* caffeMode */ true); diff --git a/paddle/gserver/layers/CostLayer.cpp b/paddle/gserver/layers/CostLayer.cpp index 6bfdea3c6e3f7cb80b620564f8229d954d773f04..ce071323ff585d28c9eaf80fec9be2394be526d1 100644 --- a/paddle/gserver/layers/CostLayer.cpp +++ b/paddle/gserver/layers/CostLayer.cpp @@ -572,13 +572,8 @@ void MultiBinaryLabelCrossEntropy::backwardImp(Matrix& output, } } -// -// Huber loss for robust 2-classes classification -// -REGISTER_LAYER(huber, HuberTwoClass); - -bool HuberTwoClass::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { +bool HuberCost::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { CostLayer::init(layerMap, parameterMap); if (useGpu_) { tmpCpuInput_.reserve(inputLayers_.size()); @@ -589,7 +584,7 @@ bool HuberTwoClass::init(const LayerMap& layerMap, return true; } -void HuberTwoClass::forwardImp(Matrix& output, Argument& label, Matrix& cost) { +void HuberCost::forwardImp(Matrix& output, Argument& label, Matrix& cost) { if (useGpu_) { for (size_t i = 0; i < inputLayers_.size(); i++) { tmpCpuInput_[i].resizeAndCopyFrom( @@ -597,13 +592,87 @@ void HuberTwoClass::forwardImp(Matrix& output, Argument& label, Matrix& cost) { } hl_stream_synchronize(HPPL_STREAM_DEFAULT); } - forwardImpIn(output, label, cost); } -void HuberTwoClass::forwardImpIn(Matrix& output, - Argument& label, - Matrix& target) { +// +// Huber loss for robust regression. +// +REGISTER_LAYER(huber_regression, HuberRegressionLoss); + +bool HuberRegressionLoss::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + HuberCost::init(layerMap, parameterMap); + delta_ = config_.delta(); + return true; +} + +void HuberRegressionLoss::forwardImp(Matrix& output, + Argument& label, + Matrix& target) { + HuberCost::forwardImp(output, label, target); + size_t numSamples = target.getHeight(); + size_t dim = output.getWidth(); + CHECK(label.value); + CHECK_EQ((*label.value).getHeight(), numSamples); + CHECK_EQ(output.getHeight(), numSamples); + CHECK_EQ(dim, (*label.value).getWidth()); + CHECK_EQ(target.getWidth(), (size_t)1); + + real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); + real* lbl = + useGpu_ ? tmpCpuInput_[1].value->getData() : (*label.value).getData(); + std::vector cost(numSamples, 0); + for (size_t i = 0; i < numSamples; ++i) { + for (size_t j = 0; j < dim; ++j) { + int index = i * dim + j; + real a = std::abs(lbl[index] - out[index]); + if (a <= delta_) + cost[i] += a * a / 2; + else + cost[i] += delta_ * (a - delta_ / 2); + } + } + target.copyFrom(cost.data(), numSamples); +} + +void HuberRegressionLoss::backwardImp(Matrix& output, + Argument& label, + Matrix& outputG) { + size_t numSamples = output.getHeight(); + size_t dim = output.getWidth(); + real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); + real* lbl = + useGpu_ ? tmpCpuInput_[1].value->getData() : (*label.value).getData(); + real* grad = useGpu_ ? tmpCpuInput_[0].grad->getData() : outputG.getData(); + for (size_t i = 0; i < numSamples; ++i) { + for (size_t j = 0; j < dim; ++j) { + int index = i * dim + j; + real a = lbl[index] - out[index]; + if (std::abs(a) <= delta_) + grad[index] += -a; + else + grad[index] += a > 0 ? -delta_ : delta_; + } + } + if (useGpu_) outputG.copyFrom(grad, numSamples * dim); +} + +// +// Huber loss for robust 2-classes classification +// +REGISTER_LAYER(huber_classification, HuberTwoClassification); + +bool HuberTwoClassification::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + return HuberCost::init(layerMap, parameterMap); +} + +void HuberTwoClassification::forwardImp(Matrix& output, + Argument& label, + Matrix& target) { + HuberCost::forwardImp(output, label, target); size_t numSamples = target.getHeight(); + CHECK(label.ids); CHECK_EQ((*label.ids).getSize(), numSamples); CHECK_EQ(output.getHeight(), numSamples); CHECK_EQ(output.getWidth(), (size_t)1); @@ -611,47 +680,35 @@ void HuberTwoClass::forwardImpIn(Matrix& output, real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData(); - std::vector cost(numSamples); + std::vector cost(numSamples, 0); for (size_t i = 0; i < numSamples; ++i) { int y = 2 * lbl[i] - 1; - if (out[i] * y < -1) - cost[i] = -4 * out[i] * y; - else if (out[i] * y < 1) - cost[i] = (1 - out[i] * y) * (1 - out[i] * y); - else - cost[i] = 0; + real a = out[i] * y; + if (a < -1) + cost[i] = -4 * a; + else if (a < 1) + cost[i] = (1 - a) * (1 - a); } target.copyFrom(cost.data(), numSamples); } -void HuberTwoClass::backwardImp(Matrix& outputValue, - Argument& label, - Matrix& outputGrad) { - if (useGpu_) { - backwardImpIn( - *tmpCpuInput_[0].value, tmpCpuInput_[1], *tmpCpuInput_[0].grad); - outputGrad.copyFrom(*tmpCpuInput_[0].grad); - } else { - backwardImpIn(outputValue, label, outputGrad); - } -} - -void HuberTwoClass::backwardImpIn(Matrix& output, - Argument& label, - Matrix& outputG) { +void HuberTwoClassification::backwardImp(Matrix& output, + Argument& label, + Matrix& outputG) { size_t numSamples = output.getHeight(); - real* out = output.getData(); - real* grad = outputG.getData(); - int* lbl = (*label.ids).getData(); + real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); + int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData(); + real* grad = useGpu_ ? tmpCpuInput_[0].grad->getData() : outputG.getData(); for (size_t i = 0; i < numSamples; ++i) { int y = 2 * lbl[i] - 1; - if (y * out[i] < -1) + real a = out[i] * y; + if (a < -1) grad[i] += -4 * y; - else if (y * out[i] < 1) - grad[i] += -2 * (1 - y * out[i]) * y; + else if (a < 1) + grad[i] += -2 * (1 - a) * y; } + if (useGpu_) outputG.copyFrom(grad, numSamples); } - /** * This cost layer compute the sum of its input as loss. * \f[ diff --git a/paddle/gserver/layers/CostLayer.h b/paddle/gserver/layers/CostLayer.h index 14c0b33ec1a628521ae2d694dda8da553c29fd38..0f655b48eea051c41ce17c0a41189b26188cc866 100644 --- a/paddle/gserver/layers/CostLayer.h +++ b/paddle/gserver/layers/CostLayer.h @@ -304,37 +304,70 @@ public: Matrix& outputGrad) override; }; -/** - * Huber loss for robust 2-classes classification. - * - * For label={0, 1}, let y=2*label-1. Given output f, the loss is: - * \f[ - * Loss = - * \left\{\begin{matrix} - * 4 * y * f & \textit{if} \ \ y* f < -1 \\ - * (1 - y * f)^2 & \textit{if} \ \ -1 < y * f < 1 \\ - * 0 & \textit{otherwise} - * \end{matrix}\right. - * \f] +/* + * A base layer for HuberRegressionLoss and HuberTwoClassification. */ -class HuberTwoClass : public CostLayer { +class HuberCost : public CostLayer { +public: std::vector tmpCpuInput_; -public: - explicit HuberTwoClass(const LayerConfig& config) : CostLayer(config) {} + explicit HuberCost(const LayerConfig& config) : CostLayer(config) {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) override; void forwardImp(Matrix& output, Argument& label, Matrix& cost) override; - void forwardImpIn(Matrix& output, Argument& label, Matrix& cost); + void backwardImp(Matrix& outputValue, + Argument& label, + Matrix& outputGrad) override {} +}; + +/** + * Huber loss for robust regression. + * + * Given output f(x), label y and delta, the loss is: + * Loss = 0.5 * (1 - y * f)^2, if abs(y - f) <= delta \\ + * Loss = delta * abs(y - f) - 0.5 * delta^2, otherwise + */ +class HuberRegressionLoss : public HuberCost { +public: + explicit HuberRegressionLoss(const LayerConfig& config) : HuberCost(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forwardImp(Matrix& output, Argument& label, Matrix& cost) override; void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad) override; - void backwardImpIn(Matrix& outputValue, Argument& label, Matrix& outputGrad); +protected: + real delta_; +}; + +/** + * Huber loss for robust 2-classes classification. + * + * For label={0, 1}, let y=2*label-1. Given output f(x), the loss is: + * Loss = 4 * y * f, if y* f < -1 \\ + * Loss = (1 - y * f)^2, if -1 < y * f < 1 \\ + * Loss = 0, otherwise + */ +class HuberTwoClassification : public HuberCost { +public: + explicit HuberTwoClassification(const LayerConfig& config) + : HuberCost(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forwardImp(Matrix& output, Argument& label, Matrix& cost) override; + + void backwardImp(Matrix& outputValue, + Argument& label, + Matrix& outputGrad) override; }; typedef std::shared_ptr CostLayerPtr; diff --git a/paddle/gserver/layers/CrossEntropyOverBeam.cpp b/paddle/gserver/layers/CrossEntropyOverBeam.cpp new file mode 100644 index 0000000000000000000000000000000000000000..578bdbbe72120abccc63ed13d11e1dec65d41e44 --- /dev/null +++ b/paddle/gserver/layers/CrossEntropyOverBeam.cpp @@ -0,0 +1,393 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "CrossEntropyOverBeam.h" + +namespace paddle { + +void CostForOneSequence::calValidExpandStep() { + validExpansionCount_ = 0; + goldAsExtraPath_ = true; + + for (size_t i = 0; i < beams_->expansionCount; ++i) { + real gold = static_cast(beams_->gold[i]); + if (i) { + real* start = beams_->candidateIds[i - 1]->getData(); + goldRowIds_[i] = std::count_if( + start, + start + goldRowIds_[i - 1] * beamSize_ + goldColIds_[i - 1], + [](const real& val) { return val != -1.; }); + } else { + goldRowIds_[i] = 0; + } + + real* start = + beams_->candidateIds[i]->getData() + goldRowIds_[i] * beamSize_; + real* findEnd = std::find(start, start + beamSize_, gold); + validExpansionCount_++; + + if (start + beamSize_ == findEnd) return; + goldColIds_[i] = findEnd - start; + } + if (goldColIds_[beams_->expansionCount - 1] != -1) goldAsExtraPath_ = false; +} + +size_t CostForOneSequence::initLastExpansion() { + int beamId = validExpansionCount_ - 1; + const MatrixPtr candidates = beams_->candidateIds[beamId]; + size_t height = candidates->getHeight(); + + /* initialization the last expansion. */ + size_t pathCount = std::count_if(candidates->getData(), + candidates->getData() + height * beamSize_, + [](const real& val) { return val != -1; }); + /* + * if the gold sequence falls off the beam during search, add the gold + * sequence as the last path into the all expanded candidates. + */ + if (goldAsExtraPath_) goldIdsInFinalExpansion_ = pathCount++; + + pathRowIdsInEachBeam_.clear(); + pathRowIdsInEachBeam_.resize(validExpansionCount_, + std::vector(pathCount, 0)); + parentIdsInBeam_.clear(); + parentIdsInBeam_.resize(pathCount, 0); + + if (goldAsExtraPath_) { + /* add gold sequence into the total expansion. */ + pathRowIdsInEachBeam_[beamId].back() = + beams_->gold[beamId] + + getSeqStartPos(beamId, goldRowIds_[validExpansionCount_ - 1]); + parentIdsInBeam_.back() = goldRowIds_[validExpansionCount_ - 1]; + } else { + size_t goldOffset = goldRowIds_[beamId] * beamSize_ + goldColIds_[beamId]; + goldIdsInFinalExpansion_ = + std::count_if(candidates->getData(), + candidates->getData() + goldOffset, + [](const real& val) { return val != -1.; }); + } + + /* + * TODO(caoying): fix this, store the indices of selected candidate + * paths into Argument.ids + */ + real* ids = candidates->getData(); + size_t curIdx = 0; + for (size_t i = 0; i < height; ++i) { + int basePos = getSeqStartPos(beamId, i); + for (size_t j = 0; j < beamSize_; ++j) { + int id = ids[i * beamSize_ + j]; + if (id == -1) continue; + pathRowIdsInEachBeam_[beamId][curIdx] = id + basePos; + parentIdsInBeam_[curIdx++] = i; + } + } + return pathCount; +} + +void CostForOneSequence::constructTotalExpansion() { + /* + * construct the entire expanded beam by begining with the last search + * in which gold falls off the beam. + */ + size_t totalPathCount = initLastExpansion(); + + for (int beamId = validExpansionCount_ - 2; beamId >= 0; --beamId) { + const MatrixPtr candidates = beams_->candidateIds[beamId]; + real* ids = candidates->getData(); + + int lastParentIdInBeam = -1; + int basePos = -1; + for (size_t i = 0; + i < (goldAsExtraPath_ ? totalPathCount - 1 : totalPathCount); + ++i) { + int id = ids[parentIdsInBeam_[i]]; + int parentRowId = std::div(parentIdsInBeam_[i], beamSize_).quot; + if (parentIdsInBeam_[i] != lastParentIdInBeam) + basePos = getSeqStartPos(beamId, parentRowId); + + pathRowIdsInEachBeam_[beamId][i] = id + basePos; + lastParentIdInBeam = parentIdsInBeam_[i]; + parentIdsInBeam_[i] = parentRowId; + + if (goldAsExtraPath_) + pathRowIdsInEachBeam_[beamId][totalPathCount - 1] = + beams_->gold[beamId] + getSeqStartPos(beamId, goldRowIds_[beamId]); + } + } +} + +real CostForOneSequence::globallyNormalizedScore() { + expandedPathScores_.resize(validExpansionCount_); + + Matrix::resizeOrCreate( + softmaxOut_, 1, pathRowIdsInEachBeam_[0].size(), false, false); + softmaxOut_->zeroMem(); + MatrixPtr tmp = Matrix::create( + softmaxOut_->getData(), softmaxOut_->getWidth(), 1, false, false); + + for (size_t i = 0; i < validExpansionCount_; ++i) { + Matrix::resizeOrCreate(expandedPathScores_[i], + pathRowIdsInEachBeam_[i].size(), + 1, + false, + false); + expandedPathScores_[i]->zeroMem(); + + IVectorPtr rowIds = IVector::create(pathRowIdsInEachBeam_[i].data(), + pathRowIdsInEachBeam_[i].size(), + false); + expandedPathScores_[i]->selectRows(*(beams_->scores[i]), *rowIds); + tmp->add(*expandedPathScores_[i]); + } + + softmaxOut_->softmax(*softmaxOut_); + return -std::log(softmaxOut_->getData()[goldIdsInFinalExpansion_]); +} + +real CostForOneSequence::forward() { + calValidExpandStep(); + constructTotalExpansion(); + return globallyNormalizedScore(); +} + +void CostForOneSequence::backward() { + /* + * when softmax layer is the output layer, and it is combined with + * cross-entropy as cost. The derivate with regard to softmax's input + * is simply: + * + * grad_i = softmax_out_i - target_i, + * + * and here hard label is used. + */ + softmaxOut_->getData()[goldIdsInFinalExpansion_] -= 1.; + + MatrixPtr tmp = Matrix::create( + softmaxOut_->getData(), softmaxOut_->getWidth(), 1, false, false); + + for (size_t i = 0; i < validExpansionCount_; ++i) { + IVectorPtr rowIds = IVector::create(pathRowIdsInEachBeam_[i].data(), + pathRowIdsInEachBeam_[i].size(), + false); + /* + beams_->scoreGrad[i] has been intialized outside this class, this + class only keeps a pointer pointing to the original input gradients, + so here does not need to allocate or initalize the memory. + */ + tmp->addToRows(*beams_->scoreGrad[i], *rowIds); + } +} + +REGISTER_LAYER(cross_entropy_over_beam, CrossEntropyOverBeam); + +bool CrossEntropyOverBeam::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + CHECK_EQ(0U, inputLayers_.size() % 3) << "Error input number."; + + beamExpanCount_ = inputLayers_.size() / 3; + + candidateScores_.resize(beamExpanCount_); + candidateScoreGrad_.resize(beamExpanCount_); + + candidateInBeam_.resize(beamExpanCount_); + goldSequence_.resize(beamExpanCount_); + gradToInputs_.resize(beamExpanCount_); + + setNeedSequenceInfo(false); + return true; +} + +void CrossEntropyOverBeam::checkInputs() { + batchSize_ = 0; + for (size_t i = 0; i < beamExpanCount_; ++i) { + const Argument& scores = getInput(i * 3); + const Argument& selCandidates = getInput(i * 3 + 1); + const Argument& goldSeq = getInput(i * 3 + 2); + + if (i) { + CHECK(scores.hasSubseq()) << "input " << i << " " + << inputLayers_[i * 3]->getName() + << " should be a nested sequence"; + CHECK_EQ(getInputValue(i * 3 + 1)->getWidth(), beamSize_); + CHECK_EQ(batchSize_, static_cast(scores.getNumSequences())); + CHECK_EQ(scores.getNumSubSequences(), selCandidates.getBatchSize()); + } else { + CHECK(scores.hasSeq()) << "input " << i << " " + << inputLayers_[i]->getName() + << " should be a sequence"; + batchSize_ = scores.getNumSequences(); + beamSize_ = getInputValue(i * 3 + 1)->getWidth(); + CHECK_EQ(batchSize_, static_cast(selCandidates.getBatchSize())); + } + CHECK_EQ(1U, scores.value->getWidth()); + CHECK_EQ(batchSize_, static_cast(goldSeq.getBatchSize())); + } +} + +void CrossEntropyOverBeam::copyInputsToCpu() { + auto copyValue = [](const MatrixPtr& src, MatrixPtr& trg) { + if (dynamic_cast(src.get())) { + Matrix::resizeOrCreate( + trg, src->getHeight(), src->getWidth(), false, false); + trg->copyFrom(*src); + } else { + trg = std::move(src); + } + }; + + auto copyIds = [](const IVectorPtr& src, IVectorPtr& trg) { + if (dynamic_cast(src.get())) { + IVector::resizeOrCreate(trg, src->getSize(), false); + trg->copyFrom(*src); + } else { + trg = std::move(src); + } + }; + + beamSplitPos_.clear(); + beamSplitPos_.resize(batchSize_, std::vector(beamExpanCount_, 0)); + for (size_t i = 0; i < beamExpanCount_; ++i) { + copyValue(getInputValue(i * 3), candidateScores_[i]); + copyValue(getInputValue(i * 3 + 1), candidateInBeam_[i]); + copyIds(getInput(i * 3 + 2).ids, goldSequence_[i]); + + if (i) { + ICpuGpuVectorPtr seqInfo = getInput(i * 3).sequenceStartPositions; + const int* seqStarts = seqInfo->getMutableData(false); + ICpuGpuVectorPtr subSeqInfo = getInput(i * 3).subSequenceStartPositions; + const int* subSeqStarts = subSeqInfo->getMutableData(false); + + size_t seqId = 1; + for (size_t subSeqId = 0; subSeqId < subSeqInfo->getSize() - 1; + ++subSeqId) { + CHECK_LT(seqId, seqInfo->getSize()); + if (subSeqStarts[subSeqId] == seqStarts[seqId]) { + beamSplitPos_[seqId][i] = beamSplitPos_[seqId - 1][i]; + seqId++; + } + beamSplitPos_[seqId - 1][i]++; + } + } else { + for (size_t j = 0; j < batchSize_; ++j) beamSplitPos_[j][i] = j + 1; + } + } +} + +void CrossEntropyOverBeam::splitBatchBeams() { + beamCosts_.resize(batchSize_); + beamPerSeq_.resize(batchSize_, BeamExpansion(beamExpanCount_)); + + for (size_t i = 0; i < beamExpanCount_; ++i) { + int* seqStarts = + getInput(i * 3).sequenceStartPositions->getMutableData(false); + + int* subSeqStarts = nullptr; + int maxLen = 0; + if (i) { + subSeqStarts = + getInput(i * 3).subSequenceStartPositions->getMutableData(false); + maxLen = getInput(i * 3).subSequenceStartPositions->getSize() - 1; + } else { + maxLen = getInput(i).sequenceStartPositions->getSize() - 1; + } + + for (size_t j = 0; j < batchSize_; ++j) { + beamPerSeq_[j].scores[i] = + Matrix::create(candidateScores_[i]->getData() + seqStarts[j], + seqStarts[j + 1] - seqStarts[j], + 1, + false, + false); + beamPerSeq_[j].scoreGrad[i] = + Matrix::create(candidateScoreGrad_[i]->getData() + seqStarts[j], + seqStarts[j + 1] - seqStarts[j], + 1, + false, + false); + + int offset = j ? beamSplitPos_[j - 1][i] : 0; + int height = beamSplitPos_[j][i] - (j ? beamSplitPos_[j - 1][i] : 0); + CHECK_GE(maxLen, offset + height); + beamPerSeq_[j].seqInfo[i] = IVector::create( + (i ? subSeqStarts : seqStarts) + offset, height + 1, false); + + beamPerSeq_[j].candidateIds[i] = + Matrix::create(candidateInBeam_[i]->getData() + offset * beamSize_, + height, + beamSize_, + false, + false); + beamPerSeq_[j].gold[i] = goldSequence_[i]->getData()[j]; + + CHECK_LE(beamPerSeq_[j].gold[i], seqStarts[j + 1] - seqStarts[j]); + } + } +} + +void CrossEntropyOverBeam::resizeOutput() { + Matrix::resizeOrCreate(output_.value, batchSize_, 1, false, false); + output_.value->zeroMem(); + + for (size_t i = 0; i < beamExpanCount_; ++i) { + MatrixPtr inGrad = getInputGrad(i * 3); + if (dynamic_cast(inGrad.get())) { + Matrix::resizeOrCreate(candidateScoreGrad_[i], + inGrad->getHeight(), + inGrad->getWidth(), + false, + false); + } else { + candidateScoreGrad_[i] = std::move(inGrad); + } + candidateScoreGrad_[i]->zeroMem(); + } +} + +void CrossEntropyOverBeam::copyGradToGpu(size_t copyCount) { + for (size_t i = 0; i < beamExpanCount_; ++i) { + if (dynamic_cast(getInputGrad(i * 3).get())) + getInputGrad(i * 3)->copyFrom(*candidateScoreGrad_[i]); + + if (i == copyCount - 1) break; + } +} + +void CrossEntropyOverBeam::forward(PassType passType) { + Layer::forward(passType); + + checkInputs(); + copyInputsToCpu(); + + resizeOutput(); + splitBatchBeams(); + + MatrixPtr outputValue = getOutputValue(); + for (size_t i = 0; i < batchSize_; ++i) { + BeamExpansionPtr ptr = std::make_shared(beamPerSeq_[i]); + beamCosts_[i].setData(std::move(ptr), beamSize_); + outputValue->getData()[i] = beamCosts_[i].forward(); + } +} + +void CrossEntropyOverBeam::backward(const UpdateCallback& callback) { + for (size_t i = 0; i < batchSize_; ++i) { + beamCosts_[i].backward(); + copyGradToGpu(beamCosts_[i].getValidExpansionCount()); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/CrossEntropyOverBeam.h b/paddle/gserver/layers/CrossEntropyOverBeam.h new file mode 100644 index 0000000000000000000000000000000000000000..5643556f43370912a730d9895658d8944f50dced --- /dev/null +++ b/paddle/gserver/layers/CrossEntropyOverBeam.h @@ -0,0 +1,135 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT 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 "CrossEntropyOverBeam.h" +#include "Layer.h" + +namespace paddle { + +/* This struct stores the beams in all search steps for a single sequence. */ +struct BeamExpansion { + std::vector scores; + std::vector seqInfo; + + std::vector candidateIds; + std::vector gold; + + std::vector scoreGrad; + + size_t expansionCount; + + explicit BeamExpansion(int n) { + expansionCount = n; + scores.resize(expansionCount); + seqInfo.resize(expansionCount); + candidateIds.resize(expansionCount); + scoreGrad.resize(expansionCount); + + gold.resize(expansionCount); + } +}; +typedef std::shared_ptr BeamExpansionPtr; + +class CostForOneSequence { +public: + CostForOneSequence() + : beamSize_(0), validExpansionCount_(0), goldAsExtraPath_(false) {} + void setData(const BeamExpansionPtr bPtr, size_t beamSize) { + beams_ = bPtr; + beamSize_ = beamSize; + + expandedPathScores_.clear(); + expandedPathScores_.resize(beams_->expansionCount); + + goldRowIds_.clear(); + goldRowIds_.resize(beams_->expansionCount, 0); + goldColIds_.clear(); + goldColIds_.resize(beams_->expansionCount, -1); + } + size_t getValidExpansionCount() { return validExpansionCount_; } + + real forward(); + void backward(); + +private: + void calValidExpandStep(); + void constructTotalExpansion(); + size_t initLastExpansion(); + real globallyNormalizedScore(); + + int getSeqStartPos(size_t beamId, size_t rowId) { + CHECK_GT(beams_->seqInfo[beamId]->getSize() - 1, rowId); + int* starts = beams_->seqInfo[beamId]->getData(); + return starts[rowId] - starts[0]; + } + + size_t beamSize_; + size_t validExpansionCount_; + bool goldAsExtraPath_; + std::vector goldRowIds_; + std::vector goldColIds_; + + BeamExpansionPtr beams_; + std::vector> pathRowIdsInEachBeam_; + std::vector parentIdsInBeam_; + size_t goldIdsInFinalExpansion_; + + std::vector expandedPathScores_; + + MatrixPtr softmaxOut_; +}; + +class CrossEntropyOverBeam : public Layer { +public: + explicit CrossEntropyOverBeam(const LayerConfig& config) : Layer(config) {} + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + void forward(PassType passType) override; + void backward(const UpdateCallback& callback) override; + +private: + void checkInputs(); + void copyInputsToCpu(); + void resizeOutput(); + void copyGradToGpu(size_t copyCount); + void splitBatchBeams(); + + size_t beamExpanCount_; + size_t batchSize_; + size_t beamSize_; + + /* + * the process of constructing beams is not friendly to GPU, currently, this + * layer only runs on CPU, if any of its inputs is on GPU memory, then copy + * it to CPU memory. + */ + std::vector candidateScores_; + std::vector candidateScoreGrad_; + std::vector candidateInBeam_; + std::vector gradToInputs_; + std::vector goldSequence_; + std::vector> beamSplitPos_; + + /* + * split entire bath of beams into beam per sequnence and store the result + * into this member. + */ + std::vector beamPerSeq_; + /* beamCosts_ is used to propagate error in one sequence. */ + std::vector beamCosts_; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/CudnnBatchNormLayer.cpp b/paddle/gserver/layers/CudnnBatchNormLayer.cpp index 44ba2c4b7d1562d2ce839b5f4b4de1af35e6925f..49a9540c0b6e36b59ed786287ff5c4569b69a6a5 100644 --- a/paddle/gserver/layers/CudnnBatchNormLayer.cpp +++ b/paddle/gserver/layers/CudnnBatchNormLayer.cpp @@ -37,7 +37,7 @@ bool CudnnBatchNormLayer::init(const LayerMap& layerMap, } void CudnnBatchNormLayer::reshape(int batchSize) { - hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_, imageW_); + hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_ * imageD_, imageW_); } void CudnnBatchNormLayer::forward(PassType passType) { @@ -104,7 +104,7 @@ void CudnnBatchNormLayer::forward(PassType passType) { EPS, batchSize, channels_, - imageH_, + imageH_ * imageD_, imageW_); } } diff --git a/paddle/gserver/layers/CudnnConvBaseLayer.cpp b/paddle/gserver/layers/CudnnConvBaseLayer.cpp index c056bbe4d1d354751d4f85f8d0743cf30486c087..9e954615cddf2566ea336d1c947985fd916e8cc4 100644 --- a/paddle/gserver/layers/CudnnConvBaseLayer.cpp +++ b/paddle/gserver/layers/CudnnConvBaseLayer.cpp @@ -46,8 +46,26 @@ bool CudnnConvBaseLayer::init(const LayerMap &layerMap, projConf_.emplace_back(conf); projections_.emplace_back( Projection::create(*projConf_[i], parameters_[i], useGpu_)); + + // create a new weight + size_t height, width; + height = filterPixels_[i] * filterChannels_[i]; + width = (!isDeconv_) ? numFilters_ : channels_[i]; + CHECK_EQ(parameters_[i]->getSize(), width * height); + Weight *w = new Weight(height, width, parameters_[i]); + weights_.emplace_back(w); } + if (biasParameter_.get()) { + if (sharedBiases_) { + CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); + biases_ = + std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); + } else { + biases_ = + std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); + } + } if (biases_.get() && sharedBiases_) { hl_create_tensor_descriptor(&biasDesc_); hl_create_tensor_descriptor(&outputDesc_); diff --git a/paddle/gserver/layers/DeConv3DLayer.cpp b/paddle/gserver/layers/DeConv3DLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..3eea638649e8ebfdd7efa18615977a9e1344c695 --- /dev/null +++ b/paddle/gserver/layers/DeConv3DLayer.cpp @@ -0,0 +1,220 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "DeConv3DLayer.h" +#include "paddle/utils/Logging.h" +#include "paddle/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(deconv3d, DeConv3DLayer); + +bool DeConv3DLayer::init(const LayerMap &layerMap, + const ParameterMap ¶meterMap) { + if (!ConvBaseLayer::init(layerMap, parameterMap)) return false; + // for Deconv, the dimension of Kernel is + // channel * output * depth * height * weigth + // Matrix storage format: (output * depth * height * weigth) x channel + for (int index = 0; index < config_.inputs().size(); ++index) { + M_.push_back(filterChannels_[index]); + K_.push_back(filterPixels_[index] * (numFilters_ / groups_[index])); + + // create a new weight + size_t height, width; + height = filterPixels_[index] * numFilters_; + width = filterChannels_[index]; + CHECK_EQ(parameters_[index]->getSize(), width * height); + Weight *w = new Weight(height, width, parameters_[index]); + weights_.emplace_back(w); + } + if (biasParameter_.get()) { + if (sharedBiases_) { + CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); + biases_ = + std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); + } else { + biases_ = + std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); + } + } + return true; +} + +size_t DeConv3DLayer::getSize() { + CHECK_NE(inputLayers_.size(), 0UL); + imgSizeW_.clear(); + imgSizeH_.clear(); + imgSizeD_.clear(); + N_.clear(); + NOut_.clear(); + size_t layerSize = 0; + for (size_t i = 0; i < inputLayers_.size(); ++i) { + imgSizeW_.push_back( + imageSize(outputW_[i], filterSize_[i], padding_[i], stride_[i], true)); + imgSizeH_.push_back(imageSize( + outputH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true)); + imgSizeD_.push_back(imageSize( + outputD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true)); + NOut_.push_back(imgSizeD_[i] * imgSizeH_[i] * imgSizeW_[i]); + N_.push_back(outputD_[i] * outputH_[i] * outputW_[i]); + CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize); + layerSize += NOut_[i] * numFilters_; + } + getOutput().setFrameHeight(imgSizeH_[0]); + getOutput().setFrameWidth(imgSizeW_[0]); + getOutput().setFrameDepth(imgSizeD_[0]); + return layerSize; +} + +void DeConv3DLayer::forward(PassType passType) { + Layer::forward(passType); + int batchSize = inputLayers_[0]->getOutputValue()->getHeight(); + int outWidth = getSize(); + resetOutput(batchSize, outWidth); + const MatrixPtr outMat = getOutputValue(); + + REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str()); + for (size_t i = 0; i != inputLayers_.size(); ++i) { + const MatrixPtr &inMat = getInputValue(i); + int M = M_[i]; + int N = N_[i]; + int K = K_[i]; + MatrixPtr wMat = weights_[i]->getW(); + Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); + for (int n = 0; n < batchSize; ++n) { + real *inData = inMat->getData() + n * inMat->getStride(); + for (int g = 0; g < groups_[i]; ++g) { + MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_); + MatrixPtr wMatSub = wMat->subMatrix(g * K, K); + MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K); + colBufDataSub->mul(*wMatSub, *inMatSub, 1.0, 0.0); + inData += M * N; + } + colBuf_->col2Vol(outMat->getData() + n * outMat->getStride(), + numFilters_, + imgSizeD_[i], + imgSizeH_[i], + imgSizeW_[i], + filterSizeZ_[i], + filterSizeY_[i], + filterSize_[i], + strideZ_[i], + strideY_[i], + stride_[i], + paddingZ_[i], + paddingY_[i], + padding_[i], + 1.0, + 1.0); + } + } + if (nullptr != this->biasParameter_) { + this->addBias(); + } + forwardActivation(); +} + +void DeConv3DLayer::backward(const UpdateCallback &callback) { + backwardActivation(); + int batchSize = getOutputGrad()->getHeight(); + if (biases_ && biases_->getWGrad()) { + bpropBiases(); + biases_->getParameterPtr()->incUpdate(callback); + } + REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str()); + for (size_t i = 0; i < inputLayers_.size(); ++i) { + if (weights_[i]->getWGrad() || this->needGradient_) { + int M = M_[i]; + int N = N_[i]; + int K = K_[i]; + Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); + const MatrixPtr &inMat = getInputValue(i); + for (int n = 0; n < batchSize; ++n) { + colBuf_->vol2Col( + getOutputGrad()->getData() + n * getOutputGrad()->getStride(), + numFilters_, + imgSizeD_[i], + imgSizeH_[i], + imgSizeW_[i], + filterSizeZ_[i], + filterSizeY_[i], + filterSize_[i], + strideZ_[i], + strideY_[i], + stride_[i], + paddingZ_[i], + paddingY_[i], + padding_[i]); + if (weights_[i]->getWGrad()) { + real *inData = inMat->getData() + n * inMat->getStride(); + for (int g = 0; g < groups_[i]; ++g) { + MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K); + MatrixPtr wGradMatSub = + weights_[i]->getWGrad()->subMatrix(g * K, K); + MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_); + wGradMatSub->mul( + *colBufDataSub, *(inMatSub->getTranspose()), 1.0, 1.0); + inData += M * N; + } + } + if (getInputGrad(i)) { + real *preGrad = + getInputGrad(i)->getData() + n * getInputGrad(i)->getStride(); + for (int g = 0; g < groups_[i]; ++g) { + MatrixPtr w = weights_[i]->getW()->subMatrix(g * K, K); + MatrixPtr outGradMat = colBuf_->subMatrix(g * K, K); + MatrixPtr inGradMatSub = + Matrix::create(preGrad, M, N, false, useGpu_); + inGradMatSub->mul(*(w->getTranspose()), *outGradMat, 1.0, 1.0); + preGrad += M * N; + } + } + } + weights_[i]->getParameterPtr()->incUpdate(callback); + } + } +} +void DeConv3DLayer::bpropWeights(int i) {} +void DeConv3DLayer::bpropData(int i) {} + +void DeConv3DLayer::bpropBiases() { + MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(), + 1, + biases_->getWGrad()->getElementCnt(), + false, + useGpu_); + const MatrixPtr &outGradMat = getOutputGrad(); + + if (this->sharedBiases_) { + biases->collectSharedBias(*outGradMat, 1.0f); + } else { + biases->collectBias(*outGradMat, 1.0f); + } +} + +void DeConv3DLayer::addBias() { + MatrixPtr outMat = getOutputValue(); + MatrixPtr bias = Matrix::create(biases_->getW()->getData(), + 1, + biases_->getW()->getElementCnt(), + false, + useGpu_); + if (this->sharedBiases_) { + outMat->addSharedBias(*(bias), 1.0f); + } else { + outMat->addBias(*(bias), 1.0f); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/DeConv3DLayer.h b/paddle/gserver/layers/DeConv3DLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..a2a3d3f8273ed77065224c27df6f711f09f34bbc --- /dev/null +++ b/paddle/gserver/layers/DeConv3DLayer.h @@ -0,0 +1,52 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT 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 "ConvBaseLayer.h" +#include "paddle/math/MathUtils.h" +#include "paddle/math/Matrix.h" + +namespace paddle { + +/** + * @brief A subclass of deconvolution3D layer. + * This layer expands input and use matrix multiplication to + * calculate deconvolution3D operation. + */ +class DeConv3DLayer : public ConvBaseLayer { +public: + explicit DeConv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {} + ~DeConv3DLayer() {} + bool init(const LayerMap& layerMap, const ParameterMap& parameterMap); + + void forward(PassType passType); + void addBias(); + void backward(const UpdateCallback& callback); + void bpropBiases(); + void bpropData(int i); + void bpropWeights(int i); + size_t getSize(); + +protected: + // Figure out the dimensions for individual gemms. + IntV M_; /// numFilters_ / filter_group_; + IntV N_; /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_ + IntV K_; /// outputD_ * outputH_ * outputW_ + IntV NOut_; + MatrixPtr colBuf_; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/DetectionOutputLayer.cpp b/paddle/gserver/layers/DetectionOutputLayer.cpp index 8ab838e191314ab25469631626c0b0564d7fffda..0cf0a92bf4bd8f9b8eba2016b2377d9dfb18c70a 100644 --- a/paddle/gserver/layers/DetectionOutputLayer.cpp +++ b/paddle/gserver/layers/DetectionOutputLayer.cpp @@ -139,7 +139,13 @@ void DetectionOutputLayer::forward(PassType passType) { allDecodedBBoxes, &allIndices); - resetOutput(numKept, 7); + if (numKept > 0) { + resetOutput(numKept, 7); + } else { + MatrixPtr outV = getOutputValue(); + outV = NULL; + return; + } MatrixPtr outV = getOutputValue(); getDetectionOutput(confBuffer_->getData(), numKept, diff --git a/paddle/gserver/layers/DetectionUtil.cpp b/paddle/gserver/layers/DetectionUtil.cpp index 3e61adc66e60c54250e4f323452aa13045310879..d83674f45a70212a8adc94a31ff58eb0e01baa00 100644 --- a/paddle/gserver/layers/DetectionUtil.cpp +++ b/paddle/gserver/layers/DetectionUtil.cpp @@ -469,7 +469,7 @@ size_t getDetectionIndices( const size_t numClasses, const size_t backgroundId, const size_t batchSize, - const size_t confThreshold, + const real confThreshold, const size_t nmsTopK, const real nmsThreshold, const size_t keepTopK, diff --git a/paddle/gserver/layers/DetectionUtil.h b/paddle/gserver/layers/DetectionUtil.h index fe4f9f075e4cf011c97f68f49598a828d62327b3..641ed873b4c8645b6455e5ef5e63593e3005b770 100644 --- a/paddle/gserver/layers/DetectionUtil.h +++ b/paddle/gserver/layers/DetectionUtil.h @@ -275,7 +275,7 @@ size_t getDetectionIndices( const size_t numClasses, const size_t backgroundId, const size_t batchSize, - const size_t confThreshold, + const real confThreshold, const size_t nmsTopK, const real nmsThreshold, const size_t keepTopK, diff --git a/paddle/gserver/layers/ExpandConvBaseLayer.cpp b/paddle/gserver/layers/ExpandConvBaseLayer.cpp index 77736e78f9349c0393e1e53ac700817a70893e53..2b7bef0a757d7c706be3815c539b036b094596cf 100644 --- a/paddle/gserver/layers/ExpandConvBaseLayer.cpp +++ b/paddle/gserver/layers/ExpandConvBaseLayer.cpp @@ -22,12 +22,31 @@ bool ExpandConvBaseLayer::init(const LayerMap &layerMap, /* Initialize the basic convolutional parent class */ ConvBaseLayer::init(layerMap, parameterMap); + int index = 0; for (auto &inputConfig : config_.inputs()) { const ConvConfig &conf = inputConfig.conv_conf(); /* Consistent caffe mode for multiple input */ caffeMode_ = conf.caffe_mode(); - } + // create a new weight + size_t height, width; + height = filterPixels_[index] * filterChannels_[index]; + width = (!isDeconv_) ? numFilters_ : channels_[index]; + CHECK_EQ(parameters_[index]->getSize(), width * height); + Weight *w = new Weight(height, width, parameters_[index]); + weights_.emplace_back(w); + index++; + } + if (biasParameter_.get()) { + if (sharedBiases_) { + CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); + biases_ = + std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); + } else { + biases_ = + std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); + } + } getOutputSize(); return true; diff --git a/paddle/gserver/layers/ExpandConvLayer.cpp b/paddle/gserver/layers/ExpandConvLayer.cpp index 0ece2799318ea5ecc91f97f71289d4d07246dcaa..20de475fc3f6b6f3c05ac26bea8363daff0cf110 100644 --- a/paddle/gserver/layers/ExpandConvLayer.cpp +++ b/paddle/gserver/layers/ExpandConvLayer.cpp @@ -29,6 +29,10 @@ namespace paddle { REGISTER_LAYER(exconv, ExpandConvLayer); REGISTER_LAYER(exconvt, ExpandConvLayer); +inline bool isDepthwiseConv(int channels, int groups) { + return channels == groups; +} + bool ExpandConvLayer::init(const LayerMap &layerMap, const ParameterMap ¶meterMap) { /* Initialize the basic convolutional parent class */ @@ -47,14 +51,27 @@ bool ExpandConvLayer::init(const LayerMap &layerMap, std::vector paddings = {(size_t)paddingY_[i], (size_t)padding_[i]}; std::vector strides = {(size_t)strideY_[i], (size_t)stride_[i]}; - if (useGpu_ && (size_t)groups_[i] == (size_t)channels_[i] && !isDeconv_) { + // Convolution Layer uses the GemmConv function by default. + convType = "GemmConv"; + convGradInputType = "GemmConvGradInput"; + convGradFilterType = "GemmConvGradFilter"; + + // If depth wise convolution and useGpu == true + if (useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) { convType = "DepthwiseConv"; convGradInputType = "DepthwiseConvGradInput"; convGradFilterType = "DepthwiseConvGradFilter"; - } else { - convType = "GemmConv"; - convGradInputType = "GemmConvGradInput"; - convGradFilterType = "GemmConvGradFilter"; + } + + // If depth wise convolution and useGpu == false and ARM-NEON + if (!useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) { +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + if ((filterSize_[i] == filterSizeY_[i]) && + (filterSize_[i] == 3 || filterSize_[i] == 4) && + (stride_[i] == strideY_[i]) && (stride_[i] == 1 || stride_[i] == 2)) { + convType = "NeonDepthwiseConv"; + } +#endif } if (FLAGS_use_nnpack && !isDeconv_) { diff --git a/paddle/gserver/layers/GruCompute.cpp b/paddle/gserver/layers/GruCompute.cpp index 06907768e98f4bad952706cffbbd65d1f86cc6df..148516391c6cad8feff34b9bd1c10c27d1a8a0e6 100644 --- a/paddle/gserver/layers/GruCompute.cpp +++ b/paddle/gserver/layers/GruCompute.cpp @@ -14,6 +14,7 @@ limitations under the License. */ #include "GruCompute.h" #include "hl_recurrent_apply.cuh" +#include "paddle/function/GruFunctor.h" #include "paddle/utils/Util.h" namespace paddle { @@ -25,13 +26,13 @@ void GruCompute::init(LayerConfig &config) { template <> void GruCompute::forward<0>(hl_gru_value value, int frameSize, int batchSize) { - hl_cpu_gru_forward(hppl::forward::gru_resetOutput(), - hppl::forward::gru_finalOutput(), - value, - frameSize, - batchSize, - activeNode_, - activeGate_); + GruFunctor::compute(hppl::forward::gru_resetOutput(), + hppl::forward::gru_finalOutput(), + value, + frameSize, + batchSize, + activeNode_, + activeGate_); } template <> @@ -39,14 +40,15 @@ void GruCompute::backward<0>(hl_gru_value value, hl_gru_grad grad, int frameSize, int batchSize) { - hl_cpu_gru_backward(hppl::backward::gru_stateGrad(), - hppl::backward::gru_resetGrad(), - value, - grad, - frameSize, - batchSize, - activeNode_, - activeGate_); + GruGradFunctor::compute( + hppl::backward::gru_stateGrad(), + hppl::backward::gru_resetGrad(), + value, + grad, + frameSize, + batchSize, + activeNode_, + activeGate_); } } // namespace paddle diff --git a/paddle/gserver/layers/KmaxSeqScoreLayer.cpp b/paddle/gserver/layers/KmaxSeqScoreLayer.cpp index 8ce591d4762466e1ed4b2970cb9cae9203bc0a2b..d5407555b248d79a5156a5ea354042d43ecda02c 100644 --- a/paddle/gserver/layers/KmaxSeqScoreLayer.cpp +++ b/paddle/gserver/layers/KmaxSeqScoreLayer.cpp @@ -80,13 +80,14 @@ void KmaxSeqScoreLayer::forward(PassType passType) { << "input of " << getName() << " must be a sequence or a nested sequence."; CHECK_EQ(input.value->getWidth(), 1UL) - << "input of " << getName() - << " is score over a sequence or a nested sequence, so its width " - << " must be 1."; + << "input of " << getName() << " are scores over a sequence or " + << "a nested sequence, so its width must be 1."; if (useGpu_) { - // this Layer runs only in CPU, if the model is runing on GPU, - // then copy the input to this layer from GPU to CPU. + /* + * currently, this Layer only runs in CPU, if the other part of the model is + * runing on GPU, then copy the input to this layer from GPU to CPU. + */ Matrix::resizeOrCreate(scores_, inputScore->getHeight(), 1, @@ -97,6 +98,14 @@ void KmaxSeqScoreLayer::forward(PassType passType) { scores_ = inputScore; } + /* + * TODO(caoying) + * In PaddePaddle, currently all matrices are real number types, + * but output of this layer which is some selected indices of the give + * sequence are actually filled with int types so that storing int types + * information in a real number matrix is dangerous, since real numbers will + * be convered to int types. + */ Matrix::resizeOrCreate( output_.value, input.hasSubseq() ? input.getNumSubSequences() : input.getNumSequences(), diff --git a/paddle/gserver/layers/Layer.cpp b/paddle/gserver/layers/Layer.cpp index d5621412caee843e24a0d0c9b7096402765738c7..2bc20eee6c452d0943dbf43b17ebe77976c97489 100644 --- a/paddle/gserver/layers/Layer.cpp +++ b/paddle/gserver/layers/Layer.cpp @@ -41,7 +41,7 @@ namespace paddle { Layer::Layer(const LayerConfig& config, bool useGpu) : config_(config), useGpu_(useGpu), - deviceId_(-1), + deviceId_(CPU_DEVICE), needSequenceInfo_(true) {} bool Layer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) { diff --git a/paddle/gserver/layers/Layer.h b/paddle/gserver/layers/Layer.h index 0ed482889d0cea884db3759620088575c5b10201..edef36194aabdb9c122ec3423deb036169a34d7c 100644 --- a/paddle/gserver/layers/Layer.h +++ b/paddle/gserver/layers/Layer.h @@ -59,7 +59,12 @@ protected: LayerConfig config_; /// whether to use GPU bool useGpu_; - /// Device Id. CPU is -1, and GPU is 0, 1, 2 ... + /// Paddle device ID, MKLDNN is -2, CPU is -1 + enum PADDLE_DEVICE_ID { + MKLDNN_DEVICE = -2, + CPU_DEVICE = -1, + }; + /// Device Id. MKLDNN is -2, CPU is -1, and GPU is 0, 1, 2 ... int deviceId_; /// Input layers std::vector inputLayers_; @@ -77,6 +82,7 @@ protected: Argument output_; /// Several outputs stored on different devices, used in 'parallel_nn' case, /// and record them by deviceId_. + /// Also used in 'use_mkldnn' case. std::vector outputOtherDevice_; /// If there are several outputs, map them by each name. std::map outputMap_; @@ -172,6 +178,13 @@ protected: return inputLayer.getOutput(deviceId_); } + /** + * Get the argument of input layer with deviceId. + */ + const Argument& getInput(size_t inputIndex, int deviceId) const { + return inputLayers_[inputIndex]->getOutput(deviceId); + } + /** * Get the forward-input value. */ @@ -186,6 +199,13 @@ protected: return inputLayer.getOutput(deviceId_).value; } + /** + * Get the forward-input value with deviceId. + */ + const MatrixPtr& getInputValue(int inputIndex, int deviceId) { + return inputLayers_[inputIndex]->getOutput(deviceId).value; + } + /** * Get the forward-input grad. */ @@ -200,6 +220,13 @@ protected: return inputLayer.getOutput(deviceId_).grad; } + /** + * Get the forward-input grad. + */ + const MatrixPtr& getInputGrad(int inputIndex, int deviceId) { + return inputLayers_[inputIndex]->getOutput(deviceId).grad; + } + /** * Get the forward-input label. */ diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp index 30f567eaf8248a8fba1b461a2bdbf2aab13f9e08..53433cef35a377a73f87b041fdcfadd848dd2ec9 100644 --- a/paddle/gserver/layers/MKLDNNFcLayer.cpp +++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp @@ -57,43 +57,28 @@ bool MKLDNNFcLayer::init(const LayerMap& layerMap, } void MKLDNNFcLayer::convertWeightsFromPaddle() { - if (FLAGS_use_mkldnn_wgt) { - return; - } - if (hasInitedWgt_) { return; } - // The weight_ is transposed from initial paddle weight - MatrixPtr paddleWgt = Matrix::create( - weight_->getW()->getData(), iLayerSize_, oc_, false, false); - - // TODO(TJ): remove this print when do not need differ weights - std::ostringstream ostr; - paddleWgt->print(ostr); - VLOG(MKLDNN_ALL) << "Initial Weight from paddle: " << std::endl << ostr.str(); - - // The mkldnn weight is transposed from initial paddle matrix - MatrixPtr paddleWgtT; - paddleWgt->transpose(paddleWgtT, true); - weight_->getW()->copyFrom(*paddleWgtT); + CHECK(wgtVal_) << "should have been initialized"; + bool hasNoSpatial_ = ih_ == 1 && iw_ == 1; + auto targetDim = wgtVal_->getDims(); + auto srcFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo; + wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim); hasInitedWgt_ = true; } void MKLDNNFcLayer::convertWeightsToPaddle() { - MatrixPtr dnnWgt = weight_->getW(); - MatrixPtr paddleWgt; - dnnWgt->transpose(paddleWgt, true); - - // copy paddle weight and override on weight_ - MatrixPtr dnnWgtT = Matrix::create( - dnnWgt->getData(), dnnWgt->getWidth(), dnnWgt->getHeight(), false, false); - dnnWgtT->copyFrom(*paddleWgt); + CHECK(wgtVal_) << "should have been initialized"; + bool hasNoSpatial_ = ih_ == 1 && iw_ == 1; + auto targetDim = wgtVal_->getDims(); + auto dstFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo; + wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim); } void MKLDNNFcLayer::reshape() { - const Argument& input = getInput(0); + const Argument& input = getInput(0, getPrev(0)->getDeviceId()); int batchSize = input.getBatchSize(); if (bs_ == batchSize) { return; @@ -107,10 +92,6 @@ void MKLDNNFcLayer::reshape() { if (iw_ == 0) { iw_ = 1; } - hasSpatial_ = true; - if (ih_ == 1 && iw_ == 1) { - hasSpatial_ = false; - } CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize()); ic_ = iLayerSize_ / (ih_ * iw_); CHECK_EQ(size_t(ic_ * ih_ * iw_), iLayerSize_) << "not divisible"; @@ -131,37 +112,56 @@ void MKLDNNFcLayer::reshape() { void MKLDNNFcLayer::resetFwd() { bool hasBias = biases_ && biases_->getW(); - real* iData = getInputValue(0)->getData(); - real* oData = getOutputValue()->getData(); - real* wData = weight_->getW()->getData(); - real* bData = hasBias ? biases_->getW()->getData() : NULL; - - // TODO(TJ): below create should be covered in MkldnnMatrix - // create memory desc - memory::desc iMD = hasSpatial_ ? createMD({bs_, ic_, ih_, iw_}, format::nchw) - : createMD({bs_, ic_}, format::nc); - memory::desc wMD = hasSpatial_ ? createMD({oc_, ic_, ih_, iw_}, format::oihw) - : createMD({oc_, ic_}, format::oi); - memory::desc bMD = bData != NULL ? createMD({oc_}, format::x) - : createMD({}, format::format_undef); - memory::desc oMD = createMD({bs_, oc_}, format::nc); - - // create memory primitive desc and memory self - inVal_.reset(new memory(memory::primitive_desc(iMD, engine_), iData)); - wgtVal_.reset(new memory(memory::primitive_desc(wMD, engine_), wData)); - outVal_.reset(new memory(memory::primitive_desc(oMD, engine_), oData)); + const MatrixPtr& wgt = weight_->getW(); + const MatrixPtr& bias = hasBias ? biases_->getW() : nullptr; + const MatrixPtr& out = output_.value; + + if (inputIsOnlyMKLDNN()) { + const MatrixPtr& in = getInputValue(0); + inVal_ = std::dynamic_pointer_cast(in); + CHECK(inVal_) << "Input should be MKLDNNMatrix"; + } else { + CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet"; + const MatrixPtr& in = getInputValue(0, CPU_DEVICE); + inVal_ = MKLDNNMatrix::create( + in, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_); + } + inVal_->downSpatial(); + wgtVal_ = MKLDNNMatrix::create( + wgt, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_); + wgtVal_->downSpatial(); + biasVal_ = + hasBias ? MKLDNNMatrix::create(bias, {oc_}, format::x, engine_) : nullptr; + outVal_ = MKLDNNMatrix::create(out, {bs_, oc_}, format::nc, engine_); + + // change original output value to mkldnn output value + output_.value = std::dynamic_pointer_cast(outVal_); + if (!outputIsOnlyMKLDNN()) { + copyOutputInfoToOtherDevice(); + // fc cpu output value do not need create convert + // just share point + getOutput(CPU_DEVICE).value->setData(output_.value->getData()); + } + // create forward handle prop_kind pk = prop_kind::forward; - fc_fwd::desc fwdDesc = bData != NULL ? fc_fwd::desc(pk, iMD, wMD, bMD, oMD) - : fc_fwd::desc(pk, iMD, wMD, oMD); + fc_fwd::desc fwdDesc = hasBias ? fc_fwd::desc(pk, + inVal_->getMemoryDesc(), + wgtVal_->getMemoryDesc(), + biasVal_->getMemoryDesc(), + outVal_->getMemoryDesc()) + : fc_fwd::desc(pk, + inVal_->getMemoryDesc(), + wgtVal_->getMemoryDesc(), + outVal_->getMemoryDesc()); fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_); - - if (bData != NULL) { - biasVal_.reset(new memory(memory::primitive_desc(bMD, engine_), bData)); + if (hasBias) { fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *biasVal_, *outVal_)); } else { fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *outVal_)); } + printValueFormatFlow(); + pipelineFwd_.clear(); pipelineFwd_.push_back(*fwd_); } @@ -171,45 +171,46 @@ void MKLDNNFcLayer::resetBwd() { return; } needResetBwd_ = false; - bool hasBias = biases_ && biases_->getWGrad(); - real* iData = getInputValue(0)->getData(); - real* iDiff = getInputGrad(0) != nullptr ? getInputGrad(0)->getData() : NULL; - real* oDiff = getOutputGrad()->getData(); - real* wDiff = weight_->getWGrad()->getData(); - real* bDiff = hasBias ? biases_->getWGrad()->getData() : NULL; /// backward weight - // create memory desc for backward memory - memory::desc iMD = hasSpatial_ ? createMD({bs_, ic_, ih_, iw_}, format::nchw) - : createMD({bs_, ic_}, format::nc); - memory::desc wMD = hasSpatial_ ? createMD({oc_, ic_, ih_, iw_}, format::oihw) - : createMD({oc_, ic_}, format::oi); - memory::desc oMD = createMD({bs_, oc_}, format::nc); - memory::desc bMD = bDiff != NULL ? createMD({oc_}, format::x) - : createMD({}, format::format_undef); - - if (inVal_) { - // update data - inVal_->set_data_handle(iData); - } else { - inVal_.reset(new memory(memory::primitive_desc(iMD, engine_), iData)); - } - - // create memory primitive desc and memory self - wgtGrad_.reset(new memory(memory::primitive_desc(wMD, engine_), wDiff)); - outGrad_.reset(new memory(memory::primitive_desc(oMD, engine_), oDiff)); - - fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward, iMD, wMD, oMD); + CHECK(inVal_) << "Should have input value"; + const MatrixPtr& wgt = weight_->getWGrad(); + const MatrixPtr& bias = hasBias ? biases_->getWGrad() : nullptr; + + // TODO(TJ): merge outgrad + int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE; + // for MKLDNN device: + // can not directly cast outputgrad to mkldnnmatrix, + // since each layer can not write the inputgrad to mkldnn inputgrad. + // So just create from matrix with outputvalue format. + // for CPU device: + // fc do not need to convert from cpu device since output is always nc format + // only need create from cpu device + const MatrixPtr& out = getOutput(device).grad; + outGrad_ = MKLDNNMatrix::create(out, outVal_->getPrimitiveDesc()); + wgtGrad_ = MKLDNNMatrix::create(wgt, wgtVal_->getPrimitiveDesc()); + biasGrad_ = hasBias ? MKLDNNMatrix::create(bias, biasVal_->getPrimitiveDesc()) + : nullptr; + + // create memory primitive desc + fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward, + inVal_->getMemoryDesc(), + wgtGrad_->getMemoryDesc(), + outGrad_->getMemoryDesc()); fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_); - fc_bwdWgt::desc bwdWgtDesc = bDiff != NULL - ? fc_bwdWgt::desc(iMD, wMD, bMD, oMD) - : fc_bwdWgt::desc(iMD, wMD, oMD); + fc_bwdWgt::desc bwdWgtDesc = hasBias + ? fc_bwdWgt::desc(inVal_->getMemoryDesc(), + wgtGrad_->getMemoryDesc(), + biasGrad_->getMemoryDesc(), + outGrad_->getMemoryDesc()) + : fc_bwdWgt::desc(inVal_->getMemoryDesc(), + wgtGrad_->getMemoryDesc(), + outGrad_->getMemoryDesc()); fc_bwdWgt::primitive_desc bwdWgtPD = fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD); - if (bDiff != NULL) { - biasGrad_.reset(new memory(memory::primitive_desc(bMD, engine_), bDiff)); + if (hasBias) { bwdWgt_.reset( new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_, *biasGrad_)); } else { @@ -219,29 +220,43 @@ void MKLDNNFcLayer::resetBwd() { pipelineBwd_.push_back(*bwdWgt_); /// backward data - if (iDiff == NULL) { + const MatrixPtr& in = inputLayers_[0]->getOutput().grad; + if (in == nullptr) { return; } - fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(iMD, wMD, oMD); + if (getInput(0, MKLDNN_DEVICE).getAllCount() > 1) { + // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done + } else { + inGrad_ = MKLDNNMatrix::create(in, inVal_->getPrimitiveDesc()); + } + + fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(inVal_->getMemoryDesc(), + wgtGrad_->getMemoryDesc(), + outGrad_->getMemoryDesc()); fc_bwdData::primitive_desc bwdDataPD = fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD); - inGrad_.reset(new memory(memory::primitive_desc(iMD, engine_), iDiff)); + CHECK(wgtVal_) << "Should have weight memory"; bwdData_.reset(new fc_bwdData(bwdDataPD, *outGrad_, *wgtVal_, *inGrad_)); + printGradFormatFlow(); pipelineBwd_.push_back(*bwdData_); } +void MKLDNNFcLayer::updateInputData() { + if (inputLayers_[0]->getType() != "data") { + return; + } + real* iData = getInputValue(0, CPU_DEVICE)->getData(); + inVal_->setData(iData); +} + void MKLDNNFcLayer::forward(PassType passType) { Layer::forward(passType); reshape(); { REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str()); - - // update input data - // since it might be changed if this is after data layer - real* iData = getInputValue(0)->getData(); - inVal_->set_data_handle(iData); + updateInputData(); // just submit forward pipeline stream_->submit(pipelineFwd_); @@ -263,10 +278,6 @@ void MKLDNNFcLayer::backward(const UpdateCallback& callback) { REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str()); resetBwd(); - // update diff - real* oDiff = getOutputGrad()->getData(); - outGrad_->set_data_handle(oDiff); - // just sumbmit backward pipeline stream_->submit(pipelineBwd_); } diff --git a/paddle/gserver/layers/MKLDNNFcLayer.h b/paddle/gserver/layers/MKLDNNFcLayer.h index 7954852a23f81d36d5fb0ae6a19768f419886fb1..4ad67a16e056a718c45a28babcf22a7cd571b15c 100644 --- a/paddle/gserver/layers/MKLDNNFcLayer.h +++ b/paddle/gserver/layers/MKLDNNFcLayer.h @@ -32,16 +32,13 @@ protected: // if has already init the weight bool hasInitedWgt_; - // if input layer has image size info (ih>1 && iw>1) - bool hasSpatial_; - // fc weight and bias std::unique_ptr weight_; std::unique_ptr biases_; public: explicit MKLDNNFcLayer(const LayerConfig& config) - : MKLDNNLayer(config), hasInitedWgt_(false), hasSpatial_(true) {} + : MKLDNNLayer(config), hasInitedWgt_(false) {} ~MKLDNNFcLayer() {} @@ -56,6 +53,8 @@ public: void backward(const UpdateCallback& callback) override; + void updateInputData() override; + protected: /** * reshape the input image sizes diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h index 63e29f447eede5ff9df8715bc9140b64ab7f7d17..543364edceff684bdcd002a8f4f10e7ce5e6953b 100644 --- a/paddle/gserver/layers/MKLDNNLayer.h +++ b/paddle/gserver/layers/MKLDNNLayer.h @@ -18,9 +18,9 @@ limitations under the License. */ #include "Layer.h" #include "MKLDNNBase.h" #include "mkldnn.hpp" +#include "paddle/math/MKLDNNMatrix.h" DECLARE_bool(use_mkldnn); -DECLARE_bool(use_mkldnn_wgt); namespace paddle { @@ -52,15 +52,15 @@ protected: std::vector pipelineFwd_; std::vector pipelineBwd_; - // TODO(TJ): change below memory as MKLDNNMatrixPtr type - std::shared_ptr inVal_; - std::shared_ptr inGrad_; - std::shared_ptr outVal_; - std::shared_ptr outGrad_; - std::shared_ptr wgtVal_; - std::shared_ptr wgtGrad_; - std::shared_ptr biasVal_; - std::shared_ptr biasGrad_; + // MKLDNNMatrixPtr + MKLDNNMatrixPtr inVal_; + MKLDNNMatrixPtr inGrad_; + MKLDNNMatrixPtr outVal_; + MKLDNNMatrixPtr outGrad_; + MKLDNNMatrixPtr wgtVal_; + MKLDNNMatrixPtr wgtGrad_; + MKLDNNMatrixPtr biasVal_; + MKLDNNMatrixPtr biasGrad_; public: explicit MKLDNNLayer(const LayerConfig& config) @@ -83,17 +83,21 @@ public: virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) { + CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn." + << "Please set WITH_MKLDNN=ON " + << "and set use_mkldnn=True"; + CHECK(!useGpu_) << "Do not support GPU yet"; + + // set device id before Layer::init + setDevice(MKLDNN_DEVICE); + // change param device to MKLDNN device + setParamsDevice(MKLDNN_DEVICE, parameterMap); if (!Layer::init(layerMap, parameterMap)) { return false; } - CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn." - << "Please set WITH_MKLDNN=ON " - << "and set use_mkldnn=True"; stream_.reset(new MKLDNNStream()); engine_ = CPUEngine::Instance().getEngine(); - - // TODO(TJ): deivecId return true; } @@ -109,6 +113,12 @@ public: */ virtual void convertWeightsToPaddle() {} + /** + * Update input value data when input layer is "data" type. + * Since the input value data address might be changed. + */ + virtual void updateInputData() {} + /** * print info about sizes */ @@ -118,14 +128,105 @@ public: << ", oh: " << oh_ << ", ow: " << ow_; } - // TODO(TJ): move to MkldnnMatrix - // create memory desc - inline mkldnn::memory::desc createMD( - mkldnn::memory::dims dims, - mkldnn::memory::format fmt, - mkldnn::memory::data_type type = mkldnn::memory::data_type::f32) { - // TODO(TJ): isFmtSuppoted(fmt) - return mkldnn::memory::desc(dims, type, fmt); + /** + * Print the mkldnn memory format flow of value + */ + virtual void printValueFormatFlow() { + if (inVal_ && outVal_) { + VLOG(MKLDNN_FMTS) << "value format flow --- " << inVal_->getFormat() + << " >>> " << outVal_->getFormat(); + } + } + + /** + * Print the mkldnn memory format flow of grad + */ + virtual void printGradFormatFlow() { + if (inGrad_ && outGrad_) { + VLOG(MKLDNN_FMTS) << "grad format flow --- " << inGrad_->getFormat() + << " <<< " << outGrad_->getFormat(); + } + } + +protected: + /** + * copy image size and sequence info to other device + * @note: can not directly use Layer::copyOutputToOtherDevice since here only + * copy base info and do not copy data value + */ + void copyOutputInfoToOtherDevice() { + int cnt = 0; + for (size_t i = 0; i < outputOtherDevice_.size(); i++) { + outputOtherDevice_[i].setFrameHeight(output_.getFrameHeight()); + outputOtherDevice_[i].setFrameWidth(output_.getFrameWidth()); + outputOtherDevice_[i].sequenceStartPositions = + output_.sequenceStartPositions; + outputOtherDevice_[i].subSequenceStartPositions = + output_.subSequenceStartPositions; + outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims; + if (outputOtherDevice_[i].deviceId == CPU_DEVICE) { + ++cnt; + } + } + if (cnt > 1) { + LOG(WARNING) << "should not have more than one CPU devie"; + } + } + + /** + * If input only has MKLDNN device. + * Otherwise, only support the previous layer using CPU device. + */ + bool inputIsOnlyMKLDNN(int index = 0) { + int prevDevice = getPrev(index)->getDeviceId(); + if (prevDevice == MKLDNN_DEVICE) { + return true; + } else { + // do not support GPU yet + CHECK_EQ(prevDevice, CPU_DEVICE) << "Only support CPU yet"; + return false; + } + } + + /** + * If output only has MKLDNN device. + * Otherwise, other devices should only using CPU device. + */ + bool outputIsOnlyMKLDNN() { + for (size_t i = 0; i < outputOtherDevice_.size(); i++) { + CHECK_EQ(outputOtherDevice_[i].deviceId, CPU_DEVICE) + << "Only support other device is CPU yet"; + } + return outputOtherDevice_.size() == 0; + } + + /** + * Set deviceId of this layer. + */ + void setDevice(int id) { deviceId_ = id; } + + /** + * Set deviceId of the params used in this layer. + */ + void setParamsDevice(int id, const ParameterMap& parameterMap) { + for (auto& inputConfig : config_.inputs()) { + if (inputConfig.has_input_parameter_name()) { + ParameterPtr parameter; + std::string name = inputConfig.input_parameter_name(); + CHECK(mapGet(name, parameterMap, ¶meter)) + << "Cannot find input parameter " << name << " for layer " + << getName(); + parameter->setDevice(id); + } + } + if (config_.has_bias_parameter_name()) { + ParameterPtr parameter; + std::string name = config_.bias_parameter_name(); + CHECK(mapGet(name, parameterMap, ¶meter)) + << "Cannot find bias parameter " << name << " for layer " + << getName(); + parameter->setDevice(id); + } } }; diff --git a/paddle/gserver/layers/Pool3DLayer.cpp b/paddle/gserver/layers/Pool3DLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..199f21adb1a5923b590e4f0e716fc67effb2a2d1 --- /dev/null +++ b/paddle/gserver/layers/Pool3DLayer.cpp @@ -0,0 +1,178 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Pool3DLayer.h" +#include "PoolProjectionLayer.h" +#include "paddle/utils/Logging.h" + +namespace paddle { + +REGISTER_LAYER(pool3d, Pool3DLayer); + +bool Pool3DLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + + /* the size of inputs for pool-layer is 1 */ + CHECK_EQ(config_.inputs_size(), 1); + + const PoolConfig& conf = config_.inputs(0).pool_conf(); + poolType_ = conf.pool_type(); + channels_ = conf.channels(); + + sizeX_ = conf.size_x(); + sizeY_ = conf.size_y(); + sizeZ_ = conf.size_z(); + + strideW_ = conf.stride(); + strideH_ = conf.stride_y(); + strideD_ = conf.stride_z(); + + imgSizeW_ = conf.img_size(); + imgSizeH_ = conf.img_size_y(); + imgSizeD_ = conf.img_size_z(); + + paddingW_ = conf.padding(); + paddingH_ = conf.padding_y(); + paddingD_ = conf.padding_z(); + + outputW_ = conf.output_x(); + outputH_ = conf.output_y(); + outputD_ = conf.output_z(); + + return true; +} + +size_t Pool3DLayer::getSize() { + CHECK_EQ(inputLayers_.size(), 1UL); + + size_t layerSize = 0; + outputD_ = outputSize(imgSizeD_, sizeZ_, paddingD_, strideD_, false); + outputH_ = outputSize(imgSizeH_, sizeY_, paddingH_, strideH_, false); + outputW_ = outputSize(imgSizeW_, sizeX_, paddingW_, strideW_, false); + + layerSize = outputD_ * outputH_ * outputW_ * channels_; + getOutput().setFrameHeight(outputH_); + getOutput().setFrameWidth(outputW_); + getOutput().setFrameDepth(outputD_); + return layerSize; +} + +void Pool3DLayer::forward(PassType passType) { + Layer::forward(passType); + const MatrixPtr& inMat = inputLayers_[0]->getOutputValue(); + size_t batchSize = inMat->getHeight(); + size_t outWidth = getSize(); + resetOutput(batchSize, outWidth); + Matrix::resizeOrCreate(maxPoolIdx_, batchSize, outWidth, false, useGpu_); + const MatrixPtr outMat = getOutputValue(); + + if (poolType_ == "avg") { + outMat->avgPool3DForward(*inMat, + channels_, + imgSizeD_, + imgSizeH_, + imgSizeW_, + outputD_, + outputH_, + outputW_, + sizeZ_, + sizeY_, + sizeX_, + strideD_, + strideH_, + strideW_, + paddingD_, + paddingH_, + paddingW_); + } else if (poolType_ == "max") { + outMat->maxPool3DForward(*inMat, + *maxPoolIdx_, + channels_, + imgSizeD_, + imgSizeH_, + imgSizeW_, + outputD_, + outputH_, + outputW_, + sizeZ_, + sizeY_, + sizeX_, + strideD_, + strideH_, + strideW_, + paddingD_, + paddingH_, + paddingW_); + } else { + LOG(FATAL) << "Unknown pool type: " << poolType_; + } + forwardActivation(); +} + +void Pool3DLayer::backward(const UpdateCallback& callback) { + backwardActivation(); + + (void)callback; + if (NULL == getInputGrad(0)) return; + MatrixPtr inMat = inputLayers_[0]->getOutputValue(); + MatrixPtr inGradMat = inputLayers_[0]->getOutputGrad(); + MatrixPtr outMat = getOutputValue(); + MatrixPtr outGradMat = getOutputGrad(); + + if (poolType_ == "avg") { + inGradMat->avgPool3DBackward(*outGradMat, + imgSizeD_, + imgSizeH_, + imgSizeW_, + outputD_, + outputH_, + outputW_, + sizeZ_, + sizeY_, + sizeZ_, + strideD_, + strideH_, + strideW_, + paddingD_, + paddingH_, + paddingW_, + 1.0, + 1.0); + } else if (poolType_ == "max") { + inGradMat->maxPool3DBackward(*outGradMat, + *maxPoolIdx_, + imgSizeD_, + imgSizeH_, + imgSizeW_, + outputD_, + outputH_, + outputW_, + sizeZ_, + sizeY_, + sizeZ_, + strideD_, + strideH_, + strideW_, + paddingD_, + paddingH_, + paddingW_, + 1.0, + 1.0); + } else { + LOG(FATAL) << "Unknown pool type: " << poolType_; + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/Pool3DLayer.h b/paddle/gserver/layers/Pool3DLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..8329a02f571bf3b5422134c756c248f77fd517b1 --- /dev/null +++ b/paddle/gserver/layers/Pool3DLayer.h @@ -0,0 +1,49 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "Layer.h" +#include "paddle/math/MathUtils.h" +#include "paddle/math/Matrix.h" + +namespace paddle { + +/** + * @brief Basic parent layer of pooling + * Pools the input within regions + */ +class Pool3DLayer : public Layer { +public: + explicit Pool3DLayer(const LayerConfig& config) : Layer(config) {} + ~Pool3DLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + void forward(PassType passType) override; + void backward(const UpdateCallback& callback) override; + size_t getSize(); + +protected: + int channels_; + int sizeX_, sizeY_, sizeZ_; + int strideW_, strideH_, strideD_; + int paddingW_, paddingH_, paddingD_; + int imgSizeW_, imgSizeH_, imgSizeD_; + int outputW_, outputH_, outputD_; + std::string poolType_; + MatrixPtr maxPoolIdx_; +}; +} // namespace paddle diff --git a/paddle/gserver/layers/PrintLayer.cpp b/paddle/gserver/layers/PrintLayer.cpp index 0a1e17b9aa57b373f0df6e079341729539f4e193..e83ae34bbe7d31b9bb7c16bc3fa84db7bd4e33d2 100644 --- a/paddle/gserver/layers/PrintLayer.cpp +++ b/paddle/gserver/layers/PrintLayer.cpp @@ -48,7 +48,16 @@ public: << inputLayers_.size() << ") at " << getName(); } s << format.substr(pos); - LOG(INFO) << s.str(); + + const std::string delimiter("\n"); + std::string content = s.str(); + std::string::size_type foundPos = 0; + std::string::size_type prevPos = 0; + while ((foundPos = content.find(delimiter, prevPos)) != std::string::npos) { + LOG(INFO) << content.substr(prevPos, foundPos - prevPos); + prevPos = foundPos + delimiter.size(); + } + LOG(INFO) << content.substr(prevPos); } void backward(const UpdateCallback& callback) override {} diff --git a/paddle/gserver/layers/ScaleShiftLayer.cpp b/paddle/gserver/layers/ScaleShiftLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..35fd038ab43a8a8b08bc328b3d1b08a7bbedd0a1 --- /dev/null +++ b/paddle/gserver/layers/ScaleShiftLayer.cpp @@ -0,0 +1,107 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" + +namespace paddle { + +/** + * A layer applies a linear transformation to each element in each row of + * the input matrix. For each element, the layer first re-scale it and then + * adds a bias to it. + * + * \f[ + * y = wx + b + * \f] + * + * Here, w is the scale and b is the bias. Both w and b are trainable scalars. + * + */ + +class ScaleShiftLayer : public Layer { +protected: + std::unique_ptr scale_; + std::unique_ptr offset_; + +public: + explicit ScaleShiftLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(scale_shift, ScaleShiftLayer); + +bool ScaleShiftLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + CHECK_EQ(inputLayers_.size(), 1U); + scale_.reset(new Weight(1, 1, parameters_[0])); + if (biasParameter_.get() != NULL) { + offset_ = std::unique_ptr(new Weight(1, 1, biasParameter_)); + } + return true; +} + +void ScaleShiftLayer::forward(PassType passType) { + Layer::forward(passType); + + MatrixPtr inV = getInputValue(0); + resetOutput(inV->getHeight(), inV->getWidth()); + MatrixPtr outV = getOutputValue(); + real scaleValue = scale_->getW()->getElement(0, 0); + outV->mulScalar(*inV, scaleValue); + if (offset_) { + real offsetValue = offset_->getW()->getElement(0, 0); + outV->add(offsetValue); + } +} + +void ScaleShiftLayer::backward(const UpdateCallback& callback) { + MatrixPtr inV = getInputValue(0); + MatrixPtr inG = getInputGrad(0); + MatrixPtr outV = getOutputValue(); + MatrixPtr outG = getOutputGrad(); + + /* Calculate the parameter gradient for the current layer */ + if (scale_->getWGrad()) { + MatrixPtr rowSumMtx; + Matrix::resizeOrCreate(rowSumMtx, outG->getHeight(), 1, false, useGpu_); + // this_i = scaleDest * this_i + scaleSum * \sum_j b_{ij} * c_{ij} + rowSumMtx->sumOfProducts( + /* b= */ *inV, /* c= */ *outG, /* scaleSum= */ 1, /* scaleDest= */ 0.); + // this_i = scaleDest * this_i + scaleSum * \sum_j b_{ji} + scale_->getWGrad()->sumCols( + /* b= */ *rowSumMtx, /* scaleSum= */ 1., /* scaleDest= */ 1.); + scale_->getParameterPtr()->incUpdate(callback); + } + if (offset_ && offset_->getWGrad()) { + MatrixPtr rowSumMtx; + Matrix::resizeOrCreate(rowSumMtx, outG->getHeight(), 1, false, useGpu_); + rowSumMtx->sumRows(*outG, 1., 0.); + offset_->getWGrad()->sumCols(*rowSumMtx, 1., 1.); + offset_->getParameterPtr()->incUpdate(callback); + } + + /* Calculate the input layers error */ + if (inG) { + real scaleValue = scale_->getW()->getElement(0, 0); + inG->add(*outG, scaleValue); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/SequenceSliceLayer.cpp b/paddle/gserver/layers/SequenceSliceLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..d3a83fad276a384ab3fddd5349912c56be6f3cc0 --- /dev/null +++ b/paddle/gserver/layers/SequenceSliceLayer.cpp @@ -0,0 +1,223 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/math/Matrix.h" +#include "paddle/math/Vector.h" +#include "paddle/utils/Logging.h" +#include "paddle/utils/Stat.h" + +namespace paddle { + +class SequenceSliceLayer : public Layer { +public: + explicit SequenceSliceLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + +private: + /* + * TODO(caoying) + * In PaddePaddle, currently all matrices are real number types, + * but the second and the (optional) third input which are some + * selected indices of the give sequence to trim the sequence, are actually + * filled with int types so that storing int types information in real number + * matrices is very dangerous, since real numbers will be convered to int + * types. If a user fills this matrix himself, invalid data may occor. + */ + + MatrixPtr startIdsOnCpu_; + MatrixPtr endIdsOnCpu_; + + std::vector selectedRows_; + IVectorPtr rowIndice_; + std::vector> inputSeqInfoVec_; + std::vector outSubSeqStartPos_; + std::vector outSeqStartPos_; + + void checkInputs(); + void copySliceIdsToCpu(); + void calSelectedRows(const MatrixPtr starts, const MatrixPtr ends); +}; + +REGISTER_LAYER(seq_slice, SequenceSliceLayer); + +bool SequenceSliceLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + CHECK_GE(inputLayers_.size(), 2U); + CHECK_LE(inputLayers_.size(), 3U); + + setNeedSequenceInfo(false); + return true; +} + +void SequenceSliceLayer::checkInputs() { + const Argument& inputSeq = getInput(0); + 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()) + << "Height of the second input should be equal to number of sequence " + << "in the first input."; + if (inputLayers_.size() == 3) { + const MatrixPtr indices2 = getInputValue(2); + CHECK_EQ(indices2->getHeight(), indices1->getHeight()) + << "start indices and end indices should have the same height."; + CHECK_EQ(indices2->getWidth(), indices1->getWidth()) + << "start indices and end indices should have the same Width."; + } +} + +void SequenceSliceLayer::copySliceIdsToCpu() { + const MatrixPtr indices1 = getInputValue(1); + if (inputLayers_.size() == 2U) { + if (config_.select_first()) { + Matrix::resizeOrCreate(startIdsOnCpu_, + indices1->getHeight(), + indices1->getWidth(), + false /* trans */, + false /* useGpu */); + startIdsOnCpu_->copyFrom(*indices1); + endIdsOnCpu_ = nullptr; + } else { + Matrix::resizeOrCreate(endIdsOnCpu_, + indices1->getHeight(), + indices1->getWidth(), + false /* trans */, + false /* useGpu */); + endIdsOnCpu_->copyFrom(*indices1); + startIdsOnCpu_ = nullptr; + } + } else if (inputLayers_.size() == 3U) { + Matrix::resizeOrCreate(startIdsOnCpu_, + indices1->getHeight(), + indices1->getWidth(), + false /* trans */, + false /* useGpu */); + startIdsOnCpu_->copyFrom(*indices1); + + const MatrixPtr indices2 = getInputValue(2); + Matrix::resizeOrCreate(endIdsOnCpu_, + indices2->getHeight(), + indices2->getWidth(), + false /* trans */, + false /* useGpu */); + endIdsOnCpu_->copyFrom(*indices2); + } +} + +void SequenceSliceLayer::calSelectedRows(const MatrixPtr starts, + const MatrixPtr ends) { + CHECK(starts || ends) << "At least one of the start or end indices " + << "should be given."; + + bool hasSubseq = getInput(0).hasSubseq(); + + outSeqStartPos_.resize(1, 0); + outSubSeqStartPos_.resize(1, 0); + selectedRows_.clear(); + + size_t beamSize = starts ? starts->getWidth() : ends->getWidth(); + size_t rowIdx = 0; + for (size_t i = 0; i < inputSeqInfoVec_.size(); ++i) { + for (size_t j = 0; j < inputSeqInfoVec_[i].size() - 1; ++j) { + for (size_t k = 0; k < beamSize; ++k) { + if (starts && starts->getElement(rowIdx, k) == -1.) break; + if (ends && ends->getElement(rowIdx, k) == -1.) break; + + int begPos = inputSeqInfoVec_[i][j]; + if (starts) begPos += starts->getElement(rowIdx, k); + + int endPos = inputSeqInfoVec_[i][j + 1] - 1; + if (ends) endPos = inputSeqInfoVec_[i][j] + ends->getElement(rowIdx, k); + + int seqLen = endPos - begPos + 1; + CHECK_GT(seqLen, 0U); + for (int m = begPos; m <= endPos; ++m) selectedRows_.push_back(m); + hasSubseq + ? outSubSeqStartPos_.push_back(outSubSeqStartPos_.back() + seqLen) + : outSeqStartPos_.push_back(outSeqStartPos_.back() + seqLen); + } + rowIdx++; + } + if (hasSubseq) outSeqStartPos_.push_back(outSubSeqStartPos_.back()); + } + + if (useGpu_) { + rowIndice_ = IVector::create(selectedRows_.size(), useGpu_); + rowIndice_->copyFrom(selectedRows_.data(), selectedRows_.size()); + } else { + rowIndice_ = + IVector::create(selectedRows_.data(), selectedRows_.size(), useGpu_); + } + + // create the sequence information for the output. + ICpuGpuVector::resizeOrCreate( + output_.sequenceStartPositions, outSeqStartPos_.size(), false); + output_.sequenceStartPositions->copyFrom( + outSeqStartPos_.data(), outSeqStartPos_.size(), false); + + if (hasSubseq) { + ICpuGpuVector::resizeOrCreate( + output_.subSequenceStartPositions, outSubSeqStartPos_.size(), false); + output_.subSequenceStartPositions->copyFrom( + outSubSeqStartPos_.data(), outSubSeqStartPos_.size(), false); + } +} + +void SequenceSliceLayer::forward(PassType passType) { + Layer::forward(passType); + checkInputs(); + + const Argument& inputSeq = getInput(0); + inputSeqInfoVec_.clear(); + Argument::reorganizeSeqInfo(inputSeq.sequenceStartPositions, + inputSeq.subSequenceStartPositions, + inputSeqInfoVec_); + if (!useGpu_) { + if (inputLayers_.size() == 2U) { + startIdsOnCpu_ = config_.select_first() ? getInputValue(1) : nullptr; + endIdsOnCpu_ = config_.select_first() ? nullptr : getInputValue(1); + } else if (inputLayers_.size() == 3U) { + startIdsOnCpu_ = getInputValue(1); + endIdsOnCpu_ = getInputValue(2); + } + } else { + copySliceIdsToCpu(); + } + + /* + * calculate the selected row indices in a batch, and build the output + * sequence information. + */ + calSelectedRows(startIdsOnCpu_, endIdsOnCpu_); + + resetOutput(selectedRows_.size(), getSize()); + + getOutputValue()->selectRows(*getInputValue(0), *rowIndice_); +} + +void SequenceSliceLayer::backward(const UpdateCallback& callback) { + getOutputGrad()->addToRows(*getInputGrad(0), *rowIndice_); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/SubNestedSequenceLayer.cpp b/paddle/gserver/layers/SubNestedSequenceLayer.cpp index 648d3908f391450f276d8a900ebb3bccb8d5532c..e9bee77212065effdac78cba590caed2e9155f0a 100644 --- a/paddle/gserver/layers/SubNestedSequenceLayer.cpp +++ b/paddle/gserver/layers/SubNestedSequenceLayer.cpp @@ -52,23 +52,34 @@ private: * ] * * ths output is saved to private member rowIndice_; - * [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15, - * 16,17,18,19,20,21,22,23,24,25,26,27] + * [0,1,2,3,4,5,6,7,8,9,15,16,17,18,19,20,21,23,24,25,26,27] */ - void calSelectedCols(const MatrixPtr selectedIndices, + void calSelectedRows(const MatrixPtr selectedIndices, const std::vector>& inputSeqInfo); - // if the second input of this layer is on GPU memory, copy it to CPU memory. + /* + * TODO(caoying) + * In PaddePaddle, currently all matrices are real number types, + * but the second is some selected indices of the give sequence to trim + * the nested sequence, are actually filled with int types so that storing + * int types information in real number matrices is very dangerous, since + * real numbers will be convered to int types. If a user fills this matrix + * himself, invalid data may occor. + * + * if the second input of this layer is on GPU memory, copy it to CPU memory. + */ MatrixPtr selIdsCpu_; - // reorganized sequenceStartPositions and subSequenceStartPositions - // into a 2d vector to facilitate the sequence selection process. + /* + * reorganize sequenceStartPositions and subSequenceStartPositions + * into a 2d vector to facilitate the sequence selection process. + */ std::vector> inputSeqInfoVec_; - // the final selected row indices in a batch, - // rowIdx_ and selectedRows_ actually share a same memory. + /* store the final selected row indices in a batch */ IVectorPtr rowIndice_; + /* rowIndice_ and selectedRows_ actually share a same memory. */ std::vector selectedRows_; }; @@ -83,7 +94,7 @@ bool SubNestedSequenceLayer::init(const LayerMap& layerMap, return true; } -void SubNestedSequenceLayer::calSelectedCols( +void SubNestedSequenceLayer::calSelectedRows( const MatrixPtr selectedIndices, const std::vector>& inputSeqInfo) { selectedRows_.clear(); @@ -160,7 +171,7 @@ void SubNestedSequenceLayer::forward(PassType passType) { Argument::reorganizeSeqInfo(inputSeq.sequenceStartPositions, inputSeq.subSequenceStartPositions, inputSeqInfoVec_); - calSelectedCols(selIdsCpu_, inputSeqInfoVec_); + calSelectedRows(selIdsCpu_, inputSeqInfoVec_); resetOutput(selectedRows_.size(), getSize()); getOutputValue()->selectRows(*getInputValue(0), *rowIndice_); diff --git a/paddle/gserver/layers/SwitchOrderLayer.cpp b/paddle/gserver/layers/SwitchOrderLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..d7eee6eaf078dab8d48adc4c7ee758a433672ac6 --- /dev/null +++ b/paddle/gserver/layers/SwitchOrderLayer.cpp @@ -0,0 +1,110 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "SwitchOrderLayer.h" +#include "paddle/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(switch_order, SwitchOrderLayer); + +bool SwitchOrderLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + auto& img_conf = config_.inputs(0).image_conf(); + size_t inD = img_conf.img_size_z(); + size_t inH = + img_conf.has_img_size_y() ? img_conf.img_size_y() : img_conf.img_size(); + size_t inW = img_conf.img_size(); + size_t inC = img_conf.channels(); + inH = inH * inD; + inDims_ = TensorShape({0, inC, inH, inW}); + outDims_ = TensorShape(4); + + auto& reshape_conf = config_.reshape_conf(); + for (int i = 0; i < reshape_conf.height_axis_size(); i++) { + heightAxis_.push_back(reshape_conf.height_axis(i)); + } + for (int i = 0; i < reshape_conf.width_axis_size(); i++) { + widthAxis_.push_back(reshape_conf.width_axis(i)); + } + createFunction(nchw2nhwc_, "NCHW2NHWC", FuncConfig()); + createFunction(nhwc2nchw_, "NHWC2NCHW", FuncConfig()); + return true; +} + +void SwitchOrderLayer::setOutDims() { + outDims_.setDim(0, inDims_[0]); + outDims_.setDim(1, inDims_[2]); + outDims_.setDim(2, inDims_[3]); + outDims_.setDim(3, inDims_[1]); + reshapeHeight_ = 1; + for (size_t i = 0; i < heightAxis_.size(); i++) { + reshapeHeight_ *= outDims_[heightAxis_[i]]; + } + output_.setFrameHeight(reshapeHeight_); + reshapeWidth_ = 1; + for (size_t i = 0; i < widthAxis_.size(); i++) { + reshapeWidth_ *= outDims_[widthAxis_[i]]; + } + output_.setFrameWidth(reshapeWidth_); +} + +void SwitchOrderLayer::setInDims() { + MatrixPtr input = inputLayers_[0]->getOutputValue(); + size_t batchSize = input->getHeight(); + inDims_.setDim(0, batchSize); + int d = inputLayers_[0]->getOutput().getFrameDepth(); + d = (d == 0 ? 1 : d); + int h = inputLayers_[0]->getOutput().getFrameHeight(); + if (h != 0) inDims_.setDim(2, h * d); + int w = inputLayers_[0]->getOutput().getFrameWidth(); + if (w != 0) inDims_.setDim(3, w); + int totalCount = input->getElementCnt(); + int channels = totalCount / (inDims_[0] * inDims_[2] * inDims_[3]); + if (channels != 0) inDims_.setDim(1, channels); +} + +void SwitchOrderLayer::forward(PassType passType) { + Layer::forward(passType); + setInDims(); + setOutDims(); + resetOutput(outDims_[0], outDims_[1] * outDims_[2] * outDims_[3]); + if (heightAxis_.size() > 0) { + getOutputValue()->reshape(reshapeHeight_, reshapeWidth_); + getOutputGrad()->reshape(reshapeHeight_, reshapeWidth_); + } + + // switch NCHW to NHWC + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getInputValue(0), inDims_); + outputs.addArg(*getOutputValue(), outDims_); + nchw2nhwc_[0]->calc(inputs, outputs); + forwardActivation(); +} + +void SwitchOrderLayer::backward(const UpdateCallback& callback) { + (void)callback; + backwardActivation(); + + // switch NHWC to NCHW + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getOutputGrad(), outDims_); + outputs.addArg(*getInputGrad(0), inDims_, ADD_TO); + nhwc2nchw_[0]->calc(inputs, outputs); +} +} // namespace paddle diff --git a/paddle/gserver/layers/SwitchOrderLayer.h b/paddle/gserver/layers/SwitchOrderLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..47b1f7f73ee783b3eae3c9cfe08b1459cef16a71 --- /dev/null +++ b/paddle/gserver/layers/SwitchOrderLayer.h @@ -0,0 +1,47 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" + +namespace paddle { + +/** + * \brief This layer calculate softmax in image channel dimension. + */ +class SwitchOrderLayer : public Layer { +public: + explicit SwitchOrderLayer(const LayerConfig& config) : Layer(config) {} + + ~SwitchOrderLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + void setInDims(); + void setOutDims(); + +protected: + std::vector> nchw2nhwc_; + std::vector> nhwc2nchw_; + TensorShape inDims_; + TensorShape outDims_; + std::vector heightAxis_; + std::vector widthAxis_; + size_t reshapeHeight_; + size_t reshapeWidth_; +}; +} // namespace paddle diff --git a/paddle/gserver/tests/CMakeLists.txt b/paddle/gserver/tests/CMakeLists.txt index c2a2993620492a9ec5dae932ff1292ced2c00064..de9b8e63dfc4291f8f42ca8c57cb5eb6baed8d8e 100644 --- a/paddle/gserver/tests/CMakeLists.txt +++ b/paddle/gserver/tests/CMakeLists.txt @@ -34,6 +34,19 @@ add_unittest_without_exec(test_CRFLayerGrad add_test(NAME test_CRFLayerGrad COMMAND test_CRFLayerGrad) +################ test_CrossEntropyOverBeam #################### +add_unittest_without_exec(test_CrossEntropyOverBeam + test_CrossEntropyOverBeamGrad.cpp + LayerGradUtil.cpp) +add_test(NAME test_CrossEntropyOverBeam + COMMAND test_CrossEntropyOverBeam) + +################ test_SeqSliceLayerGrad #################### +add_unittest_without_exec(test_SeqSliceLayerGrad + test_SeqSliceLayerGrad.cpp + LayerGradUtil.cpp) +add_test(NAME test_SeqSliceLayerGrad + COMMAND test_SeqSliceLayerGrad) add_unittest_without_exec(test_ActivationGrad test_ActivationGrad.cpp diff --git a/paddle/gserver/tests/LayerGradUtil.cpp b/paddle/gserver/tests/LayerGradUtil.cpp index fd9cfa1dc7a9028cb2c5c98baca98ffb2a837bac..a38880e14cdfcef05461dae567d198e5400c6bb1 100644 --- a/paddle/gserver/tests/LayerGradUtil.cpp +++ b/paddle/gserver/tests/LayerGradUtil.cpp @@ -388,14 +388,23 @@ void initDataLayer(TestConfig testConf, data.grad->zeroMem(); break; case INPUT_SELF_DEFINE_DATA: { - size_t height = testConf.inputDefs[i].selfDefinedData->getHeight(); - size_t width = testConf.inputDefs[i].selfDefinedData->getWidth(); - CHECK_GT(static_cast(height), 0); - CHECK_GT(static_cast(width), 0); - data.value = Matrix::create(height, width, false, useGpu); - data.grad = Matrix::create(height, width, false, useGpu); - data.value->copyFrom(*testConf.inputDefs[i].selfDefinedData); - data.grad->zeroMem(); + if (testConf.inputDefs[i].ids.size()) { + data.ids = IVector::create(testConf.inputDefs[i].ids.size(), useGpu); + data.ids->copyFrom(testConf.inputDefs[i].ids.data(), + testConf.inputDefs[i].ids.size()); + } else if (testConf.inputDefs[i].selfDefinedData) { + size_t height = testConf.inputDefs[i].selfDefinedData->getHeight(); + size_t width = testConf.inputDefs[i].selfDefinedData->getWidth(); + CHECK_GT(static_cast(height), 0); + CHECK_GT(static_cast(width), 0); + data.value = Matrix::create(height, width, false, useGpu); + data.grad = Matrix::create(height, width, false, useGpu); + data.value->copyFrom(*testConf.inputDefs[i].selfDefinedData); + data.grad->zeroMem(); + } else { + LOG(FATAL) << "No self-defined data are given."; + return; + } const std::vector& labelSeqStartPositions = testConf.inputDefs[i].labelSeqStartPositions; diff --git a/paddle/gserver/tests/LayerGradUtil.h b/paddle/gserver/tests/LayerGradUtil.h index 5debedf5ef6a3262578ca01b335e664f9a334d35..88e831f78bd165f63806df6c081d84411be51502 100644 --- a/paddle/gserver/tests/LayerGradUtil.h +++ b/paddle/gserver/tests/LayerGradUtil.h @@ -68,6 +68,7 @@ struct InputDef { std::vector labelInitValue; std::vector labelSeqStartPositions; std::vector labelSubSeqStartPositions; + std::vector ids; MatrixPtr selfDefinedData; InputDef(InputType type, string nameIn, size_t dimIn, size_t sizeIn) { @@ -95,6 +96,23 @@ struct InputDef { isStatic = false; } + InputDef(InputType type, + string nameIn, + const std::vector& ids, + const std::vector& selfDefinedSeqStartPos = {}, + const std::vector& selfDefinedSubSeqStartPos = {}) + : labelSeqStartPositions(selfDefinedSeqStartPos), + labelSubSeqStartPositions(selfDefinedSubSeqStartPos), + ids(ids) { + selfDefinedData = nullptr; + inputType = type; + name = nameIn; + dim = 0; + sparse = {""}; + paraSize = 0; + isStatic = false; + } + InputDef(InputType type, string nameIn, size_t dimIn, diff --git a/paddle/gserver/tests/MKLDNNTester.cpp b/paddle/gserver/tests/MKLDNNTester.cpp index 99c8c4948c9b05ad15d1217ebb70026bbd48453f..de1635be2af37cd0ba49010199a417090865b0e4 100644 --- a/paddle/gserver/tests/MKLDNNTester.cpp +++ b/paddle/gserver/tests/MKLDNNTester.cpp @@ -330,9 +330,7 @@ void MKLDNNTester::run(const TestConfig& dnn, log_ = log; lvl_ = level; - // Firstly test FLAGS_use_mkldnn_wgt = false - FLAGS_use_mkldnn_wgt = false; - // reset and run once + // Firstly test mkldnn init from PARAM_FORMAT_ORIGINAL weight reset(dnn, ref, batchSize); randomWgtDatas(); clearWgtDiffs(); @@ -342,17 +340,32 @@ void MKLDNNTester::run(const TestConfig& dnn, runOnce(); } - // Then test FLAGS_use_mkldnn_wgt = true - FLAGS_use_mkldnn_wgt = true; - // after run once the mkldnn weight has been stored in dnnlayer + if (parameters_[DNN].empty()) { + // has no paramters + return; + } + + // After run some iterations, the mkldnn weight has been stored in dnnLayer + // and we can also get the mkldnn weight parameter header format. + // Weight parameter should always be index 0 (and bias index 1). + // TODO(TJ): should also consider mean and var format when batchnorm ready + int dnnWgtFmt = parameters_[DNN][0]->getHeaderFormat(); + int refWgtFmt = parameters_[REF][0]->getHeaderFormat(); + if (dnnWgtFmt == refWgtFmt) { + // weight format are equal, so no need check more + return; + } + // then save the weights and restart again vector dnnWgts, refWgts; CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size()); saveWgt(parameters_[DNN], dnnWgts); saveWgt(parameters_[REF], refWgts); - // restart again with flag true + // restart again with dnn weight format reset(dnn, ref, batchSize); + // TODO(TJ): should also considerate mean and var format when batchnorm ready + parameters_[DNN][0]->setHeaderFormat(dnnWgtFmt); // restore wgt restoreWgt(dnnWgts, parameters_[DNN]); diff --git a/paddle/gserver/tests/MKLDNNTester.h b/paddle/gserver/tests/MKLDNNTester.h index 522eeaf24b1949abac057a1e59e9977610be23c0..e55e4493ffdfe45b8cfdee423febd1878b8b3d8a 100644 --- a/paddle/gserver/tests/MKLDNNTester.h +++ b/paddle/gserver/tests/MKLDNNTester.h @@ -108,7 +108,7 @@ private: * if many(>failRate) wrong(abs(dnn-ref)/abs(ref)>thres) points return the * max(diff/ref) * else return sum(abs(a-b)) / sum(abs(b)) - * The return value should smaller than eps when passing. + * The return value should be smaller than eps when passing. */ double getDelta(const real* d1, const real* d2, diff --git a/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp b/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp new file mode 100644 index 0000000000000000000000000000000000000000..538d18cdc3d262df0ddb031d9e6b38a3fea57606 --- /dev/null +++ b/paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp @@ -0,0 +1,353 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES 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 "ModelConfig.pb.h" +#include "paddle/gserver/layers/DataLayer.h" +#include "paddle/trainer/Trainer.h" + +#include "LayerGradUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT + +DECLARE_int32(gpu_id); +DECLARE_bool(thread_local_rand_use_global_seed); + +const size_t MAX_SEQ_NUM = 23; +const size_t MAX_SEQ_LEN = 50; +const size_t MAX_BEAM_SIZE = 27; + +const size_t SEED = (size_t)(time(NULL)); + +struct SingleBeamExpansion { + vector seqStartPos; + vector subSeqStartPos; + vector candidateScores; + + // TODO(caoying): store this into Argument.ids + vector selectedIndices; + + vector groundTruth; + vector inBeam; + vector rowIdxInBeam; + vector colIdxInBeam; + + void resetGroundTruth(size_t n) { + groundTruth.clear(); + groundTruth.resize(n, -1); + + inBeam.clear(); + inBeam.resize(n, 0); + + rowIdxInBeam.clear(); + rowIdxInBeam.resize(n, -1); + + colIdxInBeam.clear(); + colIdxInBeam.resize(n, -1); + } +}; + +inline float randFloat() { + return static_cast(rand()) / static_cast(RAND_MAX); +} + +void genRand(real* numbers, size_t n) { + default_random_engine generator; + uniform_real_distribution distribution(0.0, 1.0); + for (size_t i = 0; i < n; ++i) numbers[i] = distribution(generator); +} + +vector randSampling(real range, int n) { + CHECK_GE(range, n); + vector num(range); + iota(begin(num), end(num), 0.); + if (range == n) return num; + + random_shuffle(begin(num), end(num)); + num.resize(n); + sort(begin(num), end(num)); + return num; +} + +void genCandidateScores(bool hasSubseq, + size_t beamSize, + SingleBeamExpansion& prevBeam, + SingleBeamExpansion& curBeam) { + vector& seqStartPos = curBeam.seqStartPos; + seqStartPos.resize(1, 0); + vector& subSeqStartPos = curBeam.subSeqStartPos; + subSeqStartPos.resize(1, 0); + + srand(SEED); + if (prevBeam.selectedIndices.size()) { + if (prevBeam.subSeqStartPos.size() > 1) { + int seqIdx = 1; + // samples in previous beam are nested sequences. + for (size_t i = 1; i < prevBeam.subSeqStartPos.size(); ++i) { + for (size_t j = 0; j < beamSize; ++j) { + if (prevBeam.selectedIndices[(i - 1) * beamSize + j] == -1.) break; + subSeqStartPos.push_back(1 + (rand() % MAX_SEQ_LEN) + + subSeqStartPos.back()); + } + if (prevBeam.seqStartPos[seqIdx] == prevBeam.subSeqStartPos[i]) { + seqStartPos.push_back(subSeqStartPos.back()); + seqIdx++; + } + } + } else { + for (size_t i = 0; i <= prevBeam.selectedIndices.size(); ++i) { + if (i && i % beamSize == 0) { + seqStartPos.push_back(subSeqStartPos.back()); + if (i == prevBeam.selectedIndices.size()) break; + } + if (prevBeam.selectedIndices[i] == -1.) continue; + subSeqStartPos.push_back(subSeqStartPos.back() + + (1 + (rand() % MAX_SEQ_LEN))); + } + } + } else { + // the first beam expansion + int seqNum = 1 + (rand() % MAX_SEQ_NUM); + for (int i = 0; i < seqNum; ++i) { + if (hasSubseq) { + for (size_t j = 0; j < 1 + (rand() % MAX_SEQ_NUM); ++j) + subSeqStartPos.push_back(subSeqStartPos.back() + + (1 + (rand() % MAX_SEQ_LEN))); + seqStartPos.push_back(subSeqStartPos.back()); + } else { + seqStartPos.push_back(seqStartPos.back() + + (1 + (rand() % MAX_SEQ_LEN))); + } + } + } + + size_t totalSeqNum = hasSubseq ? subSeqStartPos.back() : seqStartPos.back(); + curBeam.candidateScores.resize(totalSeqNum, 0.); + genRand(curBeam.candidateScores.data(), totalSeqNum); +} + +void genSelectedIndices(size_t beamSize, + vector& seqStartPos, + vector& selectedIndices) { + size_t selectedIdsCount = beamSize * (seqStartPos.size() - 1); + selectedIndices.resize(selectedIdsCount, -1.); + + for (size_t i = 0; i < seqStartPos.size() - 1; ++i) { + int seqLen = seqStartPos[i + 1] - seqStartPos[i]; + int n = min(seqLen, static_cast(beamSize)); + vector ids = randSampling(seqLen, n); + memcpy(selectedIndices.data() + i * beamSize, + ids.data(), + sizeof(real) * ids.size()); + } +} + +void genGroundTruth(vector& beamExpansions, + size_t beamSize) { + SingleBeamExpansion& beam = beamExpansions[1]; + size_t seqNum = beam.seqStartPos.size() - 1; + for (size_t i = 2; i < beamExpansions.size(); ++i) + CHECK_EQ(seqNum, beamExpansions[i].seqStartPos.size() - 1); + + srand(SEED); + + // initialize the first beam. + beam.resetGroundTruth(seqNum); + for (size_t i = 0; i < seqNum; ++i) { + if (randFloat() > 0.5) { + /* + * force the randomly generated label falls in the beam by chance 0.5. + * otherwise, when sequence length is relatively long and beam size is + * relatively small, the gold sequences falls off the beam at in the + * first search. + */ + real* begPos = beam.selectedIndices.data() + i * beamSize; + beam.colIdxInBeam[i] = + rand() % count_if(begPos, begPos + beamSize, [](const real& val) { + return val != -1.; + }); + beam.groundTruth[i] = + beam.selectedIndices[i * beamSize + beam.colIdxInBeam[i]]; + beam.inBeam[i] = 1; + } else { + int label = rand() % (beam.seqStartPos[i + 1] - beam.seqStartPos[i]); + beam.groundTruth[i] = label; + + real* begPos = beam.selectedIndices.data() + i * beamSize; + real* endPos = begPos + beamSize; + real* lblPos = find(begPos, endPos, real(label)); + if (lblPos != endPos) { + beam.inBeam[i] = 1; + beam.colIdxInBeam[i] = lblPos - begPos; + } + } + beam.rowIdxInBeam[i] = i; + } + + // iterate over each beam expansions + for (size_t i = 2; i < beamExpansions.size(); ++i) { + SingleBeamExpansion& curBeam = beamExpansions[i]; + SingleBeamExpansion& prevBeam = beamExpansions[i - 1]; + curBeam.resetGroundTruth(seqNum); + + // iterate over each sequence + for (size_t j = 0; j < seqNum; ++j) { + if (!prevBeam.inBeam[j]) continue; + + // gold sequence falls in the beam in previous search. + real* begPos = prevBeam.selectedIndices.data(); + int offset = + prevBeam.rowIdxInBeam[j] * beamSize + prevBeam.colIdxInBeam[j]; + curBeam.rowIdxInBeam[j] = count_if( + begPos, begPos + offset, [](const real& val) { return val != -1.; }); + + if (randFloat() > 0.5) { + // force the randomly generated label falls in the beam by chance 0.5. + + real* start = + curBeam.selectedIndices.data() + curBeam.rowIdxInBeam[j] * beamSize; + int n = rand() % count_if(start, start + beamSize, [](const real& val) { + return val != -1.; + }); + curBeam.colIdxInBeam[j] = n; + curBeam.groundTruth[j] = *(start + n); + curBeam.inBeam[j] = 1; + } else { + CHECK_LE(curBeam.rowIdxInBeam[j] + 1, + curBeam.subSeqStartPos.size() - 1); + int start = curBeam.subSeqStartPos[curBeam.rowIdxInBeam[j]]; + int end = curBeam.subSeqStartPos[curBeam.rowIdxInBeam[j] + 1]; + CHECK_GT(size_t(end), size_t(start)); + int label = rand() % (end - start); + + curBeam.groundTruth[j] = label; + real* findBeg = + curBeam.selectedIndices.data() + curBeam.rowIdxInBeam[j] * beamSize; + real* lblPos = + find(findBeg, findBeg + beamSize, static_cast(label)); + if (lblPos != (findBeg + beamSize)) { + curBeam.inBeam[j] = 1; + curBeam.colIdxInBeam[j] = lblPos - findBeg; + } + } + } + } +} + +void genOneBeam(size_t beamSize, + bool hasSubseq, + SingleBeamExpansion& prevBeam, + SingleBeamExpansion& curBeam) { + genCandidateScores(hasSubseq, beamSize, prevBeam, curBeam); + genSelectedIndices(beamSize, + hasSubseq ? curBeam.subSeqStartPos : curBeam.seqStartPos, + curBeam.selectedIndices); +} + +void genRandomBeamExpansion(size_t expansionCount, + size_t beamSize, + vector& beamExpansions) { + beamExpansions.clear(); + beamExpansions.resize(expansionCount + 1); + + // beamExpansions[0] is reserved. + for (size_t i = 1; i <= expansionCount; ++i) + genOneBeam(beamSize, bool(i - 1), beamExpansions[i - 1], beamExpansions[i]); + genGroundTruth(beamExpansions, beamSize); +} + +void testCrossEntropyOverBeam(bool useGpu, + size_t beamSize, + vector& beams) { + TestConfig config; + config.layerConfig.set_type("cross_entropy_over_beam"); + + size_t seqNum = 0; + for (size_t i = 1; i < beams.size(); ++i) { + const SingleBeamExpansion& beam = beams[i]; + // create scores for all the candidates + MatrixPtr candidateScorePtr = + Matrix::create(beam.candidateScores.size(), 1, false, false); + candidateScorePtr->copyFrom(beam.candidateScores.data(), + beam.candidateScores.size()); + + ostringstream paramName; + paramName << "candidate_scores_" << i; + + if (beam.subSeqStartPos.size() > 1) { + seqNum = beam.subSeqStartPos.size() - 1; + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, + paramName.str(), + candidateScorePtr, + beam.seqStartPos, + beam.subSeqStartPos}); + } else { + seqNum = beam.seqStartPos.size() - 1; + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, + paramName.str(), + candidateScorePtr, + beam.seqStartPos}); + } + config.layerConfig.add_inputs(); + + // create indices for the selected candidates + MatrixPtr selectedCandidates = + Matrix::create(seqNum, beamSize, false, false); + selectedCandidates->copyFrom(beam.selectedIndices.data(), + beam.selectedIndices.size()); + paramName.clear(); + paramName << "selected_candidates_" << i; + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, paramName.str(), selectedCandidates}); + config.layerConfig.add_inputs(); + + // create the ground truth + paramName.clear(); + paramName << "label_" << i; + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, paramName.str(), beam.groundTruth}); + config.layerConfig.add_inputs(); + } + + testLayerGrad( + config, "cross_entropy_over_beam", seqNum, false, useGpu, false); +} + +TEST(Layer, CrossEntropyOverBeam) { + LOG(INFO) << "SEED = " << SEED; + const size_t beamSize = 1 + rand() % MAX_BEAM_SIZE; + LOG(INFO) << "beamSize = " << beamSize; + + // TODO(caoying): test with random beam expansions. + const size_t expansionCount = 3; + vector beams; + genRandomBeamExpansion(expansionCount, beamSize, beams); + + for (bool useGpu : {false, true}) + testCrossEntropyOverBeam(useGpu, beamSize, beams); +} + +int main(int argc, char** argv) { + initMain(argc, argv); + hl_start(); + hl_init(FLAGS_gpu_id); + FLAGS_thread_local_rand_use_global_seed = true; + srand(SEED); + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp index 0f312b6ca50bc1e6317251ba785f1c61a224b54e..090bde7b203652e3ffb1662b8f5b8937885d2608 100644 --- a/paddle/gserver/tests/test_LayerGrad.cpp +++ b/paddle/gserver/tests/test_LayerGrad.cpp @@ -12,6 +12,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#ifndef PADDLE_ONLY_CPU +#include +#endif #include #include #include @@ -189,10 +192,16 @@ TEST(Projection, scaling) { void testProjectionConv(size_t groups, bool isDeconv) { const int NUM_FILTERS = 18; const int FILTER_SIZE = 2; - const int FILTER_SIZE_Y = 4; + const int FILTER_SIZE_Y = 2; const int CHANNELS = 3; const int IMAGE_SIZE = 16; +#if CUDNN_VERSION >= 6000 + const int DILATION = 2; +#else + const int DILATION = 1; +#endif + ProjectionConfig conf; if (isDeconv) { conf.set_type("convt"); @@ -209,6 +218,8 @@ void testProjectionConv(size_t groups, bool isDeconv) { conv->set_padding_y(1); conv->set_stride(2); conv->set_stride_y(2); + conv->set_dilation(DILATION); + conv->set_dilation_y(DILATION); conv->set_groups(groups); if (isDeconv) { conv->set_filter_channels(NUM_FILTERS / conv->groups()); @@ -217,12 +228,12 @@ void testProjectionConv(size_t groups, bool isDeconv) { } conv->set_img_size(IMAGE_SIZE); int output_x = outputSize(conv->img_size(), - conv->filter_size(), + (conv->filter_size() - 1) * DILATION + 1, conv->padding(), conv->stride(), /* caffeMode */ true); int output_y = outputSize(conv->img_size(), - conv->filter_size_y(), + (conv->filter_size_y() - 1) * DILATION + 1, conv->padding_y(), conv->stride_y(), /* caffeMode */ true); @@ -424,27 +435,38 @@ void testConvLayer(const string& type, bool trans, bool useGpu) { config.layerConfig.set_partial_sum(1); config.layerConfig.set_shared_biases(true); - config.inputDefs.push_back({INPUT_DATA, "layer_0", 384, 288}); + int dilation = 1; + if (type == "cudnn_conv") { +#if CUDNN_VERSION >= 6000 + dilation = 2; +#else + dilation = 1; +#endif + } + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 768, 192}); LayerInputConfig* input = config.layerConfig.add_inputs(); ConvConfig* conv = input->mutable_conv_conf(); conv->set_filter_size(2); - conv->set_filter_size_y(3); + conv->set_filter_size_y(2); conv->set_channels(3); conv->set_padding(0); conv->set_padding_y(1); conv->set_stride(2); conv->set_stride_y(2); + conv->set_dilation(dilation); + conv->set_dilation_y(dilation); conv->set_groups(1); conv->set_filter_channels(conv->channels() / conv->groups()); conv->set_img_size(16); - conv->set_img_size_y(8); + conv->set_img_size_y(16); conv->set_output_x(outputSize(conv->img_size(), - conv->filter_size(), + (conv->filter_size() - 1) * dilation + 1, conv->padding(), conv->stride(), /* caffeMode */ true)); conv->set_output_y(outputSize(conv->img_size_y(), - conv->filter_size_y(), + (conv->filter_size_y() - 1) * dilation + 1, conv->padding_y(), conv->stride_y(), /* caffeMode */ true)); @@ -828,9 +850,27 @@ TEST(Layer, square_error_weighted) { } } +TEST(Layer, huber_regression_loss) { + TestConfig config; + config.layerConfig.set_type("huber_regression"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); + config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + for (auto delta : {1, 3, 5}) { + config.layerConfig.set_delta(delta); + testLayerGrad(config, "huber_regression", 100, /* trans */ false, useGpu); + } + } +} + TEST(Layer, huber_two_class) { TestConfig config; - config.layerConfig.set_type("huber"); + config.layerConfig.set_type("huber_classification"); config.biasSize = 0; config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); @@ -839,7 +879,7 @@ TEST(Layer, huber_two_class) { config.layerConfig.add_inputs(); for (auto useGpu : {false, true}) { - testLayerGrad(config, "huber", 100, /* trans */ false, useGpu); + testLayerGrad(config, "huber_two_class", 100, /* trans */ false, useGpu); } } @@ -1206,6 +1246,75 @@ TEST(Layer, PoolLayer) { #endif } +void setPool3DConfig(TestConfig* config, + PoolConfig* pool, + const string& poolType) { + // filter size + const int NUM_FILTERS = 16; + const int FILTER_SIZE = 3; + const int FILTER_SIZE_Y = 3; + const int FILTER_SIZE_Z = 3; + const int CHANNELS = 16; + + (*config).biasSize = 0; + (*config).layerConfig.set_type("pool3d"); + (*config).layerConfig.set_num_filters(NUM_FILTERS); + + int kw = FILTER_SIZE, kh = FILTER_SIZE_Y, kd = FILTER_SIZE_Z; + int pw = 0, ph = 0, pd = 0; + int sw = 2, sh = 2, sd = 2; + + pool->set_pool_type(poolType); + pool->set_pool_type("avg"); + pool->set_channels(CHANNELS); + pool->set_size_x(kw); + pool->set_size_y(kh); + pool->set_size_z(kd); + pool->set_padding(0); + pool->set_padding_y(0); + pool->set_padding_z(0); + pool->set_stride(sw); + pool->set_stride_y(sh); + pool->set_stride_z(sd); + pool->set_start(0); + int ow = outputSize(pool->img_size(), kw, pw, sw, /* caffeMode */ false); + int oh = outputSize(pool->img_size_y(), kh, ph, sh, /* caffeMode */ false); + int od = outputSize(pool->img_size_z(), kd, pd, sd, /* caffeMode */ false); + pool->set_output_x(ow); + pool->set_output_y(oh); + pool->set_output_z(od); +} + +void testPool3DLayer(const string& poolType, bool trans, bool useGpu) { + TestConfig config; + config.inputDefs.push_back({INPUT_DATA, "layer_0", 11664, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + PoolConfig* pool = input->mutable_pool_conf(); + + const int IMAGE_SIZE = 9; + const int IMAGE_SIZE_Y = 9; + const int IMAGE_SIZE_Z = 9; + + pool->set_img_size(IMAGE_SIZE); + pool->set_img_size_y(IMAGE_SIZE_Y); + pool->set_img_size_z(IMAGE_SIZE_Z); + + setPool3DConfig(&config, pool, poolType); + config.layerConfig.set_size(pool->output_x() * pool->output_y() * + pool->channels()); + + testLayerGrad(config, "pool3d", 100, trans, useGpu); +} + +TEST(Layer, Pool3DLayer) { + testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ false); + testPool3DLayer("max", /* trans= */ false, /* useGpu= */ false); +#ifndef PADDLE_ONLY_CPU + testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ true); + testPool3DLayer("max", /* trans= */ false, /* useGpu= */ true); +#endif +} + void testSppLayer(const string& poolType, const int pyramidHeight, bool trans, @@ -1594,6 +1703,55 @@ TEST(Layer, BatchNormalizationLayer) { #endif } +void testBatchNorm3DLayer(const string& type, bool trans, bool useGpu) { + TestConfig config; + const int CHANNELS = 10; + const int IMG_SIZE = 16; + const int IMG_SIZE_Y = 8; + const int IMG_SIZE_Z = 8; + size_t size = CHANNELS * IMG_SIZE * IMG_SIZE_Y * IMG_SIZE_Z; + config.layerConfig.set_type(type); + config.layerConfig.set_size(size); + config.layerConfig.set_active_type("sigmoid"); + config.biasSize = CHANNELS; + config.inputDefs.push_back({INPUT_DATA, + "layer_0", + /* dim= */ size, + /* paraSize= */ CHANNELS}); + + config.inputDefs.push_back({INPUT_DATA, "layer_1_running_mean", 1, CHANNELS}); + config.inputDefs.back().isStatic = true; + config.inputDefs.push_back({INPUT_DATA, "layer_2_running_var", 1, CHANNELS}); + config.inputDefs.back().isStatic = true; + + LayerInputConfig* input = config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + ImageConfig* img_conf = input->mutable_image_conf(); + img_conf->set_channels(CHANNELS); + img_conf->set_img_size(IMG_SIZE); + img_conf->set_img_size_y(IMG_SIZE_Y); + img_conf->set_img_size_z(IMG_SIZE_Z); + + testLayerGrad(config, + "batch_norm", + 64, + /* trans= */ trans, + useGpu, + /* useWeight */ true); +} + +TEST(Layer, testBatchNorm3DLayer) { + testBatchNorm3DLayer("batch_norm", false, false); +#ifndef PADDLE_ONLY_CPU + testBatchNorm3DLayer("batch_norm", false, true); + if (hl_get_cudnn_lib_version() >= int(4000)) { + testBatchNorm3DLayer("cudnn_batch_norm", false, true); + } +#endif +} + void testConvOperator(bool isDeconv) { TestConfig config; const int NUM_FILTERS = 16; @@ -1899,6 +2057,31 @@ TEST(Layer, CropLayer) { } } +TEST(Layer, SwitchOrderLayer) { + TestConfig config; + // config input_0 + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + ImageConfig* img = input->mutable_image_conf(); + img->set_channels(4); + img->set_img_size(16); + img->set_img_size_y(16); + + ReshapeConfig* reshape = config.layerConfig.mutable_reshape_conf(); + reshape->add_height_axis(0); + reshape->add_height_axis(1); + reshape->add_height_axis(2); + reshape->add_width_axis(3); + + // config softmax layer + config.layerConfig.set_type("switch_order"); + config.layerConfig.set_name("switchOrderLayer"); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "switch_order", 100, false, useGpu, true); + } +} + vector randSampling(real range, int n) { CHECK_GE(range, n); vector num(range); @@ -2007,6 +2190,175 @@ TEST(Layer, RowL2NormLayer) { } } +void test3DConvLayer(const string& type, bool trans, bool useGpu) { + // filter size + const int NUM_FILTERS = 6; + // const int CHANNELS = 3; + const int FILTER_SIZE = 3; + const int FILTER_SIZE_Y = 3; + const int FILTER_SIZE_Z = 3; + + // input image + const int CHANNELS = 3; + const int IMAGE_SIZE = 9; + const int IMAGE_SIZE_Y = 9; + const int IMAGE_SIZE_Z = 9; + + TestConfig config; + config.biasSize = NUM_FILTERS; + config.layerConfig.set_type(type); + config.layerConfig.set_num_filters(NUM_FILTERS); + config.layerConfig.set_partial_sum(1); + config.layerConfig.set_shared_biases(true); + + // Setting up conv3D-trans layer + LayerInputConfig* input = config.layerConfig.add_inputs(); + ConvConfig* conv = input->mutable_conv_conf(); + + conv->set_channels(CHANNELS); + conv->set_filter_size(FILTER_SIZE); + conv->set_filter_size_y(FILTER_SIZE_Y); + conv->set_filter_size_z(FILTER_SIZE_Z); + conv->set_padding(0); + conv->set_padding_y(0); + conv->set_padding_z(0); + conv->set_stride(2); + conv->set_stride_y(2); + conv->set_stride_z(2); + conv->set_img_size(IMAGE_SIZE); + conv->set_img_size_y(IMAGE_SIZE_Y); + conv->set_img_size_z(IMAGE_SIZE_Z); + conv->set_output_x(outputSize(conv->img_size(), + conv->filter_size(), + conv->padding(), + conv->stride(), + /* caffeMode */ true)); + conv->set_output_y(outputSize(conv->img_size_y(), + conv->filter_size_y(), + conv->padding_y(), + conv->stride_y(), + /* caffeMode */ true)); + conv->set_output_z(outputSize(conv->img_size_z(), + conv->filter_size_z(), + conv->padding_z(), + conv->stride_z(), + /* caffeMode */ true)); + + config.layerConfig.set_size(conv->output_x() * conv->output_y() * + conv->output_z() * NUM_FILTERS); + conv->set_groups(1); + conv->set_filter_channels(conv->channels() / conv->groups()); + config.inputDefs.push_back( + {INPUT_DATA, + "layer_0", + CHANNELS * IMAGE_SIZE * IMAGE_SIZE_Y * IMAGE_SIZE_Z, + conv->filter_channels() * FILTER_SIZE * FILTER_SIZE_Y * FILTER_SIZE_Z * + NUM_FILTERS}); + + testLayerGrad(config, "conv3D", 10, trans, useGpu); + // Use small batch_size and useWeight=true to test biasGrad + testLayerGrad(config, "conv3D", 2, trans, useGpu, true, 0.02); +} + +TEST(Layer, test3DConvLayer) { + test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ false); +#ifndef PADDLE_ONLY_CPU + test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ true); +#endif +} + +void test3DDeConvLayer(const string& type, bool trans, bool useGpu) { + // filter size + const int NUM_FILTERS = 6; + // const int CHANNELS = 3; + const int FILTER_SIZE = 3; + const int FILTER_SIZE_Y = 3; + const int FILTER_SIZE_Z = 3; + + // input image + const int CHANNELS = 3; + const int IMAGE_SIZE = 4; + const int IMAGE_SIZE_Y = 6; + const int IMAGE_SIZE_Z = 6; + + // Setting up conv-trans layer + TestConfig config; + config.biasSize = NUM_FILTERS; + config.layerConfig.set_type("deconv3d"); + config.layerConfig.set_num_filters(NUM_FILTERS); + config.layerConfig.set_partial_sum(1); + config.layerConfig.set_shared_biases(true); + + LayerInputConfig* input = config.layerConfig.add_inputs(); + ConvConfig* conv = input->mutable_conv_conf(); + + conv->set_channels(CHANNELS); + conv->set_filter_size(FILTER_SIZE); + conv->set_filter_size_y(FILTER_SIZE_Y); + conv->set_filter_size_z(FILTER_SIZE_Z); + conv->set_padding(0); + conv->set_padding_y(0); + conv->set_padding_z(0); + conv->set_stride(2); + conv->set_stride_y(2); + conv->set_stride_z(2); + conv->set_output_x(IMAGE_SIZE); + conv->set_output_y(IMAGE_SIZE_Y); + conv->set_output_z(IMAGE_SIZE_Z); + + conv->set_img_size(imageSize(conv->output_x(), + conv->filter_size(), + conv->padding(), + conv->stride(), + true)); + conv->set_img_size_y(imageSize(conv->output_y(), + conv->filter_size_y(), + conv->padding_y(), + conv->stride_y(), + true)); + conv->set_img_size_z(imageSize(conv->output_z(), + conv->filter_size_z(), + conv->padding_z(), + conv->stride_z(), + true)); + config.layerConfig.set_size(conv->img_size() * conv->img_size_y() * + conv->img_size_z() * NUM_FILTERS); + conv->set_groups(1); + conv->set_filter_channels(conv->channels() / conv->groups()); + config.inputDefs.push_back( + {INPUT_DATA, + "layer_0", + CHANNELS * IMAGE_SIZE * IMAGE_SIZE_Y * IMAGE_SIZE_Z, + conv->filter_channels() * FILTER_SIZE * FILTER_SIZE_Y * FILTER_SIZE_Z * + NUM_FILTERS}); + + testLayerGrad(config, "deconv3D", 10, trans, useGpu); + // Use small batch_size and useWeight=true to test biasGrad + testLayerGrad(config, "deconv3D", 2, trans, useGpu, true, 0.02); +} + +TEST(Layer, test3DDeConvLayer) { + test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ false); +#ifndef PADDLE_ONLY_CPU + test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ true); +#endif +} + +TEST(Layer, ScaleShiftLayer) { + const size_t batchSize = 16; + const size_t size = 32; + TestConfig config; + config.layerConfig.set_type("scale_shift"); + config.layerConfig.set_size(size); + config.biasSize = 1; + config.inputDefs.push_back( + {INPUT_DATA, "input", /* dim= */ size, /* paraSize= */ 1}); + config.layerConfig.add_inputs(); + for (auto useGpu : {false, true}) { + testLayerGrad(config, "scale_shift", batchSize, false, useGpu, false); + } +} + int main(int argc, char** argv) { testing::InitGoogleTest(&argc, argv); initMain(argc, argv); diff --git a/paddle/gserver/tests/test_NetworkCompare.cpp b/paddle/gserver/tests/test_NetworkCompare.cpp index f930c72fde3f5e0a6a45cb6bfd3507a4f48028fc..d36f72360f8ebd2033fb3e8c0e1b30911abba362 100644 --- a/paddle/gserver/tests/test_NetworkCompare.cpp +++ b/paddle/gserver/tests/test_NetworkCompare.cpp @@ -269,7 +269,8 @@ TEST(Compare, img_conv2) { bool useGpu = FLAGS_use_gpu; double eps = FLAGS_checkgrad_eps; FLAGS_use_gpu = true; - FLAGS_checkgrad_eps = 1e-2; + // Sometimes, this unit test will fail with 1e-2 + FLAGS_checkgrad_eps = 4e-2; compareNetwork(config_file_a, config_file_b); FLAGS_use_gpu = useGpu; FLAGS_checkgrad_eps = eps; diff --git a/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp b/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e1d4ae16176433b898ba88dd60550e44b4fe37be --- /dev/null +++ b/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp @@ -0,0 +1,225 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES 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 "ModelConfig.pb.h" +#include "paddle/gserver/layers/DataLayer.h" +#include "paddle/trainer/Trainer.h" + +#include "LayerGradUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_int32(gpu_id); +DECLARE_bool(thread_local_rand_use_global_seed); + +const int MAX_SEQ_NUM = 17; +const int MAX_SEQ_LEN = 23; +const int MAX_BEAM_SIZE = 13; + +const size_t SEED = (size_t)(time(NULL)); + +vector randSampling(real range, int n) { + CHECK_GE(range, n); + vector num(range); + iota(begin(num), end(num), 0.); + if (range == n) return num; + + random_shuffle(begin(num), end(num)); + num.resize(n); + sort(begin(num), end(num)); + return num; +} + +void genSeqInfo(vector& seqStartPos, vector& subSeqStartPos) { + seqStartPos.resize(1, 0); + subSeqStartPos.resize(1, 0); + + srand(SEED); + int seqNum = 1 + (rand() % MAX_SEQ_NUM); + for (int i = 0; i < seqNum; ++i) { + int subSeqNum = 1 + (rand() % MAX_SEQ_NUM); + for (int j = 0; j < subSeqNum; ++j) + subSeqStartPos.push_back(subSeqStartPos.back() + + (1 + (rand() % MAX_SEQ_LEN))); + seqStartPos.push_back(subSeqStartPos.back()); + } +} + +/* + generate start indices according to sequence start positions. + */ +void genStarts(vector& seqStartPos, + vector>& starts, + size_t beamSize) { + starts.clear(); + starts.resize(seqStartPos.size() - 1, vector(beamSize, -1.)); + + for (size_t i = 0; i < seqStartPos.size() - 1; ++i) { + int seqLen = seqStartPos[i + 1] - seqStartPos[i]; + vector randStarts = + randSampling(seqLen, min(seqLen, static_cast(beamSize))); + copy(begin(randStarts), end(randStarts), begin(starts[i])); + } +} + +/* + generate end indices according to sequence start positions and start indices. + */ +void genEnds(vector& seqStartPos, + vector>& starts, + vector>& ends, + size_t beamSize) { + CHECK_EQ(seqStartPos.size() - 1, starts.size()); + ends.clear(); + ends.resize(seqStartPos.size() - 1, vector(beamSize, -1.)); + + for (size_t i = 0; i < starts.size(); ++i) { + for (size_t j = 0; j < starts[i].size(); ++j) { + int seqLen = seqStartPos[i + 1] - seqStartPos[i]; + CHECK_GE(seqLen - 1, starts[i][j]); + if (starts[i][j] == -1.) break; + if (starts[i][j] == (seqLen - 1)) { + ends[i][j] = starts[i][j]; + } else { + ends[i][j] = starts[i][j] + randSampling(seqLen - starts[i][j], 1)[0]; + } + } + } +} + +void genTestData(vector& seqStartPos, + vector& subSeqStartPos, + vector>& starts, + vector>& ends, + bool hasSubseq) { + size_t beamSize = 1 + (rand() % MAX_BEAM_SIZE); + genSeqInfo(seqStartPos, subSeqStartPos); + + genStarts(hasSubseq ? subSeqStartPos : seqStartPos, starts, beamSize); + genEnds(hasSubseq ? subSeqStartPos : seqStartPos, starts, ends, beamSize); +} + +template +void flatten2dVector(vector>& inVec, vector& outVec) { + size_t totalSize{0}; + for (auto const& items : inVec) totalSize += items.size(); + outVec.reserve(totalSize); + + for (auto& items : inVec) + move(items.begin(), items.end(), back_inserter(outVec)); +} + +void testSeqSliceLayer(bool hasSubseq, + bool useGpu, + vector& seqStartPos, + vector& subSeqStartPos, + vector>& starts, + vector>& ends) { + // layer size is not crutial for this layer, + // so here use a small layer size in the unittest. + const size_t layerSize{4}; + TestConfig config; + config.layerConfig.set_type("seq_slice"); + config.layerConfig.set_size(layerSize); + + // add the first input + MatrixPtr seqInputPtr = + Matrix::create(hasSubseq ? subSeqStartPos.back() : seqStartPos.back(), + layerSize, + false, + false); + seqInputPtr->randomizeUniform(); + + if (hasSubseq) { + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, + "seq_input", + seqInputPtr, + seqStartPos, + subSeqStartPos}); + } else { + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, "seq_input", seqInputPtr, seqStartPos}); + } + config.layerConfig.add_inputs(); + + // add start indices + if (starts.size()) { + vector startsToVec; + flatten2dVector(starts, startsToVec); + + MatrixPtr startMatrixPtr = + Matrix::create(starts.size(), starts[0].size(), false, false); + startMatrixPtr->copyFrom(startsToVec.data(), startsToVec.size()); + + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, "starts", startMatrixPtr}); + config.layerConfig.add_inputs(); + config.layerConfig.set_select_first(true); + } + + // add end indices + if (ends.size()) { + vector endsToVec; + flatten2dVector(ends, endsToVec); + + MatrixPtr endMatrixPtr = + Matrix::create(ends.size(), ends[0].size(), false, false); + endMatrixPtr->copyFrom(endsToVec.data(), endsToVec.size()); + + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "ends", endMatrixPtr}); + config.layerConfig.add_inputs(); + config.layerConfig.set_select_first(false); + } + + testLayerGrad(config, "seq_slice", /*batchSize*/ 100, false, useGpu, false); +} + +TEST(Layer, SeqSliceLayer) { + vector seqStartPos; + vector subSeqStartPos; + vector> starts; + vector> ends; + + std::vector mode = {false}; +#ifndef PADDLE_ONLY_CPU + mode.push_back(true); +#endif + genSeqInfo(seqStartPos, subSeqStartPos); + for (bool hasSubseq : {true, false}) { + LOG(INFO) << "hasSubSeq : " << hasSubseq; + genTestData(seqStartPos, subSeqStartPos, starts, ends, hasSubseq); + for (bool useGpu : mode) { + vector> tmp; + testSeqSliceLayer( + hasSubseq, useGpu, seqStartPos, subSeqStartPos, tmp, ends); + testSeqSliceLayer( + hasSubseq, useGpu, seqStartPos, subSeqStartPos, starts, tmp); + testSeqSliceLayer( + hasSubseq, useGpu, seqStartPos, subSeqStartPos, starts, ends); + } + } +} + +int main(int argc, char** argv) { + initMain(argc, argv); + hl_start(); + hl_init(FLAGS_gpu_id); + FLAGS_thread_local_rand_use_global_seed = true; + srand(1); + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/math/Allocator.h b/paddle/math/Allocator.h index 666a8b8368e3e2ebc522902c176d7491d2920d2a..94ef561f066a127496e2849a419835e175c526d7 100644 --- a/paddle/math/Allocator.h +++ b/paddle/math/Allocator.h @@ -48,7 +48,13 @@ public: */ virtual void* alloc(size_t size) { void* ptr; +#ifdef PADDLE_USE_MKLDNN + // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp + // memory alignment + CHECK_EQ(posix_memalign(&ptr, 4096ul, size), 0); +#else CHECK_EQ(posix_memalign(&ptr, 32ul, size), 0); +#endif CHECK(ptr) << "Fail to allocate CPU memory: size=" << size; return ptr; } diff --git a/paddle/math/CMakeLists.txt b/paddle/math/CMakeLists.txt index bf28092e82b778dc904c5a2e271f76261cf5f6b6..68b5296228cd733dc3cb7ca0f762e0a69187dbff 100644 --- a/paddle/math/CMakeLists.txt +++ b/paddle/math/CMakeLists.txt @@ -14,6 +14,17 @@ # file(GLOB MATH_HEADERS . *.h) file(GLOB MATH_SOURCES . *.cpp) + +if(NOT WITH_MKLDNN) + set(DNN_HEADER "${CMAKE_CURRENT_SOURCE_DIR}/MKLDNNMatrix.h") + set(DNN_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/MKLDNNMatrix.cpp") + list(REMOVE_ITEM MATH_HEADERS "${DNN_HEADER}") + list(REMOVE_ITEM MATH_SOURCES "${DNN_SOURCE}") + message(STATUS "Skip compiling with MKLDNNMatrix") +else() + message(STATUS "Compile with MKLDNNMatrix") +endif() + set(MATH_SOURCES "${PADDLE_SOURCE_DIR}/paddle/math/BaseMatrix.cu" "${PADDLE_SOURCE_DIR}/paddle/math/TrainingAlgorithmOp.cu" diff --git a/paddle/math/MKLDNNMatrix.cpp b/paddle/math/MKLDNNMatrix.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c4063e5069854242d9f93886b66580385557ca73 --- /dev/null +++ b/paddle/math/MKLDNNMatrix.cpp @@ -0,0 +1,142 @@ +/* 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 "MKLDNNMatrix.h" + +using namespace mkldnn; // NOLINT + +namespace paddle { + +MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) { + memory::desc md = pd.desc(); + size_t ndims = md.data.ndims; + int* dims = md.data.dims; + CHECK(ndims > 0) << "Input dims should not be empty"; + size_t cnts = 1; + for (size_t i = 0; i < ndims; ++i) { + cnts *= dims[i]; + } + + if (m == nullptr) { + size_t height = dims[0]; + size_t width = cnts / dims[0]; + m = Matrix::create(height, width, false, false); + } + CHECK(m) << " Matrix should not be empty"; + + CpuMatrixPtr cpuMatrix = std::dynamic_pointer_cast(m); + CHECK(cpuMatrix) << "Only support create from CPU matrix yet"; + CHECK_EQ(cpuMatrix->getElementCnt(), cnts) << "Count size does not match"; + return std::make_shared(cpuMatrix, pd); +} + +MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, + memory::dims dims, + memory::format fmt, + engine& eg, + mkldnn::memory::data_type dtype) { + return create(m, memory::primitive_desc(memory::desc(dims, dtype, fmt), eg)); +} + +void MKLDNNMatrix::reorderDataFrom(const MKLDNNMatrixPtr& m, + memory::format srcFmt, + memory::dims targetDim) { + memory::format dstFmt = getFormat(); + if (srcFmt == dstFmt) { + return; + } + CHECK_EQ(getElementCnt(), m->getElementCnt()) << "size should equal"; + reorderOnce(getData(), m->getData(), srcFmt, dstFmt, targetDim); +} + +void MKLDNNMatrix::reorderDataTo(const MKLDNNMatrixPtr& m, + memory::format dstFmt, + memory::dims targetDim) { + memory::format srcFmt = getFormat(); + if (srcFmt == dstFmt) { + return; + } + CHECK_EQ(getElementCnt(), m->getElementCnt()) << "size should equal"; + reorderOnce(getData(), m->getData(), srcFmt, dstFmt, targetDim); +} + +void MKLDNNMatrix::reorderOnce(void* srcData, + void* dstData, + memory::format srcFmt, + memory::format dstFmt, + memory::dims dm) { + CHECK(srcData); + CHECK(dstData); + MatrixPtr tmpSrc; + if (dstData == srcData) { + // inplace data + size_t sz = 1; + for (size_t i = 0; i < dm.size(); ++i) { + sz *= dm[i]; + } + tmpSrc = Matrix::create(sz, 1, false, false); + tmpSrc->copyFrom((real*)srcData, sz); + srcData = tmpSrc->getData(); + } + + auto dtype = this->getDtype(); + auto srcMD = memory::desc(dm, dtype, srcFmt); + auto dstMD = memory::desc(dm, dtype, dstFmt); + + auto eg = this->getEngine(); + auto src = memory(memory::primitive_desc(srcMD, eg), srcData); + auto dst = memory(memory::primitive_desc(dstMD, eg), dstData); + + auto r = reorder(src, dst); + stream(stream::kind::eager).submit({r}).wait(); +} + +void MKLDNNMatrix::downSpatial() { + int fmt = getFormat(); + if (!(fmt == memory::format::nchw || fmt == memory::format::oihw)) { + // only support nchw and oihw yet, later can support more like nhwc, ihwo + return; + } + + // TODO(TJ): change H(height) and W(width) if support nhwc or more + const int H = 2, W = 3; + memory::dims srcDims = getDims(); + if (srcDims[H] != 1 || srcDims[W] != 1) { + // can not down spatial + return; + } + + memory::dims dstDims = memory::dims{srcDims[0], srcDims[1]}; + memory::format dstFmt; + switch (fmt) { + case memory::format::nchw: + dstFmt = memory::format::nc; + break; + case memory::format::oihw: + dstFmt = memory::format::oi; + break; + default: + LOG(FATAL) << "unsupported format"; + } + memory::desc md = memory::desc(dstDims, getDtype(), dstFmt); + memory::primitive_desc pd = memory::primitive_desc(md, getEngine()); + mkldnn_primitive_t result; + mkldnn::error::wrap_c_api( + mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr), + "could not create a memory primitive"); + reset(result); + set_data_handle(data_); +} + +} // namespace paddle diff --git a/paddle/math/MKLDNNMatrix.h b/paddle/math/MKLDNNMatrix.h new file mode 100644 index 0000000000000000000000000000000000000000..eef3b429e6fa0087aeac3f5aed9dff983b06e826 --- /dev/null +++ b/paddle/math/MKLDNNMatrix.h @@ -0,0 +1,169 @@ +/* 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 +#include "Matrix.h" +#include "mkldnn.hpp" +#include "paddle/parameter/Parameter.h" + +namespace paddle { + +class MKLDNNMatrix; +typedef std::shared_ptr MKLDNNMatrixPtr; + +/** + * @brief MKLDNN Matrix. + * + */ +class MKLDNNMatrix : public CpuMatrix, public mkldnn::memory { +public: + MKLDNNMatrix(CpuMatrixPtr m, mkldnn::memory::primitive_desc pd) + : CpuMatrix(m->getData(), m->getHeight(), m->getWidth(), false), + mkldnn::memory(pd, m->getData()), + m_(m) {} + + ~MKLDNNMatrix() {} + + /** + * Create MKLDNNMatrix from a MatrixPtr and memory primitive_desc + */ + static MKLDNNMatrixPtr create(MatrixPtr m, mkldnn::memory::primitive_desc pd); + + /** + * Create MKLDNNMatrix from a MatrixPtr and memory details info + */ + static MKLDNNMatrixPtr create( + MatrixPtr m, + mkldnn::memory::dims dims, + mkldnn::memory::format fmt, + mkldnn::engine& eg, + mkldnn::memory::data_type dtype = mkldnn::memory::data_type::f32); + +public: + /** + * Reorder this MKLDNNMatrix from other format. + * Support inplace reorder. + * @note: this function would only reorder the data layout. + * will NOT change this original dim or format info + */ + void reorderDataFrom(const MKLDNNMatrixPtr& m, + memory::format srcFmt, + memory::dims targetDim); + + /** + * Reorder this MKLDNNMatrix to other format. + * Support inplace reorder. + * @note: this function would only reorder the data layout. + * will NOT change the dst dim or format info + */ + void reorderDataTo(const MKLDNNMatrixPtr& m, + memory::format dstFmt, + memory::dims targetDim); + + /** + * Dimensionality reduction. + * Change format "nchw --> nc" or "oihw --> oi" if the h and w are both 1 + */ + void downSpatial(); + + /** + * set the memory data handle. + * Caution: This will not check the buffer size of the data, + * it should be coverd by user. + */ + void setData(real* data) { + set_data_handle(data); + CpuMatrix::setData(data); + m_.reset(); + } + + /** + * override Matrix::getData + * check data before return + */ + real* getData() override { + CHECK_EQ((void*)data_, get_data_handle()); + return data_; + } + + const real* getData() const override { + CHECK_EQ((void*)data_, get_data_handle()); + return data_; + } + + /** + * Get primitive descriptor. + */ + mkldnn::memory::primitive_desc getPrimitiveDesc() { + return this->get_primitive_desc(); + } + + /** + * Get memory descriptor. + */ + mkldnn::memory::desc getMemoryDesc() { return getPrimitiveDesc().desc(); } + + /** + * Get dimensions. + */ + mkldnn::memory::dims getDims() { + mkldnn::memory::desc md = getMemoryDesc(); + const int* src = md.data.dims; + int ndims = md.data.ndims; + mkldnn::memory::dims dst; + dst.resize(ndims); + for (int i = 0; i < ndims; ++i) { + dst[i] = src[i]; + } + return dst; + } + + /** + * Get format. + */ + mkldnn::memory::format getFormat() { + return (mkldnn::memory::format)(getMemoryDesc().data.format); + } + + /** + * Get memory data type. + */ + mkldnn::memory::data_type getDtype() { + return (mkldnn::memory::data_type)(getMemoryDesc().data.data_type); + } + + /** + * Get engine. + */ + mkldnn::engine getEngine() { return getPrimitiveDesc().get_engine(); } + +protected: + /** + * Do reorder once. + * Can support inplace. + */ + void reorderOnce(void* srcData, + void* dstData, + memory::format srcFmt, + memory::format dstFmt, + memory::dims dm); + +private: + // save the CpuMatrixPtr in case the buffer released outside + CpuMatrixPtr m_; +}; + +} // namespace paddle diff --git a/paddle/math/MathFunctions.cpp b/paddle/math/MathFunctions.cpp index c8ba1074a1555bbddde7e5f0fb2a046138b27c09..c2f17beeb87942ea681f5d388659c0d280157b26 100644 --- a/paddle/math/MathFunctions.cpp +++ b/paddle/math/MathFunctions.cpp @@ -84,6 +84,7 @@ LAPACK_ROUTINE_EACH(DYNAMIC_LOAD_LAPACK_WRAP) namespace paddle { +#ifndef PADDLE_USE_EIGEN_FOR_BLAS template <> void gemm(const CBLAS_TRANSPOSE transA, const CBLAS_TRANSPOSE transB, @@ -143,6 +144,7 @@ void gemm(const CBLAS_TRANSPOSE transA, C, ldc); } +#endif template <> int getrf(const CBLAS_ORDER order, @@ -182,6 +184,7 @@ int getri(const CBLAS_ORDER order, return dynload::PADDLE_DGETRI(order, N, A, lda, ipiv); } +#ifndef PADDLE_USE_EIGEN_FOR_BLAS template <> void axpy(const int n, const float alpha, const float* x, float* y) { cblas_saxpy(n, alpha, x, 1, y, 1); @@ -201,6 +204,7 @@ template <> double dotProduct(const int n, const double* x, const double* y) { return cblas_ddot(n, x, 1, y, 1); } +#endif #if defined(PADDLE_USE_MKL) || defined(PADDLE_USE_MKLML) diff --git a/paddle/math/MathFunctions.h b/paddle/math/MathFunctions.h index 637643838ff433753e0cbb9154ee069c2f7c6d15..e8ea6e37ac527a19c529d1731b94bed970211755 100644 --- a/paddle/math/MathFunctions.h +++ b/paddle/math/MathFunctions.h @@ -40,7 +40,14 @@ extern "C" { #ifndef LAPACK_FOUND extern "C" { +#ifndef PADDLE_USE_EIGEN_FOR_BLAS #include +#else +typedef enum CBLAS_ORDER { + CblasRowMajor = 101, + CblasColMajor = 102 +} CBLAS_ORDER; +#endif int LAPACKE_sgetrf( int matrix_layout, int m, int n, float* a, int lda, int* ipiv); int LAPACKE_dgetrf( @@ -56,6 +63,7 @@ int LAPACKE_dgetri( namespace paddle { +#ifndef PADDLE_USE_EIGEN_FOR_BLAS template void gemm(const CBLAS_TRANSPOSE transA, const CBLAS_TRANSPOSE transB, @@ -70,6 +78,7 @@ void gemm(const CBLAS_TRANSPOSE transA, const T beta, T* C, const int ldc); +#endif template int getrf(const CBLAS_ORDER Order, @@ -84,10 +93,21 @@ int getri( const CBLAS_ORDER Order, const int N, T* A, const int lda, const int* ipiv); template -void axpy(const int n, const T alpha, const T* x, T* y); +void axpy(const int n, const T alpha, const T* x, T* y) { + /// y = y + alpha * x + for (int i = 0; i < n; i++) { + y[i] = y[i] + alpha * x[i]; + } +} template -T dotProduct(const int n, const T* x, const T* y); +T dotProduct(const int n, const T* x, const T* y) { + T result = static_cast(0); + for (int i = 0; i < n; i++) { + result += x[i] * y[i]; + } + return result; +} template void vExp(const int n, const T* a, T* r); diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp index 27f7d95b752d4a423bf99fa425b10b2816575d6a..4a2132c8d1bfa329ced575f9b78052bdbfe3e4d5 100644 --- a/paddle/math/Matrix.cpp +++ b/paddle/math/Matrix.cpp @@ -28,6 +28,7 @@ limitations under the License. */ #include "hl_top_k.h" #include "paddle/utils/Logging.h" +#include "paddle/function/GemmFunctor.h" #include "paddle/utils/ThreadLocal.h" #include "SIMDFunctions.h" @@ -1190,6 +1191,221 @@ void GpuMatrix::avgPoolBackward(Matrix& outGrad, outGrad.getStride()); } +void GpuMatrix::maxPool3DForward(Matrix& inputMat, + Matrix& maxPoolIdx, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW) { + CHECK(inputMat.useGpu_) << "Matrix type are not correct"; + + real* inputData = inputMat.getData(); + real* maxPoolIdxData = maxPoolIdx.getData(); + size_t num = inputMat.getHeight(); + size_t width = imgSizeW; + size_t height = imgSizeH; + size_t depth = imgSizeD; + CHECK(depth * height * width * channels == inputMat.getWidth()); + CHECK(height_ == inputMat.getHeight()); + CHECK(width_ == outputD * outputH * outputW * channels); + + hl_maxpool3D_forward(num, + inputData, + channels, + depth, + height, + width, + outputD, + outputH, + outputW, + sizeZ, + sizeY, + sizeX, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + getData(), + maxPoolIdxData, + getStride()); +} + +void GpuMatrix::maxPool3DBackward(Matrix& outGrad, + Matrix& maxPoolIdx, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput) { + CHECK(outGrad.useGpu_ && maxPoolIdx.useGpu_) << "Matrix type are not equal"; + + real* outDiff = outGrad.getData(); + real* maxPoolIdxData = maxPoolIdx.getData(); + size_t frameNum = getHeight(); + size_t channels = outGrad.getWidth() / outputD / outputH / outputW; + size_t width = imgSizeW; + size_t height = imgSizeH; + size_t depth = imgSizeD; + CHECK(depth * height * width * channels == getWidth()); + CHECK(width_ == depth * width * height * channels); + CHECK(outGrad.getHeight() == maxPoolIdx.getHeight() && + outGrad.getWidth() == maxPoolIdx.getWidth()); + + hl_maxpool3D_backward(frameNum, + outDiff, + channels, + depth, + height, + width, + outputD, + outputH, + outputW, + sizeZ, + sizeY, + sizeX, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + scaleTargets, + scaleOutput, + getData(), + maxPoolIdxData, + outGrad.getStride()); +} + +void GpuMatrix::avgPool3DForward(Matrix& inputMat, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW) { + CHECK(inputMat.useGpu_) << "Matrix type are not equal"; + + real* inputData = inputMat.getData(); + size_t frameNum = inputMat.getHeight(); + size_t height = imgSizeH; + size_t width = imgSizeW; + size_t depth = imgSizeD; + CHECK(depth * height * width * channels == inputMat.getWidth()); + CHECK(height_ == inputMat.getHeight()); + CHECK(width_ == outputD * outputH * outputW * channels); + + hl_avgpool3D_forward(frameNum, + inputData, + channels, + depth, + height, + width, + outputD, + outputH, + outputW, + sizeZ, + sizeY, + sizeX, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + getData(), + getStride()); +} + +void GpuMatrix::avgPool3DBackward(Matrix& outGrad, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput) { + CHECK(outGrad.useGpu_) << "Matrix type are not equal"; + + real* outDiff = outGrad.getData(); + size_t frameNum = outGrad.getHeight(); + size_t channels = outGrad.getWidth() / outputD / outputH / outputW; + size_t height = imgSizeH; + size_t width = imgSizeW; + size_t depth = imgSizeD; + CHECK(depth * height * width * channels == width_); + CHECK(height_ == outGrad.getHeight()); + CHECK(outGrad.getWidth() == outputD * outputH * outputW * channels); + + hl_avgpool3D_backward(frameNum, + outDiff, + channels, + depth, + height, + width, + outputD, + outputH, + outputW, + sizeZ, + sizeY, + sizeX, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + scaleTargets, + scaleOutput, + getData(), + outGrad.getStride()); +} + void GpuMatrix::maxSequenceForward(Matrix& input, const IVector& sequence, IVector& index) { @@ -1389,6 +1605,72 @@ void GpuMatrix::multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label) { output_d, grad_d, mat_d, height_, width_); } +void GpuMatrix::vol2Col(real* dataSrc, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW) { + hl_matrix_vol2Col(dataSrc, + channels, + depth, + height, + width, + filterD, + filterH, + filterW, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + getData()); +} + +void GpuMatrix::col2Vol(real* dataDst, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real alpha, + real beta) { + hl_matrix_col2Vol(dataDst, + channels, + depth, + height, + width, + filterD, + filterH, + filterW, + strideD, + strideH, + strideW, + paddingD, + paddingH, + paddingW, + getData(), + alpha, + beta); +} + /** * CpuMatrix */ @@ -1930,6 +2212,276 @@ void CpuMatrix::avgPoolBackward(Matrix& input, } } +void CpuMatrix::maxPool3DForward(Matrix& inputMat, + Matrix& maxPoolIdx, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW) { + real* inputData = inputMat.getData(); + real* outData = getData(); + real* maxPoolIdxData = maxPoolIdx.getData(); + size_t num = inputMat.getHeight(); + size_t inWidth = imgSizeW; + size_t inHeight = imgSizeH; + size_t inDepth = imgSizeD; + CHECK(inHeight * inWidth * inDepth == inputMat.getWidth() / channels); + CHECK_EQ(num, this->getHeight()); + CHECK_EQ(channels * outputH * outputW * outputD, this->getWidth()); + size_t outStride = getStride(); + + /* initialize the data_ */ + for (size_t i = 0; i < height_; i++) { + for (size_t j = 0; j < width_; j++) { + outData[(i)*outStride + j] = -(real)FLT_MAX; + maxPoolIdxData[(i)*outStride + j] = -1; + } + } + + /* pool max one by one */ + for (size_t n = 0; n < num; ++n) { // frame by frame + if (!isContiguous()) { + outData = getData() + n * outStride; + maxPoolIdxData = maxPoolIdx.getData() + n * outStride; + } + for (size_t c = 0; c < channels; ++c) { // channel by channel + for (size_t pd = 0; pd < outputD; ++pd) { + for (size_t ph = 0; ph < outputH; ++ph) { + for (size_t pw = 0; pw < outputW; ++pw) { + int dstart = pd * strideD - paddingD; + int hstart = ph * strideH - paddingH; + int wstart = pw * strideW - paddingW; + int dend = std::min(dstart + sizeZ, inDepth); + int hend = std::min(hstart + sizeY, inHeight); + int wend = std::min(wstart + sizeX, inWidth); + dstart = std::max(dstart, 0); + hstart = std::max(hstart, 0); + wstart = std::max(wstart, 0); + int maxIdx = -1; + real maxOutData = outData[(pd * outputH + ph) * outputW + pw]; + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + if (maxOutData < + inputData[(d * inHeight + h) * inWidth + w]) { + maxOutData = inputData[(d * inHeight + h) * inWidth + w]; + maxIdx = (d * inHeight + h) * inWidth + w; + } + } + } + } + outData[(pd * outputH + ph) * outputW + pw] = maxOutData; + maxPoolIdxData[(pd * outputH + ph) * outputW + pw] = maxIdx; + } + } + } + // compute offset + inputData += inDepth * inHeight * inWidth; + outData += outputD * outputH * outputW; + maxPoolIdxData += outputD * outputH * outputW; + } + } +} + +void CpuMatrix::maxPool3DBackward(Matrix& outGrad, + Matrix& maxPoolIdx, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput) { + size_t num = getHeight(); + size_t channels = size_t(width_ / imgSizeD / imgSizeH / imgSizeW); + CHECK(maxPoolIdx.getHeight() == outGrad.getHeight() && + maxPoolIdx.getWidth() == outGrad.getWidth()); + + real* tgtGrad = getData(); + real* otGrad = outGrad.getData(); + real* maxPoolIdxData = maxPoolIdx.getData(); + size_t outStride = outGrad.getStride(); + + for (size_t n = 0; n < num; ++n) { + if (!outGrad.isContiguous()) { + otGrad = outGrad.getData() + n * outStride; + maxPoolIdxData = maxPoolIdx.getData() + n * outStride; + } + for (size_t c = 0; c < channels; ++c) { + for (size_t pd = 0; pd < outputD; ++pd) { + for (size_t ph = 0; ph < outputH; ++ph) { + for (size_t pw = 0; pw < outputW; ++pw) { + const size_t index = (pd * outputH + ph) * outputW + pw; + const size_t tgtIdx = static_cast(maxPoolIdxData[index]); + tgtGrad[tgtIdx] = + scaleTargets * tgtGrad[tgtIdx] + scaleOutput * otGrad[index]; + } + } + } + // offset + tgtGrad += imgSizeD * imgSizeH * imgSizeW; + otGrad += outputD * outputH * outputW; + maxPoolIdxData += outputD * outputH * outputW; + } + } +} + +void CpuMatrix::avgPool3DForward(Matrix& input, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW) { + // The main loop + size_t num = input.getHeight(); + size_t inDepth = imgSizeD; + size_t inHeight = imgSizeH; + size_t inWidth = imgSizeW; + CHECK(inDepth * inHeight * inWidth * channels == input.getWidth()); + CHECK(outputD * outputH * outputW * channels * num == height_ * width_); + real* tgtData = getData(); + real* inData = input.getData(); + + for (size_t n = 0; n < num; ++n) { + if (!isContiguous()) { + tgtData = data_ + n * getStride(); + } + for (size_t c = 0; c < channels; ++c) { + for (size_t pd = 0; pd < outputD; ++pd) { + for (size_t ph = 0; ph < outputH; ++ph) { + for (size_t pw = 0; pw < outputW; ++pw) { + int dstart = pd * strideD - paddingD; + int hstart = ph * strideH - paddingH; + int wstart = pw * strideW - paddingW; + int dend = std::min(dstart + sizeZ, inDepth + paddingD); + int hend = std::min(hstart + sizeY, inHeight + paddingH); + int wend = std::min(wstart + sizeX, inWidth + paddingW); + int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart); + dstart = std::max(dstart, 0); + hstart = std::max(hstart, 0); + wstart = std::max(wstart, 0); + dend = std::min(dend, static_cast(inDepth)); + hend = std::min(hend, static_cast(inHeight)); + wend = std::min(wend, static_cast(inWidth)); + + CHECK(poolSize); + tgtData[(pd * outputH + ph) * outputW + pw] = 0; // clear + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + tgtData[(pd * outputH + ph) * outputW + pw] += + inData[(d * inHeight + h) * inWidth + w]; + } + } + } + tgtData[(pd * outputH + ph) * outputW + pw] /= poolSize; + } + } + } + // compute offset + inData += inDepth * inHeight * inWidth; + tgtData += outputD * outputH * outputW; + } + } +} + +void CpuMatrix::avgPool3DBackward(Matrix& input, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput) { + size_t num = input.getHeight(); + size_t channels = input.getWidth() / outputD / outputH / outputW; + CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth()); + real* inData = input.getData(); + real* outData = getData(); + + for (size_t n = 0; n < num; ++n) { + if (!input.isContiguous()) { + inData = input.getData() + n * input.getStride(); + } + for (size_t c = 0; c < channels; ++c) { + for (size_t pd = 0; pd < outputD; ++pd) { + for (size_t ph = 0; ph < outputH; ++ph) { + for (size_t pw = 0; pw < outputW; ++pw) { + int dstart = pd * strideD - paddingD; + int hstart = ph * strideH - paddingH; + int wstart = pw * strideW - paddingW; + int dend = std::min(dstart + sizeZ, imgSizeD + paddingD); + int hend = std::min(hstart + sizeY, imgSizeH + paddingH); + int wend = std::min(wstart + sizeX, imgSizeW + paddingW); + int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart); + dstart = std::max(dstart, 0); + hstart = std::max(hstart, 0); + wstart = std::max(wstart, 0); + dend = std::min(dend, static_cast(imgSizeD)); + hend = std::min(hend, static_cast(imgSizeH)); + wend = std::min(wend, static_cast(imgSizeW)); + CHECK(poolSize); + for (int d = dstart; d < dend; ++d) { + for (int h = hstart; h < hend; ++h) { + for (int w = wstart; w < wend; ++w) { + outData[(d * imgSizeH + h) * imgSizeW + w] += + inData[(pd * outputH + ph) * outputW + pw] / poolSize; + } + } + } + } + } + } + // offset + outData += imgSizeD * imgSizeH * imgSizeW; + inData += outputD * outputH * outputW; + } + } +} + /** * Input: one or more sequences. Each sequence contains some instances. * Output: output size is the number of input sequences (NOT input instances). @@ -2222,24 +2774,24 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) { CHECK(!isTransposed()) << "Not supported"; size_t a_col, b_col, a_row, b_row; - CBLAS_TRANSPOSE a_trans, b_trans; + bool a_trans, b_trans; if (!a->isTransposed()) { a_col = a->getWidth(); a_row = a->getHeight(); - a_trans = CblasNoTrans; + a_trans = false; } else { a_col = a->getHeight(); a_row = a->getWidth(); - a_trans = CblasTrans; + a_trans = true; } if (!b->isTransposed()) { b_col = b->getWidth(); b_row = b->getHeight(); - b_trans = CblasNoTrans; + b_trans = false; } else { b_col = b->getHeight(); b_row = b->getWidth(); - b_trans = CblasTrans; + b_trans = true; } CHECK_EQ(a_col, b_row); @@ -2256,7 +2808,7 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) { int lda = a->getStride(); int ldb = b->getStride(); int ldc = getStride(); - gemm( + BlasGemm::compute( a_trans, b_trans, M, N, K, scaleAB, A, lda, B, ldb, scaleT, C, ldc); } @@ -3975,6 +4527,95 @@ void CpuMatrix::bilinearBackward(const Matrix& out, } } +void CpuMatrix::vol2Col(real* data, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW) { + real* outData = getData(); + int outHeight = (height + 2 * paddingH - filterH) / strideH + 1; + int outWidth = (width + 2 * paddingW - filterW) / strideW + 1; + int outDepth = (depth + 2 * paddingD - filterD) / strideD + 1; + + int channelsCol = channels * filterD * filterH * filterW; + for (int c = 0; c < channelsCol; ++c) { + int wOffset = c % filterW; + int hOffset = (c / filterW) % filterH; + int dOffset = (c / filterW / filterH) % filterD; + int cIn = c / filterW / filterH / filterD; + for (int d = 0; d < outDepth; ++d) { + for (int h = 0; h < outHeight; ++h) { + for (int w = 0; w < outWidth; ++w) { + int dPad = d * strideD - paddingD + dOffset; + int hPad = h * strideH - paddingH + hOffset; + int wPad = w * strideW - paddingW + wOffset; + + if (hPad >= 0 && hPad < height && wPad >= 0 && wPad < width && + dPad >= 0 && dPad < depth) + outData[((c * outDepth + d) * outHeight + h) * outWidth + w] = + data[((cIn * depth + dPad) * height + hPad) * width + wPad]; + else + outData[((c * outDepth + d) * outHeight + h) * outWidth + w] = 0; + } + } + } + } +} + +void CpuMatrix::col2Vol(real* trg, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real alpha, + real beta) { + real* src = getData(); + int outDepth = (depth + 2 * paddingD - filterD) / strideD + 1; + int outHeight = (height + 2 * paddingH - filterH) / strideH + 1; + int outWidth = (width + 2 * paddingW - filterW) / strideW + 1; + int channelsCol = channels * filterD * filterH * filterW; + for (int c = 0; c < channelsCol; ++c) { + int wOffset = c % filterW; + int hOffset = (c / filterW) % filterH; + int dOffset = (c / filterW / filterH) % filterD; + int cIm = c / filterW / filterH / filterD; + for (int d = 0; d < outDepth; ++d) { + for (int h = 0; h < outHeight; ++h) { + for (int w = 0; w < outWidth; ++w) { + int dPad = d * strideD - paddingD + dOffset; + int hPad = h * strideH - paddingH + hOffset; + int wPad = w * strideW - paddingW + wOffset; + if (hPad >= 0 && hPad < height && wPad >= 0 && wPad < width && + dPad >= 0 && dPad < depth) + trg[((cIm * depth + dPad) * height + hPad) * width + wPad] = + alpha * + src[((c * outDepth + d) * outHeight + h) * outWidth + w] + + beta * + trg[((cIm * depth + dPad) * height + hPad) * width + wPad]; + } + } + } + } +} + //////////////////////////////////////////////////////////////// // functions executed via cpu // //////////////////////////////////////////////////////////////// diff --git a/paddle/math/Matrix.h b/paddle/math/Matrix.h index bb802bbb2c75289a45d987b22ad41ce8b1e95c98..44180bca8bca53e74d71ce7bed3516399c01c81d 100644 --- a/paddle/math/Matrix.h +++ b/paddle/math/Matrix.h @@ -928,15 +928,102 @@ public: size_t paddingW) { LOG(FATAL) << "Not implemeted"; } - /** - * Input: one or more sequences. Each sequence contains some instances. - * - * Output: output size is the number of input sequences (NOT input - * instances). - * - * output[i] is set to max_input[i]. + * Pooling 3D forward operation, pick out the largest element + * in the sizeX of value */ + virtual void maxPool3DForward(Matrix& inputMat, + Matrix& maxPoolIdx, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW) { + LOG(FATAL) << "Not implemeted"; + } + + virtual void maxPool3DBackward(Matrix& outGrad, + Matrix& maxPoolIdx, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput) { + LOG(FATAL) << "Not implemeted"; + } + + virtual void avgPool3DForward(Matrix& input, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW) { + LOG(FATAL) << "Not implemeted"; + } + + virtual void avgPool3DBackward(Matrix& input, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput) { + LOG(FATAL) << "Not implemeted"; + } + + /** + * Input: one or more sequences. Each sequence contains some instances. + * + * Output: output size is the number of input sequences (NOT input + * instances). + * + * output[i] is set to max_input[i]. + */ virtual void maxSequenceForward(Matrix& input, const IVector& sequence, IVector& index) { @@ -1039,6 +1126,42 @@ public: LOG(FATAL) << "Not implemented"; } + virtual void vol2Col(real* data, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW) { + LOG(FATAL) << "Not implemeted"; + } + + virtual void col2Vol(real* trg, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real alpha, + real beta) { + LOG(FATAL) << "Not implemeted"; + } + virtual void bilinearForward(const Matrix& in, const size_t inImgH, const size_t inImgW, @@ -1348,6 +1471,82 @@ public: size_t paddingH, size_t paddingW); + void maxPool3DForward(Matrix& inputMat, + Matrix& maxPoolIdx, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW); + + void maxPool3DBackward(Matrix& outGrad, + Matrix& maxPoolIdx, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput); + + void avgPool3DForward(Matrix& input, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW); + + void avgPool3DBackward(Matrix& input, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput); + void maxSequenceForward(Matrix& input, const IVector& sequence, IVector& index); @@ -1374,6 +1573,38 @@ public: const real ratioH, const real ratioW); + void vol2Col(real* data, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW); + + void col2Vol(real* trg, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real alpha, + real beta); + void multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label); void multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label); @@ -1385,6 +1616,10 @@ public: }; class CpuMatrix : public Matrix { +private: + MatrixPtr sftmaxSum_; + MatrixPtr sftmaxDot_; + public: CpuMatrix(size_t height, size_t width, bool trans = false); CpuMatrix(real* data, size_t height, size_t width, bool trans = false) @@ -1507,6 +1742,82 @@ public: size_t paddingH, size_t paddingW); + void maxPool3DForward(Matrix& inputMat, + Matrix& maxPoolIdx, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW); + + void maxPool3DBackward(Matrix& outGrad, + Matrix& maxPoolIdx, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput); + + void avgPool3DForward(Matrix& input, + size_t channels, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW); + + void avgPool3DBackward(Matrix& input, + size_t imgSizeD, + size_t imgSizeH, + size_t imgSizeW, + size_t outputD, + size_t outputH, + size_t outputW, + size_t sizeZ, + size_t sizeY, + size_t sizeX, + size_t strideD, + size_t strideH, + size_t strideW, + size_t paddingD, + size_t paddingH, + size_t paddingW, + real scaleTargets, + real scaleOutput); + void maxSequenceForward(Matrix& input, const IVector& sequence, IVector& index); @@ -1715,6 +2026,38 @@ public: const real ratioH, const real ratioW); + void vol2Col(real* data, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW); + + void col2Vol(real* trg, + int channels, + int depth, + int height, + int width, + int filterD, + int filterH, + int filterW, + int strideD, + int strideH, + int strideW, + int paddingD, + int paddingH, + int paddingW, + real alpha, + real beta); + template void operator=(const ExpressionType& expr) { TensorCpuApply(*this, expr); diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp index d77478f345df97b37b214b5978f51ce47c1d791c..103f06acc57d7a23f019f5e713f6cacf2179e9e0 100644 --- a/paddle/math/tests/test_matrixCompare.cpp +++ b/paddle/math/tests/test_matrixCompare.cpp @@ -18,6 +18,7 @@ limitations under the License. */ #include #include "TensorCheck.h" +#include "paddle/math/MathUtils.h" #include "paddle/math/Matrix.h" #include "paddle/math/SparseMatrix.h" #include "paddle/testing/TestUtil.h" @@ -1203,4 +1204,497 @@ TEST(Matrix, warpCTC) { } } +void testMaxPool3DFwdBwd(int numSamples, + int channels, + int imgSizeD, + int imgSizeH, + int imgSizeW, + int ksizeD, + int ksizeH, + int ksizeW, + int strideD, + int strideH, + int strideW, + int padD, + int padH, + int padW) { + int outD = outputSize(imgSizeD, ksizeD, padD, strideD, true); + int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true); + int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true); + + int inWidth = channels * imgSizeD * imgSizeH * imgSizeW; + MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); + + int outWidth = channels * outD * outH * outW; + MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); + MatrixPtr maxIdx = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr maxIdxGpu = GpuMatrix::create(numSamples, outWidth, false, true); + + input->randomizeUniform(); + target->randomizeUniform(); + inputGpu->copyFrom(*input); + targetGpu->copyFrom(*target); + + target->maxPool3DForward(*input, + *maxIdx, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW); + targetGpu->maxPool3DForward(*inputGpu, + *maxIdxGpu, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW); + MatrixPtr targetCheck = CpuMatrix::create(numSamples, outWidth, false, false); + targetCheck->copyFrom(*targetGpu); + checkMatrixEqual(target, targetCheck); + + MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); + MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpuGrad = + GpuMatrix::create(numSamples, outWidth, false, true); + + inputGrad->randomizeUniform(); + targetGrad->randomizeUniform(); + inputGpuGrad->copyFrom(*inputGrad); + targetGpuGrad->copyFrom(*targetGrad); + + inputGrad->maxPool3DBackward(*targetGrad, + *maxIdx, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW, + 1.0, + 1.0); + inputGpuGrad->maxPool3DBackward(*targetGpuGrad, + *maxIdxGpu, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW, + 1.0, + 1.0); + MatrixPtr targetBwdCheck = + CpuMatrix::create(numSamples, inWidth, false, false); + targetBwdCheck->copyFrom(*inputGpuGrad); + checkMatrixEqual(inputGrad, targetBwdCheck); +} + +void testAvgPool3DFwdBwd(int numSamples, + int channels, + int imgSizeD, + int imgSizeH, + int imgSizeW, + int ksizeD, + int ksizeH, + int ksizeW, + int strideD, + int strideH, + int strideW, + int padD, + int padH, + int padW) { + int outD = outputSize(imgSizeD, ksizeD, padD, strideD, true); + int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true); + int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true); + + int inWidth = imgSizeD * imgSizeH * imgSizeW * channels; + MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true); + + int outWidth = channels * outD * outH * outW; + MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true); + + input->randomizeUniform(); + target->randomizeUniform(); + inputGpu->copyFrom(*input); + targetGpu->copyFrom(*target); + + target->avgPool3DForward(*input, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW); + + targetGpu->avgPool3DForward(*inputGpu, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW); + + TensorCheckErr(*target, *targetGpu); + + MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false); + MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true); + MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false); + MatrixPtr targetGpuGrad = + GpuMatrix::create(numSamples, outWidth, false, true); + + inputGrad->randomizeUniform(); + targetGrad->randomizeUniform(); + inputGpuGrad->copyFrom(*inputGrad); + targetGpuGrad->copyFrom(*targetGrad); + + inputGrad->avgPool3DBackward(*targetGrad, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW, + 1.0, + 1.0); + + inputGpuGrad->avgPool3DBackward(*targetGpuGrad, + imgSizeD, + imgSizeH, + imgSizeW, + outD, + outH, + outW, + ksizeD, + ksizeH, + ksizeW, + strideD, + strideH, + strideW, + padD, + padH, + padW, + 1.0, + 1.0); + TensorCheckErr(*inputGrad, *inputGpuGrad); +} + +// TODO(yi): I noticed many such blindly combinatorial tests in this +// file. They are no help to locate defects at all. +TEST(Matrix, Pool3DFwdBwd) { + for (auto numSamples : {1, 3}) { + for (auto channels : {3}) { + for (auto imgSizeD : {9, 16}) { + for (auto imgSizeH : {9, 32}) { + for (auto imgSizeW : {9, 32}) { + for (auto sizeX : {3}) { + for (auto sizeY : {3}) { + for (auto sizeZ : {3}) { + for (auto sD : {2}) { + for (auto sH : {2}) { + for (auto sW : {2}) { + for (auto pD : {0, (sizeZ - 1) / 2}) { + for (auto pH : {0, (sizeY - 1) / 2}) { + for (auto pW : {0, (sizeX - 1) / 2}) { + VLOG(3) << " numSamples=" << numSamples + << " channels=" << channels + << " imgSizeD=" << imgSizeD + << " imgSizeH=" << imgSizeH + << " imgSizeW=" << imgSizeW + << " sizeX=" << sizeX + << " sizeY=" << sizeY + << " sizeZ=" << sizeZ << " strideD=" << sD + << " strideH=" << sH << " strideW=" << sW + << " padingD=" << pD << " padingH=" << pH + << " padingW=" << pW; + + testMaxPool3DFwdBwd(numSamples, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + sizeX, + sizeY, + sizeZ, + sD, + sH, + sW, + pD, + pH, + pW); + testAvgPool3DFwdBwd(numSamples, + channels, + imgSizeD, + imgSizeH, + imgSizeW, + sizeX, + sizeY, + sizeZ, + sD, + sH, + sW, + pD, + pH, + pW); + } + } + } + } + } + } + } + } + } + } + } + } + } + } + + // for (auto numSamples : {1, 3}) { + // for (auto channels : {1, 3}) { + // for (auto imgSizeD : {9,16}) { + // for (auto imgSizeH : {9, 32}) { + // for (auto imgSizeW : {9, 32}) { + // for (auto sizeX : {2, 3}) { + // for (auto sizeY : {2, 3}) { + // for (auto sizeZ : {2,3}){ + // for (auto sD : {1, 2}) { + // for (auto sH : {1, 2}) { + // for (auto sW : {1, 2}) { + // for (auto pD : {0, (sizeZ - 1) / 2}){ + // for (auto pH : {0, (sizeY - 1) / 2}) { + // for (auto pW : {0, (sizeX - 1) / 2}) { + // VLOG(3) << " numSamples=" << numSamples + // << " channels=" << channels + // << " imgSizeD=" << imgSizeD + // << " imgSizeH=" << imgSizeH + // << " imgSizeW=" << imgSizeW + // << " sizeX=" << sizeX + // << " sizeY=" << sizeY + // << " sizeZ=" << sizeZ + // << " strideD=" << sD + // << " strideH=" << sH + // << " strideW=" << sW + // << " padingD=" << pD + // << " padingH=" << pH + // << " padingW=" << pW; + // + // testMaxPool3DFwdBwd(numSamples, + // channels, + // imgSizeD, + // imgSizeH, + // imgSizeW, + // sizeX, + // sizeY, + // sizeZ, + // sD, + // sH, + // sW, + // pD, + // pH, + // pW); + // testAvgPool3DFwdBwd(numSamples, + // channels, + // imgSizeD, + // imgSizeH, + // imgSizeW, + // sizeX, + // sizeY, + // sizeZ, + // sD, + // sH, + // sW, + // pD, + // pH, + // pW); + // } + // } + // } + // } + // } + // } + // } + // } + // } + // } + // } + // } + // } + // } +} + +void testMatrixCol2Vol(int depth, int height, int width) { + int channel = 3; + int filterX = 3, filterY = 4, filterZ = 5; + int strideX = 2, strideY = 2, strideZ = 2; + int padX = 1, padY = 1, padZ = 1; + + MatrixPtr cpuImage = + std::make_shared(channel, depth * height * width); + MatrixPtr gpuImage = + std::make_shared(channel, depth * height * width); + cpuImage->randomizeUniform(); + gpuImage->copyFrom(*cpuImage); + + int outD = outputSize(depth, filterZ, padZ, strideZ, true); + int outH = outputSize(height, filterY, padY, strideY, true); + int outW = outputSize(width, filterX, padX, strideX, true); + + int colBufHeight = channel * filterZ * filterY * filterX; + int colBufWidth = outD * outH * outW; + MatrixPtr cpuColBuf = std::make_shared(colBufHeight, colBufWidth); + MatrixPtr gpuColBuf = std::make_shared(colBufHeight, colBufWidth); + cpuColBuf->vol2Col(cpuImage->getData(), + channel, + depth, + height, + width, + filterZ, + filterY, + filterX, + strideZ, + strideY, + strideX, + padZ, + padY, + padX); + gpuColBuf->vol2Col(gpuImage->getData(), + channel, + depth, + height, + width, + filterZ, + filterY, + filterX, + strideZ, + strideY, + strideX, + padZ, + padY, + padX); + TensorCheckEqual(*cpuColBuf, *gpuColBuf); + + cpuColBuf->randomizeUniform(); + gpuColBuf->copyFrom(*cpuColBuf); + cpuColBuf->col2Vol(cpuImage->getData(), + channel, + depth, + height, + width, + filterZ, + filterY, + filterX, + strideZ, + strideY, + strideX, + padZ, + padY, + padX, + 1.0, + 1.0); + gpuColBuf->col2Vol(gpuImage->getData(), + channel, + depth, + height, + width, + filterZ, + filterY, + filterX, + strideZ, + strideY, + strideX, + padZ, + padY, + padX, + 1.0, + 1.0); + TensorCheckErr(*cpuImage, *gpuImage); +} + +TEST(Matrix, col2Vol) { + for (auto depth : {9, 16, 64}) { + for (auto height : {9, 11, 128}) { + for (auto width : {9, 32, 128}) { + VLOG(3) << "depth=" << depth << " height=" << height + << " width=" << width; + testMatrixCol2Vol(depth, height, width); + } + } + } +} + #endif diff --git a/paddle/memory/CMakeLists.txt b/paddle/memory/CMakeLists.txt index 8035d93bfec75b20a54c5af0521ab724cafba8ca..9cc4233e43267472d405c3e4e617f0782e1430ea 100644 --- a/paddle/memory/CMakeLists.txt +++ b/paddle/memory/CMakeLists.txt @@ -1,7 +1,7 @@ add_subdirectory(detail) cc_library(memory SRCS memory.cc) -cc_library(memcpy SRCS memcpy.cc DEPS device_context) +cc_library(memcpy SRCS memcpy.cc) cc_library(paddle_memory DEPS diff --git a/paddle/memory/detail/system_allocator.cc b/paddle/memory/detail/system_allocator.cc index f61e67a32906083881dd7f47433521876be9b355..a270bd59581520859d43cddd2fc0cfa72080f46d 100644 --- a/paddle/memory/detail/system_allocator.cc +++ b/paddle/memory/detail/system_allocator.cc @@ -27,7 +27,7 @@ limitations under the License. */ // between host and device. Allocates too much would reduce the amount // of memory available to the system for paging. So, by default, we // should set false to use_pinned_memory. -DEFINE_bool(use_pinned_memory, false, "If set, allocate cpu pinned memory."); +DEFINE_bool(use_pinned_memory, true, "If set, allocate cpu pinned memory."); namespace paddle { namespace memory { diff --git a/paddle/memory/memcpy.cc b/paddle/memory/memcpy.cc index aaab1142ca18d3319469a4d685fde9d30929113f..a19a3e3675e3e2e7cc0c3594f21191f932d6379f 100644 --- a/paddle/memory/memcpy.cc +++ b/paddle/memory/memcpy.cc @@ -16,8 +16,6 @@ limitations under the License. */ #include // for memcpy -#include "paddle/platform/device_context.h" - namespace paddle { namespace memory { diff --git a/paddle/memory/memory.cc b/paddle/memory/memory.cc index 207025f9b1c64f0f8943f9fae5edefc9328a1d26..29bc26f9d3bca0e30896657431f9a9bb1dac0d1d 100644 --- a/paddle/memory/memory.cc +++ b/paddle/memory/memory.cc @@ -13,22 +13,38 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/memory/memory.h" + +#include // for transform +#include // for memcpy +#include // for unique_ptr +#include // for call_once + +#include "glog/logging.h" + #include "paddle/memory/detail/buddy_allocator.h" #include "paddle/memory/detail/system_allocator.h" +#include "paddle/platform/gpu_info.h" -#include // for memcpy +DECLARE_double(fraction_of_gpu_memory_to_use); namespace paddle { namespace memory { -detail::BuddyAllocator* GetCPUBuddyAllocator() { - static detail::BuddyAllocator* a = nullptr; - if (a == nullptr) { - a = new detail::BuddyAllocator(new detail::CPUAllocator, - platform::CpuMinChunkSize(), - platform::CpuMaxChunkSize()); - } - return a; +using BuddyAllocator = detail::BuddyAllocator; + +std::once_flag cpu_allocator_flag; +std::once_flag gpu_allocator_flag; + +BuddyAllocator* GetCPUBuddyAllocator() { + static std::unique_ptr a{nullptr}; + + std::call_once(cpu_allocator_flag, [&]() { + a.reset(new BuddyAllocator(new detail::CPUAllocator, + platform::CpuMinChunkSize(), + platform::CpuMaxChunkSize())); + }); + + return a.get(); } template <> @@ -48,20 +64,36 @@ size_t Used(platform::CPUPlace place) { #ifndef PADDLE_ONLY_CPU -detail::BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { - static detail::BuddyAllocator** as = NULL; - if (as == NULL) { +BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { + using BuddyAllocVec = std::vector; + static std::unique_ptr as{ + new BuddyAllocVec, [](BuddyAllocVec* p) { + std::for_each(p->begin(), p->end(), + [](BuddyAllocator* p) { delete p; }); + }}; + + // GPU buddy allocators + auto& allocators = *as.get(); + + // GPU buddy allocator initialization + std::call_once(gpu_allocator_flag, [&]() { int gpu_num = platform::GetDeviceCount(); - as = new detail::BuddyAllocator*[gpu_num]; + allocators.reserve(gpu_num); for (int gpu = 0; gpu < gpu_num; gpu++) { platform::SetDeviceId(gpu); - as[gpu] = new detail::BuddyAllocator(new detail::GPUAllocator, - platform::GpuMinChunkSize(), - platform::GpuMaxChunkSize()); + allocators.emplace_back(new BuddyAllocator(new detail::GPUAllocator, + platform::GpuMinChunkSize(), + platform::GpuMaxChunkSize())); } - } + VLOG(3) << "\n\nNOTE: each GPU device use " + << FLAGS_fraction_of_gpu_memory_to_use * 100 << "% of GPU memory.\n" + << "You can set environment variable '" + << platform::kEnvFractionGpuMemoryToUse + << "' to change the fraction of GPU usage.\n\n"; + }); + platform::SetDeviceId(gpu_id); - return as[gpu_id]; + return allocators[gpu_id]; } template <> diff --git a/paddle/memory/memory.h b/paddle/memory/memory.h index 72351b9dfa63513713463bb47a3684f0dfd84ad3..11bbb881874ec50e1132547336fc6fb6b42bcc4f 100644 --- a/paddle/memory/memory.h +++ b/paddle/memory/memory.h @@ -14,7 +14,6 @@ limitations under the License. */ #pragma once -#include "paddle/platform/gpu_info.h" #include "paddle/platform/place.h" namespace paddle { diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt index c181bd7b881c08dfd80d640b1ddce10b3c74d758..f9ea25ab045a02be5ab9ed81ef9c679126d3a188 100644 --- a/paddle/operators/CMakeLists.txt +++ b/paddle/operators/CMakeLists.txt @@ -1,7 +1,10 @@ +file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc") +string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}") function(op_library TARGET) # op_library is a function to create op library. The interface is same as # cc_library. But it handle split GPU/CPU code and link some common library # for ops. + set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE) set(cc_srcs) set(cu_srcs) set(op_common_deps operator op_registry) @@ -11,27 +14,31 @@ function(op_library TARGET) cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - foreach(src ${op_library_SRCS}) - if (${src} MATCHES ".*\\.cu$") - list(APPEND cu_srcs ${src}) - elseif(${src} MATCHES ".*\\.cc$") - list(APPEND cc_srcs ${src}) - else() - message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu") + list(LENGTH op_library_SRCS op_library_SRCS_len) + if (${op_library_SRCS_len} EQUAL 0) + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc) + list(APPEND cc_srcs ${TARGET}.cc) endif() - endforeach() + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu) + list(APPEND cu_srcs ${TARGET}.cu) + endif() + else() + foreach(src ${op_library_SRCS}) + if (${src} MATCHES ".*\\.cu$") + list(APPEND cu_srcs ${src}) + elseif(${src} MATCHES ".*\\.cc$") + list(APPEND cc_srcs ${src}) + else() + message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu") + endif() + endforeach() + endif() list(LENGTH cc_srcs cc_srcs_len) if (${cc_srcs_len} EQUAL 0) message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file") endif() - list(LENGTH cu_srcs cu_srcs_len) - list(LENGTH op_library_DEPS dep_len) - if (${cu_srcs_len} EQUAL 0 AND ${dep_len} EQUAL 0) - message(WARNING "The op library ${TARGET} not support GPU!") - endif() - if (WITH_GPU) nv_library(${TARGET} SRCS ${cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS} ${op_common_deps}) @@ -41,28 +48,28 @@ function(op_library TARGET) endif() endfunction() -cc_test(gather_test SRCS gather_test.cc DEPS tensor) - -cc_library(net_op SRCS net_op.cc DEPS op_registry) -cc_test(net_op_test SRCS net_op_test.cc DEPS net_op) - -op_library(add_op SRCS add_op.cc add_op.cu) +add_subdirectory(math) -op_library(mean_op SRCS mean_op.cc mean_op.cu) - -op_library(mul_op SRCS mul_op.cc mul_op.cu) -op_library(rowwise_add_op SRCS rowwise_add_op.cu rowwise_add_op.cc) +set(DEPS_OPS + identity_op + minus_op + mul_op + recurrent_op + scale_op) +op_library(identity_op DEPS scale_op) +op_library(minus_op DEPS scale_op) +op_library(mul_op DEPS math_function) +op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc + DEPS framework_proto tensor operator net_op) +op_library(scale_op DEPS net_op) -op_library(sigmoid_op SRCS sigmoid_op.cc sigmoid_op.cu) -op_library(softmax_op SRCS softmax_op.cc softmax_op.cu) -op_library(gaussian_random_op SRCS gaussian_random_op.cc gaussian_random_op.cu) -op_library(cross_entropy_op SRCS cross_entropy_op.cc cross_entropy_op.cu) -op_library(fill_zeros_like_op SRCS fill_zeros_like_op.cc fill_zeros_like_op.cu) +list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS}) +foreach(src ${GENERAL_OPS}) + op_library(${src}) +endforeach() -op_library(sgd_op SRCS sgd_op.cc sgd_op.cu) +set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library") -op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc - DEPS op_desc tensor op_registry operator net_op) -cc_test(recurrent_op_test SRCS recurrent_op_test.cc DEPS recurrent_op gtest mul_op add_op) -op_library(uniform_random_op - SRCS uniform_random_op.cc uniform_random_op.cu) +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) diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc index b886ded9bbd97dc1942c87d7603521e8d72e3f6c..8dbd47cf0dfbc265032a9966343eed5c7bd8692e 100644 --- a/paddle/operators/add_op.cc +++ b/paddle/operators/add_op.cc @@ -18,17 +18,15 @@ namespace paddle { namespace operators { class AddOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(AddOp, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 2); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "Inputs of AddOp must all be set"); - PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr, - "Outputs of AddOp must all be set"); - PADDLE_ENFORCE(ctx.Input(0)->dims() == ctx.Input(1)->dims(), - "Two input of Add Op's dimension must be same."); - ctx.Output(0)->Resize(ctx.Input(0)->dims()); + 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()); } }; @@ -48,7 +46,9 @@ The equation is: Out = X + Y }; class AddOpGrad : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(AddOpGrad, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override {} }; @@ -57,8 +57,6 @@ class AddOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker); -REGISTER_GRADIENT_OP(add_two, add_two_grad, ops::AddOpGrad); +REGISTER_OP(add, ops::AddOp, ops::AddOpMaker, add_grad, ops::AddOpGrad); -REGISTER_OP_CPU_KERNEL(add_two, - ops::AddKernel); +REGISTER_OP_CPU_KERNEL(add, ops::AddKernel); diff --git a/paddle/operators/add_op.cu b/paddle/operators/add_op.cu index cec5f558cbc161124620ad4241d6bd8a5324277c..d9c6d20a6c320b59e57ed25da3dd8b093833f8c7 100644 --- a/paddle/operators/add_op.cu +++ b/paddle/operators/add_op.cu @@ -12,10 +12,7 @@ See the License for the specific language governing permissions and limitations under the License. */ -#define EIGEN_USE_GPU -#include "paddle/framework/op_registry.h" #include "paddle/operators/add_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(add_two, - ops::AddKernel); +REGISTER_OP_GPU_KERNEL(add, ops::AddKernel); diff --git a/paddle/operators/add_op.h b/paddle/operators/add_op.h index d76c10957e943deb970b1d79a1507a36669314e3..a7307b6818aa3d10ff215d06281e2b53196fd101 100644 --- a/paddle/operators/add_op.h +++ b/paddle/operators/add_op.h @@ -28,9 +28,9 @@ template class AddKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto input0 = context.Input(0); - auto input1 = context.Input(1); - auto output = context.Output(0); + auto* input0 = context.Input("X"); + auto* input1 = context.Input("Y"); + auto* output = context.Output("Out"); output->mutable_data(context.GetPlace()); diff --git a/paddle/operators/concat_op.cc b/paddle/operators/concat_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..0ebefbab26ec8fdf316f852fbb7f6d9f3bbc48eb --- /dev/null +++ b/paddle/operators/concat_op.cc @@ -0,0 +1,79 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/concat_op.h" +#include + +namespace paddle { +namespace operators { +using framework::Tensor; + +class ConcatOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + auto ins = ctx.MultiInput("X"); + auto *out = ctx.Output("Out"); + size_t axis = static_cast(ctx.Attr("axis")); + size_t n = ins.size(); + + PADDLE_ENFORCE_GT(n, 1, "Input tensors count should > 1."); + + auto out_dims = ins[0]->dims(); + 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]; + continue; + } + PADDLE_ENFORCE_EQ(out_dims[j], ins[i]->dims()[j], + "Input tensors should have the same " + "elements except the specify axis.") + } + } + out->Resize(out_dims); + } +}; + +class ConcatOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ConcatOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "the input tensors of concat operator.").AsDuplicable(); + AddOutput("Out", "the output tensor of concat operator."); + AddComment(R"DOC( + Join the input tensors along with the axis. + Examples: + Input[0] = [[1,2],[3,4]] + Input[1] = [[5,6]] + axis = 0 + Output = [[1,2], + [3,4], + [5,6]] + )DOC"); + AddAttr("axis", "The axis which the inputs will be joined with.") + .SetDefault(0); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(concat, ops::ConcatOp, ops::ConcatOpMaker) +REGISTER_OP_CPU_KERNEL(concat, + ops::ConcatKernel) diff --git a/paddle/operators/concat_op.cu b/paddle/operators/concat_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..38fee7473dbb2ba97fe95b6632db7a1749cf3bbe --- /dev/null +++ b/paddle/operators/concat_op.cu @@ -0,0 +1,19 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/concat_op.h" + +namespace ops = paddle::operators; +// TODO(Yancey1989) Add GPU kernel diff --git a/paddle/operators/concat_op.h b/paddle/operators/concat_op.h new file mode 100644 index 0000000000000000000000000000000000000000..f977054fdf8aa0164db726b94a21c57f770dd674 --- /dev/null +++ b/paddle/operators/concat_op.h @@ -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. */ + +#pragma once + +#include +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +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]; + } + } + size_t output_offset = 0; + 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; + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..c033af3b741ae26ad9d37b2164f87aa6e8651c6e --- /dev/null +++ b/paddle/operators/cos_sim_op.cc @@ -0,0 +1,107 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/cos_sim_op.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; + +class CosSimOp : 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("Y"), "Input(Y) must not be null."); + PADDLE_ENFORCE_EQ(ctx.Input("X")->dims(), + ctx.Input("Y")->dims(), + "Dimensions of Input(X) and Input(Y) must be the same."); + + auto dims = ctx.Input("X")->dims(); + ctx.Output("Out")->Resize({dims[0], 1}); + ctx.Output("XNorm")->Resize({dims[0], 1}); + ctx.Output("YNorm")->Resize({dims[0], 1}); + } +}; + +class CosSimOpMaker : public framework::OpProtoAndCheckerMaker { + public: + CosSimOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The first input of cos_sim op."); + AddInput("Y", "The second input of cos_sim op."); + AddOutput("Out", "The output of cos_sim op."); + AddOutput("XNorm", "Row norm of the first input.").AsIntermediate(); + AddOutput("YNorm", "Row norm of the second input.").AsIntermediate(); + + AddComment(R"DOC( +Cosine Similarity Operator. + +The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y)) +)DOC"); + } +}; + +class CosSimOpGrad : 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("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(framework::GradVarName("Out")), + "Input(Out@GRAD) must not be null."); + + 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(framework::GradVarName("Out"))->dims(); + PADDLE_ENFORCE_EQ(x_dims, y_dims, + "Dimensions of Input(X) and Input(Y) must be the same."); + PADDLE_ENFORCE_EQ(xnorm_dims[0], x_dims[0], + "1st dimension of XNorm must equal that of Input(X)."); + PADDLE_ENFORCE_EQ(xnorm_dims[1], 1, "2st dimension of XNorm must be one."); + PADDLE_ENFORCE_EQ(ynorm_dims[0], y_dims[0], + "1st dimension of YNorm must equal that of Input(Y)."); + PADDLE_ENFORCE_EQ(ynorm_dims[1], 1, "2st dimension of YNorm must be one."); + PADDLE_ENFORCE_EQ(out_dims[0], x_dims[0], + "1st dimension of Out@GRAD must equal that of Input(X)"); + PADDLE_ENFORCE_EQ(out_dims[1], 1, "1st dimension of Out@GRAD must be one."); + + 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); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(cos_sim, ops::CosSimOp, ops::CosSimOpMaker, cos_sim_grad, + ops::CosSimOpGrad); +REGISTER_OP_CPU_KERNEL(cos_sim, + ops::CosSimKernel); +REGISTER_OP_CPU_KERNEL( + cos_sim_grad, ops::CosSimGradKernel); diff --git a/paddle/operators/cos_sim_op.cu b/paddle/operators/cos_sim_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..0cb8fd26de47a4a464db98664263544e3e503d63 --- /dev/null +++ b/paddle/operators/cos_sim_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. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/cos_sim_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(cos_sim, + ops::CosSimKernel); +REGISTER_OP_GPU_KERNEL( + cos_sim_grad, ops::CosSimGradKernel); diff --git a/paddle/operators/cos_sim_op.h b/paddle/operators/cos_sim_op.h new file mode 100644 index 0000000000000000000000000000000000000000..9e2bcebe3b5432c157fac895a9bbab5164193dbb --- /dev/null +++ b/paddle/operators/cos_sim_op.h @@ -0,0 +1,107 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT 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 EigenMatrix = framework::EigenMatrix; +template +using EigenVector = framework::EigenVector; + +template +class CosSimKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* input_x = context.Input("X"); + auto* input_y = context.Input("Y"); + auto* output_z = context.Output("Out"); + auto* output_x_norm = context.Output("XNorm"); + auto* output_y_norm = context.Output("YNorm"); + + output_z->mutable_data(context.GetPlace()); + output_x_norm->mutable_data(context.GetPlace()); + output_y_norm->mutable_data(context.GetPlace()); + + auto dims = input_x->dims(); + int size = static_cast(framework::product(dims)); + auto new_dims = framework::make_ddim({dims[0], size / dims[0]}); + auto x = EigenMatrix::From(*input_x, new_dims); + auto y = EigenMatrix::From(*input_y, new_dims); + auto z = EigenVector::Flatten(*output_z); + auto x_norm = EigenVector::Flatten(*output_x_norm); + auto y_norm = EigenVector::Flatten(*output_y_norm); + + auto place = context.GetEigenDevice(); + auto xy = (x * y).sum(Eigen::array({{1}})); + x_norm.device(place) = x.square().sum(Eigen::array({{1}})).sqrt(); + y_norm.device(place) = y.square().sum(Eigen::array({{1}})).sqrt(); + z.device(place) = xy / x_norm / y_norm; + } +}; + +template +class CosSimGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* input_x = context.Input("X"); + auto* input_y = context.Input("Y"); + auto* input_z = context.Input("Out"); + auto* input_x_norm = context.Input("XNorm"); + auto* input_y_norm = context.Input("YNorm"); + auto* output_grad_x = context.Output(framework::GradVarName("X")); + auto* output_grad_y = context.Output(framework::GradVarName("Y")); + auto* input_grad_z = context.Input(framework::GradVarName("Out")); + + auto dims = input_x->dims(); + int size = static_cast(framework::product(dims)); + auto new_dims = framework::make_ddim({dims[0], size / dims[0]}); + auto x = EigenMatrix::From(*input_x, new_dims); + auto y = EigenMatrix::From(*input_y, new_dims); + auto z = EigenMatrix::From(*input_z); + auto x_norm = EigenMatrix::From(*input_x_norm); + auto y_norm = EigenMatrix::From(*input_y_norm); + auto dz = EigenMatrix::From(*input_grad_z); + + Eigen::DSizes bcast(1, new_dims[1]); + auto z_bcast = z.broadcast(bcast); + auto dz_bcast = dz.broadcast(bcast); + auto place = context.GetEigenDevice(); + auto x_snorm_bcast = x_norm.square().eval().broadcast(bcast); + auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast); + auto norm_prod_bcast = (x_norm * y_norm).eval().broadcast(bcast); + if (output_grad_x) { + output_grad_x->mutable_data(context.GetPlace()); + auto dx = EigenMatrix::From(*output_grad_x, new_dims); + dx.device(place) = + dz_bcast * (y / norm_prod_bcast - z_bcast * x / x_snorm_bcast); + } + if (output_grad_y) { + output_grad_y->mutable_data(context.GetPlace()); + auto dy = EigenMatrix::From(*output_grad_y, new_dims); + dy.device(place) = + dz_bcast * (x / norm_prod_bcast - z_bcast * y / y_snorm_bcast); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/cross_entropy_op.cc b/paddle/operators/cross_entropy_op.cc index 09aa589d3caf7ed7b790850b515d49afdd3e1467..ab1e1c101a10e09a81f7785d2f1514822e3bdf15 100644 --- a/paddle/operators/cross_entropy_op.cc +++ b/paddle/operators/cross_entropy_op.cc @@ -18,36 +18,31 @@ namespace paddle { namespace operators { class OnehotCrossEntropyOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(OnehotCrossEntropyOp, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 2, - "Input size of OnehotCrossEntropyOp must be two"); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, - "Output size of OnehotCrossEntropyOp must be one"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), - "0-th input of OnehotCrossEntropyOp should be set"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(1), - "1-th input of OnehotCrossEntropyOp should be set"); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0), - "Outputs of OnehotCrossEntropyOp must all be set"); - PADDLE_ENFORCE_EQ(ctx.Input(0)->dims().size(), 2); - PADDLE_ENFORCE_EQ(ctx.Output(0)->dims().size(), 1, - "label's dimension must be 1."); - ctx.Output(0)->Resize({ctx.Input(0)->dims()[0]}); + auto *X = ctx.Input("X"); + auto *label = ctx.Input("label"); + + PADDLE_ENFORCE_EQ(X->dims().size(), 2, "X's dimension must be 2."); + PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label's dimension must be 1."); + PADDLE_ENFORCE_EQ(X->dims()[0], label->dims()[0]); + ctx.Output("Y")->Resize({X->dims()[0]}); } }; class OnehotCrossEntropyGradientOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(OnehotCrossEntropyGradientOp, - framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - auto X_grad = ctx.Output(framework::GradVarName("X")); + auto dX = ctx.Output(framework::GradVarName("X")); auto X = ctx.Input("X"); - // TODO(superjom) add enforce here after helper functions ready - X_grad->Resize(X->dims()); + dX->Resize(X->dims()); } }; @@ -72,12 +67,9 @@ OnehotCrossEntropy Operator. namespace ops = paddle::operators; REGISTER_OP(onehot_cross_entropy, ops::OnehotCrossEntropyOp, - ops::OnehotCrossEntropyOpMaker); -REGISTER_OP_CPU_KERNEL( - onehot_cross_entropy, - ops::OnehotCrossEntropyOpKernel); -REGISTER_GRADIENT_OP(onehot_cross_entropy, onehot_cross_entropy_grad, - ops::OnehotCrossEntropyGradientOp); -REGISTER_OP_CPU_KERNEL( - onehot_cross_entropy_grad, - ops::OnehotCrossEntropyGradientOpKernel); + ops::OnehotCrossEntropyOpMaker, onehot_cross_entropy_grad, + ops::OnehotCrossEntropyGradientOp); +REGISTER_OP_CPU_KERNEL(onehot_cross_entropy, + ops::OnehotCrossEntropyOpKernel); +REGISTER_OP_CPU_KERNEL(onehot_cross_entropy_grad, + ops::OnehotCrossEntropyGradientOpKernel); diff --git a/paddle/operators/cross_entropy_op.cu b/paddle/operators/cross_entropy_op.cu index 4bbc8f093a794d46737a16488684a6a0cc25e285..d999bfce58c8a6db5c811aad677c07094b881841 100644 --- a/paddle/operators/cross_entropy_op.cu +++ b/paddle/operators/cross_entropy_op.cu @@ -12,10 +12,122 @@ See the License for the specific language governing permissions and limitations under the License. */ -#define EIGEN_USE_GPU -#include "paddle/operators/cross_entropy_op.h" +#include "paddle/framework/op_registry.h" +#include "paddle/platform/assert.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +__host__ __device__ T clipping_log(const T x) { + PADDLE_ASSERT(std::is_floating_point::value); + const T kApproInf = 1e20; + T v = log(x); + if (v == INFINITY) { + return kApproInf; + } + if (v == -INFINITY) { + return -kApproInf; + } + return v; +} + +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] = -clipping_log(X[i * D + label[i]]); + } +} + +// 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; + } +} + +template +__global__ void CrossEntropyGradientKernel(T* dX, const T* dY, 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) { + int idx = i * D + label[i]; + dX[idx] = -dY[i] / X[idx]; + } +} + +template +class OnehotCrossEntropyOpCUDAKernel : 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"); + const T* Xdata = X->data(); + const int* label_data = ctx.Input("label")->data(); + auto Y = ctx.Output("Y"); + Y->mutable_data(ctx.GetPlace()); + T* Ydata = 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. + CrossEntropyKernel<<>>(Ydata, Xdata, label_data, N, D); + } +}; + +template +class OnehotCrossEntropyGradientOpCUDAKernel : 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 dX = ctx.Output(framework::GradVarName("X")); + auto dY = ctx.Input(framework::GradVarName("Y")); + auto label = ctx.Input("label"); + + auto* dXdata = dX->template mutable_data(ctx.GetPlace()); + auto* dYdata = dY->template data(); + auto* Xdata = X->template data(); + auto* label_data = label->data(); + + int N = X->dims()[0]; + int D = X->dims()[1]; + int block = 512; + int grid = (N * D + block - 1) / block; + zero<<>>(dXdata, N * D); + + grid = (N + block - 1) / block; + // TODO(qingqing): launch kernel on specified stream + // base on ExecutionContext. + CrossEntropyGradientKernel<<>>(dXdata, dYdata, Xdata, + label_data, N, D); + } +}; + +} // namespace operators +} // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - onehot_cross_entropy, - ops::OnehotCrossEntropyOpKernel); +REGISTER_OP_GPU_KERNEL(onehot_cross_entropy, + ops::OnehotCrossEntropyOpCUDAKernel); +REGISTER_OP_GPU_KERNEL(onehot_cross_entropy_grad, + ops::OnehotCrossEntropyGradientOpCUDAKernel); diff --git a/paddle/operators/cross_entropy_op.h b/paddle/operators/cross_entropy_op.h index d1bbc2cb66d6ce84ddcdcb87648f23c6ce77b748..eb4d1348de1d940e2648c83c8ba94b289f10c5b2 100644 --- a/paddle/operators/cross_entropy_op.h +++ b/paddle/operators/cross_entropy_op.h @@ -21,7 +21,7 @@ namespace operators { using Tensor = framework::Tensor; template -T tolerable_value(T x) { +inline T tolerable_value(const T x) { static_assert(std::is_floating_point::value, "tolerable_value works only on float, " "double and double double."); @@ -39,13 +39,16 @@ T tolerable_value(T x) { return x; } -template +template class OnehotCrossEntropyOpKernel : 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"); const T* Xdata = X->data(); - const int* label_data = ctx.Input(1)->data(); + const int* label_data = ctx.Input("label")->data(); auto Y = ctx.Output("Y"); Y->mutable_data(ctx.GetPlace()); @@ -62,10 +65,13 @@ class OnehotCrossEntropyOpKernel : public framework::OpKernel { } }; -template +template class OnehotCrossEntropyGradientOpKernel : 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")); @@ -79,6 +85,8 @@ class OnehotCrossEntropyGradientOpKernel : public framework::OpKernel { const int batch_size = X->dims()[0]; const int class_num = X->dims()[1]; + // TODO(qingqing): make zero setting an common function. + memset(dXdata, 0, sizeof(T) * batch_size * class_num); for (int i = 0; i < batch_size; ++i) { int index = i * class_num + label_data[i]; dXdata[index] = -tolerable_value(dYdata[i] / Xdata[index]); diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc index eda23a0ccfacd3a620412876e18f4ec47652bf9d..9d51f6e3a16fe96125599bb440d40237aeb9a028 100644 --- a/paddle/operators/fill_zeros_like_op.cc +++ b/paddle/operators/fill_zeros_like_op.cc @@ -18,19 +18,13 @@ namespace paddle { namespace operators { class FillZerosLikeOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(FillZerosLikeOp, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 1UL, - "Input size of FillZerosLikeOp must be one."); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL, - "Output size of AddOp must be one."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), - "Input of FillZerosLikeOp must be set."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0), - "Output of FillZerosLikeOp must be set."); - ctx.Output(0)->Resize( - ctx.Input(0)->dims()); + ctx.Output("Dst")->Resize( + ctx.Input("Src")->dims()); } }; @@ -52,7 +46,8 @@ The output will have the same size with input. } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(fill_zeros_like, ops::FillZerosLikeOp, ops::FillZerosLikeOpMaker); +REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, ops::FillZerosLikeOp, + ops::FillZerosLikeOpMaker); REGISTER_OP_CPU_KERNEL( fill_zeros_like, ops::FillZerosLikeKernel); diff --git a/paddle/operators/fill_zeros_like_op.h b/paddle/operators/fill_zeros_like_op.h index f846c7a8ab15e2cd997564edb36660a1360227a8..969998ce2eae02b8ad057c6259703e51559bf98a 100644 --- a/paddle/operators/fill_zeros_like_op.h +++ b/paddle/operators/fill_zeros_like_op.h @@ -23,10 +23,10 @@ template class FillZerosLikeKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* output = context.Output(0); + auto* output = context.Output("Dst"); output->mutable_data(context.GetPlace()); auto t = framework::EigenVector::Flatten(*output); - t.device(context.GetEigenDevice()) = t.constant(T(0)); + t.device(context.GetEigenDevice()) = t.constant(static_cast(0)); } }; diff --git a/paddle/operators/gather.h b/paddle/operators/gather.h index 0c73717d38aca9f3430e66cafc3ecccdd2eec776..92fb51ec17709bc6f8abb2f516a9240fb5dc3a77 100644 --- a/paddle/operators/gather.h +++ b/paddle/operators/gather.h @@ -17,6 +17,7 @@ limitations under the License. */ #include #include "paddle/framework/ddim.h" +#include "paddle/framework/eigen.h" #include "paddle/framework/tensor.h" #include "paddle/platform/place.h" @@ -25,13 +26,13 @@ namespace operators { // Implementation of CPU copy template -void CPUGather(const T* params, const int* indices, const int slice_size, +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 (size_t i = 0; i < index_size; ++i) { + for (int i = 0; i < index_size; ++i) { int index_ = indices[i]; - memcpy(output + i * slice_size, params + index_ * slice_size, slice_bytes); + memcpy(output + i * slice_size, src + index_ * slice_size, slice_bytes); } } @@ -55,12 +56,12 @@ void Gather(const platform::Place& place, const paddle::framework::Tensor* src, int index_size = index->dims()[0]; auto src_dims = src->dims(); - paddle::framework::DDim output_dims(src_dims); + framework::DDim output_dims(src_dims); output_dims[0] = index_size; // slice size int slice_size = 1; - for (size_t i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i]; + for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i]; // Gathering if (platform::is_cpu_place(place)) { diff --git a/paddle/operators/gather_op.cc b/paddle/operators/gather_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..123bed296c462c30bddd3bfbd530098fdbfe4856 --- /dev/null +++ b/paddle/operators/gather_op.cc @@ -0,0 +1,72 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR 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/gather_op.h" +#include "paddle/framework/ddim.h" + +namespace paddle { +namespace operators { + +class GatherOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + int batch_size = ctx.Input("Index")->dims()[0]; + PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0"); + framework::DDim output_dims(ctx.Input("X")->dims()); + output_dims[0] = batch_size; + ctx.Output("Out")->Resize(output_dims); + } +}; + +class GatherGradOp : public framework::OperatorWithKernel { + public: + 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"); + + X_grad->Resize(X->dims()); + } +}; + +class GatherOpMaker : public framework::OpProtoAndCheckerMaker { + public: + 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, + +Out = X[Index] +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +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); diff --git a/paddle/operators/gather_op.h b/paddle/operators/gather_op.h new file mode 100644 index 0000000000000000000000000000000000000000..381854f301870beadb72d9e9b4eb17ff199960fb --- /dev/null +++ b/paddle/operators/gather_op.h @@ -0,0 +1,53 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT 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 "gather.h" +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "scatter.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +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"); + + Y->mutable_data(ctx.GetPlace()); + Gather(ctx.GetPlace(), X, Index, Y); + } +}; + +template +class GatherGradientOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + 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); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/gather_test.cc b/paddle/operators/gather_test.cc index 5de748ec461e4b1a34b75b57c9cd7d5bc9326059..0ae1e99452973feb6d085dd6ef51e2afca988f59 100644 --- a/paddle/operators/gather_test.cc +++ b/paddle/operators/gather_test.cc @@ -35,7 +35,7 @@ TEST(Gather, GatherData) { p_src = src->mutable_data(make_ddim({3, 4}), CPUPlace()); p_index = index->mutable_data(make_ddim({2}), CPUPlace()); - for (size_t i = 0; i < 12; ++i) p_src[i] = i; + for (int i = 0; i < 12; ++i) p_src[i] = i; p_index[0] = 1; p_index[1] = 0; @@ -43,6 +43,10 @@ TEST(Gather, GatherData) { Gather(CPUPlace(), src, index, output); - for (size_t i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4); - for (size_t i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4); + 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); + + delete src; + delete index; + delete output; } diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc index 893cf56e5cf0d99d3f3bfffe98734a868f9b7595..6574880c0eb6324b2dd175e39a364d2ef46e735e 100644 --- a/paddle/operators/gaussian_random_op.cc +++ b/paddle/operators/gaussian_random_op.cc @@ -1,11 +1,8 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. @@ -19,38 +16,44 @@ namespace paddle { namespace operators { template -class GaussianRandomKernel : public framework::OpKernel { +class CPUGaussianRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - float mean = context.op_.GetAttr("mean"); - float std = context.op_.GetAttr("std"); - auto* tensor = context.Output(0); + float mean = context.Attr("mean"); + float std = context.Attr("std"); + auto* tensor = context.Output("Out"); T* data = tensor->mutable_data(context.GetPlace()); - // TODO(dzh): attribute does not support unsigned int. - // And we need a global random seed configuration. - int seed = context.op_.GetAttr("seed"); + unsigned int seed = static_cast(context.Attr("seed")); + std::minstd_rand engine; if (seed == 0) { seed = std::random_device()(); } - std::mt19937 g(seed); - std::normal_distribution distribution(mean, std); - ssize_t size = framework::product(tensor->dims()); - for (int i = 0; i < size; ++i) { - data[i] = distribution(g); + engine.seed(seed); + std::normal_distribution dist(mean, std); + int64_t size = framework::product(tensor->dims()); + for (int64_t i = 0; i < size; ++i) { + data[i] = dist(engine); } } }; class GaussianRandomOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(GaussianRandomOp, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext& context) const override { - auto* tensor = context.Output(0); - auto dims = GetAttr>("dims"); + auto* tensor = context.Output("Out"); + auto dims = Attr>("dims"); + std::vector temp; + temp.reserve(dims.size()); + for (auto dim : dims) { + temp.push_back(static_cast(dim)); + } PADDLE_ENFORCE(dims.size() > 0UL, "dims can be one int or array. dims must be set."); - tensor->Resize(framework::make_ddim(dims)); + tensor->Resize(framework::make_ddim(temp)); } }; @@ -66,8 +69,8 @@ Use to initialize tensor with gaussian random generator. )DOC"); AddAttr>("dims", "The dimension of random tensor."); - AddAttr("mean", "mean value of random.").SetDefault(.0f); - AddAttr("std", "minimum value of random value.").SetDefault(1.0f); + AddAttr("mean", "mean of random tensor.").SetDefault(.0f); + AddAttr("std", "std of random tensor.").SetDefault(1.0f); AddAttr("seed", "Random seed of generator." "0 means use system wide seed") @@ -79,5 +82,6 @@ Use to initialize tensor with gaussian random generator. } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(gaussian_random, ops::GaussianRandomOp, ops::GaussianRandomOpMaker); -REGISTER_OP_CPU_KERNEL(gaussian_random, ops::GaussianRandomKernel); +REGISTER_OP_WITHOUT_GRADIENT(gaussian_random, ops::GaussianRandomOp, + ops::GaussianRandomOpMaker); +REGISTER_OP_CPU_KERNEL(gaussian_random, ops::CPUGaussianRandomKernel); diff --git a/paddle/operators/gaussian_random_op.cu b/paddle/operators/gaussian_random_op.cu index 1340b1e1e9f19fd96ced9e57fab75fe9d33bc84e..d9dbc1dcfe6a6676938d64be93c879ea69148018 100644 --- a/paddle/operators/gaussian_random_op.cu +++ b/paddle/operators/gaussian_random_op.cu @@ -1,53 +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 -#include -#include "paddle/platform/dynload/curand.h" -#include "paddle/platform/gpu_info.h" - +#include +#include +#include +#include #include "paddle/framework/op_registry.h" +#include "paddle/framework/operator.h" namespace paddle { namespace operators { template -class GaussianRandomKernel : public framework::OpKernel { +struct GaussianGenerator { + T mean_, std_; + unsigned int seed_; + + __host__ __device__ GaussianGenerator(T mean, T std, int seed) + : mean_(mean), std_(std), seed_(seed) {} + + __host__ __device__ T operator()(const unsigned int n) const { + thrust::minstd_rand rng; + rng.seed(seed_); + thrust::normal_distribution dist(mean_, std_); + rng.discard(n); + return dist(rng); + } +}; + +template +class GPUGaussianRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - float mean = context.op_.GetAttr("mean"); - float std = context.op_.GetAttr("std"); - auto* tensor = context.Output(0); + auto* tensor = context.Output("Out"); T* data = tensor->mutable_data(context.GetPlace()); - - int seed = context.op_.GetAttr("seed"); + unsigned int seed = static_cast(context.Attr("seed")); if (seed == 0) { std::random_device rd; seed = rd(); } - curandGenerator_t g; - PADDLE_ENFORCE(platform::dynload::curandCreateGenerator( - &g, CURAND_RNG_PSEUDO_DEFAULT)); - PADDLE_ENFORCE( - platform::dynload::curandSetPseudoRandomGeneratorSeed(g, seed)); - platform::dynload::curandGenerateNormal( - g, data, framework::product(tensor->dims()), mean, std); + T mean = static_cast(context.Attr("mean")); + T std = static_cast(context.Attr("std")); + thrust::counting_iterator index_sequence_begin(0); + ssize_t N = framework::product(tensor->dims()); + thrust::transform(index_sequence_begin, index_sequence_begin + N, + thrust::device_ptr(data), + GaussianGenerator(mean, std, seed)); } }; } // namespace operators } // namespace paddle -namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(gaussian_random, ops::GaussianRandomKernel); +REGISTER_OP_GPU_KERNEL(gaussian_random, + paddle::operators::GPUGaussianRandomKernel); diff --git a/paddle/operators/identity_op.cc b/paddle/operators/identity_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..7d9d4fa519d1c690feacbadc5175aeab49082282 --- /dev/null +++ b/paddle/operators/identity_op.cc @@ -0,0 +1,57 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR 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/net_op.h" +#include "paddle/operators/scale_op.h" + +namespace paddle { +namespace operators { + +// The identity operator is an alias of the scale operator. This is also an +// example for creating an alias for an existing operator. +template +class IdentityOpMaker : public framework::OpProtoAndCheckerMaker { + public: + IdentityOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The input tensor of identity operator."); + AddOutput("Out", "The output tensor of identity operator."); + AddComment(R"DOC( +The identity operator is an alias of the scale operator +with the attribute scale fixed to 1.0. +)DOC"); + } +}; + +template +class IdentityOp : public NetOp { + public: + IdentityOp(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("X")}}}, {{"Out", {Output("Out")}}}, + {{"scale", static_cast(1)}})); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_WITHOUT_GRADIENT(identity, ops::IdentityOp, + ops::IdentityOpMaker); diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..94d40890a765413e88a35a6ad995ca97ac84dcda --- /dev/null +++ b/paddle/operators/lookup_table_op.cc @@ -0,0 +1,72 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/lookup_table_op.h" + +namespace paddle { +namespace operators { + +class LookupTableOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &context) const override { + auto table_t = context.Input("W"); + auto ids_t = context.Input("Ids"); + auto output_t = context.Output("Out"); + + output_t->Resize({ids_t->dims()[0], table_t->dims()[1]}); + } +}; + +class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { + public: + LookupTableOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("W", + "An input represents embedding tensors," + " which is a learnable parameter."); + AddInput("Ids", + "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."); + } +}; + +class LookupTableOpGrad : public framework::OperatorWithKernel { + public: + 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()); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(lookup_table, ops::LookupTableOp, ops::LookupTableOpMaker, + lookup_table_grad, ops::LookupTableOpGrad); + +REGISTER_OP_CPU_KERNEL(lookup_table, ops::LookupTableKernel); +REGISTER_OP_CPU_KERNEL(lookup_table_grad, ops::LookupTableGradKernel); diff --git a/paddle/operators/lookup_table_op.cu b/paddle/operators/lookup_table_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..27eee3436af8107cef2aa3577ea238be49edf1af --- /dev/null +++ b/paddle/operators/lookup_table_op.cu @@ -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. */ + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/platform/assert.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +__global__ void LookupTable(T* output, const T* table, const int32_t* ids, + const int N, const int K, const int D) { + int idx = threadIdx.x; + int idy = blockIdx.x + threadIdx.y * GridDimX; + + while (idy < K) { + int id = ids[idy]; + PADDLE_ASSERT(id >= 0); + PADDLE_ASSERT(id < N); + T* out = output + idy * D; + const T* tab = table + id * D; + for (int i = idx; i < D; i += BlockDimX) { + out[i] = tab[i]; + } + idy += BlockDimY * GridDimX; + } +} + +template +__global__ void LookupTableGrad(T* table, const T* output, const int32_t* ids, + const int N, const int K, const int D) { + int idx = threadIdx.x; + int idy = blockIdx.x + threadIdx.y * GridDimX; + + while (idy < K) { + int id = ids[idy]; + PADDLE_ASSERT(id >= 0); + PADDLE_ASSERT(id < N); + const T* out = output + idy * D; + T* tab = table + id * D; + for (int i = idx; i < D; i += BlockDimX) { + paddle::platform::CudaAtomicAdd(&tab[i], out[i]); + } + idy += BlockDimY * GridDimX; + } +} + +template +class LookupTableCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto table_t = context.Input("W"); + auto ids_t = context.Input("Ids"); + auto output_t = context.Output("Out"); + + size_t N = table_t->dims()[0]; + size_t D = table_t->dims()[1]; + size_t K = product(ids_t->dims()); + auto ids = ids_t->data(); + auto table = table_t->data(); + auto output = output_t->mutable_data(context.GetPlace()); + + dim3 threads(128, 8); + dim3 grids(8, 1); + LookupTable<<>>(output, table, ids, N, K, D); + } +}; + +template +class LookupTableGradCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto ids_t = context.Input("Ids"); + auto d_output_t = context.Input(framework::GradVarName("Out")); + auto d_table_t = context.Output(framework::GradVarName("W")); + + int N = d_table_t->dims()[0]; + int D = d_table_t->dims()[1]; + int K = product(ids_t->dims()); + const int32_t* ids = ids_t->data(); + const T* d_output = d_output_t->data(); + T* d_table = d_table_t->mutable_data(context.GetPlace()); + + auto t = framework::EigenVector::Flatten(*d_table_t); + t.device(context.GetEigenDevice()) = + t.constant(static_cast(0)); + + dim3 threads(128, 8); + dim3 grids(8, 1); + LookupTableGrad<<>>(d_table, d_output, ids, N, + K, D); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(lookup_table, ops::LookupTableCUDAKernel); +REGISTER_OP_GPU_KERNEL(lookup_table_grad, + ops::LookupTableGradCUDAKernel); diff --git a/paddle/operators/lookup_table_op.h b/paddle/operators/lookup_table_op.h new file mode 100644 index 0000000000000000000000000000000000000000..877b36cef4ea9cdaaaf37c97d5e5bfce55b91436 --- /dev/null +++ b/paddle/operators/lookup_table_op.h @@ -0,0 +1,75 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class LookupTableKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto table_t = context.Input("W"); // float tensor + auto ids_t = context.Input("Ids"); // int tensor + auto output_t = context.Output("Out"); // float tensor + + int N = table_t->dims()[0]; + int D = table_t->dims()[1]; + auto ids = ids_t->data(); + auto table = table_t->data(); + auto output = output_t->mutable_data(context.GetPlace()); + for (ssize_t i = 0; i < product(ids_t->dims()); ++i) { + PADDLE_ENFORCE_LT(ids[i], N); + PADDLE_ENFORCE_GE(ids[i], 0); + memcpy(output + i * D, table + ids[i] * D, D * sizeof(T)); + } + } +}; + +template +class LookupTableGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto ids_t = context.Input("Ids"); + auto d_output_t = context.Input(framework::GradVarName("Out")); + auto d_table_t = context.Output(framework::GradVarName("W")); + + int N = d_table_t->dims()[0]; + int D = d_table_t->dims()[1]; + auto ids = ids_t->data(); + const T* d_output = d_output_t->data(); + T* d_table = d_table_t->mutable_data(context.GetPlace()); + + auto t = framework::EigenVector::Flatten(*d_table_t); + t.device(context.GetEigenDevice()) = + t.constant(static_cast(0)); + + for (ssize_t i = 0; i < product(ids_t->dims()); ++i) { + PADDLE_ENFORCE_LT(ids[i], N); + PADDLE_ENFORCE_GE(ids[i], 0); + for (int j = 0; j < D; ++j) { + d_table[ids[i] * D + j] += d_output[i * D + j]; + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/CMakeLists.txt b/paddle/operators/math/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..f8333f34f7b4c7b0f9a0f14a7a33f9d98e1d331c --- /dev/null +++ b/paddle/operators/math/CMakeLists.txt @@ -0,0 +1,10 @@ + +if(WITH_GPU) + nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc + im2col.cu DEPS cblas device_context) +else() + cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context) +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/im2col.cc b/paddle/operators/math/im2col.cc new file mode 100644 index 0000000000000000000000000000000000000000..5727c1cab16c1379ffe77f5594c057e93a042785 --- /dev/null +++ b/paddle/operators/math/im2col.cc @@ -0,0 +1,260 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR 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/im2col.h" + +namespace paddle { +namespace operators { +namespace math { + +/* + * im = [input_channels, input_height, input_width] + * col = + * [input_channels, filter_height, filter_width, output_height, output_width] + */ +template +class Im2ColFunctor { + public: + void operator()(const framework::Tensor& im, framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[1]; + int filter_width = col.dims()[2]; + int output_height = col.dims()[3]; + int output_width = col.dims()[4]; + int channels_col = input_channels * filter_height * filter_width; + + const T* im_data = im.data(); + T* col_data = col.data(); + + for (int c = 0; c < channels_col; ++c) { + int w_offset = c % filter_width; + int h_offset = (c / filter_width) % filter_height; + int c_im = c / filter_width / filter_height; + for (int h = 0; h < output_height; ++h) { + for (int w = 0; w < output_width; ++w) { + int im_row_idx = h * stride_height + h_offset; + int im_col_idx = w * stride_width + w_offset; + if ((im_row_idx - padding_height) < 0 || + (im_row_idx - padding_height) >= input_height || + (im_col_idx - padding_width) < 0 || + (im_col_idx - padding_width) >= input_width) { + col_data[(c * output_height + h) * output_width + w] = T(0); + } else { + im_row_idx += c_im * input_height - padding_height; + im_col_idx -= padding_width; + col_data[(c * output_height + h) * output_width + w] = + im_data[im_row_idx * input_width + im_col_idx]; + } + } + } + } + } +}; + +/* + * im = [input_channels, input_height, input_width] + * col = + * [input_channels, filter_height, filter_width, output_height, output_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) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[1]; + int filter_width = col.dims()[2]; + int output_height = col.dims()[3]; + int output_width = col.dims()[4]; + int channels_col = input_channels * filter_height * filter_width; + + T* im_data = im.data(); + const T* col_data = col.data(); + + for (int c = 0; c < channels_col; ++c) { + int w_offset = c % filter_width; + int h_offset = (c / filter_width) % filter_height; + int c_im = c / filter_width / filter_height; + for (int h = 0; h < output_height; ++h) { + for (int w = 0; w < output_width; ++w) { + int im_row_idx = h * stride_height + h_offset; + int im_col_idx = w * stride_width + w_offset; + if ((im_row_idx - padding_height) >= 0 && + (im_row_idx - padding_height) < input_height && + (im_col_idx - padding_width) >= 0 && + (im_col_idx - padding_width) < input_width) { + im_row_idx += c_im * input_height - padding_height; + im_col_idx -= padding_width; + im_data[im_row_idx * input_width + im_col_idx] += + col_data[(c * output_height + h) * output_width + w]; + } + } + } + } + } +}; + +template class Im2ColFunctor; +template class Im2ColFunctor; +template class Col2ImFunctor; +template class Col2ImFunctor; + +/* + * im = [input_channels, input_height, input_width] + * col = + * [output_height, output_width, input_channels, filter_height, filter_width] + */ +template +class Im2ColFunctor { + public: + void operator()(const framework::Tensor& im, framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[3]; + int filter_width = col.dims()[4]; + int output_height = col.dims()[0]; + int output_width = col.dims()[1]; + + const T* im_data = im.data(); + T* col_data = col.data(); + + for (int col_row_idx = 0; col_row_idx < output_height; ++col_row_idx) { + for (int col_col_idx = 0; col_col_idx < output_width; ++col_col_idx) { + for (int channel = 0; channel < input_channels; ++channel) { + for (int filter_row_idx = 0; filter_row_idx < filter_height; + ++filter_row_idx) { + for (int filter_col_idx = 0; filter_col_idx < filter_width; + ++filter_col_idx) { + int im_row_offset = + col_row_idx * stride_height + filter_row_idx - padding_height; + int im_col_offset = + col_col_idx * stride_width + filter_col_idx - padding_width; + int col_offset = (((col_row_idx * output_width + col_col_idx) * + input_channels + + channel) * + filter_height + + filter_row_idx) * + filter_width + + filter_col_idx; + if (im_row_offset < 0 || im_row_offset >= input_height || + im_col_offset < 0 || im_col_offset >= input_width) { + col_data[col_offset] = T(0); + } else { + int im_offset = + (channel * input_height + im_row_offset) * input_width + + im_col_offset; + col_data[col_offset] = im_data[im_offset]; + } + } + } + } + } + } + } +}; + +/* + * im = [input_channels, input_height, input_width] + * col = + * [output_height, output_width, input_channels, filter_height, filter_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) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[3]; + int filter_width = col.dims()[4]; + int output_height = col.dims()[0]; + int output_width = col.dims()[1]; + + T* im_data = im.data(); + const T* col_data = col.data(); + + for (int col_row_idx = 0; col_row_idx < output_height; ++col_row_idx) { + for (int col_col_idx = 0; col_col_idx < output_width; ++col_col_idx) { + for (int channel = 0; channel < input_channels; ++channel) { + for (int filter_row_idx = 0; filter_row_idx < filter_height; + ++filter_row_idx) { + for (int filter_col_idx = 0; filter_col_idx < filter_width; + ++filter_col_idx) { + int im_row_offset = + col_row_idx * stride_height + filter_row_idx - padding_height; + int im_col_offset = + col_col_idx * stride_width + filter_col_idx - padding_width; + int col_offset = (((col_row_idx * output_width + col_col_idx) * + input_channels + + channel) * + filter_height + + filter_row_idx) * + filter_width + + filter_col_idx; + if (im_row_offset >= 0 && im_row_offset < input_height && + im_col_offset >= 0 && im_col_offset < input_width) { + int im_offset = + (channel * input_height + im_row_offset) * input_width + + im_col_offset; + im_data[im_offset] += col_data[col_offset]; + } + } + } + } + } + } + } +}; + +template class Im2ColFunctor; +template class Im2ColFunctor; +template class Col2ImFunctor; +template class Col2ImFunctor; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/im2col.cu b/paddle/operators/math/im2col.cu new file mode 100644 index 0000000000000000000000000000000000000000..9bff7bee3c95093852305d392af0949b831e5665 --- /dev/null +++ b/paddle/operators/math/im2col.cu @@ -0,0 +1,374 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR 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/im2col.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { +namespace math { + +template +__global__ void im2col(const T* data_im, int num_outs, int height, int width, + int filter_height, int filter_width, int stride_height, + int stride_width, int padding_height, int padding_width, + int output_height, int output_width, T* data_col) { + int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; + if (index < num_outs) { + int w_out = index % output_width; + index /= output_width; + int h_out = index % output_height; + int channel_in = index / output_height; + int channel_out = channel_in * filter_height * filter_width; + int h_in = h_out * stride_height; + int w_in = w_out * stride_width; + + data_col += (channel_out * output_height + h_out) * output_width + w_out; + for (int i = 0; i < filter_height; ++i) { + for (int j = 0; j < filter_width; ++j) { + int rIdx = int(h_in + i); + int cIdx = int(w_in + j); + if ((rIdx - (int)padding_height) >= (int)height || + (rIdx - (int)padding_height) < 0 || + (cIdx - (int)padding_width) >= (int)width || + (cIdx - (int)padding_width) < 0) { + *data_col = 0; + } else { + rIdx = rIdx + channel_in * height - padding_height; + cIdx = cIdx - padding_width; + *data_col = data_im[rIdx * width + cIdx]; + } + data_col += output_height * output_width; + } + } + } +} + +/* + * im = [input_channels, input_height, input_width] + * col = + * [input_channels, filter_height, filter_width, output_height, output_width] + */ +template +class Im2ColFunctor { + public: + void operator()(const framework::Tensor& im, framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[1]; + int filter_width = col.dims()[2]; + int output_height = col.dims()[3]; + int output_width = col.dims()[4]; + + int num_outputs = input_channels * output_height * output_width; + int blocks = (num_outputs + 1024 - 1) / 1024; + int block_x = 512; + int block_y = (blocks + 512 - 1) / 512; + dim3 threads(1024, 1); + dim3 grid(block_x, block_y); + im2col<<< + grid, threads, 0, + reinterpret_cast(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()); + } +}; + +template +__global__ void col2im(size_t n, const T* data_col, size_t height, size_t width, + size_t channels, size_t filter_height, + size_t filter_width, size_t stride_height, + size_t stride_width, size_t padding_height, + size_t padding_width, size_t output_height, + size_t output_width, T* data_im) { + size_t index = + (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; + if (index < n) { + T val = 0; + int w = int(index % width); + int h = int((index / width) % height); + int c = int(index / (width * height)); + if ((w - (int)padding_width) >= 0 && + (w - (int)padding_width) < (width - 2 * padding_width) && + (h - (int)padding_height) >= 0 && + (h - padding_height) < (height - 2 * padding_height)) { + // compute the start and end of the output + int w_col_start = (w < (int)filter_width) + ? 0 + : (w - int(filter_width)) / (int)stride_width + 1; + int w_col_end = + min((int)(w / (int)stride_width + 1), (int)(output_width)); + int h_col_start = (h < (int)filter_height) + ? 0 + : (h - (int)filter_height) / (int)stride_height + 1; + int h_col_end = min(int(h / stride_height + 1), int(output_height)); + for (int h_col = h_col_start; h_col < h_col_end; ++h_col) { + for (int w_col = w_col_start; w_col < w_col_end; ++w_col) { + // the col location: [c * width * height + h_out, w_out] + int c_col = int(c * filter_height * filter_width) + + (h - h_col * (int)stride_height) * (int)filter_width + + (w - w_col * (int)stride_width); + val += + data_col[(c_col * output_height + h_col) * output_width + w_col]; + } + } + h -= padding_height; + w -= padding_width; + data_im[c * ((width - 2 * padding_width) * + (height - 2 * padding_height)) + + h * (width - 2 * padding_width) + w] += val; + } + } +} + +/* + * im = [input_channels, input_height, input_width] + * col = + * [input_channels, filter_height, filter_width, output_height, output_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) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[1]; + int filter_width = col.dims()[2]; + int output_height = col.dims()[3]; + int output_width = col.dims()[4]; + + size_t num_kernels = input_channels * (input_height + 2 * padding_height) * + (input_width + 2 * padding_width); + + size_t blocks = (num_kernels + 1024 - 1) / 1024; + size_t block_x = 512; + size_t block_y = (blocks + 512 - 1) / 512; + dim3 threads(1024, 1); + dim3 grid(block_x, block_y); + + // To avoid involving atomic operations, we will launch one kernel per + // bottom dimension, and then in the kernel add up the top dimensions. + col2im<<< + grid, threads, 0, + reinterpret_cast(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, + padding_width, output_height, output_width, im.data()); + } +}; + +template class Im2ColFunctor; +template class Im2ColFunctor; +template class Col2ImFunctor; +template class Col2ImFunctor; + +template +__global__ void im2colOCF(const T* im_data, T* col_data, int input_channels, + int input_height, int input_width, int filter_height, + int filter_width, int stride_height, int stride_width, + int padding_height, int padding_width, + int output_height, int output_width) { + int swid = blockIdx.x; + int shid = blockIdx.y; + for (int channelid = threadIdx.z; channelid < input_channels; + channelid += blockDim.z) { + for (int idy = threadIdx.y; idy < filter_height; idy += blockDim.y) { + for (int idx = threadIdx.x; idx < filter_width; idx += blockDim.x) { + int width_offset = idx + swid * stride_width - padding_width; + int height_offset = idy + shid * stride_height - padding_height; + int im_offset = width_offset + height_offset * input_width + + channelid * input_height * input_width; + + int col_offset = idx + idy * filter_width + + channelid * filter_height * filter_width + + (shid * output_width + swid) * + (input_channels * filter_height * filter_width); + + if (height_offset >= input_height || height_offset < 0 || + width_offset >= input_width || width_offset < 0) { + col_data[col_offset] = T(0); + } else { + col_data[col_offset] = im_data[im_offset]; + } + } + } + } +} + +/* + * im = [input_channels, input_height, input_width] + * col = + * [output_height, output_width, input_channels, filter_height, filter_width] + */ +template +class Im2ColFunctor { + public: + void operator()(const framework::Tensor& im, framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[3]; + int filter_width = col.dims()[4]; + int output_height = col.dims()[0]; + int output_width = col.dims()[1]; + + int block_dim_x = 0; + int block_dim_y = 0; + if (filter_height <= 4 && filter_width <= 4) { + block_dim_x = 4; + block_dim_y = 4; + } else if (filter_height <= 8 && filter_width <= 8) { + block_dim_x = 8; + block_dim_y = 8; + } else if (filter_height <= 16 && filter_width <= 16) { + block_dim_x = 16; + block_dim_y = 16; + } else { + block_dim_x = 32; + block_dim_y = 32; + } + + int block_dim_z = 1024 / block_dim_x / block_dim_y; + dim3 threads(block_dim_x, block_dim_y, + std::min(block_dim_z, input_channels)); + dim3 grid(output_width, output_height); + im2colOCF<<< + grid, threads, 0, + reinterpret_cast(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); + } +}; + +template +__global__ void col2imOCF(T* im_data, const T* col_data, int input_channels, + int input_height, int input_width, int filter_height, + int filter_width, int stride_height, int stride_width, + int padding_height, int padding_width, + int output_height, int output_width) { + int swid = blockIdx.x; + int shid = blockIdx.y; + for (int channelid = threadIdx.z; channelid < input_channels; + channelid += blockDim.z) { + for (int idy = threadIdx.y; idy < filter_height; idy += blockDim.y) { + for (int idx = threadIdx.x; idx < filter_width; idx += blockDim.x) { + int width_offset = idx + swid * stride_width - padding_width; + int height_offset = idy + shid * stride_height - padding_height; + int im_offset = width_offset + height_offset * input_width + + channelid * input_height * input_width; + + int col_offset = idx + idy * filter_width + + channelid * filter_height * filter_width + + (shid * output_width + swid) * + (input_channels * filter_height * filter_width); + + if (height_offset >= 0 && height_offset < input_height && + width_offset >= 0 && width_offset < input_width) { + paddle::platform::CudaAtomicAdd(im_data + im_offset, + col_data[col_offset]); + } + } + } + } +} + +/* + * im = [input_channels, input_height, input_width] + * col = + * [output_height, output_width, input_channels, filter_height, filter_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) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[3]; + int filter_width = col.dims()[4]; + int output_height = col.dims()[0]; + int output_width = col.dims()[1]; + + int block_dim_x = 0; + int block_dim_y = 0; + if (filter_height <= 4 && filter_width <= 4) { + block_dim_x = 4; + block_dim_y = 4; + } else if (filter_height <= 8 && filter_width <= 8) { + block_dim_x = 8; + block_dim_y = 8; + } else if (filter_height <= 16 && filter_width <= 16) { + block_dim_x = 16; + block_dim_y = 16; + } else { + block_dim_x = 32; + block_dim_y = 32; + } + + int block_dim_z = 1024 / block_dim_x / block_dim_y; + dim3 threads(block_dim_x, block_dim_y, + std::min(block_dim_z, input_channels)); + dim3 grid(output_width, output_height); + col2imOCF<<< + grid, threads, 0, + reinterpret_cast(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); + } +}; + +template class Im2ColFunctor; +template class Im2ColFunctor; +template class Col2ImFunctor; +template class Col2ImFunctor; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/im2col.h b/paddle/operators/math/im2col.h new file mode 100644 index 0000000000000000000000000000000000000000..8958c5457cc2c3034c34ca82fb2e98cc06be63c5 --- /dev/null +++ b/paddle/operators/math/im2col.h @@ -0,0 +1,90 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/framework/tensor.h" +#include "paddle/platform/device_context.h" + +namespace paddle { +namespace operators { +namespace math { + +/* The storage format of the coldata in the Im2ColFunctor and Col2ImFunctor. */ +enum class ColFormat { kCFO = 0, kOCF = 1 }; + +/* + * \brief Converts the image data of three dimensions(CHW) into a colData of + * five dimensions in the Im2ColFunctor calculation, + * And in the Col2ImFunctor calculation, it is reversed. + * + * \param imData Image data. + * \param imShape The shape of imData, + * [input_channels, input_height, input_width]. + * \param colData Column data. + * \param colShape The shape of colData. + * + * If the template argument Format is kCFO, the shape of colData is: + * [input_channels, filter_height, filter_width, output_height, output_width] + * So, it is easy to reshape into a convolution matrix for convolution + * calculation based on matrix multiplication. + * The shape of convolution matrix is [height, width], where the height is equal + * input_channels * filter_height * filter_width, and the width is equal + * output_height * output_width. + * + * Reshape: + * shape of colData shape of convolution matrix + * [input_channels, + * filter_height, + * filter_width, ======> [height, width] + * output_height, + * output_width] + * + * If the template argument Format is kOCF, the shape of colData is: + * [output_height, output_width, input_channels, filter_height, filter_width] + * So, it is easy to reshape into a sequence matrix for rnn calculation. + * The shape of sequence matrix is [seq_length, step_size], where the seq_length + * is equal output_height * output_width, and the step_size is equal + * input_channels * filter_height * filter_width. + * + * Reshape: + * shape of colData shape of sequence matrix + * [output_height, + * output_width, + * input_channels, ======> [seqLength, stepSize] + * filter_height, + * filter_width] + * + * \note The caller needs to ensure that imShape.inputChannels is equal to + * colShape.inputChannels. + */ +template +class Im2ColFunctor { + public: + void operator()(const framework::Tensor& im, framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context); +}; + +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); +}; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/im2col_test.cc b/paddle/operators/math/im2col_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..186a33edcec88bd5e51091a524a778eeb27ad526 --- /dev/null +++ b/paddle/operators/math/im2col_test.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/math/im2col.h" +#include +#include + +template +void testIm2col() { + paddle::framework::Tensor input_tmp; + paddle::framework::Tensor input; + paddle::framework::Tensor output_cfo; + paddle::framework::Tensor output_ocf; + paddle::framework::Tensor output_tmp; + + /** + * input = [0, 1, 2, + * 3, 4, 5] + * + * output_cfo = [0, 1 + * 1, 2 + * 3, 4 + * 4, 5] + * + * output_ocf = [0, 1, 3, 4 + * 1, 2, 4, 5] + */ + int input_height = 2; + int input_width = 3; + int filter_size = 2; + int stride = 1; + int padding = 0; + int output_height = (input_height - filter_size + 2 * padding) / stride + 1; + int output_width = (input_width - filter_size + 2 * padding) / stride + 1; + float* input_ptr = input_tmp.mutable_data( + {1, input_height, input_width}, paddle::platform::CPUPlace()); + float arr[6] = {0, 1, 2, 3, 4, 5}; + memcpy(input_ptr, arr, 6 * sizeof(float)); + + auto* place = new Place(); + if (paddle::platform::is_cpu_place(*place)) { + input = input_tmp; + } else { + input.CopyFrom(input_tmp, *place); + } + output_cfo.mutable_data( + {1, filter_size, filter_size, output_height, output_width}, *place); + output_ocf.mutable_data( + {output_height, output_width, 1, filter_size, filter_size}, *place); + + paddle::operators::math::Im2ColFunctor< + paddle::operators::math::ColFormat::kCFO, Place, float> + im2col; + paddle::operators::math::Im2ColFunctor< + paddle::operators::math::ColFormat::kOCF, Place, float> + im2col_ocf; + + paddle::platform::DeviceContext* context; + if (paddle::platform::is_cpu_place(*place)) { + context = + new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace()); + } else { +#ifndef PADDLE_ONLY_CPU + 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); + + float* out_cfo_ptr; + if (paddle::platform::is_cpu_place(*place)) { + out_cfo_ptr = output_cfo.data(); + } else { + output_tmp.CopyFrom(output_cfo, paddle::platform::CPUPlace()); + out_cfo_ptr = output_tmp.data(); + } + EXPECT_EQ(out_cfo_ptr[0], 0); + EXPECT_EQ(out_cfo_ptr[1], 1); + EXPECT_EQ(out_cfo_ptr[2], 1); + EXPECT_EQ(out_cfo_ptr[3], 2); + EXPECT_EQ(out_cfo_ptr[4], 3); + EXPECT_EQ(out_cfo_ptr[5], 4); + EXPECT_EQ(out_cfo_ptr[6], 4); + EXPECT_EQ(out_cfo_ptr[7], 5); + + float* out_ocf_ptr; + if (paddle::platform::is_cpu_place(*place)) { + out_ocf_ptr = output_ocf.data(); + } else { + output_tmp.CopyFrom(output_ocf, paddle::platform::CPUPlace()); + out_ocf_ptr = output_tmp.data(); + } + EXPECT_EQ(out_ocf_ptr[0], 0); + EXPECT_EQ(out_ocf_ptr[1], 1); + EXPECT_EQ(out_ocf_ptr[2], 3); + EXPECT_EQ(out_ocf_ptr[3], 4); + EXPECT_EQ(out_ocf_ptr[4], 1); + EXPECT_EQ(out_ocf_ptr[5], 2); + EXPECT_EQ(out_ocf_ptr[6], 4); + EXPECT_EQ(out_ocf_ptr[7], 5); +} + +TEST(math, im2col) { + testIm2col(); +#ifndef PADDLE_ONLY_CPU + testIm2col(); +#endif +} \ No newline at end of file diff --git a/paddle/operators/math/math_function.cc b/paddle/operators/math/math_function.cc new file mode 100644 index 0000000000000000000000000000000000000000..1e86fc3d166077265e0f433a6712b0665ea5a152 --- /dev/null +++ b/paddle/operators/math/math_function.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/math/math_function.h" + +namespace paddle { +namespace operators { +namespace math { + +template <> +void gemm(const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, const int M, + const int N, const int K, + const float alpha, const float* A, + const float* B, const float beta, float* C, + platform::DeviceContext* context) { + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + int ldc = N; + cblas_sgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, + beta, C, ldc); +} + +template <> +void gemm(const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, const int M, + const int N, const int K, + const double alpha, const double* A, + const double* B, const double beta, + double* C, + platform::DeviceContext* context) { + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + int ldc = N; + cblas_dgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, + beta, C, ldc); +} + +template <> +void matmul(const framework::Tensor& matrix_a, + bool trans_a, + const framework::Tensor& matrix_b, + bool trans_b, float alpha, + framework::Tensor* matrix_out, + float beta, + platform::DeviceContext* context) { + auto dim_a = matrix_a.dims(); + auto dim_b = matrix_b.dims(); + auto dim_out = matrix_out->dims(); + PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, + "The input and output of matmul be matrix"); + + PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) && + platform::is_cpu_place(matrix_b.place()) && + platform::is_cpu_place(matrix_out->place()), + "Matrix must all be in CPUPlace"); + + int M = dim_out[0]; + int N = dim_out[1]; + int K = (trans_a == false) ? dim_a[1] : dim_a[0]; + + CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; + CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; + + gemm( + transA, transB, M, N, K, alpha, matrix_a.data(), + matrix_b.data(), beta, matrix_out->data(), context); +} + +template <> +void matmul(const framework::Tensor& matrix_a, + bool trans_a, + const framework::Tensor& matrix_b, + bool trans_b, double alpha, + framework::Tensor* matrix_out, + double beta, + platform::DeviceContext* context) { + auto dim_a = matrix_a.dims(); + auto dim_b = matrix_b.dims(); + auto dim_out = matrix_out->dims(); + PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, + "The input and output of matmul be matrix"); + + PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) && + platform::is_cpu_place(matrix_b.place()) && + platform::is_cpu_place(matrix_out->place()), + "Matrix must all be in CPUPlace"); + + int M = dim_out[0]; + int N = dim_out[1]; + int K = (trans_a == false) ? dim_a[1] : dim_a[0]; + + CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; + CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; + + gemm( + transA, transB, M, N, K, alpha, matrix_a.data(), + matrix_b.data(), beta, matrix_out->data(), context); +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/math_function.cu b/paddle/operators/math/math_function.cu new file mode 100644 index 0000000000000000000000000000000000000000..da40b27c948918e4997f4a046d2145552296158b --- /dev/null +++ b/paddle/operators/math/math_function.cu @@ -0,0 +1,127 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/math/math_function.h" + +namespace paddle { +namespace operators { +namespace math { + +template <> +void gemm(const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, const int M, + const int N, const int K, + const float alpha, const float* A, + const float* B, const float beta, float* C, + platform::DeviceContext* context) { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + cublasOperation_t cuTransA = + (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = + (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + + PADDLE_ENFORCE(platform::dynload::cublasSgemm( + reinterpret_cast(context)->cublas_handle(), + cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N)); +} + +template <> +void gemm(const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, const int M, + const int N, const int K, + const double alpha, const double* A, + const double* B, const double beta, + double* C, + platform::DeviceContext* context) { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + cublasOperation_t cuTransA = + (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = + (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + PADDLE_ENFORCE(platform::dynload::cublasDgemm( + reinterpret_cast(context)->cublas_handle(), + cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N)); +} + +template <> +void matmul(const framework::Tensor& matrix_a, + bool trans_a, + const framework::Tensor& matrix_b, + bool trans_b, float alpha, + framework::Tensor* matrix_out, + float beta, + platform::DeviceContext* context) { + auto dim_a = matrix_a.dims(); + auto dim_b = matrix_b.dims(); + auto dim_out = matrix_out->dims(); + PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, + "The input and output of matmul be matrix"); + + PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) && + platform::is_gpu_place(matrix_b.place()) && + platform::is_gpu_place(matrix_out->place()), + "Matrix must all be in GPUPlace"); + + int M = dim_out[0]; + int N = dim_out[1]; + int K = (trans_a == false) ? dim_a[1] : dim_a[0]; + + CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; + CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; + + gemm( + transA, transB, M, N, K, alpha, matrix_a.data(), + matrix_b.data(), beta, matrix_out->data(), context); +} + +template <> +void matmul(const framework::Tensor& matrix_a, + bool trans_a, + const framework::Tensor& matrix_b, + bool trans_b, double alpha, + framework::Tensor* matrix_out, + double beta, + platform::DeviceContext* context) { + auto dim_a = matrix_a.dims(); + auto dim_b = matrix_b.dims(); + auto dim_out = matrix_out->dims(); + PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, + "The input and output of matmul be matrix"); + + PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) && + platform::is_gpu_place(matrix_b.place()) && + platform::is_gpu_place(matrix_out->place()), + "Matrix must all be in GPUPlace"); + + int M = dim_out[0]; + int N = dim_out[1]; + int K = (trans_a == false) ? dim_a[1] : dim_a[0]; + + CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; + CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; + + gemm( + transA, transB, M, N, K, alpha, matrix_a.data(), + matrix_b.data(), beta, matrix_out->data(), context); +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/math_function.h b/paddle/operators/math/math_function.h new file mode 100644 index 0000000000000000000000000000000000000000..155589fadb3ed9f59160a750d546dd8093a56cbe --- /dev/null +++ b/paddle/operators/math/math_function.h @@ -0,0 +1,82 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#ifdef PADDLE_USE_MKLML +#include +#include +#include +#endif + +#ifdef PADDLE_USE_MKL +#include +#include +#endif + +#ifdef PADDLE_USE_ATLAS +extern "C" { +#include +#include +} +#endif + +#ifdef PADDLE_USE_OPENBLAS +#include +#include +#endif + +#ifndef LAPACK_FOUND +extern "C" { +#include +int LAPACKE_sgetrf(int matrix_layout, int m, int n, float* a, int lda, + int* ipiv); +int LAPACKE_dgetrf(int matrix_layout, int m, int n, double* a, int lda, + int* ipiv); +int LAPACKE_sgetri(int matrix_layout, int n, float* a, int lda, + const int* ipiv); +int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda, + const int* ipiv); +} +#endif + +#include + +#include "paddle/framework/tensor.h" +#include "paddle/platform/device_context.h" +#include "paddle/platform/enforce.h" + +namespace paddle { +namespace operators { +namespace math { + +// Support continuous memory now +// If transA = N, and transB = N +// Then matrixA: M * K, matrixB: K * N matrixC : M * N +// For more detailed info, please refer to +// http://www.netlib.org/lapack/explore-html/d4/de2/sgemm_8f.html +template +void gemm(const 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, platform::DeviceContext* context); + +// matrix multiply with continuous memory +template +void matmul(const framework::Tensor& matrix_a, bool trans_a, + const framework::Tensor& matrix_b, bool trans_b, T alpha, + framework::Tensor* matrix_out, T beta, + platform::DeviceContext* context); + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/math_function_test.cc b/paddle/operators/math/math_function_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..6c020c4ff7285b43bc5836d80c173d3a068e72b3 --- /dev/null +++ b/paddle/operators/math/math_function_test.cc @@ -0,0 +1,75 @@ +#include "paddle/operators/math/math_function.h" +#include "gtest/gtest.h" + +#ifndef PADDLE_ONLY_CPU +TEST(math_function, notrans_mul_trans) { + paddle::framework::Tensor input1; + paddle::framework::Tensor input1_gpu; + paddle::framework::Tensor input2_gpu; + paddle::framework::Tensor out_gpu; + paddle::framework::Tensor out; + + auto* cpu_place = new paddle::platform::CPUPlace(); + float* input1_ptr = input1.mutable_data({2, 3}, *cpu_place); + float arr[6] = {0, 1, 2, 3, 4, 5}; + memcpy(input1_ptr, arr, 6 * sizeof(float)); + + auto* gpu_place = new paddle::platform::GPUPlace(0); + paddle::platform::DeviceContext* context = + new paddle::platform::CUDADeviceContext(*gpu_place); + + input1_gpu.CopyFrom(input1, *gpu_place); + input2_gpu.CopyFrom(input1, *gpu_place); + + out_gpu.mutable_data({2, 2}, *gpu_place); + + paddle::operators::math::matmul( + input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0, context); + + out.CopyFrom(out_gpu, *cpu_place); + + float* out_ptr = out.data(); + EXPECT_EQ(out_ptr[0], 5); + EXPECT_EQ(out_ptr[1], 14); + EXPECT_EQ(out_ptr[2], 14); + EXPECT_EQ(out_ptr[3], 50); +} + +TEST(math_function, trans_mul_notrans) { + paddle::framework::Tensor input1; + paddle::framework::Tensor input1_gpu; + paddle::framework::Tensor input2_gpu; + paddle::framework::Tensor out_gpu; + paddle::framework::Tensor out; + + auto* cpu_place = new paddle::platform::CPUPlace(); + float* input1_ptr = input1.mutable_data({2, 3}, *cpu_place); + float arr[6] = {0, 1, 2, 3, 4, 5}; + memcpy(input1_ptr, arr, 6 * sizeof(float)); + + auto* gpu_place = new paddle::platform::GPUPlace(0); + paddle::platform::DeviceContext* context = + new paddle::platform::CUDADeviceContext(*gpu_place); + + input1_gpu.CopyFrom(input1, *gpu_place); + input2_gpu.CopyFrom(input1, *gpu_place); + + out_gpu.mutable_data({3, 3}, *gpu_place); + + paddle::operators::math::matmul( + input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0, context); + + out.CopyFrom(out_gpu, *cpu_place); + + float* out_ptr = out.data(); + EXPECT_EQ(out_ptr[0], 9); + EXPECT_EQ(out_ptr[1], 12); + EXPECT_EQ(out_ptr[2], 15); + EXPECT_EQ(out_ptr[3], 12); + EXPECT_EQ(out_ptr[4], 17); + EXPECT_EQ(out_ptr[5], 22); + EXPECT_EQ(out_ptr[6], 15); + EXPECT_EQ(out_ptr[7], 22); + EXPECT_EQ(out_ptr[8], 29); +} +#endif diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc index f6abba7ab45728f74dcea1363035a729b2cd1d03..d3d0e55a674587fb04f43f24d0790de4358f035a 100644 --- a/paddle/operators/mean_op.cc +++ b/paddle/operators/mean_op.cc @@ -18,14 +18,14 @@ namespace paddle { namespace operators { class MeanOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(MeanOp, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 1, "Input size of AddOp must be one"); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "Output size of AddOp must be one"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "input should be set"); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0), "output should be set"); - ctx.Output(0)->Resize(framework::make_ddim({1})); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), + "Input of MeanOp must be initialized."); + ctx.Output("Out")->Resize({1}); } }; @@ -34,13 +34,15 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker { 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").IgnoreGradient(); + AddOutput("Out", "The output of mean op").NotInGradient(); AddComment("Mean Operator"); } }; class MeanGradOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(MeanGradOp, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { ctx.Output(framework::GradVarName("X")) @@ -52,9 +54,8 @@ class MeanGradOp : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker); +REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker, mean_grad, ops::MeanGradOp); REGISTER_OP_CPU_KERNEL(mean, ops::MeanKernel); -REGISTER_GRADIENT_OP(mean, mean_grad, ops::MeanGradOp); REGISTER_OP_CPU_KERNEL(mean_grad, ops::MeanGradKernel); diff --git a/paddle/operators/mean_op.h b/paddle/operators/mean_op.h index e8595a14faa7c1b03734f814c78f9cbf1819fbb5..9848af280b62729bef9243052ceae0b7d8f4c6f5 100644 --- a/paddle/operators/mean_op.h +++ b/paddle/operators/mean_op.h @@ -31,14 +31,14 @@ template class MeanKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto input = context.Input(0); - auto output = context.Output(0); + auto* input = context.Input("X"); + auto* output = context.Output("Out"); output->mutable_data(context.GetPlace()); auto X = EigenVector::Flatten(*input); auto y = EigenScalar::From(*output); - auto place = context.GetEigenDevice(); + auto& place = context.GetEigenDevice(); y.device(place) = X.mean(); } @@ -55,9 +55,10 @@ class MeanGradKernel : public framework::OpKernel { IG->mutable_data(context.GetPlace()); T ig_size = (T)framework::product(IG->dims()); + Eigen::DSizes bcast(ig_size); EigenVector::Flatten(*IG).device(context.GetEigenDevice()) = - EigenScalar::From(*OG) / ig_size; + (EigenVector::From(*OG) / ig_size).broadcast(bcast); } }; diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..069fb5e1abc657aa02a50fde352ce88d078c36e1 --- /dev/null +++ b/paddle/operators/minus_op.cc @@ -0,0 +1,87 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/minus_op.h" +#include "paddle/operators/net_op.h" + +namespace paddle { +namespace operators { + +class MinusOp : public framework::OperatorWithKernel { + public: + MinusOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + auto *left_tensor = ctx.Input("X"); + auto *right_tensor = ctx.Input("Y"); + + PADDLE_ENFORCE_EQ( + framework::product(left_tensor->dims()), + framework::product(right_tensor->dims()), + "Minus operator must take two tensor with same num of elements"); + ctx.Output("Out")->Resize(left_tensor->dims()); + } +}; + +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(); + + AddComment(R"DOC(Minus Operator + +Equation: Out = X - Y +)DOC"); + } +}; +template +class MinusGradOp : public NetOp { + 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}}}, + {{"Out", {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); + } +}; + +} // namespace operators +} // namespace paddle + +USE_OP(scale); +USE_NO_KERNEL_OP(identity); +namespace ops = paddle::operators; +REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, minus_grad, + ops::MinusGradOp); +REGISTER_OP_CPU_KERNEL(minus, + ops::MinusKernel); diff --git a/paddle/operators/minus_op.cu b/paddle/operators/minus_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..a8375cc6301b2c1a917299c3933b03226bb72907 --- /dev/null +++ b/paddle/operators/minus_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/minus_op.h" + +REGISTER_OP_GPU_KERNEL( + minus, paddle::operators::MinusKernel); diff --git a/paddle/operators/minus_op.h b/paddle/operators/minus_op.h new file mode 100644 index 0000000000000000000000000000000000000000..6310a4fd5141516cff4fc7acbe1d17913a1b5506 --- /dev/null +++ b/paddle/operators/minus_op.h @@ -0,0 +1,39 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT 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 MinusKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* left_tensor = context.Input("X"); + auto* right_tensor = context.Input("Y"); + auto* out_tensor = context.Output("Out"); + + out_tensor->mutable_data(context.GetPlace()); + auto& dev = context.GetEigenDevice(); + framework::EigenVector::Flatten(*out_tensor).device(dev) = + framework::EigenVector::Flatten(*left_tensor) - + framework::EigenVector::Flatten(*right_tensor); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc index 6115a3f3332dba419b56e74a737627483448a715..710a56a0e8e2d17162d7d000df226f1537104eb9 100644 --- a/paddle/operators/mul_op.cc +++ b/paddle/operators/mul_op.cc @@ -17,24 +17,35 @@ namespace paddle { namespace operators { +using framework::Tensor; + class MulOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(MulOp, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 2, "The mul op must take two inputs"); - auto dim0 = ctx.Input(0)->dims(); - auto dim1 = ctx.Input(1)->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")); + 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")); + + 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); + PADDLE_ENFORCE_EQ( - dim0[1], dim1[0], + x_mat_dims[1], y_mat_dims[0], "First matrix's width must be equal with second matrix's height."); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "The mul op takes only one output"); - ctx.Output(0)->Resize({dim0[0], dim1[1]}); + ctx.Output("Out")->Resize({x_mat_dims[0], y_mat_dims[1]}); } }; @@ -45,6 +56,23 @@ class MulOpMaker : public framework::OpProtoAndCheckerMaker { 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 + `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") + .SetDefault(1) + .EqualGreaterThan(1); + AddAttr( + "y_num_col_dims", + R"DOC(mul_op can take tensors with more than two dimensions as input `Y`, + in that case, tensors will be reshaped to a matrix. Just like input `X`. + )DOC") + .SetDefault(1) + .EqualGreaterThan(1); AddComment(R"DOC( Two Element Mul Operator. @@ -54,12 +82,37 @@ The equation is: Out = X * Y }; class MulOpGrad : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(MulOpGrad, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: - void InferShape(const framework::InferShapeContext &ctx) const override {} - std::string DebugString() const override { - LOG(INFO) << "MulGrad"; - return ""; + 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")); + + auto x_mat_dims = + framework::flatten_to_2d(x_dims, Attr("x_num_col_dims")); + auto y_mat_dims = + framework::flatten_to_2d(y_dims, Attr("y_num_col_dims")); + + PADDLE_ENFORCE_EQ( + x_mat_dims[0], out_dims[0], + "The first dimension of Out@GRAD must equal to the first dimension of " + "the first operand."); + PADDLE_ENFORCE_EQ( + y_mat_dims[1], out_dims[1], + "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); } }; @@ -67,7 +120,7 @@ class MulOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker); -REGISTER_GRADIENT_OP(mul, mul_grad, ops::MulOpGrad); - +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); diff --git a/paddle/operators/mul_op.cu b/paddle/operators/mul_op.cu index 43debbc21a365a15c914e60e151f7782b82080cb..a81444dbe63edeecedc5d822c65ff56c42b5db90 100644 --- a/paddle/operators/mul_op.cu +++ b/paddle/operators/mul_op.cu @@ -16,5 +16,6 @@ #include "paddle/operators/mul_op.h" namespace ops = paddle::operators; - REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel); +REGISTER_OP_GPU_KERNEL(mul_grad, + ops::MulGradKernel); diff --git a/paddle/operators/mul_op.h b/paddle/operators/mul_op.h index ab12631c03453a18fbb067e2d12c2bc332acd567..3c01f868bda8cba488b3403df456d63d6b082fa6 100644 --- a/paddle/operators/mul_op.h +++ b/paddle/operators/mul_op.h @@ -1,7 +1,7 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. + You may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 @@ -13,6 +13,9 @@ limitations under the License. */ #pragma once + +#include "paddle/operators/math/math_function.h" + #include "paddle/framework/eigen.h" #include "paddle/framework/op_registry.h" @@ -28,21 +31,66 @@ template class MulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - Eigen::array, 1> dim_pair = { - {Eigen::IndexPair(1, 0)}}; + const Tensor* x = context.Input("X"); + const Tensor* y = context.Input("Y"); + Tensor* z = context.Output("Out"); + const Tensor x_matrix = + x->dims().size() > 2 + ? framework::ReshapeToMatrix( + *x, context.template Attr("x_num_col_dims")) + : *x; + const Tensor y_matrix = + y->dims().size() > 2 + ? framework::ReshapeToMatrix( + *y, context.template Attr("y_num_col_dims")) + : *y; - auto input0 = context.Input("X"); - auto input1 = context.Input("Y"); - auto output = context.Output(0); - - output->mutable_data(context.GetPlace()); + z->mutable_data(context.GetPlace()); + auto* device_context = + const_cast(context.device_context_); + math::matmul(x_matrix, false, y_matrix, false, 1, z, 0, + device_context); + } +}; - auto X = EigenMatrix::From(*input0); - auto Y = EigenMatrix::From(*input1); - auto Z = EigenMatrix::From(*output); - auto place = context.GetEigenDevice(); +template +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"); + int y_num_col_dims = ctx.template Attr("y_num_col_dims"); + const Tensor* x = ctx.Input("X"); + const Tensor* y = ctx.Input("Y"); + const Tensor x_matrix = + x->dims().size() > 2 ? framework::ReshapeToMatrix(*x, x_num_col_dims) + : *x; + const Tensor y_matrix = + y->dims().size() > 2 ? framework::ReshapeToMatrix(*y, y_num_col_dims) + : *y; + const Tensor* dout = ctx.Input(framework::GradVarName("Out")); - Z.device(place) = X.contract(Y, dim_pair); + Tensor* dx = ctx.Output(framework::GradVarName("X")); + Tensor* dy = ctx.Output(framework::GradVarName("Y")); + auto* device_context = + const_cast(ctx.device_context_); + if (dx) { + dx->mutable_data(ctx.GetPlace()); + Tensor dx_matrix = dx->dims().size() > 2 ? framework::ReshapeToMatrix( + *dx, x_num_col_dims) + : *dx; + // dx = dout * y'. dx: M x K, dout : M x N, y : K x N + math::matmul(*dout, false, y_matrix, true, 1, &dx_matrix, 0, + device_context); + } + if (dy) { + dy->mutable_data(ctx.GetPlace()); + Tensor dy_matrix = dy->dims().size() > 2 ? framework::ReshapeToMatrix( + *dy, y_num_col_dims) + : *dy; + // dy = x' * dout. dy K x N, dout : M x N, x : M x K + math::matmul(x_matrix, true, *dout, false, 1, &dy_matrix, 0, + device_context); + } } }; diff --git a/paddle/operators/net_op.cc b/paddle/operators/net_op.cc index a466c4f30fe87db4ad2a44518e083b57f3cbc2ed..78b5e2767842312722fac3509e843a05fe194559 100644 --- a/paddle/operators/net_op.cc +++ b/paddle/operators/net_op.cc @@ -15,48 +15,45 @@ */ #include "paddle/operators/net_op.h" +#include +#include "paddle/framework/op_registry.h" namespace paddle { namespace operators { +const char NetOp::kAll[] = "all"; + void NetOp::CompleteAddOp(bool calc) { add_op_done_ = true; if (!calc) return; - std::unordered_set input_set; - std::unordered_set output_set; - std::unordered_set temp_output; + std::set input_set; + std::set output_set; for (auto& op : ops_) { - for (auto& ipt : op->inputs_) { - if (!Contains(output_set, ipt)) { // Not other op's output - input_set.insert(ipt); - } else { - temp_output.insert(ipt); + for (auto& ipt : op->Inputs()) { + for (auto& var_name : ipt.second) { + // If input variable has been in output set, then it will be + // added into intermediate_outputs_. Otherwise, it will be + // added into input set. + if (Contains(output_set, var_name)) { + intermediate_outputs_.insert(var_name); + } else { + input_set.insert(var_name); + } } } - for (auto& opt : op->outputs_) { - output_set.insert(opt); - } - } - - inputs_.reserve(input_set.size()); - std::copy(input_set.begin(), input_set.end(), std::back_inserter(inputs_)); - std::sort(inputs_.begin(), inputs_.end()); - - outputs_.reserve(output_set.size()); - std::copy(output_set.begin(), output_set.end(), std::back_inserter(outputs_)); - std::sort(outputs_.begin(), outputs_.end()); - - std::vector tmp_index; - tmp_index.reserve(temp_output.size()); - int output_len = static_cast(outputs_.size()); - for (int i = 0; i < output_len; ++i) { - if (Contains(temp_output, outputs_[i])) { - tmp_index.push_back(i); + for (auto& opt : op->Outputs()) { + for (auto& var_name : opt.second) { + output_set.insert(var_name); + } } } - - attrs_["temporary_index"] = tmp_index; + auto& inputs = inputs_[kAll]; + inputs.reserve(input_set.size()); + std::copy(input_set.begin(), input_set.end(), std::back_inserter(inputs)); + auto& outputs = outputs_[kAll]; + outputs.reserve(output_set.size()); + std::copy(output_set.begin(), output_set.end(), std::back_inserter(outputs)); } std::string NetOp::DebugString() const { @@ -73,5 +70,36 @@ std::string NetOp::DebugString() const { bool NetOp::IsNetOp() const { return true; } +std::vector NetOp::OutputVars(bool has_intermediate) const { + std::vector all; + for (auto& pair : this->outputs_) { + for (auto& var_name : pair.second) { + all.push_back(var_name); + } + } + if (has_intermediate) { + return all; + } + std::vector ret_val; + for (auto& each : all) { + if (!Contains(intermediate_outputs_, each)) { + ret_val.push_back(each); + } + } + return ret_val; +} + +NetOp::NetOp(const std::string& type, const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : framework::OperatorBase(type, inputs, outputs, attrs) {} + +std::unique_ptr NetOp::Clone() const { + PADDLE_ENFORCE( + add_op_done_, + "Must clone a sealed NetOp, invoke Net::CompleteAddOp before clone"); + return std::unique_ptr(new NetOp(*this)); +} + } // namespace operators } // namespace paddle diff --git a/paddle/operators/net_op.h b/paddle/operators/net_op.h index 24c9e61c66933c6be5bf44b3537e00b70a33922f..fcd8134b2c19cae6a4d006a4cd6fe32d2d627c34 100644 --- a/paddle/operators/net_op.h +++ b/paddle/operators/net_op.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include "paddle/framework/framework.pb.h" #include "paddle/framework/op_registry.h" namespace paddle { @@ -35,7 +36,22 @@ namespace operators { */ class NetOp : public framework::OperatorBase { public: - DEFINE_OPERATOR_CTOR(NetOp, framework::OperatorBase) + static const char kAll[]; + NetOp() : framework::OperatorBase("plain_net", {}, {}, {}) {} + + NetOp(const std::string& type, const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs); + + NetOp(const NetOp& o) : framework::OperatorBase(o.type_, {}, {}, o.attrs_) { + this->ops_.reserve(o.ops_.size()); + std::transform( + o.ops_.begin(), o.ops_.end(), std::back_inserter(this->ops_), + [](const std::unique_ptr& op) { + return std::unique_ptr(op->Clone()); + }); + this->CompleteAddOp(); + } /** * Infer all the operators' input and output variables' shapes, will be called @@ -70,21 +86,28 @@ class NetOp : public framework::OperatorBase { return true; } + void AppendOp(const framework::OperatorBase& op) { AppendOp(op.Clone()); } + /** * @brief Add an operator by ptr */ - void AddOp(const std::shared_ptr& op) { - PADDLE_ENFORCE(!add_op_done_, "Cannot AddOp when this network is sealed"); + void AppendOp(std::unique_ptr op) { + PADDLE_ENFORCE(!add_op_done_, + "Cannot AppendOp when this network is sealed"); PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op"); - ops_.push_back(op); + ops_.push_back(std::move(op)); } - void InsertOp(size_t pos, const std::shared_ptr& op) { + void InsertOp(size_t pos, std::unique_ptr op) { PADDLE_ENFORCE(!add_op_done_, "Cannot InsertOp when this network is sealed"); PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op"); PADDLE_ENFORCE_LE(pos, ops_.size(), "Out of range"); - ops_.insert(ops_.begin() + pos, op); + ops_.insert(ops_.begin() + pos, std::move(op)); + } + + void InsertOp(size_t pos, const framework::OperatorBase& op) { + InsertOp(pos, op.Clone()); } void CompleteAddOp(bool calculate = true); @@ -92,11 +115,15 @@ class NetOp : public framework::OperatorBase { std::string DebugString() const override; bool IsNetOp() const override; + std::vector OutputVars(bool has_intermediate) const override; + + std::unique_ptr Clone() const override; - std::vector> ops_; + std::vector> ops_; private: bool add_op_done_{false}; + std::set intermediate_outputs_; template static bool Contains(T container, KeyType key) { diff --git a/paddle/operators/net_op_test.cc b/paddle/operators/net_op_test.cc index 0d5c3de798d0b580860d24ea9a61a6a4ede5d0ab..f2e98ee7a1e14ee739abba01e97608845ce557f4 100644 --- a/paddle/operators/net_op_test.cc +++ b/paddle/operators/net_op_test.cc @@ -12,8 +12,8 @@ static int run_cnt = 0; class TestOp : public framework::OperatorBase { public: - DEFINE_OPERATOR_CTOR(TestOp, framework::OperatorBase) - + 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 { @@ -21,14 +21,6 @@ class TestOp : public framework::OperatorBase { } }; -class EmptyOp : public framework::OperatorBase { - public: - DEFINE_OPERATOR_CTOR(EmptyOp, framework::OperatorBase) - - void InferShape(const Scope& scope) const override {} - void Run(const Scope& scope, const DeviceContext& dev_ctx) const override {} -}; - template void AssertSameVectorWithoutOrder(const std::vector& expected, const std::vector& actual) { @@ -46,46 +38,51 @@ TEST(OpKernel, all) { auto net = std::make_shared(); ASSERT_NE(net, nullptr); - auto op1 = std::make_shared(); - op1->inputs_ = {"x", "w1", "b1"}; - op1->outputs_ = {"y"}; - net->AddOp(op1); - - auto op2 = std::make_shared(); - op2->inputs_ = {"y", "w2", "b2"}; - op2->outputs_ = {"z"}; - net->AddOp(op2); + net->AppendOp(std::unique_ptr( + new TestOp("test", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}}, + {{"Out", {"y"}}}, {}))); + net->AppendOp(std::unique_ptr( + new TestOp("test", {{"X", {"y"}}, {"W", {"w2"}}, {"b", {"b2"}}}, + {{"Out", {"z"}}}, {}))); net->CompleteAddOp(); - AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"}, net->inputs_); - AssertSameVectorWithoutOrder({"y", "z"}, net->outputs_); - auto tmp_idx_iter = net->attrs_.find("temporary_index"); - ASSERT_NE(net->attrs_.end(), tmp_idx_iter); - auto& tmp_idx = boost::get>(tmp_idx_iter->second); - ASSERT_EQ(1UL, tmp_idx.size()); - ASSERT_EQ("y", net->outputs_[tmp_idx[0]]); + AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"}, + net->Inputs(NetOp::kAll)); + AssertSameVectorWithoutOrder({"y", "z"}, net->Outputs(NetOp::kAll)); - Scope scope; - platform::CPUDeviceContext dev_ctx; + auto final_outs = net->OutputVars(false); - net->InferShape(scope); - net->Run(scope, dev_ctx); - ASSERT_EQ(2, infer_shape_cnt); - ASSERT_EQ(2, run_cnt); - ASSERT_THROW(net->AddOp(op2), platform::EnforceNotMet); + ASSERT_EQ(final_outs.size(), 1UL); + ASSERT_EQ(final_outs[0], "z"); } TEST(NetOp, insert_op) { NetOp net; - auto op1 = std::make_shared(); - op1->inputs_ = {"x", "w1", "b1"}; - op1->outputs_ = {"y"}; - net.AddOp(op1); - net.InsertOp(0, op1); + auto op1 = std::unique_ptr( + new framework::NOP("empty", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}}, + {{"Out", {"y"}}}, {})); + net.AppendOp(*op1); + net.InsertOp(0, *op1); ASSERT_EQ(2UL, net.ops_.size()); - net.InsertOp(2, op1); + net.InsertOp(2, std::move(op1)); ASSERT_EQ(3UL, net.ops_.size()); } +TEST(NetOp, Clone) { + NetOp net; + net.AppendOp( + std::unique_ptr(new framework::NOP{"empty", {}, {}, {}})); + net.AppendOp(std::unique_ptr( + new framework::NOP{"empty2", {}, {}, {}})); + net.CompleteAddOp(true); + auto new_net_op = net.Clone(); + ASSERT_NE(new_net_op, nullptr); + ASSERT_TRUE(new_net_op->IsNetOp()); + auto* new_net = static_cast(new_net_op.get()); + ASSERT_EQ(2UL, new_net->ops_.size()); + ASSERT_EQ(new_net->ops_[0]->Type(), "empty"); + ASSERT_EQ(new_net->ops_[1]->Type(), "empty2"); +} + } // namespace operators } // namespace paddle diff --git a/paddle/operators/recurrent_op.cc b/paddle/operators/recurrent_op.cc index 243837420562634c3d99fd0acf234ebd53539735..e826703c60ca82e1fe690eb78c3d4f92981ef3a2 100644 --- a/paddle/operators/recurrent_op.cc +++ b/paddle/operators/recurrent_op.cc @@ -36,15 +36,13 @@ void RecurrentAlgorithm::InferShape(const Scope& scope) const { rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, true /*infer_shape_mode*/); InitMemories(step_scopes[0], true /*infer_shape_mode*/); - Variable* net = scope.FindVar(arg_->step_net); - PADDLE_ENFORCE(net != nullptr, "failed to get step net"); for (size_t i = 0; i < seq_len_; i++) { if (i > 0) { rnn::LinkMemories(step_scopes, arg_->memories, i, -1, true /*infer_shape_mode*/); } - net->GetMutable()->InferShape(*step_scopes[i]); + (*stepnet_)->InferShape(*step_scopes[i]); } rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, true /*infer_shape_mode*/); @@ -56,7 +54,6 @@ void RecurrentAlgorithm::Run(const Scope& scope, rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, false /*infer_shape_mode*/); InitMemories(step_scopes[0], false /*infer_shape_mode*/); - Variable* net = scope.FindVar(arg_->step_net); for (size_t step_id = 0; step_id < seq_len_; step_id++) { // create output alias variables @@ -64,7 +61,7 @@ void RecurrentAlgorithm::Run(const Scope& scope, rnn::LinkMemories(step_scopes, arg_->memories, step_id, -1, false /*infer_shape_mode*/); } - net->GetMutable()->Run(*step_scopes[step_id], dev_ctx); + (*stepnet_)->Run(*step_scopes[step_id], dev_ctx); } rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, false /*infer_shape_mode*/); @@ -78,25 +75,28 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const { auto step_scopes = step_scopes_var->GetMutable>(); // Now all variables in scope must be created outside of op. - auto net_var = scope.FindVar(arg_->step_net); - PADDLE_ENFORCE(net_var != nullptr, "no stepnet called %s in scope", - arg_->step_net); - auto net_op = net_var->GetMutable(); - PADDLE_ENFORCE(!net_op->outputs_.empty(), "net_op has no outputs"); + 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) { auto& step_scope = scope.NewScope(); // create step net's temp inputs - for (auto& input : net_op->inputs_) { + for (auto& input : (*stepnet_)->Inputs()) { // the weight are located in parent scope - if (!step_scope.FindVar(input)) - step_scope.NewVar(input)->GetMutable(); + for (auto& var_name : input.second) { + if (!step_scope.FindVar(var_name)) { + step_scope.NewVar(var_name)->GetMutable(); + } + } } // create stepnet's outputs - for (const auto& output : net_op->outputs_) { - step_scope.NewVar(output); + for (const auto& output : (*stepnet_)->Outputs()) { + for (auto& var_name : output.second) { + step_scope.NewVar(var_name); + } } step_scopes->emplace_back(&step_scope); } @@ -130,11 +130,13 @@ const rnn::ArgumentName RecurrentGradientOp::kArgName{ "inlink@grad", "inlink_alias", "outlink_alias", "memories", "pre_memories", "boot_memories@grad"}; -void RecurrentOp::Init() { - OperatorBase::Init(); - std::unique_ptr arg(new rnn::Argument()); - rnn::InitArgument(kArgName, arg.get(), *this); - alg_.Init(std::move(arg)); +RecurrentOp::RecurrentOp(const std::string& type, + const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) { + rnn::InitArgument(kArgName, &arg_, *this); + alg_.Init(&arg_, &stepnet_); } class RecurrentAlgorithmProtoAndCheckerMaker @@ -147,13 +149,12 @@ class RecurrentAlgorithmProtoAndCheckerMaker // inputs and outputs stored in proto AddInput(name.inlinks, "the inputs that need to be segmented for each step.") - .SetMultiple(); + .AsDuplicable(); AddInput(name.boot_memories, "variables to initialize memories.") - .SetMultiple(); - AddInput(name.step_net, "network shared by all steps."); + .AsDuplicable(); AddOutput(name.outlinks, "the outputs that need to concated for all steps.") - .SetMultiple(); + .AsDuplicable(); AddOutput(name.step_scopes, "step scopes"); // Attributes stored in AttributeMap @@ -172,14 +173,12 @@ void RecurrentGradientAlgorithm::Run( auto step_scopes = GetStepScopes(scope); rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, false /*infer_shape_mode*/); - Variable* net = scope.FindVar(arg_->step_net); - PADDLE_ENFORCE(net != nullptr, "failed to get step net"); 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*/); } - net->GetMutable()->Run(*step_scopes[step_id], dev_ctx); + (*stepnet_)->Run(*step_scopes[step_id], dev_ctx); } LinkBootMemoryGradients(step_scopes[0], false); rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, @@ -211,29 +210,30 @@ void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const { auto step_scopes = GetStepScopes(scope); rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, true /*infer_shape_mode*/); - Variable* net = scope.FindVar(arg_->step_net); - PADDLE_ENFORCE(net != nullptr, "failed to get step net"); 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*/); } - net->GetMutable()->InferShape(*step_scopes[step_id]); + (*stepnet_)->InferShape(*step_scopes[step_id]); } rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, true /*infer_shape_mode*/); LinkBootMemoryGradients(step_scopes[0], true /*infer_shape_mode*/); } -void RecurrentGradientOp::Init() { - OperatorBase::Init(); - std::unique_ptr arg(new rnn::Argument()); - rnn::InitArgument(kArgName, arg.get(), *this); - alg_.Init(std::move(arg)); +RecurrentGradientOp::RecurrentGradientOp( + const std::string& type, const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) { + rnn::InitArgument(kArgName, &arg_, *this); + alg_.Init(&arg_, &stepnet_); } } // namespace operators } // namespace paddle -REGISTER_OP(recurrent_op, paddle::operators::RecurrentOp, - paddle::operators::RecurrentAlgorithmProtoAndCheckerMaker); +REGISTER_OP_WITHOUT_GRADIENT( + recurrent, paddle::operators::RecurrentOp, + paddle::operators::RecurrentAlgorithmProtoAndCheckerMaker); diff --git a/paddle/operators/recurrent_op.h b/paddle/operators/recurrent_op.h index fdd9d005378e63b8d44803fb2b4be83d134c6a5b..1033d657a3a8f96c8b3dae8dd93d3f1f6840b59b 100644 --- a/paddle/operators/recurrent_op.h +++ b/paddle/operators/recurrent_op.h @@ -15,6 +15,7 @@ #pragma once #include "paddle/framework/operator.h" +#include "paddle/operators/net_op.h" #include "paddle/operators/rnn/recurrent_op_utils.h" namespace paddle { @@ -33,7 +34,12 @@ class RecurrentAlgorithm { void Run(const framework::Scope& scope, const platform::DeviceContext& dev_ctx) const; - void Init(std::unique_ptr arg) { arg_ = std::move(arg); } + void Init(rnn::Argument* arg, + std::unique_ptr* stepnet) { + PADDLE_ENFORCE_NOT_NULL(stepnet, "stepnet should be set before."); + arg_ = arg; + stepnet_ = stepnet; + } /** * InferShape must be called before Run. @@ -58,7 +64,8 @@ class RecurrentAlgorithm { void InitMemories(framework::Scope* step_scopes, bool infer_shape_mode) const; private: - std::unique_ptr arg_; + std::unique_ptr* stepnet_; + rnn::Argument* arg_; mutable size_t seq_len_; }; @@ -74,7 +81,12 @@ class RecurrentGradientAlgorithm { * operator. */ public: - void Init(std::unique_ptr arg) { arg_ = std::move(arg); } + void Init(rnn::Argument* arg, + std::unique_ptr* stepnet) { + PADDLE_ENFORCE_NOT_NULL(stepnet, "stepnet should be set before."); + arg_ = std::move(arg); + stepnet_ = stepnet; + } void Run(const framework::Scope& scope, const platform::DeviceContext& dev_ctx) const; @@ -95,15 +107,23 @@ class RecurrentGradientAlgorithm { } private: - std::unique_ptr arg_; + rnn::Argument* arg_; mutable size_t seq_len_; + std::unique_ptr* stepnet_; }; -class RecurrentOp final : public framework::OperatorBase { - DEFINE_OPERATOR_CTOR(RecurrentOp, framework::OperatorBase) +class RecurrentOp : public framework::OperatorBase { public: - void Init() override; - + RecurrentOp(const std::string& type, const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs); + + RecurrentOp(const RecurrentOp& o) + : framework::OperatorBase( + static_cast(o)) { + // TODO(yuyang18): Implement copy ctor well. + PADDLE_THROW("Not implemented"); + } /** * InferShape must be called before Run. */ @@ -116,15 +136,32 @@ class RecurrentOp final : public framework::OperatorBase { alg_.Run(scope, dev_ctx); } + void set_stepnet(std::unique_ptr net) { + stepnet_ = std::move(net); + } + const OperatorBase& stepnet() const { return *stepnet_; } + static const rnn::ArgumentName kArgName; private: RecurrentAlgorithm alg_; + rnn::Argument arg_; + std::unique_ptr stepnet_; }; -class RecurrentGradientOp final : public framework::OperatorBase { +class RecurrentGradientOp : public framework::OperatorBase { public: - void Init() override; + RecurrentGradientOp(const std::string& type, + const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs); + + RecurrentGradientOp(const RecurrentGradientOp& o) + : framework::OperatorBase( + static_cast(o)) { + // TODO(yuyang18): Implement Copy ctor. + PADDLE_THROW("Not Implemented"); + } /** * InferShape must be called before Run. @@ -140,8 +177,15 @@ class RecurrentGradientOp final : public framework::OperatorBase { static const rnn::ArgumentName kArgName; + void set_stepnet(std::unique_ptr net) { + stepnet_ = std::move(net); + } + const OperatorBase& stepnet() const { return *stepnet_; } + private: RecurrentGradientAlgorithm alg_; + std::unique_ptr stepnet_; + rnn::Argument arg_; }; } // namespace operators diff --git a/paddle/operators/recurrent_op_test.cc b/paddle/operators/recurrent_op_test.cc deleted file mode 100644 index 0c9a343415835540c7543f15f40c53b78a6a55c4..0000000000000000000000000000000000000000 --- a/paddle/operators/recurrent_op_test.cc +++ /dev/null @@ -1,398 +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/recurrent_op.h" - -#include -#include - -#include "paddle/framework/ddim.h" -#include "paddle/framework/op_registry.h" -#include "paddle/framework/operator.h" -#include "paddle/framework/tensor.h" -#include "paddle/operators/net_op.h" - -namespace paddle { -namespace operators { - -using framework::make_ddim; -using framework::DDim; -using framework::Tensor; -using framework::Variable; -using framework::Scope; -using framework::OpRegistry; - -class RecurrentOpTest : public ::testing::Test { - protected: - virtual void SetUp() override { - CreateGlobalVariables(); - CreateStepNet(); - CreateRNNOp(); - } - - virtual void TearDown() override {} - - void CreateGlobalVariables() { - // create input, and init content - LOG(INFO) << "create global variable x"; - for (auto inlink : std::vector{"x", "x0", "x1", "h"}) { - Variable* x = scope_.NewVar(inlink); - DDim dims = make_ddim(std::vector{ - 10 /*sent size*/, 20 /*batch size*/, 30 /*input dim*/}); - x->GetMutable()->mutable_data(dims, platform::CPUPlace()); - } - // create output alias just for test - for (auto inlink : std::vector{"h@alias"}) { - Variable* x = scope_.NewVar(inlink); - DDim dims = - make_ddim(std::vector{20 /*batch size*/, 30 /*input dim*/}); - x->GetMutable()->mutable_data(dims, platform::CPUPlace()); - } - - LOG(INFO) << "create global variable w"; - Variable* w = scope_.NewVar("rnn/w"); - w->GetMutable()->mutable_data( - make_ddim(std::vector{30, 30}), platform::CPUPlace()); - - for (auto boot : std::vector{"h_boot"}) { - LOG(INFO) << "create global variable " << boot; - Variable* h_boot = scope_.NewVar(boot); - h_boot->GetMutable()->mutable_data( - make_ddim(std::vector{20 /*batch size*/, 30 /*input dim*/}), - platform::CPUPlace()); - } - - LOG(INFO) << "create variable step_scopes"; - scope_.NewVar("step_scopes"); - - LOG(INFO) << "create variable h"; - scope_.NewVar("h"); - } - - void CreateRNNOp() { - framework::OpDesc op_desc; - - op_desc.set_type("recurrent_op"); - // inlinks 0 - op_desc.add_inputs("x"); - op_desc.add_inputs("x0"); - op_desc.add_inputs("x1"); - // boot_memories 3 - op_desc.add_inputs("h_boot"); - // step net 5 - op_desc.add_inputs("step_net"); - // outlinks 6 - op_desc.add_outputs("h"); - // step scopes 7 - op_desc.add_outputs("step_scopes"); - - auto _input_format = std::vector{ - 0, // in_link - 3, // memories - 4 // step_net - }; - auto input_format = op_desc.add_attrs(); - input_format->set_name("input_format"); - input_format->set_type(paddle::framework::AttrType::INTS); - for (auto i : _input_format) { - input_format->add_ints(i); - } - - auto output_format = op_desc.add_attrs(); - output_format->set_name("output_format"); - output_format->set_type(paddle::framework::AttrType::INTS); - for (auto i : std::vector{0, 1, 2}) { - output_format->add_ints(i); - } - - auto inlink_alias = op_desc.add_attrs(); - inlink_alias->set_name("inlink_alias"); - inlink_alias->set_type(paddle::framework::AttrType::STRINGS); - - auto outlink_alias = op_desc.add_attrs(); - outlink_alias->set_name("outlink_alias"); - outlink_alias->set_type(paddle::framework::AttrType::STRINGS); - - auto pre_memories = op_desc.add_attrs(); - pre_memories->set_name("pre_memories"); - pre_memories->set_type(paddle::framework::AttrType::STRINGS); - - auto memories = op_desc.add_attrs(); - memories->set_name("memories"); - memories->set_type(paddle::framework::AttrType::STRINGS); - - // create inlink_alias - for (const auto& item : - std::vector{"x@alias", "x0@alias", "x1@alias"}) { - inlink_alias->add_strings(item); - } - // pre memories - for (const auto& item : std::vector{"rnn/h@pre"}) { - pre_memories->add_strings(item); - } - // memories - for (const auto& item : std::vector{"rnn/h"}) { - memories->add_strings(item); - } - // output alias - for (const auto& item : std::vector{"h@alias"}) { - outlink_alias->add_strings(item); - } - - rnn_op_ = OpRegistry::CreateOp(op_desc); - - LOG(INFO) << "rnn_op finish init"; - } - - void CreateStepNet() { - LOG(INFO) << "create variable step_net"; - Variable* var = scope_.NewVar("step_net"); - auto net = var->GetMutable(); - net->AddOp( - OpRegistry::CreateOp("mul", {"rnn/h@pre", "rnn/w"}, {"rnn/s"}, {})); - - net->AddOp( - OpRegistry::CreateOp("add_two", {"x@alias", "rnn/s"}, {"rnn/h"}, {})); - net->CompleteAddOp(); - } - - // father scope - Scope scope_; - std::shared_ptr rnn_op_; -}; - -TEST_F(RecurrentOpTest, Run) { - platform::CPUDeviceContext ctx; - rnn_op_->InferShape(scope_); - rnn_op_->Run(scope_, ctx); -} - -class RecurrentGradientAlgorithmTest : public ::testing::Test { - protected: - virtual void SetUp() override { - CreateGlobalVariables(); - CreateStepScopes(); - CreateStepNet(); - CreateRNNGradientAlgorithm(); - - // segment inputs - SegmentInputs(); - // link forward memories - LinkeMemories(); - } - - virtual void TearDown() override {} - - void CreateGlobalVariables() { - // inputs: x - LOG(INFO) << "create global variable x"; - Variable* x = scope_.NewVar("x"); - DDim dims = - make_ddim({10 /*sent size*/, 20 /*batch size*/, 30 /*input dim*/}); - x->GetMutable()->mutable_data(dims, platform::CPUPlace()); - // inputs: h_boot - LOG(INFO) << "create global variable h_boot"; - Variable* h_boot = scope_.NewVar("h_boot"); - h_boot->GetMutable()->mutable_data( - make_ddim({20 /*batch size*/, 30 /*input dim*/}), platform::CPUPlace()); - // inputs: w - LOG(INFO) << "create global variable w"; - Variable* w = scope_.NewVar("rnn/w"); - w->GetMutable()->mutable_data(make_ddim({30, 30}), - platform::CPUPlace()); - // inputs: h_grad - LOG(INFO) << "create variable h_grad"; - Variable* dh = scope_.NewVar("h_grad"); - dh->GetMutable()->mutable_data(make_ddim({10, 20, 30}), - platform::CPUPlace()); - // inputs: step_scopes - LOG(INFO) << "create variable step_scopes"; - scope_.NewVar("step_scopes"); - // inputs: step_net - LOG(INFO) << "create variable step_net"; - scope_.NewVar("step_net"); - // outputs: w_grad - LOG(INFO) << "create global variable w_grad"; - scope_.NewVar("rnn/w_grad"); - // outputs: x_grad - LOG(INFO) << "create global variable x_grad"; - scope_.NewVar("x_grad"); - // outputs: h_boot_grad - LOG(INFO) << "create global variable h_boot_grad"; - scope_.NewVar("h_boot_grad"); - } - - void CreateStepScopes() { - auto step_scopes = - scope_.FindVar("step_scopes")->GetMutable>(); - for (int i = 0; i < 10; ++i) { - auto& scope = scope_.NewScope(); - auto pre_t = scope.NewVar("rnn/pre_h")->GetMutable(); - pre_t->mutable_data({20, 30}, platform::CPUPlace()); - auto tensor = scope.NewVar("rnn/h")->GetMutable(); - tensor->mutable_data({20, 30}, platform::CPUPlace()); - - // for unit test of ConcatOutputs - auto xg = scope.NewVar("rnn/x_grad")->GetMutable(); - xg->mutable_data({20, 30}, platform::CPUPlace()); - - step_scopes->emplace_back(&scope); - } - - // last time step - auto g = (*step_scopes)[9]->NewVar("rnn/h_pre_grad")->GetMutable(); - g->mutable_data({20, 30}, platform::CPUPlace()); - } - - void CreateRNNGradientAlgorithm() { - std::unique_ptr arg(new rnn::Argument()); - arg->step_net = "step_net"; - arg->step_scopes = "step_scopes"; - rnn::Link inlink; - inlink.external = "h_grad"; - inlink.internal = "rnn/h_grad"; - arg->inlinks = std::vector{inlink}; - - rnn::Link outlink; - outlink.external = "x_grad"; - outlink.internal = "rnn/x_grad"; - arg->outlinks = std::vector{outlink}; - - rnn::MemoryAttr mem_attr; - mem_attr.pre_var = "rnn/h_pre_grad"; - mem_attr.var = "rnn/h_grad"; - mem_attr.boot_var = "h_boot_grad"; - arg->memories = std::vector{mem_attr}; - - rnn_grad_algo_.Init(std::move(arg)); - } - - void CreateStepNet() { - LOG(INFO) << "create variable step_net"; - Variable* var = scope_.NewVar("step_net"); - auto net = var->GetMutable(); - net->AddOp(OpRegistry::CreateOp("mul", {"rnn/h_pre", "rnn/w", "rnn/s_grad"}, - {"rnn/h_pre_grad", "rnn/w_grad"}, {})); - - net->AddOp(OpRegistry::CreateOp("add_two", {"rnn/h_grad"}, - {"rnn/x_grad", "rnn/s_grad"}, {})); - net->CompleteAddOp(); - } - - void SegmentInputs() { - LOG(INFO) << "segment inputs"; - std::vector inlinks = {"x"}; - std::vector inlinks_alias = {"rnn/x"}; - - rnn::Link inlink; - inlink.external = "x"; - inlink.internal = "rnn/x"; - auto step_scopes = - scope_.FindVar("step_scopes")->GetMutable>(); - rnn::SegmentInputs(*step_scopes, std::vector{inlink}, 10, - true /*infer_shape_mode*/); - } - - void LinkeMemories() { - LOG(INFO) << "link memories"; - rnn::MemoryAttr mem_attr; - mem_attr.pre_var = "rnn/h_pre"; - mem_attr.var = "rnn/h"; - mem_attr.boot_var = "boot_h"; - std::vector memories; - memories.push_back(mem_attr); - auto step_scopes = - scope_.FindVar("step_scopes")->GetMutable>(); - for (int i = 1; i < 10; ++i) { - rnn::LinkMemories(*step_scopes, memories, i, -1, - true /*infer_shape_mode*/); - } - } - - Scope scope_; - RecurrentGradientAlgorithm rnn_grad_algo_; -}; - -// TEST_F(RecurrentGradientAlgorithmTest, Run) { -// platform::CPUDeviceContext ctx; -// rnn_grad_algo_.Run(scope_, ctx); -// } - -} // namespace operators -} // namespace paddle - -TEST(RecurrentOp, LinkMemories) { - using namespace paddle::framework; - using namespace paddle::platform; - using namespace paddle::operators; - - // create and init step scopes - size_t len = 10; - std::vector step_scopes; - for (size_t i = 0; i < len; ++i) { - auto scope = new Scope(); - scope->NewVar("pre_h"); - auto tensor = scope->NewVar("h")->GetMutable(); - float* data = tensor->mutable_data({15, 20}, CPUPlace()); - for (size_t j = 0; j < 15 * 20; ++j) { - data[j] = rand() * (1. / (double)RAND_MAX); - } - step_scopes.push_back(scope); - } - - // create MemoryAttr - rnn::MemoryAttr mem_attr; - mem_attr.pre_var = "pre_h"; - mem_attr.var = "h"; - mem_attr.boot_var = "boot_h"; - std::vector memories; - memories.push_back(mem_attr); - - for (size_t i = 1; i < len; ++i) { - rnn::LinkMemories(step_scopes, memories, i, -1, false /*infer_shape_mode*/); - } - // check - for (size_t i = 0; i < len - 1; ++i) { - const float* a = - step_scopes[i]->FindVar("h")->GetMutable()->data(); - const float* b = step_scopes[i + 1] - ->FindVar("pre_h") - ->GetMutable() - ->data(); - for (size_t j = 0; j < 15 * 20; ++j) { - ASSERT_FLOAT_EQ(a[j], b[j]); - } - } - - for (int i = len - 2; i >= 0; --i) { - rnn::LinkMemories(step_scopes, memories, i, 1, false /*infer_shape_mode*/); - } - // check - for (int i = len - 2; i >= 0; --i) { - const float* a = - step_scopes[i]->FindVar("pre_h")->GetMutable()->data(); - const float* b = - step_scopes[i + 1]->FindVar("h")->GetMutable()->data(); - for (size_t j = 0; j < 15 * 20; ++j) { - ASSERT_FLOAT_EQ(a[j], b[j]); - } - } - - for (auto s : step_scopes) { - delete s; - } -} - -USE_OP(add_two); -USE_OP(mul); -USE_OP_ITSELF(recurrent_op); diff --git a/paddle/operators/rnn/recurrent_op_utils.cc b/paddle/operators/rnn/recurrent_op_utils.cc index 7e4770630ed2a49214194689aa489e6ab8e476da..97872c67ac99fbf6c9c177d52f1d4069163e8548 100644 --- a/paddle/operators/rnn/recurrent_op_utils.cc +++ b/paddle/operators/rnn/recurrent_op_utils.cc @@ -61,7 +61,7 @@ void ConcatOutputs(const std::vector& step_scopes, 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); + std::vector dims_vec = vectorize(step_dims); dims_vec.insert(dims_vec.begin(), seq_len); output->Resize(f::make_ddim(dims_vec)); } else { @@ -106,11 +106,10 @@ void LinkMemories(const std::vector& scopes, void InitArgument(const ArgumentName& name, Argument* arg, const framework::OperatorBase& op) { - arg->step_net = op.Input(name.step_net); arg->step_scopes = op.Output(name.step_scopes); auto inlinks = op.Inputs(name.inlinks); - auto inlink_alias = op.GetAttr>(name.inlink_alias); + 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()); @@ -122,7 +121,7 @@ void InitArgument(const ArgumentName& name, Argument* arg, } auto outlinks = op.Outputs(name.outlinks); - auto outlink_alias = op.GetAttr>(name.outlink_alias); + 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()); @@ -136,8 +135,8 @@ void InitArgument(const ArgumentName& name, Argument* arg, auto boot_memories = op.Inputs(name.boot_memories); // attributes - auto memories = op.GetAttr>(name.memories); - auto pre_memories = op.GetAttr>(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/rowwise_add_op.cc b/paddle/operators/rowwise_add_op.cc index 402f6340a04d9b423bb16431a99a2f2866d203bc..fa8f0ff1a858143af427b51025279c726f1628e0 100644 --- a/paddle/operators/rowwise_add_op.cc +++ b/paddle/operators/rowwise_add_op.cc @@ -17,26 +17,33 @@ namespace paddle { namespace operators { -class RowWiseAddOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(RowWiseAddOp, framework::OperatorWithKernel) +using framework::Tensor; + +class RowwiseAddOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 2UL, - "Two inputs is needed by rowwise add"); - auto dim0 = ctx.Input(0)->dims(); - auto dim1 = ctx.Input(1)->dims(); - - PADDLE_ENFORCE(dim0.size() == 2, "Input 0 must be matrix"); - PADDLE_ENFORCE(dim1.size() == 1, "The second input must be vector"); - PADDLE_ENFORCE(dim0[1] == dim1[0], "The width of two input must be same"); - PADDLE_ENFORCE(ctx.OutputSize() == 1, "The output size must be 1"); - ctx.Output(0)->Resize(ctx.Input(0)->dims()); + 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 { +class RowwiseAddOpMaker : public framework::OpProtoAndCheckerMaker { public: - RowWiseAddOpMaker(framework::OpProto *proto, + 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"); @@ -49,11 +56,41 @@ for i in xrange(X.shape[0]): )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); +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, ops::RowWiseAddKernel); + rowwise_add_grad, + ops::RowwiseAddGradKernel); diff --git a/paddle/operators/rowwise_add_op.cu b/paddle/operators/rowwise_add_op.cu index 86f80b81228a69ac4c05a4693901570f2b9966e0..4a57f64c890ce99d6060faec6a4a01b107403344 100644 --- a/paddle/operators/rowwise_add_op.cu +++ b/paddle/operators/rowwise_add_op.cu @@ -17,4 +17,7 @@ namespace ops = paddle::operators; REGISTER_OP_GPU_KERNEL( - rowwise_add, ops::RowWiseAddKernel); + rowwise_add, ops::RowwiseAddKernel); +REGISTER_OP_GPU_KERNEL( + rowwise_add_grad, + ops::RowwiseAddGradKernel); diff --git a/paddle/operators/rowwise_add_op.h b/paddle/operators/rowwise_add_op.h index 82e9d70e959441869b958c1241fa5f5beef4c50c..35774b940926f77167b8f19597027e74d3477e5b 100644 --- a/paddle/operators/rowwise_add_op.h +++ b/paddle/operators/rowwise_add_op.h @@ -1,16 +1,16 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 + http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ #pragma once #include "paddle/framework/eigen.h" @@ -28,15 +28,17 @@ template ; template -class RowWiseAddKernel : public framework::OpKernel { +class RowwiseAddKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto out = context.Output(0); + auto out = context.Output("Out"); out->mutable_data(context.GetPlace()); - - auto input = EigenMatrix::From(*context.Input(0)); - auto bias = EigenVector::From(*context.Input(1)); - auto output = EigenMatrix::From(*out); + 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; @@ -47,5 +49,32 @@ class RowWiseAddKernel : public framework::OpKernel { } }; +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 new file mode 100644 index 0000000000000000000000000000000000000000..ea991f683d841b3dc4624a0d8aa3c88367fd3c6d --- /dev/null +++ b/paddle/operators/scale_op.cc @@ -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 "paddle/operators/scale_op.h" +#include "paddle/operators/net_op.h" + +namespace paddle { +namespace operators { + +class ScaleOp : public framework::OperatorWithKernel { + public: + ScaleOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + auto *in = ctx.Input("X"); + auto *out = ctx.Output("Out"); + out->Resize(in->dims()); + } +}; + +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", "The scaling factor of the scale operator.") + .SetDefault(1.0); + } +}; + +// 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 { + 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); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP(scale, ops::ScaleOp, ops::ScaleOpMaker, scale_grad, + ops::ScaleGradOp); +REGISTER_OP_CPU_KERNEL(scale, + ops::ScaleKernel); diff --git a/paddle/operators/scale_op.cu b/paddle/operators/scale_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..63efbe0da8a90dd237d2d692076075339179acf6 --- /dev/null +++ b/paddle/operators/scale_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/scale_op.h" + +REGISTER_OP_GPU_KERNEL( + scale, paddle::operators::ScaleKernel); diff --git a/paddle/operators/scale_op.h b/paddle/operators/scale_op.h new file mode 100644 index 0000000000000000000000000000000000000000..02fbdc52bbf89c9f2acc5eeaa1197e4ccbca9d31 --- /dev/null +++ b/paddle/operators/scale_op.h @@ -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. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { +template +class ScaleKernel : public framework::OpKernel { + public: + virtual void Compute(const framework::ExecutionContext& context) const { + auto* tensor = context.Output("Out"); + auto* in = context.Input("X"); + tensor->mutable_data(in->place()); + + auto scale = static_cast(context.Attr("scale")); + + auto eigen_out = framework::EigenVector::Flatten(*tensor); + auto eigen_in = framework::EigenVector::Flatten(*in); + auto& dev = context.GetEigenDevice(); + eigen_out.device(dev) = scale * eigen_in; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/scatter.h b/paddle/operators/scatter.h new file mode 100644 index 0000000000000000000000000000000000000000..6b542675c291607b35f180123cf42fee6a783a85 --- /dev/null +++ b/paddle/operators/scatter.h @@ -0,0 +1,92 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT 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/ddim.h" +#include "paddle/framework/eigen.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +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]] + * 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) { + // check index of shape 1-D + PADDLE_ENFORCE(index->dims().size() == 1); + int index_size = index->dims()[0]; + + auto src_dims = src->dims(); + auto dst_dims = output->dims(); + + // 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]; + + if (platform::is_cpu_place(place)) { + CPUScatterUpdate(src, index->data(), index_size, output); + } else { + } +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/scatter_op.cc b/paddle/operators/scatter_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..f901edefa22dc9a252e87116df756d04767a7162 --- /dev/null +++ b/paddle/operators/scatter_op.cc @@ -0,0 +1,86 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR 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/scatter_op.h" +#include "paddle/framework/ddim.h" + +namespace paddle { +namespace operators { + +class ScatterOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + PADDLE_ENFORCE_EQ(ctx.Input("Index")->dims().size(), 1, + "Update Index should be 1-D."); + PADDLE_ENFORCE_EQ(ctx.Input("Ref")->dims().size(), + ctx.Input("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], + "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()); + } +}; + +class ScatterGradOp : public framework::OperatorWithKernel { + public: + 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()); + } +}; + +class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ScatterOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Ref", "The source input of scatter op"); + AddInput("Index", + "The index input of scatter op where Ref will be updated"); + 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, + +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); diff --git a/paddle/operators/scatter_op.h b/paddle/operators/scatter_op.h new file mode 100644 index 0000000000000000000000000000000000000000..e9595638a86a4a4536ddad4e6f20fd80a54b1608 --- /dev/null +++ b/paddle/operators/scatter_op.h @@ -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. */ + +#pragma once +#include "gather.h" +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "scatter.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class ScatterOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto *Ref = ctx.Input("Ref"); + auto *Index = ctx.Input("Index"); + auto *Updates = ctx.Input("Updates"); + auto *Out = ctx.Output("Out"); + + // In place output: Out = Ref, Out[Index] += Updates + Out->ShareDataWith(*Ref); + // Apply ScatterUpdate: Out[index] += Updates[:] + ScatterUpdate(ctx.GetPlace(), Updates, Index, Out); + } +}; + +template +class ScatterGradientOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + 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] + Gather(ctx.GetPlace(), dOut, Index, dUpdates); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/scatter_test.cc b/paddle/operators/scatter_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..26fdaff1460a297fa638181641991f732533fe52 --- /dev/null +++ b/paddle/operators/scatter_test.cc @@ -0,0 +1,56 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR 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/scatter.h" +#include "paddle/framework/ddim.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +#include +#include +#include + +TEST(scatter, ScatterUpdate) { + using namespace paddle::framework; + using namespace paddle::platform; + using namespace paddle::operators; + + Tensor* src = new Tensor(); + Tensor* index = new Tensor(); + Tensor* output = new Tensor(); + + float* p_src = nullptr; + int* p_index = nullptr; + p_src = src->mutable_data(make_ddim({1, 4}), CPUPlace()); + p_index = index->mutable_data(make_ddim({1}), CPUPlace()); + + for (size_t i = 0; i < 4; ++i) p_src[i] = float(i); + p_index[0] = 1; + + float* p_output = output->mutable_data(make_ddim({4, 4}), CPUPlace()); + + ScatterUpdate(CPUPlace(), 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)); + for (size_t i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], float(i - 4)); + for (size_t i = 4; i < 8; ++i) + EXPECT_EQ(output->data()[i], float(i - 4)); + for (size_t i = 8; i < 16; ++i) EXPECT_EQ(p_output[i], float(0)); + for (size_t i = 8; i < 16; ++i) EXPECT_EQ(output->data()[i], float(0)); + + delete src; + delete index; + delete output; +} diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc index 5b8093f0f77e0982a7ad25b42b299a6461712630..ad267e7f087943ff3b8326a7baf2ce3955fa51c2 100644 --- a/paddle/operators/sgd_op.cc +++ b/paddle/operators/sgd_op.cc @@ -18,17 +18,15 @@ namespace paddle { namespace operators { class SGDOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(SGDOp, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 2, "Input size of SGDOp must be two"); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "Output size of SGDOp must be one"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "inputs[0] mast be set"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(1), "inputs[1] mast be set"); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0), "outputs[0] mast be set"); - PADDLE_ENFORCE(ctx.Input(0)->dims() == ctx.Input(1)->dims(), - "Two input of SGD Op's dimension must be same."); - ctx.Output(0)->Resize(ctx.Input(0)->dims()); + PADDLE_ENFORCE( + ctx.Input("param")->dims() == ctx.Input("grad")->dims(), + "Two input of SGD Op's dimension must be same."); + ctx.Output("param_out")->Resize(ctx.Input("param")->dims()); } }; @@ -53,6 +51,6 @@ param_out = param - learning_rate * grad; } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(sgd, ops::SGDOp, ops::SGDOpMaker); +REGISTER_OP_WITHOUT_GRADIENT(sgd, ops::SGDOp, ops::SGDOpMaker); REGISTER_OP_CPU_KERNEL(sgd, ops::SGDOpKernel); diff --git a/paddle/operators/sgd_op.h b/paddle/operators/sgd_op.h index bfb449d0b029409eda4177fc7643810ee6a1df3d..f8888f9c362e1c39af42236bb3a23be37aa3ae15 100644 --- a/paddle/operators/sgd_op.h +++ b/paddle/operators/sgd_op.h @@ -30,8 +30,8 @@ class SGDOpKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { auto param = ctx.Input("param"); auto grad = ctx.Input("grad"); - auto param_out = ctx.Output(0); - float lr = ctx.op_.GetAttr("learning_rate"); + auto param_out = ctx.Output("param_out"); + float lr = ctx.Attr("learning_rate"); param_out->mutable_data(ctx.GetPlace()); diff --git a/paddle/operators/sigmoid_op.cc b/paddle/operators/sigmoid_op.cc index a02e2dc39e8f0d3e31c22a5cafeff111d08aa905..761c6de8d4d2150b30b97b58da95da3d5f33db63 100644 --- a/paddle/operators/sigmoid_op.cc +++ b/paddle/operators/sigmoid_op.cc @@ -18,12 +18,12 @@ namespace paddle { namespace operators { class SigmoidOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(SigmoidOp, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 1, "Sigmoid Op only have one input"); - PADDLE_ENFORCE(ctx.OutputSize() == 1, "Sigmoid Op only have one output"); - ctx.Output(0)->Resize(ctx.Input(0)->dims()); + ctx.Output("Y")->Resize(ctx.Input("X")->dims()); } }; @@ -39,10 +39,13 @@ class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker { }; class SigmoidOpGrad : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(SigmoidOpGrad, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - ctx.Output(0)->Resize(ctx.Input(0)->dims()); + ctx.Output(framework::GradVarName("X")) + ->Resize(ctx.Input("Y")->dims()); } }; @@ -50,9 +53,8 @@ class SigmoidOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker); -REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, ops::SigmoidOpGrad); - +REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker, sigmoid_grad, + ops::SigmoidOpGrad); REGISTER_OP_CPU_KERNEL(sigmoid, ops::SigmoidKernel); REGISTER_OP_CPU_KERNEL( diff --git a/paddle/operators/sigmoid_op.h b/paddle/operators/sigmoid_op.h index 7af879b2091e4a7f80a3a64be029394156650c23..b01a9b3f23283471f8846325075719ba0e75ed35 100644 --- a/paddle/operators/sigmoid_op.h +++ b/paddle/operators/sigmoid_op.h @@ -28,8 +28,8 @@ template class SigmoidKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto input = context.Input(0); - auto output = context.Output(0); + auto input = context.Input("X"); + auto output = context.Output("Y"); output->mutable_data(context.GetPlace()); // The clipping is used in Paddle's raw implenmention @@ -37,7 +37,7 @@ class SigmoidKernel : public framework::OpKernel { auto Y = EigenVector::Flatten(*output); auto place = context.GetEigenDevice(); - Y.device(place) = 1.0 / (1.0 + (-1.0 * X).exp()); + Y.device(place) = 1. / (1. + (-X).exp()); } }; diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc index 9b6a679642303a2cb34954ce16b4a5811acf0ec2..7166b2f60be8a6088ab3a81686f7bed1b7181d97 100644 --- a/paddle/operators/softmax_op.cc +++ b/paddle/operators/softmax_op.cc @@ -18,15 +18,13 @@ namespace paddle { namespace operators { class SoftmaxOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(SoftmaxOp, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 1UL, - "Only one input is need for softmax"); - PADDLE_ENFORCE_EQ(ctx.Input("X")->dims().size(), 2UL, - "The input of softmax op must be matrix"); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL, - "Only one output is need for softmax"); + PADDLE_ENFORCE(ctx.Input("X")->dims().size() == 2UL, + "The input of softmax op must be a matrix."); ctx.Output("Y")->Resize(ctx.Input("X")->dims()); } }; @@ -36,28 +34,45 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { SoftmaxOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "input of softmax"); - AddOutput("Y", "output of softmax"); - AddComment("Softmax Op"); + AddInput("X", + "The input tensor of softmax. " + "2-D with shape [batch_size, input_feature_dimensions]."); + AddOutput("Y", "The normalized values with the same shape as X."); + AddComment(R"DOC( +The input of softmax operator is a 2-D tensor with shape N x K (N is the +batch_size, K is the dimension of input feature). The output tensor has the +same shape as the input tensor. + +For each row of the input tensor, the softmax operator squashes the +K-dimensional vector of arbitrary real values to a K-dimensional vector of real +values in the range [0, 1] that add up to 1. Specifically, it computes the +exponential of the given dimension and the sum of exponential values of all +the other dimensions in the K-dimensional vector input. Then the ratio of the +exponential of the given dimension and the sum of exponential values of all +the other dimensions is the output of the softmax operator. + +For each row `i` and each column `j` in input X, we have: + Y[i, j] = exp(X[i, j]) / sum_j(exp(X[i, j])) + +)DOC"); } }; class SoftmaxOpGrad : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(SoftmaxOpGrad, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 3UL, - "Input of SoftmaxOpGrad should be 3, X, Y, YG"); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL, - "Output of SoftmaxOpGrad should be 1"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null"); + 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 not be null"); - PADDLE_ENFORCE(ctx.Input("Y")->dims() == - ctx.Input(framework::GradVarName("Y"))->dims(), - "the shape of Input(0) and Input(1) should be the same"); + "Input(Y@GRAD) should be not null."); + PADDLE_ENFORCE_EQ(ctx.Input("Y")->dims(), + ctx.Input(framework::GradVarName("Y"))->dims(), + "Input(Y) and its gradients should have a same shape."); + ctx.Output(framework::GradVarName("X")) - ->Resize(ctx.Input("Y")->dims()); + ->Resize(ctx.Input("X")->dims()); } }; @@ -66,9 +81,9 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel { namespace ops = paddle::operators; -REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker); +REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker, softmax_grad, + ops::SoftmaxOpGrad); REGISTER_OP_CPU_KERNEL(softmax, ops::SoftmaxKernel); -REGISTER_GRADIENT_OP(softmax, softmax_grad, ops::SoftmaxOpGrad); REGISTER_OP_CPU_KERNEL( softmax_grad, ops::SoftmaxGradKernel); diff --git a/paddle/operators/softmax_op.h b/paddle/operators/softmax_op.h index 4fa6b59540498638c3b7df639ae10a66c0fa1c16..8a3a5ab927c0e2937936fcc973f000d4d95c3dbc 100644 --- a/paddle/operators/softmax_op.h +++ b/paddle/operators/softmax_op.h @@ -28,12 +28,12 @@ template class SoftmaxKernel : 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()); + auto X = context.Input("X"); + auto Y = context.Output("Y"); + Y->mutable_data(context.GetPlace()); - auto logits = EigenMatrix::From(*input); - auto softmax = EigenMatrix::From(*output); + auto logits = EigenMatrix::From(*X); + auto softmax = EigenMatrix::From(*Y); const int kBatchDim = 0; const int kClassDim = 1; diff --git a/paddle/operators/squared_l2_distance_op.cc b/paddle/operators/squared_l2_distance_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..dc30644a5e7e33d4289e48cac093aa5fde7e75e7 --- /dev/null +++ b/paddle/operators/squared_l2_distance_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/squared_l2_distance_op.h" + +namespace paddle { +namespace operators { + +class SquaredL2DistanceOp : 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 of SquaredL2DistanceOp " + "must be initialized."); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), + "Target of SquaredL2DistanceOp " + "must be initialized."); + + auto* x = ctx.Input("X"); + auto x_dims = x->dims(); + auto* y = ctx.Input("Y"); + auto y_dims = y->dims(); + + PADDLE_ENFORCE_EQ(framework::arity(x_dims), framework::arity(y_dims), + "Tensor rank of both SquaredL2DistanceOp's " + "inputs must be same."); + + int rank = framework::arity(x_dims); + PADDLE_ENFORCE_GE(rank, 2, "Tensor rank should be at least equal to 2."); + PADDLE_ENFORCE_EQ(framework::product(x_dims) / x_dims[0], + framework::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({static_cast(x_dims[0]), + static_cast(framework::product(x_dims) / x_dims[0])}); + ctx.Output("Out")->Resize({x_dims[0], 1}); + } +}; + +class SquaredL2DistanceOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SquaredL2DistanceOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of SquaredL2DistanceOp."); + AddInput("Y", "Target of SquaredL2DistanceOp."); + AddOutput("sub_result", + "Buffering substraction result which " + "will be reused in backward.") + .AsIntermediate(); + AddOutput("Out", "Squared l2 distance between input and target."); + AddComment(R"DOC( + SquaredL2DistanceOp will cacluate the squared L2 distance for + input and target. Number of distance value equals to the + first dimension of input. First dimension of target could be equal to + 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. + )DOC"); + } +}; + +class SquaredL2DistanceGradOp : 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"); + auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); + auto x_dims = ctx.Input("X")->dims(); + auto y_dims = ctx.Input("Y")->dims(); + 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); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(squared_l2_distance, ops::SquaredL2DistanceOp, + ops::SquaredL2DistanceOpMaker, squared_l2_distance_grad, + ops::SquaredL2DistanceGradOp); +REGISTER_OP_CPU_KERNEL( + squared_l2_distance, + ops::SquaredL2DistanceKernel); +REGISTER_OP_CPU_KERNEL( + squared_l2_distance_grad, + ops::SquaredL2DistanceGradKernel); diff --git a/paddle/operators/squared_l2_distance_op.cu b/paddle/operators/squared_l2_distance_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..3fe62f1a9cb56722ea544b0fed052ac384e799aa --- /dev/null +++ b/paddle/operators/squared_l2_distance_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/squared_l2_distance_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL( + squared_l2_distance, + ops::SquaredL2DistanceKernel); +REGISTER_OP_GPU_KERNEL( + squared_l2_distance_grad, + ops::SquaredL2DistanceGradKernel); diff --git a/paddle/operators/squared_l2_distance_op.h b/paddle/operators/squared_l2_distance_op.h new file mode 100644 index 0000000000000000000000000000000000000000..ad3347a0b35f3385c5adbcd7ceaa94fe134105e3 --- /dev/null +++ b/paddle/operators/squared_l2_distance_op.h @@ -0,0 +1,123 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT 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 SquaredL2DistanceKernel : 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("sub_result"); + auto* out1 = context.Output("Out"); + + auto in0_dims = in0->dims(); + auto in1_dims = in1->dims(); + + int cols = framework::product(in0_dims) / in0_dims[0]; + // reduce dimensions except the first + auto x = + EigenMatrix::From(*in0, framework::make_ddim({in0_dims[0], cols})); + auto y = + EigenMatrix::From(*in1, framework::make_ddim({in1_dims[0], cols})); + + out0->mutable_data(context.GetPlace()); + out1->mutable_data(context.GetPlace()); + auto sub_result = EigenMatrix::From(*out0); + auto z = EigenVector::Flatten(*out1); + + auto place = context.GetEigenDevice(); + auto x_dims = x.dimensions(); + auto y_dims = y.dimensions(); + // buffer the substraction result + if (y_dims[0] == 1 && x_dims[0] > y_dims[0]) { + sub_result.device(place) = + x - + y.broadcast(Eigen::array({{static_cast(x_dims[0]), 1}})); + } else { + sub_result.device(place) = x - y; + } + auto sub_res_pow2 = sub_result * sub_result; + z.device(place) = sub_res_pow2.sum(Eigen::array({{1}})); + } +}; + +template +class SquaredL2DistanceGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* in0 = context.Input("sub_result"); + auto* in1 = context.Input(framework::GradVarName("Out")); + auto* x_g = context.Output(framework::GradVarName("X")); + auto* y_g = context.Output(framework::GradVarName("Y")); + + auto sub_result = EigenMatrix::From(*in0); + auto out_grad = EigenMatrix::From(*in1); + + auto x_dims = x_g->dims(); + auto y_dims = y_g->dims(); + + int cols = framework::product(x_dims) / x_dims[0]; + // calculate gradient + auto grad_mat = 2 * + (out_grad.broadcast(Eigen::array({{1, cols}}))) * + sub_result; + + // propagate back to input + auto eigen_place = context.GetEigenDevice(); + if (x_g) { + x_g->mutable_data(context.GetPlace()); + // eigen matrix + auto x_grad = + EigenMatrix::From(*x_g, framework::make_ddim({x_dims[0], cols})); + // dimensions are same with subResult + x_grad.device(eigen_place) = grad_mat; + } + + if (y_g) { + y_g->mutable_data(context.GetPlace()); + + PADDLE_ENFORCE_GE(sub_result.dimensions()[0], y_dims[0], + "First dimension of gradient must be greater or " + "equal than first dimension of target."); + + if (sub_result.dimensions()[0] == y_dims[0]) { + auto y_grad = + EigenMatrix::From(*y_g, framework::make_ddim({y_dims[0], cols})); + y_grad.device(eigen_place) = -1 * grad_mat; + } else { + auto col_sum_res = -1 * (grad_mat.sum(Eigen::array({{0}}))); + auto y_grad = EigenVector::Flatten(*y_g); + y_grad.device(eigen_place) = col_sum_res; + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..5805826ee8a555ca6dfc1ca81feaadffea9e1012 --- /dev/null +++ b/paddle/operators/sum_op.cc @@ -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. */ + +#include "paddle/operators/sum_op.h" +#include + +namespace paddle { +namespace operators { +using framework::Tensor; + +class SumOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + auto ins = ctx.MultiInput("X"); + auto *out = ctx.Output("Out"); + int N = ins.size(); + + auto in_dim = ins[0]->dims(); + + PADDLE_ENFORCE_GT(N, 1, "Input tensors count should > 1."); + for (int i = 1; i < N; i++) { + auto dim = ins[i]->dims(); + PADDLE_ENFORCE(in_dim == dim, "Input tensors must have same shape"); + } + out->Resize(in_dim); + } +}; + +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(R"DOC( + Sum the input tensors. + )DOC"); + } +}; + +class SumGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + 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); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(sum, ops::SumOp, ops::SumOpMaker, sum_grad, ops::SumGradOp); +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 new file mode 100644 index 0000000000000000000000000000000000000000..a465cf3659ba7c51338abadfc62962fb6755a39d --- /dev/null +++ b/paddle/operators/sum_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. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/sum_op.h" + +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 new file mode 100644 index 0000000000000000000000000000000000000000..0b1e9ebaa38d455fb5e3ce8c1a39cbbcdad9a940 --- /dev/null +++ b/paddle/operators/sum_op.h @@ -0,0 +1,65 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at +http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#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 SumKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto ins = context.MultiInput("X"); + auto* out = context.Output("Out"); + out->mutable_data(context.GetPlace()); + + auto place = context.GetEigenDevice(); + auto result = EigenVector::Flatten(*out); + + int N = ins.size(); + auto in = EigenVector::Flatten(*(ins[0])); + result.device(place) = in; + for (int i = 1; i < N; i++) { + auto in = EigenVector::Flatten(*(ins[i])); + result.device(place) = result + in; + } + } +}; + +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 new file mode 100644 index 0000000000000000000000000000000000000000..38d2f0a09aec751734864947a2f3cfa20107e22f --- /dev/null +++ b/paddle/operators/top_k_op.cc @@ -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. */ + +#include "paddle/operators/top_k_op.h" + +namespace paddle { +namespace operators { + +class TopkOp : 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 of TopkOP must be initialized."); + auto *input = ctx.Input("X"); + const int k = static_cast(ctx.Attr("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, + "input must have >= k columns"); + + framework::DDim dims = input->dims(); + dims[dims.size() - 1] = k; + ctx.Output("Out")->Resize(dims); + ctx.Output("Indices")->Resize(dims); + } +}; + +class TopkOpMaker : public framework::OpProtoAndCheckerMaker { + public: + TopkOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The input of Topk op"); + AddOutput("Out", "The output tensor of Topk op"); + AddOutput("Indices", "The indices of Topk elements of input"); + AddComment( + R"DOC(If the input is a vector (1d tensor), finds the k largest entries in the vector and outputs their values and indices as vectors. Thus values[j] is the j-th largest entry in input, and its index is indices[j]. + + For matrices, computes the top k entries in each row. )DOC"); + AddAttr("k", + "Number of top elements to look for along the last " + "dimension (along each row for matrices).") + .SetDefault(1); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(top_k, ops::TopkOp, ops::TopkOpMaker); +REGISTER_OP_CPU_KERNEL(top_k, + ops::TopkKernel); diff --git a/paddle/operators/top_k_op.cu b/paddle/operators/top_k_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..afe4d149c53819c45e20353bc9d16393f3f61e0f --- /dev/null +++ b/paddle/operators/top_k_op.cu @@ -0,0 +1,318 @@ +/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/framework/op_registry.h" +#include "paddle/platform/assert.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +struct Pair { + __device__ __forceinline__ Pair() {} + __device__ __forceinline__ Pair(T value, int id) : v(value), id(id) {} + + __device__ __forceinline__ void set(T value, int id) { + v = value; + id = id; + } + + __device__ __forceinline__ void operator=(const Pair& in) { + v = in.v; + id = in.id; + } + + __device__ __forceinline__ bool operator<(const T value) const { + return (v < value); + } + + __device__ __forceinline__ bool operator<(const Pair& in) const { + return (v < in.v) || ((v == in.v) && (id > in.id)); + } + + __device__ __forceinline__ bool operator>(const Pair& in) const { + return (v > in.v) || ((v == in.v) && (id < in.id)); + } + + T v; + int id; +}; + +template +__device__ __forceinline__ void AddTo(Pair topk[], const Pair& p, + int beam_size) { + for (int k = beam_size - 2; k >= 0; k--) { + if (topk[k] < p) { + topk[k + 1] = topk[k]; + } else { + topk[k + 1] = p; + return; + } + } + topk[0] = p; +} + +template +__device__ __forceinline__ void AddTo(Pair topk[], const Pair& p) { + for (int k = beam_size - 2; k >= 0; k--) { + if (topk[k] < p) { + topk[k + 1] = topk[k]; + } else { + topk[k + 1] = p; + return; + } + } + topk[0] = p; +} + +template +__device__ __forceinline__ void GetTopK(Pair topk[], const T* src, int idx, + int dim, int beam_size) { + while (idx < dim) { + if (topk[beam_size - 1] < src[idx]) { + Pair tmp(src[idx], idx); + AddTo(topk, tmp, beam_size); + } + idx += BlockSize; + } +} + +template +__device__ __forceinline__ void GetTopK(Pair topk[], const T* src, int idx, + int dim, const Pair& max, + int beam_size) { + while (idx < dim) { + if (topk[beam_size - 1] < src[idx]) { + Pair tmp(src[idx], idx); + if (tmp < max) { + AddTo(topk, tmp, beam_size); + } + } + idx += BlockSize; + } +} + +template +__device__ __forceinline__ void GetTopK(Pair topk[], const T* val, int* col, + int idx, int dim, int beam_size) { + while (idx < dim) { + if (topk[beam_size - 1] < val[idx]) { + Pair tmp(val[idx], col[idx]); + AddTo(topk, tmp, beam_size); + } + idx += BlockSize; + } +} + +template +__device__ __forceinline__ void GetTopK(Pair topk[], const T* val, int* col, + int idx, int dim, const Pair& max, + int beam_size) { + while (idx < dim) { + if (topk[beam_size - 1] < val[idx]) { + Pair tmp(val[idx], col[idx]); + if (tmp < max) { + AddTo(topk, tmp, beam_size); + } + } + idx += BlockSize; + } +} + +template +__device__ __forceinline__ void ThreadGetTopK(Pair topk[], int& beam, + int beam_size, const T* src, + bool& firstStep, bool& is_empty, + Pair& max, int dim, + const int tid) { + if (beam > 0) { + int length = beam < beam_size ? beam : beam_size; + if (firstStep) { + firstStep = false; + GetTopK(topk, src, tid, dim, length); + } else { + for (int k = 0; k < MaxLength; k++) { + if (k < MaxLength - beam) { + topk[k] = topk[k + beam]; + } else { + topk[k].set(-INFINITY, -1); + } + } + if (!is_empty) { + GetTopK(topk + MaxLength - beam, src, tid, dim, max, + length); + } + } + + max = topk[MaxLength - 1]; + if (max.v == -1) is_empty = true; + beam = 0; + } +} + +template +__device__ __forceinline__ void ThreadGetTopK(Pair topk[], int& beam, + int beam_size, const T* val, + int* col, bool& firstStep, + bool& is_empty, Pair& max, + int dim, const int tid) { + if (beam > 0) { + int length = beam < beam_size ? beam : beam_size; + if (firstStep) { + firstStep = false; + GetTopK(topk, val, col, tid, dim, length); + } else { + for (int k = 0; k < MaxLength; k++) { + if (k < MaxLength - beam) { + topk[k] = topk[k + beam]; + } else { + topk[k].set(-INFINITY, -1); + } + } + if (!is_empty) { + GetTopK(topk + MaxLength - beam, val, col, tid, dim, max, + length); + } + } + + max = topk[MaxLength - 1]; + if (max.v == -1) is_empty = true; + beam = 0; + } +} + +template +__device__ __forceinline__ void BlockReduce(Pair* sh_topk, int* maxid, + Pair topk[], T** topVal, + int** topIds, int& beam, int& k, + const int tid, const int warp) { + while (true) { + __syncthreads(); + if (tid < BlockSize / 2) { + if (sh_topk[tid] < sh_topk[tid + BlockSize / 2]) { + maxid[tid] = tid + BlockSize / 2; + } else { + maxid[tid] = tid; + } + } + __syncthreads(); + for (int stride = BlockSize / 4; stride > 0; stride = stride / 2) { + if (tid < stride) { + if (sh_topk[maxid[tid]] < sh_topk[maxid[tid + stride]]) { + maxid[tid] = maxid[tid + stride]; + } + } + __syncthreads(); + } + __syncthreads(); + + if (tid == 0) { + **topVal = sh_topk[maxid[0]].v; + **topIds = sh_topk[maxid[0]].id; + (*topVal)++; + (*topIds)++; + } + if (tid == maxid[0]) beam++; + if (--k == 0) break; + __syncthreads(); + + if (tid == maxid[0]) { + if (beam < MaxLength) { + sh_topk[tid] = topk[beam]; + } + } + if (maxid[0] / 32 == warp) { + if (__shfl(beam, (maxid[0]) % 32, 32) == MaxLength) break; + } + } +} + +/** + * Each block compute one sample. + * In a block: + * 1. every thread get top MaxLength value; + * 2. merge to sh_topk, block reduce and get max value; + * 3. go to the second setp, until one thread's topk value is null; + * 4. go to the first setp, until get the topk value. + */ +template +__global__ void KeMatrixTopK(T* output, int output_stride, int* indices, + const T* src, int lds, int dim, int k) { + __shared__ Pair sh_topk[BlockSize]; + __shared__ int maxid[BlockSize / 2]; + const int tid = threadIdx.x; + const int warp = threadIdx.x / 32; + output += blockIdx.x * output_stride; + indices += blockIdx.x * k; + + Pair topk[MaxLength]; + int beam = MaxLength; + Pair max; + bool is_empty = false; + bool firststep = true; + + for (int k = 0; k < MaxLength; k++) { + topk[k].set(-INFINITY, -1); + } + while (k) { + ThreadGetTopK(topk, beam, k, + src + blockIdx.x * lds, firststep, + is_empty, max, dim, tid); + + sh_topk[tid] = topk[0]; + BlockReduce(sh_topk, maxid, topk, &output, + &indices, beam, k, tid, warp); + } +} + +template +class TopkOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use GPUPlace."); + auto* input = ctx.Input("X"); + auto* output = ctx.Output("Out"); + auto* indices = ctx.Output("Indices"); + size_t k = static_cast(ctx.Attr("k")); + + const T* input_data = input->data(); + + T* output_data = output->mutable_data(ctx.GetPlace()); + // FIXME(typhoonzero): data is always converted to type T? + int* indices_data = indices->mutable_data(ctx.GetPlace()); + + size_t input_height = input->dims()[0]; + size_t input_width = input->dims()[1]; + if (k > input_width) k = input_width; + + // 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)); + } +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OP_GPU_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel); diff --git a/paddle/operators/top_k_op.h b/paddle/operators/top_k_op.h new file mode 100644 index 0000000000000000000000000000000000000000..ef66acc1d569282a42be64b7a5e90f3fbdb20690 --- /dev/null +++ b/paddle/operators/top_k_op.h @@ -0,0 +1,76 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +using EigenMatrix = framework::EigenMatrix; + +template +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 + // FIXME: only deal with matrix(2d tensor). + auto* input = ctx.Input("X"); + auto* output = ctx.Output("Out"); + auto* indices = ctx.Output("Indices"); + // k is determined by Attr + const size_t k = static_cast(ctx.Attr("k")); + + T* output_data = output->mutable_data(ctx.GetPlace()); + T* indices_data = indices->mutable_data(ctx.GetPlace()); + + auto eg_input = EigenMatrix::From(*input); + + // reshape input to a flattern matrix(like flat_inner_dims) + framework::DDim inputdims = input->dims(); + const size_t row = framework::product( + framework::slice_ddim(inputdims, 0, inputdims.size() - 1)); + const size_t col = inputdims[inputdims.size() - 1]; + Eigen::DSizes flat2dims(row, col); + // NOTE: eigen shape doesn't affect paddle tensor. + eg_input.reshape(flat2dims); + + for (size_t i = 0; i < row; i++) { + std::vector> vec; + for (size_t j = 0; j < col; j++) { + vec.push_back(std::pair(eg_input(i, j), j)); + } + + std::partial_sort( + vec.begin(), vec.begin() + k, vec.end(), + [](const std::pair& l, const std::pair& r) { + return l.first > r.first; + }); + for (size_t j = 0; j < k; j++) { + output_data[i * k + j] = vec[j].first; + indices_data[i * k + j] = vec[j].second; + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc index ea81ec053f8b9029114f7c98d292a778dc50c3e4..f2aeef6c310df8535e67fa3906301a87f8ec4694 100644 --- a/paddle/operators/uniform_random_op.cc +++ b/paddle/operators/uniform_random_op.cc @@ -1,11 +1,8 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. @@ -27,33 +24,40 @@ template class CPUUniformRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* tensor = context.Output(0); + auto* tensor = context.Output("Out"); T* data = tensor->mutable_data(context.GetPlace()); - unsigned int seed = - static_cast(context.op_.GetAttr("seed")); + unsigned int seed = static_cast(context.Attr("seed")); std::minstd_rand engine; if (seed == 0) { seed = std::random_device()(); } engine.seed(seed); std::uniform_real_distribution dist( - static_cast(context.op_.GetAttr("min")), - static_cast(context.op_.GetAttr("max"))); - for (ssize_t i = 0; i < framework::product(tensor->dims()); ++i) { + static_cast(context.Attr("min")), + static_cast(context.Attr("max"))); + int64_t size = framework::product(tensor->dims()); + for (int64_t i = 0; i < size; ++i) { data[i] = dist(engine); } } }; class UniformRandomOp : public framework::OperatorWithKernel { - DEFINE_OPERATOR_CTOR(UniformRandomOp, framework::OperatorWithKernel) + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE(GetAttr("min") < GetAttr("max"), + PADDLE_ENFORCE(Attr("min") < Attr("max"), "uniform_random's min must less then max"); - auto* tensor = ctx.Output(0); - auto dims = GetAttr>("dims"); - tensor->Resize(framework::make_ddim(dims)); + auto* tensor = ctx.Output("Out"); + 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)); } }; @@ -64,7 +68,6 @@ class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker { : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddOutput("Out", "The output tensor of uniform random op"); AddComment(R"DOC(Uniform random operator. - Used to initialize tensor with uniform random generator. )DOC"); AddAttr>("dims", "the dimension of random tensor"); @@ -79,7 +82,7 @@ Used to initialize tensor with uniform random generator. } // namespace operators } // namespace paddle -REGISTER_OP(uniform_random, paddle::operators::UniformRandomOp, - paddle::operators::UniformRandomOpMaker); +REGISTER_OP_WITHOUT_GRADIENT(uniform_random, paddle::operators::UniformRandomOp, + paddle::operators::UniformRandomOpMaker); REGISTER_OP_CPU_KERNEL(uniform_random, paddle::operators::CPUUniformRandomKernel); diff --git a/paddle/operators/uniform_random_op.cu b/paddle/operators/uniform_random_op.cu index b35ebe7b630be72a5856ec1d3cc32bfaf097aa8a..c2c041b144b6ca1f019f972e1301b756ec1c9301 100644 --- a/paddle/operators/uniform_random_op.cu +++ b/paddle/operators/uniform_random_op.cu @@ -1,11 +1,8 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. @@ -46,16 +43,15 @@ template class GPUUniformRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* tensor = context.Output(0); + auto* tensor = context.Output("Out"); T* data = tensor->mutable_data(context.GetPlace()); - unsigned int seed = - static_cast(context.op_.GetAttr("seed")); + unsigned int seed = static_cast(context.Attr("seed")); if (seed == 0) { std::random_device rd; seed = rd(); } - T min = static_cast(context.op_.GetAttr("min")); - T max = static_cast(context.op_.GetAttr("max")); + T min = static_cast(context.Attr("min")); + T max = static_cast(context.Attr("max")); thrust::counting_iterator index_sequence_begin(0); ssize_t N = framework::product(tensor->dims()); thrust::transform(index_sequence_begin, index_sequence_begin + N, diff --git a/paddle/parameter/Argument.cpp b/paddle/parameter/Argument.cpp index 0547ac93cd183afbcede41d280c6b4b16ed7dab1..8dbef0b22e7b2f14c62586f86e686356b6e9c68e 100644 --- a/paddle/parameter/Argument.cpp +++ b/paddle/parameter/Argument.cpp @@ -186,6 +186,7 @@ void Argument::resizeAndCopyFrom(const Argument& src, resizeAndCopy(strs, src.strs, useGpu, stream); frameWidth = src.frameWidth; frameHeight = src.frameHeight; + frameDepth = src.frameDepth; } int32_t Argument::resizeAndCopyFrom(const Argument& src, @@ -206,6 +207,7 @@ int32_t Argument::resizeAndCopyFrom(const Argument& src, dataId = src.dataId; frameWidth = src.frameWidth; frameHeight = src.frameHeight; + frameDepth = src.frameDepth; if (!src.sequenceStartPositions) { // non-sequence input, copy samples directly @@ -276,17 +278,21 @@ int32_t Argument::resizeAndCopyFrom(const Argument& src, void Argument::concat(const std::vector& args, const std::vector& selectRows, const std::vector& seqStartPos, + const std::vector& copySize, bool useGpu, hl_stream_t stream, PassType passType) { CHECK(!subSequenceStartPositions) << "undefined behavior for subsequence positions"; - size_t batchSize = selectRows.size(); + size_t batchSize = 0; + for (size_t i = 0; i < copySize.size(); ++i) + batchSize += copySize[i] * (seqStartPos[i + 1] - seqStartPos[i]); + auto copyArg = [batchSize, stream](MatrixPtr& dst, MatrixPtr src, - int startRow, - int pos, + int desStartRow, + int srcStartRow, int size, bool useGpu) { if (!src) { @@ -300,14 +306,14 @@ void Argument::concat(const std::vector& args, dst->resize(batchSize, width); } - MatrixPtr tmpMatrix = dst->subMatrix(startRow, size); - tmpMatrix->copyFrom(*src->subMatrix(pos, size), stream); + MatrixPtr tmpMatrix = dst->subMatrix(desStartRow, size); + tmpMatrix->copyFrom(*src->subMatrix(srcStartRow, size), stream); }; auto copyIds = [batchSize, stream](IVectorPtr& dst, const IVectorPtr& src, - int startRow, - int pos, + int desStartRow, + int srcStartRow, int size, bool useGpu) { if (!src) { @@ -315,13 +321,14 @@ void Argument::concat(const std::vector& args, return; } IVector::resizeOrCreate(dst, batchSize, useGpu); - dst->subVec(startRow, size)->copyFrom(*src->subVec(pos, size), stream); + dst->subVec(desStartRow, size) + ->copyFrom(*src->subVec(srcStartRow, size), stream); }; auto copyStrs = [batchSize, stream](SVectorPtr& dst, const SVectorPtr& src, - int startRow, - int pos, + int desStartRow, + int srcStartRow, int size, bool useGpu) { if (!src) { @@ -333,30 +340,31 @@ void Argument::concat(const std::vector& args, } else { dst->resize(batchSize); } - std::copy( - src->begin() + pos, src->begin() + pos + size, dst->begin() + startRow); + std::copy(src->begin() + srcStartRow, + src->begin() + srcStartRow + size, + dst->begin() + desStartRow); }; dataId = args[0].dataId; CHECK_NE(seqStartPos.size(), 0UL); - size_t sampleNum = seqStartPos.size() - 1; - for (size_t i = 0; i < sampleNum; ++i) { + int desStartRow = 0; + for (size_t i = 0; i < copySize.size(); ++i) { int startPos = seqStartPos[i]; int endPos = seqStartPos[i + 1]; CHECK_GE(args.size(), static_cast(endPos - startPos)); for (int j = startPos; j < endPos; ++j) { const Argument& arg = args[j - startPos]; - CHECK_EQ(arg.dataId, dataId) << "Arguments in concat should have" - << " same dataId"; - const int copySize = 1; - const int rowIdx = selectRows[j]; - copyArg(in, arg.in, j, rowIdx, copySize, useGpu); - copyArg(value, arg.value, j, rowIdx, copySize, useGpu); + CHECK_EQ(arg.dataId, dataId) << "Arguments to concatenate should have " + << "the same dataId."; + const int srcStartRow = selectRows[j]; + copyArg(in, arg.in, desStartRow, srcStartRow, copySize[i], useGpu); + copyArg(value, arg.value, desStartRow, srcStartRow, copySize[i], useGpu); if (passType != PASS_TEST) { - copyArg(grad, arg.grad, j, rowIdx, copySize, useGpu); + copyArg(grad, arg.grad, desStartRow, srcStartRow, copySize[i], useGpu); } - copyIds(ids, arg.ids, j, rowIdx, copySize, useGpu); - copyStrs(strs, arg.strs, j, rowIdx, copySize, useGpu); + copyIds(ids, arg.ids, desStartRow, srcStartRow, copySize[i], useGpu); + copyStrs(strs, arg.strs, desStartRow, srcStartRow, copySize[i], useGpu); + desStartRow += copySize[i]; } } ICpuGpuVector::resizeOrCreate( @@ -670,19 +678,29 @@ void Argument::reorganizeSeqInfo( const ICpuGpuVectorPtr seqStartPos, const ICpuGpuVectorPtr subSeqStartPos, std::vector>& reorganizedSeqInfo) { - int* seqStarts = seqStartPos->getMutableData(false); - int* subSeqStarts = subSeqStartPos->getMutableData(false); + CHECK(seqStartPos); + reorganizedSeqInfo.clear(); int seqNum = seqStartPos->getSize() - 1; - reorganizedSeqInfo.resize(seqNum, std::vector()); - int seqIdx = 0; - for (size_t i = 0; i < subSeqStartPos->getSize(); ++i) { - reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]); - if (subSeqStarts[i] == seqStarts[seqIdx + 1]) { - seqIdx++; - if (seqIdx == seqNum) return; + int* seqStarts = seqStartPos->getMutableData(false); + + if (subSeqStartPos) { + int* subSeqStarts = subSeqStartPos->getMutableData(false); + reorganizedSeqInfo.resize(seqNum, std::vector()); + int seqIdx = 0; + for (size_t i = 0; i < subSeqStartPos->getSize(); ++i) { reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]); + if (subSeqStarts[i] == seqStarts[seqIdx + 1]) { + seqIdx++; + if (seqIdx == seqNum) return; + reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]); + } } + } else { + reorganizedSeqInfo.resize(1, std::vector(seqNum + 1, 0)); + memcpy(reorganizedSeqInfo[0].data(), + seqStarts, + sizeof(int) * seqStartPos->getSize()); } } diff --git a/paddle/parameter/Argument.h b/paddle/parameter/Argument.h index d8d7a4398f99a2794c5d25528a7d582f5ed629ba..7b59199dded5b3f5d030e389d8bfcac1668fd127 100644 --- a/paddle/parameter/Argument.h +++ b/paddle/parameter/Argument.h @@ -1,11 +1,8 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. @@ -35,6 +32,7 @@ struct Argument { strs(nullptr), frameHeight(0), frameWidth(0), + frameDepth(0), sequenceStartPositions(nullptr), subSequenceStartPositions(nullptr), cpuSequenceDims(nullptr), @@ -64,6 +62,7 @@ struct Argument { allCount = argument.allCount; frameHeight = argument.frameHeight; frameWidth = argument.frameWidth; + frameDepth = argument.frameDepth; dataId = argument.dataId; } @@ -76,6 +75,7 @@ struct Argument { // A dataBatch includes batchSize frames, one frame maybe not only vector size_t frameHeight; size_t frameWidth; + size_t frameDepth; // If NULL, each position is treated independently. // Otherwise, its size should be #NumberOfSequences + 1. @@ -136,8 +136,10 @@ struct Argument { } size_t getFrameHeight() const { return frameHeight; } size_t getFrameWidth() const { return frameWidth; } + size_t getFrameDepth() const { return frameDepth; } void setFrameHeight(size_t h) { frameHeight = h; } void setFrameWidth(size_t w) { frameWidth = w; } + void setFrameDepth(size_t d) { frameDepth = d; } int64_t getNumSequences() const { return sequenceStartPositions ? sequenceStartPositions->getSize() - 1 @@ -240,6 +242,7 @@ struct Argument { void concat(const std::vector& args, const std::vector& selectRows, const std::vector& seqStartPos, + const std::vector& copySize, bool useGpu, hl_stream_t stream, PassType passType); diff --git a/paddle/parameter/Parameter.cpp b/paddle/parameter/Parameter.cpp index ebe36d49376882fe4c1013e19dcf71f452b3e501..f0311095012d944768d80abe423d4a9bfc0e97f5 100644 --- a/paddle/parameter/Parameter.cpp +++ b/paddle/parameter/Parameter.cpp @@ -48,7 +48,8 @@ Parameter::Parameter(const ParameterConfig& config, bool useGpu, bool doInit) deviceId_(-1), sharedCount_(0), updateCounter_(0), - updated_(false) { + updated_(false), + headerFormat_(PARAM_FORMAT_ORIGINAL) { setID(-1); /* capture uninitialized id */ if (useGpu_ && FLAGS_parallel_nn) { /* gpu environment is specified by device property */ @@ -285,7 +286,7 @@ bool Parameter::save(const std::string& filename) const { bool Parameter::save(std::ostream& s) const { CpuVector vec(*bufs_[PARAMETER_VALUE].get()); Header header; - header.version = kFormatVersion; + header.format = headerFormat_; header.valueSize = sizeof(real); header.size = getSize(); @@ -344,8 +345,9 @@ bool Parameter::load(std::istream& s) { Header header; CHECK(s.read(reinterpret_cast(&header), sizeof(header))) << "Fail to read parameter " << getName(); - CHECK_EQ(header.version, kFormatVersion) << "Incorrect format version: " - << header.version; + CHECK(isHeaderFormatSupported(header.format)) << "Incorrect format version: " + << header.format; + headerFormat_ = header.format; CHECK_EQ(header.size, getSize()) << "The size (" << header.size << ") in the file does not match the size " << "(" << getSize() << ") of the parameter: " << getName(); diff --git a/paddle/parameter/Parameter.h b/paddle/parameter/Parameter.h index 0bac76f068ec22bec52766b43e331fe109a34188..04f12efaac15a21ef54ae71074b6d474e2b66c04 100644 --- a/paddle/parameter/Parameter.h +++ b/paddle/parameter/Parameter.h @@ -34,6 +34,20 @@ limitations under the License. */ namespace paddle { +typedef enum { + /// The paddle original basic format + PARAM_FORMAT_ORIGINAL = 0, + + /// See mkldnn_memory_format_t in + /// https://github.com/01org/mkl-dnn/blob/master/include/mkldnn_types.h + /// for a detailed description. + /// 2D weights tensor in the format (output channels, input channels). + PARAM_FORMAT_MKLDNN_OI, + + /// The total format items numbers + PARAM_FORMAT_ITEMS, +} PARAM_FORMAT; + class SparsePrefetchRowCpuMatrix; class Parameter; @@ -51,7 +65,10 @@ public: size_t getSize() const { return config_.size(); } bool isFullSize() const { - return this->getSize() == bufs_[PARAMETER_VALUE]->getSize(); + if (bufs_[PARAMETER_VALUE]) { + return this->getSize() == bufs_[PARAMETER_VALUE]->getSize(); + } + return false; } inline bool useGpu() const { return useGpu_; } @@ -242,14 +259,34 @@ public: /// Initialize the value to 0 void zeroMem(); - static const int kFormatVersion = 0; /// file header structure struct Header { - int32_t version; // = 0, file format version + int32_t format; // = PARAM_FORMAT uint32_t valueSize; // = sizeof(real) uint64_t size; // = getSize() }; + /** + * @brief Is the header format supported. + */ + static bool isHeaderFormatSupported(int32_t fmt) { + return fmt < PARAM_FORMAT_ITEMS; + } + + /** + * @brief Get the format in header. + */ + int getHeaderFormat() { return headerFormat_; } + + /** + * @brief Set the format in header. + */ + void setHeaderFormat(int32_t fmt) { + CHECK(isHeaderFormatSupported(fmt)) << "Unsupported format version: " + << fmt; + headerFormat_ = fmt; + } + /** * @brief Parameter Update Hook. * @@ -321,6 +358,9 @@ protected: bool updated_; SparseFormat format_; + /// The header format for saving or loading param + int32_t headerFormat_; + std::vector> updaterHooks_; public: diff --git a/paddle/platform/CMakeLists.txt b/paddle/platform/CMakeLists.txt index 4154aad15c39119e2f155cb2c7b5177b5aa78022..17bdac8749e31565b119b2cb84aed199fac0f441 100644 --- a/paddle/platform/CMakeLists.txt +++ b/paddle/platform/CMakeLists.txt @@ -1,7 +1,7 @@ cc_library(cpu_info SRCS cpu_info.cc DEPS gflags glog) cc_test(cpu_info_test SRCS cpu_info_test.cc DEPS cpu_info) -nv_library(gpu_info SRCS gpu_info.cc DEPS gflags) +nv_library(gpu_info SRCS gpu_info.cc DEPS gflags glog) cc_library(place SRCS place.cc) cc_test(place_test SRCS place_test.cc DEPS place glog gflags) @@ -9,6 +9,7 @@ cc_test(place_test SRCS place_test.cc DEPS place glog gflags) add_subdirectory(dynload) cc_test(enforce_test SRCS enforce_test.cc DEPS stringpiece) +cc_test(environment_test SRCS environment_test.cc DEPS stringpiece) IF(WITH_GPU) set(GPU_CTX_DEPS dynload_cuda dynamic_loader) @@ -16,5 +17,10 @@ ELSE() set(GPU_CTX_DEPS) ENDIF() -cc_library(device_context SRCS device_context.cc DEPS place eigen3 ${GPU_CTX_DEPS}) +# memcpy deoends on device_context, here add deps individually for +# avoiding cycle dependencies +cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator + system_allocator memory_block meta_data meta_cache place eigen3 ${GPU_CTX_DEPS}) nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info) + +nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda) diff --git a/paddle/platform/cuda_helper.h b/paddle/platform/cuda_helper.h new file mode 100644 index 0000000000000000000000000000000000000000..6feec0d7f8bd5d32d9e5eedee962fcbeff655f1c --- /dev/null +++ b/paddle/platform/cuda_helper.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 + +namespace paddle { +namespace platform { + +#define CUDA_ATOMIC_WRAPPER(op, T) \ + __device__ __forceinline__ T CudaAtomic##op(T* address, const T val) + +#define USE_CUDA_ATOMIC(op, T) \ + CUDA_ATOMIC_WRAPPER(op, T) { return atomic##op(address, val); } + +// For atomicAdd. +USE_CUDA_ATOMIC(Add, float); + +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 600 +USE_CUDA_ATOMIC(Add, double); +#else +CUDA_ATOMIC_WRAPPER(Add, double) { + unsigned long long int* address_as_ull = + reinterpret_cast(address); + unsigned long long int old = *address_as_ull, assumed; + + do { + assumed = old; + old = atomicCAS(address_as_ull, assumed, + __double_as_longlong(val + __longlong_as_double(assumed))); + + // Note: uses integer comparison to avoid hang in case of NaN + } while (assumed != old); + + return __longlong_as_double(old); +} +#endif + +} // namespace platform +} // namespace paddle diff --git a/paddle/platform/cudnn_helper.h b/paddle/platform/cudnn_helper.h new file mode 100644 index 0000000000000000000000000000000000000000..2841d2a2dbec5c17ef098a06c976ca01247820f5 --- /dev/null +++ b/paddle/platform/cudnn_helper.h @@ -0,0 +1,201 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT 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/dynload/cudnn.h" +#include "paddle/platform/enforce.h" +#include "paddle/platform/macros.h" + +namespace paddle { +namespace platform { + +enum class DataLayout { + kNHWC, + kNCHW, + kNCHW_VECT_C, +}; + +enum class PoolingMode { + kMaximum, + kAverage, +}; + +template +class CudnnDataType; + +template <> +class CudnnDataType { + public: + static const cudnnDataType_t type = CUDNN_DATA_FLOAT; +}; + +template <> +class CudnnDataType { + public: + static const cudnnDataType_t type = CUDNN_DATA_DOUBLE; +}; + +inline cudnnTensorFormat_t GetCudnnTensorFormat(const DataLayout& order) { + switch (order) { + case DataLayout::kNHWC: + return CUDNN_TENSOR_NHWC; + case DataLayout::kNCHW: + return CUDNN_TENSOR_NCHW; + default: + PADDLE_THROW("Unknown cudnn equivalent for order"); + } + return CUDNN_TENSOR_NCHW; +} + +class ScopedTensorDescriptor { + public: + ScopedTensorDescriptor() { + PADDLE_ENFORCE(dynload::cudnnCreateTensorDescriptor(&desc_)); + } + ~ScopedTensorDescriptor() { + PADDLE_ENFORCE(dynload::cudnnDestroyTensorDescriptor(desc_)); + } + + inline cudnnTensorDescriptor_t descriptor(const cudnnTensorFormat_t format, + const cudnnDataType_t type, + const std::vector& dims) { + // the format is not used now, but it maybe useful feature + std::vector strides(dims.size()); + strides[dims.size() - 1] = 1; + for (int i = dims.size() - 2; i >= 0; i--) { + strides[i] = dims[i + 1] * strides[i + 1]; + } + PADDLE_ENFORCE(dynload::cudnnSetTensorNdDescriptor( + desc_, type, dims.size(), dims.data(), strides.data())); + return desc_; + } + + template + inline cudnnTensorDescriptor_t descriptor(const DataLayout& order, + const std::vector& dims) { + return descriptor(GetCudnnTensorFormat(order), CudnnDataType::type, + dims); + } + + private: + cudnnTensorDescriptor_t desc_; + DISABLE_COPY_AND_ASSIGN(ScopedTensorDescriptor); +}; + +class ScopedFilterDescriptor { + public: + ScopedFilterDescriptor() { + PADDLE_ENFORCE(dynload::cudnnCreateFilterDescriptor(&desc_)); + } + ~ScopedFilterDescriptor() { + PADDLE_ENFORCE(dynload::cudnnDestroyFilterDescriptor(desc_)); + } + + inline cudnnFilterDescriptor_t descriptor(const cudnnTensorFormat_t format, + const cudnnDataType_t type, + const std::vector& kernel) { + // filter layout: output input spatial_dim_y spatial_dim_x + PADDLE_ENFORCE(dynload::cudnnSetFilterNdDescriptor( + desc_, type, format, kernel.size(), kernel.data())); + return desc_; + } + + template + inline cudnnFilterDescriptor_t descriptor(const DataLayout& order, + const std::vector& kernel) { + return descriptor(GetCudnnTensorFormat(order), CudnnDataType::type, + kernel); + } + + private: + cudnnFilterDescriptor_t desc_; + DISABLE_COPY_AND_ASSIGN(ScopedFilterDescriptor); +}; + +class ScopedConvolutionDescriptor { + public: + ScopedConvolutionDescriptor() { + PADDLE_ENFORCE(dynload::cudnnCreateConvolutionDescriptor(&desc_)); + } + ~ScopedConvolutionDescriptor() { + PADDLE_ENFORCE(dynload::cudnnDestroyConvolutionDescriptor(desc_)); + } + + inline cudnnConvolutionDescriptor_t descriptor( + cudnnDataType_t type, const std::vector& pads, + const std::vector& strides, const std::vector& dilations) { + PADDLE_ENFORCE_EQ(pads.size(), strides.size()); + PADDLE_ENFORCE_EQ(pads.size(), dilations.size()); + +#if CUDNN_VERSION < 6000 + // cudnn v5 does not support dilation conv, the argument is called upscale + // instead of dilations and it is must be one. + for (size_t i = 0; i < dilations.size(); ++i) { + PADDLE_ENFORCE_EQ( + dilations[i], 1, + "Dilations conv is not supported in this cuDNN version"); + } +#endif + + PADDLE_ENFORCE(dynload::cudnnSetConvolutionNdDescriptor( + desc_, pads.size(), pads.data(), strides.data(), dilations.data(), + CUDNN_CROSS_CORRELATION, type)); + return desc_; + } + + template + inline cudnnConvolutionDescriptor_t descriptor( + const std::vector& pads, const std::vector& strides, + const std::vector& dilations) { + return descriptor(CudnnDataType::type, pads, strides, dilations); + } + + private: + cudnnConvolutionDescriptor_t desc_; + DISABLE_COPY_AND_ASSIGN(ScopedConvolutionDescriptor); +}; + +class ScopedPoolingDescriptor { + public: + ScopedPoolingDescriptor() { + PADDLE_ENFORCE(dynload::cudnnCreatePoolingDescriptor(&desc_)); + } + ~ScopedPoolingDescriptor() { + PADDLE_ENFORCE(dynload::cudnnDestroyPoolingDescriptor(desc_)); + } + + inline cudnnPoolingDescriptor_t descriptor(const PoolingMode& mode, + const std::vector& kernel, + const std::vector& pads, + const std::vector& strides) { + PADDLE_ENFORCE_EQ(kernel.size(), pads.size()); + PADDLE_ENFORCE_EQ(kernel.size(), strides.size()); + PADDLE_ENFORCE(dynload::cudnnSetPoolingNdDescriptor( + desc_, (mode == PoolingMode::kMaximum + ? CUDNN_POOLING_MAX + : CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING), + CUDNN_PROPAGATE_NAN, // Always propagate nans. + kernel.size(), kernel.data(), pads.data(), strides.data())); + return desc_; + } + + private: + cudnnPoolingDescriptor_t desc_; + DISABLE_COPY_AND_ASSIGN(ScopedPoolingDescriptor); +}; + +} // namespace platform +} // namespace paddle diff --git a/paddle/platform/cudnn_helper_test.cc b/paddle/platform/cudnn_helper_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..6bd85ae1ca8b47b203e0321e9d9224d5cfd3a586 --- /dev/null +++ b/paddle/platform/cudnn_helper_test.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/platform/cudnn_helper.h" +#include + +TEST(CudnnHelper, ScopedTensorDescriptor) { + using paddle::platform::ScopedTensorDescriptor; + using paddle::platform::DataLayout; + + ScopedTensorDescriptor tensor_desc; + std::vector shape = {2, 4, 6, 6}; + auto desc = tensor_desc.descriptor(DataLayout::kNCHW, shape); + + cudnnDataType_t type; + int nd; + std::vector dims(4); + std::vector strides(4); + paddle::platform::dynload::cudnnGetTensorNdDescriptor( + desc, 4, &type, &nd, dims.data(), strides.data()); + + EXPECT_EQ(nd, 4); + for (size_t i = 0; i < dims.size(); ++i) { + EXPECT_EQ(dims[i], shape[i]); + } + EXPECT_EQ(strides[3], 1); + EXPECT_EQ(strides[2], 6); + EXPECT_EQ(strides[1], 36); + EXPECT_EQ(strides[0], 144); +} + +TEST(CudnnHelper, ScopedFilterDescriptor) { + using paddle::platform::ScopedFilterDescriptor; + using paddle::platform::DataLayout; + + ScopedFilterDescriptor filter_desc; + std::vector shape = {2, 3, 3}; + auto desc = filter_desc.descriptor(DataLayout::kNCHW, shape); + + cudnnDataType_t type; + int nd; + cudnnTensorFormat_t format; + std::vector kernel(3); + paddle::platform::dynload::cudnnGetFilterNdDescriptor(desc, 3, &type, &format, + &nd, kernel.data()); + + EXPECT_EQ(GetCudnnTensorFormat(DataLayout::kNCHW), format); + EXPECT_EQ(nd, 3); + for (size_t i = 0; i < shape.size(); ++i) { + EXPECT_EQ(kernel[i], shape[i]); + } +} + +TEST(CudnnHelper, ScopedConvolutionDescriptor) { + using paddle::platform::ScopedConvolutionDescriptor; + + ScopedConvolutionDescriptor conv_desc; + std::vector src_pads = {2, 2, 2}; + std::vector src_strides = {1, 1, 1}; + std::vector src_dilations = {1, 1, 1}; + auto desc = conv_desc.descriptor(src_pads, src_strides, src_dilations); + + cudnnDataType_t type; + cudnnConvolutionMode_t mode; + int nd; + std::vector pads(3); + std::vector strides(3); + std::vector dilations(3); + paddle::platform::dynload::cudnnGetConvolutionNdDescriptor( + desc, 3, &nd, pads.data(), strides.data(), dilations.data(), &mode, + &type); + + EXPECT_EQ(nd, 3); + for (size_t i = 0; i < src_pads.size(); ++i) { + EXPECT_EQ(pads[i], src_pads[i]); + EXPECT_EQ(strides[i], src_strides[i]); + EXPECT_EQ(dilations[i], src_dilations[i]); + } + EXPECT_EQ(mode, CUDNN_CROSS_CORRELATION); +} + +TEST(CudnnHelper, ScopedPoolingDescriptor) { + using paddle::platform::ScopedPoolingDescriptor; + using paddle::platform::PoolingMode; + + ScopedPoolingDescriptor pool_desc; + std::vector src_kernel = {2, 2, 5}; + std::vector src_pads = {1, 1, 2}; + std::vector src_strides = {2, 2, 3}; + auto desc = pool_desc.descriptor(PoolingMode::kMaximum, src_kernel, src_pads, + src_strides); + + cudnnPoolingMode_t mode; + cudnnNanPropagation_t nan_t = CUDNN_PROPAGATE_NAN; + int nd; + std::vector kernel(3); + std::vector pads(3); + std::vector strides(3); + paddle::platform::dynload::cudnnGetPoolingNdDescriptor( + desc, 3, &mode, &nan_t, &nd, kernel.data(), pads.data(), strides.data()); + + EXPECT_EQ(nd, 3); + for (size_t i = 0; i < src_pads.size(); ++i) { + EXPECT_EQ(kernel[i], src_kernel[i]); + EXPECT_EQ(pads[i], src_pads[i]); + EXPECT_EQ(strides[i], src_strides[i]); + } + EXPECT_EQ(mode, CUDNN_POOLING_MAX); +} diff --git a/paddle/platform/device_context.cc b/paddle/platform/device_context.cc index a928e097787db9deebe1c6eab263190caacac7eb..ad212c5b2c47312743362db4926c80bf056e100d 100644 --- a/paddle/platform/device_context.cc +++ b/paddle/platform/device_context.cc @@ -10,6 +10,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/platform/device_context.h" +#include "paddle/memory/memory.h" namespace paddle { namespace platform { @@ -36,6 +37,59 @@ Place CPUDeviceContext::GetPlace() const { return CPUPlace(); } #ifndef PADDLE_ONLY_CPU +class EigenCudaStreamDevice : public Eigen::StreamInterface { + public: + EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) { + Eigen::initializeDeviceProp(); + } + ~EigenCudaStreamDevice() override {} + + void Reinitialize(const cudaStream_t* cuda_stream, GPUPlace place) { + stream_ = cuda_stream; + place_ = place; + device_prop_ = &Eigen::m_deviceProperties[place.device]; + } + + const cudaStream_t& stream() const override { return *stream_; } + + const cudaDeviceProp& deviceProperties() const override { + return *device_prop_; + } + + void* allocate(size_t num_bytes) const override { + return paddle::memory::Alloc(place_, num_bytes); + } + + void deallocate(void* buffer) const override { + paddle::memory::Free(place_, buffer); + } + + void* scratchpad() const override { + if (scratch_ == NULL) { + scratch_ = allocate(Eigen::kCudaScratchSize + sizeof(unsigned int)); + } + return scratch_; + } + + unsigned int* semaphore() const override { + if (semaphore_ == NULL) { + char* scratch = + static_cast(scratchpad()) + Eigen::kCudaScratchSize; + semaphore_ = reinterpret_cast(scratch); + PADDLE_ENFORCE( + cudaMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_)); + } + return semaphore_; + } + + private: + GPUPlace place_; + const cudaStream_t* stream_; // not owned; + const cudaDeviceProp* device_prop_; // not owned; + mutable void* scratch_; + mutable unsigned int* semaphore_; +}; + template <> Eigen::GpuDevice* DeviceContext::get_eigen_device() const { return reinterpret_cast(this)->eigen_device(); @@ -43,19 +97,9 @@ Eigen::GpuDevice* DeviceContext::get_eigen_device() const { CUDADeviceContext::CUDADeviceContext(GPUPlace place) : place_(place) { SetDeviceId(place_.device); - // TODO(qijun) Pass a created cuda stream to Eigen::CudaStreamDevice directly - // here will cause segment fault. We must implement a class derived from - // Eigen::StreamInterface, and reinitialize it with a cuda stream and a gpu id - // later. Please refer to the implementation of class EigenCudaStreamDevice - // in TensorFlow. - // - // We find that CUDA 7 introduces a new option, the per-thread default stream, - // that has two effects. Please refer to https://devblogs.nvidia.com/ - // parallelforall/gpu-pro-tip-cuda-7-streams-simplify-concurrency/ - // - // So, we decide to use default stream and add –default-stream per-thread nvcc - // flag. Than, two threads with two CUDADeviceContexts will run parallelly. - eigen_stream_.reset(new Eigen::CudaStreamDevice()); + PADDLE_ENFORCE(cudaStreamCreate(&stream_)); + eigen_stream_.reset(new EigenCudaStreamDevice()); + eigen_stream_->Reinitialize(&stream_, place); eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get())); } @@ -70,17 +114,15 @@ CUDADeviceContext::~CUDADeviceContext() { PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_)); } - if (curand_generator_) { - PADDLE_ENFORCE(dynload::curandDestroyGenerator(curand_generator_)); - } eigen_stream_.reset(); eigen_device_.reset(); + PADDLE_ENFORCE(cudaStreamDestroy(stream_)); } Place CUDADeviceContext::GetPlace() const { return place_; } void CUDADeviceContext::Wait() const { - PADDLE_ENFORCE(cudaStreamSynchronize(0)); + PADDLE_ENFORCE(cudaStreamSynchronize(stream_)); } Eigen::GpuDevice* CUDADeviceContext::eigen_device() const { @@ -91,6 +133,7 @@ cublasHandle_t CUDADeviceContext::cublas_handle() { if (!cublas_handle_) { SetDeviceId(place_.device); PADDLE_ENFORCE(dynload::cublasCreate(&cublas_handle_)); + PADDLE_ENFORCE(dynload::cublasSetStream(cublas_handle_, stream_)); } return cublas_handle_; } @@ -99,20 +142,12 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() { if (!cudnn_handle_) { SetDeviceId(place_.device); PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_)); + PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream_)); } return cudnn_handle_; } -curandGenerator_t CUDADeviceContext::curand_generator() { - if (!curand_generator_) { - SetDeviceId(place_.device); - PADDLE_ENFORCE(dynload::curandCreateGenerator(&curand_generator_, - CURAND_RNG_PSEUDO_DEFAULT)); - PADDLE_ENFORCE( - dynload::curandSetPseudoRandomGeneratorSeed(curand_generator_, seed_)); - } - return curand_generator_; -} +cudaStream_t CUDADeviceContext::stream() { return stream_; } #endif // PADDLE_ONLY_CPU diff --git a/paddle/platform/device_context.h b/paddle/platform/device_context.h index 08b5b2cff900cc4239a615fe7d7f6b5faa13510b..11528e1194e4516891034fa8febdac3ba6eed204 100644 --- a/paddle/platform/device_context.h +++ b/paddle/platform/device_context.h @@ -17,7 +17,6 @@ limitations under the License. */ #ifndef PADDLE_ONLY_CPU #include "paddle/platform/dynload/cublas.h" #include "paddle/platform/dynload/cudnn.h" -#include "paddle/platform/dynload/curand.h" #include "paddle/platform/gpu_info.h" #define EIGEN_USE_GPU #endif @@ -40,7 +39,7 @@ class DeviceContext { class CPUDeviceContext : public DeviceContext { public: CPUDeviceContext(); - explicit CPUDeviceContext(CPUPlace); + explicit CPUDeviceContext(CPUPlace place); virtual ~CPUDeviceContext() {} Eigen::DefaultDevice* eigen_device() const; @@ -52,10 +51,11 @@ class CPUDeviceContext : public DeviceContext { }; #ifndef PADDLE_ONLY_CPU +class EigenCudaStreamDevice; class CUDADeviceContext : public DeviceContext { public: - explicit CUDADeviceContext(GPUPlace); + explicit CUDADeviceContext(GPUPlace place); virtual ~CUDADeviceContext(); /*! \brief Wait for all operations completion in the stream. */ @@ -74,24 +74,20 @@ class CUDADeviceContext : public DeviceContext { /*! \brief Return cudnn handle in the device context. */ cudnnHandle_t cudnn_handle(); - /*! \brief Return curand handle in the device context. */ - curandGenerator_t curand_generator(); + /*! \brief Return cuda stream in the device context. */ + cudaStream_t stream(); // clang-format on private: GPUPlace place_; - private: std::unique_ptr eigen_device_; - std::unique_ptr eigen_stream_; - - private: - uint64_t seed_; + std::unique_ptr eigen_stream_; // clang-format off - cudnnHandle_t cudnn_handle_ = nullptr; - cublasHandle_t cublas_handle_ = nullptr; - curandGenerator_t curand_generator_ = nullptr; + cudaStream_t stream_{nullptr}; + cudnnHandle_t cudnn_handle_{nullptr}; + cublasHandle_t cublas_handle_{nullptr}; // clang-format on }; diff --git a/paddle/platform/device_context_test.cc b/paddle/platform/device_context_test.cc index 65345c433c0a328e7f89038a39312edba35eb8c7..5883a55272f0f24c94d48bc43c62ddb7bef15465 100644 --- a/paddle/platform/device_context_test.cc +++ b/paddle/platform/device_context_test.cc @@ -43,8 +43,7 @@ TEST(Device, CUDADeviceContext) { ASSERT_NE(nullptr, cudnn_handle); cublasHandle_t cublas_handle = device_context->cublas_handle(); ASSERT_NE(nullptr, cublas_handle); - curandGenerator_t curand_handle = device_context->curand_generator(); - ASSERT_NE(nullptr, curand_handle); + ASSERT_NE(nullptr, device_context->stream()); delete device_context; } } diff --git a/paddle/platform/dynload/CMakeLists.txt b/paddle/platform/dynload/CMakeLists.txt index d205ead84598e04eea523be32139959a02e0dd83..ceb66f84b6b01892cbaf61c79a47ae60d2589164 100644 --- a/paddle/platform/dynload/CMakeLists.txt +++ b/paddle/platform/dynload/CMakeLists.txt @@ -1,2 +1,2 @@ cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags) -nv_library(dynload_cuda SRCS cublas.cc cudnn.cc curand.cc) +nv_library(dynload_cuda SRCS cublas.cc cudnn.cc curand.cc DEPS dynamic_loader) diff --git a/paddle/platform/dynload/cublas.h b/paddle/platform/dynload/cublas.h index aad8097dbb33cbf6c0f2b4b3efb1376fbe96bc74..9d8343c0b5e200b390ccda760f09816959952e9d 100644 --- a/paddle/platform/dynload/cublas.h +++ b/paddle/platform/dynload/cublas.h @@ -62,12 +62,12 @@ extern void *cublas_dso_handle; DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP(__name) #define CUBLAS_BLAS_ROUTINE_EACH(__macro) \ - __macro(cublasSgemv); \ - __macro(cublasDgemv); \ - __macro(cublasSgemm); \ - __macro(cublasDgemm); \ - __macro(cublasSgeam); \ - __macro(cublasDgeam); \ + __macro(cublasSgemv_v2); \ + __macro(cublasDgemv_v2); \ + __macro(cublasSgemm_v2); \ + __macro(cublasDgemm_v2); \ + __macro(cublasSgeam_v2); \ + __macro(cublasDgeam_v2); \ __macro(cublasCreate_v2); \ __macro(cublasDestroy_v2); \ __macro(cublasSetStream_v2); \ diff --git a/paddle/platform/dynload/cudnn.h b/paddle/platform/dynload/cudnn.h index ef0dd85b083dc2335dd5c70d3dc5f59eda25daeb..0120625b7c14448f1b8deb88c24a3ee06eaf4f01 100644 --- a/paddle/platform/dynload/cudnn.h +++ b/paddle/platform/dynload/cudnn.h @@ -62,19 +62,27 @@ extern void* cudnn_dso_handle; #define CUDNN_DNN_ROUTINE_EACH(__macro) \ __macro(cudnnSetTensor4dDescriptor); \ __macro(cudnnSetTensor4dDescriptorEx); \ + __macro(cudnnSetTensorNdDescriptor); \ + __macro(cudnnGetTensorNdDescriptor); \ __macro(cudnnGetConvolutionNdForwardOutputDim); \ __macro(cudnnGetConvolutionForwardAlgorithm); \ __macro(cudnnCreateTensorDescriptor); \ __macro(cudnnDestroyTensorDescriptor); \ __macro(cudnnCreateFilterDescriptor); \ __macro(cudnnSetFilter4dDescriptor); \ + __macro(cudnnSetFilterNdDescriptor); \ + __macro(cudnnGetFilterNdDescriptor); \ __macro(cudnnSetPooling2dDescriptor); \ + __macro(cudnnSetPoolingNdDescriptor); \ + __macro(cudnnGetPoolingNdDescriptor); \ __macro(cudnnDestroyFilterDescriptor); \ __macro(cudnnCreateConvolutionDescriptor); \ __macro(cudnnCreatePoolingDescriptor); \ __macro(cudnnDestroyPoolingDescriptor); \ __macro(cudnnSetConvolution2dDescriptor); \ __macro(cudnnDestroyConvolutionDescriptor); \ + __macro(cudnnSetConvolutionNdDescriptor); \ + __macro(cudnnGetConvolutionNdDescriptor); \ __macro(cudnnCreate); \ __macro(cudnnDestroy); \ __macro(cudnnSetStream); \ diff --git a/paddle/platform/enforce.h b/paddle/platform/enforce.h index 337a059fb1494d500be0fd2437e59c863ae1563c..64fcbd93b6c4d5d9b36f2636c3ef4f7327f08d25 100644 --- a/paddle/platform/enforce.h +++ b/paddle/platform/enforce.h @@ -14,11 +14,14 @@ limitations under the License. */ #pragma once -#include +#include // for dladdr +#include // for backtrace #include +#include #include #include #include + #include "paddle/string/printf.h" #include "paddle/string/to_string.h" @@ -48,15 +51,29 @@ struct EnforceNotMet : public std::exception { std::rethrow_exception(exp_); } catch (const std::exception& exp) { std::ostringstream sout; + sout << string::Sprintf("%s at [%s:%d]", exp.what(), f, l) << std::endl; - sout << "Call Stacks: " << std::endl; + sout << "PaddlePaddle Call Stacks: " << std::endl; + void* call_stack[TRACE_STACK_LIMIT]; - int sz = backtrace(call_stack, TRACE_STACK_LIMIT); - auto line = backtrace_symbols(call_stack, sz); - for (int i = 0; i < sz; ++i) { - sout << line[i] << std::endl; + auto size = backtrace(call_stack, TRACE_STACK_LIMIT); + auto symbols = backtrace_symbols(call_stack, size); + + Dl_info info; + for (int i = 0; i < size; ++i) { + if (dladdr(call_stack[i], &info)) { + auto demangled = info.dli_sname; + auto addr_offset = static_cast(call_stack[i]) - + static_cast(info.dli_saddr); + sout << string::Sprintf("%-3d %*0p %s + %zd\n", i, + 2 + sizeof(void*) * 2, call_stack[i], + demangled, addr_offset); + } else { + sout << string::Sprintf("%-3d %*0p\n", i, 2 + sizeof(void*) * 2, + call_stack[i]); + } } - free(line); + free(symbols); err_str_ = sout.str(); } } @@ -170,7 +187,7 @@ inline void throw_on_error(T e) { * PADDLE_ENFORCE_EQ(a, b); * * will raise an expression described as follows: - * "enforce a == b failed, 1 != 2" with detailed stack infomation. + * "enforce a == b failed, 1 != 2" with detailed stack information. * * extra messages is also supported, for example: * PADDLE_ENFORCE(a, b, "some simple enforce failed between %d numbers", 2) diff --git a/paddle/platform/environment.h b/paddle/platform/environment.h new file mode 100644 index 0000000000000000000000000000000000000000..4edcce932edc61453cef74f2c4ee0f72496b3677 --- /dev/null +++ b/paddle/platform/environment.h @@ -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. */ + +#pragma once + +#include +#include +#include + +#include "paddle/platform/enforce.h" +#include "paddle/string/piece.h" + +extern char** environ; // for environment variables + +namespace paddle { +namespace platform { + +inline void SetEnvVariable(const std::string& name, const std::string& value) { + PADDLE_ENFORCE_NE(setenv(name.c_str(), value.c_str(), 1), -1, + "Failed to set environment variable %s=%s", name, value); +} + +inline void UnsetEnvVariable(const std::string& name) { + PADDLE_ENFORCE_NE(unsetenv(name.c_str()), -1, + "Failed to unset environment variable %s", name); +} + +inline bool IsEnvVarDefined(const std::string& name) { + return std::getenv(name.c_str()) != nullptr; +} + +inline std::string GetEnvValue(const std::string& name) { + PADDLE_ENFORCE(IsEnvVarDefined(name), + "Tried to access undefined environment variable %s", name); + return std::getenv(name.c_str()); +} + +inline std::vector GetAllEnvVariables() { + std::vector vars; + for (auto var = environ; *var != nullptr; ++var) { + auto tail = string::Index(*var, "="); + auto name = string::SubStr(*var, 0, tail).ToString(); + vars.push_back(name); + } + return vars; +} + +} // namespace platform +} // namespace paddle diff --git a/paddle/platform/environment_test.cc b/paddle/platform/environment_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..5f136527215d6a676cfa1a3b08f09dfd3ab24a90 --- /dev/null +++ b/paddle/platform/environment_test.cc @@ -0,0 +1,54 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/platform/environment.h" + +#include "glog/logging.h" +#include "gtest/gtest.h" + +TEST(ENVIRONMENT, ACCESS) { + namespace platform = paddle::platform; + namespace string = paddle::string; + + platform::SetEnvVariable("PADDLE_USE_ENV", "TRUE"); + + EXPECT_TRUE(platform::IsEnvVarDefined("PADDLE_USE_ENV")); + EXPECT_EQ(platform::GetEnvValue("PADDLE_USE_ENV"), "TRUE"); + + platform::UnsetEnvVariable("PADDLE_USE_ENV"); + EXPECT_FALSE(platform::IsEnvVarDefined("PADDLE_USE_ENV")); + + platform::SetEnvVariable("PADDLE_USE_ENV1", "Hello "); + platform::SetEnvVariable("PADDLE_USE_ENV2", "World, "); + platform::SetEnvVariable("PADDLE_USE_ENV3", "PaddlePaddle!"); + + std::string env_info; + auto vars = platform::GetAllEnvVariables(); + for_each(vars.begin(), vars.end(), [&](const std::string& var) { + env_info += platform::GetEnvValue(var); + }); + + EXPECT_TRUE(string::Contains(env_info, "Hello World, PaddlePaddle!")); + platform::UnsetEnvVariable("PADDLE_USE_ENV1"); + platform::UnsetEnvVariable("PADDLE_USE_ENV2"); + platform::UnsetEnvVariable("PADDLE_USE_ENV3"); + + env_info.clear(); + vars = platform::GetAllEnvVariables(); + for_each(vars.begin(), vars.end(), [&](const std::string& var) { + env_info += platform::GetEnvValue(var); + }); + + EXPECT_FALSE(string::Contains(env_info, "Hello World, PaddlePaddle!")); + EXPECT_FALSE(platform::IsEnvVarDefined("PADDLE_USE_ENV1")); + EXPECT_FALSE(platform::IsEnvVarDefined("PADDLE_USE_ENV2")); + EXPECT_FALSE(platform::IsEnvVarDefined("PADDLE_USE_ENV3")); +} diff --git a/paddle/platform/gpu_info.cc b/paddle/platform/gpu_info.cc index edeb3ecd7bf8b87333813eee5b40f71030f6609f..be381a4e26cf0eb41f5b3de88bd03ad8901683cc 100644 --- a/paddle/platform/gpu_info.cc +++ b/paddle/platform/gpu_info.cc @@ -13,8 +13,11 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/platform/gpu_info.h" + #include "gflags/gflags.h" + #include "paddle/platform/enforce.h" +#include "paddle/platform/environment.h" DEFINE_double(fraction_of_gpu_memory_to_use, 0.95, "Default use 95% of GPU memory for PaddlePaddle," @@ -70,6 +73,13 @@ size_t GpuMaxChunkSize() { GpuMemoryUsage(available, total); + if (IsEnvVarDefined(kEnvFractionGpuMemoryToUse)) { + auto val = std::stod(GetEnvValue(kEnvFractionGpuMemoryToUse)); + PADDLE_ENFORCE_GT(val, 0.0); + PADDLE_ENFORCE_LE(val, 1.0); + FLAGS_fraction_of_gpu_memory_to_use = val; + } + // Reserving the rest memory for page tables, etc. size_t reserving = (1 - FLAGS_fraction_of_gpu_memory_to_use) * total; diff --git a/paddle/platform/gpu_info.h b/paddle/platform/gpu_info.h index d3a5f5f13fdd3dd59eb43465da4a64b0d8d95e5b..ed2420b8740e583d307f6836a70fe7e1c780e28b 100644 --- a/paddle/platform/gpu_info.h +++ b/paddle/platform/gpu_info.h @@ -18,10 +18,15 @@ limitations under the License. */ #include #include +#include namespace paddle { namespace platform { +//! Environment variable: fraction of GPU memory to use on each device. +const std::string kEnvFractionGpuMemoryToUse = + "PADDLE_FRACTION_GPU_MEMORY_TO_USE"; + //! Get the total number of GPU devices in system. int GetDeviceCount(); diff --git a/paddle/framework/attribute.proto b/paddle/platform/macros.h similarity index 65% rename from paddle/framework/attribute.proto rename to paddle/platform/macros.h index 13ae312c10e934566384b8bd0f41dacd6c01fc2f..4a04a38c0c6f905639004dea2f4416ecc57c8620 100644 --- a/paddle/framework/attribute.proto +++ b/paddle/platform/macros.h @@ -12,17 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -syntax = "proto2"; -package paddle.framework; +#pragma once -// Attribute Type for paddle's Op. -// Op contains many attributes. Each type of attributes could be different. -// The AttrType will be shared between AttrDesc and AttrProto. -enum AttrType { - INT = 0; - FLOAT = 1; - STRING = 2; - INTS = 3; - FLOATS = 4; - STRINGS = 5; -} \ No newline at end of file +// 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 +#endif diff --git a/paddle/pserver/LightNetwork.cpp b/paddle/pserver/LightNetwork.cpp index 8616fd2d5aef666f16533fe062f3f40a7a2b202d..4203f2616456244df616ee2109436ab7caef9741 100644 --- a/paddle/pserver/LightNetwork.cpp +++ b/paddle/pserver/LightNetwork.cpp @@ -22,7 +22,6 @@ limitations under the License. */ #include #include -#include #include #include diff --git a/paddle/pserver/ParameterClient2.cpp b/paddle/pserver/ParameterClient2.cpp index f7e391f76324a09c203dfbbb449feb050caa8fb4..54063a809a4f9e558f8d364f5c437f2b6d98925b 100644 --- a/paddle/pserver/ParameterClient2.cpp +++ b/paddle/pserver/ParameterClient2.cpp @@ -65,7 +65,6 @@ void ParameterClient2::initThreads() { LOG(INFO) << "parallel_thread_num dosent need to set"; } syncThreadPool_.reset(new SyncThreadPool(threadNum_)); - startThreads(); } @@ -224,6 +223,14 @@ void ParameterClient2::prepareSendData( request.set_cost(cost); request.set_batch_status(batchStatus); CHECK_EQ(request.blocks_size(), 0); + VLOG(10) << "request: trainer_id: " << request.trainer_id() + << " update_mode" << request.update_mode() + << " send_back_parameter: " << request.send_back_parameter() + << " send_back_parameter_type: " + << request.send_back_parameter_type() + << " num_samples: " << request.num_samples() + << " cost: " << request.cost() + << " batch_status: " << request.batch_status(); } for (const auto& segments : parameterSegments) { const auto it = parameterMap_.find(segments.id); @@ -251,11 +258,17 @@ void ParameterClient2::prepareSendData( CHECK(sendMat != nullptr) << "sendMat is nullptr"; syncThreadPool_->exec([&](int tid, size_t numThreads) { + std::lock_guard guard(sparseAutoGrowthMutex_); const auto& localIndices = prefetchMat->getLocalIndices(); /// num of sparse rows size_t nLocalBlocks = localIndices.size(); uint64_t beginDim = 0; uint64_t endDim = 0; + + // FIXME(typhoonzero): let it resize first + prefetchMat->getLocalRow(nLocalBlocks + 1); + sendMat->getLocalRow(nLocalBlocks + 1); + for (size_t row = 0; row < nLocalBlocks; ++row) { int64_t blockId = localIndices[row]; // local row -> sparse row int serverId = std::abs((blockId + nameHash) % serviceNum_); @@ -275,7 +288,6 @@ void ParameterClient2::prepareSendData( block->set_begin_pos(row * blockSize); /// block len block->set_block_size(endDim - beginDim); - if (sendingPara) { sendJob->parallelInputIovs[serverId].push_back( {sendMat->getLocalRow(row), sizeof(real) * (size_t)blockSize}); diff --git a/paddle/pserver/ParameterClient2.h b/paddle/pserver/ParameterClient2.h index 89b3ddd502151e537b81bdbb09f171dd6e13ba26..29b9eeacddf2945dd22b7b17fc87c7c74b868896 100644 --- a/paddle/pserver/ParameterClient2.h +++ b/paddle/pserver/ParameterClient2.h @@ -583,6 +583,7 @@ protected: #ifndef PADDLE_DISABLE_TIMER uint64_t forwardbackwordTime_; #endif + std::mutex sparseAutoGrowthMutex_; /// map id to parameter used for decoding protobuf data std::unordered_map parameterMap_; diff --git a/paddle/pserver/ParameterServer2.cpp b/paddle/pserver/ParameterServer2.cpp index d7c1d4f788f44c6bfcec040ba24bdc454348c911..54f5c4c0fb4994871edc7a1e52237c9f903ce63b 100644 --- a/paddle/pserver/ParameterServer2.cpp +++ b/paddle/pserver/ParameterServer2.cpp @@ -1032,8 +1032,8 @@ void ParameterServer2::loadValueVector(const LoadValueRequest& request, Parameter::Header header; CHECK(fs.read(reinterpret_cast(&header), sizeof(header))) << "Fail to read parameters in pserver"; - CHECK_EQ(header.version, Parameter::kFormatVersion) - << "Incorrect format version: " << header.version; + CHECK(Parameter::isHeaderFormatSupported(header.format)) + << "Incorrect format version: " << header.format; CHECK_EQ(header.size, (size_t)size_) << "The size (" << header.size << ") in the file does not match the size " << "(" << size_ << ") of the pserver: " << serverId_; @@ -1063,7 +1063,8 @@ void ParameterServer2::saveValueVector(const SaveValueRequest& request, CpuVector& vec = vectors_[PARAMETER_APPLY] ? *vectors_[PARAMETER_APPLY] : *vectors_[PARAMETER_VALUE]; Parameter::Header header; - header.version = Parameter::kFormatVersion; + // TODO(TJ): save param headerFormat_ + header.format = PARAM_FORMAT_ORIGINAL; header.valueSize = sizeof(real); header.size = size_; diff --git a/paddle/pybind/CMakeLists.txt b/paddle/pybind/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..00030050700bfb2cee224124d090b0027d456ba0 --- /dev/null +++ b/paddle/pybind/CMakeLists.txt @@ -0,0 +1,6 @@ +if(WITH_PYTHON) +cc_library(paddle_pybind SHARED + SRCS pybind.cc + DEPS pybind python backward + ${GLOB_OP_LIB}) +endif(WITH_PYTHON) diff --git a/paddle/framework/pybind.cc b/paddle/pybind/pybind.cc similarity index 63% rename from paddle/framework/pybind.cc rename to paddle/pybind/pybind.cc index 75cd5bcb38e1d864358314c1c15b6fb59e9c3752..227b75aff86089d0b21bdae7e6e402292bce67d9 100644 --- a/paddle/framework/pybind.cc +++ b/paddle/pybind/pybind.cc @@ -18,10 +18,11 @@ limitations under the License. */ #include "paddle/framework/backward.h" #include "paddle/framework/op_registry.h" -#include "paddle/framework/tensor_py.h" #include "paddle/operators/net_op.h" +#include "paddle/operators/recurrent_op.h" #include "paddle/platform/enforce.h" #include "paddle/platform/place.h" +#include "paddle/pybind/tensor_py.h" #include "paddle/string/to_string.h" #include "pybind11/numpy.h" #include "pybind11/pybind11.h" @@ -29,59 +30,35 @@ limitations under the License. */ namespace py = pybind11; -USE_OP(add_two); -USE_CPU_OP(onehot_cross_entropy); -USE_NO_GRAD_OP(sgd); +USE_OP(add); +USE_OP(onehot_cross_entropy); +USE_OP(sgd); USE_OP(mul); USE_OP(mean); USE_OP(sigmoid); USE_OP(softmax); USE_OP(rowwise_add); USE_OP(fill_zeros_like); -USE_OP_ITSELF(recurrent_op); +USE_NO_KERNEL_OP(recurrent); USE_OP(gaussian_random); USE_OP(uniform_random); +USE_OP(lookup_table); +USE_OP(scale); +USE_NO_KERNEL_OP(identity); +USE_OP(minus); +USE_OP(cos_sim); +USE_CPU_ONLY_OP(gather); +USE_CPU_ONLY_OP(scatter); +USE_CPU_ONLY_OP(concat); +USE_OP(top_k); +USE_OP(squared_l2_distance); +USE_OP(sum); namespace paddle { namespace framework { using Tensor = framework::Tensor; -template -void ExposeOperator(ClassType &m) { - m.def("infer_shape", &ClassType::type::InferShape) - .def("run", &ClassType::type::Run) - .def("type", - [](const typename ClassType::type &op) -> std::string { - return op.type_; - }) - .def("outputs", - [](const typename ClassType::type &op) -> std::vector { - return op.outputs_; - }) - .def("inputs", - [](const typename ClassType::type &op) -> std::vector { - return op.inputs_; - }) - .def("support_gpu", &ClassType::type::SupportGPU) - .def("temp_outputs", - [](const typename ClassType::type &op) -> std::vector { - auto iter = op.attrs_.find("temporary_index"); - std::vector ret; - if (iter == op.attrs_.end()) { - return ret; - } else { - auto tmp_idx = boost::get>(iter->second); - ret.reserve(tmp_idx.size()); - for (auto &index : tmp_idx) { - ret.push_back(op.outputs_.at(index)); - } - return ret; - } - }) - .def("__str__", &ClassType::type::DebugString); -} - static size_t UniqueIntegerGenerator() { static std::atomic generator; return generator.fetch_add(1); @@ -104,7 +81,7 @@ PYBIND11_PLUGIN(core) { .def("get_dims", [](const Tensor &self) { return vectorize(self.dims()); }) .def("set_dims", - [](Tensor &self, const std::vector &dim) { + [](Tensor &self, const std::vector &dim) { self.Resize(make_ddim(dim)); }) .def("alloc_float", @@ -165,23 +142,24 @@ All parameter, weight, gradient are variables in Paddle. py::return_value_policy::reference) .def("find_var", &Scope::FindVar, py::return_value_policy::reference) .def(py::init<>()) - .def("new_scope", [](Scope &self) -> Scope * { return &self.NewScope(); }, + .def("new_scope", + [](Scope &self) -> Scope * { return &self.NewScope(); }, py::return_value_policy::reference) .def("drop_kids", &Scope::DropKids); //! @note: Be careful! PyBind will return std::string as an unicode, not //! Python str. If you want a str object, you should cast them in Python. m.def("get_all_op_protos", []() -> std::vector { - auto &protos = OpRegistry::protos(); std::vector ret_values; - for (auto it = protos.begin(); it != protos.end(); ++it) { - PADDLE_ENFORCE(it->second.IsInitialized(), - "OpProto must all be initialized"); + + OpInfoMap::Instance().IterAllInfo([&ret_values](const std::string &type, + const OpInfo &info) { + if (!info.HasOpProtoAndChecker()) return; std::string str; - PADDLE_ENFORCE(it->second.SerializeToString(&str), + PADDLE_ENFORCE(info.Proto().SerializeToString(&str), "Serialize OpProto Error. This could be a bug of Paddle."); - ret_values.push_back(py::bytes(str)); - } + ret_values.emplace_back(str); + }); return ret_values; }); m.def_submodule( @@ -215,47 +193,73 @@ All parameter, weight, gradient are variables in Paddle. .def(py::init<>()) .def("__str__", string::to_string); - py::class_> operator_base( - m, "Operator"); - - operator_base.def_static("create", [](py::bytes protobin) { - OpDesc desc; - PADDLE_ENFORCE(desc.ParsePartialFromString(protobin), - "Cannot parse user input to OpDesc"); - PADDLE_ENFORCE(desc.IsInitialized(), - "User OpDesc is not initialized, reason %s", - desc.InitializationErrorString()); - return OpRegistry::CreateOp(desc); - }); - - operator_base.def("backward", - [](const OperatorBase &forwardOp, - const std::unordered_set &no_grad_vars) { - return Backward(forwardOp, no_grad_vars); - }); - - ExposeOperator(operator_base); - - py::class_> net(m, "Net"); + py::class_(m, "Operator") + .def_static("create", + [](py::bytes protobin) { + OpDesc desc; + PADDLE_ENFORCE(desc.ParsePartialFromString(protobin), + "Cannot parse user input to OpDesc"); + PADDLE_ENFORCE(desc.IsInitialized(), + "User OpDesc is not initialized, reason %s", + desc.InitializationErrorString()); + return OpRegistry::CreateOp(desc); + }) + .def("backward", + [](const OperatorBase &forwardOp, + const std::unordered_set &no_grad_vars) { + return Backward(forwardOp, no_grad_vars).release(); + }) + .def("infer_shape", &OperatorBase::InferShape) + .def("run", &OperatorBase::Run) + .def("type", + [](const OperatorBase &op) -> std::string { return op.Type(); }) + .def("outputs", + [](const OperatorBase &op) + -> std::map> { + return op.Outputs(); + }) + .def("output_vars", + [](const OperatorBase &op) { return op.OutputVars(true); }) + .def("inputs", [](const OperatorBase &op) { return op.Inputs(); }) + .def("input_vars", [](const OperatorBase &op) { return op.InputVars(); }) + .def("__str__", &OperatorBase::DebugString) + .def("no_intermediate_outputs", + [](const OperatorBase &op) { return op.OutputVars(false); }) + .def("support_gpu", &OperatorBase::SupportGPU); - net.def_static("create", - []() -> std::shared_ptr { - auto retv = std::make_shared(); - retv->type_ = "plain_net"; - return retv; - }) - .def("add_op", &operators::NetOp::AddOp) - .def("add_op", - [](operators::NetOp &self, - const std::shared_ptr &net) -> void { - self.AddOp(std::static_pointer_cast(net)); + py::class_(m, "Net") + .def_static("create", + []() -> operators::NetOp * { + auto *retv = new operators::NetOp; + retv->SetType("plain_net"); + return retv; + }) + .def("append_op", + [](operators::NetOp &self, const OperatorBase &op) { + self.AppendOp(op); }) .def("complete_add_op", &operators::NetOp::CompleteAddOp) .def("complete_add_op", [](std::shared_ptr &self) { self->CompleteAddOp(); }); - ExposeOperator(net); + // recurrent_op + py::class_(m, "RecurrentOp") + .def_static( + "create", + [](py::bytes protobin) -> operators::RecurrentOp * { + OpDesc desc; + PADDLE_ENFORCE(desc.ParsePartialFromString(protobin), + "Cannot parse user input to OpDesc"); + PADDLE_ENFORCE(desc.IsInitialized(), + "User OpDesc is not initialized, reason %s", + desc.InitializationErrorString()); + auto rnn_op = OpRegistry::CreateOp(desc); + return static_cast(rnn_op.release()); + }) + .def("set_stepnet", + [](operators::RecurrentOp &self, const operators::NetOp &net) + -> void { self.set_stepnet(net.Clone()); }); m.def("unique_integer", UniqueIntegerGenerator); diff --git a/paddle/framework/tensor_py.h b/paddle/pybind/tensor_py.h similarity index 91% rename from paddle/framework/tensor_py.h rename to paddle/pybind/tensor_py.h index 4e1ab77b157fe1adaeac55c271c056236f2d40de..95171acf729a513e5c92d1e0cba15cb12b38561a 100644 --- a/paddle/framework/tensor_py.h +++ b/paddle/pybind/tensor_py.h @@ -63,8 +63,11 @@ struct CastToPyBufferImpl { } return py::buffer_info( dst_tensor.mutable_data(dst_tensor.holder_->place()), - sizeof(CUR_TYPE), py::format_descriptor::format(), - (size_t)framework::arity(dst_tensor.dims()), dims_outside, strides); + sizeof(CUR_TYPE), + py::format_descriptor::format(), + (size_t)framework::arity(dst_tensor.dims()), + dims_outside, + strides); } else { constexpr bool less = I + 1 < std::tuple_size>::value; return CastToPyBufferImpl()(tensor); @@ -82,7 +85,7 @@ void PyCPUTensorSetFromArray( framework::Tensor &self, py::array_t array, paddle::platform::CPUPlace &place) { - std::vector dims; + std::vector dims; dims.reserve(array.ndim()); for (size_t i = 0; i < array.ndim(); ++i) { dims.push_back((int)array.shape()[i]); @@ -99,7 +102,7 @@ void PyCUDATensorSetFromArray( framework::Tensor &self, py::array_t array, paddle::platform::GPUPlace &place) { - std::vector dims; + std::vector dims; dims.reserve(array.ndim()); for (size_t i = 0; i < array.ndim(); ++i) { dims.push_back((int)array.shape()[i]); @@ -107,8 +110,8 @@ void PyCUDATensorSetFromArray( self.Resize(framework::make_ddim(dims)); auto *dst = self.mutable_data(place); - paddle::platform::GpuMemcpySync(dst, array.data(), sizeof(T) * array.size(), - cudaMemcpyHostToDevice); + paddle::platform::GpuMemcpySync( + dst, array.data(), sizeof(T) * array.size(), cudaMemcpyHostToDevice); } #endif diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh index 2f0205b7702b6d73b5348430f39166ec78f6c143..2ac455d771bf78377ce4ee7d921393d3b3958e3c 100644 --- a/paddle/scripts/docker/build.sh +++ b/paddle/scripts/docker/build.sh @@ -30,6 +30,8 @@ Configuring cmake in /paddle/build ... -DCMAKE_BUILD_TYPE=Release -DWITH_DOC=OFF -DWITH_GPU=${WITH_GPU:-OFF} + -DWITH_MKLDNN=${WITH_MKLDNN:-ON} + -DWITH_MKLML=${WITH_MKLML:-ON} -DWITH_AVX=${WITH_AVX:-OFF} -DWITH_GOLANG=${WITH_GOLANG:-ON} -DWITH_SWIG_PY=ON @@ -37,8 +39,8 @@ Configuring cmake in /paddle/build ... -DWITH_PYTHON=${WITH_PYTHON:-ON} -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} -DCUDNN_ROOT=/usr/ - -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} - -DWITH_TESTING=${WITH_TESTING:-OFF} + -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON} + -DWITH_TESTING=${WITH_TESTING:-ON} -DCMAKE_EXPORT_COMPILE_COMMANDS=ON ======================================== EOF @@ -50,25 +52,26 @@ cmake .. \ -DCMAKE_BUILD_TYPE=Release \ -DWITH_DOC=OFF \ -DWITH_GPU=${WITH_GPU:-OFF} \ + -DWITH_MKLDNN=${WITH_MKLDNN:-ON} \ + -DWITH_MKLML=${WITH_MKLML:-ON} \ -DWITH_AVX=${WITH_AVX:-OFF} \ -DWITH_GOLANG=${WITH_GOLANG:-ON} \ -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} \ -DWITH_C_API=${WITH_C_API:-OFF} \ -DWITH_PYTHON=${WITH_PYTHON:-ON} \ -DCUDNN_ROOT=/usr/ \ - -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} \ - -DWITH_TESTING=${WITH_TESTING:-OFF} \ + -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON} \ + -DWITH_TESTING=${WITH_TESTING:-ON} \ -DCMAKE_EXPORT_COMPILE_COMMANDS=ON cat <> /paddle/build/Dockerfile </dev/null || true -cmake -DCMAKE_SYSTEM_NAME=Android \ - -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_STANDALONE_TOOLCHAIN \ - -DANDROID_ABI=armeabi-v7a \ - -DANDROID_ARM_NEON=ON \ - -DANDROID_ARM_MODE=ON \ - -DHOST_C_COMPILER=/usr/bin/gcc \ - -DHOST_CXX_COMPILER=/usr/bin/g++ \ - -DCMAKE_INSTALL_PREFIX=/paddle/install \ - -DCMAKE_BUILD_TYPE=RelWithDebInfo \ - -DCMAKE_C_FLAGS_RELWITHDEBINFO="-O3" \ - -DCMAKE_CXX_FLAGS_RELWITHDEBINFO="-O3" \ - -DWITH_C_API=ON \ - -DWITH_SWIG_PY=OFF \ - .. +BUILD_ROOT=/paddle/build_android +DEST_ROOT=/paddle/install + +rm -rf $BUILD_ROOT 2>/dev/null || true +mkdir -p $BUILD_ROOT +cd $BUILD_ROOT + +if [ $ANDROID_ABI == "armeabi-v7a" ]; then + cmake -DCMAKE_SYSTEM_NAME=Android \ + -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \ + -DANDROID_ABI=$ANDROID_ABI \ + -DANDROID_ARM_NEON=ON \ + -DANDROID_ARM_MODE=ON \ + -DHOST_C_COMPILER=/usr/bin/gcc \ + -DHOST_CXX_COMPILER=/usr/bin/g++ \ + -DCMAKE_INSTALL_PREFIX=$DEST_ROOT \ + -DCMAKE_BUILD_TYPE=Release \ + -DUSE_EIGEN_FOR_BLAS=ON \ + -DWITH_C_API=ON \ + -DWITH_SWIG_PY=OFF \ + -DWITH_STYLE_CHECK=OFF \ + .. +elif [ $ANDROID_ABI == "arm64-v8a" ]; then + cmake -DCMAKE_SYSTEM_NAME=Android \ + -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM64_STANDALONE_TOOLCHAIN \ + -DANDROID_ABI=$ANDROID_ABI \ + -DANDROID_ARM_MODE=ON \ + -DHOST_C_COMPILER=/usr/bin/gcc \ + -DHOST_CXX_COMPILER=/usr/bin/g++ \ + -DCMAKE_INSTALL_PREFIX=$DEST_ROOT \ + -DCMAKE_BUILD_TYPE=Release \ + -DUSE_EIGEN_FOR_BLAS=OFF \ + -DWITH_C_API=ON \ + -DWITH_SWIG_PY=OFF \ + -DWITH_STYLE_CHECK=OFF \ + .. +elif [ $ANDROID_ABI == "armeabi" ]; then + cmake -DCMAKE_SYSTEM_NAME=Android \ + -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \ + -DANDROID_ABI=$ANDROID_ABI \ + -DANDROID_ARM_MODE=ON \ + -DHOST_C_COMPILER=/usr/bin/gcc \ + -DHOST_CXX_COMPILER=/usr/bin/g++ \ + -DCMAKE_INSTALL_PREFIX=/paddle/install \ + -DCMAKE_BUILD_TYPE=Release \ + -DWITH_C_API=ON \ + -DWITH_SWIG_PY=OFF \ + -DWITH_STYLE_CHECK=OFF \ + .. +else + echo "Invalid ANDROID_ABI: $ANDROID_ABI" +fi + make -j `nproc` make install -j `nproc` diff --git a/paddle/scripts/submit_local.sh.in b/paddle/scripts/submit_local.sh.in index 12bf629ea920832f96bc5f7cc0b38abfddd34d97..26f9c0fcd4e045f5d603fc4e4b16691a418823ca 100755 --- a/paddle/scripts/submit_local.sh.in +++ b/paddle/scripts/submit_local.sh.in @@ -18,6 +18,8 @@ function version(){ echo "PaddlePaddle @PADDLE_VERSION@, compiled with" echo " with_avx: @WITH_AVX@" echo " with_gpu: @WITH_GPU@" + echo " with_mkldnn: @WITH_MKLDNN" + echo " with_mklml: @WITH_MKLML@" echo " with_double: @WITH_DOUBLE@" echo " with_python: @WITH_PYTHON@" echo " with_rdma: @WITH_RDMA@" @@ -54,8 +56,7 @@ if [ -z "${PADDLE_NO_STAT+x}" ]; then fi fi - -MYDIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )" +PADDLE_BIN_PATH="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )" if [ ! -z "${DEBUGGER}" ]; then echo "Using debug command ${DEBUGGER}" @@ -91,34 +92,16 @@ else: sys.exit(0) EOF -if [ $? -eq 1 ]; then # Older version installed, or not installed at all - echo "First time run paddle, need to install some python dependencies." - # setuptools normalizes package version, so we need to use normalized - # package version for paddle python package - PYTHON_PADDLE_VERSION=$(python -c 'import packaging.version -import setuptools -print str(packaging.version.Version("@PADDLE_VERSION@")) -' 2>/dev/null) - BASEDIR=$(dirname "$0") - pip install ${BASEDIR}/../opt/paddle/share/wheels/*-${PYTHON_PADDLE_VERSION}-*.whl - if [ $? -ne 0 ]; then - echo "pip install wheels failed. " - echo "Please use 'sudo paddle' at the first time you use PaddlePaddle" - echo "PaddlePaddle will install some python dependencies automatically." - exit 1 - fi - echo "Python dependencies are installed." -fi case "$1" in "train") - ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_trainer ${@:2} + ${DEBUGGER} $PADDLE_BIN_PATH/paddle_trainer ${@:2} ;; "merge_model") - ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_merge_model ${@:2} + ${DEBUGGER} $PADDLE_BIN_PATH/paddle_merge_model ${@:2} ;; "pserver") - ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_pserver_main ${@:2} + ${DEBUGGER} $PADDLE_BIN_PATH/paddle_pserver_main ${@:2} ;; "dump_config") python -m paddle.utils.dump_config ${@:2} @@ -127,7 +110,7 @@ case "$1" in python -m paddle.utils.make_model_diagram ${@:2} ;; "usage") - $MYDIR/../opt/paddle/bin/paddle_usage ${@:2} + $PADDLE_BIN_PATH/paddle_usage ${@:2} ;; "version") version diff --git a/paddle/scripts/travis/build_android.sh b/paddle/scripts/travis/build_android.sh index 004067a8f55351509caaf2bbf6d5c349a4698a79..9da71d1e8cdec4047167fe354973e6bac85fb9f0 100755 --- a/paddle/scripts/travis/build_android.sh +++ b/paddle/scripts/travis/build_android.sh @@ -22,6 +22,7 @@ cmake -DCMAKE_SYSTEM_NAME=Android \ -DANDROID_ABI=armeabi-v7a \ -DANDROID_ARM_NEON=ON \ -DANDROID_ARM_MODE=ON \ + -DUSE_EIGEN_FOR_BLAS=ON \ -DWITH_C_API=ON \ -DWITH_SWIG_PY=OFF \ -DWITH_STYLE_CHECK=OFF \ diff --git a/paddle/trainer/TrainerConfigHelper.cpp b/paddle/trainer/TrainerConfigHelper.cpp index eba40862b926cfe863c569e73a6a3ceabcf1f3b4..a0a365aa0bb0ac26939a02c1cd626d0c17c6a9fe 100644 --- a/paddle/trainer/TrainerConfigHelper.cpp +++ b/paddle/trainer/TrainerConfigHelper.cpp @@ -29,7 +29,6 @@ DECLARE_bool(with_gpu); DECLARE_bool(parallel_nn); DECLARE_string(config_args); DECLARE_bool(use_mkldnn); -DECLARE_bool(use_mkldnn_wgt); const char *kConfigParserModuleName = "paddle.trainer.config_parser"; const char *kConfigParserFuncName = "parse_config_and_serialize"; @@ -47,7 +46,6 @@ TrainerConfigHelper::TrainerConfigHelper(const std::string &configFilePath) << ",with_cost=" << FLAGS_with_cost << ",use_gpu=" << FLAGS_use_gpu << ",parallel_nn=" << FLAGS_parallel_nn << ",use_mkldnn=" << FLAGS_use_mkldnn - << ",use_mkldnn_wgt=" << FLAGS_use_mkldnn_wgt << ",cudnn_version=" << hl_get_cudnn_lib_version(); if (!FLAGS_config_args.empty()) { configArgs << "," << FLAGS_config_args; diff --git a/paddle/utils/Flags.cpp b/paddle/utils/Flags.cpp index 600c83a8487191895de635dd8433f6c44e86ce77..ab1c181c62cdbee8cc5f804ec9aaf63ac5464ad6 100644 --- a/paddle/utils/Flags.cpp +++ b/paddle/utils/Flags.cpp @@ -27,7 +27,6 @@ DEFINE_bool(use_mkldnn, false, "Default still keep use CPU training"); DEFINE_bool(use_mkldnn, false, "Only support CPU training"); #endif -DEFINE_bool(use_mkldnn_wgt, false, "Init weight from CPU weight"); DEFINE_bool(parallel_nn, false, "Whether to use multi-threads to calculate one neural network." diff --git a/paddle/utils/Flags.h b/paddle/utils/Flags.h index 0aca4c0ee036ee8490c0ceca7279df876dc21947..1832bb515ec85df3d7733e01b063a01ad6a3b282 100644 --- a/paddle/utils/Flags.h +++ b/paddle/utils/Flags.h @@ -41,4 +41,3 @@ DECLARE_string(predict_file); DECLARE_bool(prev_batch_state); DECLARE_string(init_model_path); DECLARE_bool(use_mkldnn); -DECLARE_bool(use_mkldnn_wgt); diff --git a/paddle/utils/Util.cpp b/paddle/utils/Util.cpp index b18b73e06a6c39c3bf9717280bc6323917c80efb..2755fdd9cd1c2509cad996557c6fb24363d42d8a 100644 --- a/paddle/utils/Util.cpp +++ b/paddle/utils/Util.cpp @@ -320,6 +320,9 @@ void loadFileList(const std::string& fileListFileName, } double getMemoryUsage() { +#if defined(__ANDROID__) + return 0.0; +#else FILE* fp = fopen("/proc/meminfo", "r"); CHECK(fp) << "failed to fopen /proc/meminfo"; size_t bufsize = 256 * sizeof(char); @@ -357,6 +360,7 @@ double getMemoryUsage() { delete[] buf; double usedMem = 1.0 - 1.0 * (freeMem + bufMem + cacheMem) / totalMem; return usedMem; +#endif } SyncThreadPool* getGlobalSyncThreadPool() { diff --git a/paddle/utils/Util.h b/paddle/utils/Util.h index 613844669d2495ada7b8f7a841f47b821b7fdeba..22ce2534d3468ded36221810aa61c15b37f13f3d 100644 --- a/paddle/utils/Util.h +++ b/paddle/utils/Util.h @@ -33,6 +33,13 @@ limitations under the License. */ #include "Flags.h" #include "hl_gpu.h" +#if defined(__ANDROID__) && (__ANDROID_API__ < 21) +inline int rand_r(unsigned int* seedp) { + (void)seedp; + return rand(); +} +#endif + /** * Loop over the elements in a container * TODO(yuyang18): It's this foreach useful? Why not use C++ 11 foreach, diff --git a/proto/ModelConfig.proto b/proto/ModelConfig.proto index 4f3d5bf3f6cb96c97285f40e3a3d100c2af47ad5..ebf0911d6ea0b39d51447859ae2aef485b50b0e6 100644 --- a/proto/ModelConfig.proto +++ b/proto/ModelConfig.proto @@ -82,6 +82,15 @@ message ConvConfig { // if not set, use img_size optional uint32 img_size_y = 14; + + optional uint32 dilation = 15 [ default = 1 ]; + optional uint32 dilation_y = 16 [ default = 1 ]; + + optional uint32 filter_size_z = 17 [ default = 1 ]; + optional uint32 padding_z = 18 [ default = 1 ]; + optional uint32 stride_z = 19 [ default = 1 ]; + optional uint32 output_z = 20 [ default = 1 ]; + optional uint32 img_size_z = 21 [ default = 1 ]; } message PoolConfig { @@ -124,6 +133,12 @@ message PoolConfig { // if not set, use padding optional uint32 padding_y = 13; + + optional uint32 size_z = 14 [ default = 1 ]; + optional uint32 stride_z = 15 [ default = 1 ]; + optional uint32 output_z = 16 [ default = 1 ]; + optional uint32 img_size_z = 17 [ default = 1 ]; + optional uint32 padding_z = 18 [ default = 1 ]; } message SppConfig { @@ -256,6 +271,7 @@ message ImageConfig { // The size of input feature map. required uint32 img_size = 8; optional uint32 img_size_y = 9; + optional uint32 img_size_z = 10 [ default = 1 ]; } message PriorBoxConfig { @@ -272,6 +288,11 @@ message PadConfig { repeated uint32 pad_w = 4; } +message ReshapeConfig { + repeated uint32 height_axis = 1; + repeated uint32 width_axis = 2; +} + message MultiBoxLossConfig { required uint32 num_classes = 1; required float overlap_threshold = 2; @@ -324,7 +345,6 @@ message LayerInputConfig { } message LayerConfig { - required string name = 1; required string type = 2; optional uint64 size = 3; @@ -496,6 +516,15 @@ message LayerConfig { optional int32 axis = 54 [ default = 2 ]; repeated uint32 offset = 55; repeated uint32 shape = 56; + + // for HuberRegressionLoss + optional double delta = 57 [ default = 1.0 ]; + + // for 3D data + optional uint64 depth = 58 [ default = 1 ]; + + // for switch order layer + optional ReshapeConfig reshape_conf = 59; } message EvaluatorConfig { diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt index 16c519d45aa62694201379b8da1ca54d8a07ee9a..7bd6d59b0096c23bb791b9b50702130057628879 100644 --- a/python/CMakeLists.txt +++ b/python/CMakeLists.txt @@ -21,6 +21,18 @@ if(WITH_GOLANG) add_dependencies(copy_paddle_master paddle_master) endif(WITH_GOLANG) +set(MKL_SHARED_LIBS "") +set(MKL_DEPENDS "") +if(WITH_MKLML) + list(APPEND MKL_SHARED_LIBS ${MKLML_LIB} ${MKLML_IOMP_LIB}) + list(APPEND MKL_DEPENDS mklml) +endif() + +if(WITH_MKLDNN) + list(APPEND MKL_SHARED_LIBS "${MKLDNN_LIB}" "${MKLDNN_LIB}.0") + list(APPEND MKL_DEPENDS mkldnn) +endif() + configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in ${CMAKE_CURRENT_BINARY_DIR}/setup.py) @@ -38,8 +50,11 @@ add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_PYTHON_BUILD_DIR}/lib* ${PADDLE_PYTHON_BUILD_DIR}/lib-python DEPENDS gen_proto_py copy_paddle_pybind framework_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER}) -add_custom_target(paddle_python ALL DEPENDS - ${PADDLE_PYTHON_BUILD_DIR}/.timestamp paddle_pserver_main paddle_trainer paddle_merge_model python_api_wheel) +set(paddle_python_deps ${PADDLE_PYTHON_BUILD_DIR}/.timestamp paddle_pserver_main paddle_trainer paddle_merge_model ${MKL_DEPENDS}) +if(WITH_SWIG_PY) + list(APPEND paddle_python_deps python_api_wheel) +endif() +add_custom_target(paddle_python ALL DEPENDS ${paddle_python_deps}) set(PADDLE_PYTHON_PACKAGE_DIR ${CMAKE_CURRENT_BINARY_DIR}/dist/) diff --git a/python/paddle/trainer/PyDataProvider2.py b/python/paddle/trainer/PyDataProvider2.py index 7e305e2cd9fbe306368a44d08f7f66b4185ae2d2..248da4ae8d1fb24652625ae8fc9ef314a028b912 100644 --- a/python/paddle/trainer/PyDataProvider2.py +++ b/python/paddle/trainer/PyDataProvider2.py @@ -27,6 +27,14 @@ class SequenceType(object): SEQUENCE = 1 SUB_SEQUENCE = 2 + @classmethod + def tostring(cls, value): + for k in cls.__dict__: + if not k.startswith('__'): + if getattr(cls, k) == value: + return cls.__name__ + '.' + k + return 'INVALID(' + str(value) + ')' + # TODO(yuyang18): Add string data type here. class DataType(object): @@ -35,6 +43,14 @@ class DataType(object): SparseValue = 2 Index = 3 + @classmethod + def tostring(cls, value): + for k in cls.__dict__: + if not k.startswith('__'): + if getattr(cls, k) == value: + return cls.__name__ + '.' + k + return 'INVALID(' + str(value) + ')' + class CacheType(object): NO_CACHE = 0 # No cache at all @@ -69,6 +85,26 @@ class InputType(object): self.seq_type = seq_type self.type = tp + def __repr__(self): + """ + Return a human readable representation like 'InputType(dim=25921, + seq_type=SequenceType.NO_SEQUENCE, type=DataType.Dense)' + """ + repr_str = type(self).__name__ + repr_str += '(' + serialize_func_map = { + 'dim': repr, + 'seq_type': SequenceType.tostring, + 'type': DataType.tostring + } + for idx, k in enumerate(self.__slots__): + if idx != 0: + repr_str += ', ' + repr_str += ( + k + '=' + serialize_func_map.get(k, repr)(getattr(self, k))) + repr_str += ')' + return repr_str + def dense_slot(dim, seq_type=SequenceType.NO_SEQUENCE): """ diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py index da99e5bd53458aa0cb201a3525e28c66ab63c52d..356e1d8b6fa9173db33a340744afd8d513a83a96 100644 --- a/python/paddle/trainer/config_parser.py +++ b/python/paddle/trainer/config_parser.py @@ -338,7 +338,8 @@ def RecurrentLayerGroupWithoutOutLinksBegin(name, in_links_count += 1 layer_name = MakeLayerNameInParentSubmodel(name) layer = g_layer_map[layer_name] - ScatterAgentLayer(name=name, size=layer.size) + ScatterAgentLayer( + name=name, size=layer.size, width=layer.width, height=layer.height) pair = g_current_submodel.in_links.add() pair.layer_name = layer_name @@ -869,18 +870,52 @@ class Conv(Cfg): caffe_mode=True, filter_size_y=None, padding_y=None, - stride_y=None): + stride_y=None, + dilation=None, + dilation_y=None): self.add_keys(locals()) if filter_size_y is None: self.filter_size_y = filter_size if padding_y is None: self.padding_y = padding + if dilation_y is None: + self.dilation_y = dilation if stride_y is None: self.stride_y = stride if output_x is not None: config_assert(output_x <= 0) +# please refer to the comments in proto/ModelConfig.proto +@config_class +class Conv3D(Cfg): + def __init__(self, + filter_size, + channels, + padding=None, + stride=None, + groups=None, + filter_channels=None, + output_x=None, + img_size=None, + caffe_mode=True, + filter_size_y=None, + padding_y=None, + stride_y=None, + filter_size_z=None, + padding_z=None, + stride_z=None): + self.add_keys(locals()) + self.filter_size_y = filter_size_y if filter_size_y else filter_size + self.filter_size_z = filter_size_z if filter_size_z else filter_size + self.padding_y = padding_y if padding_y else padding + self.padding_z = padding_z if padding_z else padding + self.stride_y = stride_y if stride_y else stride + self.stride_z = stride_z if stride_z else stride + if output_x is not None: + config_assert(output_x <= 0) + + @config_class class BilinearInterp(Cfg): def __init__(self, out_size_x=None, out_size_y=None, channels=None): @@ -903,6 +938,31 @@ class Pool(Cfg): self.add_keys(locals()) +@config_class +class Pool3d(Cfg): + def __init__( + self, + pool_type, + channels, + size_x, + size_y=None, + size_z=None, + start=None, + stride=None, # 1 by defalut in protobuf + stride_y=None, + stride_z=None, + padding=None, # 0 by defalut in protobuf + padding_y=None, + padding_z=None): + self.add_keys(locals()) + self.filter_size_y = size_y if size_y else size_x + self.filter_size_z = size_z if size_z else size_x + self.padding_y = padding_y if padding_y else padding + self.padding_z = padding_z if padding_z else padding + self.stride_y = stride_y if stride_y else stride + self.stride_z = stride_z if stride_z else stride + + @config_class class SpatialPyramidPool(Cfg): def __init__(self, pool_type, pyramid_height, channels): @@ -1167,6 +1227,20 @@ def get_img_size(input_layer_name, channels): return img_size, img_size_y +def get_img3d_size(input_layer_name, channels): + input = g_layer_map[input_layer_name] + img_pixels = input.size / channels + img_size = input.width + img_size_y = input.height + img_size_z = input.depth + + config_assert( + img_size * img_size_y * img_size_z == img_pixels, + "Input layer %s: Incorrect input image size %d * %d * %d for input image pixels %d" + % (input_layer_name, img_size, img_size_y, img_size_z, img_pixels)) + return img_size, img_size_y, img_size_z + + def parse_bilinear(bilinear, input_layer_name, bilinear_conf): parse_image(bilinear, input_layer_name, bilinear_conf.image_conf) bilinear_conf.out_size_x = bilinear.out_size_x @@ -1204,6 +1278,45 @@ def parse_pool(pool, input_layer_name, pool_conf, ceil_mode): pool_conf.stride_y, not ceil_mode) +def parse_pool3d(pool, input_layer_name, pool_conf, ceil_mode): + pool_conf.pool_type = pool.pool_type + config_assert(pool.pool_type in ['max-projection', 'avg-projection'], + "pool-type %s is not in " + "['max-projection', 'avg-projection']" % pool.pool_type) + + pool_conf.channels = pool.channels + + pool_conf.size_x = pool.size_x + pool_conf.stride = pool.stride + pool_conf.padding = pool.padding + + pool_conf.size_y = default(pool.size_y, pool_conf.size_x) + pool_conf.size_z = default(pool.size_z, pool_conf.size_x) + pool_conf.stride_y = default(pool.stride_y, pool_conf.stride) + pool_conf.stride_z = default(pool.stride_z, pool_conf.stride) + pool_conf.padding_y = default(pool.padding_y, pool_conf.padding) + pool_conf.padding_z = default(pool.padding_z, pool_conf.padding) + + pool_conf.img_size, pool_conf.img_size_y, pool_conf.img_size_z = \ + get_img3d_size(input_layer_name, pool.channels) + + config_assert(not pool.start, "start is deprecated in pooling.") + + if pool.padding is not None: + pool_conf.padding = pool.padding + pool_conf.padding_y = default(pool.padding_y, pool_conf.padding) + pool_conf.padding_z = default(pool.padding_z, pool_conf.padding) + pool_conf.output_x = cnn_output_size(pool_conf.img_size, pool_conf.size_x, + pool_conf.padding, pool_conf.stride, + not ceil_mode) + pool_conf.output_y = cnn_output_size(pool_conf.img_size_y, pool_conf.size_y, + pool_conf.padding_y, + pool_conf.stride_y, not ceil_mode) + pool_conf.output_z = cnn_output_size(pool_conf.img_size_z, pool_conf.size_z, + pool_conf.padding_z, + pool_conf.stride_z, not ceil_mode) + + def parse_spp(spp, input_layer_name, spp_conf): parse_image(spp, input_layer_name, spp_conf.image_conf) spp_conf.pool_type = spp.pool_type @@ -1219,6 +1332,12 @@ def parse_image(image, input_layer_name, image_conf): get_img_size(input_layer_name, image_conf.channels) +def parse_image3d(image, input_layer_name, image_conf): + image_conf.channels = image.channels + image_conf.img_size, image_conf.img_size_y, image_conf.img_size_z = \ + get_img3d_size(input_layer_name, image_conf.channels) + + def parse_norm(norm, input_layer_name, norm_conf): norm_conf.norm_type = norm.norm_type config_assert( @@ -1277,6 +1396,50 @@ def parse_conv(conv, input_layer_name, conv_conf, num_filters, trans=False): conv_conf.stride_y, conv_conf.caffe_mode) +#caffe_mode: compute the output size using floor instead of ceil, +# which is consistent of caffe and CuDNN's convention. +def parse_conv3d(conv, input_layer_name, conv_conf, num_filters, trans=False): + conv_conf.filter_size = conv.filter_size + conv_conf.filter_size_y = conv.filter_size_y + conv_conf.filter_size_z = conv.filter_size_z + conv_conf.channels = conv.channels + conv_conf.padding = conv.padding + conv_conf.padding_y = conv.padding_y + conv_conf.padding_z = conv.padding_z + conv_conf.stride = conv.stride + conv_conf.stride_y = conv.stride_y + conv_conf.stride_z = conv.stride_z + conv_conf.groups = conv.groups + conv_conf.caffe_mode = conv.caffe_mode + + if not trans: + conv_conf.filter_channels = conv.channels / conv.groups + conv_conf.img_size, conv_conf.img_size_y, conv_conf.img_size_z = \ + get_img3d_size(input_layer_name, conv.channels) + conv_conf.output_x = cnn_output_size( + conv_conf.img_size, conv_conf.filter_size, conv_conf.padding, + conv_conf.stride, conv_conf.caffe_mode) + conv_conf.output_y = cnn_output_size( + conv_conf.img_size_y, conv_conf.filter_size_y, conv_conf.padding_y, + conv_conf.stride_y, conv_conf.caffe_mode) + conv_conf.output_z = cnn_output_size( + conv_conf.img_size_z, conv_conf.filter_size_z, conv_conf.padding_z, + conv_conf.stride_z, conv_conf.caffe_mode) + else: + conv_conf.filter_channels = num_filters / conv.groups + conv_conf.output_x, conv_conf.output_y, conv_conf.output_z = \ + get_img3d_size(input_layer_name, conv.channels) + conv_conf.img_size = cnn_image_size( + conv_conf.output_x, conv_conf.filter_size, conv_conf.padding, + conv_conf.stride, conv_conf.caffe_mode) + conv_conf.img_size_y = cnn_image_size( + conv_conf.output_y, conv_conf.filter_size_y, conv_conf.padding_y, + conv_conf.stride_y, conv_conf.caffe_mode) + conv_conf.img_size_z = cnn_image_size( + conv_conf.output_z, conv_conf.filter_size_z, conv_conf.padding_z, + conv_conf.stride_z, conv_conf.caffe_mode) + + def parse_block_expand(block_expand, input_layer_name, block_expand_conf): block_expand_conf.channels = block_expand.channels block_expand_conf.stride_x = block_expand.stride_x @@ -1580,6 +1743,9 @@ class LayerBase(object): self.config.height = height self.config.width = width + def set_layer_depth(self, depth): + self.config.depth = depth + def set_cnn_layer(self, input_layer_name, height, @@ -1602,6 +1768,21 @@ class MultiClassCrossEntropySelfNormCostLayer(LayerBase): self.config.softmax_selfnorm_alpha = softmax_selfnorm_alpha +@config_layer('cross_entropy_over_beam') +class CrossEntropyOverBeamLayer(LayerBase): + def __init__(self, name, inputs, **xargs): + config_assert(len(inputs) % 3 == 0, "Error input number.") + super(CrossEntropyOverBeamLayer, self).__init__( + name, 'cross_entropy_over_beam', 0, inputs, **xargs) + input_num = len(inputs) / 3 + for i in range(input_num): + input_layer = self.get_input_layer(i * 3) + config_assert(input_layer.size == 1, ( + "Inputs for this layer are made up of " + "several triples, in which the first one is scores over " + "all candidate paths, whose size should be equal to 1.")) + + @config_layer('fc') class FCLayer(LayerBase): layer_type = 'fc' @@ -1783,11 +1964,19 @@ class DetectionOutputLayer(LayerBase): @config_layer('data') class DataLayer(LayerBase): - def __init__(self, name, size, height=None, width=None, device=None): + def __init__(self, + name, + size, + depth=None, + height=None, + width=None, + device=None): super(DataLayer, self).__init__( name, 'data', size, inputs=[], device=device) if height and width: self.set_layer_height_width(height, width) + if depth: + self.set_layer_depth(depth) ''' @@ -1902,7 +2091,7 @@ class ConvLayerBase(LayerBase): def calc_parameter_size(self, conv_conf): return self.config.num_filters * conv_conf.filter_channels \ - * (conv_conf.filter_size * conv_conf.filter_size_y) + * (conv_conf.filter_size * conv_conf.filter_size_y) @config_layer('exconv') @@ -1986,6 +2175,87 @@ class ConvTransLayer(ConvTransLayerBase): layer_type = 'cudnn_convt' +@config_layer('conv_3d') +class Conv3DLayerBase(LayerBase): + def __init__(self, + name, + inputs=[], + bias=True, + num_filters=None, + shared_biases=True, + **xargs): + super(Conv3DLayerBase, self).__init__( + name, self.layer_type, 0, inputs=inputs, **xargs) + + if num_filters is not None: + self.config.num_filters = num_filters + + # need to specify layer in config + self.config.type = self.layer_type + + trans = False + if self.config.type == "deconv3d": + trans = True + + if shared_biases is not None: + self.config.shared_biases = shared_biases + + for input_index in xrange(len(self.inputs)): + input_layer = self.get_input_layer(input_index) + conv_conf = self.config.inputs[input_index].conv_conf + parse_conv3d( + self.inputs[input_index].conv, + input_layer.name, + conv_conf, + num_filters, + trans=trans + ) # for z-axis pad:0, strid:1, filter_size:1, img_size:1 + psize = self.calc_parameter_size(conv_conf) + self.create_input_parameter(input_index, psize) + if trans: + self.set_cnn_layer(name, conv_conf.img_size_z, + conv_conf.img_size_y, conv_conf.img_size, + self.config.num_filters) + else: + self.set_cnn_layer(name, conv_conf.output_z, conv_conf.output_y, + conv_conf.output_x, self.config.num_filters) + + psize = self.config.size + if shared_biases: + psize = self.config.num_filters + self.create_bias_parameter(bias, psize, [psize, 1]) + + def calc_parameter_size(self, conv_conf): + return self.config.num_filters * conv_conf.filter_channels \ + * (conv_conf.filter_size * conv_conf.filter_size_y \ + * conv_conf.filter_size_z) + + def set_cnn_layer(self, + input_layer_name, + depth, + height, + width, + channels, + is_print=True): + size = depth * height * width * channels + self.set_layer_size(size) + self.set_layer_height_width(height, width) + self.set_layer_depth(depth) + if is_print: + print("output for %s: c = %d, d = %d, h = %d, w = %d, size = %d" % + (input_layer_name, channels, depth, height, width, size)) + + +@config_layer('conv3d') +class Conv3DLayer(Conv3DLayerBase): + layer_type = 'conv3d' + + +@config_layer('deconv3d') +class Conv3DLayer(Conv3DLayerBase): + layer_type = 'deconv3d' + + @config_layer('norm') class NormLayer(LayerBase): def __init__(self, name, inputs, **xargs): @@ -2015,6 +2285,35 @@ class PoolLayer(LayerBase): pool_conf.channels) +@config_layer('pool3d') +class Pool3DLayer(LayerBase): + def __init__(self, name, inputs, ceil_mode=True, **xargs): + super(Pool3DLayer, self).__init__( + name, 'pool3d', 0, inputs=inputs, **xargs) + for input_index in xrange(len(self.inputs)): + input_layer = self.get_input_layer(input_index) + pool_conf = self.config.inputs[input_index].pool_conf + parse_pool3d(self.inputs[input_index].pool, input_layer.name, + pool_conf, ceil_mode) + self.set_cnn_layer(name, pool_conf.output_z, pool_conf.output_y, + pool_conf.output_x, pool_conf.channels) + + def set_cnn_layer(self, + input_layer_name, + depth, + height, + width, + channels, + is_print=True): + size = depth * height * width * channels + self.set_layer_size(size) + self.set_layer_height_width(height, width) + self.set_layer_depth(depth) + if is_print: + print("output for %s: c = %d, d = %d, h = %d, w = %d, size = %d" % + (input_layer_name, channels, depth, height, width, size)) + + @config_layer('spp') class SpatialPyramidPoolLayer(LayerBase): def __init__(self, name, inputs, **xargs): @@ -2072,9 +2371,11 @@ class BatchNormLayer(LayerBase): name, inputs, bias=True, + img3D=False, use_global_stats=True, moving_average_fraction=0.9, batch_norm_type=None, + mean_var_names=None, **xargs): if inputs is None: inputs = [] @@ -2116,24 +2417,69 @@ class BatchNormLayer(LayerBase): input_layer = self.get_input_layer(0) image_conf = self.config.inputs[0].image_conf - parse_image(self.inputs[0].image, input_layer.name, image_conf) - - # Only pass the width and height of input to batch_norm layer - # when either of it is non-zero. - if input_layer.width != 0 or input_layer.height != 0: - self.set_cnn_layer(name, image_conf.img_size_y, image_conf.img_size, - image_conf.channels, False) + if img3D: + parse_image3d(self.inputs[0].image, input_layer.name, image_conf) + # Only pass the width and height of input to batch_norm layer + # when either of it is non-zero. + if input_layer.width != 0 or input_layer.height != 0: + self.set_cnn_layer( + input_layer_name=name, + depth=image_conf.img_size_z, + height=image_conf.img_size_y, + width=image_conf.img_size, + channels=image_conf.channels, + is_print=True) + else: + self.set_layer_size(input_layer.size) else: - self.set_layer_size(input_layer.size) + parse_image(self.inputs[0].image, input_layer.name, image_conf) + # Only pass the width and height of input to batch_norm layer + # when either of it is non-zero. + if input_layer.width != 0 or input_layer.height != 0: + self.set_cnn_layer( + input_layer_name=name, + height=image_conf.img_size_y, + width=image_conf.img_size, + channels=image_conf.channels, + is_print=True) + else: + self.set_layer_size(input_layer.size) psize = self.calc_parameter_size(image_conf) dims = [1, psize] + if mean_var_names is not None: + assert len(mean_var_names) == 2 + self.inputs[1].parameter_name = mean_var_names[0] + self.inputs[2].parameter_name = mean_var_names[1] + self.create_input_parameter(0, psize) self.create_input_parameter(1, psize, dims) self.create_input_parameter(2, psize, dims) self.create_bias_parameter(bias, psize) + def set_cnn_layer(self, + input_layer_name, + depth=None, + height=None, + width=None, + channels=None, + is_print=True): + depthIsNone = False + if depth is None: + depth = 1 + depthIsNone = True + size = depth * height * width * channels + self.set_layer_size(size) + self.set_layer_height_width(height, width) + self.set_layer_depth(depth) + if is_print and depthIsNone: + print("output for %s: c = %d, h = %d, w = %d, size = %d" % + (input_layer_name, channels, height, width, size)) + elif is_print: + print("output for %s: c = %d, d = %d, h = %d, w = %d, size = %d" % + (input_layer_name, channels, depth, height, width, size)) + def calc_parameter_size(self, image_conf): return image_conf.channels @@ -2197,8 +2543,8 @@ class MaxOutLayer(LayerBase): maxout_conf = self.config.inputs[0].maxout_conf parse_maxout(self.inputs[0].maxout, input_layer.name, maxout_conf) out_channels = maxout_conf.image_conf.channels / maxout_conf.groups - self.set_cnn_layer(name, g_layer_map[input_layer.name].height, - g_layer_map[input_layer.name].width, out_channels) + self.set_cnn_layer(name, maxout_conf.image_conf.img_size_y, + maxout_conf.image_conf.img_size, out_channels) @config_layer('row_conv') @@ -2232,6 +2578,20 @@ class ClipLayer(LayerBase): self.config.inputs[0].clip_conf.max = max +@config_layer('scale_shift') +class ScaleShiftLayer(LayerBase): + def __init__(self, name, inputs, bias=True, **xargs): + super(ScaleShiftLayer, self).__init__( + name, 'scale_shift', 0, inputs=inputs, **xargs) + config_assert( + len(self.inputs) == 1, + 'ScaleShiftLayer must have one and only one input.') + input_layer = self.get_input_layer(0) + self.set_layer_size(input_layer.size) + self.create_input_parameter(0, 1, [1, 1]) + self.create_bias_parameter(bias, 1) + + # key: cost type # value: cost class g_cost_map = {} @@ -2249,13 +2609,14 @@ def define_cost(class_name, cost_type): define_cost('MultiClassCrossEntropy', 'multi-class-cross-entropy') +define_cost('CrossEntropyOverBeamCostLayer', 'cross_entropy_over_beam') define_cost('RankingCost', 'rank-cost') define_cost('AucValidation', 'auc-validation') define_cost('PnpairValidation', 'pnpair-validation') define_cost('SumOfSquaresCostLayer', 'square_error') define_cost('MultiBinaryLabelCrossEntropy', 'multi_binary_label_cross_entropy') define_cost('SoftBinaryClassCrossEntropy', 'soft_binary_class_cross_entropy') -define_cost('HuberTwoClass', 'huber') +define_cost('HuberTwoClassification', 'huber_classification') define_cost('SumCost', 'sum_cost') define_cost('SmoothL1Cost', 'smooth_l1') @@ -2317,6 +2678,17 @@ class LambdaCost(LayerBase): self.config.max_sort_size = max_sort_size +@config_layer('huber_regression') +class HuberRegressionLoss(LayerBase): + def __init__(self, name, inputs, delta=1., coeff=1., device=None): + super(HuberRegressionLoss, self).__init__( + name, 'huber_regression', 1, inputs=inputs, device=device) + config_assert( + len(self.inputs) == 2, 'HuberRegression must have 2 inputs') + self.config.delta = delta + self.config.coeff = coeff + + @config_layer('nce') class NCELayer(LayerBase): def __init__(self, @@ -2369,9 +2741,20 @@ class AddToLayer(LayerBase): super(AddToLayer, self).__init__( name, 'addto', 0, inputs=inputs, **xargs) config_assert(len(inputs) > 0, 'inputs cannot be empty for AddToLayer') - for input_index in xrange(len(self.inputs)): - input_layer = self.get_input_layer(input_index) - self.set_layer_size(input_layer.size) + + if len(self.inputs) > 1: + for input_index in xrange(len(self.inputs)): + assert self.get_input_layer(0).height == self.get_input_layer( + input_index).height + assert self.get_input_layer(0).width == self.get_input_layer( + input_index).width + assert self.get_input_layer(0).depth == self.get_input_layer( + input_index).depth + + self.set_layer_size(self.get_input_layer(0).size) + self.set_layer_height_width(self.get_input_layer(0).height, \ + self.get_input_layer(0).width) + self.set_layer_depth(self.get_input_layer(0).depth) self.create_bias_parameter(bias, self.config.size) @@ -2391,9 +2774,11 @@ class GatherAgentLayer(LayerBase): @config_layer('scatter_agent') class ScatterAgentLayer(LayerBase): - def __init__(self, name, size, device=None): + def __init__(self, name, size, width=None, height=None, device=None): super(ScatterAgentLayer, self).__init__( name, 'scatter_agent', size, inputs=[], device=device) + if height and width: + self.set_layer_height_width(height, width) @config_layer('multiplex') @@ -2677,6 +3062,49 @@ class SubSequenceLayer(LayerBase): self.create_bias_parameter(bias, size) +@config_layer('seq_slice') +class SeqSliceLayer(LayerBase): + def __init__(self, name, inputs, starts, ends, bias=False, **xargs): + if isinstance(inputs, list): + assert len(inputs) == 1, ('the first input of sequence slice layer ' + 'is a single sequence input.') + else: + inputs = [inputs] + + if starts is not None: + if isinstance(starts, list): + assert len(starts) == 1, ( + 'the start indices for sequence slice layer cannot ' + 'be a list having more than one element.') + starts = starts[0] + inputs.append(starts) + + if ends is not None: + if isinstance(ends, list): + assert len(ends) == 1, ( + 'the end indices for sequence slice layer cannot ' + 'be a list having more than one element.') + ends = ends[0] + inputs.append(ends) + assert len(inputs) >= 2, ( + 'the sequence slice layer has at least two inputs.') + + super(SeqSliceLayer, self).__init__( + name, 'seq_slice', 0, inputs=inputs, **xargs) + + input_layer0 = self.get_input_layer(0) + size = input_layer0.size + self.set_layer_size(size) + + if len(inputs) == 3: + assert ( + self.get_input_layer(1).size == self.get_input_layer(2).size), ( + 'If start and end indices are both given to' + 'sequence slice layer, they should have the same width.') + elif len(inputs) == 2: + self.config.select_first = (starts is not None) + + @config_layer('sub_nested_seq') class SubNestedSequenceLayer(LayerBase): def __init__(self, name, inputs, selected_indices, bias=False, **xargs): @@ -3006,11 +3434,20 @@ class ConcatenateLayer(LayerBase): name, 'concat', 0, inputs=inputs, **xargs) size = 0 for input_index in xrange(len(self.inputs)): + assert self.get_input_layer(0).height == self.get_input_layer( + input_index).height + assert self.get_input_layer(0).width == self.get_input_layer( + input_index).width + assert self.get_input_layer(0).depth == self.get_input_layer( + input_index).depth input_layer = self.get_input_layer(input_index) input = self.inputs[input_index] if self.config.size == 0: size += input_layer.size + self.set_layer_height_width(self.get_input_layer(0).height, \ + self.get_input_layer(0).width) + self.set_layer_depth(self.get_input_layer(0).depth) self.set_layer_size(size) @@ -3306,6 +3743,15 @@ class RecurrentLayerGroup(LayerBase): name, 'recurrent_layer_group', 0, inputs=[], device=device) +@config_layer('switch_order') +class SwitchOrderLayer(LayerBase): + def __init__(self, name, inputs, reshape, **xargs): + super(SwitchOrderLayer, self).__init__( + name, 'switch_order', 0, inputs=inputs, **xargs) + self.config.reshape_conf.height_axis.extend(reshape['height']) + self.config.reshape_conf.width_axis.extend(reshape['width']) + + # Deprecated, use a new layer specific class instead @config_func def Layer(name, type, **xargs): diff --git a/python/paddle/trainer/recurrent_units.py b/python/paddle/trainer/recurrent_units.py old mode 100755 new mode 100644 diff --git a/python/paddle/trainer_config_helpers/evaluators.py b/python/paddle/trainer_config_helpers/evaluators.py index 44d52edfa7bae49bea196eba9387391b171840d8..57979db4de08989ab583b0ab41589c09789a0921 100644 --- a/python/paddle/trainer_config_helpers/evaluators.py +++ b/python/paddle/trainer_config_helpers/evaluators.py @@ -298,8 +298,8 @@ def pnpair_evaluator( input, label, info, - name=None, - weight=None, ): + weight=None, + name=None, ): """ Positive-negative pair rate Evaluator which adapts to rank task like learning to rank. This evaluator must contain at least three layers. @@ -308,27 +308,31 @@ def pnpair_evaluator( .. code-block:: python - eval = pnpair_evaluator(input, info, label) + eval = pnpair_evaluator(input, label, info) - :param name: Evaluator name. - :type name: None|basestring :param input: Input Layer name. The output prediction of network. :type input: LayerOutput :param label: Label layer name. :type label: LayerOutput - :param info: Label layer name. (TODO, explaination) + :param info: Info layer name. (TODO, explaination) :type info: LayerOutput :param weight: Weight Layer name. It should be a matrix with size [sample_num, 1]. (TODO, explaination) :type weight: LayerOutput + :param name: Evaluator name. + :type name: None|basestring """ + if not isinstance(input, list): + input = [input] + if label: + input.append(label) + if info: + input.append(info) evaluator_base( - name=name, - type="pnpair", input=input, - label=label, - info=info, - weight=weight) + type="pnpair", + weight=weight, + name=name, ) @evaluator(EvaluatorAttribute.FOR_CLASSIFICATION) @@ -429,12 +433,12 @@ def chunk_evaluator( .. code-block:: text - Scheme Description + Scheme Description plain Use the same label for the whole chunk. - IOB Two labels for chunk type X, B-X for chunk begining and I-X for chunk inside. + IOB Two labels for chunk type X, B-X for chunk begining and I-X for chunk inside. IOE Two labels for chunk type X, E-X for chunk ending and I-X for chunk inside. - IOBES Four labels for chunk type X, B-X for chunk begining, I-X for chunk inside, E-X for chunk end and S-X for single word chunk. - + IOBES Four labels for chunk type X, B-X for chunk begining, I-X for chunk inside, E-X for chunk end and S-X for single word chunk. + To make it clear, let's illustrate by an NER example. Assuming that there are three named entity types including ORG, PER and LOC which are called 'chunk type' here, if 'IOB' scheme were used, the label set will be extended to a set including B-ORG, I-ORG, B-PER, I-PER, B-LOC, I-LOC and O, @@ -451,7 +455,7 @@ def chunk_evaluator( tagType = label % numTagType chunkType = label / numTagType otherChunkType = numChunkTypes - + The following table shows the mapping rule between tagType and tag type in each scheme. .. code-block:: text @@ -475,7 +479,7 @@ def chunk_evaluator( O 6 In this example, chunkType has three values: 0 for ORG, 1 for PER, 2 for LOC, because the scheme is - "IOB" so tagType has two values: 0 for B and 1 for I. + "IOB" so tagType has two values: 0 for B and 1 for I. Here we will use I-LOC to explain the above mapping rules in detail. For I-LOC, the label id is 5, so we can get tagType=1 and chunkType=2, which means I-LOC is a part of NER chunk LOC and the tag is I. @@ -486,7 +490,7 @@ def chunk_evaluator( eval = chunk_evaluator(input, label, chunk_scheme, num_chunk_types) - + :param input: The input layers. :type input: LayerOutput :param label: An input layer containing the ground truth label. diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py old mode 100755 new mode 100644 index 1bc55c869601551aff5fc0311458f906385522d2..4b1d80d3db924bfa2ad0e081f785d8f5dd719fce --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -11,16 +11,17 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - import functools import collections import inspect +import paddle.trainer.config_parser as cp from paddle.trainer.config_parser import * from .activations import LinearActivation, SigmoidActivation, TanhActivation, \ ReluActivation, IdentityActivation, SoftmaxActivation, BaseActivation from .evaluators import * -from .poolings import MaxPooling, AvgPooling, BasePoolingType +from .poolings import MaxPooling, AvgPooling, BasePoolingType, \ + CudnnAvgPooling, CudnnMaxPooling from .attrs import * from .default_decorators import * @@ -52,7 +53,7 @@ __all__ = [ 'cos_sim', 'hsigmoid', 'conv_projection', - 'mse_cost', + 'square_error_cost', 'regression_cost', 'classification_cost', 'LayerOutput', @@ -104,11 +105,14 @@ __all__ = [ 'nce_layer', 'cross_entropy_with_selfnorm', 'cross_entropy', + 'BeamInput', + 'cross_entropy_over_beam', 'multi_binary_label_cross_entropy', 'sum_cost', 'rank_cost', 'lambda_cost', - 'huber_cost', + 'huber_regression_cost', + 'huber_classification_cost', 'block_expand_layer', 'maxout_layer', 'out_prod_layer', @@ -127,12 +131,17 @@ __all__ = [ 'row_conv_layer', 'dropout_layer', 'prelu_layer', + 'switch_order_layer', 'gated_unit_layer', 'crop_layer', 'sub_nested_seq_layer', 'clip_layer', 'slice_projection', - 'kmax_sequence_score_layer', + 'seq_slice_layer', + 'kmax_seq_score_layer', + 'img_pool3d_layer', + 'scale_shift_layer', + 'img_conv3d_layer', ] @@ -161,6 +170,7 @@ class LayerType(object): EXCONVTRANS_LAYER = 'exconvt' CUDNNCONV_LAYER = 'cudnn_conv' POOL_LAYER = 'pool' + POOL3D_LAYER = 'pool3d' BATCH_NORM_LAYER = 'batch_norm' NORM_LAYER = 'norm' SUM_TO_ONE_NORM_LAYER = 'sum_to_one_norm' @@ -214,22 +224,30 @@ class LayerType(object): CRF_DECODING_LAYER = 'crf_decoding' NCE_LAYER = 'nce' + CONV3D_LAYER = 'conv3d' + DECONV3D_LAYER = 'deconv3d' + RANK_COST = 'rank-cost' LAMBDA_COST = 'lambda_cost' - HUBER = 'huber' + HUBER_REGRESSION = 'huber_regression' + HUBER_CLASSIFICATION = 'huber_classification' CROSS_ENTROPY = 'multi-class-cross-entropy' CROSS_ENTROPY_WITH_SELFNORM = 'multi_class_cross_entropy_with_selfnorm' + CROSS_ENTROPY_OVER_BEAM = 'cross_entropy_over_beam' SOFT_BIN_CLASS_CROSS_ENTROPY = 'soft_binary_class_cross_entropy' MULTI_BIN_LABEL_CROSS_ENTROPY = 'multi_binary_label_cross_entropy' SUM_COST = 'sum_cost' SMOOTH_L1 = 'smooth_l1' PRELU = 'prelu' + SWITCH_ORDER_LAYER = 'switch_order' CROP_LAYER = 'crop' SUB_NESTED_SEQ = 'sub_nested_seq' CLIP_LAYER = 'clip' + SEQ_SLICE = 'seq_slice' KMAX_SEQ_SCORE = 'kmax_seq_score' + SCALE_SHIFT_LAYER = 'scale_shift' @staticmethod def is_layer_type(type_name): @@ -328,6 +346,18 @@ class LayerOutput(object): self.outputs = outputs self.reverse = reverse + @property + def width(self): + return cp.g_layer_map[self.full_name].width + + @property + def height(self): + return cp.g_layer_map[self.full_name].height + + @property + def depth(self): + return cp.g_layer_map[self.full_name].depth + def set_input(self, input): """ Set the input for a memory layer. Can only be used for memory layer @@ -878,7 +908,8 @@ def mixed_layer(size=0, @layer_support() -def data_layer(name, size, height=None, width=None, layer_attr=None): +def data_layer(name, size, depth=None, height=None, width=None, + layer_attr=None): """ Define DataLayer For NeuralNetwork. @@ -905,11 +936,20 @@ def data_layer(name, size, height=None, width=None, layer_attr=None): type=LayerType.DATA, name=name, size=size, + depth=depth, height=height, width=width, **ExtraLayerAttribute.to_kwargs(layer_attr)) - return LayerOutput(name, LayerType.DATA, size=size) + if depth is None: + depth = 1 + num_filters = None + if height is not None and width is not None: + num_filters = size / (width * height * depth) + assert num_filters * width * height * depth == size, \ + "size=%s width=%s height=%s depth=%s" % (size, width, height, depth) + + return LayerOutput(name, LayerType.DATA, size=size, num_filters=num_filters) @wrap_name_default("embedding") @@ -1183,7 +1223,8 @@ def detection_output_layer(input_loc, name=None): """ Apply the NMS to the output of network and compute the predict bounding - box location. + box location. The output of this layer could be None if there is no valid + bounding box. :param name: The Layer Name. :type name: basestring @@ -2322,6 +2363,7 @@ def img_conv_layer(input, groups=1, stride=1, padding=0, + dilation=1, bias_attr=None, param_attr=None, shared_biases=True, @@ -2329,6 +2371,7 @@ def img_conv_layer(input, filter_size_y=None, stride_y=None, padding_y=None, + dilation_y=None, trans=False, layer_type=None): """ @@ -2393,6 +2436,11 @@ def img_conv_layer(input, :type padding: int|tuple|list :param padding_y: The y dimension of the padding. :type padding_y: int + :param dilation: The x dimension of the dilation. Or input a tuple for two + image dimension + :type dilation: int|tuple|list + :param dilation_y: The y dimension of the dilation. + :type dilation_y: int :param bias_attr: Convolution bias attribute. None means default bias. False means no bias. :type bias_attr: ParameterAttribute|False @@ -2440,6 +2488,13 @@ def img_conv_layer(input, else: padding_y = padding + if dilation_y is None: + if isinstance(dilation, collections.Sequence): + assert len(dilation) == 2 + dilation, dilation_y = dilation + else: + dilation_y = dilation + if param_attr.attr.get('initial_smart'): # special initial for conv layers. init_w = (2.0 / (filter_size**2 * num_channels))**0.5 @@ -2449,6 +2504,8 @@ def img_conv_layer(input, param_attr.attr["initial_smart"] = False if layer_type: + if dilation > 1 or dilation_y > 1: + assert layer_type in ["cudnn_conv", "cudnn_convt"] if trans: assert layer_type in ["exconvt", "cudnn_convt"] else: @@ -2464,11 +2521,13 @@ def img_conv_layer(input, conv=Conv( filter_size=filter_size, padding=padding, + dilation=dilation, stride=stride, channels=num_channels, groups=groups, filter_size_y=filter_size_y, padding_y=padding_y, + dilation_y=dilation_y, stride_y=stride_y), **param_attr.attr), active_type=act.name, @@ -2574,11 +2633,14 @@ def img_pool_layer(input, elif isinstance(pool_type, AvgPooling): pool_type.name = 'avg' + assert type(pool_type) in [AvgPooling, MaxPooling, CudnnAvgPooling, + CudnnMaxPooling], \ + "only (Cudnn)AvgPooling, (Cudnn)MaxPooling are supported" + type_name = pool_type.name + '-projection' \ if ( isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \ else pool_type.name - pool_size_y = pool_size if pool_size_y is None else pool_size_y stride_y = stride if stride_y is None else stride_y padding_y = padding if padding_y is None else padding_y @@ -2610,6 +2672,146 @@ def img_pool_layer(input, size=l.config.size) +@wrap_name_default("pool3d") +@layer_support() +def img_pool3d_layer(input, + pool_size, + name=None, + num_channels=None, + pool_type=None, + stride=1, + padding=0, + layer_attr=None, + pool_size_y=None, + stride_y=None, + padding_y=None, + pool_size_z=None, + stride_z=None, + padding_z=None, + ceil_mode=True): + """ + Image pooling Layer. + + The details of pooling layer, please refer ufldl's pooling_ . + + .. _pooling: http://ufldl.stanford.edu/tutorial/supervised/Pooling/ + + - ceil_mode=True: + + .. math:: + + w = 1 + int(ceil(input\_width + 2 * padding - pool\_size) / float(stride)) + h = 1 + int(ceil(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y)) + d = 1 + int(ceil(input\_depth + 2 * padding\_z - pool\_size\_z) / float(stride\_z)) + + - ceil_mode=False: + + .. math:: + + w = 1 + int(floor(input\_width + 2 * padding - pool\_size) / float(stride)) + h = 1 + int(floor(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y)) + d = 1 + int(floor(input\_depth + 2 * padding\_z - pool\_size\_z) / float(stride\_z)) + + The example usage is: + + .. code-block:: python + + maxpool = img_pool3d_layer(input=conv, + pool_size=3, + num_channels=8, + stride=1, + padding=1, + pool_type=MaxPooling()) + + :param padding: pooling padding width. + :type padding: int|tuple|list + :param name: name of pooling layer + :type name: basestring. + :param input: layer's input + :type input: LayerOutput + :param pool_size: pooling window width + :type pool_size: int|tuple|list + :param num_channels: number of input channel. + :type num_channels: int + :param pool_type: pooling type. MaxPooling or AvgPooling. Default is + MaxPooling. + :type pool_type: BasePoolingType + :param stride: stride width of pooling. + :type stride: int|tuple|list + :param layer_attr: Extra Layer attribute. + :type layer_attr: ExtraLayerAttribute + :param ceil_mode: Wether to use ceil mode to calculate output height and with. + Defalut is True. If set false, Otherwise use floor. + + :type ceil_mode: bool + :return: LayerOutput object. + :rtype: LayerOutput + """ + if num_channels is None: + assert input.num_filters is not None + num_channels = input.num_filters + + if pool_type is None: + pool_type = MaxPooling() + elif isinstance(pool_type, AvgPooling): + pool_type.name = 'avg' + + type_name = pool_type.name + '-projection' \ + if ( + isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \ + else pool_type.name + + if isinstance(pool_size, collections.Sequence): + assert len(pool_size) == 3 + pool_size, pool_size_y, pool_size_z = pool_size + else: + pool_size_y = pool_size + pool_size_z = pool_size + + if isinstance(stride, collections.Sequence): + assert len(stride) == 3 + stride, stride_y, stride_z = stride + else: + stride_y = stride + stride_z = stride + + if isinstance(padding, collections.Sequence): + assert len(padding) == 3 + padding, padding_y, padding_y = padding + else: + padding_y = padding + padding_z = padding + + l = Layer( + name=name, + type=LayerType.POOL3D_LAYER, + inputs=[ + Input( + input.name, + pool=Pool3d( + pool_type=type_name, + channels=num_channels, + size_x=pool_size, + start=None, + stride=stride, + padding=padding, + size_y=pool_size_y, + stride_y=stride_y, + padding_y=padding_y, + size_z=pool_size_z, + stride_z=stride_z, + padding_z=padding_z)) + ], + ceil_mode=ceil_mode, + **ExtraLayerAttribute.to_kwargs(layer_attr)) + return LayerOutput( + name, + LayerType.POOL_LAYER, + parents=[input], + num_filters=num_channels, + size=l.config.size) + + @wrap_name_default("spp") @layer_support() def spp_layer(input, @@ -2756,13 +2958,15 @@ def img_cmrnorm_layer(input, def batch_norm_layer(input, act=None, name=None, + img3D=False, num_channels=None, bias_attr=None, param_attr=None, layer_attr=None, batch_norm_type=None, moving_average_fraction=0.9, - use_global_stats=None): + use_global_stats=None, + mean_var_names=None): """ Batch Normalization Layer. The notation of this layer as follow. @@ -2829,6 +3033,8 @@ def batch_norm_layer(input, :math:`runningMean = newMean*(1-factor) + runningMean*factor` :type moving_average_fraction: float. + :param mean_var_names: [mean name, variance name] + :type mean_var_names: string list :return: LayerOutput object. :rtype: LayerOutput """ @@ -2842,6 +3048,7 @@ def batch_norm_layer(input, (batch_norm_type == "cudnn_batch_norm") l = Layer( name=name, + img3D=img3D, inputs=Input( input.name, image=Image(channels=num_channels), **param_attr.attr), active_type=act.name, @@ -2850,6 +3057,7 @@ def batch_norm_layer(input, bias=ParamAttr.to_bias(bias_attr), moving_average_fraction=moving_average_fraction, use_global_stats=use_global_stats, + mean_var_names=mean_var_names, **ExtraLayerAttribute.to_kwargs(layer_attr)) return LayerOutput( @@ -4028,8 +4236,12 @@ def __cost_input__(input, label, weight=None): """ inputs and parents for cost layers. """ - ipts = [Input(input.name), Input(label.name)] - parents = [input, label] + if isinstance(input, LayerOutput): + input = [input] + if isinstance(label, LayerOutput): + label = [label] + ipts = [Input(ipt.name) for ipt in (input + label)] + parents = [ipt for ipt in (input + label)] if weight is not None: assert weight.size == 1 ipts.append(Input(weight.name)) @@ -4039,13 +4251,18 @@ def __cost_input__(input, label, weight=None): @wrap_name_default() @layer_support() -def mse_cost(input, label, weight=None, name=None, coeff=1.0, layer_attr=None): +def square_error_cost(input, + label, + weight=None, + name=None, + coeff=1.0, + layer_attr=None): """ - mean squared error cost: + sum of square error cost: .. math:: - \\frac{1}{N}\sum_{i=1}^N(t_i-y_i)^2 + cost = \\sum_{i=1}^N(t_i-y_i)^2 :param name: layer name. :type name: basestring @@ -4074,7 +4291,7 @@ def mse_cost(input, label, weight=None, name=None, coeff=1.0, layer_attr=None): return LayerOutput(name, LayerType.COST, parents=parents, size=1) -regression_cost = mse_cost +regression_cost = square_error_cost @wrap_name_default("cost") @@ -4202,8 +4419,7 @@ def conv_operator(img, num_channels = img.num_filters assert isinstance(filter, LayerOutput) - if filter.size is not None: - filter.size = filter_size * filter_size_y * num_filters * num_channels + assert filter.size is not None opCls = ConvTransOperator if trans else ConvOperator @@ -4914,7 +5130,6 @@ def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None): :return: LayerOutput object. :rtype: LayerOutput """ - assert input.layer_type == LayerType.CONV_LAYER assert isinstance(input.activation, LinearActivation) assert groups > 1 if num_channels is None: @@ -5018,17 +5233,6 @@ def warp_ctc_layer(input, building process, PaddlePaddle will clone the source codes, build and install it to :code:`third_party/install/warpctc` directory. - To use warp_ctc layer, you need to specify the path of :code:`libwarpctc.so`, - using following methods: - - 1. Set it in :code:`paddle.init` (python api) or :code:`paddle_init` (c api), - such as :code:`paddle.init(use_gpu=True, - warpctc_dir=your_paddle_source_dir/third_party/install/warpctc/lib)`. - - 2. Set environment variable LD_LIBRARY_PATH on Linux or DYLD_LIBRARY_PATH - on Mac OS. For instance, :code:`export - LD_LIBRARY_PATH=your_paddle_source_dir/third_party/install/warpctc/lib:$LD_LIBRARY_PATH`. - More details of CTC can be found by referring to `Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks 0 and reshape_axis < 4) + height = [ele for ele in xrange(reshape_axis)] + width = [ele for ele in range(reshape_axis, 4)] + reshape = {'height': height, 'width': width} + + l = Layer( + name=name, + inputs=input.name, + reshape=reshape, + type=LayerType.SWITCH_ORDER_LAYER, + active_type=act.name, + **ExtraLayerAttribute.to_kwargs(layer_attr)) + return LayerOutput( + name=name, + layer_type=LayerType.SWITCH_ORDER_LAYER, + activation=act, + parents=input, + size=l.config.size) + + @wrap_name_default() @layer_support() def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None): @@ -6174,16 +6596,82 @@ def clip_layer(input, min, max, name=None): name, LayerType.CLIP_LAYER, parents=[input], size=input.size) +@wrap_name_default() +def seq_slice_layer(input, starts, ends, name=None): + """ + seq_slice_layer will return one or several sub-sequences from the + input sequence layer given start and end indices. + + - If only start indices are given, and end indices are set to None, + this layer slices the input sequence from the given start indices + to its end. + - If only end indices are given, and start indices are set to None, + this layer slices the input sequence from its beginning to the + given end indices. + - If start and end indices are both given, they should have the same + number of elements. + + If start or end indices contains more than one elements, the input sequence + will be sliced for multiple times. + + + .. code-block:: python + + seq_silce = seq_slice_layer(input=input_seq, + starts=start_pos, ends=end_pos) + + :param name: name of this layer. + :type name: basestring + :param input: input for this layer, it should be a sequence. + :type input: LayerOutput + :param starts: start indices to slice the input sequence. + :type starts: LayerOutput|None + :param ends: end indices to slice the input sequence. + :type ends: LayerOutput|None + :return: LayerOutput object. + :rtype: LayerOutput + + """ + + assert isinstance(input, LayerOutput), ( + 'The first input of seq_slice layer must be a PaddlePaddle layer.') + + if starts is not None: + assert isinstance(starts, LayerOutput), ( + 'The start indices for seq_slice layer ' + 'must be a PaddlePaddle layer.') + if ends is not None: + assert isinstance(ends, LayerOutput), ( + 'The end indices for seq_slice layer must be a PaddlePaddle layer.') + assert starts is not None or ends is not None, ( + 'start and end indices ' + 'cannot be set to None at the same time, at least one of ' + 'them should be given.') + if starts is not None and ends is not None: + assert starts.size == ends.size, ( + 'If start and end indices are both given to seq_slice_layer, ' + 'they should have the same width.') + + Layer( + name=name, + type=LayerType.SEQ_SLICE, + inputs=input.name, + starts=starts.name if starts is not None else None, + ends=ends.name if ends is not None else None) + return LayerOutput( + name, LayerType.SEQ_SLICE, parents=[input], size=input.size) + + @wrap_name_default() @layer_support() -def kmax_sequence_score_layer(input, name=None, beam_size=1): +def kmax_seq_score_layer(input, name=None, beam_size=1): """ This layer accepts one input which are scores over a sequence or a nested sequence, and returns indices of beam_size sequences with highest scores. .. code-block:: python - kmax_indices = kmax_sequence_score_layer(input=input_layer, beam_size) + kmax_indices = kmax_seq_score_layer(input=input_layer, beam_size) :param name: The Layer Name. @@ -6196,10 +6684,10 @@ def kmax_sequence_score_layer(input, name=None, beam_size=1): :return: LayerOutput object. :rtype: LayerOutput """ - assert isinstance(input, LayerOutput), ("kmax_sequence_score_layer " + assert isinstance(input, LayerOutput), ("kmax_seq_score_layer " "accepts only one input.") assert input.size == 1, ( - "input of kmax_sequence_score_layer is a score" + "input of kmax_seq_score_layer is a score " "over a sequence or a nested sequence, so its width must be 1.") Layer( @@ -6210,3 +6698,186 @@ def kmax_sequence_score_layer(input, name=None, beam_size=1): return LayerOutput( name, LayerType.KMAX_SEQ_SCORE, parents=[input], size=input.size) + + +@wrap_name_default("conv3d") +@wrap_param_attr_default() +@wrap_bias_attr_default() +@wrap_act_default(act=ReluActivation()) +@layer_support(DROPOUT) +def img_conv3d_layer(input, + filter_size, + num_filters, + name=None, + num_channels=None, + act=None, + groups=1, + stride=1, + padding=0, + bias_attr=None, + param_attr=None, + shared_biases=True, + layer_attr=None, + trans=False, + layer_type=None): + """ + + The example usage is: + + .. code-block:: python + + conv = img_conv3d_layer(input=data, filter_size=1, + num_channels=8, + num_filters=16, stride=1, + bias_attr=False, + act=ReluActivation()) + + :param name: Layer name. + :type name: basestring + :param input: Layer Input. + :type input: LayerOutput + :param filter_size: The x dimension of a filter kernel. Or input a list. + :type filter_size: int|tuple|list + :param num_filters: Each filter group's number of filter + :param act: Activation type. Default is tanh + :type act: BaseActivation + :param groups: Group size of filters. + :type groups: int + :param stride: The x dimension of the stride. Or input a tuple for two image + dimension. + :type stride: int|tuple|list + :param padding: The x dimension of the padding. Or input a tuple for two + image dimension + :type padding: int|tuple|list + :param bias_attr: Convolution bias attribute. None means default bias. + False means no bias. + :type bias_attr: ParameterAttribute|False + :param num_channels: number of input channels. If None will be set + automatically from previous output. + :type num_channels: int + :param param_attr: Convolution param attribute. None means default attribute + :type param_attr: ParameterAttribute + :param shared_biases: Is biases will be shared between filters or not. + :type shared_biases: bool + :param layer_attr: Layer Extra Attribute. + :type layer_attr: ExtraLayerAttribute + :param trans: true if it is a convTransLayer, false if it is a convLayer + :type trans: bool + :param layer_type: specify the layer_type, default is None. If trans=True, + layer_type has to be "exconvt" or "cudnn_convt", + otherwise layer_type has to be either "exconv" or + "cudnn_conv" + :type layer_type: String + :return: LayerOutput object. + :rtype: LayerOutput + """ + if num_channels is None: + assert input.num_filters is not None + num_channels = input.num_filters + + if isinstance(filter_size, collections.Sequence): + assert len(filter_size) == 3 + filter_size, filter_size_y, filter_size_z = filter_size + else: + filter_size_y = filter_size + filter_size_z = filter_size + + if isinstance(stride, collections.Sequence): + assert len(stride) == 3 + stride, stride_y, stride_z = stride + else: + stride_y = stride + stride_z = stride + + if isinstance(padding, collections.Sequence): + assert len(padding) == 3 + padding, padding_y, padding_z = padding + else: + padding_y = padding + padding_z = padding + + if param_attr.attr.get('initial_smart'): + # special initial for conv layers. + init_w = (2.0 / (filter_size**2 * num_channels))**0.5 + param_attr.attr["initial_mean"] = 0.0 + param_attr.attr["initial_std"] = init_w + param_attr.attr["initial_strategy"] = 0 + param_attr.attr["initial_smart"] = False + + if layer_type: + if trans: + assert layer_type in ["deconv3d"] + lt = layer_type + else: + lt = LayerType.DECONV3D_LAYER if trans else LayerType.CONV3D_LAYER + + l = Layer( + name=name, + inputs=Input( + input.name, + conv=Conv3D( + filter_size=filter_size, + padding=padding, + stride=stride, + channels=num_channels, + groups=groups, + filter_size_y=filter_size_y, + padding_y=padding_y, + stride_y=stride_y, + filter_size_z=filter_size_z, + padding_z=padding_z, + stride_z=stride_z), + **param_attr.attr), + active_type=act.name, + num_filters=num_filters, + bias=ParamAttr.to_bias(bias_attr), + shared_biases=shared_biases, + type=lt, + **ExtraLayerAttribute.to_kwargs(layer_attr)) + return LayerOutput( + name, + lt, + parents=[input], + activation=act, + num_filters=num_filters, + size=l.config.size) + + +@wrap_name_default("scale_shift") +@wrap_param_attr_default() +@wrap_bias_attr_default() +def scale_shift_layer(input, name=None, param_attr=None, bias_attr=None): + """ + A layer applies a linear transformation to each element in each row of + the input matrix. For each element, the layer first re-scale it and then + adds a bias to it. + + This layer is very like the SlopeInterceptLayer, except the scale and + bias are trainable. + + .. math:: + + y = w * x + b + + .. code-block:: python + + scale_shift = scale_shift_layer(input=input_layer, bias_attr=False) + + :param name: The Layer Name. + :type name: basestring + :param input: The input layer. + :type input: LayerOutput. + :param param_attr: The parameter attribute of scaling. + :type param_attr: ParameterAttribute + :param bias_attr: The parameter attribute of shifting. + :type bias_attr: ParameterAttribute + :return: LayerOutput object. + :rtype: LayerOutput + """ + Layer( + name=name, + type=LayerType.SCALE_SHIFT_LAYER, + inputs=Input(input.name, **param_attr.attr), + bias=ParamAttr.to_bias(bias_attr)) + return LayerOutput( + name, LayerType.SCALE_SHIFT_LAYER, parents=[input], size=input.size) diff --git a/python/paddle/trainer_config_helpers/networks.py b/python/paddle/trainer_config_helpers/networks.py old mode 100755 new mode 100644 diff --git a/python/paddle/trainer_config_helpers/tests/configs/file_list.sh b/python/paddle/trainer_config_helpers/tests/configs/file_list.sh index a61beb871ad064c617fa141451afcb2a5ac64854..8a204a96f3ef57673cef65306d0bf8e8c3409751 100755 --- a/python/paddle/trainer_config_helpers/tests/configs/file_list.sh +++ b/python/paddle/trainer_config_helpers/tests/configs/file_list.sh @@ -8,6 +8,8 @@ test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops test_seq_concat_reshape test_pad test_smooth_l1 test_multiplex_layer test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_layer test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer -test_kmax_seq_socre_layer test_seq_select_layers) +test_kmax_seq_socre_layer test_sub_nested_seq_select_layer test_scale_shift_layer +test_seq_slice_layer test_cross_entropy_over_beam test_pooling3D_layer +test_conv3d_layer test_deconv3d_layer test_BatchNorm3D) export whole_configs=(test_split_datasource) diff --git a/python/paddle/trainer_config_helpers/tests/configs/img_layers.py b/python/paddle/trainer_config_helpers/tests/configs/img_layers.py index 9fda16a5407a1fe0af8c5986023a8368e5b87222..01d31ef3fad827bfd103ee00f4ddd1bde14e0f82 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/img_layers.py +++ b/python/paddle/trainer_config_helpers/tests/configs/img_layers.py @@ -12,6 +12,7 @@ img_conv = img_conv_layer( num_filters=64, filter_size=(32, 32), padding=(1, 1), + dilation=(1, 1), stride=(1, 1), act=LinearActivation()) img_bn = batch_norm_layer(input=img_conv, act=ReluActivation()) diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr index 1a577b8d9b1e1915236ba6afcfa97040d70c707a..5ddf6052df021b055390a42c25ce6c0d650e4aee 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr @@ -62,6 +62,7 @@ layers { moving_average_fraction: 0.9 height: 227 width: 227 + depth: 1 } layers { name: "__crmnorm_0__" diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr index 2818389b16cca75f5030b75fc4de8c89c06c5e02..c0252b945b4c7fd6b4dad8770e3e1dccb88df28a 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr @@ -62,6 +62,7 @@ layers { moving_average_fraction: 0.9 height: 256 width: 256 + depth: 1 } layers { name: "__crmnorm_0__" diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_BatchNorm3D.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_BatchNorm3D.protostr new file mode 100644 index 0000000000000000000000000000000000000000..832ed24a31dd2bedba9a4fce77d7a088d1796fdb --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_BatchNorm3D.protostr @@ -0,0 +1,92 @@ +type: "nn" +layers { + name: "data3D" + type: "data" + size: 360 + active_type: "" + height: 6 + width: 20 + depth: 3 +} +layers { + name: "__batch_norm_0__" + type: "batch_norm" + size: 360 + active_type: "relu" + inputs { + input_layer_name: "data3D" + input_parameter_name: "___batch_norm_0__.w0" + image_conf { + channels: 1 + img_size: 20 + img_size_y: 6 + img_size_z: 3 + } + } + inputs { + input_layer_name: "data3D" + input_parameter_name: "___batch_norm_0__.w1" + } + inputs { + input_layer_name: "data3D" + input_parameter_name: "___batch_norm_0__.w2" + } + bias_parameter_name: "___batch_norm_0__.wbias" + moving_average_fraction: 0.9 + height: 6 + width: 20 + depth: 3 +} +parameters { + name: "___batch_norm_0__.w0" + size: 1 + initial_mean: 1.0 + initial_std: 0.0 + initial_strategy: 0 + initial_smart: false +} +parameters { + name: "___batch_norm_0__.w1" + size: 1 + initial_mean: 0.0 + initial_std: 0.0 + dims: 1 + dims: 1 + initial_strategy: 0 + initial_smart: false + is_static: true + is_shared: true +} +parameters { + name: "___batch_norm_0__.w2" + size: 1 + initial_mean: 0.0 + initial_std: 0.0 + dims: 1 + dims: 1 + initial_strategy: 0 + initial_smart: false + is_static: true + is_shared: true +} +parameters { + name: "___batch_norm_0__.wbias" + size: 1 + initial_mean: 0.0 + initial_std: 0.0 + dims: 1 + dims: 1 + initial_strategy: 0 + initial_smart: false +} +input_layer_names: "data3D" +output_layer_names: "__batch_norm_0__" +sub_models { + name: "root" + layer_names: "data3D" + layer_names: "__batch_norm_0__" + input_layer_names: "data3D" + output_layer_names: "__batch_norm_0__" + is_recurrent_layer_group: false +} + diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_bi_grumemory.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_bi_grumemory.protostr index b110e91498ce7d112987714bd769868179141c54..8a1399efad0ff339e35f69400ac654a4787a6018 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_bi_grumemory.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_bi_grumemory.protostr @@ -74,6 +74,9 @@ layers { inputs { input_layer_name: "__bidirectional_gru_0___bw" } + height: 0 + width: 0 + depth: 1 } parameters { name: "___bidirectional_gru_0___fw_transform.w0" diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_conv3d_layer.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_conv3d_layer.protostr new file mode 100644 index 0000000000000000000000000000000000000000..9fe2bc29d3cd06231b67102e28f7a49c28306958 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_conv3d_layer.protostr @@ -0,0 +1,132 @@ +type: "nn" +layers { + name: "data" + type: "data" + size: 36288 + active_type: "" + height: 48 + width: 42 + depth: 6 +} +layers { + name: "conv3d_1" + type: "conv3d" + size: 24192 + active_type: "" + inputs { + input_layer_name: "data" + input_parameter_name: "_conv3d_1.w0" + conv_conf { + filter_size: 3 + channels: 3 + stride: 2 + padding: 1 + groups: 1 + filter_channels: 3 + output_x: 21 + img_size: 42 + caffe_mode: true + filter_size_y: 3 + padding_y: 1 + stride_y: 2 + output_y: 24 + img_size_y: 48 + filter_size_z: 3 + padding_z: 1 + stride_z: 2 + output_z: 3 + img_size_z: 6 + } + } + bias_parameter_name: "_conv3d_1.wbias" + num_filters: 16 + shared_biases: true + height: 24 + width: 21 + depth: 3 +} +layers { + name: "conv3d_2" + type: "conv3d" + size: 24192 + active_type: "" + inputs { + input_layer_name: "data" + input_parameter_name: "_conv3d_2.w0" + conv_conf { + filter_size: 3 + channels: 3 + stride: 2 + padding: 1 + groups: 1 + filter_channels: 3 + output_x: 21 + img_size: 42 + caffe_mode: true + filter_size_y: 3 + padding_y: 1 + stride_y: 2 + output_y: 24 + img_size_y: 48 + filter_size_z: 3 + padding_z: 1 + stride_z: 2 + output_z: 3 + img_size_z: 6 + } + } + bias_parameter_name: "_conv3d_2.wbias" + num_filters: 16 + shared_biases: true + height: 24 + width: 21 + depth: 3 +} +parameters { + name: "_conv3d_1.w0" + size: 1296 + initial_mean: 0.0 + initial_std: 0.272165526976 + initial_strategy: 0 + initial_smart: false +} +parameters { + name: "_conv3d_1.wbias" + size: 16 + initial_mean: 0.0 + initial_std: 0.0 + dims: 16 + dims: 1 + initial_strategy: 0 + initial_smart: false +} +parameters { + name: "_conv3d_2.w0" + size: 1296 + initial_mean: 0.0 + initial_std: 0.272165526976 + initial_strategy: 0 + initial_smart: false +} +parameters { + name: "_conv3d_2.wbias" + size: 16 + initial_mean: 0.0 + initial_std: 0.0 + dims: 16 + dims: 1 + initial_strategy: 0 + initial_smart: false +} +input_layer_names: "data" +output_layer_names: "conv3d_2" +sub_models { + name: "root" + layer_names: "data" + layer_names: "conv3d_1" + layer_names: "conv3d_2" + input_layer_names: "data" + output_layer_names: "conv3d_2" + is_recurrent_layer_group: false +} + diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cost_layers.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cost_layers.protostr index 05847344be60b4de42a7dd709914fd3da524d1ae..55ab464ddf88f55bfb7b93ec0a189d4e53633468 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cost_layers.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cost_layers.protostr @@ -167,6 +167,20 @@ layers { softmax_selfnorm_alpha: 0.1 coeff: 1.0 } +layers { + name: "__huber_regression_cost_0__" + type: "huber_regression" + size: 1 + active_type: "" + inputs { + input_layer_name: "input" + } + inputs { + input_layer_name: "labels" + } + coeff: 1.0 + delta: 1.0 +} layers { name: "huber_probs" type: "data" @@ -180,8 +194,8 @@ layers { active_type: "" } layers { - name: "__huber_cost_0__" - type: "huber" + name: "__huber_classification_cost_0__" + type: "huber_classification" size: 1 active_type: "" inputs { @@ -300,7 +314,8 @@ output_layer_names: "__rank_cost_0__" output_layer_names: "__lambda_cost_0__" output_layer_names: "__cross_entropy_0__" output_layer_names: "__cross_entropy_with_selfnorm_0__" -output_layer_names: "__huber_cost_0__" +output_layer_names: "__huber_regression_cost_0__" +output_layer_names: "__huber_classification_cost_0__" output_layer_names: "__multi_binary_label_cross_entropy_0__" output_layer_names: "__sum_cost_0__" output_layer_names: "__nce_layer_0__" @@ -324,9 +339,10 @@ sub_models { layer_names: "__lambda_cost_0__" layer_names: "__cross_entropy_0__" layer_names: "__cross_entropy_with_selfnorm_0__" + layer_names: "__huber_regression_cost_0__" layer_names: "huber_probs" layer_names: "huber_label" - layer_names: "__huber_cost_0__" + layer_names: "__huber_classification_cost_0__" layer_names: "__multi_binary_label_cross_entropy_0__" layer_names: "__sum_cost_0__" layer_names: "__nce_layer_0__" @@ -349,7 +365,8 @@ sub_models { output_layer_names: "__lambda_cost_0__" output_layer_names: "__cross_entropy_0__" output_layer_names: "__cross_entropy_with_selfnorm_0__" - output_layer_names: "__huber_cost_0__" + output_layer_names: "__huber_regression_cost_0__" + output_layer_names: "__huber_classification_cost_0__" output_layer_names: "__multi_binary_label_cross_entropy_0__" output_layer_names: "__sum_cost_0__" output_layer_names: "__nce_layer_0__" diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cost_layers_with_weight.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cost_layers_with_weight.protostr index 96fb1d4ebde08b1bca2ffd09e8db0895842cbfd3..cec8a73db66f6091ec971527b3a42aa9e08154eb 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cost_layers_with_weight.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cost_layers_with_weight.protostr @@ -45,7 +45,7 @@ layers { coeff: 1.0 } layers { - name: "__mse_cost_0__" + name: "__square_error_cost_0__" type: "square_error" size: 1 active_type: "" @@ -130,7 +130,7 @@ input_layer_names: "label" input_layer_names: "weight" input_layer_names: "multi_class_label" output_layer_names: "__cost_0__" -output_layer_names: "__mse_cost_0__" +output_layer_names: "__square_error_cost_0__" output_layer_names: "__nce_layer_0__" evaluators { name: "classification_error_evaluator" @@ -146,7 +146,7 @@ sub_models { layer_names: "weight" layer_names: "__fc_layer_0__" layer_names: "__cost_0__" - layer_names: "__mse_cost_0__" + layer_names: "__square_error_cost_0__" layer_names: "multi_class_label" layer_names: "__nce_layer_0__" input_layer_names: "input" @@ -154,7 +154,7 @@ sub_models { input_layer_names: "weight" input_layer_names: "multi_class_label" output_layer_names: "__cost_0__" - output_layer_names: "__mse_cost_0__" + output_layer_names: "__square_error_cost_0__" output_layer_names: "__nce_layer_0__" evaluator_names: "classification_error_evaluator" is_recurrent_layer_group: false diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cross_entropy_over_beam.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cross_entropy_over_beam.protostr new file mode 100644 index 0000000000000000000000000000000000000000..a602569697e91b11b8d421ac359c2e523a00fa98 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cross_entropy_over_beam.protostr @@ -0,0 +1,207 @@ +type: "nn" +layers { + name: "sentence_states" + type: "data" + size: 32 + active_type: "" +} +layers { + name: "sentence_scores" + type: "data" + size: 1 + active_type: "" +} +layers { + name: "__kmax_seq_score_layer_0__" + type: "kmax_seq_score" + active_type: "" + inputs { + input_layer_name: "sentence_scores" + } + beam_size: 5 +} +layers { + name: "__sub_nested_seq_layer_0__" + type: "sub_nested_seq" + size: 32 + active_type: "" + inputs { + input_layer_name: "sentence_states" + } + inputs { + input_layer_name: "__kmax_seq_score_layer_0__" + } +} +layers { + name: "__fc_layer_0__" + type: "fc" + size: 1 + active_type: "" + inputs { + input_layer_name: "__sub_nested_seq_layer_0__" + input_parameter_name: "___fc_layer_0__.w0" + } + bias_parameter_name: "___fc_layer_0__.wbias" +} +layers { + name: "__kmax_seq_score_layer_1__" + type: "kmax_seq_score" + active_type: "" + inputs { + input_layer_name: "sentence_scores" + } + beam_size: 5 +} +layers { + name: "__seq_slice_layer_0__" + type: "seq_slice" + size: 32 + active_type: "" + inputs { + input_layer_name: "__sub_nested_seq_layer_0__" + } + inputs { + input_layer_name: "__kmax_seq_score_layer_1__" + } + select_first: true +} +layers { + name: "__fc_layer_1__" + type: "fc" + size: 1 + active_type: "" + inputs { + input_layer_name: "__seq_slice_layer_0__" + input_parameter_name: "___fc_layer_1__.w0" + } + bias_parameter_name: "___fc_layer_1__.wbias" +} +layers { + name: "__kmax_seq_score_layer_2__" + type: "kmax_seq_score" + active_type: "" + inputs { + input_layer_name: "__fc_layer_1__" + } + beam_size: 5 +} +layers { + name: "sentences_ids" + type: "data" + size: 1 + active_type: "" +} +layers { + name: "start_ids" + type: "data" + size: 1 + active_type: "" +} +layers { + name: "end_ids" + type: "data" + size: 1 + active_type: "" +} +layers { + name: "__cross_entropy_over_beam_0__" + type: "cross_entropy_over_beam" + active_type: "" + inputs { + input_layer_name: "sentence_scores" + } + inputs { + input_layer_name: "__kmax_seq_score_layer_0__" + } + inputs { + input_layer_name: "sentences_ids" + } + inputs { + input_layer_name: "__fc_layer_0__" + } + inputs { + input_layer_name: "__kmax_seq_score_layer_1__" + } + inputs { + input_layer_name: "start_ids" + } + inputs { + input_layer_name: "__fc_layer_1__" + } + inputs { + input_layer_name: "__kmax_seq_score_layer_2__" + } + inputs { + input_layer_name: "end_ids" + } +} +parameters { + name: "___fc_layer_0__.w0" + size: 32 + initial_mean: 0.0 + initial_std: 0.176776695297 + dims: 32 + dims: 1 + initial_strategy: 0 + initial_smart: true +} +parameters { + name: "___fc_layer_0__.wbias" + size: 1 + initial_mean: 0.0 + initial_std: 0.0 + dims: 1 + dims: 1 + initial_strategy: 0 + initial_smart: false +} +parameters { + name: "___fc_layer_1__.w0" + size: 32 + initial_mean: 0.0 + initial_std: 0.176776695297 + dims: 32 + dims: 1 + initial_strategy: 0 + initial_smart: true +} +parameters { + name: "___fc_layer_1__.wbias" + size: 1 + initial_mean: 0.0 + initial_std: 0.0 + dims: 1 + dims: 1 + initial_strategy: 0 + initial_smart: false +} +input_layer_names: "sentence_scores" +input_layer_names: "sentences_ids" +input_layer_names: "sentence_states" +input_layer_names: "start_ids" +input_layer_names: "end_ids" +output_layer_names: "__cross_entropy_over_beam_0__" +sub_models { + name: "root" + layer_names: "sentence_states" + layer_names: "sentence_scores" + layer_names: "__kmax_seq_score_layer_0__" + layer_names: "__sub_nested_seq_layer_0__" + layer_names: "__fc_layer_0__" + layer_names: "__kmax_seq_score_layer_1__" + layer_names: "__seq_slice_layer_0__" + layer_names: "__fc_layer_1__" + layer_names: "__kmax_seq_score_layer_2__" + layer_names: "sentences_ids" + layer_names: "start_ids" + layer_names: "end_ids" + layer_names: "__cross_entropy_over_beam_0__" + input_layer_names: "sentence_scores" + input_layer_names: "sentences_ids" + input_layer_names: "sentence_states" + input_layer_names: "start_ids" + input_layer_names: "end_ids" + output_layer_names: "__cross_entropy_over_beam_0__" + is_recurrent_layer_group: false +} + diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_deconv3d_layer.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_deconv3d_layer.protostr new file mode 100644 index 0000000000000000000000000000000000000000..7bf409731cbf8d5d98341b03c7c09d91fa8328d9 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_deconv3d_layer.protostr @@ -0,0 +1,132 @@ +type: "nn" +layers { + name: "data" + type: "data" + size: 36288 + active_type: "" + height: 48 + width: 42 + depth: 6 +} +layers { + name: "deconv3d_1" + type: "deconv3d" + size: 1387760 + active_type: "" + inputs { + input_layer_name: "data" + input_parameter_name: "_deconv3d_1.w0" + conv_conf { + filter_size: 3 + channels: 3 + stride: 2 + padding: 1 + groups: 1 + filter_channels: 16 + output_x: 42 + img_size: 83 + caffe_mode: true + filter_size_y: 3 + padding_y: 1 + stride_y: 2 + output_y: 48 + img_size_y: 95 + filter_size_z: 3 + padding_z: 1 + stride_z: 2 + output_z: 6 + img_size_z: 11 + } + } + bias_parameter_name: "_deconv3d_1.wbias" + num_filters: 16 + shared_biases: true + height: 95 + width: 83 + depth: 11 +} +layers { + name: "deconv3d_2" + type: "deconv3d" + size: 1387760 + active_type: "" + inputs { + input_layer_name: "data" + input_parameter_name: "_deconv3d_2.w0" + conv_conf { + filter_size: 3 + channels: 3 + stride: 2 + padding: 1 + groups: 1 + filter_channels: 16 + output_x: 42 + img_size: 83 + caffe_mode: true + filter_size_y: 3 + padding_y: 1 + stride_y: 2 + output_y: 48 + img_size_y: 95 + filter_size_z: 3 + padding_z: 1 + stride_z: 2 + output_z: 6 + img_size_z: 11 + } + } + bias_parameter_name: "_deconv3d_2.wbias" + num_filters: 16 + shared_biases: true + height: 95 + width: 83 + depth: 11 +} +parameters { + name: "_deconv3d_1.w0" + size: 6912 + initial_mean: 0.0 + initial_std: 0.272165526976 + initial_strategy: 0 + initial_smart: false +} +parameters { + name: "_deconv3d_1.wbias" + size: 16 + initial_mean: 0.0 + initial_std: 0.0 + dims: 16 + dims: 1 + initial_strategy: 0 + initial_smart: false +} +parameters { + name: "_deconv3d_2.w0" + size: 6912 + initial_mean: 0.0 + initial_std: 0.272165526976 + initial_strategy: 0 + initial_smart: false +} +parameters { + name: "_deconv3d_2.wbias" + size: 16 + initial_mean: 0.0 + initial_std: 0.0 + dims: 16 + dims: 1 + initial_strategy: 0 + initial_smart: false +} +input_layer_names: "data" +output_layer_names: "deconv3d_2" +sub_models { + name: "root" + layer_names: "data" + layer_names: "deconv3d_1" + layer_names: "deconv3d_2" + input_layer_names: "data" + output_layer_names: "deconv3d_2" + is_recurrent_layer_group: false +} + diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_kmax_seq_socre_layer.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_kmax_seq_socre_layer.protostr index 81bd71f68eb3f2c04ccd46ee3b77a07543395c60..f93d368c8687573db80106b9cc4defa56a881e46 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_kmax_seq_socre_layer.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_kmax_seq_socre_layer.protostr @@ -1,12 +1,6 @@ type: "nn" layers { - name: "input" - type: "data" - size: 300 - active_type: "" -} -layers { - name: "data" + name: "input_seq" type: "data" size: 128 active_type: "" @@ -17,13 +11,13 @@ layers { size: 1 active_type: "exponential" inputs { - input_layer_name: "data" + input_layer_name: "input_seq" input_parameter_name: "___fc_layer_0__.w0" } bias_parameter_name: "___fc_layer_0__.wbias" } layers { - name: "__kmax_sequence_score_layer_0__" + name: "__kmax_seq_score_layer_0__" type: "kmax_seq_score" active_type: "" inputs { @@ -51,16 +45,15 @@ parameters { initial_strategy: 0 initial_smart: false } -input_layer_names: "data" -output_layer_names: "__kmax_sequence_score_layer_0__" +input_layer_names: "input_seq" +output_layer_names: "__kmax_seq_score_layer_0__" sub_models { name: "root" - layer_names: "input" - layer_names: "data" + layer_names: "input_seq" layer_names: "__fc_layer_0__" - layer_names: "__kmax_sequence_score_layer_0__" - input_layer_names: "data" - output_layer_names: "__kmax_sequence_score_layer_0__" + layer_names: "__kmax_seq_score_layer_0__" + input_layer_names: "input_seq" + output_layer_names: "__kmax_seq_score_layer_0__" is_recurrent_layer_group: false } diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_pooling3D_layer.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_pooling3D_layer.protostr new file mode 100644 index 0000000000000000000000000000000000000000..8eb98593f6f692a445cf5088e101e9da3763b41d --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_pooling3D_layer.protostr @@ -0,0 +1,123 @@ +type: "nn" +layers { + name: "data_2d" + type: "data" + size: 6000 + active_type: "" + height: 20 + width: 10 +} +layers { + name: "pool___2d" + type: "pool" + size: 840 + active_type: "" + inputs { + input_layer_name: "data_2d" + pool_conf { + pool_type: "avg-projection" + channels: 30 + size_x: 5 + stride: 3 + output_x: 4 + img_size: 10 + padding: 1 + size_y: 5 + stride_y: 3 + output_y: 7 + img_size_y: 20 + padding_y: 1 + } + } + height: 7 + width: 4 +} +layers { + name: "data_3d_1" + type: "data" + size: 60000 + active_type: "" + height: 20 + width: 10 + depth: 10 +} +layers { + name: "pool_3d_1" + type: "pool3d" + size: 3360 + active_type: "" + inputs { + input_layer_name: "data_3d_1" + pool_conf { + pool_type: "avg-projection" + channels: 30 + size_x: 5 + stride: 3 + output_x: 4 + img_size: 10 + padding: 1 + size_y: 5 + stride_y: 3 + output_y: 7 + img_size_y: 20 + padding_y: 1 + size_z: 5 + stride_z: 3 + output_z: 4 + img_size_z: 10 + padding_z: 1 + } + } + height: 7 + width: 4 + depth: 4 +} +layers { + name: "pool_3d_2" + type: "pool3d" + size: 3360 + active_type: "" + inputs { + input_layer_name: "data_3d_1" + pool_conf { + pool_type: "max-projection" + channels: 30 + size_x: 5 + stride: 3 + output_x: 4 + img_size: 10 + padding: 1 + size_y: 5 + stride_y: 3 + output_y: 7 + img_size_y: 20 + padding_y: 1 + size_z: 5 + stride_z: 3 + output_z: 4 + img_size_z: 10 + padding_z: 1 + } + } + height: 7 + width: 4 + depth: 4 +} +input_layer_names: "data_2d" +output_layer_names: "pool___2d" +output_layer_names: "pool_3d_1" +output_layer_names: "pool_3d_2" +sub_models { + name: "root" + layer_names: "data_2d" + layer_names: "pool___2d" + layer_names: "data_3d_1" + layer_names: "pool_3d_1" + layer_names: "pool_3d_2" + input_layer_names: "data_2d" + output_layer_names: "pool___2d" + output_layer_names: "pool_3d_1" + output_layer_names: "pool_3d_2" + is_recurrent_layer_group: false +} + diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_recursive_topology.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_recursive_topology.protostr index 8133aa9c8d3e7c6843d1b27b70e87d394a1e0e47..046037936a6d85f54095c65f206e468aa69065d7 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_recursive_topology.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_recursive_topology.protostr @@ -16,6 +16,9 @@ layers { inputs { input_layer_name: "data" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_1__" @@ -28,6 +31,9 @@ layers { inputs { input_layer_name: "__addto_0__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_2__" @@ -40,6 +46,9 @@ layers { inputs { input_layer_name: "__addto_1__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_3__" @@ -52,6 +61,9 @@ layers { inputs { input_layer_name: "__addto_2__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_4__" @@ -64,6 +76,9 @@ layers { inputs { input_layer_name: "__addto_3__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_5__" @@ -76,6 +91,9 @@ layers { inputs { input_layer_name: "__addto_4__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_6__" @@ -88,6 +106,9 @@ layers { inputs { input_layer_name: "__addto_5__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_7__" @@ -100,6 +121,9 @@ layers { inputs { input_layer_name: "__addto_6__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_8__" @@ -112,6 +136,9 @@ layers { inputs { input_layer_name: "__addto_7__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_9__" @@ -124,6 +151,9 @@ layers { inputs { input_layer_name: "__addto_8__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_10__" @@ -136,6 +166,9 @@ layers { inputs { input_layer_name: "__addto_9__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_11__" @@ -148,6 +181,9 @@ layers { inputs { input_layer_name: "__addto_10__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_12__" @@ -160,6 +196,9 @@ layers { inputs { input_layer_name: "__addto_11__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_13__" @@ -172,6 +211,9 @@ layers { inputs { input_layer_name: "__addto_12__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_14__" @@ -184,6 +226,9 @@ layers { inputs { input_layer_name: "__addto_13__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_15__" @@ -196,6 +241,9 @@ layers { inputs { input_layer_name: "__addto_14__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_16__" @@ -208,6 +256,9 @@ layers { inputs { input_layer_name: "__addto_15__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_17__" @@ -220,6 +271,9 @@ layers { inputs { input_layer_name: "__addto_16__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_18__" @@ -232,6 +286,9 @@ layers { inputs { input_layer_name: "__addto_17__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_19__" @@ -244,6 +301,9 @@ layers { inputs { input_layer_name: "__addto_18__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_20__" @@ -256,6 +316,9 @@ layers { inputs { input_layer_name: "__addto_19__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_21__" @@ -268,6 +331,9 @@ layers { inputs { input_layer_name: "__addto_20__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_22__" @@ -280,6 +346,9 @@ layers { inputs { input_layer_name: "__addto_21__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_23__" @@ -292,6 +361,9 @@ layers { inputs { input_layer_name: "__addto_22__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_24__" @@ -304,6 +376,9 @@ layers { inputs { input_layer_name: "__addto_23__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_25__" @@ -316,6 +391,9 @@ layers { inputs { input_layer_name: "__addto_24__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_26__" @@ -328,6 +406,9 @@ layers { inputs { input_layer_name: "__addto_25__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_27__" @@ -340,6 +421,9 @@ layers { inputs { input_layer_name: "__addto_26__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_28__" @@ -352,6 +436,9 @@ layers { inputs { input_layer_name: "__addto_27__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_29__" @@ -364,6 +451,9 @@ layers { inputs { input_layer_name: "__addto_28__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_30__" @@ -376,6 +466,9 @@ layers { inputs { input_layer_name: "__addto_29__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__addto_31__" @@ -388,6 +481,9 @@ layers { inputs { input_layer_name: "__addto_30__" } + height: 0 + width: 0 + depth: 1 } layers { name: "__fc_layer_0__" diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_scale_shift_layer.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_scale_shift_layer.protostr new file mode 100644 index 0000000000000000000000000000000000000000..35ade126a2586a8e3eee6f0ac3c7e49523c8f5c5 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_scale_shift_layer.protostr @@ -0,0 +1,72 @@ +type: "nn" +layers { + name: "data" + type: "data" + size: 100 + active_type: "" +} +layers { + name: "__scale_shift_0__" + type: "scale_shift" + size: 100 + active_type: "" + inputs { + input_layer_name: "data" + input_parameter_name: "___scale_shift_0__.w0" + } +} +layers { + name: "__scale_shift_1__" + type: "scale_shift" + size: 100 + active_type: "" + inputs { + input_layer_name: "data" + input_parameter_name: "___scale_shift_1__.w0" + } + bias_parameter_name: "___scale_shift_1__.wbias" +} +parameters { + name: "___scale_shift_0__.w0" + size: 1 + initial_mean: 0.0 + initial_std: 1.0 + dims: 1 + dims: 1 + initial_strategy: 0 + initial_smart: true +} +parameters { + name: "___scale_shift_1__.w0" + size: 1 + initial_mean: 0.0 + initial_std: 1.0 + dims: 1 + dims: 1 + initial_strategy: 0 + initial_smart: true +} +parameters { + name: "___scale_shift_1__.wbias" + size: 1 + initial_mean: 0.0 + initial_std: 0.0 + dims: 1 + dims: 1 + initial_strategy: 0 + initial_smart: false +} +input_layer_names: "data" +output_layer_names: "__scale_shift_0__" +output_layer_names: "__scale_shift_1__" +sub_models { + name: "root" + layer_names: "data" + layer_names: "__scale_shift_0__" + layer_names: "__scale_shift_1__" + input_layer_names: "data" + output_layer_names: "__scale_shift_0__" + output_layer_names: "__scale_shift_1__" + is_recurrent_layer_group: false +} + diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_slice_layer.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_slice_layer.protostr new file mode 100644 index 0000000000000000000000000000000000000000..5b73d614fe862e74c8dc5c24a776c0020334224c --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_slice_layer.protostr @@ -0,0 +1,79 @@ +type: "nn" +layers { + name: "word" + type: "data" + size: 128 + active_type: "" +} +layers { + name: "starts" + type: "data" + size: 5 + active_type: "" +} +layers { + name: "ends" + type: "data" + size: 5 + active_type: "" +} +layers { + name: "__seq_slice_layer_0__" + type: "seq_slice" + size: 128 + active_type: "" + inputs { + input_layer_name: "word" + } + inputs { + input_layer_name: "starts" + } + inputs { + input_layer_name: "ends" + } +} +layers { + name: "__seq_slice_layer_1__" + type: "seq_slice" + size: 128 + active_type: "" + inputs { + input_layer_name: "word" + } + inputs { + input_layer_name: "starts" + } + select_first: true +} +layers { + name: "__seq_slice_layer_2__" + type: "seq_slice" + size: 128 + active_type: "" + inputs { + input_layer_name: "word" + } + inputs { + input_layer_name: "ends" + } + select_first: false +} +input_layer_names: "word" +output_layer_names: "__seq_slice_layer_0__" +output_layer_names: "__seq_slice_layer_1__" +output_layer_names: "__seq_slice_layer_2__" +sub_models { + name: "root" + layer_names: "word" + layer_names: "starts" + layer_names: "ends" + layer_names: "__seq_slice_layer_0__" + layer_names: "__seq_slice_layer_1__" + layer_names: "__seq_slice_layer_2__" + input_layer_names: "word" + output_layer_names: "__seq_slice_layer_0__" + output_layer_names: "__seq_slice_layer_1__" + output_layer_names: "__seq_slice_layer_2__" + is_recurrent_layer_group: false +} + diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_select_layers.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_sub_nested_seq_select_layer.protostr similarity index 100% rename from python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_select_layers.protostr rename to python/paddle/trainer_config_helpers/tests/configs/protostr/test_sub_nested_seq_select_layer.protostr diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/util_layers.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/util_layers.protostr index d0ad388165007b8f96f059e5b003c52f756383e5..7a2f3eab38808a031c27cf7ab9d6273952e389eb 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/util_layers.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/util_layers.protostr @@ -22,6 +22,9 @@ layers { inputs { input_layer_name: "b" } + height: 0 + width: 0 + depth: 1 } layers { name: "__concat_0__" @@ -34,6 +37,9 @@ layers { inputs { input_layer_name: "b" } + height: 0 + width: 0 + depth: 1 } layers { name: "__concat_1__" diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_BatchNorm3D.py b/python/paddle/trainer_config_helpers/tests/configs/test_BatchNorm3D.py new file mode 100644 index 0000000000000000000000000000000000000000..a991b22252ba10eed895efd931108c2d8b0e52f1 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_BatchNorm3D.py @@ -0,0 +1,11 @@ +from paddle.trainer_config_helpers import * + +settings(batch_size=1000, learning_rate=1e-4) + +#data = data_layer(name='data', size=180, width=30, height=6) +#batchNorm = batch_norm_layer(data, num_channels=1) +#outputs(batchNorm) + +data3D = data_layer(name='data3D', size=120 * 3, width=20, height=6, depth=3) +batchNorm3D = batch_norm_layer(data3D, num_channels=1, img3D=True) +outputs(batchNorm3D) diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_conv3d_layer.py b/python/paddle/trainer_config_helpers/tests/configs/test_conv3d_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..aa0a2c0d5fe19b6c414acd708bb6e82d9fb6568f --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_conv3d_layer.py @@ -0,0 +1,49 @@ +from paddle.trainer_config_helpers import * + +settings(batch_size=1000, learning_rate=1e-5) + +num_channels = 3 +filter_size = 3 +filter_size_y = 3 +filter_size_z = 3 +stride = 2 +stride_y = 2 +stride_z = 2 +padding = 1 +padding_y = 1 +padding_z = 1 +groups = 1 + +data = data_layer( + name='data', size=12096 * num_channels, height=48, width=42, depth=6) +# first +conv3d_1 = img_conv3d_layer( + input=data, + name='conv3d_1', + num_filters=16, + num_channels=num_channels, + filter_size=filter_size, + stride=stride, + padding=padding, + groups=groups, + bias_attr=True, + shared_biases=True, + trans=False, + layer_type="conv3d", + act=LinearActivation()) +# second +conv3d_2 = img_conv3d_layer( + input=data, + name='conv3d_2', + num_filters=16, + num_channels=num_channels, + filter_size=[filter_size, filter_size_y, filter_size_z], + stride=[stride, stride_y, stride_z], + padding=[padding, padding_y, padding_z], + groups=groups, + bias_attr=True, + shared_biases=True, + trans=False, + layer_type="conv3d", + act=LinearActivation()) +outputs(conv3d_2) diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_cost_layers.py b/python/paddle/trainer_config_helpers/tests/configs/test_cost_layers.py index d2a3b702a1d7b650947b344e4719098f68d4dd73..7ce375c708af7b0b7ae1d700dedbdb6a4ce16c7f 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/test_cost_layers.py +++ b/python/paddle/trainer_config_helpers/tests/configs/test_cost_layers.py @@ -33,7 +33,9 @@ outputs( input=probs, label=xe_label), cross_entropy_with_selfnorm( input=probs, label=xe_label), - huber_cost( + huber_regression_cost( + input=seq_in, label=labels), + huber_classification_cost( input=data_layer( name='huber_probs', size=1), label=data_layer( diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_cost_layers_with_weight.py b/python/paddle/trainer_config_helpers/tests/configs/test_cost_layers_with_weight.py index c369062930e2b067ceab0dc3b25ba6c1eabe2450..caa6aaa9430ffaee7ade93ee04ec90103bf8cf43 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/test_cost_layers_with_weight.py +++ b/python/paddle/trainer_config_helpers/tests/configs/test_cost_layers_with_weight.py @@ -10,7 +10,7 @@ fc = fc_layer(input=data, size=10, act=SoftmaxActivation()) outputs( classification_cost( input=fc, label=lbl, weight=wt), - mse_cost( + square_error_cost( input=fc, label=lbl, weight=wt), nce_layer( input=fc, diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_cross_entropy_over_beam.py b/python/paddle/trainer_config_helpers/tests/configs/test_cross_entropy_over_beam.py new file mode 100644 index 0000000000000000000000000000000000000000..4a5bdf1181dc4538418a8b89b41a1ff713e423c8 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_cross_entropy_over_beam.py @@ -0,0 +1,45 @@ +#!/usr/bin/env python +#coding=utf-8 + +from paddle.trainer_config_helpers import * +beam_size = 5 + +# the first beam expansion. +sentence_states = data_layer(name="sentence_states", size=32) +sentence_scores = data_layer(name="sentence_scores", size=1) +topk_sentence_ids = kmax_seq_score_layer( + input=sentence_scores, beam_size=beam_size) + +# the second beam expansion. +topk_sen = sub_nested_seq_layer( + input=sentence_states, selected_indices=topk_sentence_ids) +start_pos_scores = fc_layer(input=topk_sen, size=1, act=LinearActivation()) +topk_start_pos_ids = kmax_seq_score_layer( + input=sentence_scores, beam_size=beam_size) + +# the final beam expansion. +topk_start_spans = seq_slice_layer( + input=topk_sen, starts=topk_start_pos_ids, ends=None) +end_pos_scores = fc_layer( + input=topk_start_spans, size=1, act=LinearActivation()) +topk_end_pos_ids = kmax_seq_score_layer( + input=end_pos_scores, beam_size=beam_size) + +# define the cost +sentence_idx = data_layer(name="sentences_ids", size=1) +start_idx = data_layer(name="start_ids", size=1) +end_idx = data_layer(name="end_ids", size=1) +cost = cross_entropy_over_beam(input=[ + BeamInput( + candidate_scores=sentence_scores, + selected_candidates=topk_sentence_ids, + gold=sentence_idx), BeamInput( + candidate_scores=start_pos_scores, + selected_candidates=topk_start_pos_ids, + gold=start_idx), BeamInput( + candidate_scores=end_pos_scores, + selected_candidates=topk_end_pos_ids, + gold=end_idx) +]) + +outputs(cost) diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_deconv3d_layer.py b/python/paddle/trainer_config_helpers/tests/configs/test_deconv3d_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..a113279fc17b49ad01b8860b61180af0f35694fb --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_deconv3d_layer.py @@ -0,0 +1,50 @@ +from paddle.trainer_config_helpers import * + +settings(batch_size=1000, learning_rate=1e-5) + +num_channels = 3 +filter_size = 3 +filter_size_y = 3 +filter_size_z = 3 +stride = 2 +stride_y = 2 +stride_z = 2 +padding = 1 +padding_y = 1 +padding_z = 1 +groups = 1 + +data = data_layer( + name='data', size=12096 * num_channels, height=48, width=42, depth=6) + +# first +deconv3d_1 = img_conv3d_layer( + input=data, + name='deconv3d_1', + num_filters=16, + num_channels=num_channels, + filter_size=filter_size, + stride=stride, + padding=padding, + groups=groups, + bias_attr=True, + shared_biases=True, + trans=True, + layer_type="deconv3d", + act=LinearActivation()) +# second +deconv3d_2 = img_conv3d_layer( + input=data, + name='deconv3d_2', + num_filters=16, + num_channels=num_channels, + filter_size=[filter_size, filter_size_y, filter_size_z], + stride=[stride, stride_y, stride_z], + padding=[padding, padding_y, padding_z], + groups=groups, + bias_attr=True, + shared_biases=True, + trans=True, + layer_type="deconv3d", + act=LinearActivation()) +outputs(deconv3d_2) diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_kmax_seq_socre_layer.py b/python/paddle/trainer_config_helpers/tests/configs/test_kmax_seq_socre_layer.py index d245c5a41c793e1f02f306bfe64071bd9885906e..171da10f75dae03eed7e110d0efd07d6a18e1ecf 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/test_kmax_seq_socre_layer.py +++ b/python/paddle/trainer_config_helpers/tests/configs/test_kmax_seq_socre_layer.py @@ -2,10 +2,8 @@ #coding=utf-8 from paddle.trainer_config_helpers import * -data = data_layer(name='input', size=300) - -data = data_layer(name="data", size=128) +data = data_layer(name="input_seq", size=128) scores = fc_layer(input=data, size=1, act=ExpActivation()) -kmax_seq_id = kmax_sequence_score_layer(input=scores, beam_size=5) +kmax_seq_id = kmax_seq_score_layer(input=scores, beam_size=5) outputs(kmax_seq_id) diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_pooling3D_layer.py b/python/paddle/trainer_config_helpers/tests/configs/test_pooling3D_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..0dbb921d41986e711d5b8b31caab1f8b6bdc47b8 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_pooling3D_layer.py @@ -0,0 +1,38 @@ +from paddle.trainer_config_helpers import * + +settings(batch_size=100, learning_rate=1e-5) + +data_2d = data_layer(name='data_2d', size=6000, height=20, width=10) + +pool_2d = img_pool_layer( + name="pool___2d", + input=data_2d, + num_channels=30, + pool_size=5, + stride=3, + padding=1, + pool_type=AvgPooling()) +outputs(pool_2d) + +data_3d = data_layer( + name='data_3d_1', size=60000, depth=10, height=20, width=10) + +pool_3d_1 = img_pool3d_layer( + name="pool_3d_1", + input=data_3d, + num_channels=30, + pool_size=5, + stride=3, + padding=1, + pool_type=AvgPooling()) +outputs(pool_3d_1) + +pool_3d_2 = img_pool3d_layer( + name="pool_3d_2", + input=data_3d, + num_channels=30, + pool_size=[5, 5, 5], + stride=[3, 3, 3], + padding=[1, 1, 1], + pool_type=MaxPooling()) +outputs(pool_3d_2) diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_scale_shift_layer.py b/python/paddle/trainer_config_helpers/tests/configs/test_scale_shift_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..dd589116fa9932144ca066d3fa4c929d1433a7f1 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_scale_shift_layer.py @@ -0,0 +1,9 @@ +from paddle.trainer_config_helpers import * + +data = data_layer(name='data', size=100) + +scale = scale_shift_layer(input=data, bias_attr=False) + +scale_shift = scale_shift_layer(input=data) + +outputs(scale, scale_shift) diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_seq_slice_layer.py b/python/paddle/trainer_config_helpers/tests/configs/test_seq_slice_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..510ad3220893fddac278ba691307d00d57e440a3 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_seq_slice_layer.py @@ -0,0 +1,13 @@ +#!/usr/bin/env python +#coding=utf-8 +from paddle.trainer_config_helpers import * + +input_seq = data_layer("word", size=128) +starts = data_layer("starts", size=5) +ends = data_layer("ends", size=5) + +seq_slice1 = seq_slice_layer(input=input_seq, starts=starts, ends=ends) +seq_slice2 = seq_slice_layer(input=input_seq, starts=starts, ends=None) +seq_slice3 = seq_slice_layer(input=input_seq, starts=None, ends=ends) + +outputs(seq_slice1, seq_slice2, seq_slice3) diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_seq_select_layers.py b/python/paddle/trainer_config_helpers/tests/configs/test_sub_nested_seq_select_layer.py similarity index 100% rename from python/paddle/trainer_config_helpers/tests/configs/test_seq_select_layers.py rename to python/paddle/trainer_config_helpers/tests/configs/test_sub_nested_seq_select_layer.py diff --git a/python/paddle/trainer_config_helpers/tests/layers_test.py b/python/paddle/trainer_config_helpers/tests/layers_test.py index 05902ea293df5a3e9c10f6700930ca6a343603c2..b3dd8f8fc784754e749240e1b895b11ef6aba438 100644 --- a/python/paddle/trainer_config_helpers/tests/layers_test.py +++ b/python/paddle/trainer_config_helpers/tests/layers_test.py @@ -17,3 +17,4 @@ from paddle.trainer.config_parser import parse_config_and_serialize if __name__ == '__main__': parse_config_and_serialize( 'trainer_config_helpers/tests/layers_test_config.py', '') +# layers_test_config.py diff --git a/python/paddle/v2/__init__.py b/python/paddle/v2/__init__.py index 5bea980611904b37a4a5d4e2cbbee13503a61ff0..1c8d8f4b2f626bea5d9a44d01de7c2c9c45dc2fb 100644 --- a/python/paddle/v2/__init__.py +++ b/python/paddle/v2/__init__.py @@ -78,6 +78,8 @@ def init(**kwargs): if 'use_gpu' in kwargs: cp.g_command_config_args['use_gpu'] = kwargs['use_gpu'] + if 'use_mkldnn' in kwargs: + cp.g_command_config_args['use_mkldnn'] = kwargs['use_mkldnn'] assert 'parallel_nn' not in kwargs, ("currently 'parallel_nn' is not " "supported in v2 APIs.") diff --git a/python/paddle/v2/event.py b/python/paddle/v2/event.py index 7589cc9917f26375d595e200245d5ba099bc38d7..e66bf67d7949057486eb54c46f39128fad5dae55 100644 --- a/python/paddle/v2/event.py +++ b/python/paddle/v2/event.py @@ -53,10 +53,13 @@ class BeginPass(object): class EndPass(WithMetric): """ Event On One Pass Training Complete. + To get the output of a specific layer, add "event.gm.getLayerOutputs('predict_layer')" + in your event_handler call back """ - def __init__(self, pass_id, evaluator): + def __init__(self, pass_id, evaluator, gm): self.pass_id = pass_id + self.gm = gm WithMetric.__init__(self, evaluator) @@ -73,10 +76,13 @@ class BeginIteration(object): class EndIteration(WithMetric): """ Event On One Batch Training Complete. + To get the output of a specific layer, add "event.gm.getLayerOutputs('predict_layer')" + in your event_handler call back """ - def __init__(self, pass_id, batch_id, cost, evaluator): + def __init__(self, pass_id, batch_id, cost, evaluator, gm): self.pass_id = pass_id self.batch_id = batch_id self.cost = cost + self.gm = gm WithMetric.__init__(self, evaluator) diff --git a/python/paddle/v2/framework/op.py b/python/paddle/v2/framework/op.py index 7fd8b55a5d167294d3270c79f7b64da03443afd3..9e665adad2d3ad91d183c6815fbd7135ac4e8965 100644 --- a/python/paddle/v2/framework/op.py +++ b/python/paddle/v2/framework/op.py @@ -1,64 +1,75 @@ import paddle.v2.framework.core as core -import paddle.v2.framework.proto.op_proto_pb2 as op_proto_pb2 -import paddle.v2.framework.proto.op_desc_pb2 as op_desc_pb2 -import paddle.v2.framework.proto.attribute_pb2 as attribute_pb2 +import paddle.v2.framework.proto.framework_pb2 as framework_pb2 def get_all_op_protos(): """ - Get all registered op proto from Paddle C++ - :return: list of OpProto + Get all registered op proto from PaddlePaddle C++ end. + :return: A list of registered OpProto. """ protostrs = core.get_all_op_protos() ret_values = [] for pbstr in protostrs: - op_proto = op_proto_pb2.OpProto.FromString(str(pbstr)) + op_proto = framework_pb2.OpProto.FromString(str(pbstr)) ret_values.append(op_proto) return ret_values +def is_str(s): + return isinstance(s, str) or isinstance(s, unicode) + + class OpDescCreationMethod(object): """ - A Functor object to convert user input(use key word args) to OpDesc based on - OpProto. - + Convert the user's input(only keyword arguments are supported) to OpDesc + based on the OpProto. + :param op_proto: The OpProto object. :type op_proto: op_proto_pb2.OpProto """ def __init__(self, op_proto): - if not isinstance(op_proto, op_proto_pb2.OpProto): - raise TypeError("Argument should be OpProto") + if not isinstance(op_proto, framework_pb2.OpProto): + raise TypeError( + "Type of op_proto should be OpProto in PaddlePaddle.") self.__op_proto__ = op_proto def __call__(self, *args, **kwargs): """ - Convert user input to OpDesc. Only key-word args are supported. - :return: OpDesc based on user input + Convert user's input to OpDesc. Only keyword arguments are supported. + :return: The OpDesc based on user input. :rtype: op_desc_pb2.OpDesc """ if len(args) != 0: - raise ValueError("Only keyword arguments is supported by Paddle") - op_desc = op_desc_pb2.OpDesc() - - # Inputs - ipts, ipt_format, _ = OpDescCreationMethod.extract_input_or_output( - "input", kwargs, self.__op_proto__.inputs) - op_desc.inputs.extend(ipts) - if ipt_format is not None: - op_desc.attrs.extend([ipt_format]) - - # Outputs - outs, out_format, tmp_index = OpDescCreationMethod.extract_input_or_output( - "output", kwargs, self.__op_proto__.outputs) - op_desc.outputs.extend(outs) - if out_format is not None: - op_desc.attrs.extend([out_format]) - if len(tmp_index) != 0: - tmp_index_attr = op_desc.attrs.add() - tmp_index_attr.type = attribute_pb2.INTS - tmp_index_attr.name = "temporary_index" - tmp_index_attr.ints.extend(tmp_index) + raise ValueError("Only keyword arguments are supported.") + op_desc = framework_pb2.OpDesc() + for input_parameter in self.__op_proto__.inputs: + input_arguments = kwargs.get(input_parameter.name, []) + if is_str(input_arguments): + input_arguments = [input_arguments] + + if not input_parameter.duplicable and len(input_arguments) > 1: + raise ValueError( + "Input %s expects only one input, but %d are given." % + (input_parameter.name, len(input_arguments))) + + ipt = op_desc.inputs.add() + ipt.parameter = input_parameter.name + ipt.arguments.extend(input_arguments) + + for output_parameter in self.__op_proto__.outputs: + output_arguments = kwargs.get(output_parameter.name, []) + if is_str(output_arguments): + output_arguments = [output_arguments] + + if not output_parameter.duplicable and len(output_arguments) > 1: + raise ValueError( + "Output %s expects only one output, but %d are given." % + (output_parameter.name, len(output_arguments))) + + out = op_desc.outputs.add() + out.parameter = output_parameter.name + out.arguments.extend(output_arguments) # Types op_desc.type = self.__op_proto__.type @@ -72,72 +83,35 @@ class OpDescCreationMethod(object): new_attr = op_desc.attrs.add() new_attr.name = attr.name new_attr.type = attr.type - if attr.type == attribute_pb2.INT: + if attr.type == framework_pb2.INT: new_attr.i = user_defined_attr - elif attr.type == attribute_pb2.FLOAT: + elif attr.type == framework_pb2.FLOAT: new_attr.f = user_defined_attr - elif attr.type == attribute_pb2.STRING: + elif attr.type == framework_pb2.STRING: new_attr.s = user_defined_attr - elif attr.type == attribute_pb2.INTS: + elif attr.type == framework_pb2.INTS: new_attr.ints.extend(user_defined_attr) - elif attr.type == attribute_pb2.FLOATS: + elif attr.type == framework_pb2.FLOATS: new_attr.floats.extend(user_defined_attr) - elif attr.type == attribute_pb2.STRINGS: + elif attr.type == framework_pb2.STRINGS: new_attr.strings.extend(user_defined_attr) + elif attr.type == framework_pb2.INT_PAIRS: + for p in user_defined_attr: + pair = new_attr.pairs.add() + pair.first = p[0] + pair.second = p[1] else: - raise NotImplementedError("Not support attribute type " + - attr.type) + raise NotImplementedError( + "A not supported attribute type: %s." % ( + str(attr.type))) return op_desc - @staticmethod - def extract_input_or_output(in_out, kwargs, meta): - """ - Extract input variable names or output variable names from key-word - arguments, which base on VarProtos. - - :param in_out: "input" or "output" - :param kwargs: key-word arguments that user inputted. - :param meta: a list of VarProto - :return: The three object will be return. The variable names. The - input_format or output_format attribute(None if the input or output is - not multiple). The temporary variable index list. - """ - multiple = OpDescCreationMethod.any_is_true((m.multiple for m in meta)) - tmp_index = [] - retv = [] - if multiple: - var_format = op_desc_pb2.AttrDesc() - var_format.type = attribute_pb2.INTS - var_format.name = "%s_format" % in_out - var_format.ints.append(0) - - for var in meta: - var_name = var.name - - if var.temporary: - var_name = [core.var_names.temp()] - tmp_index.append(len(retv)) - else: - var_name = kwargs.get(var_name, []) - if not isinstance(var_name, list): - var_name = [var_name] - retv.extend(var_name) - var_format.ints.append(len(var_name) + var_format.ints[-1]) - return retv, var_format, tmp_index - else: - for var in meta: - if var.temporary: - retv.append(kwargs.get(var.name, core.var_names.temp())) - tmp_index.append(len(retv)) - else: - retv.append(kwargs.get(var.name, core.var_names.empty())) - return retv, None, tmp_index - @staticmethod def any_is_true(generator): """ - Reduce a bool array to one. If any of them is True, then return True. + Reduce a boolean array to a single boolean parameter. If any element in + the array is True, this function will return True, otherwise False. """ for flag in generator: if flag: @@ -146,18 +120,17 @@ class OpDescCreationMethod(object): class OpInfo(object): - def __init__(self, name, method, inputs, outputs, attrs, no_temp_outputs): + def __init__(self, name, method, inputs, outputs, attrs): self.name = name self.method = method self.inputs = inputs self.outputs = outputs self.attrs = attrs - self.no_temp_outputs = no_temp_outputs def create_op_creation_method(op_proto): """ - Generate op creation method for an OpProto + Generate op creation method for an OpProto. """ method = OpDescCreationMethod(op_proto) @@ -168,31 +141,31 @@ def create_op_creation_method(op_proto): return OpInfo( method=__impl__, name=op_proto.type, - inputs=[var.name for var in op_proto.inputs], - outputs=[var.name for var in op_proto.outputs], - attrs=[attr.name for attr in op_proto.attrs], - no_temp_outputs=[ - var.name for var in op_proto.outputs if not var.temporary - ]) + inputs=[(var.name, var.duplicable) for var in op_proto.inputs], + outputs=[(var.name, var.duplicable) for var in op_proto.outputs], + attrs=[attr.name for attr in op_proto.attrs]) class OperatorFactory(object): def __init__(self): self.op_methods = dict() + for op_proto in get_all_op_protos(): method = create_op_creation_method(op_proto) self.op_methods[method.name] = method def __call__(self, *args, **kwargs): - if 'type' in kwargs: + if "type" in kwargs: if len(args) != 0: - raise ValueError("All Paddle argument should be key-word " - "argument except type") - t = kwargs.pop('type') + raise ValueError( + "Except the argument \"type\"," + "all of the other arguments should be keyword arguments.") + t = kwargs.pop("type") else: if len(args) != 1: - raise ValueError("All Paddle argument should be key-word " - "argument except type") + raise ValueError( + "Except the argument \"type\"," + "all of the other arguments should be keyword arguments.") t = args[0] return self.get_op_info(t).method(**kwargs) @@ -202,20 +175,45 @@ class OperatorFactory(object): def get_op_info(self, t): if t not in self.op_methods: - raise ValueError("operator %s is not registered", t) + raise ValueError("The operator: %s is not registered." % t) return self.op_methods.get(t) def get_op_input_names(self, type): + return map(lambda x: x[0], self.get_op_info(type).inputs) + + def get_op_inputs(self, type): return self.get_op_info(type).inputs def get_op_output_names(self, type): + return map(lambda x: x[0], self.get_op_info(type).outputs) + + def get_op_outputs(self, type): return self.get_op_info(type).outputs def get_op_attr_names(self, type): return self.get_op_info(type).attrs - def get_op_no_temp_output_names(self, type): - return self.get_op_info(type).no_temp_outputs + +class __RecurrentOp__(object): + __proto__ = None + type = "recurrent" + + def __init__(self): + # cache recurrent_op's proto + if self.__proto__ is None: + for op_proto in get_all_op_protos(): + if op_proto.type == self.type: + self.__proto__ = op_proto + + def __call__(self, *args, **kwargs): + if self.type not in args and "type" not in kwargs: + kwargs["type"] = self.type + # create proto + create_method = OpDescCreationMethod(self.__proto__) + proto = create_method(*args, **kwargs) + # create rnnop + return core.RecurrentOp.create(proto.SerializeToString()) -Operator = OperatorFactory() # Default global factory +Operator = OperatorFactory() # The default global factory +RecurrentOp = __RecurrentOp__() diff --git a/python/paddle/v2/framework/tests/CMakeLists.txt b/python/paddle/v2/framework/tests/CMakeLists.txt index 96fad9b42e04a88fdcbda093683b57451b2a3e41..3de9e69e34d3d2be53b597d489323466a0fe4033 100644 --- a/python/paddle/v2/framework/tests/CMakeLists.txt +++ b/python/paddle/v2/framework/tests/CMakeLists.txt @@ -4,6 +4,7 @@ py_test(test_scope SRCS test_scope.py) py_test(test_tensor SRCS test_tensor.py) py_test(test_mul_op SRCS test_mul_op.py) +py_test(test_cos_sim_op SRCS test_cos_sim_op.py) py_test(test_mean_op SRCS test_mean_op.py) @@ -13,15 +14,24 @@ py_test(test_add_two_op SRCS test_add_two_op.py) py_test(test_sigmoid_op SRCS test_sigmoid_op.py) py_test(test_softmax_op SRCS test_softmax_op.py) py_test(test_cross_entropy_op SRCS test_cross_entropy_op.py) +py_test(test_gather_op SRCS test_gather_op.py) +py_test(test_scatter_op SRCS test_scatter_op.py) py_test(test_fill_zeros_like_op SRCS test_fill_zeros_like_op.py) - -py_test(gradient_checker SRCS gradient_checker.py) +py_test(test_top_k_op SRCS test_top_k_op.py) py_test(test_rowwise_add_op SRCS test_rowwise_add_op.py) py_test(test_default_scope_funcs SRCS test_default_scope_funcs.py) py_test(test_operator SRCS test_operator.py) -# py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py) +py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py) py_test(test_uniform_random_op SRCS test_uniform_random_op.py) py_test(test_recurrent_op SRCS test_recurrent_op.py) +py_test(test_sgd_op SRCS test_sgd_op.py) +py_test(test_gradient_checker SRCS test_gradient_checker.py) +py_test(test_lookup_table SRCS test_lookup_table.py) +py_test(test_scale_and_identity_op SRCS test_scale_and_identity_op.py) +py_test(test_sum_op SRCS test_sum_op.py) +py_test(mnist SRCS mnist.py) +py_test(test_concat_op SRCS test_concat_op.py) +py_test(test_squared_l2_distance_op SRCS test_squared_l2_distance_op.py) diff --git a/python/paddle/v2/framework/tests/gradient_checker.py b/python/paddle/v2/framework/tests/gradient_checker.py deleted file mode 100644 index 015e832e82560bb8b3518cbdf605c705d77cdd99..0000000000000000000000000000000000000000 --- a/python/paddle/v2/framework/tests/gradient_checker.py +++ /dev/null @@ -1,256 +0,0 @@ -import unittest - -import numpy -import paddle.v2.framework.core as core -from paddle.v2.framework.op import Operator - -__all__ = ['get_numeric_gradient'] - - -def create_op(op_type): - kwargs = dict() - for in_name in Operator.get_op_input_names(op_type): - kwargs[in_name] = in_name - for out_name in Operator.get_op_output_names(op_type): - kwargs[out_name] = out_name - - return Operator(op_type, **kwargs) - - -def grad_var_name(var_name): - return var_name + "@GRAD" - - -def get_numeric_gradient(op, - input_values, - output_name, - input_to_check, - delta=0.005, - local_scope=None): - """ - 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 output_name: The final output variable name. - :param input_to_check: The input variable need to get 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. - :param local_scope: The local scope used for get_numeric_gradient. - :return: The gradient array in numpy format. - """ - if local_scope is None: - local_scope = core.Scope() - - # Create all input variable in local_scope - for var_name in input_values: - var = local_scope.new_var(var_name) - tensor = var.get_tensor() - tensor.set_dims(input_values[var_name].shape) - tensor.alloc_float(core.CPUPlace()) - tensor.set(input_values[var_name], core.CPUPlace()) - - # Create all output variable in local_scope - for output in op.outputs(): - if local_scope.find_var(output) is None: - local_scope.new_var(output).get_tensor() - - op.infer_shape(local_scope) - - # allocate output memory - for output in op.outputs(): - local_scope.find_var(output).get_tensor().alloc_float(core.CPUPlace()) - - # TODO(yuyang18): Only CPU is support now. - cpu_ctx = core.DeviceContext.create(core.CPUPlace()) - - def get_output(): - op.run(local_scope, cpu_ctx) - return numpy.array(local_scope.find_var(output_name).get_tensor()).sum() - - def product(dim): - return reduce(lambda a, b: a * b, dim, 1) - - # get the input tensor that we want to get it's numeric gradient. - tensor_to_check = local_scope.find_var(input_to_check).get_tensor() - tensor_size = product(tensor_to_check.get_dims()) - # prepare a numpy array to store the gradient. - gradient_flat = numpy.zeros(shape=(tensor_size, ), dtype='float32') - - # we only compute gradient of one element each time. - # we use a for loop to compute the gradient of every element. - for i in xrange(tensor_size): - # get one input element throw it's 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. - 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. - x_neg = origin - delta - tensor_to_check.set_float_element(i, x_neg) - y_neg = get_output() - - # restore old value - tensor_to_check.set_float_element(i, origin) - - # compute the gradient of this element and store it into a numpy array. - gradient_flat[i] = (y_pos - y_neg) / delta / 2 - - # reshape the gradient result to the shape of the source tensor. - return gradient_flat.reshape(tensor_to_check.get_dims()) - - -class GradientChecker(unittest.TestCase): - def assert_is_close(self, numeric_grads, scope, max_relative_error, - msg_prefix): - for name in numeric_grads: - b = numpy.array(scope.find_var(grad_var_name(name)).get_tensor()) - a = numeric_grads[name] - - abs_a = numpy.abs(a) - # if abs_a is nearly zero, then use abs error for a, not relative - # error. - abs_a[abs_a < 1e-3] = 1 - - diff_mat = numpy.abs(a - b) / abs_a - max_diff = numpy.max(diff_mat) - - def err_msg(): - offset = numpy.argmax(diff_mat > max_relative_error) - return "%s Variable %s max gradient diff %f over limit %f, the first " \ - "error element is %d" % ( - msg_prefix, name, max_diff, max_relative_error, offset) - - self.assertLessEqual(max_diff, max_relative_error, err_msg()) - - def check_grad(self, - forward_op, - input_vars, - inputs_to_check, - output_name, - no_grad_set=None, - only_cpu=False, - max_relative_error=0.005): - """ - :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 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. - :return: - """ - if no_grad_set is None: - no_grad_set = set() - - tmp_outs = forward_op.temp_outputs() - no_tmp_out = filter(lambda name: name not in tmp_outs, - forward_op.outputs()) - if len(no_tmp_out) != 1: - raise ValueError("non temp out_names should be 1") - - in_names = forward_op.inputs() - for no_grad in no_grad_set: - if no_grad not in in_names: - raise ValueError("no_grad should be in in_names") - - backward_op = core.Operator.backward(forward_op, no_grad_set) - - places = [core.CPUPlace()] - if not only_cpu and core.is_compile_gpu() and backward_op.support_gpu(): - places.append(core.GPUPlace(0)) - - numeric_grad = dict() - # get numeric gradient - for check_name in inputs_to_check: - numeric_grad[check_name] = \ - get_numeric_gradient(forward_op, input_vars, output_name, - check_name) - - # get operator gradient according to different device - for place in places: - scope = core.Scope() - ctx = core.DeviceContext.create(place) - - # create input var and set value - for name, value in input_vars.iteritems(): - if name not in in_names: - raise ValueError(name + " not in op.inputs_") - var = scope.new_var(name).get_tensor() - var.set_dims(value.shape) - var.set(value, place) - - # create output var - for out_name in forward_op.outputs(): - scope.new_var(out_name).get_tensor() - - # infer the shape of output var and compute/set value of output var - forward_op.infer_shape(scope) - forward_op.run(scope, ctx) - - # create output grad var - # set shape as the output var - # set value of this grad to ones - for name in forward_op.outputs(): - out_tensor = scope.find_var(name).get_tensor() - grad_tensor = scope.new_var(grad_var_name(name)).get_tensor() - grad_tensor.set_dims(out_tensor.shape()) - data = 1.0 * numpy.ones(out_tensor.shape()) - grad_tensor.set(data, place) - - # create input grad var - for name in backward_op.outputs(): - scope.new_var(name).get_tensor() - - # infer the shape of input gradient var and compute/set it's value - # with backward op - backward_op.infer_shape(scope) - backward_op.run(scope, ctx) - - self.assert_is_close(numeric_grad, scope, max_relative_error, - "Gradient Check On %s" % str(place)) - - -if __name__ == '__main__': - - class GetNumericGradientTest(unittest.TestCase): - def test_add_op(self): - add_op = Operator('add_two', X="X", Y="Y", Out="Z") - x = numpy.random.random((10, 1)).astype("float32") - y = numpy.random.random((10, 1)).astype("float32") - - arr = get_numeric_gradient(add_op, {'X': x, "Y": y}, 'Z', 'X') - self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-2) - - def test_softmax_op(self): - def stable_softmax(x): - """Compute the softmax of vector x in a numerically stable way.""" - shiftx = x - numpy.max(x) - exps = numpy.exp(shiftx) - return exps / numpy.sum(exps) - - def label_softmax_grad(Y, dY): - dX = Y * 0.0 - for i in range(Y.shape[0]): - d = numpy.dot(Y[i, :], dY[i, :]) - dX[i, :] = Y[i, :] * (dY[i, :] - d) - return dX - - softmax_op = Operator("softmax", X="X", Y="Y") - - X = numpy.random.random((2, 2)).astype("float32") - Y = numpy.apply_along_axis(stable_softmax, 1, X) - dY = numpy.ones(Y.shape) - dX = label_softmax_grad(Y, dY) - - arr = get_numeric_gradient(softmax_op, {"X": X}, 'Y', 'X') - numpy.testing.assert_almost_equal(arr, dX, decimal=1e-2) - - unittest.main() diff --git a/python/paddle/v2/framework/tests/mnist.py b/python/paddle/v2/framework/tests/mnist.py new file mode 100644 index 0000000000000000000000000000000000000000..f6f8f49b797fb6e5016a5e309f12f192d5096431 --- /dev/null +++ b/python/paddle/v2/framework/tests/mnist.py @@ -0,0 +1,252 @@ +import paddle.v2.framework.core as core +from paddle.v2.framework.op import Operator +import numpy +import paddle.v2 as paddle + +BATCH_SIZE = 100 + +scope = core.Scope() +place = core.CPUPlace() +# if you want to test GPU training, you can use gpu place +# place = core.GPUPlace(0) +dev_ctx = core.DeviceContext.create(place) + +init_net = core.Net.create() +forward_net = core.Net.create() +backward_net = None +optimize_net = core.Net.create() + + +def atomic_id(): + id = 0 + while True: + yield id + id += 1 + + +uniq_id = atomic_id().next + + +def data_layer(name, dims): + var = scope.new_var(name) + tensor = var.get_tensor() + tensor.set_dims(dims) # 1 is batch size holder. + return name + + +def feed_data(name, data): + assert isinstance(data, numpy.ndarray) + tensor = scope.find_var(name).get_tensor() + tensor.set_dims(data.shape) + if data.dtype == numpy.dtype("int32"): + tensor.alloc_int(place) + elif data.dtype == numpy.dtype("float32"): + tensor.alloc_float(place) + else: + raise ValueError("data type not supported") + tensor.set(data, place) + + +def grad_var_name(var_name): + return var_name + "@GRAD" + + +def sgd_optimizer(net, param_name, learning_rate=0.005): + grad_name = grad_var_name(param_name) + optimize_op = Operator( + "sgd", + param=param_name, + grad=grad_name, + param_out=param_name, + learning_rate=learning_rate) + net.append_op(optimize_op) + + +# should use operator and add these to the init_network +def init_param(net, param_name, dims): + scope.new_var(param_name) + op = Operator( + "uniform_random", Out=param_name, dims=dims, min=-0.5, max=0.5, seed=10) + op.infer_shape(scope) + net.append_op(op) + + +# fc_layer +def fc_layer(net, input, size, act="softmax", bias=True, param=None, name=None): + """ + The fully connected layer. + + :param input: The name of input variable. + :type input: str + :param size: The size of fully connected layer. + :param act: The name of activation. + :param param: The attribute of learnable parameter which can be used to + modify initialization mean and std of the parameter. + :param bias: The attribute of bias. If set False, this layer does not have + a bias. + :param name: The name of this layer. If it is not set explictly, a name + will be generated automatically. + :return: The name of the output variable. + """ + + if name is None: + name = "fc_%d" % uniq_id() + if not isinstance(name, str): + raise ValueError("The name of a layer should be a string.") + + input_dims = scope.find_var(input).get_tensor().get_dims() + + w_name = param or name + ".w" + init_param(net=init_net, param_name=w_name, dims=[input_dims[1], size]) + sgd_optimizer(net=optimize_net, param_name=w_name, learning_rate=0.01) + + pre_activation = name + ".mul.out" + scope.new_var(pre_activation) + mul_op = Operator("mul", X=input, Y=w_name, Out=pre_activation) + net.append_op(mul_op) + + # create bias variable if needed + if bias: + bias_name = name + ".b" + init_param(net=init_net, param_name=bias_name, dims=[size]) + sgd_optimizer( + net=optimize_net, param_name=bias_name, learning_rate=0.001) + bias_out = name + ".rowwise_add.out" + scope.new_var(bias_out) + rowwise_append_op = Operator( + "rowwise_add", X=pre_activation, b=bias_name, Out=bias_out) + net.append_op(rowwise_append_op) + pre_activation = bias_out + + activation_op = Operator(act, X=pre_activation, Y=name) + net.append_op(activation_op) + scope.new_var(name) + net.infer_shape(scope) + return name + + +def cross_entropy_layer(net, input, label): + cost_name = "cross_entropy_%d" % uniq_id() + cross_entropy_op = Operator( + "onehot_cross_entropy", X=input, label=label, Y=cost_name) + net.append_op(cross_entropy_op) + scope.new_var(cost_name) + net.infer_shape(scope) + return cost_name + + +def create_backward_net(forward_net): + net = core.Operator.backward(forward_net, set()) + for input in net.inputs()["all"]: + var = scope.new_var(input) + var.get_tensor() + for output in net.outputs()["all"]: + var = scope.new_var(output) + var.get_tensor() + return net + + +def debug_print_op(op): + print("===============" + op.type() + "==============") + print("***inputs:***") + for input in op.inputs()["all"]: + print input, scope.find_var(input).get_tensor().get_dims() + print("\n***outputs:***") + for output in op.outputs()["all"]: + print output, scope.find_var(output).get_tensor().get_dims() + print("") + print("") + + +def set_cost(cost): + cost_shape = numpy.array(scope.find_var(cost).get_tensor()).shape + cost_grad = \ + scope.find_var(grad_var_name(cost)).get_tensor() + cost_grad.set_dims(cost_shape) + cost_grad.alloc_float(place) + cost_grad.set(numpy.ones(cost_shape).astype("float32"), place) + + +def get_cost_mean(cost): + cost_data = numpy.array(scope.find_var(cost).get_tensor()) + return cost_data.sum() / len(cost_data) + + +def error_rate(predict, label): + predict_var = numpy.array(scope.find_var(predict).get_tensor()).argmax( + axis=1) + label = numpy.array(scope.find_var(label).get_tensor()) + error_num = numpy.sum(predict_var != label) + return error_num / float(len(label)) + + +images = data_layer(name="pixel", dims=[BATCH_SIZE, 784]) +labels = data_layer(name="label", dims=[BATCH_SIZE]) +fc1 = fc_layer(net=forward_net, input=images, size=100, act="sigmoid") +fc2 = fc_layer(net=forward_net, input=fc1, size=100, act="sigmoid") +predict = fc_layer(net=forward_net, input=fc2, size=10, act="softmax") +cost = cross_entropy_layer(net=forward_net, input=predict, label=labels) + +init_net.complete_add_op(True) +forward_net.complete_add_op(True) +backward_net = create_backward_net(forward_net) +optimize_net.complete_add_op(True) + +print(init_net) +print(forward_net) +print(backward_net) +print(optimize_net) + +debug_print_op(forward_net) +debug_print_op(backward_net) +debug_print_op(optimize_net) + +train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=8192), + batch_size=BATCH_SIZE) + + +def test(cost_name): + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=BATCH_SIZE) + cost = [] + error = [] + for data in test_reader(): + image_data = numpy.array(map(lambda x: x[0], data)).astype("float32") + label_data = numpy.array(map(lambda x: x[1], data)).astype("int32") + feed_data(images, image_data) + feed_data(labels, label_data) + + forward_net.infer_shape(scope) + forward_net.run(scope, dev_ctx) + cost.append(get_cost_mean(cost_name)) + error.append(error_rate(predict, "label")) + print("cost=" + str(sum(cost) / float(len(cost))) + " error_rate=" + str( + sum(error) / float(len(error)))) + + +PASS_NUM = 1 + +init_net.run(scope, dev_ctx) +for pass_id in range(PASS_NUM): + batch_id = 0 + + for data in train_reader(): + image_data = numpy.array(map(lambda x: x[0], data)).astype("float32") + label_data = numpy.array(map(lambda x: x[1], data)).astype("int32") + feed_data(images, image_data) + feed_data(labels, label_data) + + forward_net.infer_shape(scope) + forward_net.run(scope, dev_ctx) + set_cost(cost) + backward_net.infer_shape(scope) + backward_net.run(scope, dev_ctx) + + optimize_net.run(scope, dev_ctx) + if batch_id % 100 == 0: + print("pass[" + str(pass_id) + "] batch_id[" + str(batch_id) + "]") + test(cost) + + batch_id = batch_id + 1 diff --git a/python/paddle/v2/framework/tests/op_test.py b/python/paddle/v2/framework/tests/op_test.py new file mode 100644 index 0000000000000000000000000000000000000000..4fec4c9109bf247abb2068177583acb47a8ebd97 --- /dev/null +++ b/python/paddle/v2/framework/tests/op_test.py @@ -0,0 +1,281 @@ +import unittest +import numpy as np +import itertools +import paddle.v2.framework.core as core +from paddle.v2.framework.op import Operator + + +def grad_var_name(var_name): + return var_name + "@GRAD" + + +def create_op(scope, op_type, inputs, outputs, attrs): + kwargs = dict() + + for in_name, in_dup in Operator.get_op_inputs(op_type): + if in_name in inputs: + kwargs[in_name] = [] + if in_dup: + sub_in = inputs[in_name] + for sub_in_name, _ in sub_in: + var = scope.new_var(sub_in_name) + kwargs[in_name].append(sub_in_name) + else: + var = scope.new_var(in_name) + kwargs[in_name].append(in_name) + + for out_name, out_dup in Operator.get_op_outputs(op_type): + if out_name in outputs: + kwargs[out_name] = [] + if out_dup: + sub_in = outputs[out_name] + for sub_in_name, _ in sub_in: + var = scope.new_var(sub_in_name) + kwargs[out_name].append(sub_in_name) + else: + var = scope.new_var(out_name) + kwargs[out_name].append(out_name) + + for attr_name in Operator.get_op_attr_names(op_type): + if attr_name in attrs: + kwargs[attr_name] = attrs[attr_name] + return Operator(op_type, **kwargs) + + +def set_input(scope, op, inputs, place): + for in_name, in_dup in Operator.get_op_inputs(op.type()): + if in_name in inputs: + if in_dup: + sub_in = inputs[in_name] + for sub_in_name, sub_in_array in sub_in: + var = scope.find_var(sub_in_name) + tensor = var.get_tensor() + tensor.set_dims(sub_in_array.shape) + tensor.set(sub_in_array, place) + else: + var = scope.find_var(in_name) + tensor = var.get_tensor() + arr = inputs[in_name] + tensor.set_dims(arr.shape) + tensor.set(arr, place) + + +def set_output_grad(scope, op, outputs, place): + for out_name, out_dup in Operator.get_op_outputs(op.type()): + if out_name in outputs: + if out_dup: + sub_out = outputs[out_name] + for sub_out_name, _ in sub_out: + out_tensor = scope.find_var(sub_out_name).get_tensor() + grad_tensor = scope.new_var(grad_var_name( + sub_out_name)).get_tensor() + grad_tensor.set_dims(out_tensor.shape()) + data = np.ones(out_tensor.shape(), dtype=np.float32) + grad_tensor.set(data, place) + else: + out_tensor = scope.find_var(out_name).get_tensor() + grad_tensor = scope.new_var(grad_var_name(out_name)).get_tensor( + ) + grad_tensor.set_dims(out_tensor.shape()) + data = np.ones(out_tensor.shape(), dtype=np.float32) + grad_tensor.set(data, place) + + +def get_numeric_gradient(scope, + op, + inputs, + input_to_check, + output_name, + delta=0.005, + in_place=False): + + set_input(scope, op, inputs, core.CPUPlace()) + op.infer_shape(scope) + + tensor_to_check = scope.find_var(input_to_check).get_tensor() + + def product(dim): + return reduce(lambda a, b: a * b, dim, 1) + + ctx = core.DeviceContext.create(core.CPUPlace()) + + def get_output(): + op.run(scope, ctx) + return np.array(scope.find_var(output_name).get_tensor()).sum() + + tensor_to_check = scope.find_var(input_to_check).get_tensor() + tensor_size = product(tensor_to_check.get_dims()) + gradient_flat = np.zeros(shape=(tensor_size, ), dtype='float32') + # we only compute gradient of one element each time. + # we use a for loop to compute the gradient of every element. + for i in xrange(tensor_size): + if in_place: + set_input(scope, op, inputs, core.CPUPlace()) + + # get one input element throw it's 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. + x_pos = origin + delta + tensor_to_check.set_float_element(i, x_pos) + y_pos = get_output() + + if in_place: + set_input(scope, op, inputs, core.CPUPlace()) + + x_neg = origin - delta + tensor_to_check.set_float_element(i, x_neg) + y_neg = get_output() + + tensor_to_check.set_float_element(i, origin) + gradient_flat[i] = (y_pos - y_neg) / delta / 2 + + return gradient_flat.reshape(tensor_to_check.get_dims()) + + +def get_backward_op(scope, op, no_grad_set): + backward_op = core.Operator.backward(op, no_grad_set) + for input in backward_op.input_vars(): + var = scope.new_var(input) + var.get_tensor() + for output in backward_op.output_vars(): + var = scope.new_var(output) + var.get_tensor() + return backward_op + + +def get_gradient(scope, op, inputs, outputs, grad_name, place, + no_grad_set=None): + ctx = core.DeviceContext.create(place) + + set_input(scope, op, inputs, place) + + op.infer_shape(scope) + op.run(scope, ctx) + + if no_grad_set is None: + no_grad_set = set() + + backward_op = get_backward_op(scope, op, no_grad_set) + set_output_grad(scope, op, outputs, place) + + backward_op.infer_shape(scope) + backward_op.run(scope, ctx) + + out = np.array(scope.find_var(grad_name).get_tensor()) + return out + + +class OpTest(unittest.TestCase): + def check_output_with_place(self, place): + self.scope = core.Scope() + op_inputs = self.inputs if hasattr(self, "inputs") else dict() + op_attrs = self.attrs if hasattr(self, "attrs") else dict() + self.op = create_op(self.scope, self.op_type, op_inputs, self.outputs, + op_attrs) + if isinstance(place, core.GPUPlace) and not self.op.support_gpu(): + return + set_input(self.scope, self.op, self.inputs, place) + self.op.infer_shape(self.scope) + ctx = core.DeviceContext.create(place) + self.op.run(self.scope, ctx) + + for out_name, out_dup in Operator.get_op_outputs(self.op.type()): + if out_dup: + sub_out = self.outputs[out_name] + for sub_out_name in sub_out: + actual = np.array( + self.scope.find_var(sub_out_name).get_tensor()) + expect = sub_out[sub_out_name] + self.assertTrue( + np.allclose( + actual, expect, atol=1e-05), + "output name: " + out_name + "has diff") + else: + actual = np.array(self.scope.find_var(out_name).get_tensor()) + expect = self.outputs[out_name] + self.assertTrue( + np.allclose( + actual, expect, atol=1e-05), + "output name: " + out_name + "has diff") + + def check_output(self): + places = [core.CPUPlace()] + if core.is_compile_gpu(): + places.append(core.GPUPlace(0)) + for place in places: + self.check_output_with_place(place) + + def __assert_is_close(self, numeric_grads, analytic_grads, names, + max_relative_error, msg_prefix): + + for a, b, name in itertools.izip(numeric_grads, analytic_grads, names): + abs_a = np.abs(a) + abs_a[abs_a < 1e-3] = 1 + + diff_mat = np.abs(a - b) / abs_a + max_diff = np.max(diff_mat) + + def err_msg(): + offset = np.argmax(diff_mat > max_relative_error) + return "%s Variable %s max gradient diff %f over limit %f, the first " \ + "error element is %d" % ( + msg_prefix, name, max_diff, max_relative_error, offset) + + self.assertLessEqual(max_diff, max_relative_error, err_msg()) + + def check_grad(self, + inputs_to_check, + output_name, + no_grad_set=None, + in_place=False, + max_relative_error=0.005): + self.scope = core.Scope() + op_inputs = self.inputs if hasattr(self, "inputs") else dict() + op_attrs = self.attrs if hasattr(self, "attrs") else dict() + self.op = create_op(self.scope, self.op_type, op_inputs, self.outputs, + op_attrs) + if no_grad_set is None: + no_grad_set = set() + + numeric_grads = [ + get_numeric_gradient( + self.scope, + self.op, + self.inputs, + input_to_check, + output_name, + in_place=in_place) for input_to_check in inputs_to_check + ] + grad_names = [ + grad_var_name(input_to_check) for input_to_check in inputs_to_check + ] + + cpu_place = core.CPUPlace() + cpu_analytic_grads = [ + get_gradient(self.scope, self.op, self.inputs, self.outputs, + grad_name, cpu_place, no_grad_set) + for grad_name in grad_names + ] + + self.__assert_is_close(numeric_grads, cpu_analytic_grads, grad_names, + max_relative_error, + "Gradient Check On %s" % str(cpu_place)) + + if core.is_compile_gpu() and self.op.support_gpu(): + gpu_place = core.GPUPlace(0) + gpu_analytic_grads = [ + get_gradient(self.scope, self.op, self.inputs, self.outputs, + grad_name, gpu_place, no_grad_set) + for grad_name in grad_names + ] + + self.__assert_is_close(numeric_grads, gpu_analytic_grads, + grad_names, max_relative_error, + "Gradient Check On %s" % str(gpu_place)) + + for c_grad, g_grad, name in itertools.izip( + cpu_analytic_grads, gpu_analytic_grads, grad_names): + self.assertTrue( + np.allclose( + c_grad, g_grad, atol=1e-4), + "output name: " + name + " has diff") diff --git a/python/paddle/v2/framework/tests/op_test_util.py b/python/paddle/v2/framework/tests/op_test_util.py deleted file mode 100644 index dd65e0f2dc23d3f657ff16c55fb297dae210b2d7..0000000000000000000000000000000000000000 --- a/python/paddle/v2/framework/tests/op_test_util.py +++ /dev/null @@ -1,71 +0,0 @@ -import numpy -import paddle.v2.framework.core as core -from paddle.v2.framework.op import Operator - - -class OpTestMeta(type): - """ - Operator Test ClassMeta. - - It injects `test_all` method into user's OperatorTest class, to make Python - unittest module run that method. - - The `test_all` read what value is stored in `self`. It use self's values to - create and run a operator, and check whether that op is OK or not. - - See `test_add_two_op` for example usage. - """ - - def __new__(cls, name, bases, attrs): - obj = super(OpTestMeta, cls).__new__(cls, name, bases, attrs) - - def test_all(self): - scope = core.Scope() - kwargs = dict() - places = [core.CPUPlace()] - if core.is_compile_gpu(): - places.append(core.GPUPlace(0)) - - for place in places: - for in_name in Operator.get_op_input_names(self.type): - if hasattr(self, "inputs") and in_name in self.inputs: - kwargs[in_name] = in_name - var = scope.new_var(in_name).get_tensor() - arr = self.inputs[in_name] - var.set_dims(arr.shape) - var.set(arr, place) - else: - kwargs[in_name] = "@EMPTY@" - - for out_name in Operator.get_op_output_names(self.type): - if not hasattr(self, "outputs"): - raise ValueError( - "The test op must set self.outputs dict.") - if out_name not in self.outputs: - raise ValueError("The %s is not in self.outputs dict." % - (out_name)) - kwargs[out_name] = out_name - scope.new_var(out_name).get_tensor() - - for attr_name in Operator.get_op_attr_names(self.type): - if hasattr(self, "attrs") and attr_name in self.attrs: - kwargs[attr_name] = self.attrs[attr_name] - - op = Operator(self.type, **kwargs) - if isinstance(place, core.GPUPlace) and not op.support_gpu(): - return - - op.infer_shape(scope) - - ctx = core.DeviceContext.create(place) - op.run(scope, ctx) - - for out_name in Operator.get_op_output_names(self.type): - actual = numpy.array(scope.find_var(out_name).get_tensor()) - expect = self.outputs[out_name] - self.assertTrue( - numpy.allclose(actual, expect), - "output name: " + out_name + "has diff") - - obj.test_all = test_all - return obj diff --git a/python/paddle/v2/framework/tests/test_add_two_op.py b/python/paddle/v2/framework/tests/test_add_two_op.py index c0237830647371e14b755953345965a3eac7bfd2..3ca34d9b9fc2b7b54cc25ca0e0d1a08a71e37c52 100644 --- a/python/paddle/v2/framework/tests/test_add_two_op.py +++ b/python/paddle/v2/framework/tests/test_add_two_op.py @@ -1,32 +1,20 @@ import unittest +import numpy as np +from op_test import OpTest -import numpy -import paddle.v2.framework.core as core -from paddle.v2.framework.op import Operator - -from op_test_util import OpTestMeta - - -class TestAddOp(unittest.TestCase): - __metaclass__ = OpTestMeta +class TestAddOp(OpTest): def setUp(self): - self.type = "add_two" + self.op_type = "add" self.inputs = { - 'X': numpy.random.random((102, 105)).astype("float32"), - 'Y': numpy.random.random((102, 105)).astype("float32") + 'X': np.random.random((102, 105)).astype("float32"), + 'Y': np.random.random((102, 105)).astype("float32") } self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']} - -class TestAddGradOp(unittest.TestCase): - def test_add_grad(self): - op = Operator('add_two', X="X", Y="Y", Out="Out") - backward_op = core.Operator.backward(op, set()) - self.assertEqual(backward_op.type(), "add_two_grad") - expected = '''Op(add_two_grad), inputs:(X, Y, Out, Out@GRAD), outputs:(X@GRAD, Y@GRAD).''' - self.assertEqual(expected, str(backward_op)) + def test_check_output(self): + self.check_output() -if __name__ == '__main__': +if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/framework/tests/test_concat_op.py b/python/paddle/v2/framework/tests/test_concat_op.py new file mode 100644 index 0000000000000000000000000000000000000000..656563f96e52df30951ec0ec7042ad9c530e90b2 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_concat_op.py @@ -0,0 +1,22 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestConcatOp(OpTest): + def setUp(self): + self.op_type = "concat" + x0 = np.random.random((2, 3, 2, 5)).astype('float32') + x1 = np.random.random((2, 3, 3, 5)).astype('float32') + x2 = np.random.random((2, 3, 4, 5)).astype('float32') + axis = 2 + self.inputs = {'X': [('x0', x0), ('x1', x1), ('x2', x2)]} + self.attrs = {'axis': axis} + self.outputs = {'Out': np.concatenate((x0, x1, x2), axis=axis)} + + def test_check_output(self): + self.check_output() + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_cos_sim_op.py b/python/paddle/v2/framework/tests/test_cos_sim_op.py new file mode 100644 index 0000000000000000000000000000000000000000..797cbd8cc5cf7f73d58ca713d02667731d5c8a0e --- /dev/null +++ b/python/paddle/v2/framework/tests/test_cos_sim_op.py @@ -0,0 +1,39 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestCosSimOp(OpTest): + def setUp(self): + self.op_type = "cos_sim" + self.inputs = { + 'X': np.random.random((10, 5)).astype("float32"), + 'Y': np.random.random((10, 5)).astype("float32") + } + expect_x_norm = np.linalg.norm(self.inputs['X'], axis=1) + expect_y_norm = np.linalg.norm(self.inputs['Y'], axis=1) + expect_out = (self.inputs['X'] * self.inputs['Y']).sum(axis=1) / \ + expect_x_norm / expect_y_norm + self.outputs = { + 'XNorm': np.expand_dims(expect_x_norm, 1), + 'YNorm': np.expand_dims(expect_y_norm, 1), + 'Out': np.expand_dims(expect_out, 1) + } + + def test_check_output(self): + self.check_output() + + def test_check_grad_normal(self): + self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.05) + + def test_check_grad_ingore_x(self): + self.check_grad( + ['Y'], 'Out', max_relative_error=0.05, no_grad_set=set('X')) + + def test_check_grad_ignore_y(self): + self.check_grad( + ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y')) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_cross_entropy_op.py b/python/paddle/v2/framework/tests/test_cross_entropy_op.py index 4815192e255c6e0429db3f50918a76a773b30131..c2fc102a8b8de82da5c3fc5fee273790325908f8 100644 --- a/python/paddle/v2/framework/tests/test_cross_entropy_op.py +++ b/python/paddle/v2/framework/tests/test_cross_entropy_op.py @@ -1,37 +1,27 @@ import unittest import numpy -from op_test_util import OpTestMeta -from gradient_checker import GradientChecker, create_op +from op_test import OpTest -class TestCrossEntropy(unittest.TestCase): - __metaclass__ = OpTestMeta - +class TestCrossEntropy(OpTest): def setUp(self): - # TODO this unit test is not passed - self.type = "onehot_cross_entropy" - batch_size = 100 + self.op_type = "onehot_cross_entropy" + batch_size = 30 class_num = 10 - X = numpy.random.random((batch_size, class_num)).astype("float32") - label = 5 * numpy.ones(batch_size).astype("int32") + X = numpy.random.uniform(0.1, 1.0, + [batch_size, class_num]).astype("float32") + label = (class_num / 2) * numpy.ones(batch_size).astype("int32") self.inputs = {'X': X, 'label': label} Y = [] for i in range(0, batch_size): Y.append(-numpy.log(X[i][label[i]])) self.outputs = {'Y': numpy.array(Y).astype("float32")} + def test_check_output(self): + self.check_output() -class CrossEntropyGradOpTest(GradientChecker): - def test_softmax_grad(self): - op = create_op("onehot_cross_entropy") - batch_size = 100 - class_num = 10 - inputs = { - "X": numpy.random.uniform( - 0.1, 1.0, [batch_size, class_num]).astype("float32"), - "label": (class_num / 2) * numpy.ones(batch_size).astype("int32") - } - self.check_grad(op, inputs, set("X"), "Y") + def test_check_grad(self): + self.check_grad(['X'], 'Y') if __name__ == "__main__": diff --git a/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py b/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py index e5c862605fb11a5ea1426cf8f9054589dc377ff1..2473daaba24438819f3f55ccc40fe1c64ee59960 100644 --- a/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py +++ b/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py @@ -1,16 +1,17 @@ import unittest -from op_test_util import OpTestMeta -import numpy +import numpy as np +from op_test import OpTest -class TestFillZerosLikeOp(unittest.TestCase): - __metaclass__ = OpTestMeta - +class TestFillZerosLikeOp(OpTest): def setUp(self): - self.type = "fill_zeros_like" - self.inputs = {'Src': numpy.random.random((219, 232)).astype("float32")} - self.outputs = {'Dst': numpy.zeros_like(self.inputs['Src'])} + self.op_type = "fill_zeros_like" + self.inputs = {'Src': np.random.random((219, 232)).astype("float32")} + self.outputs = {'Dst': np.zeros_like(self.inputs["Src"])} + + def test_check_output(self): + self.check_output() -if __name__ == '__main__': +if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/framework/tests/test_gather_op.py b/python/paddle/v2/framework/tests/test_gather_op.py new file mode 100644 index 0000000000000000000000000000000000000000..b0ab429ef1b53640dfb696f6ea2f7b745564b874 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_gather_op.py @@ -0,0 +1,21 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestGatherOp(OpTest): + def setUp(self): + self.op_type = "gather" + xnp = np.random.random((10, 20)).astype("float32") + self.inputs = {'X': xnp, 'Index': np.array([1, 3, 5]).astype("int32")} + self.outputs = {'Out': self.inputs["X"][self.inputs["Index"]]} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out') + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_gaussian_random_op.py b/python/paddle/v2/framework/tests/test_gaussian_random_op.py index f95ed70b58d611b3233a21d3f2a34c864ae4d1b3..1f9e4db783c9907a22db72c8a6ff06c7ca0735da 100644 --- a/python/paddle/v2/framework/tests/test_gaussian_random_op.py +++ b/python/paddle/v2/framework/tests/test_gaussian_random_op.py @@ -14,11 +14,11 @@ class GaussianRandomTest(unittest.TestCase): def gaussian_random_test(self, place): scope = core.Scope() - scope.new_var("Out").get_tensor() + scope.new_var('Out').get_tensor() op = Operator( "gaussian_random", - Out="Out", + Out='Out', dims=[1000, 784], mean=.0, std=1., @@ -27,10 +27,10 @@ class GaussianRandomTest(unittest.TestCase): op.infer_shape(scope) context = core.DeviceContext.create(place) op.run(scope, context) - tensor = numpy.array(scope.find_var("Out").get_tensor()) + tensor = numpy.array(scope.find_var('Out').get_tensor()) self.assertAlmostEqual(numpy.mean(tensor), .0, delta=0.1) self.assertAlmostEqual(numpy.std(tensor), 1., delta=0.1) -if __name__ == '__main__': +if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/framework/tests/test_gradient_checker.py b/python/paddle/v2/framework/tests/test_gradient_checker.py new file mode 100644 index 0000000000000000000000000000000000000000..abeb01cb34158a43b5dcce5e39efc0e21e9fe638 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_gradient_checker.py @@ -0,0 +1,45 @@ +import unittest +import numpy as np +import paddle.v2.framework.core as core +from op_test import get_numeric_gradient +from op_test import create_op + + +class GetNumericGradientTest(unittest.TestCase): + def test_add_op(self): + x = np.random.random((10, 1)).astype("float32") + y = np.random.random((10, 1)).astype("float32") + z = x + y + scope = core.Scope() + add_op = create_op(scope, "add", {'X': x, 'Y': y}, {'Out': z}, dict()) + arr = get_numeric_gradient(scope, add_op, {'X': x, 'Y': y}, 'X', 'Out') + self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-4) + + def test_softmax_op(self): + def stable_softmax(x): + """Compute the softmax of vector x in a numerically stable way.""" + shiftx = x - np.max(x) + exps = np.exp(shiftx) + return exps / np.sum(exps) + + def label_softmax_grad(Y, dY): + dX = Y * 0.0 + for i in range(Y.shape[0]): + d = np.dot(Y[i, :], dY[i, :]) + dX[i, :] = Y[i, :] * (dY[i, :] - d) + return dX + + X = np.random.random((2, 2)).astype("float32") + Y = np.apply_along_axis(stable_softmax, 1, X) + dY = np.ones(Y.shape) + dX = label_softmax_grad(Y, dY) + + scope = core.Scope() + softmax_op = create_op(scope, "softmax", {"X": X}, {"Y": Y}, dict()) + + arr = get_numeric_gradient(scope, softmax_op, {"X": X}, "X", "Y") + np.testing.assert_almost_equal(arr, dX, decimal=1e-2) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_lookup_table.py b/python/paddle/v2/framework/tests/test_lookup_table.py new file mode 100644 index 0000000000000000000000000000000000000000..b259bb67e832adcb31b0ab4e992738be2b85f884 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_lookup_table.py @@ -0,0 +1,22 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestLookupTableOp(OpTest): + def setUp(self): + self.op_type = "lookup_table" + table = np.random.random((17, 31)).astype("float32") + ids = np.random.randint(0, 17, 4).astype("int32") + self.inputs = {'W': table, 'Ids': ids} + self.outputs = {'Out': table[ids]} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['W'], 'Out', no_grad_set=set('Ids')) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_mean_op.py b/python/paddle/v2/framework/tests/test_mean_op.py index b5d52b90567bcd0c9f376147145d8638049f7bab..7823abd8f813aad6462c98a9ace9a13dc286a157 100644 --- a/python/paddle/v2/framework/tests/test_mean_op.py +++ b/python/paddle/v2/framework/tests/test_mean_op.py @@ -1,16 +1,20 @@ import unittest -from op_test_util import OpTestMeta import numpy as np +from op_test import OpTest -class TestMeanOp(unittest.TestCase): - __metaclass__ = OpTestMeta - +class TestMeanOp(OpTest): def setUp(self): - self.type = "mean" - self.inputs = {'X': np.random.random((32, 784)).astype("float32")} - self.outputs = {'Out': np.mean(self.inputs['X'])} + self.op_type = "mean" + self.inputs = {'X': np.random.random((10, 10)).astype("float32")} + self.outputs = {'Out': np.mean(self.inputs["X"])} + + def test_check_output(self): + self.check_output() + + def test_checkout_grad(self): + self.check_grad(['X'], 'Out') -if __name__ == '__main__': +if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/framework/tests/test_minus_op.py b/python/paddle/v2/framework/tests/test_minus_op.py new file mode 100644 index 0000000000000000000000000000000000000000..dea797a1fea34265d0a32e097f413f421abf2521 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_minus_op.py @@ -0,0 +1,23 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class MinusOpTest(OpTest): + def setUp(self): + self.op_type = "minus" + self.inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'Y': np.random.random((32, 84)).astype("float32") + } + self.outputs = {'Out': (self.inputs['X'] - self.inputs['Y'])} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X', 'Y'], 'Out') + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_mul_op.py b/python/paddle/v2/framework/tests/test_mul_op.py index ec0ac99156a546dd3fb7b27778032bece38ab5a9..b3d95a56b88e510734da54f36ff21ccd7e1baabb 100644 --- a/python/paddle/v2/framework/tests/test_mul_op.py +++ b/python/paddle/v2/framework/tests/test_mul_op.py @@ -1,19 +1,59 @@ import unittest -from op_test_util import OpTestMeta import numpy as np +from op_test import OpTest -class TestMulOp(unittest.TestCase): - __metaclass__ = OpTestMeta - +class TestMulOp(OpTest): def setUp(self): - self.type = "mul" + self.op_type = "mul" self.inputs = { 'X': np.random.random((32, 84)).astype("float32"), 'Y': np.random.random((84, 100)).astype("float32") } self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} + def test_check_output(self): + self.check_output() + + def test_check_grad_normal(self): + self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5) + + def test_check_grad_ingore_x(self): + self.check_grad( + ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X")) + + def test_check_grad_ingore_y(self): + self.check_grad( + ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) + + +class TestMulOp2(OpTest): + def setUp(self): + self.op_type = "mul" + self.inputs = { + 'X': np.random.random((15, 4, 12, 10)).astype("float32"), + 'Y': np.random.random((4, 30, 8, 2, 9)).astype("float32") + } + self.attrs = {'x_num_col_dims': 2, 'y_num_col_dims': 2} + self.outputs = { + 'Out': np.dot(self.inputs['X'].reshape(15 * 4, 12 * 10), + self.inputs['Y'].reshape(4 * 30, 8 * 2 * 9)) + } + + def test_check_output(self): + self.check_output() + + def test_check_grad_normal(self): + self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5) + + def test_check_grad_ingore_x(self): + self.check_grad( + ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set('X')) + + def test_check_grad_ignore_y(self): + self.check_grad( + ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) + -if __name__ == '__main__': +if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/framework/tests/test_net.py b/python/paddle/v2/framework/tests/test_net.py index cc7f09e7155f5b1afa47fc4133b71ae3676b7436..50cfb855f2b01d8fd32342855d46716da7e07856 100644 --- a/python/paddle/v2/framework/tests/test_net.py +++ b/python/paddle/v2/framework/tests/test_net.py @@ -6,8 +6,8 @@ import unittest def fc(X, W, Y): ret_v = core.Net.create() - ret_v.add_op(Operator("mul", X="X", Y="W", Out="pre_activation")) - ret_v.add_op(Operator("sigmoid", X="pre_activation", Y=Y)) + ret_v.append_op(Operator("mul", X="X", Y="W", Out="pre_activation")) + ret_v.append_op(Operator("sigmoid", X="pre_activation", Y=Y)) ret_v.complete_add_op(True) return ret_v @@ -15,25 +15,25 @@ def fc(X, W, Y): class TestNet(unittest.TestCase): def test_net_all(self): net = core.Net.create() - op1 = Operator("add_two", X="X", Y="Y", Out="Out") - net.add_op(op1) + op1 = Operator("add", X="X", Y="Y", Out="Out") + net.append_op(op1) net2 = core.Net.create() - net2.add_op(fc(X="X", W="w", Y="fc.out")) + net2.append_op(fc(X="X", W="w", Y="fc.out")) net2.complete_add_op(True) - net.add_op(net2) + net.append_op(net2) net.complete_add_op(True) expected = ''' -Op(plain_net), inputs:(W, X, Y), outputs:(Out, fc.out, pre_activation). - Op(add_two), inputs:(X, Y), outputs:(Out). - Op(plain_net), inputs:(W, X), outputs:(fc.out, pre_activation). - Op(plain_net), inputs:(W, X), outputs:(fc.out, pre_activation). - Op(mul), inputs:(X, W), outputs:(pre_activation). - Op(sigmoid), inputs:(pre_activation), outputs:(fc.out). +Op(plain_net), inputs:{all[W, X, Y]}, outputs:{all[Out, fc.out, pre_activation]}. + Op(add), inputs:{X[X], Y[Y]}, outputs:{Out[Out]}. + Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}. + Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}. + Op(mul), inputs:{X[X], Y[W]}, outputs:{Out[pre_activation]}. + Op(sigmoid), inputs:{X[pre_activation]}, outputs:{Y[fc.out]}. ''' self.assertEqual(expected, "\n" + str(net)) -if __name__ == '__main__': +if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/framework/tests/test_operator.py b/python/paddle/v2/framework/tests/test_operator.py index 4f164e1a69e3fd0409f9b575a8bd9b4e423b486b..040556322d79cbb594eb9af585a5b9920d7ab625 100644 --- a/python/paddle/v2/framework/tests/test_operator.py +++ b/python/paddle/v2/framework/tests/test_operator.py @@ -1,9 +1,7 @@ import unittest import paddle.v2.framework.op as op import paddle.v2.framework.core as core -import paddle.v2.framework.proto.op_proto_pb2 as op_proto_pb2 -import paddle.v2.framework.proto.op_desc_pb2 as op_desc_pb2 -import paddle.v2.framework.proto.attribute_pb2 as attribute_pb2 +import paddle.v2.framework.proto.framework_pb2 as framework_pb2 class TestGetAllProtos(unittest.TestCase): @@ -17,7 +15,7 @@ class TestGetAllProtos(unittest.TestCase): class TestOpDescCreationMethod(unittest.TestCase): def test_plain_input_output(self): - op_proto = op_proto_pb2.OpProto() + op_proto = framework_pb2.OpProto() op_proto.type = "test" ipt = op_proto.inputs.add() ipt.name = "X" @@ -37,25 +35,32 @@ class TestOpDescCreationMethod(unittest.TestCase): method = op.OpDescCreationMethod(op_proto) output = method(X="a", Y="b", Z="c") - - expected = op_desc_pb2.OpDesc() + expected = framework_pb2.OpDesc() expected.type = "test" - expected.inputs.extend(["a", "b"]) - expected.outputs.append("c") + ipt_0 = expected.inputs.add() + ipt_0.parameter = "X" + ipt_0.arguments.extend(["a"]) + ipt_1 = expected.inputs.add() + ipt_1.parameter = 'Y' + ipt_1.arguments.extend(['b']) + opt = expected.outputs.add() + opt.parameter = "Z" + opt.arguments.extend(["c"]) + self.assertEqual(expected, output) def test_multiple_input_plain_output(self): - op_proto = op_proto_pb2.OpProto() + op_proto = framework_pb2.OpProto() op_proto.type = "fc" ipt = op_proto.inputs.add() ipt.name = "X" ipt.comment = "" - ipt.multiple = True + ipt.duplicable = True ipt = op_proto.inputs.add() ipt.name = "W" ipt.comment = "" - ipt.multiple = True + ipt.duplicable = True ipt = op_proto.inputs.add() ipt.name = "b" @@ -70,30 +75,50 @@ class TestOpDescCreationMethod(unittest.TestCase): method = op.OpDescCreationMethod(op_proto) generated1 = method(X="x", W="w", b="b", Y="y") - expected1 = op_desc_pb2.OpDesc() - expected1.inputs.extend(['x', 'w', 'b']) - expected1.outputs.extend(['y']) + expected1 = framework_pb2.OpDesc() + tmp = expected1.inputs.add() + tmp.parameter = "X" + tmp.arguments.extend(['x']) + + tmp = expected1.inputs.add() + tmp.parameter = 'W' + tmp.arguments.extend(['w']) + + tmp = expected1.inputs.add() + tmp.parameter = 'b' + tmp.arguments.extend(['b']) + + tmp = expected1.outputs.add() + tmp.parameter = 'Y' + tmp.arguments.extend(['y']) expected1.type = 'fc' - attr = expected1.attrs.add() - attr.name = 'input_format' - attr.type = attribute_pb2.INTS - attr.ints.extend([0, 1, 2, 3]) self.assertEqual(expected1, generated1) generated2 = method( X=['x1', 'x2', 'x3'], b='b', W=['w1', 'w2', 'w3'], Y='y') - expected2 = op_desc_pb2.OpDesc() - expected2.inputs.extend(['x1', 'x2', 'x3', 'w1', 'w2', 'w3', 'b']) - expected2.outputs.extend(['y']) + expected2 = framework_pb2.OpDesc() + + tmp = expected2.inputs.add() + tmp.parameter = "X" + tmp.arguments.extend(['x1', 'x2', 'x3']) + + tmp = expected2.inputs.add() + tmp.parameter = 'W' + tmp.arguments.extend(['w1', 'w2', 'w3']) + + tmp = expected2.inputs.add() + tmp.parameter = 'b' + tmp.arguments.extend(['b']) + + tmp = expected2.outputs.add() + tmp.parameter = 'Y' + tmp.arguments.extend(['y']) + expected2.type = 'fc' - attr = expected2.attrs.add() - attr.name = 'input_format' - attr.type = attribute_pb2.INTS - attr.ints.extend([0, 3, 6, 7]) self.assertEqual(expected2, generated2) def test_attrs(self): - op_proto = op_proto_pb2.OpProto() + op_proto = framework_pb2.OpProto() op_proto.type = "test" ipt = op_proto.inputs.add() ipt.name = 'X' @@ -105,12 +130,12 @@ class TestOpDescCreationMethod(unittest.TestCase): attr.comment = "" attr.type = type - __add_attr__("int_attr", attribute_pb2.INT) - __add_attr__("float_attr", attribute_pb2.FLOAT) - __add_attr__("string_attr", attribute_pb2.STRING) - __add_attr__("ints_attr", attribute_pb2.INTS) - __add_attr__("floats_attr", attribute_pb2.FLOATS) - __add_attr__("strings_attr", attribute_pb2.STRINGS) + __add_attr__("int_attr", framework_pb2.INT) + __add_attr__("float_attr", framework_pb2.FLOAT) + __add_attr__("string_attr", framework_pb2.STRING) + __add_attr__("ints_attr", framework_pb2.INTS) + __add_attr__("floats_attr", framework_pb2.FLOATS) + __add_attr__("strings_attr", framework_pb2.STRINGS) op_proto.comment = "" self.assertTrue(op_proto.IsInitialized()) @@ -126,76 +151,52 @@ class TestOpDescCreationMethod(unittest.TestCase): floats_attr=[0.2, 3.2, 4.5], strings_attr=["a", "b", "c"]) - expected = op_desc_pb2.OpDesc() + expected = framework_pb2.OpDesc() expected.type = "test" - expected.inputs.extend(['a']) + + ipt = expected.inputs.add() + ipt.parameter = "X" + ipt.arguments.extend(['a']) + attr = expected.attrs.add() attr.name = "int_attr" - attr.type = attribute_pb2.INT + attr.type = framework_pb2.INT attr.i = 10 attr = expected.attrs.add() attr.name = "float_attr" - attr.type = attribute_pb2.FLOAT + attr.type = framework_pb2.FLOAT attr.f = 3.2 attr = expected.attrs.add() attr.name = "string_attr" - attr.type = attribute_pb2.STRING + attr.type = framework_pb2.STRING attr.s = "test_str" attr = expected.attrs.add() attr.name = "ints_attr" - attr.type = attribute_pb2.INTS + attr.type = framework_pb2.INTS attr.ints.extend([0, 1, 2, 3, 4]) attr = expected.attrs.add() attr.name = 'floats_attr' - attr.type = attribute_pb2.FLOATS + attr.type = framework_pb2.FLOATS attr.floats.extend([0.2, 3.2, 4.5]) attr = expected.attrs.add() attr.name = 'strings_attr' - attr.type = attribute_pb2.STRINGS + attr.type = framework_pb2.STRINGS attr.strings.extend(['a', 'b', 'c']) self.assertEqual(expected, generated) - def test_input_temporary_output(self): - op_proto = op_proto_pb2.OpProto() - op_proto.type = "test" - out = op_proto.outputs.add() - out.name = "OUT" - out.comment = "" - - out = op_proto.outputs.add() - out.name = "TMP" - out.comment = "" - out.temporary = True - - out = op_proto.outputs.add() - out.name = "OUT2" - out.comment = "" - op_proto.comment = "" - - method = op.OpDescCreationMethod(op_proto) - generated = method(OUT="a", OUT2="b") - desc = op_desc_pb2.OpDesc() - desc.outputs.extend(["a", core.var_names.temp(), "b"]) - desc.type = "test" - attr = desc.attrs.add() - attr.name = "temporary_index" - attr.type = attribute_pb2.INTS - attr.ints.append(2) - self.assertEqual(generated, desc) - class TestOpCreations(unittest.TestCase): def test_all(self): - add_op = op.Operator("add_two", X="a", Y="b", Out="z") + add_op = op.Operator("add", X="a", Y="b", Out="z") self.assertIsNotNone(add_op) # Invoke C++ DebugString() - self.assertEqual('Op(add_two), inputs:(a, b), outputs:(z).', + self.assertEqual('Op(add), inputs:{X[a], Y[b]}, outputs:{Out[z]}.', str(add_op)) diff --git a/python/paddle/v2/framework/tests/test_protobuf.py b/python/paddle/v2/framework/tests/test_protobuf.py index 69e98e2f250a9df23b25e7e2043af29f87c996a0..848a396b3b6eec57d500b464780b64f339b09e94 100644 --- a/python/paddle/v2/framework/tests/test_protobuf.py +++ b/python/paddle/v2/framework/tests/test_protobuf.py @@ -1,11 +1,10 @@ -import paddle.v2.framework.proto.op_proto_pb2 as op_proto_lib -import paddle.v2.framework.proto.attribute_pb2 as attr_type_lib +import paddle.v2.framework.proto.framework_pb2 as framework_pb2 import unittest class TestFrameworkProto(unittest.TestCase): def test_all(self): - op_proto = op_proto_lib.OpProto() + op_proto = framework_pb2.OpProto() ipt0 = op_proto.inputs.add() ipt0.name = "a" ipt0.comment = "the input of cosine op" @@ -19,7 +18,7 @@ class TestFrameworkProto(unittest.TestCase): attr = op_proto.attrs.add() attr.name = "scale" attr.comment = "scale of cosine op" - attr.type = attr_type_lib.FLOAT + attr.type = framework_pb2.FLOAT op_proto.type = "cos" self.assertTrue(op_proto.IsInitialized()) diff --git a/python/paddle/v2/framework/tests/test_recurrent_op.py b/python/paddle/v2/framework/tests/test_recurrent_op.py index 0db66cc4e181fde10f161a323ea749fd84a5f963..22e680fd783ec681e95326fb84db34570265cffc 100644 --- a/python/paddle/v2/framework/tests/test_recurrent_op.py +++ b/python/paddle/v2/framework/tests/test_recurrent_op.py @@ -2,7 +2,7 @@ import logging import paddle.v2.framework.core as core import unittest import numpy as np -from paddle.v2.framework.op import Operator +from paddle.v2.framework.op import Operator, RecurrentOp def py_sigmoid(x): @@ -98,11 +98,11 @@ class TestRecurrentOp(unittest.TestCase): def forward(self): self.scope = core.Scope() self.create_global_variables() + self.create_rnn_op() self.create_step_net() - rnn_op = self.create_rnn_op() ctx = core.DeviceContext.create(core.CPUPlace()) - rnn_op.infer_shape(self.scope) - rnn_op.run(self.scope, ctx) + self.rnnop.infer_shape(self.scope) + self.rnnop.run(self.scope, ctx) return np.array(self.scope.find_var("h").get_tensor()) def create_global_variables(self): @@ -128,8 +128,7 @@ class TestRecurrentOp(unittest.TestCase): def create_rnn_op(self): # create RNNOp - rnnop = Operator( - "recurrent_op", + self.rnnop = RecurrentOp( # inputs inlinks=["x"], boot_memories=["h_boot"], @@ -142,22 +141,18 @@ class TestRecurrentOp(unittest.TestCase): outlink_alias=["h@alias"], pre_memories=["h@pre"], memories=["h@alias"]) - return rnnop def create_step_net(self): - var = self.scope.new_var("stepnet") - stepnet = var.get_net() - - # x_fc_op = Operator("fc", X="x@alias", W="W", Y="Wx") - # h_fc_op = Operator("fc", X="h@pre", W="U", Y="Uh") + stepnet = core.Net.create() x_fc_op = Operator("mul", X="x@alias", Y="W", Out="Wx") h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh") - sum_op = Operator("add_two", X="Wx", Y="Uh", Out="sum") + sum_op = Operator("add", X="Wx", Y="Uh", Out="sum") sig_op = Operator("sigmoid", X="sum", Y="h@alias") for op in [x_fc_op, h_fc_op, sum_op, sig_op]: - stepnet.add_op(op) + stepnet.append_op(op) stepnet.complete_add_op(True) + self.rnnop.set_stepnet(stepnet) def test_forward(self): print 'test recurrent op forward' diff --git a/python/paddle/v2/framework/tests/test_rowwise_add_op.py b/python/paddle/v2/framework/tests/test_rowwise_add_op.py index f8521eb517057fbeb104b28af7da4fffe54f37de..336645bd993ff743cbe20bb5cae5cd278db57ce7 100644 --- a/python/paddle/v2/framework/tests/test_rowwise_add_op.py +++ b/python/paddle/v2/framework/tests/test_rowwise_add_op.py @@ -1,19 +1,51 @@ import unittest -from op_test_util import OpTestMeta import numpy as np +from op_test import OpTest -class TestRowwiseAddOp(unittest.TestCase): - __metaclass__ = OpTestMeta +class TestRowwiseAddOp(OpTest): + def setUp(self): + self.op_type = "rowwise_add" + self.inputs = { + 'X': np.random.uniform(0.1, 1, [5, 10]).astype("float32"), + 'b': np.random.uniform(0.1, 1, [10]).astype("float32") + } + self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['b'])} + + def test_check_output(self): + self.check_output() + + def test_check_grad_normal(self): + self.check_grad(['X', 'b'], 'Out') + + def test_check_grad_ingore_b(self): + self.check_grad(['X'], 'Out', no_grad_set=set('b')) + def test_check_grad_ingore_x(self): + self.check_grad(['b'], 'Out', no_grad_set=set('X')) + + +class TestRowwiseAddOp2(OpTest): def setUp(self): - self.type = "rowwise_add" + self.op_type = "rowwise_add" self.inputs = { - 'X': np.random.random((32, 84)).astype("float32"), - 'b': np.random.random(84).astype("float32") + 'X': np.random.uniform(0.1, 1, [2, 3, 2, 5]).astype("float32"), + 'b': np.random.uniform(0.1, 1, [2, 5]).astype("float32") } self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['b'])} + def test_check_output(self): + self.check_output() + + def test_check_grad_normal(self): + self.check_grad(['X', 'b'], 'Out') + + def test_check_grad_ignore_b(self): + self.check_grad(['X'], 'Out', no_grad_set=set('b')) + + def test_check_grad_ignore_x(self): + self.check_grad(['b'], 'Out', no_grad_set=set('X')) + -if __name__ == '__main__': +if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/framework/tests/test_scale_and_identity_op.py b/python/paddle/v2/framework/tests/test_scale_and_identity_op.py new file mode 100644 index 0000000000000000000000000000000000000000..05d76d428299c8176d1a6adf6da15a203fa7502a --- /dev/null +++ b/python/paddle/v2/framework/tests/test_scale_and_identity_op.py @@ -0,0 +1,34 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class IdentityTest(OpTest): + def setUp(self): + self.op_type = "identity" + self.inputs = {'X': np.random.random((10, 10)).astype("float32")} + self.outputs = {'Out': self.inputs['X']} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out') + + +class ScaleTest(OpTest): + def setUp(self): + self.op_type = "scale" + self.inputs = {'X': np.random.random((10, 10)).astype("float32")} + self.attrs = {'scale': -2.3} + self.outputs = {'Out': self.inputs['X'] * self.attrs['scale']} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out') + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_scatter_op.py b/python/paddle/v2/framework/tests/test_scatter_op.py new file mode 100644 index 0000000000000000000000000000000000000000..33c73c52631a09ea0fefdeb9467991ae9c04321c --- /dev/null +++ b/python/paddle/v2/framework/tests/test_scatter_op.py @@ -0,0 +1,25 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestScatterOp(OpTest): + def setUp(self): + self.op_type = "scatter" + ref_np = np.ones((3, 3)).astype("float32") + index_np = np.array([1, 2]).astype("int32") + updates_np = np.random.random((2, 3)).astype("float32") + output_np = np.copy(ref_np) + output_np[index_np] += updates_np + self.inputs = {'Ref': ref_np, 'Index': index_np, 'Updates': updates_np} + self.outputs = {'Out': output_np} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['Updates', 'Ref'], 'Out', in_place=True) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_sgd_op.py b/python/paddle/v2/framework/tests/test_sgd_op.py index e5f9ef865e84f1a78e28884ad7e2e758f9ca8054..557cf15ace63e336462c7dcdbbc10f30aeedc6f4 100644 --- a/python/paddle/v2/framework/tests/test_sgd_op.py +++ b/python/paddle/v2/framework/tests/test_sgd_op.py @@ -1,21 +1,22 @@ import unittest -import numpy -from op_test_util import OpTestMeta +import numpy as np +from op_test import OpTest -class TestSGD(unittest.TestCase): - __metaclass__ = OpTestMeta - +class TestSGD(OpTest): def setUp(self): - self.type = "sgd" - w = numpy.random.random((102, 105)).astype("float32") - g = numpy.random.random((102, 105)).astype("float32") + self.op_type = "sgd" + w = np.random.random((102, 105)).astype("float32") + g = np.random.random((102, 105)).astype("float32") lr = 0.1 self.inputs = {'param': w, 'grad': g} self.attrs = {'learning_rate': lr} self.outputs = {'param_out': w - lr * g} + def test_check_output(self): + self.check_output() + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/framework/tests/test_sigmoid_op.py b/python/paddle/v2/framework/tests/test_sigmoid_op.py index 2a57a41ed8b718fd420062ba68e853a4861b7359..2316e49eff7bb1cdb53acb3889a6ef05060b59f3 100644 --- a/python/paddle/v2/framework/tests/test_sigmoid_op.py +++ b/python/paddle/v2/framework/tests/test_sigmoid_op.py @@ -1,19 +1,22 @@ import unittest -from op_test_util import OpTestMeta import numpy as np +from op_test import OpTest -class TestSigmoidOp(unittest.TestCase): - __metaclass__ = OpTestMeta - +class TestSigmoid(OpTest): def setUp(self): - self.type = "sigmoid" - self.inputs = {'X': np.random.random((32, 100)).astype("float32")} + self.op_type = "sigmoid" + self.inputs = { + 'X': np.random.uniform(0.1, 1, [11, 17]).astype("float32") + } self.outputs = {'Y': 1 / (1 + np.exp(-self.inputs['X']))} + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Y", max_relative_error=0.007) -#class TestSigmoidGradOp(unittest.TestCase): -#TODO(qingqing) add unit test if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/framework/tests/test_softmax_op.py b/python/paddle/v2/framework/tests/test_softmax_op.py index e670d93653e07d35e5019c9daac45c214eddf367..1b948f252fa631e9886840b377de2996e110dc91 100644 --- a/python/paddle/v2/framework/tests/test_softmax_op.py +++ b/python/paddle/v2/framework/tests/test_softmax_op.py @@ -1,9 +1,6 @@ import unittest - import numpy as np - -from gradient_checker import GradientChecker, create_op -from op_test_util import OpTestMeta +from op_test import OpTest def stable_softmax(x): @@ -13,23 +10,22 @@ def stable_softmax(x): return exps / np.sum(exps) -class TestSoftmaxOp(unittest.TestCase): - __metaclass__ = OpTestMeta - +class TestSoftmaxOp(OpTest): def setUp(self): - self.type = "softmax" - self.inputs = {'X': np.random.random((32, 100)).astype("float32")} + self.op_type = "softmax" + self.inputs = { + 'X': np.random.uniform(0.1, 1, [10, 10]).astype("float32") + } self.outputs = { 'Y': np.apply_along_axis(stable_softmax, 1, self.inputs['X']) } + def test_check_output(self): + self.check_output() -class SoftmaxGradOpTest(GradientChecker): - def test_softmax(self): - op = create_op("softmax") - inputs = {"X": np.random.uniform(0.1, 1, [10, 10]).astype("float32")} - self.check_grad(op, inputs, set("X"), "Y") + def test_check_grad(self): + self.check_grad(['X'], 'Y') -if __name__ == '__main__': +if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/framework/tests/test_squared_l2_distance_op.py b/python/paddle/v2/framework/tests/test_squared_l2_distance_op.py new file mode 100644 index 0000000000000000000000000000000000000000..dc6ebf5d30369231b4918a168bbdf25c7096c808 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_squared_l2_distance_op.py @@ -0,0 +1,71 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestSquaredL2DistanceOp_f0(OpTest): + def setUp(self): + self.op_type = "squared_l2_distance" + self.inputs = { + 'X': np.random.uniform(0.1, 0.6, (2, 3)).astype("float32"), + 'Y': np.random.uniform(0.1, 0.6, (2, 3)).astype("float32") + } + sub_res = self.inputs['X'] - self.inputs['Y'] + output = sub_res * sub_res + self.outputs = { + 'sub_result': sub_res, + 'Out': np.expand_dims(output.sum(1), 1) + } + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X', 'Y'], 'Out') + + +class TestSquaredL2DistanceOp_f1(OpTest): + def setUp(self): + self.op_type = "squared_l2_distance" + self.inputs = { + 'X': np.random.uniform(0.1, 0.6, (2, 3)).astype("float32"), + 'Y': np.random.uniform(0.1, 0.6, (1, 3)).astype("float32") + } + sub_res = self.inputs['X'] - self.inputs['Y'] + output = sub_res * sub_res + self.outputs = { + 'sub_result': sub_res, + 'Out': np.expand_dims(output.sum(1), 1) + } + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X', 'Y'], 'Out') + + +class TestSquaredL2DistanceOp_f2(OpTest): + def setUp(self): + self.op_type = "squared_l2_distance" + self.inputs = { + 'X': np.random.uniform(0.1, 0.6, (2, 3, 4)).astype("float32"), + 'Y': np.random.uniform(0.1, 0.6, (1, 3, 4)).astype("float32") + } + sub_res = self.inputs['X'] - self.inputs['Y'] + sub_res = sub_res.reshape((2, 3 * 4)) + output = sub_res * sub_res + self.outputs = { + 'sub_result': sub_res, + 'Out': np.expand_dims(output.sum(1), 1) + } + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X', 'Y'], 'Out') + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_sum_op.py b/python/paddle/v2/framework/tests/test_sum_op.py new file mode 100644 index 0000000000000000000000000000000000000000..60254291e2ab9215e2bc37c12d5e2e1ca6d33d5d --- /dev/null +++ b/python/paddle/v2/framework/tests/test_sum_op.py @@ -0,0 +1,24 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestSumOp(OpTest): + def setUp(self): + self.op_type = "sum" + x0 = np.random.random((3, 4)).astype('float32') + x1 = np.random.random((3, 4)).astype('float32') + x2 = np.random.random((3, 4)).astype('float32') + self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]} + y = x0 + x1 + x2 + self.outputs = {'Out': y} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['x0'], 'Out') + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_top_k_op.py b/python/paddle/v2/framework/tests/test_top_k_op.py new file mode 100644 index 0000000000000000000000000000000000000000..cab799256d791889c295aa7f9048080f5caaf2dc --- /dev/null +++ b/python/paddle/v2/framework/tests/test_top_k_op.py @@ -0,0 +1,47 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestTopkOp(OpTest): + def setUp(self): + self.op_type = "top_k" + k = 1 + input = np.random.random((32, 84)).astype("float32") + output = np.ndarray((32, k)) + indices = np.ndarray((32, k)) + + self.inputs = {'X': input} + self.attrs = {'k': k} + + for rowid in xrange(32): + row = input[rowid] + output[rowid] = np.sort(row)[-k:] + indices[rowid] = row.argsort()[-k:] + + self.outputs = {'Out': output, 'Indices': indices} + + +class TestTopkOp3d(OpTest): + def setUp(self): + self.op_type = "top_k" + k = 1 + input = np.random.random((32, 2, 84)).astype("float32") + input_flat_2d = input.reshape(64, 84) + output = np.ndarray((64, k)) + indices = np.ndarray((64, k)).astype("int") + + # FIXME: should use 'X': input for a 3d input + self.inputs = {'X': input_flat_2d} + self.attrs = {'k': k} + + for rowid in xrange(64): + row = input_flat_2d[rowid] + output[rowid] = np.sort(row)[-k:] + indices[rowid] = row.argsort()[-k:] + + self.outputs = {'Out': output, 'Indices': indices} + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_uniform_random_op.py b/python/paddle/v2/framework/tests/test_uniform_random_op.py index c3d2bb44da3977c0899b2609a8efe15b7e1789f2..76a5e36e56ab08230bdc2597d209fcf5d1d2acb0 100644 --- a/python/paddle/v2/framework/tests/test_uniform_random_op.py +++ b/python/paddle/v2/framework/tests/test_uniform_random_op.py @@ -14,11 +14,11 @@ class UniformRandomTest(unittest.TestCase): def uniform_random_test(self, place): scope = core.Scope() - scope.new_var("X").get_tensor() + scope.new_var('X').get_tensor() op = Operator( "uniform_random", - Out="X", + Out='X', dims=[1000, 784], min=-5.0, max=10.0, @@ -27,9 +27,9 @@ class UniformRandomTest(unittest.TestCase): op.infer_shape(scope) ctx = core.DeviceContext.create(place) op.run(scope, ctx) - tensor = numpy.array(scope.find_var("X").get_tensor()) + tensor = numpy.array(scope.find_var('X').get_tensor()) self.assertAlmostEqual(tensor.mean(), 2.5, delta=0.1) -if __name__ == '__main__': +if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/inference.py b/python/paddle/v2/inference.py index 4dcc3ab57e7e6dfbe040ac61025e55b9e48b4415..8acea6155c588f2e8e5ad009cd8f0a0c09afb92b 100644 --- a/python/paddle/v2/inference.py +++ b/python/paddle/v2/inference.py @@ -70,7 +70,7 @@ class Inference(object): item = [each_result[each_field] for each_field in field] yield item - def infer(self, input, field='value', **kwargs): + def infer(self, input, field='value', flatten_result=True, **kwargs): """ Infer a data by model. :param input: input data batch. Should be python iterable object. @@ -83,7 +83,10 @@ class Inference(object): retv = [[] for i in xrange(len(result))] for i, item in enumerate(result): retv[i].append(item) - retv = [numpy.concatenate(out) for out in retv] + + if flatten_result: + retv = [numpy.concatenate(out) for out in retv] + if len(retv) == 1: return retv[0] else: diff --git a/python/paddle/v2/parameters.py b/python/paddle/v2/parameters.py index b8af5abaeada49a3e8951c21c9065aaf4d1ab851..4cfd91882e2d5f0098d27b8897359152ddd94dda 100644 --- a/python/paddle/v2/parameters.py +++ b/python/paddle/v2/parameters.py @@ -14,6 +14,7 @@ import numpy as np from paddle.proto.ParameterConfig_pb2 import ParameterConfig +from collections import OrderedDict import paddle.trainer.config_parser as cp import struct import tarfile @@ -42,9 +43,25 @@ def create(layers): class Parameters(object): """ - Parameters is a dictionary contains Paddle's parameter. The key of - Parameters is the name of parameter. The value of Parameters is a plain - :code:`numpy.ndarry` . + `Parameters` manages all the learnable parameters in a neural network. + It stores parameters' information in an OrderedDict. The key is + the name of a parameter, and value is a parameter's configuration(in + protobuf format), such as initialization mean and std, its size, whether it + is a static parameter, and so on. + + :param __param_conf__: store the configurations of learnable parameters in + the network in an OrderedDict. Parameter is added one by one into the + dict by following their created order in the network: parameters of + the previous layers in a network are careted first. You can visit the + parameters from bottom to top by iterating over this dict. + :type __param_conf__: OrderedDict + :param __gradient_machines__: all of the parameters in a neural network are + appended to a PaddlePaddle gradient machine, which is used internally to + copy parameter values between C++ and Python end. + :type __gradient_machines__: list + :param __tmp_params__: a dict to store dummy parameters if no + __gradient_machines__ is appended to `Parameters`. + :type __tmp_params__: dict Basically usage is @@ -62,7 +79,7 @@ class Parameters(object): """ def __init__(self): - self.__param_conf__ = dict() + self.__param_conf__ = OrderedDict() self.__gradient_machines__ = [] self.__tmp_params__ = dict() @@ -231,6 +248,9 @@ class Parameters(object): :rtype: np.ndarray """ import py_paddle.swig_paddle as api + if self.__param_conf__[key].is_static: + return np.zeros(self.__param_conf__[key].size, dtype=np.float32) + return self.__getter_inner(key, api.PARAMETER_GRADIENT) def set(self, parameter_name, value): @@ -250,7 +270,7 @@ class Parameters(object): append gradient machine to parameters. This method is used internally in Trainer.train. - :param gradient_machine: Paddle C++ GradientMachine object. + :param gradient_machine: PaddlePaddle C++ GradientMachine object. :type gradient_machine: api.GradientMachine :return: """ diff --git a/python/paddle/v2/reader/creator.py b/python/paddle/v2/reader/creator.py index d0f18e4b6611fa56654e7f2a0144758339cb9e19..97e844b92c77a7c58105dc5df2b4092fa5571d6f 100644 --- a/python/paddle/v2/reader/creator.py +++ b/python/paddle/v2/reader/creator.py @@ -57,7 +57,7 @@ def text_file(path): return reader -def recordio_local(paths, buf_size=100): +def recordio(paths, buf_size=100): """ Creates a data reader from given RecordIO file paths separated by ",", glob pattern is supported. @@ -67,15 +67,19 @@ def recordio_local(paths, buf_size=100): import recordio as rec import paddle.v2.reader.decorator as dec + import cPickle as pickle def reader(): - a = ','.join(paths) - f = rec.reader(a) + if isinstance(paths, basestring): + path = paths + else: + path = ",".join(paths) + f = rec.reader(path) while True: r = f.read() if r is None: break - yield r + yield pickle.loads(r) f.close() return dec.buffered(reader, buf_size) diff --git a/python/paddle/v2/reader/tests/creator_test.py b/python/paddle/v2/reader/tests/creator_test.py index 359f3eeefbe8efeb343cc875c707c9251a7087fb..cf190aa6645f9a5bed891a3a47c03efa03813d65 100644 --- a/python/paddle/v2/reader/tests/creator_test.py +++ b/python/paddle/v2/reader/tests/creator_test.py @@ -34,5 +34,27 @@ class TestTextFile(unittest.TestCase): self.assertEqual(e, str(idx * 2) + " " + str(idx * 2 + 1)) +class TestRecordIO(unittest.TestCase): + def do_test(self, path): + reader = paddle.v2.reader.creator.recordio(path) + idx = 0 + for e in reader(): + if idx == 0: + self.assertEqual(e, (1, 2, 3)) + elif idx == 1: + self.assertEqual(e, (4, 5, 6)) + idx += 1 + self.assertEqual(idx, 2) + + def test_recordIO(self): + self.do_test( + os.path.join( + os.path.dirname(__file__), "test_reader_recordio.dat")) + self.do_test([ + os.path.join( + os.path.dirname(__file__), "test_reader_recordio.dat") + ]) + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/reader/tests/test_reader_recordio.dat b/python/paddle/v2/reader/tests/test_reader_recordio.dat new file mode 100644 index 0000000000000000000000000000000000000000..a99a35bb829e066c4845d0b85b96cd1eb3a12491 Binary files /dev/null and b/python/paddle/v2/reader/tests/test_reader_recordio.dat differ diff --git a/python/paddle/v2/tests/test_layer.py b/python/paddle/v2/tests/test_layer.py index f2097e195f41637977e71f65f36dad005d3e7941..de932ad715bea8db158393c3c192ef67502e2fa3 100644 --- a/python/paddle/v2/tests/test_layer.py +++ b/python/paddle/v2/tests/test_layer.py @@ -134,19 +134,21 @@ class CostLayerTest(unittest.TestCase): cost3 = layer.cross_entropy_cost(input=inference, label=label) cost4 = layer.cross_entropy_with_selfnorm_cost( input=inference, label=label) - cost5 = layer.mse_cost(input=inference, label=label) - cost6 = layer.mse_cost(input=inference, label=label, weight=weight) + cost5 = layer.square_error_cost(input=inference, label=label) + cost6 = layer.square_error_cost( + input=inference, label=label, weight=weight) cost7 = layer.multi_binary_label_cross_entropy_cost( input=inference, label=label) cost8 = layer.rank_cost(left=score, right=score, label=score) cost9 = layer.lambda_cost(input=inference, score=score) cost10 = layer.sum_cost(input=inference) - cost11 = layer.huber_cost(input=score, label=label) + cost11 = layer.huber_regression_cost(input=score, label=label) + cost12 = layer.huber_classification_cost(input=score, label=label) print layer.parse_network([cost1, cost2]) print layer.parse_network([cost3, cost4]) print layer.parse_network([cost5, cost6]) - print layer.parse_network([cost7, cost8, cost9, cost10, cost11]) + print layer.parse_network([cost7, cost8, cost9, cost10, cost11, cost12]) crf = layer.crf(input=inference, label=label) crf_decoding = layer.crf_decoding(input=inference, size=3) diff --git a/python/paddle/v2/trainer.py b/python/paddle/v2/trainer.py index 9c4dd5f25083d210bbd218a85d8dbb3cce2c3d0e..ca95ef13bd440ac0ba3d46f6e4680d4d7aa94c42 100644 --- a/python/paddle/v2/trainer.py +++ b/python/paddle/v2/trainer.py @@ -27,16 +27,24 @@ class SGD(object): SGD Trainer combines data reader, network topolopy and update_equation together to train/test a neural network. - :param update_equation: The optimizer object. - :type update_equation: paddle.v2.optimizer.Optimizer :param cost: Target cost that neural network should be optimized. :type cost: paddle.v2.config_base.Layer :param parameters: The parameters dictionary. :type parameters: paddle.v2.parameters.Parameters + :param update_equation: The optimizer object. + :type update_equation: paddle.v2.optimizer.Optimizer :param extra_layers: Some layers in the neural network graph are not in the path of cost layer. - :param pserver_spec: pserver location, eg: localhost:3000 :type extra_layers: paddle.v2.config_base.Layer + :param is_local: Whether trainning locally + :type is_local: bool + :param pserver_spec: comma string for pserver location, + eg:127.10.0.10:3000,127.10.0.11:3000, + and this parameter is only used for fault + tolerant mode cluster training. + :type pserver_spec: string + :param use_etcd: Whether using etcd pserver. + :param use_etcd: bool """ def __init__(self, @@ -166,13 +174,18 @@ class SGD(object): pass_id=pass_id, batch_id=batch_id, cost=cost, - evaluator=batch_evaluator)) + evaluator=batch_evaluator, + gm=self.__gradient_machine__)) self.__parameter_updater__.finishBatch(cost) batch_evaluator.finish() self.__parameter_updater__.finishPass() pass_evaluator.finish() - event_handler(v2_event.EndPass(pass_id, evaluator=pass_evaluator)) + event_handler( + v2_event.EndPass( + pass_id, + evaluator=pass_evaluator, + gm=self.__gradient_machine__)) self.__gradient_machine__.finish() def test(self, reader, feeding=None): diff --git a/python/requirements.txt b/python/requirements.txt index 3df822bd76d2a64a0a35f84b4ec309ce7150c221..e19453c25da1ec78773c00a72b8e517b0d798fff 100644 --- a/python/requirements.txt +++ b/python/requirements.txt @@ -1,7 +1,7 @@ requests==2.9.2 numpy>=1.12 protobuf==3.1 -recordio +recordio>=0.1.0 matplotlib rarfile scipy>=0.19.0 diff --git a/python/setup.py.in b/python/setup.py.in index 38728aa2fd77cf3c882479ed83e99688b9ffa541..87b3823e52604b889cdee76bc696a1ae9b9de802 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -23,6 +23,19 @@ with open('@PADDLE_SOURCE_DIR@/python/requirements.txt') as f: if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']: setup_requires+=["opencv-python"] +# the prefix is sys.prefix which should always be usr +paddle_bin_dir = 'opt/paddle/bin' +paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/scripts/paddle_usage', + '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer', + '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model', + '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main', + '${PADDLE_BINARY_DIR}/paddle/scripts/paddle'] + +paddle_rt_lib_dir = 'lib' +paddle_rt_libs = ['${WARPCTC_LIBRARIES}'] +if '${MKL_SHARED_LIBS}'!= '': + paddle_rt_libs += '${MKL_SHARED_LIBS}'.split(';') + setup(name='paddlepaddle', version='${PADDLE_VERSION}', description='Parallel Distributed Deep Learning', @@ -40,11 +53,7 @@ setup(name='paddlepaddle', 'paddle.v2.framework.proto': '${PADDLE_BINARY_DIR}/paddle/framework', 'py_paddle': '${PADDLE_SOURCE_DIR}/paddle/py_paddle' }, - scripts=['${PADDLE_BINARY_DIR}/paddle/scripts/paddle'], + scripts=paddle_bins, distclass=BinaryDistribution, - data_files=[('/usr/local/opt/paddle/bin', - ['${PADDLE_BINARY_DIR}/paddle/scripts/paddle_usage', - '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer', - '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model', - '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main'])] + data_files=[(paddle_rt_lib_dir, paddle_rt_libs)] )