diff --git a/.gitignore b/.gitignore index c84b2fc8c79d6e2c9c83e2b830ab176295846fd0..9622ab78e0e0556ec2b4cc974fee93ff680d54d2 100644 --- a/.gitignore +++ b/.gitignore @@ -24,4 +24,5 @@ cmake-build-* python/paddle/v2/framework/core.so CMakeFiles cmake_install.cmake - +paddle/.timestamp +python/paddlepaddle.egg-info/ diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index 2ca988c406ae2987e26ca37dbc17cc0a2af43743..bb8c88787d37faf9ce4d7d856a307c11f1085d98 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -24,7 +24,7 @@ description: Format files with ClangFormat. entry: clang-format -i language: system - files: \.(c|cc|cxx|cpp|h|hpp|hxx)$ + files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$ - repo: https://github.com/PaddlePaddle/pre-commit-golang sha: 8337620115c25ff8333f1b1a493bd031049bd7c0 hooks: diff --git a/.travis.yml b/.travis.yml index 376c693602b56fe719decfeb41c217497e143e12..b4b83fcdbc84ce0fb0c91c816ebc3c964acfa590 100644 --- a/.travis.yml +++ b/.travis.yml @@ -37,8 +37,8 @@ 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 numpy wheel 'protobuf==3.1' sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit requests==2.9.2 LinkChecker - - pip install rarfile + - pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt + - pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker - curl https://glide.sh/get | bash - eval "$(GIMME_GO_VERSION=1.8.3 gimme)" - go get -u github.com/alecthomas/gometalinter diff --git a/CMakeLists.txt b/CMakeLists.txt index b174831109372cb014741d63032fa6a470e74042..dcd1218a5b0b62f2739b727391aca31b48ed9ccb 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -14,8 +14,8 @@ cmake_minimum_required(VERSION 3.0) set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_CURRENT_SOURCE_DIR}/cmake") -set(PROJ_ROOT ${CMAKE_CURRENT_SOURCE_DIR}) -set(PROJ_BINARY_ROOT ${CMAKE_CURRENT_BINARY_DIR}) +set(PADDLE_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}) +set(PADDLE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR}) include(system) @@ -121,8 +121,8 @@ include(version) # set PADDLE_VERSION include(coveralls) # set code coverage -include_directories("${PROJ_ROOT}") -include_directories("${PROJ_ROOT}/paddle/cuda/include") +include_directories("${PADDLE_SOURCE_DIR}") +include_directories("${PADDLE_SOURCE_DIR}/paddle/cuda/include") include_directories("${CMAKE_CURRENT_BINARY_DIR}/proto") include_directories("${CMAKE_CURRENT_BINARY_DIR}/go/pserver/client/c") include_directories(${Boost_INCLUDE_DIRS}) @@ -144,7 +144,7 @@ if(WITH_GPU) endif(WITH_GPU) if(WITH_MKLDNN) - list(APPEND EXTERNAL_LIBS ${MKLDNN_LIBRARY} ${MKLDNN_IOMP_LIB}) + list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB} ${MKLDNN_IOMP_LIB}) endif() if(USE_NNPACK) @@ -164,10 +164,12 @@ if(WITH_GOLANG) add_subdirectory(go) endif(WITH_GOLANG) +set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build") add_subdirectory(paddle) if(WITH_PYTHON) add_subdirectory(python) endif() + if(WITH_DOC) add_subdirectory(doc) endif() diff --git a/Dockerfile b/Dockerfile index 5dd9b0be4f7e0a304108abfdfb089fea4faa4d38..da0047102572d203810d2f9e5ce8ec76063d0cba 100644 --- a/Dockerfile +++ b/Dockerfile @@ -27,25 +27,24 @@ RUN apt-get update && \ git python-pip python-dev openssh-server bison \ wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \ curl sed grep graphviz libjpeg-dev zlib1g-dev \ - python-numpy python-matplotlib gcc-4.8 g++-4.8 \ - automake locales clang-format-3.8 swig doxygen cmake \ + python-matplotlib gcc-4.8 g++-4.8 \ + automake locales clang-format swig doxygen cmake \ liblapack-dev liblapacke-dev libboost-dev \ clang-3.8 llvm-3.8 libclang-3.8-dev \ net-tools && \ 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- 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 # install glide -RUN curl -q https://glide.sh/get | sh +RUN curl -s -q https://glide.sh/get | sh # git credential to skip password typing RUN git config --global credential.helper store @@ -56,19 +55,37 @@ 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 rarfile + 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 && \ (cd /woboq \ diff --git a/cmake/configure.cmake b/cmake/configure.cmake index 69220e03fe8e337205f31cb1f45e3e19ae4f5d1e..209f9078a637ac581d90212a48216eb388c477ed 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -74,8 +74,6 @@ if(WITH_MKLDNN) set(OPENMP_FLAGS "-fopenmp") set(CMAKE_C_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS}) set(CMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS}) - set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -L${MKLDNN_IOMP_DIR} -liomp5 -Wl,--as-needed") - set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -L${MKLDNN_IOMP_DIR} -liomp5 -Wl,--as-needed") set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OPENMP_FLAGS}") set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OPENMP_FLAGS}") else() @@ -131,7 +129,7 @@ if(WITH_GOLANG) add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/glide COMMAND env GOPATH=${GOPATH} ${GLIDE} install COMMAND touch ${CMAKE_BINARY_DIR}/glide - DEPENDS ${PROJ_ROOT}/go/glide.lock + DEPENDS ${PADDLE_SOURCE_DIR}/go/glide.lock WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go" ) diff --git a/cmake/cpplint.cmake b/cmake/cpplint.cmake index 656e1a0803c6e389d70f37f592c3aa2e95a2bcd4..8d5d533126c9b7fa84c725d614cf3486126d0284 100644 --- a/cmake/cpplint.cmake +++ b/cmake/cpplint.cmake @@ -42,26 +42,21 @@ macro(add_style_check_target TARGET_NAME) if(WITH_STYLE_CHECK) set(SOURCES_LIST ${ARGN}) list(REMOVE_DUPLICATES SOURCES_LIST) - list(SORT SOURCES_LIST) - foreach(filename ${SOURCES_LIST}) - set(LINT ON) foreach(pattern ${IGNORE_PATTERN}) if(filename MATCHES ${pattern}) - message(STATUS "DROP LINT ${filename}") - set(LINT OFF) + list(REMOVE_ITEM SOURCES_LIST ${filename}) endif() endforeach() - if(LINT MATCHES ON) - # cpplint code style - get_filename_component(base_filename ${filename} NAME) - set(CUR_GEN ${CMAKE_CURRENT_BINARY_DIR}/${base_filename}.cpplint) - add_custom_command(TARGET ${TARGET_NAME} PRE_BUILD - COMMAND "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py" - "--filter=${STYLE_FILTER}" - "--write-success=${CUR_GEN}" ${filename} - WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) - endif() endforeach() + + if(SOURCES_LIST) + add_custom_command(TARGET ${TARGET_NAME} POST_BUILD + COMMAND "${PYTHON_EXECUTABLE}" "${PADDLE_SOURCE_DIR}/paddle/scripts/cpplint.py" + "--filter=${STYLE_FILTER}" + ${SOURCES_LIST} + COMMENT "cpplint: Checking source code style" + WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) + endif() endif() endmacro() diff --git a/cmake/external/any.cmake b/cmake/external/any.cmake index 45e3764e8482a4cfc8ee72fe4d79f04a3c9b74fa..85cce80b70a1fcf57015ac7a264e4950616b2717 100644 --- a/cmake/external/any.cmake +++ b/cmake/external/any.cmake @@ -7,8 +7,8 @@ INCLUDE_DIRECTORIES(${ANY_SOURCE_DIR}/src/extern_lib_any) ExternalProject_Add( extern_lib_any ${EXTERNAL_PROJECT_LOG_ARGS} - GIT_REPOSITORY "https://github.com/thelink2012/any.git" - GIT_TAG "8fef1e93710a0edf8d7658999e284a1142c4c020" + GIT_REPOSITORY "https://github.com/PaddlePaddle/any.git" + GIT_TAG "15595d8324be9e8a9a80d9ae442fdd12bd66df5d" PREFIX ${ANY_SOURCE_DIR} UPDATE_COMMAND "" CONFIGURE_COMMAND "" diff --git a/cmake/external/gflags.cmake b/cmake/external/gflags.cmake index a0d0a892c4b3cc3743ac725f3cd90444f18abf34..16e5bef4cdb8d6513de51838e3c3c8398dbad60d 100644 --- a/cmake/external/gflags.cmake +++ b/cmake/external/gflags.cmake @@ -28,7 +28,14 @@ INCLUDE_DIRECTORIES(${GFLAGS_INCLUDE_DIR}) ExternalProject_Add( extern_gflags ${EXTERNAL_PROJECT_LOG_ARGS} - GIT_REPOSITORY "https://github.com/gflags/gflags.git" + # TODO(yiwang): The annoying warnings mentioned in + # https://github.com/PaddlePaddle/Paddle/issues/3277 are caused by + # gflags. I fired a PR https://github.com/gflags/gflags/pull/230 + # to fix it. Before it gets accepted by the gflags team, we use + # my personal fork, which contains above fix, temporarily. Let's + # change this back to the official Github repo once my PR is + # merged. + GIT_REPOSITORY "https://github.com/wangkuiyi/gflags.git" PREFIX ${GFLAGS_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} diff --git a/cmake/external/mklml.cmake b/cmake/external/mklml.cmake index 17a1ca4ed04dce85ae3c7fdd5f22d6eeed03db59..e9fd3d4bedc983ae7c544cf289dc841cf22f9de4 100644 --- a/cmake/external/mklml.cmake +++ b/cmake/external/mklml.cmake @@ -17,7 +17,7 @@ IF(NOT ${WITH_MKLML}) ENDIF(NOT ${WITH_MKLML}) IF(WIN32 OR APPLE) - MESSAGE(WARNING + MESSAGE(WARNING "Windows or Mac is not supported with MKLML in Paddle yet." "Force WITH_MKLML=OFF") SET(WITH_MKLML OFF CACHE STRING "Disable MKLML package in Windows and MacOS" FORCE) @@ -43,22 +43,21 @@ SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib") INCLUDE_DIRECTORIES(${MKLML_INC_DIR}) -SET(mklml_cmakefile ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt) -FILE(WRITE ${mklml_cmakefile} "PROJECT(MKLML)\n" - "cmake_minimum_required(VERSION 3.0)\n" - "install(DIRECTORY ${MKLML_VER}\n" - " DESTINATION ${MKLML_DST_DIR})\n") +FILE(WRITE ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt + "PROJECT(MKLML)\n" + "cmake_minimum_required(VERSION 3.0)\n" + "install(DIRECTORY ${MKLML_VER}\n" + " DESTINATION ${MKLML_DST_DIR})\n") ExternalProject_Add( ${MKLML_PROJECT} ${EXTERNAL_PROJECT_LOG_ARGS} PREFIX ${MKLML_SOURCE_DIR} DOWNLOAD_DIR ${MKLML_DOWNLOAD_DIR} - DOWNLOAD_COMMAND wget --no-check-certificate -O ${MKLML_DOWNLOAD_DIR}/${MKLML_VER}.tgz ${MKLML_URL} - && tar -xzf ${MKLML_DOWNLOAD_DIR}/${MKLML_VER}.tgz + DOWNLOAD_COMMAND wget --no-check-certificate -qO- ${MKLML_URL} | tar xz -C ${MKLML_DOWNLOAD_DIR} DOWNLOAD_NO_PROGRESS 1 UPDATE_COMMAND "" - CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLML_INSTALL_ROOT} + CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLML_INSTALL_ROOT} CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLML_INSTALL_ROOT} ) diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index 60a1041936437775e0994157b8ffcb7c52b7ab87..0eeccbf7d8a1df17351c8914df6dabf005802787 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -69,9 +69,22 @@ ENDIF(NOT ${CBLAS_FOUND}) MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}") INCLUDE_DIRECTORIES(${CBLAS_INC_DIR}) -ADD_LIBRARY(cblas STATIC IMPORTED) -SET_PROPERTY(TARGET cblas PROPERTY IMPORTED_LOCATION ${CBLAS_LIBRARIES}) +# FIXME(gangliao): generate cblas target to track all high performance +# linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas) +SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c) +FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";") +IF(${CBLAS_PROVIDER} 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 ${CBLAS_FOUND}) diff --git a/cmake/external/python.cmake b/cmake/external/python.cmake index 67a359d4b5f4cca8fc8e74eab4d4acb4cc12baed..490c87d67ed79a238dd506127cd4d9855fab6626 100644 --- a/cmake/external/python.cmake +++ b/cmake/external/python.cmake @@ -24,7 +24,6 @@ IF(WITH_PYTHON) ENDIF(WITH_PYTHON) SET(py_env "") -SET(USE_VIRTUALENV_FOR_TEST 1) IF(PYTHONINTERP_FOUND) find_python_module(pip REQUIRED) find_python_module(numpy REQUIRED) diff --git a/cmake/flags.cmake b/cmake/flags.cmake index d00a9bb3a30cfb16623e073414088059481c3e1a..b27eb71550b68b5c27e47bf067ae0df329bbd628 100644 --- a/cmake/flags.cmake +++ b/cmake/flags.cmake @@ -9,10 +9,12 @@ function(CheckCompilerCXX11Flag) if(${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 4.8) message(FATAL_ERROR "Unsupported GCC version. GCC >= 4.8 required.") endif() - # 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) + 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" @@ -115,7 +117,7 @@ set(COMMON_FLAGS -Wno-error=literal-suffix -Wno-error=sign-compare -Wno-error=unused-local-typedefs - -Wno-error=parentheses-equality # Warnings in Pybind11 + -Wno-error=parentheses-equality # Warnings in pybind11 ) set(GPU_COMMON_FLAGS @@ -195,6 +197,7 @@ endif() # Modern gpu architectures: Pascal if (CUDA_VERSION VERSION_GREATER "8.0" OR CUDA_VERSION VERSION_EQUAL "8.0") list(APPEND __arch_flags " -gencode arch=compute_60,code=sm_60") + list(APPEND CUDA_NVCC_FLAGS --expt-relaxed-constexpr) endif() # Custom gpu architecture diff --git a/cmake/generic.cmake b/cmake/generic.cmake index 534be0abe246ac70950d85ad05441825c8ca768a..d2aab938d4636b1583062e27b73cb30f5d56b7b0 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -187,7 +187,13 @@ function(cc_library TARGET_NAME) endif() # cpplint code style - add_style_check_target(${TARGET_NAME} ${cc_library_SRCS}) + foreach(source_file ${cc_library_SRCS}) + string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file}) + if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) + list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) + endif() + endforeach() + add_style_check_target(${TARGET_NAME} ${cc_library_SRCS} ${cc_library_HEADERS}) else(cc_library_SRCS) if (cc_library_DEPS) @@ -239,6 +245,14 @@ function(nv_library TARGET_NAME) add_dependencies(${TARGET_NAME} ${nv_library_DEPS}) target_link_libraries(${TARGET_NAME} ${nv_library_DEPS}) endif() + # cpplint code style + foreach(source_file ${nv_library_SRCS}) + string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file}) + if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) + list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) + endif() + endforeach() + add_style_check_target(${TARGET_NAME} ${nv_library_SRCS} ${nv_library_HEADERS}) else(nv_library_SRCS) if (nv_library_DEPS) merge_static_libs(${TARGET_NAME} ${nv_library_DEPS}) @@ -389,3 +403,16 @@ function(py_proto_compile TARGET_NAME) protobuf_generate_python(py_srcs ${py_proto_compile_SRCS}) add_custom_target(${TARGET_NAME} ALL DEPENDS ${py_srcs}) endfunction() + +function(py_test TARGET_NAME) + if(WITH_TESTING) + set(options STATIC static SHARED shared) + set(oneValueArgs "") + set(multiValueArgs SRCS DEPS) + cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + add_test(NAME ${TARGET_NAME} + COMMAND env PYTHONPATH=${PADDLE_PYTHON_BUILD_DIR}/lib-python + python2 ${py_test_SRCS} + WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) + endif() +endfunction() diff --git a/cmake/package.cmake b/cmake/package.cmake index ff49a2d08e8f6004320acfce266339aa301eb9c4..79e02147f3f7cc19c1bf45d8a1d208a9a32416ff 100644 --- a/cmake/package.cmake +++ b/cmake/package.cmake @@ -12,7 +12,7 @@ set(CPACK_PACKAGE_DESCRIPTION "") set(CPACK_DEBIAN_PACKAGE_DEPENDS "libpython2.7-dev, libstdc++6, python-pip, curl, libgfortran3, python-pip-whl") set(CPACK_DEBIAN_PACKAGE_SECTION Devel) set(CPACK_DEBIAN_PACKAGE_VERSION ${PADDLE_VERSION}) -set(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${PROJ_ROOT}/paddle/scripts/deb/postinst") +set(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${PADDLE_SOURCE_DIR}/paddle/scripts/deb/postinst") #set(CPACK_GENERATOR "DEB") # Start cpack include (CMakePackageConfigHelpers) diff --git a/cmake/util.cmake b/cmake/util.cmake index 87ad9d91d8701c56255c1e7f224764998df634a7..0da4969d310368ab27b0ed65237813c07d6e59f0 100644 --- a/cmake/util.cmake +++ b/cmake/util.cmake @@ -118,7 +118,6 @@ endfunction() macro(add_unittest_without_exec TARGET_NAME) add_executable(${TARGET_NAME} ${ARGN}) link_paddle_test(${TARGET_NAME}) - add_style_check_target(${TARGET_NAME} ${ARGN}) endmacro() # add_unittest @@ -142,17 +141,20 @@ endmacro() function(create_resources res_file output_file) add_custom_command( OUTPUT ${output_file} - COMMAND python ARGS ${PROJ_ROOT}/cmake/make_resource.py ${res_file} ${output_file} - DEPENDS ${res_file} ${PROJ_ROOT}/cmake/make_resource.py) + COMMAND python ARGS ${PADDLE_SOURCE_DIR}/cmake/make_resource.py ${res_file} ${output_file} + DEPENDS ${res_file} ${PADDLE_SOURCE_DIR}/cmake/make_resource.py) endfunction() # Create a python unittest using run_python_tests.sh, # which takes care of making correct running environment function(add_python_test TEST_NAME) - add_test(NAME ${TEST_NAME} - COMMAND env PADDLE_PACKAGE_DIR=${PADDLE_PYTHON_PACKAGE_DIR} - bash ${PROJ_ROOT}/paddle/scripts/run_python_tests.sh - ${USE_VIRTUALENV_FOR_TEST} ${PYTHON_EXECUTABLE} ${ARGN} - WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) + foreach(arg ${ARGN}) + get_filename_component(py_fn ${arg} NAME_WE) + set(TRG_NAME ${TEST_NAME}_${py_fn}) + add_test(NAME ${TRG_NAME} + COMMAND env PYTHONPATH=${PADDLE_PYTHON_PACKAGE_DIR} + python2 ${arg} + WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) + endforeach() endfunction() diff --git a/cmake/version.cmake b/cmake/version.cmake index ac1583a24c828629c46cb9cf4e965f8da2273732..cde650128a068faf32f4abfff5cdfdeb656d8577 100644 --- a/cmake/version.cmake +++ b/cmake/version.cmake @@ -4,7 +4,7 @@ set(tmp_version "HEAD") while ("${PADDLE_VERSION}" STREQUAL "") execute_process( COMMAND ${GIT_EXECUTABLE} describe --tags --abbrev=0 ${tmp_version} - WORKING_DIRECTORY ${PROJ_ROOT} + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR} OUTPUT_VARIABLE GIT_TAG_NAME RESULT_VARIABLE GIT_RESULT ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) diff --git a/doc/api/v2/config/layer.rst b/doc/api/v2/config/layer.rst index 372272a53c12c314fc80eebbce5eae9fcabc55ba..cb330ea5e1b914587a725c9b90a33053f3fbbc3d 100644 --- a/doc/api/v2/config/layer.rst +++ b/doc/api/v2/config/layer.rst @@ -257,6 +257,16 @@ seq_concat .. autoclass:: paddle.v2.layer.seq_concat :noindex: +kmax_sequence_score +------------------- +.. autoclass:: paddle.v2.layer.kmax_sequence_score + :noindex: + +sub_nested_seq +-------------- +.. autoclass:: paddle.v2.layer.sub_nested_seq + :noindex: + Reshaping Layers ================ diff --git a/doc/design/auto_gradient_check.md b/doc/design/auto_gradient_check.md new file mode 100644 index 0000000000000000000000000000000000000000..1f4d4ec16f7c395005e610751d95c10f5f3adf52 --- /dev/null +++ b/doc/design/auto_gradient_check.md @@ -0,0 +1,146 @@ +## Auto Gradient Checker Design + +## Backgraound： +- Operator forward computing is easy to check if the result is right because it has a clear definition. **But** backpropagation is a notoriously difficult algorithm to debug and get right: + - 1. you should get the right backpropagation formula according to the forward computation. + - 2. you should implement it right in CPP. + - 3. it's difficult to prepare test data. + +- Auto gradient check gets a numeric gradient by forward Operator and use it as a reference of the backward Operator's result. It has several advantages: + - 1. numeric gradient checker only need forward operator. + - 2. user only need to prepare the input data for forward Operator. + +## Mathematical Theory +The following two document from stanford has a detailed explanation of how to get numeric gradient and why it's useful. + +- [Gradient checking and advanced optimization(en)](http://deeplearning.stanford.edu/wiki/index.php/Gradient_checking_and_advanced_optimization) +- [Gradient checking and advanced optimization(cn)](http://ufldl.stanford.edu/wiki/index.php/%E6%A2%AF%E5%BA%A6%E6%A3%80%E9%AA%8C%E4%B8%8E%E9%AB%98%E7%BA%A7%E4%BC%98%E5%8C%96) + + +## Numeric Gradient Implementation +### Python Interface +```python +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. + """ +``` + +### Explaination: + +- Why need `output_name` + - One Operator may have multiple Output, you can get independent gradient from each Output. So user should set one output to calculate. + +- Why need `input_to_check` + - One operator may have multiple inputs. Gradient Op can calculate the gradient of these Inputs at the same time. But Numeric Gradient needs to calculate them one by one. So `get_numeric_gradient` is designed to calculate the gradient for one input. If you need to compute multiple inputs, you can call `get_numeric_gradient` multiple times. + + +### Core Algorithm Implementation + + +```python + # we only compute gradient of one element each time. + # we use a for loop to compute the gradient of every element. + 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()) +``` + +## Auto Graident Checker Framework + +Each Operator Kernel has three kinds of Gradient: + +- 1. Numeric Gradient +- 2. CPU Operator Gradient +- 3. GPU Operator Gradient(if supported) + +Numeric Gradient Only relies on forward Operator. So we use Numeric Gradient as the reference value. + +- 1. calculate the numeric gradient. +- 2. calculate CPU kernel Gradient with the backward Operator and compare it with the numeric gradient. +- 3. calculate GPU kernel Gradient with the backward Operator and compare it with the numeric gradient.(if support GPU) + +#### Python Interface + +```python + 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: + """ +``` + +### How to check if two numpy array is close enough? +if `abs_numeric_grad` is nearly zero, then use abs error for numeric_grad, not relative + +```python +numeric_grad = ... +operator_grad = numpy.array(scope.find_var(grad_var_name(name)).get_tensor()) + +abs_numeric_grad = numpy.abs(numeric_grad) +# if abs_numeric_grad is nearly zero, then use abs error for numeric_grad, not relative +# error. +abs_numeric_grad[abs_numeric_grad < 1e-3] = 1 + +diff_mat = numpy.abs(abs_numeric_grad - operator_grad) / abs_numeric_grad +max_diff = numpy.max(diff_mat) +``` + + +#### Notes： +1，The Input data for auto gradient checker should be reasonable to avoid numeric problem. + + +#### Refs: + +- [Gradient checking and advanced optimization(en)](http://deeplearning.stanford.edu/wiki/index.php/Gradient_checking_and_advanced_optimization) +- [Gradient checking and advanced optimization(cn)](http://ufldl.stanford.edu/wiki/index.php/%E6%A2%AF%E5%BA%A6%E6%A3%80%E9%AA%8C%E4%B8%8E%E9%AB%98%E7%BA%A7%E4%BC%98%E5%8C%96) 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/mkldnn/README.MD b/doc/design/mkldnn/README.MD new file mode 100644 index 0000000000000000000000000000000000000000..e956994431fbb43438c56dcd96ad8313cf516090 --- /dev/null +++ b/doc/design/mkldnn/README.MD @@ -0,0 +1,110 @@ +# Intel® MKL-DNN on PaddlePaddle: Design Doc + +我们计划将Intel深度神经网络数学库(**MKL-DNN**\[[1](#references)\])集成到PaddlePaddle，充分展现英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。 + +我们短期内的基本目标是： + +- 完成常用layer的MKL-DNN实现。 +- 完成常见深度神经网络VGG，GoogLeNet 和 ResNet的MKL-DNN实现。 + + +## Contents + +- [Overview](#overview) +- [Actions](#actions) + - [CMake](#cmake) + - [Layers](#layers) + - [Activations](#activations) + - [Unit Tests](#unit-tests) + - [Protobuf Messages](#protobuf-messages) + - [Python API](#python-api) + - [Demos](#demos) + - [Benchmarking](#benchmarking) + - [Others](#others) +- [Design Concerns](#design-concerns) + +## Overview + +我们会把MKL-DNN作为第三方库集成进PaddlePaddle，整体框架图 +
+
+Figure 1. PaddlePaddle on IA. +
+ +## Actions +我们把集成方案大致分为了如下几个方面。 + +### CMake +我们会在`CMakeLists.txt`中会添加`WITH_MKLDNN`的选项，当设置这个值为`ON`的时候会启用编译MKL-DNN功能。同时会自动开启OpenMP用于提高MKL-DNN的性能。 + +同时，我们会引入`WITH_MKLML`选项，用于选择是否使用MKL-DNN自带的MKLML安装包。这个安装包可以独立于MKL-DNN使用，但是建议在开启MKL-DNN的同时也打开MKLML的开关，这样才能发挥最好的性能。 + +所以，我们会在`cmake/external`目录新建`mkldnn.cmake`和`mklml.cmake`文件，它们会在编译PaddlePaddle的时候下载对应的软件包，并放到PaddlePaddle的third party目录中。 + +**备注**：当`WITH_MKLML=ON`的时候，会优先使用这个包作为PaddlePaddle的CBLAS和LAPACK库，所以会稍微改动`cmake/cblas.cmake`中的逻辑。 + +### Layers +所有MKL-DNN相关的C++ layers，都会按照PaddlePaddle的目录结构存放在 +`paddle/gserver/layers`中，并且文件名都会一以*Mkldnn*开头。 + +所有MKL-DNN的layers都会继承于一个叫做`MkldnnLayer`的父类，该父类继承于PaddlePaddle的基类`Layer`。 + +### Activations +由于在PaddlePaddle中，激活函数是独立于layer概念的，所以会在`paddle/gserver/activations`目录下添加一个`MkldnnActivation.h`文件定义一些用于MKL-DNN的接口，实现方法还是会在`ActivationFunction.cpp`文件。 + +### Unit Tests +会在`paddle/gserver/test`目录下添加`test_Mkldnn.cpp`和`MkldnnTester.*`用于MKL-DNN的测试。 + +Activation的测试，计划在PaddlePaddle原有的测试文件上直接添加新的测试type。 + +### Protobuf Messages +根据具体layer的需求可能会在`proto/ModelConfig.proto`里面添加必要的选项。 + +### Python API +目前只考虑**v1 API**。 + +计划在`python/paddle/trainer/config_parser.py`里面添加`use_mkldnn`这个选择，方便用户选择使用MKL-DNN的layers。 + +具体实现方式比如： + +```python +use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0))) +if use_mkldnn + self.layer_type = mkldnn_* +``` + +所有MKL-DNN的layer type会以*mkldnn_*开头，以示区分。 + +并且可能在`python/paddle/trainer_config_helper`目录下的`activations.py `和`layers.py`里面添加必要的MKL-DNN的接口。 + +### Demos + +会在`v1_api_demo`目录下添加一个`mkldnn`的文件夹，里面放入一些用于MKL-DNN测试的demo脚本。 + +### Benchmarking +会考虑添加部分逻辑在`benchmark/paddle/image/run.sh`，添加使用MKL-DNN的测试。 + +### Others +1. 如果在使用MKL-DNN的情况下，会把CPU的Buffer对齐为64。 +2. 深入PaddlePaddle，寻找有没有其他可以优化的可能，进一步优化。比如可能会用OpenMP改进SGD的更新性能。 + +## Design Concerns + +为了更好的符合PaddlePaddle的代码风格\[[2](#references)\]，同时又尽可能少的牺牲MKL-DNN的性能\[[3](#references)\]。 + +我们总结出一些特别需要注意的点： + +1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数，我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MkldnnLayer`特有的设备ID。 +2. 重写父类Layer的**init**函数，修改`deviceId_`为`-2`，代表这个layer是用于跑在MKL-DNN的环境下。 +3. 创建`MkldnnMatrix`，用于管理MKL-DNN会用到的相关memory函数、接口以及会用的到格式信息。 +4. 创建`MkldnnBase`，定义一些除了layer和memory相关的类和函数。包括MKL-DNN会用到`MkldnnStream`和`CpuEngine`，和未来可能还会用到`FPGAEngine`等。 +5. 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。 + +## References + +1. [Intel Math Kernel Library for Deep Neural Networks (Intel MKL-DNN)](https://github.com/01org/mkl-dnn "Intel MKL-DNN") +2. [原来的方案](https://github.com/PaddlePaddle/Paddle/pull/3096)会引入**nextLayer**的信息。但是在PaddlePaddle中，无论是重构前的layer还是重构后的op，都不会想要知道next layer/op的信息。 +3. MKL-DNN的高性能格式与PaddlePaddle原有的`NCHW`不同(PaddlePaddle中的CUDNN部分使用的也是`NCHW`，所以不存在这个问题)，所以需要引入一个转换方法，并且只需要在必要的时候转换这种格式，才能更好的发挥MKL-DNN的性能。 + diff --git a/doc/design/mkldnn/image/overview.png b/doc/design/mkldnn/image/overview.png new file mode 100644 index 0000000000000000000000000000000000000000..84b455c28230703599a2529f014cfbb222138fef Binary files /dev/null and b/doc/design/mkldnn/image/overview.png differ diff --git a/doc/design/releasing_process.md b/doc/design/releasing_process.md index 3692a5248a355cfcfd1cfd0911d43d65166921b1..0c10e782808ca6456347ec54cb5e921162731ede 100644 --- a/doc/design/releasing_process.md +++ b/doc/design/releasing_process.md @@ -11,6 +11,15 @@ Paddle每次发新的版本，遵循以下流程: * 编译这个版本的Ubuntu Deb包。如果失败，修复Ubuntu Deb包编译问题，Patch号加一，返回第二步。 * 使用Regression Test List作为检查列表，测试Docker镜像/ubuntu安装包的功能正确性 * 如果失败，记录下所有失败的例子，在这个`release/版本号`分支中，修复所有bug后，Patch号加一，返回第二步 + * 编译这个版本的python wheel包，并发布到pypi。 + * 由于pypi.python.org目前遵循[严格的命名规范PEP 513](https://www.python.org/dev/peps/pep-0513)，在使用twine上传之前，需要重命名wheel包中platform相关的后缀，比如将`linux_x86_64`修改成`manylinux1_x86_64`。 + * pypi上的package名称为paddlepaddle和paddlepaddle_gpu，如果要上传GPU版本的包，需要修改build/python/setup.py中，name: "paddlepaddle_gpu"并重新打包wheel包：`python setup.py bdist_wheel`。 + * 上传方法： + ``` + cd build/python + pip install twine + twine upload dist/[package to upload] + ``` 4. 第三步完成后，将`release/版本号`分支合入master分支，并删除`release/版本号`分支。将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。最后删除`release/版本号`分支。 5. 编译master分支的Docker发行镜像，发布到dockerhub。编译ubuntu的deb包，发布到github release页面 6. 协同完成Release Note的书写 diff --git a/doc/getstarted/build_and_install/docker_install_cn.rst b/doc/getstarted/build_and_install/docker_install_cn.rst index 87c286a1af75e08313813f1373ea03b85d4af523..84e33177740ca1652efc09c8081c2519b4366906 100644 --- a/doc/getstarted/build_and_install/docker_install_cn.rst +++ b/doc/getstarted/build_and_install/docker_install_cn.rst @@ -3,6 +3,43 @@ PaddlePaddle的Docker容器使用方式 PaddlePaddle目前唯一官方支持的运行的方式是Docker容器。因为Docker能在所有主要操作系统（包括Linux，Mac OS X和Windows）上运行。 请注意，您需要更改 `Dockers设置 `_ 才能充分利用Mac OS X和Windows上的硬件资源。 +Docker使用入门 +------------------------------ + +几个基础的概念帮助理解和使用Docker： + +- *镜像*：一个Docker镜像是一个打包好的软件。它包含了这个软件本身和它所依赖的运行环境。PaddlePaddle的Docker镜像就包含了PaddlePaddle的Python库以及其依赖的多个Python库。这样我们可以直接在Docker中运行需要的程序而不需要安装后在执行。可以执行： + + .. code-block:: bash + + docker images + + 来列出当前系统中的所有镜像，同样可以执行： + + .. code-block:: bash + + docker pull paddlepaddle/paddle:0.10.0 + + 来下载Docker镜像，paddlepaddle/paddle是从官方镜像源Dockerhub.com下载的，推荐国内用户使用ocker.paddlepaddle.org/paddle下载。 + +- *容器*： 如果说一个Docker镜像就是一个程序，那容器就是这个程序运行时产生的“进程”。 + 实际上，一个容器就是一个操作系统的进程，但是是运行在独立的进程空间，文件系统以及网络之上。 + 可以执行： + + .. code-block:: bash + + docker run paddlepaddle/paddle:0.10.0 + + 来使用一个镜像启动一个容器。 + +- 默认情况下，Docker容器会运行在独立的文件系统空间之上，我们无法在Docker容器中 + 访问到主机上的文件。可以通过*挂载Volume*的方式，将主机上的文件或目录挂载到 + Docker容器中。下面的命令把当前目录挂载到了容器中的 /data 目录下，容器使用 + debian镜像，并且启动后执行 :code:`ls /data`。 + + .. code-block:: bash + + docker run --rm -v $(pwd):/data debian ls /data PaddlePaddle发布的Docker镜像使用说明 ------------------------------ @@ -12,11 +49,11 @@ PaddlePaddle需要的所有编译工具。把编译出来的PaddlePaddle也打 像，称为生产镜像，里面涵盖了PaddlePaddle运行所需的所有环境。每次 PaddlePaddle发布新版本的时候都会发布对应版本的生产镜像以及开发镜像。运 行镜像包括纯CPU版本和GPU版本以及其对应的非AVX版本。我们会在 -`dockerhub.com `_ 提供最新 -的Docker镜像，可以在"tags"标签下找到最新的Paddle镜像版本。为了方便在国 -内的开发者下载Docker镜像，我们提供了国内的镜像服务器供大家使用。如果您 -在国内，请把文档里命令中的paddlepaddle/paddle替换成 -docker.paddlepaddle.org/paddle。 +`dockerhub.com `_ +和国内镜像`docker.paddlepaddle.org` 提供最新 +的Docker镜像，可以在"tags"标签下找到最新的Paddle镜像版本。 + +**注意：为了方便在国内的开发者下载Docker镜像，我们提供了国内的镜像服务器供大家使用。如果您在国内，请把文档里命令中的paddlepaddle/paddle替换成docker.paddlepaddle.org/paddle。** 1. 开发镜像：:code:`paddlepaddle/paddle:0.10.0-dev` @@ -37,13 +74,13 @@ docker.paddlepaddle.org/paddle。 .. code-block:: bash - docker run -it --rm paddlepaddle/paddle:0.10.0-dev /bin/bash + docker run -it --rm -v $(pwd):/paddle paddlepaddle/paddle:0.10.0-dev /bin/bash 或者，可以以后台进程方式运行容器： .. code-block:: bash - docker run -d -p 2202:22 -p 8888:8888 paddledev/paddle:0.10.0-dev + docker run -d -p 2202:22 -p 8888:8888 -v $(pwd):/paddle paddlepaddle/paddle:0.10.0-dev /usr/sbin/sshd -D 然后用密码 :code:`root` SSH进入容器： @@ -68,6 +105,8 @@ docker.paddlepaddle.org/paddle。 如果输出是No，就需要选择使用no-AVX的镜像 + **注：在0.10.0之后的版本，PaddlePaddle都可以自动判断硬件是否支持AVX，所以无需判断AVX即可使用** + 以上方法在GPU镜像里也能用，只是请不要忘记提前在物理机上安装GPU最新驱动。 为了保证GPU驱动能够在镜像里面正常运行，我们推荐使用[nvidia-docker](https://github.com/NVIDIA/nvidia-docker)来运行镜像。 diff --git a/doc/getstarted/build_and_install/docker_install_en.rst b/doc/getstarted/build_and_install/docker_install_en.rst index b6fd3329b273aabe80edd5f1ff064a311648b3c2..94860240f6a4a9bed8a865684a8a79960489280e 100644 --- a/doc/getstarted/build_and_install/docker_install_en.rst +++ b/doc/getstarted/build_and_install/docker_install_en.rst @@ -63,12 +63,35 @@ CPU-only version and a CUDA GPU version and their no-AVX versions. We put the docker images on `dockerhub.com `_. You can find the -latest versions under "tags" tab at dockerhub.com. If you are in -China, you can use our Docker image registry mirror to speed up the -download process. To use it, please replace all paddlepaddle/paddle in -the commands to docker.paddlepaddle.org/paddle. +latest versions under "tags" tab at dockerhub.com. -1. Production images, this image might have multiple variants: +** NOTE: If you are in China, you can use our Docker image registry mirror to speed up the download process. To use it, please replace all paddlepaddle/paddle in the commands to docker.paddlepaddle.org/paddle.** + + +1. development image :code:`paddlepaddle/paddle:-dev` + + This image has packed related develop tools and runtime + environment. Users and developers can use this image instead of + their own local computer to accomplish development, build, + releasing, document writing etc. While different version of paddle + may depends on different version of libraries and tools, if you + want to setup a local environment, you must pay attention to the + versions. The development image contains: + + - gcc/clang + - nvcc + - Python + - sphinx + - woboq + - sshd + + Many developers use servers with GPUs, they can use ssh to login to + the server and run :code:`docker exec` to enter the docker + container and start their work. Also they can start a development + docker image with SSHD service, so they can login to the container + and start work. + +2. Production images, this image might have multiple variants: - GPU/AVX：:code:`paddlepaddle/paddle:-gpu` - GPU/no-AVX：:code:`paddlepaddle/paddle:-gpu-noavx` @@ -84,7 +107,7 @@ the commands to docker.paddlepaddle.org/paddle. if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi - + **NOTE：versions after 0.10.0 will automatically detect system AVX support, so manual detect is not needed in this case.** To run the CPU-only image as an interactive container: .. code-block:: bash @@ -103,29 +126,6 @@ the commands to docker.paddlepaddle.org/paddle. nvidia-docker run -it --rm paddlepaddle/paddle:0.10.0-gpu /bin/bash -2. development image :code:`paddlepaddle/paddle:-dev` - - This image has packed related develop tools and runtime - environment. Users and developers can use this image instead of - their own local computer to accomplish development, build, - releasing, document writing etc. While different version of paddle - may depends on different version of libraries and tools, if you - want to setup a local environment, you must pay attention to the - versions. The development image contains: - - - gcc/clang - - nvcc - - Python - - sphinx - - woboq - - sshd - - Many developers use servers with GPUs, they can use ssh to login to - the server and run :code:`docker exec` to enter the docker - container and start their work. Also they can start a development - docker image with SSHD service, so they can login to the container - and start work. - Train Model Using Python API ---------------------------- diff --git a/doc/templates/conf.py.cn.in b/doc/templates/conf.py.cn.in index 95cad835b11816f4d2e256c2abd662a545a5bad2..41b35b5b233abd737db07aaeb6c6dd4bf6d42b08 100644 --- a/doc/templates/conf.py.cn.in +++ b/doc/templates/conf.py.cn.in @@ -13,22 +13,18 @@ # serve to show the default. import sys import os, subprocess +sys.path.insert(0, os.path.abspath('@PADDLE_SOURCE_DIR@/python')) import shlex from recommonmark import parser, transform -try: - import py_paddle - import paddle - import paddle.v2 -except ImportError: - print("Must install paddle python package before generating documentation") - sys.exit(1) +import paddle +import paddle.v2 MarkdownParser = parser.CommonMarkParser AutoStructify = transform.AutoStructify # If extensions (or modules to document with autodoc) are in another directory, # add these directories to sys.path here. If the directory is relative to the # documentation root, use os.path.abspath to make it absolute, like shown here. -templates_path = ["@PROJ_ROOT@/doc_theme/templates"] +templates_path = ["@PADDLE_SOURCE_DIR@/doc_theme/templates"] # -- General configuration ------------------------------------------------ @@ -124,7 +120,7 @@ html_theme = 'sphinx_rtd_theme' # Add any paths that contain custom static files (such as style sheets) here, # relative to this directory. They are copied after the builtin static files, # so a file named "default.css" will overwrite the builtin "default.css". -html_static_path = ['@PROJ_ROOT@/doc_theme/static'] +html_static_path = ['@PADDLE_SOURCE_DIR@/doc_theme/static'] # Output file base name for HTML help builder. htmlhelp_basename = project + 'doc' diff --git a/doc/templates/conf.py.en.in b/doc/templates/conf.py.en.in index b477f0120c4fa0544012080b7cfb8572d3c44b04..5822c2481dd61da2084b0de76f6f65aa4e32e033 100644 --- a/doc/templates/conf.py.en.in +++ b/doc/templates/conf.py.en.in @@ -13,15 +13,11 @@ # serve to show the default. import sys import os, subprocess +sys.path.insert(0, os.path.abspath('@PADDLE_SOURCE_DIR@/python')) import shlex from recommonmark import parser, transform -try: - import py_paddle - import paddle - import paddle.v2 -except ImportError: - print("Must install paddle python package before generating documentation") - sys.exit(1) +import paddle +import paddle.v2 MarkdownParser = parser.CommonMarkParser @@ -29,7 +25,7 @@ AutoStructify = transform.AutoStructify # If extensions (or modules to document with autodoc) are in another directory, # add these directories to sys.path here. If the directory is relative to the # documentation root, use os.path.abspath to make it absolute, like shown here. -templates_path = ["@PROJ_ROOT@/doc_theme/templates"] +templates_path = ["@PADDLE_SOURCE_DIR@/doc_theme/templates"] # -- General configuration ------------------------------------------------ @@ -124,7 +120,7 @@ html_theme = 'sphinx_rtd_theme' # Add any paths that contain custom static files (such as style sheets) here, # relative to this directory. They are copied after the builtin static files, # so a file named "default.css" will overwrite the builtin "default.css". -html_static_path = ['@PROJ_ROOT@/doc_theme/static'] +html_static_path = ['@PADDLE_SOURCE_DIR@/doc_theme/static'] # Output file base name for HTML help builder. htmlhelp_basename = project + 'doc' diff --git a/go/cmd/pserver/pserver.go b/go/cmd/pserver/pserver.go index f9cd8f87e8f2e715c87834ee08482be0f511f681..bec5775d540729000ab2dd3002600f0a92619d70 100644 --- a/go/cmd/pserver/pserver.go +++ b/go/cmd/pserver/pserver.go @@ -32,7 +32,7 @@ import ( func main() { port := flag.Int("port", 0, "port of the pserver") - index := flag.Int("index", -1, "index of this pserver, should be larger or equal than 0") + index := flag.Int("index", -1, "index of the pserver, set to -1 if use etcd for auto pserver index registry") etcdEndpoint := flag.String("etcd-endpoint", "http://127.0.0.1:2379", "comma separated endpoint string for pserver to connect to etcd") dialTimeout := flag.Duration("dial-timeout", 5*time.Second, "dial timeout") @@ -60,12 +60,12 @@ func main() { idx, err = e.Register(*port) candy.Must(err) - cp, err = pserver.NewCheckpointFromFile(*checkpointPath, idx, e) + cp, err = pserver.LoadCheckpoint(e, idx) if err != nil { if err == pserver.ErrCheckpointNotFound { log.Infof("Could not find the pserver checkpoint.") } else { - log.Errorf("Fetch checkpoint failed, %s", err) + panic(err) } } } diff --git a/go/glide.lock b/go/glide.lock index 1f16abdf66422abcd0ab7987cab3499d02cf1b9c..1ecdd217520e0a62b546b4c7048a25f4316d3f37 100644 --- a/go/glide.lock +++ b/go/glide.lock @@ -1,5 +1,5 @@ -hash: 2a1c0eca5c07a130e3d224f9821f96cfa37a39bf6bce141c855bbc57ef569f1c -updated: 2017-07-29T07:34:48.722757905+08:00 +hash: 1b9b07408ca7fac27a374dc2ccd2433e4bff090484008a037df967284949a582 +updated: 2017-08-07T23:37:48.867469328Z imports: - name: github.com/beorn7/perks version: 4c0e84591b9aa9e6dcfdf3e020114cd81f89d5f9 @@ -10,7 +10,7 @@ imports: - name: github.com/cockroachdb/cmux version: 112f0506e7743d64a6eb8fedbcff13d9979bbf92 - name: github.com/coreos/etcd - version: c31bec0f29facff13f7c3e3d948e55dd6689ed42 + version: d0d1a87aa96ae14914751d42264262cb69eda170 subpackages: - alarm - auth @@ -24,6 +24,7 @@ imports: - error - etcdserver - etcdserver/api + - etcdserver/api/etcdhttp - etcdserver/api/v2http - etcdserver/api/v2http/httptypes - etcdserver/api/v3client @@ -145,6 +146,8 @@ imports: version: a1dba9ce8baed984a2495b658c82687f8157b98f subpackages: - xfs +- name: github.com/satori/go.uuid + version: 879c5887cd475cd7864858769793b2ceb0d44feb - name: github.com/sirupsen/logrus version: a3f95b5c423586578a4e099b11a46c2479628cac - name: github.com/topicai/candy @@ -208,11 +211,6 @@ testImports: version: 04cdfd42973bb9c8589fd6a731800cf222fde1a9 subpackages: - spew -- name: github.com/docker/docker - version: b6d164e6c46d8115b146e4c3ac93784e9ef8b49e - subpackages: - - pkg/ioutils - - pkg/longpath - name: github.com/pmezard/go-difflib version: d8ed2627bdf02c080bf22230dbb337003b7aba2d subpackages: diff --git a/go/glide.yaml b/go/glide.yaml index bc23fa6ebf2c3db61e2d63e5f7e7ddcb595dfed0..a90e71b615de92d64c79823e2a04c46001963932 100644 --- a/go/glide.yaml +++ b/go/glide.yaml @@ -14,11 +14,13 @@ import: version: ^1.0.0 - package: github.com/topicai/candy - package: golang.org/x/crypto - vcs: git repo: https://github.com/golang/crypto.git -- package: golang.org/x/sys vcs: git +- package: golang.org/x/sys repo: https://github.com/golang/sys.git -- package: golang.org/x/text vcs: git +- package: golang.org/x/text repo: https://github.com/golang/text.git + vcs: git +- package: github.com/satori/go.uuid + version: v1.1.0 diff --git a/go/master/service.go b/go/master/service.go index d30e9a33229c0aff354417771b5bf2ae6a781715..df7c6860e6ae13a5be7d0425273812208685ee9d 100644 --- a/go/master/service.go +++ b/go/master/service.go @@ -77,11 +77,12 @@ type taskEntry struct { NumFailure int } -type taskQueues struct { +type masterState struct { Todo []taskEntry Pending map[int]taskEntry // map from task ID to task entry Done []taskEntry Failed []taskEntry + CurPass int } // Service is the master server service. @@ -94,11 +95,11 @@ type Service struct { ready chan struct{} initDone bool - mu sync.Mutex - taskQueues taskQueues - currPass int - jobTasks []taskEntry - + mu sync.Mutex + // State to be persisted to snapshot. + state masterState + // The trainer that is currently saving model. This state is + // transient, does not need to be persisted to snapshot. savingTrainer string } @@ -141,8 +142,8 @@ func NewService(store Store, chunksPerTask int, timeoutDur time.Duration, failur s.chunksPerTask = chunksPerTask s.timeoutDur = timeoutDur s.failureMax = failureMax - s.taskQueues = taskQueues{} - s.taskQueues.Pending = make(map[int]taskEntry) + s.state = masterState{} + s.state.Pending = make(map[int]taskEntry) s.ready = make(chan struct{}) s.store = store recovered, err := s.recover() @@ -180,7 +181,7 @@ func (s *Service) recover() (bool, error) { } dec := gob.NewDecoder(gr) - var tqs taskQueues + var tqs masterState err = dec.Decode(&tqs) if err != nil { return false, err @@ -193,7 +194,12 @@ func (s *Service) recover() (bool, error) { log.Errorln(err) } - s.taskQueues = tqs + s.state = tqs + log.WithFields(s.logFields()).Infof("Master recovered from snapshot, scheduling pending task timeout check.") + for _, t := range s.state.Pending { + time.AfterFunc(s.timeoutDur, s.checkTimeoutFunc(t.Task.Meta.ID, t.Task.Meta.Epoch)) + } + return true, nil } @@ -208,7 +214,7 @@ func (s *Service) snapshot() error { var buf bytes.Buffer gw := gzip.NewWriter(&buf) enc := gob.NewEncoder(gw) - err := enc.Encode(s.taskQueues) + err := enc.Encode(s.state) if err != nil { return err } @@ -290,8 +296,7 @@ func (s *Service) SetDataset(globPaths []string, _ *int) error { return err } - s.jobTasks = partition(chunks, s.chunksPerTask) - s.taskQueues.Todo = s.jobTasks + s.state.Todo = partition(chunks, s.chunksPerTask) err = s.snapshot() if err != nil { @@ -319,17 +324,17 @@ func (s *Service) processFailedTask(t taskEntry, epoch int) { } }() - delete(s.taskQueues.Pending, t.Task.Meta.ID) + delete(s.state.Pending, t.Task.Meta.ID) t.NumFailure++ if t.NumFailure > s.failureMax { log.Warningf("Task %v failed %d times, discard.", t.Task, t.NumFailure) - s.taskQueues.Failed = append(s.taskQueues.Failed, t) + s.state.Failed = append(s.state.Failed, t) return } log.Warningf("Task %v failed %d times, re-dispatch.", t.Task, t.NumFailure) - s.taskQueues.Todo = append(s.taskQueues.Todo, t) + s.state.Todo = append(s.state.Todo, t) return } @@ -338,7 +343,7 @@ func (s *Service) checkTimeoutFunc(taskID int, epoch int) func() { s.mu.Lock() defer s.mu.Unlock() - t, ok := s.taskQueues.Pending[taskID] + t, ok := s.state.Pending[taskID] if !ok { return } @@ -350,10 +355,11 @@ func (s *Service) checkTimeoutFunc(taskID int, epoch int) func() { // must be called with lock held. func (s *Service) logFields() log.Fields { return log.Fields{ - "todoLen": len(s.taskQueues.Todo), - "pendingLen": len(s.taskQueues.Pending), - "doneLen": len(s.taskQueues.Done), - "failedLen": len(s.taskQueues.Failed), + "todoLen": len(s.state.Todo), + "pendingLen": len(s.state.Pending), + "doneLen": len(s.state.Done), + "failedLen": len(s.state.Failed), + "curPass": s.state.CurPass, } } @@ -366,17 +372,17 @@ func (s *Service) GetTask(passID int, task *Task) error { s.mu.Lock() defer s.mu.Unlock() - if passID < s.currPass { + if passID < s.state.CurPass { return ErrPassBefore } - if passID > s.currPass { + if passID > s.state.CurPass { // Client may get run to pass after master when one client faster than the // other return ErrPassAfter } - if len(s.taskQueues.Todo) == 0 { - if len(s.taskQueues.Done) == 0 && len(s.taskQueues.Pending) == 0 { + if len(s.state.Todo) == 0 { + if len(s.state.Done) == 0 && len(s.state.Pending) == 0 { log.WithFields(s.logFields()).Warningln("All tasks failed, may start next pass") return ErrAllTaskFailed } @@ -384,10 +390,10 @@ func (s *Service) GetTask(passID int, task *Task) error { return ErrNoMoreAvailable } - t := s.taskQueues.Todo[0] + t := s.state.Todo[0] t.Task.Meta.Epoch++ - s.taskQueues.Todo = s.taskQueues.Todo[1:] - s.taskQueues.Pending[t.Task.Meta.ID] = t + s.state.Todo = s.state.Todo[1:] + s.state.Pending[t.Task.Meta.ID] = t err := s.snapshot() if err != nil { return err @@ -409,7 +415,7 @@ func (s *Service) TaskFinished(taskID int, dummy *int) error { s.mu.Lock() defer s.mu.Unlock() - t, ok := s.taskQueues.Pending[taskID] + t, ok := s.state.Pending[taskID] if !ok { log.WithFields(s.logFields()).Warningln("Pending task #%d not found.", taskID) return nil @@ -417,18 +423,18 @@ func (s *Service) TaskFinished(taskID int, dummy *int) error { // task finished, reset timeout t.NumFailure = 0 - s.taskQueues.Done = append(s.taskQueues.Done, t) - delete(s.taskQueues.Pending, taskID) + s.state.Done = append(s.state.Done, t) + delete(s.state.Pending, taskID) log.WithFields(s.logFields()).Infof("Task #%d finished.", taskID) - if len(s.taskQueues.Todo) == 0 && len(s.taskQueues.Pending) == 0 { + if len(s.state.Todo) == 0 && len(s.state.Pending) == 0 { // increase master side pass count if all tasks finished - s.currPass++ - s.taskQueues.Todo = s.jobTasks - s.taskQueues.Done = []taskEntry{} + s.state.CurPass++ + s.state.Todo = append(s.state.Done, s.state.Failed...) + s.state.Done = []taskEntry{} // TODO(typhoonzero): deal with failed tasks - s.taskQueues.Failed = []taskEntry{} - log.WithFields(s.logFields()).Warningf("all task finished, add new pass data, newpass: %d.", s.currPass) + s.state.Failed = []taskEntry{} + log.WithFields(s.logFields()).Warningf("all task finished, add new pass data, newpass: %d.", s.state.CurPass) } err := s.snapshot() @@ -447,7 +453,7 @@ func (s *Service) TaskFailed(meta TaskMeta, dummy *int) error { s.mu.Lock() defer s.mu.Unlock() - t, ok := s.taskQueues.Pending[meta.ID] + t, ok := s.state.Pending[meta.ID] if !ok { log.WithFields(s.logFields()).Warningln("TaskFailed:Pending task #%v not found.", t.Task.Meta) return nil diff --git a/go/master/service_test.go b/go/master/service_test.go index 5f91910ecc8cf32289e71e2e41e8b283acc115e6..2d00c22d6feb7177da5c19c557fd16d7925ef6d1 100644 --- a/go/master/service_test.go +++ b/go/master/service_test.go @@ -1,24 +1,30 @@ package master_test import ( + "io/ioutil" + "net/url" "os" + "strings" "testing" "time" "github.com/PaddlePaddle/Paddle/go/master" "github.com/coreos/etcd/clientv3" "github.com/coreos/etcd/embed" - "github.com/docker/docker/pkg/ioutils" "github.com/stretchr/testify/assert" ) func TestNewServiceWithEtcd(t *testing.T) { // setup an embed etcd server - etcdDir, err := ioutils.TempDir("", "") + etcdDir, err := ioutil.TempDir("", "") if err != nil { t.Fatal(err) } cfg := embed.NewConfig() + lpurl, _ := url.Parse("http://localhost:0") + lcurl, _ := url.Parse("http://localhost:0") + cfg.LPUrls = []url.URL{*lpurl} + cfg.LCUrls = []url.URL{*lcurl} cfg.Dir = etcdDir e, err := embed.StartEtcd(cfg) if err != nil { @@ -30,15 +36,13 @@ func TestNewServiceWithEtcd(t *testing.T) { t.Fatal(err) } }() - select { - case <-e.Server.ReadyNotify(): - t.Log("Server is ready!") - case <-time.After(60 * time.Second): - e.Server.Stop() // trigger a shutdown - t.Fatal("Server took too long to start!") - } - ep := []string{"127.0.0.1:2379"} + <-e.Server.ReadyNotify() + + port := strings.Split(e.Clients[0].Addr().String(), ":")[1] + endpoint := "127.0.0.1:" + port + + ep := []string{endpoint} masterAddr := "127.0.0.1:3306" store, err := master.NewEtcdClient(ep, masterAddr, master.DefaultLockPath, master.DefaultAddrPath, master.DefaultStatePath, 30) if err != nil { diff --git a/go/pserver/client/c/cclient.go b/go/pserver/client/c/cclient.go index 14ad0774550f6e5a5d8610d6007904cd2820432c..a49cd01522b8b49a74f21fcb97e9eeb1fbb2d272 100644 --- a/go/pserver/client/c/cclient.go +++ b/go/pserver/client/c/cclient.go @@ -90,8 +90,12 @@ func cArrayToSlice(p unsafe.Pointer, len int) []byte { type selector bool -func (s selector) Select() bool { - return bool(s) +func (s selector) Select() (bool, error) { + return bool(s), nil +} + +func (s selector) Done() error { + return nil } type lister []client.Server @@ -114,11 +118,10 @@ func paddle_new_pserver_client(addrs *C.char, selected int) C.paddle_pserver_cli } //export paddle_new_etcd_pserver_client -func paddle_new_etcd_pserver_client(etcdEndpoints *C.char, selected int) C.paddle_pserver_client { - // TODO(Longfei: use etcd lock to decide which trainer to initialize the parameters) +func paddle_new_etcd_pserver_client(etcdEndpoints *C.char) C.paddle_pserver_client { addr := C.GoString(etcdEndpoints) etcdClient := client.NewEtcd(addr) - c := client.NewClient(etcdClient, etcdClient.Desired(), selector(selected != 0)) + c := client.NewClient(etcdClient, etcdClient.Desired(), etcdClient) return add(c) } @@ -136,7 +139,12 @@ func paddle_pserver_client_release(client C.paddle_pserver_client) { //export paddle_begin_init_params func paddle_begin_init_params(client C.paddle_pserver_client) C.int { c := get(client) - if selected := c.BeginInitParams(); selected { + selected, err := c.BeginInitParams() + if err != nil { + panic(err) + } + + if selected { return 1 } return 0 diff --git a/go/pserver/client/c/test/test_train.py b/go/pserver/client/c/test/test_train.py index 572a61e4ccaa9ef3d03a60d916e80eab907c6d88..8d9c6b9b20f515ed0865df8cf46b6dfc2d8ffa34 100644 --- a/go/pserver/client/c/test/test_train.py +++ b/go/pserver/client/c/test/test_train.py @@ -17,12 +17,10 @@ def main(): # network config x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13)) y_predict = paddle.layer.fc(input=x, - param_attr=paddle.attr.Param( - name='w', learning_rate=1e-3), + param_attr=paddle.attr.Param(name='w'), size=1, act=paddle.activation.Linear(), - bias_attr=paddle.attr.Param( - name='b', learning_rate=1e-3)) + bias_attr=paddle.attr.Param(name='b')) y = paddle.layer.data(name='y', type=paddle.data_type.dense_vector(1)) cost = paddle.layer.mse_cost(input=y_predict, label=y) diff --git a/go/pserver/client/client.go b/go/pserver/client/client.go index 15adda4735b022c16cb22715fb690b3740e58b76..20d91e77034e1a0c6825bc401175e6dc1afec52f 100644 --- a/go/pserver/client/client.go +++ b/go/pserver/client/client.go @@ -27,9 +27,13 @@ import ( // TODO(helin): add RPC call retry logic -// Selector selects if the client should initialize parameter servers. +// Selector selects if the client should initialize parameters and +// reports the initialization process done. type Selector interface { - Select() bool + // Select selects if the client should initialize parameter servers. + Select() (bool, error) + // Done indicates the initialization process is done. + Done() error } // Server is the identification of a parameter Server. @@ -115,7 +119,7 @@ func (c *Client) monitorPservers(l Lister, pserverNum int) { // servers. Other trainers will be blocked until the initialization is // done, and they need to get the initialized parameters from // parameter servers using GetParams. -func (c *Client) BeginInitParams() bool { +func (c *Client) BeginInitParams() (bool, error) { return c.sel.Select() } diff --git a/go/pserver/client/client_test.go b/go/pserver/client/client_test.go index b630d434dca283df67f5b850b35057870fe27529..c3d88e926d7cb5f3027be26a270bee6f2db65f31 100644 --- a/go/pserver/client/client_test.go +++ b/go/pserver/client/client_test.go @@ -59,7 +59,7 @@ func initClient() [numPserver]int { go func(l net.Listener) { var cp pserver.Checkpoint - s, err := pserver.NewService(0, 1, "", nil, cp) + s, err := pserver.NewService(0, time.Hour, "", nil, cp) if err != nil { panic(err) } @@ -124,8 +124,12 @@ func initEtcdClient() { type selector bool -func (s selector) Select() bool { - return bool(s) +func (s selector) Select() (bool, error) { + return bool(s), nil +} + +func (s selector) Done() error { + return nil } type lister []client.Server @@ -135,7 +139,11 @@ func (l lister) List() []client.Server { } func testClient(t *testing.T, c *client.Client) { - selected := c.BeginInitParams() + selected, err := c.BeginInitParams() + if err != nil { + t.Fatal(err) + } + if !selected { t.Fatal("should be selected.") } diff --git a/go/pserver/client/etcd_client.go b/go/pserver/client/etcd_client.go index b6ff1fec8a6f37f61f38cb5d004b1d2c886473ed..f9071caaa8f5ac32d426b1d4344a30262202b96d 100644 --- a/go/pserver/client/etcd_client.go +++ b/go/pserver/client/etcd_client.go @@ -16,53 +16,60 @@ package client import ( "context" + "errors" + "fmt" "strconv" "strings" "time" "github.com/PaddlePaddle/Paddle/go/pserver" "github.com/coreos/etcd/clientv3" + "github.com/coreos/etcd/clientv3/concurrency" log "github.com/sirupsen/logrus" ) const ( defaultEtcdTimeout time.Duration = 5 * time.Second + + initLockPath = "/init_ps/lock" + initDonePath = "/init_ps/done" + initDoneVal = "1" ) -// EtcdClient is used by pserver client that is a part of trainer process. +// Etcd is used by pserver client that is a part of trainer process. // TODO: -// 1. add watcher to watch the change state of pservers) -// 1. add etcd lock) -type EtcdClient struct { +// 1. add watcher to watch the change state of pservers. +type Etcd struct { client *clientv3.Client timeout time.Duration endpoints []string + lock *concurrency.Mutex } // Desired read ps desired number from etcd. -func (p *EtcdClient) Desired() int { +func (e *Etcd) Desired() int { var psDesired int for { - ctx, cancel := context.WithTimeout(context.Background(), p.timeout) - resp, err := p.client.Get(ctx, pserver.PsDesired) + ctx, cancel := context.WithTimeout(context.Background(), e.timeout) + resp, err := e.client.Get(ctx, pserver.PsDesired) cancel() if err != nil { log.Errorf("Get ps dresire number failed! recnnectiong..., %v", err) - time.Sleep(p.timeout) + time.Sleep(e.timeout) continue } kvs := resp.Kvs if len(kvs) == 0 { log.Infoln("Waiting for ps desired registered ...") - time.Sleep(p.timeout) + time.Sleep(e.timeout) continue } psDesired, err = strconv.Atoi(string(resp.Kvs[0].Value)) if err != nil { log.Errorf("psDesired %d invalid %v", psDesired, err) - time.Sleep(p.timeout) + time.Sleep(e.timeout) continue } @@ -73,26 +80,26 @@ func (p *EtcdClient) Desired() int { } // List return the pserver list read from etcd. -func (p *EtcdClient) List() []Server { - psDesired := p.Desired() +func (e *Etcd) List() []Server { + psDesired := e.Desired() servers := make([]Server, psDesired) for { for i := 0; i < psDesired; i++ { - ctx, cancel := context.WithTimeout(context.Background(), p.timeout) + ctx, cancel := context.WithTimeout(context.Background(), e.timeout) psKey := pserver.PsPath + strconv.Itoa(i) log.Debugf("checking %s", psKey) - resp, err := p.client.Get(ctx, psKey) + resp, err := e.client.Get(ctx, psKey) cancel() if err != nil { log.Infof("Get psKey= %s error, %v", psKey, err) - time.Sleep(p.timeout) + time.Sleep(e.timeout) continue } kvs := resp.Kvs if len(kvs) == 0 { log.Infof("Waiting for ps addr registered ...") - time.Sleep(p.timeout) + time.Sleep(e.timeout) continue } @@ -100,10 +107,10 @@ func (p *EtcdClient) List() []Server { // TODO(Longfei) check the ps address if psAddr == "" { log.Infof("Get psKey = %s, psAddr is empty", psKey) - time.Sleep(p.timeout) + time.Sleep(e.timeout) continue } - log.Infof("got value (%s) for key: %s", psAddr, psKey) + log.Debugf("got value (%s) for key: %s", psAddr, psKey) servers[i].Index = i servers[i].Addr = psAddr } @@ -113,7 +120,7 @@ func (p *EtcdClient) List() []Server { } // NewEtcd create a etcd client to return the state of pserver on etcd. -func NewEtcd(endpoints string) *EtcdClient { +func NewEtcd(endpoints string) *Etcd { ep := strings.Split(endpoints, ",") var cli *clientv3.Client var err error @@ -130,10 +137,118 @@ func NewEtcd(endpoints string) *EtcdClient { break } log.Infof("Connected to etcd: %s\n", endpoints) - client := &EtcdClient{ + client := &Etcd{ client: cli, timeout: defaultEtcdTimeout, endpoints: ep, } return client } + +// Select indicates if the current trainer is selected to initialize +// the pserver parameters. +func (e *Etcd) Select() (bool, error) { + sess, err := concurrency.NewSession(e.client, concurrency.WithTTL(5)) + if err != nil { + return false, err + } + + lock := concurrency.NewMutex(sess, initLockPath) + log.Infof("Trying to acquire lock at %s.", initLockPath) + // Do not use timeout context here, since we don't know how + // long does it take for other trainers to initialize the + // parameters. + err = lock.Lock(context.Background()) + if err != nil { + return false, err + } + log.Infof("Successfully acquired lock at %s.", initLockPath) + + get := clientv3.OpGet(initDonePath) + ctx, cancel := context.WithTimeout(context.Background(), e.timeout) + tresp, err := e.client.Txn(ctx).If(lock.IsOwner()).Then(get).Commit() + cancel() + if err != nil { + return false, err + } + + if !tresp.Succeeded { + return false, errors.New("no longer the owner of the lock") + } + + resp := tresp.Responses[0].GetResponseRange() + + if len(resp.Kvs) == 0 { + // Key value not set, select current trainer. + e.lock = lock + log.Infoln("Trainer selected.") + return true, nil + } + + if string(resp.Kvs[0].Value) == initDoneVal { + log.Infoln("Initialization is already done.") + ctx, cancel = context.WithTimeout(context.Background(), e.timeout) + err = lock.Unlock(ctx) + cancel() + if err != nil { + log.Errorln(err) + } + return false, nil + } + + return false, fmt.Errorf("key %s have unexpected value: %v", initDonePath, resp.Kvs[0].Value) +} + +// Done indicates the parameter initialization process is done. +func (e *Etcd) Done() error { + if e.lock == nil { + return errors.New("lock is nil, Done called unexpectedly") + } + + put := clientv3.OpPut(initDonePath, initDoneVal) + ctx, cancel := context.WithTimeout(context.Background(), e.timeout) + tresp, err := e.client.Txn(ctx).If(e.lock.IsOwner()).Then(put).Commit() + cancel() + if err != nil { + return err + } + + if !tresp.Succeeded { + return errors.New("no longer the owner of the lock") + } + + ctx, cancel = context.WithTimeout(context.Background(), e.timeout) + err = e.lock.Unlock(ctx) + cancel() + if err != nil { + log.Errorln(err) + } else { + e.lock = nil + } + + return nil +} + +// Close closes the etcd client. +func (e *Etcd) Close() error { + var err error + if e.lock != nil { + ctx, cancel := context.WithTimeout(context.Background(), e.timeout) + err = e.lock.Unlock(ctx) + cancel() + if err == nil { + e.lock = nil + } + } + + cErr := e.client.Close() + if cErr != nil { + if err != nil { + log.Errorln(cErr) + return err + } + return cErr + } + + return err +} diff --git a/go/pserver/client/etcd_client_test.go b/go/pserver/client/etcd_client_test.go new file mode 100644 index 0000000000000000000000000000000000000000..08742433e7a266fbd39e34f4b92ac4cc4caeb0fb --- /dev/null +++ b/go/pserver/client/etcd_client_test.go @@ -0,0 +1,106 @@ +package client_test + +import ( + "io/ioutil" + "net/url" + "os" + "strings" + "sync" + "testing" + + "github.com/PaddlePaddle/Paddle/go/pserver/client" + "github.com/coreos/etcd/embed" +) + +func TestSelector(t *testing.T) { + etcdDir, err := ioutil.TempDir("", "") + if err != nil { + t.Fatal(err) + } + cfg := embed.NewConfig() + lpurl, _ := url.Parse("http://localhost:0") + lcurl, _ := url.Parse("http://localhost:0") + cfg.LPUrls = []url.URL{*lpurl} + cfg.LCUrls = []url.URL{*lcurl} + cfg.Dir = etcdDir + e, err := embed.StartEtcd(cfg) + if err != nil { + t.Fatal(err) + } + + defer func() { + e.Close() + if err := os.RemoveAll(etcdDir); err != nil { + t.Fatal(err) + } + }() + + <-e.Server.ReadyNotify() + + port := strings.Split(e.Clients[0].Addr().String(), ":")[1] + endpoint := "127.0.0.1:" + port + + var mu sync.Mutex + selectedCount := 0 + var wg sync.WaitGroup + selectAndDone := func(c *client.Etcd) { + defer wg.Done() + + selected, err := c.Select() + if err != nil { + panic(err) + } + + if selected { + mu.Lock() + selectedCount++ + mu.Unlock() + err = c.Done() + if err != nil { + t.Fatal(err) + } + } + } + + c0 := client.NewEtcd(endpoint) + c1 := client.NewEtcd(endpoint) + c2 := client.NewEtcd(endpoint) + c3 := client.NewEtcd(endpoint) + wg.Add(3) + go selectAndDone(c0) + go selectAndDone(c1) + go selectAndDone(c2) + wg.Wait() + + // simulate trainer crashed and restarted after the + // initialization process. + wg.Add(1) + go selectAndDone(c3) + wg.Wait() + + mu.Lock() + if selectedCount != 1 { + t.Fatal("selected count wrong:", selectedCount) + } + mu.Unlock() + + err = c0.Close() + if err != nil { + t.Fatal(err) + } + + err = c1.Close() + if err != nil { + t.Fatal(err) + } + + err = c2.Close() + if err != nil { + t.Fatal(err) + } + + err = c3.Close() + if err != nil { + t.Fatal(err) + } +} diff --git a/go/pserver/etcd_client.go b/go/pserver/etcd_client.go index 4fb26307667295ab825d07be6c3d1d4b33f6eb8b..41f0640fc09a3265c0e11c06255c7ee834983203 100644 --- a/go/pserver/etcd_client.go +++ b/go/pserver/etcd_client.go @@ -206,6 +206,7 @@ func (e *EtcdClient) GetKey(key string, timeout time.Duration) ([]byte, error) { if err != nil { return []byte{}, err } + kvs := resp.Kvs if len(kvs) == 0 { return []byte{}, nil @@ -215,9 +216,14 @@ func (e *EtcdClient) GetKey(key string, timeout time.Duration) ([]byte, error) { } // PutKey put into etcd with value by key specified -func (e *EtcdClient) PutKey(key string, value []byte, timeout time.Duration) error { +func (e *EtcdClient) PutKey(key string, value []byte, timeout time.Duration, withLease bool) error { ctx, cancel := context.WithTimeout(context.Background(), timeout) - _, err := e.client.Put(ctx, key, string(value), clientv3.WithLease(e.sess.Lease())) + var err error + if withLease { + _, err = e.client.Put(ctx, key, string(value), clientv3.WithLease(e.sess.Lease())) + } else { + _, err = e.client.Put(ctx, key, string(value)) + } cancel() return err } diff --git a/go/pserver/optimizer.go b/go/pserver/optimizer.go index 709160d45d98b6cf6d60f52ceb3fb33e0a0bd17d..ae7359073494bd9cb6b70b12af4daca064179556 100644 --- a/go/pserver/optimizer.go +++ b/go/pserver/optimizer.go @@ -32,6 +32,7 @@ type optimizer struct { opt *C.struct_paddle_optimizer elementType ElementType contentLen int + config []byte } func cArrayToSlice(p unsafe.Pointer, len int) []byte { @@ -70,6 +71,7 @@ func newOptimizer(paramWithConfigs ParameterWithConfig, State []byte) *optimizer cstate = unsafe.Pointer(&s[0]) } + o.config = c o.opt = C.paddle_create_optimizer((*C.uchar)(&c[0]), C.int(len(c)), C.paddle_element_type(p.ElementType), cbuffer, C.int(paramBufferSize), (*C.char)(cstate), C.int(len(s))) return o diff --git a/go/pserver/service.go b/go/pserver/service.go index 7d297c46d03bf78d18ca9830a318968397119d3e..25751540a9a2dff043c14e0912bfab1aaa938ab4 100644 --- a/go/pserver/service.go +++ b/go/pserver/service.go @@ -25,11 +25,13 @@ import ( "fmt" "io/ioutil" "os" - "path/filepath" + "path" "strconv" "sync" "time" + uuid "github.com/satori/go.uuid" + log "github.com/sirupsen/logrus" ) @@ -42,9 +44,9 @@ var ErrCheckpointNotFound = errors.New("checkpoint not found") // RPC error message. const ( - AlreadyInitialized = "pserver already initialized" - Uninitialized = "pserver not fully initialized" - CheckpointMD5Failed = "checkpoint file MD5 validation failed" + AlreadyInitialized = "pserver already initialized" + Uninitialized = "pserver not fully initialized" + WrongChecksum = "checkpoint file checksum validation failed" ) // Supported element types. @@ -73,11 +75,12 @@ type ParameterWithConfig struct { // checkpointMeta saves checkpoint metadata type checkpointMeta struct { UUID string `json:"uuid"` + Path string `json:"path"` MD5 string `json:"md5"` Timestamp int64 `json:"timestamp"` } -// Checkpoint is the pserver shard persist in file +// Checkpoint is the pserver shard persist in file. type Checkpoint []parameterCheckpoint // Gradient is the gradient of the parameter. @@ -90,50 +93,58 @@ type Service struct { checkpointInterval time.Duration checkpointPath string client *EtcdClient - mu sync.Mutex - optMap map[string]*optimizer + + mu sync.Mutex + optMap map[string]*optimizer } -// parameterCheckpoint saves parameter checkpoint +// parameterCheckpoint saves parameter checkpoint. type parameterCheckpoint struct { ParameterWithConfig State []byte } -// NewCheckpointFromFile loads parameters and state from checkpoint file -func NewCheckpointFromFile(cpPath string, idx int, e *EtcdClient) (Checkpoint, error) { - v, err := e.GetKey(PsPath+string(idx), 3*time.Second) +func loadMeta(e *EtcdClient, idx int) (meta checkpointMeta, err error) { + v, err := e.GetKey(PsCheckpoint+strconv.Itoa(idx), 3*time.Second) if err != nil { - return nil, err + return } if len(v) == 0 { - return nil, ErrCheckpointNotFound + err = ErrCheckpointNotFound + return } - var cpMeta checkpointMeta - if err = json.Unmarshal(v, &cpMeta); err != nil { - return nil, err + if err = json.Unmarshal(v, &meta); err != nil { + return } - fn := filepath.Join(cpPath, cpMeta.UUID) - if _, err = os.Stat(fn); os.IsNotExist(err) { + return +} + +// LoadCheckpoint loads checkpoint from file. +func LoadCheckpoint(e *EtcdClient, idx int) (Checkpoint, error) { + cpMeta, err := loadMeta(e, idx) + if err != nil { return nil, err } - content, err := ioutil.ReadFile(fn) + + content, err := ioutil.ReadFile(cpMeta.Path) if err != nil { return nil, err } + // TODO(helin): change MD5 to CRC since CRC is better for file + // checksum in our use case (emphasize speed over security). h := md5.New() md5 := hex.EncodeToString(h.Sum(content)) if md5 != cpMeta.MD5 { - return nil, errors.New(CheckpointMD5Failed) + return nil, errors.New(WrongChecksum) } dec := gob.NewDecoder(bytes.NewReader(content)) - cp := Checkpoint{} - if err = dec.Decode(cp); err != nil { + var cp Checkpoint + if err = dec.Decode(&cp); err != nil { return nil, err } return cp, nil @@ -193,6 +204,15 @@ func (s *Service) FinishInitParams(_ int, _ *int) error { } close(s.initialized) + go func() { + t := time.Tick(s.checkpointInterval) + for range t { + err := s.checkpoint() + if err != nil { + log.Errorln(err) + } + } + }() return nil } @@ -240,23 +260,36 @@ func (s *Service) GetParam(name string, parameter *Parameter) error { return nil } -// pserver save checkpoint -func (s *Service) doCheckpoint() (err error) { - <-s.initialized - s.mu.Lock() - defer s.mu.Unlock() +func traceTime(start time.Time, name string) { + elapsed := time.Since(start) + log.Infof("%s took %v", name, elapsed) +} + +// checkpoint saves checkpoint to disk. +// +// checkpoint should be only called after the parameters are +// initialized. +func (s *Service) checkpoint() (err error) { + log.Infoln("Begin save checkpoint.") + defer traceTime(time.Now(), "save checkpoint") + s.mu.Lock() cp := make([]parameterCheckpoint, len(s.optMap)) index := 0 + // TODO(helin): write checkpoint incrementally to reduce memory + // footprint during checkpoint. for name, opt := range s.optMap { var pc parameterCheckpoint pc.Param.Name = name pc.Param.ElementType = opt.elementType pc.Param.Content = opt.GetWeights() + pc.Config = opt.config pc.State = opt.GetStates() cp[index] = pc index++ } + s.mu.Unlock() + var buf bytes.Buffer encoder := gob.NewEncoder(&buf) err = encoder.Encode(cp) @@ -264,32 +297,9 @@ func (s *Service) doCheckpoint() (err error) { return } - cpMeta := checkpointMeta{} - cpMeta.UUID = s.checkpointPath + strconv.Itoa(s.idx) - cpMeta.Timestamp = time.Now().UnixNano() - h := md5.New() - cpMeta.MD5 = hex.EncodeToString(h.Sum(buf.Bytes())) - - cpMetajson, err := json.Marshal(cpMeta) - if err != nil { - return - } - - err = s.client.PutKey(filepath.Join(PsCheckpoint, strconv.Itoa(s.idx)), cpMetajson, 3*time.Second) - if err != nil { - return - } - if _, err = os.Stat(cpMeta.UUID); os.IsNotExist(err) { - log.Info("checkpoint does not exists.") - } else { - err = os.Remove(cpMeta.UUID) - if err != nil { - log.Infof("Removing checkpoint %s failed", cpMeta.UUID) - } else { - log.Infof("checkpoint %s already exsits, removing ", cpMeta.UUID) - } - } - f, err := os.Create(cpMeta.UUID) + id := uuid.NewV4().String() + p := path.Join(s.checkpointPath, id) + f, err := os.Create(p) if err != nil { return } @@ -317,5 +327,43 @@ func (s *Service) doCheckpoint() (err error) { return } + oldMeta, err := loadMeta(s.client, s.idx) + if err == ErrCheckpointNotFound { + log.Infoln("Do not have existing checkpoint.") + err = nil + } + + if err != nil { + return + } + + h := md5.New() + md5 := hex.EncodeToString(h.Sum(buf.Bytes())) + cpMeta := checkpointMeta{ + UUID: id, + Timestamp: time.Now().UnixNano(), + MD5: md5, + Path: p, + } + + json, err := json.Marshal(cpMeta) + if err != nil { + return + } + + err = s.client.PutKey(PsCheckpoint+strconv.Itoa(s.idx), json, 3*time.Second, false) + if err != nil { + return + } + + if oldMeta.Path != "" { + rmErr := os.Remove(oldMeta.Path) + if rmErr != nil { + // log error, but still treat checkpoint as + // successful. + log.Errorln(rmErr) + } + } + return } diff --git a/go/pserver/service_test.go b/go/pserver/service_test.go index 988f3b5acb82a95aeb54af2b8b0e4d39a458291a..be648cd1e83e4f7790edac5842db432fb4870072 100644 --- a/go/pserver/service_test.go +++ b/go/pserver/service_test.go @@ -30,7 +30,7 @@ const ( func TestServiceFull(t *testing.T) { var cp pserver.Checkpoint - s, err := pserver.NewService(0, 1, "", nil, cp) + s, err := pserver.NewService(0, time.Hour, "", nil, cp) if err != nil { t.Error(err) } @@ -102,7 +102,7 @@ func TestServiceFull(t *testing.T) { func TestMultipleInit(t *testing.T) { var cp pserver.Checkpoint - s, err := pserver.NewService(0, 1, "", nil, cp) + s, err := pserver.NewService(0, time.Hour, "", nil, cp) if err != nil { t.Fatal(err) } @@ -119,7 +119,7 @@ func TestMultipleInit(t *testing.T) { func TestUninitialized(t *testing.T) { var cp pserver.Checkpoint - s, err := pserver.NewService(0, 1, "", nil, cp) + s, err := pserver.NewService(0, time.Hour, "", nil, cp) err = s.SendGrad(pserver.Gradient{}, nil) if err.Error() != pserver.Uninitialized { t.Fatal(err) @@ -128,7 +128,7 @@ func TestUninitialized(t *testing.T) { func TestBlockUntilInitialized(t *testing.T) { var cp pserver.Checkpoint - s, err := pserver.NewService(0, 1, "", nil, cp) + s, err := pserver.NewService(0, time.Hour, "", nil, cp) if err != nil { t.Error(err) } diff --git a/paddle/.set_python_path.sh b/paddle/.set_python_path.sh index fa7baccc86e0b56e57d52a40c95cfe1b98fececc..8fd58925ee4820269572176ff9496f42914652da 100755 --- a/paddle/.set_python_path.sh +++ b/paddle/.set_python_path.sh @@ -21,22 +21,15 @@ # # It same as PYTHONPATH=${YOUR_PYTHON_PATH}:$PYTHONPATH {exec...} # - -if ! python -c "import paddle" >/dev/null 2>/dev/null; then - PYPATH="" - set -x - while getopts "d:" opt; do - case $opt in - d) - PYPATH=$OPTARG - ;; - esac - done - shift $(($OPTIND - 1)) - export PYTHONPATH=$PYPATH:$PYTHONPATH - $@ -else - echo "paddle package is already in your PYTHONPATH. But unittest need a clean environment." - echo "Please uninstall paddle package before start unittest. Try to 'pip uninstall paddle'" - exit 1 -fi +PYPATH="" +set -x +while getopts "d:" opt; do + case $opt in + d) + PYPATH=$OPTARG + ;; + esac +done +shift $(($OPTIND - 1)) +export PYTHONPATH=$PYPATH:$PYTHONPATH +$@ diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index f8a88cf317aee6c5dd25e4cc25d588c6c50fcbce..cf61a243e9df2fd4a580e41f07cb0a22dcc72083 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -22,7 +22,5 @@ if(WITH_C_API) endif() if(WITH_SWIG_PY) - configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in - ${CMAKE_CURRENT_SOURCE_DIR}/setup.py) add_subdirectory(api) endif() diff --git a/paddle/api/CMakeLists.txt b/paddle/api/CMakeLists.txt index 84da89a1422b6095b995744cebb6a3af98a071c6..d7b3d2bdec1687425df804c0d56d568241f9e8b0 100644 --- a/paddle/api/CMakeLists.txt +++ b/paddle/api/CMakeLists.txt @@ -19,9 +19,9 @@ add_library(paddle_api STATIC ${API_SOURCES}) add_dependencies(paddle_api paddle_proto paddle_trainer_lib) INCLUDE(${SWIG_USE_FILE}) -INCLUDE_DIRECTORIES(${PROJ_ROOT}/paddle) +INCLUDE_DIRECTORIES(${PADDLE_SOURCE_DIR}/paddle) -FILE(GLOB PY_PADDLE_PYTHON_FILES ${PROJ_ROOT}/paddle/py_paddle/*.py) +FILE(GLOB PY_PADDLE_PYTHON_FILES ${PADDLE_SOURCE_DIR}/paddle/py_paddle/*.py) SET_SOURCE_FILES_PROPERTIES(Paddle.i PROPERTIES CPLUSPLUS ON) @@ -79,22 +79,16 @@ SWIG_LINK_LIBRARIES(swig_paddle ${START_END} ) -add_custom_command(OUTPUT ${PROJ_ROOT}/paddle/py_paddle/_swig_paddle.so - COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/swig_paddle.py ${PROJ_ROOT}/paddle/py_paddle - COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/_swig_paddle.so ${PROJ_ROOT}/paddle/py_paddle - COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel - COMMAND ${CMAKE_COMMAND} -E touch dist/.timestamp - COMMAND rm -rf py_paddle.egg-info build - WORKING_DIRECTORY ${PROJ_ROOT}/paddle +add_custom_command(OUTPUT ${PADDLE_SOURCE_DIR}/paddle/py_paddle/_swig_paddle.so + COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/swig_paddle.py ${PADDLE_SOURCE_DIR}/paddle/py_paddle + COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/_swig_paddle.so ${PADDLE_SOURCE_DIR}/paddle/py_paddle + COMMAND ${CMAKE_COMMAND} -E touch .timestamp + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle DEPENDS _swig_paddle ) # TODO(yuyang18) : make wheel name calculated by cmake -add_custom_target(python_api_wheel ALL DEPENDS ${PROJ_ROOT}/paddle/py_paddle/_swig_paddle.so) - -install(DIRECTORY ${CMAKE_SOURCE_DIR}/paddle/dist/ - DESTINATION opt/paddle/share/wheels -) +add_custom_target(python_api_wheel ALL DEPENDS ${PADDLE_SOURCE_DIR}/paddle/py_paddle/_swig_paddle.so) if(WITH_TESTING) IF(NOT PY_PIP_FOUND) @@ -108,7 +102,7 @@ if(WITH_TESTING) BUILD_COMMAND "" INSTALL_COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py install BUILD_IN_SOURCE 1 - DEPENDS python setuptools python_api_wheel + #DEPENDS python setuptools python_api_wheel ) ENDIF() add_subdirectory(test) diff --git a/paddle/api/ParameterUpdater.cpp b/paddle/api/ParameterUpdater.cpp index 5934cb898b5f6adc74c237b1733a7459d8437a28..8cd73b348c507386cd88e907f7b431ef25e793aa 100644 --- a/paddle/api/ParameterUpdater.cpp +++ b/paddle/api/ParameterUpdater.cpp @@ -41,7 +41,7 @@ ParameterUpdater *ParameterUpdater::createNewRemoteUpdater( config->m->getConfig(), pserverSpec, useEtcd)); return updater; #else - throw UnsupportError(); + throw UnsupportError("not compiled with WITH_GOLANG"); #endif } diff --git a/paddle/api/test/CMakeLists.txt b/paddle/api/test/CMakeLists.txt index f3b1c2c4d438b5d3e776ef27ce8f8b78f710f2ab..761aeb5b174105edece8880a9f5012c13a63fd11 100644 --- a/paddle/api/test/CMakeLists.txt +++ b/paddle/api/test/CMakeLists.txt @@ -1,2 +1,6 @@ -add_python_test(test_swig_api - testArguments.py testGradientMachine.py testMatrix.py testVector.py testTrain.py testTrainer.py) +py_test(testTrain SRCS testTrain.py) +py_test(testMatrix SRCS testMatrix.py) +py_test(testVector SRCS testVector.py) +py_test(testTrainer SRCS testTrainer.py) +py_test(testArguments SRCS testArguments.py) +py_test(testGradientMachine SRCS testGradientMachine.py) diff --git a/paddle/capi/Arguments.cpp b/paddle/capi/Arguments.cpp index 8b81ec69e60399af86f055d2258276ac06e0b13a..1ec403077e7ea0bc8299e6266167b50ed81c3b08 100644 --- a/paddle/capi/Arguments.cpp +++ b/paddle/capi/Arguments.cpp @@ -90,6 +90,18 @@ paddle_error paddle_arguments_set_ids(paddle_arguments args, return kPD_NO_ERROR; } +paddle_error paddle_arguments_set_frame_shape(paddle_arguments args, + uint64_t ID, + uint64_t frameHeight, + uint64_t frameWidth) { + if (args == nullptr) return kPD_NULLPTR; + auto a = castArg(args); + if (ID >= a->args.size()) return kPD_OUT_OF_RANGE; + a->args[ID].setFrameHeight(frameHeight); + a->args[ID].setFrameWidth(frameWidth); + return kPD_NO_ERROR; +} + paddle_error paddle_arguments_set_sequence_start_pos(paddle_arguments args, uint64_t ID, uint32_t nestedLevel, diff --git a/paddle/capi/arguments.h b/paddle/capi/arguments.h index d71ea26a5d1aff130d974541532fda3b09bf6fe5..7c32524a00b11573a037968cf10d6488ea6c5644 100644 --- a/paddle/capi/arguments.h +++ b/paddle/capi/arguments.h @@ -111,6 +111,20 @@ PD_API paddle_error paddle_arguments_set_ids(paddle_arguments args, uint64_t ID, paddle_ivector ids); +/** + * @brief paddle_arguments_set_frame_shape Set the fram size of one argument + * in array, which index is `ID`. + * @param [in] args arguments array + * @param [in] ID array index + * @param [in] frameHeight maximum height of input images + * @param [in] frameWidth maximum width of input images + * @return paddle_error + */ +PD_API paddle_error paddle_arguments_set_frame_shape(paddle_arguments args, + uint64_t ID, + uint64_t frameHeight, + uint64_t frameWidth); + /** * @brief PDArgsSetSequenceStartPos Set sequence start position vector of one * argument in array, which index is `ID`. diff --git a/paddle/capi/examples/model_inference/common/common.h b/paddle/capi/examples/model_inference/common/common.h index a78522e4a7c3cb34b341b7f4c89b53d32b72f114..e32f2f9836f63ba10ef5be447a4c41514e079219 100644 --- a/paddle/capi/examples/model_inference/common/common.h +++ b/paddle/capi/examples/model_inference/common/common.h @@ -3,18 +3,21 @@ #include #include -#define CHECK(stmt) \ - do { \ - paddle_error __err__ = stmt; \ - if (__err__ != kPD_NO_ERROR) { \ - fprintf(stderr, "Invoke paddle error %d \n" #stmt, __err__); \ - exit(__err__); \ - } \ +#define CHECK(stmt) \ + do { \ + paddle_error __err__ = stmt; \ + if (__err__ != kPD_NO_ERROR) { \ + fprintf(stderr, "Invoke paddle error %d in " #stmt "\n", __err__); \ + exit(__err__); \ + } \ } while (0) void* read_config(const char* filename, long* size) { FILE* file = fopen(filename, "r"); - if (file == NULL) return NULL; + if (file == NULL) { + fprintf(stderr, "Open %s error\n", filename); + return NULL; + } fseek(file, 0L, SEEK_END); *size = ftell(file); fseek(file, 0L, SEEK_SET); diff --git a/paddle/capi/gradient_machine.cpp b/paddle/capi/gradient_machine.cpp index 00f76e0152366834eafc22df710cf3d6c7b8471f..b3287552db87d25edbf6e7f3d5e68121df49e9d6 100644 --- a/paddle/capi/gradient_machine.cpp +++ b/paddle/capi/gradient_machine.cpp @@ -54,6 +54,31 @@ paddle_error paddle_gradient_machine_create_for_inference( return kPD_NO_ERROR; } +paddle_error paddle_gradient_machine_create_for_inference_with_parameters( + paddle_gradient_machine* machine, void* mergedModel, uint64_t size) { + if (mergedModel == nullptr) return kPD_NULLPTR; + std::istringstream is(std::string(static_cast(mergedModel), size)); + int64_t modelConfigSize = 0; + is.read((char*)(&modelConfigSize), sizeof(modelConfigSize)); + std::string modelConfigProtobuf; + modelConfigProtobuf.resize(modelConfigSize); + is.read(&modelConfigProtobuf[0], modelConfigSize); + paddle::TrainerConfig config; + if (!config.ParseFromString(modelConfigProtobuf) || !config.IsInitialized()) { + return kPD_PROTOBUF_ERROR; + } + auto ptr = new paddle::capi::CGradientMachine(); + ptr->machine.reset(paddle::GradientMachine::create( + config.model_config(), CREATE_MODE_TESTING, {paddle::PARAMETER_VALUE})); + std::vector& parameters = ptr->machine->getParameters(); + for (auto& para : parameters) { + para->load(is); + } + + *machine = ptr; + return kPD_NO_ERROR; +} + paddle_error paddle_gradient_machine_destroy(paddle_gradient_machine machine) { delete cast(machine); return kPD_NO_ERROR; diff --git a/paddle/capi/gradient_machine.h b/paddle/capi/gradient_machine.h index d7e2dd9bf8037ed474971624d4518160604abe4d..c613ade5b24efbbf52f21c7ee86dd3189981c5ef 100644 --- a/paddle/capi/gradient_machine.h +++ b/paddle/capi/gradient_machine.h @@ -36,6 +36,18 @@ typedef void* paddle_gradient_machine; PD_API paddle_error paddle_gradient_machine_create_for_inference( paddle_gradient_machine* machine, void* modelConfigProtobuf, int size); +/** + * @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. + * @param [in] mergedModel + * @param [in] size + * @return paddle_error + */ +PD_API paddle_error +paddle_gradient_machine_create_for_inference_with_parameters( + paddle_gradient_machine* machine, void* mergedModel, uint64_t size); + /** * @brief Load parameter from disk. * @param machine Gradient Machine. diff --git a/paddle/capi/tests/CMakeLists.txt b/paddle/capi/tests/CMakeLists.txt index d73f6b7733950bd472a46afb21694aac943fc909..8208808b94f54f2ddaf4d426a65b8db562b36aca 100644 --- a/paddle/capi/tests/CMakeLists.txt +++ b/paddle/capi/tests/CMakeLists.txt @@ -10,5 +10,5 @@ target_include_directories(capi_test_gradientMachine PUBLIC ${PADDLE_CAPI_INC_PATH}) target_link_libraries(capi_test_gradientMachine paddle_capi) add_test(NAME capi_test_gradientMachine - COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine - WORKING_DIRECTORY ${PROJ_ROOT}/paddle/capi/tests) + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/capi/tests) diff --git a/paddle/cuda/CMakeLists.txt b/paddle/cuda/CMakeLists.txt index 73ffa690d9d91b673079fc0ecf91f17cbabfdb1e..0865b02c4f275f3d5069109917b05dff1393fc1e 100755 --- a/paddle/cuda/CMakeLists.txt +++ b/paddle/cuda/CMakeLists.txt @@ -39,6 +39,7 @@ set(CUDA_CU_SOURCES src/hl_cuda_lstm.cu src/hl_top_k.cu src/hl_batch_transpose.cu + src/hl_batch_norm.cu src/hl_cuda_sequence.cu src/hl_table_apply.cu) diff --git a/paddle/cuda/include/hl_batch_norm.h b/paddle/cuda/include/hl_batch_norm.h new file mode 100644 index 0000000000000000000000000000000000000000..afc5e0b2deacc4aadf98b3f7ce115e534bbc5124 --- /dev/null +++ b/paddle/cuda/include/hl_batch_norm.h @@ -0,0 +1,48 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#ifndef HL_BATCH_NORM_H_ +#define HL_BATCH_NORM_H_ + +#include "hl_base.h" + +/** + * @brief batch norm inferece. + * + * @param[in] input input data. + * @param[out] output output data. + * @param[in] scale batch normalization scale parameter (in original + * paper scale is referred to as gamma). + * @param[in] bias batch normalization bias parameter (in original + * paper scale is referred to as beta). + * @param[in] estimatedMean + * @param[in] estimatedVar The moving mean and variance + * accumulated during the training phase are passed + * as inputs here. + * @param[in] epsilon Epsilon value used in the batch + * normalization formula. + */ +extern void hl_batch_norm_cuda_inference(const real* input, + real* output, + const real* scale, + const real* bias, + const real* estimatedMean, + const real* estimatedVar, + const double epsilon, + size_t batchSize, + size_t channel, + size_t height, + size_t width); + +#endif // HL_BATCH_NORM_H_ diff --git a/paddle/cuda/src/hl_batch_norm.cu b/paddle/cuda/src/hl_batch_norm.cu new file mode 100644 index 0000000000000000000000000000000000000000..5828ecb8e049c2f0573ab8547164794bef6db1ca --- /dev/null +++ b/paddle/cuda/src/hl_batch_norm.cu @@ -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. */ + +#include "hl_batch_norm.h" + +__global__ void batchNormInference(real* output, + const real* input, + const real* scale, + const real* bias, + const real* estimatedMean, + const real* estimatedVar, + const double epsilon, + size_t batchSize, + size_t channel, + size_t height, + size_t width) { + const int tid = threadIdx.x; + const int num = channel * height * width; + const int batch = blockIdx.x; + for (int i = tid; i < num; i += blockDim.x) { + const int c = i / (height * width); + const int id = batch * num + i; + real val = input[id] - estimatedMean[c]; + val /= sqrt(estimatedVar[c] + epsilon); + val *= scale[c]; + val += bias[c]; + output[id] = val; + } +} + +void hl_batch_norm_cuda_inference(const real* input, + real* output, + const real* scale, + const real* bias, + const real* estimatedMean, + const real* estimatedVar, + const double epsilon, + size_t batchSize, + size_t channel, + size_t height, + size_t width) { + batchNormInference<<>>(output, + input, + scale, + bias, + estimatedMean, + estimatedVar, + epsilon, + batchSize, + channel, + height, + width); + + CHECK_SYNC("hl_batch_norm_cuda_inference failed!"); +} diff --git a/paddle/cuda/src/hl_batch_transpose.cu b/paddle/cuda/src/hl_batch_transpose.cu index f047403da17e66960f029f2fee7312210009c952..f4c253df7b4be937f041f18587efd4c9d693fbe4 100644 --- a/paddle/cuda/src/hl_batch_transpose.cu +++ b/paddle/cuda/src/hl_batch_transpose.cu @@ -12,17 +12,15 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "hl_batch_transpose.h" #include "hl_base.h" +#include "hl_batch_transpose.h" const int TILE_DIM = 64; const int BLOCK_ROWS = 16; // No bank-conflict transpose for a batch of data. -__global__ void batchTransposeNoBankConflicts(real* odata, - const real* idata, - int numSamples, int width, - int height) { +__global__ void batchTransposeNoBankConflicts( + real* odata, const real* idata, int numSamples, int width, int height) { __shared__ float tile[TILE_DIM][TILE_DIM + 1]; const int x = blockIdx.x * TILE_DIM + threadIdx.x; @@ -50,12 +48,12 @@ __global__ void batchTransposeNoBankConflicts(real* odata, newX] = tile[threadIdx.x][j]; } -void batchTranspose(const real* input, real* output, int width, int height, - int batchSize) { +void batchTranspose( + const real* input, real* output, int width, int height, int batchSize) { dim3 dimBlock(TILE_DIM, BLOCK_ROWS, 1); dim3 dimGrid(DIVUP(width, TILE_DIM), DIVUP(height, TILE_DIM), batchSize); - batchTransposeNoBankConflicts<<>> - (output, input, batchSize, width, height); + batchTransposeNoBankConflicts<<>>( + output, input, batchSize, width, height); CHECK_SYNC("batchTranspose failed!"); } diff --git a/paddle/cuda/src/hl_cuda_aggregate.cu b/paddle/cuda/src/hl_cuda_aggregate.cu index 97034a917708487d1c5dc59e6ebbf45bad1c3227..16a54ad343fa140aa1f3bec311c4b712d0086082 100644 --- a/paddle/cuda/src/hl_cuda_aggregate.cu +++ b/paddle/cuda/src/hl_cuda_aggregate.cu @@ -12,27 +12,23 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ - +#include "hl_aggregate.h" #include "hl_base.h" #include "hl_cuda.h" #include "hl_cuda.ph" -#include "hl_aggregate.h" -#include "hl_thread.ph" #include "hl_matrix_base.cuh" +#include "hl_thread.ph" #include "paddle/utils/Logging.h" /** * @brief matrix row operator. */ -template -__global__ void KeMatrixRowOp(Agg agg, - real *E, - real *Sum, - int dimN) { +template +__global__ void KeMatrixRowOp(Agg agg, real *E, real *Sum, int dimN) { __shared__ real sum_s[blockSize]; - int cnt = (dimN + blockSize -1) / blockSize; - int rowId = blockIdx.x + blockIdx.y*gridDim.x; - int index = rowId*dimN; + int cnt = (dimN + blockSize - 1) / blockSize; + int rowId = blockIdx.x + blockIdx.y * gridDim.x; + int index = rowId * dimN; int tid = threadIdx.x; int lmt = tid; @@ -44,7 +40,7 @@ __global__ void KeMatrixRowOp(Agg agg, sum_s[tid] = tmp; __syncthreads(); - for (int stride = blockSize/2; stride > 0; stride = stride/2) { + for (int stride = blockSize / 2; stride > 0; stride = stride / 2) { if (tid < stride) { sum_s[tid] = agg(sum_s[tid], sum_s[tid + stride]); } @@ -58,29 +54,21 @@ __global__ void KeMatrixRowOp(Agg agg, } template -void hl_matrix_row_op(Agg agg, - real *A_d, - real *C_d, - int dimM, - int dimN) { +void hl_matrix_row_op(Agg agg, real *A_d, real *C_d, int dimM, int dimN) { int blocksX = dimM; int blocksY = 1; dim3 threads(128, 1); dim3 grid(blocksX, blocksY); - KeMatrixRowOp<<< grid, threads, 0, STREAM_DEFAULT >>> - (agg, A_d, C_d, dimN); + KeMatrixRowOp<<>>( + agg, A_d, C_d, dimN); } void hl_matrix_row_sum(real *A_d, real *C_d, int dimM, int dimN) { CHECK_NOTNULL(A_d); CHECK_NOTNULL(C_d); - hl_matrix_row_op(aggregate::sum(), - A_d, - C_d, - dimM, - dimN); + hl_matrix_row_op(aggregate::sum(), A_d, C_d, dimM, dimN); CHECK_SYNC("hl_matrix_row_sum failed"); } @@ -88,11 +76,7 @@ void hl_matrix_row_max(real *A_d, real *C_d, int dimM, int dimN) { CHECK_NOTNULL(A_d); CHECK_NOTNULL(C_d); - hl_matrix_row_op(aggregate::max(), - A_d, - C_d, - dimM, - dimN); + hl_matrix_row_op(aggregate::max(), A_d, C_d, dimM, dimN); CHECK_SYNC("hl_matrix_row_max failed"); } @@ -100,23 +84,16 @@ void hl_matrix_row_min(real *A_d, real *C_d, int dimM, int dimN) { CHECK_NOTNULL(A_d); CHECK_NOTNULL(C_d); - hl_matrix_row_op(aggregate::min(), - A_d, - C_d, - dimM, - dimN); + hl_matrix_row_op(aggregate::min(), A_d, C_d, dimM, dimN); CHECK_SYNC("hl_matrix_row_min failed"); } /** * @brief matrix column operator. */ -template -__global__ void KeMatrixColumnOp(Agg agg, - real *E, - real *Sum, - int dimM, - int dimN) { +template +__global__ void KeMatrixColumnOp( + Agg agg, real *E, real *Sum, int dimM, int dimN) { int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; real tmp = agg.init(); if (rowIdx < dimN) { @@ -127,15 +104,12 @@ __global__ void KeMatrixColumnOp(Agg agg, } } -template -__global__ void KeMatrixColumnOp_S(Agg agg, - real *E, - real *Sum, - int dimM, - int dimN) { - __shared__ real _sum[blockDimX*blockDimY]; - int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; - int index = threadIdx.y; +template +__global__ void KeMatrixColumnOp_S( + Agg agg, real *E, real *Sum, int dimM, int dimN) { + __shared__ real _sum[blockDimX * blockDimY]; + int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; + int index = threadIdx.y; real tmp = agg.init(); if (rowIdx < dimN) { @@ -144,14 +118,14 @@ __global__ void KeMatrixColumnOp_S(Agg agg, index += blockDimY; } } - _sum[threadIdx.x + threadIdx.y*blockDimX] = tmp; + _sum[threadIdx.x + threadIdx.y * blockDimX] = tmp; __syncthreads(); if (rowIdx < dimN) { - if (threadIdx.y ==0) { + if (threadIdx.y == 0) { real tmp = agg.init(); - for (int i=0; i < blockDimY; i++) { - tmp = agg(tmp, _sum[threadIdx.x + i*blockDimX]); + for (int i = 0; i < blockDimY; i++) { + tmp = agg(tmp, _sum[threadIdx.x + i * blockDimX]); } Sum[rowIdx] = tmp; } @@ -159,25 +133,21 @@ __global__ void KeMatrixColumnOp_S(Agg agg, } template -void hl_matrix_column_op(Agg agg, - real *A_d, - real *C_d, - int dimM, - int dimN) { +void hl_matrix_column_op(Agg agg, real *A_d, real *C_d, int dimM, int dimN) { if (dimN >= 8192) { - int blocksX = (dimN + 128 -1) / 128; + int blocksX = (dimN + 128 - 1) / 128; int blocksY = 1; dim3 threads(128, 1); dim3 grid(blocksX, blocksY); - KeMatrixColumnOp<<< grid, threads, 0, STREAM_DEFAULT >>> - (agg, A_d, C_d, dimM, dimN); + KeMatrixColumnOp<<>>( + agg, A_d, C_d, dimM, dimN); } else { - int blocksX = (dimN + 32 -1) / 32; + int blocksX = (dimN + 32 - 1) / 32; int blocksY = 1; dim3 threads(32, 32); dim3 grid(blocksX, blocksY); - KeMatrixColumnOp_S<<< grid, threads, 0, STREAM_DEFAULT>>> - (agg, A_d, C_d, dimM, dimN); + KeMatrixColumnOp_S<<>>( + agg, A_d, C_d, dimM, dimN); } return; @@ -187,11 +157,7 @@ void hl_matrix_column_sum(real *A_d, real *C_d, int dimM, int dimN) { CHECK_NOTNULL(A_d); CHECK_NOTNULL(C_d); - hl_matrix_column_op(aggregate::sum(), - A_d, - C_d, - dimM, - dimN); + hl_matrix_column_op(aggregate::sum(), A_d, C_d, dimM, dimN); CHECK_SYNC("hl_matrix_column_sum failed"); } @@ -200,11 +166,7 @@ void hl_matrix_column_max(real *A_d, real *C_d, int dimM, int dimN) { CHECK_NOTNULL(A_d); CHECK_NOTNULL(C_d); - hl_matrix_column_op(aggregate::max(), - A_d, - C_d, - dimM, - dimN); + hl_matrix_column_op(aggregate::max(), A_d, C_d, dimM, dimN); CHECK_SYNC("hl_matrix_column_max failed"); } @@ -213,11 +175,7 @@ void hl_matrix_column_min(real *A_d, real *C_d, int dimM, int dimN) { CHECK_NOTNULL(A_d); CHECK_NOTNULL(C_d); - hl_matrix_column_op(aggregate::min(), - A_d, - C_d, - dimM, - dimN); + hl_matrix_column_op(aggregate::min(), A_d, C_d, dimM, dimN); CHECK_SYNC("hl_matrix_column_min failed"); } @@ -226,16 +184,16 @@ template __global__ void KeVectorSum(real *E, real *Sum, int dimM) { __shared__ double sum_s[blockSize]; int tid = threadIdx.x; - int index = blockIdx.y*blockDim.x+threadIdx.x; + int index = blockIdx.y * blockDim.x + threadIdx.x; sum_s[tid] = 0.0f; while (index < dimM) { sum_s[tid] += E[index]; - index += blockDim.x*gridDim.y; + index += blockDim.x * gridDim.y; } __syncthreads(); - for (int stride = blockSize/2; stride > 0; stride = stride/2) { + for (int stride = blockSize / 2; stride > 0; stride = stride / 2) { if (tid < stride) { sum_s[tid] += sum_s[tid + stride]; } @@ -259,38 +217,39 @@ void hl_vector_sum(real *A_d, real *C_h, int dimM) { dim3 threads(blockSize, 1); dim3 grid(blocksX, blocksY); - struct _hl_event_st hl_event_st = {.cu_event = t_resource.event}; + struct _hl_event_st hl_event_st = {.cu_event = t_resource.event}; hl_event_t hl_event = &hl_event_st; - while (!hl_cuda_event_is_ready(hl_event)) {} + while (!hl_cuda_event_is_ready(hl_event)) { + } - KeVectorSum<128><<< grid, threads, 0, STREAM_DEFAULT >>> - (A_d, t_resource.gpu_mem, dimM); - KeVectorSum<128><<< 1, threads, 0, STREAM_DEFAULT >>> - (t_resource.gpu_mem, t_resource.cpu_mem, 128); + KeVectorSum<128><<>>( + A_d, t_resource.gpu_mem, dimM); + KeVectorSum<128><<<1, threads, 0, STREAM_DEFAULT>>>( + t_resource.gpu_mem, t_resource.cpu_mem, 128); hl_memcpy_async(C_h, t_resource.cpu_mem, sizeof(real), HPPL_STREAM_DEFAULT); hl_stream_record_event(HPPL_STREAM_DEFAULT, hl_event); hl_stream_synchronize(HPPL_STREAM_DEFAULT); cudaError_t err = (cudaError_t)hl_get_device_last_error(); - CHECK_EQ(cudaSuccess, err) - << "CUDA error: " << hl_get_device_error_string((size_t)err); + CHECK_EQ(cudaSuccess, err) << "CUDA error: " + << hl_get_device_error_string((size_t)err); } template __global__ void KeVectorAbsSum(real *E, real *Sum, int dimM) { __shared__ double sum_s[blockSize]; int tid = threadIdx.x; - int index = blockIdx.y*blockDim.x+threadIdx.x; + int index = blockIdx.y * blockDim.x + threadIdx.x; sum_s[tid] = 0.0f; while (index < dimM) { sum_s[tid] += abs(E[index]); - index += blockDim.x*gridDim.y; + index += blockDim.x * gridDim.y; } __syncthreads(); - for (int stride = blockSize/2; stride > 0; stride = stride/2) { + for (int stride = blockSize / 2; stride > 0; stride = stride / 2) { if (tid < stride) { sum_s[tid] += sum_s[tid + stride]; } @@ -314,20 +273,21 @@ void hl_vector_abs_sum(real *A_d, real *C_h, int dimM) { dim3 threads(blockSize, 1); dim3 grid(blocksX, blocksY); - struct _hl_event_st hl_event_st = {.cu_event = t_resource.event}; + struct _hl_event_st hl_event_st = {.cu_event = t_resource.event}; hl_event_t hl_event = &hl_event_st; - while (!hl_cuda_event_is_ready(hl_event)) {} + while (!hl_cuda_event_is_ready(hl_event)) { + } - KeVectorAbsSum<128><<< grid, threads, 0, STREAM_DEFAULT >>> - (A_d, t_resource.gpu_mem, dimM); - KeVectorAbsSum<128><<< 1, threads, 0, STREAM_DEFAULT >>> - (t_resource.gpu_mem, t_resource.cpu_mem, 128); + KeVectorAbsSum<128><<>>( + A_d, t_resource.gpu_mem, dimM); + KeVectorAbsSum<128><<<1, threads, 0, STREAM_DEFAULT>>>( + t_resource.gpu_mem, t_resource.cpu_mem, 128); hl_memcpy_async(C_h, t_resource.cpu_mem, sizeof(real), HPPL_STREAM_DEFAULT); hl_stream_record_event(HPPL_STREAM_DEFAULT, hl_event); hl_stream_synchronize(HPPL_STREAM_DEFAULT); cudaError_t err = (cudaError_t)hl_get_device_last_error(); - CHECK_EQ(cudaSuccess, err) - << "CUDA error: " << hl_get_device_error_string((size_t)err); + CHECK_EQ(cudaSuccess, err) << "CUDA error: " + << hl_get_device_error_string((size_t)err); } diff --git a/paddle/cuda/src/hl_cuda_cnn.cu b/paddle/cuda/src/hl_cuda_cnn.cu index b6e3e63a4f52261e49467bd82fdabd063e81460e..aac19b1ea566ad69f1f7374e393676c8debd9883 100644 --- a/paddle/cuda/src/hl_cuda_cnn.cu +++ b/paddle/cuda/src/hl_cuda_cnn.cu @@ -12,21 +12,27 @@ WITHOUT WARRANTIES 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 "hl_base.h" #include "hl_cnn.h" #include "hl_device_functions.cuh" -__global__ void KeMaxPoolForward(const int nthreads, const real* inputData, - const int channels, const int height, +__global__ void KeMaxPoolForward(const int nthreads, + const real* inputData, + const int channels, + const int height, const int width, - const int pooledH, const int pooledW, - const int ksizeW, const int ksizeH, - const int strideH, const int strideW, - const int offsetH, const int offsetW, - real* tgtData, const int tgtStride) { - int index = blockIdx.x * blockDim.x + threadIdx.x; + const int pooledH, + const int pooledW, + const int ksizeW, + const int ksizeH, + const int strideH, + const int strideW, + const int offsetH, + const int offsetW, + real* tgtData, + const int tgtStride) { + int index = blockIdx.x * blockDim.x + threadIdx.x; if (index < nthreads) { int pw = index % pooledW; int ph = (index / pooledW) % pooledH; @@ -46,44 +52,70 @@ __global__ void KeMaxPoolForward(const int nthreads, const real* inputData, maxval = inputData[h * width + w]; } } - int tgtIndex = index % (pooledW * pooledH * channels) + - frameNum * tgtStride; + int tgtIndex = + index % (pooledW * pooledH * channels) + frameNum * tgtStride; tgtData[tgtIndex] = maxval; } } -void hl_maxpool_forward(const int frameCnt, const real* inputData, +void hl_maxpool_forward(const int frameCnt, + const real* inputData, const int channels, - const int height, const int width, - const int pooledH, const int pooledW, - const int sizeX, const int sizeY, - const int strideH, const int strideW, - const int paddingH, const int paddingW, - real* tgtData, const int tgtStride) { - + const int height, + const int width, + const int pooledH, + const int pooledW, + const int sizeX, + const int sizeY, + const int strideH, + const int strideW, + const int paddingH, + const int paddingW, + real* tgtData, + const int tgtStride) { int num_kernels = pooledH * pooledW * channels * frameCnt; int blocks = (num_kernels + 1024 - 1) / 1024; dim3 threads(1024, 1); dim3 grid(blocks, 1); - KeMaxPoolForward<<< grid, threads, 0, STREAM_DEFAULT >>> - (num_kernels, inputData, channels, height, width, - pooledH, pooledW, sizeX, sizeY, strideH, strideW, - paddingH, paddingW, tgtData, tgtStride); + KeMaxPoolForward<<>>(num_kernels, + inputData, + channels, + height, + width, + pooledH, + pooledW, + sizeX, + sizeY, + strideH, + strideW, + paddingH, + paddingW, + tgtData, + tgtStride); CHECK_SYNC("hl_maxpool_forward failed"); } -__global__ void KeMaxPoolBackward(const int nthreads, const real* inputData, - const real* outData, const real* outGrad, - const int channels, const int height, +__global__ void KeMaxPoolBackward(const int nthreads, + const real* inputData, + const real* outData, + const real* outGrad, + const int channels, + const int height, const int width, - const int pooledH, const int pooledW, - const int sizeX, const int sizeY, - const int strideH, const int strideW, - const int padH, const int padW, - real scaleA, real scaleB, - real* targetGrad, const int outStride) { - int index = blockIdx.x * blockDim.x + threadIdx.x; + const int pooledH, + const int pooledW, + const int sizeX, + const int sizeY, + const int strideH, + const int strideW, + const int padH, + const int padW, + real scaleA, + real scaleB, + real* targetGrad, + const int outStride) { + int index = blockIdx.x * blockDim.x + threadIdx.x; if (index < nthreads) { // find out the local index // find out the local offset @@ -107,43 +139,69 @@ __global__ void KeMaxPoolBackward(const int nthreads, const real* inputData, } } } - targetGrad[index] = - scaleB * targetGrad[index] + scaleA * gradient; + targetGrad[index] = scaleB * targetGrad[index] + scaleA * gradient; } } -void hl_maxpool_backward(const int frameCnt, const real* inputData, - const real* outData, const real* outGrad, - const int channels, const int height, - const int width, - const int pooledH, const int pooledW, - const int sizeX, const int sizeY, - const int strideH, const int strideW, - const int paddingH, const int paddingW, - real scaleA, real scaleB, - real* targetGrad, const int outStride) { - +void hl_maxpool_backward(const int frameCnt, + const real* inputData, + const real* outData, + const real* outGrad, + const int channels, + const int height, + const int width, + const int pooledH, + const int pooledW, + const int sizeX, + const int sizeY, + const int strideH, + const int strideW, + const int paddingH, + const int paddingW, + real scaleA, + real scaleB, + real* targetGrad, + const int outStride) { int num_kernels = height * width * channels * frameCnt; int blocks = (num_kernels + 1024 - 1) / 1024; - KeMaxPoolBackward<<< blocks, 1024, 0, STREAM_DEFAULT >>> - (num_kernels, inputData, outData, outGrad, channels, - height, width, pooledH, pooledW, sizeX, sizeY, - strideH, strideW, - paddingH, paddingW, - scaleA, scaleB, - targetGrad, outStride); + KeMaxPoolBackward<<>>(num_kernels, + inputData, + outData, + outGrad, + channels, + height, + width, + pooledH, + pooledW, + sizeX, + sizeY, + strideH, + strideW, + paddingH, + paddingW, + scaleA, + scaleB, + targetGrad, + outStride); CHECK_SYNC("hl_maxpool_backward"); } -__global__ void KeAvgPoolForward(const int nthreads, const real* inputData, +__global__ void KeAvgPoolForward(const int nthreads, + const real* inputData, const int channels, - const int height, const int width, - const int pooledH, const int pooledW, - const int sizeX, const int sizeY, - const int strideH, const int strideW, - const int padH, const int padW, - real* tgtData, const int tgtStride) { + const int height, + const int width, + const int pooledH, + const int pooledW, + const int sizeX, + const int sizeY, + const int strideH, + const int strideW, + const int padH, + const int padW, + real* tgtData, + const int tgtStride) { int index = blockIdx.x * blockDim.x + threadIdx.x; if (index < nthreads) { int pw = index % pooledW; @@ -168,39 +226,64 @@ __global__ void KeAvgPoolForward(const int nthreads, const real* inputData, aveval += inputData[h * width + w]; } } - int tgtIndex = index % (pooledW * pooledH * channels) + - frameNum * tgtStride; + int tgtIndex = + index % (pooledW * pooledH * channels) + frameNum * tgtStride; tgtData[tgtIndex] = aveval / pool_size; } } -void hl_avgpool_forward(const int frameCnt, const real* inputData, +void hl_avgpool_forward(const int frameCnt, + const real* inputData, const int channels, - const int height, const int width, - const int pooledH, const int pooledW, - const int sizeX, const int sizeY, - const int strideH, const int strideW, - const int paddingH, const int paddingW, - real* tgtData, const int tgtStride) { + const int height, + const int width, + const int pooledH, + const int pooledW, + const int sizeX, + const int sizeY, + const int strideH, + const int strideW, + const int paddingH, + const int paddingW, + real* tgtData, + const int tgtStride) { int num_kernels = pooledH * pooledW * channels * frameCnt; int blocks = (num_kernels + 1024 - 1) / 1024; - KeAvgPoolForward<<< blocks, 1024, 0, STREAM_DEFAULT >>> - (num_kernels, inputData, channels, - height, width, pooledH, pooledW, - sizeX, sizeY, strideH, strideW, - paddingH, paddingW, tgtData, tgtStride); + KeAvgPoolForward<<>>(num_kernels, + inputData, + channels, + height, + width, + pooledH, + pooledW, + sizeX, + sizeY, + strideH, + strideW, + paddingH, + paddingW, + tgtData, + tgtStride); CHECK_SYNC("hl_avgpool_forward failed"); } -__global__ void KeAvgPoolBackward(const int nthreads, const real* outGrad, - const int channels, const int height, +__global__ void KeAvgPoolBackward(const int nthreads, + const real* outGrad, + const int channels, + const int height, const int width, - const int pooledH, const int pooledW, - const int sizeX, const int sizeY, - const int strideH, const int strideW, - const int padH, const int padW, - real scaleA, real scaleB, - real* tgtGrad, const int outStride) { + const int pooledH, + const int pooledW, + const int sizeX, + const int sizeY, + const int strideH, + const int strideW, + const int padH, + const int padW, + real scaleA, + real scaleB, + real* tgtGrad, + const int outStride) { int index = blockIdx.x * blockDim.x + threadIdx.x; if (index < nthreads) { int offsetW = index % width + padW; @@ -215,7 +298,6 @@ __global__ void KeAvgPoolBackward(const int nthreads, const real* outGrad, real gradient = 0; outGrad += (frameNum * outStride + offsetC * pooledH * pooledW); - for (int ph = phstart; ph < phend; ++ph) { for (int pw = pwstart; pw < pwend; ++pw) { // figure out the pooling size @@ -224,32 +306,50 @@ __global__ void KeAvgPoolBackward(const int nthreads, const real* outGrad, int hend = min(hstart + sizeY, height + padH); int wend = min(wstart + sizeX, width + padW); int poolsize = (hend - hstart) * (wend - wstart); - gradient += outGrad[ph * pooledW + pw]/poolsize; + gradient += outGrad[ph * pooledW + pw] / poolsize; } } tgtGrad[index] = scaleB * tgtGrad[index] + scaleA * gradient; } } -void hl_avgpool_backward(const int frameCnt, const real* outGrad, +void hl_avgpool_backward(const int frameCnt, + const real* outGrad, const int channels, - const int height, const int width, - const int pooledH, const int pooledW, - const int sizeX, const int sizeY, - const int strideH, const int strideW, - const int paddingH, const int paddingW, - real scaleA, real scaleB, - real* backGrad, const int outStride) { + const int height, + const int width, + const int pooledH, + const int pooledW, + const int sizeX, + const int sizeY, + const int strideH, + const int strideW, + const int paddingH, + const int paddingW, + real scaleA, + real scaleB, + real* backGrad, + const int outStride) { int num_kernels = height * width * channels * frameCnt; int blocks = (num_kernels + 1024 - 1) / 1024; - KeAvgPoolBackward <<< blocks, 1024, 0, STREAM_DEFAULT >>> - (num_kernels, outGrad, channels, height, width, - pooledH, pooledW, sizeX, sizeY, - strideH, strideW, - paddingH, paddingW, - scaleA, scaleB, - backGrad, outStride); + KeAvgPoolBackward<<>>(num_kernels, + outGrad, + channels, + height, + width, + pooledH, + pooledW, + sizeX, + sizeY, + strideH, + strideW, + paddingH, + paddingW, + scaleA, + scaleB, + backGrad, + outStride); CHECK_SYNC("hl_avgpool_backward failed"); } @@ -266,7 +366,7 @@ __global__ void KeBilinearInterpFw(const real* in, const size_t numChannels, const real ratioH, const real ratioW) { - int nthreads = outputH * outputW; + int nthreads = outputH * outputW; int tid = blockIdx.x * blockDim.x + threadIdx.x; if (tid < nthreads) { int outIdH = tid / outputW; @@ -287,13 +387,14 @@ __global__ void KeBilinearInterpFw(const real* in, real w1lambda = ratioW * outImgIdx - inImgIdx; real w2lambda = 1.f - w1lambda; - const real* inPos = - &in[outIdH * inputW + channelId * inImgSize + inImgIdy * inImgW + inImgIdx]; + const real* inPos = &in[outIdH * inputW + channelId * inImgSize + + inImgIdy * inImgW + inImgIdx]; // bilinear interpolation out[outIdH * outputW + outIdW] = - h2lambda * (w2lambda * inPos[0] + w1lambda * inPos[wId]) + - h1lambda * (w2lambda * inPos[hId * inImgW] + w1lambda * inPos[hId * inImgW + wId]); + h2lambda * (w2lambda * inPos[0] + w1lambda * inPos[wId]) + + h1lambda * (w2lambda * inPos[hId * inImgW] + + w1lambda * inPos[hId * inImgW + wId]); } } @@ -313,9 +414,19 @@ void hl_bilinear_forward(const real* inData, int threadNum = outputH * outputW; int blocks = (threadNum + 1024 - 1) / 1024; - KeBilinearInterpFw<<< blocks, 1024, 0, STREAM_DEFAULT>>>( - inData, inImgH, inImgW, inputH, inputW, outData, outImgH, - outImgW, outputH, outputW, numChannels, ratioH, ratioW); + KeBilinearInterpFw<<>>(inData, + inImgH, + inImgW, + inputH, + inputW, + outData, + outImgH, + outImgW, + outputH, + outputW, + numChannels, + ratioH, + ratioW); CHECK_SYNC("hl_bilinear_forward failed"); } @@ -353,13 +464,15 @@ __global__ void KeBilinearInterpBw(real* in, real w1lambda = ratioW * outImgIdx - inImgIdx; real w2lambda = 1.f - w1lambda; - real* inPos = - &in[outIdH * inputW + channelId * inImgSize + inImgIdy * inImgW + inImgIdx]; + real* inPos = &in[outIdH * inputW + channelId * inImgSize + + inImgIdy * inImgW + inImgIdx]; const real* outPos = &out[outIdH * outputW + outIdW]; paddle::paddleAtomicAdd(&inPos[0], h2lambda * w2lambda * outPos[0]); paddle::paddleAtomicAdd(&inPos[wId], h2lambda * w1lambda * outPos[0]); - paddle::paddleAtomicAdd(&inPos[hId * inImgW], h1lambda * w2lambda * outPos[0]); - paddle::paddleAtomicAdd(&inPos[hId * inImgW + wId], h1lambda * w1lambda * outPos[0]); + paddle::paddleAtomicAdd(&inPos[hId * inImgW], + h1lambda * w2lambda * outPos[0]); + paddle::paddleAtomicAdd(&inPos[hId * inImgW + wId], + h1lambda * w1lambda * outPos[0]); } } @@ -379,22 +492,37 @@ void hl_bilinear_backward(real* inGrad, int threadNum = outputH * outputW; int blocks = (threadNum + 1024 - 1) / 1024; - KeBilinearInterpBw<<< blocks, 1024, 0, STREAM_DEFAULT>>>( - inGrad, inImgH, inImgW, inputH, inputW, outGrad, outImgH, - outImgW, outputH, outputW, numChannels, ratioH, ratioW); + KeBilinearInterpBw<<>>(inGrad, + inImgH, + inImgW, + inputH, + inputW, + outGrad, + outImgH, + outImgW, + outputH, + outputW, + numChannels, + ratioH, + ratioW); CHECK_SYNC("hl_bilinear_backward failed"); } -__global__ void maxoutFpCompute(size_t nthreads, const real * inData, - real * outData, int* idData, - size_t size, size_t featLen, size_t groups) { +__global__ void maxoutFpCompute(size_t nthreads, + const real* inData, + real* outData, + int* idData, + size_t size, + size_t featLen, + size_t groups) { int index = blockIdx.x * blockDim.x + threadIdx.x; - if(index < nthreads) { + if (index < nthreads) { size_t batch_idx = index / size; size_t i = index % size; size_t channel_idx = i / featLen; size_t feat_idx = i % featLen; - size_t data_idx = (batch_idx * size + channel_idx * featLen) * groups + feat_idx; + size_t data_idx = + (batch_idx * size + channel_idx * featLen) * groups + feat_idx; real max = inData[data_idx]; int maxId = 0; for (size_t g = 1; g < groups; ++g) { @@ -409,37 +537,50 @@ __global__ void maxoutFpCompute(size_t nthreads, const real * inData, } } -void hl_maxout_forward(const real* inData, real* outData, - int* idData, size_t batchSize, size_t size, - size_t featLen, size_t groups) { +void hl_maxout_forward(const real* inData, + real* outData, + int* idData, + size_t batchSize, + size_t size, + size_t featLen, + size_t groups) { int num_kernels = size * batchSize; int blocks = (num_kernels + 1024 - 1) / 1024; - maxoutFpCompute<<< blocks, 1024, 0, STREAM_DEFAULT>>>( - num_kernels, inData, outData, idData, size, featLen, groups); + maxoutFpCompute<<>>( + num_kernels, inData, outData, idData, size, featLen, groups); CHECK_SYNC("hl_maxout_forward failed"); } -__global__ void maxoutBpCompute(size_t nthreads, real* inGrad, - const real* outGrad, const int* idData, - size_t size, size_t featLen, size_t groups) { +__global__ void maxoutBpCompute(size_t nthreads, + real* inGrad, + const real* outGrad, + const int* idData, + size_t size, + size_t featLen, + size_t groups) { int index = blockIdx.x * blockDim.x + threadIdx.x; - if(index < nthreads) { + if (index < nthreads) { size_t batch_idx = index / size; size_t i = index % size; size_t channel_idx = i / featLen; size_t feat_idx = i % featLen; size_t newIndex = batch_idx * size; - size_t gradIdx = (channel_idx * groups + (idData + newIndex)[i]) * featLen + feat_idx; + size_t gradIdx = + (channel_idx * groups + (idData + newIndex)[i]) * featLen + feat_idx; (inGrad + newIndex * groups)[gradIdx] += (outGrad + newIndex)[i]; } } -void hl_maxout_backward(real* inGrad, const real* outGrad, - const int* idData, size_t batchSize, size_t size, - size_t featLen, size_t groups) { +void hl_maxout_backward(real* inGrad, + const real* outGrad, + const int* idData, + size_t batchSize, + size_t size, + size_t featLen, + size_t groups) { int num_kernels = size * batchSize; int blocks = (num_kernels + 1024 - 1) / 1024; - maxoutBpCompute<<< blocks, 1024, 0, STREAM_DEFAULT >>>( - num_kernels, inGrad, outGrad, idData, size, featLen, groups); + maxoutBpCompute<<>>( + num_kernels, inGrad, outGrad, idData, size, featLen, groups); CHECK_SYNC("hl_maxout_backward failed"); } diff --git a/paddle/cuda/src/hl_cuda_cudnn.cc b/paddle/cuda/src/hl_cuda_cudnn.cc index 7ad8a39768a064140a08c912a5a467bc24a12adf..78642a17443b0b4d81defaa46579332ef20c71a1 100644 --- a/paddle/cuda/src/hl_cuda_cudnn.cc +++ b/paddle/cuda/src/hl_cuda_cudnn.cc @@ -1023,14 +1023,6 @@ void hl_batch_norm_forward_inference(hl_tensor_descriptor inputDesc, real beta = 1.0f; cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL; - int batch_size = ((cudnn_tensor_descriptor)inputDesc)->batch_size; - if (batch_size > 1024 && g_cudnn_lib_version < 6000) { - LOG(INFO) << " To process current batch data with size " << batch_size - << " (>1024), cudnnBatchNorm requires cuDNN version >= 6000." - << " If there is an error complaining CUDNN_STATUS_NOT_SUPPORTED," - << " just recompile PaddlePaddle with cuDNN >= 6000, replacing" - << " current version " << g_cudnn_lib_version; - } CHECK_CUDNN( dynload::cudnnBatchNormalizationForwardInference(t_resource.cudnn_handle, mode, diff --git a/paddle/cuda/src/hl_cuda_lstm.cu b/paddle/cuda/src/hl_cuda_lstm.cu index b869d903ba3cfb188f823518ba8ee7d17f9b2440..a5ce81a904ebbd655a16ef68660b81d442478575 100644 --- a/paddle/cuda/src/hl_cuda_lstm.cu +++ b/paddle/cuda/src/hl_cuda_lstm.cu @@ -12,14 +12,13 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ - +#include "hl_activation_functions.h" #include "hl_base.h" #include "hl_cuda_cublas.h" #include "hl_device_functions.cuh" -#include "hl_activation_functions.h" #include "paddle/utils/Logging.h" -typedef hppl::Active::forward t_forward; +typedef hppl::Active::forward t_forward; typedef hppl::Active::backward t_backward; bool hl_lstm_sequence_parallel(int frameSize) { @@ -42,9 +41,9 @@ public: value_ += (start + length - 1) * frameSize + idx; } } - __device__ inline real *getPtr() const {return value_;} - __device__ inline real getValue() {return *value_;} - __device__ inline void setValue(real value) {*value_ = value;} + __device__ inline real *getPtr() const { return value_; } + __device__ inline real getValue() { return *value_; } + __device__ inline void setValue(real value) { *value_ = value; } template __device__ inline void nextFrame() { if (reversed == 0) { @@ -55,28 +54,25 @@ public: } }; -__device__ __forceinline__ -void ptx_sync(const int id, const int barriers) { +__device__ __forceinline__ void ptx_sync(const int id, const int barriers) { asm volatile("bar.sync %0, %1;" : : "r"(id), "r"(barriers) : "memory"); } -__device__ __forceinline__ -void ptx_arrive(const int id, const int barriers) { +__device__ __forceinline__ void ptx_arrive(const int id, const int barriers) { asm volatile("bar.arrive %0, %1;" : : "r"(id), "r"(barriers) : "memory"); } -template -__device__ __forceinline__ real -forward_sequence(real value, - real *shValue, - real *state, - real *preOutput, - real *output, - real check, - int index, - t_forward activeNode, - t_forward activeGate, - t_forward activeState) { +template +__device__ __forceinline__ real forward_sequence(real value, + real *shValue, + real *state, + real *preOutput, + real *output, + real check, + int index, + t_forward activeNode, + t_forward activeGate, + t_forward activeState) { real out; real prevOut; real state_r; @@ -112,17 +108,20 @@ forward_sequence(real value, if (idy == 0) { ptx_sync(2, frameSize * 2); prevOut = state[idx]; - prevOut = activeState(prevOut); + prevOut = activeState(prevOut); preOutput[idx] = prevOut; ptx_arrive(3, frameSize * 2); } return value; } -#define OUTPUT_BARRIER_ID 10 -#define OUTPUT_BARRIER_ID2 11 -template +#define OUTPUT_BARRIER_ID 10 +#define OUTPUT_BARRIER_ID2 11 +template __global__ void KeLstmForward(real *gateValue, real *state, real *output, @@ -184,10 +183,16 @@ __global__ void KeLstmForward(real *gateValue, } } value = forward_sequence( - value, shValue, shState, shPrevOutput, shOutput, check, index, - hppl::gpu::forward[active_node], - hppl::gpu::forward[active_gate], - hppl::gpu::forward[active_state]); + value, + shValue, + shState, + shPrevOutput, + shOutput, + check, + index, + hppl::gpu::forward[active_node], + hppl::gpu::forward[active_gate], + hppl::gpu::forward[active_state]); const int idx = index % frameSize; const int idy = index / frameSize; if (valueSize == 128) { @@ -218,7 +223,7 @@ __global__ void KeLstmForward(real *gateValue, real B_r[frameSize]; const int computeIdx = index - valueSize; if (i == 0) { - #pragma unroll +#pragma unroll for (int n = 0; n < frameSize; n++) { B_r[n] = weight[n * valueSize + computeIdx]; } @@ -230,7 +235,7 @@ __global__ void KeLstmForward(real *gateValue, } real sum = 0.0f; for (int n = 0; n < frameSize; n++) { - sum += A_r[n]*B_r[n]; + sum += A_r[n] * B_r[n]; } shValue[computeIdx] = sum; ptx_arrive(OUTPUT_BARRIER_ID2, blockSize); @@ -239,14 +244,14 @@ __global__ void KeLstmForward(real *gateValue, if (valueSize == 256) { real B_r[frameSize]; if (i == 0) { - #pragma unroll +#pragma unroll for (int n = 0; n < frameSize; n++) { B_r[n] = weight[n * valueSize + index]; } } real sum = 0.0f; for (int n = 0; n < frameSize; n++) { - sum += shOutput[n]*B_r[n]; + sum += shOutput[n] * B_r[n]; } value += sum; } @@ -273,50 +278,81 @@ void hl_lstm_parallel_forward(real *gateValue, dim3 grid(numSequences, 1); if (!reversed) { if (frameSize == 32) { - KeLstmForward<128, 32, 0, 128, 256> - <<>> - (gateValue, stateValue, outputValue, preOutputValue, - checkIg, checkFg, checkOg, weight, sequence, - active_node, active_gate, active_state); + KeLstmForward<128, 32, 0, 128, 256><<>>( + gateValue, + stateValue, + outputValue, + preOutputValue, + checkIg, + checkFg, + checkOg, + weight, + sequence, + active_node, + active_gate, + active_state); } else if (frameSize == 64) { - KeLstmForward<256, 64, 0, 256, 256> - <<>> - (gateValue, stateValue, outputValue, preOutputValue, - checkIg, checkFg, checkOg, weight, sequence, - active_node, active_gate, active_state); + KeLstmForward<256, 64, 0, 256, 256><<>>( + gateValue, + stateValue, + outputValue, + preOutputValue, + checkIg, + checkFg, + checkOg, + weight, + sequence, + active_node, + active_gate, + active_state); } } else { if (frameSize == 32) { - KeLstmForward<128, 32, 1, 128, 256> - <<>> - (gateValue, stateValue, outputValue, preOutputValue, - checkIg, checkFg, checkOg, weight, sequence, - active_node, active_gate, active_state); + KeLstmForward<128, 32, 1, 128, 256><<>>( + gateValue, + stateValue, + outputValue, + preOutputValue, + checkIg, + checkFg, + checkOg, + weight, + sequence, + active_node, + active_gate, + active_state); } else if (frameSize == 64) { - KeLstmForward<256, 64, 1, 256, 256> - <<>> - (gateValue, stateValue, outputValue, preOutputValue, - checkIg, checkFg, checkOg, weight, sequence, - active_node, active_gate, active_state); + KeLstmForward<256, 64, 1, 256, 256><<>>( + gateValue, + stateValue, + outputValue, + preOutputValue, + checkIg, + checkFg, + checkOg, + weight, + sequence, + active_node, + active_gate, + active_state); } } CHECK_SYNC("hl_lstm_parallel_forward failed"); } -__device__ __forceinline__ -void transpose_32x32(real a[], const int idx) { +__device__ __forceinline__ void transpose_32x32(real a[], const int idx) { int addr = idx % 32; - #pragma unroll +#pragma unroll for (int k = 1; k < 32; k++) { // rSrc[k] = __shfl(rSrc[k], (threadIdx.x + k) % 32, 32); addr = __shfl(addr, (idx + 1) % 32, 32); a[k] = __shfl(a[k], addr, 32); } - #pragma unroll +#pragma unroll for (int tid = 0; tid < 31; tid++) { real tmp = (idx > tid) ? a[0] : a[1]; - #pragma unroll +#pragma unroll for (int k = 31; k > 0; k--) { a[(k + 1) % 32] = (idx > tid) ? a[k] : a[(k + 1) % 32]; } @@ -324,29 +360,28 @@ void transpose_32x32(real a[], const int idx) { } addr = (32 - idx) % 32; - #pragma unroll +#pragma unroll for (int k = 0; k < 32; k++) { a[k] = __shfl(a[k], addr, 32); addr = __shfl(addr, (idx + 31) % 32, 32); } } -template -__device__ void -backward_sequence(real rGateValue, - real rOutputGrad, - real rPreOutputValue, - real &rGateGrad, - real &rStateGrad, - real *shStateGrad, - real *shStateValue, - real *shGateValue, - real rCheck, - real &rGateValuePrev, - int index, - t_backward activeNode, - t_backward activeGate, - t_backward activeState) { +template +__device__ void backward_sequence(real rGateValue, + real rOutputGrad, + real rPreOutputValue, + real &rGateGrad, + real &rStateGrad, + real *shStateGrad, + real *shStateValue, + real *shGateValue, + real rCheck, + real &rGateValuePrev, + int index, + t_backward activeNode, + t_backward activeGate, + t_backward activeState) { const int frameIdx = index % frameSize; const int frameIdy = index / frameSize; if (frameIdy == 3) { @@ -363,8 +398,8 @@ backward_sequence(real rGateValue, rStateGrad = rGateGrad * rCheck; shStateGrad[index] = rStateGrad; ptx_sync(3, valueSize); - rStateGrad += shStateGrad[frameIdx + frameSize *2]; - rStateGrad += shStateGrad[frameIdx + frameSize *3]; + rStateGrad += shStateGrad[frameIdx + frameSize * 2]; + rStateGrad += shStateGrad[frameIdx + frameSize * 3]; rGateGrad = rStateGrad * shGateValue[frameIdx]; rGateGrad = activeGate(rGateGrad, rGateValue); } else if (frameIdy == 2) { @@ -373,7 +408,7 @@ backward_sequence(real rGateValue, shStateGrad[index] = rStateGrad; ptx_sync(3, valueSize); rStateGrad += shStateGrad[frameIdx + frameSize]; - rStateGrad += shStateGrad[frameIdx + frameSize *3]; + rStateGrad += shStateGrad[frameIdx + frameSize * 3]; rGateValuePrev = rGateValue; rGateGrad = rStateGrad * shStateValue[frameIdx]; rGateGrad = activeGate(rGateGrad, rGateValue); @@ -381,43 +416,43 @@ backward_sequence(real rGateValue, shGateValue[frameIdx] = rGateValue; ptx_sync(3, valueSize); rStateGrad = shStateGrad[frameIdx + frameSize]; - rStateGrad += shStateGrad[frameIdx + frameSize *2]; - rStateGrad += shStateGrad[frameIdx + frameSize *3]; + rStateGrad += shStateGrad[frameIdx + frameSize * 2]; + rStateGrad += shStateGrad[frameIdx + frameSize * 3]; rGateGrad = rStateGrad * shGateValue[frameIdx + frameSize]; rGateGrad = activeNode(rGateGrad, rGateValue); } } -template +template __device__ void load_weight(real rWeight[], real *weight, const int index) { if (valueSize == 128) { weight += index; - #pragma unroll +#pragma unroll for (int n = 0; n < frameSize; n++) { - rWeight[n] = weight[n*valueSize]; + rWeight[n] = weight[n * valueSize]; } transpose_32x32(rWeight, index % 32); } if (valueSize == 256) { int id = (index / 32) % 2; weight += index - id * 32 + id * 32 * valueSize; - #pragma unroll +#pragma unroll for (int n = 0; n < 32; n++) { - rWeight[n] = weight[n*valueSize]; - rWeight[n + 32] = weight[n*valueSize + 32]; + rWeight[n] = weight[n * valueSize]; + rWeight[n + 32] = weight[n * valueSize + 32]; } transpose_32x32(rWeight, index % 32); transpose_32x32(&rWeight[32], index % 32); } } -template +template __global__ void KeLstmBackward(real *gateValue, real *gateGrad, real *stateValue, - real *stateGrad, /* do not need save */ + real *stateGrad, /* do not need save */ real *preOutputValue, - real *preOutputGrad, /* do not need save */ + real *preOutputGrad, /* do not need save */ real *checkIg, real *checkIgGrad, real *checkFg, @@ -484,20 +519,27 @@ __global__ void KeLstmBackward(real *gateValue, for (int i = 0; i < length; ++i) { if (frameIdy == 3) { - if (i != length -1) { + if (i != length - 1) { frameStateValue.nextFrame(); shStateValue[frameIdx] = frameStateValue.getValue(); } else { shStateValue[frameIdx] = 0.0; } } - backward_sequence( - rGateValue, rOutputGrad, rPreOutputValue, rGateGrad, - rStateGrad, shStateGrad, shStateValue, shGateValue, - rCheck, rGateValuePrev, index, - hppl::gpu::backward[active_node], - hppl::gpu::backward[active_gate], - hppl::gpu::backward[active_state]); + backward_sequence(rGateValue, + rOutputGrad, + rPreOutputValue, + rGateGrad, + rStateGrad, + shStateGrad, + shStateValue, + shGateValue, + rCheck, + rGateValuePrev, + index, + hppl::gpu::backward[active_node], + hppl::gpu::backward[active_gate], + hppl::gpu::backward[active_state]); if (frameIdy == 3) { rCheckGrad += rGateGrad * rStateValue; rStateValue = shStateValue[frameIdx]; @@ -523,9 +565,9 @@ __global__ void KeLstmBackward(real *gateValue, shGateGrad[frameIdy][frameIdx] = rGateGrad; if (valueSize == 128) { real sum = 0.0f; - #pragma unroll +#pragma unroll for (int n = 0; n < frameSize; n++) { - sum += shGateGrad[frameIdy][n]*B_r[n]; + sum += shGateGrad[frameIdy][n] * B_r[n]; } if (frameIdy == 3) { rOutputGrad += sum; @@ -541,7 +583,7 @@ __global__ void KeLstmBackward(real *gateValue, } real sum = 0.0f; for (int n = 0; n < frameSize; n++) { - sum += A_r[n]*B_r[n]; + sum += A_r[n] * B_r[n]; } if (frameIdy == 3) { rOutputGrad += sum; @@ -552,8 +594,8 @@ __global__ void KeLstmBackward(real *gateValue, if (frameIdy == 3) { ptx_sync(6, valueSize); - #pragma unroll - for (int i = 0; i < 3; i ++) { +#pragma unroll + for (int i = 0; i < 3; i++) { rOutputGrad += shOutputGrad[i][frameIdx]; } } else { @@ -564,11 +606,14 @@ __global__ void KeLstmBackward(real *gateValue, /* TODO: Temporary save & merger in another kernel */ if (frameIdy == 1) { - if (checkIgGrad) paddle::paddleAtomicAdd(checkIgGrad+frameIdx, rCheckGrad); + if (checkIgGrad) + paddle::paddleAtomicAdd(checkIgGrad + frameIdx, rCheckGrad); } else if (frameIdy == 2) { - if (checkFgGrad) paddle::paddleAtomicAdd(checkFgGrad+frameIdx, rCheckGrad); + if (checkFgGrad) + paddle::paddleAtomicAdd(checkFgGrad + frameIdx, rCheckGrad); } else if (frameIdy == 3) { - if (checkOgGrad) paddle::paddleAtomicAdd(checkOgGrad+frameIdx, rCheckGrad); + if (checkOgGrad) + paddle::paddleAtomicAdd(checkOgGrad + frameIdx, rCheckGrad); } } @@ -593,68 +638,183 @@ void hl_lstm_parallel_backward_data(real *gateValue, hl_activation_mode_t active_node, hl_activation_mode_t active_gate, hl_activation_mode_t active_state) { - CHECK(frameSize == 32 || frameSize == 64 || - frameSize == 128 || frameSize == 256); + CHECK(frameSize == 32 || frameSize == 64 || frameSize == 128 || + frameSize == 256); dim3 grid(numSequences, 1); if (!reversed) { if (frameSize == 32) { - KeLstmBackward<128, 32, 0><<>> - (gateValue, gateGrad, stateValue, stateGrad, preOutputValue, - preOutputGrad, checkIg, checkIgGrad, checkFg, checkFgGrad, checkOg, - checkOgGrad, outputGrad, weight, sequence, - active_node, active_gate, active_state); + KeLstmBackward<128, 32, 0><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); } else if (frameSize == 64) { - KeLstmBackward<256, 64, 0><<>> - (gateValue, gateGrad, stateValue, stateGrad, preOutputValue, - preOutputGrad, checkIg, checkIgGrad, checkFg, checkFgGrad, checkOg, - checkOgGrad, outputGrad, weight, sequence, - active_node, active_gate, active_state); + KeLstmBackward<256, 64, 0><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); } else if (frameSize == 128) { - KeLstmBackward<512, 128, 0><<>> - (gateValue, gateGrad, stateValue, stateGrad, preOutputValue, - preOutputGrad, checkIg, checkIgGrad, checkFg, checkFgGrad, checkOg, - checkOgGrad, outputGrad, weight, sequence, - active_node, active_gate, active_state); + KeLstmBackward<512, 128, 0><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); } else if (frameSize == 256) { - KeLstmBackward<1024, 256, 0><<>> - (gateValue, gateGrad, stateValue, stateGrad, preOutputValue, - preOutputGrad, checkIg, checkIgGrad, checkFg, checkFgGrad, checkOg, - checkOgGrad, outputGrad, weight, sequence, - active_node, active_gate, active_state); + KeLstmBackward<1024, 256, 0><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); } } else { if (frameSize == 32) { - KeLstmBackward<128, 32, 1><<>> - (gateValue, gateGrad, stateValue, stateGrad, preOutputValue, - preOutputGrad, checkIg, checkIgGrad, checkFg, checkFgGrad, checkOg, - checkOgGrad, outputGrad, weight, sequence, - active_node, active_gate, active_state); + KeLstmBackward<128, 32, 1><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); } else if (frameSize == 64) { - KeLstmBackward<256, 64, 1><<>> - (gateValue, gateGrad, stateValue, stateGrad, preOutputValue, - preOutputGrad, checkIg, checkIgGrad, checkFg, checkFgGrad, checkOg, - checkOgGrad, outputGrad, weight, sequence, - active_node, active_gate, active_state); + KeLstmBackward<256, 64, 1><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); } else if (frameSize == 128) { - KeLstmBackward<512, 128, 1><<>> - (gateValue, gateGrad, stateValue, stateGrad, preOutputValue, - preOutputGrad, checkIg, checkIgGrad, checkFg, checkFgGrad, checkOg, - checkOgGrad, outputGrad, weight, sequence, - active_node, active_gate, active_state); + KeLstmBackward<512, 128, 1><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); } else if (frameSize == 256) { - KeLstmBackward<1024, 256, 1><<>> - (gateValue, gateGrad, stateValue, stateGrad, preOutputValue, - preOutputGrad, checkIg, checkIgGrad, checkFg, checkFgGrad, checkOg, - checkOgGrad, outputGrad, weight, sequence, - active_node, active_gate, active_state); + KeLstmBackward<1024, 256, 1><<>>( + gateValue, + gateGrad, + stateValue, + stateGrad, + preOutputValue, + preOutputGrad, + checkIg, + checkIgGrad, + checkFg, + checkFgGrad, + checkOg, + checkOgGrad, + outputGrad, + weight, + sequence, + active_node, + active_gate, + active_state); } } CHECK_SYNC("hl_lstm_parallel_backward_data"); } -template +template __global__ void KeSetGradZero(real *gateGrad, - const int *starts, int valueSize, int numSequences, bool reversed) { + const int *starts, + int valueSize, + int numSequences, + bool reversed) { // const int tid = threadIdx.x; const int frameIdx = blockIdx.x * B_X + threadIdx.x; @@ -682,19 +842,31 @@ void hl_lstm_parallel_backward_weight(real *weightGrad, int valueSize = 4 * frameSize; dim3 threads(32, 32); dim3 grid((valueSize + 32 - 1) / 32, (numSequences + 32 - 1) / 32); - KeSetGradZero<32, 32><<>> - (gateGrad, sequence, valueSize, numSequences, reversed); + KeSetGradZero<32, 32><<>>( + gateGrad, sequence, valueSize, numSequences, reversed); if (!reversed) { hl_matrix_mul(outputValue, - HPPL_OP_T, gateGrad + valueSize, HPPL_OP_N, weightGrad, - frameSize, valueSize, batchSize - 1, - 1.0, 1.0); + HPPL_OP_T, + gateGrad + valueSize, + HPPL_OP_N, + weightGrad, + frameSize, + valueSize, + batchSize - 1, + 1.0, + 1.0); } else { hl_matrix_mul(outputValue + frameSize, - HPPL_OP_T, gateGrad, HPPL_OP_N, weightGrad, - frameSize, valueSize, batchSize - 1, - 1.0, 1.0); + HPPL_OP_T, + gateGrad, + HPPL_OP_N, + weightGrad, + frameSize, + valueSize, + batchSize - 1, + 1.0, + 1.0); } CHECK_SYNC("hl_lstm_parallel_backward_weight"); } diff --git a/paddle/cuda/src/hl_cuda_matrix.cu b/paddle/cuda/src/hl_cuda_matrix.cu index 9bcc7fb7de44b2211db450fb164655f7947dcad9..39272456c394adc0509e60cf5972df832f7b3424 100644 --- a/paddle/cuda/src/hl_cuda_matrix.cu +++ b/paddle/cuda/src/hl_cuda_matrix.cu @@ -12,22 +12,21 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ - #include "hl_base.h" +#include "hl_device_functions.cuh" +#include "hl_gpu_matrix_kernel.cuh" #include "hl_matrix.h" -#include "hl_matrix_ops.cuh" #include "hl_matrix_apply.cuh" +#include "hl_matrix_ops.cuh" #include "hl_sequence.h" #include "hl_sparse.ph" #include "paddle/utils/Logging.h" -#include "hl_device_functions.cuh" -#include "hl_gpu_matrix_kernel.cuh" DEFINE_MATRIX_UNARY_OP(Zero, a = 0); -DEFINE_MATRIX_TERNARY_PARAMETER_OP(_add, TWO_PARAMETER, c = p1*a + p2*b); -void hl_matrix_add(real *A_d, - real *B_d, - real *C_d, +DEFINE_MATRIX_TERNARY_PARAMETER_OP(_add, TWO_PARAMETER, c = p1 * a + p2 * b); +void hl_matrix_add(real* A_d, + real* B_d, + real* C_d, int dimM, int dimN, real alpha, @@ -36,33 +35,32 @@ void hl_matrix_add(real *A_d, CHECK_NOTNULL(B_d); CHECK_NOTNULL(C_d); - hl_gpu_apply_ternary_op - , 0, 0>(ternary::_add(alpha, beta), - A_d, - B_d, - C_d, - dimM, - dimN, - dimN, - dimN, - dimN); + hl_gpu_apply_ternary_op, 0, 0>( + ternary::_add(alpha, beta), + A_d, + B_d, + C_d, + dimM, + dimN, + dimN, + dimN, + dimN); CHECK_SYNC("hl_matrix_add failed"); } #ifdef PADDLE_TYPE_DOUBLE - #define THRESHOLD 128 +#define THRESHOLD 128 #else - #define THRESHOLD 64 +#define THRESHOLD 64 #endif -__device__ __forceinline__ -void findMax(real* I, - real* dfMax_s, - int blockSize, - int base, - int curIdx, - int nextIdx, - int dimN, - real* max) { +__device__ __forceinline__ void findMax(real* I, + real* dfMax_s, + int blockSize, + int base, + int curIdx, + int nextIdx, + int dimN, + real* max) { dfMax_s[base] = -1.0e20; while (curIdx < dimN) { if (dfMax_s[base] < I[nextIdx]) { @@ -78,25 +76,24 @@ void findMax(real* I, if (base < stride) { nextIdx = base + stride; if (dfMax_s[base] < dfMax_s[nextIdx]) { - dfMax_s[base] = dfMax_s[nextIdx]; + dfMax_s[base] = dfMax_s[nextIdx]; } } } - if (0 == base) { + if (0 == base) { max[0] = dfMax_s[0]; } __syncthreads(); } -__device__ __forceinline__ -void subMaxAndExp(real* I, - real* O, - int curIdx, - int nextIdx, - int blockSize, - int dimN, - real max) { +__device__ __forceinline__ void subMaxAndExp(real* I, + real* O, + int curIdx, + int nextIdx, + int blockSize, + int dimN, + real max) { real val; while (curIdx < dimN) { val = I[nextIdx] - max; @@ -115,14 +112,13 @@ void subMaxAndExp(real* I, __syncthreads(); } -__device__ __forceinline__ -void valueSum(real* O, - real* dfMax_s, - int blockSize, - int base, - int curIdx, - int nextIdx, - int dimN) { +__device__ __forceinline__ void valueSum(real* O, + real* dfMax_s, + int blockSize, + int base, + int curIdx, + int nextIdx, + int dimN) { dfMax_s[base] = 0; while (curIdx < dimN) { dfMax_s[base] += O[nextIdx]; @@ -141,13 +137,8 @@ void valueSum(real* O, __syncthreads(); } -__device__ __forceinline__ -void divSum(real* O, - real sum, - int curIdx, - int nextIdx, - int blockSize, - int dimN) { +__device__ __forceinline__ void divSum( + real* O, real sum, int curIdx, int nextIdx, int blockSize, int dimN) { while (curIdx < dimN) { O[nextIdx] /= sum; nextIdx += blockSize; @@ -155,20 +146,18 @@ void divSum(real* O, } } -__device__ __forceinline__ -void softmax(real* I, - real* O, - real* dfMax_s, - int blockSize, - int base, - int curIdx, - int nextIdx, - int dimN) { +__device__ __forceinline__ void softmax(real* I, + real* O, + real* dfMax_s, + int blockSize, + int base, + int curIdx, + int nextIdx, + int dimN) { __shared__ real max; // find the max number - findMax(I, dfMax_s, blockSize, base, curIdx, - nextIdx, dimN, &max); + findMax(I, dfMax_s, blockSize, base, curIdx, nextIdx, dimN, &max); // sub max Value and do Exp operation subMaxAndExp(I, O, base, nextIdx, blockSize, dimN, max); @@ -181,8 +170,8 @@ void softmax(real* I, divSum(O, dfMax_s[0], curIdx, nextIdx, blockSize, dimN); } -template -__global__ void KeMatrixSoftMax(real *O, real *I, int dimN) { +template +__global__ void KeMatrixSoftMax(real* O, real* I, int dimN) { int base = threadIdx.x; __shared__ real dfMax_s[blockSize]; int nextIdx = blockIdx.x * dimN + base; @@ -191,19 +180,18 @@ __global__ void KeMatrixSoftMax(real *O, real *I, int dimN) { softmax(I, O, dfMax_s, blockSize, base, curIdx, nextIdx, dimN); } -void hl_matrix_softmax(real *A_d, real *C_d, int dimM, int dimN) { +void hl_matrix_softmax(real* A_d, real* C_d, int dimM, int dimN) { CHECK_NOTNULL(A_d); CHECK_NOTNULL(C_d); dim3 block(512, 1); dim3 grid(dimM, 1); - KeMatrixSoftMax<512> - <<>>(C_d, A_d, dimN); + KeMatrixSoftMax<512><<>>(C_d, A_d, dimN); CHECK_SYNC("hl_matrix_softmax failed"); } -template -__global__ void KeSequenceSoftMax(real *O, real *I, const int* index) { +template +__global__ void KeSequenceSoftMax(real* O, real* I, const int* index) { int base = threadIdx.x; int bid = blockIdx.x; __shared__ real dfMax_s[blockSize]; @@ -217,8 +205,8 @@ __global__ void KeSequenceSoftMax(real *O, real *I, const int* index) { softmax(I, O, dfMax_s, blockSize, base, curIdx, nextIdx, dimN); } -void hl_sequence_softmax_forward(real *A_d, - real *C_d, +void hl_sequence_softmax_forward(real* A_d, + real* C_d, const int* index, int numSequence) { CHECK_NOTNULL(A_d); @@ -226,59 +214,48 @@ void hl_sequence_softmax_forward(real *A_d, dim3 block(512, 1); dim3 grid(numSequence, 1); - KeSequenceSoftMax<512> - <<>>(C_d, A_d, index); + KeSequenceSoftMax<512><<>>(C_d, A_d, index); CHECK_SYNC("hl_sequence_softmax_forward failed"); } -__global__ void KeMatrixDerivative(real *grad_d, - real *output_d, - real *sftmaxSum_d, - int dimM, - int dimN) { - int rowIdx = blockIdx.x*blockDim.x + threadIdx.x; - int colIdx = blockIdx.y*blockDim.y + threadIdx.y; +__global__ void KeMatrixDerivative( + real* grad_d, real* output_d, real* sftmaxSum_d, int dimM, int dimN) { + int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; + int colIdx = blockIdx.y * blockDim.y + threadIdx.y; int index; if (rowIdx < dimM && colIdx < dimN) { - index = rowIdx*dimN + colIdx; + index = rowIdx * dimN + colIdx; grad_d[index] = output_d[index] * (grad_d[index] - sftmaxSum_d[rowIdx]); } } -void hl_matrix_softmax_derivative(real *grad_d, - real *output_d, - real *sftmaxSum_d, - int dimM, - int dimN) { +void hl_matrix_softmax_derivative( + real* grad_d, real* output_d, real* sftmaxSum_d, int dimM, int dimN) { CHECK_NOTNULL(grad_d); CHECK_NOTNULL(output_d); CHECK_NOTNULL(sftmaxSum_d); int blocksX = (dimM + 0) / 1; - int blocksY = (dimN + 1024 -1) / 1024; + int blocksY = (dimN + 1024 - 1) / 1024; dim3 threads(1, 1024); dim3 grid(blocksX, blocksY); - KeMatrixDerivative<<< grid, threads, 0, STREAM_DEFAULT >>> - (grad_d, output_d, sftmaxSum_d, dimM, dimN); + KeMatrixDerivative<<>>( + grad_d, output_d, sftmaxSum_d, dimM, dimN); CHECK_SYNC("hl_matrix_softmax_derivative failed"); } -__global__ void KeMatrixMultiBinaryCrossEntropy(real* output, - real* entropy, - int* row, - int* col, - int dimM, - int dimN) { +__global__ void KeMatrixMultiBinaryCrossEntropy( + real* output, real* entropy, int* row, int* col, int dimM, int dimN) { int index = blockIdx.x * blockDim.x + threadIdx.x; if (index < dimM) { - for (int i = 0; i < dimN; i ++) { + for (int i = 0; i < dimN; i++) { entropy[index] -= log(1 - output[index * dimN + i]); } - int *row_col = col + row[index]; + int* row_col = col + row[index]; int col_num = row[index + 1] - row[index]; - for (int i = 0; i < col_num; i ++) { + for (int i = 0; i < col_num; i++) { real o = output[index * dimN + row_col[i]]; entropy[index] -= log(o / (1 - o)); } @@ -299,37 +276,30 @@ void hl_matrix_multi_binary_cross_entropy(real* output, dim3 threads(n_threads); dim3 grid(blocks); hl_csr_matrix mat = (hl_csr_matrix)(csr_mat->matrix); - KeMatrixMultiBinaryCrossEntropy<<< grid, threads, 0, STREAM_DEFAULT >>> - (output, entropy, mat->csr_row, mat->csr_col, dimM, dimN); + KeMatrixMultiBinaryCrossEntropy<<>>( + output, entropy, mat->csr_row, mat->csr_col, dimM, dimN); CHECK_SYNC("hl_matrix_multi_binary_cross_entropy failed"); } -__global__ void KeMatrixMultiBinaryCrossEntropyBp(real* output, - real* grad, - int* row, - int* col, - int dimM, - int dimN) { +__global__ void KeMatrixMultiBinaryCrossEntropyBp( + real* output, real* grad, int* row, int* col, int dimM, int dimN) { int row_idx = blockIdx.x * blockDim.x + threadIdx.x; if (row_idx < dimM) { - for (int i = 0; i < dimN; i ++) { + for (int i = 0; i < dimN; i++) { int index = row_idx * dimN + i; grad[index] += 1.0 / (1 - output[index]); } int col_num = row[row_idx + 1] - row[row_idx]; - int *row_col = col + row[row_idx]; - for (int i = 0; i < col_num; i ++) { + int* row_col = col + row[row_idx]; + for (int i = 0; i < col_num; i++) { int index = row_idx * dimN + row_col[i]; grad[index] -= 1.0 / (output[index] * (1 - output[index])); } } } -void hl_matrix_multi_binary_cross_entropy_bp(real* output, - real* grad, - hl_sparse_matrix_s csr_mat, - int dimM, - int dimN) { +void hl_matrix_multi_binary_cross_entropy_bp( + real* output, real* grad, hl_sparse_matrix_s csr_mat, int dimM, int dimN) { CHECK_NOTNULL(output); CHECK_NOTNULL(grad); CHECK_NOTNULL(csr_mat); @@ -339,16 +309,13 @@ void hl_matrix_multi_binary_cross_entropy_bp(real* output, dim3 threads(n_threads); dim3 grid(blocks); hl_csr_matrix mat = (hl_csr_matrix)(csr_mat->matrix); - KeMatrixMultiBinaryCrossEntropyBp<<< grid, threads, 0, STREAM_DEFAULT >>> - (output, grad, mat->csr_row, mat->csr_col, dimM, dimN); + KeMatrixMultiBinaryCrossEntropyBp<<>>( + output, grad, mat->csr_row, mat->csr_col, dimM, dimN); CHECK_SYNC("hl_matrix_multi_binary_cross_entropy_bp failed"); } -__global__ void KeMatrixCrossEntropy(real* O, - real* E, - int* label, - int dimM, - int dimN) { +__global__ void KeMatrixCrossEntropy( + real* O, real* E, int* label, int dimM, int dimN) { int index = blockIdx.x * blockDim.x + threadIdx.x; int newBase; if (index < dimM) { @@ -358,59 +325,49 @@ __global__ void KeMatrixCrossEntropy(real* O, } } -void hl_matrix_cross_entropy(real* A_d, - real* C_d, - int* label_d, - int dimM, - int dimN) { +void hl_matrix_cross_entropy( + real* A_d, real* C_d, int* label_d, int dimM, int dimN) { CHECK_NOTNULL(A_d); CHECK_NOTNULL(C_d); int blocks = (dimM + 1024 - 1) / 1024; dim3 threads(1024, 1); dim3 grid(blocks, 1); - KeMatrixCrossEntropy<<< grid, threads, 0, STREAM_DEFAULT >>> - (A_d, C_d, label_d, dimM, dimN); + KeMatrixCrossEntropy<<>>( + A_d, C_d, label_d, dimM, dimN); CHECK_SYNC("hl_matrix_cross_entropy failed"); } -__global__ void KeMatrixCrossEntropyBp(real* grad_d, - real* output_d, - int* label_d, - int dimM, - int dimN) { - int rowIdx = blockIdx.x*blockDim.x + threadIdx.x; - int colIdx = blockIdx.y*blockDim.y + threadIdx.y; +__global__ void KeMatrixCrossEntropyBp( + real* grad_d, real* output_d, int* label_d, int dimM, int dimN) { + int rowIdx = blockIdx.x * blockDim.x + threadIdx.x; + int colIdx = blockIdx.y * blockDim.y + threadIdx.y; int index; if (rowIdx < dimM && colIdx < dimN) { - index = rowIdx*dimN + colIdx; + index = rowIdx * dimN + colIdx; if (label_d[rowIdx] == colIdx) { grad_d[index] -= 1.0f / output_d[index]; } } } -void hl_matrix_cross_entropy_bp(real* grad_d, - real* output_d, - int* label_d, - int dimM, - int dimN) { +void hl_matrix_cross_entropy_bp( + real* grad_d, real* output_d, int* label_d, int dimM, int dimN) { CHECK_NOTNULL(grad_d); CHECK_NOTNULL(output_d); CHECK_NOTNULL(label_d); - int blocksX = (dimM + 0)/1; - int blocksY = (dimN + 1024 -1) / 1024; + int blocksX = (dimM + 0) / 1; + int blocksY = (dimN + 1024 - 1) / 1024; dim3 threads(1, 1024); dim3 grid(blocksX, blocksY); - KeMatrixCrossEntropyBp<<< grid, threads, 0, STREAM_DEFAULT >>> - (grad_d, output_d, label_d, dimM, dimN); + KeMatrixCrossEntropyBp<<>>( + grad_d, output_d, label_d, dimM, dimN); CHECK_SYNC("hl_matrix_cross_entropy_bp failed"); } void hl_matrix_zero_mem(real* data, int num) { - hl_gpu_apply_unary_op( - unary::Zero(), data, 1, num, num); + hl_gpu_apply_unary_op(unary::Zero(), data, 1, num, num); } __global__ void KeParamReluForward(real* output, @@ -423,8 +380,8 @@ __global__ void KeParamReluForward(real* output, int ty = blockIdx.y * blockDim.y + threadIdx.y; if (tx < width && ty < height) { int index = ty * width + tx; - output[index] = input[index] > 0 ? input[index] : - input[index] * w[tx / partial_sum]; + output[index] = + input[index] > 0 ? input[index] : input[index] * w[tx / partial_sum]; } } @@ -439,14 +396,14 @@ void hl_param_relu_forward(real* output, CHECK_NOTNULL(w); dim3 threads(16, 16); int blockX = (width + 16 - 1) / 16; - int blockY = (height + 16 -1) / 16; + int blockY = (height + 16 - 1) / 16; dim3 grid(blockX, blockY); - KeParamReluForward<<>> - (output, input, w, width, height, partial_sum); + KeParamReluForward<<>>( + output, input, w, width, height, partial_sum); CHECK_SYNC("hl_param_relu_forward failed"); } -template +template __global__ void KeParamReluBackWardW(real* grad_w, real* grad_o, real* input, @@ -491,8 +448,8 @@ void hl_param_relu_backward_w(real* grad_w, int grid_num = width / partial_sum; dim3 threads(blockSize, 1); dim3 grid(grid_num, 1); - KeParamReluBackWardW<<>> - (grad_w, grad_o, input, width, height, partial_sum); + KeParamReluBackWardW<<>>( + grad_w, grad_o, input, width, height, partial_sum); CHECK_SYNC("hl_param_relu_backward_w failed"); } @@ -524,19 +481,15 @@ void hl_param_relu_backward_diff(real* grad_o, CHECK_NOTNULL(diff); dim3 threads(16, 16); int blockX = (width + 16 - 1) / 16; - int blockY = (height + 16 -1) / 16; + int blockY = (height + 16 - 1) / 16; dim3 grid(blockX, blockY); - KeParamReluBackwardDiff<<>> - (grad_o, data, w, diff, width, height, partial_sum); + KeParamReluBackwardDiff<<>>( + grad_o, data, w, diff, width, height, partial_sum); CHECK_SYNC("hl_param_relu_backward_diff failed"); } -__global__ void KeMatrixAddSharedBias(real* A, - real* B, - const int channel, - const int M, - const int N, - real scale) { +__global__ void KeMatrixAddSharedBias( + real* A, real* B, const int channel, const int M, const int N, real scale) { int index = blockIdx.x * blockDim.x + threadIdx.x; int dim = N / channel; if (index < M * N) { @@ -554,15 +507,14 @@ void hl_matrix_add_shared_bias(real* A_d, real scale) { const int blocks = 512; const int grids = DIVUP(dimM * dimN, blocks); - KeMatrixAddSharedBias<<>> - (A_d, B_d, channel, dimM, dimN, scale); + KeMatrixAddSharedBias<<>>( + A_d, B_d, channel, dimM, dimN, scale); CHECK_SYNC("hl_matrix_add_shared_bias failed"); } - template -__global__ void KeMatrixCollectSharedBias(real *B, - real *A, +__global__ void KeMatrixCollectSharedBias(real* B, + real* A, const int channel, const int M, const int N, @@ -589,7 +541,7 @@ __global__ void KeMatrixCollectSharedBias(real *B, int n = j * blockSize + tid; int m = n / dim; int w = n % dim; - smem[tid] = (m < M && w < dim) ? A[m * N + bid * dim + w] : 0.0; + smem[tid] = (m < M && w < dim) ? A[m * N + bid * dim + w] : 0.0; __syncthreads(); simpleReduce(smem, tid, blockSize); sum += smem[0]; @@ -611,33 +563,32 @@ void hl_matrix_collect_shared_bias(real* B_d, const int limit = 64; int grids = (dimM * dim) < limit ? DIVUP(channel, blocks) : channel; - KeMatrixCollectSharedBias - <<< grids, blocks, 0, STREAM_DEFAULT>>> - (B_d, A_d, channel, dimM, dimN, dim, limit, scale); + KeMatrixCollectSharedBias<<>>( + B_d, A_d, channel, dimM, dimN, dim, limit, scale); CHECK_SYNC("hl_matrix_collect_shared_bias failed"); } -__global__ void keMatrixRotate(real* mat, real* matRot, - int dimM, int dimN, bool clockWise) { - int idx = blockIdx.x * blockDim.x + threadIdx.x; - if (idx < dimM * dimN) { - int i = idx / dimN; - int j = idx % dimN; - if (clockWise) { - matRot[j * dimM + i] = mat[(dimM - i - 1) * dimN + j]; - } else { - matRot[j * dimM + i] = mat[i * dimN + (dimN - j - 1)]; - } +__global__ void keMatrixRotate( + real* mat, real* matRot, int dimM, int dimN, bool clockWise) { + int idx = blockIdx.x * blockDim.x + threadIdx.x; + if (idx < dimM * dimN) { + int i = idx / dimN; + int j = idx % dimN; + if (clockWise) { + matRot[j * dimM + i] = mat[(dimM - i - 1) * dimN + j]; + } else { + matRot[j * dimM + i] = mat[i * dimN + (dimN - j - 1)]; } + } } -void hl_matrix_rotate(real *mat, real* matRot, - int dimM, int dimN, bool clockWise) { - CHECK_NOTNULL(mat); - CHECK_NOTNULL(matRot); - const int threads = 512; - const int blocks = DIVUP(dimM * dimN, threads); - keMatrixRotate<<< blocks, threads, 0, STREAM_DEFAULT >>> - (mat, matRot, dimM, dimN, clockWise); - CHECK_SYNC("hl_matrix_rotate failed"); +void hl_matrix_rotate( + real* mat, real* matRot, int dimM, int dimN, bool clockWise) { + CHECK_NOTNULL(mat); + CHECK_NOTNULL(matRot); + const int threads = 512; + const int blocks = DIVUP(dimM * dimN, threads); + keMatrixRotate<<>>( + mat, matRot, dimM, dimN, clockWise); + CHECK_SYNC("hl_matrix_rotate failed"); } diff --git a/paddle/cuda/src/hl_cuda_sequence.cu b/paddle/cuda/src/hl_cuda_sequence.cu index eeee921db54e20ea6a017d2b83f2d7ca9e5e037e..c52780dfcaff6e5b94d3568fac4ca011b76a1442 100644 --- a/paddle/cuda/src/hl_cuda_sequence.cu +++ b/paddle/cuda/src/hl_cuda_sequence.cu @@ -16,36 +16,36 @@ limitations under the License. */ #include "hl_device_functions.cuh" #include "paddle/utils/Logging.h" -__global__ void KeMaxSequenceForward(real *input, - const int *sequence, +__global__ void KeMaxSequenceForward(real* input, + const int* sequence, real* output, - int *index, + int* index, int numSequences, int dim) { int dimIdx = threadIdx.x; int sequenceId = blockIdx.x; if (sequenceId >= numSequences) return; int start = sequence[sequenceId]; - int end = sequence[sequenceId+1]; + int end = sequence[sequenceId + 1]; for (int i = dimIdx; i < dim; i += blockDim.x) { real tmp = -HL_FLOAT_MAX; int tmpId = -1; for (int insId = start; insId < end; insId++) { - if (tmp < input[insId*dim + i]) { - tmp = input[insId*dim + i]; + if (tmp < input[insId * dim + i]) { + tmp = input[insId * dim + i]; tmpId = insId; } } - output[sequenceId*dim + i] = tmp; - index[sequenceId*dim + i] = tmpId; + output[sequenceId * dim + i] = tmp; + index[sequenceId * dim + i] = tmpId; } } void hl_max_sequence_forward(real* input, const int* sequence, real* output, - int *index, + int* index, int numSequences, int dim) { CHECK_NOTNULL(input); @@ -55,29 +55,23 @@ void hl_max_sequence_forward(real* input, dim3 threads(256, 1); dim3 grid(numSequences, 1); - KeMaxSequenceForward<<< grid, threads, 0, STREAM_DEFAULT >>> - (input, sequence, output, index, numSequences, dim); + KeMaxSequenceForward<<>>( + input, sequence, output, index, numSequences, dim); CHECK_SYNC("hl_max_sequence_forward failed"); } -__global__ void KeMaxSequenceBackward(real *outputGrad, - int *index, - real* inputGrad, - int numSequences, - int dim) { +__global__ void KeMaxSequenceBackward( + real* outputGrad, int* index, real* inputGrad, int numSequences, int dim) { int idx = threadIdx.x + blockIdx.x * blockDim.x; int colIdx = idx % dim; - if (idx < numSequences*dim) { + if (idx < numSequences * dim) { int insId = index[idx]; inputGrad[insId * dim + colIdx] += outputGrad[idx]; } } -void hl_max_sequence_backward(real* outputGrad, - int *index, - real* inputGrad, - int numSequences, - int dim) { +void hl_max_sequence_backward( + real* outputGrad, int* index, real* inputGrad, int numSequences, int dim) { CHECK_NOTNULL(outputGrad); CHECK_NOTNULL(index); CHECK_NOTNULL(inputGrad); @@ -85,12 +79,12 @@ void hl_max_sequence_backward(real* outputGrad, unsigned int blocks = (numSequences * dim + 128 - 1) / 128; dim3 threads(128, 1); dim3 grid(blocks, 1); - KeMaxSequenceBackward<<< grid, threads, 0, STREAM_DEFAULT >>> - (outputGrad, index, inputGrad, numSequences, dim); + KeMaxSequenceBackward<<>>( + outputGrad, index, inputGrad, numSequences, dim); CHECK_SYNC("hl_max_sequence_backward failed"); } -template +template __global__ void KeMatrixAddRows(real* output, real* table, int* ids, @@ -104,8 +98,8 @@ __global__ void KeMatrixAddRows(real* output, while (sampleId < numSamples) { int tableId = ids[sampleId]; if ((0 <= tableId) && (tableId < tableSize)) { - real *outputData = output + sampleId * dim; - real *tableData = table + tableId * dim; + real* outputData = output + sampleId * dim; + real* tableData = table + tableId * dim; for (int i = idx; i < dim; i += blockDimX) { if (AddRow == 0) { outputData[i] += tableData[i]; @@ -114,24 +108,27 @@ __global__ void KeMatrixAddRows(real* output, } } } - sampleId += blockDimY*gridDimX; + sampleId += blockDimY * gridDimX; } } -template -__global__ -void KeSequence2Batch(real *batch, - real *sequence, - const int *batchIndex, - int seqWidth, - int batchCount) { +template +__global__ void KeSequence2Batch(real* batch, + real* sequence, + const int* batchIndex, + int seqWidth, + int batchCount) { int idx = threadIdx.x; int idy = threadIdx.y; int id = blockIdx.x + idy * gridDimX; while (id < batchCount) { int seqId = batchIndex[id]; - real* batchData = batch + id*seqWidth; - real* seqData = sequence + seqId*seqWidth; + real* batchData = batch + id * seqWidth; + real* seqData = sequence + seqId * seqWidth; for (int i = idx; i < seqWidth; i += blockDimX) { if (seq2batch) { if (isAdd) { @@ -147,13 +144,13 @@ void KeSequence2Batch(real *batch, } } } - id += blockDimY*gridDimX; + id += blockDimY * gridDimX; } } -void hl_sequence2batch_copy(real *batch, - real *sequence, - const int *batchIndex, +void hl_sequence2batch_copy(real* batch, + real* sequence, + const int* batchIndex, int seqWidth, int batchCount, bool seq2batch) { @@ -164,18 +161,18 @@ void hl_sequence2batch_copy(real *batch, dim3 threads(128, 8); dim3 grid(8, 1); if (seq2batch) { - KeSequence2Batch<128, 8, 8, 1, 0><<< grid, threads, 0, STREAM_DEFAULT >>> - (batch, sequence, batchIndex, seqWidth, batchCount); + KeSequence2Batch<128, 8, 8, 1, 0><<>>( + batch, sequence, batchIndex, seqWidth, batchCount); } else { - KeSequence2Batch<128, 8, 8, 0, 0><<< grid, threads, 0, STREAM_DEFAULT >>> - (batch, sequence, batchIndex, seqWidth, batchCount); + KeSequence2Batch<128, 8, 8, 0, 0><<>>( + batch, sequence, batchIndex, seqWidth, batchCount); } CHECK_SYNC("hl_sequence2batch_copy failed"); } -void hl_sequence2batch_add(real *batch, - real *sequence, - int *batchIndex, +void hl_sequence2batch_add(real* batch, + real* sequence, + int* batchIndex, int seqWidth, int batchCount, bool seq2batch) { @@ -186,23 +183,22 @@ void hl_sequence2batch_add(real *batch, dim3 threads(128, 8); dim3 grid(8, 1); if (seq2batch) { - KeSequence2Batch<128, 8, 8, 1, 1><<< grid, threads, 0, STREAM_DEFAULT >>> - (batch, sequence, batchIndex, seqWidth, batchCount); + KeSequence2Batch<128, 8, 8, 1, 1><<>>( + batch, sequence, batchIndex, seqWidth, batchCount); } else { - KeSequence2Batch<128, 8, 8, 0, 1><<< grid, threads, 0, STREAM_DEFAULT >>> - (batch, sequence, batchIndex, seqWidth, batchCount); + KeSequence2Batch<128, 8, 8, 0, 1><<>>( + batch, sequence, batchIndex, seqWidth, batchCount); } CHECK_SYNC("hl_sequence2batch_add failed"); } -template -__global__ -void KeSequence2BatchPadding(real* batch, - real* sequence, - const int* sequenceStartPositions, - const size_t sequenceWidth, - const size_t maxSequenceLength, - const size_t numSequences) { +template +__global__ void KeSequence2BatchPadding(real* batch, + real* sequence, + const int* sequenceStartPositions, + const size_t sequenceWidth, + const size_t maxSequenceLength, + const size_t numSequences) { int batchIdx = blockIdx.y; int sequenceStart = sequenceStartPositions[batchIdx]; int sequenceLength = sequenceStartPositions[batchIdx + 1] - sequenceStart; @@ -276,37 +272,49 @@ void hl_sequence2batch_copy_padding(real* batch, if (seq2batch) { /* sequence -> batch */ if (normByTimes) { - KeSequence2BatchPadding<1, 1><<< grid, threads, 0, STREAM_DEFAULT >>>( - batch, sequence, sequenceStartPositions, - sequenceWidth, maxSequenceLength, numSequences); + KeSequence2BatchPadding<1, 1><<>>( + batch, + sequence, + sequenceStartPositions, + sequenceWidth, + maxSequenceLength, + numSequences); } else { - KeSequence2BatchPadding<0, 1><<< grid, threads, 0, STREAM_DEFAULT >>>( - batch, sequence, sequenceStartPositions, - sequenceWidth, maxSequenceLength, numSequences); + KeSequence2BatchPadding<0, 1><<>>( + batch, + sequence, + sequenceStartPositions, + sequenceWidth, + maxSequenceLength, + numSequences); } } else { /* batch -> sequence */ if (normByTimes) { - KeSequence2BatchPadding<1, 0><<< grid, threads, 0, STREAM_DEFAULT >>>( - batch, sequence, sequenceStartPositions, - sequenceWidth, maxSequenceLength, numSequences); + KeSequence2BatchPadding<1, 0><<>>( + batch, + sequence, + sequenceStartPositions, + sequenceWidth, + maxSequenceLength, + numSequences); } else { - KeSequence2BatchPadding<0, 0><<< grid, threads, 0, STREAM_DEFAULT >>>( - batch, sequence, sequenceStartPositions, - sequenceWidth, maxSequenceLength, numSequences); + KeSequence2BatchPadding<0, 0><<>>( + batch, + sequence, + sequenceStartPositions, + sequenceWidth, + maxSequenceLength, + numSequences); } } CHECK_SYNC("hl_sequence2batch_copy_padding failed"); } -__device__ inline float my_rsqrt(float x) { - return rsqrtf(x); -} +__device__ inline float my_rsqrt(float x) { return rsqrtf(x); } -__device__ inline double my_rsqrt(double x) { - return rsqrt(x); -} +__device__ inline double my_rsqrt(double x) { return rsqrt(x); } __global__ void KeSequenceAvgForward(real* dst, real* src, @@ -327,8 +335,8 @@ __global__ void KeSequenceAvgForward(real* dst, for (int i = start; i < end; i++) { sum += src[i * width + col]; } - sum = mode == 1 ? sum : - (mode == 0 ? sum / seqLength : sum * my_rsqrt((real)seqLength)); + sum = mode == 1 ? sum : (mode == 0 ? sum / seqLength + : sum * my_rsqrt((real)seqLength)); dst[gid] += sum; } } @@ -347,10 +355,10 @@ void hl_sequence_avg_forward(real* dst, int grid = DIVUP(width * height, 512); CHECK(mode == 0 || mode == 1 || mode == 2) - << "mode error in hl_sequence_avg_forward!"; + << "mode error in hl_sequence_avg_forward!"; - KeSequenceAvgForward<<< grid, block, 0, STREAM_DEFAULT >>> - (dst, src, starts, height, width, mode); + KeSequenceAvgForward<<>>( + dst, src, starts, height, width, mode); CHECK_SYNC("hl_sequence_avg_forward failed"); } @@ -370,8 +378,8 @@ __global__ void KeSequenceAvgBackward(real* dst, int seqLength = end - start; if (seqLength == 0) return; real grad = src[gid]; - grad = mode == 1 ? grad : - (mode == 0 ? grad / seqLength : grad * my_rsqrt((real)seqLength)); + grad = mode == 1 ? grad : (mode == 0 ? grad / seqLength + : grad * my_rsqrt((real)seqLength)); for (int i = start; i < end; i++) { dst[i * width + col] += grad; } @@ -392,9 +400,9 @@ void hl_sequence_avg_backward(real* dst, int grid = DIVUP(width * height, 512); CHECK(mode == 0 || mode == 1 || mode == 2) - << "mode error in hl_sequence_avg_backward!"; + << "mode error in hl_sequence_avg_backward!"; - KeSequenceAvgBackward<<< grid, block, 0, STREAM_DEFAULT >>> - (dst, src, starts, height, width, mode); + KeSequenceAvgBackward<<>>( + dst, src, starts, height, width, mode); CHECK_SYNC("hl_sequence_avg_backward failed"); } diff --git a/paddle/cuda/src/hl_cuda_sparse.cu b/paddle/cuda/src/hl_cuda_sparse.cu index ab9ab57c884137f117c25c2752b5603b2e8b7135..6351e7e01ee55b6303a6e48bc9ebf9834a83130e 100644 --- a/paddle/cuda/src/hl_cuda_sparse.cu +++ b/paddle/cuda/src/hl_cuda_sparse.cu @@ -12,13 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ - #include "hl_cuda.h" +#include "hl_cuda_sparse.cuh" +#include "hl_matrix_apply.cuh" +#include "hl_matrix_ops.cuh" #include "hl_sparse.h" #include "hl_sparse.ph" -#include "hl_matrix_ops.cuh" -#include "hl_matrix_apply.cuh" -#include "hl_cuda_sparse.cuh" #include "paddle/utils/Logging.h" DEFINE_MATRIX_UNARY_PARAMETER_OP(mul_scalar, ONE_PARAMETER, a = a * p); @@ -34,15 +33,15 @@ void hl_matrix_csr2dense(hl_sparse_matrix_s A_d, CHECK(A_d->format == HL_SPARSE_CSR) << "matrix format error!"; if (A_d->nnz == 0) { - hl_gpu_apply_unary_op( - unary::Zero(), C_d, dimM, dimN, dimN); + hl_gpu_apply_unary_op(unary::Zero(), C_d, dimM, dimN, dimN); return; } /* nnz != 0 */ hl_csr_matrix A_d2 = (hl_csr_matrix)(A_d->matrix); - CHECK((A_d2->csr_val || A_d->type == HL_NO_VALUE) && - A_d2->csr_row && A_d2->csr_col) << "parameter transa error!"; + CHECK((A_d2->csr_val || A_d->type == HL_NO_VALUE) && A_d2->csr_row && + A_d2->csr_col) + << "parameter transa error!"; int blocksX = (dimN + CU_CSR2DENSE_THREAD_X - 1) / CU_CSR2DENSE_THREAD_X; int blocksY = (dimM + CU_CSR2DENSE_THREAD_X - 1) / CU_CSR2DENSE_THREAD_X; @@ -50,21 +49,11 @@ void hl_matrix_csr2dense(hl_sparse_matrix_s A_d, dim3 grid(blocksX, blocksY); if (A_d->type == HL_NO_VALUE) { - KeSMatrixCsr2Dense<0> - <<>>(A_d2->csr_val, - A_d2->csr_row, - A_d2->csr_col, - C_d, - dimM, - dimN); + KeSMatrixCsr2Dense<0><<>>( + A_d2->csr_val, A_d2->csr_row, A_d2->csr_col, C_d, dimM, dimN); } else if (A_d->type == HL_FLOAT_VALUE) { - KeSMatrixCsr2Dense<1> - <<>>(A_d2->csr_val, - A_d2->csr_row, - A_d2->csr_col, - C_d, - dimM, - dimN); + KeSMatrixCsr2Dense<1><<>>( + A_d2->csr_val, A_d2->csr_row, A_d2->csr_col, C_d, dimM, dimN); } else { } CHECK_SYNC("hl_matrix_csr2dense failed"); @@ -80,15 +69,15 @@ void hl_matrix_csc2dense(hl_sparse_matrix_s A_d, CHECK(A_d->format == HL_SPARSE_CSC) << "matrix format error!"; if (A_d->nnz == 0) { - hl_gpu_apply_unary_op( - unary::Zero(), C_d, dimM, dimN, dimN); + hl_gpu_apply_unary_op(unary::Zero(), C_d, dimM, dimN, dimN); return; } /* nnz != 0 */ hl_csc_matrix A_d2 = (hl_csc_matrix)(A_d->matrix); - CHECK((A_d2->csc_val || A_d->type == HL_NO_VALUE) && - A_d2->csc_row && A_d2->csc_col) << "parameter transa error!"; + CHECK((A_d2->csc_val || A_d->type == HL_NO_VALUE) && A_d2->csc_row && + A_d2->csc_col) + << "parameter transa error!"; int blocksX = (dimN + CU_CSR2DENSE_THREAD_X - 1) / CU_CSR2DENSE_THREAD_X; int blocksY = (dimM + CU_CSR2DENSE_THREAD_X - 1) / CU_CSR2DENSE_THREAD_X; @@ -96,21 +85,11 @@ void hl_matrix_csc2dense(hl_sparse_matrix_s A_d, dim3 grid(blocksX, blocksY); if (A_d->type == HL_NO_VALUE) { - KeSMatrixCsc2Dense<0> - <<>>(A_d2->csc_val, - A_d2->csc_row, - A_d2->csc_col, - C_d, - dimM, - dimN); + KeSMatrixCsc2Dense<0><<>>( + A_d2->csc_val, A_d2->csc_row, A_d2->csc_col, C_d, dimM, dimN); } else if (A_d->type == HL_FLOAT_VALUE) { - KeSMatrixCsc2Dense<1> - <<>>(A_d2->csc_val, - A_d2->csc_row, - A_d2->csc_col, - C_d, - dimM, - dimN); + KeSMatrixCsc2Dense<1><<>>( + A_d2->csc_val, A_d2->csc_row, A_d2->csc_col, C_d, dimM, dimN); } else { } CHECK_SYNC("hl_matrix_csc2dense failed"); @@ -118,43 +97,43 @@ void hl_matrix_csc2dense(hl_sparse_matrix_s A_d, void hl_malloc_sparse_matrix(hl_sparse_matrix_s *A_d, hl_matrix_format_t format, - hl_matrix_value_t value_type, + hl_matrix_value_t value_type, int dimM, int dimN, int nnz) { CHECK_NOTNULL(A_d); CHECK(format == HL_SPARSE_CSR || format == HL_SPARSE_CSC) - << "sparse matrix format error!"; + << "sparse matrix format error!"; CHECK(value_type == HL_FLOAT_VALUE || value_type == HL_NO_VALUE) - << "sparse matrix value type error!"; + << "sparse matrix value type error!"; /* avoid malloc 0 bytes */ int nnz_s = (nnz == 0 ? 1 : nnz); if (format == HL_SPARSE_CSR) { CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; - char* tmp = (char*)malloc(sizeof(_hl_sparse_matrix_s) - + sizeof(_hl_csr_matrix)); + char *tmp = + (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csr_matrix)); CHECK_NOTNULL(tmp); - hl_csr_matrix csr = (hl_csr_matrix)(tmp+sizeof(_hl_sparse_matrix_s)); + hl_csr_matrix csr = (hl_csr_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); csr->sparsity = -1.0; if (value_type == HL_NO_VALUE) { csr->csr_val = NULL; csr->nnz_s = nnz_s; - csr->row_s = dimM+1; - csr->csr_row = (int*)hl_malloc_device((dimM+1)*sizeof(int)); - csr->csr_col = (int*)hl_malloc_device((nnz_s)*sizeof(int)); + csr->row_s = dimM + 1; + csr->csr_row = (int *)hl_malloc_device((dimM + 1) * sizeof(int)); + csr->csr_col = (int *)hl_malloc_device((nnz_s) * sizeof(int)); *A_d = (hl_sparse_matrix_s)tmp; (*A_d)->matrix = (hl_matrix_s)csr; } else if (value_type == HL_FLOAT_VALUE) { csr->nnz_s = nnz_s; - csr->row_s = dimM+1; - csr->csr_val = (real*)hl_malloc_device((nnz_s)*sizeof(real)); - csr->csr_row = (int*)hl_malloc_device((dimM+1)*sizeof(int)); - csr->csr_col = (int*)hl_malloc_device((nnz_s)*sizeof(int)); + csr->row_s = dimM + 1; + csr->csr_val = (real *)hl_malloc_device((nnz_s) * sizeof(real)); + csr->csr_row = (int *)hl_malloc_device((dimM + 1) * sizeof(int)); + csr->csr_col = (int *)hl_malloc_device((nnz_s) * sizeof(int)); *A_d = (hl_sparse_matrix_s)tmp; (*A_d)->matrix = (hl_matrix_s)csr; @@ -162,28 +141,28 @@ void hl_malloc_sparse_matrix(hl_sparse_matrix_s *A_d, } else if (format == HL_SPARSE_CSC) { CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; - char* tmp = (char*)malloc(sizeof(_hl_sparse_matrix_s) - + sizeof(_hl_csc_matrix)); + char *tmp = + (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csc_matrix)); CHECK_NOTNULL(tmp); - hl_csc_matrix csc = (hl_csc_matrix)(tmp+sizeof(_hl_sparse_matrix_s)); + hl_csc_matrix csc = (hl_csc_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); csc->sparsity = -1.0f; if (value_type == HL_NO_VALUE) { csc->csc_val = NULL; csc->nnz_s = nnz_s; - csc->col_s = dimN+1; - csc->csc_row = (int*)hl_malloc_device((nnz_s)*sizeof(int)); - csc->csc_col = (int*)hl_malloc_device((dimN+1)*sizeof(int)); + csc->col_s = dimN + 1; + csc->csc_row = (int *)hl_malloc_device((nnz_s) * sizeof(int)); + csc->csc_col = (int *)hl_malloc_device((dimN + 1) * sizeof(int)); *A_d = (hl_sparse_matrix_s)tmp; (*A_d)->matrix = (hl_matrix_s)csc; } else if (value_type == HL_FLOAT_VALUE) { csc->nnz_s = nnz_s; - csc->col_s = dimN+1; - csc->csc_val = (real*)hl_malloc_device((nnz_s)*sizeof(real)); - csc->csc_row = (int*)hl_malloc_device((nnz_s)*sizeof(int)); - csc->csc_col = (int*)hl_malloc_device((dimN+1)*sizeof(int)); + csc->col_s = dimN + 1; + csc->csc_val = (real *)hl_malloc_device((nnz_s) * sizeof(real)); + csc->csc_row = (int *)hl_malloc_device((nnz_s) * sizeof(int)); + csc->csc_col = (int *)hl_malloc_device((dimN + 1) * sizeof(int)); *A_d = (hl_sparse_matrix_s)tmp; (*A_d)->matrix = (hl_matrix_s)csc; @@ -200,7 +179,7 @@ void hl_malloc_sparse_matrix(hl_sparse_matrix_s *A_d, void hl_free_sparse_matrix(hl_sparse_matrix_s A_d) { CHECK_NOTNULL(A_d); CHECK(A_d->format == HL_SPARSE_CSR || A_d->format == HL_SPARSE_CSC) - << "sparse matrix format error!"; + << "sparse matrix format error!"; if (A_d->matrix == NULL) { free(A_d); @@ -249,77 +228,77 @@ void hl_free_sparse_matrix(hl_sparse_matrix_s A_d) { } void hl_construct_sparse_matrix(hl_sparse_matrix_s *A_d, - void * dest_d, + void *dest_d, size_t size, hl_matrix_format_t format, - hl_matrix_value_t value_type, + hl_matrix_value_t value_type, int dimM, int dimN, int nnz) { CHECK_NOTNULL(A_d); CHECK(format == HL_SPARSE_CSR || format == HL_SPARSE_CSC) - << "sparse matrix format error!"; + << "sparse matrix format error!"; if (format == HL_SPARSE_CSR) { CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; - size_t size_ = (dimM+1)*sizeof(int) + nnz*sizeof(int); + size_t size_ = (dimM + 1) * sizeof(int) + nnz * sizeof(int); if (value_type != HL_NO_VALUE) { - size_ += nnz*sizeof(real); + size_ += nnz * sizeof(real); } CHECK_LE(size_, size) << "dest_d size(" << size - << ") too small, should bigger than(" << size_ << ")!"; + << ") too small, should bigger than(" << size_ + << ")!"; - char* tmp = (char*)malloc(sizeof(_hl_sparse_matrix_s) - + sizeof(_hl_csr_matrix)); + char *tmp = + (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csr_matrix)); CHECK_NOTNULL(tmp); - hl_csr_matrix csr = (hl_csr_matrix)(tmp+sizeof(_hl_sparse_matrix_s)); + hl_csr_matrix csr = (hl_csr_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); if (value_type == HL_NO_VALUE) { csr->csr_val = NULL; - csr->csr_row = (int*)dest_d; - csr->csr_col = (int*)((char*)dest_d + (dimM+1)*sizeof(int)); + csr->csr_row = (int *)dest_d; + csr->csr_col = (int *)((char *)dest_d + (dimM + 1) * sizeof(int)); } else { - csr->csr_val = (real*)dest_d; - csr->csr_row = (int*)((char*)dest_d + nnz*sizeof(real)); - csr->csr_col = (int*)((char*)dest_d + - nnz*sizeof(real) + - (dimM+1)*sizeof(int)); + csr->csr_val = (real *)dest_d; + csr->csr_row = (int *)((char *)dest_d + nnz * sizeof(real)); + csr->csr_col = (int *)((char *)dest_d + nnz * sizeof(real) + + (dimM + 1) * sizeof(int)); } csr->nnz_s = nnz; - csr->row_s = dimM+1; + csr->row_s = dimM + 1; csr->sparsity = -1.0; *A_d = (hl_sparse_matrix_s)tmp; (*A_d)->matrix = (hl_matrix_s)csr; } else if (format == HL_SPARSE_CSC) { CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; - size_t size_ = (dimN+1)*sizeof(int) + nnz*sizeof(int); + size_t size_ = (dimN + 1) * sizeof(int) + nnz * sizeof(int); if (value_type != HL_NO_VALUE) { - size_ += nnz*sizeof(real); + size_ += nnz * sizeof(real); } CHECK_LE(size_, size) << "dest_d size(" << size - << ") too small, should bigger than(" << size_ << ")!"; + << ") too small, should bigger than(" << size_ + << ")!"; - char* tmp = (char*)malloc(sizeof(_hl_sparse_matrix_s) - + sizeof(_hl_csc_matrix)); + char *tmp = + (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csc_matrix)); CHECK_NOTNULL(tmp); - hl_csc_matrix csc = (hl_csc_matrix)(tmp+sizeof(_hl_sparse_matrix_s)); + hl_csc_matrix csc = (hl_csc_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); if (value_type == HL_NO_VALUE) { csc->csc_val = NULL; - csc->csc_col = (int*)dest_d; - csc->csc_row = (int*)((char*)dest_d + (dimN+1)*sizeof(int)); + csc->csc_col = (int *)dest_d; + csc->csc_row = (int *)((char *)dest_d + (dimN + 1) * sizeof(int)); } else { - csc->csc_val = (real*)dest_d; - csc->csc_col = (int*)((char*)dest_d + nnz*sizeof(real)); - csc->csc_row = (int*)((char*)dest_d + - nnz*sizeof(real) + - (dimN+1)*sizeof(int)); + csc->csc_val = (real *)dest_d; + csc->csc_col = (int *)((char *)dest_d + nnz * sizeof(real)); + csc->csc_row = (int *)((char *)dest_d + nnz * sizeof(real) + + (dimN + 1) * sizeof(int)); } csc->nnz_s = nnz; - csc->col_s = dimN+1; + csc->col_s = dimN + 1; csc->sparsity = -1.0f; *A_d = (hl_sparse_matrix_s)tmp; (*A_d)->matrix = (hl_matrix_s)csc; @@ -333,11 +312,11 @@ void hl_construct_sparse_matrix(hl_sparse_matrix_s *A_d, } void hl_construct_sparse_matrix(hl_sparse_matrix_s *A_d, - real* value_d, - int* rows_d, - int* cols_d, + real *value_d, + int *rows_d, + int *cols_d, hl_matrix_format_t format, - hl_matrix_value_t value_type, + hl_matrix_value_t value_type, int dimM, int dimN, int nnz) { @@ -345,11 +324,11 @@ void hl_construct_sparse_matrix(hl_sparse_matrix_s *A_d, CHECK(dimM > 0 && nnz >= 0) << "sparse matrix size error!"; CHECK(format == HL_SPARSE_CSR || format == HL_SPARSE_CSC) - << "sparse matrix format error!"; + << "sparse matrix format error!"; if (format == HL_SPARSE_CSR) { - char* tmp = (char*)malloc(sizeof(_hl_sparse_matrix_s) - + sizeof(_hl_csr_matrix)); + char *tmp = + (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csr_matrix)); CHECK_NOTNULL(tmp); hl_csr_matrix csr = (hl_csr_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); @@ -362,8 +341,8 @@ void hl_construct_sparse_matrix(hl_sparse_matrix_s *A_d, *A_d = (hl_sparse_matrix_s)tmp; (*A_d)->matrix = (hl_matrix_s)csr; } else if (format == HL_SPARSE_CSC) { - char* tmp = (char*)malloc(sizeof(_hl_sparse_matrix_s) - + sizeof(_hl_csc_matrix)); + char *tmp = + (char *)malloc(sizeof(_hl_sparse_matrix_s) + sizeof(_hl_csc_matrix)); CHECK_NOTNULL(tmp); hl_csc_matrix csc = (hl_csc_matrix)(tmp + sizeof(_hl_sparse_matrix_s)); @@ -396,35 +375,30 @@ void hl_memcpy_csr_matrix(hl_sparse_matrix_s csr_matrix, hl_stream_t stream) { CHECK_NOTNULL(csr_matrix); CHECK_EQ(csr_matrix->format, HL_SPARSE_CSR) - << "csr_matrix is not csr format!"; + << "csr_matrix is not csr format!"; CHECK_NOTNULL(csr_matrix->matrix); hl_csr_matrix csr = (hl_csr_matrix)(csr_matrix->matrix); - CHECK_LE(csr_matrix->nnz, csr->nnz_s) - << "copy size " << csr_matrix->nnz - << " is big than alloc size " << csr->nnz_s; + CHECK_LE(csr_matrix->nnz, csr->nnz_s) << "copy size " << csr_matrix->nnz + << " is big than alloc size " + << csr->nnz_s; - CHECK_LE((csr_matrix->rows+1), csr->row_s) - << "copy size " << (csr_matrix->rows + 1) - << " is big than alloc size " << csr->row_s; + CHECK_LE((csr_matrix->rows + 1), csr->row_s) + << "copy size " << (csr_matrix->rows + 1) << " is big than alloc size " + << csr->row_s; - CHECK(csr_matrix->type == HL_FLOAT_VALUE || - csr_matrix->type == HL_NO_VALUE) - << "sparse matrix value type error!"; + CHECK(csr_matrix->type == HL_FLOAT_VALUE || csr_matrix->type == HL_NO_VALUE) + << "sparse matrix value type error!"; if (csr_matrix->type == HL_NO_VALUE) { if (csr_row == NULL && csr_col == NULL) { return; } else if (csr_row != NULL && csr_col != NULL) { - hl_memcpy_async(csr->csr_row, - csr_row, - (csr_matrix->rows+1)*sizeof(int), - stream); + hl_memcpy_async( + csr->csr_row, csr_row, (csr_matrix->rows + 1) * sizeof(int), stream); - hl_memcpy_async(csr->csr_col, - csr_col, - (csr_matrix->nnz)*sizeof(int), - stream); + hl_memcpy_async( + csr->csr_col, csr_col, (csr_matrix->nnz) * sizeof(int), stream); } else { LOG(FATAL) << "parameter csr_row or csr_col is null pointer!"; } @@ -432,30 +406,21 @@ void hl_memcpy_csr_matrix(hl_sparse_matrix_s csr_matrix, if (csr_val == NULL && csr_row == NULL && csr_col == NULL) { return; } else if (csr_val != NULL && csr_row == NULL && csr_col == NULL) { - hl_memcpy_async(csr->csr_val, - csr_val, - (csr_matrix->nnz)*sizeof(real), - stream); + hl_memcpy_async( + csr->csr_val, csr_val, (csr_matrix->nnz) * sizeof(real), stream); } else if (csr_val != NULL && csr_row != NULL && csr_col != NULL) { - hl_memcpy_async(csr->csr_val, - csr_val, - (csr_matrix->nnz)*sizeof(real), - stream); - hl_memcpy_async(csr->csr_row, - csr_row, - (csr_matrix->rows+1)*sizeof(int), - stream); - hl_memcpy_async(csr->csr_col, - csr_col, - (csr_matrix->nnz)*sizeof(int), - stream); + hl_memcpy_async( + csr->csr_val, csr_val, (csr_matrix->nnz) * sizeof(real), stream); + hl_memcpy_async( + csr->csr_row, csr_row, (csr_matrix->rows + 1) * sizeof(int), stream); + hl_memcpy_async( + csr->csr_col, csr_col, (csr_matrix->nnz) * sizeof(int), stream); } else { LOG(FATAL) << "parameter csr_row or csr_col is null pointer!"; } } - csr->sparsity = ((float)csr_matrix->nnz) / - ((float)csr_matrix->rows) / + csr->sparsity = ((float)csr_matrix->nnz) / ((float)csr_matrix->rows) / ((float)csr_matrix->cols); } @@ -466,33 +431,28 @@ void hl_memcpy_csc_matrix(hl_sparse_matrix_s csc_matrix, hl_stream_t stream) { CHECK_NOTNULL(csc_matrix); CHECK_EQ(csc_matrix->format, HL_SPARSE_CSC) - << "csc_matrix is not csc format error!"; + << "csc_matrix is not csc format error!"; hl_csc_matrix csc = (hl_csc_matrix)(csc_matrix->matrix); - CHECK_LE(csc_matrix->nnz, csc->nnz_s) - << "copy size " << csc_matrix->nnz - << " is big than alloc size " << csc->nnz_s; + CHECK_LE(csc_matrix->nnz, csc->nnz_s) << "copy size " << csc_matrix->nnz + << " is big than alloc size " + << csc->nnz_s; - CHECK_LE((csc_matrix->cols+1), csc->col_s) - << "copy size " <<(csc_matrix->cols + 1) - << " is big than alloc size " << csc->col_s; + CHECK_LE((csc_matrix->cols + 1), csc->col_s) + << "copy size " << (csc_matrix->cols + 1) << " is big than alloc size " + << csc->col_s; - CHECK(csc_matrix->type == HL_FLOAT_VALUE || - csc_matrix->type == HL_NO_VALUE) - << "sparse matrix value type error!"; + CHECK(csc_matrix->type == HL_FLOAT_VALUE || csc_matrix->type == HL_NO_VALUE) + << "sparse matrix value type error!"; if (csc_matrix->type == HL_NO_VALUE) { if (csc_row == NULL && csc_col == NULL) { return; } else if (csc_row != NULL && csc_col != NULL) { - hl_memcpy_async(csc->csc_row, - csc_row, - (csc_matrix->nnz)*sizeof(int), - stream); - hl_memcpy_async(csc->csc_col, - csc_col, - (csc_matrix->cols+1)*sizeof(int), - stream); + hl_memcpy_async( + csc->csc_row, csc_row, (csc_matrix->nnz) * sizeof(int), stream); + hl_memcpy_async( + csc->csc_col, csc_col, (csc_matrix->cols + 1) * sizeof(int), stream); } else { LOG(FATAL) << "parameter csc_row or csc_col is null pointer!"; } @@ -500,30 +460,21 @@ void hl_memcpy_csc_matrix(hl_sparse_matrix_s csc_matrix, if (csc_val == NULL && csc_row == NULL && csc_col == NULL) { return; } else if (csc_val != NULL && csc_row == NULL && csc_col == NULL) { - hl_memcpy_async(csc->csc_val, - csc_val, - (csc_matrix->nnz)*sizeof(real), - stream); + hl_memcpy_async( + csc->csc_val, csc_val, (csc_matrix->nnz) * sizeof(real), stream); } else if (csc_val != NULL && csc_row != NULL && csc_col != NULL) { - hl_memcpy_async(csc->csc_val, - csc_val, - (csc_matrix->nnz)*sizeof(real), - stream); - hl_memcpy_async(csc->csc_row, - csc_row, - (csc_matrix->nnz)*sizeof(int), - stream); - hl_memcpy_async(csc->csc_col, - csc_col, - (csc_matrix->cols+1)*sizeof(int), - stream); + hl_memcpy_async( + csc->csc_val, csc_val, (csc_matrix->nnz) * sizeof(real), stream); + hl_memcpy_async( + csc->csc_row, csc_row, (csc_matrix->nnz) * sizeof(int), stream); + hl_memcpy_async( + csc->csc_col, csc_col, (csc_matrix->cols + 1) * sizeof(int), stream); } else { LOG(FATAL) << "parameter csc_row or csc_col is null pointer!"; } } - csc->sparsity = ((float)csc_matrix->nnz) / - ((float)csc_matrix->rows) / + csc->sparsity = ((float)csc_matrix->nnz) / ((float)csc_matrix->rows) / ((float)csc_matrix->cols); } @@ -531,32 +482,23 @@ void hl_memcpy_sparse_matrix(hl_sparse_matrix_s dst, hl_sparse_matrix_s src, hl_stream_t stream) { CHECK(dst && src && dst->matrix && src->matrix) - << "parameter dst or src is null pointer!"; - CHECK_EQ(dst->format, src->format) - << "sparse matrix format does not match!"; + << "parameter dst or src is null pointer!"; + CHECK_EQ(dst->format, src->format) << "sparse matrix format does not match!"; CHECK(dst->type != HL_FLOAT_VALUE || src->type != HL_NO_VALUE) - << "src sparse matrix is no value, dst sparse matrix has value!"; + << "src sparse matrix is no value, dst sparse matrix has value!"; if (dst->format == HL_SPARSE_CSR) { dst->rows = src->rows; dst->cols = src->cols; - dst->nnz = src->nnz; + dst->nnz = src->nnz; hl_csr_matrix csr = (hl_csr_matrix)src->matrix; - hl_memcpy_csr_matrix(dst, - csr->csr_val, - csr->csr_row, - csr->csr_col, - stream); + hl_memcpy_csr_matrix(dst, csr->csr_val, csr->csr_row, csr->csr_col, stream); } else if (dst->format == HL_SPARSE_CSC) { dst->rows = src->rows; dst->cols = src->cols; - dst->nnz = src->nnz; + dst->nnz = src->nnz; hl_csc_matrix csc = (hl_csc_matrix)src->matrix; - hl_memcpy_csc_matrix(dst, - csc->csc_val, - csc->csc_row, - csc->csc_col, - stream); + hl_memcpy_csc_matrix(dst, csc->csc_val, csc->csc_row, csc->csc_col, stream); } else { LOG(FATAL) << "sparse matrix format error!"; } @@ -569,20 +511,24 @@ static void _beta_mul_c(real *c, int dimM, int dimN, real beta) { if (beta == 0.0) { hl_gpu_apply_unary_op(unary::Zero(), c, dimM, dimN, dimN); } else { - if (beta != 1.0){ - hl_gpu_apply_unary_op( - unary::mul_scalar(beta), c, dimM, dimN, dimN); + if (beta != 1.0) { + hl_gpu_apply_unary_op(unary::mul_scalar(beta), c, dimM, dimN, dimN); } } return; } -void hl_matrix_csr_mul_dense(hl_sparse_matrix_s A_d, hl_trans_op_t transa, - real *B_d, hl_trans_op_t transb, +void hl_matrix_csr_mul_dense(hl_sparse_matrix_s A_d, + hl_trans_op_t transa, + real *B_d, + hl_trans_op_t transb, real *C_d, - int dimM, int dimN, int dimK, - real alpha, real beta) { + int dimM, + int dimN, + int dimK, + real alpha, + real beta) { CHECK_EQ(transb, HPPL_OP_N); CHECK_NOTNULL(A_d); CHECK_NOTNULL(B_d); @@ -592,7 +538,7 @@ void hl_matrix_csr_mul_dense(hl_sparse_matrix_s A_d, hl_trans_op_t transa, if ((HPPL_OP_N == transa && (A_d->rows != dimM || A_d->cols != dimK)) || (HPPL_OP_T == transa && (A_d->rows != dimK || A_d->cols != dimM))) { - LOG(FATAL) << "parameter error!"; + LOG(FATAL) << "parameter error!"; } if (A_d->nnz == 0) { @@ -603,8 +549,7 @@ void hl_matrix_csr_mul_dense(hl_sparse_matrix_s A_d, hl_trans_op_t transa, /* nnz != 0 */ hl_csr_matrix A_d2 = (hl_csr_matrix)(A_d->matrix); if ((A_d2->csr_val == NULL && A_d->type != HL_NO_VALUE) || - A_d2->csr_row == NULL || - A_d2->csr_col == NULL) { + A_d2->csr_row == NULL || A_d2->csr_col == NULL) { LOG(FATAL) << "parameter error!"; } @@ -617,63 +562,63 @@ void hl_matrix_csr_mul_dense(hl_sparse_matrix_s A_d, hl_trans_op_t transa, /* sparsity pattern */ // A_d->sparsity; if (A_d->type == HL_NO_VALUE) { - KeSMatrixCsrMulDense<0> - <<>>(C_d, - A_d2->csr_val, - A_d2->csr_col, - A_d2->csr_row, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixCsrMulDense<0><<>>( + C_d, + A_d2->csr_val, + A_d2->csr_col, + A_d2->csr_row, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); } else { - KeSMatrixCsrMulDense<1> - <<>>(C_d, - A_d2->csr_val, - A_d2->csr_col, - A_d2->csr_row, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixCsrMulDense<1><<>>( + C_d, + A_d2->csr_val, + A_d2->csr_col, + A_d2->csr_row, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); } } else if (HPPL_OP_T == transa) { _beta_mul_c(C_d, dimM, dimN, beta); - int blocksX = (dimN + CU_CSC_MUL_DENSE_BLOCK_N - 1) / - CU_CSC_MUL_DENSE_BLOCK_N; - int blocksY = (dimK + CU_CSC_MUL_DENSE_BLOCK_K - 1) / - CU_CSC_MUL_DENSE_BLOCK_K; + int blocksX = + (dimN + CU_CSC_MUL_DENSE_BLOCK_N - 1) / CU_CSC_MUL_DENSE_BLOCK_N; + int blocksY = + (dimK + CU_CSC_MUL_DENSE_BLOCK_K - 1) / CU_CSC_MUL_DENSE_BLOCK_K; dim3 threads(CU_CSC_MUL_DENSE_THREAD_X, CU_CSC_MUL_DENSE_THREAD_Y); dim3 grid(blocksX, blocksY); if (A_d->type == HL_NO_VALUE) { - KeSMatrixCscMulDense<0> - <<>>(C_d, - A_d2->csr_val, - A_d2->csr_col, - A_d2->csr_row, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixCscMulDense<0><<>>( + C_d, + A_d2->csr_val, + A_d2->csr_col, + A_d2->csr_row, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); } else { - KeSMatrixCscMulDense<1> - <<>>(C_d, - A_d2->csr_val, - A_d2->csr_col, - A_d2->csr_row, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixCscMulDense<1><<>>( + C_d, + A_d2->csr_val, + A_d2->csr_col, + A_d2->csr_row, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); } } else { LOG(FATAL) << "parameter transa error!"; @@ -682,11 +627,16 @@ void hl_matrix_csr_mul_dense(hl_sparse_matrix_s A_d, hl_trans_op_t transa, CHECK_SYNC("hl_matrix_csr_mul_dense failed"); } -void hl_matrix_dense_mul_csc(real *A_d, hl_trans_op_t transa, - hl_sparse_matrix_s B_d, hl_trans_op_t transb, +void hl_matrix_dense_mul_csc(real *A_d, + hl_trans_op_t transa, + hl_sparse_matrix_s B_d, + hl_trans_op_t transb, real *C_d, - int dimM, int dimN, int dimK, - real alpha, real beta) { + int dimM, + int dimN, + int dimK, + real alpha, + real beta) { CHECK_EQ(transa, HPPL_OP_N); CHECK_NOTNULL(A_d); CHECK_NOTNULL(B_d); @@ -698,8 +648,7 @@ void hl_matrix_dense_mul_csc(real *A_d, hl_trans_op_t transa, LOG(FATAL) << "parameter dims error!"; } - CHECK_EQ(B_d->format, HL_SPARSE_CSC) - << "matrix format error!"; + CHECK_EQ(B_d->format, HL_SPARSE_CSC) << "matrix format error!"; if (B_d->nnz == 0) { _beta_mul_c(C_d, dimM, dimN, beta); @@ -709,8 +658,7 @@ void hl_matrix_dense_mul_csc(real *A_d, hl_trans_op_t transa, /* nnz != 0 */ hl_csc_matrix B_d2 = (hl_csc_matrix)(B_d->matrix); if ((B_d2->csc_val == NULL && B_d->type != HL_NO_VALUE) || - B_d2->csc_row == NULL || - B_d2->csc_col == NULL) { + B_d2->csc_row == NULL || B_d2->csc_col == NULL) { LOG(FATAL) << "parameter B is null!"; } @@ -721,60 +669,60 @@ void hl_matrix_dense_mul_csc(real *A_d, hl_trans_op_t transa, dim3 grid(blocksX, blocksY); if (B_d->type == HL_NO_VALUE) { - KeSMatrixDenseMulCsc<0> - <<>>(C_d, - A_d, - B_d2->csc_val, - B_d2->csc_row, - B_d2->csc_col, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixDenseMulCsc<0><<>>( + C_d, + A_d, + B_d2->csc_val, + B_d2->csc_row, + B_d2->csc_col, + dimM, + dimN, + dimK, + alpha, + beta); } else { - KeSMatrixDenseMulCsc<1> - <<>>(C_d, - A_d, - B_d2->csc_val, - B_d2->csc_row, - B_d2->csc_col, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixDenseMulCsc<1><<>>( + C_d, + A_d, + B_d2->csc_val, + B_d2->csc_row, + B_d2->csc_col, + dimM, + dimN, + dimK, + alpha, + beta); } } else if (transb == HPPL_OP_T) { _beta_mul_c(C_d, dimM, dimN, beta); - int blocksX = 1 + (dimK-1)/CU_DM_CSR_THREAD_X; - int blocksY = 1 + (dimM-1)/CU_DM_CSR_BLOCK_M; + int blocksX = 1 + (dimK - 1) / CU_DM_CSR_THREAD_X; + int blocksY = 1 + (dimM - 1) / CU_DM_CSR_BLOCK_M; dim3 threads(CU_DM_CSR_THREAD_X, CU_DM_CSR_THREAD_Y); dim3 grid(blocksX, blocksY); if (B_d->type == HL_NO_VALUE) { - KeSMatrixDenseMulCsr<0> - <<>>(C_d, - A_d, - B_d2->csc_val, - B_d2->csc_col, - B_d2->csc_row, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixDenseMulCsr<0><<>>( + C_d, + A_d, + B_d2->csc_val, + B_d2->csc_col, + B_d2->csc_row, + dimM, + dimN, + dimK, + alpha, + beta); } else { - KeSMatrixDenseMulCsr<1> - <<>>(C_d, - A_d, - B_d2->csc_val, - B_d2->csc_col, - B_d2->csc_row, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixDenseMulCsr<1><<>>( + C_d, + A_d, + B_d2->csc_val, + B_d2->csc_col, + B_d2->csc_row, + dimM, + dimN, + dimK, + alpha, + beta); } } else { LOG(FATAL) << "parameter transb error!"; @@ -783,24 +731,28 @@ void hl_matrix_dense_mul_csc(real *A_d, hl_trans_op_t transa, CHECK_SYNC("hl_matrix_dense_mul_csc failed"); } -void hl_matrix_dense_mul_csr(real *A_d, hl_trans_op_t transa, - hl_sparse_matrix_s B_d, hl_trans_op_t transb, +void hl_matrix_dense_mul_csr(real *A_d, + hl_trans_op_t transa, + hl_sparse_matrix_s B_d, + hl_trans_op_t transb, real *C_d, - int dimM, int dimN, int dimK, - real alpha, real beta) { + int dimM, + int dimN, + int dimK, + real alpha, + real beta) { CHECK_EQ(transa, HPPL_OP_N); CHECK_NOTNULL(A_d); CHECK_NOTNULL(B_d); CHECK_NOTNULL(C_d); - if (dimM <= 0 || dimN <= 0 || dimK <= 0 - || (transb == HPPL_OP_N && (B_d->rows != dimK || B_d->cols != dimN)) - || (transb == HPPL_OP_T && (B_d->rows != dimN || B_d->cols != dimK))) { + if (dimM <= 0 || dimN <= 0 || dimK <= 0 || + (transb == HPPL_OP_N && (B_d->rows != dimK || B_d->cols != dimN)) || + (transb == HPPL_OP_T && (B_d->rows != dimN || B_d->cols != dimK))) { LOG(FATAL) << "parameter dims error!"; } - CHECK_EQ(B_d->format, HL_SPARSE_CSR) - << "matrix format error!"; + CHECK_EQ(B_d->format, HL_SPARSE_CSR) << "matrix format error!"; if (B_d->nnz == 0) { _beta_mul_c(C_d, dimM, dimN, beta); @@ -810,41 +762,40 @@ void hl_matrix_dense_mul_csr(real *A_d, hl_trans_op_t transa, /* nnz != 0 */ hl_csr_matrix B_d2 = (hl_csr_matrix)(B_d->matrix); if ((B_d2->csr_val == NULL && B_d->type != HL_NO_VALUE) || - B_d2->csr_row == NULL || - B_d2->csr_col == NULL) { + B_d2->csr_row == NULL || B_d2->csr_col == NULL) { LOG(FATAL) << "parameter transa error!"; } if (transb == HPPL_OP_N) { _beta_mul_c(C_d, dimM, dimN, beta); - int blocksX = 1 + (dimK-1)/CU_DM_CSR_THREAD_X; - int blocksY = 1 + (dimM-1)/CU_DM_CSR_BLOCK_M; + int blocksX = 1 + (dimK - 1) / CU_DM_CSR_THREAD_X; + int blocksY = 1 + (dimM - 1) / CU_DM_CSR_BLOCK_M; dim3 threads(CU_DM_CSR_THREAD_X, CU_DM_CSR_THREAD_Y); dim3 grid(blocksX, blocksY); if (B_d->type == HL_NO_VALUE) { - KeSMatrixDenseMulCsr<0> - <<>>(C_d, - A_d, - B_d2->csr_val, - B_d2->csr_row, - B_d2->csr_col, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixDenseMulCsr<0><<>>( + C_d, + A_d, + B_d2->csr_val, + B_d2->csr_row, + B_d2->csr_col, + dimM, + dimN, + dimK, + alpha, + beta); } else { - KeSMatrixDenseMulCsr<1> - <<>>(C_d, - A_d, - B_d2->csr_val, - B_d2->csr_row, - B_d2->csr_col, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixDenseMulCsr<1><<>>( + C_d, + A_d, + B_d2->csr_val, + B_d2->csr_row, + B_d2->csr_col, + dimM, + dimN, + dimK, + alpha, + beta); } } else if (transb == HPPL_OP_T) { int blocksX = (dimM + CU_CSCMM_BLOCK_M_BEST - 1) / CU_CSCMM_BLOCK_M_BEST; @@ -852,29 +803,29 @@ void hl_matrix_dense_mul_csr(real *A_d, hl_trans_op_t transa, dim3 threads(CU_CSCMM_THREAD_X_BEST, CU_CSCMM_THREAD_Y_BEST); dim3 grid(blocksX, blocksY); if (B_d->type == HL_NO_VALUE) { - KeSMatrixDenseMulCsc<0> - <<>>(C_d, - A_d, - B_d2->csr_val, - B_d2->csr_col, - B_d2->csr_row, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixDenseMulCsc<0><<>>( + C_d, + A_d, + B_d2->csr_val, + B_d2->csr_col, + B_d2->csr_row, + dimM, + dimN, + dimK, + alpha, + beta); } else { - KeSMatrixDenseMulCsc<1> - <<>>(C_d, - A_d, - B_d2->csr_val, - B_d2->csr_col, - B_d2->csr_row, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixDenseMulCsc<1><<>>( + C_d, + A_d, + B_d2->csr_val, + B_d2->csr_col, + B_d2->csr_row, + dimM, + dimN, + dimK, + alpha, + beta); } } else { LOG(FATAL) << "parameter transb error!"; @@ -883,11 +834,16 @@ void hl_matrix_dense_mul_csr(real *A_d, hl_trans_op_t transa, CHECK_SYNC("hl_matrix_dense_mul_csr failed"); } -void hl_matrix_csc_mul_dense(hl_sparse_matrix_s A_d, hl_trans_op_t transa, - real *B_d, hl_trans_op_t transb, +void hl_matrix_csc_mul_dense(hl_sparse_matrix_s A_d, + hl_trans_op_t transa, + real *B_d, + hl_trans_op_t transb, real *C_d, - int dimM, int dimN, int dimK, - real alpha, real beta) { + int dimM, + int dimN, + int dimK, + real alpha, + real beta) { CHECK_EQ(transb, HPPL_OP_N); CHECK_NOTNULL(A_d); CHECK_NOTNULL(B_d); @@ -908,42 +864,43 @@ void hl_matrix_csc_mul_dense(hl_sparse_matrix_s A_d, hl_trans_op_t transa, /* nnz != 0 */ hl_csc_matrix A_d2 = (hl_csc_matrix)(A_d->matrix); if ((A_d2->csc_val == NULL && A_d->type != HL_NO_VALUE) || - A_d2->csc_row == NULL || - A_d2->csc_col == NULL) { + A_d2->csc_row == NULL || A_d2->csc_col == NULL) { LOG(FATAL) << "parameter error!"; } if (HPPL_OP_N == transa) { _beta_mul_c(C_d, dimM, dimN, beta); - int blocksX = (dimN + CU_CSC_MUL_DENSE_BLOCK_N -1)/CU_CSC_MUL_DENSE_BLOCK_N; - int blocksY = (dimK + CU_CSC_MUL_DENSE_BLOCK_K -1)/CU_CSC_MUL_DENSE_BLOCK_K; + int blocksX = + (dimN + CU_CSC_MUL_DENSE_BLOCK_N - 1) / CU_CSC_MUL_DENSE_BLOCK_N; + int blocksY = + (dimK + CU_CSC_MUL_DENSE_BLOCK_K - 1) / CU_CSC_MUL_DENSE_BLOCK_K; dim3 threads(CU_CSC_MUL_DENSE_THREAD_X, CU_CSC_MUL_DENSE_THREAD_Y); dim3 grid(blocksX, blocksY); if (A_d->type == HL_NO_VALUE) { - KeSMatrixCscMulDense<0> - <<>>(C_d, - A_d2->csc_val, - A_d2->csc_row, - A_d2->csc_col, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixCscMulDense<0><<>>( + C_d, + A_d2->csc_val, + A_d2->csc_row, + A_d2->csc_col, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); } else { - KeSMatrixCscMulDense<1> - <<>>(C_d, - A_d2->csc_val, - A_d2->csc_row, - A_d2->csc_col, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixCscMulDense<1><<>>( + C_d, + A_d2->csc_val, + A_d2->csc_row, + A_d2->csc_col, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); } } else if (HPPL_OP_T == transa) { int blocksX = (dimN + CU_CSRMM_BLOCK_N - 1) / CU_CSRMM_BLOCK_N; @@ -954,29 +911,29 @@ void hl_matrix_csc_mul_dense(hl_sparse_matrix_s A_d, hl_trans_op_t transa, /* sparsity pattern */ // A_d->sparsity; if (A_d->type == HL_NO_VALUE) { - KeSMatrixCsrMulDense<0> - <<>>(C_d, - A_d2->csc_val, - A_d2->csc_row, - A_d2->csc_col, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixCsrMulDense<0><<>>( + C_d, + A_d2->csc_val, + A_d2->csc_row, + A_d2->csc_col, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); } else { - KeSMatrixCsrMulDense<1> - <<>>(C_d, - A_d2->csc_val, - A_d2->csc_row, - A_d2->csc_col, - B_d, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixCsrMulDense<1><<>>( + C_d, + A_d2->csc_val, + A_d2->csc_row, + A_d2->csc_col, + B_d, + dimM, + dimN, + dimK, + alpha, + beta); } } else { LOG(FATAL) << "parameter transa error!"; @@ -985,11 +942,16 @@ void hl_matrix_csc_mul_dense(hl_sparse_matrix_s A_d, hl_trans_op_t transa, CHECK_SYNC("hl_matrix_csc_mul_dense failed"); } -void hl_sparse_matrix_mul(real *A_d, hl_trans_op_t transa, - real *B_d, hl_trans_op_t transb, - hl_sparse_matrix_s C_d, - int dimM, int dimN, int dimK, - real alpha, real beta) { +void hl_sparse_matrix_mul(real *A_d, + hl_trans_op_t transa, + real *B_d, + hl_trans_op_t transb, + hl_sparse_matrix_s C_d, + int dimM, + int dimN, + int dimK, + real alpha, + real beta) { CHECK_NOTNULL(A_d); CHECK_NOTNULL(B_d); CHECK_NOTNULL(C_d); @@ -1000,18 +962,14 @@ void hl_sparse_matrix_mul(real *A_d, hl_trans_op_t transa, if (C_d->format == HL_SPARSE_CSC) { hl_csc_matrix C_d2 = (hl_csc_matrix)(C_d->matrix); - if (C_d2->csc_val == NULL || - C_d2->csc_row == NULL || + if (C_d2->csc_val == NULL || C_d2->csc_row == NULL || C_d2->csc_col == NULL) { LOG(FATAL) << "parameter error!"; } if (beta != 1.0) { - hl_gpu_apply_unary_op(unary::mul_scalar(beta), - C_d2->csc_val, - 1, - C_d->nnz, - C_d->nnz); + hl_gpu_apply_unary_op( + unary::mul_scalar(beta), C_d2->csc_val, 1, C_d->nnz, C_d->nnz); } int blocksX = dimN; @@ -1020,34 +978,30 @@ void hl_sparse_matrix_mul(real *A_d, hl_trans_op_t transa, dim3 grid(blocksX, blocksY); bool transA = transa == HPPL_OP_T ? 1 : 0; bool transB = transb == HPPL_OP_T ? 1 : 0; - KeSMatrixDenseMulDense2CSC - <<>>(C_d2->csc_val, - C_d2->csc_row, - C_d2->csc_col, - A_d, - B_d, - transA, - transB, - dimM, - dimN, - dimK, - alpha, - beta); + KeSMatrixDenseMulDense2CSC<<>>( + C_d2->csc_val, + C_d2->csc_row, + C_d2->csc_col, + A_d, + B_d, + transA, + transB, + dimM, + dimN, + dimK, + alpha, + beta); CHECK_SYNC("hl_sparse_matrix_mul failed"); } else { hl_csr_matrix C_d2 = (hl_csr_matrix)(C_d->matrix); if ((C_d2->csr_val == NULL && C_d->type != HL_NO_VALUE) || - C_d2->csr_row == NULL || - C_d2->csr_col == NULL) { + C_d2->csr_row == NULL || C_d2->csr_col == NULL) { LOG(FATAL) << "parameter error!"; } if (beta != 1.0) { - hl_gpu_apply_unary_op(unary::mul_scalar(beta), - C_d2->csr_val, - 1, - C_d->nnz, - C_d->nnz); + hl_gpu_apply_unary_op( + unary::mul_scalar(beta), C_d2->csr_val, 1, C_d->nnz, C_d->nnz); } bool transA = transa == HPPL_OP_T ? 1 : 0; @@ -1058,20 +1012,20 @@ void hl_sparse_matrix_mul(real *A_d, hl_trans_op_t transa, dim3 threads(CU_CSCMM_DMD2CSR_THREAD_X, 1); dim3 grid(blocksX, blocksY); - KeSMatrixDenseMulDense2CSR - <<>>(C_d2->csr_val, - C_d2->csr_row, - C_d2->csr_col, - A_d, - B_d, - transA, - transB, - dimM, - dimN, - dimK, - alpha, - beta); - CHECK_SYNC("hl_sparse_matrix_mul failed"); + KeSMatrixDenseMulDense2CSR<<>>( + C_d2->csr_val, + C_d2->csr_row, + C_d2->csr_col, + A_d, + B_d, + transA, + transB, + dimM, + dimN, + dimK, + alpha, + beta); + CHECK_SYNC("hl_sparse_matrix_mul failed"); } else { CHECK(!transA) << "Not supported A is trans and B is not trans!"; @@ -1080,21 +1034,21 @@ void hl_sparse_matrix_mul(real *A_d, hl_trans_op_t transa, avgNnzPerRow = avgNnzPerRow > 0 ? avgNnzPerRow : 1; int gridx = DIVUP(avgNnzPerRow, CU_BLOCK_SIZE); dim3 grid(gridx, dimM); - KeSMatrixDenseMulDenseTrans2CSR - <<>>(C_d2->csr_val, - C_d2->csr_row, - C_d2->csr_col, - A_d, - B_d, - transA, - transB, - dimM, - dimN, - dimK, - alpha, - beta); - CHECK_SYNC("hl_sparse_matrix_mul failed"); - } + KeSMatrixDenseMulDenseTrans2CSR<<>>( + C_d2->csr_val, + C_d2->csr_row, + C_d2->csr_col, + A_d, + B_d, + transA, + transB, + dimM, + dimN, + dimK, + alpha, + beta); + CHECK_SYNC("hl_sparse_matrix_mul failed"); + } } } @@ -1111,7 +1065,7 @@ void hl_memcpy_from_csc_matrix(real *csc_val, CHECK_NOTNULL(csc_col); CHECK_EQ(csc_matrix->format, HL_SPARSE_CSC) - << "csc_matrix is not csc format error!"; + << "csc_matrix is not csc format error!"; if (csc_matrix->nnz > row_size || csc_matrix->cols + 1 > static_cast(col_size)) { @@ -1119,20 +1073,20 @@ void hl_memcpy_from_csc_matrix(real *csc_val, } hl_csc_matrix csc = (hl_csc_matrix)(csc_matrix->matrix); - hl_memcpy_async((void*)csc_row, - (void*)csc->csc_row, + hl_memcpy_async((void *)csc_row, + (void *)csc->csc_row, (csc_matrix->nnz) * sizeof(int), stream); - hl_memcpy_async((void*)csc_col, - (void*)csc->csc_col, + hl_memcpy_async((void *)csc_col, + (void *)csc->csc_col, (csc_matrix->cols + 1) * sizeof(int), stream); if (csc_matrix->type == HL_FLOAT_VALUE) { if (csc_val != NULL) { CHECK_LE(csc_matrix->nnz, val_size) << "size not match!"; - hl_memcpy_async((void*)csc_val, - (void*)csc->csc_val, - (csc_matrix->nnz)*sizeof(real), + hl_memcpy_async((void *)csc_val, + (void *)csc->csc_val, + (csc_matrix->nnz) * sizeof(real), stream); } else { LOG(FATAL) << "parameter csr_val is null pointer!"; @@ -1152,7 +1106,7 @@ void hl_memcpy_from_csr_matrix(real *csr_val, CHECK_NOTNULL(csr_row); CHECK_NOTNULL(csr_col); CHECK_EQ(csr_matrix->format, HL_SPARSE_CSR) - << "csr_matrix is not csr format error!"; + << "csr_matrix is not csr format error!"; if (csr_matrix->nnz > col_size || csr_matrix->rows + 1 > static_cast(row_size)) { @@ -1160,20 +1114,20 @@ void hl_memcpy_from_csr_matrix(real *csr_val, } hl_csr_matrix csr = (hl_csr_matrix)(csr_matrix->matrix); - hl_memcpy_async((void*)csr_row, - (void*)csr->csr_row, - (csr_matrix->rows+1)*sizeof(int), + hl_memcpy_async((void *)csr_row, + (void *)csr->csr_row, + (csr_matrix->rows + 1) * sizeof(int), stream); - hl_memcpy_async((void*)csr_col, - (void*)csr->csr_col, - (csr_matrix->nnz)*sizeof(int), + hl_memcpy_async((void *)csr_col, + (void *)csr->csr_col, + (csr_matrix->nnz) * sizeof(int), stream); if (csr_matrix->type == HL_FLOAT_VALUE) { if (csr_val != NULL) { CHECK_LE(csr_matrix->nnz, val_size) << "size not match!"; - hl_memcpy_async((void*)csr_val, - (void*)csr->csr_val, - (csr_matrix->nnz)*sizeof(real), + hl_memcpy_async((void *)csr_val, + (void *)csr->csr_val, + (csr_matrix->nnz) * sizeof(real), stream); } else { LOG(FATAL) << "parameter csr_val is null pointer!"; @@ -1181,8 +1135,8 @@ void hl_memcpy_from_csr_matrix(real *csr_val, } } -void hl_sparse_matrix_column_sum(real* A_d, hl_sparse_matrix_s B_d, int dimM, - int dimN, real scale) { +void hl_sparse_matrix_column_sum( + real *A_d, hl_sparse_matrix_s B_d, int dimM, int dimN, real scale) { if (B_d->format == HL_SPARSE_CSR) { hl_matrix_csr_column_sum(A_d, B_d, dimM, dimN, scale); } else { @@ -1190,8 +1144,8 @@ void hl_sparse_matrix_column_sum(real* A_d, hl_sparse_matrix_s B_d, int dimM, } } -void hl_matrix_csr_column_sum(real* A_d, hl_sparse_matrix_s B_d, - int dimM, int dimN, real scale) { +void hl_matrix_csr_column_sum( + real *A_d, hl_sparse_matrix_s B_d, int dimM, int dimN, real scale) { CHECK_NOTNULL(A_d); CHECK_NOTNULL(B_d); @@ -1216,8 +1170,7 @@ void hl_matrix_csr_column_sum(real* A_d, hl_sparse_matrix_s B_d, CHECK_SYNC("hl_matrix_csr_column_sum failed"); } -void hl_sparse_matrix_add_bias(hl_sparse_matrix_s A_d, - real* B_d, real scale) { +void hl_sparse_matrix_add_bias(hl_sparse_matrix_s A_d, real *B_d, real scale) { if (A_d->format == HL_SPARSE_CSR) { hl_matrix_csr_add_bias(A_d, B_d, scale); } else { @@ -1225,8 +1178,7 @@ void hl_sparse_matrix_add_bias(hl_sparse_matrix_s A_d, } } -void hl_matrix_csr_add_bias(hl_sparse_matrix_s A_d, real* B_d, - real scale) { +void hl_matrix_csr_add_bias(hl_sparse_matrix_s A_d, real *B_d, real scale) { CHECK_NOTNULL(A_d); CHECK_NOTNULL(B_d); @@ -1247,8 +1199,12 @@ void hl_matrix_csr_add_bias(hl_sparse_matrix_s A_d, real* B_d, CHECK_SYNC("hl_sparse_matrix_add_bias failed"); } -void hl_sparse_matrix_add_dense(hl_sparse_matrix_s A_d, real *B_d, int dimM, - int dimN, real alpha, real beta) { +void hl_sparse_matrix_add_dense(hl_sparse_matrix_s A_d, + real *B_d, + int dimM, + int dimN, + real alpha, + real beta) { if (A_d->format == HL_SPARSE_CSR) { hl_matrix_csr_add_dense(A_d, B_d, dimM, dimN, alpha, beta); } else { @@ -1256,8 +1212,12 @@ void hl_sparse_matrix_add_dense(hl_sparse_matrix_s A_d, real *B_d, int dimM, } } -void hl_matrix_csr_add_dense(hl_sparse_matrix_s A_d, real* B_d, int dimM, - int dimN, real alpha, real beta) { +void hl_matrix_csr_add_dense(hl_sparse_matrix_s A_d, + real *B_d, + int dimM, + int dimN, + real alpha, + real beta) { CHECK_NOTNULL(A_d); CHECK_NOTNULL(B_d); @@ -1277,20 +1237,26 @@ void hl_matrix_csr_add_dense(hl_sparse_matrix_s A_d, real* B_d, int dimM, gridX = gridX > 0 ? gridX : 1; dim3 block(512, 1); dim3 grid(gridX, dimM); - KeSMatrixCsrAddDense<<>>( - A_d2->csr_val, A_d2->csr_row, A_d2->csr_col, B_d, alpha, beta, dimM, dimN); + KeSMatrixCsrAddDense<<>>(A_d2->csr_val, + A_d2->csr_row, + A_d2->csr_col, + B_d, + alpha, + beta, + dimM, + dimN); CHECK_SYNC("hl_sparse_matrix_add_dense failed"); } -int* hl_sparse_matrix_get_rows(hl_sparse_matrix_s sMat) { +int *hl_sparse_matrix_get_rows(hl_sparse_matrix_s sMat) { __sparse_get_return__(sMat, row); } -int* hl_sparse_matrix_get_cols(hl_sparse_matrix_s sMat) { +int *hl_sparse_matrix_get_cols(hl_sparse_matrix_s sMat) { __sparse_get_return__(sMat, col); } -real* hl_sparse_matrix_get_value(hl_sparse_matrix_s sMat) { +real *hl_sparse_matrix_get_value(hl_sparse_matrix_s sMat) { __sparse_get_return__(sMat, val); } diff --git a/paddle/cuda/src/hl_perturbation_util.cu b/paddle/cuda/src/hl_perturbation_util.cu index 2a945bcdb87fe49c121890128ef77b084ebe8e60..d01a91561efa2ebe8e0cabc2b4e8885f2c02ab48 100644 --- a/paddle/cuda/src/hl_perturbation_util.cu +++ b/paddle/cuda/src/hl_perturbation_util.cu @@ -12,13 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ - -#include #include -#include "hl_cuda.h" -#include "hl_time.h" +#include #include "hl_base.h" +#include "hl_cuda.h" #include "hl_perturbation_util.cuh" +#include "hl_time.h" #define _USE_MATH_DEFINES @@ -30,10 +29,16 @@ limitations under the License. */ * centerX, centerY: translation. * sourceX, sourceY: output coordinates in the original image. */ -__device__ void getTranformCoord(int x, int y, real theta, real scale, - real tgtCenter, real imgCenter, - real centerR, real centerC, - int* sourceX, int* sourceY) { +__device__ void getTranformCoord(int x, + int y, + real theta, + real scale, + real tgtCenter, + real imgCenter, + real centerR, + real centerC, + int* sourceX, + int* sourceY) { real H[4] = {cosf(-theta), -sinf(-theta), sinf(-theta), cosf(-theta)}; // compute coornidates in the rotated and scaled image @@ -57,11 +62,17 @@ __device__ void getTranformCoord(int x, int y, real theta, real scale, * created by Wei Xu (genome), converted by Jiang Wang */ -__global__ void kSamplingPatches(const real* imgs, real* targets, - int imgSize, int tgtSize, const int channels, - int samplingRate, const real* thetas, - const real* scales, const int* centerRs, - const int* centerCs, const real padValue, +__global__ void kSamplingPatches(const real* imgs, + real* targets, + int imgSize, + int tgtSize, + const int channels, + int samplingRate, + const real* thetas, + const real* scales, + const int* centerRs, + const int* centerCs, + const real padValue, const int numImages) { const int caseIdx = blockIdx.x * 4 + threadIdx.x; const int pxIdx = blockIdx.y * 128 + threadIdx.y; @@ -80,8 +91,15 @@ __global__ void kSamplingPatches(const real* imgs, real* targets, const int pxY = pxIdx / tgtSize; int srcPxX, srcPxY; - getTranformCoord(pxX, pxY, thetas[imgIdx], scales[imgIdx], tgtCenter, - imgCenter, centerCs[caseIdx], centerRs[caseIdx], &srcPxX, + getTranformCoord(pxX, + pxY, + thetas[imgIdx], + scales[imgIdx], + tgtCenter, + imgCenter, + centerCs[caseIdx], + centerRs[caseIdx], + &srcPxX, &srcPxY); imgs += (imgIdx * imgPixels + srcPxY * imgSize + srcPxX) * channels; @@ -100,10 +118,15 @@ __global__ void kSamplingPatches(const real* imgs, real* targets, * * created by Wei Xu */ -void hl_generate_disturb_params(real*& gpuAngle, real*& gpuScaleRatio, - int*& gpuCenterR, int*& gpuCenterC, - int numImages, int imgSize, real rotateAngle, - real scaleRatio, int samplingRate, +void hl_generate_disturb_params(real*& gpuAngle, + real*& gpuScaleRatio, + int*& gpuCenterR, + int*& gpuCenterC, + int numImages, + int imgSize, + real rotateAngle, + real scaleRatio, + int samplingRate, bool isTrain) { // The number of output samples. int numPatches = numImages * samplingRate; @@ -123,7 +146,8 @@ void hl_generate_disturb_params(real*& gpuAngle, real*& gpuScaleRatio, for (int i = 0; i < numImages; i++) { r_angle[i] = (rotateAngle * M_PI / 180.0) * (rand() / (RAND_MAX + 1.0) // NOLINT - - 0.5); + - + 0.5); s_ratio[i] = 1 + (rand() / (RAND_MAX + 1.0) - 0.5) * scaleRatio; // NOLINT } @@ -140,8 +164,10 @@ void hl_generate_disturb_params(real*& gpuAngle, real*& gpuScaleRatio, int pxY = (int)(real(imgSize - 1) * rand() / (RAND_MAX + 1.0)); // NOLINT - const real H[4] = {cos(-r_angle[i]), -sin(-r_angle[i]), - sin(-r_angle[i]), cos(-r_angle[i])}; + const real H[4] = {cos(-r_angle[i]), + -sin(-r_angle[i]), + sin(-r_angle[i]), + cos(-r_angle[i])}; real x = pxX - imgCenter; real y = pxY - imgCenter; real xx = H[0] * x + H[1] * y; @@ -185,9 +211,12 @@ void hl_generate_disturb_params(real*& gpuAngle, real*& gpuScaleRatio, delete[] center_c; } -void hl_conv_random_disturb_with_params(const real* images, int imgSize, - int tgtSize, int channels, - int numImages, int samplingRate, +void hl_conv_random_disturb_with_params(const real* images, + int imgSize, + int tgtSize, + int channels, + int numImages, + int samplingRate, const real* gpuRotationAngle, const real* gpuScaleRatio, const int* gpuCenterR, @@ -202,29 +231,59 @@ void hl_conv_random_disturb_with_params(const real* images, int imgSize, dim3 threadsPerBlock(4, 128); dim3 numBlocks(DIVUP(numPatches, 4), DIVUP(targetSize, 128)); - kSamplingPatches <<>> - (images, target, imgSize, tgtSize, channels, samplingRate, - gpuRotationAngle, gpuScaleRatio, gpuCenterR, gpuCenterC, - paddingValue, numImages); + kSamplingPatches<<>>(images, + target, + imgSize, + tgtSize, + channels, + samplingRate, + gpuRotationAngle, + gpuScaleRatio, + gpuCenterR, + gpuCenterC, + paddingValue, + numImages); hl_device_synchronize(); } -void hl_conv_random_disturb(const real* images, int imgSize, - int tgtSize, int channels, int numImages, - real scaleRatio, real rotateAngle, - int samplingRate, real* gpu_r_angle, - real* gpu_s_ratio, int* gpu_center_r, - int* gpu_center_c, int paddingValue, - bool isTrain, real* targets) { +void hl_conv_random_disturb(const real* images, + int imgSize, + int tgtSize, + int channels, + int numImages, + real scaleRatio, + real rotateAngle, + int samplingRate, + real* gpu_r_angle, + real* gpu_s_ratio, + int* gpu_center_r, + int* gpu_center_c, + int paddingValue, + bool isTrain, + real* targets) { // generate the random disturbance sequence and the sampling locations - hl_generate_disturb_params(gpu_r_angle, gpu_s_ratio, gpu_center_r, - gpu_center_c, numImages, imgSize, rotateAngle, - scaleRatio, samplingRate, isTrain); - - hl_conv_random_disturb_with_params( - images, imgSize, tgtSize, channels, numImages, - samplingRate, gpu_r_angle, gpu_s_ratio, - gpu_center_r, gpu_center_r, paddingValue, - targets); + hl_generate_disturb_params(gpu_r_angle, + gpu_s_ratio, + gpu_center_r, + gpu_center_c, + numImages, + imgSize, + rotateAngle, + scaleRatio, + samplingRate, + isTrain); + + hl_conv_random_disturb_with_params(images, + imgSize, + tgtSize, + channels, + numImages, + samplingRate, + gpu_r_angle, + gpu_s_ratio, + gpu_center_r, + gpu_center_r, + paddingValue, + targets); } diff --git a/paddle/cuda/src/hl_table_apply.cu b/paddle/cuda/src/hl_table_apply.cu index 61edbe3ccc7028fd8779c4119f33c4cb5afe0564..d3b71c75e6e69d48c8d98041e3d6075aa8d53610 100644 --- a/paddle/cuda/src/hl_table_apply.cu +++ b/paddle/cuda/src/hl_table_apply.cu @@ -12,15 +12,16 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ - #include "hl_base.h" -#include "hl_device_functions.cuh" #include "hl_cuda.h" +#include "hl_device_functions.cuh" #include "paddle/utils/Logging.h" -template -__global__ void KeMatrixAddRows(real* output, int ldo, - real* table, int ldt, +template +__global__ void KeMatrixAddRows(real* output, + int ldo, + real* table, + int ldt, int* ids, int numSamples, int tableSize, @@ -31,8 +32,8 @@ __global__ void KeMatrixAddRows(real* output, int ldo, while (idy < numSamples) { int tableId = ids[idy]; if ((0 <= tableId) && (tableId < tableSize)) { - real *out = output + idy * ldo; - real *tab = table + tableId * ldt; + real* out = output + idy * ldo; + real* tab = table + tableId * ldt; for (int i = idx; i < dim; i += blockDimX) { if (AddRow) { paddle::paddleAtomicAdd(&tab[i], out[i]); @@ -45,8 +46,10 @@ __global__ void KeMatrixAddRows(real* output, int ldo, } } -void hl_matrix_select_rows(real* output, int ldo, - real* table, int ldt, +void hl_matrix_select_rows(real* output, + int ldo, + real* table, + int ldt, int* ids, int numSamples, int tableSize, @@ -57,14 +60,16 @@ void hl_matrix_select_rows(real* output, int ldo, dim3 threads(128, 8); dim3 grid(8, 1); - KeMatrixAddRows<128, 8, 8, 0><<< grid, threads, 0, STREAM_DEFAULT >>> - (output, ldo, table, ldt, ids, numSamples, tableSize, dim); + KeMatrixAddRows<128, 8, 8, 0><<>>( + output, ldo, table, ldt, ids, numSamples, tableSize, dim); CHECK_SYNC("hl_matrix_select_rows failed"); } -void hl_matrix_add_to_rows(real* table, int ldt, - real* input, int ldi, +void hl_matrix_add_to_rows(real* table, + int ldt, + real* input, + int ldi, int* ids, int numSamples, int tableSize, @@ -75,16 +80,15 @@ void hl_matrix_add_to_rows(real* table, int ldt, dim3 threads(128, 8); dim3 grid(8, 1); - KeMatrixAddRows<128, 8, 8, 1><<< grid, threads, 0, STREAM_DEFAULT >>> - (input, ldi, table, ldt, ids, numSamples, tableSize, dim); + KeMatrixAddRows<128, 8, 8, 1><<>>( + input, ldi, table, ldt, ids, numSamples, tableSize, dim); CHECK_SYNC("hl_matrix_add_to_rows failed"); } -template -__global__ void KeVectorSelect(T* dst, int sized, - const T* src, int sizes, - const int* ids, int sizei) { +template +__global__ void KeVectorSelect( + T* dst, int sized, const T* src, int sizes, const int* ids, int sizei) { int idx = threadIdx.x + blockDimX * blockIdx.x; while (idx < sizei) { int index = ids[idx]; @@ -95,9 +99,8 @@ __global__ void KeVectorSelect(T* dst, int sized, } template -void hl_vector_select_from(T* dst, int sized, - const T* src, int sizes, - const int* ids, int sizei) { +void hl_vector_select_from( + T* dst, int sized, const T* src, int sizes, const int* ids, int sizei) { CHECK_NOTNULL(dst); CHECK_NOTNULL(src); CHECK_NOTNULL(ids); @@ -105,18 +108,17 @@ void hl_vector_select_from(T* dst, int sized, dim3 threads(512, 1); dim3 grid(8, 1); - KeVectorSelect<<< grid, threads, 0, STREAM_DEFAULT >>> - (dst, sized, src, sizes, ids, sizei); + KeVectorSelect<<>>( + dst, sized, src, sizes, ids, sizei); CHECK_SYNC("hl_vector_select_from failed"); } -template -void hl_vector_select_from(real* dst, int sized, - const real* src, int sizes, - const int* ids, int sizei); -template -void hl_vector_select_from(int* dst, int sized, - const int* src, int sizes, - const int* ids, int sizei); - +template void hl_vector_select_from(real* dst, + int sized, + const real* src, + int sizes, + const int* ids, + int sizei); +template void hl_vector_select_from( + int* dst, int sized, const int* src, int sizes, const int* ids, int sizei); diff --git a/paddle/cuda/src/hl_top_k.cu b/paddle/cuda/src/hl_top_k.cu index 4f0bbfcf4e3aa51dd06acf254af65c62098a1df7..1896a56634c3a75e5a2a1e08661088b263f8ee10 100644 --- a/paddle/cuda/src/hl_top_k.cu +++ b/paddle/cuda/src/hl_top_k.cu @@ -12,45 +12,37 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ - #include "hl_base.h" -#include "hl_top_k.h" #include "hl_sparse.ph" +#include "hl_top_k.h" #include "paddle/utils/Logging.h" // using namespace hppl; struct Pair { - __device__ __forceinline__ - Pair() {} + __device__ __forceinline__ Pair() {} - __device__ __forceinline__ - Pair(real value, int id) : v_(value), id_(id) {} + __device__ __forceinline__ Pair(real value, int id) : v_(value), id_(id) {} - __device__ __forceinline__ - void set(real value, int id) { + __device__ __forceinline__ void set(real value, int id) { v_ = value; id_ = id; } - __device__ __forceinline__ - void operator=(const Pair& in) { + __device__ __forceinline__ void operator=(const Pair& in) { v_ = in.v_; id_ = in.id_; } - __device__ __forceinline__ - bool operator<(const real value) const { + __device__ __forceinline__ bool operator<(const real value) const { return (v_ < value); } - __device__ __forceinline__ - bool operator<(const Pair& in) const { + __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 { + __device__ __forceinline__ bool operator>(const Pair& in) const { return (v_ > in.v_) || ((v_ == in.v_) && (id_ < in.id_)); } @@ -58,8 +50,9 @@ struct Pair { int id_; }; -__device__ __forceinline__ -void addTo(Pair topK[], const Pair &p, int beamSize) { +__device__ __forceinline__ void addTo(Pair topK[], + const Pair& p, + int beamSize) { for (int k = beamSize - 2; k >= 0; k--) { if (topK[k] < p) { topK[k + 1] = topK[k]; @@ -71,9 +64,8 @@ void addTo(Pair topK[], const Pair &p, int beamSize) { topK[0] = p; } -template -__device__ __forceinline__ -void addTo(Pair topK[], const Pair &p) { +template +__device__ __forceinline__ void addTo(Pair topK[], const Pair& p) { for (int k = beamSize - 2; k >= 0; k--) { if (topK[k] < p) { topK[k + 1] = topK[k]; @@ -85,9 +77,9 @@ void addTo(Pair topK[], const Pair &p) { topK[0] = p; } -template -__device__ __forceinline__ -void getTopK(Pair topK[], real *src, int idx, int dim, int beamSize) { +template +__device__ __forceinline__ void getTopK( + Pair topK[], real* src, int idx, int dim, int beamSize) { while (idx < dim) { if (topK[beamSize - 1] < src[idx]) { Pair tmp(src[idx], idx); @@ -97,10 +89,9 @@ void getTopK(Pair topK[], real *src, int idx, int dim, int beamSize) { } } -template -__device__ __forceinline__ -void getTopK(Pair topK[], real *src, int idx, int dim, - const Pair& max, int beamSize) { +template +__device__ __forceinline__ void getTopK( + Pair topK[], real* src, int idx, int dim, const Pair& max, int beamSize) { while (idx < dim) { if (topK[beamSize - 1] < src[idx]) { Pair tmp(src[idx], idx); @@ -112,10 +103,9 @@ void getTopK(Pair topK[], real *src, int idx, int dim, } } -template -__device__ __forceinline__ -void getTopK(Pair topK[], real *val, int *col, - int idx, int dim, int beamSize) { +template +__device__ __forceinline__ void getTopK( + Pair topK[], real* val, int* col, int idx, int dim, int beamSize) { while (idx < dim) { if (topK[beamSize - 1] < val[idx]) { Pair tmp(val[idx], col[idx]); @@ -125,10 +115,14 @@ void getTopK(Pair topK[], real *val, int *col, } } -template -__device__ __forceinline__ -void getTopK(Pair topK[], real *val, int *col, int idx, int dim, - const Pair& max, int beamSize) { +template +__device__ __forceinline__ void getTopK(Pair topK[], + real* val, + int* col, + int idx, + int dim, + const Pair& max, + int beamSize) { while (idx < dim) { if (topK[beamSize - 1] < val[idx]) { Pair tmp(val[idx], col[idx]); @@ -140,12 +134,16 @@ void getTopK(Pair topK[], real *val, int *col, int idx, int dim, } } -template -__device__ __forceinline__ -void threadGetTopK(Pair topK[], int& beam, int beamSize, - real* src, - bool& firstStep, bool& isEmpty, Pair& max, - int dim, const int tid) { +template +__device__ __forceinline__ void threadGetTopK(Pair topK[], + int& beam, + int beamSize, + real* src, + bool& firstStep, + bool& isEmpty, + Pair& max, + int dim, + const int tid) { if (beam > 0) { int length = beam < beamSize ? beam : beamSize; if (firstStep) { @@ -160,8 +158,7 @@ void threadGetTopK(Pair topK[], int& beam, int beamSize, } } if (!isEmpty) { - getTopK(topK + maxLength - beam, src, tid, dim, - max, length); + getTopK(topK + maxLength - beam, src, tid, dim, max, length); } } @@ -171,12 +168,17 @@ void threadGetTopK(Pair topK[], int& beam, int beamSize, } } -template -__device__ __forceinline__ -void threadGetTopK(Pair topK[], int& beam, int beamSize, - real* val, int* col, - bool& firstStep, bool& isEmpty, Pair& max, - int dim, const int tid) { +template +__device__ __forceinline__ void threadGetTopK(Pair topK[], + int& beam, + int beamSize, + real* val, + int* col, + bool& firstStep, + bool& isEmpty, + Pair& max, + int dim, + const int tid) { if (beam > 0) { int length = beam < beamSize ? beam : beamSize; if (firstStep) { @@ -191,8 +193,8 @@ void threadGetTopK(Pair topK[], int& beam, int beamSize, } } if (!isEmpty) { - getTopK(topK + maxLength - beam, val, col, tid, dim, - max, length); + getTopK( + topK + maxLength - beam, val, col, tid, dim, max, length); } } @@ -202,12 +204,16 @@ void threadGetTopK(Pair topK[], int& beam, int beamSize, } } -template -__device__ __forceinline__ -void blockReduce(Pair* shTopK, int* maxId, Pair topK[], - real** topVal, int** topIds, - int& beam, int& beamSize, - const int tid, const int warp) { +template +__device__ __forceinline__ void blockReduce(Pair* shTopK, + int* maxId, + Pair topK[], + real** topVal, + int** topIds, + int& beam, + int& beamSize, + const int tid, + const int warp) { while (true) { __syncthreads(); if (tid < blockSize / 2) { @@ -218,7 +224,7 @@ void blockReduce(Pair* shTopK, int* maxId, Pair topK[], } } __syncthreads(); - for (int stride = blockSize / 4; stride > 0; stride = stride/2) { + for (int stride = blockSize / 4; stride > 0; stride = stride / 2) { if (tid < stride) { if (shTopK[maxId[tid]] < shTopK[maxId[tid + stride]]) { maxId[tid] = maxId[tid + stride]; @@ -257,10 +263,12 @@ void blockReduce(Pair* shTopK, int* maxId, Pair topK[], * 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(real* topVal, int ldv, - int * topIds, - real* src, int lds, +template +__global__ void KeMatrixTopK(real* topVal, + int ldv, + int* topIds, + real* src, + int lds, int dim, int beamSize) { __shared__ Pair shTopK[blockSize]; @@ -271,7 +279,7 @@ __global__ void KeMatrixTopK(real* topVal, int ldv, topVal += blockIdx.x * ldv; topIds += blockIdx.x * beamSize; - Pair topK[maxLength]; // NOLINT + Pair topK[maxLength]; // NOLINT int beam = maxLength; Pair max; bool isEmpty = false; @@ -281,18 +289,19 @@ __global__ void KeMatrixTopK(real* topVal, int ldv, topK[k].set(-HL_FLOAT_MAX, -1); } while (beamSize) { - threadGetTopK - (topK, beam, beamSize, src, firstStep, isEmpty, max, dim, tid); + threadGetTopK( + topK, beam, beamSize, src, firstStep, isEmpty, max, dim, tid); shTopK[tid] = topK[0]; - blockReduce - (shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp); + blockReduce( + shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp); } } -template -__global__ void KeSMatrixTopK(real* topVal, int ldv, - int * topIds, +template +__global__ void KeSMatrixTopK(real* topVal, + int ldv, + int* topIds, real* val, int* row, int* col, @@ -304,7 +313,7 @@ __global__ void KeSMatrixTopK(real* topVal, int ldv, topVal += blockIdx.x * ldv; topIds += blockIdx.x * beamSize; - Pair topK[maxLength]; // NOLINT + Pair topK[maxLength]; // NOLINT int beam = maxLength; Pair max; bool isEmpty = false; @@ -330,18 +339,20 @@ __global__ void KeSMatrixTopK(real* topVal, int ldv, topK[k].set(-HL_FLOAT_MAX, -1); } while (beamSize) { - threadGetTopK - (topK, beam, beamSize, val, col, firstStep, isEmpty, max, dim, tid); + threadGetTopK( + topK, beam, beamSize, val, col, firstStep, isEmpty, max, dim, tid); shTopK[tid] = topK[0]; - blockReduce - (shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp); + blockReduce( + shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp); } } -void hl_matrix_top_k(real* topVal, int ldv, - int * topIds, - real* src, int lds, +void hl_matrix_top_k(real* topVal, + int ldv, + int* topIds, + real* src, + int lds, int dim, int beamSize, int numSamples) { @@ -353,33 +364,32 @@ void hl_matrix_top_k(real* topVal, int ldv, dim3 threads(256, 1); dim3 grid(numSamples, 1); - KeMatrixTopK<5, 256><<< grid, threads, 0, STREAM_DEFAULT >>> - (topVal, ldv, topIds, src, lds, dim, beamSize); + KeMatrixTopK<5, 256><<>>( + topVal, ldv, topIds, src, lds, dim, beamSize); CHECK_SYNC("hl_matrix_top_k failed"); } -void hl_sparse_matrix_top_k(real* topVal, int ldv, - int * topIds, +void hl_sparse_matrix_top_k(real* topVal, + int ldv, + int* topIds, hl_sparse_matrix_s src, int beamSize, int numSamples) { CHECK_NOTNULL(topVal); CHECK_NOTNULL(topIds); CHECK_NOTNULL(src); - CHECK_EQ(src->format, HL_SPARSE_CSR) - <<"sparse matrix format error!"; + CHECK_EQ(src->format, HL_SPARSE_CSR) << "sparse matrix format error!"; hl_csr_matrix csr = (hl_csr_matrix)src->matrix; - if (csr->csr_val == NULL || csr->csr_row == NULL || - csr->csr_col == NULL) { + if (csr->csr_val == NULL || csr->csr_row == NULL || csr->csr_col == NULL) { LOG(FATAL) << "parameter src is null!"; } dim3 threads(256, 1); dim3 grid(numSamples, 1); - KeSMatrixTopK<5, 256><<< grid, threads, 0, STREAM_DEFAULT >>> - (topVal, ldv, topIds, csr->csr_val, csr->csr_row, csr->csr_col, beamSize); + KeSMatrixTopK<5, 256><<>>( + topVal, ldv, topIds, csr->csr_val, csr->csr_row, csr->csr_col, beamSize); CHECK_SYNC("hl_sparse_matrix_top_k failed"); } @@ -392,10 +402,12 @@ void hl_sparse_matrix_top_k(real* topVal, int ldv, * 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 KeMatrixTopKClassificationError(real* topVal, int ldv, - int * topIds, - real* src, int lds, +template +__global__ void KeMatrixTopKClassificationError(real* topVal, + int ldv, + int* topIds, + real* src, + int lds, int dim, int beamSize, int* label, @@ -408,7 +420,7 @@ __global__ void KeMatrixTopKClassificationError(real* topVal, int ldv, topVal += blockIdx.x * ldv; topIds += blockIdx.x * beamSize; - Pair topK[maxLength]; // NOLINT + Pair topK[maxLength]; // NOLINT int beam = maxLength; Pair max; bool isEmpty = false; @@ -420,34 +432,36 @@ __global__ void KeMatrixTopKClassificationError(real* topVal, int ldv, } while (beamSize) { - threadGetTopK - (topK, beam, beamSize, src, firstStep, isEmpty, max, dim, tid); + threadGetTopK( + topK, beam, beamSize, src, firstStep, isEmpty, max, dim, tid); shTopK[tid] = topK[0]; - blockReduce - (shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp); + blockReduce( + shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp); } __syncthreads(); if (tid == 0) { for (int i = 0; i < topkSize; i++) { - if (*--topIds == label[blockIdx.x]) { - recResult[blockIdx.x] = 0; - break; - } - recResult[blockIdx.x] = 1.0f; + if (*--topIds == label[blockIdx.x]) { + recResult[blockIdx.x] = 0; + break; + } + recResult[blockIdx.x] = 1.0f; } } } -void hl_matrix_classification_error(real* topVal, int ldv, - int* topIds, - real* src, int lds, - int dim, - int topkSize, - int numSamples, - int* label, - real* recResult) { +void hl_matrix_classification_error(real* topVal, + int ldv, + int* topIds, + real* src, + int lds, + int dim, + int topkSize, + int numSamples, + int* label, + real* recResult) { CHECK_NOTNULL(topVal); CHECK_NOTNULL(topIds); CHECK_NOTNULL(src); @@ -456,9 +470,8 @@ void hl_matrix_classification_error(real* topVal, int ldv, dim3 threads(256, 1); dim3 grid(numSamples, 1); - KeMatrixTopKClassificationError<5, 256> - <<< grid, threads, 0, STREAM_DEFAULT >>> - (topVal, ldv, topIds, src, lds, dim, topkSize, label, recResult); + KeMatrixTopKClassificationError<5, 256><<>>( + topVal, ldv, topIds, src, lds, dim, topkSize, label, recResult); CHECK_SYNC("hl_matrix_top_k classification error failed"); } diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt index 9c39430835d37d5dfbe4031f29e5a6216ed8b67f..68304c9fc8b8fa13cb1f99b82517abc87c71496c 100644 --- a/paddle/framework/CMakeLists.txt +++ b/paddle/framework/CMakeLists.txt @@ -7,38 +7,53 @@ cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory device_context) cc_test(tensor_test SRCS tensor_test.cc DEPS tensor) cc_test(eigen_test SRCS eigen_test.cc DEPS tensor) +cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor) +cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor) + 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(attr_type SRCS attr_type.proto) -proto_library(op_proto SRCS op_proto.proto DEPS attr_type) -proto_library(op_desc SRCS op_desc.proto DEPS attr_type) -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 framework_proto) -cc_library(operator SRCS operator.cc DEPS op_desc device_context tensor scope) +cc_library(operator SRCS operator.cc DEPS framework_proto device_context tensor scope attribute) 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 attr_type.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) +add_custom_command(TARGET framework_py_proto POST_BUILD + COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto + COMMAND cp *.py ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto/ + COMMENT "Copy generated python proto into directory paddle/v2/framework/proto." + 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) +cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context) + +if(WITH_PYTHON) cc_library(paddle_pybind SHARED SRCS pybind.cc DEPS pybind python backward - fc_op - sgd_op - add_op - mean_op - cross_entropy_op - recurrent_op) + 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) diff --git a/paddle/framework/attr_type.proto b/paddle/framework/attr_type.proto deleted file mode 100644 index 2d8e0476d710b7ba987d085d828ca13a4ee23707..0000000000000000000000000000000000000000 --- a/paddle/framework/attr_type.proto +++ /dev/null @@ -1,28 +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; - -// 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 diff --git a/paddle/framework/attribute.cc b/paddle/framework/attribute.cc new file mode 100644 index 0000000000000000000000000000000000000000..9eb07acdff1d00dd926f1cee9c24f9f151006d7e --- /dev/null +++ b/paddle/framework/attribute.cc @@ -0,0 +1,85 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/attribute.h" + +#include + +namespace paddle { +namespace framework { + +template <> +AttrType AttrTypeID() { + return INT; +} +template <> +AttrType AttrTypeID() { + return FLOAT; +} +template <> +AttrType AttrTypeID() { + return STRING; +} +template <> +AttrType AttrTypeID>() { + return INTS; +} +template <> +AttrType AttrTypeID>() { + return FLOATS; +} +template <> +AttrType AttrTypeID>() { + return STRINGS; +} + +Attribute GetAttrValue(const OpDesc::Attr& attr_desc) { + switch (attr_desc.type()) { + case paddle::framework::AttrType::INT: { + return attr_desc.i(); + } + case paddle::framework::AttrType::FLOAT: { + return attr_desc.f(); + } + case paddle::framework::AttrType::STRING: { + return attr_desc.s(); + } + case paddle::framework::AttrType::INTS: { + std::vector val(attr_desc.ints_size()); + for (int i = 0; i < attr_desc.ints_size(); ++i) { + val[i] = attr_desc.ints(i); + } + return val; + } + case paddle::framework::AttrType::FLOATS: { + std::vector val(attr_desc.floats_size()); + for (int i = 0; i < attr_desc.floats_size(); ++i) { + val[i] = attr_desc.floats(i); + } + return val; + } + case paddle::framework::AttrType::STRINGS: { + std::vector val(attr_desc.strings_size()); + for (int i = 0; i < attr_desc.strings_size(); ++i) { + val[i] = attr_desc.strings(i); + } + return val; + } + } + PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !"); + return boost::blank(); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/attr_checker.h b/paddle/framework/attribute.h similarity index 79% rename from paddle/framework/attr_checker.h rename to paddle/framework/attribute.h index ea5614a45f3a77a851358aff80abbc276c9972ba..08b47cabd4c2225c50022bd35734dcc2663324d6 100644 --- a/paddle/framework/attr_checker.h +++ b/paddle/framework/attribute.h @@ -1,12 +1,28 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + #pragma once -#include #include #include #include #include #include + +#include "paddle/framework/framework.pb.h" #include "paddle/platform/enforce.h" +#include "paddle/platform/variant.h" namespace paddle { namespace framework { @@ -14,13 +30,19 @@ namespace framework { typedef boost::variant, std::vector, std::vector> Attribute; + typedef std::unordered_map AttributeMap; +template +AttrType AttrTypeID(); + +Attribute GetAttrValue(const OpDesc::Attr& attr_desc); + // check whether a value(attribute) fit a certain limit template class LargerThanChecker { public: - LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {} + explicit LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {} void operator()(T& value) const { PADDLE_ENFORCE(value > lower_bound_, "larger_than check fail"); } @@ -35,7 +57,8 @@ class LargerThanChecker { template class DefaultValueSetter { public: - DefaultValueSetter(T default_value) : default_value_(default_value) {} + explicit DefaultValueSetter(T default_value) + : default_value_(default_value) {} void operator()(T& value) const { value = default_value_; } private: @@ -78,7 +101,8 @@ class TypedAttrChecker { typedef std::function ValueChecker; public: - TypedAttrChecker(const std::string& attr_name) : attr_name_(attr_name) {} + explicit TypedAttrChecker(const std::string& attr_name) + : attr_name_(attr_name) {} TypedAttrChecker& InEnum(const std::unordered_set& range) { value_checkers_.push_back(EnumInContainer(range)); diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc index c034e265fe4837ca22ab969b0e6952677904e05c..c226e4e3d2a58d1a647e204c4cd26f4eb6bcd968 100644 --- a/paddle/framework/backward.cc +++ b/paddle/framework/backward.cc @@ -13,32 +13,44 @@ limitations under the License. */ #include "paddle/framework/backward.h" + #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; +} + +// 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::shared_ptr NOP() { auto net_op = std::make_shared(); - net_op->type_ = "@NOP@"; + net_op->SetType("@NOP@"); net_op->CompleteAddOp(); return 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. // @@ -46,33 +58,35 @@ 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( 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_, OperatorBase::GRAD_VAR_SUFFIX(), - 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_, OperatorBase::GRAD_VAR_SUFFIX(), - no_grad_names)) { - for (auto& name : forwardOp.inputs_) { - // Mark all input is not need - no_grad_names.insert(name + OperatorBase::GRAD_VAR_SUFFIX()); - } + 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(); } @@ -84,71 +98,101 @@ std::shared_ptr BackwardRecursive( 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 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; + }); } // Get unique ID for this method. auto uid = uniq_id++; // TODO(dzh): more comment + // 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; 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); } - } else { std::shared_ptr grad_op = OpRegistry::CreateGradOp(forwardOp); - for (std::string& grad_input : grad_op->inputs_) { + + 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() - OperatorBase::GRAD_VAR_SUFFIX().size()); - grad_input = prefix + OperatorBase::ZERO_VAR_SUFFIX(); + 0, grad_input.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1); + 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 = OperatorBase::EMPTY_VAR_NAME(); + net->AddOp(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_op") { + // 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 + auto grad_stepnet = BackwardRecursive(stepnet_op, no_grad_names, uniq_id); + rnn_grad_op->set_stepnet( + std::static_pointer_cast(grad_stepnet)); } if (net->ops_.empty()) { // Current no aux op is added to network @@ -156,10 +200,10 @@ std::shared_ptr BackwardRecursive( } net->AddOp(grad_op); } - net->type_ = "@GENERATED_BACKWARD@"; + net->SetType("@GENERATED_BACKWARD@"); net->CompleteAddOp(); return net; -} +} // namespace framework // See header for comments std::shared_ptr Backward( @@ -168,14 +212,14 @@ std::shared_ptr Backward( std::unordered_set no_grad_names; no_grad_names.reserve(no_grad_vars.size()); - no_grad_names.insert(OperatorBase::EMPTY_VAR_NAME() + - OperatorBase::GRAD_VAR_SUFFIX()); + no_grad_names.insert(std::string(kEmptyVarName) + kGradVarSuffix); for (auto& name : no_grad_vars) { - no_grad_names.insert(name + OperatorBase::GRAD_VAR_SUFFIX()); + no_grad_names.insert(name + kGradVarSuffix); } size_t uid = 0; return BackwardRecursive(forwardOp, no_grad_names, uid); } + } // namespace framework } // namespace paddle diff --git a/paddle/framework/backward_test.cc b/paddle/framework/backward_test.cc index 8f437e68041188831a17217099e0b0c96432cda4..d942604bf05998ab9e1ee147b28586e7e4e9777d 100644 --- a/paddle/framework/backward_test.cc +++ b/paddle/framework/backward_test.cc @@ -17,25 +17,24 @@ #include #include "paddle/framework/op_registry.h" #include "paddle/operators/net_op.h" -#include "paddle/operators/type_alias.h" namespace paddle { namespace framework { -class EmptyOp : public OperatorBase { - public: - void InferShape(const Scope &scope) const override {} - void Run(const Scope &scope, - const platform::DeviceContext &dev_ctx) const override {} -}; +using OperatorBase = framework::OperatorBase; +using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker; +using OpProto = framework::OpProto; +using OpAttrChecker = framework::OpAttrChecker; +using Scope = framework::Scope; +using DeviceContext = platform::DeviceContext; class RowWiseAddOpMaker : public OpProtoAndCheckerMaker { public: RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "Input X of Add").IgnoreGradient(); - AddInput("b", "Bias of Add").IgnoreGradient(); - AddOutput("Out", "Out of Add").IgnoreGradient(); + AddInput("X", "Input X of Add").AsNoGradient(); + AddInput("b", "Bias of Add").AsNoGradient(); + AddOutput("Out", "Out of Add").AsNoGradient(); AddComment("Add Op"); } }; @@ -44,8 +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"); } @@ -56,7 +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"); } }; @@ -66,31 +65,35 @@ 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 ops::NetOp { +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 VarNameMap &inputs, + const VarNameMap &outputs, const AttributeMap &attrs) + : NetOp(type, inputs, outputs, attrs) { + AddOp(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 != EMPTY_VAR_NAME()) { - AddOp(OpRegistry::CreateOp("rowwise_add", {Output("mul_result"), b_name}, - {Output("add_result")}, {})); + if (input_b.size() != 0) { + AddOp(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"); - if (out_varname != EMPTY_VAR_NAME()) { - this->Rename(out_varname, EMPTY_VAR_NAME()); + if (out_varname != kEmptyVarName) { + this->Rename(out_varname, kEmptyVarName); } } - AddOp(OpRegistry::CreateOp("sigmoid", {Output(before_act)}, {Output("Out")}, - {})); + AddOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}}, + {{"Out", {Output("Out")}}}, {})); CompleteAddOp(false); } }; @@ -102,8 +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(""); } @@ -134,7 +137,7 @@ class AddOpMaker : public OpProtoAndCheckerMaker { public: AddOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "x").SetMultiple(); + AddInput("X", "x").AsDuplicable(); AddOutput("Y", "y"); AddComment(""); } @@ -143,44 +146,38 @@ class AddOpMaker : public OpProtoAndCheckerMaker { } // namespace paddle 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::OperatorBase::EMPTY_VAR_NAME(), gop->inputs_[0]); - ASSERT_EQ("rowwise_add_grad", gop->type_); - ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->outputs_[0]); - ASSERT_EQ("b" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->outputs_[1]); - - ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(), - gop->Output("X" + f::OperatorBase::GRAD_VAR_SUFFIX())); + 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(), - "X" + f::OperatorBase::GRAD_VAR_SUFFIX()), - 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, @@ -188,8 +185,12 @@ TEST(Backward, simple_op_not_need_grad) { } 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()); @@ -200,19 +201,22 @@ 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::OperatorBase::EMPTY_VAR_NAME()}, - {"mul_result", "add_result", "tmp"}, {}); + std::shared_ptr fwd = + 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()); @@ -223,168 +227,176 @@ 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.AddOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"x"}}, {"W", {"W1"}}, {"b", {"b1"}}}, + {{"mul_result", {"mul_tmp_0"}}, + {"add_result", {"add_tmp_0"}}, + {"Out", {"hidden0"}}}, + {})); + net.AddOp(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::OperatorBase::EMPTY_VAR_NAME()); + 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(out + f::OperatorBase::GRAD_VAR_SUFFIX()), - all_output.end()); + ASSERT_NE(all_outputs.find(f::GradVarName(out)), all_outputs.end()); } // Not Generated X - ASSERT_EQ(all_output.find("X" + f::OperatorBase::GRAD_VAR_SUFFIX()), - 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::OperatorBase::EMPTY_VAR_NAME(), - first_fc_grad->ops_[2]->Output("A" + f::OperatorBase::GRAD_VAR_SUFFIX())); + ASSERT_EQ(f::kEmptyVarName, + 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.AddOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}}, + {{"Out", {"out"}}}, {})); + net.AddOp(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("Z" + f::OperatorBase::ZERO_VAR_SUFFIX(), 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("Z" + f::OperatorBase::ZERO_VAR_SUFFIX(), - d_many_out.Input("z" + f::OperatorBase::GRAD_VAR_SUFFIX())); - ASSERT_EQ("Y" + f::OperatorBase::GRAD_VAR_SUFFIX(), - d_many_out.Input("y" + f::OperatorBase::GRAD_VAR_SUFFIX())); - ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(), - d_many_out.Output("x" + f::OperatorBase::GRAD_VAR_SUFFIX())); + 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"))); + ASSERT_EQ(f::GradVarName("X"), d_many_out.Output(f::GradVarName("x"))); } 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("A" + f::OperatorBase::GRAD_VAR_SUFFIX()), - f::OperatorBase::EMPTY_VAR_NAME()); - ASSERT_EQ(grad_mul.Output("B" + f::OperatorBase::GRAD_VAR_SUFFIX()), - "b" + f::OperatorBase::GRAD_VAR_SUFFIX()); - ASSERT_EQ(grad_mul.Input("Out" + f::OperatorBase::GRAD_VAR_SUFFIX()), - "out" + f::OperatorBase::GRAD_VAR_SUFFIX()); - ASSERT_EQ(grad_mul.Input("A"), "a"); - ASSERT_EQ(grad_mul.Input("B"), "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("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.AddOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"x1"}}, {"W", {"w1"}}, {"b", {"b1"}}}, + {{"mul_result", {"mul_out1"}}, + {"add_result", {"add_out1"}}, + {"Out", {"out1"}}}, + {})); + net.AddOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"out1"}}, {"W", {"w2"}}, {"b", {"b2"}}}, + {{"mul_result", {"mul_out2"}}, + {"add_result", {"tmp_out2"}}, + {"Out", {"out2"}}}, + {})); + net.AddOp(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.Output("X" + f::OperatorBase::GRAD_VAR_SUFFIX()), - f::OperatorBase::EMPTY_VAR_NAME()); - EXPECT_EQ(grad_fc.Output("W" + f::OperatorBase::GRAD_VAR_SUFFIX()), - "w3" + f::OperatorBase::GRAD_VAR_SUFFIX()); - EXPECT_EQ(grad_fc.Output("b" + f::OperatorBase::GRAD_VAR_SUFFIX()), - "b3" + f::OperatorBase::GRAD_VAR_SUFFIX()); - EXPECT_EQ(grad_fc.Output("mul_result" + f::OperatorBase::GRAD_VAR_SUFFIX()), - "mul_out3" + f::OperatorBase::GRAD_VAR_SUFFIX()); - - EXPECT_EQ(grad_fc.Input("Out" + f::OperatorBase::GRAD_VAR_SUFFIX()), - "out3" + f::OperatorBase::GRAD_VAR_SUFFIX()); - EXPECT_EQ(grad_fc.Input("X"), "out2"); - EXPECT_EQ(grad_fc.Input("W"), "w3"); - EXPECT_EQ(grad_fc.Input("mul_result"), "mul_out3"); - EXPECT_EQ(grad_fc.Input("add_result"), "tmp_out3"); - EXPECT_EQ(grad_fc.Input("Out"), "out3"); - */ + 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..cfd3e8dfdec0e92620aef5cd246b4622b779ce19 100644 --- a/paddle/framework/ddim.cc +++ b/paddle/framework/ddim.cc @@ -283,6 +283,5 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) { DDim::DDim(std::initializer_list init_list) { *this = make_ddim(init_list); } - } // namespace framework } // namespace paddle diff --git a/paddle/framework/ddim.h b/paddle/framework/ddim.h index 9fcc657edcd5459d0a42a64d708603a4bcd53cf0..95f294b62737be5c3eac39303148ac35da29fe7d 100644 --- a/paddle/framework/ddim.h +++ b/paddle/framework/ddim.h @@ -14,29 +14,25 @@ limitations under the License. */ #pragma once -#include #include #include #include #include "paddle/framework/dim.h" #include "paddle/platform/enforce.h" -#include "unsupported/Eigen/CXX11/Tensor" +#include "paddle/platform/variant.h" namespace paddle { namespace framework { -namespace { -typedef boost::variant, Dim<2>, Dim<3>, Dim<4>, Dim<5>, Dim<6>, Dim<7>, - Dim<8>, Dim<9>> - DDimVar; -} - /** * \brief A dynamically sized dimension. * * The number of dimensions must be between [1, 9]. */ struct DDim { + typedef boost::variant, Dim<2>, Dim<3>, Dim<4>, Dim<5>, Dim<6>, Dim<7>, + Dim<8>, Dim<9>> + DDimVar; DDimVar var; DDim() : var(Dim<1>()) {} diff --git a/paddle/framework/framework.proto b/paddle/framework/framework.proto new file mode 100644 index 0000000000000000000000000000000000000000..7077e8aa2c77c24efdbb34ed3a13821fe7678455 --- /dev/null +++ b/paddle/framework/framework.proto @@ -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. */ + +syntax = "proto2"; +package paddle.framework; + +enum AttrType { + INT = 0; + FLOAT = 1; + STRING = 2; + INTS = 3; + FLOATS = 4; + STRINGS = 5; +} + +// 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; + }; + + 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 no_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; +} diff --git a/paddle/framework/grad_op_builder.cc b/paddle/framework/grad_op_builder.cc index ea5e939c6e26514c2f3c515da5581b29103f75b6..b73dac22d029876de9a012de533647db3dd74cbb 100644 --- a/paddle/framework/grad_op_builder.cc +++ b/paddle/framework/grad_op_builder.cc @@ -13,92 +13,53 @@ 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 { - -class OpRegistry; - -using VarIndexMap = std::unordered_map; - enum class OpArgType { IN, OUT }; -static std::vector* GetOpFormat(OperatorBase* op, const OpArgType& type) { - std::string key = type == OpArgType::IN ? "input_format" : "output_format"; - return op->attrs_.count(key) - ? &boost::get>(op->attrs_.at(key)) - : nullptr; -} - -static const std::vector* GetOpFormat(const OperatorBase* op, - const OpArgType& type) { - std::string key = type == OpArgType::IN ? "input_format" : "output_format"; - return op->attrs_.count(key) - ? &boost::get>(op->attrs_.at(key)) - : nullptr; -} - -static void TransOpArg(const OperatorBase* src_op, OperatorBase* dst_op, - const OpArgType& src_type, const OpArgType& dst_type, - int& idx, bool is_grad) { - const std::vector& src_inout = - src_type == OpArgType::IN ? src_op->inputs_ : src_op->outputs_; - const std::vector* src_format = GetOpFormat(src_op, src_type); - - std::vector& dst_inout = - dst_type == OpArgType::IN ? dst_op->inputs_ : dst_op->outputs_; - std::vector* dst_format = GetOpFormat(dst_op, 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, OperatorBase::VarNameMap* vars) { + const auto& src_inout = + src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs(); + auto& dst_inout = *vars; + const OpProto* proto = OpRegistry::op_info_map().at(src_op->Type()).proto_; const auto& src_arg_list = - src_type == OpArgType::IN ? 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 + OperatorBase::GRAD_VAR_SUFFIX() : src_name; - (*dst_op->in_out_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] + OperatorBase::GRAD_VAR_SUFFIX() - : arg.ignore_gradient() - ? OperatorBase::EMPTY_VAR_NAME() - : src_inout[i]; - dst_inout.emplace_back(s); - } - if (dst_format != nullptr) { - dst_format->push_back(dst_inout.size()); + if (arg.no_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) { - std::string grad_op_type = OpRegistry::grad_ops().at(op->type_); - OperatorBase* grad_op = OpRegistry::op_creators().at(grad_op_type)(); - grad_op->type_ = grad_op_type; - grad_op->attrs_ = op->attrs_; - grad_op->attrs_.erase("input_format"); - grad_op->attrs_.erase("output_format"); - if (GetOpFormat(op, OpArgType::IN) != nullptr) { - grad_op->attrs_["output_format"] = std::vector({0}); - } - if (GetOpFormat(op, OpArgType::IN) != nullptr || - GetOpFormat(op, OpArgType::OUT) != nullptr) { - grad_op->attrs_["input_format"] = std::vector({0}); - } - grad_op->in_out_idxs_.reset(new VarIndexMap()); - int in_idx = 0; - int out_idx = 0; - TransOpArg(op, grad_op, OpArgType::IN, OpArgType::IN, in_idx, false); // I - TransOpArg(op, grad_op, OpArgType::OUT, OpArgType::IN, in_idx, false); // G - TransOpArg(op, grad_op, OpArgType::OUT, OpArgType::IN, in_idx, true); // OG - TransOpArg(op, grad_op, OpArgType::IN, OpArgType::OUT, out_idx, true); // IG - return grad_op; + auto it = OpRegistry::op_info_map().find(op->Type()); + PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(), + "'%s' has not been registered.", op->Type()); + PADDLE_ENFORCE(it->second.proto_ != nullptr, "'%s' has no OpProto.", + op->Type()); + std::string grad_op_type = it->second.grad_op_type_; + PADDLE_ENFORCE(!grad_op_type.empty(), "'%s' has no gradient operator.", + op->Type()); + + OperatorBase::VarNameMap inputs; + OperatorBase::VarNameMap 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 + + it = OpRegistry::op_info_map().find(grad_op_type); + PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(), + "'%s' has not been registered.", grad_op_type); + return it->second.creator_(grad_op_type, inputs, outputs, op->Attrs()); } } // namespace framework diff --git a/paddle/framework/grad_op_builder.h b/paddle/framework/grad_op_builder.h index cf235de6c267a4a1feb7afd3e4dbe7a6a668ee5e..998f8ebbb5f2f4fb8b7e938b5916afd0f8a7930d 100644 --- a/paddle/framework/grad_op_builder.h +++ b/paddle/framework/grad_op_builder.h @@ -1,3 +1,17 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + #pragma once #include "paddle/framework/operator.h" diff --git a/paddle/framework/grad_op_builder_test.cc b/paddle/framework/grad_op_builder_test.cc index 96d7f309d67b15c000ab8ce3769931322fbca880..0c26293fd29d24a7a40c47bdf055d2758846709b 100644 --- a/paddle/framework/grad_op_builder_test.cc +++ b/paddle/framework/grad_op_builder_test.cc @@ -8,22 +8,15 @@ USE_OP(add_two); namespace paddle { namespace framework { -class NOP : public OperatorBase { - public: - 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"); } }; @@ -33,10 +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().AsNoGradient(); + AddInput("In3_mult", "another multiple input").AsDuplicable(); + AddOutput("Out1_mult", "a multiple output").AsDuplicable(); + AddOutput("Out2", "a single output").AsNoGradient(); AddComment("op with inputs and outputs ignored in gradient calculating"); } }; @@ -47,35 +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_two", {{"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"})); @@ -83,63 +74,49 @@ TEST(GradOpBuilder, MutiInOut) { EXPECT_EQ(grad_test_op->Input("Out1"), "out1"); EXPECT_EQ(grad_test_op->Inputs("Out2_mult"), std::vector({"out2_1", "out2_2"})); - EXPECT_EQ(grad_test_op->Input("Out1" + f::OperatorBase::GRAD_VAR_SUFFIX()), - "out1" + f::OperatorBase::GRAD_VAR_SUFFIX()); - EXPECT_EQ( - grad_test_op->Inputs("Out2_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()), - std::vector( - {"out2_1" + f::OperatorBase::GRAD_VAR_SUFFIX(), - "out2_2" + f::OperatorBase::GRAD_VAR_SUFFIX()})); + EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out1")), + f::GradVarName("out1")); + EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out2_mult")), + std::vector( + {f::GradVarName("out2_1"), f::GradVarName("out2_2")})); - ASSERT_EQ(grad_test_op->outputs_.size(), 5UL); - EXPECT_EQ(grad_test_op->Output("In1" + f::OperatorBase::GRAD_VAR_SUFFIX()), - "in1" + f::OperatorBase::GRAD_VAR_SUFFIX()); - EXPECT_EQ( - grad_test_op->Outputs("In2_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()), - std::vector({"in2_1" + f::OperatorBase::GRAD_VAR_SUFFIX(), - "in2_2" + f::OperatorBase::GRAD_VAR_SUFFIX(), - "in2_3" + f::OperatorBase::GRAD_VAR_SUFFIX()})); - EXPECT_EQ(grad_test_op->Output("In3" + f::OperatorBase::GRAD_VAR_SUFFIX()), - "in3" + f::OperatorBase::GRAD_VAR_SUFFIX()); + ASSERT_EQ(grad_test_op->Outputs().size(), 3UL); + EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1")); + EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")), + std::vector({f::GradVarName("in2_1"), + f::GradVarName("in2_2"), + f::GradVarName("in2_3")})); + EXPECT_EQ(grad_test_op->Output(f::GradVarName("In3")), f::GradVarName("in3")); } TEST(GradOpBuilder, IOIgnoredInGradient) { - 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::OperatorBase::EMPTY_VAR_NAME(), - f::OperatorBase::EMPTY_VAR_NAME()})); 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::OperatorBase::EMPTY_VAR_NAME()); - EXPECT_EQ( - grad_test_op->Inputs("Out1_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()), - std::vector( - {"out1_1" + f::OperatorBase::GRAD_VAR_SUFFIX(), - "out1_2" + f::OperatorBase::GRAD_VAR_SUFFIX()})); - EXPECT_EQ(grad_test_op->Input("Out2" + f::OperatorBase::GRAD_VAR_SUFFIX()), - "out2" + f::OperatorBase::GRAD_VAR_SUFFIX()); + 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); - EXPECT_EQ(grad_test_op->Output("In1" + f::OperatorBase::GRAD_VAR_SUFFIX()), - "in1" + f::OperatorBase::GRAD_VAR_SUFFIX()); - EXPECT_EQ( - grad_test_op->Outputs("In2_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()), - std::vector({"in2_1" + f::OperatorBase::GRAD_VAR_SUFFIX(), - "in2_2" + f::OperatorBase::GRAD_VAR_SUFFIX()})); - EXPECT_EQ( - grad_test_op->Outputs("In3_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()), - std::vector({"in3_1" + f::OperatorBase::GRAD_VAR_SUFFIX(), - "in3_2" + f::OperatorBase::GRAD_VAR_SUFFIX()})); + ASSERT_EQ(grad_test_op->Outputs().size(), 3UL); + EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1")); + EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")), + std::vector( + {f::GradVarName("in2_1"), f::GradVarName("in2_2")})); + EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In3_mult")), + std::vector( + {f::GradVarName("in3_1"), f::GradVarName("in3_2")})); } diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc new file mode 100644 index 0000000000000000000000000000000000000000..2b178907747b3911292b070b65160a24c120b726 --- /dev/null +++ b/paddle/framework/lod_tensor.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/framework/lod_tensor.h" + +#include + +namespace paddle { +namespace framework { + +LODTensor::LOD LODTensor::LOD::SliceLevels(size_t level_begin, + size_t level_end) const { + 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)); + } + return new_lod; +} + +LODTensor::LOD LODTensor::LOD::SliceInLevel(size_t level, size_t elem_begin, + size_t elem_end) const { + // 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]; + + for (auto it = this->begin() + level; it != this->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"); + PADDLE_ENFORCE(it_end != it->end(), "error in parsing lod info"); + new_lod.emplace_back(it_begin, it_end + 1); + // reset offset if tensor is copyed and sliced. + 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_LE(new_lod.size(), this->size()); + return new_lod; +} + +bool operator==(const LODTensor::LOD& a, const LODTensor::LOD& b) { + if (a.size() != b.size()) { + return false; + } + + for (size_t i = 0; i < a.size(); i++) { + const auto& a_level = a[i]; + const auto& b_level = b[i]; + if (a_level.size() != b_level.size()) { + return false; + } + for (size_t j = 0; j < a_level.size(); j++) { + if (a_level[j] != b_level[j]) { + return false; + } + } + } + + return true; +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h new file mode 100644 index 0000000000000000000000000000000000000000..9e27aec38d336db8a4f0adbed098d299aa741356 --- /dev/null +++ b/paddle/framework/lod_tensor.h @@ -0,0 +1,149 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT 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 +#if !defined(PADDLE_ONLY_CPU) +#include +#include +#endif + +#include "paddle/framework/ddim.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/enforce.h" + +namespace paddle { +namespace framework { + +/* + * LODTensor (Level of details Tensor) + * see https://en.wikipedia.org/wiki/Level_of_details for reference. + */ +class LODTensor : public Tensor { + 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() {} + explicit LODTensor(const LOD &lod) : lod_(lod) {} + + virtual Tensor *Clone() const { return new LODTensor(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, + NumLevels()); + PADDLE_ENFORCE(elem < NumElements(level), + "element begin [%d] out of range [%d]", elem, + NumElements(level)); + return (lod_)[level][elem]; + } + + /* + * 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(); } + /* + * Number of elements in a level. + */ + size_t NumElements(size_t level = 0) const { + PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level, + NumLevels()); + // the last offset is the end of last element + return lod_[level].size() - 1; + } + + /* + * Slice of levels[level_begin:level_end], with tensor shared. + */ + template + LODTensor SliceLevels(size_t level_begin, size_t level_end) const; + + /* + * Slice of elements of a level, [elem_begin: elem_end], with tensor shared. + * @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() {} + + private: + LOD lod_; +}; + +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_test.cc b/paddle/framework/lod_tensor_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..2881136ced6ef957a192e303e529b9b2867b3dda --- /dev/null +++ b/paddle/framework/lod_tensor_test.cc @@ -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/lod_tensor.h" + +#include +#include +#include +#include + +namespace paddle { +namespace framework { + +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.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}); + + ASSERT_EQ(lod.size(), 3UL); + + tensor.Resize({20 /*batch size*/, 128 /*dim*/}); + // 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); + } + + protected: + std::unique_ptr lod_tensor; + platform::CPUPlace place; + Tensor tensor; +}; + +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, SliceLevels) { + // slice 1 level + for (size_t level = 0; level < 3UL; ++level) { + auto new_lod_tensor = 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); + } + // slice 2 level + for (size_t level = 0; level < 2UL; ++level) { + auto new_lod_tensor = 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()); + } +} + +TEST_F(LODTensorTester, SliceInLevel) { + size_t level = 0; + auto new_lod_tensor = 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()); + + level = 1; + new_lod_tensor = 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); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/op_desc.proto b/paddle/framework/op_desc.proto deleted file mode 100644 index 89497f3c16bc28aa93b25a83c1f2eccafdf1c5b4..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 "attr_type.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_proto.proto b/paddle/framework/op_proto.proto deleted file mode 100644 index 366c84e53dc29e41eefbaef0a6452e01c4fe37bd..0000000000000000000000000000000000000000 --- a/paddle/framework/op_proto.proto +++ /dev/null @@ -1,114 +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 "attr_type.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 1d14535c50b542733663a6900a8b5f2033290ea6..1caa02a2a1d046778f875d04eeaef957be741302 100644 --- a/paddle/framework/op_registry.cc +++ b/paddle/framework/op_registry.cc @@ -14,37 +14,8 @@ limitations under the License. */ #include -namespace paddle { -namespace framework { - -template <> -void AttrTypeHelper::SetAttrType(AttrProto* attr) { - attr->set_type(paddle::framework::AttrType::INT); -} - -template <> -void AttrTypeHelper::SetAttrType(AttrProto* attr) { - attr->set_type(paddle::framework::AttrType::FLOAT); -} - -template <> -void AttrTypeHelper::SetAttrType(AttrProto* attr) { - attr->set_type(paddle::framework::AttrType::STRING); -} +#include -template <> -void AttrTypeHelper::SetAttrType>(AttrProto* attr) { - attr->set_type(paddle::framework::AttrType::INTS); -} - -template <> -void AttrTypeHelper::SetAttrType>(AttrProto* attr) { - attr->set_type(paddle::framework::AttrType::FLOATS); -} - -template <> -void AttrTypeHelper::SetAttrType>(AttrProto* attr) { - attr->set_type(paddle::framework::AttrType::STRINGS); -} -} // namespace framework +namespace paddle { +namespace framework {} // namespace framework } // namespace paddle diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h index 9a975185f04da8df5ba22e457936218756e7c4bc..120f4ede6ba21e31683a8d19f0b39072c3f5c309 100644 --- a/paddle/framework/op_registry.h +++ b/paddle/framework/op_registry.h @@ -17,59 +17,18 @@ limitations under the License. */ #include #include #include +#include #include #include -#include "paddle/framework/attr_checker.h" +#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/operator.h" #include "paddle/framework/scope.h" namespace paddle { namespace framework { -// helper class to set attribute type -struct AttrTypeHelper { - template - static void SetAttrType(AttrProto* attr); - - static Attribute GetAttrValue(const AttrDesc& attr_desc) { - switch (attr_desc.type()) { - case paddle::framework::AttrType::INT: { - return attr_desc.i(); - } - case paddle::framework::AttrType::FLOAT: { - return attr_desc.f(); - } - case paddle::framework::AttrType::STRING: { - return attr_desc.s(); - } - case paddle::framework::AttrType::INTS: { - std::vector val(attr_desc.ints_size()); - for (int i = 0; i < attr_desc.ints_size(); ++i) { - val[i] = attr_desc.ints(i); - } - return val; - } - case paddle::framework::AttrType::FLOATS: { - std::vector val(attr_desc.floats_size()); - for (int i = 0; i < attr_desc.floats_size(); ++i) { - val[i] = attr_desc.floats(i); - } - return val; - } - case paddle::framework::AttrType::STRINGS: { - std::vector val(attr_desc.strings_size()); - for (int i = 0; i < attr_desc.strings_size(); ++i) { - val[i] = attr_desc.strings(i); - } - return val; - } - } - PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !"); - return boost::blank(); - } -}; - // this class not only make proto but also init attribute checkers. class OpProtoAndCheckerMaker { public: @@ -87,111 +46,58 @@ class OpProtoAndCheckerMaker { protected: struct VariableBuilder { - VarProto* var_; - std::function on_multiple_; - std::function on_temporary_; + OpProto::Var* var_; - VariableBuilder& SetMultiple() { - var_->set_multiple(true); - on_multiple_(); + VariableBuilder& AsDuplicable() { + var_->set_duplicable(true); return *this; } - VariableBuilder& SetTemporary() { - PADDLE_ENFORCE(bool(on_temporary_), "Cannot set temporary"); - var_->set_temporary(true); - on_temporary_(); + VariableBuilder& AsIntermediate() { + var_->set_intermediate(true); return *this; } - VariableBuilder& IgnoreGradient() { - var_->set_ignore_gradient(true); + // TODO(FengJiayi, yuyang18): `AsNoGradient` is a very bad name, because it + // means that input/output is not needed when calculate gradient. It does + // not mean no gradient when backward. It should be changed soon. + VariableBuilder& AsNoGradient() { + var_->set_no_gradient(true); return *this; } }; VariableBuilder AddInput(const std::string& name, const std::string& comment) { - auto input = proto_->mutable_inputs()->Add(); - *input->mutable_name() = name; - *input->mutable_comment() = comment; - return VariableBuilder{input, [=] { this->SetHasMultipleInput(); }, - nullptr}; + auto* input = proto_->add_inputs(); + input->set_name(name); + input->set_comment(comment); + return VariableBuilder{input}; } VariableBuilder AddOutput(const std::string& name, const std::string& comment) { - auto output = proto_->mutable_outputs()->Add(); - *output->mutable_name() = name; - *output->mutable_comment() = comment; - return VariableBuilder{output, [=] { this->SetHasMultipleOutput(); }, - [=] { this->SetHasTemporaryOutput(); }}; + auto* output = proto_->add_outputs(); + output->set_name(name); + output->set_comment(comment); + return VariableBuilder{output}; } template TypedAttrChecker& AddAttr(const std::string& name, const std::string& comment, bool generated = false) { - auto attr = proto_->mutable_attrs()->Add(); - *attr->mutable_name() = name; - *attr->mutable_comment() = comment; + auto* attr = proto_->add_attrs(); + attr->set_name(name); + attr->set_comment(comment); attr->set_generated(generated); - AttrTypeHelper::SetAttrType(attr); + attr->set_type(AttrTypeID()); return op_checker_->AddAttrChecker(name); } - void AddComment(const std::string& comment) { - *(proto_->mutable_comment()) = comment; - } + 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) { @@ -212,92 +118,93 @@ Add a mark to which output is temporary is helpful for future optimization. OpProto* proto_; OpAttrChecker* op_checker_; bool validated_{false}; - bool has_multiple_input_{false}; - bool has_multiple_output_{false}; - bool has_temporary_output_{false}; +}; + +class NOPMaker : public OpProtoAndCheckerMaker { + public: + NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) {} }; class OpRegistry { - using OpCreator = std::function; - using VarIndexMap = std::unordered_map; - using VarNameList = std::vector; + using VarNameMap = OperatorBase::VarNameMap; + using OpCreator = std::function; 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.mutable_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++; + struct OpInfo { + OpCreator creator_; + std::string grad_op_type_; + OpProto* proto_; + OpAttrChecker* checker_; + }; + + template + static void RegisterOp(const std::string& op_type, + const std::string& grad_op_type) { + PADDLE_ENFORCE(op_info_map().count(op_type) == 0, + "'%s' is registered more than once.", op_type); + OpInfo op_info; + op_info.creator_ = [](const std::string& type, const VarNameMap& inputs, + const VarNameMap& 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; } - idx = 0; - for (auto& var : op_proto.outputs()) { - varmap[var.name()] = idx++; + op_info_map().insert(std::make_pair(op_type, op_info)); + // register gradient op + if (!grad_op_type.empty()) { + RegisterOp(grad_op_type, ""); } } - 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 VarNameMap& inputs, + const VarNameMap& outputs, + AttributeMap attrs) { + auto it = op_info_map().find(type); + PADDLE_ENFORCE(it != op_info_map().end(), + "Operator '%s' has not been registered.", type); + it->second.checker_->Check(attrs); + auto op = it->second.creator_(type, inputs, outputs, attrs); + return std::shared_ptr(op); } - 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; - } + static VarNameMap ConvertOpDescVarsToVarNameMap( + const google::protobuf::RepeatedPtrField& op_desc_vars) { + VarNameMap 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)); } - - op->Init(); - return std::shared_ptr(op); + return ret_val; } 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)); - + VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs()); + VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs()); AttributeMap attrs; for (auto& attr : op_desc.attrs()) { - attrs[attr.name()] = AttrTypeHelper::GetAttrValue(attr); + attrs[attr.name()] = GetAttrValue(attr); } return CreateOp(op_desc.type(), inputs, outputs, attrs); @@ -307,62 +214,45 @@ class OpRegistry { 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& op_info_map() { + static std::unordered_map op_info_map_; + return op_info_map_; } +}; - 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_; - } - - private: - static std::unordered_map& op_checkers() { - static std::unordered_map op_checkers_; - return op_checkers_; - }; - - static void GenerateTempVariableName(OperatorBase* op) { - static std::atomic gUniqId(0UL); - for (auto& outname : op->outputs_) { - if (outname == OperatorBase::TMP_VAR_NAME()) { - outname += op->type_; - outname += "@"; - outname += std::to_string(gUniqId.fetch_add(1)); - } - } - } +class Registrar { + public: + // In our design, various kinds of classes, e.g., operators and kernels, + // have their corresponding registry and registrar. The action of + // registration is in the constructor of a global registrar variable, which, + // however, are not used in the code that calls package framework, and would + // be removed from the generated binary file by the linker. To avoid such + // removal, we add Touch to all registrar classes and make USE_OP macros to + // call this method. So, as long as the callee code calls USE_OP, the global + // registrar variable won't be removed by the linker. + void Touch() {} }; -template -class OpRegisterHelper { +template +class OpRegistrar : public Registrar { public: - OpRegisterHelper(const char* op_type) { - OpRegistry::RegisterOp(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); } }; -template -class GradOpRegisterHelper { +template +class OpKernelRegistrar : public Registrar { public: - GradOpRegisterHelper(const char* op_type, const char* grad_op_type) { - OpRegistry::RegisterGradOp(op_type, grad_op_type); + explicit OpKernelRegistrar(const char* op_type) { + OperatorWithKernel::OpKernelKey key; + key.place_ = PlaceType(); + OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KernelType); } }; @@ -376,97 +266,81 @@ class GradOpRegisterHelper { msg) /** - * Macro to Register Operator. + * Macro to register Operator. */ -#define REGISTER_OP(__op_type, __op_class, __op_maker_class) \ - STATIC_ASSERT_GLOBAL_NAMESPACE(__reg_op__##__op_type, \ - "REGISTER_OP must be in global namespace"); \ - static ::paddle::framework::OpRegisterHelper<__op_class, __op_maker_class> \ - __op_register_##__op_type##__(#__op_type); \ - int __op_register_##__op_type##_handle__() { 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 in global namespace"); \ - static ::paddle::framework::GradOpRegisterHelper<__grad_op_class> \ - __op_gradient_register_##__op_type##__grad_op_type##__(#__op_type, \ - #__grad_op_type); \ - int __op_gradient_register_##__op_type##__grad_op_type##_handle__() { \ - return 0; \ +#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, #grad_op_type); \ + int TouchOpRegistrar_##op_type() { \ + __op_registrar_##op_type##__.Touch(); \ + 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 in global namespace") +#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. + * Macro to register OperatorKernel. */ -#define REGISTER_OP_KERNEL(type, DEVICE_TYPE, PlaceType, ...) \ +#define REGISTER_OP_KERNEL(op_type, DEVICE_TYPE, place_class, ...) \ STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __reg_op_kernel_##type##_##DEVICE_TYPE##__, \ - "REGISTER_OP_KERNEL must be in global namespace"); \ - struct __op_kernel_register__##type##__##DEVICE_TYPE##__ { \ - __op_kernel_register__##type##__##DEVICE_TYPE##__() { \ - ::paddle::framework::OperatorWithKernel::OpKernelKey key; \ - key.place_ = PlaceType(); \ - ::paddle::framework::OperatorWithKernel::AllOpKernels()[#type][key] \ - .reset(new __VA_ARGS__()); \ - } \ - }; \ - static __op_kernel_register__##type##__##DEVICE_TYPE##__ \ - __reg_kernel_##type##__##DEVICE_TYPE##__; \ - int __op_kernel_register_##type##_handle_##DEVICE_TYPE##__() { return 0; } - -// (type, KernelType) -#define REGISTER_OP_GPU_KERNEL(type, ...) \ - REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__) - -// (type, KernelType) -#define REGISTER_OP_CPU_KERNEL(type, ...) \ - REGISTER_OP_KERNEL(type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__) + __reg_op_kernel_##op_type##_##DEVICE_TYPE##__, \ + "REGISTER_OP_KERNEL must be called in global namespace"); \ + static ::paddle::framework::OpKernelRegistrar \ + __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__(#op_type); \ + int TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE() { \ + __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__.Touch(); \ + return 0; \ + } + +#define REGISTER_OP_GPU_KERNEL(op_type, ...) \ + REGISTER_OP_KERNEL(op_type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__) + +#define REGISTER_OP_CPU_KERNEL(op_type, ...) \ + 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 * link them into target. */ -#define USE_OP_WITHOUT_KERNEL(op_type) \ - STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __use_op_without_kernel_##op_type, \ - "USE_OP_WITHOUT_KERNEL must be in global namespace"); \ - extern int __op_register_##op_type##_handle__(); \ - static int __use_op_ptr_##op_type##_without_kernel__ \ - __attribute__((unused)) = __op_register_##op_type##_handle__() - -#define USE_OP_KERNEL(op_type, DEVICE_TYPE) \ - STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __use_op_kernel_##op_type##_##DEVICE_TYPE##__, \ - "USE_OP_KERNEL must be in global namespace"); \ - extern int __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__(); \ - static int __use_op_ptr_##op_type##_##DEVICE_TYPE##_kernel__ \ - __attribute__((unused)) = \ - __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__() - -// use Operator with only cpu kernel. -#define USE_OP_CPU(op_type) \ - USE_OP_WITHOUT_KERNEL(op_type); \ - USE_OP_KERNEL(op_type, CPU) +#define USE_OP_ITSELF(op_type) \ + STATIC_ASSERT_GLOBAL_NAMESPACE( \ + __use_op_itself_##op_type, \ + "USE_OP_ITSELF must be called in global namespace"); \ + extern int TouchOpRegistrar_##op_type(); \ + static int use_op_itself_##op_type##_ __attribute__((unused)) = \ + TouchOpRegistrar_##op_type() + +#define USE_OP_DEVICE_KERNEL(op_type, DEVICE_TYPE) \ + STATIC_ASSERT_GLOBAL_NAMESPACE( \ + __use_op_kernel_##op_type##_##DEVICE_TYPE##__, \ + "USE_OP_DEVICE_KERNEL must be in global namespace"); \ + extern int TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE(); \ + static int use_op_kernel_##op_type##_##DEVICE_TYPE##_ \ + __attribute__((unused)) = \ + TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE() + +// TODO(fengjiayi): The following macros seems ugly, do we have better method? #ifdef PADDLE_ONLY_CPU -#define USE_OP(op_type) USE_OP_CPU(op_type) +#define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU) #else -#define USE_OP(op_type) \ - USE_OP_CPU(op_type); \ - USE_OP_KERNEL(op_type, GPU) +#define USE_OP_KERNEL(op_type) \ + USE_OP_DEVICE_KERNEL(op_type, CPU); \ + USE_OP_DEVICE_KERNEL(op_type, GPU) #endif +#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_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 9894928a7aa19bc6c7ad8b230562fb9a681cfebd..1a85d568350dc04ca1df28129de19cd45b5204b8 100644 --- a/paddle/framework/op_registry_test.cc +++ b/paddle/framework/op_registry_test.cc @@ -7,6 +7,7 @@ namespace paddle { namespace framework { class CosineOp : public OperatorBase { public: + using OperatorBase::OperatorBase; void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override {} void InferShape(const Scope& scope) const override {} @@ -27,6 +28,7 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { class MyTestOp : public OperatorBase { public: + using OperatorBase::OperatorBase; void InferShape(const Scope& scope) const override {} void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override {} @@ -36,8 +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!"); }; @@ -49,16 +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(); @@ -78,8 +88,8 @@ TEST(OpRegistry, CreateOp) { 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"); @@ -103,8 +113,8 @@ 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()); @@ -116,20 +126,11 @@ TEST(OpRegistry, DefaultValue) { 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); -} - 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; @@ -169,7 +170,6 @@ 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; diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc index cb86e6be2be3624bf54ee28193ca5d4c7bafa0eb..0daf12e7f5f3539d460ce67d39ca1c06f5aa2237 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 { @@ -22,95 +22,146 @@ namespace framework { template <> Eigen::DefaultDevice& ExecutionContext::GetEigenDevice< platform::CPUPlace, Eigen::DefaultDevice>() const { - return *device_context_.get_eigen_device(); + return *device_context_->get_eigen_device(); } #ifndef PADDLE_ONLY_CPU template <> Eigen::GpuDevice& ExecutionContext::GetEigenDevice() const { - return *device_context_.get_eigen_device(); + return *device_context_->get_eigen_device(); } #endif const std::string& OperatorBase::Input(const std::string& name) const { - PADDLE_ENFORCE(in_out_idxs_ != nullptr, - "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); - } + auto& ins = Inputs(name); + PADDLE_ENFORCE_EQ(ins.size(), 1UL, + "Op %s input %s should contain only one variable", type_, + name); + return ins[0]; } -std::vector OperatorBase::Inputs(const std::string& name) const { - PADDLE_ENFORCE(in_out_idxs_ != nullptr, "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(in_out_idxs_ != nullptr, "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); - } + auto& outs = Outputs(name); + PADDLE_ENFORCE_EQ(outs.size(), 1UL, + "Op %s output %s should contain only one variable", type_, + name); + return outs[0]; } -std::vector OperatorBase::Outputs(const std::string& name) const { - PADDLE_ENFORCE(in_out_idxs_ != nullptr, "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 OperatorBase::VarNameMap& inputs, + const OperatorBase::VarNameMap& outputs, + const AttributeMap& attrs) + : type_(type), inputs_(inputs), outputs_(outputs), attrs_(attrs) { + 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)); + } + } + } +} + +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 it = OpRegistry::op_info_map().find(type_); + PADDLE_ENFORCE( + it != OpRegistry::op_info_map().end(), + "Operator %s not registered, cannot figure out intermediate outputs", + type_); + PADDLE_ENFORCE( + it->second.proto_ != nullptr, + "Operator %s has no OpProto, cannot figure out intermediate outputs", + type_); + + // get all OpProto::Var for outputs + for (auto& o : it->second.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; } } // namespace framework diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h index 0b588297169540417586d7c167a1265827b683ac..60d4f06c7e6f8849800f238242a340ef5dbf771e 100644 --- a/paddle/framework/operator.h +++ b/paddle/framework/operator.h @@ -15,26 +15,45 @@ limitations under the License. */ #pragma once #include -#include #include #include #include -#include "paddle/framework/attr_checker.h" -#include "paddle/framework/op_desc.pb.h" -#include "paddle/framework/op_proto.pb.h" +#include "paddle/framework/attribute.h" +#include "paddle/framework/framework.pb.h" #include "paddle/framework/scope.h" #include "paddle/framework/tensor.h" #include "paddle/platform/device_context.h" #include "paddle/platform/place.h" +#include "paddle/platform/variant.h" #include "paddle/utils/Error.h" namespace paddle { namespace framework { +/// If a variable is a empty variable, that name will be used. +constexpr char kEmptyVarName[] = "@EMPTY@"; + +/// If a variable is a temporary variable, that name will be set in Python, +/// but it will be convert to a unique name in scope after OpCreator. +constexpr char kTempVarName[] = "@TEMP@"; + +/// If a variable's name has a certain suffix, it means that the +/// variable is the gradient of another varibale. +/// e.g. Variable "x@GRAD" is the gradient of varibale "x". +constexpr char kGradVarSuffix[] = "@GRAD"; + +/// Variables with this suffix are supposed to be filled up with zeros. +constexpr char kZeroVarSuffix[] = "@ZERO"; + +inline std::string GradVarName(const std::string& var_name) { + return var_name + kGradVarSuffix; +} + class OperatorBase; class InferShapeContext; class ExecutionContext; + /** * OperatorBase has the basic element that Net will call to do computation. * Only CreateOperator from OpRegistry will new Operator directly. User @@ -43,24 +62,14 @@ class ExecutionContext; */ class OperatorBase { public: - /// If a variable is a empty variable, that name will be used. - static std::string EMPTY_VAR_NAME() { return "@EMPTY@"; } - - /// If a variable is a temporary variable, that name will be set in Python, - /// but it will be convert to a unique name in scope after OpCreator. - static std::string TMP_VAR_NAME() { return "@TEMP@"; } + using VarNameMap = std::map>; - /// If a variable's name has a certain suffix, it means that the - /// variable is the gradient of another varibale. - /// e.g. Variable "x@GRAD" is the gradient of varibale "x". - static std::string GRAD_VAR_SUFFIX() { return "@GRAD"; } + OperatorBase(const std::string& type, const VarNameMap& inputs, + const VarNameMap& outputs, const AttributeMap& attrs); - static std::string GRAD_VAR_NAME(const std::string& name) { - return name + GRAD_VAR_SUFFIX(); - } - - /// Variables with this suffix are supposed to be filled up with zeros. - static std::string ZERO_VAR_SUFFIX() { return "@ZERO"; } + OperatorBase(const OperatorBase& o) = delete; + OperatorBase& operator=(const OperatorBase& o) = delete; + OperatorBase(OperatorBase&& o) = delete; virtual ~OperatorBase() {} @@ -73,10 +82,6 @@ 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; @@ -87,51 +92,63 @@ class OperatorBase { virtual bool IsNetOp() const { return false; } + virtual bool SupportGPU() const { return false; } + /// rename inputs outputs name void Rename(const std::string& old_name, const std::string& new_name); + const VarNameMap& Inputs() const { return inputs_; } + const VarNameMap& Outputs() const { return outputs_; } //! Get a input with argument's name described in `op_proto` 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; + //! Get a output with argument's name described in `op_proto` 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; - public: + 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_; } + + protected: std::string type_; // NOTE: in case of OpGrad, inputs_ contains: // I (Inputs) // O (Outputs) // OG (Output Gradients) - std::vector inputs_; + VarNameMap inputs_; + // NOTE: in case of OpGrad, outputs_ contains // IG (Inputs Gradients) - std::vector outputs_; + VarNameMap outputs_; AttributeMap attrs_; - // store the arguments' offset described in op_desc. - std::shared_ptr> in_out_idxs_; }; -class OperatorContext { +class NOP : public OperatorBase { public: - OperatorContext(const OperatorBase* op, const Scope& scope) - : op_(*op), scope_(scope) {} - - size_t InputSize() const { return op_.inputs_.size(); } + using OperatorBase::OperatorBase; + void InferShape(const Scope& scope) const override {} + void Run(const Scope& scope, + const platform::DeviceContext& dev_ctx) const override {} +}; - size_t OutputSize() const { return op_.outputs_.size(); } +class InferShapeContext { + public: + InferShapeContext(const OperatorBase& op, const Scope& scope) + : op_(op), scope_(scope) {} - const Variable* InputVar(const size_t index) const { - return scope_.FindVar(op_.inputs_.at(index)); + 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 { @@ -163,31 +180,17 @@ class OperatorContext { return res; } - template - const T* Input(const size_t index) const { - auto var = InputVar(index); - PADDLE_ENFORCE(var != nullptr, "Input(%d) should not be nullptr", index); - return &var->Get(); - } - - template - T* Output(const size_t index) const { - auto var = OutputVar(index); - PADDLE_ENFORCE(var != nullptr, "Output(%d) should not be nullptr", index); - return var->GetMutable(); - } - template const T* Input(const std::string& name) const { - auto var = InputVar(name); - PADDLE_ENFORCE(var != nullptr, "Input(%s) should not be nullptr", name); + auto* var = InputVar(name); + PADDLE_ENFORCE_NOT_NULL(var, "Input(%s) should not be nullptr", name); return &var->Get(); } template T* Output(const std::string& name) const { auto var = OutputVar(name); - PADDLE_ENFORCE(var != nullptr, "Output(%s) should not be nullptr", name); + PADDLE_ENFORCE_NOT_NULL(var, "Output(%s) should not be nullptr", name); return var->GetMutable(); } @@ -199,9 +202,9 @@ class OperatorContext { std::transform(names.begin(), names.end(), std::back_inserter(res), [&](const std::string& sub_name) { auto var = scope_.FindVar(sub_name); - PADDLE_ENFORCE(var != nullptr, - "MultiInput(%s:%s) should not be nullptr", - name, sub_name); + PADDLE_ENFORCE_NOT_NULL( + var, "MultiInput(%s:%s) should not be nullptr", name, + sub_name); return &var->Get(); }); return res; @@ -215,9 +218,9 @@ class OperatorContext { std::transform(names.begin(), names.end(), std::back_inserter(res), [&](const std::string& sub_name) { auto var = scope_.FindVar(sub_name); - PADDLE_ENFORCE(var != nullptr, - "MultiOutput(%s:%s) should not be nullptr", - name, sub_name); + PADDLE_ENFORCE_NOT_NULL( + var, "MultiOutput(%s:%s) should not be nullptr.", name, + sub_name); return var->GetMutable(); }); return res; @@ -227,12 +230,6 @@ class OperatorContext { const Scope& scope_; }; -class InferShapeContext : public OperatorContext { - public: - InferShapeContext(const OperatorBase* op, const Scope& scope) - : OperatorContext(op, scope) {} -}; - template struct EigenDeviceConverter; @@ -248,20 +245,24 @@ struct EigenDeviceConverter { }; #endif -class ExecutionContext : public OperatorContext { +class ExecutionContext : public InferShapeContext { public: - ExecutionContext(const OperatorBase* op, const Scope& scope, - const platform::DeviceContext& device_context) - : OperatorContext(op, scope), device_context_(device_context) {} + ExecutionContext(const OperatorBase& op, const Scope& scope, + const platform::DeviceContext* device_context) + : InferShapeContext(op, scope), device_context_(device_context) {} template ::EigenDeviceType> DeviceType& GetEigenDevice() const; - platform::Place GetPlace() const { return device_context_.GetPlace(); } + platform::Place GetPlace() const { return device_context_->GetPlace(); } - const platform::DeviceContext& device_context_; + const platform::DeviceContext* device_context() const { + return device_context_; + } + + const platform::DeviceContext* device_context_; }; class OpKernel { @@ -284,7 +285,7 @@ class OperatorWithKernel : public OperatorBase { platform::Place place_; OpKernelKey() = default; - OpKernelKey(const platform::DeviceContext& dev_ctx) { + explicit OpKernelKey(const platform::DeviceContext& dev_ctx) { place_ = dev_ctx.GetPlace(); } @@ -303,14 +304,18 @@ class OperatorWithKernel : public OperatorBase { using OpKernelMap = std::unordered_map, OpKernelHash>; - void InferShape(const Scope& scope) const { - InferShape(InferShapeContext(this, scope)); + OperatorWithKernel(const std::string& type, const VarNameMap& inputs, + const VarNameMap& outputs, const AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + void InferShape(const Scope& scope) const override { + InferShape(InferShapeContext(*this, scope)); } void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const final { auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx)); - opKernel->Compute(ExecutionContext(this, scope, dev_ctx)); + opKernel->Compute(ExecutionContext(*this, scope, &dev_ctx)); } static std::unordered_map& @@ -319,6 +324,12 @@ class OperatorWithKernel : public OperatorBase { return g_all_op_kernels; } + bool SupportGPU() const override { + OperatorWithKernel::OpKernelKey key; + key.place_ = platform::GPUPlace(); + return OperatorWithKernel::AllOpKernels().at(type_).count(key) != 0; + } + protected: virtual void InferShape(const InferShapeContext& ctx) const = 0; }; diff --git a/paddle/framework/operator_test.cc b/paddle/framework/operator_test.cc index 6a6a802b7da05c37a317540030836baa28a89cd7..0441cec9f6d10246fba38b02b4de3cbe2ee4766b 100644 --- a/paddle/framework/operator_test.cc +++ b/paddle/framework/operator_test.cc @@ -23,20 +23,22 @@ static int op_run_num = 0; class OpWithoutKernelTest : public OperatorBase { public: - void Init() override { x = 1; } + OpWithoutKernelTest(const std::string& type, const VarNameMap& inputs, + const VarNameMap& 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 { @@ -54,14 +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); @@ -97,6 +110,9 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { static int cpu_kernel_run_num = 0; class OpWithKernelTest : public OperatorWithKernel { + public: + using OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext& ctx) const override {} }; @@ -113,33 +129,15 @@ class CPUKernelTest : public OpKernel { } }; -// multiple inputs test -class OperatorMultiInputsTest : public OperatorBase { - public: - 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"); - } - - public: - float x = 0; -}; - class OpKernelTestMultiInputsProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: 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); @@ -157,22 +155,22 @@ class CPUKernalMultiInputsTest : public OpKernel { ASSERT_EQ(xs[2], "x2"); auto inVar0 = ctx.MultiInputVar("xs"); - ASSERT_EQ(inVar0.size(), 3); + ASSERT_EQ(inVar0.size(), 3U); auto intVar1 = ctx.InputVar("k"); ASSERT_NE(intVar1, nullptr); auto outVar0 = ctx.MultiOutputVar("ys"); - ASSERT_EQ(outVar0.size(), 2); + ASSERT_EQ(outVar0.size(), 2U); auto inTensor0 = ctx.MultiInput("xs"); - ASSERT_EQ(inTensor0.size(), 3); + ASSERT_EQ(inTensor0.size(), 3U); auto intTensor1 = ctx.Input("k"); ASSERT_NE(intTensor1, nullptr); auto outTensor0 = ctx.MultiOutput("ys"); - ASSERT_EQ(outTensor0.size(), 2); + ASSERT_EQ(outTensor0.size(), 2U); auto k = ctx.op_.Input("k"); ASSERT_EQ(k, "k0"); @@ -187,8 +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); @@ -196,8 +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); @@ -212,8 +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); @@ -223,32 +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(); diff --git a/paddle/framework/pybind.cc b/paddle/framework/pybind.cc index b4f0f3ef7e3a4230c09ea6f766c4017946ac0b5a..fe0c87bc570825014222807cb90a3bb341b44e8e 100644 --- a/paddle/framework/pybind.cc +++ b/paddle/framework/pybind.cc @@ -18,13 +18,12 @@ limitations under the License. */ #include "paddle/framework/backward.h" #include "paddle/framework/op_registry.h" -#include "paddle/framework/operator.h" -#include "paddle/framework/scope.h" #include "paddle/framework/tensor_py.h" #include "paddle/operators/net_op.h" -#include "paddle/operators/type_alias.h" +#include "paddle/operators/recurrent_op.h" #include "paddle/platform/enforce.h" #include "paddle/platform/place.h" +#include "paddle/string/to_string.h" #include "pybind11/numpy.h" #include "pybind11/pybind11.h" #include "pybind11/stl.h" @@ -32,30 +31,44 @@ limitations under the License. */ namespace py = pybind11; USE_OP(add_two); -USE_OP(onehot_cross_entropy); -USE_OP_WITHOUT_KERNEL(fc); +USE_CPU_ONLY_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_WITHOUT_KERNEL(recurrent_op); +USE_OP(fill_zeros_like); +USE_OP_ITSELF(recurrent_op); +USE_OP(gaussian_random); +USE_OP(uniform_random); + 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_; + return op.Type(); }) .def("outputs", - [](const typename ClassType::type &op) -> std::vector { - return op.outputs_; + [](const typename ClassType::type &op) + -> std::map> { + return op.Outputs(); + }) + .def("inputs", + [](const typename ClassType::type &op) { return op.Inputs(); }) + .def("__str__", &ClassType::type::DebugString) + .def("no_intermediate_outputs", + [](const typename ClassType::type &op) { + return op.OutputVars(false); }) - .def("__str__", &ClassType::type::DebugString); + .def("support_gpu", &ClassType::type::SupportGPU); } static size_t UniqueIntegerGenerator() { @@ -105,7 +118,16 @@ PYBIND11_PLUGIN(core) { .def("set", PyCUDATensorSetFromArray) .def("set", PyCUDATensorSetFromArray) #endif - .def("shape", [](Tensor &self) { return vectorize(self.dims()); }); + .def("shape", [](Tensor &self) { return vectorize(self.dims()); }) + .def("set_float_element", + [](Tensor &self, size_t offset, float f) { + // TODO(yuyang18): Only support GPU now. + self.data()[offset] = f; + }) + .def("get_float_element", [](Tensor &self, size_t offset) -> float { + // TODO(yuyang18): Only support GPU now. + return self.data()[offset]; + }); py::class_(m, "Variable", R"DOC(Variable Class. @@ -119,8 +141,8 @@ All parameter, weight, gradient are variables in Paddle. [](Variable &self) -> Tensor * { return self.GetMutable(); }, py::return_value_policy::reference) .def("get_net", - [](Variable &self) -> ops::NetOp * { - return self.GetMutable(); + [](Variable &self) -> operators::NetOp * { + return self.GetMutable(); }, py::return_value_policy::reference); @@ -139,13 +161,16 @@ All parameter, weight, gradient are variables in Paddle. //! @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(); + auto &op_info_map = OpRegistry::op_info_map(); std::vector ret_values; - for (auto it = protos.begin(); it != protos.end(); ++it) { - PADDLE_ENFORCE(it->second.IsInitialized(), - "OpProto must all be initialized"); + for (auto it = op_info_map.begin(); it != op_info_map.end(); ++it) { + const OpProto *proto = it->second.proto_; + if (proto == nullptr) { + continue; + } + PADDLE_ENFORCE(proto->IsInitialized(), "OpProto must all be initialized"); std::string str; - PADDLE_ENFORCE(it->second.SerializeToString(&str), + PADDLE_ENFORCE(proto->SerializeToString(&str), "Serialize OpProto Error. This could be a bug of Paddle."); ret_values.push_back(py::bytes(str)); } @@ -154,8 +179,8 @@ All parameter, weight, gradient are variables in Paddle. m.def_submodule( "var_names", "The module will return special predefined variable name in Paddle") - .def("empty", OperatorBase::EMPTY_VAR_NAME) - .def("temp", OperatorBase::TMP_VAR_NAME); + .def("empty", []() { return kEmptyVarName; }) + .def("temp", []() { return kTempVarName; }); // clang-format off py::class_(m, "DeviceContext") .def_static("create", @@ -174,9 +199,13 @@ All parameter, weight, gradient are variables in Paddle. }); // clang-format on - py::class_(m, "GPUPlace").def(py::init()); + py::class_(m, "GPUPlace") + .def(py::init()) + .def("__str__", string::to_string); - py::class_(m, "CPUPlace").def(py::init<>()); + py::class_(m, "CPUPlace") + .def(py::init<>()) + .def("__str__", string::to_string); py::class_> operator_base( m, "Operator"); @@ -199,26 +228,55 @@ All parameter, weight, gradient are variables in Paddle. ExposeOperator(operator_base); - py::class_> net(m, "Net"); + py::class_> net(m, "Net"); net.def_static("create", - []() -> std::shared_ptr { - auto retv = std::make_shared(); - retv->type_ = "plain_net"; + []() -> std::shared_ptr { + auto retv = std::make_shared(); + retv->SetType("plain_net"); return retv; }) - .def("add_op", &ops::NetOp::AddOp) - .def( - "add_op", - [](ops::NetOp &self, const std::shared_ptr &net) -> void { - self.AddOp(std::static_pointer_cast(net)); - }) - .def("complete_add_op", &ops::NetOp::CompleteAddOp) - .def("complete_add_op", - [](std::shared_ptr &self) { self->CompleteAddOp(); }); + .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)); + }) + .def("add_op", + [](operators::NetOp &self, + const std::shared_ptr &rnn) -> void { + self.AddOp(std::static_pointer_cast(rnn)); + }) + .def("complete_add_op", &operators::NetOp::CompleteAddOp) + .def("complete_add_op", [](std::shared_ptr &self) { + self->CompleteAddOp(); + }); ExposeOperator(net); + // recurrent_op + py::class_> + rnn(m, "RecurrentOp"); + + rnn.def_static( + "create", + [](py::bytes protobin) -> std::shared_ptr { + 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 std::dynamic_pointer_cast(rnn_op); + }) + .def("set_stepnet", + [](operators::RecurrentOp &self, + const std::shared_ptr &net) -> void { + self.set_stepnet(net); + }); + ExposeOperator(rnn); + m.def("unique_integer", UniqueIntegerGenerator); m.def("is_compile_gpu", IsCompileGPU); diff --git a/paddle/framework/tensor.h b/paddle/framework/tensor.h index 4c3b14b83d841e88683a13634c93f51c012128b6..b8c779f4e5fc7bc51298cdd35b26c2c8ac98edf6 100644 --- a/paddle/framework/tensor.h +++ b/paddle/framework/tensor.h @@ -18,6 +18,8 @@ limitations under the License. */ #include #include #include +#include + #include "paddle/framework/ddim.h" #include "paddle/memory/memory.h" #include "paddle/platform/device_context.h" @@ -77,11 +79,11 @@ class Tensor { inline const DDim& dims() const; /*! Resize the dimensions of the memory block. */ - inline void Resize(const DDim& dims); + inline Tensor& Resize(const DDim& dims); /*! The internal of two tensors share the same memory block. */ template - inline void ShareDataWith(const Tensor& src); + inline Tensor& ShareDataWith(const Tensor& src); /** * @brief Copy the content of external tensor to a new place. @@ -103,6 +105,8 @@ class Tensor { template inline Tensor Slice(const int& begin_idx, const int& end_idx) const; + platform::Place place() const { return holder_->place(); } + private: template inline void check_memory_size() const; @@ -127,8 +131,8 @@ class Tensor { memory::PODDeleter(place)), place_(place), size_(size) { - PADDLE_ENFORCE(ptr_ != nullptr, "Insufficient %s memory to allocation.", - is_cpu_place(place_) ? "CPU" : "GPU"); + PADDLE_ENFORCE_NOT_NULL(ptr_, "Insufficient %s memory to allocation.", + (is_cpu_place(place_) ? "CPU" : "GPU")); } virtual size_t size() const { return size_; } diff --git a/paddle/framework/tensor_impl.h b/paddle/framework/tensor_impl.h index 92621f8c18ec0d03160a23c462830d14272c7f64..7d7263b899afb7a2128548f264065a8013b6f0c9 100644 --- a/paddle/framework/tensor_impl.h +++ b/paddle/framework/tensor_impl.h @@ -14,17 +14,20 @@ limitations under the License. */ #pragma once #include "paddle/memory/memcpy.h" +#include "paddle/platform/enforce.h" namespace paddle { namespace framework { template inline void Tensor::check_memory_size() const { - PADDLE_ENFORCE(holder_ != nullptr, - "Tenosr holds no memory. Call Tensor::mutable_data first."); - PADDLE_ENFORCE(holder_->size() >= product(dims_) * sizeof(T) + offset_, - "Tensor's dims_ is out of bound. Call Tensor::mutable_data " - "first to re-allocate memory."); + PADDLE_ENFORCE_NOT_NULL( + holder_, "Tenosr holds no memory. Call Tensor::mutable_data first."); + PADDLE_ENFORCE_GE( + holder_->size(), product(dims_) * sizeof(T) + offset_, + "Tensor's dims_ is out of bound. Call Tensor::mutable_data " + "first to re-allocate memory.\n" + "or maybe the required data-type mismatches the data already stored."); } template @@ -51,9 +54,9 @@ inline T* Tensor::mutable_data(DDim dims, platform::Place place) { template inline T* Tensor::mutable_data(platform::Place place) { static_assert(std::is_pod::value, "T must be POD"); - PADDLE_ENFORCE(product(dims_) > 0, - "Tensor's numel must be larger than zero to call " - "Tensor::mutable_data. Call Tensor::set_dim first."); + PADDLE_ENFORCE_GT(product(dims_), 0, + "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); if (holder_ == nullptr || !(holder_->place() == place) || @@ -77,9 +80,10 @@ inline T* Tensor::mutable_data(platform::Place place) { } template -inline void Tensor::ShareDataWith(const Tensor& src) { +inline Tensor& Tensor::ShareDataWith(const Tensor& src) { src.check_memory_size(); *this = src; + return *this; } template @@ -120,11 +124,11 @@ inline void Tensor::CopyFrom(const Tensor& src, template inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const { check_memory_size(); - PADDLE_ENFORCE(begin_idx >= 0, "Slice begin index is less than zero."); - PADDLE_ENFORCE(end_idx <= dims_[0], "Slice end index is out of bound."); - PADDLE_ENFORCE(begin_idx < end_idx, - "Begin index must be less than end index."); - PADDLE_ENFORCE(dims_[0] != 1, "Can not slice a tensor with dims_[0] = 1."); + PADDLE_ENFORCE_GE(begin_idx, 0, "Slice begin index is less than zero."); + PADDLE_ENFORCE_LE(end_idx, dims_[0], "Slice end index is out of bound."); + PADDLE_ENFORCE_LT(begin_idx, end_idx, + "Begin index must be less than end index."); + PADDLE_ENFORCE_NE(dims_[0], 1, "Can not slice a tensor with dims_[0] = 1."); int base = product(dims_) / dims_[0]; Tensor dst; dst.holder_ = holder_; @@ -135,7 +139,10 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const { return dst; } -inline void Tensor::Resize(const DDim& dims) { dims_ = dims; } +inline Tensor& Tensor::Resize(const DDim& dims) { + dims_ = dims; + return *this; +} inline const DDim& Tensor::dims() const { return dims_; } diff --git a/paddle/framework/tensor_test.cc b/paddle/framework/tensor_test.cc index ef1cc10b840896d9ab97f963fc12a4971cd74e1f..7db38d5caeebccf710334e854faf785ef0f64063 100644 --- a/paddle/framework/tensor_test.cc +++ b/paddle/framework/tensor_test.cc @@ -19,7 +19,7 @@ TEST(Tensor, Dims) { using namespace paddle::framework; using namespace paddle::platform; Tensor tt; - tt.Resize(make_ddim({2, 3, 4})); + tt.Resize({2, 3, 4}); DDim dims = tt.dims(); ASSERT_EQ(arity(dims), 3); for (int i = 0; i < 3; ++i) { @@ -36,7 +36,8 @@ TEST(Tensor, DataAssert) { } catch (paddle::platform::EnforceNotMet err) { caught = true; std::string msg = - "Tenosr holds no memory. Call Tensor::mutable_data first."; + "holder_ should not be null\nTenosr holds no memory. Call " + "Tensor::mutable_data first."; const char* what = err.what(); for (size_t i = 0; i < msg.length(); ++i) { ASSERT_EQ(what[i], msg[i]); @@ -111,7 +112,8 @@ TEST(Tensor, ShareDataWith) { } catch (paddle::platform::EnforceNotMet err) { caught = true; std::string msg = - "Tenosr holds no memory. Call Tensor::mutable_data first."; + "holder_ should not be null\nTenosr holds no memory. Call " + "Tensor::mutable_data first."; const char* what = err.what(); for (size_t i = 0; i < msg.length(); ++i) { ASSERT_EQ(what[i], msg[i]); diff --git a/paddle/function/BlockExpandOpTest.cpp b/paddle/function/BlockExpandOpTest.cpp index 5e4897e72ba9fab2dd9e25d90313dc1b4d38e2d4..59193a3ec3d0fabe7c841372394204ab568f5a2b 100644 --- a/paddle/function/BlockExpandOpTest.cpp +++ b/paddle/function/BlockExpandOpTest.cpp @@ -18,10 +18,10 @@ limitations under the License. */ namespace paddle { TEST(BlockExpandForward, real) { - for (size_t batchSize : {5, 32}) { - for (size_t channels : {1, 5, 32}) { - for (size_t inputHeight : {5, 33, 100}) { - for (size_t inputWidth : {5, 32, 96}) { + for (size_t batchSize : {5}) { + for (size_t channels : {1, 5}) { + for (size_t inputHeight : {5, 33}) { + for (size_t inputWidth : {5, 32}) { for (size_t block : {1, 3, 5}) { for (size_t stride : {1, 2}) { for (size_t padding : {0, 1}) { @@ -61,10 +61,10 @@ TEST(BlockExpandForward, real) { } TEST(BlockExpandBackward, real) { - for (size_t batchSize : {5, 32}) { - for (size_t channels : {1, 5, 32}) { - for (size_t inputHeight : {5, 33, 100}) { - for (size_t inputWidth : {5, 32, 96}) { + for (size_t batchSize : {5}) { + for (size_t channels : {1, 5}) { + for (size_t inputHeight : {5, 33}) { + for (size_t inputWidth : {5, 32}) { for (size_t block : {1, 3, 5}) { for (size_t stride : {1, 2}) { for (size_t padding : {0, 1}) { diff --git a/paddle/function/BufferArgTest.cpp b/paddle/function/BufferArgTest.cpp index 1744f377808f137dcda4a28acce336dc22be3d01..6b8e1e2da9775ccd03c84cc86ad226f3c00ab7fe 100644 --- a/paddle/function/BufferArgTest.cpp +++ b/paddle/function/BufferArgTest.cpp @@ -32,7 +32,7 @@ TEST(BufferTest, SequenceIdArg) { sizeOfValuType(VALUE_TYPE_INT32)); SequenceIdArg buffer(memory.getBuf(), shape); EXPECT_EQ(buffer.data(), memory.getBuf()); - EXPECT_EQ(buffer.numSeqs(), 9); + EXPECT_EQ(buffer.numSeqs(), 9U); } } // namespace paddle diff --git a/paddle/function/CMakeLists.txt b/paddle/function/CMakeLists.txt index 93304f73037690b5cf3ac8189aabc28f51316a77..7dfb6f61c50959f7269725a00dbc4f9c27474bdf 100644 --- a/paddle/function/CMakeLists.txt +++ b/paddle/function/CMakeLists.txt @@ -38,10 +38,11 @@ if(WITH_GPU) add_simple_unittest(RowConvOpTest) add_simple_unittest(BlockExpandOpTest) add_simple_unittest(CropOpTest) + add_simple_unittest(DepthwiseConvOpTest) endif() -add_simple_unittest(ConvOpTest) add_simple_unittest(Im2ColTest) +add_simple_unittest(GemmConvOpTest) endif() add_style_check_target(paddle_function ${h_files}) diff --git a/paddle/function/ContextProjectionOpGpu.cu b/paddle/function/ContextProjectionOpGpu.cu index 1a5b4042402df3081a493962a5e080d72b7f40b2..4492dea5d8a6f8580a13f3059401c87fa2164085 100644 --- a/paddle/function/ContextProjectionOpGpu.cu +++ b/paddle/function/ContextProjectionOpGpu.cu @@ -12,8 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "hl_base.h" #include "ContextProjectionOp.h" +#include "hl_base.h" namespace paddle { @@ -30,7 +30,7 @@ __global__ void KeContextProjectionForward(const real* input, int block_size = blockDim.x; int sequenceId = blockIdx.x; int seq_start = sequence[sequenceId]; - int seq_end = sequence[sequenceId+1]; + int seq_end = sequence[sequenceId + 1]; real value = 0; int instances = seq_end - seq_start + context_length - 1; @@ -49,8 +49,9 @@ __global__ void KeContextProjectionForward(const real* input, } else if ((i + context_start) >= (seq_end - seq_start)) { if (padding) { value = - weight[(begin_pad + i + context_start - (seq_end - seq_start)) * - input_dim + idx]; + weight[(begin_pad + i + context_start - (seq_end - seq_start)) * + input_dim + + idx]; } else { continue; } @@ -61,7 +62,7 @@ __global__ void KeContextProjectionForward(const real* input, int outx = (i - context_length) < 0 ? i : (context_length - 1); int outy = (i - context_length) < 0 ? 0 : (i - (context_length - 1)); real* output_r = - output + outy * input_dim * context_length + outx * input_dim; + output + outy * input_dim * context_length + outx * input_dim; for (int j = outy; j < seq_end - seq_start; j++) { output_r[idx] += value; if (j - outy == outx) break; @@ -108,13 +109,25 @@ void hl_context_projection_forward(const real* input, dim3 grid(blocks_x, blocks_y); if (weight) { - KeContextProjectionForward<<< grid, threads, 0, STREAM_DEFAULT >>> - (input, sequence, weight, output, input_dim, - context_length, context_start, begin_pad); - } else { - KeContextProjectionForward<<< grid, threads, 0, STREAM_DEFAULT >>> - (input, sequence, weight, output, input_dim, - context_length, context_start, begin_pad); + KeContextProjectionForward<<>>( + input, + sequence, + weight, + output, + input_dim, + context_length, + context_start, + begin_pad); + } else { + KeContextProjectionForward<<>>( + input, + sequence, + weight, + output, + input_dim, + context_length, + context_start, + begin_pad); } CHECK_SYNC("hl_context_projection_forward failed"); } @@ -148,7 +161,7 @@ __global__ void KeContextProjectionBackwardData(const real* out_grad, int block_size = blockDim.x; int sequenceId = blockIdx.x; int seq_start = sequence[sequenceId]; - int seq_end = sequence[sequenceId+1]; + int seq_end = sequence[sequenceId + 1]; real value = 0; int instances = seq_end - seq_start + context_length - 1; @@ -170,7 +183,7 @@ __global__ void KeContextProjectionBackwardData(const real* out_grad, int outx = (i - context_length) < 0 ? i : (context_length - 1); int outy = (i - context_length) < 0 ? 0 : (i - (context_length - 1)); real* output_r = - out + outy * input_dim * context_length + outx * input_dim; + out + outy * input_dim * context_length + outx * input_dim; for (int j = outy; j < seq_end - seq_start; j++) { value += output_r[idx]; if (j - outy == outx) break; @@ -211,8 +224,8 @@ void hl_context_projection_backward_data(const real* out_grad, int blocks_y = 1; dim3 threads(block_size, 1); dim3 grid(blocks_x, blocks_y); - KeContextProjectionBackwardData<<< grid, threads, 0, STREAM_DEFAULT >>> - (out_grad, sequence, input_grad, input_dim, context_length, context_start); + KeContextProjectionBackwardData<<>>( + out_grad, sequence, input_grad, input_dim, context_length, context_start); CHECK_SYNC("hl_context_projection_backward_data failed"); } @@ -231,7 +244,7 @@ void ContextProjectionBackwardData(const GpuMatrix& out_grad, context_start); } -template +template __global__ void KeContextProjectionBackwardWeight(const real* out_grad, const int* sequence, real* w_grad, @@ -254,17 +267,17 @@ __global__ void KeContextProjectionBackwardWeight(const real* out_grad, if (weight_idx < w_dim) { for (int seqId = idy; seqId < num_sequences; seqId += THREADS_Y) { int seq_start = sequence[seqId]; - int seq_end = sequence[seqId+1]; - output_r = const_cast(out_grad) - + seq_start * w_dim * context_length; + int seq_end = sequence[seqId + 1]; + output_r = + const_cast(out_grad) + seq_start * w_dim * context_length; if (context_start < 0) { if (padId + context_start < 0) { instanceId = padId; } else { // begin_pad > 0; - instanceId = (padId - begin_pad) + - (seq_end - seq_start) - context_start; + instanceId = + (padId - begin_pad) + (seq_end - seq_start) - context_start; } } else { if (padId + (seq_end - seq_start) < context_start) { @@ -275,10 +288,11 @@ __global__ void KeContextProjectionBackwardWeight(const real* out_grad, } } - int outx = (instanceId - context_length) < 0 ? - instanceId : (context_length - 1); - int outy = (instanceId - context_length) < 0 ? - 0 : (instanceId - (context_length - 1)); + int outx = + (instanceId - context_length) < 0 ? instanceId : (context_length - 1); + int outy = (instanceId - context_length) < 0 + ? 0 + : (instanceId - (context_length - 1)); output_r += outy * w_dim * context_length + outx * w_dim; for (int j = outy; j < seq_end - seq_start; j++) { value += output_r[weight_idx]; @@ -290,7 +304,7 @@ __global__ void KeContextProjectionBackwardWeight(const real* out_grad, } __syncthreads(); - for (int stride = THREADS_Y/2; stride > 0; stride = stride/2) { + for (int stride = THREADS_Y / 2; stride > 0; stride = stride / 2) { if (idy < stride) { sum_s[idy][idx] += sum_s[idy + stride][idx]; } @@ -339,22 +353,27 @@ void hl_context_projection_backward_weight(const real* out_grad, dim3 threads(threads_x, threads_y); dim3 grid(blocks_x, 1); - KeContextProjectionBackwardWeight<32, 32> - <<< grid, threads, 0, STREAM_DEFAULT >>> - (out_grad, sequence, w_grad, num_sequences, w_dim, - context_length, context_start, begin_pad); + KeContextProjectionBackwardWeight<32, + 32><<>>( + out_grad, + sequence, + w_grad, + num_sequences, + w_dim, + context_length, + context_start, + begin_pad); CHECK_SYNC("hl_context_projection_backward_weight failed"); } template <> -void ContextProjectionBackwardWeight( - const GpuMatrix& out_grad, - GpuMatrix& w_grad, - const GpuIVector& seq_vec, - size_t context_length, - int context_start, - size_t total_pad, - size_t begin_pad) { +void ContextProjectionBackwardWeight(const GpuMatrix& out_grad, + GpuMatrix& w_grad, + const GpuIVector& seq_vec, + size_t context_length, + int context_start, + size_t total_pad, + size_t begin_pad) { hl_context_projection_backward_weight(out_grad.getData(), seq_vec.getData(), w_grad.getData(), @@ -376,23 +395,18 @@ void ContextProjectionBackward(const GpuMatrix& out_grad, size_t begin_pad, bool is_padding, size_t total_pad) { - if (in_grad) { - ContextProjectionBackwardData( - out_grad, - in_grad, - sequence, - context_length, - context_start); - } - if (is_padding && w_grad) { - ContextProjectionBackwardWeight( - out_grad, - w_grad, - sequence, - context_length, - context_start, - total_pad, - begin_pad); + if (in_grad) { + ContextProjectionBackwardData( + out_grad, in_grad, sequence, context_length, context_start); + } + if (is_padding && w_grad) { + ContextProjectionBackwardWeight(out_grad, + w_grad, + sequence, + context_length, + context_start, + total_pad, + begin_pad); } } diff --git a/paddle/function/ConvOpTest.cpp b/paddle/function/ConvOpTest.cpp deleted file mode 100644 index 7f32c734791853a8cd0287a80a7955dbd1bd7571..0000000000000000000000000000000000000000 --- a/paddle/function/ConvOpTest.cpp +++ /dev/null @@ -1,306 +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 -#include "Function.h" -#include "FunctionTest.h" - -namespace paddle { - -enum TestType { - kForwardTest = 0, - kBackwardInputTest = 1, - kBackwardFilterTest = 2, -}; - -template -class ConvolutionTest { -public: - ConvolutionTest(const std::string& conv1, - const std::string& conv2, - TestType type, - bool useGroups = true, - std::string algo = "auto") { - for (size_t batchSize : {1, 32}) { - for (size_t inputSize : {7, 14, 54}) { - for (size_t filterSize : {1, 3, 5}) { - for (size_t inputChannels : {3, 64}) { - for (size_t outputChannels : {3, 64}) { - if (inputChannels > outputChannels) break; - size_t groups; - if (!useGroups) { - groups = 1; - } else { - if (outputChannels % inputChannels != 0) continue; - groups = inputChannels; - } - - for (size_t stride : {1, 2}) { - for (size_t padding : {0, 1}) { - if (padding >= filterSize) break; - size_t outputSize = - (inputSize - filterSize + 2 * padding + stride) / stride; - VLOG(3) << " batchSize=" << batchSize - << " inputChannels=" << inputChannels - << " inputHeight=" << inputSize - << " inputWidth=" << inputSize - << " outputChannels=" << outputChannels - << " filterHeight=" << filterSize - << " filterWidth=" << filterSize - << " outputHeight=" << outputSize - << " outputWidth=" << outputSize - << " stride=" << stride << " padding=" << padding; - - std::vector paddings = {padding, padding}; - std::vector strides = {stride, stride}; - Compare2Function test( - conv1, - conv2, - FuncConfig() - .set("paddings", paddings) - .set("strides", strides) - .set("groups", groups) - .set("algo", algo)); - - TensorShape input{ - batchSize, inputChannels, inputSize, inputSize}; - - TensorShape filter; - if (groups > 1) - filter = TensorShape({groups, - outputChannels / groups, - inputChannels / groups, - filterSize, - filterSize}); - else - filter = TensorShape({outputChannels, - inputChannels, - filterSize, - filterSize}); - TensorShape output{ - batchSize, outputChannels, outputSize, outputSize}; - - if (type == kForwardTest) { - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input)); - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter)); - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, output)); - test.run(); - } else if (type == kBackwardInputTest) { - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output)); - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter)); - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, input), ADD_TO); - test.run(); - } else if (type == kBackwardFilterTest) { - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output)); - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input)); - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter), - ADD_TO); - test.run(); - } - } - } - } - } - } - } - } - } -}; - -// Mainly used to test cases where the height and width (input, filter) -// are not equal. -template -class ConvolutionTest2 { -public: - ConvolutionTest2(const std::string& conv1, - const std::string& conv2, - TestType type, - bool useGroups = true, - std::string algo = "auto") { - for (size_t batchSize : {16}) { - for (size_t inputHeight : {7, 31}) { - for (size_t inputWidth : {10, 54}) { - for (size_t filterHeight : {1, 5}) { - for (size_t filterWidth : {3, 7}) { - for (size_t inputChannels : {7}) { - for (size_t outputChannels : {7}) { - size_t groups; - if (!useGroups) { - groups = 1; - } else { - if (outputChannels % inputChannels != 0) continue; - groups = inputChannels; - } - - size_t stride = 1; - size_t padding = 0; - size_t outputHeight = - (inputHeight - filterHeight + 2 * padding + stride) / - stride; - size_t outputWidth = - (inputWidth - filterWidth + 2 * padding + stride) / - stride; - VLOG(3) << " batchSize=" << batchSize - << " inputChannels=" << inputChannels - << " inputHeight=" << inputHeight - << " inputWidth=" << inputWidth - << " outputChannels=" << outputChannels - << " filterHeight=" << filterHeight - << " filterWidth=" << filterWidth - << " outputHeight=" << outputHeight - << " outputWidth=" << outputWidth - << " stride=" << stride << " padding=" << padding; - - std::vector paddings = {padding, padding}; - std::vector strides = {stride, stride}; - Compare2Function test( - conv1, - conv2, - FuncConfig() - .set("paddings", paddings) - .set("strides", strides) - .set("groups", groups) - .set("algo", algo)); - - TensorShape input{ - batchSize, inputChannels, inputHeight, inputWidth}; - - TensorShape filter; - if (groups > 1) - filter = TensorShape({groups, - outputChannels / groups, - inputChannels / groups, - filterHeight, - filterWidth}); - else - filter = TensorShape({outputChannels, - inputChannels, - filterHeight, - filterWidth}); - TensorShape output{ - batchSize, outputChannels, outputHeight, outputWidth}; - - if (type == kForwardTest) { - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input)); - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter)); - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, output)); - test.run(); - } else if (type == kBackwardInputTest) { - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output)); - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter)); - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, input), ADD_TO); - test.run(); - } else if (type == kBackwardFilterTest) { - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output)); - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input)); - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter), - ADD_TO); - test.run(); - } - } - } - } - } - } - } - } - } -}; - -// ======Start Convolution TEST====== - -TEST(Forward, GEMM) { - ConvolutionTest test( - "NaiveConv-CPU", "GemmConv-CPU", kForwardTest, false); - ConvolutionTest2 test2( - "NaiveConv-CPU", "GemmConv-CPU", kForwardTest, false); -} - -#ifndef PADDLE_ONLY_CPU -TEST(Forward, GEMM2) { - ConvolutionTest test( - "GemmConv-CPU", "GemmConv-GPU", kForwardTest, false); - ConvolutionTest2 test2( - "GemmConv-CPU", "GemmConv-GPU", kForwardTest, false); -} - -TEST(BackwardInput, GEMM) { - ConvolutionTest test( - "GemmConvGradInput-CPU", - "GemmConvGradInput-GPU", - kBackwardInputTest, - false); - ConvolutionTest2 test2( - "GemmConvGradInput-CPU", - "GemmConvGradInput-GPU", - kBackwardInputTest, - false); -} - -TEST(BackwardFilter, GEMM) { - ConvolutionTest test( - "GemmConvGradFilter-CPU", - "GemmConvGradFilter-GPU", - kBackwardFilterTest, - false); - ConvolutionTest2 test2( - "GemmConvGradFilter-CPU", - "GemmConvGradFilter-GPU", - kBackwardFilterTest, - false); -} -#endif -// ======End Convolution TEST====== - -// ======Start DepthwiseConvolution TEST====== - -// TODO(zhaolong) The depthwise convolution cpu test will be added when the cpu -// version of depthwiseConv is implemented. - -#ifndef PADDLE_ONLY_CPU - -TEST(DepthwiseConvForward, GEMM2) { - ConvolutionTest test( - "GemmConv-CPU", "DepthwiseConv-GPU", kForwardTest); - ConvolutionTest2 test2( - "GemmConv-CPU", "DepthwiseConv-GPU", kForwardTest); -} - -TEST(DepthwiseConvBackwardInput, GEMM) { - ConvolutionTest test( - "GemmConvGradInput-CPU", - "DepthwiseConvGradInput-GPU", - kBackwardInputTest); - ConvolutionTest2 test2( - "GemmConvGradInput-CPU", - "DepthwiseConvGradInput-GPU", - kBackwardInputTest); -} - -TEST(DepthwiseConvBackwardFilter, GEMM) { - ConvolutionTest test( - "GemmConvGradFilter-CPU", - "DepthwiseConvGradFilter-GPU", - kBackwardFilterTest); - ConvolutionTest2 test2( - "GemmConvGradFilter-CPU", - "DepthwiseConvGradFilter-GPU", - kBackwardFilterTest); -} - -#endif -// ======End DepthwiseConvolution TEST====== - -} // namespace paddle diff --git a/paddle/function/ConvOpTest.h b/paddle/function/ConvOpTest.h new file mode 100644 index 0000000000000000000000000000000000000000..cb02a96d0dbef6f64fd9e7576179572e68bf5513 --- /dev/null +++ b/paddle/function/ConvOpTest.h @@ -0,0 +1,256 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "FunctionTest.h" + +namespace paddle { + +template +void forward(Compare2Function& test, + const TensorShape& input, + const TensorShape& filter, + const TensorShape& output) { + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input)); + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter)); + test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, output)); + test.run(); +} + +template +void backward_input(Compare2Function& test, + const TensorShape& input, + const TensorShape& filter, + const TensorShape& output) { + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output)); + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter)); + test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, input), ADD_TO); + test.run(); +} + +template +void backward_filter(Compare2Function& test, + const TensorShape& input, + const TensorShape& filter, + const TensorShape& output) { + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output)); + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input)); + test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter), ADD_TO); + test.run(); +} + +template +using Function = void (*)(Compare2Function& test, + const TensorShape& input, + const TensorShape& filter, + const TensorShape& output); + +/** + * \brief A basic convolution function test interface. + * + * \param conv1 type name of convolution function 1. + * \param conv2 type name of convolution function 2. + * \param function test function, can be one of the forward, backward_input + * backward_filter function. + * Example: + * 1. Compare GemmConv's CPU and GPU implementation: + * Convolution( + * "GemmConv-CPU", "GemmConv-GPU", forward); + */ +template +void Convolution(const std::string& conv1, + const std::string& conv2, + Function function) { + for (size_t batchSize : {1, 5}) { + for (size_t inputSize : {7, 14, 31}) { + for (size_t filterSize : {1, 3, 5}) { + for (size_t inputChannels : {3, 16}) { + for (size_t outputChannels : {3, 16}) { + if (outputChannels < inputChannels) continue; + for (size_t stride : {1, 2}) { + for (size_t padding : {0, 1}) { + if (padding >= filterSize) break; + + // NNPACK only supports stride = 1 if batchSize > 1 + if ((conv1 == "NNPACKConv-CPU" || conv2 == "NNPACKConv-CPU") && + batchSize > 1 && stride > 1) + break; + + size_t outputSize = + (inputSize - filterSize + 2 * padding + stride) / stride; + VLOG(3) << " batchSize=" << batchSize + << " inputChannels=" << inputChannels + << " inputHeight=" << inputSize + << " inputWidth=" << inputSize + << " outputChannels=" << outputChannels + << " filterHeight=" << filterSize + << " filterWidth=" << filterSize + << " outputHeight=" << outputSize + << " outputWidth=" << outputSize << " stride=" << stride + << " padding=" << padding; + + std::vector paddings = {padding, padding}; + std::vector strides = {stride, stride}; + Compare2Function test( + conv1, + conv2, + FuncConfig() + .set("paddings", paddings) + .set("strides", strides) + .set("groups", (size_t)1) + .set("algo", (std::string) "auto")); + + TensorShape input{ + batchSize, inputChannels, inputSize, inputSize}; + TensorShape filter{ + outputChannels, inputChannels, filterSize, filterSize}; + TensorShape output{ + batchSize, outputChannels, outputSize, outputSize}; + + function(test, input, filter, output); + } + } + } + } + } + } + } +} + +/** + * \brief A convolution function test interface for + * image height is not equal image width. + */ +template +void Convolution2(const std::string& conv1, + const std::string& conv2, + Function function) { + for (size_t batchSize : {4}) { + for (size_t inputHeight : {7, 31}) { + for (size_t inputWidth : {10, 54}) { + for (size_t filterHeight : {1, 5}) { + for (size_t filterWidth : {3, 7}) { + for (size_t inputChannels : {7}) { + for (size_t outputChannels : {7}) { + size_t stride = 1; + size_t padding = 0; + size_t outputHeight = + (inputHeight - filterHeight + 2 * padding + stride) / + stride; + size_t outputWidth = + (inputWidth - filterWidth + 2 * padding + stride) / stride; + VLOG(3) << " batchSize=" << batchSize + << " inputChannels=" << inputChannels + << " inputHeight=" << inputHeight + << " inputWidth=" << inputWidth + << " outputChannels=" << outputChannels + << " filterHeight=" << filterHeight + << " filterWidth=" << filterWidth + << " outputHeight=" << outputHeight + << " outputWidth=" << outputWidth + << " stride=" << stride << " padding=" << padding; + + std::vector paddings = {padding, padding}; + std::vector strides = {stride, stride}; + Compare2Function test( + conv1, + conv2, + FuncConfig() + .set("paddings", paddings) + .set("strides", strides) + .set("groups", (size_t)1) + .set("algo", (std::string) "auto")); + + TensorShape input{ + batchSize, inputChannels, inputHeight, inputWidth}; + TensorShape filter{ + outputChannels, inputChannels, filterHeight, filterWidth}; + TensorShape output{ + batchSize, outputChannels, outputHeight, outputWidth}; + + function(test, input, filter, output); + } + } + } + } + } + } + } +} + +/** + * \brief A convolution function test interface for depthwise convolution. + */ +template +void DepthwiseConvolution(const std::string& conv1, + const std::string& conv2, + Function function) { + for (size_t batchSize : {1, 32}) { + for (size_t inputSize : {7, 14, 54}) { + for (size_t filterSize : {3, 4}) { + for (size_t inputChannels : {32}) { + for (size_t outputChannels : {32, 64}) { + for (size_t stride : {1, 2}) { + for (size_t padding : {0, 1}) { + // NNPACK only supports stride = 1 if batchSize > 1, + // and there has some bug when batchSize > 1 and groups != 1 + if ((conv1 == "NNPACKConv-CPU" || conv2 == "NNPACKConv-CPU") && + batchSize > 1) + break; + + size_t outputSize = + (inputSize - filterSize + 2 * padding + stride) / stride; + VLOG(3) << " batchSize=" << batchSize + << " inputChannels=" << inputChannels + << " inputHeight=" << inputSize + << " inputWidth=" << inputSize + << " outputChannels=" << outputChannels + << " filterHeight=" << filterSize + << " filterWidth=" << filterSize + << " outputHeight=" << outputSize + << " outputWidth=" << outputSize << " stride=" << stride + << " padding=" << padding; + + std::vector paddings = {padding, padding}; + std::vector strides = {stride, stride}; + size_t groups = inputChannels; + Compare2Function test( + conv1, + conv2, + FuncConfig() + .set("paddings", paddings) + .set("strides", strides) + .set("groups", groups) + .set("algo", (std::string) "auto")); + + TensorShape input{ + batchSize, inputChannels, inputSize, inputSize}; + TensorShape filter{groups, + outputChannels / groups, + inputChannels / groups, + filterSize, + filterSize}; + TensorShape output{ + batchSize, outputChannels, outputSize, outputSize}; + + function(test, input, filter, output); + } + } + } + } + } + } + } +} + +} // namespace paddle diff --git a/paddle/function/CosSimOpGpu.cu b/paddle/function/CosSimOpGpu.cu index c62ab39551f02288618244871ae31c6800df5b42..a1f88f479b5818e3864129a4dac723bceed76fcf 100644 --- a/paddle/function/CosSimOpGpu.cu +++ b/paddle/function/CosSimOpGpu.cu @@ -12,13 +12,13 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include "CosSimOp.h" #include "hl_base.h" #include "hl_device_functions.cuh" -#include "CosSimOp.h" namespace paddle { -template +template __global__ void KeCosSim(real* output, const real* input1, const real* input2, @@ -78,8 +78,8 @@ void hlCossim(real* output, dim3 threads(block_size, 1); dim3 grid(1, input1_height); - KeCosSim<<>> - (output, input1, input2, width, input1_height, input2_height, scale); + KeCosSim<<>>( + output, input1, input2, width, input1_height, input2_height, scale); CHECK_SYNC("hlCossim failed"); } @@ -99,7 +99,7 @@ void CosSimForward(GpuMatrix& out_mat, hlCossim(out, x, y, dim, in1_mat.getHeight(), in2_mat.getHeight(), scale); } -template +template __global__ void KeCosSimDerivative(const real* grad, const real* output, const real* prev_out_x, @@ -148,14 +148,13 @@ __global__ void KeCosSimDerivative(const real* grad, if (xy[0] == 0) { real reciprocal = 1.0 / (sqrt(xx[0]) * sqrt(yy[0])); for (int index = tid; index < width; index += block_size) { - prev_grad_x[index] += - scale * grad[ty] * prev_out_y[index] * reciprocal; + prev_grad_x[index] += scale * grad[ty] * prev_out_y[index] * reciprocal; if (input2_height > 1) { - prev_grad_y[index] += - scale * grad[ty] * prev_out_x[index] * reciprocal; + prev_grad_y[index] += scale * grad[ty] * prev_out_x[index] * reciprocal; } else { - paddle::paddleAtomicAdd(prev_grad_y + index, - scale * grad[ty] * prev_out_x[index] * reciprocal); + paddle::paddleAtomicAdd( + prev_grad_y + index, + scale * grad[ty] * prev_out_x[index] * reciprocal); } } } else { @@ -163,17 +162,18 @@ __global__ void KeCosSimDerivative(const real* grad, real reciprocalSquareSumX = 1.0 / xx[0]; real reciprocalSquareSumY = 1.0 / yy[0]; for (int index = tid; index < width; index += block_size) { - prev_grad_x[index] += output[ty] * grad[ty] * - (prev_out_y[index] * reciprocalXY - - prev_out_x[index] * reciprocalSquareSumX); + prev_grad_x[index] += + output[ty] * grad[ty] * (prev_out_y[index] * reciprocalXY - + prev_out_x[index] * reciprocalSquareSumX); if (input2_height > 1) { - prev_grad_y[index] += output[ty] * grad[ty] * - (prev_out_x[index] * reciprocalXY - - prev_out_y[index] * reciprocalSquareSumY); + prev_grad_y[index] += + output[ty] * grad[ty] * (prev_out_x[index] * reciprocalXY - + prev_out_y[index] * reciprocalSquareSumY); } else { - paddle::paddleAtomicAdd(prev_grad_y + index, output[ty] * grad[ty] * - (prev_out_x[index] * reciprocalXY - - prev_out_y[index] * reciprocalSquareSumY)); + paddle::paddleAtomicAdd( + prev_grad_y + index, + output[ty] * grad[ty] * (prev_out_x[index] * reciprocalXY - + prev_out_y[index] * reciprocalSquareSumY)); } } } @@ -198,9 +198,17 @@ void hlCossimDerivative(const real* grad, const int block_size = 256; dim3 threads(block_size, 1); dim3 grid(1, input1_height); - KeCosSimDerivative<<>> - (grad, output, prev_out_x, prev_out_y, prev_grad_x, prev_grad_y, width, - input1_height, input2_height, scale); + KeCosSimDerivative<<>>( + grad, + output, + prev_out_x, + prev_out_y, + prev_grad_x, + prev_grad_y, + width, + input1_height, + input2_height, + scale); CHECK_SYNC("hlCossimDerivate failed"); } @@ -214,9 +222,9 @@ void CosSimBackward(const GpuMatrix& out_grad, real scale) { CHECK(out_grad.getData() && out_val.getData() && in1_val.getData() && in2_val.getData() && in1_grad.getData() && in2_grad.getData()); - CHECK(out_grad.useGpu_ && out_val.useGpu_ && in1_val.useGpu_ - && in2_val.useGpu_ && in1_grad.useGpu_ && in2_grad.useGpu_) - << "Matrix types are not equally GPU"; + CHECK(out_grad.useGpu_ && out_val.useGpu_ && in1_val.useGpu_ && + in2_val.useGpu_ && in1_grad.useGpu_ && in2_grad.useGpu_) + << "Matrix types are not equally GPU"; size_t dim = in1_val.getWidth(); const real* grad = out_grad.getData(); diff --git a/paddle/function/CropOpGpu.cu b/paddle/function/CropOpGpu.cu index 786eb268d45aadee0c1f6fcbbafc23173cf0bc77..241356a9ca0b673c86ff4c39594722211e2d224e 100644 --- a/paddle/function/CropOpGpu.cu +++ b/paddle/function/CropOpGpu.cu @@ -12,15 +12,23 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "hl_base.h" #include "CropOp.h" +#include "hl_base.h" namespace paddle { -__global__ void KeCrop(real* outputs, const real* inputs, - int inC, int inH, int inW, - int cropC, int cropH, int cropW, - int outC, int outH, int outW, int nthreads) { +__global__ void KeCrop(real* outputs, + const real* inputs, + int inC, + int inH, + int inW, + int cropC, + int cropH, + int cropW, + int outC, + int outH, + int outW, + int nthreads) { const int idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < nthreads) { const int w = idx % outW; @@ -35,12 +43,12 @@ __global__ void KeCrop(real* outputs, const real* inputs, template <> void Crop(real* outputs, - const real* inputs, - const TensorShape inShape, - const TensorShape outShape, - const FuncConfig& conf) { + const real* inputs, + const TensorShape inShape, + const TensorShape outShape, + const FuncConfig& conf) { std::vector crop_corner = - conf.get>("crop_corner"); + conf.get>("crop_corner"); int cropC = crop_corner[1]; int cropH = crop_corner[2]; int cropW = crop_corner[3]; @@ -58,16 +66,33 @@ void Crop(real* outputs, int blockSize = 1024; int gridSize = (nth + blockSize - 1) / blockSize; - KeCrop<<>> - (outputs, inputs, inC, inH, inW, cropC, cropH, cropW, - outC, outH, outW, nth); + KeCrop<<>>(outputs, + inputs, + inC, + inH, + inW, + cropC, + cropH, + cropW, + outC, + outH, + outW, + nth); CHECK_SYNC("Crop"); } -__global__ void KeCropDiff(const real* inGrad, real* outGrad, - int inC, int inH, int inW, - int cropC, int cropH, int cropW, - int outC, int outH, int outW, int nthreads) { +__global__ void KeCropDiff(const real* inGrad, + real* outGrad, + int inC, + int inH, + int inW, + int cropC, + int cropH, + int cropW, + int outC, + int outH, + int outW, + int nthreads) { const int idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < nthreads) { const int w = idx % inW; @@ -84,12 +109,12 @@ __global__ void KeCropDiff(const real* inGrad, real* outGrad, template <> void CropGrad(const real* inGrad, - real* outGrad, - const TensorShape inShape, - const TensorShape outShape, - const FuncConfig& conf) { + real* outGrad, + const TensorShape inShape, + const TensorShape outShape, + const FuncConfig& conf) { std::vector crop_corner = - conf.get>("crop_corner"); + conf.get>("crop_corner"); int cropC = crop_corner[1]; int cropH = crop_corner[2]; int cropW = crop_corner[3]; @@ -107,9 +132,18 @@ void CropGrad(const real* inGrad, int blockSize = 1024; int gridSize = (nth + blockSize - 1) / blockSize; - KeCropDiff <<>> - (inGrad, outGrad, inC, inH, inW, cropC, cropH, cropW, - outC, outH, outW, nth); + KeCropDiff<<>>(inGrad, + outGrad, + inC, + inH, + inW, + cropC, + cropH, + cropW, + outC, + outH, + outW, + nth); CHECK_SYNC("CropGrad"); } diff --git a/paddle/function/CrossMapNormalOpGpu.cu b/paddle/function/CrossMapNormalOpGpu.cu index b33dd108348b7789c6e73bfe3b1ffbc448163ef7..88b991ff6a1f028b333e82e2801ed2e9251aa36d 100644 --- a/paddle/function/CrossMapNormalOpGpu.cu +++ b/paddle/function/CrossMapNormalOpGpu.cu @@ -12,14 +12,18 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "hl_base.h" #include "CrossMapNormalOp.h" +#include "hl_base.h" namespace paddle { -__global__ void KeCMRNormFillScale(size_t imageSize, const real* in, - real* scale, size_t channels, - size_t height, size_t width, size_t size, +__global__ void KeCMRNormFillScale(size_t imageSize, + const real* in, + real* scale, + size_t channels, + size_t height, + size_t width, + size_t size, real alpha) { const int idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < imageSize) { @@ -51,8 +55,10 @@ __global__ void KeCMRNormFillScale(size_t imageSize, const real* in, } } -__global__ void KeCMRNormOutput(size_t inputSize, const real* in, - const real* scale, real negative_beta, +__global__ void KeCMRNormOutput(size_t inputSize, + const real* in, + const real* scale, + real negative_beta, real* out) { const int index = threadIdx.x + blockIdx.x * blockDim.x; if (index < inputSize) { @@ -74,24 +80,30 @@ void CrossMapNormal(real* outputs, size_t imageSize = numSamples * height * width; int blockSize = 1024; int gridSize = (imageSize + 1024 - 1) / 1024; - KeCMRNormFillScale<<>> - (imageSize, inputs, denoms, channels, height, width, size, scale); + KeCMRNormFillScale<<>>( + imageSize, inputs, denoms, channels, height, width, size, scale); - size_t inputSize = numSamples * height * width *channels; + size_t inputSize = numSamples * height * width * channels; blockSize = 1024; gridSize = (inputSize + 1024 - 1) / 1024; - KeCMRNormOutput<<>> - (inputSize, inputs, denoms, -pow, outputs); + KeCMRNormOutput<<>>( + inputSize, inputs, denoms, -pow, outputs); CHECK_SYNC("CrossMapNormal"); } -__global__ void KeCMRNormDiff(size_t imageSize, const real* bottom_data, - const real* top_data, const real* scale, - const real* top_diff, size_t channels, - size_t height, size_t width, size_t size, - real negative_beta, real cache_ratio, - real* bottom_diff ) { +__global__ void KeCMRNormDiff(size_t imageSize, + const real* bottom_data, + const real* top_data, + const real* scale, + const real* top_diff, + size_t channels, + size_t height, + size_t width, + size_t size, + real negative_beta, + real cache_ratio, + real* bottom_diff) { const int idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < imageSize) { const int w = idx % width; @@ -113,17 +125,17 @@ __global__ void KeCMRNormDiff(size_t imageSize, const real* bottom_data, while (index < channels + post_pad) { if (index < channels) { accum += top_diff[index * step] * top_data[index * step] / - scale[index * step]; + scale[index * step]; } if (index >= size) { accum -= top_diff[(index - size) * step] * - top_data[(index - size) * step] / scale[(index - size) * step]; + top_data[(index - size) * step] / scale[(index - size) * step]; } if (index >= post_pad) { bottom_diff[(index - post_pad) * step] += - top_diff[(index - post_pad) * step] * - pow(scale[(index - post_pad) * step], negative_beta) - cache_ratio * - bottom_data[(index - post_pad) * step] * accum; + top_diff[(index - post_pad) * step] * + pow(scale[(index - post_pad) * step], negative_beta) - + cache_ratio * bottom_data[(index - post_pad) * step] * accum; } ++index; } @@ -147,9 +159,18 @@ void CrossMapNormalGrad(real* inputsGrad, int blockSize = 1024; int gridSize = (imageSize + 1024 - 1) / 1024; - KeCMRNormDiff <<>> - (imageSize, inputsValue, outputsValue, denoms, outputsGrad, channels, - height, width, size, -pow, 2.0f * pow * scale, inputsGrad); + KeCMRNormDiff<<>>(imageSize, + inputsValue, + outputsValue, + denoms, + outputsGrad, + channels, + height, + width, + size, + -pow, + 2.0f * pow * scale, + inputsGrad); CHECK_SYNC("CrossMapNormalGrad"); } diff --git a/paddle/function/CrossMapNormalOpTest.cpp b/paddle/function/CrossMapNormalOpTest.cpp index ed17b17da616db9d52318f21c133458d698b0dd8..3b390db77f085aecfd65a9aa64e68ecc189163c1 100644 --- a/paddle/function/CrossMapNormalOpTest.cpp +++ b/paddle/function/CrossMapNormalOpTest.cpp @@ -18,11 +18,11 @@ limitations under the License. */ namespace paddle { TEST(CrossMapNormal, real) { - for (size_t numSamples : {5, 32}) { - for (size_t channels : {1, 5, 32}) { - for (size_t imgSizeH : {5, 33, 100}) { - for (size_t imgSizeW : {5, 32, 96}) { - for (size_t size : {1, 2, 3, 5, 7}) { + for (size_t numSamples : {5}) { + for (size_t channels : {1, 5}) { + for (size_t imgSizeH : {5, 33}) { + for (size_t imgSizeW : {5, 32}) { + for (size_t size : {1, 3}) { VLOG(3) << " numSamples=" << numSamples << " channels=" << channels << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW << " size=" << size; @@ -48,11 +48,11 @@ TEST(CrossMapNormal, real) { } TEST(CrossMapNormalGrad, real) { - for (size_t numSamples : {5, 32}) { - for (size_t channels : {1, 5, 32}) { - for (size_t imgSizeH : {5, 33, 100}) { - for (size_t imgSizeW : {5, 32, 96}) { - for (size_t size : {1, 2, 3, 5, 7}) { + for (size_t numSamples : {5}) { + for (size_t channels : {1, 5}) { + for (size_t imgSizeH : {5, 33}) { + for (size_t imgSizeW : {5, 32}) { + for (size_t size : {1, 3}) { VLOG(3) << " numSamples=" << numSamples << " channels=" << channels << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW << " size=" << size; diff --git a/paddle/function/DepthwiseConvOpGpu.cu b/paddle/function/DepthwiseConvOpGpu.cu index ede0d27aa82e7d71ff5bc33df110fec260e06463..33463805cbd4746c05548028e0bc4a0e2a90453e 100644 --- a/paddle/function/DepthwiseConvOpGpu.cu +++ b/paddle/function/DepthwiseConvOpGpu.cu @@ -20,17 +20,25 @@ namespace paddle { // CUDA kernel to compute the depthwise convolution forward pass template -__global__ -void ConvolutionDepthwiseForward(const int nthreads, - const T* const inputData, const T* const filterData, - const int batchSize, const int outputChannels, const int outputHeight, - const int outputWidth, const int inputChannels, const int inputHeight, - const int inputWidth, const int filterMultiplier, const int filterHeight, - const int filterWidth, const int strideH, const int strideW, - const int paddingH, const int paddingW, T* const outputData) { - - int index = - (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; +__global__ void ConvolutionDepthwiseForward(const int nthreads, + const T* const inputData, + const T* const filterData, + const int batchSize, + const int outputChannels, + const int outputHeight, + const int outputWidth, + const int inputChannels, + const int inputHeight, + const int inputWidth, + const int filterMultiplier, + const int filterHeight, + const int filterWidth, + const int strideH, + const int strideW, + const int paddingH, + const int paddingW, + T* const outputData) { + int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; if (index < nthreads) { const int batch = index / outputChannels / outputHeight / outputWidth; @@ -45,32 +53,36 @@ void ConvolutionDepthwiseForward(const int nthreads, const int w_in_start = -paddingW + w_out * strideW; const int h_in_end = -paddingH + h_out * strideH + filterHeight - 1; const int w_in_end = -paddingW + w_out * strideW + filterWidth - 1; - if ((h_in_start >= 0) && (h_in_end < inputHeight) - && (w_in_start >= 0) && (w_in_end < inputWidth)) { - for (int kh = 0; kh < filterHeight; ++kh) { - for (int kw = 0; kw < filterWidth; ++kw) { - const int h_in = -paddingH + h_out * strideH + kh; - const int w_in = -paddingW + w_out * strideW + kw; - const int offset = ((batch * inputChannels + c_in) - * inputHeight + h_in) * inputWidth + w_in; - value += (*weight) * inputData[offset]; - ++weight; - } + if ((h_in_start >= 0) && (h_in_end < inputHeight) && (w_in_start >= 0) && + (w_in_end < inputWidth)) { + for (int kh = 0; kh < filterHeight; ++kh) { + for (int kw = 0; kw < filterWidth; ++kw) { + const int h_in = -paddingH + h_out * strideH + kh; + const int w_in = -paddingW + w_out * strideW + kw; + const int offset = + ((batch * inputChannels + c_in) * inputHeight + h_in) * + inputWidth + + w_in; + value += (*weight) * inputData[offset]; + ++weight; } + } } else { - for (int kh = 0; kh < filterHeight; ++kh) { - for (int kw = 0; kw < filterWidth; ++kw) { - const int h_in = -paddingH + h_out * strideH + kh; - const int w_in = -paddingW + w_out * strideW + kw; - if ((h_in >= 0) && (h_in < inputHeight) - && (w_in >= 0) && (w_in < inputWidth)) { - const int offset = ((batch * inputChannels + c_in) - * inputHeight + h_in) * inputWidth + w_in; - value += (*weight) * inputData[offset]; - } - ++weight; - } - } + for (int kh = 0; kh < filterHeight; ++kh) { + for (int kw = 0; kw < filterWidth; ++kw) { + const int h_in = -paddingH + h_out * strideH + kh; + const int w_in = -paddingW + w_out * strideW + kw; + if ((h_in >= 0) && (h_in < inputHeight) && (w_in >= 0) && + (w_in < inputWidth)) { + const int offset = + ((batch * inputChannels + c_in) * inputHeight + h_in) * + inputWidth + + w_in; + value += (*weight) * inputData[offset]; + } + ++weight; + } + } } outputData[index] = value; } @@ -78,16 +90,25 @@ void ConvolutionDepthwiseForward(const int nthreads, // CUDA kernel to compute the depthwise convolution backprop w.r.t input. template -__global__ -void ConvolutionDepthwiseInputBackward(const int nthreads, - const T* const top_diff, const T* const weight_data, - const int num, const int outputChannels, const int outputHeight, - const int outputWidth, const int inputChannels, const int inputHeight, - const int inputWidth, const int filterMultiplier, const int filterHeight, - const int filterWidth, const int strideH, const int strideW, - const int paddingH, const int paddingW, T* const bottom_diff) { - int index = - (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; +__global__ void ConvolutionDepthwiseInputBackward(const int nthreads, + const T* const top_diff, + const T* const weight_data, + const int num, + const int outputChannels, + const int outputHeight, + const int outputWidth, + const int inputChannels, + const int inputHeight, + const int inputWidth, + const int filterMultiplier, + const int filterHeight, + const int filterWidth, + const int strideH, + const int strideW, + const int paddingH, + const int paddingW, + T* const bottom_diff) { + int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; if (index < nthreads) { const int batch = index / inputChannels / inputHeight / inputWidth; const int c_in = (index / inputHeight / inputWidth) % inputChannels; @@ -96,65 +117,80 @@ void ConvolutionDepthwiseInputBackward(const int nthreads, const int c_out_start = c_in * filterMultiplier; - int h_out_start = (h_in - filterHeight + paddingH + strideH)/strideH; + int h_out_start = (h_in - filterHeight + paddingH + strideH) / strideH; h_out_start = 0 > h_out_start ? 0 : h_out_start; - int h_out_end = (h_in + paddingH)/strideH; - h_out_end = outputHeight - 1 < h_out_end? outputHeight - 1 : h_out_end; - int w_out_start = (w_in - filterWidth + paddingW + strideW)/strideW; + int h_out_end = (h_in + paddingH) / strideH; + h_out_end = outputHeight - 1 < h_out_end ? outputHeight - 1 : h_out_end; + int w_out_start = (w_in - filterWidth + paddingW + strideW) / strideW; w_out_start = 0 > w_out_start ? 0 : w_out_start; - int w_out_end = (w_in + paddingW)/strideW; - w_out_end = outputWidth - 1 < w_out_end? outputWidth - 1 : w_out_end; + int w_out_end = (w_in + paddingW) / strideW; + w_out_end = outputWidth - 1 < w_out_end ? outputWidth - 1 : w_out_end; T value = 0; - for (int c_out = c_out_start; - c_out < c_out_start + filterMultiplier; c_out ++) { - for (int h_out = h_out_start; h_out <= h_out_end; ++h_out) { - const int filter_h = h_in + paddingH - h_out * strideH; - for (int w_out = w_out_start; w_out <= w_out_end; ++w_out) { - const int filter_w = w_in + paddingW - w_out * strideW; - const int filter_offset = c_out * filterHeight * filterWidth - + filter_h * filterWidth + filter_w; - const int top_diff_offset = ((batch * outputChannels + c_out) * - outputHeight + h_out)* outputWidth + w_out; - value += top_diff[top_diff_offset] * weight_data[filter_offset]; - } + for (int c_out = c_out_start; c_out < c_out_start + filterMultiplier; + c_out++) { + for (int h_out = h_out_start; h_out <= h_out_end; ++h_out) { + const int filter_h = h_in + paddingH - h_out * strideH; + for (int w_out = w_out_start; w_out <= w_out_end; ++w_out) { + const int filter_w = w_in + paddingW - w_out * strideW; + const int filter_offset = c_out * filterHeight * filterWidth + + filter_h * filterWidth + filter_w; + const int top_diff_offset = + ((batch * outputChannels + c_out) * outputHeight + h_out) * + outputWidth + + w_out; + value += top_diff[top_diff_offset] * weight_data[filter_offset]; } + } } bottom_diff[index] += value; - } + } } // CUDA kernel to compute the depthwise convolution backprop w.r.t filter. template -__global__ -void ConvolutionDepthwiseFilterBackward(const int num_i, const int nthreads, - const T* const top_diff, const T* const inputData, - const int num, const int outputChannels, const int outputHeight, - const int outputWidth, const int inputChannels, const int inputHeight, - const int inputWidth, const int filterMultiplier, const int filterHeight, - const int filterWidth, const int strideH, const int strideW, - const int paddingH, const int paddingW, T* const buffer_data) { - int index = - (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; +__global__ void ConvolutionDepthwiseFilterBackward(const int num_i, + const int nthreads, + const T* const top_diff, + const T* const inputData, + const int num, + const int outputChannels, + const int outputHeight, + const int outputWidth, + const int inputChannels, + const int inputHeight, + const int inputWidth, + const int filterMultiplier, + const int filterHeight, + const int filterWidth, + const int strideH, + const int strideW, + const int paddingH, + const int paddingW, + T* const buffer_data) { + int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; if (index < nthreads) { const int h_out = (index / outputWidth) % outputHeight; const int w_out = index % outputWidth; - const int kh = (index / filterWidth / outputHeight / outputWidth) - % filterHeight; + const int kh = + (index / filterWidth / outputHeight / outputWidth) % filterHeight; const int kw = (index / outputHeight / outputWidth) % filterWidth; const int h_in = -paddingH + h_out * strideH + kh; const int w_in = -paddingW + w_out * strideW + kw; - if ((h_in >= 0) && (h_in < inputHeight) - && (w_in >= 0) && (w_in < inputWidth)) { - const int c_out = index / - (filterHeight * filterWidth * outputHeight * outputWidth); + if ((h_in >= 0) && (h_in < inputHeight) && (w_in >= 0) && + (w_in < inputWidth)) { + const int c_out = + index / (filterHeight * filterWidth * outputHeight * outputWidth); const int c_in = c_out / filterMultiplier; const int batch = num_i; - const int top_offset = ((batch * outputChannels + c_out) * - outputHeight + h_out) * outputWidth + w_out; - const int bottom_offset = ((batch * inputChannels + c_in) - * inputHeight + h_in) * inputWidth + w_in; + const int top_offset = + ((batch * outputChannels + c_out) * outputHeight + h_out) * + outputWidth + + w_out; + const int bottom_offset = + ((batch * inputChannels + c_in) * inputHeight + h_in) * inputWidth + + w_in; buffer_data[index] = top_diff[top_offset] * inputData[bottom_offset]; } else { buffer_data[index] = 0; @@ -163,170 +199,169 @@ void ConvolutionDepthwiseFilterBackward(const int num_i, const int nthreads, } template -class DepthwiseConvFunctor{ +class DepthwiseConvFunctor { public: void operator()(const T* inputData, - const T* filterData, - int batchSize, - int outputChannels, - int outputHeight, - int outputWidth, - int inputChannels, - int inputHeight, - int inputWidth, - int filterMultiplier, - int filterHeight, - int filterWidth, - int strideH, - int strideW, - int paddingH, - int paddingW, - T* outputData){ + const T* filterData, + int batchSize, + int outputChannels, + int outputHeight, + int outputWidth, + int inputChannels, + int inputHeight, + int inputWidth, + int filterMultiplier, + int filterHeight, + int filterWidth, + int strideH, + int strideW, + int paddingH, + int paddingW, + T* outputData) { int outputSize = batchSize * outputChannels * outputHeight * outputWidth; - size_t blocks = (outputSize + 1024 -1) / 1024; + size_t blocks = (outputSize + 1024 - 1) / 1024; size_t blockX = 512; - size_t blockY = (blocks+512-1)/512; + size_t blockY = (blocks + 512 - 1) / 512; dim3 threads(1024, 1); dim3 grid(blockX, blockY); - ConvolutionDepthwiseForward - <<< grid, threads, 0, STREAM_DEFAULT >>>( - outputSize, - inputData, - filterData, - batchSize, - outputChannels, - outputHeight, - outputWidth, - inputChannels, - inputHeight, - inputWidth, - filterMultiplier, - filterHeight, - filterWidth, - strideH, - strideW, - paddingH, - paddingW, - outputData); - } + ConvolutionDepthwiseForward<<>>( + outputSize, + inputData, + filterData, + batchSize, + outputChannels, + outputHeight, + outputWidth, + inputChannels, + inputHeight, + inputWidth, + filterMultiplier, + filterHeight, + filterWidth, + strideH, + strideW, + paddingH, + paddingW, + outputData); + } }; template -class DepthwiseConvGradInputFunctor{ +class DepthwiseConvGradInputFunctor { public: void operator()(const T* outputGrad, - const T* filterData, - int batchSize, - int outputChannels, - int outputHeight, - int outputWidth, - int inputChannels, - int inputHeight, - int inputWidth, - int filterMultiplier, - int filterHeight, - int filterWidth, - int strideH, - int strideW, - int paddingH, - int paddingW, - T* inputGrad){ + const T* filterData, + int batchSize, + int outputChannels, + int outputHeight, + int outputWidth, + int inputChannels, + int inputHeight, + int inputWidth, + int filterMultiplier, + int filterHeight, + int filterWidth, + int strideH, + int strideW, + int paddingH, + int paddingW, + T* inputGrad) { int inputSize = batchSize * inputChannels * inputHeight * inputWidth; - size_t blocks = (inputSize + 1024 -1) / 1024; + size_t blocks = (inputSize + 1024 - 1) / 1024; size_t blockX = 512; - size_t blockY = (blocks+512-1)/512; + size_t blockY = (blocks + 512 - 1) / 512; dim3 threads(1024, 1); dim3 grid(blockX, blockY); - ConvolutionDepthwiseInputBackward - // NOLINT_NEXT_LINE(whitespace/operators) - <<< grid, threads, 0, STREAM_DEFAULT >>>( - inputSize, - outputGrad, - filterData, - batchSize, - outputChannels, - outputHeight, - outputWidth, - inputChannels, - inputHeight, - inputWidth, - filterMultiplier, - filterHeight, - filterWidth, - strideH, - strideW, - paddingH, - paddingW, - inputGrad); - } + // NOLINT_NEXT_LINE(whitespace/operators) + <<>>(inputSize, + outputGrad, + filterData, + batchSize, + outputChannels, + outputHeight, + outputWidth, + inputChannels, + inputHeight, + inputWidth, + filterMultiplier, + filterHeight, + filterWidth, + strideH, + strideW, + paddingH, + paddingW, + inputGrad); + } }; template class DepthwiseConvGradFilterFunctor { public: void operator()(const T* outputGrad, - const T* inputData, - int batchSize, - int outputChannels, - int outputHeight, - int outputWidth, - int inputChannels, - int inputHeight, - int inputWidth, - int filterMultiplier, - int filterHeight, - int filterWidth, - int strideH, - int strideW, - int paddingH, - int paddingW, - T* colData, - T* filterGrad){ - int colDataSize = outputChannels * filterHeight * filterWidth - * outputHeight * outputWidth; + const T* inputData, + int batchSize, + int outputChannels, + int outputHeight, + int outputWidth, + int inputChannels, + int inputHeight, + int inputWidth, + int filterMultiplier, + int filterHeight, + int filterWidth, + int strideH, + int strideW, + int paddingH, + int paddingW, + T* colData, + T* filterGrad) { + int colDataSize = outputChannels * filterHeight * filterWidth * + outputHeight * outputWidth; - size_t blocks = (colDataSize + 1024 -1) / 1024; - size_t blockX = 512; - size_t blockY = (blocks+512-1)/512; - dim3 threads(1024, 1); - dim3 grid(blockX, blockY); - BaseMatrix filterGradMatrix(outputChannels * filterHeight * filterWidth, - 1, filterGrad, false, true); + size_t blocks = (colDataSize + 1024 - 1) / 1024; + size_t blockX = 512; + size_t blockY = (blocks + 512 - 1) / 512; + dim3 threads(1024, 1); + dim3 grid(blockX, blockY); + BaseMatrix filterGradMatrix(outputChannels * filterHeight * filterWidth, + 1, + filterGrad, + false, + true); - for (int i = 0; i < batchSize; i++) { - ConvolutionDepthwiseFilterBackward - <<< grid, threads, 0, STREAM_DEFAULT >>>( - i, - colDataSize, - outputGrad, - inputData, - batchSize, - outputChannels, - outputHeight, - outputWidth, - inputChannels, - inputHeight, - inputWidth, - filterMultiplier, - filterHeight, - filterWidth, - strideH, - strideW, - paddingH, - paddingW, - colData); - int K = outputHeight * outputWidth; - int M = colDataSize / K; + for (int i = 0; i < batchSize; i++) { + ConvolutionDepthwiseFilterBackward< + T><<>>(i, + colDataSize, + outputGrad, + inputData, + batchSize, + outputChannels, + outputHeight, + outputWidth, + inputChannels, + inputHeight, + inputWidth, + filterMultiplier, + filterHeight, + filterWidth, + strideH, + strideW, + paddingH, + paddingW, + colData); + int K = outputHeight * outputWidth; + int M = colDataSize / K; - BaseMatrix colMatrix(M, K, colData, false, true); - filterGradMatrix.sumRows(colMatrix, (T)1.0, (T)1.0); - } + BaseMatrix colMatrix(M, K, colData, false, true); + filterGradMatrix.sumRows(colMatrix, (T)1.0, (T)1.0); } + } }; #ifdef PADDLE_TYPE_DOUBLE diff --git a/paddle/operators/mean_op_test.cc b/paddle/function/DepthwiseConvOpTest.cpp similarity index 51% rename from paddle/operators/mean_op_test.cc rename to paddle/function/DepthwiseConvOpTest.cpp index 375dcd50e130355c60f82b9d39d1b94fb2c911b0..f44ae0c342e9536366e2b537694cee81fcb1a6ed 100644 --- a/paddle/operators/mean_op_test.cc +++ b/paddle/function/DepthwiseConvOpTest.cpp @@ -13,13 +13,25 @@ See the License for the specific language governing permissions and limitations under the License. */ #include +#include "ConvOpTest.h" -#include +namespace paddle { -USE_OP(mean); +#ifndef PADDLE_ONLY_CPU +TEST(DepthwiseConv, Forward) { + DepthwiseConvolution( + "GemmConv-CPU", "DepthwiseConv-GPU", forward); +} + +TEST(DepthwiseConv, BackwardInput) { + DepthwiseConvolution( + "GemmConvGradInput-CPU", "DepthwiseConvGradInput-GPU", backward_input); +} -TEST(MeanOp, GetOpProto) { - auto& protos = paddle::framework::OpRegistry::protos(); - auto it = protos.find("mean"); - ASSERT_NE(it, protos.end()); +TEST(DepthwiseConv, BackwardFilter) { + DepthwiseConvolution( + "GemmConvGradFilter-CPU", "DepthwiseConvGradFilter-GPU", backward_filter); } +#endif + +} // namespace paddle diff --git a/paddle/function/FunctionTest.cpp b/paddle/function/FunctionTest.cpp index fdf7e631e5ab8c67eb5cf906bd0af49740d60112..7b0b1c6adbd36a6cbe9b89e9518fbe07fb1db368 100644 --- a/paddle/function/FunctionTest.cpp +++ b/paddle/function/FunctionTest.cpp @@ -24,14 +24,14 @@ void FunctionApi(typename Tensor::Matrix& output, template <> void FunctionApi(CpuMatrix& output, const CpuMatrix& input) { - EXPECT_EQ(output.getHeight(), 100); - EXPECT_EQ(output.getWidth(), 200); + EXPECT_EQ(output.getHeight(), 100U); + EXPECT_EQ(output.getWidth(), 200U); } template <> void FunctionApi(GpuMatrix& output, const GpuMatrix& input) { - EXPECT_EQ(output.getHeight(), 10); - EXPECT_EQ(output.getWidth(), 20); + EXPECT_EQ(output.getHeight(), 10U); + EXPECT_EQ(output.getWidth(), 20U); } template @@ -85,16 +85,16 @@ void testBufferArgs(const BufferArgs& inputs, } void testBufferArgs(const BufferArgs& inputs, const CheckBufferArg& check) { - EXPECT_EQ(inputs.size(), 1); + EXPECT_EQ(inputs.size(), 1U); check(inputs[0]); } TEST(Arguments, Matrix) { MatrixPtr matrix = Matrix::create(100, 200); CheckBufferArg check = [=](const BufferArg& arg) { - EXPECT_EQ(arg.shape().ndims(), 2); - EXPECT_EQ(arg.shape()[0], 100); - EXPECT_EQ(arg.shape()[1], 200); + EXPECT_EQ(arg.shape().ndims(), 2U); + EXPECT_EQ(arg.shape()[0], 100U); + EXPECT_EQ(arg.shape()[1], 200U); EXPECT_EQ(arg.data(), matrix->getData()); EXPECT_EQ(arg.matrix().getHeight(), matrix->getHeight()); @@ -112,8 +112,8 @@ TEST(Arguments, Matrix) { TEST(Arguments, Vector) { VectorPtr vector = Vector::create(100, false); CheckBufferArg check = [=](const BufferArg& arg) { - EXPECT_EQ(arg.shape().ndims(), 1); - EXPECT_EQ(arg.shape()[0], 100); + EXPECT_EQ(arg.shape().ndims(), 1U); + EXPECT_EQ(arg.shape()[0], 100U); EXPECT_EQ(arg.data(), vector->getData()); CpuVector inVector = arg.vector(); @@ -131,9 +131,9 @@ TEST(Arguments, Vector) { TEST(Arguments, CpuSparseMatrix) { CpuSparseMatrix sparse(200, 300, 50); CheckBufferArg check = [=](const BufferArg& arg) { - EXPECT_EQ(arg.shape().ndims(), 2); - EXPECT_EQ(arg.shape()[0], 200); - EXPECT_EQ(arg.shape()[1], 300); + EXPECT_EQ(arg.shape().ndims(), 2U); + EXPECT_EQ(arg.shape()[0], 200U); + EXPECT_EQ(arg.shape()[1], 300U); EXPECT_EQ(arg.data(), sparse.getData()); // CHECK_EQ(arg.sparse().nnz(), 50); // CHECK_EQ(arg.sparse().dataFormat(), SPARSE_CSR_FORMAT); @@ -152,10 +152,10 @@ TEST(Arguments, CpuSparseMatrix) { TEST(Arguments, BufferArg) { BufferArg arg(nullptr, VALUE_TYPE_FLOAT, {1, 2, 3}); CheckBufferArg check = [=](const BufferArg& arg) { - EXPECT_EQ(arg.shape().ndims(), 3); - EXPECT_EQ(arg.shape()[0], 1); - EXPECT_EQ(arg.shape()[1], 2); - EXPECT_EQ(arg.shape()[2], 3); + EXPECT_EQ(arg.shape().ndims(), 3U); + EXPECT_EQ(arg.shape()[0], 1U); + EXPECT_EQ(arg.shape()[1], 2U); + EXPECT_EQ(arg.shape()[2], 3U); }; BufferArgs argments; diff --git a/paddle/function/GemmConvOpTest.cpp b/paddle/function/GemmConvOpTest.cpp new file mode 100644 index 0000000000000000000000000000000000000000..5283d79a5a53d979ae4e134f7e46b7ee106e9c44 --- /dev/null +++ b/paddle/function/GemmConvOpTest.cpp @@ -0,0 +1,50 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES 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 "ConvOpTest.h" + +namespace paddle { + +TEST(GemmConv, NaiveConv) { + Convolution( + "NaiveConv-CPU", "GemmConv-CPU", forward); + Convolution2( + "NaiveConv-CPU", "GemmConv-CPU", forward); +} + +#ifndef PADDLE_ONLY_CPU +TEST(GemmConv, Forward) { + Convolution( + "GemmConv-CPU", "GemmConv-GPU", forward); + Convolution2( + "GemmConv-CPU", "GemmConv-GPU", forward); +} + +TEST(GemmConv, BackwardInput) { + Convolution( + "GemmConvGradInput-CPU", "GemmConvGradInput-GPU", backward_input); + Convolution2( + "GemmConvGradInput-CPU", "GemmConvGradInput-GPU", backward_input); +} + +TEST(GemmConv, BackwardFilter) { + Convolution( + "GemmConvGradFilter-CPU", "GemmConvGradFilter-GPU", backward_filter); + Convolution2( + "GemmConvGradFilter-CPU", "GemmConvGradFilter-GPU", backward_filter); +} +#endif + +} // namespace paddle diff --git a/paddle/function/Im2ColOpGpu.cu b/paddle/function/Im2ColOpGpu.cu index 15ba854009636d027447d104071163100d5e3f4b..bd98610498b1af003574129118be4684d38e5813 100644 --- a/paddle/function/Im2ColOpGpu.cu +++ b/paddle/function/Im2ColOpGpu.cu @@ -17,16 +17,21 @@ limitations under the License. */ namespace paddle { -template -__global__ -void im2col(const T* data_im, int numOuts, int height, int width, - int blockH, int blockW, - int strideH, int strideW, - int paddingH, int paddingW, - int height_col, int width_col, - T* data_col) { - int index = - (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; +template +__global__ void im2col(const T* data_im, + int numOuts, + int height, + int width, + int blockH, + int blockW, + int strideH, + int strideW, + int paddingH, + int paddingW, + int height_col, + int width_col, + T* data_col) { + int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; if (index < numOuts) { int w_out = index % width_col; index /= width_col; @@ -39,17 +44,17 @@ void im2col(const T* data_im, int numOuts, int height, int width, data_col += (channel_out * height_col + h_out) * width_col + w_out; for (int i = 0; i < blockH; ++i) { for (int j = 0; j < blockW; ++j) { - int rIdx = int(h_in+i); - int cIdx = int(w_in+j); - if ((rIdx-(int)paddingH) >= (int)height || - (rIdx-(int)paddingH) < 0 || - (cIdx-(int)paddingW) >= (int)width || - (cIdx-(int)paddingW) < 0) { + int rIdx = int(h_in + i); + int cIdx = int(w_in + j); + if ((rIdx - (int)paddingH) >= (int)height || + (rIdx - (int)paddingH) < 0 || + (cIdx - (int)paddingW) >= (int)width || + (cIdx - (int)paddingW) < 0) { *data_col = 0; } else { - rIdx = rIdx + channel_in*height - paddingH; + rIdx = rIdx + channel_in * height - paddingH; cIdx = cIdx - paddingW; - *data_col = data_im[rIdx* width + cIdx]; + *data_col = data_im[rIdx * width + cIdx]; } data_col += height_col * width_col; } @@ -82,60 +87,73 @@ public: int outputWidth = colShape[4]; int numKernels = inputChannels * outputHeight * outputWidth; - int blocks = (numKernels + 1024 -1) / 1024; + int blocks = (numKernels + 1024 - 1) / 1024; int blockX = 512; int blockY = (blocks + 512 - 1) / 512; dim3 threads(1024, 1); dim3 grid(blockX, blockY); - im2col<<< grid, threads, 0, STREAM_DEFAULT >>> - (imData, numKernels, inputHeight, inputWidth, filterHeight, filterWidth, - strideHeight, strideWidth, paddingHeight, paddingWidth, - outputHeight, outputWidth, colData); + im2col<<>>(imData, + numKernels, + inputHeight, + inputWidth, + filterHeight, + filterWidth, + strideHeight, + strideWidth, + paddingHeight, + paddingWidth, + outputHeight, + outputWidth, + colData); CHECK_SYNC("Im2ColFunctor GPU failed"); } }; -template -__global__ -void col2im(size_t n, const T* data_col, size_t height, - size_t width, size_t channels, - size_t blockH, size_t blockW, - size_t strideH, size_t strideW, - size_t paddingH, size_t paddingW, - size_t height_col, size_t width_col, - T* data_im) { +template +__global__ void col2im(size_t n, + const T* data_col, + size_t height, + size_t width, + size_t channels, + size_t blockH, + size_t blockW, + size_t strideH, + size_t strideW, + size_t paddingH, + size_t paddingW, + size_t height_col, + size_t width_col, + T* data_im) { size_t index = - (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; + (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)paddingW) >= 0 && - (w - (int)paddingW) < (width-2 * paddingW) && - (h - (int)paddingH) >= 0 && - (h - paddingH) < (height - 2 * paddingH)) { + (w - (int)paddingW) < (width - 2 * paddingW) && + (h - (int)paddingH) >= 0 && (h - paddingH) < (height - 2 * paddingH)) { // compute the start and end of the output int w_col_start = - (w < (int)blockW) ? 0 : (w - int(blockW)) / (int)strideW + 1; - int w_col_end = - min((int)(w / (int)strideW + 1), (int)(width_col)); + (w < (int)blockW) ? 0 : (w - int(blockW)) / (int)strideW + 1; + int w_col_end = min((int)(w / (int)strideW + 1), (int)(width_col)); int h_col_start = - (h < (int)blockH) ? 0 : (h - (int)blockH) / (int)strideH + 1; + (h < (int)blockH) ? 0 : (h - (int)blockH) / (int)strideH + 1; int h_col_end = min(int(h / strideH + 1), int(height_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) { // the col location: [c * width * height + h_out, w_out] - int c_col = int(c * blockH* blockW) + \ - (h - h_col * (int)strideH) * (int)blockW + - (w - w_col * (int)strideW); + int c_col = int(c * blockH * blockW) + + (h - h_col * (int)strideH) * (int)blockW + + (w - w_col * (int)strideW); val += data_col[(c_col * height_col + h_col) * width_col + w_col]; } } h -= paddingH; w -= paddingW; - data_im[c*((width-2*paddingW) * (height-2*paddingH)) + - h*(width-2*paddingW) + w] += val; + data_im[c * ((width - 2 * paddingW) * (height - 2 * paddingH)) + + h * (width - 2 * paddingW) + w] += val; } } } @@ -164,32 +182,32 @@ public: int outputHeight = colShape[3]; int outputWidth = colShape[4]; - size_t numKernels = inputChannels * (inputHeight + 2*paddingHeight) - * (inputWidth + 2*paddingWidth); + size_t numKernels = inputChannels * (inputHeight + 2 * paddingHeight) * + (inputWidth + 2 * paddingWidth); - size_t blocks = (numKernels + 1024 -1) / 1024; + size_t blocks = (numKernels + 1024 - 1) / 1024; size_t blockX = 512; - size_t blockY = (blocks+512-1)/512; + size_t blockY = (blocks + 512 - 1) / 512; dim3 threads(1024, 1); dim3 grid(blockX, blockY); // 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, STREAM_DEFAULT >>> - (numKernels, - colData, - inputHeight + 2*paddingHeight, - inputWidth + 2*paddingWidth, - inputChannels, - filterHeight, - filterWidth, - strideHeight, - strideWidth, - paddingHeight, - paddingWidth, - outputHeight, - outputWidth, - imData); + col2im<<>>( + numKernels, + colData, + inputHeight + 2 * paddingHeight, + inputWidth + 2 * paddingWidth, + inputChannels, + filterHeight, + filterWidth, + strideHeight, + strideWidth, + paddingHeight, + paddingWidth, + outputHeight, + outputWidth, + imData); CHECK_SYNC("Col2ImFunctor GPU failed"); } }; @@ -199,31 +217,35 @@ template class Im2ColFunctor; template class Col2ImFunctor; template class Col2ImFunctor; -template -__global__ -void im2colOCF(const T* imData, T* colData, - int inputChannels, - int inputHeight, int inputWidth, - int filterHeight, int filterWidth, - int strideHeight, int strideWidth, - int paddingHeight, int paddingWidth, - int outputHeight, int outputWidth) { +template +__global__ void im2colOCF(const T* imData, + T* colData, + int inputChannels, + int inputHeight, + int inputWidth, + int filterHeight, + int filterWidth, + int strideHeight, + int strideWidth, + int paddingHeight, + int paddingWidth, + int outputHeight, + int outputWidth) { int swId = blockIdx.x; int shId = blockIdx.y; - for (int channelId = threadIdx.z; - channelId < inputChannels; + for (int channelId = threadIdx.z; channelId < inputChannels; channelId += blockDim.z) { for (int idy = threadIdx.y; idy < filterHeight; idy += blockDim.y) { for (int idx = threadIdx.x; idx < filterWidth; idx += blockDim.x) { int widthOffset = idx + swId * strideWidth - paddingWidth; int heightOffset = idy + shId * strideHeight - paddingHeight; - int imOffset = widthOffset + heightOffset * inputWidth - + channelId * inputHeight * inputWidth; + int imOffset = widthOffset + heightOffset * inputWidth + + channelId * inputHeight * inputWidth; - int colOffset = idx + idy * filterWidth - + channelId * filterHeight * filterWidth - + (shId * outputWidth + swId) - * (inputChannels * filterHeight * filterWidth); + int colOffset = idx + idy * filterWidth + + channelId * filterHeight * filterWidth + + (shId * outputWidth + swId) * + (inputChannels * filterHeight * filterWidth); if (heightOffset >= inputHeight || heightOffset < 0 || widthOffset >= inputWidth || widthOffset < 0) { @@ -279,39 +301,52 @@ public: int blockDimZ = 1024 / blockDimX / blockDimY; dim3 threads(blockDimX, blockDimY, std::min(blockDimZ, inputChannels)); dim3 grid(outputWidth, outputHeight); - im2colOCF<<< grid, threads, 0, STREAM_DEFAULT >>> - (imData, colData, inputChannels, inputHeight, inputWidth, - filterHeight, filterWidth, strideHeight, strideWidth, - paddingHeight, paddingWidth, outputHeight, outputWidth); + im2colOCF<<>>(imData, + colData, + inputChannels, + inputHeight, + inputWidth, + filterHeight, + filterWidth, + strideHeight, + strideWidth, + paddingHeight, + paddingWidth, + outputHeight, + outputWidth); CHECK_SYNC("Im2ColFunctor GPU failed"); } }; -template -__global__ -void col2imOCF(T* imData, const T* colData, - int inputChannels, - int inputHeight, int inputWidth, - int filterHeight, int filterWidth, - int strideHeight, int strideWidth, - int paddingHeight, int paddingWidth, - int outputHeight, int outputWidth) { +template +__global__ void col2imOCF(T* imData, + const T* colData, + int inputChannels, + int inputHeight, + int inputWidth, + int filterHeight, + int filterWidth, + int strideHeight, + int strideWidth, + int paddingHeight, + int paddingWidth, + int outputHeight, + int outputWidth) { int swId = blockIdx.x; int shId = blockIdx.y; - for (int channelId = threadIdx.z; - channelId < inputChannels; + for (int channelId = threadIdx.z; channelId < inputChannels; channelId += blockDim.z) { for (int idy = threadIdx.y; idy < filterHeight; idy += blockDim.y) { for (int idx = threadIdx.x; idx < filterWidth; idx += blockDim.x) { int widthOffset = idx + swId * strideWidth - paddingWidth; int heightOffset = idy + shId * strideHeight - paddingHeight; - int imOffset = widthOffset + heightOffset * inputWidth - + channelId * inputHeight * inputWidth; + int imOffset = widthOffset + heightOffset * inputWidth + + channelId * inputHeight * inputWidth; - int colOffset = idx + idy * filterWidth - + channelId * filterHeight * filterWidth - + (shId * outputWidth + swId) - * (inputChannels * filterHeight * filterWidth); + int colOffset = idx + idy * filterWidth + + channelId * filterHeight * filterWidth + + (shId * outputWidth + swId) * + (inputChannels * filterHeight * filterWidth); if (heightOffset >= 0 && heightOffset < inputHeight && widthOffset >= 0 && widthOffset < inputWidth) { @@ -365,10 +400,19 @@ public: int blockDimZ = 1024 / blockDimX / blockDimY; dim3 threads(blockDimX, blockDimY, std::min(blockDimZ, inputChannels)); dim3 grid(outputWidth, outputHeight); - col2imOCF<<< grid, threads, 0, STREAM_DEFAULT >>> - (imData, colData, inputChannels, inputHeight, inputWidth, - filterHeight, filterWidth, strideHeight, strideWidth, - paddingHeight, paddingWidth, outputHeight, outputWidth); + col2imOCF<<>>(imData, + colData, + inputChannels, + inputHeight, + inputWidth, + filterHeight, + filterWidth, + strideHeight, + strideWidth, + paddingHeight, + paddingWidth, + outputHeight, + outputWidth); CHECK_SYNC("Col2ImFunctor GPU failed"); } }; diff --git a/paddle/function/MulOpGpu.cu b/paddle/function/MulOpGpu.cu index dcfcb2325d7dae22e0e0e78fc0bddf061fc0940c..9449b89056b4b1740cb4c3de630348b1b361d61e 100644 --- a/paddle/function/MulOpGpu.cu +++ b/paddle/function/MulOpGpu.cu @@ -12,8 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "hl_base.h" #include "MulOp.h" +#include "hl_base.h" #include "paddle/math/Matrix.h" #include "paddle/math/SparseMatrix.h" diff --git a/paddle/function/PadOpGpu.cu b/paddle/function/PadOpGpu.cu index 9094f1528433fdcaad3397a991aa8ac6fa04bc01..5b6f4e6832aea4bcfe22e530f5f25ef5815729f1 100644 --- a/paddle/function/PadOpGpu.cu +++ b/paddle/function/PadOpGpu.cu @@ -12,15 +12,23 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "hl_base.h" #include "PadOp.h" +#include "hl_base.h" namespace paddle { -__global__ void KePad(real* outputs, const real* inputs, - int inC, int inH, int inW, - int padc, int padh, int padw, - int outC, int outH, int outW, int nthreads) { +__global__ void KePad(real* outputs, + const real* inputs, + int inC, + int inH, + int inW, + int padc, + int padh, + int padw, + int outC, + int outH, + int outW, + int nthreads) { const int idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < nthreads) { const int w = idx % inW; @@ -50,16 +58,33 @@ void Pad(real* outputs, int outC = inC + cstart + cend; int outH = inH + hstart + hend; int outW = inW + wstart + wend; - KePad<<>> - (outputs, inputs, inC, inH, inW, cstart, hstart, wstart, - outC, outH, outW, nth); + KePad<<>>(outputs, + inputs, + inC, + inH, + inW, + cstart, + hstart, + wstart, + outC, + outH, + outW, + nth); CHECK_SYNC("Pad"); } -__global__ void KePadDiff(real* inGrad, const real* outGrad, - int inC, int inH, int inW, - int padc, int padh, int padw, - int outC, int outH, int outW, int nthreads) { +__global__ void KePadDiff(real* inGrad, + const real* outGrad, + int inC, + int inH, + int inW, + int padc, + int padh, + int padw, + int outC, + int outH, + int outW, + int nthreads) { const int idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < nthreads) { const int w = idx % inW; @@ -89,9 +114,18 @@ void PadGrad(real* inGrad, int outC = inC + cstart + cend; int outH = inH + hstart + hend; int outW = inW + wstart + wend; - KePadDiff <<>> - (inGrad, outGrad, inC, inH, inW, cstart, hstart, wstart, - outC, outH, outW, nth); + KePadDiff<<>>(inGrad, + outGrad, + inC, + inH, + inW, + cstart, + hstart, + wstart, + outC, + outH, + outW, + nth); CHECK_SYNC("PadGrad"); } diff --git a/paddle/function/RowConvOpGpu.cu b/paddle/function/RowConvOpGpu.cu index d9dcc7d59d1e3c222f5a7ce448daa8d7edb6c978..b0cbd9fd1df9a35d6cc1cb5312099d8b45197944 100644 --- a/paddle/function/RowConvOpGpu.cu +++ b/paddle/function/RowConvOpGpu.cu @@ -12,16 +12,20 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "hl_base.h" #include "RowConvOp.h" +#include "hl_base.h" namespace paddle { -template -__global__ void KeRowConv(real* y, const real* x, const real* w, - const int* starts, const int height, const int width, - const int numSeq, const int context) { - +template +__global__ void KeRowConv(real* y, + const real* x, + const real* w, + const int* starts, + const int height, + const int width, + const int numSeq, + const int context) { const int tidx = threadIdx.x; const int tidy = threadIdx.y; const int blky = blockDim.y; @@ -30,7 +34,7 @@ __global__ void KeRowConv(real* y, const real* x, const real* w, __shared__ real sw[BLOCK_H][BLOCK_W]; for (int i = tidy; i < context; i += blky) { - sw[i][tidx] = gidx + tidx < width ? w[i*width + gidx + tidx] : 0.0; + sw[i][tidx] = gidx + tidx < width ? w[i * width + gidx + tidx] : 0.0; } __syncthreads(); @@ -56,9 +60,14 @@ __global__ void KeRowConv(real* y, const real* x, const real* w, } } -__global__ void KeRowConv2(real* y, const real* x, const real* w, - const int* starts, const int height, const int width, - const int numSeq, const int context) { +__global__ void KeRowConv2(real* y, + const real* x, + const real* w, + const int* starts, + const int height, + const int width, + const int numSeq, + const int context) { const int tidx = threadIdx.x; const int tidy = threadIdx.y; const int blky = blockDim.y; @@ -84,8 +93,6 @@ __global__ void KeRowConv2(real* y, const real* x, const real* w, } } - - template <> void RowConv(GpuMatrix& out, const GpuMatrix& in, @@ -105,21 +112,24 @@ void RowConv(GpuMatrix& out, dim3 dimGrid(DIVUP(width, dimBlock.x), 1); if (contextLength <= 32) { - KeRowConv<32, 32><<>> - (y, x, w, starts, height, width, numSeq, contextLength); + KeRowConv<32, 32><<>>( + y, x, w, starts, height, width, numSeq, contextLength); } else { - KeRowConv2<<>> - (y, x, w, starts, height, width, numSeq, contextLength); + KeRowConv2<<>>( + y, x, w, starts, height, width, numSeq, contextLength); } CHECK_SYNC("RowConv"); } - -template -__global__ void KeRowConvBwWeight(real* dw, const real* x, const real* dy, - const int* starts, const int height, const int width, const int numSeq, - const int context) { - +template +__global__ void KeRowConvBwWeight(real* dw, + const real* x, + const real* dy, + const int* starts, + const int height, + const int width, + const int numSeq, + const int context) { const int tidx = threadIdx.x; const int tidy = threadIdx.y; const int blky = blockDim.y; @@ -138,21 +148,21 @@ __global__ void KeRowConvBwWeight(real* dw, const real* x, const real* dy, const int start = starts[i]; const int end = starts[i + 1]; const int steps = end - start; - const int size = ((steps + BLOCK_H - 1)/BLOCK_H) * BLOCK_H; + const int size = ((steps + BLOCK_H - 1) / BLOCK_H) * BLOCK_H; for (int j = tidy; j < size; j += BLOCK_H) { int xoff = gidx + tidx; int yoff = start + j; // transpose - sh_x[tidx][tidy] = (xoff < width && yoff < end) ? - x[yoff * width + xoff] : 0.0; - sh_dy[tidx][tidy + context - 1] = (xoff < width && yoff < end) ? - dy[yoff * width + xoff] : 0.0; + sh_x[tidx][tidy] = + (xoff < width && yoff < end) ? x[yoff * width + xoff] : 0.0; + sh_dy[tidx][tidy + context - 1] = + (xoff < width && yoff < end) ? dy[yoff * width + xoff] : 0.0; __syncthreads(); if (tidy < (context - 1)) { yoff = yoff - context + 1; - sh_dy[tidx][tidy] = (xoff < width && yoff >= start) ? - dy[yoff * width + xoff] : 0.0; + sh_dy[tidx][tidy] = + (xoff < width && yoff >= start) ? dy[yoff * width + xoff] : 0.0; } __syncthreads(); @@ -179,11 +189,15 @@ __global__ void KeRowConvBwWeight(real* dw, const real* x, const real* dy, } } -template -__global__ void KeRowConvBwWeight2(real* dw, const real* x, const real* dy, - const int* starts, const int height, const int width, const int numSeq, - const int context) { - +template +__global__ void KeRowConvBwWeight2(real* dw, + const real* x, + const real* dy, + const int* starts, + const int height, + const int width, + const int numSeq, + const int context) { const int tidx = threadIdx.x; const int tidy = threadIdx.y; const int gidx = blockIdx.x * blockDim.x; @@ -196,19 +210,21 @@ __global__ void KeRowConvBwWeight2(real* dw, const real* x, const real* dy, const int end = starts[i + 1]; const int steps = end - start; - const int size = ((steps + BLOCK_H - 1)/BLOCK_H) * BLOCK_H; + const int size = ((steps + BLOCK_H - 1) / BLOCK_H) * BLOCK_H; for (int j = tidy; j < size; j += BLOCK_H) { int xoff = gidx + tidx; int yoff = start + j; // transpose - sh_x[tidx][tidy] = (xoff < width && yoff < end) ? - x[yoff * width + xoff] : 0.0; + sh_x[tidx][tidy] = + (xoff < width && yoff < end) ? x[yoff * width + xoff] : 0.0; __syncthreads(); for (int t = 0; t < context; t++) { - sh_dy[tidx][tidy] = (xoff < width && (yoff - t) >= start && - yoff - t < end) ? dy[(yoff - t) * width + xoff] : 0.0; + sh_dy[tidx][tidy] = + (xoff < width && (yoff - t) >= start && yoff - t < end) + ? dy[(yoff - t) * width + xoff] + : 0.0; __syncthreads(); real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx]; @@ -222,18 +238,22 @@ __global__ void KeRowConvBwWeight2(real* dw, const real* x, const real* dy, __syncthreads(); if (tidx == 0 && (gidx + tidy) < width) { - dw[t*width + gidx + tidy] += val; + dw[t * width + gidx + tidy] += val; } } } } } -template -__global__ void KeRowConvBwData(real* dx, const real* w, const real* dy, - const int* starts, const int height, const int width, const int numSeq, - const int context) { - +template +__global__ void KeRowConvBwData(real* dx, + const real* w, + const real* dy, + const int* starts, + const int height, + const int width, + const int numSeq, + const int context) { const int tidx = threadIdx.x; const int tidy = threadIdx.y; const int blky = blockDim.y; @@ -242,7 +262,7 @@ __global__ void KeRowConvBwData(real* dx, const real* w, const real* dy, __shared__ real sw[BLOCK_H][BLOCK_W]; for (int i = tidy; i < context; i += blky) { - sw[i][tidx] = gidx + tidx < width ? w[i*width + gidx + tidx] : 0.0; + sw[i][tidx] = gidx + tidx < width ? w[i * width + gidx + tidx] : 0.0; } __syncthreads(); @@ -266,10 +286,14 @@ __global__ void KeRowConvBwData(real* dx, const real* w, const real* dy, } } -__global__ void KeRowConvBwData2(real* dx, const real* w, const real* dy, - const int* starts, const int height, const int width, const int numSeq, - const int context) { - +__global__ void KeRowConvBwData2(real* dx, + const real* w, + const real* dy, + const int* starts, + const int height, + const int width, + const int numSeq, + const int context) { const int tidx = threadIdx.x; const int tidy = threadIdx.y; const int blky = blockDim.y; @@ -295,14 +319,13 @@ __global__ void KeRowConvBwData2(real* dx, const real* w, const real* dy, } } - template <> void RowConvGrad(const GpuMatrix& outG, - const GpuMatrix& in, - const GpuMatrix& filter, - GpuMatrix& inG, - GpuMatrix& filterG, - const GpuIVector& seq) { + const GpuMatrix& in, + const GpuMatrix& filter, + GpuMatrix& inG, + GpuMatrix& filterG, + const GpuIVector& seq) { const size_t numSeq = seq.getSize() - 1; const size_t contextLength = filter.getHeight(); const size_t height = in.getHeight(); @@ -318,13 +341,11 @@ void RowConvGrad(const GpuMatrix& outG, dim3 dimGrid(DIVUP(width, dimBlock.x), 1); real* dw = filterG.getData(); if (contextLength <= 32) { - KeRowConvBwWeight<32, 32, 32> - <<>> - (dw, x, dy, starts, height, width, numSeq, contextLength); + KeRowConvBwWeight<32, 32, 32><<>>( + dw, x, dy, starts, height, width, numSeq, contextLength); } else { - KeRowConvBwWeight2<32, 32> - <<>> - (dw, x, dy, starts, height, width, numSeq, contextLength); + KeRowConvBwWeight2<32, 32><<>>( + dw, x, dy, starts, height, width, numSeq, contextLength); } } @@ -333,13 +354,11 @@ void RowConvGrad(const GpuMatrix& outG, dim3 dimBlock2(32, 32); dim3 dimGrid2(DIVUP(width, dimBlock2.x), 1); if (contextLength <= 64) { - KeRowConvBwData<32, 64> - <<>> - (dx, w, dy, starts, height, width, numSeq, contextLength); + KeRowConvBwData<32, 64><<>>( + dx, w, dy, starts, height, width, numSeq, contextLength); } else { - KeRowConvBwData2 - <<>> - (dx, w, dy, starts, height, width, numSeq, contextLength); + KeRowConvBwData2<<>>( + dx, w, dy, starts, height, width, numSeq, contextLength); } } diff --git a/paddle/function/TensorShapeTest.cpp b/paddle/function/TensorShapeTest.cpp index 45a2e106e7fc3f0e9e57cf8c2bb549d747f4f49b..e55d516d4ac8e86b05911ac1a4e7095f53e2ea3c 100644 --- a/paddle/function/TensorShapeTest.cpp +++ b/paddle/function/TensorShapeTest.cpp @@ -19,35 +19,35 @@ namespace paddle { TEST(TensorShape, Constructor) { TensorShape t1; - EXPECT_EQ(t1.ndims(), 0); - EXPECT_EQ(t1.getElements(), 0); + EXPECT_EQ(t1.ndims(), 0U); + EXPECT_EQ(t1.getElements(), 0U); TensorShape t2(3); - EXPECT_EQ(t2.ndims(), 3); - EXPECT_EQ(t2.getElements(), 1); + EXPECT_EQ(t2.ndims(), 3U); + EXPECT_EQ(t2.getElements(), 1U); TensorShape t3({8, 10}); - EXPECT_EQ(t3.ndims(), 2); - EXPECT_EQ(t3.getElements(), 80); + EXPECT_EQ(t3.ndims(), 2U); + EXPECT_EQ(t3.getElements(), 80U); TensorShape t4(t3); EXPECT_EQ(t4.ndims(), t3.ndims()); EXPECT_EQ(t4.getElements(), t3.getElements()); TensorShape t5({1, 2, 3, 4, 5}); - EXPECT_EQ(t5.ndims(), 5); - EXPECT_EQ(t5.getElements(), 120); + EXPECT_EQ(t5.ndims(), 5U); + EXPECT_EQ(t5.getElements(), 120U); } TEST(TensorShape, GetAndSet) { TensorShape t({1, 2, 3}); - EXPECT_EQ(t.ndims(), 3); - EXPECT_EQ(t.getElements(), 6); + EXPECT_EQ(t.ndims(), 3U); + EXPECT_EQ(t.getElements(), 6U); - EXPECT_EQ(t[1], 2); + EXPECT_EQ(t[1], 2U); t.setDim(1, 100); - EXPECT_EQ(t.getElements(), 300); - EXPECT_EQ(t[1], 100); + EXPECT_EQ(t.getElements(), 300U); + EXPECT_EQ(t[1], 100U); } } // namespace paddle diff --git a/paddle/function/TensorTypeTest.cpp b/paddle/function/TensorTypeTest.cpp index e50e46f3e99111731d9587f3e4ddfd4b26ae27e9..d1c559a91e294853fa6e19f9115bc008ae56915c 100644 --- a/paddle/function/TensorTypeTest.cpp +++ b/paddle/function/TensorTypeTest.cpp @@ -19,9 +19,9 @@ namespace paddle { TEST(TensorType, Matrix) { Tensor::Matrix matrix(100, 200); - EXPECT_EQ(matrix.getHeight(), 100); - EXPECT_EQ(matrix.getWidth(), 200); - EXPECT_EQ(matrix.getElementCnt(), 100 * 200); + EXPECT_EQ(matrix.getHeight(), 100U); + EXPECT_EQ(matrix.getWidth(), 200U); + EXPECT_EQ(matrix.getElementCnt(), 100U * 200U); EXPECT_EQ(matrix.useGpu(), false); Tensor::Matrix testGpu(100, 200); @@ -33,15 +33,15 @@ TEST(TensorType, Vector) { Tensor::Vector gpuVector(100); EXPECT_EQ(cpuVector.useGpu(), false); EXPECT_EQ(gpuVector.useGpu(), true); - EXPECT_EQ(cpuVector.getSize(), 100); - EXPECT_EQ(gpuVector.getSize(), 100); + EXPECT_EQ(cpuVector.getSize(), 100U); + EXPECT_EQ(gpuVector.getSize(), 100U); Tensor::Vector cpuIVector(100); Tensor::Vector gpuIVector(100); EXPECT_EQ(cpuIVector.useGpu(), false); EXPECT_EQ(gpuIVector.useGpu(), true); - EXPECT_EQ(cpuIVector.getSize(), 100); - EXPECT_EQ(gpuIVector.getSize(), 100); + EXPECT_EQ(cpuIVector.getSize(), 100U); + EXPECT_EQ(gpuIVector.getSize(), 100U); } TEST(TensorType, EmptyMatrix) { diff --git a/paddle/function/nnpack/NNPACKConvOp.cpp b/paddle/function/nnpack/NNPACKConvOp.cpp index f0ec77a5d00333993427fb8d0bc938c884e50c95..6ccc487cf1c26b181b025cc62c93807c8a2848ef 100644 --- a/paddle/function/nnpack/NNPACKConvOp.cpp +++ b/paddle/function/nnpack/NNPACKConvOp.cpp @@ -49,9 +49,7 @@ class NNPACKConvFunction : public ConvFunctionBase { public: void init(const FuncConfig& config) override { ConvFunctionBase::init(config); - CHECK_EQ(groups_, (size_t)1); algorithm_ = get_nnp_convolution_algorithm(config.get("algo")); - // algorithm_ = nnp_convolution_algorithm_auto; transform_strategy_ = nnp_convolution_transform_strategy_compute; nnp_status status = nnp_initialize(); CHECK_EQ(status, nnp_status_success); @@ -67,8 +65,7 @@ public: } } - virtual void check(const BufferArgs& inputs, - const BufferArgs& outputs) override { + 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(); @@ -91,8 +88,8 @@ public: size_t filterHeight = getFilterHeight(filter); size_t filterWidth = getFilterWidth(filter); size_t outputChannels = output[1]; - // size_t outputHeight = output[2]; - // size_t outputWidth = output[3]; + size_t outputHeight = output[2]; + size_t outputWidth = output[3]; nnp_size inputSize = {.width = inputWidth, .height = inputHeight}; nnp_padding padding = {.top = (size_t)paddingH(), @@ -171,31 +168,40 @@ public: } } + size_t inputOffset = inputChannels / groups_ * inputHeight * inputWidth; + size_t outputOffset = outputChannels / groups_ * outputHeight * outputWidth; + size_t filterOffset = filter.getElements() / groups_; + if (batchSize == 1) { - nnp_status status = - nnp_convolution_inference(algorithm_, - transform_strategy_, - inputChannels, - outputChannels, - inputSize, - padding, - kernelSize, - outputSubsampling, - inputData, - filterData, - nullptr, /* bias */ - outputData, - bufferPtr, - sizePtr, - nnp_activation_identity, - nullptr, - threadpool_, /* threadpool */ - nullptr); - CHECK_EQ(status, nnp_status_success); + for (size_t g = 0; g < groups_; g++) { + nnp_status status = + nnp_convolution_inference(algorithm_, + transform_strategy_, + inputChannels / groups_, + outputChannels / groups_, + inputSize, + padding, + kernelSize, + outputSubsampling, + inputData + inputOffset * g, + filterData + filterOffset * g, + nullptr, /* bias */ + outputData + outputOffset * g, + bufferPtr, + sizePtr, + nnp_activation_identity, + nullptr, + threadpool_, /* threadpool */ + nullptr); + CHECK_EQ(status, nnp_status_success); + } } else { // only supports stride = 1 CHECK_EQ(strideH(), 1); CHECK_EQ(strideW(), 1); + + // TODO(hedaoyuan): There has some bug when batchSize > 1 and groups_ > 1. + CHECK_EQ(groups_, static_cast(1)); nnp_status status = nnp_convolution_output(algorithm_, batchSize, inputChannels, diff --git a/paddle/function/nnpack/NNPACKConvOpTest.cpp b/paddle/function/nnpack/NNPACKConvOpTest.cpp index 48180112111c67f36ddd425008187201655089c9..4dd3982487f3567f461ddaea8c5dc719fff04736 100644 --- a/paddle/function/nnpack/NNPACKConvOpTest.cpp +++ b/paddle/function/nnpack/NNPACKConvOpTest.cpp @@ -13,87 +13,18 @@ See the License for the specific language governing permissions and limitations under the License. */ #include -#include "paddle/function/Function.h" -#include "paddle/function/FunctionTest.h" - -DEFINE_string(algo, - "auto", - "The algorithm (auto, ft8x8, ft16x16, wt8x8, " - "implicit-gemm, or direct) for computing convolution of NNPACK."); +#include "paddle/function/ConvOpTest.h" namespace paddle { -#define IS_NNPACK_SUPPORT(algo, filterSize, stride) \ - if (algo == "direct" && filterSize != 1) continue; \ - if (algo == "direct" && batchSize != 1) continue; \ - if (algo == "wt8x8" && filterSize != 3) continue; \ - if (algo == "implicit-gemm" && batchSize != 1) continue; \ - if (algo != "auto" && algo != "implicit-gemm" && stride > 1) continue; - -class ConvolutionTest { -public: - ConvolutionTest(const std::string& conv1, - const std::string& conv2, - std::string algo = "auto") { - for (size_t batchSize : {1, 32}) { - for (size_t inputSize : {7, 14, 54}) { - for (size_t filterSize : {1, 3, 5}) { - for (size_t inputChannels : {3, 64}) { - for (size_t outputChannels : {3, 64, 128}) { - if (inputChannels < outputChannels) break; - for (size_t stride : {1, 2}) { - // if batchSize > 1 NNPACKConv only supports stride = 1 - if (batchSize > 1 && stride > 1) break; - for (size_t padding : {0, 1}) { - if (padding >= filterSize) break; - size_t outputSize = - (inputSize - filterSize + 2 * padding + stride) / stride; - IS_NNPACK_SUPPORT(algo, filterSize, stride); - LOG(INFO) << " batchSize=" << batchSize - << " inputChannels=" << inputChannels - << " inputHeight=" << inputSize - << " inputWidth=" << inputSize - << " outputChannels=" << outputChannels - << " filterHeight=" << filterSize - << " filterWidth=" << filterSize - << " outputHeight=" << outputSize - << " outputWidth=" << outputSize - << " stride=" << stride << " padding=" << padding; - - std::vector paddings = {padding, padding}; - std::vector strides = {stride, stride}; - Compare2Function test( - conv1, - conv2, - FuncConfig() - .set("paddings", paddings) - .set("strides", strides) - .set("groups", (size_t)1) - .set("algo", algo)); - - TensorShape shape0{ - batchSize, inputChannels, inputSize, inputSize}; - TensorShape shape1{ - outputChannels, inputChannels, filterSize, filterSize}; - TensorShape shape2{ - batchSize, outputChannels, outputSize, outputSize}; - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape0)); - test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape1)); - test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, shape2)); - test.run(); - } - } - } - } - } - } - } - } -}; +TEST(NNPACK, Forward) { + Convolution( + "GemmConv-CPU", "NNPACKConv-CPU", forward); +} -TEST(Convolution, NNPACK) { - // NNPACK only supports stride = 1 - ConvolutionTest test("GemmConv-CPU", "NNPACKConv-CPU", FLAGS_algo); +TEST(NNPACK, Depthwise) { + DepthwiseConvolution( + "GemmConv-CPU", "NNPACKConv-CPU", forward); } } // namespace paddle diff --git a/paddle/gserver/CMakeLists.txt b/paddle/gserver/CMakeLists.txt index 0012636b8f618a1b45cfc801c04781e67694956f..62cff9361ccba3ae3b9359ddb932f5b26146eb97 100644 --- a/paddle/gserver/CMakeLists.txt +++ b/paddle/gserver/CMakeLists.txt @@ -23,6 +23,17 @@ endmacro() filter_test(GSERVER_HEADER) filter_test(GSERVER_SOURCES) + +if(NOT WITH_MKLDNN) + file(GLOB_RECURSE DNN_HEADER RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "MKLDNN*.h") + file(GLOB_RECURSE DNN_SOURCES RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "MKLDNN*.cpp") + list(REMOVE_ITEM GSERVER_HEADER ${DNN_HEADER}) + list(REMOVE_ITEM GSERVER_SOURCES ${DNN_SOURCES}) + message(STATUS "Skip compiling with MKLDNNLayers and MKLDNNActivations") +else() + message(STATUS "Compile with MKLDNNLayers and MKLDNNActivations") +endif() + if(NOT WITH_GPU) list(REMOVE_ITEM GSERVER_HEADER layers/CudnnConvBaseLayer.h diff --git a/paddle/gserver/activations/ActivationFunction.cpp b/paddle/gserver/activations/ActivationFunction.cpp index 81cc3c890b6d4ad048e4edc03208c85778244078..78e958e06fac84fa956abc9faea60157bf6132eb 100644 --- a/paddle/gserver/activations/ActivationFunction.cpp +++ b/paddle/gserver/activations/ActivationFunction.cpp @@ -112,7 +112,6 @@ BEGIN_DEFINE_ACTIVATION(softmax) private: MatrixPtr sftMaxSum_; MatrixPtr sftMaxDot_; -MatrixPtr one_; public: Error __must_check forward(Argument& act) { @@ -138,14 +137,6 @@ Error __must_check backward(Argument& act) { 1, /* trans */ false, useGpu(act.deviceId)); - if (!one_ || one_->getWidth() != outputG->getWidth()) { - Matrix::resizeOrCreate(one_, - 1, - outputG->getWidth(), - /* trans */ false, - useGpu(act.deviceId)); - one_->one(); - } sftMaxDot_->dotMul(*outputG, *outputV); sftMaxSum_->colMerge(*sftMaxDot_); @@ -186,7 +177,10 @@ Error __must_check forward(Argument& act) { useGpu(act.deviceId)); } - auto starts = act.sequenceStartPositions->getVector(useGpu(act.deviceId)); + auto starts = + act.hasSubseq() + ? act.subSequenceStartPositions->getVector(useGpu(act.deviceId)) + : act.sequenceStartPositions->getVector(useGpu(act.deviceId)); act.value->sequenceSoftmax(*act.value, *starts); return Error(); } @@ -197,8 +191,9 @@ Error __must_check backward(Argument& act) { "Input width for each timestep of sequence softmax should be 1"); } - size_t numSequences = act.getNumSequences(); - const int* starts = act.sequenceStartPositions->getData(false); + size_t numSequences = + act.hasSubseq() ? act.getNumSubSequences() : act.getNumSequences(); + const int* starts = act.getCpuStartPositions(); for (size_t i = 0; i < numSequences; ++i) { // TODO(Dangqingqing) optimization for GPU diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp index 9ddd449de7500f5682d59469328f06971c6e83bf..f98bf95064fa539b990309dfe0bff10c1e99d096 100644 --- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp +++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp @@ -967,8 +967,9 @@ void RecurrentGradientMachine::generateSequence() { size_t numSequences = getGenBatchSize(); resizeBootFrame(numSequences); - // We create only two sub-network in generation for alternate use. - // Thus, we can reduce total memory of output_ in layer forward. + // We create only two sub-network in generation, one stores states of all + // layers in previous time step and the other storing the states at current + // time step. resizeOrCreateFrames(2); // outFrameLines_.size() > 1UL @@ -1001,10 +1002,9 @@ void RecurrentGradientMachine::generateSequence() { // init outArg size_t resultNum = generator_.config.num_results_per_sample(); - IVector::resizeOrCreate( - generator_.outArg.ids, - generator_.config.max_num_frames() * numSequences * resultNum, - false); + size_t maxGenWordCount = + generator_.config.max_num_frames() * numSequences * resultNum; + IVector::resizeOrCreate(generator_.outArg.ids, maxGenWordCount, false); if (resultNum > 1) { CHECK_LE(resultNum, static_cast(generator_.config.beam_size())); Matrix::resizeOrCreate(generator_.outArg.in, @@ -1012,6 +1012,11 @@ 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, @@ -1313,13 +1318,20 @@ void RecurrentGradientMachine::fillGenOutputs() { starts[0] = 0; if (numResults > 1) { real* probs = generator_.outArg.in->getData(); + real* idsProb = generator_.outArg.value->getData(); + size_t curPos = 0; for (size_t i = 0; i < finalPaths_.size(); ++i) { for (size_t j = 0; j < finalPaths_[i].size(); ++j) { Path& path = finalPaths_[i][j]; - generator_.ids.push_back(path.ids.size()); // sequence size + size_t genLen = path.ids.size(); + generator_.ids.push_back(genLen); // sequence size generator_.ids.insert( generator_.ids.end(), path.ids.begin(), path.ids.end()); generator_.ids.push_back(-1); // end of sequence + + memcpy(idsProb + curPos, path.idsProb.data(), sizeof(real) * genLen); + curPos += genLen; + idsProb[curPos++] = -1.0; probs[i * numResults + j] = path.logProb; if (!j && dataArgsSize_) { diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h index f245620cf668bb341df99cf498105cbd996a6b24..fb3fc5877ac96323e891f800db80af83b6809831 100644 --- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h +++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h @@ -189,6 +189,11 @@ public: */ std::vector ids; + /** + * @brief idsProb, log probability of each generated words. + */ + std::vector idsProb; + /** * @brief logProb, current probability of path. */ @@ -228,11 +233,13 @@ public: */ Path(Path& old, int newId, real logProb, int machineId, int topIndex) : ids(old.ids), + idsProb(old.idsProb), logProb(old.logProb + logProb), machineId(machineId), topIndex(topIndex), seqId(old.seqId) { ids.push_back(newId); + idsProb.push_back(logProb); if (!old.probHistory.empty()) { this->probHistory = old.probHistory; // probHistory store current prob, not sum @@ -411,8 +418,9 @@ protected: struct Generator { GeneratorConfig config; - std::vector ids; // store generated sequences - Argument outArg; // final output argument + std::vector ids; // store generated sequences + std::vector idsProb; // log probability of each generated word + Argument outArg; // final output argument }; bool generating_; Generator generator_; diff --git a/paddle/gserver/layers/CudnnBatchNormLayer.cpp b/paddle/gserver/layers/CudnnBatchNormLayer.cpp index 09dac05a7ad7a80bd6b9e12e8f7f060310d516c8..44ba2c4b7d1562d2ce839b5f4b4de1af35e6925f 100644 --- a/paddle/gserver/layers/CudnnBatchNormLayer.cpp +++ b/paddle/gserver/layers/CudnnBatchNormLayer.cpp @@ -14,6 +14,7 @@ limitations under the License. */ #include "CudnnBatchNormLayer.h" #include "Layer.h" +#include "paddle/cuda/include/hl_batch_norm.h" #include "paddle/utils/Stat.h" namespace paddle { @@ -79,16 +80,33 @@ void CudnnBatchNormLayer::forward(PassType passType) { savedInvVar); } else { // used movingMean and movingVar in testing - hl_batch_norm_forward_inference(ioDesc_, - input, - ioDesc_, - output, - bnParamDesc_, - gamma, - beta, - movingMean, - movingVar, - EPS); + if (batchSize <= 1024) { + hl_batch_norm_forward_inference(ioDesc_, + input, + ioDesc_, + output, + bnParamDesc_, + gamma, + beta, + movingMean, + movingVar, + EPS); + } else { + // There is a limitation in cudnn library. + // When the batch size is larger than 1024 in cuDNN v5.1, + // the cudnnBatchNormalizationForwardInference will fail. + hl_batch_norm_cuda_inference(input, + output, + gamma, + beta, + movingMean, + movingVar, + EPS, + batchSize, + channels_, + imageH_, + imageW_); + } } /* activation */ { diff --git a/paddle/gserver/layers/ExpandConvLayer.cpp b/paddle/gserver/layers/ExpandConvLayer.cpp index 783e02e47cb91e28eb88b079f1e94439d34fa775..0ece2799318ea5ecc91f97f71289d4d07246dcaa 100644 --- a/paddle/gserver/layers/ExpandConvLayer.cpp +++ b/paddle/gserver/layers/ExpandConvLayer.cpp @@ -57,8 +57,7 @@ bool ExpandConvLayer::init(const LayerMap &layerMap, convGradFilterType = "GemmConvGradFilter"; } - if (FLAGS_use_nnpack) { - CHECK_EQ(isDeconv_, false); + if (FLAGS_use_nnpack && !isDeconv_) { createFunction(forward_, "NNPACKConv", FuncConfig() diff --git a/paddle/gserver/layers/GruCompute.cu b/paddle/gserver/layers/GruCompute.cu index d5e547dce347c824f959425551afea66dfd94e5a..b4f5c54b14767586cb7b7e2c86cc069e2063ccfd 100644 --- a/paddle/gserver/layers/GruCompute.cu +++ b/paddle/gserver/layers/GruCompute.cu @@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ - #include "GruCompute.h" #include "hl_recurrent_apply.cuh" @@ -31,8 +30,10 @@ void GruCompute::forward<1>(hl_gru_value value, int frameSize, int batchSize) { } template <> -void GruCompute::backward<1>(hl_gru_value value, hl_gru_grad grad, - int frameSize, int batchSize) { +void GruCompute::backward<1>(hl_gru_value value, + hl_gru_grad grad, + int frameSize, + int batchSize) { hl_gpu_gru_backward(hppl::backward::gru_stateGrad(), hppl::backward::gru_resetGrad(), value, diff --git a/paddle/gserver/layers/KmaxSeqScoreLayer.cpp b/paddle/gserver/layers/KmaxSeqScoreLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..8ce591d4762466e1ed4b2970cb9cae9203bc0a2b --- /dev/null +++ b/paddle/gserver/layers/KmaxSeqScoreLayer.cpp @@ -0,0 +1,117 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" + +namespace paddle { + +class KmaxSeqScoreLayer : public Layer { +private: + MatrixPtr scores_; + size_t beamSize_; + void kmaxScorePerSeq(const real* score, + real* sortedRes, + const ICpuGpuVectorPtr seqStartPos); + +public: + explicit KmaxSeqScoreLayer(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(kmax_seq_score, KmaxSeqScoreLayer); + +bool KmaxSeqScoreLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + bool ret = Layer::init(layerMap, parameterMap); + CHECK_EQ(1U, inputLayers_.size()); + + beamSize_ = config_.beam_size(); + CHECK_GE(beamSize_, 1U); + + setNeedSequenceInfo(false); + setNeedGradient(false); + return ret; +} + +void KmaxSeqScoreLayer::kmaxScorePerSeq(const real* scores, + real* sortedIds, + const ICpuGpuVectorPtr seqStartPos) { + int* starts = seqStartPos->getMutableData(false); + std::vector indices; + for (size_t i = 0; i < seqStartPos->getSize() - 1; ++i) { + int seqLen = starts[i + 1] - starts[i]; + int k = std::min(static_cast(beamSize_), seqLen); + + indices.resize(seqLen, 0); + std::iota(begin(indices), end(indices), 0.); + std::vector tmpScore(scores + starts[i], scores + starts[i + 1]); + std::partial_sort( + begin(indices), + begin(indices) + k, + end(indices), + [&](size_t a, size_t b) { return tmpScore[a] > tmpScore[b]; }); + memcpy(sortedIds + (i * beamSize_), indices.data(), k * sizeof(real)); + } +} + +void KmaxSeqScoreLayer::forward(PassType passType) { + Layer::forward(passType); + + const Argument& input = getInput(0); + const MatrixPtr inputScore = getInputValue(0); + + CHECK(input.hasSeq() || input.hasSubseq()) + << "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."; + + 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. + Matrix::resizeOrCreate(scores_, + inputScore->getHeight(), + 1, + false /* trans */, + false /* useGpu */); + scores_->copyFrom(*inputScore); + } else { + scores_ = inputScore; + } + + Matrix::resizeOrCreate( + output_.value, + input.hasSubseq() ? input.getNumSubSequences() : input.getNumSequences(), + beamSize_, + false, + false); + output_.value->one(); + output_.value->mulScalar(-1.); + + kmaxScorePerSeq(scores_->getData(), + output_.value->getData(), + input.hasSubseq() ? input.subSequenceStartPositions + : input.sequenceStartPositions); +} + +void KmaxSeqScoreLayer::backward(const UpdateCallback& callback) {} + +} // namespace paddle diff --git a/paddle/gserver/layers/LstmCompute.cu b/paddle/gserver/layers/LstmCompute.cu index f75c0c40ccc833e35f8fe8f21c12b3d3f68d5eb6..d3f59b52a4b3163f47a969d9a08ecd139a099e33 100644 --- a/paddle/gserver/layers/LstmCompute.cu +++ b/paddle/gserver/layers/LstmCompute.cu @@ -12,41 +12,62 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ - #include "LstmCompute.h" #include "hl_recurrent_apply.cuh" namespace paddle { template <> -void LstmCompute::forwardBatch<1>(hl_lstm_value value, int frameSize, - int batchSize) { - hl_gpu_lstm_forward(hppl::forward::lstm(), value, frameSize, - batchSize, activeNode_, activeGate_, +void LstmCompute::forwardBatch<1>(hl_lstm_value value, + int frameSize, + int batchSize) { + hl_gpu_lstm_forward(hppl::forward::lstm(), + value, + frameSize, + batchSize, + activeNode_, + activeGate_, activeState_); } template <> -void LstmCompute::backwardBatch<1>(hl_lstm_value value, hl_lstm_grad grad, - int frameSize, int batchSize) { - hl_gpu_lstm_backward(hppl::backward::lstm(), value, grad, - frameSize, batchSize, activeNode_, - activeGate_, activeState_); +void LstmCompute::backwardBatch<1>(hl_lstm_value value, + hl_lstm_grad grad, + int frameSize, + int batchSize) { + hl_gpu_lstm_backward(hppl::backward::lstm(), + value, + grad, + frameSize, + batchSize, + activeNode_, + activeGate_, + activeState_); } template <> void LstmCompute::forwardOneSequence<1>(hl_lstm_value value, int frameSize) { - hl_gpu_lstm_forward(hppl::forward::lstm(), value, - frameSize, /* batchSize */ 1, - activeNode_, activeGate_, activeState_); + hl_gpu_lstm_forward(hppl::forward::lstm(), + value, + frameSize, + /* batchSize */ 1, + activeNode_, + activeGate_, + activeState_); } template <> -void LstmCompute::backwardOneSequence<1>(hl_lstm_value value, hl_lstm_grad grad, +void LstmCompute::backwardOneSequence<1>(hl_lstm_value value, + hl_lstm_grad grad, int frameSize) { - hl_gpu_lstm_backward(hppl::backward::lstm(), value, grad, - frameSize, /* batchSize */ 1, - activeNode_, activeGate_, activeState_); + hl_gpu_lstm_backward(hppl::backward::lstm(), + value, + grad, + frameSize, + /* batchSize */ 1, + activeNode_, + activeGate_, + activeState_); } } // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNBase.h b/paddle/gserver/layers/MKLDNNBase.h new file mode 100644 index 0000000000000000000000000000000000000000..4c0234e7b3a91053596c32cea581fa5d1e26b9d5 --- /dev/null +++ b/paddle/gserver/layers/MKLDNNBase.h @@ -0,0 +1,97 @@ +/* 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 "mkldnn.hpp" + +namespace paddle { + +typedef enum { + MKLDNN_BASE = 1, // basical info of MKLDNN + MKLDNN_TESTS = 1, // gtest info of MKLDNN + MKLDNN_SIZES = 2, // size info of MKLDNN + MKLDNN_FMTS = 3, // format info of MKLDNN + MKLDNN_ALL = 4, // show all info of MKLDNN +} MKLDNN_LOG_LEVEL; + +/** + * @brief MKLDNN CPU engine. + * + */ +class CPUEngine { +public: + static CPUEngine& Instance() { + // Thread-safe in C++11. + static CPUEngine myInstance; + return myInstance; + } + + // Disallow copy or move + CPUEngine(const CPUEngine&) = delete; // Copy constructor + CPUEngine(CPUEngine&&) = delete; // Move constructor + CPUEngine& operator=(const CPUEngine&) = delete; // Copy assignment + CPUEngine& operator=(CPUEngine&&) = delete; // Move assignment + + mkldnn::engine& getEngine() { return cpuEngine_; } + +protected: + CPUEngine() : cpuEngine_(mkldnn::engine::cpu, 0) {} + // CPUEngine() : cpuEngine_(mkldnn::engine::cpu_lazy, 0) {} + ~CPUEngine() {} + +private: + mkldnn::engine cpuEngine_; +}; + +/** + * @brief MKLDNN Stream. + * + */ +class MKLDNNStream { +public: + MKLDNNStream() : ready_(false) { resetState(); } + + virtual ~MKLDNNStream() {} + + /** + * @brief Submit stream + * @param prims The primitives vector + * @param block Waiting for the stream to complete + */ + void submit(std::vector& prims, bool block = true) { + resetState(); + stream_->submit(prims).wait(block); + ready_ = false; + } + + /** + * @brief Reset the mkldnn stream + */ + void resetState() { + if (ready_) { + return; + } + // TODO(TJ): change me when mkldnn have method to reset this state + // stream_.reset(new mkldnn::stream(mkldnn::stream::kind::lazy)); + stream_.reset(new mkldnn::stream(mkldnn::stream::kind::eager)); + ready_ = true; + } + +private: + bool ready_; + std::shared_ptr stream_; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..30f567eaf8248a8fba1b461a2bdbf2aab13f9e08 --- /dev/null +++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp @@ -0,0 +1,282 @@ +/* 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 "MKLDNNFcLayer.h" +#include "paddle/utils/Logging.h" +#include "paddle/utils/Stat.h" + +using namespace mkldnn; // NOLINT +typedef memory::format format; +typedef inner_product_forward fc_fwd; +typedef inner_product_backward_weights fc_bwdWgt; +typedef inner_product_backward_data fc_bwdData; + +namespace paddle { + +REGISTER_LAYER(mkldnn_fc, MKLDNNFcLayer); + +bool MKLDNNFcLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + if (!MKLDNNLayer::init(layerMap, parameterMap)) { + return false; + } + + CHECK_EQ(inputLayers_.size(), 1) << "Only support one input layer yet"; + CHECK_EQ(inputLayers_.size(), parameters_.size()); + CHECK(!parameters_[0]->isSparse()) << "Do not support sparse yet"; + + // output size, cat not be changed + oc_ = getSize(); + oh_ = 1; + ow_ = 1; + + // input size can not change in FC + iLayerSize_ = inputLayers_[0]->getSize(); + CHECK_EQ(parameters_[0]->getSize(), iLayerSize_ * oc_); + + // create weight + weight_ = + std::unique_ptr(new Weight(oc_, iLayerSize_, parameters_[0], 0)); + + // create biases + if (biasParameter_.get() != NULL) { + biases_ = std::unique_ptr(new Weight(1, oc_, biasParameter_)); + } + return true; +} + +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); + 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); +} + +void MKLDNNFcLayer::reshape() { + const Argument& input = getInput(0); + int batchSize = input.getBatchSize(); + if (bs_ == batchSize) { + return; + } + bs_ = batchSize; + ih_ = input.getFrameHeight(); + iw_ = input.getFrameWidth(); + if (ih_ == 0) { + ih_ = 1; + } + 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"; + CHECK_EQ(size_t(oc_), getSize()); + printSizeInfo(); + + // reset output + output_.setFrameHeight(oh_); + output_.setFrameWidth(ow_); + resetOutput(bs_, oc_); + + // reset mkldnn forward + resetFwd(); + needResetBwd_ = true; + + convertWeightsFromPaddle(); +} + +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)); + + 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::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_); + + if (bData != NULL) { + biasVal_.reset(new memory(memory::primitive_desc(bMD, engine_), bData)); + fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *biasVal_, *outVal_)); + } else { + fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *outVal_)); + } + pipelineFwd_.clear(); + pipelineFwd_.push_back(*fwd_); +} + +void MKLDNNFcLayer::resetBwd() { + if (!needResetBwd_) { + 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); + 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::primitive_desc bwdWgtPD = + fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD); + + if (bDiff != NULL) { + biasGrad_.reset(new memory(memory::primitive_desc(bMD, engine_), bDiff)); + bwdWgt_.reset( + new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_, *biasGrad_)); + } else { + bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_)); + } + pipelineBwd_.clear(); + pipelineBwd_.push_back(*bwdWgt_); + + /// backward data + if (iDiff == NULL) { + return; + } + fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(iMD, wMD, oMD); + 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_)); + pipelineBwd_.push_back(*bwdData_); +} + +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); + + // just submit forward pipeline + stream_->submit(pipelineFwd_); + } + + /* activation */ { + REGISTER_TIMER_INFO("FwActTimer", getName().c_str()); + forwardActivation(); + } +} + +void MKLDNNFcLayer::backward(const UpdateCallback& callback) { + /* Do derivation */ { + REGISTER_TIMER_INFO("BpActTimer", getName().c_str()); + backwardActivation(); + } + + { + 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_); + } + + { + REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); + weight_->getParameterPtr()->incUpdate(callback); + if (biases_ && biases_->getWGrad()) { + biases_->getParameterPtr()->incUpdate(callback); + } + } +} +} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNFcLayer.h b/paddle/gserver/layers/MKLDNNFcLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..7954852a23f81d36d5fb0ae6a19768f419886fb1 --- /dev/null +++ b/paddle/gserver/layers/MKLDNNFcLayer.h @@ -0,0 +1,80 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "MKLDNNLayer.h" +#include "mkldnn.hpp" + +namespace paddle { + +/** + * @brief A subclass of MKLDNNLayer fc layer. + * + * The config file api is mkldnn_fc + */ +class MKLDNNFcLayer : public MKLDNNLayer { +protected: + // input layer size, can not be change after init + size_t iLayerSize_; // == ic * ih * iw + + // 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) {} + + ~MKLDNNFcLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void convertWeightsFromPaddle() override; + + void convertWeightsToPaddle() override; + + void forward(PassType passType) override; + + void backward(const UpdateCallback& callback) override; + +protected: + /** + * reshape the input image sizes + * and reset output buffer size + * and reset mkldnn forward + */ + void reshape(); + + /** + * reset the forward primitve and memory + * only would be called when input size changes + */ + void resetFwd(); + + /** + * reset the backward primitve and memory for mkldnn fc + * only would be called when needed + */ + void resetBwd(); +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..63e29f447eede5ff9df8715bc9140b64ab7f7d17 --- /dev/null +++ b/paddle/gserver/layers/MKLDNNLayer.h @@ -0,0 +1,132 @@ +/* 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 "Layer.h" +#include "MKLDNNBase.h" +#include "mkldnn.hpp" + +DECLARE_bool(use_mkldnn); +DECLARE_bool(use_mkldnn_wgt); + +namespace paddle { + +class MKLDNNLayer; +typedef std::shared_ptr MKLDNNLayerPtr; + +/** + * @brief Base class of MKLDNNlayer. + * + */ +class MKLDNNLayer : public Layer { +protected: + // batch size + int bs_; + // input image channel, height and width + int ic_, ih_, iw_; + // output image channel, height and width + int oc_, oh_, ow_; + + // backward also need reset after reset forward handle + bool needResetBwd_; + + // mkldnn engine, stream and primivtives + mkldnn::engine engine_; + std::shared_ptr stream_; + std::shared_ptr fwd_; + std::shared_ptr bwdWgt_; + std::shared_ptr bwdData_; + 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_; + +public: + explicit MKLDNNLayer(const LayerConfig& config) + : Layer(config), + bs_(0), + ic_(0), + ih_(0), + iw_(0), + oc_(0), + oh_(0), + ow_(0), + needResetBwd_(true), + engine_(mkldnn::engine::cpu, 0), + stream_(nullptr), + fwd_(nullptr), + bwdWgt_(nullptr), + bwdData_(nullptr) {} + + ~MKLDNNLayer() {} + + virtual bool init(const LayerMap& layerMap, + const ParameterMap& 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; + } + + /** + * convert weight from paddle format to mkldnn format + * weight_ will be override + */ + virtual void convertWeightsFromPaddle() {} + + /** + * convert mkldnn weight to paddle format + * weight_ will be override + */ + virtual void convertWeightsToPaddle() {} + + /** + * print info about sizes + */ + virtual void printSizeInfo() { + VLOG(MKLDNN_SIZES) << getName() << ": bs: " << bs_ << ", ic: " << ic_ + << ", ih: " << ih_ << ", iw: " << iw_ << ", oc: " << oc_ + << ", 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); + } +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/PrintLayer.cpp b/paddle/gserver/layers/PrintLayer.cpp index a97fa6bf78fce27a4e0cf329bf3309ba4a439965..0a1e17b9aa57b373f0df6e079341729539f4e193 100644 --- a/paddle/gserver/layers/PrintLayer.cpp +++ b/paddle/gserver/layers/PrintLayer.cpp @@ -29,7 +29,7 @@ public: vals.push_back(s.str()); } size_t pos = 0; - int i = 0; + size_t i = 0; std::ostringstream s; const std::string& format = config_.user_arg(); while (true) { diff --git a/paddle/gserver/layers/SubNestedSequenceLayer.cpp b/paddle/gserver/layers/SubNestedSequenceLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..648d3908f391450f276d8a900ebb3bccb8d5532c --- /dev/null +++ b/paddle/gserver/layers/SubNestedSequenceLayer.cpp @@ -0,0 +1,176 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES 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 SubNestedSequenceLayer : public Layer { +public: + explicit SubNestedSequenceLayer(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: + /* + * This functions generates the indices of rows in a batch according to the + * indices of selected sub-sequence in each sequence. + * + * Examples: + * selectedIndices: + * [ + * [0, 1, -1], + * [0, 1, 2], + * [0, -1, -1], + * [0, 2, 3], + * ] + * inputSeqInfo: + * [ + * [0,3,4], + * [4,5,7,10,15], + * [15,20], + * [20,22,23,25,28] + * ] + * + * 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] + */ + + void calSelectedCols(const MatrixPtr selectedIndices, + const std::vector>& inputSeqInfo); + + // 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. + std::vector> inputSeqInfoVec_; + + // the final selected row indices in a batch, + // rowIdx_ and selectedRows_ actually share a same memory. + IVectorPtr rowIndice_; + std::vector selectedRows_; +}; + +REGISTER_LAYER(sub_nested_seq, SubNestedSequenceLayer); + +bool SubNestedSequenceLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + CHECK_EQ(2U, inputLayers_.size()); + setNeedSequenceInfo(false); + return true; +} + +void SubNestedSequenceLayer::calSelectedCols( + const MatrixPtr selectedIndices, + const std::vector>& inputSeqInfo) { + selectedRows_.clear(); + + std::vector outSeqStartInfo(1, 0); + std::vector outSubSeqStartInfo(1, 0); + + size_t seqNum = selectedIndices->getHeight(); + size_t beamSize = selectedIndices->getWidth(); + for (size_t i = 0; i < seqNum; ++i) { + for (size_t j = 0; j < beamSize; ++j) { + if (selectedIndices->getElement(i, j) == -1.) break; + size_t selSubSeqIdx = selectedIndices->getElement(i, j); + CHECK_GT(inputSeqInfoVec_[i].size() - 1, selSubSeqIdx); + + size_t subSeqLen = inputSeqInfoVec_[i][selSubSeqIdx + 1] - + inputSeqInfoVec_[i][selSubSeqIdx]; + for (size_t k = 0; k < subSeqLen; ++k) + selectedRows_.push_back(inputSeqInfoVec_[i][selSubSeqIdx] + k); + outSubSeqStartInfo.push_back(outSubSeqStartInfo.back() + subSeqLen); + } + outSeqStartInfo.push_back(outSubSeqStartInfo.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, outSeqStartInfo.size(), false); + output_.sequenceStartPositions->copyFrom( + outSeqStartInfo.data(), outSeqStartInfo.size(), false); + + ICpuGpuVector::resizeOrCreate( + output_.subSequenceStartPositions, outSubSeqStartInfo.size(), false); + output_.subSequenceStartPositions->copyFrom( + outSubSeqStartInfo.data(), outSubSeqStartInfo.size(), false); +} + +void SubNestedSequenceLayer::forward(PassType passType) { + Layer::forward(passType); + + const Argument& inputSeq = getInput(0); + CHECK(inputSeq.hasSubseq()) << "The first input of SubNestSequence layer " + << "must be a nested sequence."; + const MatrixPtr selectedIndices = getInputValue(1); + CHECK_EQ(size_t(inputSeq.getNumSequences()), selectedIndices->getHeight()); + + if (dynamic_cast(selectedIndices.get())) { + /* + * Currently, the second input for this layer is generated by + * kmax_sequence_score_layer whose output is always stored on CPU, + * or a data_layer which canbe on GPU. + * + * If the second input is on GPU, copy it to CPU memory, because this + * input always uses very few memory, and operations related to it are + * all logic control, not computations. + */ + Matrix::resizeOrCreate(selIdsCpu_, + selectedIndices->getHeight(), + selectedIndices->getWidth(), + false /* trans */, + false /* useGpu */); + selIdsCpu_->copyFrom(*selectedIndices); + } else { + selIdsCpu_ = selectedIndices; + } + + Argument::reorganizeSeqInfo(inputSeq.sequenceStartPositions, + inputSeq.subSequenceStartPositions, + inputSeqInfoVec_); + calSelectedCols(selIdsCpu_, inputSeqInfoVec_); + + resetOutput(selectedRows_.size(), getSize()); + getOutputValue()->selectRows(*getInputValue(0), *rowIndice_); +} + +void SubNestedSequenceLayer::backward(const UpdateCallback& callback) { + MatrixPtr inputSeqGrad = getInputGrad(0); + MatrixPtr outputGrad = getOutputGrad(); + + if (inputSeqGrad) outputGrad->addToRows(*inputSeqGrad, *rowIndice_); +} + +} // namespace paddle diff --git a/paddle/gserver/tests/CMakeLists.txt b/paddle/gserver/tests/CMakeLists.txt index a43adc7ce7db937bd62ea9bf1533b8a5899c259a..c2a2993620492a9ec5dae932ff1292ced2c00064 100644 --- a/paddle/gserver/tests/CMakeLists.txt +++ b/paddle/gserver/tests/CMakeLists.txt @@ -9,7 +9,7 @@ add_unittest_without_exec(test_ProtoDataProvider # mkdir will get error. add_test(NAME test_ProtoDataProvider COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoDataProvider - WORKING_DIRECTORY ${PROJ_ROOT}/paddle) + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) ################# test_LayerGrad ####################### add_unittest_without_exec(test_LayerGrad @@ -18,6 +18,15 @@ add_unittest_without_exec(test_LayerGrad add_test(NAME test_LayerGrad COMMAND test_LayerGrad) +########## test_Mkldnn layers and activations ########## +if(WITH_MKLDNN) + add_unittest_without_exec(test_MKLDNN + test_MKLDNN.cpp + MKLDNNTester.cpp + LayerGradUtil.cpp) + add_test(NAME test_MKLDNN COMMAND test_MKLDNN) +endif() + ################ test_CRFLayerGrad #################### add_unittest_without_exec(test_CRFLayerGrad test_CRFLayerGrad.cpp @@ -50,7 +59,7 @@ add_unittest_without_exec(test_DetectionOutput test_DetectionOutput.cpp LayerGradUtil.cpp) -add_test(NAME test_DetectionOutput +add_test(NAME test_DetectionOutput COMMAND test_DetectionOutput) ################# test_ConvUnify ####################### add_unittest_without_exec(test_ConvUnify @@ -66,6 +75,16 @@ add_unittest_without_exec(test_BatchNorm add_test(NAME test_BatchNorm COMMAND test_BatchNorm) + + +################# test_KmaxSeqScore ####################### +add_unittest_without_exec(test_KmaxSeqScore + test_KmaxSeqScore.cpp + LayerGradUtil.cpp) + +add_test(NAME test_KmaxSeqScore + COMMAND test_KmaxSeqScore) + ################## test_Evaluator ####################### add_unittest(test_Evaluator test_Evaluator.cpp) @@ -82,8 +101,8 @@ if(WITH_PYTHON) test_PyDataProvider.cpp) add_test(NAME test_PyDataProvider - COMMAND .set_python_path.sh -d ./gserver/tests:${PROJ_ROOT}/python/ ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider - WORKING_DIRECTORY ${PROJ_ROOT}/paddle) + COMMAND .set_python_path.sh -d ./gserver/tests:${PADDLE_SOURCE_DIR}/python/ ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) endif() ############### test_RecurrentLayer ####################### @@ -96,7 +115,7 @@ if(NOT WITH_DOUBLE) add_test(NAME test_WarpCTCLayer COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_WarpCTCLayer --warpctc_dir=${WARPCTC_LIB_DIR} - WORKING_DIRECTORY ${PROJ_ROOT}/paddle) + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) endif() ############### test_RecurrentGradientMachine ############### @@ -106,20 +125,20 @@ add_unittest_without_exec(test_RecurrentGradientMachine test_RecurrentGradientMachine.cpp) add_test(NAME test_RecurrentGradientMachine COMMAND .set_python_path.sh -d - ${PROJ_ROOT}/python:${PROJ_ROOT}/paddle/gserver/tests + ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests ${CMAKE_CURRENT_BINARY_DIR}/test_RecurrentGradientMachine - WORKING_DIRECTORY ${PROJ_ROOT}/paddle) + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) add_unittest_without_exec(test_NetworkCompare test_NetworkCompare.cpp) if(WITH_GPU) add_test(NAME test_NetworkCompare - COMMAND .set_python_path.sh -d ${PROJ_ROOT}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=true - WORKING_DIRECTORY ${PROJ_ROOT}/paddle) + COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=true + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) else() add_test(NAME test_NetworkCompare - COMMAND .set_python_path.sh -d ${PROJ_ROOT}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=false - WORKING_DIRECTORY ${PROJ_ROOT}/paddle) + COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=false + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle) endif() @@ -127,6 +146,6 @@ add_unittest_without_exec(test_PyDataProvider2 test_PyDataProvider2.cpp) add_test(NAME test_PyDataProvider2 - COMMAND .set_python_path.sh -d ${PROJ_ROOT}/paddle/gserver/tests:${PROJ_ROOT}/python ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider2 - WORKING_DIRECTORY ${PROJ_ROOT}/paddle + COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/paddle/gserver/tests:${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider2 + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle ) diff --git a/paddle/gserver/tests/LayerGradUtil.cpp b/paddle/gserver/tests/LayerGradUtil.cpp index 9eca58f1a1baa6fb1c404a91a345bc7f9d6b4acc..a38880e14cdfcef05461dae567d198e5400c6bb1 100644 --- a/paddle/gserver/tests/LayerGradUtil.cpp +++ b/paddle/gserver/tests/LayerGradUtil.cpp @@ -388,19 +388,27 @@ 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; if (labelSeqStartPositions.size() != 0) { - CHECK(!sequenceStartPositions); CHECK_GE(static_cast(labelSeqStartPositions.size()), 2); sequenceStartPositions = @@ -410,6 +418,19 @@ void initDataLayer(TestConfig testConf, useGpu); data.sequenceStartPositions = sequenceStartPositions; } + + const std::vector& labelSubSeqStartPositions = + testConf.inputDefs[i].labelSubSeqStartPositions; + if (labelSubSeqStartPositions.size() != 0) { + CHECK_GE(static_cast(labelSubSeqStartPositions.size()), 2); + + subSequenceStartPositions = + ICpuGpuVector::create(labelSubSeqStartPositions.size(), useGpu); + subSequenceStartPositions->copyFrom(labelSubSeqStartPositions.data(), + labelSubSeqStartPositions.size(), + useGpu); + data.subSequenceStartPositions = subSequenceStartPositions; + } break; } default: diff --git a/paddle/gserver/tests/LayerGradUtil.h b/paddle/gserver/tests/LayerGradUtil.h index d299b4dd09418589514d99a72f83e1103ace7de1..88e831f78bd165f63806df6c081d84411be51502 100644 --- a/paddle/gserver/tests/LayerGradUtil.h +++ b/paddle/gserver/tests/LayerGradUtil.h @@ -67,6 +67,8 @@ struct InputDef { bool isStatic; 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) { @@ -81,8 +83,10 @@ struct InputDef { InputDef(InputType type, string nameIn, MatrixPtr selfDefinedData, - std::vector selfDefinedSeqStartPos = {}) + std::vector selfDefinedSeqStartPos = {}, + std::vector selfDefinedSubSeqStartPos = {}) : labelSeqStartPositions(selfDefinedSeqStartPos), + labelSubSeqStartPositions(selfDefinedSubSeqStartPos), selfDefinedData(selfDefinedData) { inputType = type; name = nameIn; @@ -92,6 +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 new file mode 100644 index 0000000000000000000000000000000000000000..99c8c4948c9b05ad15d1217ebb70026bbd48453f --- /dev/null +++ b/paddle/gserver/tests/MKLDNNTester.cpp @@ -0,0 +1,369 @@ +/* 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 "MKLDNNTester.h" +#include "paddle/gserver/layers/MKLDNNBase.h" +#include "paddle/gserver/layers/MKLDNNLayer.h" + +namespace paddle { + +// init data layer and test layer of both dnn and reference +void MKLDNNTester::reset(const TestConfig& dnn, + const TestConfig& ref, + size_t batchSize) { + const bool trans = false; + const bool useGpu = false; + + // clear + configs_.clear(); + layerNames_.clear(); + dataLayers_.clear(); + datas_.clear(); + layerMaps_.clear(); + parameters_.clear(); + testLayers_.clear(); + + // resize + configs_.resize(NUM); + layerNames_.resize(NUM); + dataLayers_.resize(NUM); + datas_.resize(NUM); + layerMaps_.resize(NUM); + parameters_.resize(NUM); + testLayers_.resize(NUM); + + // reset configs and layer names + configs_[DNN] = dnn; + configs_[REF] = ref; + layerNames_[DNN] = "mkldnn"; // the first is mkldnn layer + layerNames_[REF] = "reference"; // second is reference layer + + // reset others + for (size_t i = 0; i < NUM; ++i) { + configs_[i].layerConfig.set_name(layerNames_[i]); + initDataLayer(configs_[i], + &(dataLayers_[i]), + &(datas_[i]), + &(layerMaps_[i]), + layerNames_[i], + batchSize, + trans, + useGpu); + initTestLayer( + configs_[i], &(layerMaps_[i]), &(parameters_[i]), &(testLayers_[i])); + } + dnnLayer_ = testLayers_[DNN]; + refLayer_ = testLayers_[REF]; + EXPECT_EQ(dataLayers_[DNN].size(), dataLayers_[REF].size()); + EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size()); + + setInputImgSize(); +} + +void MKLDNNTester::setInputImgSize() { + for (size_t n = 0; n < dataLayers_.size(); ++n) { + for (size_t i = 0; i < dataLayers_[n].size(); ++i) { + // TODO(TJ): fix me when concat and elewise ready + dataLayers_[n][i]->getOutput().setFrameHeight(ih_); + dataLayers_[n][i]->getOutput().setFrameWidth(iw_); + } + } +} + +// init randome parameters of ref, and copy to mkldnn +void MKLDNNTester::randomWgtDatas() { + EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size()); + for (size_t i = 0; i < parameters_[REF].size(); ++i) { + const VectorPtr& dnnValue = parameters_[DNN][i]->getBuf(PARAMETER_VALUE); + const VectorPtr& refValue = parameters_[REF][i]->getBuf(PARAMETER_VALUE); + parameters_[REF][i]->randomize(); + dnnValue->copyFrom(*refValue); + + VLOG(lvl_) << "Random weight data " << parameters_[DNN][i]->getName(); + printVector(dnnValue); + } +} + +// random botdata of ref layer and copy same to mkldnn +void MKLDNNTester::randomBotDatas() { + CHECK_EQ(dataLayers_.size(), NUM); + for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) { + dataLayers_[REF][i]->getOutputValue()->randomizeUniform(); + dataLayers_[DNN][i]->getOutputValue()->copyFrom( + *(dataLayers_[REF][i]->getOutputValue())); + VLOG(lvl_) << "Input " << i << " data:"; + printMatrix(dataLayers_[REF][i]->getOutputValue()); + } +} + +void MKLDNNTester::randomTopDiffs() { + refLayer_->getOutputGrad()->randomizeUniform(); + dnnLayer_->getOutputGrad()->copyFrom(*(refLayer_->getOutputGrad())); + VLOG(lvl_) << "Random dom Backward Input, TopDiff: "; + printMatrix(refLayer_->getOutputGrad()); +} + +void MKLDNNTester::checkForward() { + printTopDatas(); + double delta = compareMatrix(testLayers_[DNN]->getOutputValue(), + testLayers_[REF]->getOutputValue()); + VLOG(MKLDNN_ALL) << "Check Forward"; + EXPECT_LE(fabs(delta), eps_); +} + +void MKLDNNTester::checkBackwardData() { + // TODO(TJ): uncomment me when batch norm ready + // const bool isBN = dnnLayer_->getType() == "mkldnn_batch_norm"; + for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) { + const MatrixPtr& dnnDiff = dataLayers_[DNN][i]->getOutputGrad(); + const MatrixPtr& refDiff = dataLayers_[REF][i]->getOutputGrad(); + VLOG(lvl_) << "Mkldnn Backward Output BotDiff " << i; + printMatrix(dnnDiff); + VLOG(lvl_) << "Reference Backward Output BotDiff " << i; + printMatrix(refDiff); + + double delta = compareMatrix(dnnDiff, refDiff); + EXPECT_LE(fabs(delta), eps_); + // TODO(TJ): uncomment me when batch norm ready + // if (isBN) { + // // the other two inputs in batch norm are for moving mean and var + // break; + // } + } +} + +void MKLDNNTester::checkBackwardWgts() { + CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size()); + vector dnnWgts; // used to temply save mkldnn weights + saveWgt(parameters_[DNN], dnnWgts); + + const MKLDNNLayerPtr dnnlayer = + std::dynamic_pointer_cast(dnnLayer_); + CHECK(dnnlayer); + dnnlayer->convertWeightsToPaddle(); + for (size_t i = 0; i < parameters_[DNN].size(); ++i) { + const VectorPtr& dnn = parameters_[DNN][i]->getBuf(PARAMETER_VALUE); + const VectorPtr& ref = parameters_[REF][i]->getBuf(PARAMETER_VALUE); + VLOG(lvl_) << "Mkldnn Output weight " << parameters_[DNN][i]->getName(); + printVector(dnn); + VLOG(lvl_) << "Reference Output weight " << parameters_[REF][i]->getName(); + printVector(ref); + + double delta = compareVector(dnn, ref); + EXPECT_LE(fabs(delta), eps_); + } + + VLOG(MKLDNN_ALL) << "Restore dnn weights before comapre"; + restoreWgt(dnnWgts, parameters_[DNN]); +} + +void MKLDNNTester::saveWgt(const vector& from, + vector& to) { + const bool useGpu = false; + to.resize(from.size()); + for (size_t i = 0; i < to.size(); ++i) { + const VectorPtr& wgt = from[i]->getBuf(PARAMETER_VALUE); + to[i] = Vector::create(wgt->getSize(), useGpu); + to[i]->copyFrom(*wgt); + } +} + +void MKLDNNTester::restoreWgt(const vector& from, + vector& to) { + CHECK_EQ(from.size(), to.size()); + for (size_t i = 0; i < from.size(); ++i) { + const VectorPtr& wgt = to[i]->getBuf(PARAMETER_VALUE); + wgt->copyFrom(*from[i]); + } +} + +// clear parameters grad +void MKLDNNTester::clearWgtDiffs() { + for (size_t n = 0; n < parameters_.size(); ++n) { + for (size_t i = 0; i < parameters_[n].size(); ++i) { + const VectorPtr& grad = parameters_[n][i]->getBuf(PARAMETER_GRADIENT); + if (grad) { + grad->zeroMem(); + } + } + } +} + +void MKLDNNTester::clearBotDiffs() { + // dnn and ref + for (size_t n = 0; n < dataLayers_.size(); ++n) { + // all inputs layers + for (size_t i = 0; i < dataLayers_[n].size(); ++i) { + dataLayers_[n][i]->getOutputGrad()->zeroMem(); + } + } +} + +void MKLDNNTester::clearBotDiffs(int n) { + CHECK_LT(n, NUM); + // all inputs layers + for (size_t i = 0; i < dataLayers_[n].size(); ++i) { + dataLayers_[n][i]->getOutputGrad()->zeroMem(); + } +} + +void MKLDNNTester::clearTopDatas() { + for (size_t i = 0; i < testLayers_.size(); ++i) { + testLayers_[i]->getOutputValue()->zeroMem(); + } +} + +void MKLDNNTester::printTopDatas() { + if (!log_) { + return; + } + + for (int n = 0; n < NUM; ++n) { + VLOG(lvl_) << testLayers_[n]->getType() << " forward output TopData: "; + printMatrix(testLayers_[n]->getOutputValue()); + } +} + +void MKLDNNTester::printMatrix(const MatrixPtr& m) { + if (!log_) { + return; + } + + std::ostringstream ostr; + m->print(ostr); + VLOG(lvl_) << std::endl << ostr.str(); +} + +void MKLDNNTester::printVector(const VectorPtr& v) { + if (!log_) { + return; + } + + std::ostringstream ostr; + v->print(ostr, v->getSize()); + VLOG(lvl_) << std::endl << ostr.str(); +} + +double MKLDNNTester::getDelta(const real* d1, + const real* d2, + size_t len, + const float failRate, + const float thres) { + double delta = 0, sum = 0; + int failCnt = 0; + const double eps = 1e-5; + double maxOut = 0; + for (size_t i = 0; i < len; ++i) { + double ref = fabs(d2[i]); + double diff = fabs(d1[i] - d2[i]); + delta += diff; + sum += ref; + if (ref > eps && fabs(d1[i]) > eps && diff / ref > thres) { + maxOut = std::max(maxOut, diff / ref); + failCnt++; + } + } + EXPECT_TRUE(std::isnormal(sum)); + EXPECT_FALSE(std::isinf(sum)); + EXPECT_FALSE(std::isnan(delta)); + VLOG(MKLDNN_ALL) << "reference avg data: " << sum / len + << ", delta: " << delta / sum << ", failCnt:" << failCnt; + return (failCnt / (float)len) > failRate ? maxOut : delta / sum; +} + +double MKLDNNTester::compareMatrix(const MatrixPtr& m1, const MatrixPtr& m2) { + CHECK_EQ(m1->getElementCnt(), m2->getElementCnt()); + return getDelta(m1->getData(), m2->getData(), m1->getElementCnt()); +} + +double MKLDNNTester::compareVector(const VectorPtr& v1, const VectorPtr& v2) { + CHECK_EQ(v1->getSize(), v2->getSize()); + return getDelta(v1->getData(), v2->getData(), v1->getSize()); +} + +void MKLDNNTester::runOnce() { + // test forward + randomBotDatas(); + dnnLayer_->forward(PASS_TRAIN); + refLayer_->forward(PASS_TRAIN); + checkForward(); + + // test backward + randomTopDiffs(); + dnnLayer_->backward(nullptr); + refLayer_->backward(nullptr); + checkBackwardData(); + checkBackwardWgts(); + + // clear buffers + // ref code will addto the diff, dnn code will writeto it + // and clearTopDatas() and clearWgtDiffs() should be coverd by test layers + clearBotDiffs(REF); +} + +void MKLDNNTester::run(const TestConfig& dnn, + const TestConfig& ref, + size_t batchSize, + size_t inputImgH, + size_t inputImgW, + size_t iter, + float epsilon, + bool log, + int level) { + VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: " << dnn.layerConfig.type() + << " vs " << ref.layerConfig.type(); + ih_ = inputImgH; + iw_ = inputImgW; + iter_ = iter; + eps_ = epsilon; + log_ = log; + lvl_ = level; + + // Firstly test FLAGS_use_mkldnn_wgt = false + FLAGS_use_mkldnn_wgt = false; + // reset and run once + reset(dnn, ref, batchSize); + randomWgtDatas(); + clearWgtDiffs(); + clearBotDiffs(); + for (size_t i = 0; i < iter_; ++i) { + VLOG(MKLDNN_TESTS) << "Check Iteration " << i; + runOnce(); + } + + // Then test FLAGS_use_mkldnn_wgt = true + FLAGS_use_mkldnn_wgt = true; + // after run once the mkldnn weight has been stored in dnnlayer + // 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 + reset(dnn, ref, batchSize); + + // restore wgt + restoreWgt(dnnWgts, parameters_[DNN]); + restoreWgt(refWgts, parameters_[REF]); + clearWgtDiffs(); + clearBotDiffs(); + + for (size_t i = 0; i < iter_; ++i) { + VLOG(MKLDNN_TESTS) << "Check Iteration " << i; + runOnce(); + } +} + +} // namespace paddle diff --git a/paddle/gserver/tests/MKLDNNTester.h b/paddle/gserver/tests/MKLDNNTester.h new file mode 100644 index 0000000000000000000000000000000000000000..522eeaf24b1949abac057a1e59e9977610be23c0 --- /dev/null +++ b/paddle/gserver/tests/MKLDNNTester.h @@ -0,0 +1,120 @@ +/* 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 +#include "LayerGradUtil.h" +#include "paddle/gserver/layers/MKLDNNBase.h" + +namespace paddle { + +/** + * @brief test the functionality of Mkldnnlayers + * refer to paddle original function + */ +class MKLDNNTester { + enum { + DNN = 0, // MKLDNN layer + REF = 1, // Reference layer + NUM = 2, // Number of total + }; + +protected: + std::vector configs_; + vector layerNames_; + vector> dataLayers_; + vector> datas_; + vector layerMaps_; + vector> parameters_; + vector testLayers_; + LayerPtr dnnLayer_, refLayer_; + + /// run some iterations, all the result should pass + size_t iter_; + /// whether to print out the details + bool log_; + /// vlog level to print the matrix details datas + int lvl_; + /// epsilon + float eps_; + /// input image size, default 1 + size_t ih_, iw_; + +public: + explicit MKLDNNTester(size_t iter = 3, float epsilon = 1e-4) { + iter_ = iter; + eps_ = epsilon; + log_ = false; + lvl_ = MKLDNN_ALL; + } + + ~MKLDNNTester() {} + +public: + void run(const TestConfig& dnn, + const TestConfig& ref, + size_t batchSize, + size_t inputImgH = 1, + size_t inputImgW = 1, + size_t iter = 3, + float epsilon = 1e-4, + bool log = false, + int level = MKLDNN_ALL); + void setLogLevel(int lvl) { lvl_ = lvl; } + +private: + void reset(const TestConfig& dnn, const TestConfig& ref, size_t batchSize); + void setInputImgSize(); + void runOnce(); + + void randomWgtDatas(); + void randomBotDatas(); + void randomTopDiffs(); + + void checkForward(); + void checkBackwardData(); + void checkBackwardWgts(); + + void clearWgtDiffs(); + void clearBotDiffs(); + void clearBotDiffs(int n); // clear specific layer + void clearTopDatas(); + + void printTopDatas(); + void printMatrix(const MatrixPtr& m); + void printVector(const VectorPtr& v); + + void saveWgt(const vector& from, vector& to); + void restoreWgt(const vector& from, vector& to); + + double compareMatrix(const MatrixPtr& m1, const MatrixPtr& m2); + double compareVector(const VectorPtr& v1, const VectorPtr& v2); + + /** + * Get delta percent + * 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. + */ + double getDelta(const real* d1, + const real* d2, + size_t len, + const float failRate = 1e-3, + const float thres = 0.1); +}; + +} // namespace paddle diff --git a/paddle/gserver/tests/test_ActivationGrad.cpp b/paddle/gserver/tests/test_ActivationGrad.cpp index b201ba8a5a4146ab28cd96454f434f889d72a968..de93972a5880518dfbfb9f8582e17c594e54b9b8 100644 --- a/paddle/gserver/tests/test_ActivationGrad.cpp +++ b/paddle/gserver/tests/test_ActivationGrad.cpp @@ -57,6 +57,39 @@ TEST(Activation, activation) { } } +void testSequenceSoftmaxAct(bool hasSubseq) { + LOG(INFO) << "test activation: sequence softmax"; + + const size_t size = 1; + TestConfig config; + config.biasSize = 0; + config.layerConfig.set_type("addto"); + config.layerConfig.set_size(size); + config.layerConfig.set_active_type("sequence_softmax"); + config.inputDefs.push_back( + {hasSubseq ? INPUT_HASSUB_SEQUENCE_DATA : INPUT_SEQUENCE_DATA, + "layer_0", + 1, + 0}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, + "sequence_softmax", + 100, + /* trans= */ false, + useGpu, + /* useWeight */ true); + } +} + +TEST(SequenceSoftmaxActivation, activation) { + for (auto hasSubseq : {false, true}) { + LOG(INFO) << "hasSubseq = " << hasSubseq; + testSequenceSoftmaxAct(hasSubseq); + } +} + int main(int argc, char** argv) { testing::InitGoogleTest(&argc, argv); initMain(argc, argv); diff --git a/paddle/gserver/tests/test_BatchNorm.cpp b/paddle/gserver/tests/test_BatchNorm.cpp index 83fcfed46cd568d22237eeef9c0215e4e3ad2666..659eefa31bdb1f2433d03a59d5bf4782c71bdecf 100644 --- a/paddle/gserver/tests/test_BatchNorm.cpp +++ b/paddle/gserver/tests/test_BatchNorm.cpp @@ -21,6 +21,8 @@ limitations under the License. */ #include "paddle/utils/GlobalConstants.h" #include "LayerGradUtil.h" +#include "paddle/cuda/include/hl_batch_norm.h" +#include "paddle/math/tests/TensorCheck.h" #include "paddle/testing/TestUtil.h" using namespace paddle; // NOLINT @@ -117,6 +119,74 @@ TEST(Layer, batchNorm) { CHECK_EQ(static_cast(convLayer->getOutputValue()->getWidth()), 576); } +#ifndef PADDLE_ONLY_CPU +void batchNormInference(int n, int c, int h, int w) { + MatrixPtr input = std::make_shared(n, c * h * w); + MatrixPtr cudnnOut = std::make_shared(n, c * h * w); + MatrixPtr cudaOut = std::make_shared(n, c * h * w); + MatrixPtr cudnnCheck = std::make_shared(n, c * h * w); + MatrixPtr cudaCheck = std::make_shared(n, c * h * w); + input->randomizeUniform(); + cudnnOut->zeroMem(); + cudaOut->zeroMem(); + + MatrixPtr scale = std::make_shared(1, c); + scale->randomizeUniform(); + MatrixPtr bias = std::make_shared(1, c); + bias->randomizeUniform(); + + MatrixPtr movingMean = std::make_shared(1, c); + movingMean->randomizeUniform(); + + MatrixPtr movingVar = std::make_shared(1, c); + movingVar->randomizeUniform(); + movingVar->clip(0.01, 50); + + hl_tensor_descriptor ioDesc; + hl_tensor_descriptor bnDesc; + hl_create_tensor_descriptor(&ioDesc); + hl_create_tensor_descriptor(&bnDesc); + hl_tensor_reshape(ioDesc, n, c, h, w); + hl_tensor_reshape(bnDesc, 1, c, 1, 1); + + double EPS = 1E-5; + hl_batch_norm_forward_inference(ioDesc, + input->getData(), + ioDesc, + cudnnOut->getData(), + bnDesc, + scale->getData(), + bias->getData(), + movingMean->getData(), + movingVar->getData(), + EPS); + + hl_batch_norm_cuda_inference(input->getData(), + cudaOut->getData(), + scale->getData(), + bias->getData(), + movingMean->getData(), + movingVar->getData(), + EPS, + n, + c, + h, + w); + + cudnnCheck->copyFrom(*cudnnOut); + cudaCheck->copyFrom(*cudaOut); + autotest::TensorCheckErr(*cudnnCheck, *cudaCheck); + + hl_destroy_tensor_descriptor(ioDesc); + hl_destroy_tensor_descriptor(bnDesc); +} + +TEST(BatchNorm, Inference) { + batchNormInference(33, 267, 1, 1); + batchNormInference(19, 105, 4, 4); +} +#endif + int main(int argc, char** argv) { testing::InitGoogleTest(&argc, argv); initMain(argc, argv); diff --git a/paddle/gserver/tests/test_KmaxSeqScore.cpp b/paddle/gserver/tests/test_KmaxSeqScore.cpp new file mode 100644 index 0000000000000000000000000000000000000000..308abe6816428bc0f98ec32e892622fa4a23b1ae --- /dev/null +++ b/paddle/gserver/tests/test_KmaxSeqScore.cpp @@ -0,0 +1,165 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include "ModelConfig.pb.h" +#include "paddle/gserver/layers/DataLayer.h" +#include "paddle/trainer/Trainer.h" +#include "paddle/utils/GlobalConstants.h" + +#include "LayerGradUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_bool(use_gpu); +DECLARE_int32(gpu_id); +DECLARE_bool(thread_local_rand_use_global_seed); + +vector randSampling(int 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); + return num; +} + +void genRandomSeqInfo(vector& seqStartPosition, + vector& subSeqStartPosition) { + const int maxSeqNum = 100; + // generate random start position information + int seqNum = 1 + (rand() % maxSeqNum); + seqStartPosition.resize(seqNum + 1, 0); + subSeqStartPosition.resize(1, 0); + + for (int i = 0; i < seqNum; ++i) { + int subSeqLen = 1 + (rand() % maxSeqNum); + for (int j = 0; j < subSeqLen; ++j) + subSeqStartPosition.push_back(subSeqStartPosition.back() + subSeqLen); + seqStartPosition[i + 1] = subSeqStartPosition.back(); + } +} + +void genRandomGroundTruth(real* values, + vector>& groundTruth, + vector& startPos, + size_t beamSize) { + groundTruth.resize(startPos.size() - 1, vector(beamSize, -1)); + for (size_t i = 0; i < startPos.size() - 1; ++i) { + int seqLen = startPos[i + 1] - startPos[i]; + vector pos = + randSampling(seqLen, min(static_cast(beamSize), seqLen)); + for (size_t j = 0; j < pos.size(); ++j) { + groundTruth[i][j] = pos[j]; + values[startPos[i] + pos[j]] = 1.; + } + } +} + +void checkLayerOut(vector> groundTruth, + real* layerOut, + size_t beamSize) { + for (size_t i = 0; i < groundTruth.size(); ++i) { + int begPos = i * beamSize; + vector tmp(layerOut + begPos, layerOut + begPos + beamSize); + sort(begin(tmp), end(tmp)); + sort(begin(groundTruth[i]), end(groundTruth[i])); + for (size_t j = 0; j < beamSize; ++j) CHECK_EQ(tmp[j], groundTruth[i][j]); + } +} + +TEST(Layer, kmaxSeqScoreLayer) { + const size_t maxBeamSize = 100; + size_t beamSize = 1 + (rand() % maxBeamSize); + + vector seqStartPosition; + vector subSeqStartPosition; + genRandomSeqInfo(seqStartPosition, subSeqStartPosition); + MatrixPtr inValue = + Matrix::create(subSeqStartPosition.back(), 1, false, false); + + std::vector mode = {false}; +#ifndef PADDLE_ONLY_CPU + mode.push_back(true); +#endif + + for (auto hasSubseq : {false, true}) { + vector> groundTruth; + inValue->randomizeUniform(); + genRandomGroundTruth(inValue->getData(), + groundTruth, + hasSubseq ? subSeqStartPosition : seqStartPosition, + beamSize); + + for (auto useGpu : mode) { + TestConfig config; + config.layerConfig.set_type("kmax_seq_score"); + config.layerConfig.set_beam_size(beamSize); + + if (hasSubseq) { + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, + "scores", + inValue, + seqStartPosition, + subSeqStartPosition}); + } else { + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, "scores", inValue, seqStartPosition}); + } + config.layerConfig.add_inputs(); + + // data layer initialize + std::vector dataLayers; + LayerMap layerMap; + vector datas; + initDataLayer( + config, + &dataLayers, + &datas, + &layerMap, + "kmax_seq_score", + 100 /* actually this parameter is unused in self-defined input*/, + false, + useGpu); + // test layer initialize + std::vector parameters; + LayerPtr kmaxSeqScoreLayer; + FLAGS_use_gpu = useGpu; + initTestLayer(config, &layerMap, ¶meters, &kmaxSeqScoreLayer); + kmaxSeqScoreLayer->forward(PASS_TRAIN); + + const MatrixPtr outValue = kmaxSeqScoreLayer->getOutputValue(); + CHECK_EQ(outValue->getHeight(), + hasSubseq ? subSeqStartPosition.size() - 1 + : seqStartPosition.size() - 1); + CHECK_EQ(outValue->getWidth(), beamSize); + checkLayerOut(groundTruth, outValue->getData(), beamSize); + } + } +} + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + initMain(argc, argv); + FLAGS_thread_local_rand_use_global_seed = true; + srand((size_t)(time(NULL))); + return RUN_ALL_TESTS(); +} diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp index fe11278f41c0118ee0bdb34f17fbf9602e0fa76b..0f312b6ca50bc1e6317251ba785f1c61a224b54e 100644 --- a/paddle/gserver/tests/test_LayerGrad.cpp +++ b/paddle/gserver/tests/test_LayerGrad.cpp @@ -1899,6 +1899,84 @@ TEST(Layer, CropLayer) { } } +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; +} + +TEST(Layer, SubNestedSequenceLayer) { + // layer size is not crutial for this layer, + // so use a small layer size in unittest + const int layerSize = 4; + + const int maxSeqNum = 50; + const int maxSeqLen = 50; + const int maxBeamSize = 32; + + srand((size_t)(time(NULL))); + int beamSize = 1 + (rand() % maxBeamSize); + + TestConfig config; + config.layerConfig.set_type("sub_nested_seq"); + config.layerConfig.set_name("sub_nested_seq_layer"); + config.layerConfig.set_size(layerSize); + + int seqNum = 1 + (rand() % maxSeqNum); + + // sequence information for the first input, it is a nested sequence + vector seqStartPos(seqNum + 1, 0); + vector subSeqStartPos(1, 0); + + // selected indices + MatrixPtr selectedIndices = Matrix::create(seqNum, beamSize, false, false); + selectedIndices->one(); + selectedIndices->mulScalar(-1.); + real* indicesData = selectedIndices->getData(); + + for (int i = 0; i < seqNum; ++i) { + int subSeqNum = 1 + (rand() % maxSeqNum); + for (int j = 0; j < subSeqNum; ++j) { + subSeqStartPos.push_back(subSeqStartPos.back() + + (1 + (rand() % maxSeqLen))); + } + vector selSeqs = + randSampling(static_cast(subSeqNum), min(beamSize, subSeqNum)); + memcpy(indicesData + (i * beamSize), + selSeqs.data(), + selSeqs.size() * sizeof(real)); + seqStartPos[i + 1] = subSeqStartPos.back(); + } + + MatrixPtr seqInputPtr = + Matrix::create(seqStartPos.back(), layerSize, false, false); + seqInputPtr->randomizeUniform(); + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, + "nested_seq_input", + seqInputPtr, + seqStartPos, + subSeqStartPos}); + config.layerConfig.add_inputs(); + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, "selected_indices", selectedIndices}); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, + "sub_nested_seq", + /* batchSize */ seqNum, + /* trans */ false, + /* useGpu*/ useGpu, + /* useWeight */ false); + } +} + TEST(Layer, ClipLayer) { const size_t batchSize = 128; const size_t size = 512; diff --git a/paddle/gserver/tests/test_MKLDNN.cpp b/paddle/gserver/tests/test_MKLDNN.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e1d2270df24331914f3a51acc90a518084b3ce4e --- /dev/null +++ b/paddle/gserver/tests/test_MKLDNN.cpp @@ -0,0 +1,76 @@ +/* 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 +#include +#include +#include "MKLDNNTester.h" +#include "ModelConfig.pb.h" + +using namespace paddle; // NOLINT + +DECLARE_bool(thread_local_rand_use_global_seed); +DECLARE_bool(use_gpu); +DECLARE_bool(use_mkldnn); + +struct testFCDesc { + int bs; + int ic; + int oc; + int ih, iw; // oh == ow == 1 +}; + +void testFcLayer(const testFCDesc& pm) { + const std::string compareTypes[] = {"mkldnn_fc", "fc"}; + TestConfig cfg; + cfg.layerConfig.set_type(compareTypes[0]); + cfg.layerConfig.set_size(pm.oc); + cfg.inputDefs.push_back( + {INPUT_DATA, + "layer_0", + /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw), + /* size of weight= */ size_t(pm.oc * pm.ic * pm.ih * pm.iw)}); + cfg.layerConfig.add_inputs(); + + MKLDNNTester tester; + for (auto biasSize : {pm.oc, 0}) { + cfg.biasSize = biasSize; + TestConfig ref = cfg; + ref.layerConfig.set_type(compareTypes[1]); + for (auto bs : {pm.bs, 1}) { + tester.run(cfg, ref, bs, pm.ih, pm.iw); + } + } +} + +TEST(MKLDNNLayer, FcLayer) { + testFcLayer({/*bs*/ 2, /*ic*/ 2, /*oc*/ 3, /*ih*/ 1, /*iw*/ 1}); + testFcLayer({/*bs*/ 3, /*ic*/ 7, /*oc*/ 19, /*ih*/ 1, /*iw*/ 1}); + testFcLayer({/*bs*/ 8, /*ic*/ 16, /*oc*/ 32, /*ih*/ 13, /*iw*/ 13}); + testFcLayer({/*bs*/ 4, /*ic*/ 12, /*oc*/ 18, /*ih*/ 13, /*iw*/ 11}); + testFcLayer({/*bs*/ 2, /*ic*/ 64, /*oc*/ 32, /*ih*/ 16, /*iw*/ 16}); + testFcLayer({/*bs*/ 15, /*ic*/ 3, /*oc*/ 6, /*ih*/ 16, /*iw*/ 16}); +} + +// TODO(TJ): add branch test + +int main(int argc, char** argv) { + testing::InitGoogleTest(&argc, argv); + FLAGS_use_gpu = false; + FLAGS_use_mkldnn = true; + initMain(argc, argv); + FLAGS_thread_local_rand_use_global_seed = true; + srand(1); + return RUN_ALL_TESTS(); +} diff --git a/paddle/math/BaseMatrix.cu b/paddle/math/BaseMatrix.cu index ba2b47d6cc6961a380b7db2781b4d214dea829db..5435808fb7f70fdf1ac98815f7fe8890fb85527c 100644 --- a/paddle/math/BaseMatrix.cu +++ b/paddle/math/BaseMatrix.cu @@ -12,21 +12,21 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include -#include #include +#include +#include #include "BaseMatrix.h" -#include "hl_matrix_ops.cuh" -#include "hl_matrix_base.cuh" -#include "hl_matrix_apply.cuh" -#include "SIMDFunctions.h" #include "MathFunctions.h" +#include "SIMDFunctions.h" +#include "hl_matrix_apply.cuh" +#include "hl_matrix_base.cuh" +#include "hl_matrix_ops.cuh" namespace paddle { const char* SPARSE_SUPPORT_ERROR = "Sparse Matrix/Vector is not supported."; -template +template template int BaseMatrixT::applyUnary(Op op) { MatrixOffset offset(0, 0); @@ -34,9 +34,11 @@ int BaseMatrixT::applyUnary(Op op) { return 0; } -template +template template -int BaseMatrixT::applyUnary(Op op, int numRows, int numCols, +int BaseMatrixT::applyUnary(Op op, + int numRows, + int numCols, MatrixOffset& offset) { CHECK(!this->isSparse()) << SPARSE_SUPPORT_ERROR; int dimM = numRows; @@ -56,7 +58,7 @@ int BaseMatrixT::applyUnary(Op op, int numRows, int numCols, return 0; } -template +template template int BaseMatrixT::applyBinary(Op op, BaseMatrixT& b) { CHECK(height_ == b.height_ && width_ == b.width_) @@ -67,18 +69,23 @@ int BaseMatrixT::applyBinary(Op op, BaseMatrixT& b) { return 0; } -template +template template -int BaseMatrixT::applyBinary(Op op, BaseMatrixT& b, int numRows, int numCols, - MatrixOffset& offset) { +int BaseMatrixT::applyBinary( + Op op, BaseMatrixT& b, int numRows, int numCols, MatrixOffset& offset) { applyBinary(op, b, numRows, numCols, offset, false_type(), false_type()); return 0; } -template +template template -int BaseMatrixT::applyBinary(Op op, BaseMatrixT& b, int numRows, int numCols, - MatrixOffset& offset, bAsRowVector, bAsColVector) { +int BaseMatrixT::applyBinary(Op op, + BaseMatrixT& b, + int numRows, + int numCols, + MatrixOffset& offset, + bAsRowVector, + bAsColVector) { CHECK(!this->isSparse()) << SPARSE_SUPPORT_ERROR; CHECK(!b.isSparse()) << SPARSE_SUPPORT_ERROR; CHECK(useGpu_ == b.useGpu_) << "Matrix type mismatch"; @@ -91,8 +98,8 @@ int BaseMatrixT::applyBinary(Op op, BaseMatrixT& b, int numRows, int numCols, T* A = data_; T* B = b.data_; CAL_MATRIX_START_ADDRESS(A, height_, width_, lda, offset.aCol_, offset.aRow_); - CAL_MATRIX_START_ADDRESS(B, b.height_, b.width_, ldb, offset.bCol_, - offset.bRow_); + CAL_MATRIX_START_ADDRESS( + B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); CHECK_LE(dimM + offset.aRow_, this->height_); CHECK_LE(dimN + offset.aCol_, this->width_); if (!bAsRowVector::value && !bAsColVector::value) { @@ -115,7 +122,7 @@ int BaseMatrixT::applyBinary(Op op, BaseMatrixT& b, int numRows, int numCols, return 0; } -template +template template int BaseMatrixT::applyTernary(Op op, BaseMatrixT& b, BaseMatrixT& c) { CHECK_EQ(height_, b.height_); @@ -129,21 +136,29 @@ int BaseMatrixT::applyTernary(Op op, BaseMatrixT& b, BaseMatrixT& c) { return 0; } -template +template template -int BaseMatrixT::applyTernary(Op op, BaseMatrixT& b, BaseMatrixT& c, - int numRows, int numCols, +int BaseMatrixT::applyTernary(Op op, + BaseMatrixT& b, + BaseMatrixT& c, + int numRows, + int numCols, MatrixOffset& offset) { applyTernary(op, b, c, numRows, numCols, offset, false_type(), false_type()); return 0; } -template +template template -int BaseMatrixT::applyTernary(Op op, BaseMatrixT& b, BaseMatrixT& c, - int numRows, int numCols, MatrixOffset& offset, - cAsRowVector, cAsColVector) { +int BaseMatrixT::applyTernary(Op op, + BaseMatrixT& b, + BaseMatrixT& c, + int numRows, + int numCols, + MatrixOffset& offset, + cAsRowVector, + cAsColVector) { CHECK(!this->isSparse()) << SPARSE_SUPPORT_ERROR; CHECK(!b.isSparse()) << SPARSE_SUPPORT_ERROR; CHECK(!c.isSparse()) << SPARSE_SUPPORT_ERROR; @@ -160,10 +175,10 @@ int BaseMatrixT::applyTernary(Op op, BaseMatrixT& b, BaseMatrixT& c, T* B = b.data_; T* C = c.data_; CAL_MATRIX_START_ADDRESS(A, height_, width_, lda, offset.aCol_, offset.aRow_); - CAL_MATRIX_START_ADDRESS(B, b.height_, b.width_, ldb, offset.bCol_, - offset.bRow_); - CAL_MATRIX_START_ADDRESS(C, c.height_, c.width_, ldc, offset.cCol_, - offset.cRow_); + CAL_MATRIX_START_ADDRESS( + B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); + CAL_MATRIX_START_ADDRESS( + C, c.height_, c.width_, ldc, offset.cCol_, offset.cRow_); CHECK_LE(dimM + offset.aRow_, this->height_); CHECK_LE(dimN + offset.aCol_, this->width_); @@ -180,21 +195,21 @@ int BaseMatrixT::applyTernary(Op op, BaseMatrixT& b, BaseMatrixT& c, } if (true == useGpu_) { - hl_gpu_apply_ternary_op - ( + hl_gpu_apply_ternary_op( op, A, B, C, dimM, dimN, lda, ldb, ldc); } else { - hl_cpu_apply_ternary_op - ( + hl_cpu_apply_ternary_op( op, A, B, C, dimM, dimN, lda, ldb, ldc); } return 0; } -template +template template -int BaseMatrixT::applyQuaternary(Op op, BaseMatrixT& b, BaseMatrixT& c, +int BaseMatrixT::applyQuaternary(Op op, + BaseMatrixT& b, + BaseMatrixT& c, BaseMatrixT& d) { CHECK_EQ(height_, b.height_); CHECK_EQ(width_, b.width_); @@ -209,10 +224,14 @@ int BaseMatrixT::applyQuaternary(Op op, BaseMatrixT& b, BaseMatrixT& c, return 0; } -template +template template -int BaseMatrixT::applyQuaternary(Op op, BaseMatrixT& b, BaseMatrixT& c, - BaseMatrixT& d, int numRows, int numCols, +int BaseMatrixT::applyQuaternary(Op op, + BaseMatrixT& b, + BaseMatrixT& c, + BaseMatrixT& d, + int numRows, + int numCols, MatrixOffset& offset) { CHECK(!this->isSparse()) << SPARSE_SUPPORT_ERROR; CHECK(!b.isSparse()) << SPARSE_SUPPORT_ERROR; @@ -234,12 +253,12 @@ int BaseMatrixT::applyQuaternary(Op op, BaseMatrixT& b, BaseMatrixT& c, T* C = c.data_; T* D = d.data_; CAL_MATRIX_START_ADDRESS(A, height_, width_, lda, offset.aCol_, offset.aRow_); - CAL_MATRIX_START_ADDRESS(B, b.height_, b.width_, ldb, offset.bCol_, - offset.bRow_); - CAL_MATRIX_START_ADDRESS(C, c.height_, c.width_, ldc, offset.cCol_, - offset.cRow_); - CAL_MATRIX_START_ADDRESS(D, d.height_, d.width_, ldd, offset.dCol_, - offset.dRow_); + CAL_MATRIX_START_ADDRESS( + B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); + CAL_MATRIX_START_ADDRESS( + C, c.height_, c.width_, ldc, offset.cCol_, offset.cRow_); + CAL_MATRIX_START_ADDRESS( + D, d.height_, d.width_, ldd, offset.dCol_, offset.dRow_); CHECK_LE(dimM + offset.aRow_, this->height_); CHECK_LE(dimN + offset.aCol_, this->width_); @@ -250,22 +269,29 @@ int BaseMatrixT::applyQuaternary(Op op, BaseMatrixT& b, BaseMatrixT& c, CHECK_LE(dimM + offset.dRow_, d.height_); CHECK_LE(dimN + offset.dCol_, d.width_); if (true == useGpu_) { - hl_gpu_apply_quaternary_op(op, A, B, C, D, dimM, dimN, lda, ldb, - ldc, ldd); + hl_gpu_apply_quaternary_op(op, A, B, C, D, dimM, dimN, lda, ldb, ldc, ldd); } else { - hl_cpu_apply_quaternary_op(op, A, B, C, D, dimM, dimN, lda, ldb, - ldc, ldd); + hl_cpu_apply_quaternary_op(op, A, B, C, D, dimM, dimN, lda, ldb, ldc, ldd); } return 0; } -template -template +template -int BaseMatrixT::aggregate(Agg agg, Op op, Saver sv, BaseMatrixT& b, - int numRows, int numCols, MatrixOffset& offset, - aAsRowVector, aAsColVector) { +int BaseMatrixT::aggregate(Agg agg, + Op op, + Saver sv, + BaseMatrixT& b, + int numRows, + int numCols, + MatrixOffset& offset, + aAsRowVector, + aAsColVector) { CHECK_EQ(useGpu_, b.useGpu_); int ld = stride_; @@ -273,10 +299,10 @@ int BaseMatrixT::aggregate(Agg agg, Op op, Saver sv, BaseMatrixT& b, T* dst = data_; T* B = b.data_; - CAL_MATRIX_START_ADDRESS(dst, height_, width_, ld, offset.aCol_, - offset.aRow_); - CAL_MATRIX_START_ADDRESS(B, b.height_, b.width_, ldb, offset.bCol_, - offset.bRow_); + CAL_MATRIX_START_ADDRESS( + dst, height_, width_, ld, offset.aCol_, offset.aRow_); + CAL_MATRIX_START_ADDRESS( + B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); if (aAsRowVector::value && !aAsColVector::value) { if (useGpu_) { @@ -297,12 +323,21 @@ int BaseMatrixT::aggregate(Agg agg, Op op, Saver sv, BaseMatrixT& b, return 0; } -template -template +template -int BaseMatrixT::aggregate(Agg agg, Op op, Saver sv, BaseMatrixT& b, - BaseMatrixT& c, int numRows, int numCols, - MatrixOffset& offset, aAsRowVector, +int BaseMatrixT::aggregate(Agg agg, + Op op, + Saver sv, + BaseMatrixT& b, + BaseMatrixT& c, + int numRows, + int numCols, + MatrixOffset& offset, + aAsRowVector, aAsColVector) { CHECK_EQ(useGpu_, b.useGpu_); CHECK_EQ(useGpu_, c.useGpu_); @@ -314,28 +349,28 @@ int BaseMatrixT::aggregate(Agg agg, Op op, Saver sv, BaseMatrixT& b, T* dst = data_; T* B = b.data_; T* C = c.data_; - CAL_MATRIX_START_ADDRESS(dst, height_, width_, ld, offset.aCol_, - offset.aRow_); - CAL_MATRIX_START_ADDRESS(B, b.height_, b.width_, ldb, offset.bCol_, - offset.bRow_); - CAL_MATRIX_START_ADDRESS(C, c.height_, c.width_, ldc, offset.cCol_, - offset.cRow_); + CAL_MATRIX_START_ADDRESS( + dst, height_, width_, ld, offset.aCol_, offset.aRow_); + CAL_MATRIX_START_ADDRESS( + B, b.height_, b.width_, ldb, offset.bCol_, offset.bRow_); + CAL_MATRIX_START_ADDRESS( + C, c.height_, c.width_, ldc, offset.cCol_, offset.cRow_); if (aAsRowVector::value && !aAsColVector::value) { if (useGpu_) { - hl_gpu_matrix_column_op(agg, op, sv, numRows, numCols, dst, B, - ldb, C, ldc); + hl_gpu_matrix_column_op( + agg, op, sv, numRows, numCols, dst, B, ldb, C, ldc); } else { - hl_cpu_matrix_column_op(agg, op, sv, numRows, numCols, dst, B, - ldb, C, ldc); + hl_cpu_matrix_column_op( + agg, op, sv, numRows, numCols, dst, B, ldb, C, ldc); } } else if (!aAsRowVector::value && aAsColVector::value) { if (useGpu_) { - hl_gpu_matrix_row_op(agg, op, sv, numRows, numCols, dst, ld, B, - ldb, C, ldc); + hl_gpu_matrix_row_op( + agg, op, sv, numRows, numCols, dst, ld, B, ldb, C, ldc); } else { - hl_cpu_matrix_row_op(agg, op, sv, numRows, numCols, dst, ld, B, - ldb, C, ldc); + hl_cpu_matrix_row_op( + agg, op, sv, numRows, numCols, dst, ld, B, ldb, C, ldc); } } else { LOG(FATAL) << "not supported"; @@ -350,15 +385,19 @@ int BaseMatrixT::aggregate(Agg agg, Op op, Saver sv, BaseMatrixT& b, */ DEFINE_MATRIX_UNARY_OP(Neg, a = -a); -template -void BaseMatrixT::neg() { applyUnary(unary::Neg()); } +template +void BaseMatrixT::neg() { + applyUnary(unary::Neg()); +} DEFINE_MATRIX_UNARY_OP(Exp, a = exp(a)); -template<> -void BaseMatrixT::exp2() { applyUnary(unary::Exp()); } +template <> +void BaseMatrixT::exp2() { + applyUnary(unary::Exp()); +} DEFINE_MATRIX_UNARY_OP(Log, a = log(a)); -template<> +template <> void BaseMatrixT::log2() { if (useGpu_) { applyUnary(unary::Log()); @@ -368,30 +407,42 @@ void BaseMatrixT::log2() { } DEFINE_MATRIX_UNARY_OP(Sqrt, a = sqrt(a)); -template<> -void BaseMatrixT::sqrt2() { applyUnary(unary::Sqrt()); } +template <> +void BaseMatrixT::sqrt2() { + applyUnary(unary::Sqrt()); +} DEFINE_MATRIX_UNARY_OP(Square, a = a * a); -template -void BaseMatrixT::square2() { applyUnary(unary::Square()); } +template +void BaseMatrixT::square2() { + applyUnary(unary::Square()); +} DEFINE_MATRIX_UNARY_OP(Reciprocal, a = 1.0f / a); -template -void BaseMatrixT::reciprocal2() { applyUnary(unary::Reciprocal()); } +template +void BaseMatrixT::reciprocal2() { + applyUnary(unary::Reciprocal()); +} DEFINE_MATRIX_UNARY_OP(Abs, a = a > 0 ? a : -a); -template -void BaseMatrixT::abs2() { applyUnary(unary::Abs()); } +template +void BaseMatrixT::abs2() { + applyUnary(unary::Abs()); +} DEFINE_MATRIX_UNARY_OP(Sign, a = (a > 0) - (a < 0)); -template -void BaseMatrixT::sign2() { applyUnary(unary::Sign()); } +template +void BaseMatrixT::sign2() { + applyUnary(unary::Sign()); +} DEFINE_MATRIX_UNARY_OP(Zero, a = 0); -template -void BaseMatrixT::zero() { applyUnary(unary::Zero()); } +template +void BaseMatrixT::zero() { + applyUnary(unary::Zero()); +} -template +template void BaseMatrixT::zeroAtOffset(int64_t columnOffset, int64_t numColumns) { int numRows = height_; int numCols = numColumns; @@ -400,11 +451,13 @@ void BaseMatrixT::zeroAtOffset(int64_t columnOffset, int64_t numColumns) { } DEFINE_MATRIX_UNARY_OP(One, a = 1); -template -void BaseMatrixT::one() { applyUnary(unary::One()); } +template +void BaseMatrixT::one() { + applyUnary(unary::One()); +} DEFINE_MATRIX_UNARY_PARAMETER_OP(Pow, ONE_PARAMETER, a = pow(a, p)); -template<> +template <> void BaseMatrixT::pow2(real p) { if (useGpu_) { applyUnary(unary::Pow(p)); @@ -414,51 +467,67 @@ void BaseMatrixT::pow2(real p) { } DEFINE_MATRIX_UNARY_PARAMETER_OP(SubScalar, ONE_PARAMETER, a -= p); -template -void BaseMatrixT::subScalar(T p) { applyUnary(unary::SubScalar(p)); } +template +void BaseMatrixT::subScalar(T p) { + applyUnary(unary::SubScalar(p)); +} DEFINE_MATRIX_UNARY_PARAMETER_OP(MulScalar, ONE_PARAMETER, a *= p); -template -void BaseMatrixT::mulScalar(T p) { applyUnary(unary::MulScalar(p)); } +template +void BaseMatrixT::mulScalar(T p) { + applyUnary(unary::MulScalar(p)); +} DEFINE_MATRIX_UNARY_PARAMETER_OP(DivScalar, ONE_PARAMETER, a /= p); -template -void BaseMatrixT::divScalar(T p) { applyUnary(unary::DivScalar(p)); } +template +void BaseMatrixT::divScalar(T p) { + applyUnary(unary::DivScalar(p)); +} DEFINE_MATRIX_UNARY_PARAMETER_OP(Assign, ONE_PARAMETER, a = p); -template -void BaseMatrixT::assign(T p) { applyUnary(unary::Assign(p)); } +template +void BaseMatrixT::assign(T p) { + applyUnary(unary::Assign(p)); +} DEFINE_MATRIX_UNARY_PARAMETER_OP(Add, ONE_PARAMETER, a += p); -template -void BaseMatrixT::add(T p) { applyUnary(unary::Add(p)); } +template +void BaseMatrixT::add(T p) { + applyUnary(unary::Add(p)); +} DEFINE_MATRIX_UNARY_PARAMETER_OP(Add2, TWO_PARAMETER, a = a * p1 + p2); -template -void BaseMatrixT::add(T p1, T p2) { applyUnary(unary::Add2(p1, p2)); } +template +void BaseMatrixT::add(T p1, T p2) { + applyUnary(unary::Add2(p1, p2)); +} -DEFINE_MATRIX_UNARY_PARAMETER_OP(Clip, TWO_PARAMETER, +DEFINE_MATRIX_UNARY_PARAMETER_OP(Clip, + TWO_PARAMETER, a = a < p1 ? p1 : (a > p2 ? p2 : a)); -template -void BaseMatrixT::clip(T p1, T p2) { applyUnary(unary::Clip(p1, p2)); } +template +void BaseMatrixT::clip(T p1, T p2) { + applyUnary(unary::Clip(p1, p2)); +} -DEFINE_MATRIX_BINARY_PARAMETER_OP(ClipDerivative, TWO_PARAMETER, - a = b < p1 ? 0 : (b > p2 ? 0 : 1)); -template +DEFINE_MATRIX_BINARY_PARAMETER_OP(ClipDerivative, + TWO_PARAMETER, + a = b < p1 ? 0 : (b > p2 ? 0 : 1)); +template void BaseMatrixT::clipDerivative(BaseMatrixT& b, T p1, T p2) { applyBinary(binary::ClipDerivative(p1, p2), b); } -DEFINE_MATRIX_UNARY_PARAMETER_OP(BiggerThanScalar, ONE_PARAMETER, +DEFINE_MATRIX_UNARY_PARAMETER_OP(BiggerThanScalar, + ONE_PARAMETER, a = a > p ? 1.0f : 0.0f); -template +template void BaseMatrixT::biggerThanScalar(T p) { applyUnary(unary::BiggerThanScalar(p)); } -DEFINE_MATRIX_UNARY_PARAMETER_OP(DownClip, ONE_PARAMETER, - a = a > p ? a : p); -template +DEFINE_MATRIX_UNARY_PARAMETER_OP(DownClip, ONE_PARAMETER, a = a > p ? a : p); +template void BaseMatrixT::downClip(T p) { applyUnary(unary::DownClip(p)); } @@ -469,12 +538,12 @@ void BaseMatrixT::downClip(T p) { */ DEFINE_MATRIX_BINARY_OP(Add, a += b); -template +template void BaseMatrixT::add(BaseMatrixT& b) { applyBinary(binary::Add(), b); } -template<> +template <> void BaseMatrixT::add(BaseMatrixT& b) { if (useGpu_) { applyBinary(binary::Add(), b); @@ -485,7 +554,7 @@ void BaseMatrixT::add(BaseMatrixT& b) { } } -template +template void BaseMatrixT::addAtOffset(BaseMatrixT& b, int64_t columnOffset) { if (columnOffset + b.width_ <= width_) { int numRows = height_; @@ -504,43 +573,53 @@ void BaseMatrixT::addAtOffset(BaseMatrixT& b, int64_t columnOffset) { } } -template +template void BaseMatrixT::addP2P(BaseMatrixT& b) { T* A = data_; T* B = b.data_; int dimM = height_; int dimN = width_; - hl_gpu_apply_binary_op, 0, 0> - (binary::Add(), A, B, dimM, dimN, dimN, dimN); + hl_gpu_apply_binary_op, 0, 0>( + binary::Add(), A, B, dimM, dimN, dimN, dimN); } -template +template void BaseMatrixT::addColVector(BaseMatrixT& b) { MatrixOffset offset(0, 0, 0, 0); int numRows = height_; int numCols = width_; - applyBinary(binary::Add(), b, numRows, numCols, offset, false_type(), + applyBinary(binary::Add(), + b, + numRows, + numCols, + offset, + false_type(), true_type() /* bAsColVector */); } -template +template void BaseMatrixT::addRowVector(BaseMatrixT& b) { MatrixOffset offset(0, 0, 0, 0); int numRows = height_; int numCols = width_; - applyBinary(binary::Add(), b, numRows, numCols, offset, - true_type() /* bAsRowVector */, false_type()); + applyBinary(binary::Add(), + b, + numRows, + numCols, + offset, + true_type() /* bAsRowVector */, + false_type()); } DEFINE_MATRIX_BINARY_PARAMETER_OP(Add1, ONE_PARAMETER, a += b * p); -template +template void BaseMatrixT::add(BaseMatrixT& b, T p) { applyBinary(binary::Add1(p), b); } DEFINE_MATRIX_BINARY_PARAMETER_OP(Pow, ONE_PARAMETER, a = pow(b, p)); -template<> +template <> void BaseMatrixT::pow2(BaseMatrixT& b, real p) { if (useGpu_) { applyBinary(binary::Pow(p), b); @@ -550,36 +629,45 @@ void BaseMatrixT::pow2(BaseMatrixT& b, real p) { } DEFINE_MATRIX_BINARY_PARAMETER_OP(Add2, TWO_PARAMETER, a = p1 * a + p2 * b); -template +template void BaseMatrixT::add(BaseMatrixT& b, T p1, T p2) { applyBinary(binary::Add2(p1, p2), b); } -template +template void BaseMatrixT::addBias(BaseMatrixT& b, T scale) { MatrixOffset offset(0, 0, 0, 0); int numRows = height_; int numCols = width_; - applyBinary(binary::Add1(scale), b, numRows, numCols, offset, - true_type() /* bAsRowVector */, false_type()); + applyBinary(binary::Add1(scale), + b, + numRows, + numCols, + offset, + true_type() /* bAsRowVector */, + false_type()); } DEFINE_MATRIX_BINARY_OP(Sub, a -= b); -template -void BaseMatrixT::sub(BaseMatrixT& b) { applyBinary(binary::Sub(), b); } +template +void BaseMatrixT::sub(BaseMatrixT& b) { + applyBinary(binary::Sub(), b); +} DEFINE_MATRIX_BINARY_PARAMETER_OP(Sub1, ONE_PARAMETER, a -= b * p); -template +template void BaseMatrixT::sub(BaseMatrixT& b, T p) { applyBinary(binary::Sub1(p), b); } DEFINE_MATRIX_BINARY_OP(Relu, b = a > 0.0f ? a : 0.0f); -template -void BaseMatrixT::relu(BaseMatrixT& b) { applyBinary(binary::Relu(), b); } +template +void BaseMatrixT::relu(BaseMatrixT& b) { + applyBinary(binary::Relu(), b); +} DEFINE_MATRIX_BINARY_OP(ReluDerivative, a *= (b > 0.0f ? 1.0f : 0.0f)); -template +template void BaseMatrixT::reluDerivative(BaseMatrixT& b) { applyBinary(binary::ReluDerivative(), b); } @@ -589,7 +677,7 @@ DEFINE_MATRIX_BINARY_OP(Softrelu, const T THRESHOLD = 40.0; ? THRESHOLD : ((a < -THRESHOLD) ? (-THRESHOLD) : a)))); -template<> +template <> void BaseMatrixT::softrelu(BaseMatrixT& b) { applyBinary(binary::Softrelu(), b); } @@ -599,97 +687,100 @@ DEFINE_MATRIX_BINARY_OP( a *= (1.0 - exp(-1.0 * ((b > THRESHOLD) ? THRESHOLD : ((b < -THRESHOLD) ? (-THRESHOLD) : b))))); -template<> +template <> void BaseMatrixT::softreluDerivative(BaseMatrixT& b) { applyBinary(binary::SoftreluDerivative(), b); } DEFINE_MATRIX_BINARY_PARAMETER_OP(Brelu, TWO_PARAMETER, b = a > p1 ? a : p1; b = b < p2 ? b : p2); -template +template void BaseMatrixT::brelu(BaseMatrixT& b) { - int p1 = 0, p2 = 24; //! TODO(yuyang18): Make p1,p2 configuable. + int p1 = 0, p2 = 24; //! TODO(yuyang18): Make p1,p2 configuable. applyBinary(binary::Brelu(p1, p2), b); } -DEFINE_MATRIX_BINARY_PARAMETER_OP(BreluDerivative, TWO_PARAMETER, +DEFINE_MATRIX_BINARY_PARAMETER_OP(BreluDerivative, + TWO_PARAMETER, a *= (b > p1 && b < p2) ? 1.0 : 0.0); -template +template void BaseMatrixT::breluDerivative(BaseMatrixT& b) { int p1 = 0, p2 = 24; applyBinary(binary::BreluDerivative(p1, p2), b); } DEFINE_MATRIX_BINARY_OP(Square, b = a * a); -template +template void BaseMatrixT::square2(BaseMatrixT& b) { applyBinary(binary::Square(), b); } DEFINE_MATRIX_BINARY_OP(SquareDerivative, a *= 2.0 * b); -template +template void BaseMatrixT::squareDerivative(BaseMatrixT& b) { applyBinary(binary::SquareDerivative(), b); } -DEFINE_MATRIX_BINARY_OP(Tanh, - T tmp = -2.0 * a; - tmp = (tmp > EXP_MAX_INPUT) ? EXP_MAX_INPUT : tmp; - b = 2.0 / (1.0 + std::exp(tmp)) - 1.0); -template<> +DEFINE_MATRIX_BINARY_OP(Tanh, T tmp = -2.0 * a; + tmp = (tmp > EXP_MAX_INPUT) ? EXP_MAX_INPUT : tmp; + b = 2.0 / (1.0 + std::exp(tmp)) - 1.0); +template <> void BaseMatrixT::tanh(BaseMatrixT& b) { applyBinary(binary::Tanh(), b); } DEFINE_MATRIX_BINARY_OP(TanhDerivative, a *= 1 - b * b); -template +template void BaseMatrixT::tanhDerivative(BaseMatrixT& b) { applyBinary(binary::TanhDerivative(), b); } -DEFINE_MATRIX_BINARY_PARAMETER_OP(ScaledTanh, TWO_PARAMETER, - b = p1 * - (2.0 / (1.0 + exp(-2 * p2 * a)) - 1.0)); -template<> +DEFINE_MATRIX_BINARY_PARAMETER_OP( + ScaledTanh, TWO_PARAMETER, b = p1 * (2.0 / (1.0 + exp(-2 * p2 * a)) - 1.0)); +template <> void BaseMatrixT::scaledTanh(BaseMatrixT& b, real p1, real p2) { applyBinary(binary::ScaledTanh(p1, p2), b); } -DEFINE_MATRIX_BINARY_PARAMETER_OP(ScaledTanhDerivative, TWO_PARAMETER, +DEFINE_MATRIX_BINARY_PARAMETER_OP(ScaledTanhDerivative, + TWO_PARAMETER, a *= p2 * (p1 - b * b)); -template +template void BaseMatrixT::scaledTanhDerivative(BaseMatrixT& b, T p1, T p2) { applyBinary(binary::ScaledTanhDerivative(p1 * p1, p2 / p1), b); } DEFINE_MATRIX_BINARY_OP(Reciprocal, b = 1.0f / a); -template +template void BaseMatrixT::reciprocal2(BaseMatrixT& b) { applyBinary(binary::Reciprocal(), b); } DEFINE_MATRIX_BINARY_OP(ReciprocalDerivative, a *= -b * b); -template +template void BaseMatrixT::reciprocalDerivative(BaseMatrixT& b) { applyBinary(binary::ReciprocalDerivative(), b); } DEFINE_MATRIX_BINARY_OP(Abs, b = a > 0.0f ? a : -a); -template -void BaseMatrixT::abs2(BaseMatrixT& b) { applyBinary(binary::Abs(), b); } +template +void BaseMatrixT::abs2(BaseMatrixT& b) { + applyBinary(binary::Abs(), b); +} DEFINE_MATRIX_BINARY_OP(AbsDerivative, a = (b > 0) ? a : (b < 0) ? -a : 0); -template +template void BaseMatrixT::absDerivative(BaseMatrixT& b) { applyBinary(binary::AbsDerivative(), b); } -DEFINE_MATRIX_BINARY_OP( - Sigmoid, const T THRESHOLD_MIN = -40.0; const T THRESHOLD_MAX = 13.0; - T tmp = (a < THRESHOLD_MIN) ? THRESHOLD_MIN - : ((a > THRESHOLD_MAX) ? THRESHOLD_MAX : a); - b = 1.0f / (1.0f + exp(-tmp))); -template<> +DEFINE_MATRIX_BINARY_OP(Sigmoid, const T THRESHOLD_MIN = -40.0; + const T THRESHOLD_MAX = 13.0; + T tmp = (a < THRESHOLD_MIN) + ? THRESHOLD_MIN + : ((a > THRESHOLD_MAX) ? THRESHOLD_MAX : a); + b = 1.0f / (1.0f + exp(-tmp))); +template <> void BaseMatrixT::sigmoid(BaseMatrixT& b) { if (useGpu_) { applyBinary(binary::Sigmoid(), b); @@ -723,31 +814,31 @@ void BaseMatrixT::sigmoid(BaseMatrixT& b) { } DEFINE_MATRIX_BINARY_OP(SigmoidDerivative, a *= b * (1 - b)); -template +template void BaseMatrixT::sigmoidDerivative(BaseMatrixT& b) { applyBinary(binary::SigmoidDerivative(), b); } DEFINE_MATRIX_BINARY_OP(ExpDerivative, a *= b); -template +template void BaseMatrixT::expDerivative(BaseMatrixT& b) { applyBinary(binary::ExpDerivative(), b); } DEFINE_MATRIX_BINARY_OP(Sign, b = a > 0.0f ? 1.0f : -1.0f); -template +template void BaseMatrixT::sign2(BaseMatrixT& b) { applyBinary(binary::Sign(), b); } DEFINE_MATRIX_BINARY_OP(Exp, a = exp(b)); -template<> +template <> void BaseMatrixT::exp2(BaseMatrixT& b) { applyBinary(binary::Exp(), b); } DEFINE_MATRIX_BINARY_OP(Log, a = log(b)); -template<> +template <> void BaseMatrixT::log2(BaseMatrixT& b) { if (useGpu_) { applyBinary(binary::Log(), b); @@ -757,13 +848,13 @@ void BaseMatrixT::log2(BaseMatrixT& b) { } DEFINE_MATRIX_BINARY_OP(Sqrt, a = sqrt(b)); -template<> +template <> void BaseMatrixT::sqrt2(BaseMatrixT& b) { applyBinary(binary::Sqrt(), b); } DEFINE_MATRIX_BINARY_OP(InvSqrt, a = 1.0f / sqrt(b)); -template<> +template <> void BaseMatrixT::invSqrt(BaseMatrixT& b) { if (useGpu_) { applyBinary(binary::InvSqrt(), b); @@ -775,37 +866,37 @@ void BaseMatrixT::invSqrt(BaseMatrixT& b) { } DEFINE_MATRIX_BINARY_PARAMETER_OP(IsEqual, ONE_PARAMETER, a = (b == p)); -template +template void BaseMatrixT::isEqualTo(BaseMatrixT& b, T value) { applyBinary(binary::IsEqual(value), b); } DEFINE_MATRIX_BINARY_PARAMETER_OP(AddScalar, ONE_PARAMETER, a = b + p); -template +template void BaseMatrixT::addScalar(BaseMatrixT& b, T p) { applyBinary(binary::AddScalar(p), b); } DEFINE_MATRIX_BINARY_PARAMETER_OP(SubScalar, ONE_PARAMETER, a = b - p); -template +template void BaseMatrixT::subScalar(BaseMatrixT& b, T p) { applyBinary(binary::SubScalar(p), b); } DEFINE_MATRIX_BINARY_PARAMETER_OP(MulScalar, ONE_PARAMETER, a = b * p); -template +template void BaseMatrixT::mulScalar(BaseMatrixT& b, T p) { applyBinary(binary::MulScalar(p), b); } DEFINE_MATRIX_BINARY_PARAMETER_OP(DivScalar, ONE_PARAMETER, a = b / p); -template +template void BaseMatrixT::divScalar(BaseMatrixT& b, T p) { applyBinary(binary::DivScalar(p), b); } DEFINE_MATRIX_BINARY_PARAMETER_OP(ScalarDiv, ONE_PARAMETER, a = p / b); -template +template void BaseMatrixT::scalarDiv(BaseMatrixT& b, T p) { applyBinary(binary::ScalarDiv(p), b); } @@ -817,20 +908,20 @@ void BaseMatrixT::scalarDiv(BaseMatrixT& b, T p) { DEFINE_MATRIX_TERNARY_OP(SoftCrossEntropy, a = -c * log(b) - (1 - c) * log(1 - b)); -template<> +template <> void BaseMatrixT::softCrossEntropy(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::SoftCrossEntropy(), b, c); } DEFINE_MATRIX_TERNARY_OP(SoftCrossEntropyBp, a += (b - c) / (b * (1 - b))); -template +template void BaseMatrixT::softCrossEntropyBp(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::SoftCrossEntropyBp(), b, c); } DEFINE_MATRIX_TERNARY_OP(BinaryCrossEntropy, a = c > 0.5 ? -log(b) : -log(1.0 - b)); -template<> +template <> void BaseMatrixT::binaryLabelCrossEntropy(BaseMatrixT& b, BaseMatrixT& c) { if (useGpu_) { @@ -858,70 +949,73 @@ void BaseMatrixT::binaryLabelCrossEntropy(BaseMatrixT& b, DEFINE_MATRIX_TERNARY_OP(BinaryCrossEntropyBp, a += c > 0.5 ? -1.0 / b : 1.0 / (1.0 - b)); -template +template void BaseMatrixT::binaryLabelCrossEntropyBp(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::BinaryCrossEntropyBp(), b, c); } DEFINE_MATRIX_TERNARY_OP(Add, a = b + c); -template +template void BaseMatrixT::add(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::Add(), b, c); } DEFINE_MATRIX_TERNARY_PARAMETER_OP(Add1, TWO_PARAMETER, a = p1 * b + p2 * c); -template +template void BaseMatrixT::add(BaseMatrixT& b, T p1, BaseMatrixT& c, T p2) { applyTernary(ternary::Add1(p1, p2), b, c); } DEFINE_MATRIX_TERNARY_OP(Sub, a = b - c); -template +template void BaseMatrixT::sub(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::Sub(), b, c); } DEFINE_MATRIX_TERNARY_PARAMETER_OP(Sub1, TWO_PARAMETER, a = p1 * b - p2 * c); -template +template void BaseMatrixT::sub(BaseMatrixT& b, T p1, BaseMatrixT& c, T p2) { applyTernary(ternary::Sub1(p1, p2), b, c); } DEFINE_MATRIX_TERNARY_OP(Add2, a = a + b + c); -template +template void BaseMatrixT::add2(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::Add2(), b, c); } -DEFINE_MATRIX_TERNARY_PARAMETER_OP(Add3, THREE_PARAMETER, +DEFINE_MATRIX_TERNARY_PARAMETER_OP(Add3, + THREE_PARAMETER, a = p1 * a + p2 * b + p3 * c); -template +template void BaseMatrixT::add2(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2, T p3) { applyTernary(ternary::Add3(p1, p2, p3), b, c); } -DEFINE_MATRIX_TERNARY_PARAMETER_OP(SgdUpdate, THREE_PARAMETER, +DEFINE_MATRIX_TERNARY_PARAMETER_OP(SgdUpdate, + THREE_PARAMETER, c = p2 * c - p1 * (b + p3 * a); a = a + c); -template +template void BaseMatrixT::sgdUpdate(BaseMatrixT& b, // grad BaseMatrixT& c, // mom - T p1, // learningRate, - T p2, // momentum, - T p3) { // decayRate + T p1, // learningRate, + T p2, // momentum, + T p3) { // decayRate applyTernary(ternary::SgdUpdate(p1, p2, p3), b, c); } -DEFINE_MATRIX_QUATERNARY_PARAMETER_OP(SgdUpdate, THREE_PARAMETER, +DEFINE_MATRIX_QUATERNARY_PARAMETER_OP(SgdUpdate, + THREE_PARAMETER, c = p2 * c - p1 * d * (b + p3 * a); a += c); -template +template void BaseMatrixT::sgdUpdate(BaseMatrixT& b, // grad, BaseMatrixT& c, // mom, BaseMatrixT& d, // lr, - T p1, // learningRate, - T p2, // momentum, - T p3) { // decayRate + T p1, // learningRate, + T p2, // momentum, + T p3) { // decayRate applyQuaternary(quaternary::SgdUpdate(p1, p2, p3), b, c, d); } @@ -929,19 +1023,22 @@ DEFINE_MATRIX_BINARY_PARAMETER_OP(ApplyL1, ONE_PARAMETER, T lambda = p * b; a = (a > lambda) ? (a - lambda) : (a < -lambda) ? (a + lambda) : 0); -template +template void BaseMatrixT::applyL1(BaseMatrixT& lr, T learningRate, T decayRate) { applyBinary(binary::ApplyL1(learningRate * decayRate), lr); } -template<> +template <> void BaseMatrixT::applyL1(BaseMatrixT& lr, real learningRate, real decayRate) { if (useGpu_) { applyBinary(binary::ApplyL1(learningRate * decayRate), lr); } else { - simd::decayL1(this->data_, this->data_, lr.data_, learningRate * decayRate, + simd::decayL1(this->data_, + this->data_, + lr.data_, + learningRate * decayRate, height_ * width_); } } @@ -950,24 +1047,25 @@ DEFINE_MATRIX_UNARY_PARAMETER_OP(ApplyL1, ONE_PARAMETER, T lambda = p; a = (a > lambda) ? (a - lambda) : (a < -lambda) ? (a + lambda) : 0); -template +template void BaseMatrixT::applyL1(T learningRate, T decayRate) { applyUnary(unary::ApplyL1(learningRate * decayRate)); } -template<> +template <> void BaseMatrixT::applyL1(real learningRate, real decayRate) { if (useGpu_) { applyUnary(unary::ApplyL1(learningRate * decayRate)); } else { - simd::decayL1(this->data_, this->data_, learningRate * decayRate, - height_ * width_); + simd::decayL1( + this->data_, this->data_, learningRate * decayRate, height_ * width_); } } -DEFINE_MATRIX_BINARY_PARAMETER_OP(ApplyL2, ONE_PARAMETER, +DEFINE_MATRIX_BINARY_PARAMETER_OP(ApplyL2, + ONE_PARAMETER, a *= (1.0f / (1.0f + p * b))); -template +template void BaseMatrixT::applyL2(BaseMatrixT& lr, T learningRate, T decayRate) { if (useGpu_) { applyBinary(binary::ApplyL2(learningRate * decayRate), lr); @@ -980,32 +1078,33 @@ void BaseMatrixT::applyL2(BaseMatrixT& lr, T learningRate, T decayRate) { } } -template +template void BaseMatrixT::applyL2(T learningRate, T decayRate) { BaseMatrixT::mulScalar(1.0f / (1.0f + learningRate * decayRate)); } DEFINE_MATRIX_BINARY_OP(DotMul, a *= b); -template +template void BaseMatrixT::dotMul(BaseMatrixT& b) { applyBinary(binary::DotMul(), b); } DEFINE_MATRIX_TERNARY_OP(DotMul, a = b * c); -template +template void BaseMatrixT::dotMul(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::DotMul(), b, c); } DEFINE_MATRIX_TERNARY_OP(DotDiv, a = (b == 0.0) ? 0.0 : b / c); -template +template void BaseMatrixT::dotDiv(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::DotDiv(), b, c); } -DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotDiv2P, TWO_PARAMETER, +DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotDiv2P, + TWO_PARAMETER, a = (b + p1) / (c + p2)); -template +template void BaseMatrixT::dotDiv(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2) { applyTernary(ternary::DotDiv2P(p1, p2), b, c); } @@ -1015,7 +1114,7 @@ DEFINE_MATRIX_QUATERNARY_OP(RankLoss, const T THRESHOLD = 40.0; a = b - c; ? THRESHOLD : ((a < -THRESHOLD) ? (-THRESHOLD) : a); a = log(1 + exp(a)) - a * d); -template<> +template <> void BaseMatrixT::rankLoss(BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d) { @@ -1026,8 +1125,9 @@ DEFINE_MATRIX_QUATERNARY_OP(RankLossBp, const T THRESHOLD = 40.0; a = b - c; a = (a > THRESHOLD) ? THRESHOLD : ((a < -THRESHOLD) ? (-THRESHOLD) : a); - a = exp(a); a = (a / (1 + a) - d)); -template<> + a = exp(a); + a = (a / (1 + a) - d)); +template <> void BaseMatrixT::rankLossBp(BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d) { @@ -1040,7 +1140,7 @@ DEFINE_MATRIX_TERNARY_OP(LogisticRegressionLoss, const T THRESHOLD = 40.0; ? -THRESHOLD : b; a = log(1 + exp(x)) - c * x); -template<> +template <> void BaseMatrixT::logisticRegressionLoss(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::LogisticRegressionLoss(), b, c); } @@ -1050,22 +1150,23 @@ DEFINE_MATRIX_TERNARY_OP(LogisticRegressionLossBp, const T THRESHOLD = 40.0; T x = (b > THRESHOLD) ? THRESHOLD : (b < -THRESHOLD) ? -THRESHOLD : b; - x = exp(x); a = x / (1 + x) - c); -template<> + x = exp(x); + a = x / (1 + x) - c); +template <> void BaseMatrixT::logisticRegressionLossBp(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::LogisticRegressionLossBp(), b, c); } DEFINE_MATRIX_TERNARY_OP(BiggerThan, a = (b > c) ? 1.0f : 0.0f); -template +template void BaseMatrixT::biggerThan(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::BiggerThan(), b, c); } DEFINE_MATRIX_QUATERNARY_OP( BiggerThan, a = ((b > c && d > 0.5f) || (b < c && d < 0.5f)) ? 1.0f : 0.0f); -template +template void BaseMatrixT::biggerThan(BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d) { @@ -1073,25 +1174,34 @@ void BaseMatrixT::biggerThan(BaseMatrixT& b, } DEFINE_MATRIX_TERNARY_OP(Max, a = (b > c) ? b : c); -template +template void BaseMatrixT::max2(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::Max(), b, c); } -DEFINE_MATRIX_TERNARY_PARAMETER_OP(BinaryClassificationError, ONE_PARAMETER, +DEFINE_MATRIX_TERNARY_PARAMETER_OP(BinaryClassificationError, + ONE_PARAMETER, c += ((a > p) == (b > p)) ? 0.0f : 1.0f); -template -void BaseMatrixT::binaryClassificationError2(size_t destCol, BaseMatrixT& b, - BaseMatrixT& c, T p) { +template +void BaseMatrixT::binaryClassificationError2(size_t destCol, + BaseMatrixT& b, + BaseMatrixT& c, + T p) { CHECK(!useGpu_) << "do not support gpu"; MatrixOffset offset(0, 0, 0, 0, destCol, 0); int numRows = b.height_; int numCols = b.width_; - b.applyTernary(ternary::BinaryClassificationError(p), c, *this, numRows, - numCols, offset, false_type(), true_type() /*cAsColVector*/); + b.applyTernary(ternary::BinaryClassificationError(p), + c, + *this, + numRows, + numCols, + offset, + false_type(), + true_type() /*cAsColVector*/); } -template<> +template <> void BaseMatrixT::binaryClassificationError(size_t destCol, BaseMatrixT& b, BaseMatrixT& c, @@ -1099,127 +1209,148 @@ void BaseMatrixT::binaryClassificationError(size_t destCol, MatrixOffset offset(destCol, 0, 0, 0, 0, 0); int numRows = b.height_; int numCols = b.width_; - aggregate(aggregate::sum(), base::binary::classificationError(p), - base::binary::add(), b, c, numRows, numCols, offset, false_type(), + aggregate(aggregate::sum(), + base::binary::classificationError(p), + base::binary::add(), + b, + c, + numRows, + numCols, + offset, + false_type(), true_type() /*aAsColVector*/); } -DEFINE_MATRIX_QUATERNARY_PARAMETER_OP(Add3, THREE_PARAMETER, +DEFINE_MATRIX_QUATERNARY_PARAMETER_OP(Add3, + THREE_PARAMETER, a = p1 * b + p2 * c + p3 * d); -template -void BaseMatrixT::add3(BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d, T p1, - T p2, T p3) { +template +void BaseMatrixT::add3( + BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT& d, T p1, T p2, T p3) { applyQuaternary(quaternary::Add3(p1, p2, p3), b, c, d); } DEFINE_MATRIX_TERNARY_OP(DotMulSquare, a = b * c * c); -template +template void BaseMatrixT::dotMulSquare(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::DotMulSquare(), b, c); } DEFINE_MATRIX_TERNARY_OP(DotSquareSquare, a = b * b * c * c); -template +template void BaseMatrixT::dotSquareSquare(BaseMatrixT& b, BaseMatrixT& c) { applyTernary(ternary::DotSquareSquare(), b, c); } DEFINE_MATRIX_BINARY_OP(DotMulSquare, a *= b * b); -template +template void BaseMatrixT::dotMulSquare(BaseMatrixT& b) { applyBinary(binary::DotMulSquare(), b); } DEFINE_MATRIX_BINARY_OP(DotSquareMul, a = a * a * b); -template +template void BaseMatrixT::dotSquareMul(BaseMatrixT& b) { applyBinary(binary::DotSquareMul(), b); } -DEFINE_MATRIX_QUATERNARY_PARAMETER_OP(AddSquareSum, THREE_PARAMETER, +DEFINE_MATRIX_QUATERNARY_PARAMETER_OP(AddSquareSum, + THREE_PARAMETER, T tmp = p1 * b + p2 * c + p3 * d; a += tmp * tmp); -template -void BaseMatrixT::addSquareSum(BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT d, - T p1, T p2, T p3) { +template +void BaseMatrixT::addSquareSum( + BaseMatrixT& b, BaseMatrixT& c, BaseMatrixT d, T p1, T p2, T p3) { applyQuaternary(quaternary::AddSquareSum(p1, p2, p3), b, c, d); } DEFINE_MATRIX_BINARY_PARAMETER_OP(AddSquare, ONE_PARAMETER, a += p * b * b); -template +template void BaseMatrixT::addSquare(BaseMatrixT& b, T p) { applyBinary(binary::AddSquare(p), b); } -DEFINE_MATRIX_BINARY_PARAMETER_OP(DecayAddSquare, TWO_PARAMETER, +DEFINE_MATRIX_BINARY_PARAMETER_OP(DecayAddSquare, + TWO_PARAMETER, a = p1 * a + p2 * b * b); -template +template void BaseMatrixT::decayAddSquare(BaseMatrixT& b, T p1, T p2) { applyBinary(binary::DecayAddSquare(p1, p2), b); } -DEFINE_MATRIX_TERNARY_PARAMETER_OP(DecayAddSquareMul, TWO_PARAMETER, +DEFINE_MATRIX_TERNARY_PARAMETER_OP(DecayAddSquareMul, + TWO_PARAMETER, a = p1 * a + p2 * b * b * c * c); -template -void BaseMatrixT::decayAddSquareMul(BaseMatrixT& b, BaseMatrixT& c, T p1, +template +void BaseMatrixT::decayAddSquareMul(BaseMatrixT& b, + BaseMatrixT& c, + T p1, T p2) { applyTernary(ternary::DecayAddSquareMul(p1, p2), b, c); } -DEFINE_MATRIX_TERNARY_PARAMETER_OP(ReciprocalSum, THREE_PARAMETER, +DEFINE_MATRIX_TERNARY_PARAMETER_OP(ReciprocalSum, + THREE_PARAMETER, a = 1 / (p1 * b + p2 * c + p3)); -template -void BaseMatrixT::reciprocalSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2, - T p3) { +template +void BaseMatrixT::reciprocalSum( + BaseMatrixT& b, BaseMatrixT& c, T p1, T p2, T p3) { applyTernary(ternary::ReciprocalSum(p1, p2, p3), b, c); } -DEFINE_MATRIX_BINARY_PARAMETER_OP(Reciprocal2, TWO_PARAMETER, +DEFINE_MATRIX_BINARY_PARAMETER_OP(Reciprocal2, + TWO_PARAMETER, a = 1 / (p1 * b + p2)); -template +template void BaseMatrixT::reciprocal2(BaseMatrixT& b, T p1, T p2) { applyBinary(binary::Reciprocal2(p1, p2), b); } -DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotMulSquareSum, TWO_PARAMETER, +DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotMulSquareSum, + TWO_PARAMETER, T tmp = p1 * b + p2 * c; a *= tmp * tmp); -template -void BaseMatrixT::dotMulSquareSum(BaseMatrixT& b, BaseMatrixT& c, T p1, +template +void BaseMatrixT::dotMulSquareSum(BaseMatrixT& b, + BaseMatrixT& c, + T p1, T p2) { applyTernary(ternary::DotMulSquareSum(p1, p2), b, c); } -DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotSquareSum, TWO_PARAMETER, +DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotSquareSum, + TWO_PARAMETER, T tmp = p1 * b + p2 * c; a = tmp * tmp); -template +template void BaseMatrixT::dotSquareSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2) { applyTernary(ternary::DotSquareSum(p1, p2), b, c); } -DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotMulSum, TWO_PARAMETER, +DEFINE_MATRIX_TERNARY_PARAMETER_OP(DotMulSum, + TWO_PARAMETER, a *= p1 * b + p2 * c); -template +template void BaseMatrixT::dotMulSum(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2) { applyTernary(ternary::DotMulSum(p1, p2), b, c); } DEFINE_MATRIX_BINARY_OP(CopyAndClear, b = a; a = 0); -template +template void BaseMatrixT::copyAndClear(BaseMatrixT& b) { applyBinary(binary::CopyAndClear(), b); } -DEFINE_MATRIX_TERNARY_PARAMETER_OP(AddDotMul, TWO_PARAMETER, +DEFINE_MATRIX_TERNARY_PARAMETER_OP(AddDotMul, + TWO_PARAMETER, a = p1 * a + p2 * b * c); -template +template void BaseMatrixT::addDotMul(BaseMatrixT& b, BaseMatrixT& c, T p1, T p2) { applyTernary(ternary::AddDotMul(p1, p2), b, c); } DEFINE_MATRIX_BINARY_OP(Assign, a = b;); -template +template void BaseMatrixT::assign(BaseMatrixT& b) { if (useGpu_) { applyBinary(binary::Assign(), b); @@ -1230,7 +1361,7 @@ void BaseMatrixT::assign(BaseMatrixT& b) { } } -template +template void BaseMatrixT::assignAtOffset(BaseMatrixT& b, int64_t columnOffset) { if (columnOffset + b.width_ <= width_) { int numRows = height_; @@ -1250,24 +1381,31 @@ void BaseMatrixT::assignAtOffset(BaseMatrixT& b, int64_t columnOffset) { } DEFINE_MATRIX_BINARY_OP(DeepSwap, T tmp = a; a = b; b = tmp); -template +template void BaseMatrixT::deepSwap(BaseMatrixT& b) { - applyBinary(binary::DeepSwap(), b); + applyBinary(binary::DeepSwap(), b); } -template<> +template <> void BaseMatrixT::rowDotMul(size_t destCol, BaseMatrixT& b, BaseMatrixT& c) { int numRows = b.height_; int numCols = b.width_; MatrixOffset offset(destCol, 0, 0, 0, 0, 0); - aggregate(aggregate::sum(), base::binary::mul(), base::binary::add(), b, c, - numRows, numCols, offset, false_type(), + aggregate(aggregate::sum(), + base::binary::mul(), + base::binary::add(), + b, + c, + numRows, + numCols, + offset, + false_type(), true_type() /*aAsColVector*/); } -template +template void BaseMatrixT::rowDotMul2(size_t destCol, BaseMatrixT& b, BaseMatrixT& c) { @@ -1290,17 +1428,24 @@ void BaseMatrixT::rowDotMul2(size_t destCol, } } -template<> +template <> void BaseMatrixT::addDotMulVMM(BaseMatrixT& b, BaseMatrixT& c) { MatrixOffset offset(0, 0, 0, 0, 0, 0); int numRows = b.height_; int numCols = b.width_; - aggregate(aggregate::sum(), base::binary::mul(), base::binary::add(), b, c, - numRows, numCols, offset, true_type() /*aAsRowVector*/, + aggregate(aggregate::sum(), + base::binary::mul(), + base::binary::add(), + b, + c, + numRows, + numCols, + offset, + true_type() /*aAsRowVector*/, false_type()); } -template +template void BaseMatrixT::addDotMulVMM2(BaseMatrixT& b, BaseMatrixT& c) { CHECK(!useGpu_) << "do not support gpu"; @@ -1321,16 +1466,22 @@ void BaseMatrixT::addDotMulVMM2(BaseMatrixT& b, BaseMatrixT& c) { } DEFINE_MATRIX_TERNARY_OP(addDotMulMMV, a += b * c); -template +template void BaseMatrixT::addDotMulMMV(BaseMatrixT& b, BaseMatrixT& c) { MatrixOffset offset(0, 0, 0, 0, 0, 0); int numRows = height_; int numCols = width_; - applyTernary(ternary::addDotMulMMV(), b, c, numRows, numCols, offset, - true_type() /*cAsRowVector*/, false_type()); + applyTernary(ternary::addDotMulMMV(), + b, + c, + numRows, + numCols, + offset, + true_type() /*cAsRowVector*/, + false_type()); } -template +template void BaseMatrixT::addDotMulMMV2(BaseMatrixT& b, BaseMatrixT& c) { CHECK(!useGpu_) << "do not support gpu"; @@ -1350,16 +1501,22 @@ void BaseMatrixT::addDotMulMMV2(BaseMatrixT& b, BaseMatrixT& c) { } } -template +template void BaseMatrixT::rowScale(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { MatrixOffset offset(0, 0, 0, 0, cCol, 0); int numRows = height_; int numCols = width_; - applyTernary(ternary::DotMul(), b, c, numRows, numCols, offset, - false_type(), true_type() /*cAsColVector*/); + applyTernary(ternary::DotMul(), + b, + c, + numRows, + numCols, + offset, + false_type(), + true_type() /*cAsColVector*/); } -template +template void BaseMatrixT::rowScale2(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { CHECK(!useGpu_) << "do not support gpu"; @@ -1379,52 +1536,82 @@ void BaseMatrixT::rowScale2(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { } } -template +template void BaseMatrixT::colScale(size_t cRow, BaseMatrixT& b, BaseMatrixT& c) { MatrixOffset offset(0, 0, 0, 0, 0, cRow); int numRows = height_; int numCols = width_; - applyTernary(ternary::DotMul(), b, c, numRows, numCols, offset, - true_type() /* cAsRowVector */, false_type() /* cAsColVector */); + applyTernary(ternary::DotMul(), + b, + c, + numRows, + numCols, + offset, + true_type() /* cAsRowVector */, + false_type() /* cAsColVector */); } -template +template void BaseMatrixT::addColScale(size_t cRow, BaseMatrixT& b, BaseMatrixT& c) { MatrixOffset offset(0, 0, 0, 0, 0, cRow); int numRows = height_; int numCols = width_; - applyTernary(ternary::addDotMulMMV(), b, c, numRows, numCols, offset, - true_type() /* cAsRowVector */, false_type() /* cAsColVector */); + applyTernary(ternary::addDotMulMMV(), + b, + c, + numRows, + numCols, + offset, + true_type() /* cAsRowVector */, + false_type() /* cAsColVector */); } -template +template void BaseMatrixT::addRowScale(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { MatrixOffset offset(0, 0, 0, 0, cCol, 0); int numRows = height_; int numCols = width_; - applyTernary(ternary::addDotMulMMV(), b, c, numRows, numCols, offset, - false_type(), true_type() /*cAsColVector*/); + applyTernary(ternary::addDotMulMMV(), + b, + c, + numRows, + numCols, + offset, + false_type(), + true_type() /*cAsColVector*/); } DEFINE_MATRIX_TERNARY_PARAMETER_OP(RowAdd, ONE_PARAMETER, a = b + p * c); -template +template void BaseMatrixT::rowAdd(size_t cCol, BaseMatrixT& b, BaseMatrixT& c, T p) { MatrixOffset offset(0, 0, 0, 0, cCol, 0); int numRows = height_; int numCols = width_; - applyTernary(ternary::RowAdd(p), b, c, numRows, numCols, offset, - false_type(), true_type() /*cAsColVector*/); + applyTernary(ternary::RowAdd(p), + b, + c, + numRows, + numCols, + offset, + false_type(), + true_type() /*cAsColVector*/); } DEFINE_MATRIX_TERNARY_OP(RowPow, a = pow(b, c)); -template<> +template <> void BaseMatrixT::rowPow(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { if (useGpu_) { MatrixOffset offset(0, 0, 0, 0, cCol, 0); int numRows = height_; int numCols = width_; - applyTernary(ternary::RowPow(), b, c, numRows, numCols, offset, - false_type(), true_type() /*cAsColVector*/); + applyTernary(ternary::RowPow(), + b, + c, + numRows, + numCols, + offset, + false_type(), + true_type() /*cAsColVector*/); } else { size_t height = this->height_; size_t width = this->width_; @@ -1441,44 +1628,64 @@ void BaseMatrixT::rowPow(size_t cCol, BaseMatrixT& b, BaseMatrixT& c) { } } -template +template void BaseMatrixT::mulRowVector(BaseMatrixT& b) { MatrixOffset offset(0, 0, 0, 0); int numRows = height_; int numCols = width_; - applyBinary(binary::DotMul(), b, numRows, numCols, offset, - true_type() /* bAsRowVector */, false_type()); + applyBinary(binary::DotMul(), + b, + numRows, + numCols, + offset, + true_type() /* bAsRowVector */, + false_type()); } DEFINE_MATRIX_BINARY_OP(DotDiv, a /= b); -template +template void BaseMatrixT::divRowVector(BaseMatrixT& b) { MatrixOffset offset(0, 0, 0, 0); int numRows = height_; int numCols = width_; - applyBinary(binary::DotDiv(), b, numRows, numCols, offset, - true_type() /* bAsRowVector */, false_type()); + applyBinary(binary::DotDiv(), + b, + numRows, + numCols, + offset, + true_type() /* bAsRowVector */, + false_type()); } -template +template void BaseMatrixT::mulColVector(BaseMatrixT& b) { MatrixOffset offset(0, 0, 0, 0); int numRows = height_; int numCols = width_; - applyBinary(binary::DotMul(), b, numRows, numCols, offset, - false_type(), true_type() /* bAsColVector */); + applyBinary(binary::DotMul(), + b, + numRows, + numCols, + offset, + false_type(), + true_type() /* bAsColVector */); } -template +template void BaseMatrixT::divColVector(BaseMatrixT& b) { MatrixOffset offset(0, 0, 0, 0); int numRows = height_; int numCols = width_; - applyBinary(binary::DotDiv(), b, numRows, numCols, offset, - false_type(), true_type() /* bAsColVector */); + applyBinary(binary::DotDiv(), + b, + numRows, + numCols, + offset, + false_type(), + true_type() /* bAsColVector */); } -template<> +template <> template int BaseMatrixT::applyRow(Agg agg, BaseMatrixT& b) { MatrixOffset offset(0, 0, 0, 0, 0, 0); @@ -1486,13 +1693,20 @@ int BaseMatrixT::applyRow(Agg agg, BaseMatrixT& b) { size_t numCols = b.width_; CHECK_EQ(height_, numRows); CHECK_EQ(width_, 1UL); - aggregate(agg, base::unary::identity(), base::binary::second(), b, numRows, - numCols, offset, false_type(), true_type() /*aAsColVector*/); + aggregate(agg, + base::unary::identity(), + base::binary::second(), + b, + numRows, + numCols, + offset, + false_type(), + true_type() /*aAsColVector*/); return 0; } -template<> +template <> template int BaseMatrixT::applyRow(Agg agg, Saver sv, BaseMatrixT& b) { MatrixOffset offset(0, 0, 0, 0, 0, 0); @@ -1500,16 +1714,25 @@ int BaseMatrixT::applyRow(Agg agg, Saver sv, BaseMatrixT& b) { size_t numCols = b.width_; CHECK_EQ(height_, numRows); CHECK_EQ(width_, 1UL); - aggregate(agg, base::unary::identity(), sv, b, numRows, numCols, offset, - false_type(), true_type() /*aAsColVector*/); + aggregate(agg, + base::unary::identity(), + sv, + b, + numRows, + numCols, + offset, + false_type(), + true_type() /*aAsColVector*/); return 0; } -template<> +template <> template -int BaseMatrixT::applyRow( - Agg agg, real scaleDest, real scaleAgg, BaseMatrixT& b) { +int BaseMatrixT::applyRow(Agg agg, + real scaleDest, + real scaleAgg, + BaseMatrixT& b) { if (scaleDest != 0) { applyRow(agg, base::binary::add2(scaleDest, scaleAgg), b); } else { @@ -1521,10 +1744,10 @@ int BaseMatrixT::applyRow( return 0; } -template<> +template <> template -int BaseMatrixT::applyRow(Agg agg, Op op, Saver sv, - BaseMatrixT& b, BaseMatrixT& c) { +int BaseMatrixT::applyRow( + Agg agg, Op op, Saver sv, BaseMatrixT& b, BaseMatrixT& c) { MatrixOffset offset(0, 0, 0, 0, 0, 0); size_t numRows = b.height_; size_t numCols = b.width_; @@ -1532,16 +1755,27 @@ int BaseMatrixT::applyRow(Agg agg, Op op, Saver sv, CHECK_EQ(width_, 1UL); CHECK_EQ(c.height_, numRows); CHECK_EQ(c.width_, numCols); - aggregate(agg, op, sv, - b, c, numRows, numCols, offset, - false_type(), true_type() /*aAsColVector*/); + aggregate(agg, + op, + sv, + b, + c, + numRows, + numCols, + offset, + false_type(), + true_type() /*aAsColVector*/); return 0; } -template<> +template <> template -int BaseMatrixT::applyRow(Agg agg, Op op, real scaleDest, real scaleAgg, - BaseMatrixT& b, BaseMatrixT& c) { +int BaseMatrixT::applyRow(Agg agg, + Op op, + real scaleDest, + real scaleAgg, + BaseMatrixT& b, + BaseMatrixT& c) { if (scaleDest != 0) { applyRow(agg, op, base::binary::add2(scaleDest, scaleAgg), b, c); } else { @@ -1553,7 +1787,7 @@ int BaseMatrixT::applyRow(Agg agg, Op op, real scaleDest, real scaleAgg, return 0; } -template<> +template <> template int BaseMatrixT::applyCol(Agg agg, BaseMatrixT& b) { MatrixOffset offset(0, 0, 0, 0, 0, 0); @@ -1561,13 +1795,20 @@ int BaseMatrixT::applyCol(Agg agg, BaseMatrixT& b) { size_t numCols = b.width_; CHECK_EQ(width_, numCols); CHECK_EQ(height_, 1UL); - aggregate(agg, base::unary::identity(), base::binary::second(), b, numRows, - numCols, offset, true_type() /*aAsRowVector*/, false_type()); + aggregate(agg, + base::unary::identity(), + base::binary::second(), + b, + numRows, + numCols, + offset, + true_type() /*aAsRowVector*/, + false_type()); return 0; } -template<> +template <> template int BaseMatrixT::applyCol(Agg agg, Saver sv, BaseMatrixT& b) { MatrixOffset offset(0, 0, 0, 0, 0, 0); @@ -1575,16 +1816,25 @@ int BaseMatrixT::applyCol(Agg agg, Saver sv, BaseMatrixT& b) { size_t numCols = b.width_; CHECK_EQ(width_, numCols); CHECK_EQ(height_, 1UL); - aggregate(agg, base::unary::identity(), sv, b, numRows, numCols, offset, - true_type() /*aAsRowVector*/, false_type()); + aggregate(agg, + base::unary::identity(), + sv, + b, + numRows, + numCols, + offset, + true_type() /*aAsRowVector*/, + false_type()); return 0; } -template<> +template <> template -int BaseMatrixT::applyCol( - Agg agg, real scaleDest, real scaleAgg, BaseMatrixT& b) { +int BaseMatrixT::applyCol(Agg agg, + real scaleDest, + real scaleAgg, + BaseMatrixT& b) { if (scaleDest != 0) { applyCol(agg, base::binary::add2(scaleDest, scaleAgg), b); } else { @@ -1596,48 +1846,51 @@ int BaseMatrixT::applyCol( return 0; } -template<> +template <> void BaseMatrixT::sumRows(BaseMatrixT& b, real scaleSum, real scaleDest) { applyRow(aggregate::sum(), scaleDest, scaleSum, b); } -template<> +template <> void BaseMatrixT::maxRows(BaseMatrixT& b) { applyRow(aggregate::max(), b); } -template<> +template <> void BaseMatrixT::minRows(BaseMatrixT& b) { applyRow(aggregate::min(), b); } -template<> +template <> void BaseMatrixT::maxCols(BaseMatrixT& b) { applyCol(aggregate::max(), b); } -template<> +template <> void BaseMatrixT::minCols(BaseMatrixT& b) { applyCol(aggregate::min(), b); } -template<> +template <> void BaseMatrixT::sumCols(BaseMatrixT& b, real scaleSum, real scaleDest) { applyCol(aggregate::sum(), scaleDest, scaleSum, b); } -template<> -void BaseMatrixT::sumOfSquaredDiffs( - BaseMatrixT& b, BaseMatrixT& c, real scaleSum, real scaleDest) { - applyRow(aggregate::sum(), base::binary::squaredDiff(), - scaleDest, scaleSum, b, c); +template <> +void BaseMatrixT::sumOfSquaredDiffs(BaseMatrixT& b, + BaseMatrixT& c, + real scaleSum, + real scaleDest) { + applyRow( + aggregate::sum(), base::binary::squaredDiff(), scaleDest, scaleSum, b, c); } -template<> -void BaseMatrixT::sumOfProducts( - BaseMatrixT& b, BaseMatrixT& c, real scaleSum, real scaleDest) { - applyRow(aggregate::sum(), base::binary::mul(), - scaleDest, scaleSum, b, c); +template <> +void BaseMatrixT::sumOfProducts(BaseMatrixT& b, + BaseMatrixT& c, + real scaleSum, + real scaleDest) { + applyRow(aggregate::sum(), base::binary::mul(), scaleDest, scaleSum, b, c); } template class BaseMatrixT; diff --git a/paddle/math/CMakeLists.txt b/paddle/math/CMakeLists.txt index 9981de61606bda6baac103592125b929d4c12a3d..bf28092e82b778dc904c5a2e271f76261cf5f6b6 100644 --- a/paddle/math/CMakeLists.txt +++ b/paddle/math/CMakeLists.txt @@ -15,13 +15,13 @@ file(GLOB MATH_HEADERS . *.h) file(GLOB MATH_SOURCES . *.cpp) set(MATH_SOURCES - "${PROJ_ROOT}/paddle/math/BaseMatrix.cu" - "${PROJ_ROOT}/paddle/math/TrainingAlgorithmOp.cu" + "${PADDLE_SOURCE_DIR}/paddle/math/BaseMatrix.cu" + "${PADDLE_SOURCE_DIR}/paddle/math/TrainingAlgorithmOp.cu" ${MATH_SOURCES}) if(NOT WITH_GPU) # then compile BaseMatrix.cu as c++ file - compile_cu_as_cpp("${PROJ_ROOT}/paddle/math/BaseMatrix.cu") - compile_cu_as_cpp("${PROJ_ROOT}/paddle/math/TrainingAlgorithmOp.cu") + compile_cu_as_cpp("${PADDLE_SOURCE_DIR}/paddle/math/BaseMatrix.cu") + compile_cu_as_cpp("${PADDLE_SOURCE_DIR}/paddle/math/TrainingAlgorithmOp.cu") add_library(paddle_math STATIC ${MATH_SOURCES}) else() diff --git a/paddle/math/CpuSparseMatrix.h b/paddle/math/CpuSparseMatrix.h index 860cad1047fc343b13efa901186ea218d0855151..36d57bbb65245de6b0de5909b55fbc4be3eccd78 100644 --- a/paddle/math/CpuSparseMatrix.h +++ b/paddle/math/CpuSparseMatrix.h @@ -302,6 +302,10 @@ public: bool isSparse() const { return true; } private: + using Matrix::mul; using Matrix::copyFrom; + using Matrix::rowMax; + using Matrix::print; + using Matrix::subMatrix; }; } // namespace paddle diff --git a/paddle/math/MathUtils.cpp b/paddle/math/MathUtils.cpp index 5bbc3e4e3725f186373072440a93f967178e0b27..980b6e138873046468f278c2f0b16938be82b81c 100644 --- a/paddle/math/MathUtils.cpp +++ b/paddle/math/MathUtils.cpp @@ -25,7 +25,7 @@ namespace paddle { */ void sparseRand( int* major, int* minor, int nnz, int majorLen, int minorMax, bool useGpu) { - CHECK(size_t(nnz) > size_t(1)); + CHECK(size_t(nnz) >= size_t(1)); int* cpuMajor; int* cpuMinor; CpuIVector cpuMinorVec(nnz); diff --git a/paddle/math/SparseMatrix.h b/paddle/math/SparseMatrix.h index f6cd5df338965b55ca17636de097d2401dc057f9..16300db081f89182faa82ea5798e8ec2f1cd93f9 100644 --- a/paddle/math/SparseMatrix.h +++ b/paddle/math/SparseMatrix.h @@ -231,6 +231,9 @@ public: private: using Matrix::mul; using Matrix::copyFrom; + using Matrix::rowMax; + using Matrix::print; + using Matrix::subMatrix; }; } // namespace paddle diff --git a/paddle/math/TrainingAlgorithmOp.cu b/paddle/math/TrainingAlgorithmOp.cu index 72ff077270382d52bfcd340cc64d9abf49d1705d..fc746b85339de596d5ddc5811a8164094c13f63f 100644 --- a/paddle/math/TrainingAlgorithmOp.cu +++ b/paddle/math/TrainingAlgorithmOp.cu @@ -12,9 +12,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/utils/Logging.h" #include "BaseMatrix.h" #include "TrainingAlgorithmOp.h" +#include "paddle/utils/Logging.h" #if __cplusplus > 199711L @@ -32,10 +32,10 @@ void sparseMomentumApply(BaseMatrix& value, real tau, real learningRate) { auto expr1 = momU.lazyAssign(momU - (alpha * gamma * learningRate) * grad); - auto expr2 = momV.lazyAssign( - momV + (tau * alpha * gamma * learningRate) * grad); - auto expr3 = value.lazyAssign( - (tau / beta + (real)1 / alpha) * momU + ((real)1 / beta) * momV); + auto expr2 = + momV.lazyAssign(momV + (tau * alpha * gamma * learningRate) * grad); + auto expr3 = value.lazyAssign((tau / beta + (real)1 / alpha) * momU + + ((real)1 / beta) * momV); AssignEvaluate(expr1, expr2, expr3); } @@ -52,12 +52,12 @@ void adadeltaApply(BaseMatrix& value, real momentum, real decayRate) { auto expr1 = accum.lazyAssign(rou * accum + ((real)1 - rou) * grad.square()); - auto expr2 = lr.lazyAssign( - ((accum_update + epsilon) / (accum + epsilon)).sqrt()); - auto expr3 = accum_update.lazyAssign( - rou * accum_update + ((real)1 - rou) * (grad * lr).square()); - auto expr4 = mom.lazyAssign( - mom * momentum - learningRate * lr * (grad + value * decayRate)); + auto expr2 = + lr.lazyAssign(((accum_update + epsilon) / (accum + epsilon)).sqrt()); + auto expr3 = accum_update.lazyAssign(rou * accum_update + + ((real)1 - rou) * (grad * lr).square()); + auto expr4 = mom.lazyAssign(mom * momentum - + learningRate * lr * (grad + value * decayRate)); auto expr5 = value.lazyAssign(value + mom); AssignEvaluate(expr1, expr2, expr3, expr4, expr5); @@ -74,10 +74,10 @@ void adagradApply(BaseMatrix& value, real momentum, real decayRate) { auto expr1 = accum.lazyAssign(accum + grad.square()); - auto expr2 = lr.lazyAssign( - (accum_buffer + accum + epsilon).sqrt().reciprocal()); - auto expr3 = mom.lazyAssign( - mom * momentum - learningRate * lr * (grad + value * decayRate)); + auto expr2 = + lr.lazyAssign((accum_buffer + accum + epsilon).sqrt().reciprocal()); + auto expr3 = mom.lazyAssign(mom * momentum - + learningRate * lr * (grad + value * decayRate)); auto expr4 = value.lazyAssign(value + mom); AssignEvaluate(expr1, expr2, expr3, expr4); @@ -98,8 +98,8 @@ void rmspropApply(BaseMatrix& value, bool firstTime) { auto expr2 = f.lazyAssign(accumulatedRou * f + ((real)1 - rou) * grad); auto expr3 = lr.lazyAssign((g - f.square() + epsilon).sqrt().reciprocal()); - auto expr4 = mom.lazyAssign( - mom * momentum - learningRate * lr * (grad + value * decayRate)); + auto expr4 = mom.lazyAssign(mom * momentum - + learningRate * lr * (grad + value * decayRate)); auto expr5 = value.lazyAssign(value + mom); if (firstTime) { @@ -107,8 +107,8 @@ void rmspropApply(BaseMatrix& value, AssignEvaluate(expr1, expr2, expr3, expr4, expr5); } else { - auto expr1 = g.lazyAssign( - accumulatedRou * g + ((real)1 - rou) * grad.square()); + auto expr1 = + g.lazyAssign(accumulatedRou * g + ((real)1 - rou) * grad.square()); AssignEvaluate(expr1, expr2, expr3, expr4, expr5); } @@ -127,8 +127,8 @@ void decayedAdagradApply(BaseMatrix& value, real decayRate, bool firstTime) { auto expr2 = lr.lazyAssign((accum + epsilon).sqrt().reciprocal()); - auto expr3 = mom.lazyAssign( - mom * momentum - learningRate * lr * (grad + value * decayRate)); + auto expr3 = mom.lazyAssign(mom * momentum - + learningRate * lr * (grad + value * decayRate)); auto expr4 = value.lazyAssign(value + mom); if (firstTime) { @@ -136,8 +136,8 @@ void decayedAdagradApply(BaseMatrix& value, AssignEvaluate(expr1, expr2, expr3, expr4); } else { - auto expr1 = accum.lazyAssign( - accumulatedRou * accum + ((real)1 - rou) * grad.square()); + auto expr1 = accum.lazyAssign(accumulatedRou * accum + + ((real)1 - rou) * grad.square()); AssignEvaluate(expr1, expr2, expr3, expr4); } @@ -153,13 +153,12 @@ void adamApply(BaseMatrix& value, real beta2_power, real epsilon, real learningRate) { - real alpha = learningRate * - std::sqrt((real)1 - beta2_power) / ((real)1 - beta1_power); + real alpha = + learningRate * std::sqrt((real)1 - beta2_power) / ((real)1 - beta1_power); auto expr1 = mom.lazyAssign(beta1 * mom + ((real)1 - beta1) * grad); auto expr2 = v.lazyAssign(beta2 * v + ((real)1 - beta2) * grad.square()); - auto expr3 = value.lazyAssign( - value - (mom * alpha) / (v.sqrt() + epsilon)); + auto expr3 = value.lazyAssign(value - (mom * alpha) / (v.sqrt() + epsilon)); AssignEvaluate(expr1, expr2, expr3); } @@ -173,10 +172,10 @@ void adamaxApply(BaseMatrix& value, int64_t step, real alpha) { auto expr1 = mom.lazyAssign(beta1 * mom + ((real)1 - beta1) * grad); - auto expr2 = u.lazyAssign( - (beta2 * u > grad.abs()).condition(beta2 * u, grad.abs())); + auto expr2 = + u.lazyAssign((beta2 * u > grad.abs()).condition(beta2 * u, grad.abs())); auto expr3 = value.lazyAssign( - value - (alpha / ((real)1 - (real)std::pow(beta1, step))) * (mom / u)); + value - (alpha / ((real)1 - (real)std::pow(beta1, step))) * (mom / u)); AssignEvaluate(expr1, expr2, expr3); } @@ -322,8 +321,8 @@ void adamApply(BaseMatrix& value, real beta2_power, real epsilon, real learningRate) { - real alpha = learningRate * - std::sqrt((real)1 - beta2_power) / ((real)1 - beta1_power); + real alpha = + learningRate * std::sqrt((real)1 - beta2_power) / ((real)1 - beta1_power); // m_t = \beta_1 * m_{t-1} + (1-\beta_1)* g_t; mom = beta1 * mom + ((real)1 - beta1) * grad; @@ -331,7 +330,7 @@ void adamApply(BaseMatrix& value, // v_t = \beta_2 * v_{t-1} + (1-\beta_2)* g_{t-1}^2 v = beta2 * v + ((real)1 - beta2) * grad.square(); - value -= (mom * alpha) / (v.sqrt() + epsilon); + value -= (mom * alpha) / (v.sqrt() + epsilon); } void adamaxApply(BaseMatrix& value, diff --git a/paddle/math/tests/test_Tensor.cu b/paddle/math/tests/test_Tensor.cu index 40e38434fa328bba8be6e1b8e509023d615899c1..31b693afa8bd50f77a8efb67769e6215dd755bd3 100644 --- a/paddle/math/tests/test_Tensor.cu +++ b/paddle/math/tests/test_Tensor.cu @@ -13,8 +13,8 @@ See the License for the specific language governing permissions and limitations under the License. */ #include -#include "paddle/math/Matrix.h" #include "TensorCheck.h" +#include "paddle/math/Matrix.h" using paddle::Matrix; using paddle::CpuMatrix; @@ -26,25 +26,25 @@ using paddle::GpuIVector; using autotest::TensorCheckEqual; using autotest::TensorCheckErr; -#define INIT_UNARY(A1, A2) \ - Tensor A1(height, width); \ - Tensor A2(height, width); \ - A1.randomizeUniform(); \ - A2.copyFrom(A1) -#define INIT_BINARY(A1, A2, B) \ - INIT_UNARY(A1, A2); \ - Tensor B(height, width); \ - B.randomizeUniform() -#define INIT_TERNARY(A1, A2, B, C) \ - INIT_BINARY(A1, A2, B); \ - Tensor C(height, width); \ - C.randomizeUniform() -#define INIT_QUATERNARY(A1, A2, B, C, D) \ - INIT_TERNARY(A1, A2, B, C); \ - Tensor D(height, width); \ - D.randomizeUniform() - -template +#define INIT_UNARY(A1, A2) \ + Tensor A1(height, width); \ + Tensor A2(height, width); \ + A1.randomizeUniform(); \ + A2.copyFrom(A1) +#define INIT_BINARY(A1, A2, B) \ + INIT_UNARY(A1, A2); \ + Tensor B(height, width); \ + B.randomizeUniform() +#define INIT_TERNARY(A1, A2, B, C) \ + INIT_BINARY(A1, A2, B); \ + Tensor C(height, width); \ + C.randomizeUniform() +#define INIT_QUATERNARY(A1, A2, B, C, D) \ + INIT_TERNARY(A1, A2, B, C); \ + Tensor D(height, width); \ + D.randomizeUniform() + +template struct TestUnaryMatrix { typedef std::function UnaryFunc; @@ -59,7 +59,7 @@ struct TestUnaryMatrix { } }; -template +template struct TestBinaryMatrix { typedef std::function BinaryFunc; @@ -74,10 +74,10 @@ struct TestBinaryMatrix { } }; -template +template struct TestTernaryMatrix { - typedef std::function TernaryFunc; + typedef std::function + TernaryFunc; explicit TestTernaryMatrix(TernaryFunc testTernaryFunc) { for (auto height : {1, 11, 73, 128, 200, 330}) { @@ -90,10 +90,11 @@ struct TestTernaryMatrix { } }; -template +template struct TestQuaternaryMatrix { typedef std::function QuaternaryFunc; + Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D)> + QuaternaryFunc; explicit TestQuaternaryMatrix(QuaternaryFunc testQuaternaryFunc) { for (auto height : {1, 11, 73, 128, 200, 330}) { @@ -106,7 +107,7 @@ struct TestQuaternaryMatrix { } }; -template +template struct TestUnaryVectorT { typedef std::function UnaryFunc; @@ -142,11 +143,11 @@ void SetTensorValue(Matrix& matrix, real value) { } } -template +template void testTensorAddScalar(Tensor& A1, Tensor& A2) { real p1 = 2.5; real p2 = 3.0; - A1.add(p1); // a += p + A1.add(p1); // a += p A2 += p1; TensorCheckEqual(A1, A2); @@ -155,7 +156,7 @@ void testTensorAddScalar(Tensor& A1, Tensor& A2) { TensorCheckEqual(A1, A2); } -template +template void testTensorSubScalar(Tensor& A1, Tensor& A2) { real p = 2.5; A1.subScalar(p); // a -= p @@ -163,7 +164,7 @@ void testTensorSubScalar(Tensor& A1, Tensor& A2) { TensorCheckEqual(A1, A2); } -template +template void testTensorMulScalar(Tensor& A1, Tensor& A2) { real p = 2.5; A1.mulScalar(p); // a *= p @@ -177,7 +178,7 @@ void testTensorMulScalar(Tensor& A1, Tensor& A2) { TensorCheckEqual(A1, A2); } -template +template void testTensorDivScalar(Tensor& A1, Tensor& A2) { real p = 2.5; A1.divScalar(p); // a /= p @@ -185,44 +186,44 @@ void testTensorDivScalar(Tensor& A1, Tensor& A2) { TensorCheckEqual(A1, A2); } -template +template void testTensorNeg(Tensor& A1, Tensor& A2) { A1.neg(); // a = -a A2 = -A2; TensorCheckEqual(A1, A2); } -template +template void testTensorAbs(Tensor& A1, Tensor& A2) { A1.abs2(); // a = a > 0 ? a : -a A2 = A2.abs(); TensorCheckEqual(A1, A2); } -template +template void testTensorSquare(Tensor& A1, Tensor& A2) { A1.square2(); // a = a * a A2 = A2.square(); TensorCheckEqual(A1, A2); } -template +template void testTensorReciprocal(Tensor& A1, Tensor& A2) { A1.reciprocal2(); // a = 1.0f / a A2 = A2.reciprocal(); TensorCheckEqual(A1, A2); } -template +template void testTensorSign(Tensor& A1, Tensor& A2) { A1.sign2(); // a = (a > 0) - (a < 0) A2 = A2.sign(); TensorCheckEqual(A1, A2); } -template +template void testTensorAssign(Tensor& A1, Tensor& A2) { - A1.assign(1.5); // a = p + A1.assign(1.5); // a = p A2 = A2.constant(1.5); TensorCheckEqual(A1, A2); @@ -235,7 +236,7 @@ void testTensorAssign(Tensor& A1, Tensor& A2) { TensorCheckEqual(A1, A2); } -template +template void testUnaryBaseOp(Tensor& A1, Tensor& A2) { testTensorAddScalar(A1, A2); testTensorSubScalar(A1, A2); @@ -249,9 +250,9 @@ void testUnaryBaseOp(Tensor& A1, Tensor& A2) { testTensorAssign(A1, A2); } -template +template void testUnaryBaseOpInt(Tensor& A1, Tensor& A2) { - A1.add(2); // a += p + A1.add(2); // a += p A2 += 2; TensorCheckEqual(A1, A2); @@ -266,46 +267,46 @@ void testUnaryBaseOpInt(Tensor& A1, Tensor& A2) { TEST(Unary, BaseOp) { TestUnaryMatrix testCpuMatrix(testUnaryBaseOp); TestUnaryVectorT testCpuVector(testUnaryBaseOp); - TestUnaryVectorT - testCpuIVector(testUnaryBaseOpInt); + TestUnaryVectorT testCpuIVector( + testUnaryBaseOpInt); #ifndef PADDLE_ONLY_CPU TestUnaryMatrix testGpuMatrix(testUnaryBaseOp); TestUnaryVectorT testGpuVector(testUnaryBaseOp); - TestUnaryVectorT - testGpuIVector(testUnaryBaseOpInt); + TestUnaryVectorT testGpuIVector( + testUnaryBaseOpInt); #endif } -template +template void testTensorExp(Tensor& A1, Tensor& A2) { A1.exp2(); // a = exp(a) A2 = A2.exp(); TensorCheckErr(A1, A2); } -template +template void testTensorLog(Tensor& A1, Tensor& A2) { A1.log2(); // a = log(a) A2 = A2.log(); TensorCheckErr(A1, A2); } -template +template void testTensorSqrt(Tensor& A1, Tensor& A2) { A1.sqrt2(); // a = sqrt(a) A2 = A2.sqrt(); TensorCheckErr(A1, A2); } -template +template void testTensorPow(Tensor& A1, Tensor& A2) { A1.pow2(3.2); // a = pow(a, p) A2 = A2.pow(3.2); TensorCheckErr(A1, A2); } -template +template void testUnayrMathOp(Tensor& A1, Tensor& A2) { testTensorExp(A1, A2); testTensorLog(A1, A2); @@ -321,7 +322,7 @@ TEST(Unary, MathOp) { #endif } -template +template void testTensorClip(Tensor& A1, Tensor& A2) { real p1 = 0.003f; real p2 = 0.877f; @@ -331,7 +332,7 @@ void testTensorClip(Tensor& A1, Tensor& A2) { TensorCheckEqual(A1, A2); } -template +template void testTensorBiggerThanScalar(Tensor& A1, Tensor& A2) { real p = 0.5f; A1.biggerThanScalar(p); // a = a > p ? 1.0f : 0.0f @@ -339,7 +340,7 @@ void testTensorBiggerThanScalar(Tensor& A1, Tensor& A2) { TensorCheckEqual(A1, A2); } -template +template void testTensorapplyL1(Tensor& A1, Tensor& A2) { /** * T lambda = p; @@ -351,14 +352,15 @@ void testTensorapplyL1(Tensor& A1, Tensor& A2) { real learningRate = 0.7f; real decayRate = 0.6f; A1.applyL1(learningRate, decayRate); - A2 = (A2 > (learningRate * decayRate)).condition( - (A2 - (learningRate * decayRate)), - (A2 < -(learningRate * decayRate)).condition( - (A2 + (learningRate * decayRate)), (real)0.0)); + A2 = (A2 > (learningRate * decayRate)) + .condition( + (A2 - (learningRate * decayRate)), + (A2 < -(learningRate * decayRate)) + .condition((A2 + (learningRate * decayRate)), (real)0.0)); TensorCheckEqual(A1, A2); } -template +template void testUnayrCompareOp(Tensor& A1, Tensor& A2) { testTensorClip(A1, A2); testTensorBiggerThanScalar(A1, A2); @@ -377,7 +379,7 @@ TEST(Unary, CompareOp) { #endif } -template +template void testTensorAdd(Tensor& A1, Tensor& A2, Tensor& B) { real p1 = 2.5; real p2 = 3.2; @@ -406,7 +408,7 @@ void testTensorAdd(Tensor& A1, Tensor& A2, Tensor& B) { TensorCheckEqual(A1, A2); } -template +template void testTensorSub(Tensor& A1, Tensor& A2, Tensor& B) { real p = 2.5; A1.sub(B); // a -= b @@ -422,7 +424,7 @@ void testTensorSub(Tensor& A1, Tensor& A2, Tensor& B) { TensorCheckEqual(A1, A2); } -template +template void testTensorMul(Tensor& A1, Tensor& A2, Tensor& B) { real p = 2.5; A1.mulScalar(B, p); // a = b * p @@ -442,7 +444,7 @@ void testTensorMul(Tensor& A1, Tensor& A2, Tensor& B) { TensorCheckEqual(A1, A2); } -template +template void testTensorDiv(Tensor& A1, Tensor& A2, Tensor& B) { real p = 2.5; A1.divScalar(B, p); // a = b / p @@ -454,28 +456,28 @@ void testTensorDiv(Tensor& A1, Tensor& A2, Tensor& B) { TensorCheckEqual(A1, A2); } -template +template void testTensorAssign(Tensor& A1, Tensor& A2, Tensor& B) { A1.assign(B); // a = b A2 = B; TensorCheckEqual(A1, A2); } -template +template void testTensorSquare(Tensor& A1, Tensor& A2, Tensor& B) { - B.square2(A1); // b = a * a + B.square2(A1); // b = a * a A2 = B.square(); TensorCheckEqual(A1, A2); } -template +template void testTensorSquareDerivative(Tensor& A1, Tensor& A2, Tensor& B) { A1.squareDerivative(B); // a *= 2.0 * b A2 = A2 * (real)2.0 * B; TensorCheckEqual(A1, A2); } -template +template void testTensorReciprocal(Tensor& A1, Tensor& A2, Tensor& B) { B.reciprocal2(A1); // b = 1.0f / a A2 = B.reciprocal(); @@ -490,33 +492,33 @@ void testTensorReciprocal(Tensor& A1, Tensor& A2, Tensor& B) { real learningRate = 0.7f; real decayRate = 1.2f; A1.applyL2(B, learningRate, decayRate); // a *= (1.0f / (1.0f + p * b)) - A2 *= (B.constant(1.0f) + - B.constant(learningRate * decayRate) * B).reciprocal(); + A2 *= (B.constant(1.0f) + B.constant(learningRate * decayRate) * B) + .reciprocal(); TensorCheckEqual(A1, A2); } -template +template void testTensorReciprocalDerivative(Tensor& A1, Tensor& A2, Tensor& B) { A1.reciprocalDerivative(B); // a *= -b * b A2 *= (-B) * B; TensorCheckEqual(A1, A2); } -template +template void testTensorSign(Tensor& A1, Tensor& A2, Tensor& B) { B.sign2(A1); // b = a > 0.0f ? 1.0f : -1.0f A2 = B.sign(); TensorCheckEqual(A1, A2); } -template +template void testTensorAbs(Tensor& A1, Tensor& A2, Tensor& B) { B.abs2(A1); // b = a > 0.0f ? a : -a A2 = B.abs(); TensorCheckEqual(A1, A2); } -template +template void testBinaryBaseOp(Tensor& A1, Tensor& A2, Tensor& B) { testTensorAdd(A1, A2, B); testTensorSub(A1, A2, B); @@ -539,7 +541,7 @@ TEST(Binary, BaseOp) { #endif } -template +template void testTensorExp(Tensor& A1, Tensor& A2, Tensor& B) { // a = exp(b) A1.exp2(B); @@ -547,14 +549,14 @@ void testTensorExp(Tensor& A1, Tensor& A2, Tensor& B) { TensorCheckErr(A1, A2); } -template +template void testTensorExpDerivative(Tensor& A1, Tensor& A2, Tensor& B) { A1.expDerivative(B); // a *= b A2 *= B; TensorCheckEqual(A1, A2); } -template +template void testTensorLog(Tensor& A1, Tensor& A2, Tensor& B) { // a = log(b) A1.log2(B); @@ -562,7 +564,7 @@ void testTensorLog(Tensor& A1, Tensor& A2, Tensor& B) { TensorCheckErr(A1, A2); } -template +template void testTensorSqrt(Tensor& A1, Tensor& A2, Tensor& B) { // a = sqrt(b) A1.sqrt2(B); @@ -570,7 +572,7 @@ void testTensorSqrt(Tensor& A1, Tensor& A2, Tensor& B) { TensorCheckErr(A1, A2); } -template +template void testTensorInvSqrt(Tensor& A1, Tensor& A2, Tensor& B) { // a = 1.0f / sqrt(b) A1.invSqrt(B); @@ -578,14 +580,14 @@ void testTensorInvSqrt(Tensor& A1, Tensor& A2, Tensor& B) { TensorCheckErr(A1, A2); } -template +template void testTensorPow(Tensor& A1, Tensor& A2, Tensor& B) { A1.pow2(B, 2.5f); // a = pow(b, p) A2 = B.pow(2.5f); TensorCheckErr(A1, A2); } -template +template void testTensorSoftrelu(Tensor& A1, Tensor& A2, Tensor& B) { /* * const T THRESHOLD = 40.0; @@ -597,12 +599,14 @@ void testTensorSoftrelu(Tensor& A1, Tensor& A2, Tensor& B) { real THRESHOLD = 40.0; A2 = (B.constant(1.0f) + - (B > THRESHOLD).condition( - THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B)).exp()).log(); + (B > THRESHOLD) + .condition(THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B)) + .exp()) + .log(); TensorCheckErr(A1, A2); } -template +template void testTensorSoftreluDerivative(Tensor& A1, Tensor& A2, Tensor& B) { /* * const T THRESHOLD = 40.0; @@ -612,14 +616,16 @@ void testTensorSoftreluDerivative(Tensor& A1, Tensor& A2, Tensor& B) { */ A1.softreluDerivative(B); real THRESHOLD = 40.0; - A2 = A2 * (B.constant(1.0f) - - (B.constant(-1.0f) * - (B > THRESHOLD).condition( - THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B))).exp()); + A2 = A2 * + (B.constant(1.0f) - + (B.constant(-1.0f) * + (B > THRESHOLD) + .condition(THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B))) + .exp()); TensorCheckErr(A1, A2); } -template +template void testTensorSigmoid(Tensor& A1, Tensor& A2, Tensor& B) { /* const T THRESHOLD_MIN = -40.0; @@ -632,46 +638,47 @@ void testTensorSigmoid(Tensor& A1, Tensor& A2, Tensor& B) { const real THRESHOLD_MIN = -40.0; const real THRESHOLD_MAX = 13.0; - auto tmp = (B < THRESHOLD_MIN).condition( - THRESHOLD_MIN, (B > THRESHOLD_MAX).condition(THRESHOLD_MAX, B)); + auto tmp = (B < THRESHOLD_MIN) + .condition(THRESHOLD_MIN, + (B > THRESHOLD_MAX).condition(THRESHOLD_MAX, B)); A2 = (B.constant(1.0f) + (-tmp).exp()).reciprocal(); TensorCheckErr(A1, A2); } -template +template void testTensorSigmoidDerivative(Tensor& A1, Tensor& A2, Tensor& B) { A1.sigmoidDerivative(B); // a *= b * (1 - b) A2 *= B * (B.constant(1.0f) - B); TensorCheckEqual(A1, A2); } -template +template void testTensorTanh(Tensor& A1, Tensor& A2, Tensor& B) { B.tanh(A1); // b = 2.0 / (1.0 + exp(-2 * a)) - 1.0 A2 = B.constant(2.0f) / ((B * ((real)-2.0f)).exp() + (real)1.0f) - (real)1.0f; TensorCheckErr(A1, A2); } -template +template void testTensorTanhDerivative(Tensor& A1, Tensor& A2, Tensor& B) { A1.tanhDerivative(B); // a *= 1 - b * b A2 *= B.constant(1.0f) - B * B; TensorCheckEqual(A1, A2); } -template +template void testTensorScaledTanh(Tensor& A1, Tensor& A2, Tensor& B) { real p1 = 2.5; real p2 = 3.1; // b = p1 * (2.0 / (1.0 + exp(-2 * p2 * a)) - 1.0) B.scaledTanh(A1, p1, p2); A2 = B.constant(p1) * - (B.constant(2.0f) / ((B.constant(-2.0f) * p2 * B).exp() + (real)1.0) - - (real)1.0); + (B.constant(2.0f) / ((B.constant(-2.0f) * p2 * B).exp() + (real)1.0) - + (real)1.0); TensorCheckErr(A1, A2); } -template +template void testTensorScaledTanhDerivative(Tensor& A1, Tensor& A2, Tensor& B) { real p1 = 2.5; real p2 = 3.1; @@ -681,7 +688,7 @@ void testTensorScaledTanhDerivative(Tensor& A1, Tensor& A2, Tensor& B) { TensorCheckEqual(A1, A2); } -template +template void testBinaryMathOp(Tensor& A1, Tensor& A2, Tensor& B) { testTensorTanhDerivative(A1, A2, B); testTensorScaledTanhDerivative(A1, A2, B); @@ -708,21 +715,21 @@ TEST(Binary, MathOp) { #endif } -template +template void testTensorRelu(Tensor& A1, Tensor& A2, Tensor& B) { B.relu(A1); // b = a > 0.0f ? a : 0.0f A2 = (B > (real)0.0f).condition(B, (real)0.0f); TensorCheckEqual(A1, A2); } -template +template void testTensorReluDerivative(Tensor& A1, Tensor& A2, Tensor& B) { A1.reluDerivative(B); // a *= (b > 0.0f ? 1.0f : 0.0f) A2 *= (B > (real)0.0).condition((real)1.0, (real)0.0); TensorCheckEqual(A1, A2); } -template +template void testTensorBrelu(Tensor& A1, Tensor& A2, Tensor& B) { /* * b = a > p1 ? a : p1 @@ -736,7 +743,7 @@ void testTensorBrelu(Tensor& A1, Tensor& A2, Tensor& B) { TensorCheckEqual(A1, A2); } -template +template void testTensorBreluDerivative(Tensor& A1, Tensor& A2, Tensor& B) { SetTensorValue(B, 32.0f); /* @@ -748,15 +755,15 @@ void testTensorBreluDerivative(Tensor& A1, Tensor& A2, Tensor& B) { TensorCheckEqual(A1, A2); } -template +template void testTensorAbsDerivative(Tensor& A1, Tensor& A2, Tensor& B) { A1.absDerivative(B); // a = (b > 0) ? a : (b < 0) ? -a : 0 - A2 = (B > (real)0.0f).condition(A2, - (B < (real)0.0f).condition(-A2, (real)0.0f)); + A2 = (B > (real)0.0f) + .condition(A2, (B < (real)0.0f).condition(-A2, (real)0.0f)); TensorCheckEqual(A1, A2); } -template +template void testTensorIsEqualTo(Tensor& A1, Tensor& A2, Tensor& B) { real p = 0.613; SetTensorValue(B, p); @@ -765,7 +772,7 @@ void testTensorIsEqualTo(Tensor& A1, Tensor& A2, Tensor& B) { TensorCheckEqual(A1, A2); } -template +template void testTensorapplyL1(Tensor& A1, Tensor& A2, Tensor& B) { /** * T lambda = p * b; @@ -778,12 +785,13 @@ void testTensorapplyL1(Tensor& A1, Tensor& A2, Tensor& B) { real decayRate = 0.6f; A1.applyL1(B, learningRate, decayRate); auto lambda = B.constant(learningRate * decayRate) * B; - A2 = (A2 > lambda).condition( - (A2 - lambda), (A2 < -lambda).condition((A2 + lambda), (real)0.0f)); + A2 = (A2 > lambda) + .condition((A2 - lambda), + (A2 < -lambda).condition((A2 + lambda), (real)0.0f)); TensorCheckEqual(A1, A2); } -template +template void testBinaryCompareOp(Tensor& A1, Tensor& A2, Tensor& B) { B.subScalar(0.5f); SetTensorValue(B, 0.0f); @@ -807,7 +815,7 @@ TEST(Binary, CompareOp) { #endif } -template +template void testTensorAdd(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { A1.add(B, C); // a = b + c A2 = B + C; @@ -833,7 +841,7 @@ void testTensorAdd(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { TensorCheckEqual(A1, A2); } -template +template void testTensorSub(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { A1.sub(B, C); // a = b - c A2 = B - C; @@ -846,7 +854,7 @@ void testTensorSub(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { TensorCheckEqual(A1, A2); } -template +template void testTensorMul(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { A1.dotMul(B, C); // a = b * c A2 = B * C; @@ -892,7 +900,7 @@ void testTensorMul(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { TensorCheckEqual(A1, A2); } -template +template void testTensorDiv(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { A1.dotDiv(B, C); // a = (b == 0.0) ? 0.0 : b / c A2 = (B == (real)0.0).condition((real)0.0, B / C); @@ -905,7 +913,7 @@ void testTensorDiv(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { TensorCheckEqual(A1, A2); } -template +template void testTensorReciprocal(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { real p1 = 1.5; real p2 = 2.5; @@ -915,14 +923,14 @@ void testTensorReciprocal(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { TensorCheckEqual(A1, A2); } -template +template void testTensorSoftCrossEntropy(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { A1.softCrossEntropy(B, C); // a = -c * log(b) - (1 - c) * log(1 - b) A2 = -C * B.log() - (C.constant(1.0f) - C) * (B.constant(1.0f) - B).log(); TensorCheckErr(A1, A2); } -template +template void testTensorSoftCrossEntropyBp(Tensor& A1, Tensor& A2, Tensor& B, @@ -932,7 +940,7 @@ void testTensorSoftCrossEntropyBp(Tensor& A1, TensorCheckEqual(A1, A2); } -template +template void testTernaryBaseOp(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { testTensorAdd(A1, A2, B, C); testTensorSub(A1, A2, B, C); @@ -952,30 +960,30 @@ TEST(Ternary, BaseOp) { #endif } -template +template void testTensorBinaryLabelCrossEntropy(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { A1.binaryLabelCrossEntropy(B, C); // a = c > 0.5 ? -log(b) : -log(1.0 - b) - A2 = (C > (real)0.5).condition( - -(B.log()), -((B.constant(1.0f) - B).log())); + A2 = (C > (real)0.5).condition(-(B.log()), -((B.constant(1.0f) - B).log())); TensorCheckErr(A1, A2); } -template +template void testTensorBinaryLabelCrossEntropyBp(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { // a += c > 0.5 ? -1.0 / b : 1.0 / (1.0 - b) A1.binaryLabelCrossEntropyBp(B, C); - A2 += (C > (real)0.5).condition( - (B.constant(-1.0f) / B), (B.constant(1.0f) - B).reciprocal()); + A2 += (C > (real)0.5) + .condition((B.constant(-1.0f) / B), + (B.constant(1.0f) - B).reciprocal()); TensorCheckErr(A1, A2); } -template +template void testTensorLogisticRegressionLoss(Tensor& A1, Tensor& A2, Tensor& B, @@ -991,13 +999,14 @@ void testTensorLogisticRegressionLoss(Tensor& A1, */ A1.logisticRegressionLoss(B, C); real THRESHOLD = 40.0; - auto tmp = (B > THRESHOLD).condition( - THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B)); + auto tmp = + (B > THRESHOLD) + .condition(THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B)); A2 = (C.constant(1.0f) + tmp.exp()).log() - C * tmp; TensorCheckErr(A1, A2); } -template +template void testTensorLogisticRegressionLossBp(Tensor& A1, Tensor& A2, Tensor& B, @@ -1013,28 +1022,29 @@ void testTensorLogisticRegressionLossBp(Tensor& A1, */ A1.logisticRegressionLossBp(B, C); real THRESHOLD = 40.0; - auto tmp = (B > THRESHOLD).condition( - THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B)); + auto tmp = + (B > THRESHOLD) + .condition(THRESHOLD, (B < -THRESHOLD).condition(-THRESHOLD, B)); auto tmp2 = tmp.exp(); A2 = tmp2 / (C.constant(1.0) + tmp2) - C; TensorCheckErr(A1, A2); } -template +template void testTensorBiggerThan(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { A1.biggerThan(B, C); // a = (b > c) ? 1.0f : 0.0f A2 = (B > C).condition((real)1.0f, (real)0.0f); TensorCheckEqual(A1, A2); } -template +template void testTensorMax(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { A1.max2(B, C); // a = (b > c) ? b : c A2 = (B > C).condition(B, C); TensorCheckEqual(A1, A2); } -template +template void testTernaryCompareOp(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) { testTensorBinaryLabelCrossEntropyBp(A1, A2, B, C); testTensorBinaryLabelCrossEntropy(A1, A2, B, C); @@ -1053,12 +1063,9 @@ TEST(Ternary, CompareOp) { #endif } -template -void testQuaternaryAdd(Tensor& A1, - Tensor& A2, - Tensor& B, - Tensor& C, - Tensor& D) { +template +void testQuaternaryAdd( + Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { // A1.add3(B, C, D, 1.5f, 2.5f, 3.5f); // a = p1 * b + p2 * c + p3 * d // A2 = B * 1.5f + C * 2.5f + D * 3.5f; // TensorCheckEqual(A1, A2); @@ -1084,25 +1091,19 @@ TEST(Quaternary, BaseOp) { #endif } -template -void testTensorBiggerThan(Tensor& A1, - Tensor& A2, - Tensor& B, - Tensor& C, - Tensor& D) { +template +void testTensorBiggerThan( + Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { // a = ((b > c && d > 0.5f) || (b < c && d < 0.5f)) ? 1.0f : 0.0f); A1.biggerThan(B, C, D); - A2 = ((B > C && D > (real)0.5) - || (B < C && D < (real)0.5)).condition((real)1.0, (real)0.0); + A2 = ((B > C && D > (real)0.5) || (B < C && D < (real)0.5)) + .condition((real)1.0, (real)0.0); TensorCheckEqual(A1, A2); } -template -void testTensorRankLoss(Tensor& A1, - Tensor& A2, - Tensor& B, - Tensor& C, - Tensor& D) { +template +void testTensorRankLoss( + Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { /** * const T THRESHOLD = 40.0; a = b - c; * a = (a > THRESHOLD) @@ -1114,19 +1115,17 @@ void testTensorRankLoss(Tensor& A1, real THRESHOLD = 40.0; auto tmp = B - C; - auto tmp2 = (tmp > THRESHOLD).condition( - THRESHOLD, (tmp < -THRESHOLD).condition(-THRESHOLD, tmp)); + auto tmp2 = + (tmp > THRESHOLD) + .condition(THRESHOLD, (tmp < -THRESHOLD).condition(-THRESHOLD, tmp)); A2 = (D.constant(1.0f) + tmp2.exp()).log() - tmp2 * D; TensorCheckErr(A1, A2); } -template -void testTensorRankLossBp(Tensor& A1, - Tensor& A2, - Tensor& B, - Tensor& C, - Tensor& D) { +template +void testTensorRankLossBp( + Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { /** * const T THRESHOLD = 40.0; a = b - c; * a = (a > THRESHOLD) @@ -1137,20 +1136,18 @@ void testTensorRankLossBp(Tensor& A1, A1.rankLossBp(B, C, D); real THRESHOLD = 40.0; auto tmp = B - C; - auto tmp2 = (tmp > THRESHOLD).condition( - THRESHOLD, (tmp < -THRESHOLD).condition(-THRESHOLD, tmp)); + auto tmp2 = + (tmp > THRESHOLD) + .condition(THRESHOLD, (tmp < -THRESHOLD).condition(-THRESHOLD, tmp)); auto tmp3 = tmp2.exp(); A2 = tmp3 / (D.constant(1.0f) + tmp3) - D; TensorCheckErr(A1, A2); } -template -void testQuaternaryCompareOp(Tensor& A1, - Tensor& A2, - Tensor& B, - Tensor& C, - Tensor& D) { +template +void testQuaternaryCompareOp( + Tensor& A1, Tensor& A2, Tensor& B, Tensor& C, Tensor& D) { testTensorBiggerThan(A1, A2, B, C, D); testTensorRankLoss(A1, A2, B, C, D); testTensorRankLossBp(A1, A2, B, C, D); diff --git a/paddle/math/tests/test_lazyAssign.cu b/paddle/math/tests/test_lazyAssign.cu index 786d863a533b58ea9856300aaa0cd8f5a10a4dd9..92afab4ff7f5ff4acc219c5ac783733340c5726a 100644 --- a/paddle/math/tests/test_lazyAssign.cu +++ b/paddle/math/tests/test_lazyAssign.cu @@ -13,10 +13,10 @@ See the License for the specific language governing permissions and limitations under the License. */ #include +#include "PerfUtils.h" +#include "TensorCheck.h" #include "paddle/math/Matrix.h" #include "paddle/math/TensorAssign.h" -#include "TensorCheck.h" -#include "PerfUtils.h" using paddle::BaseMatrix; using paddle::CpuMatrix; @@ -27,14 +27,28 @@ using autotest::TensorCheckErr; typedef std::function testMatrixFunc; void testMatrixCase(testMatrixFunc matrixFunc) { for (auto height : {1}) { - for (auto width : {1, 32, 64, 128, 512, 1024, 4096, 32768, 65536, 131072, - 262144, 524288, 1048576, 2097152, 4194304, 8388608}) { + for (auto width : {1, + 32, + 64, + 128, + 512, + 1024, + 4096, + 32768, + 65536, + 131072, + 262144, + 524288, + 1048576, + 2097152, + 4194304, + 8388608}) { matrixFunc(height, width); } } } -template +template void testLazyAssign(int height, int width) { Tensor A1(height, width); Tensor A2(height, width); @@ -49,40 +63,39 @@ void testLazyAssign(int height, int width) { EXPRESSION_PERFORMANCE(A1 = B + C; A1 = A1 * D;); - EXPRESSION_PERFORMANCE( - auto expr1 = A2.lazyAssign(B + C); - auto expr2 = A2.lazyAssign(A2 * D); - AssignEvaluate(expr1, expr2);); + EXPRESSION_PERFORMANCE(auto expr1 = A2.lazyAssign(B + C); + auto expr2 = A2.lazyAssign(A2 * D); + AssignEvaluate(expr1, expr2);); TensorCheckErr(A1, A2); } -TEST(lazyAssign, CPU) { - testMatrixCase(testLazyAssign); -} +TEST(lazyAssign, CPU) { testMatrixCase(testLazyAssign); } #ifndef PADDLE_ONLY_CPU -TEST(lazyAssign, GPU) { - testMatrixCase(testLazyAssign); -} +TEST(lazyAssign, GPU) { testMatrixCase(testLazyAssign); } #endif -template -void sgdUpdateTensor(Tensor& A, Tensor& B, Tensor& C, Tensor& D, - real p1, real p2, real p3) { +template +void sgdUpdateTensor( + Tensor& A, Tensor& B, Tensor& C, Tensor& D, real p1, real p2, real p3) { C = C * p2 - D * (B + A * p3) * p1; A += C; } -void sgdUpdateLazyAssign(BaseMatrix& A, BaseMatrix& B, - BaseMatrix& C, BaseMatrix& D, - real p1, real p2, real p3) { +void sgdUpdateLazyAssign(BaseMatrix& A, + BaseMatrix& B, + BaseMatrix& C, + BaseMatrix& D, + real p1, + real p2, + real p3) { auto expr1 = C.lazyAssign(C * p2 - D * (B + A * p3) * p1); auto expr2 = A.lazyAssign(A + C); AssignEvaluate(expr1, expr2); } -template +template void testSgdUpdate(int height, int width) { Tensor A1(height, width); Tensor A2(height, width); @@ -113,16 +126,13 @@ void testSgdUpdate(int height, int width) { * a = a + c; */ // BaseMatrix API - EXPRESSION_PERFORMANCE( - A1.sgdUpdate(B, C1, D, p1, p2, p3);); + EXPRESSION_PERFORMANCE(A1.sgdUpdate(B, C1, D, p1, p2, p3);); // Tensor expression - EXPRESSION_PERFORMANCE( - sgdUpdateTensor(A2, B, C2, D, p1, p2, p3)); + EXPRESSION_PERFORMANCE(sgdUpdateTensor(A2, B, C2, D, p1, p2, p3)); // lazyAssign - EXPRESSION_PERFORMANCE( - sgdUpdateLazyAssign(A3, B, C3, D, p1, p2, p3)); + EXPRESSION_PERFORMANCE(sgdUpdateLazyAssign(A3, B, C3, D, p1, p2, p3)); TensorCheckErr(A1, A2); TensorCheckErr(A1, A3); @@ -130,12 +140,8 @@ void testSgdUpdate(int height, int width) { TensorCheckErr(C1, C3); } -TEST(sgdUpdate, CPU) { - testMatrixCase(testSgdUpdate); -} +TEST(sgdUpdate, CPU) { testMatrixCase(testSgdUpdate); } #ifndef PADDLE_ONLY_CPU -TEST(sgdUpdate, GPU) { - testMatrixCase(testSgdUpdate); -} +TEST(sgdUpdate, GPU) { testMatrixCase(testSgdUpdate); } #endif diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp index 4980208e659233d50cd464dfeb213adfd2be3f38..d77478f345df97b37b214b5978f51ce47c1d791c 100644 --- a/paddle/math/tests/test_matrixCompare.cpp +++ b/paddle/math/tests/test_matrixCompare.cpp @@ -79,8 +79,8 @@ void testMatrixMaxSequence(int batchSize, int inputDim) { } TEST(Matrix, maxSequence) { - for (auto batchSize : {1, 10, 128, 1000, 6000}) { - for (auto inputDim : {1, 32, 100, 512}) { + for (auto batchSize : {1, 3, 997}) { // prime numbers close to 1, 4, 1024 + for (auto inputDim : {1, 7, 131}) { // prime numbers close to 1, 8, 128 VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim; testMatrixMaxSequence(batchSize, inputDim); } @@ -240,14 +240,10 @@ TEST(Matrix, unary) { // inverse matrix testMatrixInverse(height); #else - LOG(WARNING) << "Cannot run Matrix Inverse Unit Test.\n" - << "Failed to find lapack library in current system.\n" - << "To address this issue, Please adopt one of the following " - "approaches: \n" - << "1. Simply issue `sudo apt-get install liblapacke-dev` to " - "avoid re-build source code. \n" - << "2. Install MKL/Openblas/ATLAS and re-build PaddlePaddle " - "source code."; + LOG(WARNING) << "This version of PaddlePaddle was not built with LAPACK" + << "support so we cannot test matrix inverse. To test " + << "matrix inverse, please install LAPACKE " + << "and MKL/Openblas/ATLAS, and re-build PaddlePaddle."; #endif } } @@ -341,8 +337,8 @@ void testMatrixSoftmaxBp(int height, int width) { } TEST(Matrix, softmax) { - for (auto height : {1, 11, 73, 128, 200}) { - for (auto width : {1, 32, 100, 512, 1000}) { + for (auto height : {1, 3, 131}) { // prime numbers close to 1, 4, 127 + for (auto width : {1, 17, 251}) { // prime numbers close to 1, 16, 256 VLOG(3) << " height=" << height << " width=" << width; testMatrixSoftmax(height, width); @@ -527,7 +523,7 @@ void testVectorRowFunc(int size) { } TEST(Vector, rowFunc) { - for (auto size : {1, 5, 31, 90, 150, 500, 1000, 4000}) { + for (auto size : {1, 3, 997}) { // prime numbers close to 1, 4, 1024 VLOG(3) << " size=" << size; testVectorRowFunc(size); } @@ -604,7 +600,7 @@ void testVectorIsEqual(int size) { } TEST(Vector, Equal) { - for (auto size : {1, 5, 31, 90, 150, 500, 1000, 4000}) { + for (auto size : {1, 3, 997}) { // prime numbers close to 1, 4, 1024 VLOG(3) << " size=" << size; testVectorReset(size); testVectorReset(size); @@ -635,9 +631,8 @@ void testMatrixTopK(int samples, int dim, int beamSize) { } TEST(Matrix, topK) { - for (auto samples : {1, 5, 31, 90, 150, 500}) { - for (auto dim : - {1, 5, 8, 10, 15, 64, 80, 120, 256, 300, 1280, 5120, 50000}) { + for (auto samples : {1, 17, 131}) { // prime numbers close to 1, 16, 127 + for (auto dim : {1, 3, 997}) { // prime numbers close to 1, 4, 1024 for (auto beamSize : {1, 5, 10, 20, 40, (int)rand() % dim + 1}) { if (beamSize > dim) continue; VLOG(3) << " samples=" << samples << " beamSize=" << beamSize @@ -650,6 +645,7 @@ TEST(Matrix, topK) { void testSMatrixTopK(int samples, int dim, int beamSize, real ratio) { int nnz = samples * dim * ratio; + if (nnz < 1) nnz = 1; // Because sparseRand in MathUtil.cpp requires this. MatrixPtr cpuSrc = std::make_shared(samples, dim, nnz); MatrixPtr gpuSrc = std::make_shared(samples, dim, nnz); MatrixPtr cpuVal = std::make_shared(samples, beamSize); @@ -683,9 +679,9 @@ void testSMatrixTopK(int samples, int dim, int beamSize, real ratio) { } TEST(SMatrix, topK) { - for (auto samples : {1, 5, 100}) { - for (auto dim : {10000, 10000, 50000}) { - for (auto beamSize : {1, 5, 40, 100, 500}) { + for (auto samples : {1, 3, 61}) { + for (auto dim : {1, 3, 61}) { + for (auto beamSize : {1, 3, 61}) { for (auto ratio : {0.01, 0.001}) { if (beamSize > dim) continue; VLOG(3) << " samples=" << samples << " beamSize=" << beamSize @@ -806,10 +802,9 @@ void testClassificationError(int numSamples, int dim, int topkSize) { } TEST(Matrix, classificationError) { - for (auto numSamples : {1, 5, 31, 90, 150, 300}) { - for (auto dim : - {1, 5, 8, 10, 15, 64, 80, 120, 256, 300, 1280, 5120, 50000}) { - for (auto topkSize : {1, 5, 10, 20, 40, (int)rand() % dim + 1}) { + for (auto numSamples : {1, 3, 31}) { + for (auto dim : {1, 3, 31}) { + for (auto topkSize : {1, 3, (int)rand() % dim + 1}) { if (topkSize > dim) continue; VLOG(3) << " sample= " << numSamples << " topkSize= " << topkSize << " dim= " << dim; @@ -1016,13 +1011,15 @@ void testAvgPoolFwdBwd(int numSamples, 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, PoolFwdBwd) { - for (auto numSamples : {5, 32}) { - for (auto channels : {1, 9, 32}) { - for (auto imgSizeH : {14, 28}) { - for (auto imgSizeW : {16, 30}) { - for (auto sizeX : {2, 5}) { - for (auto sizeY : {2, 5}) { + for (auto numSamples : {1, 3}) { + for (auto channels : {1, 3}) { + for (auto imgSizeH : {13, 17}) { + for (auto imgSizeW : {17, 19}) { + for (auto sizeX : {2, 3}) { + for (auto sizeY : {2, 3}) { for (auto sH : {1, 2}) { for (auto sW : {1, 2}) { for (auto pH : {0, (sizeY - 1) / 2}) { @@ -1128,8 +1125,8 @@ TEST(Matrix, MaxOutFwdBwd) { } TEST(CpuMatrix, copyFrom) { - const size_t height = 1000; - const size_t width = 1000; + const size_t height = 31; + const size_t width = 53; CpuMatrix cpu(height, width); GpuMatrix gpu(height, width); CpuMatrix copy(height, width); @@ -1149,6 +1146,10 @@ void testBatch2seqPadding(int batchSize, int inputDim) { IVectorPtr cpuSequence; generateSequenceStartPositions(batchSize, cpuSequence); + for (int i = 0; i < int(cpuSequence->getSize()); ++i) { + (cpuSequence->getData())[i] += 1; // so no way that maxSeqLen is 0; + } + IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true); gpuSequence->copyFrom(*cpuSequence); @@ -1156,45 +1157,46 @@ void testBatch2seqPadding(int batchSize, int inputDim) { size_t maxSeqLen = *std::max_element(cpuSequence->getData(), cpuSequence->getData() + numSeq); + printf("numSeq = %ld, maxSeqLen = %ld\n", numSeq, maxSeqLen); MatrixPtr cBatch = std::make_shared(numSeq * maxSeqLen, inputDim); MatrixPtr gBatch = std::make_shared(numSeq * maxSeqLen, inputDim); MatrixPtr cCheck = std::make_shared(numSeq * maxSeqLen, inputDim); - hl_sequence2batch_copy_padding(gBatch->getData(), - gpuInput->getData(), - cpuSequence->getData(), - inputDim, - maxSeqLen, - numSeq, - false, - true); - cCheck->copyFrom(*gBatch); - - int* seqStart = cpuSequence->getData(); - float* batchData = cBatch->getData(); - float* seqData = cpuInput->getData(); - for (size_t i = 0; i < maxSeqLen; i++) { - for (size_t j = 0; j < numSeq; j++) { - size_t sequenceStart = seqStart[j]; - size_t sequenceLength = seqStart[j + 1] - seqStart[j]; - if (i < sequenceLength) { - memcpy(batchData + (i * numSeq + j) * inputDim, - seqData + (sequenceStart + i) * inputDim, - inputDim * sizeof(real)); - } else { - memset(batchData + (i * numSeq + j) * inputDim, - 0, - inputDim * sizeof(real)); - } - } - } - - TensorCheckErr(*cBatch, *cCheck); + // hl_sequence2batch_copy_padding(gBatch->getData(), + // gpuInput->getData(), + // cpuSequence->getData(), + // inputDim, + // maxSeqLen, + // numSeq, + // false, + // true); + // cCheck->copyFrom(*gBatch); + + // int* seqStart = cpuSequence->getData(); + // float* batchData = cBatch->getData(); + // float* seqData = cpuInput->getData(); + // for (size_t i = 0; i < maxSeqLen; i++) { + // for (size_t j = 0; j < numSeq; j++) { + // size_t sequenceStart = seqStart[j]; + // size_t sequenceLength = seqStart[j + 1] - seqStart[j]; + // if (i < sequenceLength) { + // memcpy(batchData + (i * numSeq + j) * inputDim, + // seqData + (sequenceStart + i) * inputDim, + // inputDim * sizeof(real)); + // } else { + // memset(batchData + (i * numSeq + j) * inputDim, + // 0, + // inputDim * sizeof(real)); + // } + // } + // } + + // TensorCheckErr(*cBatch, *cCheck); } TEST(Matrix, warpCTC) { - for (auto batchSize : {51, 526, 2884}) { - for (auto inputDim : {32, 512, 2026}) { + for (auto batchSize : {1, 3, 17}) { + for (auto inputDim : {1, 3, 31}) { VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim; testBatch2seqPadding(batchSize, inputDim); } diff --git a/paddle/memory/detail/buddy_allocator.h b/paddle/memory/detail/buddy_allocator.h index 4fa3fb0ee5f826d2b084c0ba184c505aee3acc48..9c41378483993101a098fc4ad1068c1ef908e566 100644 --- a/paddle/memory/detail/buddy_allocator.h +++ b/paddle/memory/detail/buddy_allocator.h @@ -39,7 +39,7 @@ class BuddyAllocator { public: void* Alloc(size_t unaligned_size); - void Free(void*); + void Free(void* ptr); size_t Used(); public: diff --git a/paddle/memory/detail/meta_cache.h b/paddle/memory/detail/meta_cache.h index ca0789779e273fb71c3d6282c0a921cda2d776cc..cf5815644284c23a1d2abc904f8c5053ce107a72 100644 --- a/paddle/memory/detail/meta_cache.h +++ b/paddle/memory/detail/meta_cache.h @@ -33,17 +33,17 @@ namespace detail { */ class MetadataCache { public: - MetadataCache(bool uses_gpu); + explicit MetadataCache(bool uses_gpu); public: /*! \brief Load the associated metadata for the specified memory block. */ - Metadata load(const MemoryBlock*); + Metadata load(const MemoryBlock* memory_block); /*! \brief Store the associated metadata for the specified memory block. */ - void store(MemoryBlock*, const Metadata&); + void store(MemoryBlock* memory_block, const Metadata& meta_data); /*! \brief Indicate that the specified metadata will no longer be used. */ - void invalidate(MemoryBlock*); + void invalidate(MemoryBlock* memory_block); public: MetadataCache(const MetadataCache&) = delete; diff --git a/paddle/memory/memory.h b/paddle/memory/memory.h index 44f567caf9c19775f17988b5142b7693b41a126d..72351b9dfa63513713463bb47a3684f0dfd84ad3 100644 --- a/paddle/memory/memory.h +++ b/paddle/memory/memory.h @@ -68,7 +68,7 @@ class PODDeleter { static_assert(std::is_pod::value, "T must be POD"); public: - PODDeleter(Place place) : place_(place) {} + explicit PODDeleter(Place place) : place_(place) {} void operator()(T* ptr) { Free(place_, static_cast(ptr)); } private: diff --git a/paddle/operators/.clang-format b/paddle/operators/.clang-format new file mode 100644 index 0000000000000000000000000000000000000000..47b8a85206ab457e2b3cb90a68b7a82a0753d327 --- /dev/null +++ b/paddle/operators/.clang-format @@ -0,0 +1,5 @@ +--- +Language: Cpp +BasedOnStyle: Google +Standard: Cpp11 +... diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt index 96c76e22e9814682008f2e6c7ae98e2599d391c2..a7c89787e43df6173791bc54b3faffc034867f7d 100644 --- a/paddle/operators/CMakeLists.txt +++ b/paddle/operators/CMakeLists.txt @@ -41,27 +41,30 @@ function(op_library TARGET) endif() endfunction() +add_subdirectory(math) +cc_test(gather_test SRCS gather_test.cc DEPS tensor) + +cc_test(scatter_test SRCS scatter_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) -cc_test(add_op_test SRCS add_op_test.cc DEPS add_op) op_library(mean_op SRCS mean_op.cc mean_op.cu) -cc_test(mean_op_test SRCS mean_op_test.cc DEPS mean_op) -op_library(mul_op SRCS mul_op.cc mul_op.cu) +op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function) op_library(rowwise_add_op SRCS rowwise_add_op.cu rowwise_add_op.cc) 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) op_library(sgd_op SRCS sgd_op.cc sgd_op.cu) -op_library(fc_op - SRCS fc_op.cc - DEPS mul_op rowwise_add_op sigmoid_op softmax_op net_op) -op_library(recurrent_op SRCS recurrent_op.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(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc + DEPS framework_proto tensor op_registry operator net_op) +op_library(uniform_random_op + SRCS uniform_random_op.cc uniform_random_op.cu) diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc index 85269a5f7445a1745d9be68417789e33eb725d5c..8ab748ed71e9a5dc0ee0259a78a2b886870bec5b 100644 --- a/paddle/operators/add_op.cc +++ b/paddle/operators/add_op.cc @@ -17,24 +17,22 @@ limitations under the License. */ namespace paddle { namespace operators { -class AddOp : public OperatorWithKernel { -protected: - void InferShape(const InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 2, "Input size of AddOp must be two"); - PADDLE_ENFORCE(ctx.OutputSize() == 1, "Output size of AddOp must be one"); - PADDLE_ENFORCE(ctx.InputVar(0) != nullptr && ctx.InputVar(1) != nullptr, - "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()); +class AddOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + 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()); } }; -class AddOpMaker : public OpProtoAndCheckerMaker { -public: - AddOpMaker(OpProto *proto, OpAttrChecker *op_checker) +class AddOpMaker : public framework::OpProtoAndCheckerMaker { + public: + AddOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The first input of add op"); AddInput("Y", "The second input of add op"); @@ -47,14 +45,19 @@ The equation is: Out = X + Y } }; -class AddOpGrad : public OperatorWithKernel { -protected: - void InferShape(const InferShapeContext &ctx) const override {} +class AddOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override {} }; } // namespace operators } // namespace paddle -REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker); -REGISTER_GRADIENT_OP(add_two, add_two_grad, ops::AddOpGrad); -REGISTER_OP_CPU_KERNEL(add_two, ops::AddKernel); +namespace ops = paddle::operators; +REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker, add_two_grad, ops::AddOpGrad); + +REGISTER_OP_CPU_KERNEL(add_two, + ops::AddKernel); diff --git a/paddle/operators/add_op.cu b/paddle/operators/add_op.cu index f961b37565f400b5c26844b9e7a3cff5e682340b..cec5f558cbc161124620ad4241d6bd8a5324277c 100644 --- a/paddle/operators/add_op.cu +++ b/paddle/operators/add_op.cu @@ -1,5 +1,21 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + #define EIGEN_USE_GPU #include "paddle/framework/op_registry.h" #include "paddle/operators/add_op.h" -REGISTER_OP_GPU_KERNEL(add_two, ops::AddKernel); +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(add_two, + ops::AddKernel); diff --git a/paddle/operators/add_op.h b/paddle/operators/add_op.h index 54d2231425293f6cfb3adc9cb34d903a75fcdcd0..a7307b6818aa3d10ff215d06281e2b53196fd101 100644 --- a/paddle/operators/add_op.h +++ b/paddle/operators/add_op.h @@ -13,18 +13,24 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once -#include "paddle/operators/type_alias.h" +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" namespace paddle { namespace operators { +using Tensor = framework::Tensor; +template +using EigenVector = framework::EigenVector; + template -class AddKernel : public OpKernel { -public: - void Compute(const ExecutionContext& context) const override { - auto input0 = context.Input(0); - auto input1 = context.Input(1); - auto output = context.Output(0); +class AddKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* input0 = context.Input("X"); + auto* input1 = context.Input("Y"); + auto* output = context.Output("Out"); output->mutable_data(context.GetPlace()); diff --git a/paddle/operators/add_op_test.cc b/paddle/operators/add_op_test.cc deleted file mode 100644 index 3d52f5498323dbb7ca0ff25d038947f0ddb2017e..0000000000000000000000000000000000000000 --- a/paddle/operators/add_op_test.cc +++ /dev/null @@ -1,28 +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 -#define private public -#include -USE_OP(add_two); -// USE_OP(add_two_grad); - -TEST(AddOp, GetOpProto) { - auto& protos = paddle::framework::OpRegistry::protos(); - auto it = protos.find("add_two"); - ASSERT_NE(it, protos.end()); - auto& op_creators = paddle::framework::OpRegistry::op_creators(); - auto it1 = op_creators.find("add_two_grad"); - ASSERT_NE(it1, op_creators.end()); -} diff --git a/paddle/operators/cross_entropy_op.cc b/paddle/operators/cross_entropy_op.cc index 4f5b935fde4d5b0d9efae66554cf890291e26941..a623c551e1088365ade6f73bc6149977b6ef017e 100644 --- a/paddle/operators/cross_entropy_op.cc +++ b/paddle/operators/cross_entropy_op.cc @@ -17,28 +17,40 @@ limitations under the License. */ namespace paddle { namespace operators { -class OnehotCrossEntropyOp : public OperatorWithKernel { -protected: - void InferShape(const InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 2, - "Input size of OnehotCrossEntropyOp must be two"); - PADDLE_ENFORCE(ctx.OutputSize() == 1, - "Output size of OnehotCrossEntropyOp must be one"); - PADDLE_ENFORCE(ctx.InputVar(0) != nullptr && ctx.InputVar(1) != nullptr, - "Inputs of OnehotCrossEntropyOp must all be set"); - PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr, - "Outputs of OnehotCrossEntropyOp must all be set"); - PADDLE_ENFORCE(ctx.Input(0)->dims().size() == 2, - "X's dimension must be 2."); - PADDLE_ENFORCE(ctx.Output(0)->dims().size() == 1, - "label's dimension must be 1."); - ctx.Output(0)->Resize({ctx.Input(0)->dims()[0]}); +class OnehotCrossEntropyOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + 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 { + 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"); + + // TODO(superjom) add enforce here after helper functions ready + X_grad->Resize(X->dims()); } }; -class OnehotCrossEntropyOpMaker : public OpProtoAndCheckerMaker { -public: - OnehotCrossEntropyOpMaker(OpProto *proto, OpAttrChecker *op_checker) +class OnehotCrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker { + public: + OnehotCrossEntropyOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The first input of OnehotCrossEntropyOp"); AddInput("label", "The second input of OnehotCrossEntropyOp"); @@ -54,8 +66,13 @@ OnehotCrossEntropy Operator. } // namespace operators } // namespace paddle -REGISTER_OP(onehot_cross_entropy, - ops::OnehotCrossEntropyOp, - ops::OnehotCrossEntropyOpMaker); -REGISTER_OP_CPU_KERNEL(onehot_cross_entropy, - ops::OnehotCrossEntropyOpKernel); +namespace ops = paddle::operators; +REGISTER_OP(onehot_cross_entropy, ops::OnehotCrossEntropyOp, + 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 926a0c616b957d8e542c1f3dee227a718fb29f07..4bbc8f093a794d46737a16488684a6a0cc25e285 100644 --- a/paddle/operators/cross_entropy_op.cu +++ b/paddle/operators/cross_entropy_op.cu @@ -1,5 +1,21 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + #define EIGEN_USE_GPU #include "paddle/operators/cross_entropy_op.h" -REGISTER_OP_GPU_KERNEL(onehot_cross_entropy, - ops::OnehotCrossEntropyOpKernel); \ No newline at end of file +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL( + onehot_cross_entropy, + ops::OnehotCrossEntropyOpKernel); diff --git a/paddle/operators/cross_entropy_op.h b/paddle/operators/cross_entropy_op.h index c3a3728149950a5c7f2195122e8e0ff728492bdb..b7df92c9a98ebf12b72a8d3d8e8e4e1a950f06c9 100644 --- a/paddle/operators/cross_entropy_op.h +++ b/paddle/operators/cross_entropy_op.h @@ -13,33 +13,75 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once -#include "paddle/operators/type_alias.h" +#include "paddle/framework/op_registry.h" namespace paddle { namespace operators { -template -class OnehotCrossEntropyOpKernel : public OpKernel { -public: - constexpr T LOG_THRESHOLD() const { return static_cast(1e-20); } +using Tensor = framework::Tensor; + +template +T tolerable_value(T x) { + static_assert(std::is_floating_point::value, + "tolerable_value works only on float, " + "double and double double."); + + const T kApproInf = 1e20; + + if (x == INFINITY) { + return kApproInf; + } - void Compute(const ExecutionContext& ctx) const override { - auto X = ctx.Input(0); - const T* X_data = X->data(); - const int* label_data = ctx.Input(1)->data(); - auto Y = ctx.Output(0); + if (x == -INFINITY) { + return -kApproInf; + } + + return x; +} + +template +class OnehotCrossEntropyOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + 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* Y_data = Y->data(); + T* Ydata = Y->data(); int batch_size = X->dims()[0]; int class_num = X->dims()[1]; - // Y[i] = -log(X[i][j]) for (int i = 0; i < batch_size; ++i) { - Y_data[i] = -std::log( - std::max(X_data[i * class_num + label_data[i]], LOG_THRESHOLD())); + int index = i * class_num + label_data[i]; + Ydata[i] = -tolerable_value(std::log(Xdata[index])); + } + } +}; + +template +class OnehotCrossEntropyGradientOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + 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(); + + const int batch_size = X->dims()[0]; + const int class_num = X->dims()[1]; + + 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/fc_op.cc b/paddle/operators/fc_op.cc deleted file mode 100644 index 71ceda958770796693265c08cb1fcae27e79bcd9..0000000000000000000000000000000000000000 --- a/paddle/operators/fc_op.cc +++ /dev/null @@ -1,71 +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 "type_alias.h" - -namespace paddle { -namespace operators { - -class FullyConnectedOp : public NetOp { -public: - void Init() override { - AddOp(OpRegistry::CreateOp("mul", - { - Input("X"), Input("W"), - }, - {Output("before_act")}, - {})); - auto b = Input("b"); - if (b != EMPTY_VAR_NAME()) { - AddOp(OpRegistry::CreateOp("rowwise_add", - {Output("before_act"), Input("b")}, - {Output("before_act")}, - {})); - } - - auto activation = GetAttr("activation"); - AddOp(OpRegistry::CreateOp( - activation, {Output("before_act")}, {Output("Y")}, {})); - CompleteAddOp(false); - } -}; - -class FullyConnectedOpMaker : public OpProtoAndCheckerMaker { -public: - FullyConnectedOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "the input of fc operator"); - AddInput("W", "the weight of fc operator"); - AddInput("b", "the bias of fc operator"); - - AddOutput("Y", "the output of fc operator"); - AddOutput("before_act", "the before activation output of fc operator") - .SetTemporary(); - AddAttr("activation", "The activation key for fc layer") - .SetDefault("sigmoid") - .InEnum({"sigmoid", "softmax"}); - - //! TODO(yuyang18): Complete comment; - AddComment("FullyConnected Operator"); - } -}; -} // namespace operators -} // namespace paddle - -USE_OP(mul); -USE_OP(rowwise_add); -USE_OP(sigmoid); -USE_OP(softmax); - -REGISTER_OP(fc, ops::FullyConnectedOp, ops::FullyConnectedOpMaker); diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc index 79a0e3d7e911b728a7a96ceff573976ba2b2e37f..9d51f6e3a16fe96125599bb440d40237aeb9a028 100644 --- a/paddle/operators/fill_zeros_like_op.cc +++ b/paddle/operators/fill_zeros_like_op.cc @@ -13,30 +13,23 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/operators/fill_zeros_like_op.h" -#include "paddle/framework/op_registry.h" -#include "paddle/framework/tensor.h" namespace paddle { namespace operators { class FillZerosLikeOp : public framework::OperatorWithKernel { -protected: + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 1UL, - "Input size of FillZerosLikeOp must be one."); - PADDLE_ENFORCE(ctx.OutputSize() == 1UL, - "Output size of AddOp must be one."); - PADDLE_ENFORCE(ctx.InputVar(0) != nullptr, - "Input of FillZerosLikeOp must be set."); - PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr, - "Output of FillZerosLikeOp must be set."); - ctx.Output(0)->Resize( - ctx.Input(0)->dims()); + ctx.Output("Dst")->Resize( + ctx.Input("Src")->dims()); } }; class FillZerosLikeOpMaker : public framework::OpProtoAndCheckerMaker { -public: + public: FillZerosLikeOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : framework::OpProtoAndCheckerMaker(proto, op_checker) { @@ -52,9 +45,9 @@ The output will have the same size with input. } // namespace operators } // namespace paddle -REGISTER_OP(fill_zeros_like, - paddle::operators::FillZerosLikeOp, - paddle::operators::FillZerosLikeOpMaker); +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, ops::FillZerosLikeOp, + ops::FillZerosLikeOpMaker); REGISTER_OP_CPU_KERNEL( fill_zeros_like, - paddle::operators::FillZerosLikeKernel); + ops::FillZerosLikeKernel); diff --git a/paddle/operators/fill_zeros_like_op.cu b/paddle/operators/fill_zeros_like_op.cu index 55ad58f4f17cd4a3e737c01b001675d2690d273e..fdbcf520a0d7b4ddfe3fc1837a21e0ce88b8e8fa 100644 --- a/paddle/operators/fill_zeros_like_op.cu +++ b/paddle/operators/fill_zeros_like_op.cu @@ -1,6 +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/framework/op_registry.h" #include "paddle/operators/fill_zeros_like_op.h" +namespace ops = paddle::operators; REGISTER_OP_GPU_KERNEL( fill_zeros_like, - paddle::operators::FillZerosLikeKernel); \ No newline at end of file + ops::FillZerosLikeKernel); diff --git a/paddle/operators/fill_zeros_like_op.h b/paddle/operators/fill_zeros_like_op.h index 05272964abd43bdc2bd5c3cae8b128099e1c888c..fd380ca8514b0ac50f39613368a4836bd485668b 100644 --- a/paddle/operators/fill_zeros_like_op.h +++ b/paddle/operators/fill_zeros_like_op.h @@ -13,20 +13,20 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once -#include "glog/logging.h" #include "paddle/framework/eigen.h" -#include "paddle/framework/operator.h" +#include "paddle/framework/op_registry.h" namespace paddle { namespace operators { template class FillZerosLikeKernel : public framework::OpKernel { -public: + public: void Compute(const framework::ExecutionContext& context) const override { - auto* output = context.Output(0); + auto* output = context.Output("Dst"); output->mutable_data(context.GetPlace()); - framework::EigenVector::Flatten(*output).setZero(); + auto t = framework::EigenVector::Flatten(*output); + t.device(context.GetEigenDevice()) = t.constant(T(0)); } }; diff --git a/paddle/operators/gather.h b/paddle/operators/gather.h new file mode 100644 index 0000000000000000000000000000000000000000..d6e6990394e46ba06c4bacfe33ca522f3ff1413a --- /dev/null +++ b/paddle/operators/gather.h @@ -0,0 +1,73 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include + +#include "paddle/framework/ddim.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +namespace paddle { +namespace operators { + +// Implementation of CPU copy +template +void CPUGather(const T* params, const int* indices, const int slice_size, + const int index_size, T* output) { + const size_t slice_bytes = slice_size * sizeof(T); + + for (int i = 0; i < index_size; ++i) { + int index_ = indices[i]; + memcpy(output + i * slice_size, params + index_ * slice_size, slice_bytes); + } +} + +// Implementation of GPU copy: +template +void GPUGather(const T* src, const int* index, const int slice_size, + const int index_size, T* output); + +/** + * Return a new tensor from source tensor, gathered according to index + * input[src]: type-T source Tensor + * input[index]: type-int index Tensor (1-D) + * return: output tensor + */ +template +void Gather(const platform::Place& place, const paddle::framework::Tensor* src, + const paddle::framework::Tensor* index, + paddle::framework::Tensor* output) { + // check index of shape 1-D + PADDLE_ENFORCE(index->dims().size() == 1); + int index_size = index->dims()[0]; + + auto src_dims = src->dims(); + paddle::framework::DDim output_dims(src_dims); + output_dims[0] = index_size; + + // slice size + int slice_size = 1; + for (int i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i]; + + // Gathering + if (platform::is_cpu_place(place)) { + CPUGather(src->data(), index->data(), slice_size, index_size, + output->data()); + } +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/gather_test.cc b/paddle/operators/gather_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..d24d83f299fdb071e60fa3cc7b223c0228cb29af --- /dev/null +++ b/paddle/operators/gather_test.cc @@ -0,0 +1,48 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/gather.h" +#include "paddle/framework/ddim.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +#include +#include +#include + +TEST(Gather, GatherData) { + using namespace paddle::framework; + using namespace paddle::platform; + using namespace paddle::operators; + + Tensor* src = new Tensor(); + Tensor* index = new Tensor(); + Tensor* output = new Tensor(); + + int* p_src = nullptr; + int* p_index = nullptr; + p_src = src->mutable_data(make_ddim({3, 4}), CPUPlace()); + p_index = index->mutable_data(make_ddim({2}), CPUPlace()); + + for (int i = 0; i < 12; ++i) p_src[i] = i; + p_index[0] = 1; + p_index[1] = 0; + + int* p_output = output->mutable_data(make_ddim({2, 4}), CPUPlace()); + + Gather(CPUPlace(), src, index, output); + + for (int i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4); + for (int i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4); +} diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..f30bbce9586d61063b4b61d98695bb568ef73c8d --- /dev/null +++ b/paddle/operators/gaussian_random_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 +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class GaussianRandomKernel : 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); + 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"); + 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); + } + } +}; + +class GaussianRandomOp : public 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"); + PADDLE_ENFORCE(dims.size() > 0UL, + "dims can be one int or array. dims must be set."); + tensor->Resize(framework::make_ddim(dims)); + } +}; + +class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker { + public: + GaussianRandomOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { + AddOutput("Out", "output matrix of random op"); + AddComment(R"DOC( +GaussianRandom operator. +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("seed", + "Random seed of generator." + "0 means use system wide seed") + .SetDefault(0); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(gaussian_random, ops::GaussianRandomOp, + ops::GaussianRandomOpMaker); +REGISTER_OP_CPU_KERNEL(gaussian_random, ops::GaussianRandomKernel); diff --git a/paddle/operators/gaussian_random_op.cu b/paddle/operators/gaussian_random_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..1340b1e1e9f19fd96ced9e57fab75fe9d33bc84e --- /dev/null +++ b/paddle/operators/gaussian_random_op.cu @@ -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. */ + +#include +#include +#include "paddle/platform/dynload/curand.h" +#include "paddle/platform/gpu_info.h" + +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class GaussianRandomKernel : 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); + T* data = tensor->mutable_data(context.GetPlace()); + + int seed = context.op_.GetAttr("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); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(gaussian_random, ops::GaussianRandomKernel); diff --git a/paddle/operators/math/CMakeLists.txt b/paddle/operators/math/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..ed51d416ed9497eee45ba826ad672b8fb1ad3678 --- /dev/null +++ b/paddle/operators/math/CMakeLists.txt @@ -0,0 +1,8 @@ + +if(WITH_GPU) + nv_library(math_function SRCS math_function.cc math_function.cu DEPS cblas device_context) +else() + cc_library(math_function SRCS math_function.cc DEPS cblas device_context) +endif() + +nv_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor) diff --git a/paddle/operators/math/math_function.cc b/paddle/operators/math/math_function.cc new file mode 100644 index 0000000000000000000000000000000000000000..affdd1ac2cd486930881ee6b34a4b32f41df7ee9 --- /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 = K; + int ldb = N; + 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 = K; + int ldb = N; + 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 78131b26808b183ee107313374493ae870f1b641..49d0f43508b1ee3df0c6b5987942970e1649e310 100644 --- a/paddle/operators/mean_op.cc +++ b/paddle/operators/mean_op.cc @@ -17,31 +17,35 @@ limitations under the License. */ namespace paddle { namespace operators { -class MeanOp : public OperatorWithKernel { -protected: - void InferShape(const InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 1, "Input size of AddOp must be one"); - PADDLE_ENFORCE(ctx.OutputSize() == 1, "Output size of AddOp must be one"); - PADDLE_ENFORCE(ctx.InputVar(0) != nullptr && ctx.OutputVar(0) != nullptr, - "Input/Output of MeanOp must be initialized."); - ctx.Output(0)->Resize(framework::make_ddim({1})); +class MeanOp : 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 MeanOp must be initialized."); + ctx.Output("Out")->Resize({1}); } }; -class MeanOpMaker : public OpProtoAndCheckerMaker { -public: - MeanOpMaker(OpProto *proto, OpAttrChecker *op_checker) +class MeanOpMaker : public framework::OpProtoAndCheckerMaker { + public: + 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").AsNoGradient(); AddComment("Mean Operator"); } }; -class MeanGradOp : public OperatorWithKernel { -protected: - void InferShape(const InferShapeContext &ctx) const override { - ctx.Output("X" + GRAD_VAR_SUFFIX()) +class MeanGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + ctx.Output(framework::GradVarName("X")) ->Resize(ctx.Input("X")->dims()); } }; @@ -49,7 +53,9 @@ protected: } // namespace operators } // namespace paddle -REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker); -REGISTER_OP_CPU_KERNEL(mean, ops::MeanKernel); -REGISTER_GRADIENT_OP(mean, mean_grad, ops::MeanGradOp); -REGISTER_OP_CPU_KERNEL(mean_grad, ops::MeanGradKernel); +namespace ops = paddle::operators; +REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker, mean_grad, ops::MeanGradOp); +REGISTER_OP_CPU_KERNEL(mean, + ops::MeanKernel); +REGISTER_OP_CPU_KERNEL(mean_grad, + ops::MeanGradKernel); diff --git a/paddle/operators/mean_op.cu b/paddle/operators/mean_op.cu index e15de2fd0dd84e4015ee0e3b5343d7651b027a88..7af624d81dc5ffbb5c31b4d6f6eb8f9f8652a431 100644 --- a/paddle/operators/mean_op.cu +++ b/paddle/operators/mean_op.cu @@ -1,6 +1,23 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + #define EIGEN_USE_GPU #include "paddle/operators/mean_op.h" -REGISTER_OP_GPU_KERNEL(mean, ops::MeanKernel); -REGISTER_OP_GPU_KERNEL(mean_grad, ops::MeanGradKernel); \ No newline at end of file +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(mean, + ops::MeanKernel); +REGISTER_OP_GPU_KERNEL(mean_grad, + ops::MeanGradKernel); diff --git a/paddle/operators/mean_op.h b/paddle/operators/mean_op.h index e712dee6a785749e51be7b233e85dbf39c835218..fcb703e63bd5a82f9ffac2bf64e06fd0218dbdaa 100644 --- a/paddle/operators/mean_op.h +++ b/paddle/operators/mean_op.h @@ -13,36 +13,45 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once -#include "paddle/operators/type_alias.h" +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" namespace paddle { namespace operators { +using Tensor = framework::Tensor; +template +using EigenScalar = framework::EigenScalar; +template +using EigenVector = framework::EigenVector; + template -class MeanKernel : public OpKernel { -public: - void Compute(const ExecutionContext& context) const override { - auto input = context.Input(0); - auto output = context.Output(0); +class MeanKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + 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(); } }; template -class MeanGradKernel : public OpKernel { -public: - void Compute(const ExecutionContext& context) const override { - auto OG = context.Input("Out" + OperatorBase::GRAD_VAR_SUFFIX()); +class MeanGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto OG = context.Input(framework::GradVarName("Out")); PADDLE_ENFORCE(framework::product(OG->dims()) == 1, "Mean Gradient should be scalar"); - auto IG = context.Output("X" + OperatorBase::GRAD_VAR_SUFFIX()); + auto IG = context.Output(framework::GradVarName("X")); IG->mutable_data(context.GetPlace()); T ig_size = (T)framework::product(IG->dims()); diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc index d127f3a302a340fe7558f918d6eeb2ea0a3fafe7..95d19fb6aad37143e65759b03e12e3e78bce5915 100644 --- a/paddle/operators/mul_op.cc +++ b/paddle/operators/mul_op.cc @@ -13,29 +13,35 @@ limitations under the License. */ #include "paddle/operators/mul_op.h" +#include "paddle/operators/math/math_function.h" namespace paddle { namespace operators { -class MulOp : public OperatorWithKernel { -protected: - void InferShape(const 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(dim0.size() == 2 && dim1.size() == 2, - "The input of mul op must be matrix"); - PADDLE_ENFORCE( - dim0[1] == dim1[0], +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."); - PADDLE_ENFORCE(ctx.OutputSize() == 1, "The mul op must take one output"); - ctx.Output(0)->Resize({dim0[0], dim1[1]}); + ctx.Output("Out")->Resize({dim0[0], dim1[1]}); } }; -class MulOpMaker : public OpProtoAndCheckerMaker { -public: - MulOpMaker(OpProto *proto, OpAttrChecker *op_checker) +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"); @@ -48,9 +54,12 @@ The equation is: Out = X * Y } }; -class MulOpGrad : public OperatorWithKernel { -protected: - void InferShape(const InferShapeContext &ctx) const override {} +class MulOpGrad : public 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 ""; @@ -60,7 +69,6 @@ protected: } // namespace operators } // namespace paddle -REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker); -REGISTER_GRADIENT_OP(mul, mul_grad, ops::MulOpGrad); - -REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); +namespace ops = paddle::operators; +REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad); +REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); diff --git a/paddle/operators/mul_op.cu b/paddle/operators/mul_op.cu index dc9236701627dc9335b844d2a82e18eb1f7dfd42..346a7e505d123b5e4e831daa39a1f6349b3dcccf 100644 --- a/paddle/operators/mul_op.cu +++ b/paddle/operators/mul_op.cu @@ -15,4 +15,5 @@ #define EIGEN_USE_GPU #include "paddle/operators/mul_op.h" -REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel); \ No newline at end of file +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel); diff --git a/paddle/operators/mul_op.h b/paddle/operators/mul_op.h index c7b78ad39045d25d73bfc2c930063c255a514864..b7812fd1a7a72f5ce543e18c8b7b5b51deff2204 100644 --- a/paddle/operators/mul_op.h +++ b/paddle/operators/mul_op.h @@ -14,31 +14,37 @@ #pragma once -#include "paddle/operators/type_alias.h" +#include "paddle/operators/math/math_function.h" + +#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 MulKernel : public OpKernel { -public: - void Compute(const ExecutionContext& context) const override { +class MulKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { Eigen::array, 1> dim_pair = { {Eigen::IndexPair(1, 0)}}; - - auto input0 = context.Input("X"); - auto input1 = context.Input("Y"); - 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()); - auto X = EigenMatrix::From(*input0); auto Y = EigenMatrix::From(*input1); auto Z = EigenMatrix::From(*output); - auto place = context.GetEigenDevice(); + auto& place = context.GetEigenDevice(); Z.device(place) = X.contract(Y, dim_pair); } }; + } // namespace operators } // namespace paddle diff --git a/paddle/operators/net_op.cc b/paddle/operators/net_op.cc index fbc98e09923bda7f3baee04e02df9076247bff0b..c36fe8d6b58a0afa568e31e43567baa5f261c7d0 100644 --- a/paddle/operators/net_op.cc +++ b/paddle/operators/net_op.cc @@ -15,49 +15,42 @@ */ #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 (!Contains(output_set, var_name)) { // Not other op's output + input_set.insert(var_name); + } else { + intermediate_outputs_.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 { @@ -74,5 +67,25 @@ std::string NetOp::DebugString() const { bool NetOp::IsNetOp() const { return true; } +std::vector NetOp::OutputVars(bool has_intermediate) const { + if (has_intermediate) { + return this->outputs_.at(kAll); + } + auto& all = this->outputs_.at(kAll); + 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::OperatorBase::VarNameMap& inputs, + const framework::OperatorBase::VarNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + } // namespace operators } // namespace paddle diff --git a/paddle/operators/net_op.h b/paddle/operators/net_op.h index 13611e1ee83170db43e17d6088e4b04588ce6255..4a3408c158a029a96740717280c1562671fa938f 100644 --- a/paddle/operators/net_op.h +++ b/paddle/operators/net_op.h @@ -14,13 +14,8 @@ limitations under the License. */ #pragma once -#include "paddle/framework/op_desc.pb.h" -#include "paddle/framework/op_proto.pb.h" +#include "paddle/framework/framework.pb.h" #include "paddle/framework/op_registry.h" -#include "paddle/framework/operator.h" -#include "paddle/framework/scope.h" -#include "paddle/operators/type_alias.h" -#include "paddle/platform/device_context.h" namespace paddle { namespace operators { @@ -40,7 +35,12 @@ namespace operators { * it defines. */ class NetOp : public framework::OperatorBase { -public: + public: + static const char kAll[]; + NetOp() : framework::OperatorBase("plain_net", {}, {}, {}) {} + NetOp(const std::string& type, const VarNameMap& inputs, + const VarNameMap& outputs, const framework::AttributeMap& attrs); + /** * Infer all the operators' input and output variables' shapes, will be called * before every mini-batch @@ -65,20 +65,29 @@ public: } } + bool SupportGPU() const override { + for (auto& op : ops_) { + if (!op->SupportGPU()) { + return false; + } + } + return true; + } + /** * @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"); - PADDLE_ENFORCE(op != nullptr, "Cannot Insert Null op"); + PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op"); ops_.push_back(op); } void InsertOp(size_t pos, const std::shared_ptr& op) { PADDLE_ENFORCE(!add_op_done_, "Cannot InsertOp when this network is sealed"); - PADDLE_ENFORCE(op != nullptr, "Cannot Insert Null op"); - PADDLE_ENFORCE(pos <= ops_.size(), "Out of range"); + PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op"); + PADDLE_ENFORCE_LE(pos, ops_.size(), "Out of range"); ops_.insert(ops_.begin() + pos, op); } @@ -87,11 +96,13 @@ public: std::string DebugString() const override; bool IsNetOp() const override; + std::vector OutputVars(bool has_intermediate) const override; std::vector> ops_; -private: + 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 18c5c60eb43250c23e2819a3c79ab8a96fec103e..0cef71de6a032674a54387986f65f17ca99b400e 100644 --- a/paddle/operators/net_op_test.cc +++ b/paddle/operators/net_op_test.cc @@ -2,33 +2,24 @@ #include -#include "paddle/framework/op_registry.h" -#include "paddle/framework/operator.h" - namespace paddle { namespace operators { +using Scope = framework::Scope; +using DeviceContext = platform::DeviceContext; static int infer_shape_cnt = 0; static int run_cnt = 0; -class TestOp : public OperatorBase { -public: - void InferShape(const framework::Scope& scope) const override { - ++infer_shape_cnt; - } - void Run(const framework::Scope& scope, - const paddle::platform::DeviceContext& dev_ctx) const override { +class TestOp : public framework::OperatorBase { + public: + using framework::OperatorBase::OperatorBase; + void InferShape(const Scope& scope) const override { ++infer_shape_cnt; } + void Run(const Scope& scope, + const platform::DeviceContext& dev_ctx) const override { ++run_cnt; } }; -class EmptyOp : public OperatorBase { -public: - void InferShape(const Scope& scope) const override {} - void Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const override {} -}; - template void AssertSameVectorWithoutOrder(const std::vector& expected, const std::vector& actual) { @@ -46,40 +37,32 @@ 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"}; + auto op1 = std::shared_ptr( + new TestOp("test", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}}, + {{"Out", {"y"}}}, {})); net->AddOp(op1); - auto op2 = std::make_shared(); - op2->inputs_ = {"y", "w2", "b2"}; - op2->outputs_ = {"z"}; + auto op2 = std::shared_ptr( + new TestOp("test", {{"X", {"y"}}, {"W", {"w2"}}, {"b", {"b2"}}}, + {{"Out", {"z"}}}, {})); net->AddOp(op2); 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), paddle::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"}; + auto op1 = std::shared_ptr( + new framework::NOP("empty", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}}, + {{"Out", {"y"}}}, {})); net.AddOp(op1); net.InsertOp(0, op1); ASSERT_EQ(2UL, net.ops_.size()); diff --git a/paddle/operators/recurrent_op.cc b/paddle/operators/recurrent_op.cc index aeb95569b728f53b288a0c9a28220be8b5f7aaa4..78ce0ba3c0fa4fe380e49a848c2434fe593cd00b 100644 --- a/paddle/operators/recurrent_op.cc +++ b/paddle/operators/recurrent_op.cc @@ -14,166 +14,18 @@ #include "paddle/operators/recurrent_op.h" -#include #include #include #include "paddle/framework/op_registry.h" #include "paddle/operators/net_op.h" -#include "paddle/platform/enforce.h" namespace paddle { namespace operators { -namespace rnn { - -void SegmentInputs(const std::vector& step_scopes, - const std::vector& inlinks, - const size_t seq_len, - bool infer_shape_mode) { - PADDLE_ENFORCE(!inlinks.empty(), "no in links are provided."); - for (size_t i = 0; i < inlinks.size(); ++i) { - auto input_var = step_scopes[0]->FindVar(inlinks[i].external); - PADDLE_ENFORCE(input_var != nullptr, - "input link [%s] is not in scope.", - inlinks[i].external); - Tensor* input = input_var->GetMutable(); - DDim dims = input->dims(); - PADDLE_ENFORCE(static_cast(dims[0]) == seq_len, - "all the inlinks must have same length"); - DDim step_dims = slice_ddim(dims, 1, dims.size()); - for (size_t j = 0; j < seq_len; j++) { - Tensor* step_input = - step_scopes[j]->NewVar(inlinks[i].internal)->GetMutable(); - if (!infer_shape_mode) { - *step_input = input->Slice(j, j + 1); - } - step_input->Resize(step_dims); - } - } -} - -void ConcatOutputs(const std::vector& step_scopes, - const std::vector& outlinks, - const size_t seq_len, - bool infer_shape_mode) { - for (size_t i = 0; i < outlinks.size(); i++) { - auto output_var = step_scopes[0]->FindVar(outlinks[i].external); - PADDLE_ENFORCE(output_var != nullptr, - "output link [%s] is not in scope.", - outlinks[i].external); - Tensor* output = output_var->GetMutable(); - if (infer_shape_mode) { - DDim step_dims = step_scopes[0] - ->FindVar(outlinks[i].internal) - ->GetMutable() - ->dims(); - std::vector dims_vec = vectorize(step_dims); - dims_vec.insert(dims_vec.begin(), seq_len); - output->Resize(make_ddim(dims_vec)); - } else { - output->mutable_data(platform::CPUPlace()); - for (size_t j = 0; j < seq_len; j++) { - Tensor* step_output = - step_scopes[j]->FindVar(outlinks[i].internal)->GetMutable(); - // TODO(luotao02) data type and platform::DeviceContext() should set - // correctly - (output->Slice(j, j + 1)) - .CopyFrom(*step_output, platform::CPUPlace()); - } - } - } -} - -void LinkMemories(const std::vector& scopes, - const std::vector& memories, - const size_t step_id, - const int offset, - bool infer_shape_mode) { - PADDLE_ENFORCE(step_id < scopes.size(), - "step [%d] is out of range of step scopes' size [%d]", - step_id, - scopes.size()); - PADDLE_ENFORCE(static_cast(step_id) + offset >= 0, - "offset [%d] must be large than -[%d]", - offset, - step_id); - PADDLE_ENFORCE(step_id + offset < scopes.size(), - "offset [%d] is out of range, it must be less than (%d - %d)", - offset, - scopes.size(), - step_id); - auto scope = scopes[step_id]; - auto linked_scope = scopes[step_id + offset]; - for (auto& attr : memories) { - auto mem = scope->FindVar(attr.pre_var)->GetMutable(); - auto linked_mem = linked_scope->FindVar(attr.var)->GetMutable(); - if (infer_shape_mode) { - mem->Resize(linked_mem->dims()); - } else { - mem->ShareDataWith(*linked_mem); - } - } -} - -void InitArgument(const ArgumentName& name, - Argument* arg, - const 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); - PADDLE_ENFORCE(inlinks.size() == inlink_alias.size(), - "the size of inlinks and inlink_alias don't match:%d,%d", - inlinks.size(), - inlink_alias.size()); - for (size_t i = 0; i < inlinks.size(); ++i) { - rnn::Link link; - link.external = inlinks[i]; - link.internal = inlink_alias[i]; - (arg->inlinks).push_back(link); - } - - auto outlinks = op.Outputs(name.outlinks); - auto outlink_alias = op.GetAttr>(name.outlink_alias); - PADDLE_ENFORCE(outlinks.size() == outlink_alias.size(), - "the size of outlinks and outlink_alias don't match:%d,%d", - outlinks.size(), - outlink_alias.size()); - for (size_t i = 0; i < outlinks.size(); ++i) { - rnn::Link link; - link.external = outlinks[i]; - link.internal = outlink_alias[i]; - (arg->outlinks).push_back(link); - } - - auto boot_memories = op.Inputs(name.boot_memories); - - // attributes - auto memories = op.GetAttr>(name.memories); - auto pre_memories = op.GetAttr>(name.pre_memories); - - PADDLE_ENFORCE(memories.size() == boot_memories.size(), - "the size of memories, boot_memories don't match:%d,%d", - memories.size(), - boot_memories.size()); - PADDLE_ENFORCE(pre_memories.size() == boot_memories.size(), - "the size of pre_memories, boot_memories don't match:%d,%d", - pre_memories.size(), - boot_memories.size()); - PADDLE_ENFORCE(memories.size() > 0, "more than 1 memories should be set"); - - for (size_t i = 0; i < memories.size(); ++i) { - rnn::MemoryAttr mem_attr; - mem_attr.var = memories[i]; - mem_attr.pre_var = pre_memories[i]; - mem_attr.boot_var = boot_memories[i]; - (arg->memories).push_back(mem_attr); - } -} - -} // namespace rnn +using Scope = framework::Scope; +using Variable = framework::Variable; +using Tensor = framework::Tensor; void RecurrentAlgorithm::InferShape(const Scope& scope) const { seq_len_ = scope.FindVar((arg_->inlinks[0]).external) @@ -181,59 +33,70 @@ void RecurrentAlgorithm::InferShape(const Scope& scope) const { ->dims()[0]; CreateScopes(scope); auto step_scopes = GetStepScopes(scope); - rnn::SegmentInputs( - step_scopes, arg_->inlinks, seq_len_, true /*infer_shape_mode*/); + 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*/); + 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*/); + rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, + true /*infer_shape_mode*/); } void RecurrentAlgorithm::Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const { auto step_scopes = GetStepScopes(scope); - rnn::SegmentInputs( - step_scopes, arg_->inlinks, seq_len_, false /*infer_shape_mode*/); + 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 if (step_id > 0) { - rnn::LinkMemories( - step_scopes, arg_->memories, step_id, -1, false /*infer_shape_mode*/); + rnn::LinkMemories(step_scopes, arg_->memories, step_id, -1, + 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*/); + rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, + false /*infer_shape_mode*/); } void RecurrentAlgorithm::CreateScopes(const Scope& scope) const { - // TODO(xxx) Only two scopes are needed for inference, this case will be + // TODO(superjom) Only two scopes are needed for inference, this case will be // supported later. - auto step_scopes = - scope.FindVar(arg_->step_scopes)->GetMutable>(); + auto step_scopes_var = scope.FindVar(arg_->step_scopes); + PADDLE_ENFORCE(step_scopes_var != nullptr, ""); + auto step_scopes = step_scopes_var->GetMutable>(); + + // Now all variables in scope must be created outside of op. + PADDLE_ENFORCE_NOT_NULL(stepnet_); + PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "stepnet_ op has no outputs"); + PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "net_op has no outputs"); if (seq_len_ > step_scopes->size()) { for (size_t i = step_scopes->size(); i < seq_len_; ++i) { auto& step_scope = scope.NewScope(); - // Now all variables in scope must be created outside of op. - auto net_op = scope.FindVar(arg_->step_net)->GetMutable(); - for (auto& input : net_op->inputs_) { + // create step net's temp inputs + for (auto& input : (*stepnet_)->Inputs()) { // the weight are located in parent scope - if (!step_scope.FindVar(input)) step_scope.NewVar(input); + for (auto& var_name : input.second) { + if (!step_scope.FindVar(var_name)) { + step_scope.NewVar(var_name)->GetMutable(); + } + } } - for (auto& output : net_op->outputs_) { - step_scope.NewVar(output); + // create stepnet's outputs + for (const auto& output : (*stepnet_)->Outputs()) { + for (auto& var_name : output.second) { + step_scope.NewVar(var_name); + } } step_scopes->emplace_back(&step_scope); } @@ -245,61 +108,53 @@ void RecurrentAlgorithm::InitMemories(Scope* step_scope, for (auto& attr : arg_->memories) { Tensor* pre_mem = step_scope->NewVar(attr.pre_var)->GetMutable(); PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr, - "memory [%s]'s boot variable [%s] not exists", - attr.var, + "memory [%s]'s boot variable [%s] not exists", attr.var, attr.boot_var); Tensor* boot_mem = step_scope->FindVar(attr.boot_var)->GetMutable(); if (infer_shape_mode) { pre_mem->Resize(boot_mem->dims()); + PADDLE_ENFORCE_EQ(pre_mem->dims().size(), 2); } else { pre_mem->ShareDataWith(*boot_mem); } } } -const rnn::ArgumentName RecurrentOp::kArgName{"step_net", - "step_scopes", - "inlinks", - "outlinks", - "inlink_alias", - "outlink_alias", - "memories", - "pre_memories", - "boot_memories"}; - -const rnn::ArgumentName RecurrentGradientOp::kArgName{"step_net", - "step_scopes", - "outlink@grad", - "inlink@grad", - "inlink_alias", - "outlink_alias", - "memories", - "pre_memories", - "boot_memories@grad"}; - -void RecurrentOp::Init() { - OperatorBase::Init(); - std::unique_ptr arg(new rnn::Argument()); - rnn::InitArgument(kArgName, arg.get(), *this); - alg_.Init(std::move(arg)); +const rnn::ArgumentName RecurrentOp::kArgName{ + "step_net", "step_scopes", "inlinks", + "outlinks", "inlink_alias", "outlink_alias", + "memories", "pre_memories", "boot_memories"}; + +const rnn::ArgumentName RecurrentGradientOp::kArgName{ + "step_net", "step_scopes", "outlink@grad", + "inlink@grad", "inlink_alias", "outlink_alias", + "memories", "pre_memories", "boot_memories@grad"}; + +RecurrentOp::RecurrentOp(const std::string& type, + const framework::OperatorBase::VarNameMap& inputs, + const framework::OperatorBase::VarNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) { + rnn::InitArgument(kArgName, &arg_, *this); + alg_.Init(&arg_, &stepnet_); } -class RecurrentAlgorithmProtoAndCheckerMaker : public OpProtoAndCheckerMaker { -public: - RecurrentAlgorithmProtoAndCheckerMaker(OpProto* proto, - OpAttrChecker* op_checker) +class RecurrentAlgorithmProtoAndCheckerMaker + : public framework::OpProtoAndCheckerMaker { + public: + RecurrentAlgorithmProtoAndCheckerMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { const auto& name = RecurrentOp::kArgName; // 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 @@ -316,31 +171,27 @@ public: void RecurrentGradientAlgorithm::Run( const Scope& scope, const platform::DeviceContext& dev_ctx) const { auto step_scopes = GetStepScopes(scope); - rnn::SegmentInputs( - step_scopes, arg_->inlinks, seq_len_, false /*infer_shape_mode*/); - Variable* net = scope.FindVar(arg_->step_net); - PADDLE_ENFORCE(net != nullptr, "failed to get step net"); + rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, + false /*infer_shape_mode*/); for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) { if (static_cast(step_id) != seq_len_ - 1) { - rnn::LinkMemories( - step_scopes, arg_->memories, step_id, 1, false /*infer_shape_mode*/); + 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_, false /*infer_shape_mode*/); + rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, + false /*infer_shape_mode*/); } void RecurrentGradientAlgorithm::LinkBootMemoryGradients( Scope* step_scope, bool infer_shape_mode) const { for (auto& attr : arg_->memories) { PADDLE_ENFORCE(step_scope->FindVar(attr.var) != nullptr, - "memory variable [%s] does not exists", - attr.var); + "memory variable [%s] does not exists", attr.var); PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr, - "boot variable [%s] does not exists", - attr.boot_var); + "boot variable [%s] does not exists", attr.boot_var); Tensor* mem_grad = step_scope->NewVar(attr.var)->GetMutable(); Tensor* boot_mem_grad = step_scope->NewVar(attr.boot_var)->GetMutable(); @@ -357,32 +208,32 @@ void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const { ->GetMutable() ->dims()[0]; 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"); + rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, + true /*infer_shape_mode*/); for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) { if (static_cast(step_id) != seq_len_ - 1) { - rnn::LinkMemories( - step_scopes, arg_->memories, step_id, 1, true /*infer_shape_mode*/); + 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*/); + 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::OperatorBase::VarNameMap& inputs, + const framework::OperatorBase::VarNameMap& 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_op, paddle::operators::RecurrentOp, + paddle::operators::RecurrentAlgorithmProtoAndCheckerMaker); diff --git a/paddle/operators/recurrent_op.h b/paddle/operators/recurrent_op.h index 2a0964fff326500b6215dd4afac63c75d64c4a06..171a0bd2ae80245939a9237f51d195e8bc9178fc 100644 --- a/paddle/operators/recurrent_op.h +++ b/paddle/operators/recurrent_op.h @@ -15,86 +15,14 @@ #pragma once #include "paddle/framework/operator.h" +#include "paddle/operators/net_op.h" +#include "paddle/operators/rnn/recurrent_op_utils.h" namespace paddle { namespace operators { -using namespace paddle::framework; - -namespace rnn { - -/** - * Memory of a RNN (same as the role of `Momory` in PaddlePaddle). - * - * Memory attributes cached by this op, dims will be infered from - * boot memories in father scope. Other attributes are copied from Op's proto - * attributes. - */ -struct MemoryAttr { - // name of current state variable - std::string var; - // name of previous step's state variable - std::string pre_var; - // name of the variables to init this memory (same role of `boot_layer` in - // PaddlePaddle), which is store in father's scope. - std::string boot_var; -}; - -struct Link { - // input or output links name. - std::string internal; - // alias to avoid duplicate keys in scopes. - std::string external; -}; - -struct Argument { - std::string step_net; - std::string step_scopes; - std::vector inlinks; - std::vector outlinks; - std::vector memories; -}; - -struct ArgumentName { - std::string step_net; - std::string step_scopes; - std::string inlinks; - std::string outlinks; - std::string inlink_alias; // the alias of inlinks in step net. - std::string outlink_alias; // the alias of outlinks in step net. - std::string memories; // the memory name - std::string pre_memories; // the previous memory name - std::string boot_memories; // the boot memory name -}; - -/** - * Prepare inputs for each step net. - */ -void SegmentInputs(const std::vector& step_scopes, - const std::vector& inlinks, - const size_t seq_len, - bool infer_shape_mode); - -/** - * Process outputs of step nets and merge to variables. - */ -void ConcatOutputs(const std::vector& step_scopes, - const std::vector& outlinks, - const size_t seq_len, - bool infer_shape_mode); - -void LinkMemories(const std::vector& step_scopes, - const std::vector& memories, - const size_t step_id, - const int offset, - bool infer_shape_mode); - -void InitArgument(const ArgumentName& name, Argument* arg); - -}; // namespace rnn - // The sequence format in RecurrentOp is Tensor now. -// TODO: +// TODO(Yan Chunwei): // 1. No-padding computing for sequences with indifinite length in one batch. // 2. Hierarchical RNN for sequence with sub-sequence. // 3. Internal Memory. @@ -102,33 +30,41 @@ void InitArgument(const ArgumentName& name, Argument* arg); // Refer to: https://arxiv.org/pdf/1502.02367.pdf class RecurrentAlgorithm { -public: - void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const; - - void Init(std::unique_ptr arg) { arg_ = std::move(arg); } + public: + void Run(const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const; + + void Init(rnn::Argument* arg, std::shared_ptr* stepnet) { + PADDLE_ENFORCE_NOT_NULL(stepnet, "stepnet should be set before."); + arg_ = arg; + stepnet_ = stepnet; + } /** * InferShape must be called before Run. */ - void InferShape(const Scope& scope) const; + void InferShape(const framework::Scope& scope) const; -protected: + protected: /* * The step scopes will be stored in the father scope as a variable. * * NOTE the scopes are reused in both the forward and backward, so just * create once and expand its size if more steps need. */ - void CreateScopes(const Scope& scope) const; + void CreateScopes(const framework::Scope& scope) const; - const std::vector& GetStepScopes(const Scope& scope) const { - return *scope.FindVar(arg_->step_scopes)->GetMutable>(); + const std::vector& GetStepScopes( + const framework::Scope& scope) const { + return *scope.FindVar(arg_->step_scopes) + ->GetMutable>(); } - void InitMemories(Scope* step_scopes, bool infer_shape_mode) const; + void InitMemories(framework::Scope* step_scopes, bool infer_shape_mode) const; -private: - std::unique_ptr arg_; + private: + std::shared_ptr* stepnet_; + rnn::Argument* arg_; mutable size_t seq_len_; }; @@ -143,70 +79,91 @@ class RecurrentGradientAlgorithm { * lot, and the latter is a wrapper acts like an dapter for it to make RNN an * operator. */ -public: - void Init(std::unique_ptr arg) { arg_ = std::move(arg); } + public: + void Init(rnn::Argument* arg, std::shared_ptr* stepnet) { + PADDLE_ENFORCE_NOT_NULL(stepnet, "stepnet should be set before."); + arg_ = std::move(arg); + stepnet_ = stepnet; + } - void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const; + void Run(const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const; - void LinkBootMemoryGradients(Scope* step_scopes, bool infer_shape_mode) const; + void LinkBootMemoryGradients(framework::Scope* step_scopes, + bool infer_shape_mode) const; /** * InferShape must be called before Run. */ - void InferShape(const Scope& scope) const; + void InferShape(const framework::Scope& scope) const; -protected: - inline const std::vector& GetStepScopes(const Scope& scope) const { - return *scope.FindVar(arg_->step_scopes)->GetMutable>(); + protected: + inline const std::vector& GetStepScopes( + const framework::Scope& scope) const { + return *scope.FindVar(arg_->step_scopes) + ->GetMutable>(); } -private: - std::unique_ptr arg_; + private: + rnn::Argument* arg_; mutable size_t seq_len_; + std::shared_ptr* stepnet_; }; -class RecurrentOp final : public OperatorBase { -public: - void Init() override; - +class RecurrentOp final : public framework::OperatorBase { + public: + RecurrentOp(const std::string& type, const VarNameMap& inputs, + const VarNameMap& outputs, const framework::AttributeMap& attrs); /** - * InferShape must be called before Run. - */ - virtual void InferShape(const Scope& scope) const override { + * InferShape must be called before Run. + */ + void InferShape(const framework::Scope& scope) const override { alg_.InferShape(scope); } - virtual void Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const override { + void Run(const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const override { alg_.Run(scope, dev_ctx); } + void set_stepnet(std::shared_ptr net) { stepnet_ = net; } + const NetOp& stepnet() const { return *stepnet_; } + static const rnn::ArgumentName kArgName; -private: + private: RecurrentAlgorithm alg_; + rnn::Argument arg_; + std::shared_ptr stepnet_; }; -class RecurrentGradientOp final : public OperatorBase { -public: - void Init() override; +class RecurrentGradientOp final : public framework::OperatorBase { + public: + RecurrentGradientOp(const std::string& type, const VarNameMap& inputs, + const VarNameMap& outputs, + const framework::AttributeMap& attrs); /** * InferShape must be called before Run. */ - virtual void InferShape(const Scope& scope) const override { + void InferShape(const framework::Scope& scope) const override { alg_.InferShape(scope); } - virtual void Run(const Scope& scope, - const platform::DeviceContext& dev_ctx) const override { + void Run(const framework::Scope& scope, + const platform::DeviceContext& dev_ctx) const override { alg_.Run(scope, dev_ctx); } static const rnn::ArgumentName kArgName; -private: + void set_stepnet(const std::shared_ptr& net) { stepnet_ = net; } + const NetOp& stepnet() const { return *stepnet_; } + + private: RecurrentGradientAlgorithm alg_; + std::shared_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 08a6d9fe5681fdea180de2e9361734ade8564775..0000000000000000000000000000000000000000 --- a/paddle/operators/recurrent_op_test.cc +++ /dev/null @@ -1,393 +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/op_registry.h" -#include "paddle/framework/operator.h" -#include "paddle/framework/tensor.h" -#include "paddle/operators/net_op.h" - -namespace paddle { -namespace operators { - -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() { - 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); diff --git a/paddle/operators/rnn/recurrent_op_utils.cc b/paddle/operators/rnn/recurrent_op_utils.cc new file mode 100644 index 0000000000000000000000000000000000000000..a9b65c30f25554e54e9fd7103f240946a93566e2 --- /dev/null +++ b/paddle/operators/rnn/recurrent_op_utils.cc @@ -0,0 +1,160 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/rnn/recurrent_op_utils.h" + +namespace paddle { +namespace operators { +namespace rnn { + +namespace f = paddle::framework; + +using Tensor = framework::Tensor; + +void SegmentInputs(const std::vector& step_scopes, + const std::vector& inlinks, const size_t seq_len, + bool infer_shape_mode) { + PADDLE_ENFORCE(!inlinks.empty(), "no in links are provided."); + for (size_t i = 0; i < inlinks.size(); ++i) { + auto input_var = step_scopes[0]->FindVar(inlinks[i].external); + PADDLE_ENFORCE(input_var != nullptr, "input link [%s] is not in scope.", + inlinks[i].external); + + Tensor* input = input_var->GetMutable(); + f::DDim dims = input->dims(); + PADDLE_ENFORCE(static_cast(dims[0]) == seq_len, + "all the inlinks must have same length"); + f::DDim step_dims = slice_ddim(dims, 1, dims.size()); + for (size_t j = 0; j < seq_len; j++) { + Tensor* step_input = + step_scopes[j]->NewVar(inlinks[i].internal)->GetMutable(); + if (!infer_shape_mode) { + *step_input = input->Slice(j, j + 1); + } + step_input->Resize(step_dims); + } + } +} + +void ConcatOutputs(const std::vector& step_scopes, + const std::vector& outlinks, const size_t seq_len, + bool infer_shape_mode) { + for (size_t i = 0; i < outlinks.size(); i++) { + auto output_var = step_scopes[0]->FindVar(outlinks[i].external); + PADDLE_ENFORCE(output_var != nullptr, "output link [%s] is not in scope.", + outlinks[i].external); + Tensor* output = output_var->GetMutable(); + + if (infer_shape_mode) { + auto step_scope_var = step_scopes[0]->FindVar(outlinks[i].internal); + PADDLE_ENFORCE(step_scope_var != nullptr, "%s not in scope", + outlinks[i].internal); + f::DDim step_dims = step_scope_var->template GetMutable()->dims(); + std::vector dims_vec = vectorize(step_dims); + dims_vec.insert(dims_vec.begin(), seq_len); + output->Resize(f::make_ddim(dims_vec)); + } else { + output->mutable_data(platform::CPUPlace()); + for (size_t j = 0; j < seq_len; j++) { + Tensor* step_output = + step_scopes[j]->FindVar(outlinks[i].internal)->GetMutable(); + // TODO(luotao02) data type and platform::DeviceContext() should set + // correctly + (output->Slice(j, j + 1)) + .CopyFrom(*step_output, platform::CPUPlace()); + } + } + } +} + +void LinkMemories(const std::vector& scopes, + const std::vector& memories, + const size_t step_id, const int offset, + bool infer_shape_mode) { + PADDLE_ENFORCE_LT(step_id, scopes.size(), + "step [%d] is out of range of step scopes' size [%d]", + step_id, scopes.size()); + PADDLE_ENFORCE_GE(static_cast(step_id) + offset, 0, + "offset [%d] must be large than -[%d]", offset, step_id); + PADDLE_ENFORCE_LT( + step_id + offset, scopes.size(), + "offset [%d] is out of range, it must be less than (%d - %d)", offset, + scopes.size(), step_id); + auto scope = scopes[step_id]; + auto linked_scope = scopes[step_id + offset]; + for (auto& attr : memories) { + auto mem = scope->FindVar(attr.pre_var)->GetMutable(); + auto linked_mem = linked_scope->FindVar(attr.var)->GetMutable(); + if (infer_shape_mode) { + mem->Resize(linked_mem->dims()); + } else { + mem->ShareDataWith(*linked_mem); + } + } +} + +void InitArgument(const ArgumentName& name, Argument* arg, + const framework::OperatorBase& op) { + arg->step_scopes = op.Output(name.step_scopes); + + auto inlinks = op.Inputs(name.inlinks); + auto inlink_alias = op.GetAttr>(name.inlink_alias); + PADDLE_ENFORCE(inlinks.size() == inlink_alias.size(), + "the size of inlinks and inlink_alias don't match:%d,%d", + inlinks.size(), inlink_alias.size()); + for (size_t i = 0; i < inlinks.size(); ++i) { + rnn::Link link; + link.external = inlinks[i]; + link.internal = inlink_alias[i]; + (arg->inlinks).push_back(link); + } + + auto outlinks = op.Outputs(name.outlinks); + auto outlink_alias = op.GetAttr>(name.outlink_alias); + PADDLE_ENFORCE(outlinks.size() == outlink_alias.size(), + "the size of outlinks and outlink_alias don't match:%d,%d", + outlinks.size(), outlink_alias.size()); + for (size_t i = 0; i < outlinks.size(); ++i) { + rnn::Link link; + link.external = outlinks[i]; + link.internal = outlink_alias[i]; + (arg->outlinks).push_back(link); + } + + auto boot_memories = op.Inputs(name.boot_memories); + + // attributes + auto memories = op.GetAttr>(name.memories); + auto pre_memories = op.GetAttr>(name.pre_memories); + + PADDLE_ENFORCE(memories.size() == boot_memories.size(), + "the size of memories, boot_memories don't match:%d,%d", + memories.size(), boot_memories.size()); + PADDLE_ENFORCE(pre_memories.size() == boot_memories.size(), + "the size of pre_memories, boot_memories don't match:%d,%d", + pre_memories.size(), boot_memories.size()); + PADDLE_ENFORCE(memories.size() > 0, "more than 1 memories should be set"); + + for (size_t i = 0; i < memories.size(); ++i) { + rnn::MemoryAttr mem_attr; + mem_attr.var = memories[i]; + mem_attr.pre_var = pre_memories[i]; + mem_attr.boot_var = boot_memories[i]; + (arg->memories).push_back(mem_attr); + } +} + +} // namespace rnn +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/rnn/recurrent_op_utils.h b/paddle/operators/rnn/recurrent_op_utils.h new file mode 100644 index 0000000000000000000000000000000000000000..17941c503cfcc83415b8bc635623a2c2ce2981c3 --- /dev/null +++ b/paddle/operators/rnn/recurrent_op_utils.h @@ -0,0 +1,94 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include + +#include "paddle/framework/operator.h" + +namespace paddle { +namespace operators { +namespace rnn { + +using Scope = framework::Scope; + +/** + * Memory of a RNN (same as the role of `Momory` in PaddlePaddle). + * + * Memory attributes cached by this op, dims will be infered from + * boot memories in father scope. Other attributes are copied from Op's proto + * attributes. + */ +struct MemoryAttr { + // name of current state variable + std::string var; + // name of previous step's state variable + std::string pre_var; + // name of the variables to init this memory (same role of `boot_layer` in + // PaddlePaddle), which is store in father's scope. + std::string boot_var; +}; + +struct Link { + // input or output links name. + std::string internal; + // alias to avoid duplicate keys in scopes. + std::string external; +}; + +struct Argument { + std::string step_net; + std::string step_scopes; + std::vector inlinks; + std::vector outlinks; + std::vector memories; +}; + +struct ArgumentName { + std::string step_net; + std::string step_scopes; + std::string inlinks; + std::string outlinks; + std::string inlink_alias; // the alias of inlinks in step net. + std::string outlink_alias; // the alias of outlinks in step net. + std::string memories; // the memory name + std::string pre_memories; // the previous memory name + std::string boot_memories; // the boot memory name +}; + +/** + * Prepare inputs for each step net. + */ +void SegmentInputs(const std::vector& step_scopes, + const std::vector& inlinks, const size_t seq_len, + bool infer_shape_mode); + +/** + * Process outputs of step nets and merge to variables. + */ +void ConcatOutputs(const std::vector& step_scopes, + const std::vector& outlinks, const size_t seq_len, + bool infer_shape_mode); + +void LinkMemories(const std::vector& step_scopes, + const std::vector& memories, const size_t step_id, + const int offset, bool infer_shape_mode); + +void InitArgument(const ArgumentName& name, Argument* arg, + const framework::OperatorBase& op); + +} // namespace rnn +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/rowwise_add_op.cc b/paddle/operators/rowwise_add_op.cc index 2ad2b66c8f385c858eb34c7ea766f168de9c817e..8375d988045dc24fa1109646b46ff477e2a78132 100644 --- a/paddle/operators/rowwise_add_op.cc +++ b/paddle/operators/rowwise_add_op.cc @@ -13,28 +13,31 @@ limitations under the License. */ #include "paddle/operators/rowwise_add_op.h" + namespace paddle { namespace operators { -class RowWiseAddOp : public OperatorWithKernel { -protected: - void InferShape(const 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(); +class RowWiseAddOp : 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("b")->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()); + PADDLE_ENFORCE(ctx.OutputSize("Out") == 1, "The output size must be 1"); + ctx.Output("Out")->Resize(ctx.Input("X")->dims()); } }; -class RowWiseAddOpMaker : public OpProtoAndCheckerMaker { -public: - RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker) +class RowWiseAddOpMaker : public framework::OpProtoAndCheckerMaker { + public: + RowWiseAddOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The left input of row-wise add op, must be matrix"); AddInput("b", "The right input of row-wise add op, must be vector"); @@ -50,6 +53,8 @@ for i in xrange(X.shape[0]): } // namespace operators } // namespace paddle -REGISTER_OP(rowwise_add, ops::RowWiseAddOp, ops::RowWiseAddOpMaker); -REGISTER_OP_CPU_KERNEL(rowwise_add, - ops::RowWiseAddKernel); +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(rowwise_add, ops::RowWiseAddOp, + ops::RowWiseAddOpMaker); +REGISTER_OP_CPU_KERNEL( + rowwise_add, ops::RowWiseAddKernel); diff --git a/paddle/operators/rowwise_add_op.cu b/paddle/operators/rowwise_add_op.cu index 82338ceccc06653791b26472e18d804f62735649..86f80b81228a69ac4c05a4693901570f2b9966e0 100644 --- a/paddle/operators/rowwise_add_op.cu +++ b/paddle/operators/rowwise_add_op.cu @@ -1,5 +1,20 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + #define EIGEN_USE_GPU #include "paddle/operators/rowwise_add_op.h" -REGISTER_OP_GPU_KERNEL(rowwise_add, - ops::RowWiseAddKernel); +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL( + rowwise_add, ops::RowWiseAddKernel); diff --git a/paddle/operators/rowwise_add_op.h b/paddle/operators/rowwise_add_op.h index bd4d1128955fb718d3a84dfd96d8c68d7196e9cc..01f88f2198774fbaa4c98ff9bf286f2f08496a9a 100644 --- a/paddle/operators/rowwise_add_op.h +++ b/paddle/operators/rowwise_add_op.h @@ -13,20 +13,29 @@ limitations under the License. */ #pragma once -#include "paddle/operators/type_alias.h" +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" namespace paddle { namespace operators { +using Tensor = framework::Tensor; +template +using EigenVector = framework::EigenVector; +template +using EigenMatrix = framework::EigenMatrix; + template -class RowWiseAddKernel : public OpKernel { -public: - void Compute(const ExecutionContext& context) const override { - auto out = context.Output(0); +class RowWiseAddKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + 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 input = EigenMatrix::From(*context.Input("X")); + auto bias = EigenVector::From(*context.Input("b")); auto output = EigenMatrix::From(*out); const int bias_size = bias.dimension(0); 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_test.cc b/paddle/operators/scatter_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..4449ce6564396f1971506efb7458c00c834db19f --- /dev/null +++ b/paddle/operators/scatter_test.cc @@ -0,0 +1,52 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#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)); +} diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc index 9a84dc8af3b3e649b776ca8a97dedba1fa3ff48d..ad267e7f087943ff3b8326a7baf2ce3955fa51c2 100644 --- a/paddle/operators/sgd_op.cc +++ b/paddle/operators/sgd_op.cc @@ -17,23 +17,22 @@ limitations under the License. */ namespace paddle { namespace operators { -class SGDOp : public OperatorWithKernel { -protected: - void InferShape(const InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 2, "Input size of SGDOp must be two"); - PADDLE_ENFORCE(ctx.OutputSize() == 1, "Output size of SGDOp must be one"); - PADDLE_ENFORCE(ctx.InputVar(0) != nullptr, "inputs[0] mast be set"); - PADDLE_ENFORCE(ctx.InputVar(1) != nullptr, "inputs[1] mast be set"); - PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr, "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()); +class SGDOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + 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()); } }; -class SGDOpMaker : public OpProtoAndCheckerMaker { -public: - SGDOpMaker(OpProto *proto, OpAttrChecker *op_checker) +class SGDOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SGDOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("param", "input parameter"); AddInput("grad", "input gradient"); @@ -51,5 +50,7 @@ param_out = param - learning_rate * grad; } // namespace operators } // namespace paddle -REGISTER_OP(sgd, ops::SGDOp, ops::SGDOpMaker); -REGISTER_OP_CPU_KERNEL(sgd, ops::SGDOpKernel); +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(sgd, ops::SGDOp, ops::SGDOpMaker); +REGISTER_OP_CPU_KERNEL(sgd, + ops::SGDOpKernel); diff --git a/paddle/operators/sgd_op.cu b/paddle/operators/sgd_op.cu index d79258cbf13c699cfb2afaee229cf96a3e377b5e..f5ba6d3c29f8dfbfdea4fbf2c3d5fd7f5b358666 100644 --- a/paddle/operators/sgd_op.cu +++ b/paddle/operators/sgd_op.cu @@ -1,4 +1,20 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + #define EIGEN_USE_GPU #include "paddle/operators/sgd_op.h" -REGISTER_OP_GPU_KERNEL(sgd, ops::SGDOpKernel); \ No newline at end of file +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(sgd, + ops::SGDOpKernel); diff --git a/paddle/operators/sgd_op.h b/paddle/operators/sgd_op.h index 0c3a240f9a4a5fc7bc4898e82786810cee2f7010..bfb449d0b029409eda4177fc7643810ee6a1df3d 100644 --- a/paddle/operators/sgd_op.h +++ b/paddle/operators/sgd_op.h @@ -13,15 +13,21 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once -#include "paddle/operators/type_alias.h" +#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 SGDOpKernel : public OpKernel { -public: - void Compute(const ExecutionContext& ctx) const override { +class SGDOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { auto param = ctx.Input("param"); auto grad = ctx.Input("grad"); auto param_out = ctx.Output(0); diff --git a/paddle/operators/sgd_op_test.cc b/paddle/operators/sgd_op_test.cc deleted file mode 100644 index 75137259f5e608b259b073101353e5818bb17c92..0000000000000000000000000000000000000000 --- a/paddle/operators/sgd_op_test.cc +++ /dev/null @@ -1,22 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -USE_OP(sgd); -TEST(SGDOp, GetOpProto) { - auto& protos = paddle::framework::OpRegistry::protos(); - auto it = protos.find("sgd"); - ASSERT_NE(it, protos.end()); -} diff --git a/paddle/operators/sigmoid_op.cc b/paddle/operators/sigmoid_op.cc index a81ab262cc6fe7bdff0045259e0030f3d46f503f..d773a4f2d50e82146a729b1cda085ce86ade89cc 100644 --- a/paddle/operators/sigmoid_op.cc +++ b/paddle/operators/sigmoid_op.cc @@ -13,21 +13,24 @@ limitations under the License. */ #include "paddle/operators/sigmoid_op.h" + namespace paddle { namespace operators { -class SigmoidOp : public OperatorWithKernel { -protected: - void InferShape(const 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()); +class SigmoidOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + ctx.Output("Y")->Resize(ctx.Input("X")->dims()); } }; -class SigmoidOpMaker : public OpProtoAndCheckerMaker { -public: - SigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker) +class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SigmoidOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "sigmoid input"); AddOutput("Y", "sigmoid output"); @@ -35,19 +38,23 @@ public: } }; -class SigmoidOpGrad : public OperatorWithKernel { -protected: - void InferShape(const InferShapeContext &ctx) const override {} - std::string DebugString() const override { - LOG(INFO) << "SigmoidGrad"; - return ""; +class SigmoidOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + ctx.Output(0)->Resize(ctx.Input(0)->dims()); } }; } // namespace operators } // namespace paddle -REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker); -REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, ops::SigmoidOpGrad); - -REGISTER_OP_CPU_KERNEL(sigmoid, ops::SigmoidKernel); +namespace ops = paddle::operators; +REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker, sigmoid_grad, + ops::SigmoidOpGrad); +REGISTER_OP_CPU_KERNEL(sigmoid, + ops::SigmoidKernel); +REGISTER_OP_CPU_KERNEL( + sigmoid_grad, ops::SigmoidGradKernel); diff --git a/paddle/operators/sigmoid_op.cu b/paddle/operators/sigmoid_op.cu index c9d11a2e1f9dcc563765c9e8cc1bae6beff57f18..1a50dfe14a7b9e2614aadb7729de9f9e461e9905 100644 --- a/paddle/operators/sigmoid_op.cu +++ b/paddle/operators/sigmoid_op.cu @@ -1,4 +1,23 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + #define EIGEN_USE_GPU #include "paddle/operators/sigmoid_op.h" -REGISTER_OP_GPU_KERNEL(sigmoid, ops::SigmoidKernel); +namespace ops = paddle::operators; + +REGISTER_OP_GPU_KERNEL(sigmoid, + ops::SigmoidKernel); +REGISTER_OP_GPU_KERNEL( + sigmoid_grad, ops::SigmoidGradKernel); diff --git a/paddle/operators/sigmoid_op.h b/paddle/operators/sigmoid_op.h index 1412e4398440c8e946d3ab434a50e978079637ab..11ab923eb346c1f8de3a6bbebdfa874b6530004a 100644 --- a/paddle/operators/sigmoid_op.h +++ b/paddle/operators/sigmoid_op.h @@ -13,20 +13,26 @@ limitations under the License. */ #pragma once - -#include "paddle/operators/type_alias.h" +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" namespace paddle { namespace operators { +using Tensor = framework::Tensor; +template +using EigenVector = framework::EigenVector; + template -class SigmoidKernel : public OpKernel { -public: - void Compute(const ExecutionContext& context) const override { - auto input = context.Input(0); - auto output = context.Output(0); +class SigmoidKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto input = context.Input("X"); + auto output = context.Output("Y"); output->mutable_data(context.GetPlace()); + // The clipping is used in Paddle's raw implenmention auto X = EigenVector::Flatten(*input); auto Y = EigenVector::Flatten(*output); auto place = context.GetEigenDevice(); @@ -34,5 +40,23 @@ public: Y.device(place) = 1.0 / (1.0 + (-1.0 * X).exp()); } }; + +template +class SigmoidGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto Y_t = context.Input("Y"); + auto dY_t = context.Input(framework::GradVarName("Y")); + auto dX_t = context.Output(framework::GradVarName("X")); + + dX_t->mutable_data(context.GetPlace()); + + auto dX = EigenVector::Flatten(*dX_t); + auto Y = EigenVector::Flatten(*Y_t); + auto dY = EigenVector::Flatten(*dY_t); + dX.device(context.GetEigenDevice()) = dY * Y * (1. - Y); + } +}; + } // namespace operators } // namespace paddle diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc index 5cbb96ab754467ea6ddab9380ca25987c9376980..40c51a64c49bc064f55975ef6ced1d54070f1291 100644 --- a/paddle/operators/softmax_op.cc +++ b/paddle/operators/softmax_op.cc @@ -17,22 +17,22 @@ limitations under the License. */ namespace paddle { namespace operators { -class SoftmaxOp : public OperatorWithKernel { -protected: - void InferShape(const InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 1UL, - "Only one input is need for softmax"); +class SoftmaxOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { PADDLE_ENFORCE(ctx.Input("X")->dims().size() == 2UL, "The input of softmax op must be matrix"); - PADDLE_ENFORCE(ctx.OutputSize() == 1UL, - "Only one output is need for softmax"); ctx.Output("Y")->Resize(ctx.Input("X")->dims()); } }; -class SoftmaxOpMaker : public OpProtoAndCheckerMaker { -public: - SoftmaxOpMaker(OpProto *proto, OpAttrChecker *op_checker) +class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SoftmaxOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "input of softmax"); AddOutput("Y", "output of softmax"); @@ -40,20 +40,19 @@ public: } }; -class SoftmaxOpGrad : public OperatorWithKernel { -protected: - void InferShape(const InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 3UL, - "Input of SoftmaxOpGrad should be 3, X, Y, YG"); - PADDLE_ENFORCE(ctx.OutputSize() == 1UL, - "Output of SoftmaxOpGrad should be 1"); +class SoftmaxOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { PADDLE_ENFORCE(ctx.InputVar("Y") != nullptr, "Input(Y) should not be null"); - PADDLE_ENFORCE(ctx.InputVar(GRAD_VAR_NAME("Y")) != nullptr, - "Input(Y@GRAD) should not be 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(GRAD_VAR_NAME("Y"))->dims(), + ctx.Input(framework::GradVarName("Y"))->dims(), "the shape of Input(0) and Input(1) should be the same"); - ctx.Output(GRAD_VAR_NAME("X")) + ctx.Output(framework::GradVarName("X")) ->Resize(ctx.Input("Y")->dims()); } }; @@ -61,8 +60,11 @@ protected: } // namespace operators } // namespace paddle -REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker); -REGISTER_OP_CPU_KERNEL(softmax, ops::SoftmaxKernel); -REGISTER_GRADIENT_OP(softmax, softmax_grad, ops::SoftmaxOpGrad); -REGISTER_OP_CPU_KERNEL(softmax_grad, - ops::SoftmaxGradKernel); +namespace ops = paddle::operators; + +REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker, softmax_grad, + ops::SoftmaxOpGrad); +REGISTER_OP_CPU_KERNEL(softmax, + ops::SoftmaxKernel); +REGISTER_OP_CPU_KERNEL( + softmax_grad, ops::SoftmaxGradKernel); diff --git a/paddle/operators/softmax_op.cu b/paddle/operators/softmax_op.cu index 8c652213f2e4c0e0ea1a31987fcb37c86374cd2a..2e99a89699dbdcafc8055c47debf9e49f10507e6 100644 --- a/paddle/operators/softmax_op.cu +++ b/paddle/operators/softmax_op.cu @@ -1,6 +1,23 @@ +/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + #define EIGEN_USE_GPU -#include "paddle/framework/op_registry.h" #include "paddle/operators/softmax_op.h" -REGISTER_OP_GPU_KERNEL(softmax, ops::SoftmaxKernel); -REGISTER_OP_GPU_KERNEL(softmax_grad, ops::SoftmaxGradKernel); +namespace ops = paddle::operators; + +REGISTER_OP_GPU_KERNEL(softmax, + ops::SoftmaxKernel); +REGISTER_OP_GPU_KERNEL( + softmax_grad, ops::SoftmaxGradKernel); diff --git a/paddle/operators/softmax_op.h b/paddle/operators/softmax_op.h index 13e74a79077982e9fba5d90f40986e699c1ed897..4fa6b59540498638c3b7df639ae10a66c0fa1c16 100644 --- a/paddle/operators/softmax_op.h +++ b/paddle/operators/softmax_op.h @@ -13,19 +13,21 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once - -#include "paddle/framework/ddim.h" -#include "paddle/framework/operator.h" -#include "paddle/framework/tensor.h" -#include "paddle/operators/type_alias.h" +#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 SoftmaxKernel : public OpKernel { -public: - void Compute(const ExecutionContext& context) const override { +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()); @@ -62,14 +64,14 @@ public: }; template -class SoftmaxGradKernel : public OpKernel { -public: - void Compute(const ExecutionContext& context) const override { +class SoftmaxGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { std::shared_ptr scale_ = std::make_shared(); auto Y = context.Input("Y"); - auto dY = context.Input(OperatorBase::GRAD_VAR_NAME("Y")); - auto dX = context.Output(OperatorBase::GRAD_VAR_NAME("X")); + auto dY = context.Input(framework::GradVarName("Y")); + auto dX = context.Output(framework::GradVarName("X")); dX->mutable_data(context.GetPlace()); const int batch_size = Y->dims()[0]; diff --git a/paddle/operators/type_alias.h b/paddle/operators/type_alias.h deleted file mode 100644 index 931740e150946a939b8656be5a30185c6ee1cb8f..0000000000000000000000000000000000000000 --- a/paddle/operators/type_alias.h +++ /dev/null @@ -1,60 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ - -#pragma once - -#include "paddle/framework/eigen.h" -#include "paddle/framework/op_registry.h" -#include "paddle/operators/net_op.h" - -namespace paddle { -namespace operators { - -using OpKernel = framework::OpKernel; -using OperatorBase = framework::OperatorBase; -using InferShapeContext = framework::InferShapeContext; -using ExecutionContext = framework::ExecutionContext; -using Variable = framework::Variable; -template -using EigenScalar = framework::EigenScalar; -template -using EigenVector = framework::EigenVector; -template -using EigenMatrix = framework::EigenMatrix; -template -using EigenTensor = framework::EigenTensor; -using Tensor = framework::Tensor; -using Scope = framework::Scope; -using OperatorWithKernel = framework::OperatorWithKernel; -using OperatorBase = framework::OperatorBase; -using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker; -using OpProto = framework::OpProto; -using OpAttrChecker = framework::OpAttrChecker; -using CPUPlace = platform::CPUPlace; -using GPUPlace = platform::GPUPlace; -using OpRegistry = framework::OpRegistry; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a0a0d4d914b37fca4250e5218a953f573611a086 --- /dev/null +++ b/paddle/operators/uniform_random_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 +#include +#include "paddle/framework/op_registry.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace operators { + +// It seems that Eigen::Tensor::random in GPU will SEGFAULT. +// Use std::random and thrust::random(thrust is a std library in CUDA) to +// implement uniform random. +template +class CPUUniformRandomKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* tensor = context.Output("Out"); + T* data = tensor->mutable_data(context.GetPlace()); + unsigned int seed = + static_cast(context.op_.GetAttr("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) { + data[i] = dist(engine); + } + } +}; + +class UniformRandomOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext& ctx) const override { + PADDLE_ENFORCE(GetAttr("min") < GetAttr("max"), + "uniform_random's min must less then max"); + auto* tensor = ctx.Output("Out"); + auto dims = GetAttr>("dims"); + tensor->Resize(framework::make_ddim(dims)); + } +}; + +class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker { + public: + UniformRandomOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : 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"); + AddAttr("min", "Minimum value of uniform random").SetDefault(-1.0f); + AddAttr("max", "Maximun value of uniform random").SetDefault(1.0f); + AddAttr("seed", + "Random seed of uniform random. " + "0 means generate a seed by system") + .SetDefault(0); + } +}; +} // namespace operators +} // namespace paddle + +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 new file mode 100644 index 0000000000000000000000000000000000000000..7a243555b6385af690e9632dfa81bf96d70f925d --- /dev/null +++ b/paddle/operators/uniform_random_op.cu @@ -0,0 +1,71 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include +#include +#include +#include +#include "paddle/framework/op_registry.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace operators { + +template +struct UniformGenerator { + T min_, max_; + unsigned int seed_; + + __host__ __device__ UniformGenerator(T min, T max, int seed) + : min_(min), max_(max), seed_(seed) {} + + __host__ __device__ T operator()(const unsigned int n) const { + thrust::minstd_rand rng; + rng.seed(seed_); + thrust::uniform_real_distribution dist(min_, max_); + rng.discard(n); + return dist(rng); + } +}; + +// It seems that Eigen::Tensor::random in GPU will SEGFAULT. +// Use std::random and thrust::random(thrust is a std library in CUDA) to +// implement uniform random. +template +class GPUUniformRandomKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* tensor = context.Output("Out"); + T* data = tensor->mutable_data(context.GetPlace()); + unsigned int seed = + static_cast(context.op_.GetAttr("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")); + 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), + UniformGenerator(min, max, seed)); + } +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OP_GPU_KERNEL(uniform_random, + paddle::operators::GPUUniformRandomKernel); diff --git a/paddle/parameter/Argument.cpp b/paddle/parameter/Argument.cpp index ef72b973c1a465a8ac03cae1070429160eac0ac1..0547ac93cd183afbcede41d280c6b4b16ed7dab1 100644 --- a/paddle/parameter/Argument.cpp +++ b/paddle/parameter/Argument.cpp @@ -666,4 +666,24 @@ void Argument::subArgFrom(const Argument& input, } } +void Argument::reorganizeSeqInfo( + const ICpuGpuVectorPtr seqStartPos, + const ICpuGpuVectorPtr subSeqStartPos, + std::vector>& reorganizedSeqInfo) { + int* seqStarts = seqStartPos->getMutableData(false); + int* subSeqStarts = subSeqStartPos->getMutableData(false); + + 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; + reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]); + } + } +} + } // namespace paddle diff --git a/paddle/parameter/Argument.h b/paddle/parameter/Argument.h index 0ccdef802e71b659788cfd24f28ebe43e1917db1..d8d7a4398f99a2794c5d25528a7d582f5ed629ba 100644 --- a/paddle/parameter/Argument.h +++ b/paddle/parameter/Argument.h @@ -317,6 +317,30 @@ struct Argument { */ void printValueString(std::ostream& stream, const std::string& prefix = "") const; + + /** + * @brief reorganizeSeqInfo will reorganize sequenceStartPositions and + * subSequenceStartPositions into a 2 dimensional arrary: reorganizedSeqInfo. + * + * @param seqStartPos: sequenceStartPositions of an Argument. + * @param subSeqStartPos: subSequenceStartPositions of an Argument. + * @param the reorganized sequence start position information. + * + * Examples: + * seqStartPos: [0, 4, 15, 20, 28] + * subSeqStartPos: [0, 3, 4, 5, 7, 10, 15, 20, 22, 23, 25, 28] + * reorganizedSeqInfo: + * [ + * [0,3,4], + * [4,5,7,10,15], + * [15,20], + * [20,22,23,25,28] + * ] + */ + static void reorganizeSeqInfo( + const ICpuGpuVectorPtr seqStartPos, + const ICpuGpuVectorPtr subSeqStartPos, + std::vector>& reorganizedSeqInfo); }; } // namespace paddle diff --git a/paddle/platform/CMakeLists.txt b/paddle/platform/CMakeLists.txt index bd77bb7daa50e0b273f110624ddf6f4b79a3ceab..4154aad15c39119e2f155cb2c7b5177b5aa78022 100644 --- a/paddle/platform/CMakeLists.txt +++ b/paddle/platform/CMakeLists.txt @@ -8,7 +8,7 @@ cc_test(place_test SRCS place_test.cc DEPS place glog gflags) add_subdirectory(dynload) -cc_test(enforce_test SRCS enforce_test.cc) +cc_test(enforce_test SRCS enforce_test.cc DEPS stringpiece) IF(WITH_GPU) set(GPU_CTX_DEPS dynload_cuda dynamic_loader) diff --git a/paddle/platform/device_context.h b/paddle/platform/device_context.h index 2038fafe2e15ec2631726643695ac6cbc317fed9..08b5b2cff900cc4239a615fe7d7f6b5faa13510b 100644 --- a/paddle/platform/device_context.h +++ b/paddle/platform/device_context.h @@ -40,7 +40,7 @@ class DeviceContext { class CPUDeviceContext : public DeviceContext { public: CPUDeviceContext(); - CPUDeviceContext(CPUPlace); + explicit CPUDeviceContext(CPUPlace); virtual ~CPUDeviceContext() {} Eigen::DefaultDevice* eigen_device() const; @@ -69,10 +69,10 @@ class CUDADeviceContext : public DeviceContext { // clang-format off /*! \brief Return cublas handle in the device context. */ - cublasHandle_t cublas_handle (); + cublasHandle_t cublas_handle(); /*! \brief Return cudnn handle in the device context. */ - cudnnHandle_t cudnn_handle (); + cudnnHandle_t cudnn_handle(); /*! \brief Return curand handle in the device context. */ curandGenerator_t curand_generator(); diff --git a/paddle/platform/device_context_test.cc b/paddle/platform/device_context_test.cc index af2ce17fc2238dda62e9888ebe9426edcd55d2bc..65345c433c0a328e7f89038a39312edba35eb8c7 100644 --- a/paddle/platform/device_context_test.cc +++ b/paddle/platform/device_context_test.cc @@ -15,24 +15,28 @@ limitations under the License. */ #include "paddle/platform/device_context.h" #include "gtest/gtest.h" -using DEVICE_GPU = Eigen::GpuDevice; TEST(Device, Init) { + using paddle::platform::DeviceContext; + using paddle::platform::CUDADeviceContext; + using paddle::platform::GPUPlace; + int count = paddle::platform::GetDeviceCount(); for (int i = 0; i < count; i++) { - paddle::platform::DeviceContext* device_context = - new paddle::platform::CUDADeviceContext(i); + DeviceContext* device_context = new CUDADeviceContext(GPUPlace(i)); Eigen::GpuDevice* gpu_device = - device_context->template get_eigen_device(); + device_context->template get_eigen_device(); ASSERT_NE(nullptr, gpu_device); delete device_context; } } TEST(Device, CUDADeviceContext) { + using paddle::platform::CUDADeviceContext; + using paddle::platform::GPUPlace; + int count = paddle::platform::GetDeviceCount(); for (int i = 0; i < count; i++) { - paddle::platform::CUDADeviceContext* device_context = - new paddle::platform::CUDADeviceContext(i); + CUDADeviceContext* device_context = new CUDADeviceContext(GPUPlace(i)); Eigen::GpuDevice* gpu_device = device_context->eigen_device(); ASSERT_NE(nullptr, gpu_device); cudnnHandle_t cudnn_handle = device_context->cudnn_handle(); diff --git a/paddle/platform/dynload/cublas.cc b/paddle/platform/dynload/cublas.cc index 4e3dfdaefb2348346e8f917b1f6c758bf6d91a1a..9cd2a1f565526f8dc45932ba6168f4e25c6ad238 100644 --- a/paddle/platform/dynload/cublas.cc +++ b/paddle/platform/dynload/cublas.cc @@ -1,3 +1,17 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + #include namespace paddle { diff --git a/paddle/platform/dynload/cublas.h b/paddle/platform/dynload/cublas.h index c44b7240a885c2ef71e550df645dbaded69f9944..9d8343c0b5e200b390ccda760f09816959952e9d 100644 --- a/paddle/platform/dynload/cublas.h +++ b/paddle/platform/dynload/cublas.h @@ -48,13 +48,13 @@ extern void *cublas_dso_handle; }; \ extern DynLoad__##__name __name #else -#define DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP(__name) \ - struct DynLoad__##__name { \ - inline template \ - cublasStatus_t operator()(Args... args) { \ - return __name(args...); \ - } \ - }; \ +#define DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP(__name) \ + struct DynLoad__##__name { \ + template \ + inline cublasStatus_t operator()(Args... args) { \ + return __name(args...); \ + } \ + }; \ extern DynLoad__##__name __name #endif @@ -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.cc b/paddle/platform/dynload/cudnn.cc index 8b5e15b5efcdae6a1eed09f002eb2f4f2163035f..d3e4cb567d71b987724366b6a0896f5df0eb6055 100644 --- a/paddle/platform/dynload/cudnn.cc +++ b/paddle/platform/dynload/cudnn.cc @@ -1,3 +1,17 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + #include namespace paddle { @@ -25,4 +39,4 @@ CUDNN_DNN_ROUTINE_EACH_R5(DEFINE_WRAP); } // namespace dynload } // namespace platform -} // namespace paddle \ No newline at end of file +} // namespace paddle diff --git a/paddle/platform/dynload/curand.cc b/paddle/platform/dynload/curand.cc index 5c1fab992c98569d4a95b6e699d97d428511e48e..d05dd88126bfee7278e553710a717b8f2eb02ae0 100644 --- a/paddle/platform/dynload/curand.cc +++ b/paddle/platform/dynload/curand.cc @@ -1,3 +1,17 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + #include namespace paddle { @@ -10,6 +24,7 @@ void *curand_dso_handle; #define DEFINE_WRAP(__name) DynLoad__##__name __name CURAND_RAND_ROUTINE_EACH(DEFINE_WRAP); -} -} -} \ No newline at end of file + +} // namespace dynload +} // namespace platform +} // namespace paddle diff --git a/paddle/platform/dynload/curand.h b/paddle/platform/dynload/curand.h index d8c46bc41e18d013a80cd0a9116a4b1a52bf5854..7bfe0778c78f6075ec8a284d478a1f9d5ee66ae9 100644 --- a/paddle/platform/dynload/curand.h +++ b/paddle/platform/dynload/curand.h @@ -55,6 +55,7 @@ extern void *curand_dso_handle; __macro(curandSetPseudoRandomGeneratorSeed); \ __macro(curandGenerateUniform); \ __macro(curandGenerateUniformDouble); \ + __macro(curandGenerateNormal); \ __macro(curandDestroyGenerator); CURAND_RAND_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CURAND_WRAP); diff --git a/paddle/platform/enforce.h b/paddle/platform/enforce.h index 60a42c777d1c2ebbc22fdb77b1100cc6fcf7ff35..15fdf7a94f462a87f7edae1429eb0c4da0b17a84 100644 --- a/paddle/platform/enforce.h +++ b/paddle/platform/enforce.h @@ -14,13 +14,21 @@ limitations under the License. */ #pragma once -#include -#include +#include // for dladdr +#include // for backtrace #include +#include #include #include #include +#include "paddle/string/printf.h" +#include "paddle/string/to_string.h" + +#ifdef __GNUC__ +#include // for __cxa_demangle +#endif + #ifndef PADDLE_ONLY_CPU #include "paddle/platform/dynload/cublas.h" @@ -38,6 +46,19 @@ limitations under the License. */ namespace paddle { namespace platform { +namespace { +#ifdef __GNUC__ +inline std::string demangle(std::string name) { + int status = -4; // some arbitrary value to eliminate the compiler warning + std::unique_ptr res{ + abi::__cxa_demangle(name.c_str(), NULL, NULL, &status), std::free}; + return (status == 0) ? res.get() : name; +} +#else +inline std::string demangle(std::string name) { return name; } +#endif +} + struct EnforceNotMet : public std::exception { std::exception_ptr exp_; std::string err_str_; @@ -47,15 +68,29 @@ 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 = demangle(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 %s\n", i, 2 + sizeof(void*) * 2, + call_stack[i]); + } } - free(line); + free(symbols); err_str_ = sout.str(); } } @@ -162,5 +197,41 @@ inline void throw_on_error(T e) { } \ } while (0) +/* + * Some enforce helpers here, usage: + * int a = 1; + * int b = 2; + * PADDLE_ENFORCE_EQ(a, b); + * + * will raise an expression described as follows: + * "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) + */ + +#define PADDLE_ENFORCE_EQ(__VAL0, __VAL1, ...) \ + __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, ==, !=, __VA_ARGS__) +#define PADDLE_ENFORCE_NE(__VAL0, __VAL1, ...) \ + __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, !=, ==, __VA_ARGS__) +#define PADDLE_ENFORCE_GT(__VAL0, __VAL1, ...) \ + __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, >, <=, __VA_ARGS__) +#define PADDLE_ENFORCE_GE(__VAL0, __VAL1, ...) \ + __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, >=, <, __VA_ARGS__) +#define PADDLE_ENFORCE_LT(__VAL0, __VAL1, ...) \ + __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, <, >=, __VA_ARGS__) +#define PADDLE_ENFORCE_LE(__VAL0, __VAL1, ...) \ + __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, <=, >, __VA_ARGS__) +#define PADDLE_ENFORCE_NOT_NULL(__VAL, ...) \ + PADDLE_ENFORCE(nullptr != (__VAL), #__VAL " should not be null\n%s", \ + paddle::string::Sprintf("" __VA_ARGS__)); + +#define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...) \ + PADDLE_ENFORCE(__VAL0 __CMP __VAL1, \ + "enforce %s " #__CMP " %s failed, %s " #__INV_CMP " %s\n%s", \ + #__VAL0, #__VAL1, paddle::string::to_string(__VAL0), \ + paddle::string::to_string(__VAL1), \ + paddle::string::Sprintf("" __VA_ARGS__)); + } // namespace platform } // namespace paddle diff --git a/paddle/platform/enforce_test.cc b/paddle/platform/enforce_test.cc index 2ac31812a80d8dd57ce82234cb5835e029a46067..80bdee3d9dfbe38ef707a6ba60cdb7f7b99714de 100644 --- a/paddle/platform/enforce_test.cc +++ b/paddle/platform/enforce_test.cc @@ -9,8 +9,16 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/platform/enforce.h" +#include +#include +#include + #include "gtest/gtest.h" +#include "paddle/platform/enforce.h" +#include "paddle/string/piece.h" + +using StringPiece = paddle::string::Piece; +using paddle::string::HasPrefix; TEST(ENFORCE, OK) { PADDLE_ENFORCE(true, "Enforce is ok %d now %f", 123, 0.345); @@ -20,17 +28,189 @@ TEST(ENFORCE, OK) { } TEST(ENFORCE, FAILED) { - bool in_catch = false; + bool caught_exception = false; try { PADDLE_ENFORCE(false, "Enforce is not ok %d at all", 123); } catch (paddle::platform::EnforceNotMet error) { - // your error handling code here - in_catch = true; - std::string msg = "Enforce is not ok 123 at all"; - const char* what = error.what(); - for (size_t i = 0; i < msg.length(); ++i) { - ASSERT_EQ(what[i], msg[i]); + caught_exception = true; + EXPECT_TRUE( + HasPrefix(StringPiece(error.what()), "Enforce is not ok 123 at all")); + } + EXPECT_TRUE(caught_exception); +} + +TEST(ENFORCE, NO_ARG_OK) { + int a = 2; + int b = 2; + PADDLE_ENFORCE_EQ(a, b); + // test enforce with extra message. + PADDLE_ENFORCE_EQ(a, b, "some thing wrong %s", "info"); +} + +TEST(ENFORCE_EQ, NO_EXTRA_MSG_FAIL) { + int a = 2; + bool caught_exception = false; + try { + PADDLE_ENFORCE_EQ(a, 1 + 3); + } catch (paddle::platform::EnforceNotMet error) { + caught_exception = true; + HasPrefix(StringPiece(error.what()), "enforce a == 1 + 3 failed, 2 != 4"); + } + EXPECT_TRUE(caught_exception); +} + +TEST(ENFORCE_EQ, EXTRA_MSG_FAIL) { + int a = 2; + bool caught_exception = false; + try { + PADDLE_ENFORCE_EQ(a, 1 + 3, "%s size not match", "their"); + } catch (paddle::platform::EnforceNotMet error) { + caught_exception = true; + HasPrefix(StringPiece(error.what()), + "enforce a == 1 + 3 failed, 2 != 4\ntheir size not match"); + } + EXPECT_TRUE(caught_exception); +} + +TEST(ENFORCE_NE, OK) { + PADDLE_ENFORCE_NE(1, 2); + PADDLE_ENFORCE_NE(1.0, 2UL); +} +TEST(ENFORCE_NE, FAIL) { + bool caught_exception = false; + + try { + // 2UL here to check data type compatible + PADDLE_ENFORCE_NE(1.0, 1UL); + } catch (paddle::platform::EnforceNotMet error) { + caught_exception = true; + EXPECT_TRUE(HasPrefix(StringPiece(error.what()), + "enforce 1.0 != 1UL failed, 1 == 1")) + << error.what() << " does not have expected prefix"; + } + EXPECT_TRUE(caught_exception); +} + +TEST(ENFORCE_GT, OK) { PADDLE_ENFORCE_GT(2, 1); } +TEST(ENFORCE_GT, FAIL) { + bool caught_exception = false; + try { + PADDLE_ENFORCE_GT(1, 2UL); + + } catch (paddle::platform::EnforceNotMet error) { + caught_exception = true; + EXPECT_TRUE( + HasPrefix(StringPiece(error.what()), "enforce 1 > 2UL failed, 1 <= 2")); + } + EXPECT_TRUE(caught_exception); +} + +TEST(ENFORCE_GE, OK) { + PADDLE_ENFORCE_GE(2, 2UL); + PADDLE_ENFORCE_GE(3, 2UL); + PADDLE_ENFORCE_GE(3, 2); + PADDLE_ENFORCE_GE(3.21, 2UL); +} +TEST(ENFORCE_GE, FAIL) { + bool caught_exception = false; + try { + PADDLE_ENFORCE_GE(1, 2UL); + + } catch (paddle::platform::EnforceNotMet error) { + caught_exception = true; + EXPECT_TRUE( + HasPrefix(StringPiece(error.what()), "enforce 1 >= 2UL failed, 1 < 2")); + } + EXPECT_TRUE(caught_exception); +} + +TEST(ENFORCE_LE, OK) { + PADDLE_ENFORCE_LE(1, 1); + PADDLE_ENFORCE_LE(1, 1UL); + PADDLE_ENFORCE_LE(2, 3UL); + PADDLE_ENFORCE_LE(2UL, 3); + PADDLE_ENFORCE_LE(2UL, 3.2); +} +TEST(ENFORCE_LE, FAIL) { + bool caught_exception = false; + try { + PADDLE_ENFORCE_GT(1, 2UL); + + } catch (paddle::platform::EnforceNotMet error) { + caught_exception = true; + EXPECT_TRUE( + HasPrefix(StringPiece(error.what()), "enforce 1 > 2UL failed, 1 <= 2")); + } + EXPECT_TRUE(caught_exception); +} + +TEST(ENFORCE_LT, OK) { + PADDLE_ENFORCE_LT(3, 10); + PADDLE_ENFORCE_LT(2, 3UL); + PADDLE_ENFORCE_LT(2UL, 3); +} +TEST(ENFORCE_LT, FAIL) { + bool caught_exception = false; + try { + PADDLE_ENFORCE_LT(1UL, 0.12); + } catch (paddle::platform::EnforceNotMet error) { + caught_exception = true; + EXPECT_TRUE(HasPrefix(StringPiece(error.what()), + "enforce 1UL < 0.12 failed, 1 >= 0.12")); + } + EXPECT_TRUE(caught_exception); +} + +TEST(ENFORCE_NOT_NULL, OK) { + int* a = new int; + PADDLE_ENFORCE_NOT_NULL(a); + delete a; +} +TEST(ENFORCE_NOT_NULL, FAIL) { + bool caught_exception = false; + try { + int* a = nullptr; + PADDLE_ENFORCE_NOT_NULL(a); + + } catch (paddle::platform::EnforceNotMet error) { + caught_exception = true; + EXPECT_TRUE(HasPrefix(StringPiece(error.what()), "a should not be null")); + } + EXPECT_TRUE(caught_exception); +} + +struct Dims { + size_t dims_[4]; + + bool operator==(const Dims& o) const { + for (size_t i = 0; i < 4; ++i) { + if (dims_[i] != o.dims_[i]) return false; + } + return true; + } +}; + +std::ostream& operator<<(std::ostream& os, const Dims& d) { + for (size_t i = 0; i < 4; ++i) { + if (i == 0) { + os << "["; + } + os << d.dims_[i]; + if (i == 4 - 1) { + os << "]"; + } else { + os << ", "; } } - ASSERT_TRUE(in_catch); + return os; +} + +TEST(ENFORCE_USER_DEFINED_CLASS, EQ) { + Dims a{{1, 2, 3, 4}}, b{{1, 2, 3, 4}}; + PADDLE_ENFORCE_EQ(a, b); } + +TEST(ENFORCE_USER_DEFINED_CLASS, NE) { + Dims a{{1, 2, 3, 4}}, b{{5, 6, 7, 8}}; + ASSERT_THROW(PADDLE_ENFORCE_EQ(a, b), paddle::platform::EnforceNotMet); +} \ No newline at end of file diff --git a/paddle/platform/place.h b/paddle/platform/place.h index 7cead183884bc9379355cd931921b40d6c11ce90..1117476bb37f1b0f3876c55e610803d5ee2558ce 100644 --- a/paddle/platform/place.h +++ b/paddle/platform/place.h @@ -14,8 +14,8 @@ limitations under the License. */ #pragma once -#include #include +#include "paddle/platform/variant.h" namespace paddle { namespace platform { @@ -32,7 +32,7 @@ struct CPUPlace { struct GPUPlace { GPUPlace() : GPUPlace(0) {} - GPUPlace(int d) : device(d) {} + explicit GPUPlace(int d) : device(d) {} // needed for variant equality comparison inline bool operator==(const GPUPlace &o) const { return device == o.device; } diff --git a/paddle/platform/variant.h b/paddle/platform/variant.h new file mode 100644 index 0000000000000000000000000000000000000000..c2257af1b5dd1a1e284979bf17e1a947072baa85 --- /dev/null +++ b/paddle/platform/variant.h @@ -0,0 +1,32 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include + +#ifndef PADDLE_ONLY_CPU + +// Because boost's variadic templates has bug on nvcc, boost will disable +// variadic template support when GPU enabled on nvcc. +// Define BOOST_NO_CXX11_VARIADIC_TEMPLATES on gcc/clang to generate same +// function symbols. +// +// https://github.com/PaddlePaddle/Paddle/issues/3386 +#ifndef BOOST_NO_CXX11_VARIADIC_TEMPLATES +#define BOOST_NO_CXX11_VARIADIC_TEMPLATES +#endif +#endif + +#include diff --git a/paddle/pserver/test/CMakeLists.txt b/paddle/pserver/test/CMakeLists.txt index 6e8f9c37f64b70921e09241089a5a480fd8ca47f..b66a00ba0652dfe1afbb877eca06cacdfe2ca343 100644 --- a/paddle/pserver/test/CMakeLists.txt +++ b/paddle/pserver/test/CMakeLists.txt @@ -3,7 +3,7 @@ add_unittest_without_exec(socket_test SocketTest.cpp) add_test(NAME socket_test - COMMAND ${PROJ_ROOT}/paddle/.set_port.sh -p port + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port ${CMAKE_CURRENT_BINARY_DIR}/socket_test --loop_time=10) ####################### test_ProtoServer #################### @@ -12,7 +12,7 @@ add_unittest_without_exec(test_ProtoServer IF(NOT ON_TRAVIS) add_test(NAME test_ProtoServer - COMMAND ${PROJ_ROOT}/paddle/.set_port.sh -p port + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoServer) ENDIF(NOT ON_TRAVIS) @@ -24,5 +24,5 @@ ENDIF(NOT ON_TRAVIS) add_unittest_without_exec(test_ParameterServer2 test_ParameterServer2.cpp) add_test(NAME test_ParameterServer2 - COMMAND ${PROJ_ROOT}/paddle/.set_port.sh -p port -n 4 + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port -n 4 ${CMAKE_CURRENT_BINARY_DIR}/test_ParameterServer2) diff --git a/paddle/pybind/CMakeLists.txt b/paddle/pybind/CMakeLists.txt deleted file mode 100644 index 29dd0ded0ac75893da7e244d92725cd5e285efce..0000000000000000000000000000000000000000 --- a/paddle/pybind/CMakeLists.txt +++ /dev/null @@ -1,9 +0,0 @@ -cc_library(paddle_pybind SHARED - SRCS pybind.cc - DEPS pybind python backward - fc_op - sgd_op - add_op - mean_op - cross_entropy_op - recurrent_op) diff --git a/paddle/scripts/CMakeLists.txt b/paddle/scripts/CMakeLists.txt index 66a46e1883a49d491f0cb3056a7039407d72e337..a52f06fe497dac467e4ef2543ebda7a423ca326d 100644 --- a/paddle/scripts/CMakeLists.txt +++ b/paddle/scripts/CMakeLists.txt @@ -1,17 +1,15 @@ configure_file(submit_local.sh.in - submit_local.sh + paddle @ONLY) -install(FILES ${CMAKE_CURRENT_BINARY_DIR}/submit_local.sh DESTINATION bin +install(FILES ${CMAKE_CURRENT_BINARY_DIR}/paddle DESTINATION bin PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ - GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ - RENAME paddle) + GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ) configure_file(tools/usage_stat/usage.sh - usage.sh + paddle_usage @ONLY) -install(FILES ${CMAKE_CURRENT_BINARY_DIR}/usage.sh DESTINATION opt/paddle/bin +install(FILES ${CMAKE_CURRENT_BINARY_DIR}/paddle_usage DESTINATION opt/paddle/bin PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ - GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ - RENAME paddle_usage) + GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ) diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh index 8de0e608c1f482e4553c07ff7ffd572d65a772aa..6c2f5fed405ecc80d5388084c7774d0ffb45f9af 100644 --- a/paddle/scripts/docker/build.sh +++ b/paddle/scripts/docker/build.sh @@ -31,54 +31,58 @@ Configuring cmake in /paddle/build ... -DWITH_DOC=OFF -DWITH_GPU=${WITH_GPU:-OFF} -DWITH_AVX=${WITH_AVX:-OFF} - -DWITH_GOLANG=${WITH_GOLANG:-OFF} + -DWITH_GOLANG=${WITH_GOLANG:-ON} -DWITH_SWIG_PY=ON + -DWITH_C_API=${WITH_C_API:-OFF} + -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} -DCMAKE_EXPORT_COMPILE_COMMANDS=ON ======================================== EOF + +# Disable UNITTEST_USE_VIRTUALENV in docker because +# docker environment is fully controlled by this script. +# See /Paddle/CMakeLists.txt, UNITTEST_USE_VIRTUALENV option. cmake .. \ -DCMAKE_BUILD_TYPE=Release \ -DWITH_DOC=OFF \ -DWITH_GPU=${WITH_GPU:-OFF} \ -DWITH_AVX=${WITH_AVX:-OFF} \ - -DWITH_GOLANG=${WITH_GOLANG:-OFF} \ - -DWITH_SWIG_PY=ON \ + -DWITH_GOLANG=${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=${WITH_TESTING:-OFF} \ -DCMAKE_EXPORT_COMPILE_COMMANDS=ON cat < /dev/null -SCRIPTPATH=$PWD -popd > /dev/null - -USE_VIRTUALENV_FOR_TEST=$1; shift -PYTHON=$1; shift - -if [ $USE_VIRTUALENV_FOR_TEST -ne 0 ]; then - rm -rf .test_env - virtualenv .test_env - unset PYTHONHOME - unset PYTHONPATH - source .test_env/bin/activate - PYTHON=python -fi - -$PYTHON -m pip install $SCRIPTPATH/../dist/*.whl - -if [ "X${PADDLE_PACKAGE_DIR}" != "X" ]; then - $PYTHON -m pip install ${PADDLE_PACKAGE_DIR}/*.whl -else - export PYTHONPATH=$SCRIPTPATH/../../python/ -fi - -$PYTHON -m pip install ipython==5.3 - -for fn in "$@" -do - echo "test $fn" - $PYTHON $fn - if [ $? -ne 0 ]; then - exit 1 - fi -done - -if [ $USE_VIRTUALENV_FOR_TEST -ne 0 ]; then - deactivate - rm -rf .test_env -fi diff --git a/paddle/scripts/submit_local.sh.in b/paddle/scripts/submit_local.sh.in old mode 100644 new mode 100755 index 12bf629ea920832f96bc5f7cc0b38abfddd34d97..2ab7d5b52f68641999603d8b675d8fabf24fd574 --- 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@" diff --git a/paddle/scripts/travis/build_doc.sh b/paddle/scripts/travis/build_doc.sh index 33fb5d84e2701c163b5d1b1bb3362ee81ebb34ea..dfcff38302703066e868c60e213f0f7cbc55a31e 100755 --- a/paddle/scripts/travis/build_doc.sh +++ b/paddle/scripts/travis/build_doc.sh @@ -5,15 +5,9 @@ set -e mkdir -p $TRAVIS_BUILD_DIR/build cd $TRAVIS_BUILD_DIR/build -# Compile paddle binaries first -cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_DOC=OFF -DWITH_MKLDNN=OFF -DWITH_MKLML=OFF -DWITH_GOLANG=ON -DWITH_STYLE_CHECK=OFF - -mkdir output -make -j `nproc` -find .. -name '*whl' | xargs pip install # install all wheels. -rm -rf * # Compile Documentation only. cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKLDNN=OFF -DWITH_MKLML=OFF -DWITH_DOC=ON +make -j `nproc` gen_proto_py make -j `nproc` paddle_docs paddle_docs_cn # check websites for broken links @@ -35,6 +29,7 @@ TARGET_BRANCH="gh-pages" SOURCE_BRANCH="master" # Clone the repo to output directory +mkdir output git clone $REPO output cd output diff --git a/paddle/setup.py.in b/paddle/setup.py.in deleted file mode 100644 index 06d55d3abc6097fa7d4b2b2ac9e29681e0fddfd5..0000000000000000000000000000000000000000 --- a/paddle/setup.py.in +++ /dev/null @@ -1,30 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - - -from setuptools import setup, Extension - -setup(name="py_paddle", - version="${PADDLE_VERSION}", - packages=['py_paddle'], - include_package_data=True, - package_data={'py_paddle':['*.py','_swig_paddle.so']}, - install_requires = [ - 'nltk>=3.2.2', - 'numpy>=1.8.0', # The numpy is required. - 'protobuf==${PROTOBUF_VERSION}' # The paddle protobuf version - ], - url='http://www.paddlepaddle.org/', - license='Apache 2.0', -) diff --git a/paddle/string/CMakeLists.txt b/paddle/string/CMakeLists.txt index 5becf62672d0c606c98ea1a1a4383df97088ab05..60667b72873f9422aec1807972a81ab680de2e64 100644 --- a/paddle/string/CMakeLists.txt +++ b/paddle/string/CMakeLists.txt @@ -2,3 +2,4 @@ cc_library(stringpiece SRCS piece.cc) cc_test(stringpiece_test SRCS piece_test.cc DEPS stringpiece glog gflags) cc_test(stringprintf_test SRCS printf_test.cc DEPS glog gflags) +cc_test(to_string_test SRCS to_string_test.cc) diff --git a/paddle/string/piece.h b/paddle/string/piece.h index 0272529d1c9b2cb6000a26f1d4d80276d06bf27b..03ae9243a4cc4e9e92e376bf46ab2b1d7162dfcb 100644 --- a/paddle/string/piece.h +++ b/paddle/string/piece.h @@ -39,8 +39,8 @@ public: // size_ is 0. Piece(); Piece(const char* d, size_t n); - Piece(const char* d); - Piece(const std::string& s); + Piece(const char* d); // NOLINT: accept C string into Piece. + Piece(const std::string& s); // NOLINT: accept C++ string into Piece. const char* data() const { return data_; } size_t len() const { return size_; } diff --git a/paddle/string/to_string.h b/paddle/string/to_string.h new file mode 100644 index 0000000000000000000000000000000000000000..4f478b6a36b23bdba8ef3ddae94b3eadf18716c2 --- /dev/null +++ b/paddle/string/to_string.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 +#include + +namespace paddle { +namespace string { +template +inline std::string to_string(T v) { + std::ostringstream sout; + sout << v; + return sout.str(); +} + +// Faster std::string/const char* type +template <> +inline std::string to_string(std::string v) { + return v; +} + +template <> +inline std::string to_string(const char* v) { + return std::string(v); +} + +} // namespace string +} // namespace paddle diff --git a/paddle/string/to_string_test.cc b/paddle/string/to_string_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..5ff1b007f1875c7b920a08bd13b8d98cdc5138d3 --- /dev/null +++ b/paddle/string/to_string_test.cc @@ -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. */ + +#include "paddle/string/to_string.h" +#include + +constexpr char kOutputString[] = "User Defined Output"; +class UserDefinedClass { +public: +}; + +std::ostream& operator<<(std::ostream& s, const UserDefinedClass& ins) { + s << kOutputString; + return s; +} + +TEST(to_string, normal) { + using namespace paddle::string; + ASSERT_EQ("10", to_string(10)); + ASSERT_EQ("abc", to_string("abc")); + ASSERT_EQ("1.2", to_string(1.2)); +} + +TEST(to_string, user_defined) { + using namespace paddle::string; + UserDefinedClass instance; + ASSERT_EQ(kOutputString, to_string(instance)); +} \ No newline at end of file diff --git a/paddle/trainer/NewRemoteParameterUpdater.cpp b/paddle/trainer/NewRemoteParameterUpdater.cpp index e1558e3fdfbcf296be0ee64202132f53bf901be9..35dcb235e7e8b65f7d1623a1ec66d963b1283385 100644 --- a/paddle/trainer/NewRemoteParameterUpdater.cpp +++ b/paddle/trainer/NewRemoteParameterUpdater.cpp @@ -50,8 +50,8 @@ void NewRemoteParameterUpdater::init( // create parameter server client. if (useEtcd_) { - parameterClient_ = paddle_new_etcd_pserver_client( - (char *)pserverSpec_.c_str(), FLAGS_trainer_id == 0); + parameterClient_ = + paddle_new_etcd_pserver_client((char *)pserverSpec_.c_str()); } else { parameterClient_ = paddle_new_pserver_client((char *)pserverSpec_.c_str(), FLAGS_trainer_id == 0); @@ -66,28 +66,92 @@ void NewRemoteParameterUpdater::init( // from parameter server if (paddle_begin_init_params(parameterClient_)) { LOG(INFO) << "paddle_begin_init_params start"; + // NOTE: convert V1 OptimizatioinConfig proto to V2 OptimizerConfig. + // This makes golang pserver compatible with handy V1 demos. + // TODO(wuyi): Refine or remove these ugly converting lines + OptimizerConfig optimizerConfigV2; + if (trainerConfig_.learning_method() == "momentum") { + optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::SGD); + } else if (trainerConfig_.learning_method() == "adagrad") { + optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::Adagrad); + optimizerConfigV2.mutable_adagrad()->set_epsilon( + trainerConfig_.ada_epsilon()); + } else if (trainerConfig_.learning_method() == "adadelta") { + optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::Adagrad); + optimizerConfigV2.mutable_adadelta()->set_epsilon( + trainerConfig_.ada_epsilon()); + optimizerConfigV2.mutable_adadelta()->set_rho(trainerConfig_.ada_rou()); + } else if (trainerConfig_.learning_method() == "adam") { + optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::Adam); + optimizerConfigV2.mutable_adam()->set_beta_1(trainerConfig_.adam_beta1()); + optimizerConfigV2.mutable_adam()->set_beta_2(trainerConfig_.adam_beta2()); + optimizerConfigV2.mutable_adam()->set_epsilon( + trainerConfig_.adam_epsilon()); + } else { + LOG(ERROR) << "got unsupported v1 optimizer config: " + << trainerConfig_.learning_method(); + optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::SGD); + } + + if (trainerConfig_.learning_rate_schedule() == "constant") { + optimizerConfigV2.set_lr_policy(paddle::OptimizerConfig::Const); + optimizerConfigV2.mutable_const_lr()->set_learning_rate( + trainerConfig_.learning_rate()); + } else if (trainerConfig_.learning_rate_schedule() == "linear") { + optimizerConfigV2.set_lr_policy(paddle::OptimizerConfig::Linear); + optimizerConfigV2.mutable_linear_lr()->set_learning_rate( + trainerConfig_.learning_rate()); + optimizerConfigV2.mutable_linear_lr()->set_lr_decay_a( + trainerConfig_.learning_rate_decay_a()); + optimizerConfigV2.mutable_linear_lr()->set_lr_decay_b( + trainerConfig_.learning_rate_decay_b()); + } else { + LOG(ERROR) << "got unsupported v1 learning_rate_schedule config: " + << trainerConfig_.learning_rate_schedule() << ", set to const"; + optimizerConfigV2.set_lr_policy(paddle::OptimizerConfig::Const); + } + + // overwrite optimizerConfigV2 for per-parameter(layer) configs for (int i = 0; i < parameterSize(); ++i) { auto paramConfig = parameters_[i]->getConfig(); - LOG(INFO) << "old param config: " << paramConfig.DebugString(); - // FIXME(typhoonzero): convert old paramConfig to optimizerConfig - OptimizerConfig optimizeConfigV2; - auto sgdConfigV2 = optimizeConfigV2.mutable_sgd(); - sgdConfigV2->set_momentum(paramConfig.momentum()); - sgdConfigV2->set_decay(paramConfig.decay_rate()); - optimizeConfigV2.set_lr_policy(paddle::OptimizerConfig::Const); - auto constlr = optimizeConfigV2.mutable_const_lr(); + if (paramConfig.has_momentum() && + trainerConfig_.learning_method() == "momentum") { + optimizerConfigV2.mutable_sgd()->set_momentum(paramConfig.momentum()); + } if (paramConfig.has_learning_rate()) { - constlr->set_learning_rate(paramConfig.learning_rate()); - } else { - constlr->set_learning_rate(trainerConfig_.learning_rate()); + switch (optimizerConfigV2.lr_policy()) { + case 0: + optimizerConfigV2.mutable_const_lr()->set_learning_rate( + paramConfig.learning_rate()); + break; + case 1: + optimizerConfigV2.mutable_linear_lr()->set_learning_rate( + paramConfig.learning_rate()); + break; + } } - if (trainerConfig_.algorithm() == "sgd") { - optimizeConfigV2.set_optimizer(paddle::OptimizerConfig::SGD); - // FIXME: config all algorithms - } else { - optimizeConfigV2.set_optimizer(paddle::OptimizerConfig::SGD); + if (paramConfig.has_decay_rate()) { + switch (optimizerConfigV2.optimizer()) { + case 1: // SGD + optimizerConfigV2.mutable_sgd()->set_decay( + paramConfig.decay_rate()); + break; + case 2: // Adadelta + optimizerConfigV2.mutable_adadelta()->set_decay( + paramConfig.decay_rate()); + break; + case 3: // Adagrad + optimizerConfigV2.mutable_adagrad()->set_decay( + paramConfig.decay_rate()); + break; + case 4: // Adam + optimizerConfigV2.mutable_adam()->set_decay( + paramConfig.decay_rate()); + break; + } } - std::string bytes = optimizeConfigV2.SerializeAsString(); + // send param and config to pserver + std::string bytes = optimizerConfigV2.SerializeAsString(); const char *array = bytes.data(); int size = (int)bytes.size(); paddle_init_param( diff --git a/paddle/trainer/TrainerConfigHelper.cpp b/paddle/trainer/TrainerConfigHelper.cpp index 133e2be104c6fbfddefd8698d2b6aa8315c56c70..eba40862b926cfe863c569e73a6a3ceabcf1f3b4 100644 --- a/paddle/trainer/TrainerConfigHelper.cpp +++ b/paddle/trainer/TrainerConfigHelper.cpp @@ -28,6 +28,8 @@ DECLARE_bool(with_cost); 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"; @@ -44,6 +46,8 @@ TrainerConfigHelper::TrainerConfigHelper(const std::string &configFilePath) configArgs << "trainer_id=" << FLAGS_trainer_id << ",local=" << FLAGS_local << ",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/trainer/tests/CMakeLists.txt b/paddle/trainer/tests/CMakeLists.txt index 08b2d8a38e2d20a357752269bd3ee3f515116abd..f01ad4142d4fe7c7f7d7aac60d967ea114b93e56 100644 --- a/paddle/trainer/tests/CMakeLists.txt +++ b/paddle/trainer/tests/CMakeLists.txt @@ -2,19 +2,19 @@ add_unittest_without_exec(test_Compare test_Compare.cpp) add_test(NAME test_Compare - COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_Compare - WORKING_DIRECTORY ${PROJ_ROOT}/paddle/) + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) ################# test_Trainer ########################### add_unittest_without_exec(test_Trainer test_Trainer.cpp) add_test(NAME test_Trainer - COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/ - ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/paddle/trainer/tests/gen_proto_data.py && - ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/ + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ + ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/paddle/trainer/tests/gen_proto_data.py && + ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ ${CMAKE_CURRENT_BINARY_DIR}/test_Trainer - WORKING_DIRECTORY ${PROJ_ROOT}/paddle/) + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) ############### test_TrainerOnePass ########################## if(WITH_PYTHON) @@ -23,60 +23,60 @@ if(WITH_PYTHON) add_unittest_without_exec(test_TrainerOnePass test_TrainerOnePass.cpp) add_test(NAME test_TrainerOnePass - COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d - ${PROJ_ROOT}/python/:${PROJ_ROOT}/paddle/trainer/tests - ${PROJ_ROOT}/paddle/.set_port.sh -p port ${CMAKE_CURRENT_BINARY_DIR}/test_TrainerOnePass - WORKING_DIRECTORY ${PROJ_ROOT}/paddle/) + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d + ${PADDLE_SOURCE_DIR}/python/:${PADDLE_SOURCE_DIR}/paddle/trainer/tests + ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port ${CMAKE_CURRENT_BINARY_DIR}/test_TrainerOnePass + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) endif() ################ test_CompareTwoNets ###################### add_unittest_without_exec(test_CompareTwoNets test_CompareTwoNets.cpp) add_test(NAME test_CompareTwoNets - COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/ + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ ${CMAKE_CURRENT_BINARY_DIR}/test_CompareTwoNets --config_file_a=trainer/tests/sample_trainer_config_qb_rnn.conf --config_file_b=trainer/tests/sample_trainer_config_rnn.conf - WORKING_DIRECTORY ${PROJ_ROOT}/paddle/) + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) ############### test_CompareTwoOpts ################### add_unittest_without_exec(test_CompareTwoOpts test_CompareTwoOpts.cpp) add_test(NAME test_CompareTwoOpts - COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/ + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ ${CMAKE_CURRENT_BINARY_DIR}/test_CompareTwoOpts --config_file_a=trainer/tests/sample_trainer_config_opt_a.conf --config_file_b=trainer/tests/sample_trainer_config_opt_b.conf --num_passes=1 --need_high_accuracy=0 - WORKING_DIRECTORY ${PROJ_ROOT}/paddle/) + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) ################# test_CompareSparse ################## add_unittest_without_exec(test_CompareSparse test_CompareSparse.cpp) if(NOT ON_TRAVIS) add_test(NAME test_CompareSparse - COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/ + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ ./.set_port.sh -p port -n 6 ${CMAKE_CURRENT_BINARY_DIR}/test_CompareSparse - WORKING_DIRECTORY ${PROJ_ROOT}/paddle/) + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) endif() ################# test_recurrent_machine_generation ############### add_unittest_without_exec(test_recurrent_machine_generation test_recurrent_machine_generation.cpp) add_test(NAME test_recurrent_machine_generation - COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/ + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ ${CMAKE_CURRENT_BINARY_DIR}/test_recurrent_machine_generation - WORKING_DIRECTORY ${PROJ_ROOT}/paddle/) + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) #################### test_PyDataProviderWrapper ######################### add_unittest_without_exec(test_PyDataProviderWrapper test_PyDataProviderWrapper.cpp) add_test(NAME test_PyDataProviderWrapper - COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d - ${PROJ_ROOT}/python/:${PROJ_ROOT}/paddle/trainer/tests + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d + ${PADDLE_SOURCE_DIR}/python/:${PADDLE_SOURCE_DIR}/paddle/trainer/tests ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProviderWrapper - WORKING_DIRECTORY ${PROJ_ROOT}/paddle/) + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) #################### test_config_parser ######################### add_test(NAME test_config_parser - COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/ - ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/paddle/trainer/tests/config_parser_test.py - WORKING_DIRECTORY ${PROJ_ROOT}/paddle/) + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ + ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/paddle/trainer/tests/config_parser_test.py + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/) diff --git a/paddle/trainer/tests/compare_sparse_data b/paddle/trainer/tests/compare_sparse_data new file mode 100644 index 0000000000000000000000000000000000000000..18fc6541383d8e8e1687b8fe1abd57aece3d4cfc Binary files /dev/null and b/paddle/trainer/tests/compare_sparse_data differ diff --git a/paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.proto b/paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.proto_data similarity index 100% rename from paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.proto rename to paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.proto_data diff --git a/paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.protolist b/paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.protolist index 8b041cd66416862a78dba27368a65860a68ef1a5..6b406dff0ba91b5f310d7eafa111c0d21d6542c3 100644 --- a/paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.protolist +++ b/paddle/trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.protolist @@ -1 +1 @@ -./trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.proto +./trainer/tests/pydata_provider_wrapper_dir/test_pydata_provider_wrapper.proto_data diff --git a/paddle/trainer/tests/sample_trainer_config_compare_sparse.conf b/paddle/trainer/tests/sample_trainer_config_compare_sparse.conf new file mode 100644 index 0000000000000000000000000000000000000000..92f32a18c0068ab4672034a270aa8c52f2716d59 --- /dev/null +++ b/paddle/trainer/tests/sample_trainer_config_compare_sparse.conf @@ -0,0 +1,154 @@ +#edit-mode: -*- python -*- +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +#Todo(luotao02) This config is only used for unitest. It is out of date now, and will be updated later. + +# Note: when making change to this file, please make sure +# sample_trainer_config_rnn.conf is changed accordingly so that the uniitest +# for comparing these two nets can pass (test_CompareTwoNets) + +default_initial_std(0.1) +default_device(0) + +word_dim = 999 +l1 = 0 +l2 = 0 + +model_type("nn") + +sparse_update = get_config_arg("sparse_update", bool, False) + +TrainData(ProtoData( + type = "proto_sequence", + files = ('trainer/tests/train_sparse.list'), + )) + +Settings( + algorithm='sgd', + batch_size=100, + learning_rate=0.0001, + learning_rate_decay_a=4e-08, + learning_rate_decay_b=0.0, + learning_rate_schedule='poly', +) + + +wordvec_dim = 32 +layer2_dim = 16 +layer3_dim = 16 +hidden_dim = 32 + +slot_names = ["qb", "qw", "tb", "tw"] + +def ltr_network(network_name, + word_dim=word_dim, + wordvec_dim=wordvec_dim, + layer2_dim=layer2_dim, + layer3_dim=layer3_dim, + hidden_dim=hidden_dim, + slot_names=slot_names, + l1=l1, + l2=l2): + + slotnum = len(slot_names) + for i in xrange(slotnum): + Inputs(slot_names[i] + network_name) + for i in xrange(slotnum): + Layer( + name = slot_names[i] + network_name, + type = "data", + size = word_dim, + device = -1, + ) + Layer( + name = slot_names[i] + "_embedding_" + network_name, + type = "mixed", + size = wordvec_dim, + bias = False, + device = -1, + inputs = TableProjection(slot_names[i] + network_name, + parameter_name = "embedding.w0", + decay_rate_l1=l1, + sparse_remote_update = True, + sparse_update = sparse_update, + ), + ) + Layer( + name = slot_names[i] + "_rnn1_" + network_name, + type = "recurrent", + active_type = "tanh", + bias = Bias(initial_std = 0, + parameter_name = "rnn1.bias"), + inputs = Input(slot_names[i] + "_embedding_" + network_name, + parameter_name = "rnn1.w0") + ) + Layer( + name = slot_names[i] + "_rnnlast_" + network_name, + type = "seqlastins", + inputs = [ + slot_names[i] + "_rnn1_" + network_name, + ], + ) + + Layer( + name = "layer2_" + network_name, + type = "fc", + active_type = "tanh", + size = layer2_dim, + bias = Bias(parameter_name = "layer2.bias"), + inputs = [Input(slot_name + "_rnnlast_" + network_name, + parameter_name = "_layer2_" + slot_name + ".w", + decay_rate = l2, + initial_smart = True) for slot_name in slot_names] + ) + Layer( + name = "layer3_" + network_name, + type = "fc", + active_type = "tanh", + size = layer3_dim, + bias = Bias(parameter_name = "layer3.bias"), + inputs = [ + Input("layer2_" + network_name, + parameter_name = "_layer3.w", + decay_rate = l2, + initial_smart = True), + ] + ) + Layer( + name = "output_" + network_name, + type = "fc", + size = 1, + bias = False, + inputs = [ + Input("layer3_" + network_name, + parameter_name = "_layerO.w"), + ], + ) + + +ltr_network("left") +ltr_network("right") +Inputs("label") +Layer( + name = "label", + type = "data", + size = 1, + ) +Outputs("cost", "qb_rnnlast_left") +Layer( + name = "cost", + type = "rank-cost", + inputs = ["output_left", "output_right", "label"], + ) diff --git a/paddle/trainer/tests/simple_sparse_neural_network.py b/paddle/trainer/tests/simple_sparse_neural_network.py index 9604e1b9b45e571130c2f1bdc6d6a5fbd9c177c4..30346ef299d0bc8585ccff7f2fc4885b0d9f9dfc 100644 --- a/paddle/trainer/tests/simple_sparse_neural_network.py +++ b/paddle/trainer/tests/simple_sparse_neural_network.py @@ -1,6 +1,6 @@ from paddle.trainer_config_helpers import * -settings(batch_size=128, learning_method=AdaGradOptimizer(), learning_rate=1e-4) +settings(batch_size=17, learning_method=AdaGradOptimizer(), learning_rate=1e-4) file_list = 'trainer/tests/fake_file_list.list' @@ -12,7 +12,7 @@ define_py_data_sources2( embedding = embedding_layer( input=data_layer( - name="word_ids", size=65536), + name="word_ids", size=8191), size=128, param_attr=ParamAttr(sparse_update=True)) prediction = fc_layer(input=embedding, size=10, act=SoftmaxActivation()) diff --git a/paddle/trainer/tests/simple_sparse_neural_network_dp.py b/paddle/trainer/tests/simple_sparse_neural_network_dp.py index 8bfd1f37e7114f2dcd0798ff1e8180b111ad988f..86b272edfe1bbb23c45cffe282f6475ceaa0cc41 100644 --- a/paddle/trainer/tests/simple_sparse_neural_network_dp.py +++ b/paddle/trainer/tests/simple_sparse_neural_network_dp.py @@ -7,15 +7,15 @@ def init_hook(settings, is_train, **kwargs): @provider( - input_types={'word_ids': integer_value(65536), + input_types={'word_ids': integer_value(8191), 'label': integer_value(10)}, min_pool_size=0, init_hook=init_hook) def process(settings, filename): if settings.is_train: - data_size = 2**20 - else: data_size = 2**10 + else: + data_size = 2**5 for _ in xrange(data_size): - yield random.randint(0, 65535), random.randint(0, 9) + yield random.randint(0, 8190), random.randint(0, 9) diff --git a/paddle/trainer/tests/test_CompareSparse.cpp b/paddle/trainer/tests/test_CompareSparse.cpp index a7000eb77e1bbeab4f6e38c0322f82bde7164080..813275518e411d6e963e23df634541f771096e0f 100644 --- a/paddle/trainer/tests/test_CompareSparse.cpp +++ b/paddle/trainer/tests/test_CompareSparse.cpp @@ -23,7 +23,7 @@ using namespace paddle; // NOLINT using namespace std; // NOLINT static const string& configFile1 = - "trainer/tests/sample_trainer_config_qb_rnn.conf"; + "trainer/tests/sample_trainer_config_compare_sparse.conf"; DECLARE_bool(use_gpu); DECLARE_string(config); diff --git a/paddle/trainer/tests/test_TrainerOnePass.cpp b/paddle/trainer/tests/test_TrainerOnePass.cpp index 4d0174f784a0dc7314977d586c3ad1f0f9c69f6d..00ba61377aeff17d82e03f7560c0d71b3570d14f 100644 --- a/paddle/trainer/tests/test_TrainerOnePass.cpp +++ b/paddle/trainer/tests/test_TrainerOnePass.cpp @@ -100,25 +100,25 @@ TEST(average_window, gpu) { } TEST(average_window, gpu2) { - FLAGS_num_passes = 100; + FLAGS_num_passes = 20; trainerOnePassTest(configFile1, true, false, 2, 0.01); FLAGS_num_passes = 1; } TEST(average_window, gpu4) { - FLAGS_num_passes = 100; + FLAGS_num_passes = 20; trainerOnePassTest(configFile1, true, false, 4, 0.01); FLAGS_num_passes = 1; } TEST(average_window_cpu, gpu2) { - FLAGS_num_passes = 100; + FLAGS_num_passes = 20; trainerOnePassTest(configFile1, true, false, 2, 0.01, true); FLAGS_num_passes = 1; } TEST(average_window_cpu, gpu4) { - FLAGS_num_passes = 100; + FLAGS_num_passes = 20; trainerOnePassTest(configFile1, true, false, 4, 0.01, true); FLAGS_num_passes = 1; } diff --git a/paddle/trainer/tests/train_sparse.list b/paddle/trainer/tests/train_sparse.list new file mode 100644 index 0000000000000000000000000000000000000000..6ea020e2202f8464f8a647cd96c84a9d17a03ae3 --- /dev/null +++ b/paddle/trainer/tests/train_sparse.list @@ -0,0 +1 @@ +trainer/tests/compare_sparse_data diff --git a/paddle/utils/Flags.cpp b/paddle/utils/Flags.cpp index 320f671ed97dbadc4fa1b4b52d5611cf9239e7dd..600c83a8487191895de635dd8433f6c44e86ce77 100644 --- a/paddle/utils/Flags.cpp +++ b/paddle/utils/Flags.cpp @@ -20,6 +20,14 @@ DEFINE_bool(use_gpu, false, "Only support CPU training"); DEFINE_bool(use_gpu, true, "Whether to use GPU for training"); #endif +#ifdef PADDLE_USE_MKLDNN +// TODO(TJ): change to true when MKLDNN layers support multi-inputs +DEFINE_bool(use_mkldnn, false, "Default still keep use CPU training"); +#else +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 dc4faef8331ed47b9ce3e952389b6469cd9fda2e..0aca4c0ee036ee8490c0ceca7279df876dc21947 100644 --- a/paddle/utils/Flags.h +++ b/paddle/utils/Flags.h @@ -40,3 +40,5 @@ DECLARE_bool(show_layer_stat); 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/tests/CMakeLists.txt b/paddle/utils/tests/CMakeLists.txt index aa923b355377752f9b297a125f5c43c364ba9b06..c770ce169878d9998e559b1d417fc1acc88cde97 100644 --- a/paddle/utils/tests/CMakeLists.txt +++ b/paddle/utils/tests/CMakeLists.txt @@ -13,6 +13,6 @@ add_executable( link_paddle_exe(test_CustomStackTracePrint) if(NOT APPLE) add_test(NAME test_CustomStackTracePrint - COMMAND ${PROJ_ROOT}/paddle/utils/tests/test_CustomStackTracePrint.sh + COMMAND ${PADDLE_SOURCE_DIR}/paddle/utils/tests/test_CustomStackTracePrint.sh WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) endif() diff --git a/proto/CMakeLists.txt b/proto/CMakeLists.txt index 18584cafe7971bad281b498908c54780250791b7..6212c2e60a8ed94ecc1d6e58535a2b3d365e3eb8 100644 --- a/proto/CMakeLists.txt +++ b/proto/CMakeLists.txt @@ -9,15 +9,15 @@ foreach(filename ${proto_filenames}) get_filename_component(ABS_FIL ${filename} ABSOLUTE) get_filename_component(FIL_WE ${filename} NAME_WE) set(CUR_PROTO_GEN_PY - ${PROJ_ROOT}/paddle/python/paddle/proto/${FIL_WE}_pb2.py) + ${PADDLE_SOURCE_DIR}/paddle/python/paddle/proto/${FIL_WE}_pb2.py) set(PROTO_GEN_PY ${CUR_PROTO_GEN_PY} ${PROTO_GEN_PY}) add_custom_command(OUTPUT ${CUR_PROTO_GEN_PY} COMMAND ${PROTOBUF_PROTOC_EXECUTABLE} - ARGS "--python_out=${PROJ_ROOT}/python/paddle/proto" + ARGS "--python_out=${PADDLE_SOURCE_DIR}/python/paddle/proto" "-I" ${CMAKE_CURRENT_SOURCE_DIR} ${ABS_FIL} - DEPENDS ${ABS_FIL} ${external_project_dependencies}) + DEPENDS ${ABS_FIL} protoc) endforeach() add_custom_target(gen_proto_py ALL DEPENDS ${PROTO_GEN_PY}) diff --git a/proto/DataConfig.proto b/proto/DataConfig.proto index e895c184d9f95dba1449e6467a2566712837600b..0cb5d7afbb3e1cb4abe45c0ed677e09b27b870fa 100644 --- a/proto/DataConfig.proto +++ b/proto/DataConfig.proto @@ -15,14 +15,13 @@ syntax = "proto2"; package paddle; - message FileGroupConf { - optional uint32 queue_capacity = 1 [default = 1]; + optional uint32 queue_capacity = 1 [ default = 1 ]; // how many files to load for a load file thread - optional int32 load_file_count = 2 [default = 1]; + optional int32 load_file_count = 2 [ default = 1 ]; // how many threads to load files // Setting to be 5~10 is appropriate when loading files by hadoop vfs - optional int32 load_thread_num = 3 [default = 1]; + optional int32 load_thread_num = 3 [ default = 1 ]; }; message DataConfig { @@ -32,26 +31,28 @@ message DataConfig { // name of a text file which contains a list of file names at each line optional string files = 3; - optional int32 feat_dim = 4;//feature dimension of one frame - repeated int32 slot_dims = 5;//feature slot dims - optional int32 context_len = 6;//max neibour frame numbers - optional uint64 buffer_capacity = 7;//the number of samples + optional int32 feat_dim = 4; // feature dimension of one frame + repeated int32 slot_dims = 5; // feature slot dims + optional int32 context_len = 6; // max neibour frame numbers + optional uint64 buffer_capacity = 7; // the number of samples - //part of data used in training - //if not -1, part of train data is used in training - optional int64 train_sample_num = 8 [default = -1]; + // part of data used in training + // if not -1, part of train data is used in training + optional int64 train_sample_num = 8 [ default = -1 ]; - //The number of documents processed once - optional int32 file_load_num = 9 [default = -1]; - optional bool async_load_data = 12 [default = false]; + // The number of documents processed once + optional int32 file_load_num = 9 [ default = -1 ]; + optional bool async_load_data = 12 [ default = false ]; /// Note the field number 10, 11 and 13 have been deprecated. - optional bool for_test = 14 [default = false]; // whether this data is for test + optional bool for_test = 14 + [ default = false ]; // whether this data is for test optional FileGroupConf file_group_conf = 15; repeated int32 float_slot_dims = 16; /// Note the field number 17, 18 and 19 have been deprecated. - // a list of values which will be used to create additional one dimensional float + // a list of values which will be used to create additional one dimensional + // float // values slots. These one dimensional slots can be used as the weight input // for cost layers. // Currently this is only supported by ProtoDataProvider. @@ -65,21 +66,21 @@ message DataConfig { // for MultiDataProvider repeated DataConfig sub_data_configs = 24; // sub dataproviders - /* - * the ratio of each sub dataproviders: - * e.g. sub dataprovider A's ratio is 1, B's ratio is 9, batch_size is 100, - * then each mini-batch is combined by 10 instance from A and 90 instances - * from B. - */ + /* + * the ratio of each sub dataproviders: + * e.g. sub dataprovider A's ratio is 1, B's ratio is 9, batch_size is 100, + * then each mini-batch is combined by 10 instance from A and 90 instances + * from B. + */ optional int32 data_ratio = 25; /* * if one of the sub dataproviders is running out of data, then * (1) it is "main data", then finish current pass. * (2) it is not "main data", then reset it, and try getNextBatch again. */ - optional bool is_main_data = 26 [default = true]; + optional bool is_main_data = 26 [ default = true ]; - // the usage ratio of instances. Setting to 1.0 means the use of all instances. - optional double usage_ratio = 27 [default = 1.0]; + // the usage ratio of instances. Setting to 1.0 means the use of all + // instances. + optional double usage_ratio = 27 [ default = 1.0 ]; }; - diff --git a/proto/DataFormat.proto b/proto/DataFormat.proto index 19b1499b0281a1b92028cc8944c27ee4d56b8dd2..7d963bc29f7c6b9895323b0d57ba4ee4cb4387d0 100644 --- a/proto/DataFormat.proto +++ b/proto/DataFormat.proto @@ -17,27 +17,32 @@ package paddle; /* If values is not empty and ids is empty, this is a dense vector. - If values is not empty and ids is not empty, this is a sparse vector. The position of each value + If values is not empty and ids is not empty, this is a sparse vector. The + position of each value is specified by ids. - If values is empty and ids is not empty, this is a sparse vector whose non-zero values are 1. + If values is empty and ids is not empty, this is a sparse vector whose non-zero + values are 1. The position of each 1 is specified by ids. */ message VectorSlot { - repeated float values = 1 [packed = true]; - repeated uint32 ids = 2 [packed = true]; + repeated float values = 1 [ packed = true ]; + repeated uint32 ids = 2 [ packed = true ]; /* For multidimensional data, for example "image width height depth" */ - repeated uint32 dims = 3 [packed = true]; - repeated string strs = 4; + repeated uint32 dims = 3 [ packed = true ]; + repeated string strs = 4; }; /* - SubseqSlot use to record whether VectorSlot or any other slot in future has subseq. - If not all VectorSlot have subseq, we only store the one who has subseq, and use *slot_id* to record it. - One vector_slots has one sequence, and it may have N subseq, thus the number of *lens* will be N too. + SubseqSlot use to record whether VectorSlot or any other slot in future has + subseq. + If not all VectorSlot have subseq, we only store the one who has subseq, and + use *slot_id* to record it. + One vector_slots has one sequence, and it may have N subseq, thus the number of + *lens* will be N too. */ message SubseqSlot { - required uint32 slot_id = 1; //the id of slot who has subseq - repeated uint32 lens = 2; // lengths of sub-sequence in the slot + required uint32 slot_id = 1; // the id of slot who has subseq + repeated uint32 lens = 2; // lengths of sub-sequence in the slot }; message SlotDef { @@ -45,13 +50,14 @@ message SlotDef { VECTOR_DENSE = 0; VECTOR_SPARSE_NON_VALUE = 1; VECTOR_SPARSE_VALUE = 2; - INDEX = 3; // This can be used as label, or word id, etc. + INDEX = 3; // This can be used as label, or word id, etc. VAR_MDIM_DENSE = 4; VAR_MDIM_INDEX = 5; STRING = 6; } required SlotType type = 1; - required uint32 dim = 2; // For INDEX slots, this means the maximal index plus 1. + required uint32 dim = + 2; // For INDEX slots, this means the maximal index plus 1. }; message DataHeader { @@ -60,11 +66,11 @@ message DataHeader { }; message DataSample { - optional bool is_beginning = 1 [default = true]; // is the beginning of a sequence + optional bool is_beginning = 1 + [ default = true ]; // is the beginning of a sequence repeated VectorSlot vector_slots = 2; - repeated uint32 id_slots = 3 [packed = true]; + repeated uint32 id_slots = 3 [ packed = true ]; /* use ids of VectorSlot */ repeated VectorSlot var_id_slots = 4; repeated SubseqSlot subseq_slots = 5; }; - diff --git a/proto/ModelConfig.proto b/proto/ModelConfig.proto index b50b73c7e169f3e8ae75322d9a0a3cad5072a9c7..4f3d5bf3f6cb96c97285f40e3a3d100c2af47ad5 100644 --- a/proto/ModelConfig.proto +++ b/proto/ModelConfig.proto @@ -21,7 +21,6 @@ package paddle; * Various structs for the configuration of a neural network */ - message ExternalConfig { repeated string layer_names = 1; repeated string input_layer_names = 2; @@ -68,7 +67,7 @@ message ConvConfig { required uint32 img_size = 8; // caffe mode for output size coherence - required bool caffe_mode = 9 [default = true]; + required bool caffe_mode = 9 [ default = true ]; // if filter_size_y is set , this convolutional layer will use // filters of size filter_size * filter_size_y pixels. @@ -99,7 +98,7 @@ message PoolConfig { optional uint32 start = 4; // Defines the stride size between successive pooling squares. - required uint32 stride = 5 [default = 1]; + required uint32 stride = 5 [ default = 1 ]; // The size of output feature map. required uint32 output_x = 6; @@ -109,7 +108,7 @@ message PoolConfig { // padding = 4, instructs the net to implicitly // pad the images with a 4-pixel border of zeros. - optional uint32 padding = 8 [default = 0]; + optional uint32 padding = 8 [ default = 0 ]; // if not set, use size_x optional uint32 size_y = 9; @@ -194,9 +193,7 @@ message MaxOutConfig { required uint32 groups = 2; } -message RowConvConfig { - required uint32 context_length = 1; -} +message RowConvConfig { required uint32 context_length = 1; } message SliceConfig { required uint32 start = 1; @@ -212,14 +209,14 @@ message ProjectionConfig { // For ShiftProjection optional int32 context_start = 5; optional int32 context_length = 6; - optional bool trainable_padding = 7 [default = false]; + optional bool trainable_padding = 7 [ default = false ]; // For convolution optional ConvConfig conv_conf = 8; optional int32 num_filters = 9; // For IdentityOffsetProjection - optional uint64 offset = 11 [default = 0]; + optional uint64 offset = 11 [ default = 0 ]; // For pool optional PoolConfig pool_conf = 12; @@ -236,7 +233,7 @@ message OperatorConfig { required uint64 output_size = 4; // For DotMulOperator - optional double dotmul_scale = 5 [default = 1.0]; + optional double dotmul_scale = 5 [ default = 1.0 ]; // For ConvOperator optional ConvConfig conv_conf = 6; @@ -282,8 +279,8 @@ message MultiBoxLossConfig { required float neg_overlap = 4; required uint32 background_id = 5; required uint32 input_num = 6; - optional uint32 height = 7 [default = 1]; - optional uint32 width = 8 [default = 1]; + optional uint32 height = 7 [ default = 1 ]; + optional uint32 width = 8 [ default = 1 ]; } message DetectionOutputConfig { @@ -294,8 +291,8 @@ message DetectionOutputConfig { required uint32 input_num = 5; required uint32 keep_top_k = 6; required float confidence_threshold = 7; - optional uint32 height = 8 [default = 1]; - optional uint32 width = 9 [default = 1]; + optional uint32 height = 8 [ default = 1 ]; + optional uint32 width = 9 [ default = 1 ]; } message ClipConfig { @@ -331,7 +328,7 @@ message LayerConfig { required string name = 1; required string type = 2; optional uint64 size = 3; - //optional ActivationConfig activation = 4; + // optional ActivationConfig activation = 4; optional string active_type = 4; repeated LayerInputConfig inputs = 5; optional string bias_parameter_name = 6; @@ -344,7 +341,7 @@ message LayerConfig { // (which is how convnets are usually trained). Setting this to // false will untie the biases, yielding a separate bias for // every location at which the filter is applied. - optional bool shared_biases = 8 [default = false]; + optional bool shared_biases = 8 [ default = false ]; // Valid values are ones that divide the area of the output // grid in this convolutional layer. For example if this layer @@ -362,33 +359,35 @@ message LayerConfig { // the gpu device which the Layer's data in. // Only used by ParallelNeuralNetork. Ignored otherwise. - optional int32 device = 12 [default = -1]; + optional int32 device = 12 [ default = -1 ]; - // for recurrent layer. If true, the recurrence runs from the end to the beginning. - optional bool reversed = 13 [default = false]; + // for recurrent layer. If true, the recurrence runs from the end to the + // beginning. + optional bool reversed = 13 [ default = false ]; - // for lstmemory layer. Different types of nodes have different activation type. - optional string active_gate_type = 14; + // for lstmemory layer. Different types of nodes have different activation + // type. + optional string active_gate_type = 14; optional string active_state_type = 15; // For NCELayer // The number of random negative labels for each sample - optional int32 num_neg_samples = 16 [default = 10]; + optional int32 num_neg_samples = 16 [ default = 10 ]; // For NCELayer // The distribution for generating the random negative labels. // A uniform distribution will be used if not provided - repeated double neg_sampling_dist = 17 [packed = true]; + repeated double neg_sampling_dist = 17 [ packed = true ]; // For MaxLayer // default: output VALUE of MaxLayer. set this flag to true for output INDEX // INDEX will be put in Argument::value as double values. - optional bool output_max_index = 19 [default = false]; + optional bool output_max_index = 19 [ default = false ]; /// The filed number 20 have been deprecated. // For self-normalized estimation - optional double softmax_selfnorm_alpha = 21 [default = 0.1]; + optional double softmax_selfnorm_alpha = 21 [ default = 0.1 ]; /// The filed numbers 22 and 23 have been deprecated. @@ -399,14 +398,14 @@ message LayerConfig { optional bool norm_by_times = 25; // for CostLayers - optional double coeff = 26 [default = 1.0]; + optional double coeff = 26 [ default = 1.0 ]; // for AverageLayer // can be set to: 'average', 'sum' or 'squarerootn' optional string average_strategy = 27; // for error clipping - optional double error_clipping_threshold = 28 [default = 0.0]; + optional double error_clipping_threshold = 28 [ default = 0.0 ]; // for operators used by mixed layer repeated OperatorConfig operator_confs = 29; @@ -434,43 +433,44 @@ message LayerConfig { optional uint32 beam_size = 39; // for seqlastins layer, whether select first instead last - optional bool select_first = 40 [default = false]; + optional bool select_first = 40 [ default = false ]; // for seqlastins layer, AverageLayer, MaxLayer and ExpandLayer // can be set to: 'non-seq','seq' - optional string trans_type = 41 [default = 'non-seq']; + optional string trans_type = 41 [ default = 'non-seq' ]; // to indicate whether selective_fc layer // is used in sequence generation or not - optional bool selective_fc_pass_generation = 42 [default = false]; + optional bool selective_fc_pass_generation = 42 [ default = false ]; // to indicate whether selective_fc layer take its last input to // selected several columns and only compute the multiplications // between the input matrices and the selected columns of // the parameter matrices of this layer. // if set false, selective_fc degrades into fc. - optional bool has_selected_colums = 43 [default = true]; + optional bool has_selected_colums = 43 [ default = true ]; // this parameter is for speed consideration. // if number of the selected columns is less than // sample number * selective_fc output size * selective_fc_mull_mull_ratio // sparse multiplication is used, otherwise, using full multiplication. - optional double selective_fc_full_mul_ratio = 44 [default = 0.02]; + optional double selective_fc_full_mul_ratio = 44 [ default = 0.02 ]; // to indicate how many threads selective_fc use to to accelate // the plain_mul period // leave empty or set to 0 to disable multi-thread accleleration - optional uint32 selective_fc_parallel_plain_mul_thread_num = 45 [default = 0]; + optional uint32 selective_fc_parallel_plain_mul_thread_num = 45 + [ default = 0 ]; // for batch normalization layer // if set use_global_stats true, will use the loaded mean and variance. optional bool use_global_stats = 46; // use to compute moving mean and variance. - optional double moving_average_fraction = 47 [default = 0.9]; + optional double moving_average_fraction = 47 [ default = 0.9 ]; // bias size - optional uint32 bias_size = 48 [default = 0]; + optional uint32 bias_size = 48 [ default = 0 ]; // this parameter can be used as a user-defined parameter when necessary, // without changing the proto file. @@ -485,18 +485,17 @@ message LayerConfig { optional uint64 width = 51; // blank label used in ctc loss - optional uint32 blank = 52 [default = 0]; + optional uint32 blank = 52 [ default = 0 ]; // stride parameter for seqlastins layer, AverageLayer, MaxLayer, which // controls the scope of pooling operation. can be set > 0. // leave empty or set to -1 to disable this stride pooling. - optional int32 seq_pool_stride = 53 [default = -1]; + optional int32 seq_pool_stride = 53 [ default = -1 ]; // for crop layer - optional int32 axis = 54 [default = 2]; + optional int32 axis = 54 [ default = 2 ]; repeated uint32 offset = 55; repeated uint32 shape = 56; - } message EvaluatorConfig { @@ -512,9 +511,9 @@ message EvaluatorConfig { // Used by PrecisionRecallEvaluator and ClassificationErrorEvaluator // For multi binary labels: true if output > classification_threshold - optional double classification_threshold = 6 [default = 0.5]; + optional double classification_threshold = 6 [ default = 0.5 ]; // The positive label. -1 means average precision and recall - optional int32 positive_label = 7 [default = -1]; + optional int32 positive_label = 7 [ default = -1 ]; // load dict from this file optional string dict_file = 8; @@ -523,10 +522,10 @@ message EvaluatorConfig { optional string result_file = 9; // top # results for max id printer - optional int32 num_results = 10 [default = 1]; + optional int32 num_results = 10 [ default = 1 ]; // whether to delimit the sequence in the seq_text_printer - optional bool delimited = 11 [default = true]; + optional bool delimited = 11 [ default = true ]; // Used by ChunkEvaluator // chunk of these types are not counted @@ -534,23 +533,23 @@ message EvaluatorConfig { // Used by ClassificationErrorEvaluator // top # classification error - optional int32 top_k = 13 [default = 1]; + optional int32 top_k = 13 [ default = 1 ]; // Used by DetectionMAPEvaluator - optional double overlap_threshold = 14 [default = 0.5]; + optional double overlap_threshold = 14 [ default = 0.5 ]; - optional int32 background_id = 15 [default = 0]; + optional int32 background_id = 15 [ default = 0 ]; - optional bool evaluate_difficult = 16 [default = false]; + optional bool evaluate_difficult = 16 [ default = false ]; - optional string ap_type = 17 [default = "11point"]; + optional string ap_type = 17 [ default = "11point" ]; } message LinkConfig { required string layer_name = 1; required string link_name = 2; // If true, this link has sub-sequence - optional bool has_subseq = 3 [default = false]; + optional bool has_subseq = 3 [ default = false ]; } message MemoryConfig { @@ -563,18 +562,18 @@ message MemoryConfig { optional uint32 boot_with_const_id = 7; // memory is a sequence, initailized by a sequence boot layer - optional bool is_sequence = 6 [default = false]; + optional bool is_sequence = 6 [ default = false ]; } message GeneratorConfig { required uint32 max_num_frames = 1; required string eos_layer_name = 2; - optional int32 num_results_per_sample = 3 [default = 1]; + optional int32 num_results_per_sample = 3 [ default = 1 ]; // for beam search - optional int32 beam_size = 4 [default = 1]; + optional int32 beam_size = 4 [ default = 1 ]; - optional bool log_prob = 5 [default = true]; + optional bool log_prob = 5 [ default = true ]; } message SubModelConfig { @@ -584,10 +583,10 @@ message SubModelConfig { repeated string output_layer_names = 4; repeated string evaluator_names = 5; - optional bool is_recurrent_layer_group = 6 [default = false]; + optional bool is_recurrent_layer_group = 6 [ default = false ]; // If true, the recurrence runs from the end to the beginning. - optional bool reversed = 7 [default = false]; + optional bool reversed = 7 [ default = false ]; // name and link name of memory repeated MemoryConfig memories = 8; @@ -601,14 +600,15 @@ message SubModelConfig { optional GeneratorConfig generator = 11; - // the id of inlink which share info with outlinks, used in recurrent layer group + // the id of inlink which share info with outlinks, used in recurrent layer + // group optional int32 target_inlinkid = 12; } message ModelConfig { // type of the model. // Currently, "nn", "recurrent_nn" and "recursive_nn" are supported - required string type = 1 [default = "nn"]; + required string type = 1 [ default = "nn" ]; // layers should be ordered in such a way that the forward propagation // can be correctly executed by going from the first layer to the last layer diff --git a/proto/OptimizerConfig.proto b/proto/OptimizerConfig.proto index 2a87e293f64d3398dea2641c3ff292eceec7e154..d27b1bcf80045216a5807812d39f7a248a956076 100644 --- a/proto/OptimizerConfig.proto +++ b/proto/OptimizerConfig.proto @@ -1,5 +1,5 @@ syntax = "proto2"; - + option optimize_for = LITE_RUNTIME; package paddle; @@ -9,13 +9,11 @@ message SGDConfig { // momentum: float >= 0. Parameter updates momentum. // decay: float >= 0. Learning rate decay over each update. // nesterov: boolean. Whether to apply Nesterov momentum. - optional double momentum = 21 [default = 0.0]; - optional double decay = 23 [default = 0.0]; - optional bool nesterov =24 [default = false]; - + optional double momentum = 21 [ default = 0.0 ]; + optional double decay = 23 [ default = 0.0 ]; + optional bool nesterov = 24 [ default = false ]; } - message AdadeltaConfig { // Adadelta // It is recommended to leave it at the default value. @@ -23,21 +21,23 @@ message AdadeltaConfig { // epsilon: float >= 0. Fuzz factor. // decay: float >= 0. Learning rate decay over each update. - // reference : [Adadelta - an adaptive learning rate method](http://arxiv.org/abs/1212.5701) - optional double rho = 33 [default = 0.90]; - optional double epsilon = 31 [default = 1e-5]; - optional double decay = 32 [default = 0.0]; - + // reference : [Adadelta - an adaptive learning rate + // method](http://arxiv.org/abs/1212.5701) + optional double rho = 33 [ default = 0.90 ]; + optional double epsilon = 31 [ default = 1e-5 ]; + optional double decay = 32 [ default = 0.0 ]; } message AdagradConfig { -// Adagrad -// epsilon: float >= 0. -// decay: float >= 0. Learning rate decay over each update. + // Adagrad + // epsilon: float >= 0. + // decay: float >= 0. Learning rate decay over each update. -// reference : [Adaptive Subgradient Methods for Online Learning and Stochastic Optimization](http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) - optional double epsilon = 41 [default = 1e-5]; - optional double decay = 42 [default = 0.0]; + // reference : [Adaptive Subgradient Methods for Online Learning and + // Stochastic + // Optimization](http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) + optional double epsilon = 41 [ default = 1e-5 ]; + optional double decay = 42 [ default = 0.0 ]; } message AdamConfig { @@ -46,7 +46,8 @@ message AdamConfig { // beta_2: float, 0 < beta < 1. Generally close to 1. // epsilon: float >= 0. Fuzz factor. // decay: float >= 0. Learning rate decay over each update. - // reference : [Adam - A Method for Stochastic Optimization](http://arxiv.org/abs/1412.6980v8) + // reference : [Adam - A Method for Stochastic + // Optimization](http://arxiv.org/abs/1412.6980v8) optional double beta_1 = 41; optional double beta_2 = 42; optional double epsilon = 43; @@ -55,32 +56,32 @@ message AdamConfig { message ConstLrConfig { // learninRate Policy - optional double learning_rate = 1 [default = 1.0]; + optional double learning_rate = 1 [ default = 1.0 ]; } message LinearLrConfig { // learninRate Policy - optional double learning_rate = 1 [default = 1.0]; + optional double learning_rate = 1 [ default = 1.0 ]; optional double lr_decay_a = 2; optional double lr_decay_b = 3; } message TensorProto { -enum DataType { - PADDLE_ELEMENT_TYPE_INT32 = 0; - PADDLE_ELEMENT_TYPE_UINT32 = 1; - PADDLE_ELEMENT_TYPE_INT64 = 2; - PADDLE_ELEMENT_TYPE_UINT64 = 3; - PADDLE_ELEMENT_TYPE_FLOAT32 = 4; - PADDLE_ELEMENT_TYPE_FLOAT64 = 5; -} + enum DataType { + PADDLE_ELEMENT_TYPE_INT32 = 0; + PADDLE_ELEMENT_TYPE_UINT32 = 1; + PADDLE_ELEMENT_TYPE_INT64 = 2; + PADDLE_ELEMENT_TYPE_UINT64 = 3; + PADDLE_ELEMENT_TYPE_FLOAT32 = 4; + PADDLE_ELEMENT_TYPE_FLOAT64 = 5; + } optional DataType data_type = 1; repeated bytes content = 2; } message LrPolicyState { // learninRate Policy - optional double learning_rate = 1 [default = 1.0]; + optional double learning_rate = 1 [ default = 1.0 ]; optional double lr_decay_a = 2; optional double lr_decay_b = 3; } @@ -104,7 +105,6 @@ message AdadeltaOptimizerState { optional TensorProto update_delta = 4; } - message AdagradOptimizerState { optional LrPolicyState lr_state = 101; optional double num_sample_passed = 104; @@ -124,10 +124,10 @@ message AdamOptimizerState { message OptimizerConfig { enum Optimizer { - SGD = 1; - Adadelta = 2; - Adagrad = 3; - Adam = 4; + SGD = 1; + Adadelta = 2; + Adagrad = 3; + Adam = 4; } optional Optimizer optimizer = 1; optional SGDConfig sgd = 3; @@ -136,8 +136,8 @@ message OptimizerConfig { optional AdamConfig adam = 6; enum LrPolicy { - Const = 0; - Linear = 1; + Const = 0; + Linear = 1; } optional LrPolicy lr_policy = 11; optional ConstLrConfig const_lr = 12; diff --git a/proto/ParameterConfig.proto b/proto/ParameterConfig.proto index 580d66324602df4c655dd2f1e1cd87159b5b346b..b13570a2c6e7b16e45892a31bb496a9dd2099df0 100644 --- a/proto/ParameterConfig.proto +++ b/proto/ParameterConfig.proto @@ -27,56 +27,57 @@ enum ParameterInitStrategy { message ParameterUpdaterHookConfig { // hook type such as 'pruning' required string type = 1; - // this represents the ratio of zero element to be set by the Parameter - optional double sparsity_ratio = 2 [default = 0.6]; + // this represents the ratio of zero element to be set by the Parameter + optional double sparsity_ratio = 2 [ default = 0.6 ]; } message ParameterConfig { required string name = 1; required uint64 size = 2; - optional double learning_rate = 3 [default = 1.0]; - optional double momentum = 4 [default = 0.0]; - optional double initial_mean = 5 [default = 0.0]; - optional double initial_std = 6 [default = 0.01]; + optional double learning_rate = 3 [ default = 1.0 ]; + optional double momentum = 4 [ default = 0.0 ]; + optional double initial_mean = 5 [ default = 0.0 ]; + optional double initial_std = 6 [ default = 0.01 ]; // use L2-regularization if decay_rate set and decay_rate_l1 not set - optional double decay_rate = 7 [default = 0.0]; + optional double decay_rate = 7 [ default = 0.0 ]; // use L1-regularization if decay_rate_l1 set - optional double decay_rate_l1 = 8 [default = 0.0]; + optional double decay_rate_l1 = 8 [ default = 0.0 ]; // dims of Parameter, e.g. dims[0] as height, dims[1] as width.. repeated uint64 dims = 9; // the gpu device which the parameter in. // Only used by ParallelNeuralNetork. Ignored otherwise. - optional int32 device = 10 [default = -1]; + optional int32 device = 10 [ default = -1 ]; // how to init the parameter: 0 -> normal, 1 -> uniform // 0: treat initial_mean as mean, intial_std as standard deviation // 1: range is (initial_mean - initial_std) to (initial_mean + initial_std) - optional int32 initial_strategy = 11 [default = 0]; + optional int32 initial_strategy = 11 [ default = 0 ]; // define the variance when init the parameter, by height of the Matrix - optional bool initial_smart = 12 [default = false]; + optional bool initial_smart = 12 [ default = false ]; // apply regularization every # batches - optional int32 num_batches_regularization = 13 [default = 1]; + optional int32 num_batches_regularization = 13 [ default = 1 ]; // if is_sparse is true, para is sparse, else para is dense - optional bool is_sparse = 14[default = false]; - // if para is sparse, format should be "csc" or "csr", empty means is not sparse - optional string format = 15 [default = ""]; + optional bool is_sparse = 14 [ default = false ]; + // if para is sparse, format should be "csc" or "csr", empty means is not + // sparse + optional string format = 15 [ default = "" ]; // sparse remote update or not - optional bool sparse_remote_update = 16 [default = false]; + optional bool sparse_remote_update = 16 [ default = false ]; // gradient clipping threshold, no clipping by default - optional double gradient_clipping_threshold = 17 [default = 0.0]; + optional double gradient_clipping_threshold = 17 [ default = 0.0 ]; // static parameters are fixed when training - optional bool is_static = 18 [default = false]; + optional bool is_static = 18 [ default = false ]; // para_id should NOT be set by config_parser. It is for // internal use. optional uint64 para_id = 19; repeated ParameterUpdaterHookConfig update_hooks = 20; // setup load mat -> csr - optional bool need_compact = 21 [default = false]; + optional bool need_compact = 21 [ default = false ]; // whether to do sparse update for this parameter - optional bool sparse_update = 22 [default = false]; + optional bool sparse_update = 22 [ default = false ]; // whether this parameter is shared or not. - optional bool is_shared = 23 [default = false]; + optional bool is_shared = 23 [ default = false ]; // parameter block size - optional uint64 parameter_block_size = 24 [default = 0]; + optional uint64 parameter_block_size = 24 [ default = 0 ]; } diff --git a/proto/ParameterServerConfig.proto b/proto/ParameterServerConfig.proto index 404f9613792653dda72eeb98f022851adedbfbfd..bd63cf35b1483a45f21de6f0d0d883e4d8432296 100644 --- a/proto/ParameterServerConfig.proto +++ b/proto/ParameterServerConfig.proto @@ -15,13 +15,10 @@ syntax = "proto2"; package paddle; - /** * Configuration structure for ParameterClient2. */ -message ParameterClientConfig { - required int32 trainer_id = 1; -} +message ParameterClientConfig { required int32 trainer_id = 1; } /** * Configuration structure for ParameterServer2. @@ -30,24 +27,24 @@ message ParameterServerConfig { // Number of ports for sending dense parameter, // following ports on parameter server will be visited // for sending dense parameter: [port, port+ports_num-1] - required int32 ports_num = 1 [default = 1]; + required int32 ports_num = 1 [ default = 1 ]; // Number of ports for sending sparse parameter, // following ports on parameter server will be visited // for sending sparse parameter: // [port+ports_num, port+ports_num+ports_num_for_sparse-1] - required int32 ports_num_for_sparse = 2 [default = 0]; + required int32 ports_num_for_sparse = 2 [ default = 0 ]; // network device name for pservers - required string nics = 3 [default = "xgbe0,xgbe1"]; - required string rdma_tcp = 4 [default = "tcp"]; + required string nics = 3 [ default = "xgbe0,xgbe1" ]; + required string rdma_tcp = 4 [ default = "tcp" ]; // Listening port for pserver - required int32 port = 5 [default = 20134]; + required int32 port = 5 [ default = 20134 ]; // number of gradient servers - required int32 num_gradient_servers = 6 [default = 1]; + required int32 num_gradient_servers = 6 [ default = 1 ]; // number of threads for sync op exec - required int32 pserver_num_threads = 7 [default = 1]; + required int32 pserver_num_threads = 7 [ default = 1 ]; // control config_.async_lagged_grad_discard_ratio() min value - required double async_lagged_ratio_min = 8 [default = 1.0]; + required double async_lagged_ratio_min = 8 [ default = 1.0 ]; // if async_lagged_grad_discard_ratio is not set in trainer_config.conf // use it as defalut value - required double async_lagged_ratio_default = 9 [default = 1.5]; + required double async_lagged_ratio_default = 9 [ default = 1.5 ]; } \ No newline at end of file diff --git a/proto/ParameterService.proto b/proto/ParameterService.proto index c1c04d8cc5bdedd09173d5dfa10b82c7ee7ed6a4..e3c180ccc3f2a9bfa13c443944cc5ae3398818a9 100644 --- a/proto/ParameterService.proto +++ b/proto/ParameterService.proto @@ -23,8 +23,8 @@ package paddle; */ enum ParameterUpdateMode { // Set parameter - PSERVER_UPDATE_MODE_SET_PARAM = 0;//use local param - PSERVER_UPDATE_MODE_SET_PARAM_ZERO = 1;//set zero param + PSERVER_UPDATE_MODE_SET_PARAM = 0; // use local param + PSERVER_UPDATE_MODE_SET_PARAM_ZERO = 1; // set zero param // Update parameter once a gradient is received PSERVER_UPDATE_MODE_ASYNC_SGD = 2; @@ -37,7 +37,7 @@ enum ParameterUpdateMode { // No update. Only get parameters back. PSERVER_UPDATE_MODE_GET_PARAM = 5; - PSERVER_UPDATE_MODE_GET_PARAM_SPARSE = 6;//only get sparse rows + PSERVER_UPDATE_MODE_GET_PARAM_SPARSE = 6; // only get sparse rows }; message ParameterBlock { @@ -80,42 +80,34 @@ message SendParameterRequest { optional int32 trainer_id = 7; // send back parameter type on pserver, PARAMETER_VALUE by default - optional int32 send_back_parameter_type = 8 [default = 0]; + optional int32 send_back_parameter_type = 8 [ default = 0 ]; // forwardbackward time in usec optional uint64 forwardbackward_time = 9; - } -message WaitPassStartRequest { -} +message WaitPassStartRequest {} -message WaitPassStartResponse { -} +message WaitPassStartResponse {} -message WaitPassFinishRequest { -} +message WaitPassFinishRequest {} -message WaitPassFinishResponse { -} +message WaitPassFinishResponse {} enum SyncObject { SYNC_DEFAULT = 0; // wait for the synchronizeBarrier_ - SYNC_DATA = 1; // wait for the synchronizeDataBarrier_ + SYNC_DATA = 1; // wait for the synchronizeDataBarrier_ } message SynchronizeRequest { - required SyncObject sync_object_id = 1 [default = SYNC_DEFAULT]; + required SyncObject sync_object_id = 1 [ default = SYNC_DEFAULT ]; optional int32 trainer_id = 2; } -message SynchronizeResponse { -} +message SynchronizeResponse {} -message SendParameterResponse { - repeated ParameterBlock blocks = 1; -} +message SendParameterResponse { repeated ParameterBlock blocks = 1; } message SetConfigRequest { repeated ParameterConfig param_configs = 1; @@ -125,26 +117,18 @@ message SetConfigRequest { required bool is_sparse_server = 6; } -message SetConfigResponse{ -} +message SetConfigResponse {} -message GetStatusRequest { -} +message GetStatusRequest {} -message GetStatusResponse { - required PServerStatus status = 1; -} +message GetStatusResponse { required PServerStatus status = 1; } -message SetStatusRequest { - required PServerStatus status = 1; -} +message SetStatusRequest { required PServerStatus status = 1; } -message SetStatusResponse { -} +message SetStatusResponse {} // create a column vector. The size is the dimension of parameter -message CreateVectorRequest { -} +message CreateVectorRequest {} message CreateVectorResponse { // error message. Empty if success @@ -153,9 +137,7 @@ message CreateVectorResponse { required int64 handle = 2; } -message ReleaseVectorRequest { - required int64 handle = 1; -} +message ReleaseVectorRequest { required int64 handle = 1; } message ReleaseVectorResponse { // error message. Empty if success @@ -164,9 +146,7 @@ message ReleaseVectorResponse { // Create a column major matrix. The number of rows is the dimension // of parameter. The number of columns is specifed by num_cols -message CreateMatrixRequest { - required int32 num_cols = 1; -} +message CreateMatrixRequest { required int32 num_cols = 1; } message CreateMatrixResponse { // error message. Empty if success @@ -175,16 +155,13 @@ message CreateMatrixResponse { required int64 handle = 2; } -message ReleaseMatrixRequest { - required int64 handle = 1; -} +message ReleaseMatrixRequest { required int64 handle = 1; } message ReleaseMatrixResponse { // error message. Empty if success optional string return_message = 1; } - /** * The operations are defined using the variables commented at Operation * and OperationResult @@ -245,36 +222,36 @@ enum MatrixVectorOperation { message ProtoVector { required int64 dim = 1; - repeated double values = 2 [packed = true]; + repeated double values = 2 [ packed = true ]; } message ProtoMatrix { required int64 num_rows = 1; required int64 num_cols = 2; - repeated double values = 3 [packed = true]; + repeated double values = 3 [ packed = true ]; } message Operation { required MatrixVectorOperation operation = 1; // vector handles created on the pserver - repeated int64 pvectors = 2; // u, v, w + repeated int64 pvectors = 2; // u, v, w // matrix handles created on the pserver - repeated int64 pmatrices = 3; // A, B, C + repeated int64 pmatrices = 3; // A, B, C - repeated double scalars = 4; // a, b, c - repeated ProtoVector vectors = 5; // x, y, z - repeated ProtoMatrix matrices = 6; // X, Y, Z + repeated double scalars = 4; // a, b, c + repeated ProtoVector vectors = 5; // x, y, z + repeated ProtoMatrix matrices = 6; // X, Y, Z } message OperationResult { // error message. Empty if success optional string return_message = 1; -// - repeated double scalars = 2; // d, e, f + // + repeated double scalars = 2; // d, e, f repeated ProtoVector vectors = 3; // p, q, r - repeated ProtoMatrix matrices = 4; // P, Q, R + repeated ProtoMatrix matrices = 4; // P, Q, R } message DoOperationRequest { @@ -301,18 +278,14 @@ message DoOperationResponse { required bool pass_finish = 3; } -message LoadValueRequest { - required string dir_name = 1; -} +message LoadValueRequest { required string dir_name = 1; } message LoadValueResponse { // error message. Empty if success optional string return_message = 1; } -message SaveValueRequest { - required string dir_name = 1; -} +message SaveValueRequest { required string dir_name = 1; } message SaveValueResponse { // error message. Empty if success @@ -331,11 +304,11 @@ enum DataUpdateMode { // Client send it's own ref label to pserver DATA_UPDATE_MODE_SET_REF_LABEL = 4; // Client get all ref labels from all pservers - DATA_UPDATE_MODE_GET_REF_LABEL =5; + DATA_UPDATE_MODE_GET_REF_LABEL = 5; // Client send it's own ref grad to pserver - DATA_UPDATE_MODE_SET_REF_GRAD =6; + DATA_UPDATE_MODE_SET_REF_GRAD = 6; // Client get all ref grad from all pservers - DATA_UPDATE_MODE_GET_REF_GRAD =7; + DATA_UPDATE_MODE_GET_REF_GRAD = 7; } enum SendDataType { @@ -360,7 +333,7 @@ message DataBlock { // byte size of one data type required int32 data_size = 2; // data_type - optional TransDataType data_type = 3 [default = TRANS_DOUBLE]; + optional TransDataType data_type = 3 [ default = TRANS_DOUBLE ]; } message SendDataRequest { diff --git a/proto/TrainerConfig.proto b/proto/TrainerConfig.proto index a819d20d11ff3932d331801007b8cfb9c77a3f2b..b7c2355159e66be0a1550d3c8fde9a15346ff7e4 100644 --- a/proto/TrainerConfig.proto +++ b/proto/TrainerConfig.proto @@ -20,14 +20,14 @@ package paddle; message OptimizationConfig { required int32 batch_size = 3; - required string algorithm = 4 [default = "async_sgd"]; - optional int32 num_batches_per_send_parameter = 5 [default = 1]; - optional int32 num_batches_per_get_parameter = 6 [default = 1]; + required string algorithm = 4 [ default = "async_sgd" ]; + optional int32 num_batches_per_send_parameter = 5 [ default = 1 ]; + optional int32 num_batches_per_get_parameter = 6 [ default = 1 ]; required double learning_rate = 7; - optional double learning_rate_decay_a = 8 [default = 0]; - optional double learning_rate_decay_b = 9 [default = 0]; - optional string learning_rate_schedule = 27 [default = "constant"]; + optional double learning_rate_decay_a = 8 [ default = 0 ]; + optional double learning_rate_decay_b = 9 [ default = 0 ]; + optional string learning_rate_schedule = 27 [ default = "constant" ]; // learning rate will be scaled according to learning_rate_schedule // 1), constant: // lr = learning_rate @@ -49,88 +49,92 @@ message OptimizationConfig { // owlqn related // L1-regularization - optional double l1weight = 10 [default = 0.1]; + optional double l1weight = 10 [ default = 0.1 ]; // L2-regularization - optional double l2weight = 11 [default = 0]; + optional double l2weight = 11 [ default = 0 ]; // "c1" in wolfe condition: if (newobj <= oldobj + c1 * origDirDeriv * step) // then accept the step - optional double c1 = 12 [default = 0.0001]; + optional double c1 = 12 [ default = 0.0001 ]; // multiply the step with "backoff", when wolfe condition doesn't satisfy - optional double backoff = 13 [default = 0.5]; + optional double backoff = 13 [ default = 0.5 ]; // how many "s"s and "y"s are kept in owlqn - optional int32 owlqn_steps = 14 [default = 10]; + optional int32 owlqn_steps = 14 [ default = 10 ]; // accept the step if encountered "max_backoff" times of "reduce the step" - optional int32 max_backoff = 15 [default = 5]; + optional int32 max_backoff = 15 [ default = 5 ]; // L2-regularization coefficient is reduced linearly from iteration 0 to // "l2weight_zero_iter", and set to 0 after "l2weight_zero_iter" // iterations. set "l2weight_zero_iter" to 0 to disable this strategy. - optional int32 l2weight_zero_iter = 17 [default = 0]; + optional int32 l2weight_zero_iter = 17 [ default = 0 ]; // averaged sgd // About average_window * numBatchProcessed parameter are used // for average. To be accurate, between average_window * numBatchProcessed // and 2 * average_window * numBatchProcessed parameters are used for // average. - optional double average_window = 18 [default = 0]; - optional int64 max_average_window = 19 [default = 0x7fffffffffffffff]; + optional double average_window = 18 [ default = 0 ]; + optional int64 max_average_window = 19 [ default = 0x7fffffffffffffff ]; ////////////////////////// // Options Adaptive SGD // ////////////////////////// - // learning method for sgd/asgd, such as "momentum", "adagrad", "adadelta", "rmsprop" - // default learning method("momentum") use global decayed learning rate with momentum. + // learning method for sgd/asgd, such as "momentum", "adagrad", "adadelta", + // "rmsprop" + // default learning method("momentum") use global decayed learning rate with + // momentum. // "adagrad", "adadelta" and "rmsprop" can set momentum too. - optional string learning_method = 23 [default = "momentum"]; - optional double ada_epsilon = 24 [default = 1e-6]; - optional double ada_rou = 26 [default = 0.95]; + optional string learning_method = 23 [ default = "momentum" ]; + optional double ada_epsilon = 24 [ default = 1e-6 ]; + optional double ada_rou = 26 [ default = 0.95 ]; // Force to do average in cpu in order to save gpu memory usage - optional bool do_average_in_cpu = 25 [default = false]; + optional bool do_average_in_cpu = 25 [ default = false ]; // delta add rate in pserver, used while num_batches_per_send_parameter>1 // will be divided by #machines automatically. - optional double delta_add_rate = 28 [default = 1.0]; + optional double delta_add_rate = 28 [ default = 1.0 ]; // We split a large size into smaller mini-batches, whose sizes are // determined by mini_batch_size. It only takes effect when there is // an ExternalMachine. - optional int32 mini_batch_size = 29 [default = 128]; + optional int32 mini_batch_size = 29 [ default = 128 ]; // automatically set if any one of parameters set sparse remote update flag - optional bool use_sparse_remote_updater = 30 [default = false]; + optional bool use_sparse_remote_updater = 30 [ default = false ]; - // how to update center parameter and feedback to local parameter, + // how to update center parameter and feedback to local parameter, // when use local sgd update in cluster training. - // A option is elastic_average, proposed by the paper: Deep learning with elastic averaging SGD. - // If use elastic_average method, every trainer node should sample from whole data sets. - optional string center_parameter_update_method = 31 [default = "average"]; + // A option is elastic_average, proposed by the paper: Deep learning with + // elastic averaging SGD. + // If use elastic_average method, every trainer node should sample from whole + // data sets. + optional string center_parameter_update_method = 31 [ default = "average" ]; // shrink sparse parameter value // only works if parameter is remote sparse update and has L1 decay rate - optional double shrink_parameter_value = 32 [default = 0]; + optional double shrink_parameter_value = 32 [ default = 0 ]; //////////////////////////// // Options Adam Optimizer // //////////////////////////// - optional double adam_beta1 = 33 [default = 0.9]; - optional double adam_beta2 = 34 [default = 0.999]; - optional double adam_epsilon = 35 [default = 1e-8]; + optional double adam_beta1 = 33 [ default = 0.9 ]; + optional double adam_beta2 = 34 [ default = 0.999 ]; + optional double adam_epsilon = 35 [ default = 1e-8 ]; // arguments for learning rate scheduler // Format: num1:rate1,num2:rate2,...,numK:rateK // For learning_rate_schedule="manual", num is the number of samples, // For learning_rate_schedule="pass_manual", // num is the number of passes (starting from 0) - optional string learning_rate_args = 36 [default = ""]; - + optional string learning_rate_args = 36 [ default = "" ]; + // for async sgd gradient commit control. // when async_lagged_grad_discard_ratio * num_gradient_servers commit passed, // current async gradient will be discard silently. - optional double async_lagged_grad_discard_ratio = 37 [default = 1.5]; + optional double async_lagged_grad_discard_ratio = 37 [ default = 1.5 ]; - // global threshold for gradient clipping - optional double gradient_clipping_threshold = 38 [default = 0.0]; + // global threshold for gradient clipping + optional double gradient_clipping_threshold = 38 [ default = 0.0 ]; }; message TrainerConfig { @@ -141,7 +145,7 @@ message TrainerConfig { repeated string config_files = 5; // the directory to save/load model files for each training path - optional string save_dir = 6 [default = "./output/model"]; + optional string save_dir = 6 [ default = "./output/model" ]; // Path of the initial model parameters. // If it was set, start_pass will be ignored. @@ -149,7 +153,7 @@ message TrainerConfig { // Start training from this pass. // Will load parameter from the previous pass. - optional int32 start_pass = 8 [default = 0]; + optional int32 start_pass = 8 [ default = 0 ]; // file path to the trainer config file optional string config_file = 9; diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt index 0171f9d8ccd6045cb876d57684269a2a49e77f96..7bd6d59b0096c23bb791b9b50702130057628879 100644 --- a/python/CMakeLists.txt +++ b/python/CMakeLists.txt @@ -1,5 +1,3 @@ -set(OUTPUT_DIR - "${CMAKE_CURRENT_BINARY_DIR}/build") file(GLOB TRAINER_PY_FILES . ./paddle/trainer/*.py) file(GLOB HELPERS_PY_FILES . ./paddle/trainer_config_helpers/*.py) @@ -18,28 +16,45 @@ SET(COPY_PADDLE_MASTER "") if(WITH_GOLANG) SET(COPY_PADDLE_MASTER "copy_paddle_master") add_custom_command(TARGET ${COPY_PADDLE_MASTER} - COMMAND cp ${paddle_master_LIB_PATH} ${PROJ_ROOT}/python/paddle/v2/master/ + COMMAND cp ${paddle_master_LIB_PATH} ${PADDLE_SOURCE_DIR}/python/paddle/v2/master/ ) 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) -add_custom_command(OUTPUT ${PROJ_ROOT}/python/paddle/v2/framework/core.so - COMMAND cmake -E copy $ ${PROJ_ROOT}/python/paddle/v2/framework/core.so +add_custom_command(OUTPUT ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/core.so + COMMAND cmake -E copy $ ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/core.so DEPENDS paddle_pybind) -add_custom_target(copy_paddle_pybind ALL DEPENDS ${PROJ_ROOT}/python/paddle/v2/framework/core.so) +add_custom_target(copy_paddle_pybind ALL DEPENDS ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/core.so) -add_custom_command(OUTPUT ${OUTPUT_DIR}/.timestamp +add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel - COMMAND ${CMAKE_COMMAND} -E touch ${OUTPUT_DIR}/.timestamp + COMMAND ${CMAKE_COMMAND} -E touch ${PADDLE_PYTHON_BUILD_DIR}/.timestamp + COMMAND ${CMAKE_COMMAND} -E remove_directory ${PADDLE_PYTHON_BUILD_DIR}/lib-python + 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 - ${OUTPUT_DIR}/.timestamp) +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/config_parser.py b/python/paddle/trainer/config_parser.py index 9ea69fc5e57636c22fb20d5d97de760b9cc3bcde..da99e5bd53458aa0cb201a3525e28c66ab63c52d 100644 --- a/python/paddle/trainer/config_parser.py +++ b/python/paddle/trainer/config_parser.py @@ -1604,6 +1604,8 @@ class MultiClassCrossEntropySelfNormCostLayer(LayerBase): @config_layer('fc') class FCLayer(LayerBase): + layer_type = 'fc' + def __init__(self, name, size, @@ -1611,14 +1613,27 @@ class FCLayer(LayerBase): bias=True, error_clipping_threshold=None, **xargs): - super(FCLayer, self).__init__(name, 'fc', size, inputs=inputs, **xargs) + use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0))) + use_mkldnn_wgt = bool( + int(g_command_config_args.get("use_mkldnn_wgt", 0))) + if use_mkldnn: + self.layer_type = 'mkldnn_fc' + config_assert( + len(inputs) == 1, + "MkldnnFCLayer support one and only one input!") + super(FCLayer, self).__init__( + name, self.layer_type, size, inputs=inputs, **xargs) for input_index in xrange(len(self.inputs)): input_layer = self.get_input_layer(input_index) psize = self.config.size * input_layer.size dims = [input_layer.size, self.config.size] format = self.inputs[input_index].format sparse = format == "csr" or format == "csc" - + if use_mkldnn: + config_assert(not sparse, + "MkldnnFCLayer do not support sparse format yet") + if use_mkldnn_wgt: + dims = [self.config.size, input_layer.size] if sparse: psize = self.inputs[input_index].nnz else: @@ -1631,6 +1646,11 @@ class FCLayer(LayerBase): self.config.error_clipping_threshold = error_clipping_threshold +@config_layer('mkldnn_fc') +class MkldnnFcLayer(FCLayer): + layer_type = 'mkldnn_fc' + + @config_layer('selective_fc') class SelectiveFCLayer(LayerBase): def __init__(self, @@ -2657,6 +2677,31 @@ class SubSequenceLayer(LayerBase): self.create_bias_parameter(bias, size) +@config_layer('sub_nested_seq') +class SubNestedSequenceLayer(LayerBase): + def __init__(self, name, inputs, selected_indices, bias=False, **xargs): + if isinstance(inputs, list): + assert len(inputs) == 1, ('the first input of sub_nested_seq ' + 'layer is a single nested sequence.') + inputs = inputs[0] + if isinstance(selected_indices, list): + assert len(selected_indices) == 1, ( + 'the second input of ' + 'sub_nested_seq layer is a single layer which is a ' + 'set of selected indices.') + selected_indices = selected_indices[0] + + super(SubNestedSequenceLayer, self).__init__( + name, + 'sub_nested_seq', + 0, + inputs=[inputs, selected_indices], + **xargs) + input_layer0 = self.get_input_layer(0) + size = input_layer0.size + self.set_layer_size(size) + + @config_layer('out_prod') class OuterProdLayer(LayerBase): def __init__(self, name, inputs, device=None): @@ -3223,6 +3268,16 @@ class CTCLayer(LayerBase): config_assert(len(self.inputs) == 2, 'CTCLayer must have 2 inputs') +@config_layer('kmax_seq_score') +class KmaxSeqScoreLayer(LayerBase): + def __init__(self, name, inputs, beam_size, **xargs): + super(KmaxSeqScoreLayer, self).__init__( + name, 'kmax_seq_score', 0, inputs=inputs, **xargs) + config_assert( + len(self.inputs) == 1, 'KmaxSeqScoreLayer has only one input.') + self.config.beam_size = beam_size + + @config_layer('warp_ctc') class WarpCTCLayer(LayerBase): def __init__(self, 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 index ea5fdcc50f6abbc67fb61b7fd56c100d9f9811d0..1bc55c869601551aff5fc0311458f906385522d2 100755 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -129,8 +129,10 @@ __all__ = [ 'prelu_layer', 'gated_unit_layer', 'crop_layer', + 'sub_nested_seq_layer', 'clip_layer', 'slice_projection', + 'kmax_sequence_score_layer', ] @@ -224,8 +226,11 @@ class LayerType(object): PRELU = 'prelu' CROP_LAYER = 'crop' + SUB_NESTED_SEQ = 'sub_nested_seq' CLIP_LAYER = 'clip' + KMAX_SEQ_SCORE = 'kmax_seq_score' + @staticmethod def is_layer_type(type_name): """ @@ -6088,6 +6093,53 @@ def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None): size=l.config.size) +@wrap_name_default() +@layer_support() +def sub_nested_seq_layer(input, selected_indices, name=None): + """ + The sub_nested_seq_layer accepts two inputs: the first one is a nested + sequence; the second one is a set of selceted indices in the nested sequence. + + Then sub_nest_seq_layer trims the first nested sequence input according + to the selected indices to form a new output. This layer is useful in + beam training. + + The example usage is: + + .. code-block:: python + + sub_nest_seq = sub_nested_seq_layer(input=[data, selected_indices]) + + + :param input: A nested sequence. + :type input: LayerOutput + :param selected_indices: a set of sequence indices in the nested sequence. + :type input: LayerOutput + :param name: name of this layer. + :type name: basestring + :return: LayerOutput object. + :rtype: LayerOutput + """ + + assert isinstance(input, LayerOutput), ( + 'The first input of ' + 'sub_nested_seq_layer must be a Paddle layer.') + assert isinstance(selected_indices, LayerOutput), ( + 'The second input of ' + 'sub_nested_seq_layer must be a Paddle layer.') + + l = Layer( + inputs=input.name, + selected_indices=selected_indices.name, + name=name, + type=LayerType.SUB_NESTED_SEQ) + return LayerOutput( + name=name, + layer_type=LayerType.SUB_NESTED_SEQ, + parents=input, + size=l.config.size) + + @wrap_name_default("clip") def clip_layer(input, min, max, name=None): """ @@ -6109,7 +6161,8 @@ def clip_layer(input, min, max, name=None): :type min: double :param max: The upper threshold for clipping. :type max: double - :return: LayerOutput + :return: LayerOutput object. + :rtype: LayerOutput """ Layer( name=name, @@ -6119,3 +6172,41 @@ def clip_layer(input, min, max, name=None): max=max) return LayerOutput( name, LayerType.CLIP_LAYER, parents=[input], size=input.size) + + +@wrap_name_default() +@layer_support() +def kmax_sequence_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) + + + :param name: The Layer Name. + :type name: basestring + :param input: The input layer. It stores scores over a sequence or a nested + sequence and its size must be 1. + :type input: LayerOutput. + :param beam_size: squence indices with top beam_size scores are returned. + :type beam_size: double + :return: LayerOutput object. + :rtype: LayerOutput + """ + assert isinstance(input, LayerOutput), ("kmax_sequence_score_layer " + "accepts only one input.") + assert input.size == 1, ( + "input of kmax_sequence_score_layer is a score" + "over a sequence or a nested sequence, so its width must be 1.") + + Layer( + name=name, + type=LayerType.KMAX_SEQ_SCORE, + inputs=[input.name], + beam_size=beam_size) + + return LayerOutput( + name, LayerType.KMAX_SEQ_SCORE, parents=[input], size=input.size) diff --git a/python/paddle/trainer_config_helpers/tests/CMakeLists.txt b/python/paddle/trainer_config_helpers/tests/CMakeLists.txt index 6c860fd49702ebc93612114011361efb885c62ec..580aef935b5cec385a88fb0b4f5b9a5ddeddb40c 100644 --- a/python/paddle/trainer_config_helpers/tests/CMakeLists.txt +++ b/python/paddle/trainer_config_helpers/tests/CMakeLists.txt @@ -1,17 +1,17 @@ #################### test_config_parser ######################### add_test(NAME layers_test - COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/ - ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/python/paddle/trainer_config_helpers/tests/layers_test.py - WORKING_DIRECTORY ${PROJ_ROOT}/python/paddle) + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ + ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/python/paddle/trainer_config_helpers/tests/layers_test.py + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/python/paddle) add_test(NAME test_reset_hook - COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/ - ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/python/paddle/trainer_config_helpers/tests/test_reset_hook.py - WORKING_DIRECTORY ${PROJ_ROOT}/python/paddle) + COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/ + ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/python/paddle/trainer_config_helpers/tests/test_reset_hook.py + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/python/paddle) add_paddle_exe(protobuf_equal ProtobufEqualMain.cpp) add_test(NAME test_layerHelpers COMMAND - ${PROJ_ROOT}/python/paddle/trainer_config_helpers/tests/configs/run_tests.sh ${PYTHON_EXECUTABLE} + ${PADDLE_SOURCE_DIR}/python/paddle/trainer_config_helpers/tests/configs/run_tests.sh ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_BINARY_DIR}/protobuf_equal ) 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 0ffa58bc1e2088f75e7cd25c7ecdffbe270825a4..a61beb871ad064c617fa141451afcb2a5ac64854 100755 --- a/python/paddle/trainer_config_helpers/tests/configs/file_list.sh +++ b/python/paddle/trainer_config_helpers/tests/configs/file_list.sh @@ -7,6 +7,7 @@ test_rnn_group shared_fc shared_lstm shared_gru test_cost_layers_with_weight 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_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer +test_kmax_seq_socre_layer test_seq_select_layers) export whole_configs=(test_split_datasource) 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 new file mode 100644 index 0000000000000000000000000000000000000000..81bd71f68eb3f2c04ccd46ee3b77a07543395c60 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_kmax_seq_socre_layer.protostr @@ -0,0 +1,66 @@ +type: "nn" +layers { + name: "input" + type: "data" + size: 300 + active_type: "" +} +layers { + name: "data" + type: "data" + size: 128 + active_type: "" +} +layers { + name: "__fc_layer_0__" + type: "fc" + size: 1 + active_type: "exponential" + inputs { + input_layer_name: "data" + input_parameter_name: "___fc_layer_0__.w0" + } + bias_parameter_name: "___fc_layer_0__.wbias" +} +layers { + name: "__kmax_sequence_score_layer_0__" + type: "kmax_seq_score" + active_type: "" + inputs { + input_layer_name: "__fc_layer_0__" + } + beam_size: 5 +} +parameters { + name: "___fc_layer_0__.w0" + size: 128 + initial_mean: 0.0 + initial_std: 0.0883883476483 + dims: 128 + 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 +} +input_layer_names: "data" +output_layer_names: "__kmax_sequence_score_layer_0__" +sub_models { + name: "root" + layer_names: "input" + layer_names: "data" + layer_names: "__fc_layer_0__" + layer_names: "__kmax_sequence_score_layer_0__" + input_layer_names: "data" + output_layer_names: "__kmax_sequence_score_layer_0__" + 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_seq_select_layers.protostr new file mode 100644 index 0000000000000000000000000000000000000000..4b906b113e3c0569d5576127e100d097e4923436 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_select_layers.protostr @@ -0,0 +1,37 @@ +type: "nn" +layers { + name: "input_seq" + type: "data" + size: 300 + active_type: "" +} +layers { + name: "input" + type: "data" + size: 5 + active_type: "" +} +layers { + name: "__sub_nested_seq_layer_0__" + type: "sub_nested_seq" + size: 300 + active_type: "" + inputs { + input_layer_name: "input_seq" + } + inputs { + input_layer_name: "input" + } +} +input_layer_names: "input_seq" +output_layer_names: "__sub_nested_seq_layer_0__" +sub_models { + name: "root" + layer_names: "input_seq" + layer_names: "input" + layer_names: "__sub_nested_seq_layer_0__" + input_layer_names: "input_seq" + output_layer_names: "__sub_nested_seq_layer_0__" + is_recurrent_layer_group: false +} + 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 new file mode 100644 index 0000000000000000000000000000000000000000..d245c5a41c793e1f02f306bfe64071bd9885906e --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_kmax_seq_socre_layer.py @@ -0,0 +1,11 @@ +#!/usr/bin/env python +#coding=utf-8 +from paddle.trainer_config_helpers import * + +data = data_layer(name='input', size=300) + +data = data_layer(name="data", size=128) +scores = fc_layer(input=data, size=1, act=ExpActivation()) +kmax_seq_id = kmax_sequence_score_layer(input=scores, beam_size=5) + +outputs(kmax_seq_id) diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_seq_select_layers.py b/python/paddle/trainer_config_helpers/tests/configs/test_seq_select_layers.py new file mode 100644 index 0000000000000000000000000000000000000000..6d1c3175ba9801d69f3f9cb9e754858253192270 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_seq_select_layers.py @@ -0,0 +1,11 @@ +#!/usr/bin/env python +#coding=utf-8 +from paddle.trainer_config_helpers import * + +beam_size = 5 + +data = data_layer(name='input_seq', size=300) +selected_ids = data_layer(name='input', size=beam_size) +sub_nest_seq = sub_nested_seq_layer(input=data, selected_indices=selected_ids) + +outputs(sub_nest_seq) diff --git a/python/paddle/v2/dataset/cifar.py b/python/paddle/v2/dataset/cifar.py index f885b2834e8ad502b752c6fd53daf7ef1693433f..0a2a1ced11ee5cb2fb407b229ce810d553c2fa46 100644 --- a/python/paddle/v2/dataset/cifar.py +++ b/python/paddle/v2/dataset/cifar.py @@ -133,7 +133,7 @@ def convert(path): """ Converts dataset to recordio format """ - paddle.v2.dataset.common.convert(path, train100(), 10, "cifar_train100") - paddle.v2.dataset.common.convert(path, test100(), 10, "cifar_test100") - paddle.v2.dataset.common.convert(path, train10(), 10, "cifar_train10") - paddle.v2.dataset.common.convert(path, test10(), 10, "cifar_test10") + paddle.v2.dataset.common.convert(path, train100(), 1000, "cifar_train100") + paddle.v2.dataset.common.convert(path, test100(), 1000, "cifar_test100") + paddle.v2.dataset.common.convert(path, train10(), 1000, "cifar_train10") + paddle.v2.dataset.common.convert(path, test10(), 1000, "cifar_test10") diff --git a/python/paddle/v2/dataset/common.py b/python/paddle/v2/dataset/common.py index 111496618dfa997246d0a067b0cd4c7dad74f9dc..053ae151c571e5557c9f2f9f4ec866f546a77797 100644 --- a/python/paddle/v2/dataset/common.py +++ b/python/paddle/v2/dataset/common.py @@ -32,17 +32,22 @@ __all__ = [ DATA_HOME = os.path.expanduser('~/.cache/paddle/dataset') + # When running unit tests, there could be multiple processes that # trying to create DATA_HOME directory simultaneously, so we cannot # use a if condition to check for the existence of the directory; # instead, we use the filesystem as the synchronization mechanism by # catching returned errors. -try: - os.makedirs(DATA_HOME) -except OSError as exc: - if exc.errno != errno.EEXIST: - raise - pass +def must_mkdirs(path): + try: + os.makedirs(DATA_HOME) + except OSError as exc: + if exc.errno != errno.EEXIST: + raise + pass + + +must_mkdirs(DATA_HOME) def md5file(fname): @@ -93,6 +98,19 @@ def fetch_all(): "fetch")() +def fetch_all_recordio(path): + for module_name in filter(lambda x: not x.startswith("__"), + dir(paddle.v2.dataset)): + if "convert" in dir( + importlib.import_module("paddle.v2.dataset.%s" % module_name)) and \ + not module_name == "common": + ds_path = os.path.join(path, module_name) + must_mkdirs(ds_path) + getattr( + importlib.import_module("paddle.v2.dataset.%s" % module_name), + "convert")(ds_path) + + def split(reader, line_count, suffix="%05d.pickle", dumper=cPickle.dump): """ you can call the function as: diff --git a/python/paddle/v2/dataset/conll05.py b/python/paddle/v2/dataset/conll05.py index f8aae52e7c29d86c7da9c1da0dd1d093634d4567..23f5a24a1cea7f665fb65e802e1a7811df78208d 100644 --- a/python/paddle/v2/dataset/conll05.py +++ b/python/paddle/v2/dataset/conll05.py @@ -233,5 +233,5 @@ def convert(path): """ Converts dataset to recordio format """ - paddle.v2.dataset.common.convert(path, test(), 10, "conl105_train") - paddle.v2.dataset.common.convert(path, test(), 10, "conl105_test") + paddle.v2.dataset.common.convert(path, test(), 1000, "conl105_train") + paddle.v2.dataset.common.convert(path, test(), 1000, "conl105_test") diff --git a/python/paddle/v2/dataset/imdb.py b/python/paddle/v2/dataset/imdb.py index c0ec5992e0e6b0a2fd2359910d0f7a6c690c2ec3..93dd3e8f7d3a569eaf56335f0f92bed04c0ee26c 100644 --- a/python/paddle/v2/dataset/imdb.py +++ b/python/paddle/v2/dataset/imdb.py @@ -173,5 +173,5 @@ def convert(path): Converts dataset to recordio format """ w = word_dict() - paddle.v2.dataset.common.convert(path, lambda: train(w), 10, "imdb_train") - paddle.v2.dataset.common.convert(path, lambda: test(w), 10, "imdb_test") + paddle.v2.dataset.common.convert(path, lambda: train(w), 1000, "imdb_train") + paddle.v2.dataset.common.convert(path, lambda: test(w), 1000, "imdb_test") diff --git a/python/paddle/v2/dataset/imikolov.py b/python/paddle/v2/dataset/imikolov.py index b18ee8e9ba91e0e8ccf061223b3c0d4636442956..617c722c4165cdfed9e650fc968d623ef6ed4391 100644 --- a/python/paddle/v2/dataset/imikolov.py +++ b/python/paddle/v2/dataset/imikolov.py @@ -155,6 +155,7 @@ def convert(path): N = 5 word_dict = build_dict() paddle.v2.dataset.common.convert(path, - train(word_dict, N), 10, "imikolov_train") + train(word_dict, N), 1000, + "imikolov_train") paddle.v2.dataset.common.convert(path, - test(word_dict, N), 10, "imikolov_test") + test(word_dict, N), 1000, "imikolov_test") diff --git a/python/paddle/v2/dataset/mnist.py b/python/paddle/v2/dataset/mnist.py index ea5891f4f3f6ee1c5023cccee9732cbd9d78b881..9f675bed895223e054cd3bb6e504fe1607f19858 100644 --- a/python/paddle/v2/dataset/mnist.py +++ b/python/paddle/v2/dataset/mnist.py @@ -119,5 +119,5 @@ def convert(path): """ Converts dataset to recordio format """ - paddle.v2.dataset.common.convert(path, train(), 10, "minist_train") - paddle.v2.dataset.common.convert(path, test(), 10, "minist_test") + paddle.v2.dataset.common.convert(path, train(), 1000, "minist_train") + paddle.v2.dataset.common.convert(path, test(), 1000, "minist_test") diff --git a/python/paddle/v2/dataset/movielens.py b/python/paddle/v2/dataset/movielens.py index d9372d422a3293eddeb7c0d5b7c8980f55c44690..5b61a9420af1bb81e1d826f8a7b69f34c306d382 100644 --- a/python/paddle/v2/dataset/movielens.py +++ b/python/paddle/v2/dataset/movielens.py @@ -254,8 +254,8 @@ def convert(path): """ Converts dataset to recordio format """ - paddle.v2.dataset.common.convert(path, train(), 10, "movielens_train") - paddle.v2.dataset.common.convert(path, test(), 10, "movielens_test") + paddle.v2.dataset.common.convert(path, train(), 1000, "movielens_train") + paddle.v2.dataset.common.convert(path, test(), 1000, "movielens_test") if __name__ == '__main__': diff --git a/python/paddle/v2/dataset/sentiment.py b/python/paddle/v2/dataset/sentiment.py index e33f120c8734621fd60497298d993e6e43bd06e0..b0b9757c1a75d215cf8945b5cedbb1239fd43af7 100644 --- a/python/paddle/v2/dataset/sentiment.py +++ b/python/paddle/v2/dataset/sentiment.py @@ -137,5 +137,5 @@ def convert(path): """ Converts dataset to recordio format """ - paddle.v2.dataset.common.convert(path, train, 10, "sentiment_train") - paddle.v2.dataset.common.convert(path, test, 10, "sentiment_test") + paddle.v2.dataset.common.convert(path, train, 1000, "sentiment_train") + paddle.v2.dataset.common.convert(path, test, 1000, "sentiment_test") diff --git a/python/paddle/v2/dataset/uci_housing.py b/python/paddle/v2/dataset/uci_housing.py index ec10ce646ebf3eca2c2a6423b69ee11b6a2b99cf..ce60aa21c2ad1fb8f089d19d548b59a8c806d1ee 100644 --- a/python/paddle/v2/dataset/uci_housing.py +++ b/python/paddle/v2/dataset/uci_housing.py @@ -119,5 +119,5 @@ def convert(path): """ Converts dataset to recordio format """ - paddle.v2.dataset.common.convert(path, train(), 10, "uci_housing_train") - paddle.v2.dataset.common.convert(path, test(), 10, "uci_houseing_test") + paddle.v2.dataset.common.convert(path, train(), 1000, "uci_housing_train") + paddle.v2.dataset.common.convert(path, test(), 1000, "uci_houseing_test") diff --git a/python/paddle/v2/dataset/wmt14.py b/python/paddle/v2/dataset/wmt14.py index 2a631c365f27a6039021a56268a62017638c2739..95a35d97ce9d9503153974cc167ee60829244d5f 100644 --- a/python/paddle/v2/dataset/wmt14.py +++ b/python/paddle/v2/dataset/wmt14.py @@ -169,5 +169,6 @@ def convert(path): Converts dataset to recordio format """ dict_size = 30000 - paddle.v2.dataset.common.convert(path, train(dict_size), 10, "wmt14_train") - paddle.v2.dataset.common.convert(path, test(dict_size), 10, "wmt14_test") + paddle.v2.dataset.common.convert(path, + train(dict_size), 1000, "wmt14_train") + paddle.v2.dataset.common.convert(path, test(dict_size), 1000, "wmt14_test") diff --git a/python/paddle/v2/framework/.gitignore b/python/paddle/v2/framework/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..2ff540d5764b76cf7bac64fc2bb9df6e9c1b398a --- /dev/null +++ b/python/paddle/v2/framework/.gitignore @@ -0,0 +1 @@ +proto diff --git a/python/paddle/v2/framework/create_op_creation_methods.py b/python/paddle/v2/framework/create_op_creation_methods.py deleted file mode 100644 index b034efffb69030cb09e09ea545e9bff6f1744671..0000000000000000000000000000000000000000 --- a/python/paddle/v2/framework/create_op_creation_methods.py +++ /dev/null @@ -1,253 +0,0 @@ -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.attr_type_pb2 as attr_type_pb2 -import cStringIO - - -def get_all_op_protos(): - """ - Get all registered op proto from Paddle C++ - :return: list of OpProto - """ - protostrs = core.get_all_op_protos() - ret_values = [] - for pbstr in protostrs: - op_proto = op_proto_pb2.OpProto.FromString(str(pbstr)) - ret_values.append(op_proto) - return ret_values - - -class OpDescCreationMethod(object): - """ - A Functor object to convert user input(use key word args) to OpDesc based on - 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") - 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 - :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 = attr_type_pb2.INTS - tmp_index_attr.name = "temporary_index" - tmp_index_attr.ints.extend(tmp_index) - - # Types - op_desc.type = self.__op_proto__.type - - # Attrs - for attr in self.__op_proto__.attrs: - if attr.generated: - continue - user_defined_attr = kwargs.get(attr.name, None) - if user_defined_attr is not None: - new_attr = op_desc.attrs.add() - new_attr.name = attr.name - new_attr.type = attr.type - if attr.type == attr_type_pb2.INT: - new_attr.i = user_defined_attr - elif attr.type == attr_type_pb2.FLOAT: - new_attr.f = user_defined_attr - elif attr.type == attr_type_pb2.STRING: - new_attr.s = user_defined_attr - elif attr.type == attr_type_pb2.INTS: - new_attr.ints.extend(user_defined_attr) - elif attr.type == attr_type_pb2.FLOATS: - new_attr.floats.extend(user_defined_attr) - elif attr.type == attr_type_pb2.STRINGS: - new_attr.strings.extend(user_defined_attr) - else: - raise NotImplementedError("Not support attribute type " + - 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 = attr_type_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. - """ - for flag in generator: - if flag: - return True - return False - - -def get_docstring_from_op_proto(op_proto): - """ - Generate docstring from a OpProto - :param op_proto: a OpProto instance. - :type op_proto: op_proto_pb2.OpProto - :return: docstring - """ - if not isinstance(op_proto, op_proto_pb2.OpProto): - raise TypeError("Input must be OpProto") - f = cStringIO.StringIO() - f.write(op_proto.comment) - f.write("\n") - - def __append_param__(name, comment, type): - # Maybe replace the following line with template engine is better. - f.write(":param ") - f.write(name) - f.write(": ") - f.write(comment) - f.write("\n") - f.write(":type ") - f.write(name) - f.write(": ") - f.write(type) - f.write("\n") - - for ipt in op_proto.inputs: - __append_param__(ipt.name, ipt.comment, "list | basestr" - if ipt.multiple else "basestr") - - temp_var_prefix = \ - "This is a temporary variable. It does not have to set by user. " - for opt in op_proto.outputs: - __append_param__(opt.name, opt.comment if not opt.temporary else - temp_var_prefix + opt.comment, "list | basestr" - if opt.multiple else "basestr") - - for attr in op_proto.attrs: - attr_type = None - if attr.type == attr_type_pb2.INT: - attr_type = "int" - elif attr.type == attr_type_pb2.FLOAT: - attr_type = "float" - elif attr.type == attr_type_pb2.STRING: - attr_type = "basestr" - elif attr.type == attr_type_pb2.INTS: - attr_type = "list of int" - elif attr.type == attr_type_pb2.FLOATS: - attr_type = "list of float" - elif attr.type == attr_type_pb2.STRINGS: - attr_type = "list of basestr" - - if attr_type is None: - raise RuntimeError("Not supported attribute type " + attr.type) - - __append_param__(attr.name, attr.comment, attr_type) - - return f.getvalue() - - -def create_op_creation_method(op_proto): - """ - Generate op creation method for an OpProto - """ - method = OpDescCreationMethod(op_proto) - - def __impl__(*args, **kwargs): - opdesc = method(*args, **kwargs) - return core.Operator.create(opdesc.SerializeToString()) - - __impl__.__doc__ = get_docstring_from_op_proto(op_proto) - __impl__.all_input_args = [var.name for var in op_proto.inputs] - __impl__.all_output_args = [var.name for var in op_proto.outputs] - __impl__.all_attr_args = [attr.name for attr in op_proto.attrs] - __impl__.all_not_temp_output_args = [ - var.name for var in op_proto.outputs if not var.temporary - ] - - return __impl__ - - -class OpCreationsHolder(object): - """ - A object will holds all op creation methods. - - Use `op_creations.xxx_op` to access them. - """ - pass - - -op_creations = OpCreationsHolder() - - -def __bootstrap__(): - """ - Bootstrap function for this module. It will dynamic create all op creation - methods in runtime. - """ - for op_proto in get_all_op_protos(): - func = create_op_creation_method(op_proto) - func.__name__ = str(op_proto.type) - setattr(op_creations, func.__name__, func) - - -__bootstrap__() diff --git a/python/paddle/v2/framework/network.py b/python/paddle/v2/framework/network.py deleted file mode 100644 index cfeb0e3dec0fd2c6ad4d2d2501f97932495fdd41..0000000000000000000000000000000000000000 --- a/python/paddle/v2/framework/network.py +++ /dev/null @@ -1,131 +0,0 @@ -import paddle.v2.framework.core as core -from paddle.v2.framework.create_op_creation_methods import op_creations -from default_scope_funcs import new_var, find_var, get_cur_scope - -__all__ = ['Network'] # Only expose Network - - -class NetworkFunctor(object): - """ - Network Op Creation Function. Used internally in this module. - It convert string input to Variable. If it is not created before, just - create in scope. - - It is a functor object. means the instances are callable. - - :param func: The op creation function which generated in Python. - :param net: The Network instance. - """ - - def __init__(self, func, net): - self.func = func - self.net = net - - def __call__(self, *args, **kwargs): - if len(args) != 0: - raise ValueError("Paddle must use keyword argument") - inputs = self.func.all_input_args - for ipt in inputs: - if ipt in kwargs: - var = kwargs[ipt] - if isinstance(var, basestring): - tmp = new_var(var) - self.net.var_names[tmp] = var - var = tmp - - if not isinstance(var, core.Variable): - raise TypeError( - "Input of op creation must be string or variable") - - kwargs[ipt] = self.net.var_names[var] - - notemp_outputs = self.func.all_not_temp_output_args - - for name in notemp_outputs: - if name not in kwargs: - kwargs[ - name] = self.func.__name__ + "@OUT@%d" % core.unique_integer( - ) - - outputs = self.func.all_output_args - for opt in outputs: - if opt in kwargs: - var = kwargs[opt] - if isinstance(var, basestring): - tmp = new_var(var) - self.net.var_names[tmp] = var - var = tmp - - if not isinstance(var, core.Variable): - raise TypeError( - "Output of op creation must be string or variable") - kwargs[opt] = self.net.var_names[var] - - op = self.func(**kwargs) - - self.net.net.add_op(op) - - lst = [find_var(kwargs[opt]) for opt in notemp_outputs] - if len(lst) == 1: - return lst[0] - elif len(lst) == 0: - return None - else: - return lst - - -class Network(object): - """ - The network concept. It avoid user to manually create operator, create - variable, and combine them into a Net. Just use Network.xxx can create the - operator, create variables in default scope, and add them into `self.net`. - - For example: - - .. code-block: python - - net = Network() - out = net.add_two(X="a", Y="b") - fc_out = net.fc(X="out", W="fc.w") - - net.run(...) - """ - - def __init__(self): - self.net = core.Net.create() - funcs = (func_name for func_name in dir(op_creations) - if not func_name.startswith("__")) - self.var_names = dict() - - # TODO(yuyang18): This code can work, but do not generate a good - # docstring, try to give a better way generate function in runtime - # later. - for func_name in funcs: - func = getattr(op_creations, func_name) - impl = NetworkFunctor(func, self) - setattr(self, func_name, impl.__call__) - self.__complete_add_op__ = False - - def infer_shape(self): - self.complete_add_op() - self.net.infer_shape(get_cur_scope()) - - def run(self, device_context): - self.complete_add_op() - self.net.run(get_cur_scope(), device_context) - - def __str__(self): - return str(self.net) - - def complete_add_op(self): - if not self.__complete_add_op__: - self.net.complete_add_op() - self.__complete_add_op__ = True - - -if __name__ == '__main__': - net = Network() - out = net.add_two(X="a", Y="b") - fc_out = net.fc(X=out, W="fc.w", b="fc.b", activation="softmax") - net.complete_add_op() - print net diff --git a/python/paddle/v2/framework/op.py b/python/paddle/v2/framework/op.py new file mode 100644 index 0000000000000000000000000000000000000000..6ac656321e72f5b0c91008091753ee50ac8200a6 --- /dev/null +++ b/python/paddle/v2/framework/op.py @@ -0,0 +1,202 @@ +import paddle.v2.framework.core as core +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 + """ + protostrs = core.get_all_op_protos() + ret_values = [] + for pbstr in protostrs: + 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. + + :param op_proto: The OpProto object. + :type op_proto: op_proto_pb2.OpProto + """ + + def __init__(self, op_proto): + if not isinstance(op_proto, framework_pb2.OpProto): + raise TypeError("Argument should be OpProto") + 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 + :rtype: op_desc_pb2.OpDesc + """ + if len(args) != 0: + raise ValueError("Only keyword arguments is supported by Paddle") + 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 only accepts one input, but give %d" + % (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 only accepts one output, but give %d" % + (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 + + # Attrs + for attr in self.__op_proto__.attrs: + if attr.generated: + continue + user_defined_attr = kwargs.get(attr.name, None) + if user_defined_attr is not None: + new_attr = op_desc.attrs.add() + new_attr.name = attr.name + new_attr.type = attr.type + if attr.type == framework_pb2.INT: + new_attr.i = user_defined_attr + elif attr.type == framework_pb2.FLOAT: + new_attr.f = user_defined_attr + elif attr.type == framework_pb2.STRING: + new_attr.s = user_defined_attr + elif attr.type == framework_pb2.INTS: + new_attr.ints.extend(user_defined_attr) + elif attr.type == framework_pb2.FLOATS: + new_attr.floats.extend(user_defined_attr) + elif attr.type == framework_pb2.STRINGS: + new_attr.strings.extend(user_defined_attr) + else: + raise NotImplementedError("Not support attribute type " + + attr.type) + + return op_desc + + @staticmethod + def any_is_true(generator): + """ + Reduce a bool array to one. If any of them is True, then return True. + """ + for flag in generator: + if flag: + return True + return False + + +class OpInfo(object): + def __init__(self, name, method, inputs, outputs, attrs): + self.name = name + self.method = method + self.inputs = inputs + self.outputs = outputs + self.attrs = attrs + + +def create_op_creation_method(op_proto): + """ + Generate op creation method for an OpProto + """ + method = OpDescCreationMethod(op_proto) + + def __impl__(*args, **kwargs): + opdesc = method(*args, **kwargs) + return core.Operator.create(opdesc.SerializeToString()) + + 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]) + + +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 len(args) != 0: + raise ValueError("All Paddle argument should be key-word " + "argument except type") + t = kwargs.pop('type') + else: + if len(args) != 1: + raise ValueError("All Paddle argument should be key-word " + "argument except type") + t = args[0] + + return self.get_op_info(t).method(**kwargs) + + def types(self): + return self.op_methods.keys() + + def get_op_info(self, t): + if t not in self.op_methods: + raise ValueError("operator %s is not registered", t) + return self.op_methods.get(t) + + def get_op_input_names(self, type): + return self.get_op_info(type).inputs + + def get_op_output_names(self, type): + return self.get_op_info(type).outputs + + def get_op_attr_names(self, type): + return self.get_op_info(type).attrs + + +class __RecurrentOp__(object): + __proto__ = None + type = 'recurrent_op' + + 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 +RecurrentOp = __RecurrentOp__() diff --git a/python/paddle/v2/framework/tests/CMakeLists.txt b/python/paddle/v2/framework/tests/CMakeLists.txt index 4619b0edc3dd7e253e01f7fee5e6a8641340d291..96fad9b42e04a88fdcbda093683b57451b2a3e41 100644 --- a/python/paddle/v2/framework/tests/CMakeLists.txt +++ b/python/paddle/v2/framework/tests/CMakeLists.txt @@ -1,16 +1,27 @@ -add_python_test(test_framework - test_protobuf.py - test_scope.py - test_default_scope_funcs.py - test_op_creation_methods.py - test_net.py - test_tensor.py - test_fc_op.py - test_add_two_op.py - test_sgd_op.py - test_mul_op.py - test_mean_op.py - test_sigmoid_op.py - test_softmax_op.py - test_rowwise_add_op.py - test_network.py) +py_test(test_net SRCS test_net.py) + +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_mean_op SRCS test_mean_op.py) + +py_test(test_protobuf SRCS test_protobuf.py) + +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_fill_zeros_like_op SRCS test_fill_zeros_like_op.py) + +py_test(gradient_checker SRCS gradient_checker.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_uniform_random_op SRCS test_uniform_random_op.py) +py_test(test_recurrent_op SRCS test_recurrent_op.py) diff --git a/python/paddle/v2/framework/tests/gradient_checker.py b/python/paddle/v2/framework/tests/gradient_checker.py new file mode 100644 index 0000000000000000000000000000000000000000..501cf6110ff745b8a6022b463bc9cc3a70145c60 --- /dev/null +++ b/python/paddle/v2/framework/tests/gradient_checker.py @@ -0,0 +1,264 @@ +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 + opts = op.outputs() + for key in opts: + for output in opts[key]: + if local_scope.find_var(output) is None: + local_scope.new_var(output).get_tensor() + op.infer_shape(local_scope) + + # allocate output memory + for key in opts: + for output in opts[key]: + 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() + + no_tmp_out = forward_op.no_intermediate_outputs() + if len(no_tmp_out) != 1: + raise ValueError("non temp out_names should be 1") + + inputs = forward_op.inputs() + in_names = [item for k in inputs for item in inputs[k]] + outputs = forward_op.outputs() + out_names = [item for k in outputs for item in outputs[k]] + + 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) + + bwd_outputs = backward_op.outputs() + bwd_out_names = [item for k in bwd_outputs for item in bwd_outputs[k]] + + 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 out_names: + 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 out_names: + 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 bwd_out_names: + 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/op_test_util.py b/python/paddle/v2/framework/tests/op_test_util.py index 98fae1b975ad6243b20e5c19ec6ff68d5536cd74..dd65e0f2dc23d3f657ff16c55fb297dae210b2d7 100644 --- a/python/paddle/v2/framework/tests/op_test_util.py +++ b/python/paddle/v2/framework/tests/op_test_util.py @@ -1,7 +1,6 @@ -import paddle.v2.framework.core as core -import unittest import numpy -import paddle.v2.framework.create_op_creation_methods as creation +import paddle.v2.framework.core as core +from paddle.v2.framework.op import Operator class OpTestMeta(type): @@ -21,50 +20,52 @@ class OpTestMeta(type): obj = super(OpTestMeta, cls).__new__(cls, name, bases, attrs) def test_all(self): - func = getattr(creation.op_creations, self.type, None) - self.assertIsNotNone(func) - scope = core.Scope() kwargs = dict() - places = [] - places.append(core.CPUPlace()) + places = [core.CPUPlace()] if core.is_compile_gpu(): places.append(core.GPUPlace(0)) for place in places: - for in_name in func.all_input_args: - if hasattr(self, in_name): + 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 = getattr(self, in_name) + arr = self.inputs[in_name] var.set_dims(arr.shape) var.set(arr, place) else: kwargs[in_name] = "@EMPTY@" - for out_name in func.all_output_args: - if hasattr(self, out_name): - kwargs[out_name] = out_name - scope.new_var(out_name).get_tensor() + 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 func.all_attr_args: - if hasattr(self, attr_name): - kwargs[attr_name] = getattr(self, attr_name) + 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 = func(**kwargs) + 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 func.all_output_args: + for out_name in Operator.get_op_output_names(self.type): actual = numpy.array(scope.find_var(out_name).get_tensor()) - expect = getattr(self, out_name) - # TODO(qijun) The default decimal is 7, but numpy.dot and eigen.mul - # has some diff, and could not pass unittest. So I set decimal 3 here. - # And I will check this in future. - numpy.testing.assert_almost_equal(actual, expect, decimal=3) + 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 6e6643201bf361fce1bad7de10b2562f0525e00a..0def484eddb88604398ee10390d3f28058714a57 100644 --- a/python/paddle/v2/framework/tests/test_add_two_op.py +++ b/python/paddle/v2/framework/tests/test_add_two_op.py @@ -2,7 +2,7 @@ import unittest import numpy import paddle.v2.framework.core as core -import paddle.v2.framework.create_op_creation_methods as creation +from paddle.v2.framework.op import Operator from op_test_util import OpTestMeta @@ -12,18 +12,11 @@ class TestAddOp(unittest.TestCase): def setUp(self): self.type = "add_two" - self.X = numpy.random.random((102, 105)).astype("float32") - self.Y = numpy.random.random((102, 105)).astype("float32") - self.Out = self.X + self.Y - - -class TestAddGradOp(unittest.TestCase): - def test_add_grad(self): - op = creation.op_creations.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)) + self.inputs = { + 'X': numpy.random.random((102, 105)).astype("float32"), + 'Y': numpy.random.random((102, 105)).astype("float32") + } + self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']} if __name__ == '__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 609c56535ef0365dda728cba334d8b4d96312192..4815192e255c6e0429db3f50918a76a773b30131 100644 --- a/python/paddle/v2/framework/tests/test_cross_entropy_op.py +++ b/python/paddle/v2/framework/tests/test_cross_entropy_op.py @@ -1,21 +1,37 @@ import unittest import numpy from op_test_util import OpTestMeta +from gradient_checker import GradientChecker, create_op -class TestSGD(unittest.TestCase): +class TestCrossEntropy(unittest.TestCase): __metaclass__ = OpTestMeta def setUp(self): + # TODO this unit test is not passed self.type = "onehot_cross_entropy" batch_size = 100 class_num = 10 - self.X = numpy.random.random((batch_size, class_num)).astype("float32") - self.label = 5 * numpy.ones(batch_size).astype("int32") + X = numpy.random.random((batch_size, class_num)).astype("float32") + label = 5 * numpy.ones(batch_size).astype("int32") + self.inputs = {'X': X, 'label': label} Y = [] for i in range(0, batch_size): - Y.append(-numpy.log(self.X[i][self.label[i]])) - self.Y = numpy.array(Y).astype("float32") + Y.append(-numpy.log(X[i][label[i]])) + self.outputs = {'Y': numpy.array(Y).astype("float32")} + + +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") if __name__ == "__main__": diff --git a/python/paddle/v2/framework/tests/test_fc_op.py b/python/paddle/v2/framework/tests/test_fc_op.py deleted file mode 100644 index 00dc4399aaf59e6382692c3a4356f89a7e79a0c5..0000000000000000000000000000000000000000 --- a/python/paddle/v2/framework/tests/test_fc_op.py +++ /dev/null @@ -1,45 +0,0 @@ -import paddle.v2.framework.core as core -import unittest -import numpy -import paddle.v2.framework.create_op_creation_methods as creation - - -class TestFc(unittest.TestCase): - def test_fc(self): - scope = core.Scope() - place = core.CPUPlace() - x = scope.new_var("X") - - x_tensor = x.get_tensor() - x_tensor.set_dims([1000, 784]) - x_tensor.alloc_float(place) - - w = scope.new_var("W") - w_tensor = w.get_tensor() - w_tensor.set_dims([784, 100]) - w_tensor.alloc_float(place) - - w_tensor.set(numpy.random.random((784, 100)).astype("float32"), place) - - # Set a real numpy array here. - # x_tensor.set(numpy.array([])) - - op = creation.op_creations.fc(X="X", Y="Y", W="W") - - for out in op.outputs(): - if scope.find_var(out) is None: - scope.new_var(out).get_tensor() - - tensor = scope.find_var("Y").get_tensor() - op.infer_shape(scope) - self.assertEqual([1000, 100], tensor.shape()) - - ctx = core.DeviceContext.create(place) - - op.run(scope, ctx) - - # After complete all ops, check Y is expect or not. - - -if __name__ == '__main__': - unittest.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 new file mode 100644 index 0000000000000000000000000000000000000000..e5c862605fb11a5ea1426cf8f9054589dc377ff1 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py @@ -0,0 +1,16 @@ +import unittest +from op_test_util import OpTestMeta +import numpy + + +class TestFillZerosLikeOp(unittest.TestCase): + __metaclass__ = OpTestMeta + + 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'])} + + +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 new file mode 100644 index 0000000000000000000000000000000000000000..f95ed70b58d611b3233a21d3f2a34c864ae4d1b3 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_gaussian_random_op.py @@ -0,0 +1,36 @@ +import unittest +import paddle.v2.framework.core as core +from paddle.v2.framework.op import Operator +import numpy + + +class GaussianRandomTest(unittest.TestCase): + def test_cpu(self): + self.gaussian_random_test(place=core.CPUPlace()) + + def test_gpu(self): + if core.is_compile_gpu(): + self.gaussian_random_test(place=core.GPUPlace(0)) + + def gaussian_random_test(self, place): + scope = core.Scope() + scope.new_var("Out").get_tensor() + + op = Operator( + "gaussian_random", + Out="Out", + dims=[1000, 784], + mean=.0, + std=1., + seed=10) + + op.infer_shape(scope) + context = core.DeviceContext.create(place) + op.run(scope, context) + 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__': + 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 78fff1eeff998109a51ea662f963a102eff49d3a..b5d52b90567bcd0c9f376147145d8638049f7bab 100644 --- a/python/paddle/v2/framework/tests/test_mean_op.py +++ b/python/paddle/v2/framework/tests/test_mean_op.py @@ -8,8 +8,8 @@ class TestMeanOp(unittest.TestCase): def setUp(self): self.type = "mean" - self.X = np.random.random((32, 784)).astype("float32") - self.Out = np.mean(self.X) + self.inputs = {'X': np.random.random((32, 784)).astype("float32")} + self.outputs = {'Out': np.mean(self.inputs['X'])} if __name__ == '__main__': diff --git a/python/paddle/v2/framework/tests/test_mul_op.py b/python/paddle/v2/framework/tests/test_mul_op.py index e1ac66d3a4d23d617f7c5a4d97d070b2660954c8..ec0ac99156a546dd3fb7b27778032bece38ab5a9 100644 --- a/python/paddle/v2/framework/tests/test_mul_op.py +++ b/python/paddle/v2/framework/tests/test_mul_op.py @@ -8,9 +8,11 @@ class TestMulOp(unittest.TestCase): def setUp(self): self.type = "mul" - self.X = np.random.random((32, 84)).astype("float32") - self.Y = np.random.random((84, 100)).astype("float32") - self.Out = np.dot(self.X, self.Y) + 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'])} if __name__ == '__main__': diff --git a/python/paddle/v2/framework/tests/test_net.py b/python/paddle/v2/framework/tests/test_net.py index db776d6b643dc4014da9f5dded8219180af639e3..b42cadd11ab75abbc35763c8d12e8c27e995f0dc 100644 --- a/python/paddle/v2/framework/tests/test_net.py +++ b/python/paddle/v2/framework/tests/test_net.py @@ -1,27 +1,36 @@ import paddle.v2.framework.core as core -from paddle.v2.framework.create_op_creation_methods import op_creations +from paddle.v2.framework.op import Operator 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.complete_add_op(True) + return ret_v + + class TestNet(unittest.TestCase): def test_net_all(self): net = core.Net.create() - op1 = op_creations.add_two(X="X", Y="Y", Out="Out") + op1 = Operator("add_two", X="X", Y="Y", Out="Out") net.add_op(op1) net2 = core.Net.create() - net2.add_op(op_creations.fc(X="X", W="w", Y="fc.out")) + net2.add_op(fc(X="X", W="w", Y="fc.out")) net2.complete_add_op(True) net.add_op(net2) net.complete_add_op(True) expected = ''' -Op(plain_net), inputs:(@EMPTY@, X, Y, w), outputs:(@TEMP@fc@0, Out, fc.out). - Op(add_two), inputs:(X, Y), outputs:(Out). - Op(plain_net), inputs:(@EMPTY@, X, w), outputs:(@TEMP@fc@0, fc.out). - Op(fc), inputs:(X, w, @EMPTY@), outputs:(fc.out, @TEMP@fc@0). - Op(mul), inputs:(X, w), outputs:(@TEMP@fc@0). - Op(sigmoid), inputs:(@TEMP@fc@0), outputs:(fc.out). +Op(plain_net), inputs:{all[W, X, Y]}, outputs:{all[Out, fc.out, pre_activation]}. + Op(add_two), 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)) diff --git a/python/paddle/v2/framework/tests/test_network.py b/python/paddle/v2/framework/tests/test_network.py deleted file mode 100644 index 6d53e233e959bd39b558ac97cdca381135505f8d..0000000000000000000000000000000000000000 --- a/python/paddle/v2/framework/tests/test_network.py +++ /dev/null @@ -1,32 +0,0 @@ -from paddle.v2.framework.network import Network -import paddle.v2.framework.core as core -import unittest - - -class TestNet(unittest.TestCase): - def test_net_all(self): - net = Network() - out = net.add_two(X="X", Y="Y") - fc_out = net.fc(X=out, W="w") - net.complete_add_op() - self.assertTrue(isinstance(fc_out, core.Variable)) - self.assertEqual( - '''Op(plain_net), inputs:(@EMPTY@, X, Y, w), outputs:(@TEMP@fc@0, add_two@OUT@0, fc@OUT@1). - Op(add_two), inputs:(X, Y), outputs:(add_two@OUT@0). - Op(fc), inputs:(add_two@OUT@0, w, @EMPTY@), outputs:(fc@OUT@1, @TEMP@fc@0). - Op(mul), inputs:(add_two@OUT@0, w), outputs:(@TEMP@fc@0). - Op(sigmoid), inputs:(@TEMP@fc@0), outputs:(fc@OUT@1). -''', str(net)) - - net2 = Network() - tmp = net2.add_two(X="X", Y="Y") - self.assertTrue(isinstance(tmp, core.Variable)) - net2.complete_add_op() - self.assertEqual( - '''Op(plain_net), inputs:(X, Y), outputs:(add_two@OUT@2). - Op(add_two), inputs:(X, Y), outputs:(add_two@OUT@2). -''', str(net2)) - - -if __name__ == '__main__': - unittest.main() diff --git a/python/paddle/v2/framework/tests/test_op_creation_methods.py b/python/paddle/v2/framework/tests/test_op_creation_methods.py deleted file mode 100644 index 41db7c0d535aa920b34d6cc346090a8c15bfb110..0000000000000000000000000000000000000000 --- a/python/paddle/v2/framework/tests/test_op_creation_methods.py +++ /dev/null @@ -1,254 +0,0 @@ -import unittest -import paddle.v2.framework.create_op_creation_methods as creation -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.attr_type_pb2 as attr_type_pb2 - - -class TestGetAllProtos(unittest.TestCase): - def test_all(self): - all_protos = creation.get_all_op_protos() - self.assertNotEqual(0, len(all_protos)) - - for each in all_protos: - self.assertTrue(each.IsInitialized()) - - -class TestOpDescCreationMethod(unittest.TestCase): - def test_plain_input_output(self): - op = op_proto_pb2.OpProto() - op.type = "test" - ipt = op.inputs.add() - ipt.name = "X" - ipt.comment = "not matter" - - ipt = op.inputs.add() - ipt.name = "Y" - ipt.comment = "not matter" - - opt = op.outputs.add() - opt.name = "Z" - opt.comment = "not matter" - - op.comment = "not matter" - - self.assertTrue(op.IsInitialized()) - - method = creation.OpDescCreationMethod(op) - output = method(X="a", Y="b", Z="c") - - expected = op_desc_pb2.OpDesc() - expected.type = "test" - expected.inputs.extend(["a", "b"]) - expected.outputs.append("c") - self.assertEqual(expected, output) - - def test_multiple_input_plain_output(self): - op = op_proto_pb2.OpProto() - op.type = "fc" - ipt = op.inputs.add() - ipt.name = "X" - ipt.comment = "" - ipt.multiple = True - - ipt = op.inputs.add() - ipt.name = "W" - ipt.comment = "" - ipt.multiple = True - - ipt = op.inputs.add() - ipt.name = "b" - ipt.comment = "" - - out = op.outputs.add() - out.name = "Y" - out.comment = "" - - op.comment = "" - self.assertTrue(op.IsInitialized()) - method = creation.OpDescCreationMethod(op) - - 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.type = 'fc' - attr = expected1.attrs.add() - attr.name = 'input_format' - attr.type = attr_type_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.type = 'fc' - attr = expected2.attrs.add() - attr.name = 'input_format' - attr.type = attr_type_pb2.INTS - attr.ints.extend([0, 3, 6, 7]) - self.assertEqual(expected2, generated2) - - def test_attrs(self): - op = op_proto_pb2.OpProto() - op.type = "test" - ipt = op.inputs.add() - ipt.name = 'X' - ipt.comment = "" - - def __add_attr__(name, type): - attr = op.attrs.add() - attr.name = name - attr.comment = "" - attr.type = type - - __add_attr__("int_attr", attr_type_pb2.INT) - __add_attr__("float_attr", attr_type_pb2.FLOAT) - __add_attr__("string_attr", attr_type_pb2.STRING) - __add_attr__("ints_attr", attr_type_pb2.INTS) - __add_attr__("floats_attr", attr_type_pb2.FLOATS) - __add_attr__("strings_attr", attr_type_pb2.STRINGS) - - op.comment = "" - self.assertTrue(op.IsInitialized()) - - method = creation.OpDescCreationMethod(op) - - generated = method( - X="a", - int_attr=10, - float_attr=3.2, - string_attr="test_str", - ints_attr=[0, 1, 2, 3, 4], - floats_attr=[0.2, 3.2, 4.5], - strings_attr=["a", "b", "c"]) - - expected = op_desc_pb2.OpDesc() - expected.type = "test" - expected.inputs.extend(['a']) - attr = expected.attrs.add() - attr.name = "int_attr" - attr.type = attr_type_pb2.INT - attr.i = 10 - - attr = expected.attrs.add() - attr.name = "float_attr" - attr.type = attr_type_pb2.FLOAT - attr.f = 3.2 - - attr = expected.attrs.add() - attr.name = "string_attr" - attr.type = attr_type_pb2.STRING - attr.s = "test_str" - - attr = expected.attrs.add() - attr.name = "ints_attr" - attr.type = attr_type_pb2.INTS - attr.ints.extend([0, 1, 2, 3, 4]) - - attr = expected.attrs.add() - attr.name = 'floats_attr' - attr.type = attr_type_pb2.FLOATS - attr.floats.extend([0.2, 3.2, 4.5]) - - attr = expected.attrs.add() - attr.name = 'strings_attr' - attr.type = attr_type_pb2.STRINGS - attr.strings.extend(['a', 'b', 'c']) - - self.assertEqual(expected, generated) - - def test_input_temporary_output(self): - op = op_proto_pb2.OpProto() - op.type = "test" - out = op.outputs.add() - out.name = "OUT" - out.comment = "" - - out = op.outputs.add() - out.name = "TMP" - out.comment = "" - out.temporary = True - - out = op.outputs.add() - out.name = "OUT2" - out.comment = "" - op.comment = "" - - method = creation.OpDescCreationMethod(op) - 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 = attr_type_pb2.INTS - attr.ints.append(2) - self.assertEqual(generated, desc) - - -class TestOpCreationDocStr(unittest.TestCase): - def test_all(self): - op = op_proto_pb2.OpProto() - op.type = "test" - op.comment = """Test Op. - -This op is used for unit test, not a real op. -""" - a = op.inputs.add() - a.name = "a" - a.comment = "Input a for test op" - a.multiple = True - - b = op.inputs.add() - b.name = "b" - b.comment = "Input b for test op" - self.assertTrue(op.IsInitialized()) - - o1 = op.outputs.add() - o1.name = "output" - o1.comment = "The output of test op" - - o2 = op.outputs.add() - o2.name = "temp output" - o2.comment = "The temporary output of test op" - o2.temporary = True - - test_str = op.attrs.add() - test_str.name = "str_attr" - test_str.type = attr_type_pb2.STRING - test_str.comment = "A string attribute for test op" - - actual = creation.get_docstring_from_op_proto(op) - expected_docstring = '''Test Op. - -This op is used for unit test, not a real op. - -:param a: Input a for test op -:type a: list | basestr -:param b: Input b for test op -:type b: basestr -:param output: The output of test op -:type output: basestr -:param temp output: This is a temporary variable. It does not have to set by user. The temporary output of test op -:type temp output: basestr -:param str_attr: A string attribute for test op -:type str_attr: basestr -''' - self.assertEqual(expected_docstring, actual) - - -class TestOpCreations(unittest.TestCase): - def test_all(self): - add_op = creation.op_creations.add_two(X="a", Y="b", Out="z") - self.assertIsNotNone(add_op) - # Invoke C++ DebugString() - self.assertEqual('Op(add_two), inputs:(a, b), outputs:(z).', - str(add_op)) - - -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 new file mode 100644 index 0000000000000000000000000000000000000000..1abc4eeb57bcedc81e34b0e156048ee4f5cfdc2d --- /dev/null +++ b/python/paddle/v2/framework/tests/test_operator.py @@ -0,0 +1,204 @@ +import unittest +import paddle.v2.framework.op as op +import paddle.v2.framework.core as core +import paddle.v2.framework.proto.framework_pb2 as framework_pb2 + + +class TestGetAllProtos(unittest.TestCase): + def test_all(self): + all_protos = op.get_all_op_protos() + self.assertNotEqual(0, len(all_protos)) + + for each in all_protos: + self.assertTrue(each.IsInitialized()) + + +class TestOpDescCreationMethod(unittest.TestCase): + def test_plain_input_output(self): + op_proto = framework_pb2.OpProto() + op_proto.type = "test" + ipt = op_proto.inputs.add() + ipt.name = "X" + ipt.comment = "not matter" + + ipt = op_proto.inputs.add() + ipt.name = "Y" + ipt.comment = "not matter" + + opt = op_proto.outputs.add() + opt.name = "Z" + opt.comment = "not matter" + + op_proto.comment = "not matter" + + self.assertTrue(op_proto.IsInitialized()) + + method = op.OpDescCreationMethod(op_proto) + output = method(X="a", Y="b", Z="c") + expected = framework_pb2.OpDesc() + expected.type = "test" + 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 = framework_pb2.OpProto() + op_proto.type = "fc" + ipt = op_proto.inputs.add() + ipt.name = "X" + ipt.comment = "" + ipt.duplicable = True + + ipt = op_proto.inputs.add() + ipt.name = "W" + ipt.comment = "" + ipt.duplicable = True + + ipt = op_proto.inputs.add() + ipt.name = "b" + ipt.comment = "" + + out = op_proto.outputs.add() + out.name = "Y" + out.comment = "" + + op_proto.comment = "" + self.assertTrue(op_proto.IsInitialized()) + method = op.OpDescCreationMethod(op_proto) + + generated1 = method(X="x", W="w", b="b", Y="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' + self.assertEqual(expected1, generated1) + + generated2 = method( + X=['x1', 'x2', 'x3'], b='b', W=['w1', 'w2', 'w3'], Y='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' + self.assertEqual(expected2, generated2) + + def test_attrs(self): + op_proto = framework_pb2.OpProto() + op_proto.type = "test" + ipt = op_proto.inputs.add() + ipt.name = 'X' + ipt.comment = "" + + def __add_attr__(name, type): + attr = op_proto.attrs.add() + attr.name = name + attr.comment = "" + attr.type = type + + __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()) + + method = op.OpDescCreationMethod(op_proto) + + generated = method( + X="a", + int_attr=10, + float_attr=3.2, + string_attr="test_str", + ints_attr=[0, 1, 2, 3, 4], + floats_attr=[0.2, 3.2, 4.5], + strings_attr=["a", "b", "c"]) + + expected = framework_pb2.OpDesc() + expected.type = "test" + + ipt = expected.inputs.add() + ipt.parameter = "X" + ipt.arguments.extend(['a']) + + attr = expected.attrs.add() + attr.name = "int_attr" + attr.type = framework_pb2.INT + attr.i = 10 + + attr = expected.attrs.add() + attr.name = "float_attr" + attr.type = framework_pb2.FLOAT + attr.f = 3.2 + + attr = expected.attrs.add() + attr.name = "string_attr" + attr.type = framework_pb2.STRING + attr.s = "test_str" + + attr = expected.attrs.add() + attr.name = "ints_attr" + attr.type = framework_pb2.INTS + attr.ints.extend([0, 1, 2, 3, 4]) + + attr = expected.attrs.add() + attr.name = 'floats_attr' + attr.type = framework_pb2.FLOATS + attr.floats.extend([0.2, 3.2, 4.5]) + + attr = expected.attrs.add() + attr.name = 'strings_attr' + attr.type = framework_pb2.STRINGS + attr.strings.extend(['a', 'b', 'c']) + + self.assertEqual(expected, generated) + + +class TestOpCreations(unittest.TestCase): + def test_all(self): + add_op = op.Operator("add_two", X="a", Y="b", Out="z") + self.assertIsNotNone(add_op) + # Invoke C++ DebugString() + self.assertEqual('Op(add_two), inputs:{X[a], Y[b]}, outputs:{Out[z]}.', + str(add_op)) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_protobuf.py b/python/paddle/v2/framework/tests/test_protobuf.py index b8702477e64203e735bff05b115eafbb2a52172d..848a396b3b6eec57d500b464780b64f339b09e94 100644 --- a/python/paddle/v2/framework/tests/test_protobuf.py +++ b/python/paddle/v2/framework/tests/test_protobuf.py @@ -1,13 +1,10 @@ -import paddle.v2.framework.proto.op_proto_pb2 -import paddle.v2.framework.proto.attr_type_pb2 +import paddle.v2.framework.proto.framework_pb2 as framework_pb2 import unittest class TestFrameworkProto(unittest.TestCase): def test_all(self): - op_proto_lib = paddle.v2.framework.proto.op_proto_pb2 - attr_type_lib = paddle.v2.framework.proto.attr_type_pb2 - 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" @@ -21,7 +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 0457e3f16a709140180ce433c1d56d146f0b6974..3d4a34d8d713ff1beeeba8ac48ad95176f7a29f2 100644 --- a/python/paddle/v2/framework/tests/test_recurrent_op.py +++ b/python/paddle/v2/framework/tests/test_recurrent_op.py @@ -1,20 +1,75 @@ +import logging import paddle.v2.framework.core as core import unittest import numpy as np -import paddle.v2.framework.create_op_creation_methods as creation +from paddle.v2.framework.op import Operator, RecurrentOp -ops = creation.op_creations +def py_sigmoid(x): + return 1. / (1. + np.exp(-x)) -def create_tensor(scope, name, shape): - tensor = scope.create_var(name).get_tensor() + +class PySimpleRNN(object): + ''' + A simple implementation of RNN based on numpy, to futhur test RecurrentOp's alogorithm + ''' + + def __init__(self, input_dim=30, batch_size=50, weight_dim=15, sent_len=11): + self.x = np.random.normal(size=(sent_len, batch_size, input_dim)) + self.W = np.random.normal(size=(input_dim, input_dim)) + self.U = np.random.normal(size=(input_dim, input_dim)) + self.h_boot = np.random.normal(size=(batch_size, input_dim)) + + # memories + self.mems = [ + np.zeros(shape=(batch_size, input_dim)) for i in range(sent_len) + ] + + def forward(self): + xs = self.segment_inputs() + for step_id in range(self.x.shape[0]): + self.step(step_id, xs[step_id]) + return self.concat_outputs() + + def segment_inputs(self): + return [self.x[i] for i in range(self.x.shape[0])] + + def concat_outputs(self): + return np.array(self.mems) + + def step(self, step_id, x): + ''' + run a step + ''' + mem = self.mems[step_id] + if step_id > 0: + pre_mem = self.mems[step_id - 1] + else: + pre_mem = self.h_boot + xW = np.matmul(x, self.W) + hU = np.matmul(mem, self.U) + + sum = xW + hU + self.mems[step_id] = py_sigmoid(sum) + + +class PySimpleRNNTest(unittest.TestCase): + def setUp(self): + self.rnn = PySimpleRNN() + + def test_forward(self): + output = self.rnn.forward() + print 'output', output + + +def create_tensor(scope, name, shape, np_data): + tensor = scope.new_var(name).get_tensor() tensor.set_dims(shape) - tensor.alloc_float() - tensor.set(np.random.random(shape)) + tensor.set(np_data, core.CPUPlace()) return tensor -class TestRNN(unittest.TestCase): +class TestRecurrentOp(unittest.TestCase): ''' Test RNNOp @@ -28,64 +83,85 @@ class TestRNN(unittest.TestCase): memories: - h outputs: - - h + - h ''' - def init(self): - input_dim = 30 - batch_size = 50 - weight_dim = 15 - - self.scope = core.Scope(None) - - # create vars - create_tensor(self.scope, "x", [batch_size, input_dim]) - create_tensor(self.scope, "W", [input_dim, weight_dim]) - create_tensor(self.scope, "U", [weight_dim, weight_dim]) - create_tensor(self.scope, "h_boot", [batch_size, weight_dim]) - - x_alias = "x@alias" - y_alias = "y@alias" - memory = "h@alias" - prememory = "h@pre" - output = "rnn_out" - output_alias = "rnn_out@alias" - - # create step net - stepnet_var = self.scope.create_var("stepnet") - stepnet = stepnet_var.get_net() - # stepnet = core.Net.create() - x_fc_op = ops.fc(X=x_alias, W="W", Y="Wx") - h_fc_op = ops.fc(X=prememory, W="U", Y="Uh") - sum_op = ops.add_two(X="Wx", Y="Uh", Out="sum") - sig_op = ops.sigmoid(X="sum", Y=memory) - stepnet.add_op(x_fc_op) - stepnet.add_op(h_fc_op) - stepnet.add_op(sum_op) - stepnet.add_op(sig_op) - stepnet.complete_add_op(True) - + input_dim = 30 + batch_size = 50 + weight_dim = 15 + sent_len = 11 + + def setUp(self): + self.py_rnn = PySimpleRNN(self.input_dim, self.batch_size, + self.weight_dim, self.sent_len) + + def forward(self): + self.scope = core.Scope() + self.create_global_variables() + self.create_rnn_op() + self.create_step_net() + ctx = core.DeviceContext.create(core.CPUPlace()) + self.rnnop.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): + # create inlink + x_np_data = self.py_rnn.x + create_tensor(self.scope, "x", + [self.sent_len, self.batch_size, self.input_dim], + x_np_data) + W_np_data = self.py_rnn.W + create_tensor(self.scope, "W", [self.input_dim, self.input_dim], + W_np_data) + + U_np_data = self.py_rnn.U + create_tensor(self.scope, "U", [self.input_dim, self.input_dim], + U_np_data) + + h_boot_np_data = self.py_rnn.h_boot + create_tensor(self.scope, "h_boot", [self.batch_size, self.input_dim], + h_boot_np_data) + self.scope.new_var("step_scopes") + self.scope.new_var("h@alias") + self.scope.new_var("h") + + def create_rnn_op(self): # create RNNOp - rnnop = ops.recurrent_op( + self.rnnop = RecurrentOp( # inputs inlinks=["x"], boot_memories=["h_boot"], step_net="stepnet", # outputs - outlinks=[output], + outlinks=["h"], step_scopes="step_scopes", # attributes inlink_alias=["x@alias"], - outlink_alias=[output_alias], - pre_memories=[prememory], - memories=[memory]) - - ctx = core.DeviceContext.cpu_context() - rnnop.infer_shape(self.scope) - rnnop.run(self.scope, ctx) - - def test_recurrent(self): - self.init() + outlink_alias=["h@alias"], + pre_memories=["h@pre"], + memories=["h@alias"]) + + def create_step_net(self): + 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") + 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.complete_add_op(True) + self.rnnop.set_stepnet(stepnet) + + def test_forward(self): + print 'test recurrent op forward' + pd_output = self.forward() + py_output = self.py_rnn.forward() + print 'pd_output', pd_output + print + print 'py_output', py_output + self.assertEqual(pd_output.shape, py_output.shape) if __name__ == '__main__': 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 04abc14ee198fe4e2307e009c696a2b40ec271b6..f8521eb517057fbeb104b28af7da4fffe54f37de 100644 --- a/python/paddle/v2/framework/tests/test_rowwise_add_op.py +++ b/python/paddle/v2/framework/tests/test_rowwise_add_op.py @@ -8,9 +8,11 @@ class TestRowwiseAddOp(unittest.TestCase): def setUp(self): self.type = "rowwise_add" - self.X = np.random.random((32, 84)).astype("float32") - self.b = np.random.random(84).astype("float32") - self.Out = np.add(self.X, self.b) + self.inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'b': np.random.random(84).astype("float32") + } + self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['b'])} if __name__ == '__main__': diff --git a/python/paddle/v2/framework/tests/test_sgd_op.py b/python/paddle/v2/framework/tests/test_sgd_op.py index ca03cc11abe2ceb31b33a87797aa752943dd2a7d..e5f9ef865e84f1a78e28884ad7e2e758f9ca8054 100644 --- a/python/paddle/v2/framework/tests/test_sgd_op.py +++ b/python/paddle/v2/framework/tests/test_sgd_op.py @@ -8,10 +8,13 @@ class TestSGD(unittest.TestCase): def setUp(self): self.type = "sgd" - self.param = numpy.random.random((102, 105)).astype("float32") - self.grad = numpy.random.random((102, 105)).astype("float32") - self.learning_rate = 0.1 - self.param_out = self.param - self.learning_rate * self.grad + w = numpy.random.random((102, 105)).astype("float32") + g = numpy.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} if __name__ == "__main__": diff --git a/python/paddle/v2/framework/tests/test_sigmoid_op.py b/python/paddle/v2/framework/tests/test_sigmoid_op.py index 50044a122f1d66dd54a24f6cce76074a60ee2262..2a57a41ed8b718fd420062ba68e853a4861b7359 100644 --- a/python/paddle/v2/framework/tests/test_sigmoid_op.py +++ b/python/paddle/v2/framework/tests/test_sigmoid_op.py @@ -8,9 +8,12 @@ class TestSigmoidOp(unittest.TestCase): def setUp(self): self.type = "sigmoid" - self.X = np.random.random((32, 100)).astype("float32") - self.Y = 1 / (1 + np.exp(-self.X)) + self.inputs = {'X': np.random.random((32, 100)).astype("float32")} + self.outputs = {'Y': 1 / (1 + np.exp(-self.inputs['X']))} +#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 c80888128781d98e4ed30d845a30b39121f66459..e670d93653e07d35e5019c9daac45c214eddf367 100644 --- a/python/paddle/v2/framework/tests/test_softmax_op.py +++ b/python/paddle/v2/framework/tests/test_softmax_op.py @@ -1,9 +1,8 @@ import unittest import numpy as np -import paddle.v2.framework.core as core -import paddle.v2.framework.create_op_creation_methods as creation +from gradient_checker import GradientChecker, create_op from op_test_util import OpTestMeta @@ -19,66 +18,17 @@ class TestSoftmaxOp(unittest.TestCase): def setUp(self): self.type = "softmax" - self.X = np.random.random((32, 100)).astype("float32") - self.Y = np.apply_along_axis(stable_softmax, 1, self.X) + self.inputs = {'X': np.random.random((32, 100)).astype("float32")} + self.outputs = { + 'Y': np.apply_along_axis(stable_softmax, 1, self.inputs['X']) + } -class TestSoftmaxGradOp(unittest.TestCase): - def test_softmax_grad(self): - op = creation.op_creations.softmax(X="X", Y="Y") - backward_op = core.Operator.backward(op, set()) - self.assertEqual(backward_op.type(), "softmax_grad") - expected = '''Op(softmax_grad), inputs:(X, Y, Y@GRAD), outputs:(X@GRAD).''' - self.assertEqual(expected, str(backward_op)) - - batch_size = 3 - class_num = 5 - # Initialize X and add 1e-2 for numerical stability - Y = np.random.rand(batch_size, class_num).astype(np.float32) - Y = Y + 1e-2 - dY = np.random.rand(batch_size, class_num).astype(np.float32) - - # Reference implementation of cross entropy with soft labels - def label_softmax_grad(Y, dY): - dX = Y * 0.0 - for i in range(batch_size): - d = np.dot(Y[i, :], dY[i, :]) - dX[i, :] = Y[i, :] * (dY[i, :] - d) - return dX - - expected = label_softmax_grad(Y, dY) - - scope = core.Scope() - places = [] - places.append(core.CPUPlace()) - if core.is_compile_gpu(): - places.append(core.GPUPlace(0)) - - for place in places: - y = scope.new_var("Y") - y_tensor = y.get_tensor() - y_tensor.set_dims([batch_size, class_num]) - y_tensor.alloc_float(place) - y_tensor.set(Y, place) - - dy = scope.new_var("Y@GRAD") - dy_tensor = dy.get_tensor() - dy_tensor.set_dims([batch_size, class_num]) - dy_tensor.alloc_float(place) - dy_tensor.set(dY, place) - - x = scope.new_var("X") - dx = scope.new_var("X@GRAD") - - tensor = scope.find_var("X@GRAD").get_tensor() - backward_op.infer_shape(scope) - self.assertEqual([batch_size, class_num], tensor.shape()) - - ctx = core.DeviceContext.create(place) - backward_op.run(scope, ctx) - actual = np.array(tensor) - - np.testing.assert_almost_equal(actual, expected, decimal=3) +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") if __name__ == '__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 new file mode 100644 index 0000000000000000000000000000000000000000..c3d2bb44da3977c0899b2609a8efe15b7e1789f2 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_uniform_random_op.py @@ -0,0 +1,35 @@ +import unittest +from paddle.v2.framework.op import Operator +import paddle.v2.framework.core as core +import numpy + + +class UniformRandomTest(unittest.TestCase): + def test_uniform_random_cpu(self): + self.uniform_random_test(place=core.CPUPlace()) + + def test_uniform_random_gpu(self): + if core.is_compile_gpu(): + self.uniform_random_test(place=core.GPUPlace(0)) + + def uniform_random_test(self, place): + scope = core.Scope() + scope.new_var("X").get_tensor() + + op = Operator( + "uniform_random", + Out="X", + dims=[1000, 784], + min=-5.0, + max=10.0, + seed=10) + + op.infer_shape(scope) + ctx = core.DeviceContext.create(place) + op.run(scope, ctx) + tensor = numpy.array(scope.find_var("X").get_tensor()) + self.assertAlmostEqual(tensor.mean(), 2.5, delta=0.1) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/optimizer.py b/python/paddle/v2/optimizer.py index ba581980334fec6226a537af2cf53b3465d32c1e..29f0945eb4c88eab8fa9ee83f455190dfd473aa4 100644 --- a/python/paddle/v2/optimizer.py +++ b/python/paddle/v2/optimizer.py @@ -1,13 +1,26 @@ -import paddle.trainer_config_helpers.config_parser_utils as config_parser_utils -import paddle.trainer_config_helpers.optimizers as v1_optimizers +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. """ Optimizers(update equation) for SGD method. -TODO(zhihong) : create new optimizer with proto config, add new optimizer here - TODO(yuyang18): Complete comments. """ +import paddle.trainer_config_helpers.config_parser_utils as config_parser_utils +import paddle.trainer_config_helpers.optimizers as v1_optimizers +from paddle.proto.OptimizerConfig_pb2 import OptimizerConfig + __all__ = [ 'Momentum', 'Adam', 'Adamax', 'AdaGrad', 'DecayedAdaGrad', 'AdaDelta', 'RMSProp', 'ModelAverage', 'L2Regularization' @@ -70,7 +83,8 @@ class Optimizer(object): gradient_machine.prefetch(in_args) parameter_updater.getParametersRemote() - :param pserver_spec: pserver location, eg: localhost:3000 + :param pserver_spec: pserver location, eg: localhost:3000, if use etcd, + pserver_spec should be the etcd endpoints, eg: http://localhost:2379 :return: parameter_updater """ if is_local: diff --git a/python/paddle/v2/parameters.py b/python/paddle/v2/parameters.py index a9cba8ca0b1efd4149463f6c7bf2dcdfbea350c9..b8af5abaeada49a3e8951c21c9065aaf4d1ab851 100644 --- a/python/paddle/v2/parameters.py +++ b/python/paddle/v2/parameters.py @@ -1,3 +1,17 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + import numpy as np from paddle.proto.ParameterConfig_pb2 import ParameterConfig import paddle.trainer.config_parser as cp @@ -113,16 +127,7 @@ class Parameters(object): """ return iter(self.__param_conf__) - def __getitem__(self, key): - """ - Get parameter by parameter name. It uses Python dict syntax. - - :note: It will always copy the parameter from C++ side. - :param key: Parameter name - :type key: basestring - :return: parameter value - :rtype: np.ndarray - """ + def __getter_inner(self, key, param_type): import py_paddle.swig_paddle as api shape = self.get_shape(key) @@ -138,7 +143,7 @@ class Parameters(object): each_gradient_machine, key) # for simplify implementation now, we always copy from C++ assert isinstance(param, api.Parameter) - val = param.getBuf(api.PARAMETER_VALUE) + val = param.getBuf(param_type) assert isinstance(val, api.Vector) val = val.copyToNumpyArray() return val @@ -146,6 +151,19 @@ class Parameters(object): raise RuntimeError("Unexpected branch") + def __getitem__(self, key): + """ + Get parameter by parameter name. It uses Python dict syntax. + + :note: It will always copy the parameter from C++ side. + :param key: Parameter name + :type key: basestring + :return: parameter value + :rtype: np.ndarray + """ + import py_paddle.swig_paddle as api + return self.__getter_inner(key, api.PARAMETER_VALUE) + def get_shape(self, key): """ get shape of the parameter. @@ -202,6 +220,19 @@ class Parameters(object): """ return self.__getitem__(key=parameter_name) + def get_grad(self, key): + """ + Get grandient by parameter name. + + :note: It will always copy the parameter from C++ side. + :param key: parameter name + :type key: basestring + :return: The grandient matrix. + :rtype: np.ndarray + """ + import py_paddle.swig_paddle as api + return self.__getter_inner(key, api.PARAMETER_GRADIENT) + def set(self, parameter_name, value): """ Set parameter by parameter name & matrix. @@ -250,7 +281,13 @@ class Parameters(object): size = reduce(lambda a, b: a * b, param.shape) f.write(struct.pack("IIQ", 0, 4, size)) param = param.astype(np.float32) - f.write(param.tostring()) + s = param.tostring() + wrote_size = 0 + buf = buffer(s, wrote_size, 65535) + while buf: # f.write crashes with big data blog. + f.write(buf) + wrote_size += 65535 + buf = buffer(s, wrote_size, 65535) def deserialize(self, name, f): """ diff --git a/python/paddle/v2/plot/tests/CMakeLists.txt b/python/paddle/v2/plot/tests/CMakeLists.txt index da5cd764889b48a3af8461a2793d948aa609d6c1..4b6c1c80969182ccf6e0189b18bade8758bbbc30 100644 --- a/python/paddle/v2/plot/tests/CMakeLists.txt +++ b/python/paddle/v2/plot/tests/CMakeLists.txt @@ -1,5 +1,5 @@ if (NOT APPLE) # The Mac OS X backend will not be able to function correctly if Python is # not installed as a framework. - add_python_test(test_ploter test_ploter.py) + py_test(test_ploter SRCS test_ploter.py) endif() 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/CMakeLists.txt b/python/paddle/v2/reader/tests/CMakeLists.txt index 6a1d337b232c7a849a8793894bf16d26d609d3dd..107d5912e1567e0c8721987a281272c7feb51e63 100644 --- a/python/paddle/v2/reader/tests/CMakeLists.txt +++ b/python/paddle/v2/reader/tests/CMakeLists.txt @@ -1 +1,2 @@ -add_python_test(reader_tests creator_test.py decorator_test.py) +py_test(creator_test SRCS creator_test.py) +py_test(decorator_test SRCS decorator_test.py) 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/CMakeLists.txt b/python/paddle/v2/tests/CMakeLists.txt index 058f22befd0657d06ff130ace55fe7322148213d..b7791559594321a85f41b508b69efeb077d69595 100644 --- a/python/paddle/v2/tests/CMakeLists.txt +++ b/python/paddle/v2/tests/CMakeLists.txt @@ -1,2 +1,7 @@ -add_python_test(test_v2_api test_data_feeder.py test_op.py test_parameters.py -test_layer.py test_rnn_layer.py test_topology.py test_image.py) +py_test(test_op SRCS test_op.py) +py_test(test_image SRCS test_image.py) +py_test(test_layer SRCS test_layer.py) +py_test(test_topology SRCS test_topology.py) +py_test(test_rnn_layer SRCS test_rnn_layer.py) +py_test(test_parameters SRCS test_parameters.py) +py_test(test_data_feeder SRCS test_data_feeder.py) diff --git a/python/paddle/v2/trainer.py b/python/paddle/v2/trainer.py index 76bae0bb12b6c33f88530386f9cc19ae9b59f457..9c4dd5f25083d210bbd218a85d8dbb3cce2c3d0e 100644 --- a/python/paddle/v2/trainer.py +++ b/python/paddle/v2/trainer.py @@ -161,14 +161,14 @@ class SGD(object): self.__parameter_updater__.update(each_param) cost_sum = out_args.sum() cost = cost_sum / len(data_batch) - self.__parameter_updater__.finishBatch(cost) - batch_evaluator.finish() event_handler( v2_event.EndIteration( pass_id=pass_id, batch_id=batch_id, cost=cost, evaluator=batch_evaluator)) + self.__parameter_updater__.finishBatch(cost) + batch_evaluator.finish() self.__parameter_updater__.finishPass() pass_evaluator.finish() diff --git a/python/requirements.txt b/python/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..e19453c25da1ec78773c00a72b8e517b0d798fff --- /dev/null +++ b/python/requirements.txt @@ -0,0 +1,9 @@ +requests==2.9.2 +numpy>=1.12 +protobuf==3.1 +recordio>=0.1.0 +matplotlib +rarfile +scipy>=0.19.0 +Pillow +nltk>=3.2.2 diff --git a/python/setup.py.in b/python/setup.py.in index 65a26940d4d703ea4fbb5022523a90716982ec10..287442e013f91df1eed9c629b7767a660d5e30d7 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -1,4 +1,7 @@ -from setuptools import setup +from setuptools import setup, Distribution +class BinaryDistribution(Distribution): + def has_ext_modules(foo): + return True packages=['paddle', 'paddle.proto', @@ -11,33 +14,44 @@ packages=['paddle', 'paddle.v2.master', 'paddle.v2.plot', 'paddle.v2.framework', - 'paddle.v2.framework.proto'] + 'paddle.v2.framework.proto', + 'py_paddle'] -setup_requires=["requests", - "numpy", - "protobuf==3.1", - "recordio", - "matplotlib", - "rarfile", - "scipy>=0.19.0", - "Pillow", - "nltk"] +with open('@PADDLE_SOURCE_DIR@/python/requirements.txt') as f: + setup_requires = f.read().splitlines() if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']: setup_requires+=["opencv-python"] -setup(name='paddle', +# the prefix is sys.prefix which should always be usr +paddle_bin_dir = 'local/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_rt_lib_dir = 'local/lib' +paddle_rt_libs = [] if '${MKL_SHARED_LIBS}'== '' else '${MKL_SHARED_LIBS}'.split(';') + +setup(name='paddlepaddle', version='${PADDLE_VERSION}', description='Parallel Distributed Deep Learning', install_requires=setup_requires, packages=packages, - package_data={'paddle.v2.master': ['libpaddle_master.so'], - 'paddle.v2.framework': ['core.so'] + package_data={ + 'paddle.v2.master': ['libpaddle_master.so'], + 'paddle.v2.framework': ['core.so'], + 'py_paddle':['*.py','_swig_paddle.so'] }, package_dir={ '': '${CMAKE_CURRENT_SOURCE_DIR}', # The paddle.v2.framework.proto will be generated while compiling. # So that package points to other directory. - 'paddle.v2.framework.proto': '${PROJ_BINARY_ROOT}/paddle/framework' + 'paddle.v2.framework.proto': '${PADDLE_BINARY_DIR}/paddle/framework', + 'py_paddle': '${PADDLE_SOURCE_DIR}/paddle/py_paddle' }, + scripts=['${PADDLE_BINARY_DIR}/paddle/scripts/paddle'], + distclass=BinaryDistribution, + data_files=[(paddle_bin_dir, paddle_bins), + (paddle_rt_lib_dir, paddle_rt_libs)] )