diff --git a/.clang_format.hook b/.clang_format.hook new file mode 100755 index 0000000000..1d92821686 --- /dev/null +++ b/.clang_format.hook @@ -0,0 +1,15 @@ +#!/bin/bash +set -e + +readonly VERSION="3.8" + +version=$(clang-format -version) + +if ! [[ $version == *"$VERSION"* ]]; then + echo "clang-format version check failed." + echo "a version contains '$VERSION' is needed, but get '$version'" + echo "you can install the right version, and make an soft-link to '\$PATH' env" + exit -1 +fi + +clang-format $@ diff --git a/.gitignore b/.gitignore index c84b2fc8c7..9622ab78e0 100644 --- a/.gitignore +++ b/.gitignore @@ -24,4 +24,5 @@ cmake-build-* python/paddle/v2/framework/core.so CMakeFiles cmake_install.cmake - +paddle/.timestamp +python/paddlepaddle.egg-info/ diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index bb8c88787d..83fe9af768 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -19,10 +19,10 @@ - id: end-of-file-fixer - repo: local hooks: - - id: clang-format + - id: clang-format-with-version-check name: clang-format description: Format files with ClangFormat. - entry: clang-format -i + entry: bash ./.clang_format.hook -i language: system files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$ - repo: https://github.com/PaddlePaddle/pre-commit-golang diff --git a/.travis.yml b/.travis.yml index 8c8c6699d3..b4b83fcdbc 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 nltk==3.2.2 scipy==0.19.0 recordio matplotlib Pillow + - 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 b174831109..ad559672ad 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 ${CMAKE_CURRENT_BINARY_DIR}) include(system) @@ -55,6 +55,7 @@ option(WITH_C_API "Compile PaddlePaddle with C-API(Prediction)" OFF) option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF) option(GLIDE_INSTALL "Download and install go dependencies " ON) option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF) +option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) # CMAKE_BUILD_TYPE if(NOT CMAKE_BUILD_TYPE) @@ -121,8 +122,8 @@ include(version) # set PADDLE_VERSION include(coveralls) # set code coverage -include_directories("${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}) @@ -137,14 +138,14 @@ set(EXTERNAL_LIBS ) if(WITH_GPU) - list(APPEND EXTERNAL_LIB ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY}) + list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY}) if(NOT WITH_DSO) - list(APPEND EXTERNAL_LIB ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY}) + list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY}) endif(NOT WITH_DSO) endif(WITH_GPU) if(WITH_MKLDNN) - list(APPEND EXTERNAL_LIBS ${MKLDNN_LIBRARY} ${MKLDNN_IOMP_LIB}) + list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB} ${MKLDNN_IOMP_LIB}) endif() if(USE_NNPACK) @@ -164,10 +165,12 @@ if(WITH_GOLANG) add_subdirectory(go) endif(WITH_GOLANG) +set(PADDLE_PYTHON_BUILD_DIR "${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 8ac123bf9c..136db772cc 100644 --- a/Dockerfile +++ b/Dockerfile @@ -10,13 +10,11 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub ARG WITH_GPU ARG WITH_AVX ARG WITH_DOC -ARG WITH_STYLE_CHECK ENV WOBOQ OFF -ENV WITH_GPU=${WITH_GPU:-OFF} +ENV WITH_GPU=${WITH_GPU:-ON} ENV WITH_AVX=${WITH_AVX:-ON} ENV WITH_DOC=${WITH_DOC:-OFF} -ENV WITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} ENV HOME /root # Add bash enhancements @@ -34,9 +32,6 @@ RUN apt-get update && \ net-tools && \ apt-get clean -y -# paddle is using numpy.flip, which is introduced since 1.12.0 -RUN pip --no-cache-dir install 'numpy>=1.12.0' - # Install Go and glide RUN wget -qO- https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \ tar -xz -C /usr/local && \ @@ -58,19 +53,23 @@ 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] + # 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 2ac0989546..51c3b918cc 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -28,6 +28,10 @@ if(NOT WITH_TIMER) add_definitions(-DPADDLE_DISABLE_TIMER) endif(NOT WITH_TIMER) +if(USE_EIGEN_FOR_BLAS) + add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS) +endif(USE_EIGEN_FOR_BLAS) + if(NOT WITH_PROFILER) add_definitions(-DPADDLE_DISABLE_PROFILER) endif(NOT WITH_PROFILER) @@ -129,7 +133,7 @@ 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 5184f0815f..8d5d533126 100644 --- a/cmake/cpplint.cmake +++ b/cmake/cpplint.cmake @@ -52,7 +52,7 @@ macro(add_style_check_target TARGET_NAME) if(SOURCES_LIST) add_custom_command(TARGET ${TARGET_NAME} POST_BUILD - COMMAND "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py" + COMMAND "${PYTHON_EXECUTABLE}" "${PADDLE_SOURCE_DIR}/paddle/scripts/cpplint.py" "--filter=${STYLE_FILTER}" ${SOURCES_LIST} COMMENT "cpplint: Checking source code style" diff --git a/cmake/cudnn.cmake b/cmake/cudnn.cmake index 69f40df516..2c84061ff5 100644 --- a/cmake/cudnn.cmake +++ b/cmake/cudnn.cmake @@ -2,7 +2,7 @@ if(NOT WITH_GPU) return() endif() -set(CUDNN_ROOT "" CACHE PATH "CUDNN ROOT") +set(CUDNN_ROOT "/usr" CACHE PATH "CUDNN ROOT") find_path(CUDNN_INCLUDE_DIR cudnn.h PATHS ${CUDNN_ROOT} ${CUDNN_ROOT}/include $ENV{CUDNN_ROOT} $ENV{CUDNN_ROOT}/include ${CUDA_TOOLKIT_INCLUDE} diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake index 25c6b4ef52..9686df0021 100644 --- a/cmake/external/mkldnn.cmake +++ b/cmake/external/mkldnn.cmake @@ -51,7 +51,7 @@ ExternalProject_Add( ${EXTERNAL_PROJECT_LOG_ARGS} DEPENDS ${MKLDNN_DEPENDS} GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git" - GIT_TAG "v0.9" + GIT_TAG "v0.10" PREFIX ${MKLDNN_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR} diff --git a/cmake/external/mklml.cmake b/cmake/external/mklml.cmake index e9fd3d4bed..51fafb9479 100644 --- a/cmake/external/mklml.cmake +++ b/cmake/external/mklml.cmake @@ -28,7 +28,7 @@ INCLUDE(ExternalProject) SET(MKLML_PROJECT "extern_mklml") SET(MKLML_VER "mklml_lnx_2018.0.20170720") -SET(MKLML_URL "https://github.com/01org/mkl-dnn/releases/download/v0.9/${MKLML_VER}.tgz") +SET(MKLML_URL "https://github.com/01org/mkl-dnn/releases/download/v0.10/${MKLML_VER}.tgz") SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml") SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}") SET(MKLML_DST_DIR "mklml") diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index db09232c0e..0eeccbf7d8 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -73,10 +73,18 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR}) # linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas) SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c) FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";") -ADD_LIBRARY(cblas STATIC ${dummyfile}) +IF(${CBLAS_PROVIDER} MATCHES MKL) + ADD_LIBRARY(cblas SHARED ${dummyfile}) +ELSE() + ADD_LIBRARY(cblas STATIC ${dummyfile}) +ENDIF() TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES}) IF(NOT ${CBLAS_FOUND}) ADD_DEPENDENCIES(cblas extern_openblas) LIST(APPEND external_project_dependencies cblas) +ELSE() + IF("${CBLAS_PROVIDER}" STREQUAL "MKLML") + ADD_DEPENDENCIES(cblas mklml) + ENDIF() ENDIF(NOT ${CBLAS_FOUND}) diff --git a/cmake/flags.cmake b/cmake/flags.cmake index e26d8d9df3..ff246b2eb4 100644 --- a/cmake/flags.cmake +++ b/cmake/flags.cmake @@ -9,11 +9,6 @@ function(CheckCompilerCXX11Flag) if(${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) - endif() elseif(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang") # cmake >= 3.0 compiler id "AppleClang" on Mac OS X, otherwise "Clang" # Apple Clang is a different compiler than upstream Clang which havs different version numbers. @@ -158,7 +153,7 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF) # Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc. # So, don't set these flags here. -LIST(APPEND CUDA_NVCC_FLAGS -std=c++11 --default-stream per-thread) +LIST(APPEND CUDA_NVCC_FLAGS -std=c++11) LIST(APPEND CUDA_NVCC_FLAGS --use_fast_math) if(CMAKE_BUILD_TYPE STREQUAL "Debug") diff --git a/cmake/generic.cmake b/cmake/generic.cmake index 957c20bcf6..d2aab938d4 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -411,7 +411,7 @@ function(py_test TARGET_NAME) set(multiValueArgs SRCS DEPS) cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) add_test(NAME ${TARGET_NAME} - COMMAND env PYTHONPATH=${PADDLE_PYTHON_PACKAGE_DIR} + COMMAND env PYTHONPATH=${PADDLE_PYTHON_BUILD_DIR}/lib-python python2 ${py_test_SRCS} WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) endif() diff --git a/cmake/package.cmake b/cmake/package.cmake index ff49a2d08e..79e02147f3 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 4a27623b7f..0da4969d31 100644 --- a/cmake/util.cmake +++ b/cmake/util.cmake @@ -141,8 +141,8 @@ 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} ${PADDLE_SOURCE_DIR}/cmake/make_resource.py) endfunction() diff --git a/cmake/version.cmake b/cmake/version.cmake index ac1583a24c..cde650128a 100644 --- a/cmake/version.cmake +++ b/cmake/version.cmake @@ -4,7 +4,7 @@ set(tmp_version "HEAD") while ("${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/about/index_cn.md b/doc/about/index_cn.md deleted file mode 100644 index 3bf030004d..0000000000 --- a/doc/about/index_cn.md +++ /dev/null @@ -1,11 +0,0 @@ -关于PaddlePaddle -================ - -PaddlePaddle是一个最早由百度科学家和工程师共同研发的并行分布式深度学习平台，兼备易用性、高效性、灵活性和可扩展性，目前已被百度内部多个产品线广泛使用。 -PaddlePaddle目前已经开放源码, 但是远未完善，我们希望能在这个基础上不断的改进、扩展和延伸。 -同时我们希望广大开发者积极提供反馈和贡献源代码，建立一个活跃的开源社区。 - -致谢 --------- - -在此，特别感谢PaddlePaddle的[所有贡献者](https://github.com/PaddlePaddle/Paddle/graphs/contributors)。 diff --git a/doc/about/index_en.rst b/doc/about/index_en.rst deleted file mode 100644 index 065c430cde..0000000000 --- a/doc/about/index_en.rst +++ /dev/null @@ -1,14 +0,0 @@ -ABOUT -======= - -PaddlPaddle is an easy-to-use, efficient, flexible and scalable deep learning platform, -which is originally developed by Baidu scientists and engineers for the purpose of applying deep learning to many products at Baidu. - -PaddlePaddle is now open source but far from complete, which is intended to be built upon, improved, scaled, and extended. -We hope to build an active open source community both by providing feedback and by actively contributing to the source code. - - -Credits --------- - -We owe many thanks to `all contributors and developers `_ of PaddlePaddle! diff --git a/doc/api/v2/config/layer.rst b/doc/api/v2/config/layer.rst index cb330ea5e1..1329b77bb4 100644 --- a/doc/api/v2/config/layer.rst +++ b/doc/api/v2/config/layer.rst @@ -257,6 +257,11 @@ seq_concat .. autoclass:: paddle.v2.layer.seq_concat :noindex: +seq_slice +--------- +.. autoclass:: paddle.v2.layer.seq_slice + :noindex: + kmax_sequence_score ------------------- .. autoclass:: paddle.v2.layer.kmax_sequence_score @@ -362,6 +367,11 @@ trans .. autoclass:: paddle.v2.layer.trans :noindex: +scale_shift +----------- +.. autoclass:: paddle.v2.layer.scale_shift + :noindex: + Sampling Layers =============== @@ -409,9 +419,14 @@ multi_binary_label_cross_entropy_cost .. autoclass:: paddle.v2.layer.multi_binary_label_cross_entropy_cost :noindex: -huber_cost ----------- -.. autoclass:: paddle.v2.layer.huber_cost +huber_regression_cost +------------------------- +.. autoclass:: paddle.v2.layer.huber_regression_cost + :noindex: + +huber_classification_cost +------------------------- +.. autoclass:: paddle.v2.layer.huber_classification_cost :noindex: lambda_cost diff --git a/doc/design/auto_gradient_check.md b/doc/design/auto_gradient_check.md new file mode 100644 index 0000000000..1f4d4ec16f --- /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/README.md b/doc/design/cluster_train/README.md index 74961f8005..177a5f5d54 100644 --- a/doc/design/cluster_train/README.md +++ b/doc/design/cluster_train/README.md @@ -54,17 +54,18 @@ The life cycle of a single task is illustrated below: 1. When a new pass of training starts, all tasks will be placed in the todo queue. -1. The master server will dispatch few tasks to each trainer at a time, puts them in the pending queue and waits for completion. -1. The trainer will work on its tasks and tell the master server once a task is completed. The master server will dispatch a new task to that trainer. -1. If a task timeout. the master server will move it back to the todo queue. The timeout count will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, so it will be discarded. +1. Upon trainer requests for new task, the master server will dispatch a task from todo queue to it, put the task in the pending queue and wait for completion. +1. The trainer will work on its task and tell the master server once the task is completed and ask for new task. The master server will dispatch a new task to that trainer. +1. If a task fails for any reason in trainer, or takes longer than a specific period of time, the master server will move the task back to the todo queue. The timeout count for that task will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, then it will be discarded. 1. The master server will move completed task to the done queue. When the todo queue is empty, the master server will start a new pass by moving all tasks in the done queue to todo queue and reset the timeout counter of all tasks to zero. ### Trainer Process The trainer process will: -- Receive tasks from the master. -- Work on the tasks: calculate and upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers. +- Request tasks from the master. +- Work on the tasks +- Upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers. ### Parameter Server Process @@ -119,8 +120,8 @@ When the master is started by the Kubernetes, it executes the following steps at 1. Grabs a unique *master* lock in etcd, which prevents concurrent master instantiations. 1. Recovers the task queues from etcd if they already exist, otherwise, the master will create them. -1. Watches the trainer prefix keys `/trainer/` on etcd to find the live trainers. -1. Starts dispatching the tasks to the trainers, and updates task queue using an etcd transaction to ensure lock is held during the update. +1. Write its ip address to */master/addr* so that trainers can discover it. +1. Listens to trainers' request of task, dispatch one upon request, and updates task queue using an etcd transaction to ensure lock is held during the update. When the master server process is dead for any reason, Kubernetes will restart it. It will be online again with all states recovered from etcd in few minutes. @@ -128,13 +129,11 @@ When the master server process is dead for any reason, Kubernetes will restart i When the trainer is started by the Kubernetes, it executes the following steps at startup: -1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count. -1. Generates a unique ID, and sets key `/trainer/` with its contact address as value. The key will be deleted when the lease expires, so the master will be aware of the trainer being online and offline. -1. Waits for tasks from the master to start training. +1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count */ps_desired*. +1. Finds and watches */master/addr* to get master's address. +1. Requests for tasks from the master to start training. -If trainer's etcd lease expires, it will try set key `/trainer/` again so that the master server can discover the trainer again. - -When a trainer fails, Kuberentes would try to restart it. The recovered trainer would fetch tasks from the TODO queue and go on training. +When a trainer fails, Kuberentes would try to restart it. The recovered trainer would fetch tasks from master and go on training. ### Parameter Server Process diff --git a/doc/design/cluster_train/large_model_dist_train.md b/doc/design/cluster_train/large_model_dist_train.md new file mode 100644 index 0000000000..0c4b5bc24c --- /dev/null +++ b/doc/design/cluster_train/large_model_dist_train.md @@ -0,0 +1,101 @@ +# Alalysis of large model distributed training in Paddle + +***NOTE: This is only some note for how we implemeted this scheme in V1, not a new design.*** + +## What is it + +We often encounter cases that the embedding layer parameters(sparse) are so large that we can not store it in the trainer's memory when training. So we need to put them to several servers, and fetch them row by row instead of fetch all of the parameters. + +## How to use + +Specify command-line argument like `--loadsave_parameters_in_pserver=true --ports_num_for_sparse=1 --use_old_updater=1` when starting the paddle trainer. And also add something like `--ports_num_for_sparse=1 --pserver_num_threads=5` when starting pserver processes. + +Accrodingly, configure your embedding layers like: + +```python +SPARSE_REMOTE=True + +w1 = data_layer(name="w1", size=dict_size) +emb1 = embedding_layer(input=w1, size=32, param_attr=ParameterAttribute(sparse_update=SPARSE_REMOTE)) +w2 = data_layer(name="w2", size=dict_size) +emb2 = embedding_layer(input=w2, size=32, param_attr=ParameterAttribute(sparse_update=SPARSE_REMOTE)) +... +``` + +## Implementation details + +```c++ +enum MatType { + MAT_NORMAL, + MAT_NORMAL_SHARED, + MAT_VALUE_SHARED, + MAT_SPARSE_ROW_IDS, + MAT_SPARSE_ROW_AUTO_GROW, + MAT_CACHE_ROW, + MAT_SPARSE_ROW, + MAT_SPARSE_ROW_PREFETCH, + MAT_SPARSE_ROW_PREFETCH_FULL_SIZE, +}; +``` + +`MAT_SPARSE_ROW_PREFETCH` is what we use when configured to fetch only row of matrix when training. + +In `trainer_internal.cpp:L93 trainOneBatch`: + +```c++ + if (config_->getOptConfig().use_sparse_remote_updater()) { + REGISTER_TIMER("prefetch"); + gradientMachine_->prefetch(inArgs); + parameterUpdater_->getParametersRemote(); + } +``` + +When doing actual network forward and backward, at the beginning of each batch, the trainer will try to download one row of data from pserver. + +In `trainer/RemoteParameterUpdater.cpp`: `parameterUpdater_->getParametersRemote();`: + +```c++ +if (fullSize) { + ... +} else { +getParams = [&] { + parameterClient_->getParameterSparse( + /* recvParameterType= */ PARAMETER_VALUE, sendBackParameterType); +}; +applyL1 = [](Parameter& para, real decayRate) { + para.getMat(PARAMETER_VALUE)->applyL1(/*lr=*/1.0f, decayRate); +}; +} +``` + +Calling `parameterClient_->getParameterSparse` will do remote call to pserver's `getParameterSparse`: + +```c++ +void ParameterServer2::getParameterSparse(const SendParameterRequest& request, + std::vector& inputBuffers, + SendParameterResponse* response, + std::vector* outputBuffers) { + (void)inputBuffers; + auto& buffer = *readWriteBuffer_; + size_t numReals = 0; + for (const auto& block : request.blocks()) { + numReals += getParameterConfig(block).dims(1); + } + buffer.resize(numReals); + + VLOG(3) << "pserver: getParameterSparse, numReals=" << numReals; + + ReadLockGuard guard(parameterMutex_); + size_t offset = 0; + for (const auto& block : request.blocks()) { + size_t width = getParameterConfig(block).dims(1); + Buffer buf = {buffer.data() + offset, width}; + int type = request.send_back_parameter_type(); + sendBackParameterSparse(block, type, response, &buf, width, outputBuffers); + offset += width; + } +} +``` + +`getParameterConfig(block).dims(1)` returns the width of the current "parameter block"(a shard of parameter object), +then `getParameterSparse` remote call returns only one row of data to the client. diff --git a/doc/design/cluster_train/src/paddle-etcd.graffle b/doc/design/cluster_train/src/paddle-etcd.graffle index 56681ae5bb..f973dc9b9d 100644 Binary files a/doc/design/cluster_train/src/paddle-etcd.graffle and b/doc/design/cluster_train/src/paddle-etcd.graffle differ diff --git a/doc/design/cluster_train/src/paddle-etcd.png b/doc/design/cluster_train/src/paddle-etcd.png index 4f9c9762b3..57981ceb4b 100644 Binary files a/doc/design/cluster_train/src/paddle-etcd.png and b/doc/design/cluster_train/src/paddle-etcd.png differ diff --git a/doc/design/mkldnn/README.MD b/doc/design/mkldnn/README.MD index e956994431..fe8da907d9 100644 --- a/doc/design/mkldnn/README.MD +++ b/doc/design/mkldnn/README.MD @@ -101,6 +101,7 @@ if use_mkldnn 5. 在**Argument**里添加两个`MkldnnMatrixPtr`，取名为`mkldnnValue`和`mkldnnGrad`，用于存放`MkldnnLayer`会用到的memory buffer。 并且添加函数cvt(会修改为一个更加合适的函数名)，用于处理"CPU device"和"MKL-DNN device"之间memory的相互转化。 6. 在父类`Layer`中的`getOutput`函数中添加一段逻辑，用于判断`deviceId`，并针对device在MKL-DNN和CPU之间不统一的情况，做一个前期转换。 也就是调用`Argument`的cvt函数把output统一到需要的device上。 7. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。 +8. 关于MKLDNN参数的保存。由于MKLDNN参数的格式与PaddlePaddle原有的格式存在不一样的情况，所以需要在保存参数时同时保存该格式信息。目前准备扩展[Header](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/parameter/Parameter.h#L247)里面的`int32_t version`。这个值不管是在v1还是在v2里面，一直保存的是0，所以可以充分利用这个信息，定义一个枚举处理所有MKLDNN的参数格式，从而`MKLDNNLayer`就可以从输入的参数中获取需要的格式信息。 ## References diff --git a/doc/getstarted/build_and_install/build_from_source_en.md b/doc/getstarted/build_and_install/build_from_source_en.md index c0608ede8e..2f14614894 100644 --- a/doc/getstarted/build_and_install/build_from_source_en.md +++ b/doc/getstarted/build_and_install/build_from_source_en.md @@ -68,7 +68,7 @@ As a simple example, consider the following: 1. **BLAS Dependencies(optional)** - CMake will search BLAS libraries from system. If not found, OpenBLAS will be downloaded, built and installed automatically. + CMake will search BLAS libraries from the system. If not found, OpenBLAS will be downloaded, built and installed automatically. To utilize preinstalled BLAS， you can simply specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`. ```bash @@ -131,9 +131,9 @@ As a simple example, consider the following: To build GPU version, you will need the following installed: 1. a CUDA-capable GPU - 2. A supported version of Linux with a gcc compiler and toolchain + 2. A supported version of Linux with a GCC compiler and toolchain 3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads) - 4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn) + 4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn) The CUDA development environment relies on tight integration with the host development environment, including the host compiler and C runtime libraries, and is therefore only supported on @@ -172,6 +172,7 @@ export PATH=/bin:$PATH # install PaddlePaddle Python modules. sudo pip install /opt/paddle/share/wheels/*.whl ``` + ## Build on Centos 7 ### Install Dependencies @@ -192,9 +193,9 @@ sudo pip install /opt/paddle/share/wheels/*.whl To build GPU version, you will need the following installed: 1. a CUDA-capable GPU - 2. A supported version of Linux with a gcc compiler and toolchain + 2. A supported version of Linux with a GCC compiler and toolchain 3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads) - 4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn) + 4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn) The CUDA development environment relies on tight integration with the host development environment, including the host compiler and C runtime libraries, and is therefore only supported on @@ -222,7 +223,7 @@ mkdir build && cd build ``` Finally, you can build and install PaddlePaddle: - + ```bash # you can add build option here, such as: cmake3 .. -DCMAKE_INSTALL_PREFIX= diff --git a/doc/getstarted/build_and_install/docker_install_cn.rst b/doc/getstarted/build_and_install/docker_install_cn.rst index 02b96bb413..84e3317774 100644 --- a/doc/getstarted/build_and_install/docker_install_cn.rst +++ b/doc/getstarted/build_and_install/docker_install_cn.rst @@ -74,13 +74,13 @@ PaddlePaddle发布新版本的时候都会发布对应版本的生产镜像以 .. 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： diff --git a/doc/getstarted/build_and_install/index_cn.rst b/doc/getstarted/build_and_install/index_cn.rst index a24df6c518..dd9923697a 100644 --- a/doc/getstarted/build_and_install/index_cn.rst +++ b/doc/getstarted/build_and_install/index_cn.rst @@ -6,14 +6,12 @@ 安装流程 ++++++++ -PaddlePaddle提供数个预编译的二进制来进行安装，包括Docker镜像，ubuntu的deb安装包等。我们推荐使用Docker镜像来部署环境，同时欢迎贡献更多的安装包。 +PaddlePaddle提供Docker镜像来部署环境。 .. toctree:: :maxdepth: 1 docker_install_cn.rst - ubuntu_install_cn.rst - 编译流程 diff --git a/doc/getstarted/build_and_install/index_en.rst b/doc/getstarted/build_and_install/index_en.rst index 1bfd4f75c0..8a53588e04 100644 --- a/doc/getstarted/build_and_install/index_en.rst +++ b/doc/getstarted/build_and_install/index_en.rst @@ -8,14 +8,13 @@ Install PaddlePaddle :maxdepth: 1 docker_install_en.rst - ubuntu_install_en.rst Build from Source ----------------- .. warning:: - Please use :code:`deb` package or :code:`docker` image to install paddle. The building guide is used for hacking or contributing PaddlePaddle source code. + Please use :code:`docker` image to install paddle. The building guide is used for hacking or contributing PaddlePaddle source code. .. toctree:: :maxdepth: 1 diff --git a/doc/getstarted/build_and_install/ubuntu_install_cn.rst b/doc/getstarted/build_and_install/ubuntu_install_cn.rst deleted file mode 100644 index 9e39ccb00f..0000000000 --- a/doc/getstarted/build_and_install/ubuntu_install_cn.rst +++ /dev/null @@ -1,71 +0,0 @@ -Ubuntu部署PaddlePaddle -=================================== - -PaddlePaddle提供了ubuntu 14.04 deb安装包。 - -安装 ------- - -安装包的下载地址是\: https://github.com/PaddlePaddle/Paddle/releases - -它包含四个版本\: - -* cpu版本: 支持主流x86处理器平台, 使用了avx指令集。 - -* cpu-noavx版本：支持主流x86处理器平台，没有使用avx指令集。 - -* gpu版本：支持主流x86处理器平台，支持nvidia cuda平台，使用了avx指令集。 - -* gpu-noavx版本：支持主流x86处理器平台，支持nvidia cuda平台，没有使用avx指令集。 - -下载完相关安装包后，执行: - -.. code-block:: shell - - sudo apt-get install gdebi - gdebi paddle-*-cpu.deb - -或者: - -.. code-block:: shell - - dpkg -i paddle-*-cpu.deb - apt-get install -f - - -在 :code:`dpkg -i` 的时候如果报一些依赖未找到的错误是正常的， -在 :code:`apt-get install -f` 里会继续安装 PaddlePaddle。 - -安装完成后，可以使用命令 :code:`paddle version` 查看安装后的paddle 版本: - -.. code-block:: shell - - PaddlePaddle 0.8.0b1, compiled with - with_avx: ON - with_gpu: OFF - with_double: OFF - with_python: ON - with_rdma: OFF - with_timer: OFF - with_predict_sdk: - - -可能遇到的问题 --------------- - -libcudart.so/libcudnn.so找不到 -++++++++++++++++++++++++++++++ - -安装完成后，运行 :code:`paddle train` 报错\: - -.. code-block:: shell - - 0831 12:36:04.151525 1085 hl_dso_loader.cc:70] Check failed: nullptr != *dso_handle For Gpu version of PaddlePaddle, it couldn't find CUDA library: libcudart.so Please make sure you already specify its path.Note: for training data on Cpu using Gpu version of PaddlePaddle,you must specify libcudart.so via LD_LIBRARY_PATH. - -原因是未设置cuda运行时环境变量。 如果使用GPU版本的PaddlePaddle，请安装CUDA 7.5 和CUDNN 5到本地环境中，并设置： - -.. code-block:: shell - - export LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda/lib:$LD_LIBRARY_PATH - export PATH=/usr/local/cuda/bin:$PATH - diff --git a/doc/getstarted/build_and_install/ubuntu_install_en.rst b/doc/getstarted/build_and_install/ubuntu_install_en.rst deleted file mode 100644 index ea8042085b..0000000000 --- a/doc/getstarted/build_and_install/ubuntu_install_en.rst +++ /dev/null @@ -1,25 +0,0 @@ -Debian Package installation guide -================================= - -PaddlePaddle supports :code:`deb` pacakge. The installation of this :code:`deb` package is tested in ubuntu 14.04, but it should be support other debian based linux, too. - -There are four versions of debian package, :code:`cpu`, :code:`gpu`, :code:`cpu-noavx`, :code:`gpu-noavx`. And :code:`noavx` version is used to support CPU which does not contain :code:`AVX` instructions. The download url of :code:`deb` package is \: https://github.com/baidu/Paddle/releases/ - - -After downloading PaddlePaddle deb packages, you can use :code:`gdebi` install. - -.. code-block:: bash - - gdebi paddle-*.deb - -If :code:`gdebi` is not installed, you can use :code:`sudo apt-get install gdebi` to install it. - -Or you can use following commands to install PaddlePaddle. - -.. code-block:: bash - - dpkg -i paddle-*.deb - apt-get install -f - -And if you use GPU version deb package, you need to install CUDA toolkit and cuDNN, and set related environment variables(such as LD_LIBRARY_PATH) first. It is normal when `dpkg -i` get errors. `apt-get install -f` will continue install paddle, and install dependences. - diff --git a/doc/howto/dev/build_cn.md b/doc/howto/dev/build_cn.md new file mode 100644 index 0000000000..0b911f7b75 --- /dev/null +++ b/doc/howto/dev/build_cn.md @@ -0,0 +1,124 @@ +# 编译PaddlePaddle和运行单元测试 + +## 需要的软硬件 + +为了开发PaddlePaddle，我们需要 + +1. 一台电脑，可以装的是 Linux, BSD, Windows 或者 MacOS 操作系统，以及 +1. Docker。 + +不需要依赖其他任何软件了。即便是 Python 和 GCC 都不需要，因为我们会把所有编译工具都安装进一个 Docker image 里。 + +## 总体流程 + +1. 获取源码 + + ```bash + git clone https://github.com/paddlepaddle/paddle + ``` + +2. 安装开发工具到 Docker image 里 + + ```bash + cd paddle; docker build -t paddle:dev . + ``` + + 请注意这个命令结尾处的 `.`；它表示 `docker build` 应该读取当前目录下的 [`Dockerfile`文件](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile)，按照其内容创建一个名为 `paddle:dev` 的 Docker image，并且把各种开发工具安装进去。 + +3. 编译 + + 以下命令启动一个 Docker container 来执行 `paddle:dev` 这个 Docker image，同时把当前目录（源码树根目录）映射为 container 里的 `/paddle` 目录，并且运行 `Dockerfile` 描述的默认入口程序 [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `cmake` 和 `make` 来编译 `/paddle` 里的源码，结果输出到 `/paddle/build`，也就是本地的源码树根目录里的 `build` 子目录。 + + ```bash + docker run --rm -v $PWD:/paddle paddle:dev + ``` + + 上述命令编译出一个 CUDA-enabled 版本。如果我们只需要编译一个只支持 CPU 的版本，可以用 + + ```bash + docker run --rm -e WITH_GPU=OFF -v $PWD:/paddle paddle:dev + ``` + +4. 运行单元测试 + + 用本机的第一个 GPU 来运行包括 GPU 单元测试在内的所有单元测试： + + ```bash + NV_GPU=0 nvidia-docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest" + ``` + + 如果编译的时候我们用了 `WITH_GPU=OFF` 选项，那么编译过程只会产生 CPU-based 单元测试，那么我们也就不需要 nvidia-docker 来运行单元测试了。我们只需要： + + ```bash + docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest" + ``` + + 有时候我们只想运行一个特定的单元测试，比如 `memory_test`，我们可以 + + ```bash + nvidia-docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest -V -R memory_test" + ``` + +5. 清理 + + 有时候我们会希望清理掉已经下载的第三方依赖以及已经编译的二进制文件。此时只需要： + + ```bash + rm -rf build + ``` + +## 为什么要 Docker 呀？ + +- 什么是 Docker? + + 如果您没有听说 Docker，可以把它想象为一个类似 virtualenv 的系统，但是虚拟的不仅仅是 Python 的运行环境。 + +- Docker 还是虚拟机？ + + 有人用虚拟机来类比 Docker。需要强调的是：Docker 不会虚拟任何硬件，Docker container 里运行的编译工具实际上都是在本机的 CPU 和操作系统上直接运行的，性能和把编译工具安装在本机运行一样。 + +- 为什么用 Docker? + + 把工具和配置都安装在一个 Docker image 里可以标准化编译环境。这样如果遇到问题，其他人可以复现问题以便帮助。 + + 另外，对于习惯使用Windows和MacOS的开发者来说，使用Docker就不用配置交叉编译环境了。 + +- 我可以选择不用Docker吗？ + + 当然可以。大家可以用把开发工具安装进入 Docker image 一样的方式，把这些工具安装到本机。这篇文档介绍基于 Docker 的开发流程，是因为这个流程比其他方法都更简便。 + +- 学习 Docker 有多难？ + + 理解 Docker 并不难，大概花十分钟看一下[这篇文章](https://zhuanlan.zhihu.com/p/19902938)。这可以帮您省掉花一小时安装和配置各种开发工具，以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。 + +- 我可以用 IDE 吗？ + + 当然可以，因为源码就在本机上。IDE 默认调用 make 之类的程序来编译源码，我们只需要配置 IDE 来调用 Docker 命令编译源码即可。 + + 很多 PaddlePaddle 开发者使用 Emacs。他们在自己的 `~/.emacs` 配置文件里加两行 + + ```emacs + (global-set-key "\C-cc" 'compile) + (setq compile-command + "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev") + ``` + + 就可以按 `Ctrl-C` 和 `c` 键来启动编译了。 + +- 可以并行编译吗？ + + 是的。我们的 Docker image 运行一个 [Bash 脚本](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。 + +## 可能碰到的问题 + +- Docker 需要 sudo + + 如果用自己的电脑开发，自然也就有管理员权限（sudo）了。如果用公用的电脑开发，需要请管理员安装和配置好 Docker。此外，PaddlePaddle 项目在努力开始支持其他不需要 sudo 的集装箱技术，比如 rkt。 + +- 在 Windows/MacOS 上编译很慢 + + Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存，以保证编译高效。具体做法请参考[这个issue](https://github.com/PaddlePaddle/Paddle/issues/627)。 + +- 磁盘不够 + + 本文中的例子里，`docker run` 命令里都用了 `--rm` 参数，这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 `docker ps -a` 命令看到停止后但是没有删除的 containers。`docker build` 命令有时候会产生一些中间结果，是没有名字的 images，也会占用磁盘。可以参考[这篇文章](https://zaiste.net/posts/removing_docker_containers/)来清理这些内容。 diff --git a/doc/howto/dev/build_en.md b/doc/howto/dev/build_en.md new file mode 100644 index 0000000000..d0048e3714 --- /dev/null +++ b/doc/howto/dev/build_en.md @@ -0,0 +1,124 @@ +# Build PaddlePaddle from Source Code and Run Unit Test + +## What Developers Need + +To contribute to PaddlePaddle, you need + +1. A computer -- Linux, BSD, Windows, MacOS, and +1. Docker. + +Nothing else. Not even Python and GCC, because you can install all build tools into a Docker image. We run all the tools by running this image. + +## General Process + +1. Retrieve source code. + + ```bash + git clone https://github.com/paddlepaddle/paddle + ``` + +2. Install build tools into a Docker image. + + ```bash + cd paddle; docker build -t paddle:dev . + ``` + + Please be aware of the `.` at the end of the command, which refers to the [`./Dockerfile` file](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile). `docker build` follows instructions in this file to create a Docker image named `paddle:dev`, and installs building tools into it. + +3. Build from source. + + This following command starts a Docker container that executes the Docker image `paddle:dev`, mapping the current directory to `/paddle/` in the container, and runs the default entry-point [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh) as specified in the Dockefile. `build.sh` invokes `cmake` and `make` to build PaddlePaddle source code, which had been mapped to `/paddle`, and writes outputs to `/paddle/build`, which maps to `build` in the current source directory on the computer. + + ```bash + docker run -v $PWD:/paddle paddle:dev + ``` + + Above command builds a CUDA-enabled version. If we want to build a CPU-only version, we can type + + ```bash + docker run -e WITH_GPU=OFF -v $PWD:/paddle paddle:dev + ``` + +4. Run unit tests. + + To run all unit tests using the first GPU of a node: + + ```bash + NV_GPU=0 nvidia-docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest" + ``` + + If we used `WITH_GPU=OFF` at build time, it generates only CPU-based unit tests, and we don't need nvidia-docker to run them. We can just run + + ```bash + docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest" + ``` + + Sometimes we want to run a specific unit test, say `memory_test`, we can run + + ```bash + nvidia-docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest -V -R memory_test" + ``` + +5. Clean Build. + + Sometimes, we might want to clean all thirt-party dependents and built binaries. To do so, just + + ```bash + rm -rf build + ``` + +## Docker, Or Not? + +- What is Docker? + + If you haven't heard of it, consider it something like Python's virtualenv. + +- Docker or virtual machine? + + Some people compare Docker with VMs, but Docker doesn't virtualize any hardware nor running a guest OS, which means there is no compromise on the performance. + +- Why Docker? + + Using a Docker image of build tools standardizes the building environment, which makes it easier for others to reproduce your problems and to help. + + Also, some build tools don't run on Windows or Mac or BSD, but Docker runs almost everywhere, so developers can use whatever computer they want. + +- Can I choose not to use Docker? + + Sure, you don't have to install build tools into a Docker image; instead, you can install them in your local computer. This document exists because Docker would make the development way easier. + +- How difficult is it to learn Docker? + + It takes you ten minutes to read [an introductory article](https://docs.docker.com/get-started) and saves you more than one hour to install all required build tools, configure them, especially when new versions of PaddlePaddle require some new tools. Not even to mention the time saved when other people trying to reproduce the issue you have. + +- Can I use my favorite IDE? + + Yes, of course. The source code resides on your local computer, and you can edit it using whatever editor you like. + + Many PaddlePaddle developers are using Emacs. They add the following few lines into their `~/.emacs` configure file: + + ```emacs + (global-set-key "\C-cc" 'compile) + (setq compile-command + "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev") + ``` + + so they could type `Ctrl-C` and `c` to build PaddlePaddle from source. + +- Does Docker do parallel building? + + Our building Docker image runs a [Bash script](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh), which calls `make -j$(nproc)` to starts as many processes as the number of your CPU cores. + +## Some Gotchas + +- Docker requires sudo + + An owner of a computer has the administrative privilege, a.k.a., sudo, and Docker requires this privilege to work properly. If you use a shared computer for development, please ask the administrator to install and configure Docker. We will do our best to support rkt, another container technology that doesn't require sudo. + +- Docker on Windows/MacOS builds slowly + + On Windows and MacOS, Docker containers run in a Linux VM. You might want to give this VM some more memory and CPUs so to make the building efficient. Please refer to [this issue](https://github.com/PaddlePaddle/Paddle/issues/627) for details. + +- Not enough disk space + + Examples in this article uses option `--rm` with the `docker run` command. This option ensures that stopped containers do not exist on hard disks. We can use `docker ps -a` to list all containers, including stopped. Sometimes `docker build` generates some intermediate dangling images, which also take disk space. To clean them, please refer to [this article](https://zaiste.net/posts/removing_docker_containers/). diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md new file mode 100644 index 0000000000..7f8da2da5a --- /dev/null +++ b/doc/howto/dev/new_op_cn.md @@ -0,0 +1,324 @@ +# 如何写新的Operator + + - [概念简介](#概念简介) + - [实现C++类](#实现C++类) + - [定义ProtoMaker类](#定义ProtoMaker类) + - [定义Operator类](#定义Operator类) + - [定义OpKernel类](#定义OpKernel类) + - [注册Operator](#注册Operator) + - [编译](#编译) + - [绑定Python](#绑定Python) + - [实现单元测试](#实现单元测试) + - [前向Operator单测](#前向Operator单测) + - [反向Operator单测](#反向Operator单测) + - [编译和执行](#编译和执行) + + +## 概念简介 + +简单介绍需要用到基类，详细介绍请参考设计文档。 + +- `framework::OperatorBase`: Operator(简写，Op)基类。 +- `framework::OpKernel`: Op计算函数的基类，称作Kernel。 +- `framework::OperatorWithKernel`：继承自OperatorBase，Op有计算函数，称作有Kernel。 +- `class OpProtoAndCheckerMaker`：描述该Op的输入、输出、属性、注释,主要用于Python API接口生成 + +依据是否包含kernel，将Op分为两种：包含Kernel的Op和不包含kernel的Op，前者Op的定义继承自`OperatorBase`，后者继承自`OperatorWithKernel`。本教程主要介绍带Kernel的Op如何写，简单总结Op需要包含的内容如下： + + + 内容 | 定义位置 +-------------- | :---------------------- +OpProtoMake定义 | `.cc`文件，Backward Op不需要定义OpProtoMake +Op定义 | `.cc`文件 +Kernel实现 | CPU、GPU共享Kernel在`.h`文件，否则，CPU可以在`.cc`文件，GPU可在`.cu`文件。 +注册Op | Op注册在`.cc`文件；Kernel注册CPU在`.cc`文件，GPU在`.cu`文件 + + +下面以矩阵乘操作，即[MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc)为例来介绍如何写带Kernel的Operator。 + + +## 实现C++类 + + +### 1. 定义ProtoMaker类 + +矩阵乘的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。首先定义`ProtoMaker`来描述该Op的输入、输出及注释： + +``` +class MulOpMaker : public framework::OpProtoAndCheckerMaker { + public: + MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The first input of mul op"); + AddInput("Y", "The second input of mul op"); + AddOutput("Out", "The output of mul op"); + AddComment(R"DOC( +Two Element Mul Operator. +The equation is: Out = X * Y +)DOC"); + } +}; +``` + +[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L43)继承自`framework::OpProtoAndCheckerMaker`，构造函数包括2个： + + - `framework::OpProto` ： 前者存储Op的输入输出和参数属性，将用于Python API接口的生成。 + - `framework::OpAttrChecker` ：后者用于检查参数属性的合法性。 + +构造函数里通过`AddInput`添加输入参数，通过`AddOutput`添加输出参数，通过`AddComment`添加该Op的注释，这些函数会将对应内容添加到`OpProto`中。 + +在`MulOp`中添加两个输入`X`和`Y`，添加了一个输出`Out`，并解释了各自含义，该命名尽可能的规范。 + + +再举个[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)的例子： + +``` +template +class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The input tensor of scale operator.").NotInGradient(); + AddOutput("Out", "The output tensor of scale operator.").NotInGradient(); + AddComment(R"DOC(Scale operator +The equation is: Out = scale*X +)DOC"); + AddAttr("scale", "scale of scale operator.").SetDefault(1.0); + } +}; +``` + + 在这个例子里，两处不同： + + - `AddInput("X","...").NotInGradient()` : 表示`X`这个输入不参与`ScaleOp`对应的梯度Op计算之中。 + - `AddAttr("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。 + + +### 2. 定义Operator类 + + +```c++ +class MulOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + auto dim0 = ctx.Input("X")->dims(); + auto dim1 = ctx.Input("Y")->dims(); + PADDLE_ENFORCE_EQ(dim0.size(), 2, + "input X(%s) should be a tensor with 2 dims, a matrix", + ctx.op_.Input("X")); + PADDLE_ENFORCE_EQ(dim1.size(), 2, + "input Y(%s) should be a tensor with 2 dims, a matrix", + ctx.op_.Input("Y")); + PADDLE_ENFORCE_EQ( + dim0[1], dim1[0], + "First matrix's width must be equal with second matrix's height."); + ctx.Output("Out")->Resize({dim0[0], dim1[1]}); + } +}; +``` + +[`MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L22)继承自`OperatorWithKernel`。`public`成员： + +```c++ +using framework::OperatorWithKernel::OperatorWithKernel; +``` + +这句表示使用基类`OperatorWithKernel`的构造函数，也可写成： + +```c++ +MulOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} +``` + +还需要重写`InferShape`接口。`InferShape`为const函数，不能修改Op的成员变量，参数为`const framework::InferShapeContext &ctx`，通过该参数可获取到输入输出以及属性。它的功能是： + + - 1). 做检查， 尽早报错：检查输入数据维度、类型等是否合法。 + - 2). 设置输出Tensor的形状。 + +通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和要讲到的注册函数一起放在`.cc`中 + +### 3. 定义OpKernel类 + +```C++ +template +class MulKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* X = context.Input("X"); + auto* Y = context.Input("Y"); + auto* Z = context.Output("Out"); + Z->mutable_data(context.GetPlace()); + auto* device_context = + const_cast(context.device_context_); + math::matmul(*X, false, *Y, false, 1, Z, 0, device_context); + } +}; +``` + +`MulKernel`继承自`framework::OpKernel`，带有模板参数: + + - `typename Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 + + - `typename T` : 表示数据类型，如`float`, `double`等。 + +`MulKernel`需要重写`Compute`接口，该接口参数为`const framework::ExecutionContext& context`, `ExecutionContext`相比`InferShapeContext`增加了设备类型，同样可获取到输入输出和属性参数，`Compute`函数里写具体实现时。 + +注意，不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。`MulOp`的CPU、GPU实现共享同一个`Kernel`，`OpKernel`不共享的例子可以参考[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 + +到此前向Op实现完成，需要在`.cc`文件中注册该op和kernel。反向Op类的定义和Kernel定义与前向Op类似，这里不再重复。但注意，反向Op没有`ProtoMaker`。 + +### 4. 注册Operator + +在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。 + +```c++ +namespace ops = paddle::operators; +REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad); +REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); +REGISTER_OP_CPU_KERNEL(mul_grad, + ops::MulGradKernel); +``` + + - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`， + - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。 + - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。 + +在 `.cu`文件中注册GPU Kernel。 + +```c++ +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel); +REGISTER_OP_GPU_KERNEL(mul_grad, + ops::MulGradKernel); +``` + +### 5. 编译 + +在[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)文件中添加编译。 + +``` +op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function) +``` + +下面命令可以编译： + +``` +make mul_op +``` + +## 绑定Python + +- 绑定Python + + 在 [`paddle/pybind/pybind.cc +`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc)文件中添加该类： + + ``` + USE_OP(mul); + ``` + 如果只实现了CPU版本，则使用`USE_CPU_ONLY_OP`: + + ``` + USE_CPU_ONLY_OP(gather); + ``` + + 使用`USE_OP`告知编译器需要链接该Op的目标文件，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。 + + + - 生成库 + + 在 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件添加类到`DEPS`中，使得该Op可以链接到生成的lib库中。 + + ``` + if(WITH_PYTHON) + cc_library(paddle_pybind SHARED + SRCS pybind.cc + DEPS pybind python backward + mul_op + minus_op) + endif(WITH_PYTHON) + ``` + +## 实现单元测试 + +单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单测](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。 + +### 前向Operator单测 + +前向Op单测继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`，具体单测流程在`OpTestMeta`里完成。需在`setUp`函数定义输入输出和属性参数，以及Python对比的输出值。 + +``` +import unittest +import numpy as np +from gradient_checker import GradientChecker, create_op +from op_test_util import OpTestMeta + +class TestMulOp(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "mul" + self.inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'Y': np.random.random((84, 100)).astype("float32") + } + self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} +``` + 首先需要`import`必要的包,下面详细解释其他值： + + - `self.type = "mul" ` : 定义类型，和注册的类型一致。 + - `self.inputs` : 定义输入，类型为Numpy.array，并初始化。 + - `self.outputs` : 定义输出，并得到Python结算结果。 + + +### 反向Operator单测 + +反向Op单测继承自`GradientChecker`，而`GradientChecker`集成自`unittest.TestCase`，所以反向单测函数需要`test_`开头。 + + ``` + class MulGradOpTest(GradientChecker): + def test_mul(self): + op = create_op("mul") + inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'Y': np.random.random((84, 100)).astype("float32") + } + self.compare_grad(op, inputs) + # mul op will enlarge the relative error + self.check_grad( + op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.5) + ``` + + - 调用`create_op("mul")`创建反向Op对应的前向Op。 + - 定义输入`inputs`。 + - 调用`compare_grad`函数对比CPU、GPU计算结果。 + - 调用`check_grad`检查梯度稳定性，这里采用数值法检测梯度正确性。 + - 第一个参数`op` : 前向op。 + - 第二个参数`inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。 + - 第三个参数`set(["X", "Y"])` : 指定对输入变量`X`、`Y`做梯度检测。 + - 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out` + + +### 编译和执行 + +单测完成之后，在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)里添加编译： + +``` +py_test(test_mul_op SRCS test_mul_op.py) +``` + +编译时需要打开`WITH_TESTING`, 即 `cmake paddle_dir -DWITH_TESTING=ON`，编译成功之后执行单测命令为： + +``` +make test ARGS="-R test_mul_op -V" +``` +或者: + +``` +ctest -R test_mul_op +``` diff --git a/doc/howto/index_cn.rst b/doc/howto/index_cn.rst index 26449a6365..0608aa3096 100644 --- a/doc/howto/index_cn.rst +++ b/doc/howto/index_cn.rst @@ -19,6 +19,7 @@ .. toctree:: :maxdepth: 1 + dev/build_cn.rst dev/write_docs_cn.rst dev/contribute_to_paddle_cn.md diff --git a/doc/howto/index_en.rst b/doc/howto/index_en.rst index 1fbfcd260b..1b6034be4e 100644 --- a/doc/howto/index_en.rst +++ b/doc/howto/index_en.rst @@ -18,6 +18,7 @@ Development .. toctree:: :maxdepth: 1 + dev/build_en.rst dev/new_layer_en.rst dev/contribute_to_paddle_en.md diff --git a/doc/index_en.rst b/doc/index_en.rst index 168c7667c6..64684b8b9b 100644 --- a/doc/index_en.rst +++ b/doc/index_en.rst @@ -7,4 +7,3 @@ PaddlePaddle Documentation getstarted/index_en.rst howto/index_en.rst api/index_en.rst - about/index_en.rst diff --git a/doc/templates/conf.py.cn.in b/doc/templates/conf.py.cn.in index 673948dfe7..41b35b5b23 100644 --- a/doc/templates/conf.py.cn.in +++ b/doc/templates/conf.py.cn.in @@ -13,7 +13,7 @@ # serve to show the default. import sys import os, subprocess -sys.path.insert(0, os.path.abspath('@PROJ_ROOT@/python')) +sys.path.insert(0, os.path.abspath('@PADDLE_SOURCE_DIR@/python')) import shlex from recommonmark import parser, transform import paddle @@ -24,7 +24,7 @@ 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 ------------------------------------------------ @@ -120,7 +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 b6b50b7dcd..5822c2481d 100644 --- a/doc/templates/conf.py.en.in +++ b/doc/templates/conf.py.en.in @@ -13,7 +13,7 @@ # serve to show the default. import sys import os, subprocess -sys.path.insert(0, os.path.abspath('@PROJ_ROOT@/python')) +sys.path.insert(0, os.path.abspath('@PADDLE_SOURCE_DIR@/python')) import shlex from recommonmark import parser, transform import paddle @@ -25,7 +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 ------------------------------------------------ @@ -120,7 +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/master/client.go b/go/master/client.go index 62801b9b7f..f04cf50ce3 100644 --- a/go/master/client.go +++ b/go/master/client.go @@ -63,13 +63,24 @@ func WithAddr(addr string) func(c *Client) error { // WithEtcd sets the client to use etcd for master discovery. func WithEtcd(endpoints []string, timeout time.Duration) func(*Client) error { return func(c *Client) error { - cli, err := clientv3.New(clientv3.Config{ - Endpoints: endpoints, - DialTimeout: timeout, - }) - if err != nil { + var cli *clientv3.Client + f := func() error { + var err error + cli, err = clientv3.New(clientv3.Config{ + Endpoints: endpoints, + DialTimeout: timeout, + }) return err } + for { + err := f() + if err != nil { + log.Warningln(err) + } else { + break + } + time.Sleep(time.Second) + } ch := make(chan string, 1) a, err := GetKey(cli, DefaultAddrPath, timeout) @@ -101,9 +112,6 @@ func NewClient(opts ...func(*Client) error) (*Client, error) { } } c.ch = make(chan record, c.bufSize) - // FIXME: connection is created asyncrosly in monitorMaster go routine, - // ensure the connection is ready for use before calling c.addClient. - time.Sleep(time.Second) return c, nil } diff --git a/go/pserver/client/c/test/test_train.py b/go/pserver/client/c/test/test_train.py index 572a61e4cc..8d9c6b9b20 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/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index cf61a243e9..ec866b2907 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -15,6 +15,7 @@ if(Boost_FOUND) add_subdirectory(platform) add_subdirectory(framework) add_subdirectory(operators) + add_subdirectory(pybind) endif() if(WITH_C_API) diff --git a/paddle/api/CMakeLists.txt b/paddle/api/CMakeLists.txt index 7a1e8b8b26..d7b3d2bdec 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) @@ -79,16 +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 +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 ${PROJ_ROOT}/paddle + 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) +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) diff --git a/paddle/api/ParameterUpdater.cpp b/paddle/api/ParameterUpdater.cpp index 5934cb898b..8cd73b348c 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/capi/Arguments.cpp b/paddle/capi/Arguments.cpp index 8b81ec69e6..1ec403077e 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/CMakeLists.txt b/paddle/capi/CMakeLists.txt index 11022d1754..dde99ab340 100644 --- a/paddle/capi/CMakeLists.txt +++ b/paddle/capi/CMakeLists.txt @@ -53,7 +53,10 @@ add_custom_target(paddle_capi_whole ALL set_target_properties(paddle_capi_whole PROPERTIES IMPORTED_LOCATION ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library}) +set(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map") +# TODO: merge mkl into paddle_capi_shared add_library(paddle_capi_shared SHARED ${CAPI_SOURCES}) +set_target_properties(paddle_capi_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}") target_include_directories(paddle_capi_shared PUBLIC ${CMAKE_CURRENT_BINARY_DIR}) link_paddle_exe(paddle_capi_shared) diff --git a/paddle/capi/arguments.h b/paddle/capi/arguments.h index d71ea26a5d..7c32524a00 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 a78522e4a7..e32f2f9836 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/export.map b/paddle/capi/export.map new file mode 100644 index 0000000000..8d673f675d --- /dev/null +++ b/paddle/capi/export.map @@ -0,0 +1,6 @@ +{ + global: + paddle_*; + local: + *; +}; diff --git a/paddle/capi/export.sym b/paddle/capi/export.sym new file mode 100644 index 0000000000..e69de29bb2 diff --git a/paddle/capi/gradient_machine.cpp b/paddle/capi/gradient_machine.cpp index 00f76e0152..629449bbd4 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; @@ -121,3 +146,19 @@ paddle_error paddle_gradient_machine_randomize_param( m->machine->randParameters(); return kPD_NO_ERROR; } + +paddle_error paddle_gradient_machine_get_layer_output( + paddle_gradient_machine machine, + const char* layerName, + paddle_arguments args) { + auto m = cast(machine); + auto out = paddle::capi::cast(args); + if (m == nullptr || layerName == nullptr || out == nullptr || + m->machine == nullptr) { + return kPD_NULLPTR; + } + + auto layerOutput = m->machine->getLayerOutput(layerName); + out->args.push_back(layerOutput); + return kPD_NO_ERROR; +} diff --git a/paddle/capi/gradient_machine.h b/paddle/capi/gradient_machine.h index d7e2dd9bf8..28eeb23e3b 100644 --- a/paddle/capi/gradient_machine.h +++ b/paddle/capi/gradient_machine.h @@ -36,6 +36,22 @@ 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`. + * Example: + * paddle merge_model \ + * --model_dir="pass-00000" \ + * --model_file="merged_model.paddle" + * @param [out] machine that used for model inference + * @param [in] mergedModel + * @param [in] size + * @return paddle_error + */ +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. @@ -85,6 +101,18 @@ paddle_gradient_machine_randomize_param(paddle_gradient_machine machine); PD_API paddle_error paddle_gradient_machine_destroy(paddle_gradient_machine machine); +/** + * @brief Get the output of the layer named `layerName`. + * @param [in] gradient machine that have run a inference + * @param [in] layerName name of specified layer + * @param [out] args output of the specified layer + * @return paddle_error + */ +PD_API paddle_error +paddle_gradient_machine_get_layer_output(paddle_gradient_machine machine, + const char* layerName, + paddle_arguments args); + #ifdef __cplusplus } #endif diff --git a/paddle/capi/tests/CMakeLists.txt b/paddle/capi/tests/CMakeLists.txt index d73f6b7733..8208808b94 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/include/hl_cuda_cudnn.h b/paddle/cuda/include/hl_cuda_cudnn.h index db18e4912b..3f68c62de6 100644 --- a/paddle/cuda/include/hl_cuda_cudnn.h +++ b/paddle/cuda/include/hl_cuda_cudnn.h @@ -214,7 +214,8 @@ extern void hl_conv_workspace(hl_tensor_descriptor input, int* convBwdDataAlgo, size_t* bwdDataLimitBytes, int* convBwdFilterAlgo, - size_t* bwdFilterLimitBytes); + size_t* bwdFilterLimitBytes, + bool useDilation); /** * @brief destroy filter descriptor. @@ -242,7 +243,9 @@ extern void hl_create_convolution_descriptor(hl_convolution_descriptor* conv, int padding_height, int padding_width, int stride_height, - int stride_width); + int stride_width, + int dilation_h = 1, + int dilation_w = 1); /** * @brief reset convolution descriptor. @@ -262,7 +265,9 @@ extern void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, int padding_height, int padding_width, int stride_height, - int stride_width); + int stride_width, + int dilation_h = 1, + int dilation_w = 1); /** * @brief destroy convolution descriptor. diff --git a/paddle/cuda/include/stub/hl_cuda_cudnn_stub.h b/paddle/cuda/include/stub/hl_cuda_cudnn_stub.h index abd0d6b099..3afcc6fa85 100644 --- a/paddle/cuda/include/stub/hl_cuda_cudnn_stub.h +++ b/paddle/cuda/include/stub/hl_cuda_cudnn_stub.h @@ -78,7 +78,9 @@ inline void hl_create_convolution_descriptor(hl_convolution_descriptor* conv, int padding_height, int padding_width, int stride_height, - int stride_width) {} + int stride_width, + int dilation_h, + int dilation_w) {} inline void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, hl_tensor_descriptor image, @@ -86,7 +88,9 @@ inline void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, int padding_height, int padding_width, int stride_height, - int stride_width) {} + int stride_width, + int dilation_h, + int dilation_w) {} inline void hl_destroy_convolution_descriptor(hl_convolution_descriptor conv) {} @@ -99,7 +103,8 @@ inline void hl_conv_workspace(hl_tensor_descriptor input, int* convBwdDataAlgo, size_t* bwdDataLimitBytes, int* convBwdFilterAlgo, - size_t* bwdFilterLimitBytes) {} + size_t* bwdFilterLimitBytes, + bool useDilation) {} inline void hl_convolution_forward(hl_tensor_descriptor input, real* input_data, diff --git a/paddle/cuda/src/hl_cuda_cudnn.cc b/paddle/cuda/src/hl_cuda_cudnn.cc index 78642a1744..f38ef69255 100644 --- a/paddle/cuda/src/hl_cuda_cudnn.cc +++ b/paddle/cuda/src/hl_cuda_cudnn.cc @@ -201,7 +201,8 @@ void hl_conv_workspace(hl_tensor_descriptor input, int* convBwdDataAlgo, size_t* bwdDataLimitBytes, int* convBwdFilterAlgo, - size_t* bwdFilterLimitBytes) { + size_t* bwdFilterLimitBytes, + bool useDilation) { #if CUDNN_VERSION >= 4000 CHECK_NOTNULL(input); @@ -213,21 +214,60 @@ void hl_conv_workspace(hl_tensor_descriptor input, size_t memoryLimitBytes = (1LL << 20) * FLAGS_cudnn_conv_workspace_limit_in_mb; + // For dilation + int algo = 0; + // cudnn convolution forward configuration cudnnTensorDescriptor_t fwd_src_desc = GET_TENSOR_DESCRIPTOR(input); cudnnTensorDescriptor_t fwd_dest_desc = GET_TENSOR_DESCRIPTOR(output); cudnnFilterDescriptor_t fwd_filter_desc = GET_FILTER_DESCRIPTOR(filter); cudnnConvolutionDescriptor_t fwd_conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv); + // cudnn convolution backward data configuration + cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter); + cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output); + cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input); + cudnnConvolutionDescriptor_t bwd_data_conv_desc = + GET_CONVOLUTION_DESCRIPTOR(conv); + // cudnn convolution backward filter configuration + cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input); + cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output); + cudnnConvolutionDescriptor_t bwd_filter_conv_desc = + GET_CONVOLUTION_DESCRIPTOR(conv); + cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter); - CHECK_CUDNN(dynload::cudnnGetConvolutionForwardAlgorithm( - t_resource.cudnn_handle, - fwd_src_desc, - fwd_filter_desc, - fwd_conv_desc, - fwd_dest_desc, - CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, - memoryLimitBytes, - reinterpret_cast(convFwdAlgo))); + if (useDilation) { + convFwdAlgo = &algo; + convBwdDataAlgo = &algo; + convBwdFilterAlgo = &algo; + } else { + CHECK_CUDNN(dynload::cudnnGetConvolutionForwardAlgorithm( + t_resource.cudnn_handle, + fwd_src_desc, + fwd_filter_desc, + fwd_conv_desc, + fwd_dest_desc, + CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, + memoryLimitBytes, + reinterpret_cast(convFwdAlgo))); + CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataAlgorithm( + t_resource.cudnn_handle, + bwd_data_filter_desc, + bwd_data_diff_desc, + bwd_data_conv_desc, + bwd_data_grad_desc, + CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT, + memoryLimitBytes, + reinterpret_cast(convBwdDataAlgo))); + CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterAlgorithm( + t_resource.cudnn_handle, + bwd_filter_src_desc, + bwd_filter_diff_desc, + bwd_filter_conv_desc, + bwd_filter_grad_desc, + CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT, + memoryLimitBytes, + reinterpret_cast(convBwdFilterAlgo))); + } CHECK_CUDNN(dynload::cudnnGetConvolutionForwardWorkspaceSize( t_resource.cudnn_handle, @@ -238,23 +278,6 @@ void hl_conv_workspace(hl_tensor_descriptor input, static_cast(*convFwdAlgo), fwdLimitBytes)); - // cudnn convolution backward data configuration - cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter); - cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output); - cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input); - cudnnConvolutionDescriptor_t bwd_data_conv_desc = - GET_CONVOLUTION_DESCRIPTOR(conv); - - CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataAlgorithm( - t_resource.cudnn_handle, - bwd_data_filter_desc, - bwd_data_diff_desc, - bwd_data_conv_desc, - bwd_data_grad_desc, - CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT, - memoryLimitBytes, - reinterpret_cast(convBwdDataAlgo))); - CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataWorkspaceSize( t_resource.cudnn_handle, bwd_data_filter_desc, @@ -264,23 +287,6 @@ void hl_conv_workspace(hl_tensor_descriptor input, static_cast(*convBwdDataAlgo), bwdDataLimitBytes)); - // cudnn convolution backward filter configuration - cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input); - cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output); - cudnnConvolutionDescriptor_t bwd_filter_conv_desc = - GET_CONVOLUTION_DESCRIPTOR(conv); - cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter); - - CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterAlgorithm( - t_resource.cudnn_handle, - bwd_filter_src_desc, - bwd_filter_diff_desc, - bwd_filter_conv_desc, - bwd_filter_grad_desc, - CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT, - memoryLimitBytes, - reinterpret_cast(convBwdFilterAlgo))); - CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize( t_resource.cudnn_handle, bwd_filter_src_desc, @@ -603,7 +609,9 @@ void hl_create_convolution_descriptor(hl_convolution_descriptor* conv, int padding_height, int padding_width, int stride_height, - int stride_width) { + int stride_width, + int dilation_h, + int dilation_w) { CHECK_NOTNULL(conv); cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)malloc( @@ -625,18 +633,24 @@ void hl_create_convolution_descriptor(hl_convolution_descriptor* conv, padding_width, stride_height, stride_width, - 1, - 1, + dilation_h, + dilation_w, mode, data_type)); #else + if (dilation_h > 1 || dilation_w > 1) { + LOG(FATAL) + << "Current cuDNN version does't support for dilation convolution. " + << "The dilation convolution requires cuDNN >= v6.0."; + } + CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(hl_conv->desc, padding_height, padding_width, stride_height, stride_width, - 1, - 1, + dilation_h, + dilation_w, mode)); #endif @@ -659,7 +673,9 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, int padding_height, int padding_width, int stride_height, - int stride_width) { + int stride_width, + int dilation_h, + int dilation_w) { CHECK_NOTNULL(conv); CHECK_NOTNULL(image); CHECK_NOTNULL(filter); @@ -678,8 +694,8 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, padding_width, stride_height, stride_width, - 1, - 1, + dilation_h, + dilation_w, mode, data_type)); #else @@ -688,8 +704,8 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv, padding_width, stride_height, stride_width, - 1, - 1, + dilation_h, + dilation_w, mode)); #endif diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt index 6601918c90..c0838d9b75 100644 --- a/paddle/framework/CMakeLists.txt +++ b/paddle/framework/CMakeLists.txt @@ -7,7 +7,7 @@ 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 details/lod_tensor.cc DEPS ddim place 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) @@ -15,40 +15,27 @@ cc_test(variable_test SRCS variable_test.cc) cc_library(scope SRCS scope.cc) cc_test(scope_test SRCS scope_test.cc DEPS scope) -proto_library(attribute_proto SRCS attribute.proto) -proto_library(op_proto SRCS op_proto.proto DEPS attribute_proto) -proto_library(op_desc SRCS op_desc.proto DEPS attribute_proto) -cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf) -cc_test(op_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf) +proto_library(framework_proto SRCS framework.proto) -cc_library(attribute SRCS attribute.cc DEPS op_desc op_proto) - -cc_library(operator SRCS operator.cc DEPS op_desc device_context tensor scope attribute) +cc_library(attribute SRCS attribute.cc DEPS framework_proto) +cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto) +cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope) cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry) -cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS op_proto operator) -cc_library(op_registry SRCS op_registry.cc DEPS op_desc grad_op_builder) +cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS operator) +cc_library(op_registry SRCS op_registry.cc DEPS grad_op_builder) cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry) cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry add_op) -py_proto_compile(framework_py_proto SRCS attribute.proto op_proto.proto op_desc.proto) +py_proto_compile(framework_py_proto SRCS framework.proto) # Generate an empty __init__.py to make framework_py_proto as a valid python module. add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py) add_dependencies(framework_py_proto framework_py_proto_init) +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) - -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 - uniform_random_op - fill_zeros_like_op) -endif(WITH_PYTHON) +cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context) diff --git a/paddle/framework/attribute.cc b/paddle/framework/attribute.cc index 4c5790693b..9eb07acdff 100644 --- a/paddle/framework/attribute.cc +++ b/paddle/framework/attribute.cc @@ -44,7 +44,7 @@ AttrType AttrTypeID>() { return STRINGS; } -Attribute GetAttrValue(const AttrDesc& attr_desc) { +Attribute GetAttrValue(const OpDesc::Attr& attr_desc) { switch (attr_desc.type()) { case paddle::framework::AttrType::INT: { return attr_desc.i(); diff --git a/paddle/framework/attribute.h b/paddle/framework/attribute.h index 3a5820e9c6..08b47cabd4 100644 --- a/paddle/framework/attribute.h +++ b/paddle/framework/attribute.h @@ -14,16 +14,15 @@ limitations under the License. */ #pragma once -#include #include #include #include #include #include -#include "paddle/framework/attribute.pb.h" -#include "paddle/framework/op_desc.pb.h" +#include "paddle/framework/framework.pb.h" #include "paddle/platform/enforce.h" +#include "paddle/platform/variant.h" namespace paddle { namespace framework { @@ -37,7 +36,7 @@ typedef std::unordered_map AttributeMap; template AttrType AttrTypeID(); -Attribute GetAttrValue(const AttrDesc& attr_desc); +Attribute GetAttrValue(const OpDesc::Attr& attr_desc); // check whether a value(attribute) fit a certain limit template diff --git a/paddle/framework/attribute.proto b/paddle/framework/attribute.proto deleted file mode 100644 index 13ae312c10..0000000000 --- a/paddle/framework/attribute.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/backward.cc b/paddle/framework/backward.cc index 47983110fa..6b4c612cd8 100644 --- a/paddle/framework/backward.cc +++ b/paddle/framework/backward.cc @@ -15,31 +15,44 @@ #include "paddle/framework/backward.h" #include +#include + #include "paddle/framework/op_registry.h" #include "paddle/operators/net_op.h" +#include "paddle/operators/recurrent_op.h" namespace paddle { namespace framework { -static bool AllInSet(const std::vector& names, - const std::string& suffix, - const std::unordered_set& set) { +template +static void ForEachVarName(const Map& names, T callback) { for (auto& name : names) { - if (set.find(name + suffix) == set.end()) { - return false; + for (auto& n : name.second) { + if (callback(n)) return; } } - return true; } -static std::shared_ptr NOP() { - auto net_op = std::make_shared(); - net_op->type_ = "@NOP@"; +// return whether all the names + suffixes in the set +static bool AllInSet( + const std::map>& names, + const std::string& suffix, const std::unordered_set& set) { + bool all_in_set = true; + ForEachVarName(names, [&all_in_set, &set, &suffix](const std::string& n) { + all_in_set = set.find(n + suffix) != set.end(); + return !all_in_set; + }); + return all_in_set; +} + +static std::unique_ptr NOP() { + auto net_op = new operators::NetOp(); + net_op->SetType("@NOP@"); net_op->CompleteAddOp(); - return net_op; + return std::unique_ptr(net_op); } -// Get backward operator from a forward operator, recursively implementation. +// Get backward operator from a forward operator, a recursive implementation. // // no_grad_names the gradient variable names without gradient calculating. // @@ -47,127 +60,161 @@ static std::shared_ptr NOP() { // BackwardRecursive. use `uid = uniq_id++;` to get the unique index, and // pass `uniq_id` through recursive calling. // -// returns The backward operator. For simple situation, it is a simple -// operator. For complex situation, it is a NetOp. +// returns The backward operator. In a simple situation, it may be a simple +// operator, in a complex situation, it maybe a NetOp. // // See Backward.h for details -static std::shared_ptr BackwardRecursive( - const OperatorBase& forwardOp, - std::unordered_set& no_grad_names, size_t& uniq_id); -std::shared_ptr BackwardRecursive( +static std::unique_ptr BackwardRecursive( const OperatorBase& forwardOp, std::unordered_set& no_grad_names, size_t& uniq_id) { // If all input gradients of forwarding operator do not need to calculate, // just return an NOP. Not return null ptr because NOP does not take // too much time for calculation, but it is useful for simplifying logic. - if (AllInSet(forwardOp.inputs_, kGradVarSuffix, no_grad_names)) { + if (AllInSet(forwardOp.Inputs() /*names*/, kGradVarSuffix /*suffix*/, + no_grad_names /*set*/)) { return NOP(); } // All output gradients of forwarding operator do not need to calculate. // Then all input gradients cannot be computed at all, and we put them into // `no_grad_names` set. Return an NOP. - if (AllInSet(forwardOp.outputs_, kGradVarSuffix, no_grad_names)) { - for (auto& name : forwardOp.inputs_) { - // Mark all input is not need - no_grad_names.insert(name + kGradVarSuffix); - } + if (AllInSet(forwardOp.Outputs() /*names*/, kGradVarSuffix /*suffix*/, + no_grad_names /*set*/)) { + ForEachVarName(forwardOp.Inputs(), + [&no_grad_names](const std::string& name) -> bool { + no_grad_names.insert(GradVarName(name)); + return false; + }); return NOP(); } // Returned gradient network - auto net = std::make_shared(); + auto net = std::unique_ptr(new operators::NetOp()); if (forwardOp.IsNetOp()) { // Because forwardOp is a net op, it can static_cast. auto& forwardNet = static_cast(forwardOp); // Map from output gradient variable name to operator's indices in - // backward net. That operator generates that variable. + // backward net's ops_. That operator generates that variable. std::unordered_map> dup_output_ops; size_t local_op_id = 0; - // reversely travel forwardNet + // reversely travel forwardNet and collect all duplicate outputs. for (auto it = forwardNet.ops_.rbegin(); it != forwardNet.ops_.rend(); ++it, ++local_op_id) { - auto fwd = *it; + auto& fwd = *it; auto bwd = BackwardRecursive(*fwd, no_grad_names, uniq_id); - net->AddOp(bwd); - for (auto& out : bwd->outputs_) { - dup_output_ops[out].emplace_back(local_op_id); - } + ForEachVarName(bwd->Outputs(), + [&dup_output_ops, local_op_id](const std::string& out) { + dup_output_ops[out].emplace_back(local_op_id); + return false; + }); + net->AppendOp(std::move(bwd)); } // Get unique ID for this method. auto uid = uniq_id++; // TODO(dzh): more comment - using Pos = std::pair>; + // multiple operators which have the same output (y for example) may + // overwrite the same y variable when backward, special operations are token + // to handle this case. For each duplicate output, rename it to an alias + // (original name with a offset), append an `add` op for its operator, + // and finally sum all the alias variable to the final output variable y. + using Pos = std::pair>; std::list insert_position; for (auto& dup_output_op : dup_output_ops) { const std::string& name = dup_output_op.first; + // duplicate @Empty@ don't need to be added + if (name == kEmptyVarName) continue; + auto& dup_op = dup_output_op.second; + // no duplicate output if (dup_op.size() == 1) continue; - std::vector dup_outputs; + // process the duplicate outputs + std::vector dup_outputs; for (size_t i = 0; i < dup_op.size(); ++i) { + // rename each duplicate output to an alias auto op_offset = dup_op[i]; dup_outputs.push_back(name + "@RENAME@" + std::to_string(uid) + "@" + std::to_string(i)); net->ops_[op_offset]->Rename(name, dup_outputs.back()); } + // collect all the offset to append `add` op for each alias insert_position.push_back( - {dup_op.back(), - OpRegistry::CreateOp( - "add", {dup_outputs}, {name}, - {{"input_format", - std::vector{0, static_cast(dup_outputs.size())}}})}); + {dup_op.back(), OpRegistry::CreateOp("add", {{"X", {dup_outputs}}}, + {{"Out", {name}}}, {})}); } + // make sure the inserted `add` ops follow the BFS order. insert_position.sort( [](const Pos& l, const Pos& r) { return l.first > r.first; }); for (auto& pos : insert_position) { - net->InsertOp(pos.first + 1, pos.second); + net->InsertOp(pos.first + 1, std::move(pos.second)); } - } else { - std::shared_ptr grad_op = OpRegistry::CreateGradOp(forwardOp); - for (std::string& grad_input : grad_op->inputs_) { + std::unique_ptr grad_op(OpRegistry::CreateGradOp(forwardOp)); + + ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op]( + const std::string& grad_input) { if (no_grad_names.count(grad_input)) { - std::string prefix = - grad_input.substr(0, grad_input.size() - kGradVarSuffix.size()); - grad_input = prefix + kZeroVarSuffix; + // +1 for \0 + std::string prefix = grad_input.substr( + 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 = kEmptyVarName; + net->AppendOp(OpRegistry::CreateOp("fill_zeros_like", + {{"Src", {prefix}}}, + {{"Dst", {grad_input}}}, {})); } + return false; + }); + + ForEachVarName(grad_op->Outputs(), + [&no_grad_names, &grad_op](const std::string& grad_output) { + if (no_grad_names.count(grad_output)) { + grad_op->Rename(grad_output, kEmptyVarName); + } + return false; + }); + + // process recurrent gradient op as a special operator. + if (forwardOp.Type() == "recurrent_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 + rnn_grad_op->set_stepnet( + BackwardRecursive(stepnet_op, no_grad_names, uniq_id)); } if (net->ops_.empty()) { // Current no aux op is added to network return grad_op; } - net->AddOp(grad_op); + net->AppendOp(std::move(grad_op)); } - net->type_ = "@GENERATED_BACKWARD@"; + net->SetType("@GENERATED_BACKWARD@"); net->CompleteAddOp(); - return net; + return std::unique_ptr( + static_cast(net.release())); } // See header for comments -std::shared_ptr Backward( +std::unique_ptr Backward( const OperatorBase& forwardOp, const std::unordered_set& no_grad_vars) { std::unordered_set no_grad_names; - no_grad_names.reserve(no_grad_vars.size()); + no_grad_names.reserve(no_grad_vars.size() + 1); - no_grad_names.insert(kEmptyVarName + kGradVarSuffix); + no_grad_names.insert(std::string(kEmptyVarName) + kGradVarSuffix); for (auto& name : no_grad_vars) { no_grad_names.insert(name + kGradVarSuffix); diff --git a/paddle/framework/backward.h b/paddle/framework/backward.h index c181919dc1..1ecf69881b 100644 --- a/paddle/framework/backward.h +++ b/paddle/framework/backward.h @@ -20,7 +20,7 @@ namespace framework { // Create the backward operator from a forward operator. // TODO(yuyang18): Add more API reference comment. -extern std::shared_ptr Backward( +extern std::unique_ptr Backward( const OperatorBase& forwardOp, const std::unordered_set& no_grad_vars); } // namespace framework diff --git a/paddle/framework/backward.md b/paddle/framework/backward.md index 74c001b06a..8aa6728a95 100644 --- a/paddle/framework/backward.md +++ b/paddle/framework/backward.md @@ -1,38 +1,82 @@ -## Operator/expression 's Backward +# Operator/expression 's Backward -### Motivation +## Motivation -In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation lineage, the operator/ expression's Backward feature will generate the backward pass respect to forward pass. +In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass. + +## Backward Operator Registry -### Implement : gradient operator registry +A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs and output gradients and then calculate its input gradients. -| | forward operator | backward operator | -| ---------------------- | ---------------- | -------------------------------- | -| **Operator::inputs_** | Inputs | Inputs, Outputs, OutputGradients | -| **Operator::outputs_** | Outputs | InputGradients | +| | forward operator | backward operator +| ---------------------- | ---------------- |------------------------- | +| **Operator::inputs_** | Inputs | Inputs, Outputs, OutputGradients | +| **Operator::outputs_** | Outputs | InputGradients | -Inputs/Outputs means the input/output of the operator, InputGradients/OutputGradients is the gradient respect to forward opeartor. Forward operator and Backward operator are isomorphic, save their corresponding needs into member attribute. + In most cases, there is a one-to-one correspondence between forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced. -We use a global hash map record the gradient operators available, follow the philosophy of minimum core, make operator pluggable unit. Each gradient is an operator and it needs to regist itself. +For example, we have got a `mul_op`, and we can register it's information and corresponding backward operator by the following macro: -grad_op_builder(fengjiayi) +```cpp +REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad); +``` -### Implement : Backward network +`mul` is the operator's type. `MulOp` and `MulOpMaker` are the operator class and the operator maker class respectively. + +`mul_grad` is the type of backward operator, and `MulOpGrad` is its class name. + +## Backward Opeartor Creating + +Given a certain forward operator, we can get its corresponding backward opeartor by calling: + +```cpp +OperatorBase* bwd_op = BuildGradOp(const OperatorBase* fwd_op); +``` + +The function `BuildGradOp` will sequentially execute following processes: + +1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`. + +2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these are not necessary for gradient computing. + +3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`. + +4. Building backward operator with `inputs`, `outputs` and forward operator's attributes. + +## Backward Network Building + +A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and put them together. + +In our design, the network itself is also a kind of operator. So the operators contained by a big network may be some small network. given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`,`InputGradients`. -1. bla bla bla (yuyang) +1. Op + + when the input forward network is a Op, return its gradient Operator Immediately. 2. NetOp - when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively and ensure them done. During the process, we need to collect the `OutputGradients` name. + when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to forward NetOp. + + **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwirte their shared input variable. + +

+
+ + 1. shared variable in two operators. + +

+ + Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator replace the overwirte links. + +

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

- Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator instead. -![./images/duplicate_op2]() -​ Then collect the sub graph OutputGradients/InputGradients as the NetOp's and return it. +​ Then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it. diff --git a/paddle/framework/backward_test.cc b/paddle/framework/backward_test.cc index 6d5835bd22..f100c4d054 100644 --- a/paddle/framework/backward_test.cc +++ b/paddle/framework/backward_test.cc @@ -28,19 +28,13 @@ using OpAttrChecker = framework::OpAttrChecker; using Scope = framework::Scope; using DeviceContext = platform::DeviceContext; -class EmptyOp : public OperatorBase { - public: - void InferShape(const Scope &scope) const override {} - void Run(const Scope &scope, const DeviceContext &dev_ctx) const override {} -}; - class RowWiseAddOpMaker : public OpProtoAndCheckerMaker { public: RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "Input X of Add").IgnoreGradient(); - AddInput("b", "Bias of Add").IgnoreGradient(); - AddOutput("Out", "Out of Add").IgnoreGradient(); + AddInput("X", "Input X of Add").NotInGradient(); + AddInput("b", "Bias of Add").NotInGradient(); + AddOutput("Out", "Out of Add").NotInGradient(); AddComment("Add Op"); } }; @@ -49,8 +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"); } @@ -61,7 +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"); } }; @@ -71,21 +65,25 @@ class NoGradOpMaker : public OpProtoAndCheckerMaker { NoGradOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "X input"); - AddOutput("Y", "Y output"); + AddOutput("Out", "Y output"); AddComment("NoGradOp, same input output. no Grad"); } }; class FcOp : public operators::NetOp { public: - void Init() override { - AddOp(OpRegistry::CreateOp("mul", {Input("X"), Input("W")}, - {Output("mul_result")}, {})); - auto b_name = Input("b"); + FcOp(const std::string &type, const VariableNameMap &inputs, + const VariableNameMap &outputs, const AttributeMap &attrs) + : NetOp(type, inputs, outputs, attrs) { + AppendOp(OpRegistry::CreateOp("mul", + {{"X", {Input("X")}}, {"Y", {Input("W")}}}, + {{"Out", {Output("mul_result")}}}, {})); + auto input_b = Inputs("b"); std::string before_act = "mul_result"; - if (b_name != kEmptyVarName) { - AddOp(OpRegistry::CreateOp("rowwise_add", {Output("mul_result"), b_name}, - {Output("add_result")}, {})); + if (input_b.size() != 0) { + AppendOp(OpRegistry::CreateOp( + "rowwise_add", {{"X", {Output("mul_result")}}, {"b", {input_b[0]}}}, + {{"Out", {Output("add_result")}}}, {})); before_act = "add_result"; } else { auto out_varname = Output("add_result"); @@ -94,8 +92,8 @@ class FcOp : public operators::NetOp { } } - AddOp(OpRegistry::CreateOp("sigmoid", {Output(before_act)}, {Output("Out")}, - {})); + AppendOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}}, + {{"Out", {Output("Out")}}}, {})); CompleteAddOp(false); } }; @@ -107,8 +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(""); } @@ -139,7 +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(""); } @@ -150,51 +148,48 @@ class AddOpMaker : public OpProtoAndCheckerMaker { namespace f = paddle::framework; namespace ops = paddle::operators; using EnforceNotMet = paddle::platform::EnforceNotMet; -REGISTER_OP(rowwise_add, f::EmptyOp, f::RowWiseAddOpMaker); -REGISTER_GRADIENT_OP(rowwise_add, rowwise_add_grad, f::EmptyOp); -REGISTER_OP(mul, f::EmptyOp, f::MulOpMaker); -REGISTER_GRADIENT_OP(mul, mul_grad, f::EmptyOp); -REGISTER_OP(sigmoid, f::EmptyOp, f::SigmoidOpMaker); -REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, f::EmptyOp); -REGISTER_OP(nograd, f::EmptyOp, f::NoGradOpMaker); -REGISTER_OP(fill_zeros_like, f::EmptyOp, f::FillZeroOpMaker); -REGISTER_OP(add, f::EmptyOp, f::AddOpMaker); -REGISTER_GRADIENT_OP(add, add_grad, f::EmptyOp); -REGISTER_OP(fc, f::FcOp, f::FcOpMaker); -REGISTER_OP(many_output_op, f::EmptyOp, f::ManyOutputOpMaker); -REGISTER_GRADIENT_OP(many_output_op, many_output_op_grad, f::EmptyOp); +REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad, + f::NOP); +REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP); +REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP); +REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker); +REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker); +REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP); +REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker); +REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad, + f::NOP); TEST(Backward, simple_op_grad) { - auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {}); + auto fwd = f::OpRegistry::CreateOp( + "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {}); ASSERT_NE(fwd, nullptr); auto gop = f::OpRegistry::CreateGradOp(*fwd); - ASSERT_EQ(4UL, gop->inputs_.size()); - ASSERT_EQ(f::kEmptyVarName, gop->inputs_[0]); - ASSERT_EQ("rowwise_add_grad", gop->type_); - ASSERT_EQ("X" + f::kGradVarSuffix, gop->outputs_[0]); - ASSERT_EQ("b" + f::kGradVarSuffix, gop->outputs_[1]); - - ASSERT_EQ("X" + f::kGradVarSuffix, gop->Output("X" + f::kGradVarSuffix)); + 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::kGradVarSuffix), - gop->outputs_.end()); + auto gop = f::Backward(*fwd, {"x"}); + ASSERT_EQ(gop->Output(f::GradVarName("X")), f::kEmptyVarName); - auto no_input_gop = f::Backward(*fwd, {"X", "b"}); + auto no_input_gop = f::Backward(*fwd, {"x", "b"}); ASSERT_NE(no_input_gop, nullptr); ASSERT_TRUE(no_input_gop->IsNetOp()); - ASSERT_EQ(0UL, - std::static_pointer_cast(no_input_gop)->ops_.size()); + ASSERT_EQ(0UL, static_cast(no_input_gop.get())->ops_.size()); } TEST(Backward, net_fc_backward_normal) { - std::shared_ptr fwd = f::OpRegistry::CreateOp( - "fc", {"X", "w", "b"}, {"mul_result", "add_result", "out"}, {}); + std::shared_ptr fwd = + f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}}, + {{"mul_result", {"mul_res"}}, + {"add_result", {"add_re"}}, + {"Out", {"out"}}}, + {}); ASSERT_NE(fwd, nullptr); std::shared_ptr gop = f::Backward(*fwd, {}); ASSERT_TRUE(gop->IsNetOp()); @@ -205,19 +200,22 @@ TEST(Backward, net_fc_backward_normal) { ASSERT_EQ(3UL, net->ops_.size()); f::OperatorBase &d_sigmoid = *net->ops_[0]; - ASSERT_EQ("sigmoid_grad", d_sigmoid.type_); + ASSERT_EQ("sigmoid_grad", d_sigmoid.Type()); f::OperatorBase &d_add = *net->ops_[1]; - ASSERT_EQ("rowwise_add_grad", d_add.type_); + ASSERT_EQ("rowwise_add_grad", d_add.Type()); f::OperatorBase &d_mul = *net->ops_[2]; - ASSERT_EQ("mul_grad", d_mul.type_); + ASSERT_EQ("mul_grad", d_mul.Type()); } TEST(Backward, net_fc_backward_not_have_b) { std::shared_ptr fwd = - f::OpRegistry::CreateOp("fc", {"X", "w", f::kEmptyVarName}, - {"mul_result", "add_result", "tmp"}, {}); + f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {}}}, + {{"mul_result", {"mul_res"}}, + {"add_result", {"add_res"}}, + {"Out", {"tmp"}}}, + {}); ASSERT_NE(fwd, nullptr); std::shared_ptr gop = f::Backward(*fwd, {}); ASSERT_TRUE(gop->IsNetOp()); @@ -228,142 +226,176 @@ 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.AppendOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"x"}}, {"W", {"W1"}}, {"b", {"b1"}}}, + {{"mul_result", {"mul_tmp_0"}}, + {"add_result", {"add_tmp_0"}}, + {"Out", {"hidden0"}}}, + {})); + net.AppendOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"hidden0"}}, {"W", {"W2"}}, {"b", {"b2"}}}, + {{"mul_result", {"mul_tmp_1"}}, + {"add_result", {"add_tmp_1"}}, + {"Out", {"hidden1"}}}, + {})); net.CompleteAddOp(); - auto bwd = Backward(net, {"X"}); // X@GRAD is not need. + auto bwd = Backward(net, {"x"}); // x@GRAD is not need. ASSERT_TRUE(bwd->IsNetOp()); auto bwd_net = static_cast(bwd.get()); - std::unordered_set all_output = std::unordered_set( - bwd_net->outputs_.begin(), bwd_net->outputs_.end()); - all_output.erase(f::kEmptyVarName); + auto output_vars = bwd_net->OutputVars(true); + std::unordered_set all_outputs = + std::unordered_set(output_vars.begin(), output_vars.end()); + all_outputs.erase(f::kEmptyVarName); for (auto &out : {"W1", "b1", "hidden0", "W2", "b2"}) { - ASSERT_NE(all_output.find(out + f::kGradVarSuffix), all_output.end()); + ASSERT_NE(all_outputs.find(f::GradVarName(out)), all_outputs.end()); } // Not Generated X - ASSERT_EQ(all_output.find("X" + f::kGradVarSuffix), all_output.end()); + ASSERT_EQ(all_outputs.find(f::GradVarName("X")), all_outputs.end()); ASSERT_EQ(2UL, bwd_net->ops_.size()); ASSERT_TRUE(bwd_net->ops_[1]->IsNetOp()); auto first_fc_grad = static_cast(bwd_net->ops_[1].get()); ASSERT_EQ(3UL, first_fc_grad->ops_.size()); ASSERT_EQ(f::kEmptyVarName, - first_fc_grad->ops_[2]->Output("A" + f::kGradVarSuffix)); + first_fc_grad->ops_[2]->Output(f::GradVarName("X"))); } TEST(Backward, net_shared_weight) { ops::NetOp net; - net.AddOp(f::OpRegistry::CreateOp("mul", {"X", "W"}, {"Out"}, {})); - net.AddOp(f::OpRegistry::CreateOp("mul", {"Out", "W"}, {"FinalOut"}, {})); + net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}}, + {{"Out", {"out"}}}, {})); + net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}}, + {{"Out", {"FinalOut"}}}, {})); net.CompleteAddOp(); auto bwd = f::Backward(net, {}); ASSERT_TRUE(bwd->IsNetOp()); auto bwd_net = static_cast(bwd.get()); ASSERT_EQ(3UL, bwd_net->ops_.size()); - ASSERT_EQ("add", bwd_net->ops_[2]->type_); + ASSERT_EQ("add", bwd_net->ops_[2]->Type()); } TEST(Backward, op_register_grad_not_for_network) { - auto fwd = f::OpRegistry::CreateOp( - "fc", {"X", "W", "b"}, {"mul_out", "add_out", "out1"}, - {{"temporary_index", std::vector{0, 1}}}); + auto fwd = + f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}}, + {{"mul_result", {"mul_out"}}, + {"add_result", {"add_out"}}, + {"Out", {"out1"}}}, + {{"temporary_index", std::vector{0, 1}}}); ASSERT_THROW(f::OpRegistry::CreateGradOp(*fwd), EnforceNotMet); } TEST(Backward, op_all_input_are_not_need) { - auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {}); - auto backward = f::Backward(*fwd, {"X", "b"}); + auto fwd = f::OpRegistry::CreateOp( + "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {}); + auto backward = f::Backward(*fwd, {"x", "b"}); ASSERT_TRUE(backward->IsNetOp()); auto net = static_cast(backward.get()); ASSERT_TRUE(net->ops_.empty()); } TEST(Backward, op_all_output_are_not_need) { - auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {}); - auto backward = f::Backward(*fwd, {"Out"}); + auto fwd = f::OpRegistry::CreateOp( + "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {}); + auto backward = f::Backward(*fwd, {"out"}); ASSERT_TRUE(backward->IsNetOp()); auto net = static_cast(backward.get()); ASSERT_TRUE(net->ops_.empty()); } TEST(Backward, op_part_of_output_are_not_need) { - auto fwd = f::OpRegistry::CreateOp("many_output_op", {"X"}, {"Y", "Z"}, {}); + auto fwd = f::OpRegistry::CreateOp("many_output_op", {{"x", {"X"}}}, + {{"y", {"Y"}}, {"z", {"Z"}}}, {}); auto backward = f::Backward(*fwd, {"Z"}); ASSERT_TRUE(backward->IsNetOp()); auto net = static_cast(backward.get()); ASSERT_EQ(net->ops_.size(), 2UL); auto &fill_zero = *net->ops_[0]; - ASSERT_EQ("fill_zeros_like", fill_zero.type_); - ASSERT_EQ(1UL, fill_zero.inputs_.size()); - ASSERT_EQ("Z", fill_zero.inputs_[0]); - ASSERT_EQ(1UL, fill_zero.outputs_.size()); - ASSERT_EQ("Z" + f::kZeroVarSuffix, fill_zero.outputs_[0]); + ASSERT_EQ("fill_zeros_like", fill_zero.Type()); + ASSERT_EQ(1UL, fill_zero.Inputs("Src").size()); + ASSERT_EQ("Z", fill_zero.Input("Src")); + ASSERT_EQ(1UL, fill_zero.Outputs("Dst").size()); + ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Dst")); auto &d_many_out = *net->ops_[1]; - ASSERT_EQ("many_output_op_grad", d_many_out.type_); - ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.inputs_.size()); // I/O/OG - ASSERT_EQ("Z" + f::kZeroVarSuffix, d_many_out.Input("z" + f::kGradVarSuffix)); - ASSERT_EQ("Y" + f::kGradVarSuffix, d_many_out.Input("y" + f::kGradVarSuffix)); - ASSERT_EQ("X" + f::kGradVarSuffix, - d_many_out.Output("x" + f::kGradVarSuffix)); + 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::kGradVarSuffix), f::kEmptyVarName); - ASSERT_EQ(grad_mul.Output("B" + f::kGradVarSuffix), "b" + f::kGradVarSuffix); - ASSERT_EQ(grad_mul.Input("Out" + f::kGradVarSuffix), - "out" + f::kGradVarSuffix); - 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.AppendOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"x1"}}, {"W", {"w1"}}, {"b", {"b1"}}}, + {{"mul_result", {"mul_out1"}}, + {"add_result", {"add_out1"}}, + {"Out", {"out1"}}}, + {})); + net.AppendOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"out1"}}, {"W", {"w2"}}, {"b", {"b2"}}}, + {{"mul_result", {"mul_out2"}}, + {"add_result", {"tmp_out2"}}, + {"Out", {"out2"}}}, + {})); + net.AppendOp(f::OpRegistry::CreateOp( + "fc", {{"X", {"out2"}}, {"W", {"w3"}}, {"b", {"b3"}}}, + {{"mul_result", {"mul_out3"}}, + {"add_result", {"tmp_out3"}}, + {"Out", {"out3"}}}, + {})); net.CompleteAddOp(); + auto backward = f::Backward(net, {"mul_out2", "tmp_out2", "out2"}); ASSERT_TRUE(backward->IsNetOp()); auto bwd_net = static_cast(backward.get()); ASSERT_EQ(bwd_net->ops_.size(), 3UL); auto &grad_fc = *bwd_net->ops_[0]; - EXPECT_EQ(grad_fc.inputs_.size(), - 3UL /* external input number */ + + const char *all = paddle::operators::NetOp::kAll; + EXPECT_EQ(grad_fc.Inputs(all).size(), + 2UL /* external input number */ + 1UL /* external output number*/ + 1UL /* number of gradient of external output*/ + 2U /* internal variable number*/); - EXPECT_EQ(grad_fc.outputs_.size(), 2UL /* input number of mul*/ - + 2UL /* input number of rowwise_add */ - + 1UL /* input number of sigmod */); - EXPECT_EQ(bwd_net->ops_[1]->inputs_.size(), 0UL); - EXPECT_EQ(bwd_net->ops_[1]->outputs_.size(), 0UL); - EXPECT_EQ(bwd_net->ops_[2]->inputs_.size(), 0UL); - EXPECT_EQ(bwd_net->ops_[2]->outputs_.size(), 0UL); + EXPECT_EQ(grad_fc.Outputs(all).size(), + 2UL /* input number of mul*/ + + 2UL /* input number of rowwise_add + */ + + 1UL /* input number of sigmod */); + EXPECT_EQ(bwd_net->ops_[1]->Inputs(all).size(), 0UL); + EXPECT_EQ(bwd_net->ops_[1]->Outputs(all).size(), 0UL); + EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL); + EXPECT_EQ(bwd_net->ops_[2]->Outputs(all).size(), 0UL); } diff --git a/paddle/framework/ddim.cc b/paddle/framework/ddim.cc index 545c1dcc2a..cfd3e8dfde 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 5aa5af0c19..95f294b627 100644 --- a/paddle/framework/ddim.h +++ b/paddle/framework/ddim.h @@ -14,13 +14,12 @@ 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 { diff --git a/paddle/framework/details/lod_tensor.cc b/paddle/framework/details/lod_tensor.cc deleted file mode 100644 index 9ad3979e5b..0000000000 --- a/paddle/framework/details/lod_tensor.cc +++ /dev/null @@ -1,62 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ - -#include "paddle/framework/lod_tensor.h" - -#include - -namespace paddle { -namespace framework { -namespace details { - -using LOD = LODTensor::LOD; - -std::shared_ptr SliceLOD(const LOD &lod, size_t level_begin, - size_t level_end) { - auto new_lod = std::make_shared(); - new_lod->reserve(level_end - level_begin); - for (size_t i = level_begin; i < level_end; i++) { - new_lod->emplace_back(lod[i]); - } - return new_lod; -} - -std::shared_ptr SliceLOD(const LOD &lod, size_t level, size_t elem_begin, - size_t elem_end, bool tensor_shared) { - // slice the lod. - auto new_lod = std::make_shared(); - new_lod->reserve(lod.size() - level); - auto start = lod.at(level)[elem_begin]; - auto end = lod.at(level)[elem_end]; - - for (auto it = lod.begin() + level; it != lod.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); - if (!tensor_shared) { - // 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(new_lod->back().front() == 0, "error in slice LOD"); - } - } - return new_lod; -} - -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/framework/details/lod_tensor.h b/paddle/framework/details/lod_tensor.h deleted file mode 100644 index 9a6a6cd2ea..0000000000 --- a/paddle/framework/details/lod_tensor.h +++ /dev/null @@ -1,46 +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 - -namespace paddle { -namespace framework { -namespace details { - -/* - * Slice levels from LOD. - * - * @lod: LOD to slice. - * @level_begin: level to begin slice. - * @level_end: level to end slice. - */ -std::shared_ptr SliceLOD(const LODTensor::LOD &lod, - size_t level_begin, size_t level_end); - -/* - * Slice elements from a level of LOD. - * - * @lod: LOD to slice. - * @level: which level to slice. - * @elem_begin: element's index to begin slice. - * @elem_end: element's index to end slice. - */ -std::shared_ptr SliceLOD(const LODTensor::LOD &lod, - size_t level, size_t elem_begin, - size_t elem_end, bool tensor_shared); -} // namespace details -} // namespace framework -} // namespace paddle diff --git a/paddle/framework/framework.proto b/paddle/framework/framework.proto new file mode 100644 index 0000000000..ae44a1ffd4 --- /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 not_in_gradient = 5 [ default = false ]; + } + + // AttrProto describes the C++ type Attribute. + message Attr { + required string name = 1; + required AttrType type = 2; + required string comment = 3; + // If that attribute is generated, it means the Paddle third + // language binding has responsibility to fill that + // attribute. End-User should not set that attribute. + optional bool generated = 4 [ default = false ]; + } + + required string type = 1; + repeated Var inputs = 2; + repeated Var outputs = 3; + repeated Attr attrs = 4; + required string comment = 5; +} diff --git a/paddle/framework/grad_op_builder.cc b/paddle/framework/grad_op_builder.cc index 6d032fb78f..b02a599a80 100644 --- a/paddle/framework/grad_op_builder.cc +++ b/paddle/framework/grad_op_builder.cc @@ -13,90 +13,45 @@ express or implied. See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/framework/grad_op_builder.h" -#include "paddle/framework/op_proto.pb.h" #include "paddle/framework/op_registry.h" namespace paddle { namespace framework { - -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, VariableNameMap* vars) { + const auto& src_inout = + src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs(); + auto& dst_inout = *vars; + auto& proto = OpInfoMap::Instance().Get(src_op->Type()).Proto(); const auto& src_arg_list = src_type == OpArgType::IN ? proto.inputs() : proto.outputs(); - for (const auto& arg : src_arg_list) { - std::string src_name = arg.name(); - std::string dst_name = is_grad ? src_name + kGradVarSuffix : 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] + kGradVarSuffix - : (arg.ignore_gradient() ? kEmptyVarName : src_inout[i]); - dst_inout.emplace_back(s); - } - if (dst_format != nullptr) { - dst_format->push_back(dst_inout.size()); + if (arg.not_in_gradient() && !is_grad) continue; + const std::string src_name = arg.name(); + std::string dst_name = is_grad ? GradVarName(src_name) : src_name; + dst_inout[dst_name].reserve(src_inout.at(src_name).size()); + for (auto& var_name : src_inout.at(src_name)) { + std::string s = is_grad ? GradVarName(var_name) : var_name; + dst_inout[dst_name].emplace_back(s); } } } OperatorBase* BuildGradOp(const OperatorBase* op) { - 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& info = OpInfoMap::Instance().Get(op->Type()); + PADDLE_ENFORCE(info.HasGradientOp()); + + VariableNameMap inputs; + VariableNameMap outputs; + TransOpArg(op, OpArgType::IN, false, &inputs); // I + TransOpArg(op, OpArgType::OUT, false, &inputs); // O + TransOpArg(op, OpArgType::OUT, true, &inputs); // OG + TransOpArg(op, OpArgType::IN, true, &outputs); // IG + + auto& grad_info = OpInfoMap::Instance().Get(info.grad_op_type_); + return grad_info.Creator()(info.grad_op_type_, inputs, outputs, op->Attrs()); } } // namespace framework diff --git a/paddle/framework/grad_op_builder_test.cc b/paddle/framework/grad_op_builder_test.cc index cf7143eba4..902c2655e9 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().NotInGradient(); + AddInput("In3_mult", "another multiple input").AsDuplicable(); + AddOutput("Out1_mult", "a multiple output").AsDuplicable(); + AddOutput("Out2", "a single output").NotInGradient(); AddComment("op with inputs and outputs ignored in gradient calculating"); } }; @@ -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"})); @@ -83,55 +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::kGradVarSuffix), - "out1" + f::kGradVarSuffix); - EXPECT_EQ(grad_test_op->Inputs("Out2_mult" + f::kGradVarSuffix), + EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out1")), + f::GradVarName("out1")); + EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out2_mult")), std::vector( - {"out2_1" + f::kGradVarSuffix, "out2_2" + f::kGradVarSuffix})); + {f::GradVarName("out2_1"), f::GradVarName("out2_2")})); - ASSERT_EQ(grad_test_op->outputs_.size(), 5UL); - EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix), - "in1" + f::kGradVarSuffix); - EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix), - std::vector({"in2_1" + f::kGradVarSuffix, - "in2_2" + f::kGradVarSuffix, - "in2_3" + f::kGradVarSuffix})); - EXPECT_EQ(grad_test_op->Output("In3" + f::kGradVarSuffix), - "in3" + f::kGradVarSuffix); + 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::kEmptyVarName, f::kEmptyVarName})); EXPECT_EQ(grad_test_op->Inputs("In3_mult"), std::vector({"in3_1", "in3_2"})); EXPECT_EQ(grad_test_op->Inputs("Out1_mult"), std::vector({"out1_1", "out1_2"})); - EXPECT_EQ(grad_test_op->Input("Out2"), f::kEmptyVarName); - EXPECT_EQ(grad_test_op->Inputs("Out1_mult" + f::kGradVarSuffix), + EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out1_mult")), std::vector( - {"out1_1" + f::kGradVarSuffix, "out1_2" + f::kGradVarSuffix})); - EXPECT_EQ(grad_test_op->Input("Out2" + f::kGradVarSuffix), - "out2" + f::kGradVarSuffix); + {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::kGradVarSuffix), - "in1" + f::kGradVarSuffix); - EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix), + 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( - {"in2_1" + f::kGradVarSuffix, "in2_2" + f::kGradVarSuffix})); - EXPECT_EQ(grad_test_op->Outputs("In3_mult" + f::kGradVarSuffix), + {f::GradVarName("in2_1"), f::GradVarName("in2_2")})); + EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In3_mult")), std::vector( - {"in3_1" + f::kGradVarSuffix, "in3_2" + f::kGradVarSuffix})); + {f::GradVarName("in3_1"), f::GradVarName("in3_2")})); } diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc index 70045dbf7a..2b17890774 100644 --- a/paddle/framework/lod_tensor.cc +++ b/paddle/framework/lod_tensor.cc @@ -19,32 +19,59 @@ namespace paddle { namespace framework { -LODTensor LODTensor::SliceShared(size_t level_begin, size_t level_end) const { - PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced."); - auto new_lod = details::SliceLOD(*lod_start_pos_, 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_. - return LODTensor(tensor_, new_lod); +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 LODTensor::SliceShared(size_t level, size_t elem_begin, - size_t elem_end) const { - PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced."); - 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 = details::SliceLOD(*lod_start_pos_, level, elem_begin, elem_end, - true /*tensor_shared*/); - - // slice elements just need to update LOD info, because offsets are not - // changed, so the original tensor_ can be reused. - return LODTensor(tensor_, 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 diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h index 4933479b10..9e27aec38d 100644 --- a/paddle/framework/lod_tensor.h +++ b/paddle/framework/lod_tensor.h @@ -15,7 +15,7 @@ #pragma once #include -#if (!PADDLE_ONLY_CPU) +#if !defined(PADDLE_ONLY_CPU) #include #include #endif @@ -31,30 +31,29 @@ namespace framework { * LODTensor (Level of details Tensor) * see https://en.wikipedia.org/wiki/Level_of_details for reference. */ -class LODTensor { +class LODTensor : public Tensor { public: // Level save offsets of each unit. #ifdef PADDLE_ONLY_CPU - using Level = std::vector; + template + using Vector = std::vector; #else - using Level = thrust::device_vector; + template + using Vector = thrust::host_vector; #endif - // LOD stores offsets of each level of units, the largest units level first, + // 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. - typedef std::vector LOD; + class LOD : public std::vector> { + public: + LOD SliceLevels(size_t level_begin, size_t level_end) const; + LOD SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) const; + }; LODTensor() {} - LODTensor(const std::shared_ptr &tensor, - const std::shared_ptr &lod) { - Reset(tensor, lod); - } + explicit LODTensor(const LOD &lod) : lod_(lod) {} - void Reset(const std::shared_ptr &tensor, - const std::shared_ptr &lod) { - tensor_ = tensor; - lod_start_pos_ = lod; - } + virtual Tensor *Clone() const { return new LODTensor(lod_); } /* * Get a element from LOD. @@ -65,16 +64,14 @@ class LODTensor { PADDLE_ENFORCE(elem < NumElements(level), "element begin [%d] out of range [%d]", elem, NumElements(level)); - return (*lod_start_pos_)[level][elem]; + 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_start_pos_ ? lod_start_pos_->size() : 0UL; - } + size_t NumLevels() const { return lod_.size(); } /* * Number of elements in a level. */ @@ -82,64 +79,71 @@ class LODTensor { PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level, NumLevels()); // the last offset is the end of last element - return lod_start_pos_->at(level).size() - 1; + return lod_[level].size() - 1; } - /* - * Slice of levels[level_begin:level_end], with tensor copied. - */ - template - LODTensor SliceCopied(size_t level_begin, size_t level_end, - const platform::Place &dst_place) const; - /* * Slice of levels[level_begin:level_end], with tensor shared. */ - LODTensor SliceShared(size_t level_begin, size_t level_end) const; - - /* - * Slice of elements of a level, [elem_begin: elem_end], with tensor copied. - * @note: low performance in slice lod_start_pos_. - */ template - LODTensor SliceCopied(size_t level, size_t elem_begin, size_t elem_end, - const platform::Place &dst_place) const; + 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_start_pos_. - */ - LODTensor SliceShared(size_t level, size_t elem_begin, size_t elem_end) const; - - /* - * Copy other's lod_start_pos_, to share LOD info. - * @note: the LOD info should not be changed. + * @note: low performance in slice lod_. */ - void ShareLOD(const LODTensor &other) { - lod_start_pos_ = other.lod_start_pos_; - } + template + LODTensor SliceInLevel(size_t level, size_t elem_begin, + size_t elem_end) const; /* - * Copy other's lod_start_pos_'s content, free to mutate. + * Copy other's lod_'s content, free to mutate. */ - void CopyLOD(const LODTensor &other) { - lod_start_pos_ = std::make_shared(*other.lod_start_pos_); - } + void CopyLOD(const LODTensor &other) { lod_ = other.lod_; } /* * Determine whether LODTensor has a valid LOD info. */ - bool HasLOD() const { return bool(lod_start_pos_); } - LOD *lod() const { return lod_start_pos_.get(); } + const LOD &lod() const { return lod_; } + LOD *mutable_lod() { return &lod_; } - std::shared_ptr &tensor() { return tensor_; } - Tensor *raw_tensor() { return tensor_.get(); } + virtual ~LODTensor() {} private: - std::shared_ptr lod_start_pos_; - std::shared_ptr tensor_; + 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 - -#include "paddle/framework/lod_tensor_impl.h" diff --git a/paddle/framework/lod_tensor_impl.h b/paddle/framework/lod_tensor_impl.h deleted file mode 100644 index 0eb6469aea..0000000000 --- a/paddle/framework/lod_tensor_impl.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/details/lod_tensor.h" - -namespace paddle { -namespace framework { - -template -LODTensor LODTensor::SliceCopied(size_t level_begin, size_t level_end, - const platform::Place &dst_place) const { - PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced."); - auto new_lod = details::SliceLOD(*lod_start_pos_, level_begin, level_end); - auto new_tensor = std::make_shared(); - new_tensor->CopyFrom(*tensor_, dst_place); - - return LODTensor(new_tensor, new_lod); -} - -template -LODTensor LODTensor::SliceCopied(size_t level, size_t elem_begin, - size_t elem_end, - const platform::Place &dst_place) const { - PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced."); - 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 = details::SliceLOD(*lod_start_pos_, level, elem_begin, elem_end, - false /*tensor_shared*/); - - auto start_idx = new_lod->front().front(); - auto end_idx = new_lod->front().back() - 1 /*the next element's start*/; - auto sliced_tensor = tensor_->Slice(start_idx, end_idx); - auto new_tensor = std::make_shared(); - new_tensor->CopyFrom(sliced_tensor, dst_place); - - return LODTensor(new_tensor, new_lod); -} - -} // namespace framework -} // namespace paddle diff --git a/paddle/framework/lod_tensor_test.cc b/paddle/framework/lod_tensor_test.cc index 511716375e..2881136ced 100644 --- a/paddle/framework/lod_tensor_test.cc +++ b/paddle/framework/lod_tensor_test.cc @@ -15,6 +15,7 @@ #include #include +#include #include namespace paddle { @@ -29,22 +30,28 @@ class LODTensorTester : public ::testing::Test { // 0 10 20 // 0 5 10 15 20 // 0 2 5 7 10 12 15 20 - auto lod = std::make_shared(); - 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}); + 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}); - auto tensor = std::make_shared(); - tensor->Resize({20 /*batch size*/, 128 /*dim*/}); + ASSERT_EQ(lod.size(), 3UL); + + tensor.Resize({20 /*batch size*/, 128 /*dim*/}); // malloc memory - tensor->mutable_data(place); + tensor.mutable_data(place); + + lod_tensor.reset(new LODTensor(lod)); + lod_tensor->Resize({20 /*batch size*/, 128 /*dim*/}); - lod_tensor->Reset(tensor, lod); + 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); } @@ -55,110 +62,54 @@ TEST_F(LODTensorTester, NumElements) { ASSERT_EQ(lod_tensor->NumElements(2), 8UL); } -TEST_F(LODTensorTester, SliceShared_Level) { - // slice 1 level - for (size_t level = 0; level < 3UL; ++level) { - auto new_lod_tensor = lod_tensor->SliceShared(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.tensor(), lod_tensor->tensor()); - } - // slice 2 level - for (size_t level = 0; level < 2UL; ++level) { - auto new_lod_tensor = lod_tensor->SliceShared(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.tensor(), lod_tensor->tensor()); - } -} - -TEST_F(LODTensorTester, SliceCopied_Level) { +TEST_F(LODTensorTester, SliceLevels) { // slice 1 level for (size_t level = 0; level < 3UL; ++level) { - auto new_lod_tensor = - lod_tensor->SliceCopied(level, level + 1, place); + 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.tensor(), lod_tensor->tensor()); - // TODO(superjom) add tensor comparation here. + // ASSERT_EQ(new_lod_tensor, *lod_tensor); } // slice 2 level for (size_t level = 0; level < 2UL; ++level) { - auto new_lod_tensor = - lod_tensor->SliceCopied(level, level + 2, place); + 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.tensor(), lod_tensor->tensor()); - // TODO(superjom) add tensor comparation here. + ASSERT_EQ(new_lod_tensor.data(), lod_tensor->data()); } } -TEST_F(LODTensorTester, SliceShared_Element) { - size_t level = 0; - auto new_lod_tensor = lod_tensor->SliceShared(level, 0, 2); - ASSERT_EQ(new_lod_tensor.NumLevels(), 3UL); - ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL); - ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL); - ASSERT_EQ(new_lod_tensor.NumElements(2), 8UL); - ASSERT_EQ(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor()); - - level = 1; - new_lod_tensor = lod_tensor->SliceShared(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.raw_tensor(), lod_tensor->raw_tensor()); -} - -TEST_F(LODTensorTester, SliceCopied_Element) { +TEST_F(LODTensorTester, SliceInLevel) { size_t level = 0; - auto new_lod_tensor = lod_tensor->SliceCopied(level, 0, 2, place); - ASSERT_EQ(new_lod_tensor.NumLevels(), 3UL); - ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL); - ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL); - ASSERT_EQ(new_lod_tensor.NumElements(2), 8UL); - ASSERT_NE(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor()); + 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->SliceCopied(level, 0, 2, place); + 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_NE(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor()); - - level = 1; - // LOD is - // 0 5 10 - // 0 2 5 7 10 - new_lod_tensor = lod_tensor->SliceCopied(level, 1, 3, place); - 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.lod_element(0, 0), 0UL); - ASSERT_EQ(new_lod_tensor.lod_element(0, 1), 5UL); - ASSERT_EQ(new_lod_tensor.lod_element(1, 0), 0UL); - ASSERT_EQ(new_lod_tensor.lod_element(1, 1), 2UL); - ASSERT_EQ(new_lod_tensor.lod_element(1, 2), 5UL); - ASSERT_EQ(new_lod_tensor.lod_element(1, 3), 7UL); - - // TODO(superjom) compare the content of these tensors + ASSERT_EQ(new_lod_tensor.data(), lod_tensor->data()); } TEST_F(LODTensorTester, ShareLOD) { LODTensor new_lod_tensor; - new_lod_tensor.ShareLOD(*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); - ASSERT_NE(new_lod_tensor.lod(), lod_tensor->lod()); + bool equals = std::equal(lod_tensor->lod().begin(), lod_tensor->lod().end(), + new_lod_tensor.lod().begin()); + ASSERT_TRUE(equals); } } // namespace framework diff --git a/paddle/framework/op_desc.proto b/paddle/framework/op_desc.proto deleted file mode 100644 index d95ba26f88..0000000000 --- a/paddle/framework/op_desc.proto +++ /dev/null @@ -1,56 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -syntax = "proto2"; -package paddle.framework; - -import "attribute.proto"; - -// AttrDesc is used to describe Attributes of an Operator. It contain's -// name, type, and value of Attribute. -// -// e.g, for scale=3.0: name=scala, type=AttrType.FLOAT, value=3.0 -message AttrDesc { - required string name = 1; - required AttrType type = 2; - optional int32 i = 3; - optional float f = 4; - optional string s = 5; - repeated int32 ints = 6; - repeated float floats = 7; - repeated string strings = 8; -}; - -// Protocol Message to describe an Operator. -// -// In PaddlePaddle, Operator is used to do a certain computation such -// as "add", "sub", "cosine", etc. -// (1) Operator needs to know the input and output variable names. -// (2) Some ops may have special attributes such as "scale" in "CosineOp". -// -// 3rd-party language can build this proto message and call -// AddOp(const OpDesc& op_desc) of Paddle core to create an Operator. -message OpDesc { - // input names of this Operator. - repeated string inputs = 1; - - // output names of this Operator. - repeated string outputs = 2; - - // type of this Operator, such as "add", "sub", "fc". - required string type = 3; - - // Attributes of this Operator. e.g., scale=3.0 in cosine op. - repeated AttrDesc attrs = 4; -}; \ No newline at end of file diff --git a/paddle/framework/op_desc_test.cc b/paddle/framework/op_desc_test.cc deleted file mode 100644 index d0c52523b6..0000000000 --- a/paddle/framework/op_desc_test.cc +++ /dev/null @@ -1,35 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include - -TEST(OpDesc, Create) { - paddle::framework::OpDesc op_desc; - op_desc.set_type("add"); - op_desc.add_inputs("X"); - op_desc.add_inputs("Y"); - op_desc.add_outputs("Z"); - - auto attr = op_desc.mutable_attrs()->Add(); - attr->set_type(paddle::framework::AttrType::FLOAT); - attr->set_f(3.14); - - // required field name is not set, so IsInitialized should be false. - ASSERT_FALSE(op_desc.IsInitialized()); - - attr->set_name("add"); - // after all required fields are set, IsInitialized should be true now. - ASSERT_TRUE(op_desc.IsInitialized()); -} \ No newline at end of file diff --git a/paddle/framework/op_info.cc b/paddle/framework/op_info.cc new file mode 100644 index 0000000000..81ba29797c --- /dev/null +++ b/paddle/framework/op_info.cc @@ -0,0 +1,29 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/framework/op_info.h" + +namespace paddle { +namespace framework { + +static OpInfoMap* g_op_info_map = nullptr; + +OpInfoMap& OpInfoMap::Instance() { + if (g_op_info_map == nullptr) { + g_op_info_map = new OpInfoMap(); + } + return *g_op_info_map; +} +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/op_info.h b/paddle/framework/op_info.h new file mode 100644 index 0000000000..94245c6c44 --- /dev/null +++ b/paddle/framework/op_info.h @@ -0,0 +1,101 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include +#include +#include +#include + +#include "paddle/framework/attribute.h" + +namespace paddle { +namespace framework { +class OperatorBase; +using VariableNameMap = std::map>; + +using OpCreator = std::function; + +struct OpInfo { + OpCreator creator_; + std::string grad_op_type_; + OpProto* proto_; + OpAttrChecker* checker_; + + bool HasOpProtoAndChecker() const { + return proto_ != nullptr && checker_ != nullptr; + } + + const OpProto& Proto() const { + PADDLE_ENFORCE_NOT_NULL(proto_, "Operator Proto has not been registered"); + PADDLE_ENFORCE(proto_->IsInitialized(), + "Operator Proto must be initialized in op info"); + return *proto_; + } + + const OpAttrChecker& Checker() const { + PADDLE_ENFORCE_NOT_NULL(checker_, + "Operator Checker has not been registered"); + return *checker_; + } + + const OpCreator& Creator() const { + PADDLE_ENFORCE_NOT_NULL(creator_, + "Operator Creator has not been registered"); + return creator_; + } + + bool HasGradientOp() const { return !grad_op_type_.empty(); } +}; + +class OpInfoMap { + public: + static OpInfoMap& Instance(); + + OpInfoMap(const OpInfoMap& o) = delete; + OpInfoMap(OpInfoMap&& o) = delete; + OpInfoMap& operator=(const OpInfoMap& o) = delete; + OpInfoMap& operator=(OpInfoMap&& o) = delete; + + bool Has(const std::string& op_type) const { + return map_.find(op_type) != map_.end(); + } + + void Insert(const std::string& type, const OpInfo& info) { + PADDLE_ENFORCE(!Has(type), "Operator %s has been registered", type); + map_.insert({type, info}); + } + + const OpInfo& Get(const std::string& type) const { + auto it = map_.find(type); + PADDLE_ENFORCE(it != map_.end(), "Operator %s are not found", type); + return it->second; + } + + template + void IterAllInfo(Callback callback) { + for (auto& it : map_) { + callback(it.first, it.second); + } + } + + private: + OpInfoMap() = default; + std::unordered_map map_; +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/op_proto.proto b/paddle/framework/op_proto.proto deleted file mode 100644 index 5229216287..0000000000 --- a/paddle/framework/op_proto.proto +++ /dev/null @@ -1,116 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -// Protocol Message for 3rd-party language binding. -// -// Paddle Python package will use `OpProto` to generate op creation methods. -// The op creation methods take user's input and generate `OpDesc` proto -// message, -// then pass `OpDesc` to C++ side and create Op pointer. -// -syntax = "proto2"; -package paddle.framework; - -import "attribute.proto"; - -// Attribute protocol message for 3rd-party language binding. -// It will store the Op support what attribute and what type. -message AttrProto { - // Supported attribute name. e.g. `scale` for cosine op. - required string name = 1; - - // Supported attribute type. - required AttrType type = 2; - - // Supported attribute comments. It helps 3rd-party language generate - // doc-string. - required string comment = 3; - - // If that attribute is generated, it means the Paddle third language - // binding has responsibility to fill that attribute. End-User should - // not set that attribute. - optional bool generated = 4 [ default = false ]; -} - -// Input or output message for 3rd-party language binding. -// It contains parameter name and its comments. -message VarProto { - // Input or output name in that op creation function. - // e.g. `cos(a, b, output, ...)`, "a", "b", "output" are names. - required string name = 1; - - // The comment for that input. It helps 3rd-party language generate - // doc-string. - required string comment = 2; - - // Is that input/output could be a list or not. - // If so, that Op should write a attributed named `input_format` or - // `output_format`. - // - // e.g. - // If the op is a fc op, the inputs are `X`, `W`, `b`. The `X` and `W` - // could be multiple, so the multiple of `X` and `W` is True, and OpDesc - // will hold a attribute of them. - // - // The Op desc of same fc could be - // { - // "type": "fc", - // "input": ["X1", "X2", "W1", "W2", "b"], - // "output": "fc.out", - // "attrs" : { - // "input_format": [0, 2, 4, 5] - // } - // } - // - optional bool multiple = 3 [ default = false ]; - - // It marks that output is a temporary output. That output is not used by - // user, but used by other op internally as input. If other op is not use - // that output, it could be optimized early. - // - // Attribute temporary_index will be set in OpDesc if there is some - // outputs are temporary. - // - // output = [ "xxx.out1", "xxx.tmp", "xxx.out2"], - // attrs = { - // "temporary_index": [1] - // } - optional bool temporary = 4 [ default = false ]; - - // The gradient of operator can be ignored immediately - // e.g. operator AddOp, y = x1 + x2, the gradient of dy/dx1, dy/dx2 - // can be ignored for the future optimized on graph. - optional bool ignore_gradient = 6; -} - -// Op protocol message for 3rd-party language binding. -// It contains all information for generating op creation method. -message OpProto { - // The input information to generate op creation method. - repeated VarProto inputs = 1; - - // The output information to generate op creation method. - repeated VarProto outputs = 2; - - // The attribute information to generate op creation method. - repeated AttrProto attrs = 3; - - // The comments for that Op. It helps 3rd-party language generate - // doc-string. The whole documentation of that Op is generated by comment, - // inputs, outputs, attrs together. - required string comment = 4; - - // The type of that Op. - required string type = 5; -} diff --git a/paddle/framework/op_proto_test.cc b/paddle/framework/op_proto_test.cc deleted file mode 100644 index 9c054bde44..0000000000 --- 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 1caa02a2a1..b0e85dd49f 100644 --- a/paddle/framework/op_registry.cc +++ b/paddle/framework/op_registry.cc @@ -17,5 +17,45 @@ limitations under the License. */ #include namespace paddle { -namespace framework {} // namespace framework +namespace framework { + +std::unique_ptr OpRegistry::CreateOp( + const std::string& type, const VariableNameMap& inputs, + const VariableNameMap& outputs, AttributeMap attrs) { + auto& info = OpInfoMap::Instance().Get(type); + info.Checker().Check(attrs); + auto op = info.Creator()(type, inputs, outputs, attrs); + return std::unique_ptr(op); +} + +static VariableNameMap ConvertOpDescVarsToVarNameMap( + const google::protobuf::RepeatedPtrField& op_desc_vars) { + VariableNameMap ret_val; + for (auto& var : op_desc_vars) { + auto& var_names = ret_val[var.parameter()]; + auto& var_names_in_proto = var.arguments(); + var_names.reserve(static_cast(var_names_in_proto.size())); + std::copy(var_names_in_proto.begin(), var_names_in_proto.end(), + std::back_inserter(var_names)); + } + return ret_val; +} + +std::unique_ptr OpRegistry::CreateOp(const OpDesc& op_desc) { + VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs()); + VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs()); + AttributeMap attrs; + for (auto& attr : op_desc.attrs()) { + attrs[attr.name()] = GetAttrValue(attr); + } + + return CreateOp(op_desc.type(), inputs, outputs, attrs); +} + +std::unique_ptr OpRegistry::CreateGradOp(const OperatorBase& op) { + PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops"); + return std::unique_ptr(BuildGradOp(&op)); +} + +} // namespace framework } // namespace paddle diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h index 6c26183818..2d09cde41e 100644 --- a/paddle/framework/op_registry.h +++ b/paddle/framework/op_registry.h @@ -17,309 +17,95 @@ limitations under the License. */ #include #include #include +#include #include #include #include "paddle/framework/attribute.h" +#include "paddle/framework/framework.pb.h" #include "paddle/framework/grad_op_builder.h" -#include "paddle/framework/op_desc.pb.h" +#include "paddle/framework/op_info.h" +#include "paddle/framework/operator.h" #include "paddle/framework/scope.h" namespace paddle { namespace framework { -// this class not only make proto but also init attribute checkers. -class OpProtoAndCheckerMaker { +class OpRegistry { public: - OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : proto_(proto), op_checker_(op_checker) {} - - ~OpProtoAndCheckerMaker() { - PADDLE_ENFORCE(validated_, "should call Validate after build"); - } - - void Validate() { - validated_ = true; - CheckNoDuplicatedInOutAttrs(); - } - - protected: - struct VariableBuilder { - VarProto* var_; - std::function on_multiple_; - std::function on_temporary_; - - VariableBuilder& SetMultiple() { - var_->set_multiple(true); - on_multiple_(); - return *this; - } - - VariableBuilder& SetTemporary() { - PADDLE_ENFORCE(bool(on_temporary_), "Cannot set temporary"); - var_->set_temporary(true); - on_temporary_(); - return *this; + template + static void RegisterOp(const std::string& op_type, + const std::string& grad_op_type) { + PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type), + "'%s' is registered more than once.", op_type); + OpInfo op_info; + op_info.creator_ = []( + const std::string& type, const VariableNameMap& inputs, + const VariableNameMap& outputs, const AttributeMap& attrs) { + return new OpType(type, inputs, outputs, attrs); + }; + op_info.grad_op_type_ = grad_op_type; + if (std::type_index(typeid(ProtoMakerType)) != + std::type_index(typeid(NOPMaker))) { + op_info.proto_ = new OpProto; + op_info.checker_ = new OpAttrChecker; + auto maker = ProtoMakerType(op_info.proto_, op_info.checker_); + maker.Validate(); + op_info.proto_->set_type(op_type); + PADDLE_ENFORCE( + op_info.proto_->IsInitialized(), + "Fail to initialize %s's OpProto, because %s is not initialized", + op_type, op_info.proto_->InitializationErrorString()); + } else { + op_info.proto_ = nullptr; + op_info.checker_ = nullptr; } - - VariableBuilder& IgnoreGradient() { - var_->set_ignore_gradient(true); - return *this; + OpInfoMap::Instance().Insert(op_type, op_info); + // register gradient op + if (!grad_op_type.empty()) { + RegisterOp(grad_op_type, ""); } - }; - - 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}; - } - - 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(); }}; } - 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; - attr->set_generated(generated); - attr->set_type(AttrTypeID()); - return op_checker_->AddAttrChecker(name); - } - - void AddComment(const std::string& comment) { - *(proto_->mutable_comment()) = comment; - } + static std::unique_ptr CreateOp(const std::string& type, + const VariableNameMap& inputs, + const VariableNameMap& outputs, + AttributeMap attrs); - 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; - } - } + static std::unique_ptr CreateOp(const OpDesc& op_desc); - void SetHasMultipleInput() { SetHasMultiple("input", &has_multiple_input_); } - void SetHasMultipleOutput() { - SetHasMultiple("output", &has_multiple_output_); - } - - void SetHasTemporaryOutput() { - if (!has_temporary_output_) { - AddAttr>("temporary_index", - R"DOC(The temporary index of output. - -Not all output of Paddle Op is used by user. For faster computation, each op -could output some its internal state to other op, other op could take that -output to make compute faster. - -Add a mark to which output is temporary is helpful for future optimization. -)DOC", - /*generated*/ true) - .SetDefault(std::vector()); - has_temporary_output_ = true; - } - } - - void CheckNoDuplicatedInOutAttrs() { - std::unordered_set names; - auto checker = [&](const std::string& name) { - PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name); - names.insert(name); - }; - for (auto& attr : proto_->attrs()) { - checker(attr.name()); - } - for (auto& input : proto_->inputs()) { - checker(input.name()); - } - for (auto& output : proto_->outputs()) { - checker(output.name()); - } - } - - OpProto* proto_; - OpAttrChecker* op_checker_; - bool validated_{false}; - bool has_multiple_input_{false}; - bool has_multiple_output_{false}; - bool has_temporary_output_{false}; + static std::unique_ptr CreateGradOp(const OperatorBase& op); }; -class OpRegistry { - using OpCreator = std::function; - using VarIndexMap = std::unordered_map; - using VarNameList = std::vector; - +class Registrar { 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++; - } - idx = 0; - for (auto& var : op_proto.outputs()) { - varmap[var.name()] = idx++; - } - } - - template - static void RegisterGradOp(const std::string& op_type, - const std::string& grad_op_type) { - op_creators()[grad_op_type] = [] { return new GradOpType; }; - grad_ops()[op_type] = grad_op_type; - } - - static std::shared_ptr CreateOp(const std::string& type, - const VarNameList& inputs, - const VarNameList& outputs, - const AttributeMap& attrs) { - auto op_create_it = op_creators().find(type); - PADDLE_ENFORCE(op_create_it != op_creators().end(), - "Operator %s cannot be found.", type); - - auto op = op_create_it->second(); - op->type_ = type; - op->inputs_ = inputs; - op->outputs_ = outputs; - - op->attrs_ = attrs; - op_checkers().at(type).Check(op->attrs_); - - GenerateTempVariableName(op); - - { - auto var_index_it = VarIndexMaps().find(type); - if (var_index_it != VarIndexMaps().end()) { - op->in_out_idxs_ = var_index_it->second; - } - } - - op->Init(); - return std::shared_ptr(op); - } - - static std::shared_ptr CreateOp(const OpDesc& op_desc) { - std::vector inputs; - inputs.reserve((size_t)op_desc.inputs_size()); - std::copy(op_desc.inputs().begin(), op_desc.inputs().end(), - std::back_inserter(inputs)); - - std::vector outputs; - outputs.reserve((size_t)op_desc.outputs_size()); - std::copy(op_desc.outputs().begin(), op_desc.outputs().end(), - std::back_inserter(outputs)); - - AttributeMap attrs; - for (auto& attr : op_desc.attrs()) { - attrs[attr.name()] = GetAttrValue(attr); - } - - return CreateOp(op_desc.type(), inputs, outputs, attrs); - } - - static std::shared_ptr CreateGradOp(const OperatorBase& op) { - PADDLE_ENFORCE(!op.IsNetOp(), - "Use framework::Backward to get backward ops"); - std::shared_ptr grad_op(BuildGradOp(&op)); - grad_op->Init(); - return grad_op; - } - - static std::unordered_map& protos() { - static std::unordered_map protos_; - return protos_; - } - - static std::unordered_map& grad_ops() { - static std::unordered_map grad_ops_; - return grad_ops_; - } - - static std::unordered_map>& - VarIndexMaps() { - static std::unordered_map> maps_; - return maps_; - } - - static std::unordered_map& op_creators() { - static std::unordered_map op_creators_; - return op_creators_; - } - - 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 == kTempVarName) { - outname += op->type_; - outname += "@"; - outname += std::to_string(gUniqId.fetch_add(1)); - } - } - } + // 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: - explicit 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); } }; @@ -333,97 +119,93 @@ 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; } +#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, \ + grad_op_class) \ + STATIC_ASSERT_GLOBAL_NAMESPACE( \ + __reg_op__##op_type, "REGISTER_OP must be called in global namespace"); \ + class _OpClass_##op_type##_ : public op_class { \ + public: \ + DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_); \ + DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class); \ + }; \ + class _OpGradClass_##op_type##_ : public grad_op_class { \ + public: \ + DEFINE_OP_CLONE_METHOD(_OpGradClass_##op_type##_); \ + DEFINE_OP_CONSTRUCTOR(_OpGradClass_##op_type##_, grad_op_class); \ + }; \ + static ::paddle::framework::OpRegistrar< \ + _OpClass_##op_type##_, op_maker_class, _OpGradClass_##op_type##_> \ + __op_registrar_##op_type##__(#op_type, #grad_op_type); \ + int TouchOpRegistrar_##op_type() { \ + __op_registrar_##op_type##__.Touch(); \ + return 0; \ + } + +#define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \ + REGISTER_OP(op_type, op_class, op_maker_class, , ::paddle::framework::NOP) /** - * Macro to Register Gradient Operator. + * Macro to register OperatorKernel. */ -#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; \ - } - -/** - * 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") - -/** - * 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; } + __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; \ + } -// (type, KernelType) -#define REGISTER_OP_GPU_KERNEL(type, ...) \ - REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__) +#define REGISTER_OP_GPU_KERNEL(op_type, ...) \ + REGISTER_OP_KERNEL(op_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__) +#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 + * 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 9894928a7a..50c45919c5 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(); @@ -66,8 +76,7 @@ TEST(OpRegistry, CreateOp) { attr->set_type(paddle::framework::AttrType::FLOAT); attr->set_f(scale); - std::shared_ptr op = - paddle::framework::OpRegistry::CreateOp(op_desc); + auto op = paddle::framework::OpRegistry::CreateOp(op_desc); paddle::framework::Scope scope; paddle::platform::CPUDeviceContext dev_ctx; op->Run(scope, dev_ctx); @@ -78,8 +87,8 @@ 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"); @@ -103,33 +112,23 @@ TEST(OpRegistry, IllegalAttr) { TEST(OpRegistry, DefaultValue) { paddle::framework::OpDesc op_desc; op_desc.set_type("cos_sim"); - op_desc.add_inputs("aa"); - op_desc.add_outputs("bb"); + BuildVar("input", {"aa"}, op_desc.add_inputs()); + BuildVar("output", {"bb"}, op_desc.add_outputs()); ASSERT_TRUE(op_desc.IsInitialized()); - std::shared_ptr op = - paddle::framework::OpRegistry::CreateOp(op_desc); + auto op = paddle::framework::OpRegistry::CreateOp(op_desc); paddle::framework::Scope scope; paddle::platform::CPUDeviceContext dev_ctx; op->Run(scope, dev_ctx); ASSERT_EQ(op->GetAttr("scale"), 1.0); } -static void SetInputFormat(paddle::framework::OpDesc* desc) { - auto attr = desc->add_attrs(); - attr->set_name("input_format"); - attr->set_type(paddle::framework::INTS); - attr->mutable_ints()->Add(0); - attr->mutable_ints()->Add(1); -} - 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 +168,6 @@ 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 d9a013b883..7abbde610f 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 { @@ -34,83 +34,164 @@ ExecutionContext::GetEigenDevice() const { #endif const std::string& OperatorBase::Input(const std::string& name) const { - PADDLE_ENFORCE_NOT_NULL(in_out_idxs_, - "Input Output Indices could not be nullptr"); - auto it = in_out_idxs_->find(name); - PADDLE_ENFORCE(it != in_out_idxs_->end(), "no key [%s] in in_out_idxs_", - name); - if (attrs_.count("input_format") == 0) { - return inputs_.at((size_t)it->second); - } else { - const auto& input_format = GetAttr>("input_format"); - int idx = input_format[it->second]; - return inputs_.at((size_t)idx); - } + 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_NOT_NULL(in_out_idxs_, "IO Idx could not be nullptr"); - auto input_format = GetAttr>("input_format"); - auto offset = in_out_idxs_->at(name); - PADDLE_ENFORCE(input_format.at(static_cast(offset) + 1) <= - static_cast(inputs_.size()), - "Input Out Of Range"); - - return std::vector{ - inputs_.begin() + input_format.at(offset), - inputs_.begin() + input_format.at(offset + 1)}; +const std::vector& OperatorBase::Inputs( + const std::string& name) const { + auto it = inputs_.find(name); + PADDLE_ENFORCE(it != inputs_.end(), "Op %s do not have input %s", type_, + name); + return it->second; } const std::string& OperatorBase::Output(const std::string& name) const { - PADDLE_ENFORCE_NOT_NULL(in_out_idxs_, "InOut Indice could not be nullptr"); - auto it = in_out_idxs_->find(name); - PADDLE_ENFORCE(it != in_out_idxs_->end(), "no key [%s] in in_out_idxs_", - name); - if (attrs_.count("output_format") == 0) { - return outputs_.at((size_t)it->second); - } else { - const auto& output_format = GetAttr>("output_format"); - int idx = output_format[it->second]; - return outputs_.at((size_t)idx); - } + 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_NOT_NULL(in_out_idxs_, "InOut Indice could not be nullptr"); - auto output_format = GetAttr>("output_format"); - auto offset = in_out_idxs_->at(name); - PADDLE_ENFORCE(output_format.at(static_cast(offset) + 1) <= - static_cast(outputs_.size()), - "Output Out of Range"); - return std::vector{ - outputs_.begin() + output_format.at(offset), - outputs_.begin() + output_format.at(offset + 1)}; +const std::vector& OperatorBase::Outputs( + const std::string& name) const { + auto it = outputs_.find(name); + PADDLE_ENFORCE(it != outputs_.end(), "Op %s does not have output %s", type_, + name); + return it->second; } std::string OperatorBase::DebugString() const { std::stringstream ss; - ss << "Op(" << type_ << "), inputs:("; - for (size_t i = 0; i < inputs_.size(); ++i) { - ss << inputs_[i]; - if (i != inputs_.size() - 1) { + ss << "Op(" << type_ << "), inputs:{"; + for (auto it = inputs_.begin(); it != inputs_.end();) { + auto& input = *it; + ss << input.first << "["; + for (size_t i = 0; i < input.second.size(); ++i) { + ss << input.second[i]; + if (i != input.second.size() - 1) { + ss << ", "; + } + } + ss << "]"; + ++it; + if (it != inputs_.end()) { ss << ", "; } } - ss << "), outputs:("; - for (size_t i = 0; i < outputs_.size(); ++i) { - ss << outputs_[i]; - if (i != outputs_.size() - 1) { + ss << "}, outputs:{"; + for (auto it = outputs_.begin(); it != outputs_.end();) { + auto& output = *it; + ss << output.first << "["; + for (size_t i = 0; i < output.second.size(); ++i) { + ss << output.second[i]; + if (i != output.second.size() - 1) { + ss << ", "; + } + } + ss << "]"; + ++it; + if (it != outputs_.end()) { ss << ", "; } } - ss << ")."; + ss << "}."; return ss.str(); } void OperatorBase::Rename(const std::string& old_name, const std::string& new_name) { - std::replace(inputs_.begin(), inputs_.end(), old_name, new_name); - std::replace(outputs_.begin(), outputs_.end(), old_name, new_name); + for (auto& input : inputs_) { + std::replace(input.second.begin(), input.second.end(), old_name, new_name); + } + for (auto& output : outputs_) { + std::replace(output.second.begin(), output.second.end(), old_name, + new_name); + } +} + +OperatorBase::OperatorBase(const std::string& type, + const VariableNameMap& inputs, + const VariableNameMap& outputs, + const AttributeMap& attrs) + : type_(type), inputs_(inputs), outputs_(outputs), attrs_(attrs) { + 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& info = OpInfoMap::Instance().Get(Type()); + + // get all OpProto::Var for outputs + for (auto& o : info.Proto().outputs()) { + // ignore all intermediate output + if (o.intermediate()) continue; + auto out = outputs_.find(o.name()); + if (out != outputs_.end()) { + ret_val.reserve(ret_val.size() + out->second.size()); + ret_val.insert(ret_val.end(), out->second.begin(), out->second.end()); + } + } + return ret_val; +} + +void OpProtoAndCheckerMaker::Validate() { + validated_ = true; + CheckNoDuplicatedInOutAttrs(); +} + +OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddInput( + const std::string& name, const std::string& comment) { + auto* input = proto_->add_inputs(); + input->set_name(name); + input->set_comment(comment); + return OpProtoAndCheckerMaker::VariableBuilder{input}; +} + +OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddOutput( + const std::string& name, const std::string& comment) { + auto* output = proto_->add_outputs(); + output->set_name(name); + output->set_comment(comment); + return OpProtoAndCheckerMaker::VariableBuilder{output}; +} + +void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() { + std::unordered_set names; + auto checker = [&](const std::string& name) { + PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name); + names.insert(name); + }; + for (auto& attr : proto_->attrs()) { + checker(attr.name()); + } + for (auto& input : proto_->inputs()) { + checker(input.name()); + } + for (auto& output : proto_->outputs()) { + checker(output.name()); + } } } // namespace framework diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h index c324fa6702..8397570d26 100644 --- a/paddle/framework/operator.h +++ b/paddle/framework/operator.h @@ -15,37 +15,37 @@ limitations under the License. */ #pragma once #include -#include #include #include #include +#include "op_info.h" #include "paddle/framework/attribute.h" -#include "paddle/framework/op_desc.pb.h" -#include "paddle/framework/op_proto.pb.h" +#include "paddle/framework/framework.pb.h" #include "paddle/framework/scope.h" #include "paddle/framework/tensor.h" #include "paddle/platform/device_context.h" #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. -const std::string kEmptyVarName = "@EMPTY@"; +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. -const std::string kTempVarName = "@TEMP@"; +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". -const std::string kGradVarSuffix = "@GRAD"; +constexpr char kGradVarSuffix[] = "@GRAD"; /// Variables with this suffix are supposed to be filled up with zeros. -const std::string kZeroVarSuffix = "@ZERO"; +constexpr char kZeroVarSuffix[] = "@ZERO"; inline std::string GradVarName(const std::string& var_name) { return var_name + kGradVarSuffix; @@ -63,6 +63,9 @@ class ExecutionContext; */ class OperatorBase { public: + OperatorBase(const std::string& type, const VariableNameMap& inputs, + const VariableNameMap& outputs, const AttributeMap& attrs); + virtual ~OperatorBase() {} template @@ -74,10 +77,6 @@ class OperatorBase { virtual std::string DebugString() const; - /// Init will be called after CreateOperator, you can put some initialization - /// logic here. - virtual void Init() {} - /// InferShape infer the size of Variables used by this Operator with /// information inside scope virtual void InferShape(const Scope& scope) const = 0; @@ -93,48 +92,149 @@ class OperatorBase { /// rename inputs outputs name void Rename(const std::string& old_name, const std::string& new_name); + const VariableNameMap& Inputs() const { return inputs_; } + const VariableNameMap& Outputs() const { return outputs_; } //! Get a input with argument's name described in `op_proto` const std::string& Input(const std::string& name) const; - //! 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_; } + + // Return a new operator instance, which is as same as this. + // Use unique_ptr to prevent caller forget to delete this pointer. + virtual std::unique_ptr Clone() const = 0; + + protected: std::string type_; // NOTE: in case of OpGrad, inputs_ contains: - // I (Inputs) + // I (Inputs)opear // O (Outputs) // OG (Output Gradients) - std::vector inputs_; + VariableNameMap inputs_; + // NOTE: in case of OpGrad, outputs_ contains // IG (Inputs Gradients) - std::vector outputs_; + VariableNameMap outputs_; AttributeMap attrs_; - // store the arguments' offset described in op_desc. - std::shared_ptr> in_out_idxs_; }; -class OperatorContext { +// Macro for define a clone method. +// If you are writing an kernel operator, `Clone` will be defined when you +// register it. i.e. `Clone` method is not needed to define by yourself. +#define DEFINE_OP_CLONE_METHOD(cls) \ + std::unique_ptr Clone() const final { \ + return std::unique_ptr(new cls(*this)); \ + } + +// Macro for define a default constructor for Operator. +// You can also use +// using PARENT_CLASS::PARENT_CLASS; +// to use parent's constructor. +#define DEFINE_OP_CONSTRUCTOR(cls, parent_cls) \ + cls(const std::string& type, \ + const ::paddle::framework::VariableNameMap& inputs, \ + const ::paddle::framework::VariableNameMap& outputs, \ + const paddle::framework::AttributeMap& attrs) \ + : parent_cls(type, inputs, outputs, attrs) {} + +class NOP : public OperatorBase { + public: + using OperatorBase::OperatorBase; + void InferShape(const Scope& scope) const override {} + void Run(const Scope& scope, + const platform::DeviceContext& dev_ctx) const override {} + std::unique_ptr Clone() const override { + return std::unique_ptr(new NOP(*this)); + } +}; + +// this class not only make proto but also init attribute checkers. +class OpProtoAndCheckerMaker { public: - OperatorContext(const OperatorBase* op, const Scope& scope) - : op_(*op), scope_(scope) {} + OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) + : proto_(proto), op_checker_(op_checker) {} + + ~OpProtoAndCheckerMaker() { + PADDLE_ENFORCE(validated_, "should call Validate after build"); + } + + void Validate(); + + protected: + struct VariableBuilder { + OpProto::Var* var_; + + VariableBuilder& AsDuplicable() { + var_->set_duplicable(true); + return *this; + } + + VariableBuilder& AsIntermediate() { + var_->set_intermediate(true); + return *this; + } + + VariableBuilder& NotInGradient() { + var_->set_not_in_gradient(true); + return *this; + } + }; + + VariableBuilder AddInput(const std::string& name, const std::string& comment); + + VariableBuilder AddOutput(const std::string& name, + const std::string& comment); + + template + TypedAttrChecker& AddAttr(const std::string& name, + const std::string& comment, + bool generated = false) { + auto* attr = proto_->add_attrs(); + attr->set_name(name); + attr->set_comment(comment); + attr->set_generated(generated); + attr->set_type(AttrTypeID()); + return op_checker_->AddAttrChecker(name); + } - size_t InputSize() const { return op_.inputs_.size(); } + void AddComment(const std::string& comment) { proto_->set_comment(comment); } - size_t OutputSize() const { return op_.outputs_.size(); } + private: + void CheckNoDuplicatedInOutAttrs(); + + OpProto* proto_; + OpAttrChecker* op_checker_; + bool validated_{false}; +}; + +class NOPMaker : public OpProtoAndCheckerMaker { + public: + NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) {} +}; + +class InferShapeContext { + public: + InferShapeContext(const OperatorBase& op, const Scope& scope) + : 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 { @@ -166,27 +266,9 @@ class OperatorContext { return res; } - template - const T* Input(const size_t index) const { - auto var = InputVar(index); - PADDLE_ENFORCE_NOT_NULL(var, "Input(%d) should not be nullptr", index); - return &var->Get(); - } - - template - T* Output(const size_t index) const { - auto var = OutputVar(index); - PADDLE_ENFORCE_NOT_NULL( - var, - "Output(%d) not be nullptr, which means variable [%s] does not " - "exist in scope", - index, op_.outputs_[index]); - return var->GetMutable(); - } - template const T* Input(const std::string& name) const { - auto var = InputVar(name); + auto* var = InputVar(name); PADDLE_ENFORCE_NOT_NULL(var, "Input(%s) should not be nullptr", name); return &var->Get(); } @@ -223,7 +305,7 @@ class OperatorContext { [&](const std::string& sub_name) { auto var = scope_.FindVar(sub_name); PADDLE_ENFORCE_NOT_NULL( - var, "MultiOutput(%s:%s) should not be nullptr", name, + var, "MultiOutput(%s:%s) should not be nullptr.", name, sub_name); return var->GetMutable(); }); @@ -234,12 +316,6 @@ class OperatorContext { const Scope& scope_; }; -class InferShapeContext : public OperatorContext { - public: - InferShapeContext(const OperatorBase* op, const Scope& scope) - : OperatorContext(op, scope) {} -}; - template struct EigenDeviceConverter; @@ -255,11 +331,11 @@ struct EigenDeviceConverter { }; #endif -class ExecutionContext : public OperatorContext { +class ExecutionContext : public InferShapeContext { public: - ExecutionContext(const OperatorBase* op, const Scope& scope, + ExecutionContext(const OperatorBase& op, const Scope& scope, const platform::DeviceContext* device_context) - : OperatorContext(op, scope), device_context_(device_context) {} + : InferShapeContext(op, scope), device_context_(device_context) {} template GetPlace(); } + const platform::DeviceContext* device_context() const { + return device_context_; + } + const platform::DeviceContext* device_context_; }; @@ -310,14 +390,18 @@ class OperatorWithKernel : public OperatorBase { using OpKernelMap = std::unordered_map, OpKernelHash>; + OperatorWithKernel(const std::string& type, const VariableNameMap& inputs, + const VariableNameMap& outputs, const AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + void InferShape(const Scope& scope) const override { - InferShape(InferShapeContext(this, scope)); + 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& diff --git a/paddle/framework/operator_test.cc b/paddle/framework/operator_test.cc index 387aada749..1d7efb7b94 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 VariableNameMap& inputs, + const VariableNameMap& outputs, const AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs), x(1) {} void InferShape(const Scope& scope) const override {} void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override { - op_run_num++; - ASSERT_EQ((int)inputs_.size(), 1); - ASSERT_EQ((int)outputs_.size(), 1); - ASSERT_EQ(scope.FindVar(inputs_[0]), nullptr); + ++op_run_num; + ASSERT_EQ(static_cast(inputs_.size()), 1); + ASSERT_EQ(static_cast(outputs_.size()), 1); + ASSERT_EQ(scope.FindVar(inputs_.at("input")[0]), nullptr); ASSERT_EQ(x, 1); - ASSERT_NE(scope.FindVar(outputs_[0]), nullptr); + ASSERT_NE(scope.FindVar(outputs_.at("output")[0]), nullptr); } public: - float x = 0; + int x{0}; }; class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { @@ -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); @@ -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(); @@ -261,3 +245,22 @@ TEST(OpKernel, multi_inputs) { auto op = paddle::framework::OpRegistry::CreateOp(op_desc); op->Run(scope, cpu_device_context); } + +class OperatorClone : public paddle::framework::OperatorBase { + public: + DEFINE_OP_CLONE_METHOD(OperatorClone); + OperatorClone(const std::string& type, + const paddle::framework::VariableNameMap& inputs, + const paddle::framework::VariableNameMap& outputs, + const paddle::framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + void InferShape(const paddle::framework::Scope& scope) const override {} + void Run(const paddle::framework::Scope& scope, + const paddle::platform::DeviceContext& dev_ctx) const override {} +}; + +TEST(Operator, Clone) { + OperatorClone a("ABC", {}, {}, {}); + auto b = a.Clone(); + ASSERT_EQ(a.Type(), b->Type()); +} \ No newline at end of file diff --git a/paddle/framework/tensor.h b/paddle/framework/tensor.h index b57958591f..643f875491 100644 --- a/paddle/framework/tensor.h +++ b/paddle/framework/tensor.h @@ -79,11 +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. @@ -105,6 +105,11 @@ class Tensor { template inline Tensor Slice(const int& begin_idx, const int& end_idx) const; + platform::Place place() const { + PADDLE_ENFORCE_NOT_NULL(holder_, "Tensor get place() must contains holder"); + return holder_->place(); + } + private: template inline void check_memory_size() const; diff --git a/paddle/framework/tensor_impl.h b/paddle/framework/tensor_impl.h index 8d9bec6dc9..7d7263b899 100644 --- a/paddle/framework/tensor_impl.h +++ b/paddle/framework/tensor_impl.h @@ -23,9 +23,11 @@ template inline void Tensor::check_memory_size() const { 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."); + 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 @@ -78,9 +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 @@ -136,7 +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/function/CMakeLists.txt b/paddle/function/CMakeLists.txt index 93304f7303..f43f15e5ca 100644 --- a/paddle/function/CMakeLists.txt +++ b/paddle/function/CMakeLists.txt @@ -4,6 +4,10 @@ file(GLOB cpp_files . *Op.cpp) list(APPEND h_files Function.h) list(APPEND cpp_files Function.cpp) list(APPEND cpp_files BufferArg.cpp) +list(APPEND cpp_files GemmFunctor.cpp) +if(USE_EIGEN_FOR_BLAS) + list(APPEND cpp_files EigenGemm.cpp) +endif(USE_EIGEN_FOR_BLAS) if(WITH_GPU) file(GLOB cu_files . *OpGpu.cu) @@ -17,6 +21,8 @@ if(USE_NNPACK) endif() endif() +list(APPEND cpp_files neon/NeonDepthwiseConv.cpp) + add_library(paddle_function STATIC ${cpp_files} ${cu_objs}) add_dependencies(paddle_function ${external_project_dependencies}) add_dependencies(paddle_function paddle_proto) @@ -40,8 +46,9 @@ if(WITH_GPU) add_simple_unittest(CropOpTest) endif() -add_simple_unittest(ConvOpTest) add_simple_unittest(Im2ColTest) +add_simple_unittest(GemmConvOpTest) +add_simple_unittest(DepthwiseConvOpTest) endif() add_style_check_target(paddle_function ${h_files}) diff --git a/paddle/function/ConvOpTest.cpp b/paddle/function/ConvOpTest.cpp deleted file mode 100644 index 7f32c73479..0000000000 --- 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 0000000000..cb02a96d0d --- /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/DepthwiseConvOp.cpp b/paddle/function/DepthwiseConvOp.cpp index 490e8d546c..2f3112fe65 100644 --- a/paddle/function/DepthwiseConvOp.cpp +++ b/paddle/function/DepthwiseConvOp.cpp @@ -14,7 +14,6 @@ limitations under the License. */ #include "DepthwiseConvOp.h" #include "ConvOp.h" -#include "GemmFunctor.h" namespace paddle { diff --git a/paddle/function/DepthwiseConvOpGpu.cu b/paddle/function/DepthwiseConvOpGpu.cu index 33463805cb..2d722dfcfc 100644 --- a/paddle/function/DepthwiseConvOpGpu.cu +++ b/paddle/function/DepthwiseConvOpGpu.cu @@ -13,7 +13,6 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "DepthwiseConvOp.h" -#include "GemmFunctor.h" #include "paddle/math/BaseMatrix.h" namespace paddle { diff --git a/paddle/function/DepthwiseConvOpTest.cpp b/paddle/function/DepthwiseConvOpTest.cpp new file mode 100644 index 0000000000..d8e8c889d5 --- /dev/null +++ b/paddle/function/DepthwiseConvOpTest.cpp @@ -0,0 +1,46 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "ConvOpTest.h" + +namespace paddle { + +#ifndef PADDLE_ONLY_CPU +TEST(DepthwiseConv, Forward) { + DepthwiseConvolution( + "GemmConv-CPU", "DepthwiseConv-GPU", forward); +} + +TEST(DepthwiseConv, BackwardInput) { + DepthwiseConvolution( + "GemmConvGradInput-CPU", "DepthwiseConvGradInput-GPU", backward_input); +} + +TEST(DepthwiseConv, BackwardFilter) { + DepthwiseConvolution( + "GemmConvGradFilter-CPU", "DepthwiseConvGradFilter-GPU", backward_filter); +} +#endif + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + +TEST(DepthwiseConv, Forward) { + DepthwiseConvolution( + "GemmConv-CPU", "NeonDepthwiseConv-CPU", forward); +} + +#endif + +} // namespace paddle diff --git a/paddle/function/EigenGemm.cpp b/paddle/function/EigenGemm.cpp new file mode 100644 index 0000000000..674141ed39 --- /dev/null +++ b/paddle/function/EigenGemm.cpp @@ -0,0 +1,91 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "unsupported/Eigen/CXX11/Tensor" + +namespace paddle { + +template +struct EigenBlasGemm { + typedef Eigen::TensorMap, + Eigen::Aligned> + Matrix; + + static void compute(const bool transA, + const bool transB, + const int M, + const int N, + const int K, + const T alpha, + const T* A, + const int lda, + const T* B, + const int ldb, + const T beta, + T* C, + const int ldc) { + Eigen::array sizeA; + if (transA) { + sizeA[0] = K; + sizeA[1] = M; + CHECK_EQ(M, lda); + } else { + sizeA[0] = M; + sizeA[1] = K; + CHECK_EQ(K, lda); + } + Eigen::array sizeB; + if (transB) { + sizeB[0] = N; + sizeB[1] = K; + CHECK_EQ(K, ldb); + } else { + sizeB[0] = K; + sizeB[1] = N; + CHECK_EQ(N, ldb); + } + Eigen::array sizeC; + sizeC[0] = M; + sizeC[1] = N; + CHECK_EQ(N, ldc); + + const Matrix a(const_cast(A), sizeA); + const Matrix b(const_cast(B), sizeB); + Matrix c(C, sizeC); + + typedef typename Eigen::Tensor::DimensionPair DimPair; + Eigen::array dims; + dims[0] = DimPair(1, 0); + dims[0].first = transA ? 0 : 1; + dims[0].second = transB ? 1 : 0; + + Eigen::DefaultDevice device; + if (alpha == T(1) && beta == T(0)) { + c.device(device) = a.contract(b, dims); + } else if (alpha == T(1) && beta == T(1)) { + c.device(device) += a.contract(b, dims); + } else { + c.device(device) = alpha * a.contract(b, dims) + beta * c; + } + } +}; + +#ifdef PADDLE_TYPE_DOUBLE +template class EigenBlasGemm; +#else +template class EigenBlasGemm; +#endif + +} // namespace paddle diff --git a/paddle/function/GemmConvOp.cpp b/paddle/function/GemmConvOp.cpp index 0ada4d70a0..f8cf4ebea8 100644 --- a/paddle/function/GemmConvOp.cpp +++ b/paddle/function/GemmConvOp.cpp @@ -85,7 +85,6 @@ public: } Im2ColFunctor im2col; - GemmFunctor gemm; size_t inputOffset = imShape.getElements(); size_t outputOffset = (outputChannels / groups_) * outputHeight * outputWidth; @@ -108,19 +107,19 @@ public: int M = outputChannels / groups_; int N = outputHeight * outputWidth; int K = inputChannels / groups_ * filterHeight * filterWidth; - gemm(CblasNoTrans, - CblasNoTrans, - M, - N, - K, - 1.0f, - filterData + g * filterOffset, - K, - colData, - N, - beta, - outputData + g * outputOffset, - N); + BlasGemm::compute(false, + false, + M, + N, + K, + 1.0f, + filterData + g * filterOffset, + K, + colData, + N, + beta, + outputData + g * outputOffset, + N); } inputData += inputChannels * inputHeight * inputWidth; outputData += outputChannels * outputHeight * outputWidth; @@ -188,8 +187,6 @@ public: } Col2ImFunctor col2im; - GemmFunctor gemm; - size_t inputOffset = imShape.getElements(); size_t outputOffset = (outputChannels / groups_) * outputHeight * outputWidth; @@ -205,19 +202,19 @@ public: colData = inputGrad + g * inputOffset; scale = 1.0f; } - gemm(CblasTrans, - CblasNoTrans, - M, - N, - K, - 1.0f, - filterData + g * filterOffset, - M, - outputGrad + g * outputOffset, - N, - scale, - colData, - N); + BlasGemm::compute(true, + false, + M, + N, + K, + 1.0f, + filterData + g * filterOffset, + M, + outputGrad + g * outputOffset, + N, + scale, + colData, + N); if (needIm2col) { col2im(inputGrad + g * inputOffset, imShape, @@ -299,7 +296,6 @@ public: } Im2ColFunctor im2col; - GemmFunctor gemm; size_t inputOffset = imShape.getElements(); size_t outputOffset = (outputChannels / groups_) * outputHeight * outputWidth; @@ -321,19 +317,19 @@ public: int M = outputChannels / groups_; int K = outputHeight * outputWidth; int N = inputChannels / groups_ * filterHeight * filterWidth; - gemm(CblasNoTrans, - CblasTrans, - M, - N, - K, - 1.0f, - outputGrad + g * outputOffset, - K, - colData, - K, - i == 0 ? beta : 1.0f, - filterGrad + g * filterOffset, - N); + BlasGemm::compute(false, + true, + M, + N, + K, + 1.0f, + outputGrad + g * outputOffset, + K, + colData, + K, + i == 0 ? beta : 1.0f, + filterGrad + g * filterOffset, + N); } inputData += inputChannels * inputHeight * inputWidth; outputGrad += outputChannels * outputHeight * outputWidth; diff --git a/paddle/function/GemmConvOpTest.cpp b/paddle/function/GemmConvOpTest.cpp new file mode 100644 index 0000000000..5283d79a5a --- /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/GemmFunctor.cpp b/paddle/function/GemmFunctor.cpp new file mode 100644 index 0000000000..9e25ee58a1 --- /dev/null +++ b/paddle/function/GemmFunctor.cpp @@ -0,0 +1,90 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "GemmFunctor.h" +#include "paddle/math/MathFunctions.h" + +namespace paddle { + +template +struct BlasGemm { + static void compute(const bool transA, + const bool transB, + const int M, + const int N, + const int K, + const T alpha, + const T* A, + const int lda, + const T* B, + const int ldb, + const T beta, + T* C, + const int ldc) { +#ifdef PADDLE_USE_EIGEN_FOR_BLAS + EigenBlasGemm::compute( + transA, transB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc); +#else + gemm(transA == false ? CblasNoTrans : CblasTrans, + transB == false ? CblasNoTrans : CblasTrans, + M, + N, + K, + alpha, + A, + lda, + B, + ldb, + beta, + C, + ldc); +#endif + } +}; + +template +struct BlasGemm { + static void compute(const bool transA, + const bool transB, + const int M, + const int N, + const int K, + const T alpha, + const T* A, + const int lda, + const T* B, + const int ldb, + const T beta, + T* C, + const int ldc) { + hl_matrix_mul((T*)A, + transA == false ? HPPL_OP_N : HPPL_OP_T, + (T*)B, + transB == false ? HPPL_OP_N : HPPL_OP_T, + C, + M, + N, + K, + alpha, + beta, + lda, + ldb, + ldc); + } +}; + +template struct BlasGemm; +template struct BlasGemm; + +} // namespace paddle diff --git a/paddle/function/GemmFunctor.h b/paddle/function/GemmFunctor.h index d5db5cf5e7..0809953b4e 100644 --- a/paddle/function/GemmFunctor.h +++ b/paddle/function/GemmFunctor.h @@ -14,7 +14,7 @@ limitations under the License. */ #pragma once -#include "paddle/math/MathFunctions.h" +#include "TensorType.h" namespace paddle { @@ -24,73 +24,42 @@ namespace paddle { // of MatMulFunction, we need to consider the reconstruction of hl_matrix_mul // interface. template -class GemmFunctor { -public: - void operator()(const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE TransB, - const int M, - const int N, - const int K, - const T alpha, - const T* A, - const int lda, - const T* B, - const int ldb, - const T beta, - T* C, - const int ldc); +struct BlasGemm { + static void compute(const bool transA, + const bool transB, + const int M, + const int N, + const int K, + const T alpha, + const T* A, + const int lda, + const T* B, + const int ldb, + const T beta, + T* C, + const int ldc); }; +// TODO(hedaoyuan): Since the definition of the real type in the Paddle +// conflicts with the Eigen library, so compile the Eigen code can not +// include the Paddle header file. And need an EigenBlasGemm template class +// that does not contain the DeviceType parameter. +// I will fix this problem and merge BlasGemm and EigenBlasGemm into one. template -class GemmFunctor { -public: - void operator()(const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE TransB, - const int M, - const int N, - const int K, - const T alpha, - const T* A, - const int lda, - const T* B, - const int ldb, - const T beta, - T* C, - const int ldc) { - gemm(transA, TransB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc); - } -}; - -template -class GemmFunctor { -public: - void operator()(const CBLAS_TRANSPOSE transA, - const CBLAS_TRANSPOSE TransB, - const int M, - const int N, - const int K, - const T alpha, - const T* A, - const int lda, - const T* B, - const int ldb, - const T beta, - T* C, - const int ldc) { - hl_matrix_mul((T*)A, - transA == CblasNoTrans ? HPPL_OP_N : HPPL_OP_T, - (T*)B, - TransB == CblasNoTrans ? HPPL_OP_N : HPPL_OP_T, - C, - M, - N, - K, - alpha, - beta, - lda, - ldb, - ldc); - } +struct EigenBlasGemm { + static void compute(const bool transA, + const bool transB, + const int M, + const int N, + const int K, + const T alpha, + const T* A, + const int lda, + const T* B, + const int ldb, + const T beta, + T* C, + const int ldc); }; } // namespace paddle diff --git a/paddle/function/Im2Col.h b/paddle/function/Im2Col.h index 48e2e32f92..9b91e223a6 100644 --- a/paddle/function/Im2Col.h +++ b/paddle/function/Im2Col.h @@ -16,6 +16,7 @@ limitations under the License. */ #include "TensorShape.h" #include "TensorType.h" +#include "neon/neon_util.h" namespace paddle { @@ -93,4 +94,95 @@ public: int paddingWidth); }; +template +struct Padding { + static void run(const T* src, + T* dest, + int channels, + int inputHeight, + int inputWidth, + int paddingHeight, + int paddingWidth) { + const int destWidth = inputWidth + 2 * paddingWidth; + for (int c = 0; c < channels; c++) { + if (paddingHeight > 0) { + memset(dest, 0, destWidth * paddingHeight * sizeof(T)); + dest += destWidth * paddingHeight; + } + + for (int i = 0; i < inputHeight; i++) { + // padding head + for (int j = 0; j < paddingWidth; j++) { + *dest++ = T(0); + } + + memcpy(dest, src, inputWidth * sizeof(T)); + dest += inputWidth; + src += inputWidth; + + // padding tail + for (int j = 0; j < paddingWidth; j++) { + *dest++ = T(0); + } + } + + if (paddingHeight > 0) { + memset(dest, 0, destWidth * paddingHeight * sizeof(T)); + dest += destWidth * paddingHeight; + } + } + } +}; + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) +template <> +struct Padding { + static void run(const float* src, + float* dest, + int channels, + int inputHeight, + int inputWidth, + int paddingHeight, + int paddingWidth) { + const int destWidth = inputWidth + 2 * paddingWidth; + for (int c = 0; c < channels; c++) { + if (paddingHeight > 0) { + memset(dest, 0, destWidth * paddingHeight * sizeof(float)); + dest += destWidth * paddingHeight; + } + + for (int i = 0; i < inputHeight; i++) { + // padding head + for (int j = 0; j < paddingWidth; j++) { + *dest++ = float(0); + } + + int step = inputWidth >> 2; + int remain = inputWidth & 3; + for (int s = 0; s < step; s++) { + float32x4_t s0 = vld1q_f32(src); + vst1q_f32(dest, s0); + src += 4; + dest += 4; + } + for (int r = 0; r < remain; r++) { + *dest++ = *src++; + } + + // padding tail + for (int j = 0; j < paddingWidth; j++) { + *dest++ = float(0); + } + } + + if (paddingHeight > 0) { + memset(dest, 0, destWidth * paddingHeight * sizeof(float)); + dest += destWidth * paddingHeight; + } + } + } +}; + +#endif + } // namespace paddle diff --git a/paddle/function/neon/NeonDepthwiseConv.cpp b/paddle/function/neon/NeonDepthwiseConv.cpp new file mode 100644 index 0000000000..f09e98587d --- /dev/null +++ b/paddle/function/neon/NeonDepthwiseConv.cpp @@ -0,0 +1,577 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "neon_util.h" +#include "paddle/function/ConvOp.h" +#include "paddle/function/Im2Col.h" + +namespace paddle { + +namespace neon { + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + +template +struct DepthwiseConvKernel {}; + +inline float32_t conv3x3(float32x4_t r0, + float32x4_t r1, + float32x4_t r2, + float32x4_t k0, + float32x4_t k1, + float32x4_t k2) { + float32x4_t tmp; + tmp = vmulq_f32(r0, k0); + tmp = vmlaq_f32(tmp, r1, k1); + tmp = vmlaq_f32(tmp, r2, k2); + return vaddvq_f32(tmp); +} + +inline float32_t conv4x4(float32x4_t r0, + float32x4_t r1, + float32x4_t r2, + float32x4_t r3, + float32x4_t k0, + float32x4_t k1, + float32x4_t k2, + float32x4_t k3) { + float32x4_t tmp; + tmp = vmulq_f32(r0, k0); + tmp = vmlaq_f32(tmp, r1, k1); + tmp = vmlaq_f32(tmp, r2, k2); + tmp = vmlaq_f32(tmp, r3, k3); + return vaddvq_f32(tmp); +} + +/** + * Each step calculates four elements of the output. + * First step: + * R0[0, 1, 2, 3...] * K[0][0] + * R0[1, 2, 3, 4...] * K[0][1] + * R0[2, 3, 4, 5...] * K[0][2] + * R1[0, 1, 2, 3...] * K[1][0] + * R1[1, 2, 3, 4...] * K[1][1] + * R1[2, 3, 4, 5...] * K[1][2] + * R2[0, 1, 2, 3...] * K[2][0] + * R2[1, 2, 3, 4...] * K[2][1] + * + R2[2, 3, 4, 5...] * K[2][2] + * ------------------------------ + * Output[0, 1, 2, 3] + */ +template <> +struct DepthwiseConvKernel<3, 1> { + static void run(const float* inputData, + const float* filterData, + int inputHeight, + int inputWidth, + int outputChannels, + int outputHeight, + int outputWidth, + int filterMultiplier, + float* outputData) { + const int steps = outputWidth >> 2; + const int remain = outputWidth & 3; + for (int c = 0; c < outputChannels; c++, filterData += 9) { + // Load the filters + float32x4_t k[3]; + k[0] = vld1q_f32(filterData); + k[1] = vld1q_f32(filterData + 3); + k[2] = vld1q_f32(filterData + 6); + k[0] = vsetq_lane_f32(0.f, k[0], 3); + k[1] = vsetq_lane_f32(0.f, k[1], 3); + k[2] = vsetq_lane_f32(0.f, k[2], 3); + + const float* r0 = + inputData + (c / filterMultiplier) * (inputHeight * inputWidth); + const float* r1 = r0 + inputWidth; + const float* r2 = r0 + inputWidth * 2; + float32x4_t input[3][3]; + for (int h = 0; h < outputHeight; h++) { + for (int s = 0; s < steps; s++) { + // Load the inputs + float32x4_t tmp; + input[0][0] = vld1q_f32(r0); + tmp = vld1q_f32(r0 + 4); + input[0][1] = vextq_f32(input[0][0], tmp, 1); + input[0][2] = vextq_f32(input[0][0], tmp, 2); + input[1][0] = vld1q_f32(r1); + tmp = vld1q_f32(r1 + 4); + input[1][1] = vextq_f32(input[1][0], tmp, 1); + input[1][2] = vextq_f32(input[1][0], tmp, 2); + input[2][0] = vld1q_f32(r2); + tmp = vld1q_f32(r2 + 4); + input[2][1] = vextq_f32(input[2][0], tmp, 1); + input[2][2] = vextq_f32(input[2][0], tmp, 2); + + float32x4_t tmp1 = vdupq_n_f32(0.f); + float32x4_t tmp2 = vdupq_n_f32(0.f); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][0], k[1], 0); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][1], k[1], 1); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][2], k[1], 2); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); + tmp1 = vaddq_f32(tmp1, tmp2); + + vst1q_f32(outputData, tmp1); + r0 += 4; + r1 += 4; + r2 += 4; + outputData += 4; + } + + for (int r = 0; r < remain; r++) { + float32x4_t i0 = vld1q_f32(r0); + float32x4_t i1 = vld1q_f32(r1); + float32x4_t i2 = vld1q_f32(r2); + *outputData = conv3x3(i0, i1, i2, k[0], k[1], k[2]); + r0++; + r1++; + r2++; + outputData++; + } + + r0 += 2; + r1 += 2; + r2 += 2; + } + } + } +}; + +/** + * Each step calculates four elements of the output. + * First step: + * R0[0, 2, 4, 6...] * K[0][0] + * R0[1, 3, 5, 7...] * K[0][1] + * R0[2, 4, 6, 8...] * K[0][2] + * R1[0, 2, 4, 6...] * K[1][0] + * R1[1, 3, 5, 7...] * K[1][1] + * R1[2, 4, 6, 8...] * K[1][2] + * R2[0, 2, 4, 6...] * K[2][0] + * R2[1, 3, 5, 7...] * K[2][1] + * R2[2, 4, 6, 8...] * K[2][2] + * ------------------------------ + * Output[0, 1, 2, 3] + */ +template <> +struct DepthwiseConvKernel<3, 2> { + static void run(const float* inputData, + const float* filterData, + int inputHeight, + int inputWidth, + int outputChannels, + int outputHeight, + int outputWidth, + int filterMultiplier, + float* outputData) { + const int steps = outputWidth >> 2; + const int remain = outputWidth & 3; + for (int c = 0; c < outputChannels; c++, filterData += 9) { + // Load the filters + float32x4_t k[3]; + k[0] = vld1q_f32(filterData); + k[1] = vld1q_f32(filterData + 3); + k[2] = vld1q_f32(filterData + 6); + k[0] = vsetq_lane_f32(0.f, k[0], 3); + k[1] = vsetq_lane_f32(0.f, k[1], 3); + k[2] = vsetq_lane_f32(0.f, k[2], 3); + + const float* start = + inputData + (c / filterMultiplier) * (inputHeight * inputWidth); + float32x4_t input[3][3]; + for (int h = 0; h < outputHeight; h++) { + const float* r0 = start + 2 * h * inputWidth; + const float* r1 = start + (2 * h + 1) * inputWidth; + const float* r2 = start + (2 * h + 2) * inputWidth; + for (int s = 0; s < steps; s++) { + // Load the inputs + float32x4_t data1; + float32x4x2_t data2; + + data2 = vld2q_f32(r0); + input[0][0] = data2.val[0]; + input[0][1] = data2.val[1]; + data1 = vld1q_f32(r0 + 8); + input[0][2] = vextq_f32(data2.val[0], data1, 1); + + data2 = vld2q_f32(r1); + input[1][0] = data2.val[0]; + input[1][1] = data2.val[1]; + data1 = vld1q_f32(r1 + 8); + input[1][2] = vextq_f32(data2.val[0], data1, 1); + + data2 = vld2q_f32(r2); + input[2][0] = data2.val[0]; + input[2][1] = data2.val[1]; + data1 = vld1q_f32(r2 + 8); + input[2][2] = vextq_f32(data2.val[0], data1, 1); + + float32x4_t tmp1 = vdupq_n_f32(0.f); + float32x4_t tmp2 = vdupq_n_f32(0.f); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][0], k[1], 0); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][1], k[1], 1); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][2], k[1], 2); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); + tmp1 = vaddq_f32(tmp1, tmp2); + + vst1q_f32(outputData, tmp1); + r0 += 8; + r1 += 8; + r2 += 8; + outputData += 4; + } + + for (int r = 0; r < remain; r++) { + float32x4_t i0 = vld1q_f32(r0); + float32x4_t i1 = vld1q_f32(r1); + float32x4_t i2 = vld1q_f32(r2); + *outputData = conv3x3(i0, i1, i2, k[0], k[1], k[2]); + r0 += 2; + r1 += 2; + r2 += 2; + outputData++; + } + } + } + } +}; + +/** + * Each step calculates four elements of the output. + */ +template <> +struct DepthwiseConvKernel<4, 1> { + static void run(const float* inputData, + const float* filterData, + int inputHeight, + int inputWidth, + int outputChannels, + int outputHeight, + int outputWidth, + int filterMultiplier, + float* outputData) { + const int steps = outputWidth >> 2; + const int remain = outputWidth & 3; + for (int c = 0; c < outputChannels; c++, filterData += 16) { + // Load the filters + float32x4_t k[4]; + k[0] = vld1q_f32(filterData); + k[1] = vld1q_f32(filterData + 4); + k[2] = vld1q_f32(filterData + 8); + k[3] = vld1q_f32(filterData + 12); + + const float* r0 = + inputData + (c / filterMultiplier) * (inputHeight * inputWidth); + const float* r1 = r0 + inputWidth; + const float* r2 = r0 + inputWidth * 2; + const float* r3 = r0 + inputWidth * 3; + float32x4_t input[4][4]; + for (int h = 0; h < outputHeight; h++) { + for (int s = 0; s < steps; s++) { + // Load the inputs + float32x4_t tmp; + input[0][0] = vld1q_f32(r0); + tmp = vld1q_f32(r0 + 4); + input[0][1] = vextq_f32(input[0][0], tmp, 1); + input[0][2] = vextq_f32(input[0][0], tmp, 2); + input[0][3] = vextq_f32(input[0][0], tmp, 3); + + input[1][0] = vld1q_f32(r1); + tmp = vld1q_f32(r1 + 4); + input[1][1] = vextq_f32(input[1][0], tmp, 1); + input[1][2] = vextq_f32(input[1][0], tmp, 2); + input[1][3] = vextq_f32(input[1][0], tmp, 3); + + input[2][0] = vld1q_f32(r2); + tmp = vld1q_f32(r2 + 4); + input[2][1] = vextq_f32(input[2][0], tmp, 1); + input[2][2] = vextq_f32(input[2][0], tmp, 2); + input[2][3] = vextq_f32(input[2][0], tmp, 3); + + input[3][0] = vld1q_f32(r3); + tmp = vld1q_f32(r3 + 4); + input[3][1] = vextq_f32(input[3][0], tmp, 1); + input[3][2] = vextq_f32(input[3][0], tmp, 2); + input[3][3] = vextq_f32(input[3][0], tmp, 3); + + float32x4_t tmp1 = vdupq_n_f32(0.f); + float32x4_t tmp2 = vdupq_n_f32(0.f); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3); + tmp1 = vaddq_f32(tmp1, tmp2); + + vst1q_f32(outputData, tmp1); + r0 += 4; + r1 += 4; + r2 += 4; + r3 += 4; + outputData += 4; + } + + for (int r = 0; r < remain; r++) { + float32x4_t i0 = vld1q_f32(r0); + float32x4_t i1 = vld1q_f32(r1); + float32x4_t i2 = vld1q_f32(r2); + float32x4_t i3 = vld1q_f32(r3); + *outputData = conv4x4(i0, i1, i2, i3, k[0], k[1], k[2], k[3]); + r0++; + r1++; + r2++; + r3++; + outputData++; + } + + r0 += 3; + r1 += 3; + r2 += 3; + r3 += 3; + } + } + } +}; + +/** + * Each step calculates four elements of the output. + */ +template <> +struct DepthwiseConvKernel<4, 2> { + static void run(const float* inputData, + const float* filterData, + int inputHeight, + int inputWidth, + int outputChannels, + int outputHeight, + int outputWidth, + int filterMultiplier, + float* outputData) { + const int steps = outputWidth >> 2; + const int remain = outputWidth & 3; + for (int c = 0; c < outputChannels; c++, filterData += 16) { + // Load the filters + float32x4_t k[4]; + k[0] = vld1q_f32(filterData); + k[1] = vld1q_f32(filterData + 4); + k[2] = vld1q_f32(filterData + 8); + k[3] = vld1q_f32(filterData + 12); + + const float* start = + inputData + (c / filterMultiplier) * (inputHeight * inputWidth); + float32x4_t input[4][4]; + for (int h = 0; h < outputHeight; h++) { + const float* r0 = start + 2 * h * inputWidth; + const float* r1 = start + (2 * h + 1) * inputWidth; + const float* r2 = start + (2 * h + 2) * inputWidth; + const float* r3 = start + (2 * h + 3) * inputWidth; + for (int s = 0; s < steps; s++) { + // Load the inputs + float32x4x2_t data1; + float32x4x2_t data2; + + data1 = vld2q_f32(r0); + data2 = vld2q_f32(r0 + 8); + input[0][0] = data1.val[0]; + input[0][1] = data1.val[1]; + input[0][2] = vextq_f32(data1.val[0], data2.val[0], 1); + input[0][3] = vextq_f32(data1.val[1], data2.val[1], 1); + + data1 = vld2q_f32(r1); + data2 = vld2q_f32(r1 + 8); + input[1][0] = data1.val[0]; + input[1][1] = data1.val[1]; + input[1][2] = vextq_f32(data1.val[0], data2.val[0], 1); + input[1][3] = vextq_f32(data1.val[1], data2.val[1], 1); + + data1 = vld2q_f32(r2); + data2 = vld2q_f32(r2 + 8); + input[2][0] = data1.val[0]; + input[2][1] = data1.val[1]; + input[2][2] = vextq_f32(data1.val[0], data2.val[0], 1); + input[2][3] = vextq_f32(data1.val[1], data2.val[1], 1); + + data1 = vld2q_f32(r3); + data2 = vld2q_f32(r3 + 8); + input[3][0] = data1.val[0]; + input[3][1] = data1.val[1]; + input[3][2] = vextq_f32(data1.val[0], data2.val[0], 1); + input[3][3] = vextq_f32(data1.val[1], data2.val[1], 1); + + float32x4_t tmp1 = vdupq_n_f32(0.f); + float32x4_t tmp2 = vdupq_n_f32(0.f); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3); + tmp1 = vaddq_f32(tmp1, tmp2); + + vst1q_f32(outputData, tmp1); + r0 += 8; + r1 += 8; + r2 += 8; + r3 += 8; + outputData += 4; + } + + for (int r = 0; r < remain; r++) { + float32x4_t i0 = vld1q_f32(r0); + float32x4_t i1 = vld1q_f32(r1); + float32x4_t i2 = vld1q_f32(r2); + float32x4_t i3 = vld1q_f32(r3); + *outputData = conv4x4(i0, i1, i2, i3, k[0], k[1], k[2], k[3]); + r0 += 2; + r1 += 2; + r2 += 2; + r3 += 2; + outputData++; + } + } + } + } +}; + +template +class NeonDepthwiseConvFunction : public ConvFunctionBase { +public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + check(inputs, outputs); + + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + + size_t batchSize = input[0]; + size_t inputChannels = input[1]; + size_t inputHeight = input[2]; + size_t inputWidth = input[3]; + size_t filterHeight = getFilterHeight(filter); + size_t filterWidth = getFilterWidth(filter); + size_t outputChannels = output[1]; + size_t outputHeight = output[2]; + size_t outputWidth = output[3]; + size_t filterMultiplier = outputChannels / groups_; + CHECK_EQ(inputChannels, groups_); + + // only support strideH() == strideW() and filterHeight == filterWidth. + CHECK_EQ(strideH(), strideW()); + CHECK_EQ(filterHeight, filterWidth); + + float* inputData = inputs[0].data(); + float* filterData = inputs[1].data(); + float* outputData = outputs[0].data(); + + // padding the input + float* inputPadding = inputData; + if (paddingH() > 0 || paddingW() > 0) { + int newSize = batchSize * inputChannels * (inputHeight + 2 * paddingH()) * + (inputWidth + 2 * paddingW()); + resizeBuffer(newSize); + inputPadding = reinterpret_cast(memory_->getBuf()); + Padding::run(inputData, + inputPadding, + batchSize * inputChannels, + inputHeight, + inputWidth, + paddingH(), + paddingW()); + + // height and width of padding data + inputHeight += 2 * paddingH(); + inputWidth += 2 * paddingW(); + } + + std::function + DepthWiseConv; + + if (filterWidth == 3 && strideW() == 1) { + DepthWiseConv = DepthwiseConvKernel<3, 1>::run; + } else if (filterWidth == 3 && strideW() == 2) { + DepthWiseConv = DepthwiseConvKernel<3, 2>::run; + } else if (filterWidth == 4 && strideW() == 1) { + DepthWiseConv = DepthwiseConvKernel<4, 1>::run; + } else if (filterWidth == 4 && strideW() == 2) { + DepthWiseConv = DepthwiseConvKernel<4, 2>::run; + } else { + LOG(FATAL) << "Not supported"; + } + + for (size_t i = 0; i < batchSize; i++) { + DepthWiseConv(inputPadding, + filterData, + inputHeight, + inputWidth, + outputChannels, + outputHeight, + outputWidth, + filterMultiplier, + outputData); + inputPadding += inputChannels * inputHeight * inputWidth; + outputData += outputChannels * outputHeight * outputWidth; + } + } +}; + +REGISTER_TYPED_FUNC(NeonDepthwiseConv, CPU, NeonDepthwiseConvFunction); + +#endif + +} // namespace neon +} // namespace paddle diff --git a/paddle/function/neon/neon_util.h b/paddle/function/neon/neon_util.h new file mode 100644 index 0000000000..56b3febe2d --- /dev/null +++ b/paddle/function/neon/neon_util.h @@ -0,0 +1,47 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + +#include + +namespace paddle { + +namespace neon { + +inline float32x4_t vld1q_f32_aligned(const float* p) { + return vld1q_f32( + (const float*)__builtin_assume_aligned(p, sizeof(float32x4_t))); +} + +#ifndef __aarch64__ +inline float32_t vaddvq_f32(float32x4_t a) { + float32x2_t v = vadd_f32(vget_high_f32(a), vget_low_f32(a)); + return vget_lane_f32(vpadd_f32(v, v), 0); +} + +inline float32x4_t vmlaq_laneq_f32(float32x4_t a, + float32x4_t b, + float32x4_t v, + const int lane) { + return vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane)); +} +#endif + +} // namespace neon +} // namespace paddle + +#endif diff --git a/paddle/function/nnpack/NNPACKConvOp.cpp b/paddle/function/nnpack/NNPACKConvOp.cpp index 00d048eb21..6ccc487cf1 100644 --- a/paddle/function/nnpack/NNPACKConvOp.cpp +++ b/paddle/function/nnpack/NNPACKConvOp.cpp @@ -196,30 +196,30 @@ public: CHECK_EQ(status, nnp_status_success); } } else { - for (size_t g = 0; g < groups_; g++) { - // only supports stride = 1 - CHECK_EQ(strideH(), 1); - CHECK_EQ(strideW(), 1); - nnp_status status = - nnp_convolution_output(algorithm_, - batchSize, - inputChannels / groups_, - outputChannels / groups_, - inputSize, - padding, - kernelSize, - 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); - } + // 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, + outputChannels, + inputSize, + padding, + kernelSize, + inputData, + filterData, + nullptr, /* bias */ + outputData, + bufferPtr, + sizePtr, + nnp_activation_identity, + nullptr, + threadpool_, /* threadpool */ + nullptr); + CHECK_EQ(status, nnp_status_success); } } diff --git a/paddle/function/nnpack/NNPACKConvOpTest.cpp b/paddle/function/nnpack/NNPACKConvOpTest.cpp index 4818011211..4dd3982487 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 0012636b8f..62cff9361c 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 5de2170877..78e958e06f 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_); diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.cpp b/paddle/gserver/gradientmachines/NeuralNetwork.cpp index cfa80a8936..26cff3e677 100644 --- a/paddle/gserver/gradientmachines/NeuralNetwork.cpp +++ b/paddle/gserver/gradientmachines/NeuralNetwork.cpp @@ -202,7 +202,7 @@ void NeuralNetwork::prefetch(const std::vector& inArgs) { auto mat = dynamic_cast( para->getMat(PARAMETER_VALUE).get()); para->clearGradient(); - mat->clearIndices(); + if (mat) mat->clearIndices(); } } } diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp index f98bf95064..9f29b97466 100644 --- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp +++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp @@ -184,7 +184,7 @@ public: } void backward(const UpdateCallback& callback) override { - if (biases_) { + if (biases_ && biases_->getWGrad()) { backwardActivation(); biases_->getWGrad()->collectBias(*getOutputGrad(), 1); biases_->getParameterPtr()->incUpdate(callback); @@ -1012,11 +1012,6 @@ void RecurrentGradientMachine::generateSequence() { /* width */ resultNum, false, /* useGpu */ false); - Matrix::resizeOrCreate(generator_.outArg.value, - /* height */ maxGenWordCount, - /* width */ 1, - false, - /* useGpu */ false); } ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions, numSequences + 1, @@ -1026,7 +1021,7 @@ void RecurrentGradientMachine::generateSequence() { } else { oneWaySearch(numSequences); } - if (dataArgsSize_) createDataOutlink(batchMachineIdVec_); + if (dataArgsSize_) createDataOutlink(); size_t size = generator_.ids.size(); generator_.outArg.ids->resize(size); @@ -1106,6 +1101,7 @@ void RecurrentGradientMachine::oneWaySearch(size_t batchSize) { } batchMachineIdVec_.clear(); + batchMachineStartPos_.clear(); int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false); starts[0] = 0; generator_.ids.clear(); @@ -1312,13 +1308,20 @@ void RecurrentGradientMachine::fillGenOutputs() { finalPaths_[i].resize(minFinalPathsSize); } - batchMachineIdVec_.clear(); generator_.ids.clear(); int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false); starts[0] = 0; if (numResults > 1) { - real* probs = generator_.outArg.in->getData(); + int idsProbSaveSize = 0; + for (auto inSeq : finalPaths_) { + for (auto path : inSeq) idsProbSaveSize += path.ids.size(); + idsProbSaveSize += inSeq.size(); + } + Matrix::resizeOrCreate( + generator_.outArg.value, idsProbSaveSize, 1, false, false); real* idsProb = generator_.outArg.value->getData(); + + real* probs = generator_.outArg.in->getData(); size_t curPos = 0; for (size_t i = 0; i < finalPaths_.size(); ++i) { for (size_t j = 0; j < finalPaths_[i].size(); ++j) { @@ -1333,24 +1336,16 @@ void RecurrentGradientMachine::fillGenOutputs() { curPos += genLen; idsProb[curPos++] = -1.0; probs[i * numResults + j] = path.logProb; - - if (!j && dataArgsSize_) { - // in beam search, here only reserved the top 1 generated result - // for out_links that are not the generated word indices. - batchMachineIdVec_.insert(batchMachineIdVec_.end(), - path.machineIdVec.begin(), - path.machineIdVec.end()); - } } starts[i + 1] = generator_.ids.size(); } } else { for (size_t i = 0; i < finalPaths_.size(); ++i) { CHECK(!finalPaths_[i].empty()); - generator_.ids.insert(generator_.ids.begin(), - finalPaths_[i][0].ids.begin(), - finalPaths_[i][0].ids.end()); - starts[i + 1] = starts[i] + finalPaths_[i][0].ids.size(); + Path& path = finalPaths_[i][0]; + generator_.ids.insert( + generator_.ids.end(), path.ids.begin(), path.ids.end()); + starts[i + 1] = starts[i] + path.ids.size(); } } } @@ -1364,25 +1359,76 @@ void RecurrentGradientMachine::copyDataOutlinkFrame(size_t machineCur) { } } -void RecurrentGradientMachine::createDataOutlink( - std::vector& machineIdVec) { - size_t seqNum = - getBeamSize() > 1UL ? finalPaths_.size() : finalPaths_[0].size(); - std::vector starts(seqNum + 1, 0); - for (size_t i = 0; i < seqNum; ++i) { - size_t seqLen = getBeamSize() > 1UL ? finalPaths_[i][0].ids.size() - : finalPaths_[0][i].ids.size(); - starts[i + 1] = starts[i] + seqLen; +void RecurrentGradientMachine::createDataOutlinkSelRowsInfo( + bool isSeq, std::vector& outArgs) { + batchMachineIdVec_.clear(); + + size_t seqIdx = 0; + for (size_t i = 0; i < finalPaths_.size(); ++i) { + for (size_t j = 0; j < finalPaths_[i].size(); ++j) { + std::vector& machineIdVec = finalPaths_[i][j].machineIdVec; + if (isSeq) { + for (size_t i = 0; i < machineIdVec.size(); ++i) { + size_t rowId = machineIdVec[i]; + int* seqPos = + outArgs[i].sequenceStartPositions->getMutableData(false); + batchMachineIdVec_.push_back(seqPos[rowId]); + } + } else { + batchMachineIdVec_.insert( + batchMachineIdVec_.end(), machineIdVec.begin(), machineIdVec.end()); + } + seqIdx++; + } + } +} + +void RecurrentGradientMachine::createDataOutlinkCopySizeInfo( + bool isSeq, std::vector& outArgs, std::vector& copySize) { + size_t totalSeqNum = std::accumulate( + finalPaths_.begin(), + finalPaths_.end(), + 0UL, + [](size_t a, const std::vector& b) { return a + b.size(); }); + copySize.resize(totalSeqNum, 1); + + batchMachineStartPos_.resize(totalSeqNum + 1, 0); + if (isSeq) { + ICpuGpuVectorPtr inputSeqStartPos = outArgs[0].sequenceStartPositions; + CHECK_EQ(static_cast(inputSeqStartPos->getSize() - 1), + getBeamSize() > 1 ? finalPaths_.size() : finalPaths_[0].size()); + int* starts = inputSeqStartPos->getMutableData(false); + int seqId = 0; + for (size_t i = 0; i < finalPaths_.size(); ++i) { + for (size_t j = 0; j < finalPaths_[i].size(); ++j) { + copySize[seqId] = getBeamSize() > 1 ? starts[i + 1] - starts[i] + : starts[j + 1] - starts[j]; + batchMachineStartPos_[seqId + 1] = + batchMachineStartPos_[seqId] + finalPaths_[i][j].ids.size(); + seqId++; + } + } + } else { + for (size_t i = 0; i < finalPaths_[0].size(); ++i) + batchMachineStartPos_[i + 1] = + batchMachineStartPos_[i] + finalPaths_[0][i].ids.size(); } +} +void RecurrentGradientMachine::createDataOutlink() { for (size_t i = 0; i < dataArgsSize_; i++) { + bool isSeq = dataArgsFrame_[i][0].hasSeq(); + std::vector copySize; + createDataOutlinkCopySizeInfo(isSeq, dataArgsFrame_[i], copySize); + createDataOutlinkSelRowsInfo(isSeq, dataArgsFrame_[i]); + dataArgs_[i].concat(dataArgsFrame_[i], - machineIdVec, - starts, + batchMachineIdVec_, + batchMachineStartPos_, + copySize, useGpu_, HPPL_STREAM_1, PASS_TEST); - auto dataAgent = dynamic_cast(outFrameLines_[i + 1].agentLayer.get()); CHECK_NOTNULL(dataAgent); diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h index fb3fc5877a..c16fae6d17 100644 --- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h +++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h @@ -190,7 +190,7 @@ public: std::vector ids; /** - * @brief idsProb, log probability of each generated words. + * @brief idsProb, log probability of each generated word. */ std::vector idsProb; @@ -472,15 +472,43 @@ private: void copyDataOutlinkFrame(size_t machineCur); /* - * @brief In generation, if the layer group has more than 1 outlink, outlinks - * except the first one are data outlinks. This function creates the data - * outlinks. - * @note In beam search, only one generated sequence with the hightest log - * probabilites are retained. - * @param machineIdVec : select a row of output matrix in each frame - * that the generation process expanded. + * @brief In generation, if the layer group has more than 1 outlink, outlink + * except the first one is a data outlink. In RecurrentLayerGroup, each time + * step is a separate Network, outputs of a layer inside the + * RecurrentLayerGroup are stored in separate Arguments. If one layer is + * specified as an outlink of RecurrentLayerGroup. This function will + * collect outputs in each time step of each generated sequence which are + * dispersed in separate Arguments to form a new single Argument as output of + * RecurrentLayerGroup. */ - void createDataOutlink(std::vector& machineIdVec); + void createDataOutlink(); + + /* + * @brief decide to select how many rows from the Matrix stored the forward + * pass results from a start position. + * + * @param isSeq: a flag indicating whetehr the layer to be output of the + * RecurrentGradientMachine is a sequence or not + * @param outArgs: all of the the returned Arguments of the forward pass + * during the generation process. + * @param copySize: the returned result, number of rows to select from the + * Matrix stored the forward pass results from a start position. + */ + void createDataOutlinkCopySizeInfo(bool isSeq, + std::vector& outArgs, + std::vector& copySize); + + /* + * @brief decide index of the start row for each time step of a generated + * sequence in Matrix stored the entire beam search batch's forward pass + * results. + * + * @param isSeq: a flag indicating whether the layer to be output of the + * RecurrentGradientMachine is a sequence or not + * @param outArgs: all of the returned Arguments of the forward pass + * during the generation process. + */ + void createDataOutlinkSelRowsInfo(bool isSeq, std::vector& outArgs); /* * @brief used in beam search, connect previous frame to form recurrent link @@ -543,6 +571,7 @@ private: std::vector topIds_; std::vector seqIds_; std::vector batchMachineIdVec_; + std::vector batchMachineStartPos_; std::vector> finalPaths_; std::vector minFinalPathLogProb_; BeamSearchControlCallbacks* beamSearchCtrlCallbacks_; diff --git a/paddle/gserver/layers/ConvBaseLayer.cpp b/paddle/gserver/layers/ConvBaseLayer.cpp index e161d89c38..a5328ef834 100644 --- a/paddle/gserver/layers/ConvBaseLayer.cpp +++ b/paddle/gserver/layers/ConvBaseLayer.cpp @@ -32,9 +32,11 @@ bool ConvBaseLayer::init(const LayerMap& layerMap, const ConvConfig& conf = inputConfig.conv_conf(); padding_.push_back(conf.padding()); stride_.push_back(conf.stride()); + dilation_.push_back(conf.dilation()); filterSize_.push_back(conf.filter_size()); paddingY_.push_back(conf.padding_y()); strideY_.push_back(conf.stride_y()); + dilationY_.push_back(conf.dilation_y()); filterSizeY_.push_back(conf.filter_size_y()); filterPixels_.push_back(filterSize_.back() * filterSizeY_.back()); channels_.push_back(conf.channels()); @@ -89,7 +91,11 @@ size_t ConvBaseLayer::calOutputSize() { size_t layerSize = 0; auto setLayerSize = [&](IntV& inH, IntV& inW, IntV& outH, IntV& outW) { + size_t filterSizeY; + size_t filterSize; for (size_t i = 0; i < inputLayers_.size(); i++) { + filterSizeY = (filterSizeY_[i] - 1) * dilationY_[i] + 1; + filterSize = (filterSize_[i] - 1) * dilation_[i] + 1; inH.push_back(inputLayers_[i]->getOutput().getFrameHeight()); inW.push_back(inputLayers_[i]->getOutput().getFrameWidth()); const ConvConfig& conf = config_.inputs(i).conv_conf(); @@ -98,17 +104,17 @@ size_t ConvBaseLayer::calOutputSize() { inH[i] = conf.has_output_y() ? conf.output_y() : conf.output_x(); if (inW[i] == 0) inW[i] = conf.output_x(); outH.push_back(imageSize( - inH[i], filterSizeY_[i], paddingY_[i], strideY_[i], caffeMode_)); - outW.push_back(imageSize( - inW[i], filterSize_[i], padding_[i], stride_[i], caffeMode_)); + inH[i], filterSizeY, paddingY_[i], strideY_[i], caffeMode_)); + outW.push_back( + imageSize(inW[i], filterSize, padding_[i], stride_[i], caffeMode_)); } else { if (inH[i] == 0) inH[i] = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size(); if (inW[i] == 0) inW[i] = conf.img_size(); outH.push_back(outputSize( - inH[i], filterSizeY_[i], paddingY_[i], strideY_[i], caffeMode_)); + inH[i], filterSizeY, paddingY_[i], strideY_[i], caffeMode_)); outW.push_back(outputSize( - inW[i], filterSize_[i], padding_[i], stride_[i], caffeMode_)); + inW[i], filterSize, padding_[i], stride_[i], caffeMode_)); } CHECK_EQ(outH[i], outH[0]); CHECK_EQ(outW[i], outW[0]); diff --git a/paddle/gserver/layers/ConvBaseLayer.h b/paddle/gserver/layers/ConvBaseLayer.h index e9d15d94f8..223bce8e29 100644 --- a/paddle/gserver/layers/ConvBaseLayer.h +++ b/paddle/gserver/layers/ConvBaseLayer.h @@ -40,6 +40,10 @@ protected: IntV stride_; /// The y dimension of the stride. IntV strideY_; + /// The x dimension of the dilation. + IntV dilation_; + /// The y dimension of the dilation. + IntV dilationY_; /// The x dimension of a filter kernel. IntV filterSize_; /// The y dimension of a filter kernel. diff --git a/paddle/gserver/layers/ConvBaseOperator.cpp b/paddle/gserver/layers/ConvBaseOperator.cpp index 5c23198629..5469c41c87 100644 --- a/paddle/gserver/layers/ConvBaseOperator.cpp +++ b/paddle/gserver/layers/ConvBaseOperator.cpp @@ -59,7 +59,8 @@ void ConvBaseOperator::allocConvWorkSpace() { &bwdDataAlgo_, &bwdDataLimitBytes_, &bwdFilterAlgo_, - &bwdFilterLimitBytes_); + &bwdFilterLimitBytes_, + /*useDilation*/ false); size_t maxWorkSpace = 0; maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_); diff --git a/paddle/gserver/layers/ConvBaseProjection.cpp b/paddle/gserver/layers/ConvBaseProjection.cpp index eb6b0445c9..08f36c516c 100644 --- a/paddle/gserver/layers/ConvBaseProjection.cpp +++ b/paddle/gserver/layers/ConvBaseProjection.cpp @@ -41,6 +41,11 @@ void ConvBaseProjection::getConvParams() { strideH_ = conf.stride_y(); strideW_ = conf.stride(); + dilationH_ = conf.dilation_y(); + dilationW_ = conf.dilation(); + CHECK_GT(dilationH_, 0); + CHECK_GT(dilationW_, 0); + filterH_ = conf.filter_size_y(); filterW_ = conf.filter_size(); @@ -77,7 +82,9 @@ void ConvBaseProjection::initCudnn() { paddingH_, paddingW_, strideH_, - strideW_); + strideW_, + dilationH_, + dilationW_); // initialize all to default algorithms fwdAlgo_ = 0; @@ -131,7 +138,9 @@ void ConvBaseProjection::reshapeTensorDesc(int batchSize) { paddingH_, paddingW_, strideH_, - strideW_); + strideW_, + dilationH_, + dilationW_); } void ConvBaseProjection::reshape(int batchSize) { @@ -140,6 +149,10 @@ void ConvBaseProjection::reshape(int batchSize) { CHECK_EQ(calInputSize(), in_->value->getWidth()); reshapeTensorDesc(batchSize); + bool useDilation = false; + if (dilationH_ > 1 || dilationW_ > 1) { + useDilation = true; + } hl_conv_workspace(imageDesc_, outputDesc_, filterDesc_, @@ -149,7 +162,8 @@ void ConvBaseProjection::reshape(int batchSize) { &bwdDataAlgo_, &bwdDataLimitBytes_, &bwdFilterAlgo_, - &bwdFilterLimitBytes_); + &bwdFilterLimitBytes_, + useDilation); size_t maxWorkSpace = 0; maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_); diff --git a/paddle/gserver/layers/ConvBaseProjection.h b/paddle/gserver/layers/ConvBaseProjection.h index e9d9f8f1b2..ebdb57845b 100644 --- a/paddle/gserver/layers/ConvBaseProjection.h +++ b/paddle/gserver/layers/ConvBaseProjection.h @@ -63,6 +63,7 @@ protected: int configChannels_, configNumFilters_; int paddingH_, paddingW_; int strideH_, strideW_; + int dilationH_, dilationW_; int filterH_, filterW_; /// One group offset of input data. int inputOffset_; diff --git a/paddle/gserver/layers/ConvProjection.cpp b/paddle/gserver/layers/ConvProjection.cpp index 5b7ecc5560..6f0106b713 100644 --- a/paddle/gserver/layers/ConvProjection.cpp +++ b/paddle/gserver/layers/ConvProjection.cpp @@ -25,12 +25,12 @@ size_t ConvProjection::calOutputSize() { if (imageH_ == 0) imageH_ = configImgH_; if (imageW_ == 0) imageW_ = configImgW_; outputH_ = outputSize(imageH_, - filterH_, + (filterH_ - 1) * dilationH_ + 1, paddingH_, strideH_, /* caffeMode */ true); outputW_ = outputSize(imageW_, - filterW_, + (filterW_ - 1) * dilationW_ + 1, paddingW_, strideW_, /* caffeMode */ true); diff --git a/paddle/gserver/layers/CostLayer.cpp b/paddle/gserver/layers/CostLayer.cpp index 6bfdea3c6e..ce071323ff 100644 --- a/paddle/gserver/layers/CostLayer.cpp +++ b/paddle/gserver/layers/CostLayer.cpp @@ -572,13 +572,8 @@ void MultiBinaryLabelCrossEntropy::backwardImp(Matrix& output, } } -// -// Huber loss for robust 2-classes classification -// -REGISTER_LAYER(huber, HuberTwoClass); - -bool HuberTwoClass::init(const LayerMap& layerMap, - const ParameterMap& parameterMap) { +bool HuberCost::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { CostLayer::init(layerMap, parameterMap); if (useGpu_) { tmpCpuInput_.reserve(inputLayers_.size()); @@ -589,7 +584,7 @@ bool HuberTwoClass::init(const LayerMap& layerMap, return true; } -void HuberTwoClass::forwardImp(Matrix& output, Argument& label, Matrix& cost) { +void HuberCost::forwardImp(Matrix& output, Argument& label, Matrix& cost) { if (useGpu_) { for (size_t i = 0; i < inputLayers_.size(); i++) { tmpCpuInput_[i].resizeAndCopyFrom( @@ -597,13 +592,87 @@ void HuberTwoClass::forwardImp(Matrix& output, Argument& label, Matrix& cost) { } hl_stream_synchronize(HPPL_STREAM_DEFAULT); } - forwardImpIn(output, label, cost); } -void HuberTwoClass::forwardImpIn(Matrix& output, - Argument& label, - Matrix& target) { +// +// Huber loss for robust regression. +// +REGISTER_LAYER(huber_regression, HuberRegressionLoss); + +bool HuberRegressionLoss::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + HuberCost::init(layerMap, parameterMap); + delta_ = config_.delta(); + return true; +} + +void HuberRegressionLoss::forwardImp(Matrix& output, + Argument& label, + Matrix& target) { + HuberCost::forwardImp(output, label, target); + size_t numSamples = target.getHeight(); + size_t dim = output.getWidth(); + CHECK(label.value); + CHECK_EQ((*label.value).getHeight(), numSamples); + CHECK_EQ(output.getHeight(), numSamples); + CHECK_EQ(dim, (*label.value).getWidth()); + CHECK_EQ(target.getWidth(), (size_t)1); + + real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); + real* lbl = + useGpu_ ? tmpCpuInput_[1].value->getData() : (*label.value).getData(); + std::vector cost(numSamples, 0); + for (size_t i = 0; i < numSamples; ++i) { + for (size_t j = 0; j < dim; ++j) { + int index = i * dim + j; + real a = std::abs(lbl[index] - out[index]); + if (a <= delta_) + cost[i] += a * a / 2; + else + cost[i] += delta_ * (a - delta_ / 2); + } + } + target.copyFrom(cost.data(), numSamples); +} + +void HuberRegressionLoss::backwardImp(Matrix& output, + Argument& label, + Matrix& outputG) { + size_t numSamples = output.getHeight(); + size_t dim = output.getWidth(); + real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); + real* lbl = + useGpu_ ? tmpCpuInput_[1].value->getData() : (*label.value).getData(); + real* grad = useGpu_ ? tmpCpuInput_[0].grad->getData() : outputG.getData(); + for (size_t i = 0; i < numSamples; ++i) { + for (size_t j = 0; j < dim; ++j) { + int index = i * dim + j; + real a = lbl[index] - out[index]; + if (std::abs(a) <= delta_) + grad[index] += -a; + else + grad[index] += a > 0 ? -delta_ : delta_; + } + } + if (useGpu_) outputG.copyFrom(grad, numSamples * dim); +} + +// +// Huber loss for robust 2-classes classification +// +REGISTER_LAYER(huber_classification, HuberTwoClassification); + +bool HuberTwoClassification::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + return HuberCost::init(layerMap, parameterMap); +} + +void HuberTwoClassification::forwardImp(Matrix& output, + Argument& label, + Matrix& target) { + HuberCost::forwardImp(output, label, target); size_t numSamples = target.getHeight(); + CHECK(label.ids); CHECK_EQ((*label.ids).getSize(), numSamples); CHECK_EQ(output.getHeight(), numSamples); CHECK_EQ(output.getWidth(), (size_t)1); @@ -611,47 +680,35 @@ void HuberTwoClass::forwardImpIn(Matrix& output, real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData(); - std::vector cost(numSamples); + std::vector cost(numSamples, 0); for (size_t i = 0; i < numSamples; ++i) { int y = 2 * lbl[i] - 1; - if (out[i] * y < -1) - cost[i] = -4 * out[i] * y; - else if (out[i] * y < 1) - cost[i] = (1 - out[i] * y) * (1 - out[i] * y); - else - cost[i] = 0; + real a = out[i] * y; + if (a < -1) + cost[i] = -4 * a; + else if (a < 1) + cost[i] = (1 - a) * (1 - a); } target.copyFrom(cost.data(), numSamples); } -void HuberTwoClass::backwardImp(Matrix& outputValue, - Argument& label, - Matrix& outputGrad) { - if (useGpu_) { - backwardImpIn( - *tmpCpuInput_[0].value, tmpCpuInput_[1], *tmpCpuInput_[0].grad); - outputGrad.copyFrom(*tmpCpuInput_[0].grad); - } else { - backwardImpIn(outputValue, label, outputGrad); - } -} - -void HuberTwoClass::backwardImpIn(Matrix& output, - Argument& label, - Matrix& outputG) { +void HuberTwoClassification::backwardImp(Matrix& output, + Argument& label, + Matrix& outputG) { size_t numSamples = output.getHeight(); - real* out = output.getData(); - real* grad = outputG.getData(); - int* lbl = (*label.ids).getData(); + real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData(); + int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData(); + real* grad = useGpu_ ? tmpCpuInput_[0].grad->getData() : outputG.getData(); for (size_t i = 0; i < numSamples; ++i) { int y = 2 * lbl[i] - 1; - if (y * out[i] < -1) + real a = out[i] * y; + if (a < -1) grad[i] += -4 * y; - else if (y * out[i] < 1) - grad[i] += -2 * (1 - y * out[i]) * y; + else if (a < 1) + grad[i] += -2 * (1 - a) * y; } + if (useGpu_) outputG.copyFrom(grad, numSamples); } - /** * This cost layer compute the sum of its input as loss. * \f[ diff --git a/paddle/gserver/layers/CostLayer.h b/paddle/gserver/layers/CostLayer.h index 14c0b33ec1..0ce72ef40a 100644 --- a/paddle/gserver/layers/CostLayer.h +++ b/paddle/gserver/layers/CostLayer.h @@ -304,37 +304,68 @@ public: Matrix& outputGrad) override; }; -/** - * Huber loss for robust 2-classes classification. - * - * For label={0, 1}, let y=2*label-1. Given output f, the loss is: - * \f[ - * Loss = - * \left\{\begin{matrix} - * 4 * y * f & \textit{if} \ \ y* f < -1 \\ - * (1 - y * f)^2 & \textit{if} \ \ -1 < y * f < 1 \\ - * 0 & \textit{otherwise} - * \end{matrix}\right. - * \f] +/* + * A base layer for HuberRegressionLoss and HuberTwoClassification. */ -class HuberTwoClass : public CostLayer { +class HuberCost : public CostLayer { +public: std::vector tmpCpuInput_; -public: - explicit HuberTwoClass(const LayerConfig& config) : CostLayer(config) {} + explicit HuberCost(const LayerConfig& config) : CostLayer(config) {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) override; void forwardImp(Matrix& output, Argument& label, Matrix& cost) override; - void forwardImpIn(Matrix& output, Argument& label, Matrix& cost); + void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad) {} +}; + +/** + * Huber loss for robust regression. + * + * Given output f(x), label y and delta, the loss is: + * Loss = 0.5 * (1 - y * f)^2, if abs(y - f) <= delta \\ + * Loss = delta * abs(y - f) - 0.5 * delta^2, otherwise + */ +class HuberRegressionLoss : public HuberCost { +public: + explicit HuberRegressionLoss(const LayerConfig& config) : HuberCost(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forwardImp(Matrix& output, Argument& label, Matrix& cost) override; void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad) override; - void backwardImpIn(Matrix& outputValue, Argument& label, Matrix& outputGrad); +protected: + real delta_; +}; + +/** + * Huber loss for robust 2-classes classification. + * + * For label={0, 1}, let y=2*label-1. Given output f(x), the loss is: + * Loss = 4 * y * f, if y* f < -1 \\ + * Loss = (1 - y * f)^2, if -1 < y * f < 1 \\ + * Loss = 0, otherwise + */ +class HuberTwoClassification : public HuberCost { +public: + explicit HuberTwoClassification(const LayerConfig& config) + : HuberCost(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forwardImp(Matrix& output, Argument& label, Matrix& cost) override; + + void backwardImp(Matrix& outputValue, + Argument& label, + Matrix& outputGrad) override; }; typedef std::shared_ptr CostLayerPtr; diff --git a/paddle/gserver/layers/ExpandConvLayer.cpp b/paddle/gserver/layers/ExpandConvLayer.cpp index 0ece279931..20de475fc3 100644 --- a/paddle/gserver/layers/ExpandConvLayer.cpp +++ b/paddle/gserver/layers/ExpandConvLayer.cpp @@ -29,6 +29,10 @@ namespace paddle { REGISTER_LAYER(exconv, ExpandConvLayer); REGISTER_LAYER(exconvt, ExpandConvLayer); +inline bool isDepthwiseConv(int channels, int groups) { + return channels == groups; +} + bool ExpandConvLayer::init(const LayerMap &layerMap, const ParameterMap ¶meterMap) { /* Initialize the basic convolutional parent class */ @@ -47,14 +51,27 @@ bool ExpandConvLayer::init(const LayerMap &layerMap, std::vector paddings = {(size_t)paddingY_[i], (size_t)padding_[i]}; std::vector strides = {(size_t)strideY_[i], (size_t)stride_[i]}; - if (useGpu_ && (size_t)groups_[i] == (size_t)channels_[i] && !isDeconv_) { + // Convolution Layer uses the GemmConv function by default. + convType = "GemmConv"; + convGradInputType = "GemmConvGradInput"; + convGradFilterType = "GemmConvGradFilter"; + + // If depth wise convolution and useGpu == true + if (useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) { convType = "DepthwiseConv"; convGradInputType = "DepthwiseConvGradInput"; convGradFilterType = "DepthwiseConvGradFilter"; - } else { - convType = "GemmConv"; - convGradInputType = "GemmConvGradInput"; - convGradFilterType = "GemmConvGradFilter"; + } + + // If depth wise convolution and useGpu == false and ARM-NEON + if (!useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) { +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + if ((filterSize_[i] == filterSizeY_[i]) && + (filterSize_[i] == 3 || filterSize_[i] == 4) && + (stride_[i] == strideY_[i]) && (stride_[i] == 1 || stride_[i] == 2)) { + convType = "NeonDepthwiseConv"; + } +#endif } if (FLAGS_use_nnpack && !isDeconv_) { diff --git a/paddle/gserver/layers/KmaxSeqScoreLayer.cpp b/paddle/gserver/layers/KmaxSeqScoreLayer.cpp index 8ce591d476..d5407555b2 100644 --- a/paddle/gserver/layers/KmaxSeqScoreLayer.cpp +++ b/paddle/gserver/layers/KmaxSeqScoreLayer.cpp @@ -80,13 +80,14 @@ void KmaxSeqScoreLayer::forward(PassType passType) { << "input of " << getName() << " must be a sequence or a nested sequence."; CHECK_EQ(input.value->getWidth(), 1UL) - << "input of " << getName() - << " is score over a sequence or a nested sequence, so its width " - << " must be 1."; + << "input of " << getName() << " are scores over a sequence or " + << "a nested sequence, so its width must be 1."; if (useGpu_) { - // this Layer runs only in CPU, if the model is runing on GPU, - // then copy the input to this layer from GPU to CPU. + /* + * currently, this Layer only runs in CPU, if the other part of the model is + * runing on GPU, then copy the input to this layer from GPU to CPU. + */ Matrix::resizeOrCreate(scores_, inputScore->getHeight(), 1, @@ -97,6 +98,14 @@ void KmaxSeqScoreLayer::forward(PassType passType) { scores_ = inputScore; } + /* + * TODO(caoying) + * In PaddePaddle, currently all matrices are real number types, + * but output of this layer which is some selected indices of the give + * sequence are actually filled with int types so that storing int types + * information in a real number matrix is dangerous, since real numbers will + * be convered to int types. + */ Matrix::resizeOrCreate( output_.value, input.hasSubseq() ? input.getNumSubSequences() : input.getNumSequences(), diff --git a/paddle/gserver/layers/Layer.cpp b/paddle/gserver/layers/Layer.cpp index d5621412ca..2bc20eee6c 100644 --- a/paddle/gserver/layers/Layer.cpp +++ b/paddle/gserver/layers/Layer.cpp @@ -41,7 +41,7 @@ namespace paddle { Layer::Layer(const LayerConfig& config, bool useGpu) : config_(config), useGpu_(useGpu), - deviceId_(-1), + deviceId_(CPU_DEVICE), needSequenceInfo_(true) {} bool Layer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) { diff --git a/paddle/gserver/layers/Layer.h b/paddle/gserver/layers/Layer.h index 0ed482889d..edef36194a 100644 --- a/paddle/gserver/layers/Layer.h +++ b/paddle/gserver/layers/Layer.h @@ -59,7 +59,12 @@ protected: LayerConfig config_; /// whether to use GPU bool useGpu_; - /// Device Id. CPU is -1, and GPU is 0, 1, 2 ... + /// Paddle device ID, MKLDNN is -2, CPU is -1 + enum PADDLE_DEVICE_ID { + MKLDNN_DEVICE = -2, + CPU_DEVICE = -1, + }; + /// Device Id. MKLDNN is -2, CPU is -1, and GPU is 0, 1, 2 ... int deviceId_; /// Input layers std::vector inputLayers_; @@ -77,6 +82,7 @@ protected: Argument output_; /// Several outputs stored on different devices, used in 'parallel_nn' case, /// and record them by deviceId_. + /// Also used in 'use_mkldnn' case. std::vector outputOtherDevice_; /// If there are several outputs, map them by each name. std::map outputMap_; @@ -172,6 +178,13 @@ protected: return inputLayer.getOutput(deviceId_); } + /** + * Get the argument of input layer with deviceId. + */ + const Argument& getInput(size_t inputIndex, int deviceId) const { + return inputLayers_[inputIndex]->getOutput(deviceId); + } + /** * Get the forward-input value. */ @@ -186,6 +199,13 @@ protected: return inputLayer.getOutput(deviceId_).value; } + /** + * Get the forward-input value with deviceId. + */ + const MatrixPtr& getInputValue(int inputIndex, int deviceId) { + return inputLayers_[inputIndex]->getOutput(deviceId).value; + } + /** * Get the forward-input grad. */ @@ -200,6 +220,13 @@ protected: return inputLayer.getOutput(deviceId_).grad; } + /** + * Get the forward-input grad. + */ + const MatrixPtr& getInputGrad(int inputIndex, int deviceId) { + return inputLayers_[inputIndex]->getOutput(deviceId).grad; + } + /** * Get the forward-input label. */ diff --git a/paddle/gserver/layers/MKLDNNBase.h b/paddle/gserver/layers/MKLDNNBase.h new file mode 100644 index 0000000000..4c0234e7b3 --- /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 0000000000..8318c8c519 --- /dev/null +++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp @@ -0,0 +1,302 @@ +/* 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 (hasInitedWgt_) { + return; + } + + CHECK(wgtVal_) << "should have been initialized"; + bool hasNoSpatial_ = ih_ == 1 && iw_ == 1; + auto targetDim = wgtVal_->getDims(); + auto srcFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo; + wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim); + hasInitedWgt_ = true; +} + +void MKLDNNFcLayer::convertWeightsToPaddle() { + CHECK(wgtVal_) << "should have been initialized"; + bool hasNoSpatial_ = ih_ == 1 && iw_ == 1; + auto targetDim = wgtVal_->getDims(); + auto dstFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo; + wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim); +} + +void MKLDNNFcLayer::convertOutputToOtherDevice() { + copyOutputInfoToOtherDevice(); + // find other cpu device and reorder output to cpu device + int cnt = 0; + for (size_t i = 0; i < outputOtherDevice_.size(); i++) { + if (outputOtherDevice_[i].deviceId == CPU_DEVICE) { + // fc cpu output value do not need convert + // just share point + outputOtherDevice_[i].value = output_.value; + ++cnt; + } + } + + if (cnt > 1) { + LOG(WARNING) << "should not have more than one CPU devie"; + } +} + +void MKLDNNFcLayer::reshape() { + const Argument& input = getInput(0, getPrev(0)->getDeviceId()); + 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; + } + 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(); + const MatrixPtr& wgt = weight_->getW(); + const MatrixPtr& bias = hasBias ? biases_->getW() : nullptr; + const MatrixPtr& out = output_.value; + + if (inputIsOnlyMKLDNN()) { + const MatrixPtr& in = getInputValue(0); + inVal_ = std::dynamic_pointer_cast(in); + CHECK(inVal_) << "Input should be MKLDNNMatrix"; + } else { + CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet"; + const MatrixPtr& in = getInputValue(0, CPU_DEVICE); + inVal_ = MKLDNNMatrix::create( + in, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_); + } + inVal_->downSpatial(); + wgtVal_ = MKLDNNMatrix::create( + wgt, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_); + wgtVal_->downSpatial(); + biasVal_ = + hasBias ? MKLDNNMatrix::create(bias, {oc_}, format::x, engine_) : nullptr; + outVal_ = MKLDNNMatrix::create(out, {bs_, oc_}, format::nc, engine_); + + // change original output value to mkldnn output value + output_.value = std::dynamic_pointer_cast(outVal_); + if (!outputIsOnlyMKLDNN()) { + convertOutputToOtherDevice(); + } + + // create forward handle + prop_kind pk = prop_kind::forward; + fc_fwd::desc fwdDesc = hasBias ? fc_fwd::desc(pk, + inVal_->getMemoryDesc(), + wgtVal_->getMemoryDesc(), + biasVal_->getMemoryDesc(), + outVal_->getMemoryDesc()) + : fc_fwd::desc(pk, + inVal_->getMemoryDesc(), + wgtVal_->getMemoryDesc(), + outVal_->getMemoryDesc()); + fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_); + if (hasBias) { + fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *biasVal_, *outVal_)); + } else { + fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *outVal_)); + } + printValueFormatFlow(); + + pipelineFwd_.clear(); + pipelineFwd_.push_back(*fwd_); +} + +void MKLDNNFcLayer::resetBwd() { + if (!needResetBwd_) { + return; + } + needResetBwd_ = false; + bool hasBias = biases_ && biases_->getWGrad(); + + /// backward weight + CHECK(inVal_) << "Should have input value"; + const MatrixPtr& wgt = weight_->getWGrad(); + const MatrixPtr& bias = hasBias ? biases_->getWGrad() : nullptr; + + // TODO(TJ): merge outgrad + int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE; + // for MKLDNN device: + // can not directly cast outputgrad to mkldnnmatrix, + // since each layer can not write the inputgrad to mkldnn inputgrad. + // So just create from matrix with outputvalue format. + // for CPU device: + // fc do not need to convert from cpu device since output is always nc format + // only need create from cpu device + const MatrixPtr& out = getOutput(device).grad; + outGrad_ = MKLDNNMatrix::create(out, outVal_->getPrimitiveDesc()); + wgtGrad_ = MKLDNNMatrix::create(wgt, wgtVal_->getPrimitiveDesc()); + biasGrad_ = hasBias ? MKLDNNMatrix::create(bias, biasVal_->getPrimitiveDesc()) + : nullptr; + + // create memory primitive desc + fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward, + inVal_->getMemoryDesc(), + wgtGrad_->getMemoryDesc(), + outGrad_->getMemoryDesc()); + fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_); + fc_bwdWgt::desc bwdWgtDesc = hasBias + ? fc_bwdWgt::desc(inVal_->getMemoryDesc(), + wgtGrad_->getMemoryDesc(), + biasGrad_->getMemoryDesc(), + outGrad_->getMemoryDesc()) + : fc_bwdWgt::desc(inVal_->getMemoryDesc(), + wgtGrad_->getMemoryDesc(), + outGrad_->getMemoryDesc()); + fc_bwdWgt::primitive_desc bwdWgtPD = + fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD); + + if (hasBias) { + 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 + device = inputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE; + const MatrixPtr& in = getInputGrad(0, device); + if (in == nullptr) { + return; + } + if (getInput(0, device).getAllCount() > 1) { + // TODO(TJ): use outputMaps_ ways when merge outgrad done + } else { + inGrad_ = MKLDNNMatrix::create(in, inVal_->getPrimitiveDesc()); + } + + fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(inVal_->getMemoryDesc(), + wgtGrad_->getMemoryDesc(), + outGrad_->getMemoryDesc()); + fc_bwdData::primitive_desc bwdDataPD = + fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD); + + CHECK(wgtVal_) << "Should have weight memory"; + bwdData_.reset(new fc_bwdData(bwdDataPD, *outGrad_, *wgtVal_, *inGrad_)); + printGradFormatFlow(); + pipelineBwd_.push_back(*bwdData_); +} + +void MKLDNNFcLayer::forward(PassType passType) { + Layer::forward(passType); + reshape(); + + { + REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str()); + syncInputValue(); + + // 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(); + + syncOutputGrad(); + // 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 0000000000..e138a6faf1 --- /dev/null +++ b/paddle/gserver/layers/MKLDNNFcLayer.h @@ -0,0 +1,79 @@ +/* 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_; + + // fc weight and bias + std::unique_ptr weight_; + std::unique_ptr biases_; + +public: + explicit MKLDNNFcLayer(const LayerConfig& config) + : MKLDNNLayer(config), hasInitedWgt_(false) {} + + ~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(); + + void convertOutputToOtherDevice() override; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h new file mode 100644 index 0000000000..b983b833d5 --- /dev/null +++ b/paddle/gserver/layers/MKLDNNLayer.h @@ -0,0 +1,252 @@ +/* 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" +#include "paddle/math/MKLDNNMatrix.h" + +DECLARE_bool(use_mkldnn); + +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_; + + // MKLDNNMatrixPtr + MKLDNNMatrixPtr inVal_; + MKLDNNMatrixPtr inGrad_; + MKLDNNMatrixPtr outVal_; + MKLDNNMatrixPtr outGrad_; + MKLDNNMatrixPtr wgtVal_; + MKLDNNMatrixPtr wgtGrad_; + MKLDNNMatrixPtr biasVal_; + MKLDNNMatrixPtr 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) { + CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn." + << "Please set WITH_MKLDNN=ON " + << "and set use_mkldnn=True"; + CHECK(!useGpu_) << "Do not support GPU yet"; + + // set device id before Layer::init + setDevice(MKLDNN_DEVICE); + // change param device to MKLDNN device + setParamsDevice(MKLDNN_DEVICE, parameterMap); + if (!Layer::init(layerMap, parameterMap)) { + return false; + } + + stream_.reset(new MKLDNNStream()); + engine_ = CPUEngine::Instance().getEngine(); + 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() {} + + /** + * convert MKLDNN output to other device. + * only support CPU device yet + */ + virtual void convertOutputToOtherDevice() {} + + /** + * 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_; + } + + /** + * Print the mkldnn memory format flow of value + */ + virtual void printValueFormatFlow() { + if (inVal_ && outVal_) { + VLOG(MKLDNN_FMTS) << "value format flow --- " << inVal_->getFormat() + << " >>> " << outVal_->getFormat(); + } + } + + /** + * Print the mkldnn memory format flow of grad + */ + virtual void printGradFormatFlow() { + if (inGrad_ && outGrad_) { + VLOG(MKLDNN_FMTS) << "grad format flow --- " << inGrad_->getFormat() + << " <<< " << outGrad_->getFormat(); + } + } + +protected: + /** + * copy image size and sequence info to other device + * @note: can not directly use Layer::copyOutputToOtherDevice since here only + * copy base info and do not copy data value + */ + void copyOutputInfoToOtherDevice() { + for (size_t i = 0; i < outputOtherDevice_.size(); i++) { + outputOtherDevice_[i].setFrameHeight(output_.getFrameHeight()); + outputOtherDevice_[i].setFrameWidth(output_.getFrameWidth()); + outputOtherDevice_[i].sequenceStartPositions = + output_.sequenceStartPositions; + outputOtherDevice_[i].subSequenceStartPositions = + output_.subSequenceStartPositions; + outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims; + } + } + + /** + * If input only has MKLDNN device. + * Otherwise, only support the previous layer using CPU device. + */ + bool inputIsOnlyMKLDNN(int index = 0) { + int prevDevice = getPrev(index)->getDeviceId(); + if (prevDevice == MKLDNN_DEVICE) { + return true; + } else { + // do not support GPU yet + CHECK_EQ(prevDevice, CPU_DEVICE) << "Only support CPU yet"; + return false; + } + } + + /** + * If output only has MKLDNN device. + * Otherwise, other devices should only using CPU device. + */ + bool outputIsOnlyMKLDNN() { + for (size_t i = 0; i < outputOtherDevice_.size(); i++) { + CHECK_EQ(outputOtherDevice_[i].deviceId, CPU_DEVICE) + << "Only support other device is CPU yet"; + } + return outputOtherDevice_.size() == 0; + } + + /** + * Sync input value data + */ + void syncInputValue() { + if (inputIsOnlyMKLDNN()) { + return; + } + real* iData = getInputValue(0, CPU_DEVICE)->getData(); + // update input data + // since it might be changed if this is after data layer + inVal_->updateData(iData); + } + + /** + * Sync output grad data + */ + void syncOutputGrad() { + if (outputIsOnlyMKLDNN()) { + return; + } + + // update diff + real* oDiff = getOutput(CPU_DEVICE).grad->getData(); + outGrad_->updateData(oDiff); + } + + /** + * Set deviceId of this layer. + */ + void setDevice(int id) { deviceId_ = id; } + + /** + * Set deviceId of the params used in this layer. + */ + void setParamsDevice(int id, const ParameterMap& parameterMap) { + for (auto& inputConfig : config_.inputs()) { + if (inputConfig.has_input_parameter_name()) { + ParameterPtr parameter; + std::string name = inputConfig.input_parameter_name(); + CHECK(mapGet(name, parameterMap, ¶meter)) + << "Cannot find input parameter " << name << " for layer " + << getName(); + parameter->setDevice(id); + } + } + if (config_.has_bias_parameter_name()) { + ParameterPtr parameter; + std::string name = config_.bias_parameter_name(); + CHECK(mapGet(name, parameterMap, ¶meter)) + << "Cannot find bias parameter " << name << " for layer " + << getName(); + parameter->setDevice(id); + } + } +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/ScaleShiftLayer.cpp b/paddle/gserver/layers/ScaleShiftLayer.cpp new file mode 100644 index 0000000000..35fd038ab4 --- /dev/null +++ b/paddle/gserver/layers/ScaleShiftLayer.cpp @@ -0,0 +1,107 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" + +namespace paddle { + +/** + * A layer applies a linear transformation to each element in each row of + * the input matrix. For each element, the layer first re-scale it and then + * adds a bias to it. + * + * \f[ + * y = wx + b + * \f] + * + * Here, w is the scale and b is the bias. Both w and b are trainable scalars. + * + */ + +class ScaleShiftLayer : public Layer { +protected: + std::unique_ptr scale_; + std::unique_ptr offset_; + +public: + explicit ScaleShiftLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +REGISTER_LAYER(scale_shift, ScaleShiftLayer); + +bool ScaleShiftLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + Layer::init(layerMap, parameterMap); + CHECK_EQ(inputLayers_.size(), 1U); + scale_.reset(new Weight(1, 1, parameters_[0])); + if (biasParameter_.get() != NULL) { + offset_ = std::unique_ptr(new Weight(1, 1, biasParameter_)); + } + return true; +} + +void ScaleShiftLayer::forward(PassType passType) { + Layer::forward(passType); + + MatrixPtr inV = getInputValue(0); + resetOutput(inV->getHeight(), inV->getWidth()); + MatrixPtr outV = getOutputValue(); + real scaleValue = scale_->getW()->getElement(0, 0); + outV->mulScalar(*inV, scaleValue); + if (offset_) { + real offsetValue = offset_->getW()->getElement(0, 0); + outV->add(offsetValue); + } +} + +void ScaleShiftLayer::backward(const UpdateCallback& callback) { + MatrixPtr inV = getInputValue(0); + MatrixPtr inG = getInputGrad(0); + MatrixPtr outV = getOutputValue(); + MatrixPtr outG = getOutputGrad(); + + /* Calculate the parameter gradient for the current layer */ + if (scale_->getWGrad()) { + MatrixPtr rowSumMtx; + Matrix::resizeOrCreate(rowSumMtx, outG->getHeight(), 1, false, useGpu_); + // this_i = scaleDest * this_i + scaleSum * \sum_j b_{ij} * c_{ij} + rowSumMtx->sumOfProducts( + /* b= */ *inV, /* c= */ *outG, /* scaleSum= */ 1, /* scaleDest= */ 0.); + // this_i = scaleDest * this_i + scaleSum * \sum_j b_{ji} + scale_->getWGrad()->sumCols( + /* b= */ *rowSumMtx, /* scaleSum= */ 1., /* scaleDest= */ 1.); + scale_->getParameterPtr()->incUpdate(callback); + } + if (offset_ && offset_->getWGrad()) { + MatrixPtr rowSumMtx; + Matrix::resizeOrCreate(rowSumMtx, outG->getHeight(), 1, false, useGpu_); + rowSumMtx->sumRows(*outG, 1., 0.); + offset_->getWGrad()->sumCols(*rowSumMtx, 1., 1.); + offset_->getParameterPtr()->incUpdate(callback); + } + + /* Calculate the input layers error */ + if (inG) { + real scaleValue = scale_->getW()->getElement(0, 0); + inG->add(*outG, scaleValue); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/SequenceSliceLayer.cpp b/paddle/gserver/layers/SequenceSliceLayer.cpp new file mode 100644 index 0000000000..d3a83fad27 --- /dev/null +++ b/paddle/gserver/layers/SequenceSliceLayer.cpp @@ -0,0 +1,223 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "Layer.h" +#include "paddle/math/Matrix.h" +#include "paddle/math/Vector.h" +#include "paddle/utils/Logging.h" +#include "paddle/utils/Stat.h" + +namespace paddle { + +class SequenceSliceLayer : public Layer { +public: + explicit SequenceSliceLayer(const LayerConfig& config) : Layer(config) {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + +private: + /* + * TODO(caoying) + * In PaddePaddle, currently all matrices are real number types, + * but the second and the (optional) third input which are some + * selected indices of the give sequence to trim the sequence, are actually + * filled with int types so that storing int types information in real number + * matrices is very dangerous, since real numbers will be convered to int + * types. If a user fills this matrix himself, invalid data may occor. + */ + + MatrixPtr startIdsOnCpu_; + MatrixPtr endIdsOnCpu_; + + std::vector selectedRows_; + IVectorPtr rowIndice_; + std::vector> inputSeqInfoVec_; + std::vector outSubSeqStartPos_; + std::vector outSeqStartPos_; + + void checkInputs(); + void copySliceIdsToCpu(); + void calSelectedRows(const MatrixPtr starts, const MatrixPtr ends); +}; + +REGISTER_LAYER(seq_slice, SequenceSliceLayer); + +bool SequenceSliceLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + CHECK_GE(inputLayers_.size(), 2U); + CHECK_LE(inputLayers_.size(), 3U); + + setNeedSequenceInfo(false); + return true; +} + +void SequenceSliceLayer::checkInputs() { + const Argument& inputSeq = getInput(0); + CHECK(inputSeq.hasSeq()) << "The first input of sequence slice layer " + << "must be a sequence."; + const MatrixPtr indices1 = getInputValue(1); + CHECK_EQ(static_cast(indices1->getHeight()), + inputSeq.hasSubseq() ? inputSeq.getNumSubSequences() + : inputSeq.getNumSequences()) + << "Height of the second input should be equal to number of sequence " + << "in the first input."; + if (inputLayers_.size() == 3) { + const MatrixPtr indices2 = getInputValue(2); + CHECK_EQ(indices2->getHeight(), indices1->getHeight()) + << "start indices and end indices should have the same height."; + CHECK_EQ(indices2->getWidth(), indices1->getWidth()) + << "start indices and end indices should have the same Width."; + } +} + +void SequenceSliceLayer::copySliceIdsToCpu() { + const MatrixPtr indices1 = getInputValue(1); + if (inputLayers_.size() == 2U) { + if (config_.select_first()) { + Matrix::resizeOrCreate(startIdsOnCpu_, + indices1->getHeight(), + indices1->getWidth(), + false /* trans */, + false /* useGpu */); + startIdsOnCpu_->copyFrom(*indices1); + endIdsOnCpu_ = nullptr; + } else { + Matrix::resizeOrCreate(endIdsOnCpu_, + indices1->getHeight(), + indices1->getWidth(), + false /* trans */, + false /* useGpu */); + endIdsOnCpu_->copyFrom(*indices1); + startIdsOnCpu_ = nullptr; + } + } else if (inputLayers_.size() == 3U) { + Matrix::resizeOrCreate(startIdsOnCpu_, + indices1->getHeight(), + indices1->getWidth(), + false /* trans */, + false /* useGpu */); + startIdsOnCpu_->copyFrom(*indices1); + + const MatrixPtr indices2 = getInputValue(2); + Matrix::resizeOrCreate(endIdsOnCpu_, + indices2->getHeight(), + indices2->getWidth(), + false /* trans */, + false /* useGpu */); + endIdsOnCpu_->copyFrom(*indices2); + } +} + +void SequenceSliceLayer::calSelectedRows(const MatrixPtr starts, + const MatrixPtr ends) { + CHECK(starts || ends) << "At least one of the start or end indices " + << "should be given."; + + bool hasSubseq = getInput(0).hasSubseq(); + + outSeqStartPos_.resize(1, 0); + outSubSeqStartPos_.resize(1, 0); + selectedRows_.clear(); + + size_t beamSize = starts ? starts->getWidth() : ends->getWidth(); + size_t rowIdx = 0; + for (size_t i = 0; i < inputSeqInfoVec_.size(); ++i) { + for (size_t j = 0; j < inputSeqInfoVec_[i].size() - 1; ++j) { + for (size_t k = 0; k < beamSize; ++k) { + if (starts && starts->getElement(rowIdx, k) == -1.) break; + if (ends && ends->getElement(rowIdx, k) == -1.) break; + + int begPos = inputSeqInfoVec_[i][j]; + if (starts) begPos += starts->getElement(rowIdx, k); + + int endPos = inputSeqInfoVec_[i][j + 1] - 1; + if (ends) endPos = inputSeqInfoVec_[i][j] + ends->getElement(rowIdx, k); + + int seqLen = endPos - begPos + 1; + CHECK_GT(seqLen, 0U); + for (int m = begPos; m <= endPos; ++m) selectedRows_.push_back(m); + hasSubseq + ? outSubSeqStartPos_.push_back(outSubSeqStartPos_.back() + seqLen) + : outSeqStartPos_.push_back(outSeqStartPos_.back() + seqLen); + } + rowIdx++; + } + if (hasSubseq) outSeqStartPos_.push_back(outSubSeqStartPos_.back()); + } + + if (useGpu_) { + rowIndice_ = IVector::create(selectedRows_.size(), useGpu_); + rowIndice_->copyFrom(selectedRows_.data(), selectedRows_.size()); + } else { + rowIndice_ = + IVector::create(selectedRows_.data(), selectedRows_.size(), useGpu_); + } + + // create the sequence information for the output. + ICpuGpuVector::resizeOrCreate( + output_.sequenceStartPositions, outSeqStartPos_.size(), false); + output_.sequenceStartPositions->copyFrom( + outSeqStartPos_.data(), outSeqStartPos_.size(), false); + + if (hasSubseq) { + ICpuGpuVector::resizeOrCreate( + output_.subSequenceStartPositions, outSubSeqStartPos_.size(), false); + output_.subSequenceStartPositions->copyFrom( + outSubSeqStartPos_.data(), outSubSeqStartPos_.size(), false); + } +} + +void SequenceSliceLayer::forward(PassType passType) { + Layer::forward(passType); + checkInputs(); + + const Argument& inputSeq = getInput(0); + inputSeqInfoVec_.clear(); + Argument::reorganizeSeqInfo(inputSeq.sequenceStartPositions, + inputSeq.subSequenceStartPositions, + inputSeqInfoVec_); + if (!useGpu_) { + if (inputLayers_.size() == 2U) { + startIdsOnCpu_ = config_.select_first() ? getInputValue(1) : nullptr; + endIdsOnCpu_ = config_.select_first() ? nullptr : getInputValue(1); + } else if (inputLayers_.size() == 3U) { + startIdsOnCpu_ = getInputValue(1); + endIdsOnCpu_ = getInputValue(2); + } + } else { + copySliceIdsToCpu(); + } + + /* + * calculate the selected row indices in a batch, and build the output + * sequence information. + */ + calSelectedRows(startIdsOnCpu_, endIdsOnCpu_); + + resetOutput(selectedRows_.size(), getSize()); + + getOutputValue()->selectRows(*getInputValue(0), *rowIndice_); +} + +void SequenceSliceLayer::backward(const UpdateCallback& callback) { + getOutputGrad()->addToRows(*getInputGrad(0), *rowIndice_); +} + +} // namespace paddle diff --git a/paddle/gserver/layers/SubNestedSequenceLayer.cpp b/paddle/gserver/layers/SubNestedSequenceLayer.cpp index 648d3908f3..e9bee77212 100644 --- a/paddle/gserver/layers/SubNestedSequenceLayer.cpp +++ b/paddle/gserver/layers/SubNestedSequenceLayer.cpp @@ -52,23 +52,34 @@ private: * ] * * ths output is saved to private member rowIndice_; - * [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15, - * 16,17,18,19,20,21,22,23,24,25,26,27] + * [0,1,2,3,4,5,6,7,8,9,15,16,17,18,19,20,21,23,24,25,26,27] */ - void calSelectedCols(const MatrixPtr selectedIndices, + void calSelectedRows(const MatrixPtr selectedIndices, const std::vector>& inputSeqInfo); - // if the second input of this layer is on GPU memory, copy it to CPU memory. + /* + * TODO(caoying) + * In PaddePaddle, currently all matrices are real number types, + * but the second is some selected indices of the give sequence to trim + * the nested sequence, are actually filled with int types so that storing + * int types information in real number matrices is very dangerous, since + * real numbers will be convered to int types. If a user fills this matrix + * himself, invalid data may occor. + * + * if the second input of this layer is on GPU memory, copy it to CPU memory. + */ MatrixPtr selIdsCpu_; - // reorganized sequenceStartPositions and subSequenceStartPositions - // into a 2d vector to facilitate the sequence selection process. + /* + * reorganize sequenceStartPositions and subSequenceStartPositions + * into a 2d vector to facilitate the sequence selection process. + */ std::vector> inputSeqInfoVec_; - // the final selected row indices in a batch, - // rowIdx_ and selectedRows_ actually share a same memory. + /* store the final selected row indices in a batch */ IVectorPtr rowIndice_; + /* rowIndice_ and selectedRows_ actually share a same memory. */ std::vector selectedRows_; }; @@ -83,7 +94,7 @@ bool SubNestedSequenceLayer::init(const LayerMap& layerMap, return true; } -void SubNestedSequenceLayer::calSelectedCols( +void SubNestedSequenceLayer::calSelectedRows( const MatrixPtr selectedIndices, const std::vector>& inputSeqInfo) { selectedRows_.clear(); @@ -160,7 +171,7 @@ void SubNestedSequenceLayer::forward(PassType passType) { Argument::reorganizeSeqInfo(inputSeq.sequenceStartPositions, inputSeq.subSequenceStartPositions, inputSeqInfoVec_); - calSelectedCols(selIdsCpu_, inputSeqInfoVec_); + calSelectedRows(selIdsCpu_, inputSeqInfoVec_); resetOutput(selectedRows_.size(), getSize()); getOutputValue()->selectRows(*getInputValue(0), *rowIndice_); diff --git a/paddle/gserver/tests/CMakeLists.txt b/paddle/gserver/tests/CMakeLists.txt index 209d0ab9c8..346c01ced6 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 ${PADDLE_SOURCE_DIR}/paddle) ################# test_LayerGrad ####################### add_unittest_without_exec(test_LayerGrad @@ -18,6 +18,15 @@ add_unittest_without_exec(test_LayerGrad add_test(NAME test_LayerGrad COMMAND test_LayerGrad) +########## test_Mkldnn layers and activations ########## +if(WITH_MKLDNN) + add_unittest_without_exec(test_MKLDNN + test_MKLDNN.cpp + MKLDNNTester.cpp + LayerGradUtil.cpp) + add_test(NAME test_MKLDNN COMMAND test_MKLDNN) +endif() + ################ test_CRFLayerGrad #################### add_unittest_without_exec(test_CRFLayerGrad test_CRFLayerGrad.cpp @@ -25,6 +34,12 @@ add_unittest_without_exec(test_CRFLayerGrad add_test(NAME test_CRFLayerGrad COMMAND test_CRFLayerGrad) +################ test_SeqSliceLayerGrad #################### +add_unittest_without_exec(test_SeqSliceLayerGrad + test_SeqSliceLayerGrad.cpp + LayerGradUtil.cpp) +add_test(NAME test_SeqSliceLayerGrad + COMMAND test_SeqSliceLayerGrad) add_unittest_without_exec(test_ActivationGrad test_ActivationGrad.cpp @@ -92,8 +107,8 @@ if(WITH_PYTHON) test_PyDataProvider.cpp) add_test(NAME test_PyDataProvider - COMMAND .set_python_path.sh -d ./gserver/tests:${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 ####################### @@ -106,7 +121,7 @@ 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 ############### @@ -116,20 +131,20 @@ 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() @@ -137,6 +152,6 @@ 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 fd9cfa1dc7..a38880e14c 100644 --- a/paddle/gserver/tests/LayerGradUtil.cpp +++ b/paddle/gserver/tests/LayerGradUtil.cpp @@ -388,14 +388,23 @@ void initDataLayer(TestConfig testConf, data.grad->zeroMem(); break; case INPUT_SELF_DEFINE_DATA: { - size_t height = testConf.inputDefs[i].selfDefinedData->getHeight(); - size_t width = testConf.inputDefs[i].selfDefinedData->getWidth(); - CHECK_GT(static_cast(height), 0); - CHECK_GT(static_cast(width), 0); - data.value = Matrix::create(height, width, false, useGpu); - data.grad = Matrix::create(height, width, false, useGpu); - data.value->copyFrom(*testConf.inputDefs[i].selfDefinedData); - data.grad->zeroMem(); + if (testConf.inputDefs[i].ids.size()) { + data.ids = IVector::create(testConf.inputDefs[i].ids.size(), useGpu); + data.ids->copyFrom(testConf.inputDefs[i].ids.data(), + testConf.inputDefs[i].ids.size()); + } else if (testConf.inputDefs[i].selfDefinedData) { + size_t height = testConf.inputDefs[i].selfDefinedData->getHeight(); + size_t width = testConf.inputDefs[i].selfDefinedData->getWidth(); + CHECK_GT(static_cast(height), 0); + CHECK_GT(static_cast(width), 0); + data.value = Matrix::create(height, width, false, useGpu); + data.grad = Matrix::create(height, width, false, useGpu); + data.value->copyFrom(*testConf.inputDefs[i].selfDefinedData); + data.grad->zeroMem(); + } else { + LOG(FATAL) << "No self-defined data are given."; + return; + } const std::vector& labelSeqStartPositions = testConf.inputDefs[i].labelSeqStartPositions; diff --git a/paddle/gserver/tests/LayerGradUtil.h b/paddle/gserver/tests/LayerGradUtil.h index 5debedf5ef..88e831f78b 100644 --- a/paddle/gserver/tests/LayerGradUtil.h +++ b/paddle/gserver/tests/LayerGradUtil.h @@ -68,6 +68,7 @@ struct InputDef { std::vector labelInitValue; std::vector labelSeqStartPositions; std::vector labelSubSeqStartPositions; + std::vector ids; MatrixPtr selfDefinedData; InputDef(InputType type, string nameIn, size_t dimIn, size_t sizeIn) { @@ -95,6 +96,23 @@ struct InputDef { isStatic = false; } + InputDef(InputType type, + string nameIn, + const std::vector& ids, + const std::vector& selfDefinedSeqStartPos = {}, + const std::vector& selfDefinedSubSeqStartPos = {}) + : labelSeqStartPositions(selfDefinedSeqStartPos), + labelSubSeqStartPositions(selfDefinedSubSeqStartPos), + ids(ids) { + selfDefinedData = nullptr; + inputType = type; + name = nameIn; + dim = 0; + sparse = {""}; + paraSize = 0; + isStatic = false; + } + InputDef(InputType type, string nameIn, size_t dimIn, diff --git a/paddle/gserver/tests/MKLDNNTester.cpp b/paddle/gserver/tests/MKLDNNTester.cpp new file mode 100644 index 0000000000..de1635be2a --- /dev/null +++ b/paddle/gserver/tests/MKLDNNTester.cpp @@ -0,0 +1,382 @@ +/* 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 mkldnn init from PARAM_FORMAT_ORIGINAL weight + reset(dnn, ref, batchSize); + randomWgtDatas(); + clearWgtDiffs(); + clearBotDiffs(); + for (size_t i = 0; i < iter_; ++i) { + VLOG(MKLDNN_TESTS) << "Check Iteration " << i; + runOnce(); + } + + if (parameters_[DNN].empty()) { + // has no paramters + return; + } + + // After run some iterations, the mkldnn weight has been stored in dnnLayer + // and we can also get the mkldnn weight parameter header format. + // Weight parameter should always be index 0 (and bias index 1). + // TODO(TJ): should also consider mean and var format when batchnorm ready + int dnnWgtFmt = parameters_[DNN][0]->getHeaderFormat(); + int refWgtFmt = parameters_[REF][0]->getHeaderFormat(); + if (dnnWgtFmt == refWgtFmt) { + // weight format are equal, so no need check more + return; + } + + // then save the weights and restart again + vector dnnWgts, refWgts; + CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size()); + saveWgt(parameters_[DNN], dnnWgts); + saveWgt(parameters_[REF], refWgts); + + // restart again with dnn weight format + reset(dnn, ref, batchSize); + // TODO(TJ): should also considerate mean and var format when batchnorm ready + parameters_[DNN][0]->setHeaderFormat(dnnWgtFmt); + + // restore wgt + restoreWgt(dnnWgts, parameters_[DNN]); + 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 0000000000..e55e4493ff --- /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 be 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_KmaxSeqScore.cpp b/paddle/gserver/tests/test_KmaxSeqScore.cpp index a51fe390c7..308abe6816 100644 --- a/paddle/gserver/tests/test_KmaxSeqScore.cpp +++ b/paddle/gserver/tests/test_KmaxSeqScore.cpp @@ -96,6 +96,11 @@ TEST(Layer, kmaxSeqScoreLayer) { 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(); @@ -104,7 +109,7 @@ TEST(Layer, kmaxSeqScoreLayer) { hasSubseq ? subSeqStartPosition : seqStartPosition, beamSize); - for (auto useGpu : {false, true}) { + for (auto useGpu : mode) { TestConfig config; config.layerConfig.set_type("kmax_seq_score"); config.layerConfig.set_beam_size(beamSize); diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp index 6418772584..894aeca1cc 100644 --- a/paddle/gserver/tests/test_LayerGrad.cpp +++ b/paddle/gserver/tests/test_LayerGrad.cpp @@ -12,6 +12,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#ifndef PADDLE_ONLY_CPU +#include +#endif #include #include #include @@ -189,10 +192,16 @@ TEST(Projection, scaling) { void testProjectionConv(size_t groups, bool isDeconv) { const int NUM_FILTERS = 18; const int FILTER_SIZE = 2; - const int FILTER_SIZE_Y = 4; + const int FILTER_SIZE_Y = 2; const int CHANNELS = 3; const int IMAGE_SIZE = 16; +#if CUDNN_VERSION >= 6000 + const int DILATION = 2; +#else + const int DILATION = 1; +#endif + ProjectionConfig conf; if (isDeconv) { conf.set_type("convt"); @@ -209,6 +218,8 @@ void testProjectionConv(size_t groups, bool isDeconv) { conv->set_padding_y(1); conv->set_stride(2); conv->set_stride_y(2); + conv->set_dilation(DILATION); + conv->set_dilation_y(DILATION); conv->set_groups(groups); if (isDeconv) { conv->set_filter_channels(NUM_FILTERS / conv->groups()); @@ -217,12 +228,12 @@ void testProjectionConv(size_t groups, bool isDeconv) { } conv->set_img_size(IMAGE_SIZE); int output_x = outputSize(conv->img_size(), - conv->filter_size(), + (conv->filter_size() - 1) * DILATION + 1, conv->padding(), conv->stride(), /* caffeMode */ true); int output_y = outputSize(conv->img_size(), - conv->filter_size_y(), + (conv->filter_size_y() - 1) * DILATION + 1, conv->padding_y(), conv->stride_y(), /* caffeMode */ true); @@ -424,27 +435,38 @@ void testConvLayer(const string& type, bool trans, bool useGpu) { config.layerConfig.set_partial_sum(1); config.layerConfig.set_shared_biases(true); - config.inputDefs.push_back({INPUT_DATA, "layer_0", 384, 288}); + int dilation = 1; + if (type == "cudnn_conv") { +#if CUDNN_VERSION >= 6000 + dilation = 2; +#else + dilation = 1; +#endif + } + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 768, 192}); LayerInputConfig* input = config.layerConfig.add_inputs(); ConvConfig* conv = input->mutable_conv_conf(); conv->set_filter_size(2); - conv->set_filter_size_y(3); + conv->set_filter_size_y(2); conv->set_channels(3); conv->set_padding(0); conv->set_padding_y(1); conv->set_stride(2); conv->set_stride_y(2); + conv->set_dilation(dilation); + conv->set_dilation_y(dilation); conv->set_groups(1); conv->set_filter_channels(conv->channels() / conv->groups()); conv->set_img_size(16); - conv->set_img_size_y(8); + conv->set_img_size_y(16); conv->set_output_x(outputSize(conv->img_size(), - conv->filter_size(), + (conv->filter_size() - 1) * dilation + 1, conv->padding(), conv->stride(), /* caffeMode */ true)); conv->set_output_y(outputSize(conv->img_size_y(), - conv->filter_size_y(), + (conv->filter_size_y() - 1) * dilation + 1, conv->padding_y(), conv->stride_y(), /* caffeMode */ true)); @@ -828,9 +850,27 @@ TEST(Layer, square_error_weighted) { } } +TEST(Layer, huber_regression_loss) { + TestConfig config; + config.layerConfig.set_type("huber_regression"); + config.biasSize = 0; + + config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0}); + config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 10, 0}); + config.layerConfig.add_inputs(); + config.layerConfig.add_inputs(); + + for (auto useGpu : {false, true}) { + for (auto delta : {1, 3, 5}) { + config.layerConfig.set_delta(delta); + testLayerGrad(config, "huber_regression", 100, /* trans */ false, useGpu); + } + } +} + TEST(Layer, huber_two_class) { TestConfig config; - config.layerConfig.set_type("huber"); + config.layerConfig.set_type("huber_classification"); config.biasSize = 0; config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0}); @@ -839,7 +879,7 @@ TEST(Layer, huber_two_class) { config.layerConfig.add_inputs(); for (auto useGpu : {false, true}) { - testLayerGrad(config, "huber", 100, /* trans */ false, useGpu); + testLayerGrad(config, "huber_two_class", 100, /* trans */ false, useGpu); } } @@ -2056,6 +2096,21 @@ TEST(Layer, RowL2NormLayer) { } } +TEST(Layer, ScaleShiftLayer) { + const size_t batchSize = 16; + const size_t size = 32; + TestConfig config; + config.layerConfig.set_type("scale_shift"); + config.layerConfig.set_size(size); + config.biasSize = 1; + config.inputDefs.push_back( + {INPUT_DATA, "input", /* dim= */ size, /* paraSize= */ 1}); + config.layerConfig.add_inputs(); + for (auto useGpu : {false, true}) { + testLayerGrad(config, "scale_shift", batchSize, false, useGpu, false); + } +} + int main(int argc, char** argv) { testing::InitGoogleTest(&argc, argv); initMain(argc, argv); diff --git a/paddle/gserver/tests/test_MKLDNN.cpp b/paddle/gserver/tests/test_MKLDNN.cpp new file mode 100644 index 0000000000..e1d2270df2 --- /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/gserver/tests/test_NetworkCompare.cpp b/paddle/gserver/tests/test_NetworkCompare.cpp index f930c72fde..d36f72360f 100644 --- a/paddle/gserver/tests/test_NetworkCompare.cpp +++ b/paddle/gserver/tests/test_NetworkCompare.cpp @@ -269,7 +269,8 @@ TEST(Compare, img_conv2) { bool useGpu = FLAGS_use_gpu; double eps = FLAGS_checkgrad_eps; FLAGS_use_gpu = true; - FLAGS_checkgrad_eps = 1e-2; + // Sometimes, this unit test will fail with 1e-2 + FLAGS_checkgrad_eps = 4e-2; compareNetwork(config_file_a, config_file_b); FLAGS_use_gpu = useGpu; FLAGS_checkgrad_eps = eps; diff --git a/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp b/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp new file mode 100644 index 0000000000..e1d4ae1617 --- /dev/null +++ b/paddle/gserver/tests/test_SeqSliceLayerGrad.cpp @@ -0,0 +1,225 @@ +/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "ModelConfig.pb.h" +#include "paddle/gserver/layers/DataLayer.h" +#include "paddle/trainer/Trainer.h" + +#include "LayerGradUtil.h" +#include "paddle/testing/TestUtil.h" + +using namespace paddle; // NOLINT +using namespace std; // NOLINT + +DECLARE_int32(gpu_id); +DECLARE_bool(thread_local_rand_use_global_seed); + +const int MAX_SEQ_NUM = 17; +const int MAX_SEQ_LEN = 23; +const int MAX_BEAM_SIZE = 13; + +const size_t SEED = (size_t)(time(NULL)); + +vector randSampling(real range, int n) { + CHECK_GE(range, n); + vector num(range); + iota(begin(num), end(num), 0.); + if (range == n) return num; + + random_shuffle(begin(num), end(num)); + num.resize(n); + sort(begin(num), end(num)); + return num; +} + +void genSeqInfo(vector& seqStartPos, vector& subSeqStartPos) { + seqStartPos.resize(1, 0); + subSeqStartPos.resize(1, 0); + + srand(SEED); + int seqNum = 1 + (rand() % MAX_SEQ_NUM); + for (int i = 0; i < seqNum; ++i) { + int subSeqNum = 1 + (rand() % MAX_SEQ_NUM); + for (int j = 0; j < subSeqNum; ++j) + subSeqStartPos.push_back(subSeqStartPos.back() + + (1 + (rand() % MAX_SEQ_LEN))); + seqStartPos.push_back(subSeqStartPos.back()); + } +} + +/* + generate start indices according to sequence start positions. + */ +void genStarts(vector& seqStartPos, + vector>& starts, + size_t beamSize) { + starts.clear(); + starts.resize(seqStartPos.size() - 1, vector(beamSize, -1.)); + + for (size_t i = 0; i < seqStartPos.size() - 1; ++i) { + int seqLen = seqStartPos[i + 1] - seqStartPos[i]; + vector randStarts = + randSampling(seqLen, min(seqLen, static_cast(beamSize))); + copy(begin(randStarts), end(randStarts), begin(starts[i])); + } +} + +/* + generate end indices according to sequence start positions and start indices. + */ +void genEnds(vector& seqStartPos, + vector>& starts, + vector>& ends, + size_t beamSize) { + CHECK_EQ(seqStartPos.size() - 1, starts.size()); + ends.clear(); + ends.resize(seqStartPos.size() - 1, vector(beamSize, -1.)); + + for (size_t i = 0; i < starts.size(); ++i) { + for (size_t j = 0; j < starts[i].size(); ++j) { + int seqLen = seqStartPos[i + 1] - seqStartPos[i]; + CHECK_GE(seqLen - 1, starts[i][j]); + if (starts[i][j] == -1.) break; + if (starts[i][j] == (seqLen - 1)) { + ends[i][j] = starts[i][j]; + } else { + ends[i][j] = starts[i][j] + randSampling(seqLen - starts[i][j], 1)[0]; + } + } + } +} + +void genTestData(vector& seqStartPos, + vector& subSeqStartPos, + vector>& starts, + vector>& ends, + bool hasSubseq) { + size_t beamSize = 1 + (rand() % MAX_BEAM_SIZE); + genSeqInfo(seqStartPos, subSeqStartPos); + + genStarts(hasSubseq ? subSeqStartPos : seqStartPos, starts, beamSize); + genEnds(hasSubseq ? subSeqStartPos : seqStartPos, starts, ends, beamSize); +} + +template +void flatten2dVector(vector>& inVec, vector& outVec) { + size_t totalSize{0}; + for (auto const& items : inVec) totalSize += items.size(); + outVec.reserve(totalSize); + + for (auto& items : inVec) + move(items.begin(), items.end(), back_inserter(outVec)); +} + +void testSeqSliceLayer(bool hasSubseq, + bool useGpu, + vector& seqStartPos, + vector& subSeqStartPos, + vector>& starts, + vector>& ends) { + // layer size is not crutial for this layer, + // so here use a small layer size in the unittest. + const size_t layerSize{4}; + TestConfig config; + config.layerConfig.set_type("seq_slice"); + config.layerConfig.set_size(layerSize); + + // add the first input + MatrixPtr seqInputPtr = + Matrix::create(hasSubseq ? subSeqStartPos.back() : seqStartPos.back(), + layerSize, + false, + false); + seqInputPtr->randomizeUniform(); + + if (hasSubseq) { + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, + "seq_input", + seqInputPtr, + seqStartPos, + subSeqStartPos}); + } else { + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, "seq_input", seqInputPtr, seqStartPos}); + } + config.layerConfig.add_inputs(); + + // add start indices + if (starts.size()) { + vector startsToVec; + flatten2dVector(starts, startsToVec); + + MatrixPtr startMatrixPtr = + Matrix::create(starts.size(), starts[0].size(), false, false); + startMatrixPtr->copyFrom(startsToVec.data(), startsToVec.size()); + + config.inputDefs.push_back( + {INPUT_SELF_DEFINE_DATA, "starts", startMatrixPtr}); + config.layerConfig.add_inputs(); + config.layerConfig.set_select_first(true); + } + + // add end indices + if (ends.size()) { + vector endsToVec; + flatten2dVector(ends, endsToVec); + + MatrixPtr endMatrixPtr = + Matrix::create(ends.size(), ends[0].size(), false, false); + endMatrixPtr->copyFrom(endsToVec.data(), endsToVec.size()); + + config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "ends", endMatrixPtr}); + config.layerConfig.add_inputs(); + config.layerConfig.set_select_first(false); + } + + testLayerGrad(config, "seq_slice", /*batchSize*/ 100, false, useGpu, false); +} + +TEST(Layer, SeqSliceLayer) { + vector seqStartPos; + vector subSeqStartPos; + vector> starts; + vector> ends; + + std::vector mode = {false}; +#ifndef PADDLE_ONLY_CPU + mode.push_back(true); +#endif + genSeqInfo(seqStartPos, subSeqStartPos); + for (bool hasSubseq : {true, false}) { + LOG(INFO) << "hasSubSeq : " << hasSubseq; + genTestData(seqStartPos, subSeqStartPos, starts, ends, hasSubseq); + for (bool useGpu : mode) { + vector> tmp; + testSeqSliceLayer( + hasSubseq, useGpu, seqStartPos, subSeqStartPos, tmp, ends); + testSeqSliceLayer( + hasSubseq, useGpu, seqStartPos, subSeqStartPos, starts, tmp); + testSeqSliceLayer( + hasSubseq, useGpu, seqStartPos, subSeqStartPos, starts, ends); + } + } +} + +int main(int argc, char** argv) { + initMain(argc, argv); + hl_start(); + hl_init(FLAGS_gpu_id); + FLAGS_thread_local_rand_use_global_seed = true; + srand(1); + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/math/Allocator.h b/paddle/math/Allocator.h index 666a8b8368..94ef561f06 100644 --- a/paddle/math/Allocator.h +++ b/paddle/math/Allocator.h @@ -48,7 +48,13 @@ public: */ virtual void* alloc(size_t size) { void* ptr; +#ifdef PADDLE_USE_MKLDNN + // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp + // memory alignment + CHECK_EQ(posix_memalign(&ptr, 4096ul, size), 0); +#else CHECK_EQ(posix_memalign(&ptr, 32ul, size), 0); +#endif CHECK(ptr) << "Fail to allocate CPU memory: size=" << size; return ptr; } diff --git a/paddle/math/CMakeLists.txt b/paddle/math/CMakeLists.txt index 9981de6160..68b5296228 100644 --- a/paddle/math/CMakeLists.txt +++ b/paddle/math/CMakeLists.txt @@ -14,14 +14,25 @@ # file(GLOB MATH_HEADERS . *.h) file(GLOB MATH_SOURCES . *.cpp) + +if(NOT WITH_MKLDNN) + set(DNN_HEADER "${CMAKE_CURRENT_SOURCE_DIR}/MKLDNNMatrix.h") + set(DNN_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/MKLDNNMatrix.cpp") + list(REMOVE_ITEM MATH_HEADERS "${DNN_HEADER}") + list(REMOVE_ITEM MATH_SOURCES "${DNN_SOURCE}") + message(STATUS "Skip compiling with MKLDNNMatrix") +else() + message(STATUS "Compile with MKLDNNMatrix") +endif() + set(MATH_SOURCES - "${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 860cad1047..36d57bbb65 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/MKLDNNMatrix.cpp b/paddle/math/MKLDNNMatrix.cpp new file mode 100644 index 0000000000..0a355e2644 --- /dev/null +++ b/paddle/math/MKLDNNMatrix.cpp @@ -0,0 +1,144 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "MKLDNNMatrix.h" + +using namespace mkldnn; // NOLINT + +namespace paddle { + +MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) { + memory::desc md = pd.desc(); + size_t ndims = md.data.ndims; + int* dims = md.data.dims; + CHECK(ndims > 0) << "Input dims should not be empty"; + size_t cnts = 1; + for (size_t i = 0; i < ndims; ++i) { + cnts *= dims[i]; + } + + if (m == nullptr) { + size_t height = dims[0]; + size_t width = cnts / dims[0]; + m = Matrix::create(height, width, false, false); + } + + CHECK(m) << " Matrix should not be empty"; + CpuMatrixPtr cpuMatrix = std::dynamic_pointer_cast(m); + CHECK(cpuMatrix) << "Only support create from CPU matrix yet"; + + CHECK_EQ(cnts, m->getElementCnt()) << "Count size does not match"; + return std::make_shared( + m->getData(), m->getHeight(), m->getWidth(), pd); +} + +MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, + memory::dims dims, + memory::format fmt, + engine& eg, + mkldnn::memory::data_type dtype) { + return create(m, memory::primitive_desc(memory::desc(dims, dtype, fmt), eg)); +} + +void MKLDNNMatrix::reorderDataFrom(const MKLDNNMatrixPtr& m, + memory::format srcFmt, + memory::dims targetDim) { + memory::format dstFmt = getFormat(); + if (srcFmt == dstFmt) { + return; + } + CHECK_EQ(getElementCnt(), m->getElementCnt()) << "size should equal"; + reorderOnce(getData(), m->getData(), srcFmt, dstFmt, targetDim); +} + +void MKLDNNMatrix::reorderDataTo(const MKLDNNMatrixPtr& m, + memory::format dstFmt, + memory::dims targetDim) { + memory::format srcFmt = getFormat(); + if (srcFmt == dstFmt) { + return; + } + CHECK_EQ(getElementCnt(), m->getElementCnt()) << "size should equal"; + reorderOnce(getData(), m->getData(), srcFmt, dstFmt, targetDim); +} + +void MKLDNNMatrix::reorderOnce(void* srcData, + void* dstData, + memory::format srcFmt, + memory::format dstFmt, + memory::dims dm) { + CHECK(srcData); + CHECK(dstData); + MatrixPtr tmpSrc; + if (dstData == srcData) { + // inplace data + size_t sz = 1; + for (size_t i = 0; i < dm.size(); ++i) { + sz *= dm[i]; + } + tmpSrc = Matrix::create(sz, 1, false, false); + tmpSrc->copyFrom((real*)srcData, sz); + srcData = tmpSrc->getData(); + } + + auto dtype = this->getDtype(); + auto srcMD = memory::desc(dm, dtype, srcFmt); + auto dstMD = memory::desc(dm, dtype, dstFmt); + + auto eg = this->getEngine(); + auto src = memory(memory::primitive_desc(srcMD, eg), srcData); + auto dst = memory(memory::primitive_desc(dstMD, eg), dstData); + + auto r = reorder(src, dst); + stream(stream::kind::eager).submit({r}).wait(); +} + +void MKLDNNMatrix::downSpatial() { + int fmt = getFormat(); + if (!(fmt == memory::format::nchw || fmt == memory::format::oihw)) { + // only support nchw and oihw yet, later can support more like nhwc, ihwo + return; + } + + // TODO(TJ): change H(height) and W(width) if support nhwc or more + const int H = 2, W = 3; + memory::dims srcDims = getDims(); + if (srcDims[H] != 1 || srcDims[W] != 1) { + // can not down spatial + return; + } + + memory::dims dstDims = memory::dims{srcDims[0], srcDims[1]}; + memory::format dstFmt; + switch (fmt) { + case memory::format::nchw: + dstFmt = memory::format::nc; + break; + case memory::format::oihw: + dstFmt = memory::format::oi; + break; + default: + LOG(FATAL) << "unsupported format"; + } + memory::desc md = memory::desc(dstDims, getDtype(), dstFmt); + memory::primitive_desc pd = memory::primitive_desc(md, getEngine()); + mkldnn_primitive_t result; + mkldnn::error::wrap_c_api( + mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr), + "could not create a memory primitive"); + reset(result); + set_data_handle(getData()); +} + +} // namespace paddle diff --git a/paddle/math/MKLDNNMatrix.h b/paddle/math/MKLDNNMatrix.h new file mode 100644 index 0000000000..e50f698b49 --- /dev/null +++ b/paddle/math/MKLDNNMatrix.h @@ -0,0 +1,148 @@ +/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "Matrix.h" +#include "mkldnn.hpp" +#include "paddle/parameter/Parameter.h" + +namespace paddle { + +class MKLDNNMatrix; +typedef std::shared_ptr MKLDNNMatrixPtr; + +/** + * @brief MKLDNN Matrix. + * + */ +class MKLDNNMatrix : public CpuMatrix, public mkldnn::memory { +public: + MKLDNNMatrix(real* data, + size_t height, + size_t width, + mkldnn::memory::primitive_desc pd) + : CpuMatrix(data, height, width, false), mkldnn::memory(pd, data) {} + + ~MKLDNNMatrix() {} + + /** + * Create MKLDNNMatrix from a MatrixPtr and memory primitive_desc + */ + static MKLDNNMatrixPtr create(MatrixPtr m, mkldnn::memory::primitive_desc pd); + + /** + * Create MKLDNNMatrix from a MatrixPtr and memory details info + */ + static MKLDNNMatrixPtr create( + MatrixPtr m, + mkldnn::memory::dims dims, + mkldnn::memory::format fmt, + mkldnn::engine& eg, + mkldnn::memory::data_type dtype = mkldnn::memory::data_type::f32); + +public: + /** + * Reorder this MKLDNNMatrix from other format. + * Support inplace reorder. + * @note: this function would only reorder the data layout. + * will NOT change this original dim or format info + */ + void reorderDataFrom(const MKLDNNMatrixPtr& m, + memory::format srcFmt, + memory::dims targetDim); + + /** + * Reorder this MKLDNNMatrix to other format. + * Support inplace reorder. + * @note: this function would only reorder the data layout. + * will NOT change the dst dim or format info + */ + void reorderDataTo(const MKLDNNMatrixPtr& m, + memory::format dstFmt, + memory::dims targetDim); + + /** + * Dimensionality reduction. + * Change format "nchw --> nc" or "oihw --> oi" if the h and w are both 1 + */ + void downSpatial(); + + /** + * Update the memory data handle. + * Caution: This will not check the buffer size of the data, + * it should be coverd by user. + */ + void updateData(void* data) { set_data_handle(data); } + + /** + * Get primitive descriptor. + */ + mkldnn::memory::primitive_desc getPrimitiveDesc() { + return this->get_primitive_desc(); + } + + /** + * Get memory descriptor. + */ + mkldnn::memory::desc getMemoryDesc() { return getPrimitiveDesc().desc(); } + + /** + * Get dimensions. + */ + mkldnn::memory::dims getDims() { + mkldnn::memory::desc md = getMemoryDesc(); + const int* src = md.data.dims; + int ndims = md.data.ndims; + mkldnn::memory::dims dst; + dst.resize(ndims); + for (int i = 0; i < ndims; ++i) { + dst[i] = src[i]; + } + return dst; + } + + /** + * Get format. + */ + mkldnn::memory::format getFormat() { + return (mkldnn::memory::format)(getMemoryDesc().data.format); + } + + /** + * Get memory data type. + */ + mkldnn::memory::data_type getDtype() { + return (mkldnn::memory::data_type)(getMemoryDesc().data.data_type); + } + + /** + * Get engine. + */ + mkldnn::engine getEngine() { return getPrimitiveDesc().get_engine(); } + +protected: + /** + * Do reorder once. + * Can support inplace. + */ + void reorderOnce(void* srcData, + void* dstData, + memory::format srcFmt, + memory::format dstFmt, + memory::dims dm); +}; + +} // namespace paddle diff --git a/paddle/math/SparseMatrix.h b/paddle/math/SparseMatrix.h index f6cd5df338..16300db081 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/memory/CMakeLists.txt b/paddle/memory/CMakeLists.txt index 8035d93bfe..9cc4233e43 100644 --- a/paddle/memory/CMakeLists.txt +++ b/paddle/memory/CMakeLists.txt @@ -1,7 +1,7 @@ add_subdirectory(detail) cc_library(memory SRCS memory.cc) -cc_library(memcpy SRCS memcpy.cc DEPS device_context) +cc_library(memcpy SRCS memcpy.cc) cc_library(paddle_memory DEPS diff --git a/paddle/memory/detail/system_allocator.cc b/paddle/memory/detail/system_allocator.cc index f61e67a329..a270bd5958 100644 --- a/paddle/memory/detail/system_allocator.cc +++ b/paddle/memory/detail/system_allocator.cc @@ -27,7 +27,7 @@ limitations under the License. */ // between host and device. Allocates too much would reduce the amount // of memory available to the system for paging. So, by default, we // should set false to use_pinned_memory. -DEFINE_bool(use_pinned_memory, false, "If set, allocate cpu pinned memory."); +DEFINE_bool(use_pinned_memory, true, "If set, allocate cpu pinned memory."); namespace paddle { namespace memory { diff --git a/paddle/memory/memcpy.cc b/paddle/memory/memcpy.cc index aaab1142ca..a19a3e3675 100644 --- a/paddle/memory/memcpy.cc +++ b/paddle/memory/memcpy.cc @@ -16,8 +16,6 @@ limitations under the License. */ #include // for memcpy -#include "paddle/platform/device_context.h" - namespace paddle { namespace memory { diff --git a/paddle/memory/memory.cc b/paddle/memory/memory.cc index 207025f9b1..29bc26f9d3 100644 --- a/paddle/memory/memory.cc +++ b/paddle/memory/memory.cc @@ -13,22 +13,38 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/memory/memory.h" + +#include // for transform +#include // for memcpy +#include // for unique_ptr +#include // for call_once + +#include "glog/logging.h" + #include "paddle/memory/detail/buddy_allocator.h" #include "paddle/memory/detail/system_allocator.h" +#include "paddle/platform/gpu_info.h" -#include // for memcpy +DECLARE_double(fraction_of_gpu_memory_to_use); namespace paddle { namespace memory { -detail::BuddyAllocator* GetCPUBuddyAllocator() { - static detail::BuddyAllocator* a = nullptr; - if (a == nullptr) { - a = new detail::BuddyAllocator(new detail::CPUAllocator, - platform::CpuMinChunkSize(), - platform::CpuMaxChunkSize()); - } - return a; +using BuddyAllocator = detail::BuddyAllocator; + +std::once_flag cpu_allocator_flag; +std::once_flag gpu_allocator_flag; + +BuddyAllocator* GetCPUBuddyAllocator() { + static std::unique_ptr a{nullptr}; + + std::call_once(cpu_allocator_flag, [&]() { + a.reset(new BuddyAllocator(new detail::CPUAllocator, + platform::CpuMinChunkSize(), + platform::CpuMaxChunkSize())); + }); + + return a.get(); } template <> @@ -48,20 +64,36 @@ size_t Used(platform::CPUPlace place) { #ifndef PADDLE_ONLY_CPU -detail::BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { - static detail::BuddyAllocator** as = NULL; - if (as == NULL) { +BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { + using BuddyAllocVec = std::vector; + static std::unique_ptr as{ + new BuddyAllocVec, [](BuddyAllocVec* p) { + std::for_each(p->begin(), p->end(), + [](BuddyAllocator* p) { delete p; }); + }}; + + // GPU buddy allocators + auto& allocators = *as.get(); + + // GPU buddy allocator initialization + std::call_once(gpu_allocator_flag, [&]() { int gpu_num = platform::GetDeviceCount(); - as = new detail::BuddyAllocator*[gpu_num]; + allocators.reserve(gpu_num); for (int gpu = 0; gpu < gpu_num; gpu++) { platform::SetDeviceId(gpu); - as[gpu] = new detail::BuddyAllocator(new detail::GPUAllocator, - platform::GpuMinChunkSize(), - platform::GpuMaxChunkSize()); + allocators.emplace_back(new BuddyAllocator(new detail::GPUAllocator, + platform::GpuMinChunkSize(), + platform::GpuMaxChunkSize())); } - } + VLOG(3) << "\n\nNOTE: each GPU device use " + << FLAGS_fraction_of_gpu_memory_to_use * 100 << "% of GPU memory.\n" + << "You can set environment variable '" + << platform::kEnvFractionGpuMemoryToUse + << "' to change the fraction of GPU usage.\n\n"; + }); + platform::SetDeviceId(gpu_id); - return as[gpu_id]; + return allocators[gpu_id]; } template <> diff --git a/paddle/memory/memory.h b/paddle/memory/memory.h index 72351b9dfa..11bbb88187 100644 --- a/paddle/memory/memory.h +++ b/paddle/memory/memory.h @@ -14,7 +14,6 @@ limitations under the License. */ #pragma once -#include "paddle/platform/gpu_info.h" #include "paddle/platform/place.h" namespace paddle { diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt index 9e4026d1c6..f0fd12f1b5 100644 --- a/paddle/operators/CMakeLists.txt +++ b/paddle/operators/CMakeLists.txt @@ -41,31 +41,35 @@ function(op_library TARGET) endif() endfunction() +add_subdirectory(math) + +cc_test(gather_test SRCS gather_test.cc DEPS tensor) +op_library(gather_op SRCS gather_op.cc gather_op.cu) + +cc_test(scatter_test SRCS scatter_test.cc DEPS tensor) +op_library(scatter_op SRCS scatter_op.cc scatter_op.cu) + 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) -cc_test(sgd_op_test SRCS sgd_op_test.cc DEPS sgd_op) -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 rnn/recurrent_op_utils.cc - DEPS op_desc tensor op_registry operator net_op) -cc_test(recurrent_op_test SRCS recurrent_op_test.cc DEPS recurrent_op gtest mul_op add_op) -op_library(uniform_random_op - SRCS uniform_random_op.cc uniform_random_op.cu) + DEPS framework_proto tensor op_registry operator net_op) +op_library(uniform_random_op SRCS uniform_random_op.cc uniform_random_op.cu) +op_library(lookup_table_op SRCS lookup_table_op.cc lookup_table_op.cu) +op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op) +op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op) diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc index 086245ef62..8ab748ed71 100644 --- a/paddle/operators/add_op.cc +++ b/paddle/operators/add_op.cc @@ -18,16 +18,15 @@ namespace paddle { namespace operators { class AddOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 2); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "Inputs of AddOp must all be set"); - PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr, - "Outputs of AddOp must all be set"); - PADDLE_ENFORCE(ctx.Input(0)->dims() == ctx.Input(1)->dims(), - "Two input of Add Op's dimension must be same."); - ctx.Output(0)->Resize(ctx.Input(0)->dims()); + PADDLE_ENFORCE_EQ(ctx.Input("X")->dims(), + ctx.Input("Y")->dims(), + "Two input of Add Op's dimension must be same."); + ctx.Output("Out")->Resize(ctx.Input("X")->dims()); } }; @@ -47,6 +46,9 @@ The equation is: Out = X + Y }; class AddOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override {} }; @@ -55,8 +57,7 @@ class AddOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker); -REGISTER_GRADIENT_OP(add_two, add_two_grad, ops::AddOpGrad); +REGISTER_OP(add_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.h b/paddle/operators/add_op.h index d76c10957e..a7307b6818 100644 --- a/paddle/operators/add_op.h +++ b/paddle/operators/add_op.h @@ -28,9 +28,9 @@ template class AddKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto input0 = context.Input(0); - auto input1 = context.Input(1); - auto output = context.Output(0); + auto* input0 = context.Input("X"); + auto* input1 = context.Input("Y"); + auto* output = context.Output("Out"); output->mutable_data(context.GetPlace()); diff --git a/paddle/operators/add_op_test.cc b/paddle/operators/add_op_test.cc deleted file mode 100644 index bf529defb2..0000000000 --- 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 "paddle/framework/op_registry.h" - -USE_OP(add_two); - -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 c813d54e17..ab1e1c101a 100644 --- a/paddle/operators/cross_entropy_op.cc +++ b/paddle/operators/cross_entropy_op.cc @@ -18,33 +18,31 @@ namespace paddle { namespace operators { class OnehotCrossEntropyOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 2, - "Input size of OnehotCrossEntropyOp must be two"); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, - "Output size of OnehotCrossEntropyOp must be one"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), - "0-th input of OnehotCrossEntropyOp should be set"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(1), - "1-th input of OnehotCrossEntropyOp should be set"); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0), - "Outputs of OnehotCrossEntropyOp must all be set"); - PADDLE_ENFORCE_EQ(ctx.Input(0)->dims().size(), 2); - PADDLE_ENFORCE_EQ(ctx.Output(0)->dims().size(), 1, - "label's dimension must be 1."); - ctx.Output(0)->Resize({ctx.Input(0)->dims()[0]}); + auto *X = ctx.Input("X"); + auto *label = ctx.Input("label"); + + PADDLE_ENFORCE_EQ(X->dims().size(), 2, "X's dimension must be 2."); + PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label's dimension must be 1."); + PADDLE_ENFORCE_EQ(X->dims()[0], label->dims()[0]); + ctx.Output("Y")->Resize({X->dims()[0]}); } }; class OnehotCrossEntropyGradientOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - auto X_grad = ctx.Output(framework::GradVarName("X")); + auto dX = ctx.Output(framework::GradVarName("X")); auto X = ctx.Input("X"); - // TODO(superjom) add enforce here after helper functions ready - X_grad->Resize(X->dims()); + dX->Resize(X->dims()); } }; @@ -69,12 +67,9 @@ OnehotCrossEntropy Operator. namespace ops = paddle::operators; REGISTER_OP(onehot_cross_entropy, ops::OnehotCrossEntropyOp, - ops::OnehotCrossEntropyOpMaker); -REGISTER_OP_CPU_KERNEL( - onehot_cross_entropy, - ops::OnehotCrossEntropyOpKernel); -REGISTER_GRADIENT_OP(onehot_cross_entropy, onehot_cross_entropy_grad, - ops::OnehotCrossEntropyGradientOp); -REGISTER_OP_CPU_KERNEL( - onehot_cross_entropy_grad, - ops::OnehotCrossEntropyGradientOpKernel); + ops::OnehotCrossEntropyOpMaker, onehot_cross_entropy_grad, + ops::OnehotCrossEntropyGradientOp); +REGISTER_OP_CPU_KERNEL(onehot_cross_entropy, + ops::OnehotCrossEntropyOpKernel); +REGISTER_OP_CPU_KERNEL(onehot_cross_entropy_grad, + ops::OnehotCrossEntropyGradientOpKernel); diff --git a/paddle/operators/cross_entropy_op.cu b/paddle/operators/cross_entropy_op.cu index 4bbc8f093a..d999bfce58 100644 --- a/paddle/operators/cross_entropy_op.cu +++ b/paddle/operators/cross_entropy_op.cu @@ -12,10 +12,122 @@ See the License for the specific language governing permissions and limitations under the License. */ -#define EIGEN_USE_GPU -#include "paddle/operators/cross_entropy_op.h" +#include "paddle/framework/op_registry.h" +#include "paddle/platform/assert.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +__host__ __device__ T clipping_log(const T x) { + PADDLE_ASSERT(std::is_floating_point::value); + const T kApproInf = 1e20; + T v = log(x); + if (v == INFINITY) { + return kApproInf; + } + if (v == -INFINITY) { + return -kApproInf; + } + return v; +} + +template +__global__ void CrossEntropyKernel(T* Y, const T* X, const int* label, + const int N, const int D) { + // TOOD(qingqing) define CUDA_1D_KERNEL_LOOP macro in a common file. + // CUDA_1D_KERNEL_LOOP(i, N) { + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; + i += blockDim.x * gridDim.x) { + PADDLE_ASSERT(label[i] >= 0 && label[i] < D); + Y[i] = -clipping_log(X[i * D + label[i]]); + } +} + +// TODO(qingqing): make zero setting an common function. +template +__global__ void zero(T* X, const int N) { + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; + i += blockDim.x * gridDim.x) { + X[i] = 0.0; + } +} + +template +__global__ void CrossEntropyGradientKernel(T* dX, const T* dY, const T* X, + const int* label, const int N, + const int D) { + // TOOD(qingqing) define CUDA_1D_KERNEL_LOOP macro in a common file. + // CUDA_1D_KERNEL_LOOP(i, N) { + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; + i += blockDim.x * gridDim.x) { + int idx = i * D + label[i]; + dX[idx] = -dY[i] / X[idx]; + } +} + +template +class OnehotCrossEntropyOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use GPUPlace."); + + auto X = ctx.Input("X"); + const T* Xdata = X->data(); + const int* label_data = ctx.Input("label")->data(); + auto Y = ctx.Output("Y"); + Y->mutable_data(ctx.GetPlace()); + T* Ydata = Y->data(); + + int N = X->dims()[0]; + int D = X->dims()[1]; + int block = 512; + int grid = (N + block - 1) / block; + // TODO(qingqing) launch kernel on specified stream + // base on ExecutionContext. + CrossEntropyKernel<<>>(Ydata, Xdata, label_data, N, D); + } +}; + +template +class OnehotCrossEntropyGradientOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use GPUPlace."); + + auto X = ctx.Input("X"); + auto dX = ctx.Output(framework::GradVarName("X")); + auto dY = ctx.Input(framework::GradVarName("Y")); + auto label = ctx.Input("label"); + + auto* dXdata = dX->template mutable_data(ctx.GetPlace()); + auto* dYdata = dY->template data(); + auto* Xdata = X->template data(); + auto* label_data = label->data(); + + int N = X->dims()[0]; + int D = X->dims()[1]; + int block = 512; + int grid = (N * D + block - 1) / block; + zero<<>>(dXdata, N * D); + + grid = (N + block - 1) / block; + // TODO(qingqing): launch kernel on specified stream + // base on ExecutionContext. + CrossEntropyGradientKernel<<>>(dXdata, dYdata, Xdata, + label_data, N, D); + } +}; + +} // namespace operators +} // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL( - onehot_cross_entropy, - ops::OnehotCrossEntropyOpKernel); +REGISTER_OP_GPU_KERNEL(onehot_cross_entropy, + ops::OnehotCrossEntropyOpCUDAKernel); +REGISTER_OP_GPU_KERNEL(onehot_cross_entropy_grad, + ops::OnehotCrossEntropyGradientOpCUDAKernel); diff --git a/paddle/operators/cross_entropy_op.h b/paddle/operators/cross_entropy_op.h index d1bbc2cb66..eb4d1348de 100644 --- a/paddle/operators/cross_entropy_op.h +++ b/paddle/operators/cross_entropy_op.h @@ -21,7 +21,7 @@ namespace operators { using Tensor = framework::Tensor; template -T tolerable_value(T x) { +inline T tolerable_value(const T x) { static_assert(std::is_floating_point::value, "tolerable_value works only on float, " "double and double double."); @@ -39,13 +39,16 @@ T tolerable_value(T x) { return x; } -template +template class OnehotCrossEntropyOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + auto X = ctx.Input("X"); const T* Xdata = X->data(); - const int* label_data = ctx.Input(1)->data(); + const int* label_data = ctx.Input("label")->data(); auto Y = ctx.Output("Y"); Y->mutable_data(ctx.GetPlace()); @@ -62,10 +65,13 @@ class OnehotCrossEntropyOpKernel : public framework::OpKernel { } }; -template +template class OnehotCrossEntropyGradientOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + auto X = ctx.Input("X"); auto dX = ctx.Output(framework::GradVarName("X")); auto dY = ctx.Input(framework::GradVarName("Y")); @@ -79,6 +85,8 @@ class OnehotCrossEntropyGradientOpKernel : public framework::OpKernel { const int batch_size = X->dims()[0]; const int class_num = X->dims()[1]; + // TODO(qingqing): make zero setting an common function. + memset(dXdata, 0, sizeof(T) * batch_size * class_num); for (int i = 0; i < batch_size; ++i) { int index = i * class_num + label_data[i]; dXdata[index] = -tolerable_value(dYdata[i] / Xdata[index]); diff --git a/paddle/operators/fc_op.cc b/paddle/operators/fc_op.cc deleted file mode 100644 index 01a1a81206..0000000000 --- a/paddle/operators/fc_op.cc +++ /dev/null @@ -1,76 +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/net_op.h" - -#include "paddle/framework/eigen.h" -#include "paddle/framework/op_registry.h" - -namespace paddle { -namespace operators { - -using OpRegistry = framework::OpRegistry; - -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 != framework::kEmptyVarName) { - 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 framework::OpProtoAndCheckerMaker { - public: - FullyConnectedOpMaker(framework::OpProto *proto, - framework::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); - -namespace ops = paddle::operators; -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 3759a88678..9d51f6e3a1 100644 --- a/paddle/operators/fill_zeros_like_op.cc +++ b/paddle/operators/fill_zeros_like_op.cc @@ -18,18 +18,13 @@ namespace paddle { namespace operators { class FillZerosLikeOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 1UL, - "Input size of FillZerosLikeOp must be one."); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL, - "Output size of AddOp must be one."); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), - "Input of FillZerosLikeOp must be set."); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0), - "Output of FillZerosLikeOp must be set."); - ctx.Output(0)->Resize( - ctx.Input(0)->dims()); + ctx.Output("Dst")->Resize( + ctx.Input("Src")->dims()); } }; @@ -51,7 +46,8 @@ The output will have the same size with input. } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(fill_zeros_like, ops::FillZerosLikeOp, ops::FillZerosLikeOpMaker); +REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, ops::FillZerosLikeOp, + ops::FillZerosLikeOpMaker); REGISTER_OP_CPU_KERNEL( fill_zeros_like, ops::FillZerosLikeKernel); diff --git a/paddle/operators/fill_zeros_like_op.h b/paddle/operators/fill_zeros_like_op.h index f846c7a8ab..969998ce2e 100644 --- a/paddle/operators/fill_zeros_like_op.h +++ b/paddle/operators/fill_zeros_like_op.h @@ -23,10 +23,10 @@ template class FillZerosLikeKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* output = context.Output(0); + auto* output = context.Output("Dst"); output->mutable_data(context.GetPlace()); auto t = framework::EigenVector::Flatten(*output); - t.device(context.GetEigenDevice()) = t.constant(T(0)); + t.device(context.GetEigenDevice()) = t.constant(static_cast(0)); } }; diff --git a/paddle/operators/gather.h b/paddle/operators/gather.h new file mode 100644 index 0000000000..92fb51ec17 --- /dev/null +++ b/paddle/operators/gather.h @@ -0,0 +1,74 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT 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/eigen.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +namespace paddle { +namespace operators { + +// Implementation of CPU copy +template +void CPUGather(const T* src, const int* indices, const int slice_size, + const int index_size, T* output) { + const size_t slice_bytes = slice_size * sizeof(T); + + for (int i = 0; i < index_size; ++i) { + int index_ = indices[i]; + memcpy(output + i * slice_size, src + index_ * slice_size, slice_bytes); + } +} + +// Implementation of GPU copy: +template +void GPUGather(const T* src, const int* index, const int slice_size, + const int index_size, T* output); + +/** + * 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(); + 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_op.cc b/paddle/operators/gather_op.cc new file mode 100644 index 0000000000..123bed296c --- /dev/null +++ b/paddle/operators/gather_op.cc @@ -0,0 +1,72 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/gather_op.h" +#include "paddle/framework/ddim.h" + +namespace paddle { +namespace operators { + +class GatherOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + int batch_size = ctx.Input("Index")->dims()[0]; + PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0"); + framework::DDim output_dims(ctx.Input("X")->dims()); + output_dims[0] = batch_size; + ctx.Output("Out")->Resize(output_dims); + } +}; + +class GatherGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + auto X_grad = ctx.Output(framework::GradVarName("X")); + auto X = ctx.Input("X"); + + X_grad->Resize(X->dims()); + } +}; + +class GatherOpMaker : public framework::OpProtoAndCheckerMaker { + public: + GatherOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The source input of gather op"); + AddInput("Index", "The index input of gather op"); + AddOutput("Out", "The output of add op"); + AddComment(R"DOC( +Gather Operator by selecting from the first axis, + +Out = X[Index] +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(gather, ops::GatherOp, ops::GatherOpMaker, gather_grad, + ops::GatherGradOp); +REGISTER_OP_CPU_KERNEL(gather, + ops::GatherOpKernel); +REGISTER_OP_CPU_KERNEL( + gather_grad, + ops::GatherGradientOpKernel); diff --git a/paddle/operators/gather_op.cu b/paddle/operators/gather_op.cu new file mode 100644 index 0000000000..3f04a7b3f8 --- /dev/null +++ b/paddle/operators/gather_op.cu @@ -0,0 +1,20 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/gather_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(gather, + ops::GatherOpKernel); diff --git a/paddle/operators/gather_op.h b/paddle/operators/gather_op.h new file mode 100644 index 0000000000..381854f301 --- /dev/null +++ b/paddle/operators/gather_op.h @@ -0,0 +1,53 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "gather.h" +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "scatter.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class GatherOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto *X = ctx.Input("X"); + auto *Index = ctx.Input("Index"); + auto *Y = ctx.Output("Out"); + + Y->mutable_data(ctx.GetPlace()); + Gather(ctx.GetPlace(), X, Index, Y); + } +}; + +template +class GatherGradientOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto *Index = ctx.Input("Index"); + auto *dX = ctx.Output(framework::GradVarName("X")); + auto *dO = ctx.Input(framework::GradVarName("Out")); + + dX->mutable_data(ctx.GetPlace()); + ScatterUpdate(ctx.GetPlace(), dO, Index, dX); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/gather_test.cc b/paddle/operators/gather_test.cc new file mode 100644 index 0000000000..0ae1e99452 --- /dev/null +++ b/paddle/operators/gather_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/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); + + delete src; + delete index; + delete output; +} diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc new file mode 100644 index 0000000000..a85363ad81 --- /dev/null +++ b/paddle/operators/gaussian_random_op.cc @@ -0,0 +1,83 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class CPUGaussianRandomKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + float mean = context.op_.GetAttr("mean"); + float std = context.op_.GetAttr("std"); + auto* tensor = context.Output("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::normal_distribution dist(mean, std); + ssize_t size = framework::product(tensor->dims()); + for (ssize_t i = 0; i < size; ++i) { + data[i] = dist(engine); + } + } +}; + +class GaussianRandomOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext& context) const override { + auto* tensor = context.Output("Out"); + 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 of random tensor.").SetDefault(.0f); + AddAttr("std", "std of random tensor.").SetDefault(1.0f); + AddAttr("seed", + "Random seed of generator." + "0 means use system wide seed") + .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::CPUGaussianRandomKernel); diff --git a/paddle/operators/gaussian_random_op.cu b/paddle/operators/gaussian_random_op.cu new file mode 100644 index 0000000000..018a4bfcb2 --- /dev/null +++ b/paddle/operators/gaussian_random_op.cu @@ -0,0 +1,65 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include +#include +#include +#include +#include "paddle/framework/op_registry.h" +#include "paddle/framework/operator.h" + +namespace paddle { +namespace operators { + +template +struct GaussianGenerator { + T mean_, std_; + unsigned int seed_; + + __host__ __device__ GaussianGenerator(T mean, T std, int seed) + : mean_(mean), std_(std), seed_(seed) {} + + __host__ __device__ T operator()(const unsigned int n) const { + thrust::minstd_rand rng; + rng.seed(seed_); + thrust::normal_distribution dist(mean_, std_); + rng.discard(n); + return dist(rng); + } +}; + +template +class GPUGaussianRandomKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + 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 mean = static_cast(context.op_.GetAttr("mean")); + T std = static_cast(context.op_.GetAttr("std")); + thrust::counting_iterator index_sequence_begin(0); + ssize_t N = framework::product(tensor->dims()); + thrust::transform(index_sequence_begin, index_sequence_begin + N, + thrust::device_ptr(data), + GaussianGenerator(mean, std, seed)); + } +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OP_GPU_KERNEL(gaussian_random, + paddle::operators::GPUGaussianRandomKernel); diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc new file mode 100644 index 0000000000..94d40890a7 --- /dev/null +++ b/paddle/operators/lookup_table_op.cc @@ -0,0 +1,72 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/lookup_table_op.h" + +namespace paddle { +namespace operators { + +class LookupTableOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &context) const override { + auto table_t = context.Input("W"); + auto ids_t = context.Input("Ids"); + auto output_t = context.Output("Out"); + + output_t->Resize({ids_t->dims()[0], table_t->dims()[1]}); + } +}; + +class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { + public: + LookupTableOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("W", + "An input represents embedding tensors," + " which is a learnable parameter."); + AddInput("Ids", + "An input with type int32 or int64" + "contains the ids to be looked up in W."); + AddOutput("Out", "The lookup results, which have the same type with W."); + AddComment( + "This operator is used to perform lookups on the parameter W," + "then concatenated into a dense tensor."); + } +}; + +class LookupTableOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &context) const override { + auto table = context.Input("W"); + auto d_table = context.Output(framework::GradVarName("W")); + d_table->Resize(table->dims()); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(lookup_table, ops::LookupTableOp, ops::LookupTableOpMaker, + lookup_table_grad, ops::LookupTableOpGrad); + +REGISTER_OP_CPU_KERNEL(lookup_table, ops::LookupTableKernel); +REGISTER_OP_CPU_KERNEL(lookup_table_grad, ops::LookupTableGradKernel); diff --git a/paddle/operators/lookup_table_op.cu b/paddle/operators/lookup_table_op.cu new file mode 100644 index 0000000000..27eee3436a --- /dev/null +++ b/paddle/operators/lookup_table_op.cu @@ -0,0 +1,116 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "paddle/platform/assert.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +__global__ void LookupTable(T* output, const T* table, const int32_t* ids, + const int N, const int K, const int D) { + int idx = threadIdx.x; + int idy = blockIdx.x + threadIdx.y * GridDimX; + + while (idy < K) { + int id = ids[idy]; + PADDLE_ASSERT(id >= 0); + PADDLE_ASSERT(id < N); + T* out = output + idy * D; + const T* tab = table + id * D; + for (int i = idx; i < D; i += BlockDimX) { + out[i] = tab[i]; + } + idy += BlockDimY * GridDimX; + } +} + +template +__global__ void LookupTableGrad(T* table, const T* output, const int32_t* ids, + const int N, const int K, const int D) { + int idx = threadIdx.x; + int idy = blockIdx.x + threadIdx.y * GridDimX; + + while (idy < K) { + int id = ids[idy]; + PADDLE_ASSERT(id >= 0); + PADDLE_ASSERT(id < N); + const T* out = output + idy * D; + T* tab = table + id * D; + for (int i = idx; i < D; i += BlockDimX) { + paddle::platform::CudaAtomicAdd(&tab[i], out[i]); + } + idy += BlockDimY * GridDimX; + } +} + +template +class LookupTableCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto table_t = context.Input("W"); + auto ids_t = context.Input("Ids"); + auto output_t = context.Output("Out"); + + size_t N = table_t->dims()[0]; + size_t D = table_t->dims()[1]; + size_t K = product(ids_t->dims()); + auto ids = ids_t->data(); + auto table = table_t->data(); + auto output = output_t->mutable_data(context.GetPlace()); + + dim3 threads(128, 8); + dim3 grids(8, 1); + LookupTable<<>>(output, table, ids, N, K, D); + } +}; + +template +class LookupTableGradCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto ids_t = context.Input("Ids"); + auto d_output_t = context.Input(framework::GradVarName("Out")); + auto d_table_t = context.Output(framework::GradVarName("W")); + + int N = d_table_t->dims()[0]; + int D = d_table_t->dims()[1]; + int K = product(ids_t->dims()); + const int32_t* ids = ids_t->data(); + const T* d_output = d_output_t->data(); + T* d_table = d_table_t->mutable_data(context.GetPlace()); + + auto t = framework::EigenVector::Flatten(*d_table_t); + t.device(context.GetEigenDevice()) = + t.constant(static_cast(0)); + + dim3 threads(128, 8); + dim3 grids(8, 1); + LookupTableGrad<<>>(d_table, d_output, ids, N, + K, D); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(lookup_table, ops::LookupTableCUDAKernel); +REGISTER_OP_GPU_KERNEL(lookup_table_grad, + ops::LookupTableGradCUDAKernel); diff --git a/paddle/operators/lookup_table_op.h b/paddle/operators/lookup_table_op.h new file mode 100644 index 0000000000..4da8079b91 --- /dev/null +++ b/paddle/operators/lookup_table_op.h @@ -0,0 +1,75 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class LookupTableKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto table_t = context.Input("W"); // float tensor + auto ids_t = context.Input("Ids"); // int tensor + auto output_t = context.Output("Out"); // float tensor + + size_t N = table_t->dims()[0]; + size_t D = table_t->dims()[1]; + auto ids = ids_t->data(); + auto table = table_t->data(); + auto output = output_t->mutable_data(context.GetPlace()); + for (size_t i = 0; i < product(ids_t->dims()); ++i) { + PADDLE_ENFORCE_LT(ids[i], N); + PADDLE_ENFORCE_GE(ids[i], 0); + memcpy(output + i * D, table + ids[i] * D, D * sizeof(T)); + } + } +}; + +template +class LookupTableGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto ids_t = context.Input("Ids"); + auto d_output_t = context.Input(framework::GradVarName("Out")); + auto d_table_t = context.Output(framework::GradVarName("W")); + + size_t N = d_table_t->dims()[0]; + size_t D = d_table_t->dims()[1]; + auto ids = ids_t->data(); + const T* d_output = d_output_t->data(); + T* d_table = d_table_t->mutable_data(context.GetPlace()); + + auto t = framework::EigenVector::Flatten(*d_table_t); + t.device(context.GetEigenDevice()) = + t.constant(static_cast(0)); + + for (size_t i = 0; i < product(ids_t->dims()); ++i) { + PADDLE_ENFORCE_LT(ids[i], N); + PADDLE_ENFORCE_GE(ids[i], 0); + for (size_t j = 0; j < D; ++j) { + d_table[ids[i] * D + j] += d_output[i * D + j]; + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/CMakeLists.txt b/paddle/operators/math/CMakeLists.txt new file mode 100644 index 0000000000..ed51d416ed --- /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 0000000000..1e86fc3d16 --- /dev/null +++ b/paddle/operators/math/math_function.cc @@ -0,0 +1,114 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/math/math_function.h" + +namespace paddle { +namespace operators { +namespace math { + +template <> +void gemm(const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, const int M, + const int N, const int K, + const float alpha, const float* A, + const float* B, const float beta, float* C, + platform::DeviceContext* context) { + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + int ldc = N; + cblas_sgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, + beta, C, ldc); +} + +template <> +void gemm(const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, const int M, + const int N, const int K, + const double alpha, const double* A, + const double* B, const double beta, + double* C, + platform::DeviceContext* context) { + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + int ldc = N; + cblas_dgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb, + beta, C, ldc); +} + +template <> +void matmul(const framework::Tensor& matrix_a, + bool trans_a, + const framework::Tensor& matrix_b, + bool trans_b, float alpha, + framework::Tensor* matrix_out, + float beta, + platform::DeviceContext* context) { + auto dim_a = matrix_a.dims(); + auto dim_b = matrix_b.dims(); + auto dim_out = matrix_out->dims(); + PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, + "The input and output of matmul be matrix"); + + PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) && + platform::is_cpu_place(matrix_b.place()) && + platform::is_cpu_place(matrix_out->place()), + "Matrix must all be in CPUPlace"); + + int M = dim_out[0]; + int N = dim_out[1]; + int K = (trans_a == false) ? dim_a[1] : dim_a[0]; + + CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; + CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; + + gemm( + transA, transB, M, N, K, alpha, matrix_a.data(), + matrix_b.data(), beta, matrix_out->data(), context); +} + +template <> +void matmul(const framework::Tensor& matrix_a, + bool trans_a, + const framework::Tensor& matrix_b, + bool trans_b, double alpha, + framework::Tensor* matrix_out, + double beta, + platform::DeviceContext* context) { + auto dim_a = matrix_a.dims(); + auto dim_b = matrix_b.dims(); + auto dim_out = matrix_out->dims(); + PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, + "The input and output of matmul be matrix"); + + PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) && + platform::is_cpu_place(matrix_b.place()) && + platform::is_cpu_place(matrix_out->place()), + "Matrix must all be in CPUPlace"); + + int M = dim_out[0]; + int N = dim_out[1]; + int K = (trans_a == false) ? dim_a[1] : dim_a[0]; + + CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; + CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; + + gemm( + transA, transB, M, N, K, alpha, matrix_a.data(), + matrix_b.data(), beta, matrix_out->data(), context); +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/math_function.cu b/paddle/operators/math/math_function.cu new file mode 100644 index 0000000000..da40b27c94 --- /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 0000000000..155589fadb --- /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 0000000000..6c020c4ff7 --- /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 997b0c514e..d3d0e55a67 100644 --- a/paddle/operators/mean_op.cc +++ b/paddle/operators/mean_op.cc @@ -18,13 +18,14 @@ namespace paddle { namespace operators { class MeanOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 1, "Input size of AddOp must be one"); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "Output size of AddOp must be one"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "input should be set"); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0), "output should be set"); - ctx.Output(0)->Resize(framework::make_ddim({1})); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), + "Input of MeanOp must be initialized."); + ctx.Output("Out")->Resize({1}); } }; @@ -33,15 +34,18 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker { MeanOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The input of mean op"); - AddOutput("Out", "The output of mean op").IgnoreGradient(); + AddOutput("Out", "The output of mean op").NotInGradient(); AddComment("Mean Operator"); } }; class MeanGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - ctx.Output("X" + framework::kGradVarSuffix) + ctx.Output(framework::GradVarName("X")) ->Resize(ctx.Input("X")->dims()); } }; @@ -50,9 +54,8 @@ class MeanGradOp : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker); +REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker, mean_grad, ops::MeanGradOp); REGISTER_OP_CPU_KERNEL(mean, ops::MeanKernel); -REGISTER_GRADIENT_OP(mean, mean_grad, ops::MeanGradOp); REGISTER_OP_CPU_KERNEL(mean_grad, ops::MeanGradKernel); diff --git a/paddle/operators/mean_op.h b/paddle/operators/mean_op.h index f3db0a29bb..9848af280b 100644 --- a/paddle/operators/mean_op.h +++ b/paddle/operators/mean_op.h @@ -31,14 +31,14 @@ template class MeanKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto input = context.Input(0); - auto output = context.Output(0); + auto* input = context.Input("X"); + auto* output = context.Output("Out"); output->mutable_data(context.GetPlace()); auto X = EigenVector::Flatten(*input); auto y = EigenScalar::From(*output); - auto place = context.GetEigenDevice(); + auto& place = context.GetEigenDevice(); y.device(place) = X.mean(); } @@ -48,16 +48,17 @@ template class MeanGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto OG = context.Input("Out" + framework::kGradVarSuffix); + auto OG = context.Input(framework::GradVarName("Out")); PADDLE_ENFORCE(framework::product(OG->dims()) == 1, "Mean Gradient should be scalar"); - auto IG = context.Output("X" + framework::kGradVarSuffix); + auto IG = context.Output(framework::GradVarName("X")); IG->mutable_data(context.GetPlace()); T ig_size = (T)framework::product(IG->dims()); + Eigen::DSizes bcast(ig_size); EigenVector::Flatten(*IG).device(context.GetEigenDevice()) = - EigenScalar::From(*OG) / ig_size; + (EigenVector::From(*OG) / ig_size).broadcast(bcast); } }; diff --git a/paddle/operators/mean_op_test.cc b/paddle/operators/mean_op_test.cc deleted file mode 100644 index 375dcd50e1..0000000000 --- a/paddle/operators/mean_op_test.cc +++ /dev/null @@ -1,25 +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(mean); - -TEST(MeanOp, GetOpProto) { - auto& protos = paddle::framework::OpRegistry::protos(); - auto it = protos.find("mean"); - ASSERT_NE(it, protos.end()); -} diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc new file mode 100644 index 0000000000..1eee9644ba --- /dev/null +++ b/paddle/operators/minus_op.cc @@ -0,0 +1,87 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/minus_op.h" +#include "paddle/operators/net_op.h" + +namespace paddle { +namespace operators { + +class MinusOp : public framework::OperatorWithKernel { + public: + MinusOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + auto *left_tensor = ctx.Input("X"); + auto *right_tensor = ctx.Input("Y"); + + PADDLE_ENFORCE_EQ( + framework::product(left_tensor->dims()), + framework::product(right_tensor->dims()), + "Minus operator must take two tensor with same num of elements"); + ctx.Output("Out")->Resize(left_tensor->dims()); + } +}; + +class MinusOpMaker : public framework::OpProtoAndCheckerMaker { + public: + MinusOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The left tensor of minus operator.").NotInGradient(); + AddInput("Y", "The right tensor of minus operator.").NotInGradient(); + AddOutput("Out", "The output tensor of minus operator.").NotInGradient(); + + AddComment(R"DOC(Minus Operator + +Equation: Out = X - Y +)DOC"); + } +}; +template +class MinusGradOp : public NetOp { + public: + MinusGradOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : NetOp(type, inputs, outputs, attrs) { + auto out_grad = Input(framework::GradVarName("Out")); + auto x_grad = Output(framework::GradVarName("X")); + auto y_grad = Output(framework::GradVarName("Y")); + + // x_grad = out_grad + AppendOp(framework::OpRegistry::CreateOp("identity", {{"X", {out_grad}}}, + {{"Out", {x_grad}}}, {})); + + framework::AttributeMap scale_attr; + scale_attr["scale"] = static_cast(-1); + AppendOp(framework::OpRegistry::CreateOp("scale", {{"X", {out_grad}}}, + {{"Out", {y_grad}}}, scale_attr)); + CompleteAddOp(false); + } +}; + +} // namespace operators +} // namespace paddle + +USE_OP(scale); +USE_OP_ITSELF(identity); +namespace ops = paddle::operators; +REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, minus_grad, + ops::MinusGradOp); +REGISTER_OP_CPU_KERNEL(minus, + ops::MinusKernel); diff --git a/paddle/operators/minus_op.cu b/paddle/operators/minus_op.cu new file mode 100644 index 0000000000..a8375cc630 --- /dev/null +++ b/paddle/operators/minus_op.cu @@ -0,0 +1,18 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/minus_op.h" + +REGISTER_OP_GPU_KERNEL( + minus, paddle::operators::MinusKernel); diff --git a/paddle/operators/minus_op.h b/paddle/operators/minus_op.h new file mode 100644 index 0000000000..6310a4fd51 --- /dev/null +++ b/paddle/operators/minus_op.h @@ -0,0 +1,39 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class MinusKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* left_tensor = context.Input("X"); + auto* right_tensor = context.Input("Y"); + auto* out_tensor = context.Output("Out"); + + out_tensor->mutable_data(context.GetPlace()); + auto& dev = context.GetEigenDevice(); + framework::EigenVector::Flatten(*out_tensor).device(dev) = + framework::EigenVector::Flatten(*left_tensor) - + framework::EigenVector::Flatten(*right_tensor); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc index db81fd555d..173cc3850c 100644 --- a/paddle/operators/mul_op.cc +++ b/paddle/operators/mul_op.cc @@ -17,12 +17,16 @@ namespace paddle { namespace operators { +using framework::Tensor; + class MulOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 2, "The mul op must take two inputs"); - auto dim0 = ctx.Input(0)->dims(); - auto dim1 = ctx.Input(1)->dims(); + 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")); @@ -32,8 +36,7 @@ class MulOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_EQ( dim0[1], dim1[0], "First matrix's width must be equal with second matrix's height."); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "The mul op takes only one output"); - ctx.Output(0)->Resize({dim0[0], dim1[1]}); + ctx.Output("Out")->Resize({dim0[0], dim1[1]}); } }; @@ -53,11 +56,27 @@ The equation is: Out = X * Y }; 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 ""; + void InferShape(const framework::InferShapeContext &ctx) const override { + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null"); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null"); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto x_dims = ctx.Input("X")->dims(); + auto y_dims = ctx.Input("Y")->dims(); + auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); + auto *x_grad = ctx.Output(framework::GradVarName("X")); + auto *y_grad = ctx.Output(framework::GradVarName("Y")); + PADDLE_ENFORCE(x_dims[0] == out_dims[0], + "Out@GRAD M X N must equal to X dims 0, M "); + PADDLE_ENFORCE(y_dims[1] == out_dims[1], + "Out@GRAD M X N must equal to Y dims 1, N "); + + x_grad->Resize(x_dims); + y_grad->Resize(y_dims); } }; @@ -65,7 +84,7 @@ class MulOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker); -REGISTER_GRADIENT_OP(mul, mul_grad, ops::MulOpGrad); - +REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad); REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); +REGISTER_OP_CPU_KERNEL(mul_grad, + ops::MulGradKernel); diff --git a/paddle/operators/mul_op.cu b/paddle/operators/mul_op.cu index 43debbc21a..a81444dbe6 100644 --- a/paddle/operators/mul_op.cu +++ b/paddle/operators/mul_op.cu @@ -16,5 +16,6 @@ #include "paddle/operators/mul_op.h" namespace ops = paddle::operators; - REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel); +REGISTER_OP_GPU_KERNEL(mul_grad, + ops::MulGradKernel); diff --git a/paddle/operators/mul_op.h b/paddle/operators/mul_op.h index ab12631c03..8facc02814 100644 --- a/paddle/operators/mul_op.h +++ b/paddle/operators/mul_op.h @@ -13,6 +13,9 @@ limitations under the License. */ #pragma once + +#include "paddle/operators/math/math_function.h" + #include "paddle/framework/eigen.h" #include "paddle/framework/op_registry.h" @@ -28,21 +31,34 @@ template 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); - - output->mutable_data(context.GetPlace()); + auto* X = context.Input("X"); + auto* Y = context.Input("Y"); + auto* Z = context.Output("Out"); + Z->mutable_data(context.GetPlace()); + auto* device_context = + const_cast(context.device_context_); + math::matmul(*X, false, *Y, false, 1, Z, 0, device_context); + } +}; - auto X = EigenMatrix::From(*input0); - auto Y = EigenMatrix::From(*input1); - auto Z = EigenMatrix::From(*output); - auto place = context.GetEigenDevice(); +template +class MulGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* X = ctx.Input("X"); + auto* Y = ctx.Input("Y"); + auto* dOut = ctx.Input(framework::GradVarName("Out")); - Z.device(place) = X.contract(Y, dim_pair); + auto* dX = ctx.Output(framework::GradVarName("X")); + auto* dY = ctx.Output(framework::GradVarName("Y")); + dX->mutable_data(ctx.GetPlace()); + dY->mutable_data(ctx.GetPlace()); + auto* device_context = + const_cast(ctx.device_context_); + // dX = dOut * Y'. dX: M x K, dOut : M x N, Y : K x N + math::matmul(*dOut, false, *Y, true, 1, dX, 0, device_context); + // dY = X' * dOut. dY: K x N, dOut : M x N, X : M x K + math::matmul(*X, true, *dOut, false, 1, dY, 0, device_context); } }; diff --git a/paddle/operators/net_op.cc b/paddle/operators/net_op.cc index a466c4f30f..78b5e27678 100644 --- a/paddle/operators/net_op.cc +++ b/paddle/operators/net_op.cc @@ -15,48 +15,45 @@ */ #include "paddle/operators/net_op.h" +#include +#include "paddle/framework/op_registry.h" namespace paddle { namespace operators { +const char NetOp::kAll[] = "all"; + void NetOp::CompleteAddOp(bool calc) { add_op_done_ = true; if (!calc) return; - std::unordered_set input_set; - std::unordered_set output_set; - std::unordered_set temp_output; + std::set input_set; + std::set output_set; for (auto& op : ops_) { - for (auto& ipt : op->inputs_) { - if (!Contains(output_set, ipt)) { // Not other op's output - input_set.insert(ipt); - } else { - temp_output.insert(ipt); + for (auto& ipt : op->Inputs()) { + for (auto& var_name : ipt.second) { + // If input variable has been in output set, then it will be + // added into intermediate_outputs_. Otherwise, it will be + // added into input set. + if (Contains(output_set, var_name)) { + intermediate_outputs_.insert(var_name); + } else { + input_set.insert(var_name); + } } } - for (auto& opt : op->outputs_) { - output_set.insert(opt); - } - } - - inputs_.reserve(input_set.size()); - std::copy(input_set.begin(), input_set.end(), std::back_inserter(inputs_)); - std::sort(inputs_.begin(), inputs_.end()); - - outputs_.reserve(output_set.size()); - std::copy(output_set.begin(), output_set.end(), std::back_inserter(outputs_)); - std::sort(outputs_.begin(), outputs_.end()); - - std::vector tmp_index; - tmp_index.reserve(temp_output.size()); - int output_len = static_cast(outputs_.size()); - for (int i = 0; i < output_len; ++i) { - if (Contains(temp_output, outputs_[i])) { - tmp_index.push_back(i); + for (auto& opt : op->Outputs()) { + for (auto& var_name : opt.second) { + output_set.insert(var_name); + } } } - - attrs_["temporary_index"] = tmp_index; + auto& inputs = inputs_[kAll]; + inputs.reserve(input_set.size()); + std::copy(input_set.begin(), input_set.end(), std::back_inserter(inputs)); + auto& outputs = outputs_[kAll]; + outputs.reserve(output_set.size()); + std::copy(output_set.begin(), output_set.end(), std::back_inserter(outputs)); } std::string NetOp::DebugString() const { @@ -73,5 +70,36 @@ std::string NetOp::DebugString() const { bool NetOp::IsNetOp() const { return true; } +std::vector NetOp::OutputVars(bool has_intermediate) const { + std::vector all; + for (auto& pair : this->outputs_) { + for (auto& var_name : pair.second) { + all.push_back(var_name); + } + } + if (has_intermediate) { + return all; + } + std::vector ret_val; + for (auto& each : all) { + if (!Contains(intermediate_outputs_, each)) { + ret_val.push_back(each); + } + } + return ret_val; +} + +NetOp::NetOp(const std::string& type, const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : framework::OperatorBase(type, inputs, outputs, attrs) {} + +std::unique_ptr NetOp::Clone() const { + PADDLE_ENFORCE( + add_op_done_, + "Must clone a sealed NetOp, invoke Net::CompleteAddOp before clone"); + return std::unique_ptr(new NetOp(*this)); +} + } // namespace operators } // namespace paddle diff --git a/paddle/operators/net_op.h b/paddle/operators/net_op.h index 792b336675..fcd8134b2c 100644 --- a/paddle/operators/net_op.h +++ b/paddle/operators/net_op.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include "paddle/framework/framework.pb.h" #include "paddle/framework/op_registry.h" namespace paddle { @@ -35,6 +36,23 @@ namespace operators { */ class NetOp : public framework::OperatorBase { public: + static const char kAll[]; + NetOp() : framework::OperatorBase("plain_net", {}, {}, {}) {} + + NetOp(const std::string& type, const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs); + + NetOp(const NetOp& o) : framework::OperatorBase(o.type_, {}, {}, o.attrs_) { + this->ops_.reserve(o.ops_.size()); + std::transform( + o.ops_.begin(), o.ops_.end(), std::back_inserter(this->ops_), + [](const std::unique_ptr& op) { + return std::unique_ptr(op->Clone()); + }); + this->CompleteAddOp(); + } + /** * Infer all the operators' input and output variables' shapes, will be called * before every mini-batch @@ -68,21 +86,28 @@ class NetOp : public framework::OperatorBase { return true; } + void AppendOp(const framework::OperatorBase& op) { AppendOp(op.Clone()); } + /** * @brief Add an operator by ptr */ - void AddOp(const std::shared_ptr& op) { - PADDLE_ENFORCE(!add_op_done_, "Cannot AddOp when this network is sealed"); + void AppendOp(std::unique_ptr op) { + PADDLE_ENFORCE(!add_op_done_, + "Cannot AppendOp when this network is sealed"); PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op"); - ops_.push_back(op); + ops_.push_back(std::move(op)); } - void InsertOp(size_t pos, const std::shared_ptr& op) { + void InsertOp(size_t pos, std::unique_ptr op) { PADDLE_ENFORCE(!add_op_done_, "Cannot InsertOp when this network is sealed"); PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op"); PADDLE_ENFORCE_LE(pos, ops_.size(), "Out of range"); - ops_.insert(ops_.begin() + pos, op); + ops_.insert(ops_.begin() + pos, std::move(op)); + } + + void InsertOp(size_t pos, const framework::OperatorBase& op) { + InsertOp(pos, op.Clone()); } void CompleteAddOp(bool calculate = true); @@ -90,11 +115,15 @@ class NetOp : public framework::OperatorBase { std::string DebugString() const override; bool IsNetOp() const override; + std::vector OutputVars(bool has_intermediate) const override; + + std::unique_ptr Clone() const override; - std::vector> ops_; + std::vector> ops_; private: bool add_op_done_{false}; + std::set intermediate_outputs_; template static bool Contains(T container, KeyType key) { diff --git a/paddle/operators/net_op_test.cc b/paddle/operators/net_op_test.cc index 76bf79f9b5..f2e98ee7a1 100644 --- a/paddle/operators/net_op_test.cc +++ b/paddle/operators/net_op_test.cc @@ -12,6 +12,8 @@ static int run_cnt = 0; class TestOp : public framework::OperatorBase { public: + using framework::OperatorBase::OperatorBase; + DEFINE_OP_CLONE_METHOD(TestOp); void InferShape(const Scope& scope) const override { ++infer_shape_cnt; } void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const override { @@ -19,12 +21,6 @@ class TestOp : public framework::OperatorBase { } }; -class EmptyOp : public framework::OperatorBase { - public: - void InferShape(const Scope& scope) const override {} - void Run(const Scope& scope, const DeviceContext& dev_ctx) const override {} -}; - template void AssertSameVectorWithoutOrder(const std::vector& expected, const std::vector& actual) { @@ -42,46 +38,51 @@ TEST(OpKernel, all) { auto net = std::make_shared(); ASSERT_NE(net, nullptr); - auto op1 = std::make_shared(); - op1->inputs_ = {"x", "w1", "b1"}; - op1->outputs_ = {"y"}; - net->AddOp(op1); - - auto op2 = std::make_shared(); - op2->inputs_ = {"y", "w2", "b2"}; - op2->outputs_ = {"z"}; - net->AddOp(op2); + net->AppendOp(std::unique_ptr( + new TestOp("test", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}}, + {{"Out", {"y"}}}, {}))); + net->AppendOp(std::unique_ptr( + new TestOp("test", {{"X", {"y"}}, {"W", {"w2"}}, {"b", {"b2"}}}, + {{"Out", {"z"}}}, {}))); net->CompleteAddOp(); - AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"}, net->inputs_); - AssertSameVectorWithoutOrder({"y", "z"}, net->outputs_); - auto tmp_idx_iter = net->attrs_.find("temporary_index"); - ASSERT_NE(net->attrs_.end(), tmp_idx_iter); - auto& tmp_idx = boost::get>(tmp_idx_iter->second); - ASSERT_EQ(1UL, tmp_idx.size()); - ASSERT_EQ("y", net->outputs_[tmp_idx[0]]); + AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"}, + net->Inputs(NetOp::kAll)); + AssertSameVectorWithoutOrder({"y", "z"}, net->Outputs(NetOp::kAll)); - Scope scope; - platform::CPUDeviceContext dev_ctx; + auto final_outs = net->OutputVars(false); - net->InferShape(scope); - net->Run(scope, dev_ctx); - ASSERT_EQ(2, infer_shape_cnt); - ASSERT_EQ(2, run_cnt); - ASSERT_THROW(net->AddOp(op2), platform::EnforceNotMet); + ASSERT_EQ(final_outs.size(), 1UL); + ASSERT_EQ(final_outs[0], "z"); } TEST(NetOp, insert_op) { NetOp net; - auto op1 = std::make_shared(); - op1->inputs_ = {"x", "w1", "b1"}; - op1->outputs_ = {"y"}; - net.AddOp(op1); - net.InsertOp(0, op1); + auto op1 = std::unique_ptr( + new framework::NOP("empty", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}}, + {{"Out", {"y"}}}, {})); + net.AppendOp(*op1); + net.InsertOp(0, *op1); ASSERT_EQ(2UL, net.ops_.size()); - net.InsertOp(2, op1); + net.InsertOp(2, std::move(op1)); ASSERT_EQ(3UL, net.ops_.size()); } +TEST(NetOp, Clone) { + NetOp net; + net.AppendOp( + std::unique_ptr(new framework::NOP{"empty", {}, {}, {}})); + net.AppendOp(std::unique_ptr( + new framework::NOP{"empty2", {}, {}, {}})); + net.CompleteAddOp(true); + auto new_net_op = net.Clone(); + ASSERT_NE(new_net_op, nullptr); + ASSERT_TRUE(new_net_op->IsNetOp()); + auto* new_net = static_cast(new_net_op.get()); + ASSERT_EQ(2UL, new_net->ops_.size()); + ASSERT_EQ(new_net->ops_[0]->Type(), "empty"); + ASSERT_EQ(new_net->ops_[1]->Type(), "empty2"); +} + } // namespace operators } // namespace paddle diff --git a/paddle/operators/recurrent_op.cc b/paddle/operators/recurrent_op.cc index 2438374205..16bd249cb3 100644 --- a/paddle/operators/recurrent_op.cc +++ b/paddle/operators/recurrent_op.cc @@ -36,15 +36,13 @@ void RecurrentAlgorithm::InferShape(const Scope& scope) const { rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, true /*infer_shape_mode*/); InitMemories(step_scopes[0], true /*infer_shape_mode*/); - Variable* net = scope.FindVar(arg_->step_net); - PADDLE_ENFORCE(net != nullptr, "failed to get step net"); for (size_t i = 0; i < seq_len_; i++) { if (i > 0) { rnn::LinkMemories(step_scopes, arg_->memories, i, -1, true /*infer_shape_mode*/); } - net->GetMutable()->InferShape(*step_scopes[i]); + (*stepnet_)->InferShape(*step_scopes[i]); } rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, true /*infer_shape_mode*/); @@ -56,7 +54,6 @@ void RecurrentAlgorithm::Run(const Scope& scope, rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, false /*infer_shape_mode*/); InitMemories(step_scopes[0], false /*infer_shape_mode*/); - Variable* net = scope.FindVar(arg_->step_net); for (size_t step_id = 0; step_id < seq_len_; step_id++) { // create output alias variables @@ -64,7 +61,7 @@ void RecurrentAlgorithm::Run(const Scope& scope, rnn::LinkMemories(step_scopes, arg_->memories, step_id, -1, false /*infer_shape_mode*/); } - net->GetMutable()->Run(*step_scopes[step_id], dev_ctx); + (*stepnet_)->Run(*step_scopes[step_id], dev_ctx); } rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, false /*infer_shape_mode*/); @@ -78,25 +75,28 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const { auto step_scopes = step_scopes_var->GetMutable>(); // Now all variables in scope must be created outside of op. - auto net_var = scope.FindVar(arg_->step_net); - PADDLE_ENFORCE(net_var != nullptr, "no stepnet called %s in scope", - arg_->step_net); - auto net_op = net_var->GetMutable(); - PADDLE_ENFORCE(!net_op->outputs_.empty(), "net_op has no outputs"); + PADDLE_ENFORCE_NOT_NULL(stepnet_); + PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "stepnet_ op has no outputs"); + PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "net_op has no outputs"); if (seq_len_ > step_scopes->size()) { for (size_t i = step_scopes->size(); i < seq_len_; ++i) { auto& step_scope = scope.NewScope(); // create step net's temp inputs - for (auto& input : net_op->inputs_) { + for (auto& input : (*stepnet_)->Inputs()) { // the weight are located in parent scope - if (!step_scope.FindVar(input)) - step_scope.NewVar(input)->GetMutable(); + for (auto& var_name : input.second) { + if (!step_scope.FindVar(var_name)) { + step_scope.NewVar(var_name)->GetMutable(); + } + } } // create stepnet's outputs - for (const auto& output : net_op->outputs_) { - step_scope.NewVar(output); + for (const auto& output : (*stepnet_)->Outputs()) { + for (auto& var_name : output.second) { + step_scope.NewVar(var_name); + } } step_scopes->emplace_back(&step_scope); } @@ -130,11 +130,13 @@ const rnn::ArgumentName RecurrentGradientOp::kArgName{ "inlink@grad", "inlink_alias", "outlink_alias", "memories", "pre_memories", "boot_memories@grad"}; -void RecurrentOp::Init() { - OperatorBase::Init(); - std::unique_ptr arg(new rnn::Argument()); - rnn::InitArgument(kArgName, arg.get(), *this); - alg_.Init(std::move(arg)); +RecurrentOp::RecurrentOp(const std::string& type, + const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) { + rnn::InitArgument(kArgName, &arg_, *this); + alg_.Init(&arg_, &stepnet_); } class RecurrentAlgorithmProtoAndCheckerMaker @@ -147,13 +149,12 @@ class RecurrentAlgorithmProtoAndCheckerMaker // inputs and outputs stored in proto AddInput(name.inlinks, "the inputs that need to be segmented for each step.") - .SetMultiple(); + .AsDuplicable(); AddInput(name.boot_memories, "variables to initialize memories.") - .SetMultiple(); - AddInput(name.step_net, "network shared by all steps."); + .AsDuplicable(); AddOutput(name.outlinks, "the outputs that need to concated for all steps.") - .SetMultiple(); + .AsDuplicable(); AddOutput(name.step_scopes, "step scopes"); // Attributes stored in AttributeMap @@ -172,14 +173,12 @@ void RecurrentGradientAlgorithm::Run( auto step_scopes = GetStepScopes(scope); rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, false /*infer_shape_mode*/); - Variable* net = scope.FindVar(arg_->step_net); - PADDLE_ENFORCE(net != nullptr, "failed to get step net"); for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) { if (static_cast(step_id) != seq_len_ - 1) { rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1, false /*infer_shape_mode*/); } - net->GetMutable()->Run(*step_scopes[step_id], dev_ctx); + (*stepnet_)->Run(*step_scopes[step_id], dev_ctx); } LinkBootMemoryGradients(step_scopes[0], false); rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, @@ -211,29 +210,30 @@ void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const { auto step_scopes = GetStepScopes(scope); rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_, true /*infer_shape_mode*/); - Variable* net = scope.FindVar(arg_->step_net); - PADDLE_ENFORCE(net != nullptr, "failed to get step net"); for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) { if (static_cast(step_id) != seq_len_ - 1) { rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1, true /*infer_shape_mode*/); } - net->GetMutable()->InferShape(*step_scopes[step_id]); + (*stepnet_)->InferShape(*step_scopes[step_id]); } rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_, true /*infer_shape_mode*/); LinkBootMemoryGradients(step_scopes[0], true /*infer_shape_mode*/); } -void RecurrentGradientOp::Init() { - OperatorBase::Init(); - std::unique_ptr arg(new rnn::Argument()); - rnn::InitArgument(kArgName, arg.get(), *this); - alg_.Init(std::move(arg)); +RecurrentGradientOp::RecurrentGradientOp( + const std::string& type, const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) { + rnn::InitArgument(kArgName, &arg_, *this); + alg_.Init(&arg_, &stepnet_); } } // namespace operators } // namespace paddle -REGISTER_OP(recurrent_op, paddle::operators::RecurrentOp, - paddle::operators::RecurrentAlgorithmProtoAndCheckerMaker); +REGISTER_OP_WITHOUT_GRADIENT( + recurrent_op, paddle::operators::RecurrentOp, + paddle::operators::RecurrentAlgorithmProtoAndCheckerMaker); diff --git a/paddle/operators/recurrent_op.h b/paddle/operators/recurrent_op.h index d1e60fed9c..1033d657a3 100644 --- a/paddle/operators/recurrent_op.h +++ b/paddle/operators/recurrent_op.h @@ -15,6 +15,7 @@ #pragma once #include "paddle/framework/operator.h" +#include "paddle/operators/net_op.h" #include "paddle/operators/rnn/recurrent_op_utils.h" namespace paddle { @@ -33,7 +34,12 @@ class RecurrentAlgorithm { void Run(const framework::Scope& scope, const platform::DeviceContext& dev_ctx) const; - void Init(std::unique_ptr arg) { arg_ = std::move(arg); } + void Init(rnn::Argument* arg, + std::unique_ptr* stepnet) { + PADDLE_ENFORCE_NOT_NULL(stepnet, "stepnet should be set before."); + arg_ = arg; + stepnet_ = stepnet; + } /** * InferShape must be called before Run. @@ -58,7 +64,8 @@ class RecurrentAlgorithm { void InitMemories(framework::Scope* step_scopes, bool infer_shape_mode) const; private: - std::unique_ptr arg_; + std::unique_ptr* stepnet_; + rnn::Argument* arg_; mutable size_t seq_len_; }; @@ -74,7 +81,12 @@ class RecurrentGradientAlgorithm { * operator. */ public: - void Init(std::unique_ptr arg) { arg_ = std::move(arg); } + void Init(rnn::Argument* arg, + std::unique_ptr* stepnet) { + PADDLE_ENFORCE_NOT_NULL(stepnet, "stepnet should be set before."); + arg_ = std::move(arg); + stepnet_ = stepnet; + } void Run(const framework::Scope& scope, const platform::DeviceContext& dev_ctx) const; @@ -95,14 +107,23 @@ class RecurrentGradientAlgorithm { } private: - std::unique_ptr arg_; + rnn::Argument* arg_; mutable size_t seq_len_; + std::unique_ptr* stepnet_; }; -class RecurrentOp final : public framework::OperatorBase { +class RecurrentOp : public framework::OperatorBase { public: - void Init() override; - + RecurrentOp(const std::string& type, const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs); + + RecurrentOp(const RecurrentOp& o) + : framework::OperatorBase( + static_cast(o)) { + // TODO(yuyang18): Implement copy ctor well. + PADDLE_THROW("Not implemented"); + } /** * InferShape must be called before Run. */ @@ -115,15 +136,32 @@ class RecurrentOp final : public framework::OperatorBase { alg_.Run(scope, dev_ctx); } + void set_stepnet(std::unique_ptr net) { + stepnet_ = std::move(net); + } + const OperatorBase& stepnet() const { return *stepnet_; } + static const rnn::ArgumentName kArgName; private: RecurrentAlgorithm alg_; + rnn::Argument arg_; + std::unique_ptr stepnet_; }; -class RecurrentGradientOp final : public framework::OperatorBase { +class RecurrentGradientOp : public framework::OperatorBase { public: - void Init() override; + RecurrentGradientOp(const std::string& type, + const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs); + + RecurrentGradientOp(const RecurrentGradientOp& o) + : framework::OperatorBase( + static_cast(o)) { + // TODO(yuyang18): Implement Copy ctor. + PADDLE_THROW("Not Implemented"); + } /** * InferShape must be called before Run. @@ -139,8 +177,15 @@ class RecurrentGradientOp final : public framework::OperatorBase { static const rnn::ArgumentName kArgName; + void set_stepnet(std::unique_ptr net) { + stepnet_ = std::move(net); + } + const OperatorBase& stepnet() const { return *stepnet_; } + private: RecurrentGradientAlgorithm alg_; + std::unique_ptr stepnet_; + rnn::Argument arg_; }; } // namespace operators diff --git a/paddle/operators/recurrent_op_test.cc b/paddle/operators/recurrent_op_test.cc deleted file mode 100644 index 6ce28a2b52..0000000000 --- a/paddle/operators/recurrent_op_test.cc +++ /dev/null @@ -1,398 +0,0 @@ -/* - Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. -*/ - -#include "paddle/operators/recurrent_op.h" - -#include -#include - -#include "paddle/framework/ddim.h" -#include "paddle/framework/op_registry.h" -#include "paddle/framework/operator.h" -#include "paddle/framework/tensor.h" -#include "paddle/operators/net_op.h" - -namespace paddle { -namespace operators { - -using framework::make_ddim; -using framework::DDim; -using framework::Tensor; -using framework::Variable; -using framework::Scope; -using framework::OpRegistry; - -class RecurrentOpTest : public ::testing::Test { - protected: - virtual void SetUp() override { - CreateGlobalVariables(); - CreateStepNet(); - CreateRNNOp(); - } - - virtual void TearDown() override {} - - void CreateGlobalVariables() { - // create input, and init content - LOG(INFO) << "create global variable x"; - for (auto inlink : std::vector{"x", "x0", "x1", "h"}) { - Variable* x = scope_.NewVar(inlink); - DDim dims = make_ddim(std::vector{ - 10 /*sent size*/, 20 /*batch size*/, 30 /*input dim*/}); - x->GetMutable()->mutable_data(dims, platform::CPUPlace()); - } - // create output alias just for test - for (auto inlink : std::vector{"h@alias"}) { - Variable* x = scope_.NewVar(inlink); - DDim dims = - make_ddim(std::vector{20 /*batch size*/, 30 /*input dim*/}); - x->GetMutable()->mutable_data(dims, platform::CPUPlace()); - } - - LOG(INFO) << "create global variable w"; - Variable* w = scope_.NewVar("rnn/w"); - w->GetMutable()->mutable_data( - make_ddim(std::vector{30, 30}), platform::CPUPlace()); - - for (auto boot : std::vector{"h_boot"}) { - LOG(INFO) << "create global variable " << boot; - Variable* h_boot = scope_.NewVar(boot); - h_boot->GetMutable()->mutable_data( - make_ddim(std::vector{20 /*batch size*/, 30 /*input dim*/}), - platform::CPUPlace()); - } - - LOG(INFO) << "create variable step_scopes"; - scope_.NewVar("step_scopes"); - - LOG(INFO) << "create variable h"; - scope_.NewVar("h"); - } - - void CreateRNNOp() { - framework::OpDesc op_desc; - - op_desc.set_type("recurrent_op"); - // inlinks 0 - op_desc.add_inputs("x"); - op_desc.add_inputs("x0"); - op_desc.add_inputs("x1"); - // boot_memories 3 - op_desc.add_inputs("h_boot"); - // step net 5 - op_desc.add_inputs("step_net"); - // outlinks 6 - op_desc.add_outputs("h"); - // step scopes 7 - op_desc.add_outputs("step_scopes"); - - auto _input_format = std::vector{ - 0, // in_link - 3, // memories - 4 // step_net - }; - auto input_format = op_desc.add_attrs(); - input_format->set_name("input_format"); - input_format->set_type(paddle::framework::AttrType::INTS); - for (auto i : _input_format) { - input_format->add_ints(i); - } - - auto output_format = op_desc.add_attrs(); - output_format->set_name("output_format"); - output_format->set_type(paddle::framework::AttrType::INTS); - for (auto i : std::vector{0, 1, 2}) { - output_format->add_ints(i); - } - - auto inlink_alias = op_desc.add_attrs(); - inlink_alias->set_name("inlink_alias"); - inlink_alias->set_type(paddle::framework::AttrType::STRINGS); - - auto outlink_alias = op_desc.add_attrs(); - outlink_alias->set_name("outlink_alias"); - outlink_alias->set_type(paddle::framework::AttrType::STRINGS); - - auto pre_memories = op_desc.add_attrs(); - pre_memories->set_name("pre_memories"); - pre_memories->set_type(paddle::framework::AttrType::STRINGS); - - auto memories = op_desc.add_attrs(); - memories->set_name("memories"); - memories->set_type(paddle::framework::AttrType::STRINGS); - - // create inlink_alias - for (const auto& item : - std::vector{"x@alias", "x0@alias", "x1@alias"}) { - inlink_alias->add_strings(item); - } - // pre memories - for (const auto& item : std::vector{"rnn/h@pre"}) { - pre_memories->add_strings(item); - } - // memories - for (const auto& item : std::vector{"rnn/h"}) { - memories->add_strings(item); - } - // output alias - for (const auto& item : std::vector{"h@alias"}) { - outlink_alias->add_strings(item); - } - - rnn_op_ = OpRegistry::CreateOp(op_desc); - - LOG(INFO) << "rnn_op finish init"; - } - - void CreateStepNet() { - LOG(INFO) << "create variable step_net"; - Variable* var = scope_.NewVar("step_net"); - auto net = var->GetMutable(); - net->AddOp( - OpRegistry::CreateOp("mul", {"rnn/h@pre", "rnn/w"}, {"rnn/s"}, {})); - - net->AddOp( - OpRegistry::CreateOp("add_two", {"x@alias", "rnn/s"}, {"rnn/h"}, {})); - net->CompleteAddOp(); - } - - // father scope - Scope scope_; - std::shared_ptr rnn_op_; -}; - -TEST_F(RecurrentOpTest, Run) { - platform::CPUDeviceContext ctx; - rnn_op_->InferShape(scope_); - rnn_op_->Run(scope_, ctx); -} - -class RecurrentGradientAlgorithmTest : public ::testing::Test { - protected: - virtual void SetUp() override { - CreateGlobalVariables(); - CreateStepScopes(); - CreateStepNet(); - CreateRNNGradientAlgorithm(); - - // segment inputs - SegmentInputs(); - // link forward memories - LinkeMemories(); - } - - virtual void TearDown() override {} - - void CreateGlobalVariables() { - // inputs: x - LOG(INFO) << "create global variable x"; - Variable* x = scope_.NewVar("x"); - DDim dims = - make_ddim({10 /*sent size*/, 20 /*batch size*/, 30 /*input dim*/}); - x->GetMutable()->mutable_data(dims, platform::CPUPlace()); - // inputs: h_boot - LOG(INFO) << "create global variable h_boot"; - Variable* h_boot = scope_.NewVar("h_boot"); - h_boot->GetMutable()->mutable_data( - make_ddim({20 /*batch size*/, 30 /*input dim*/}), platform::CPUPlace()); - // inputs: w - LOG(INFO) << "create global variable w"; - Variable* w = scope_.NewVar("rnn/w"); - w->GetMutable()->mutable_data(make_ddim({30, 30}), - platform::CPUPlace()); - // inputs: h_grad - LOG(INFO) << "create variable h_grad"; - Variable* dh = scope_.NewVar("h_grad"); - dh->GetMutable()->mutable_data(make_ddim({10, 20, 30}), - platform::CPUPlace()); - // inputs: step_scopes - LOG(INFO) << "create variable step_scopes"; - scope_.NewVar("step_scopes"); - // inputs: step_net - LOG(INFO) << "create variable step_net"; - scope_.NewVar("step_net"); - // outputs: w_grad - LOG(INFO) << "create global variable w_grad"; - scope_.NewVar("rnn/w_grad"); - // outputs: x_grad - LOG(INFO) << "create global variable x_grad"; - scope_.NewVar("x_grad"); - // outputs: h_boot_grad - LOG(INFO) << "create global variable h_boot_grad"; - scope_.NewVar("h_boot_grad"); - } - - void CreateStepScopes() { - auto step_scopes = - scope_.FindVar("step_scopes")->GetMutable>(); - for (int i = 0; i < 10; ++i) { - auto& scope = scope_.NewScope(); - auto pre_t = scope.NewVar("rnn/pre_h")->GetMutable(); - pre_t->mutable_data({20, 30}, platform::CPUPlace()); - auto tensor = scope.NewVar("rnn/h")->GetMutable(); - tensor->mutable_data({20, 30}, platform::CPUPlace()); - - // for unit test of ConcatOutputs - auto xg = scope.NewVar("rnn/x_grad")->GetMutable(); - xg->mutable_data({20, 30}, platform::CPUPlace()); - - step_scopes->emplace_back(&scope); - } - - // last time step - auto g = (*step_scopes)[9]->NewVar("rnn/h_pre_grad")->GetMutable(); - g->mutable_data({20, 30}, platform::CPUPlace()); - } - - void CreateRNNGradientAlgorithm() { - std::unique_ptr arg(new rnn::Argument()); - arg->step_net = "step_net"; - arg->step_scopes = "step_scopes"; - rnn::Link inlink; - inlink.external = "h_grad"; - inlink.internal = "rnn/h_grad"; - arg->inlinks = std::vector{inlink}; - - rnn::Link outlink; - outlink.external = "x_grad"; - outlink.internal = "rnn/x_grad"; - arg->outlinks = std::vector{outlink}; - - rnn::MemoryAttr mem_attr; - mem_attr.pre_var = "rnn/h_pre_grad"; - mem_attr.var = "rnn/h_grad"; - mem_attr.boot_var = "h_boot_grad"; - arg->memories = std::vector{mem_attr}; - - rnn_grad_algo_.Init(std::move(arg)); - } - - void CreateStepNet() { - LOG(INFO) << "create variable step_net"; - Variable* var = scope_.NewVar("step_net"); - auto net = var->GetMutable(); - net->AddOp(OpRegistry::CreateOp("mul", {"rnn/h_pre", "rnn/w", "rnn/s_grad"}, - {"rnn/h_pre_grad", "rnn/w_grad"}, {})); - - net->AddOp(OpRegistry::CreateOp("add_two", {"rnn/h_grad"}, - {"rnn/x_grad", "rnn/s_grad"}, {})); - net->CompleteAddOp(); - } - - void SegmentInputs() { - LOG(INFO) << "segment inputs"; - std::vector inlinks = {"x"}; - std::vector inlinks_alias = {"rnn/x"}; - - rnn::Link inlink; - inlink.external = "x"; - inlink.internal = "rnn/x"; - auto step_scopes = - scope_.FindVar("step_scopes")->GetMutable>(); - rnn::SegmentInputs(*step_scopes, std::vector{inlink}, 10, - true /*infer_shape_mode*/); - } - - void LinkeMemories() { - LOG(INFO) << "link memories"; - rnn::MemoryAttr mem_attr; - mem_attr.pre_var = "rnn/h_pre"; - mem_attr.var = "rnn/h"; - mem_attr.boot_var = "boot_h"; - std::vector memories; - memories.push_back(mem_attr); - auto step_scopes = - scope_.FindVar("step_scopes")->GetMutable>(); - for (int i = 1; i < 10; ++i) { - rnn::LinkMemories(*step_scopes, memories, i, -1, - true /*infer_shape_mode*/); - } - } - - Scope scope_; - RecurrentGradientAlgorithm rnn_grad_algo_; -}; - -// TEST_F(RecurrentGradientAlgorithmTest, Run) { -// platform::CPUDeviceContext ctx; -// rnn_grad_algo_.Run(scope_, ctx); -// } - -} // namespace operators -} // namespace paddle - -TEST(RecurrentOp, LinkMemories) { - using namespace paddle::framework; - using namespace paddle::platform; - using namespace paddle::operators; - - // create and init step scopes - size_t len = 10; - std::vector step_scopes; - for (size_t i = 0; i < len; ++i) { - auto scope = new Scope(); - scope->NewVar("pre_h"); - auto tensor = scope->NewVar("h")->GetMutable(); - float* data = tensor->mutable_data({15, 20}, CPUPlace()); - for (size_t j = 0; j < 15 * 20; ++j) { - data[j] = rand() * (1. / (double)RAND_MAX); - } - step_scopes.push_back(scope); - } - - // create MemoryAttr - rnn::MemoryAttr mem_attr; - mem_attr.pre_var = "pre_h"; - mem_attr.var = "h"; - mem_attr.boot_var = "boot_h"; - std::vector memories; - memories.push_back(mem_attr); - - for (size_t i = 1; i < len; ++i) { - rnn::LinkMemories(step_scopes, memories, i, -1, false /*infer_shape_mode*/); - } - // check - for (size_t i = 0; i < len - 1; ++i) { - const float* a = - step_scopes[i]->FindVar("h")->GetMutable()->data(); - const float* b = step_scopes[i + 1] - ->FindVar("pre_h") - ->GetMutable() - ->data(); - for (size_t j = 0; j < 15 * 20; ++j) { - ASSERT_FLOAT_EQ(a[j], b[j]); - } - } - - for (int i = len - 2; i >= 0; --i) { - rnn::LinkMemories(step_scopes, memories, i, 1, false /*infer_shape_mode*/); - } - // check - for (int i = len - 2; i >= 0; --i) { - const float* a = - step_scopes[i]->FindVar("pre_h")->GetMutable()->data(); - const float* b = - step_scopes[i + 1]->FindVar("h")->GetMutable()->data(); - for (size_t j = 0; j < 15 * 20; ++j) { - ASSERT_FLOAT_EQ(a[j], b[j]); - } - } - - for (auto s : step_scopes) { - delete s; - } -} - -USE_OP(add_two); -USE_OP(mul); -USE_OP_WITHOUT_KERNEL(recurrent_op); diff --git a/paddle/operators/rnn/recurrent_op_utils.cc b/paddle/operators/rnn/recurrent_op_utils.cc index 7e4770630e..a9b65c30f2 100644 --- a/paddle/operators/rnn/recurrent_op_utils.cc +++ b/paddle/operators/rnn/recurrent_op_utils.cc @@ -106,7 +106,6 @@ void LinkMemories(const std::vector& scopes, void InitArgument(const ArgumentName& name, Argument* arg, const framework::OperatorBase& op) { - arg->step_net = op.Input(name.step_net); arg->step_scopes = op.Output(name.step_scopes); auto inlinks = op.Inputs(name.inlinks); diff --git a/paddle/operators/rowwise_add_op.cc b/paddle/operators/rowwise_add_op.cc index 55ed1c2f4c..6825dce332 100644 --- a/paddle/operators/rowwise_add_op.cc +++ b/paddle/operators/rowwise_add_op.cc @@ -17,25 +17,28 @@ namespace paddle { namespace operators { -class RowWiseAddOp : public framework::OperatorWithKernel { +using framework::Tensor; + +class RowwiseAddOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 2UL, - "Two inputs is needed by rowwise add"); - auto dim0 = ctx.Input(0)->dims(); - auto dim1 = ctx.Input(1)->dims(); + 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 framework::OpProtoAndCheckerMaker { +class RowwiseAddOpMaker : public framework::OpProtoAndCheckerMaker { public: - RowWiseAddOpMaker(framework::OpProto *proto, + RowwiseAddOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "The left input of row-wise add op, must be matrix"); @@ -48,11 +51,32 @@ for i in xrange(X.shape[0]): )DOC"); } }; +class RowwiseAddGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "X should not be null"); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("b"), "b should not be null"); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto dims0 = ctx.Input("X")->dims(); + auto dims1 = ctx.Input("b")->dims(); + PADDLE_ENFORCE_EQ(1, dims1.size(), "b dims should be 1") + ctx.Output(framework::GradVarName("X"))->Resize(dims0); + ctx.Output(framework::GradVarName("b"))->Resize(dims1); + } +}; } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(rowwise_add, ops::RowWiseAddOp, ops::RowWiseAddOpMaker); +REGISTER_OP(rowwise_add, ops::RowwiseAddOp, ops::RowwiseAddOpMaker, + rowwise_add_grad, ops::RowwiseAddGradOp); +REGISTER_OP_CPU_KERNEL( + rowwise_add, ops::RowwiseAddKernel); REGISTER_OP_CPU_KERNEL( - rowwise_add, ops::RowWiseAddKernel); + rowwise_add_grad, + ops::RowwiseAddGradKernel); diff --git a/paddle/operators/rowwise_add_op.cu b/paddle/operators/rowwise_add_op.cu index 86f80b8122..4a57f64c89 100644 --- a/paddle/operators/rowwise_add_op.cu +++ b/paddle/operators/rowwise_add_op.cu @@ -17,4 +17,7 @@ namespace ops = paddle::operators; REGISTER_OP_GPU_KERNEL( - rowwise_add, ops::RowWiseAddKernel); + rowwise_add, ops::RowwiseAddKernel); +REGISTER_OP_GPU_KERNEL( + rowwise_add_grad, + ops::RowwiseAddGradKernel); diff --git a/paddle/operators/rowwise_add_op.h b/paddle/operators/rowwise_add_op.h index 82e9d70e95..1cbd8bb31a 100644 --- a/paddle/operators/rowwise_add_op.h +++ b/paddle/operators/rowwise_add_op.h @@ -1,16 +1,16 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 + http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ #pragma once #include "paddle/framework/eigen.h" @@ -28,14 +28,14 @@ template ; template -class RowWiseAddKernel : public framework::OpKernel { +class RowwiseAddKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto out = context.Output(0); + auto out = context.Output("Out"); out->mutable_data(context.GetPlace()); - auto input = EigenMatrix::From(*context.Input(0)); - auto bias = EigenVector::From(*context.Input(1)); + auto 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); @@ -47,5 +47,25 @@ class RowWiseAddKernel : public framework::OpKernel { } }; +template +class RowwiseAddGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* dOut = context.Input(framework::GradVarName("Out")); + auto* dX = context.Output(framework::GradVarName("X")); + auto* db = context.Output(framework::GradVarName("b")); + dX->mutable_data(context.GetPlace()); + db->mutable_data(context.GetPlace()); + + auto OutGrad = EigenMatrix::From(*dOut); + auto place = context.GetEigenDevice(); + EigenMatrix::From(*dX).device(place) = OutGrad; + + // https://eigen.tuxfamily.org/dox/unsupported/TensorBase_8h_source.html + // colwise add + Eigen::array dims{{0}}; /* dimension to reduce */ + EigenVector::Flatten(*db).device(place) = OutGrad.sum(dims); + } +}; } // namespace operators } // namespace paddle diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc new file mode 100644 index 0000000000..8e96a74c94 --- /dev/null +++ b/paddle/operators/scale_op.cc @@ -0,0 +1,105 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/scale_op.h" +#include "paddle/operators/net_op.h" + +namespace paddle { +namespace operators { + +class ScaleOp : public framework::OperatorWithKernel { + public: + ScaleOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + auto *in = ctx.Input("X"); + auto *out = ctx.Output("Out"); + out->Resize(in->dims()); + } +}; + +template +class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The input tensor of scale operator.").NotInGradient(); + AddOutput("Out", "The output tensor of scale operator.").NotInGradient(); + AddComment(R"DOC(Scale operator + +The equation is: Out = scale*X +)DOC"); + AddAttr("scale", "scale of scale operator.").SetDefault(1.0); + } +}; + +// Identity Op's gradient is identity op, too. +// Grad(Out=scale(X)) => Grad(X) = scale(Grad(Out)) +template +class ScaleGradOp : public NetOp { + public: + ScaleGradOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : NetOp(type, inputs, outputs, attrs) { + AppendOp(framework::OpRegistry::CreateOp( + "scale", {{"X", {Input(framework::GradVarName("Out"))}}}, + {{"Out", {Output(framework::GradVarName("X"))}}}, + {{"scale", GetAttr("scale")}})); + CompleteAddOp(false); + } +}; + +// identity is a alias of scale op. This is also a example for creating a alias +// operator. +template +class IdentityOpMaker : public framework::OpProtoAndCheckerMaker { + public: + IdentityOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "input tensor of identity op"); + AddOutput("Out", "output tensor of identity op"); + AddComment("identity operator. Just a alias of scale op which scale = 1.0"); + } +}; + +template +class IdentityOp : public NetOp { + public: + IdentityOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : NetOp(type, inputs, outputs, attrs) { + AppendOp(framework::OpRegistry::CreateOp( + "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}}, + {{"scale", static_cast(1)}})); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP(scale, ops::ScaleOp, ops::ScaleOpMaker, scale_grad, + ops::ScaleGradOp); +REGISTER_OP_CPU_KERNEL(scale, + ops::ScaleKernel); +REGISTER_OP_WITHOUT_GRADIENT(identity, ops::IdentityOp, + ops::IdentityOpMaker); diff --git a/paddle/operators/scale_op.cu b/paddle/operators/scale_op.cu new file mode 100644 index 0000000000..63efbe0da8 --- /dev/null +++ b/paddle/operators/scale_op.cu @@ -0,0 +1,18 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/operators/scale_op.h" + +REGISTER_OP_GPU_KERNEL( + scale, paddle::operators::ScaleKernel); diff --git a/paddle/operators/scale_op.h b/paddle/operators/scale_op.h new file mode 100644 index 0000000000..aea64f1b04 --- /dev/null +++ b/paddle/operators/scale_op.h @@ -0,0 +1,40 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { +template +class ScaleKernel : public framework::OpKernel { + public: + virtual void Compute(const framework::ExecutionContext& context) const { + auto* tensor = context.Output("Out"); + auto* in = context.Input("X"); + tensor->mutable_data(in->place()); + + auto scale = static_cast(context.op_.GetAttr("scale")); + + auto eigen_out = framework::EigenVector::Flatten(*tensor); + auto eigen_in = framework::EigenVector::Flatten(*in); + auto& dev = context.GetEigenDevice(); + eigen_out.device(dev) = scale * eigen_in; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/scatter.h b/paddle/operators/scatter.h new file mode 100644 index 0000000000..6b542675c2 --- /dev/null +++ b/paddle/operators/scatter.h @@ -0,0 +1,92 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include + +#include "paddle/framework/ddim.h" +#include "paddle/framework/eigen.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenVector = framework::EigenVector; + +// Implementation of CPU copy +template +void CPUScatterUpdate(const paddle::framework::Tensor* src, const int* index, + const size_t index_size, + paddle::framework::Tensor* output) { + paddle::framework::DDim output_dims = output->dims(); + + for (size_t i = 0; i < index_size; ++i) { + int index_ = index[i]; + + paddle::framework::Tensor src_ = *src; + paddle::framework::Tensor output_ = *output; + if (index_size > 1) src_ = src->Slice(i, i + 1); + if (output_dims[0] > 1) output_ = output->Slice(index_, index_ + 1); + + auto X = EigenVector::Flatten(src_); + auto Y = EigenVector::Flatten(output_); + + Y = X + Y; + } +} + +// Implementation of GPU scatter: +template +void GPUScatterUpdate(const T* src, const int* index, const int slice_size, + const int index_size, T* output); + +/** + * Return a updated tensor from source tensor, scattered according to index: + * dst[i] += src[index[i]] + * input[src]: type-T source Tensor + * input[index]: type-int index Tensor (1-D) + * return: output tensor + */ +template +void ScatterUpdate(const platform::Place& place, + const paddle::framework::Tensor* src, + const paddle::framework::Tensor* index, + paddle::framework::Tensor* output) { + // check index of shape 1-D + PADDLE_ENFORCE(index->dims().size() == 1); + int index_size = index->dims()[0]; + + auto src_dims = src->dims(); + auto dst_dims = output->dims(); + + // check src shape and dst shape should match + for (int i = 1; i < src_dims.size(); i++) + PADDLE_ENFORCE(src_dims[i] == dst_dims[i]); + + // slice size + size_t slice_size = 1; + for (int i = 0; i < src_dims.size(); ++i) slice_size *= src_dims[i]; + + if (platform::is_cpu_place(place)) { + CPUScatterUpdate(src, index->data(), index_size, output); + } else { + } +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/scatter_op.cc b/paddle/operators/scatter_op.cc new file mode 100644 index 0000000000..f901edefa2 --- /dev/null +++ b/paddle/operators/scatter_op.cc @@ -0,0 +1,86 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/scatter_op.h" +#include "paddle/framework/ddim.h" + +namespace paddle { +namespace operators { + +class ScatterOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + PADDLE_ENFORCE_EQ(ctx.Input("Index")->dims().size(), 1, + "Update Index should be 1-D."); + PADDLE_ENFORCE_EQ(ctx.Input("Ref")->dims().size(), + ctx.Input("Updates")->dims().size(), + "Reference and Updates should have the same shape size"); + PADDLE_ENFORCE_EQ(ctx.Input("Updates")->dims()[0], + ctx.Input("Index")->dims()[0], + "Updates and Index should have same batch-size."); + framework::DDim data_dim(ctx.Input("Updates")->dims()); + for (int i = 1; i < data_dim.size(); ++i) + PADDLE_ENFORCE_EQ(data_dim[i], ctx.Input("Updates")->dims()[i]); + ctx.Output("Out")->Resize(ctx.Input("Ref")->dims()); + } +}; + +class ScatterGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + auto *dUpdates = ctx.Output(framework::GradVarName("Updates")); + auto *Updates = ctx.Input("Updates"); + auto *dRef = ctx.Output(framework::GradVarName("Ref")); + auto *Ref = ctx.Input("Ref"); + + dRef->Resize(Ref->dims()); + dUpdates->Resize(Updates->dims()); + } +}; + +class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { + public: + ScatterOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Ref", "The source input of scatter op"); + AddInput("Index", + "The index input of scatter op where Ref will be updated"); + AddInput("Updates", "The updated value of updates op"); + AddOutput("Out", "The output of add op"); + AddComment(R"DOC( +Scatter Operator by selecting from the first axis, + +Out = Ref +Out[Index] = Ref[Index] + Updates +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(scatter, ops::ScatterOp, ops::ScatterOpMaker, scatter_grad, + ops::ScatterGradOp); +REGISTER_OP_CPU_KERNEL(scatter, + ops::ScatterOpKernel); +REGISTER_OP_CPU_KERNEL( + scatter_grad, + ops::ScatterGradientOpKernel); diff --git a/paddle/operators/sgd_op_test.cc b/paddle/operators/scatter_op.cu similarity index 72% rename from paddle/operators/sgd_op_test.cc rename to paddle/operators/scatter_op.cu index 75137259f5..6716b47883 100644 --- a/paddle/operators/sgd_op_test.cc +++ b/paddle/operators/scatter_op.cu @@ -12,11 +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 -USE_OP(sgd); -TEST(SGDOp, GetOpProto) { - auto& protos = paddle::framework::OpRegistry::protos(); - auto it = protos.find("sgd"); - ASSERT_NE(it, protos.end()); -} +#define EIGEN_USE_GPU +#include "paddle/operators/scatter_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(scatter, + ops::ScatterOpKernel); diff --git a/paddle/operators/scatter_op.h b/paddle/operators/scatter_op.h new file mode 100644 index 0000000000..e9595638a8 --- /dev/null +++ b/paddle/operators/scatter_op.h @@ -0,0 +1,60 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "gather.h" +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" +#include "scatter.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class ScatterOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto *Ref = ctx.Input("Ref"); + auto *Index = ctx.Input("Index"); + auto *Updates = ctx.Input("Updates"); + auto *Out = ctx.Output("Out"); + + // In place output: Out = Ref, Out[Index] += Updates + Out->ShareDataWith(*Ref); + // Apply ScatterUpdate: Out[index] += Updates[:] + ScatterUpdate(ctx.GetPlace(), Updates, Index, Out); + } +}; + +template +class ScatterGradientOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto *dRef = ctx.Output(framework::GradVarName("Ref")); + auto *dUpdates = ctx.Output(framework::GradVarName("Updates")); + auto *Index = ctx.Input("Index"); + auto *dOut = ctx.Input(framework::GradVarName("Out")); + + // In place gradient: dRef = dO + dRef->ShareDataWith(*dOut); + dUpdates->mutable_data(ctx.GetPlace()); + // Gradient by Gather: dUpdates += dO[Index] + Gather(ctx.GetPlace(), dOut, Index, dUpdates); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/scatter_test.cc b/paddle/operators/scatter_test.cc new file mode 100644 index 0000000000..26fdaff146 --- /dev/null +++ b/paddle/operators/scatter_test.cc @@ -0,0 +1,56 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/scatter.h" +#include "paddle/framework/ddim.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +#include +#include +#include + +TEST(scatter, ScatterUpdate) { + using namespace paddle::framework; + using namespace paddle::platform; + using namespace paddle::operators; + + Tensor* src = new Tensor(); + Tensor* index = new Tensor(); + Tensor* output = new Tensor(); + + float* p_src = nullptr; + int* p_index = nullptr; + p_src = src->mutable_data(make_ddim({1, 4}), CPUPlace()); + p_index = index->mutable_data(make_ddim({1}), CPUPlace()); + + for (size_t i = 0; i < 4; ++i) p_src[i] = float(i); + p_index[0] = 1; + + float* p_output = output->mutable_data(make_ddim({4, 4}), CPUPlace()); + + ScatterUpdate(CPUPlace(), src, index, output); + + for (size_t i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], float(0)); + for (size_t i = 0; i < 4; ++i) EXPECT_EQ(output->data()[i], float(0)); + for (size_t i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], float(i - 4)); + for (size_t i = 4; i < 8; ++i) + EXPECT_EQ(output->data()[i], float(i - 4)); + for (size_t i = 8; i < 16; ++i) EXPECT_EQ(p_output[i], float(0)); + for (size_t i = 8; i < 16; ++i) EXPECT_EQ(output->data()[i], float(0)); + + delete src; + delete index; + delete output; +} diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc index f9a28ff8a6..ad267e7f08 100644 --- a/paddle/operators/sgd_op.cc +++ b/paddle/operators/sgd_op.cc @@ -18,16 +18,15 @@ namespace paddle { namespace operators { class SGDOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 2, "Input size of SGDOp must be two"); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "Output size of SGDOp must be one"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "inputs[0] mast be set"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(1), "inputs[1] mast be set"); - PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0), "outputs[0] mast be set"); - PADDLE_ENFORCE(ctx.Input(0)->dims() == ctx.Input(1)->dims(), - "Two input of SGD Op's dimension must be same."); - ctx.Output(0)->Resize(ctx.Input(0)->dims()); + PADDLE_ENFORCE( + ctx.Input("param")->dims() == ctx.Input("grad")->dims(), + "Two input of SGD Op's dimension must be same."); + ctx.Output("param_out")->Resize(ctx.Input("param")->dims()); } }; @@ -52,6 +51,6 @@ param_out = param - learning_rate * grad; } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(sgd, ops::SGDOp, ops::SGDOpMaker); +REGISTER_OP_WITHOUT_GRADIENT(sgd, ops::SGDOp, ops::SGDOpMaker); REGISTER_OP_CPU_KERNEL(sgd, ops::SGDOpKernel); diff --git a/paddle/operators/sgd_op.h b/paddle/operators/sgd_op.h index bfb449d0b0..a0b5000ffb 100644 --- a/paddle/operators/sgd_op.h +++ b/paddle/operators/sgd_op.h @@ -30,7 +30,7 @@ class SGDOpKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { auto param = ctx.Input("param"); auto grad = ctx.Input("grad"); - auto param_out = ctx.Output(0); + auto param_out = ctx.Output("param_out"); float lr = ctx.op_.GetAttr("learning_rate"); param_out->mutable_data(ctx.GetPlace()); diff --git a/paddle/operators/sigmoid_op.cc b/paddle/operators/sigmoid_op.cc index bc5e0bbb18..761c6de8d4 100644 --- a/paddle/operators/sigmoid_op.cc +++ b/paddle/operators/sigmoid_op.cc @@ -18,11 +18,12 @@ namespace paddle { namespace operators { class SigmoidOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputSize() == 1, "Sigmoid Op only have one input"); - PADDLE_ENFORCE(ctx.OutputSize() == 1, "Sigmoid Op only have one output"); - ctx.Output(0)->Resize(ctx.Input(0)->dims()); + ctx.Output("Y")->Resize(ctx.Input("X")->dims()); } }; @@ -38,9 +39,13 @@ class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker { }; 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()); + ctx.Output(framework::GradVarName("X")) + ->Resize(ctx.Input("Y")->dims()); } }; @@ -48,9 +53,8 @@ class SigmoidOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker); -REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, ops::SigmoidOpGrad); - +REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker, sigmoid_grad, + ops::SigmoidOpGrad); REGISTER_OP_CPU_KERNEL(sigmoid, ops::SigmoidKernel); REGISTER_OP_CPU_KERNEL( diff --git a/paddle/operators/sigmoid_op.h b/paddle/operators/sigmoid_op.h index 7af879b209..b01a9b3f23 100644 --- a/paddle/operators/sigmoid_op.h +++ b/paddle/operators/sigmoid_op.h @@ -28,8 +28,8 @@ template class SigmoidKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto input = context.Input(0); - auto output = context.Output(0); + auto input = context.Input("X"); + auto output = context.Output("Y"); output->mutable_data(context.GetPlace()); // The clipping is used in Paddle's raw implenmention @@ -37,7 +37,7 @@ class SigmoidKernel : public framework::OpKernel { auto Y = EigenVector::Flatten(*output); auto place = context.GetEigenDevice(); - Y.device(place) = 1.0 / (1.0 + (-1.0 * X).exp()); + Y.device(place) = 1. / (1. + (-X).exp()); } }; diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc index 3dd4e86918..40c51a64c4 100644 --- a/paddle/operators/softmax_op.cc +++ b/paddle/operators/softmax_op.cc @@ -18,14 +18,13 @@ namespace paddle { namespace operators { class SoftmaxOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 1UL, - "Only one input is need for softmax"); - PADDLE_ENFORCE_EQ(ctx.Input("X")->dims().size(), 2UL, - "The input of softmax op must be matrix"); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL, - "Only one output is need for softmax"); + PADDLE_ENFORCE(ctx.Input("X")->dims().size() == 2UL, + "The input of softmax op must be matrix"); ctx.Output("Y")->Resize(ctx.Input("X")->dims()); } }; @@ -42,13 +41,12 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { }; class SoftmaxOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE_EQ(ctx.InputSize(), 3UL, - "Input of SoftmaxOpGrad should be 3, X, Y, YG"); - PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL, - "Output of SoftmaxOpGrad should be 1"); - PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null"); + PADDLE_ENFORCE(ctx.InputVar("Y") != nullptr, "Input(Y) 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() == @@ -64,9 +62,9 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel { namespace ops = paddle::operators; -REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker); +REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker, softmax_grad, + ops::SoftmaxOpGrad); REGISTER_OP_CPU_KERNEL(softmax, ops::SoftmaxKernel); -REGISTER_GRADIENT_OP(softmax, softmax_grad, ops::SoftmaxOpGrad); REGISTER_OP_CPU_KERNEL( softmax_grad, ops::SoftmaxGradKernel); diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc index 405b84b76d..29491137e6 100644 --- a/paddle/operators/uniform_random_op.cc +++ b/paddle/operators/uniform_random_op.cc @@ -1,11 +1,8 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. @@ -27,7 +24,7 @@ template class CPUUniformRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* tensor = context.Output(0); + auto* tensor = context.Output("Out"); T* data = tensor->mutable_data(context.GetPlace()); unsigned int seed = static_cast(context.op_.GetAttr("seed")); @@ -39,18 +36,22 @@ class CPUUniformRandomKernel : public framework::OpKernel { 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) { + ssize_t size = framework::product(tensor->dims()); + for (ssize_t i = 0; i < size; ++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(0); + auto* tensor = ctx.Output("Out"); auto dims = GetAttr>("dims"); tensor->Resize(framework::make_ddim(dims)); } @@ -63,7 +64,6 @@ class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker { : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddOutput("Out", "The output tensor of uniform random op"); AddComment(R"DOC(Uniform random operator. - Used to initialize tensor with uniform random generator. )DOC"); AddAttr>("dims", "the dimension of random tensor"); @@ -78,7 +78,7 @@ Used to initialize tensor with uniform random generator. } // namespace operators } // namespace paddle -REGISTER_OP(uniform_random, paddle::operators::UniformRandomOp, - paddle::operators::UniformRandomOpMaker); +REGISTER_OP_WITHOUT_GRADIENT(uniform_random, paddle::operators::UniformRandomOp, + paddle::operators::UniformRandomOpMaker); REGISTER_OP_CPU_KERNEL(uniform_random, paddle::operators::CPUUniformRandomKernel); diff --git a/paddle/operators/uniform_random_op.cu b/paddle/operators/uniform_random_op.cu index f1a63e52ec..1d6709934c 100644 --- a/paddle/operators/uniform_random_op.cu +++ b/paddle/operators/uniform_random_op.cu @@ -1,11 +1,8 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. @@ -46,12 +43,13 @@ template class GPUUniformRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* tensor = context.Output(0); + auto* tensor = context.Output("Out"); T* data = tensor->mutable_data(context.GetPlace()); unsigned int seed = static_cast(context.op_.GetAttr("seed")); if (seed == 0) { - seed = std::random_device()(); + std::random_device rd; + seed = rd(); } T min = static_cast(context.op_.GetAttr("min")); T max = static_cast(context.op_.GetAttr("max")); diff --git a/paddle/parameter/Argument.cpp b/paddle/parameter/Argument.cpp index 77fd0c5890..94c1bda5bb 100644 --- a/paddle/parameter/Argument.cpp +++ b/paddle/parameter/Argument.cpp @@ -278,17 +278,21 @@ int32_t Argument::resizeAndCopyFrom(const Argument& src, void Argument::concat(const std::vector& args, const std::vector& selectRows, const std::vector& seqStartPos, + const std::vector& copySize, bool useGpu, hl_stream_t stream, PassType passType) { CHECK(!subSequenceStartPositions) << "undefined behavior for subsequence positions"; - size_t batchSize = selectRows.size(); + size_t batchSize = 0; + for (size_t i = 0; i < copySize.size(); ++i) + batchSize += copySize[i] * (seqStartPos[i + 1] - seqStartPos[i]); + auto copyArg = [batchSize, stream](MatrixPtr& dst, MatrixPtr src, - int startRow, - int pos, + int desStartRow, + int srcStartRow, int size, bool useGpu) { if (!src) { @@ -302,14 +306,14 @@ void Argument::concat(const std::vector& args, dst->resize(batchSize, width); } - MatrixPtr tmpMatrix = dst->subMatrix(startRow, size); - tmpMatrix->copyFrom(*src->subMatrix(pos, size), stream); + MatrixPtr tmpMatrix = dst->subMatrix(desStartRow, size); + tmpMatrix->copyFrom(*src->subMatrix(srcStartRow, size), stream); }; auto copyIds = [batchSize, stream](IVectorPtr& dst, const IVectorPtr& src, - int startRow, - int pos, + int desStartRow, + int srcStartRow, int size, bool useGpu) { if (!src) { @@ -317,13 +321,14 @@ void Argument::concat(const std::vector& args, return; } IVector::resizeOrCreate(dst, batchSize, useGpu); - dst->subVec(startRow, size)->copyFrom(*src->subVec(pos, size), stream); + dst->subVec(desStartRow, size) + ->copyFrom(*src->subVec(srcStartRow, size), stream); }; auto copyStrs = [batchSize, stream](SVectorPtr& dst, const SVectorPtr& src, - int startRow, - int pos, + int desStartRow, + int srcStartRow, int size, bool useGpu) { if (!src) { @@ -335,30 +340,31 @@ void Argument::concat(const std::vector& args, } else { dst->resize(batchSize); } - std::copy( - src->begin() + pos, src->begin() + pos + size, dst->begin() + startRow); + std::copy(src->begin() + srcStartRow, + src->begin() + srcStartRow + size, + dst->begin() + desStartRow); }; dataId = args[0].dataId; CHECK_NE(seqStartPos.size(), 0UL); - size_t sampleNum = seqStartPos.size() - 1; - for (size_t i = 0; i < sampleNum; ++i) { + int desStartRow = 0; + for (size_t i = 0; i < copySize.size(); ++i) { int startPos = seqStartPos[i]; int endPos = seqStartPos[i + 1]; CHECK_GE(args.size(), static_cast(endPos - startPos)); for (int j = startPos; j < endPos; ++j) { const Argument& arg = args[j - startPos]; - CHECK_EQ(arg.dataId, dataId) << "Arguments in concat should have" - << " same dataId"; - const int copySize = 1; - const int rowIdx = selectRows[j]; - copyArg(in, arg.in, j, rowIdx, copySize, useGpu); - copyArg(value, arg.value, j, rowIdx, copySize, useGpu); + CHECK_EQ(arg.dataId, dataId) << "Arguments to concatenate should have " + << "the same dataId."; + const int srcStartRow = selectRows[j]; + copyArg(in, arg.in, desStartRow, srcStartRow, copySize[i], useGpu); + copyArg(value, arg.value, desStartRow, srcStartRow, copySize[i], useGpu); if (passType != PASS_TEST) { - copyArg(grad, arg.grad, j, rowIdx, copySize, useGpu); + copyArg(grad, arg.grad, desStartRow, srcStartRow, copySize[i], useGpu); } - copyIds(ids, arg.ids, j, rowIdx, copySize, useGpu); - copyStrs(strs, arg.strs, j, rowIdx, copySize, useGpu); + copyIds(ids, arg.ids, desStartRow, srcStartRow, copySize[i], useGpu); + copyStrs(strs, arg.strs, desStartRow, srcStartRow, copySize[i], useGpu); + desStartRow += copySize[i]; } } ICpuGpuVector::resizeOrCreate( @@ -672,19 +678,28 @@ void Argument::reorganizeSeqInfo( const ICpuGpuVectorPtr seqStartPos, const ICpuGpuVectorPtr subSeqStartPos, std::vector>& reorganizedSeqInfo) { - int* seqStarts = seqStartPos->getMutableData(false); - int* subSeqStarts = subSeqStartPos->getMutableData(false); + CHECK(seqStartPos); int seqNum = seqStartPos->getSize() - 1; - reorganizedSeqInfo.resize(seqNum, std::vector()); - int seqIdx = 0; - for (size_t i = 0; i < subSeqStartPos->getSize(); ++i) { - reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]); - if (subSeqStarts[i] == seqStarts[seqIdx + 1]) { - seqIdx++; - if (seqIdx == seqNum) return; + int* seqStarts = seqStartPos->getMutableData(false); + + if (subSeqStartPos) { + int* subSeqStarts = subSeqStartPos->getMutableData(false); + reorganizedSeqInfo.resize(seqNum, std::vector()); + int seqIdx = 0; + for (size_t i = 0; i < subSeqStartPos->getSize(); ++i) { reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]); + if (subSeqStarts[i] == seqStarts[seqIdx + 1]) { + seqIdx++; + if (seqIdx == seqNum) return; + reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]); + } } + } else { + reorganizedSeqInfo.resize(1, std::vector(seqNum + 1, 0)); + memcpy(reorganizedSeqInfo[0].data(), + seqStarts, + sizeof(int) * seqStartPos->getSize()); } } diff --git a/paddle/parameter/Argument.h b/paddle/parameter/Argument.h index ba3ad2fd4d..7b59199dde 100644 --- a/paddle/parameter/Argument.h +++ b/paddle/parameter/Argument.h @@ -242,6 +242,7 @@ struct Argument { void concat(const std::vector& args, const std::vector& selectRows, const std::vector& seqStartPos, + const std::vector& copySize, bool useGpu, hl_stream_t stream, PassType passType); diff --git a/paddle/parameter/Parameter.cpp b/paddle/parameter/Parameter.cpp index ebe36d4937..f031109501 100644 --- a/paddle/parameter/Parameter.cpp +++ b/paddle/parameter/Parameter.cpp @@ -48,7 +48,8 @@ Parameter::Parameter(const ParameterConfig& config, bool useGpu, bool doInit) deviceId_(-1), sharedCount_(0), updateCounter_(0), - updated_(false) { + updated_(false), + headerFormat_(PARAM_FORMAT_ORIGINAL) { setID(-1); /* capture uninitialized id */ if (useGpu_ && FLAGS_parallel_nn) { /* gpu environment is specified by device property */ @@ -285,7 +286,7 @@ bool Parameter::save(const std::string& filename) const { bool Parameter::save(std::ostream& s) const { CpuVector vec(*bufs_[PARAMETER_VALUE].get()); Header header; - header.version = kFormatVersion; + header.format = headerFormat_; header.valueSize = sizeof(real); header.size = getSize(); @@ -344,8 +345,9 @@ bool Parameter::load(std::istream& s) { Header header; CHECK(s.read(reinterpret_cast(&header), sizeof(header))) << "Fail to read parameter " << getName(); - CHECK_EQ(header.version, kFormatVersion) << "Incorrect format version: " - << header.version; + CHECK(isHeaderFormatSupported(header.format)) << "Incorrect format version: " + << header.format; + headerFormat_ = header.format; CHECK_EQ(header.size, getSize()) << "The size (" << header.size << ") in the file does not match the size " << "(" << getSize() << ") of the parameter: " << getName(); diff --git a/paddle/parameter/Parameter.h b/paddle/parameter/Parameter.h index 0bac76f068..04f12efaac 100644 --- a/paddle/parameter/Parameter.h +++ b/paddle/parameter/Parameter.h @@ -34,6 +34,20 @@ limitations under the License. */ namespace paddle { +typedef enum { + /// The paddle original basic format + PARAM_FORMAT_ORIGINAL = 0, + + /// See mkldnn_memory_format_t in + /// https://github.com/01org/mkl-dnn/blob/master/include/mkldnn_types.h + /// for a detailed description. + /// 2D weights tensor in the format (output channels, input channels). + PARAM_FORMAT_MKLDNN_OI, + + /// The total format items numbers + PARAM_FORMAT_ITEMS, +} PARAM_FORMAT; + class SparsePrefetchRowCpuMatrix; class Parameter; @@ -51,7 +65,10 @@ public: size_t getSize() const { return config_.size(); } bool isFullSize() const { - return this->getSize() == bufs_[PARAMETER_VALUE]->getSize(); + if (bufs_[PARAMETER_VALUE]) { + return this->getSize() == bufs_[PARAMETER_VALUE]->getSize(); + } + return false; } inline bool useGpu() const { return useGpu_; } @@ -242,14 +259,34 @@ public: /// Initialize the value to 0 void zeroMem(); - static const int kFormatVersion = 0; /// file header structure struct Header { - int32_t version; // = 0, file format version + int32_t format; // = PARAM_FORMAT uint32_t valueSize; // = sizeof(real) uint64_t size; // = getSize() }; + /** + * @brief Is the header format supported. + */ + static bool isHeaderFormatSupported(int32_t fmt) { + return fmt < PARAM_FORMAT_ITEMS; + } + + /** + * @brief Get the format in header. + */ + int getHeaderFormat() { return headerFormat_; } + + /** + * @brief Set the format in header. + */ + void setHeaderFormat(int32_t fmt) { + CHECK(isHeaderFormatSupported(fmt)) << "Unsupported format version: " + << fmt; + headerFormat_ = fmt; + } + /** * @brief Parameter Update Hook. * @@ -321,6 +358,9 @@ protected: bool updated_; SparseFormat format_; + /// The header format for saving or loading param + int32_t headerFormat_; + std::vector> updaterHooks_; public: diff --git a/paddle/platform/CMakeLists.txt b/paddle/platform/CMakeLists.txt index 4154aad15c..120eb1e4af 100644 --- a/paddle/platform/CMakeLists.txt +++ b/paddle/platform/CMakeLists.txt @@ -1,7 +1,7 @@ cc_library(cpu_info SRCS cpu_info.cc DEPS gflags glog) cc_test(cpu_info_test SRCS cpu_info_test.cc DEPS cpu_info) -nv_library(gpu_info SRCS gpu_info.cc DEPS gflags) +nv_library(gpu_info SRCS gpu_info.cc DEPS gflags glog) cc_library(place SRCS place.cc) cc_test(place_test SRCS place_test.cc DEPS place glog gflags) @@ -9,6 +9,7 @@ cc_test(place_test SRCS place_test.cc DEPS place glog gflags) add_subdirectory(dynload) cc_test(enforce_test SRCS enforce_test.cc DEPS stringpiece) +cc_test(environment_test SRCS environment_test.cc DEPS stringpiece) IF(WITH_GPU) set(GPU_CTX_DEPS dynload_cuda dynamic_loader) @@ -16,5 +17,8 @@ ELSE() set(GPU_CTX_DEPS) ENDIF() -cc_library(device_context SRCS device_context.cc DEPS place eigen3 ${GPU_CTX_DEPS}) +# memcpy deoends on device_context, here add deps individually for +# avoiding cycle dependencies +cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator + system_allocator memory_block meta_data meta_cache place eigen3 ${GPU_CTX_DEPS}) nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info) diff --git a/paddle/platform/cuda_helper.h b/paddle/platform/cuda_helper.h new file mode 100644 index 0000000000..6feec0d7f8 --- /dev/null +++ b/paddle/platform/cuda_helper.h @@ -0,0 +1,51 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include + +namespace paddle { +namespace platform { + +#define CUDA_ATOMIC_WRAPPER(op, T) \ + __device__ __forceinline__ T CudaAtomic##op(T* address, const T val) + +#define USE_CUDA_ATOMIC(op, T) \ + CUDA_ATOMIC_WRAPPER(op, T) { return atomic##op(address, val); } + +// For atomicAdd. +USE_CUDA_ATOMIC(Add, float); + +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 600 +USE_CUDA_ATOMIC(Add, double); +#else +CUDA_ATOMIC_WRAPPER(Add, double) { + unsigned long long int* address_as_ull = + reinterpret_cast(address); + unsigned long long int old = *address_as_ull, assumed; + + do { + assumed = old; + old = atomicCAS(address_as_ull, assumed, + __double_as_longlong(val + __longlong_as_double(assumed))); + + // Note: uses integer comparison to avoid hang in case of NaN + } while (assumed != old); + + return __longlong_as_double(old); +} +#endif + +} // namespace platform +} // namespace paddle diff --git a/paddle/platform/device_context.cc b/paddle/platform/device_context.cc index a928e09778..ad212c5b2c 100644 --- a/paddle/platform/device_context.cc +++ b/paddle/platform/device_context.cc @@ -10,6 +10,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/platform/device_context.h" +#include "paddle/memory/memory.h" namespace paddle { namespace platform { @@ -36,6 +37,59 @@ Place CPUDeviceContext::GetPlace() const { return CPUPlace(); } #ifndef PADDLE_ONLY_CPU +class EigenCudaStreamDevice : public Eigen::StreamInterface { + public: + EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) { + Eigen::initializeDeviceProp(); + } + ~EigenCudaStreamDevice() override {} + + void Reinitialize(const cudaStream_t* cuda_stream, GPUPlace place) { + stream_ = cuda_stream; + place_ = place; + device_prop_ = &Eigen::m_deviceProperties[place.device]; + } + + const cudaStream_t& stream() const override { return *stream_; } + + const cudaDeviceProp& deviceProperties() const override { + return *device_prop_; + } + + void* allocate(size_t num_bytes) const override { + return paddle::memory::Alloc(place_, num_bytes); + } + + void deallocate(void* buffer) const override { + paddle::memory::Free(place_, buffer); + } + + void* scratchpad() const override { + if (scratch_ == NULL) { + scratch_ = allocate(Eigen::kCudaScratchSize + sizeof(unsigned int)); + } + return scratch_; + } + + unsigned int* semaphore() const override { + if (semaphore_ == NULL) { + char* scratch = + static_cast(scratchpad()) + Eigen::kCudaScratchSize; + semaphore_ = reinterpret_cast(scratch); + PADDLE_ENFORCE( + cudaMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_)); + } + return semaphore_; + } + + private: + GPUPlace place_; + const cudaStream_t* stream_; // not owned; + const cudaDeviceProp* device_prop_; // not owned; + mutable void* scratch_; + mutable unsigned int* semaphore_; +}; + template <> Eigen::GpuDevice* DeviceContext::get_eigen_device() const { return reinterpret_cast(this)->eigen_device(); @@ -43,19 +97,9 @@ Eigen::GpuDevice* DeviceContext::get_eigen_device() const { CUDADeviceContext::CUDADeviceContext(GPUPlace place) : place_(place) { SetDeviceId(place_.device); - // TODO(qijun) Pass a created cuda stream to Eigen::CudaStreamDevice directly - // here will cause segment fault. We must implement a class derived from - // Eigen::StreamInterface, and reinitialize it with a cuda stream and a gpu id - // later. Please refer to the implementation of class EigenCudaStreamDevice - // in TensorFlow. - // - // We find that CUDA 7 introduces a new option, the per-thread default stream, - // that has two effects. Please refer to https://devblogs.nvidia.com/ - // parallelforall/gpu-pro-tip-cuda-7-streams-simplify-concurrency/ - // - // So, we decide to use default stream and add –default-stream per-thread nvcc - // flag. Than, two threads with two CUDADeviceContexts will run parallelly. - eigen_stream_.reset(new Eigen::CudaStreamDevice()); + PADDLE_ENFORCE(cudaStreamCreate(&stream_)); + eigen_stream_.reset(new EigenCudaStreamDevice()); + eigen_stream_->Reinitialize(&stream_, place); eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get())); } @@ -70,17 +114,15 @@ CUDADeviceContext::~CUDADeviceContext() { PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_)); } - if (curand_generator_) { - PADDLE_ENFORCE(dynload::curandDestroyGenerator(curand_generator_)); - } eigen_stream_.reset(); eigen_device_.reset(); + PADDLE_ENFORCE(cudaStreamDestroy(stream_)); } Place CUDADeviceContext::GetPlace() const { return place_; } void CUDADeviceContext::Wait() const { - PADDLE_ENFORCE(cudaStreamSynchronize(0)); + PADDLE_ENFORCE(cudaStreamSynchronize(stream_)); } Eigen::GpuDevice* CUDADeviceContext::eigen_device() const { @@ -91,6 +133,7 @@ cublasHandle_t CUDADeviceContext::cublas_handle() { if (!cublas_handle_) { SetDeviceId(place_.device); PADDLE_ENFORCE(dynload::cublasCreate(&cublas_handle_)); + PADDLE_ENFORCE(dynload::cublasSetStream(cublas_handle_, stream_)); } return cublas_handle_; } @@ -99,20 +142,12 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() { if (!cudnn_handle_) { SetDeviceId(place_.device); PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_)); + PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream_)); } return cudnn_handle_; } -curandGenerator_t CUDADeviceContext::curand_generator() { - if (!curand_generator_) { - SetDeviceId(place_.device); - PADDLE_ENFORCE(dynload::curandCreateGenerator(&curand_generator_, - CURAND_RNG_PSEUDO_DEFAULT)); - PADDLE_ENFORCE( - dynload::curandSetPseudoRandomGeneratorSeed(curand_generator_, seed_)); - } - return curand_generator_; -} +cudaStream_t CUDADeviceContext::stream() { return stream_; } #endif // PADDLE_ONLY_CPU diff --git a/paddle/platform/device_context.h b/paddle/platform/device_context.h index 08b5b2cff9..11528e1194 100644 --- a/paddle/platform/device_context.h +++ b/paddle/platform/device_context.h @@ -17,7 +17,6 @@ limitations under the License. */ #ifndef PADDLE_ONLY_CPU #include "paddle/platform/dynload/cublas.h" #include "paddle/platform/dynload/cudnn.h" -#include "paddle/platform/dynload/curand.h" #include "paddle/platform/gpu_info.h" #define EIGEN_USE_GPU #endif @@ -40,7 +39,7 @@ class DeviceContext { class CPUDeviceContext : public DeviceContext { public: CPUDeviceContext(); - explicit CPUDeviceContext(CPUPlace); + explicit CPUDeviceContext(CPUPlace place); virtual ~CPUDeviceContext() {} Eigen::DefaultDevice* eigen_device() const; @@ -52,10 +51,11 @@ class CPUDeviceContext : public DeviceContext { }; #ifndef PADDLE_ONLY_CPU +class EigenCudaStreamDevice; class CUDADeviceContext : public DeviceContext { public: - explicit CUDADeviceContext(GPUPlace); + explicit CUDADeviceContext(GPUPlace place); virtual ~CUDADeviceContext(); /*! \brief Wait for all operations completion in the stream. */ @@ -74,24 +74,20 @@ class CUDADeviceContext : public DeviceContext { /*! \brief Return cudnn handle in the device context. */ cudnnHandle_t cudnn_handle(); - /*! \brief Return curand handle in the device context. */ - curandGenerator_t curand_generator(); + /*! \brief Return cuda stream in the device context. */ + cudaStream_t stream(); // clang-format on private: GPUPlace place_; - private: std::unique_ptr eigen_device_; - std::unique_ptr eigen_stream_; - - private: - uint64_t seed_; + std::unique_ptr eigen_stream_; // clang-format off - cudnnHandle_t cudnn_handle_ = nullptr; - cublasHandle_t cublas_handle_ = nullptr; - curandGenerator_t curand_generator_ = nullptr; + cudaStream_t stream_{nullptr}; + cudnnHandle_t cudnn_handle_{nullptr}; + cublasHandle_t cublas_handle_{nullptr}; // clang-format on }; diff --git a/paddle/platform/device_context_test.cc b/paddle/platform/device_context_test.cc index 65345c433c..5883a55272 100644 --- a/paddle/platform/device_context_test.cc +++ b/paddle/platform/device_context_test.cc @@ -43,8 +43,7 @@ TEST(Device, CUDADeviceContext) { ASSERT_NE(nullptr, cudnn_handle); cublasHandle_t cublas_handle = device_context->cublas_handle(); ASSERT_NE(nullptr, cublas_handle); - curandGenerator_t curand_handle = device_context->curand_generator(); - ASSERT_NE(nullptr, curand_handle); + ASSERT_NE(nullptr, device_context->stream()); delete device_context; } } diff --git a/paddle/platform/dynload/cublas.h b/paddle/platform/dynload/cublas.h index c44b7240a8..9d8343c0b5 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/curand.h b/paddle/platform/dynload/curand.h index d8c46bc41e..7bfe0778c7 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 d2adb997de..81448897e9 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\n", i, 2 + sizeof(void*) * 2, + call_stack[i]); + } } - free(line); + free(symbols); err_str_ = sout.str(); } } @@ -169,7 +204,7 @@ inline void throw_on_error(T e) { * PADDLE_ENFORCE_EQ(a, b); * * will raise an expression described as follows: - * "enforce a == b failed, 1 != 2" with detailed stack infomation. + * "enforce a == b failed, 1 != 2" with detailed stack information. * * extra messages is also supported, for example: * PADDLE_ENFORCE(a, b, "some simple enforce failed between %d numbers", 2) @@ -194,8 +229,8 @@ inline void throw_on_error(T e) { #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, std::to_string(__VAL0), \ - std::to_string(__VAL1), \ + #__VAL0, #__VAL1, paddle::string::to_string(__VAL0), \ + paddle::string::to_string(__VAL1), \ paddle::string::Sprintf("" __VA_ARGS__)); } // namespace platform diff --git a/paddle/platform/enforce_test.cc b/paddle/platform/enforce_test.cc index 5408fce558..80bdee3d9d 100644 --- a/paddle/platform/enforce_test.cc +++ b/paddle/platform/enforce_test.cc @@ -9,6 +9,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include +#include #include #include "gtest/gtest.h" @@ -83,7 +85,7 @@ TEST(ENFORCE_NE, FAIL) { } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; EXPECT_TRUE(HasPrefix(StringPiece(error.what()), - "enforce 1.0 != 1UL failed, 1.000000 == 1")) + "enforce 1.0 != 1UL failed, 1 == 1")) << error.what() << " does not have expected prefix"; } EXPECT_TRUE(caught_exception); @@ -176,3 +178,39 @@ TEST(ENFORCE_NOT_NULL, FAIL) { } 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 << ", "; + } + } + 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/environment.h b/paddle/platform/environment.h new file mode 100644 index 0000000000..4edcce932e --- /dev/null +++ b/paddle/platform/environment.h @@ -0,0 +1,60 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include + +#include "paddle/platform/enforce.h" +#include "paddle/string/piece.h" + +extern char** environ; // for environment variables + +namespace paddle { +namespace platform { + +inline void SetEnvVariable(const std::string& name, const std::string& value) { + PADDLE_ENFORCE_NE(setenv(name.c_str(), value.c_str(), 1), -1, + "Failed to set environment variable %s=%s", name, value); +} + +inline void UnsetEnvVariable(const std::string& name) { + PADDLE_ENFORCE_NE(unsetenv(name.c_str()), -1, + "Failed to unset environment variable %s", name); +} + +inline bool IsEnvVarDefined(const std::string& name) { + return std::getenv(name.c_str()) != nullptr; +} + +inline std::string GetEnvValue(const std::string& name) { + PADDLE_ENFORCE(IsEnvVarDefined(name), + "Tried to access undefined environment variable %s", name); + return std::getenv(name.c_str()); +} + +inline std::vector GetAllEnvVariables() { + std::vector vars; + for (auto var = environ; *var != nullptr; ++var) { + auto tail = string::Index(*var, "="); + auto name = string::SubStr(*var, 0, tail).ToString(); + vars.push_back(name); + } + return vars; +} + +} // namespace platform +} // namespace paddle diff --git a/paddle/platform/environment_test.cc b/paddle/platform/environment_test.cc new file mode 100644 index 0000000000..5f13652721 --- /dev/null +++ b/paddle/platform/environment_test.cc @@ -0,0 +1,54 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/platform/environment.h" + +#include "glog/logging.h" +#include "gtest/gtest.h" + +TEST(ENVIRONMENT, ACCESS) { + namespace platform = paddle::platform; + namespace string = paddle::string; + + platform::SetEnvVariable("PADDLE_USE_ENV", "TRUE"); + + EXPECT_TRUE(platform::IsEnvVarDefined("PADDLE_USE_ENV")); + EXPECT_EQ(platform::GetEnvValue("PADDLE_USE_ENV"), "TRUE"); + + platform::UnsetEnvVariable("PADDLE_USE_ENV"); + EXPECT_FALSE(platform::IsEnvVarDefined("PADDLE_USE_ENV")); + + platform::SetEnvVariable("PADDLE_USE_ENV1", "Hello "); + platform::SetEnvVariable("PADDLE_USE_ENV2", "World, "); + platform::SetEnvVariable("PADDLE_USE_ENV3", "PaddlePaddle!"); + + std::string env_info; + auto vars = platform::GetAllEnvVariables(); + for_each(vars.begin(), vars.end(), [&](const std::string& var) { + env_info += platform::GetEnvValue(var); + }); + + EXPECT_TRUE(string::Contains(env_info, "Hello World, PaddlePaddle!")); + platform::UnsetEnvVariable("PADDLE_USE_ENV1"); + platform::UnsetEnvVariable("PADDLE_USE_ENV2"); + platform::UnsetEnvVariable("PADDLE_USE_ENV3"); + + env_info.clear(); + vars = platform::GetAllEnvVariables(); + for_each(vars.begin(), vars.end(), [&](const std::string& var) { + env_info += platform::GetEnvValue(var); + }); + + EXPECT_FALSE(string::Contains(env_info, "Hello World, PaddlePaddle!")); + EXPECT_FALSE(platform::IsEnvVarDefined("PADDLE_USE_ENV1")); + EXPECT_FALSE(platform::IsEnvVarDefined("PADDLE_USE_ENV2")); + EXPECT_FALSE(platform::IsEnvVarDefined("PADDLE_USE_ENV3")); +} diff --git a/paddle/platform/gpu_info.cc b/paddle/platform/gpu_info.cc index edeb3ecd7b..be381a4e26 100644 --- a/paddle/platform/gpu_info.cc +++ b/paddle/platform/gpu_info.cc @@ -13,8 +13,11 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/platform/gpu_info.h" + #include "gflags/gflags.h" + #include "paddle/platform/enforce.h" +#include "paddle/platform/environment.h" DEFINE_double(fraction_of_gpu_memory_to_use, 0.95, "Default use 95% of GPU memory for PaddlePaddle," @@ -70,6 +73,13 @@ size_t GpuMaxChunkSize() { GpuMemoryUsage(available, total); + if (IsEnvVarDefined(kEnvFractionGpuMemoryToUse)) { + auto val = std::stod(GetEnvValue(kEnvFractionGpuMemoryToUse)); + PADDLE_ENFORCE_GT(val, 0.0); + PADDLE_ENFORCE_LE(val, 1.0); + FLAGS_fraction_of_gpu_memory_to_use = val; + } + // Reserving the rest memory for page tables, etc. size_t reserving = (1 - FLAGS_fraction_of_gpu_memory_to_use) * total; diff --git a/paddle/platform/gpu_info.h b/paddle/platform/gpu_info.h index d3a5f5f13f..ed2420b874 100644 --- a/paddle/platform/gpu_info.h +++ b/paddle/platform/gpu_info.h @@ -18,10 +18,15 @@ limitations under the License. */ #include #include +#include namespace paddle { namespace platform { +//! Environment variable: fraction of GPU memory to use on each device. +const std::string kEnvFractionGpuMemoryToUse = + "PADDLE_FRACTION_GPU_MEMORY_TO_USE"; + //! Get the total number of GPU devices in system. int GetDeviceCount(); diff --git a/paddle/platform/place.h b/paddle/platform/place.h index a82e8c942f..1117476bb3 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 { diff --git a/paddle/platform/variant.h b/paddle/platform/variant.h new file mode 100644 index 0000000000..c2257af1b5 --- /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/ParameterClient2.cpp b/paddle/pserver/ParameterClient2.cpp index f7e391f763..54063a809a 100644 --- a/paddle/pserver/ParameterClient2.cpp +++ b/paddle/pserver/ParameterClient2.cpp @@ -65,7 +65,6 @@ void ParameterClient2::initThreads() { LOG(INFO) << "parallel_thread_num dosent need to set"; } syncThreadPool_.reset(new SyncThreadPool(threadNum_)); - startThreads(); } @@ -224,6 +223,14 @@ void ParameterClient2::prepareSendData( request.set_cost(cost); request.set_batch_status(batchStatus); CHECK_EQ(request.blocks_size(), 0); + VLOG(10) << "request: trainer_id: " << request.trainer_id() + << " update_mode" << request.update_mode() + << " send_back_parameter: " << request.send_back_parameter() + << " send_back_parameter_type: " + << request.send_back_parameter_type() + << " num_samples: " << request.num_samples() + << " cost: " << request.cost() + << " batch_status: " << request.batch_status(); } for (const auto& segments : parameterSegments) { const auto it = parameterMap_.find(segments.id); @@ -251,11 +258,17 @@ void ParameterClient2::prepareSendData( CHECK(sendMat != nullptr) << "sendMat is nullptr"; syncThreadPool_->exec([&](int tid, size_t numThreads) { + std::lock_guard guard(sparseAutoGrowthMutex_); const auto& localIndices = prefetchMat->getLocalIndices(); /// num of sparse rows size_t nLocalBlocks = localIndices.size(); uint64_t beginDim = 0; uint64_t endDim = 0; + + // FIXME(typhoonzero): let it resize first + prefetchMat->getLocalRow(nLocalBlocks + 1); + sendMat->getLocalRow(nLocalBlocks + 1); + for (size_t row = 0; row < nLocalBlocks; ++row) { int64_t blockId = localIndices[row]; // local row -> sparse row int serverId = std::abs((blockId + nameHash) % serviceNum_); @@ -275,7 +288,6 @@ void ParameterClient2::prepareSendData( block->set_begin_pos(row * blockSize); /// block len block->set_block_size(endDim - beginDim); - if (sendingPara) { sendJob->parallelInputIovs[serverId].push_back( {sendMat->getLocalRow(row), sizeof(real) * (size_t)blockSize}); diff --git a/paddle/pserver/ParameterClient2.h b/paddle/pserver/ParameterClient2.h index 89b3ddd502..29b9eeacdd 100644 --- a/paddle/pserver/ParameterClient2.h +++ b/paddle/pserver/ParameterClient2.h @@ -583,6 +583,7 @@ protected: #ifndef PADDLE_DISABLE_TIMER uint64_t forwardbackwordTime_; #endif + std::mutex sparseAutoGrowthMutex_; /// map id to parameter used for decoding protobuf data std::unordered_map parameterMap_; diff --git a/paddle/pserver/ParameterServer2.cpp b/paddle/pserver/ParameterServer2.cpp index d7c1d4f788..54f5c4c0fb 100644 --- a/paddle/pserver/ParameterServer2.cpp +++ b/paddle/pserver/ParameterServer2.cpp @@ -1032,8 +1032,8 @@ void ParameterServer2::loadValueVector(const LoadValueRequest& request, Parameter::Header header; CHECK(fs.read(reinterpret_cast(&header), sizeof(header))) << "Fail to read parameters in pserver"; - CHECK_EQ(header.version, Parameter::kFormatVersion) - << "Incorrect format version: " << header.version; + CHECK(Parameter::isHeaderFormatSupported(header.format)) + << "Incorrect format version: " << header.format; CHECK_EQ(header.size, (size_t)size_) << "The size (" << header.size << ") in the file does not match the size " << "(" << size_ << ") of the pserver: " << serverId_; @@ -1063,7 +1063,8 @@ void ParameterServer2::saveValueVector(const SaveValueRequest& request, CpuVector& vec = vectors_[PARAMETER_APPLY] ? *vectors_[PARAMETER_APPLY] : *vectors_[PARAMETER_VALUE]; Parameter::Header header; - header.version = Parameter::kFormatVersion; + // TODO(TJ): save param headerFormat_ + header.format = PARAM_FORMAT_ORIGINAL; header.valueSize = sizeof(real); header.size = size_; diff --git a/paddle/pserver/test/CMakeLists.txt b/paddle/pserver/test/CMakeLists.txt index 6e8f9c37f6..b66a00ba06 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 new file mode 100644 index 0000000000..37e186a408 --- /dev/null +++ b/paddle/pybind/CMakeLists.txt @@ -0,0 +1,22 @@ +if(WITH_PYTHON) +cc_library(paddle_pybind SHARED + SRCS pybind.cc + DEPS pybind python backward + sgd_op + gather_op + scatter_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 + lookup_table_op + scale_op + minus_op) +endif(WITH_PYTHON) diff --git a/paddle/framework/pybind.cc b/paddle/pybind/pybind.cc similarity index 62% rename from paddle/framework/pybind.cc rename to paddle/pybind/pybind.cc index 915ffb1c00..3bc150ccb7 100644 --- a/paddle/framework/pybind.cc +++ b/paddle/pybind/pybind.cc @@ -18,10 +18,12 @@ limitations under the License. */ #include "paddle/framework/backward.h" #include "paddle/framework/op_registry.h" -#include "paddle/framework/tensor_py.h" #include "paddle/operators/net_op.h" +#include "paddle/operators/recurrent_op.h" #include "paddle/platform/enforce.h" #include "paddle/platform/place.h" +#include "paddle/pybind/tensor_py.h" +#include "paddle/string/to_string.h" #include "pybind11/numpy.h" #include "pybind11/pybind11.h" #include "pybind11/stl.h" @@ -29,8 +31,7 @@ limitations under the License. */ namespace py = pybind11; USE_OP(add_two); -USE_OP_CPU(onehot_cross_entropy); -USE_OP_WITHOUT_KERNEL(fc); +USE_OP(onehot_cross_entropy); USE_OP(sgd); USE_OP(mul); USE_OP(mean); @@ -38,48 +39,21 @@ USE_OP(sigmoid); USE_OP(softmax); USE_OP(rowwise_add); USE_OP(fill_zeros_like); -USE_OP_WITHOUT_KERNEL(recurrent_op); +USE_OP_ITSELF(recurrent_op); +USE_OP(gaussian_random); USE_OP(uniform_random); +USE_OP(lookup_table); +USE_OP(scale); +USE_OP_ITSELF(identity); +USE_OP(minus); +USE_CPU_ONLY_OP(gather); +USE_CPU_ONLY_OP(scatter); + namespace paddle { namespace framework { using Tensor = framework::Tensor; -template -void ExposeOperator(ClassType &m) { - m.def("infer_shape", &ClassType::type::InferShape) - .def("run", &ClassType::type::Run) - .def("type", - [](const typename ClassType::type &op) -> std::string { - return op.type_; - }) - .def("outputs", - [](const typename ClassType::type &op) -> std::vector { - return op.outputs_; - }) - .def("inputs", - [](const typename ClassType::type &op) -> std::vector { - return op.inputs_; - }) - .def("support_gpu", &ClassType::type::SupportGPU) - .def("temp_outputs", - [](const typename ClassType::type &op) -> std::vector { - auto iter = op.attrs_.find("temporary_index"); - std::vector ret; - if (iter == op.attrs_.end()) { - return ret; - } else { - auto tmp_idx = boost::get>(iter->second); - ret.reserve(tmp_idx.size()); - for (auto &index : tmp_idx) { - ret.push_back(op.outputs_.at(index)); - } - return ret; - } - }) - .def("__str__", &ClassType::type::DebugString); -} - static size_t UniqueIntegerGenerator() { static std::atomic generator; return generator.fetch_add(1); @@ -163,23 +137,24 @@ All parameter, weight, gradient are variables in Paddle. py::return_value_policy::reference) .def("find_var", &Scope::FindVar, py::return_value_policy::reference) .def(py::init<>()) - .def("new_scope", [](Scope &self) -> Scope * { return &self.NewScope(); }, + .def("new_scope", + [](Scope &self) -> Scope * { return &self.NewScope(); }, py::return_value_policy::reference) .def("drop_kids", &Scope::DropKids); //! @note: Be careful! PyBind will return std::string as an unicode, not //! Python str. If you want a str object, you should cast them in Python. m.def("get_all_op_protos", []() -> std::vector { - auto &protos = OpRegistry::protos(); std::vector ret_values; - for (auto it = protos.begin(); it != protos.end(); ++it) { - PADDLE_ENFORCE(it->second.IsInitialized(), - "OpProto must all be initialized"); + + OpInfoMap::Instance().IterAllInfo([&ret_values](const std::string &type, + const OpInfo &info) { + if (!info.HasOpProtoAndChecker()) return; std::string str; - PADDLE_ENFORCE(it->second.SerializeToString(&str), + PADDLE_ENFORCE(info.Proto().SerializeToString(&str), "Serialize OpProto Error. This could be a bug of Paddle."); - ret_values.push_back(py::bytes(str)); - } + ret_values.emplace_back(str); + }); return ret_values; }); m.def_submodule( @@ -205,51 +180,78 @@ All parameter, weight, gradient are variables in Paddle. }); // clang-format on - py::class_(m, "GPUPlace").def(py::init()); - - py::class_(m, "CPUPlace").def(py::init<>()); - - py::class_> operator_base( - m, "Operator"); + py::class_(m, "GPUPlace") + .def(py::init()) + .def("__str__", string::to_string); - operator_base.def_static("create", [](py::bytes protobin) { - OpDesc desc; - PADDLE_ENFORCE(desc.ParsePartialFromString(protobin), - "Cannot parse user input to OpDesc"); - PADDLE_ENFORCE(desc.IsInitialized(), - "User OpDesc is not initialized, reason %s", - desc.InitializationErrorString()); - return OpRegistry::CreateOp(desc); - }); - - operator_base.def("backward", - [](const OperatorBase &forwardOp, - const std::unordered_set &no_grad_vars) { - return Backward(forwardOp, no_grad_vars); - }); - - ExposeOperator(operator_base); + py::class_(m, "CPUPlace") + .def(py::init<>()) + .def("__str__", string::to_string); - py::class_> net(m, "Net"); + py::class_(m, "Operator") + .def_static("create", + [](py::bytes protobin) { + OpDesc desc; + PADDLE_ENFORCE(desc.ParsePartialFromString(protobin), + "Cannot parse user input to OpDesc"); + PADDLE_ENFORCE(desc.IsInitialized(), + "User OpDesc is not initialized, reason %s", + desc.InitializationErrorString()); + return OpRegistry::CreateOp(desc); + }) + .def("backward", + [](const OperatorBase &forwardOp, + const std::unordered_set &no_grad_vars) { + return Backward(forwardOp, no_grad_vars).release(); + }) + .def("infer_shape", &OperatorBase::InferShape) + .def("run", &OperatorBase::Run) + .def("type", + [](const OperatorBase &op) -> std::string { return op.Type(); }) + .def("outputs", + [](const OperatorBase &op) + -> std::map> { + return op.Outputs(); + }) + .def("inputs", [](const OperatorBase &op) { return op.Inputs(); }) + .def("__str__", &OperatorBase::DebugString) + .def("no_intermediate_outputs", + [](const OperatorBase &op) { return op.OutputVars(false); }) + .def("support_gpu", &OperatorBase::SupportGPU); - net.def_static("create", - []() -> std::shared_ptr { - auto retv = std::make_shared(); - retv->type_ = "plain_net"; - return retv; - }) - .def("add_op", &operators::NetOp::AddOp) - .def("add_op", - [](operators::NetOp &self, - const std::shared_ptr &net) -> void { - self.AddOp(std::static_pointer_cast(net)); + py::class_(m, "Net") + .def_static("create", + []() -> operators::NetOp * { + auto *retv = new operators::NetOp; + retv->SetType("plain_net"); + return retv; + }) + .def("append_op", + [](operators::NetOp &self, const OperatorBase &op) { + self.AppendOp(op); }) .def("complete_add_op", &operators::NetOp::CompleteAddOp) .def("complete_add_op", [](std::shared_ptr &self) { self->CompleteAddOp(); }); - ExposeOperator(net); + // recurrent_op + py::class_(m, "RecurrentOp") + .def_static( + "create", + [](py::bytes protobin) -> operators::RecurrentOp * { + OpDesc desc; + PADDLE_ENFORCE(desc.ParsePartialFromString(protobin), + "Cannot parse user input to OpDesc"); + PADDLE_ENFORCE(desc.IsInitialized(), + "User OpDesc is not initialized, reason %s", + desc.InitializationErrorString()); + auto rnn_op = OpRegistry::CreateOp(desc); + return static_cast(rnn_op.release()); + }) + .def("set_stepnet", + [](operators::RecurrentOp &self, const operators::NetOp &net) + -> void { self.set_stepnet(net.Clone()); }); m.def("unique_integer", UniqueIntegerGenerator); diff --git a/paddle/framework/tensor_py.h b/paddle/pybind/tensor_py.h similarity index 92% rename from paddle/framework/tensor_py.h rename to paddle/pybind/tensor_py.h index 4e1ab77b15..39ba60b4dc 100644 --- a/paddle/framework/tensor_py.h +++ b/paddle/pybind/tensor_py.h @@ -63,8 +63,11 @@ struct CastToPyBufferImpl { } return py::buffer_info( dst_tensor.mutable_data(dst_tensor.holder_->place()), - sizeof(CUR_TYPE), py::format_descriptor::format(), - (size_t)framework::arity(dst_tensor.dims()), dims_outside, strides); + sizeof(CUR_TYPE), + py::format_descriptor::format(), + (size_t)framework::arity(dst_tensor.dims()), + dims_outside, + strides); } else { constexpr bool less = I + 1 < std::tuple_size>::value; return CastToPyBufferImpl()(tensor); @@ -107,8 +110,8 @@ void PyCUDATensorSetFromArray( self.Resize(framework::make_ddim(dims)); auto *dst = self.mutable_data(place); - paddle::platform::GpuMemcpySync(dst, array.data(), sizeof(T) * array.size(), - cudaMemcpyHostToDevice); + paddle::platform::GpuMemcpySync( + dst, array.data(), sizeof(T) * array.size(), cudaMemcpyHostToDevice); } #endif diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh index 44442be472..1798642022 100644 --- a/paddle/scripts/docker/build.sh +++ b/paddle/scripts/docker/build.sh @@ -31,14 +31,14 @@ 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} + -DWITH_TESTING=${WITH_TESTING:-ON} -DCMAKE_EXPORT_COMPILE_COMMANDS=ON ======================================== EOF @@ -51,41 +51,36 @@ cmake .. \ -DWITH_DOC=OFF \ -DWITH_GPU=${WITH_GPU:-OFF} \ -DWITH_AVX=${WITH_AVX:-OFF} \ - -DWITH_GOLANG=${WITH_GOLANG:-OFF} \ + -DWITH_GOLANG=${WITH_GOLANG:-ON} \ -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} \ -DWITH_C_API=${WITH_C_API:-OFF} \ -DWITH_PYTHON=${WITH_PYTHON:-ON} \ -DCUDNN_ROOT=/usr/ \ - -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} \ - -DWITH_TESTING=${WITH_TESTING:-OFF} \ + -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON} \ + -DWITH_TESTING=${WITH_TESTING:-ON} \ -DCMAKE_EXPORT_COMPILE_COMMANDS=ON cat <> /paddle/build/Dockerfile </dev/null) - BASEDIR=$(dirname "$0") - pip install ${BASEDIR}/../opt/paddle/share/wheels/*-${PYTHON_PADDLE_VERSION}-*.whl - if [ $? -ne 0 ]; then - echo "pip install wheels failed. " - echo "Please use 'sudo paddle' at the first time you use PaddlePaddle" - echo "PaddlePaddle will install some python dependencies automatically." - exit 1 - fi - echo "Python dependencies are installed." -fi case "$1" in "train") - ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_trainer ${@:2} + ${DEBUGGER} $PADDLE_BIN_PATH/paddle_trainer ${@:2} ;; "merge_model") - ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_merge_model ${@:2} + ${DEBUGGER} $PADDLE_BIN_PATH/paddle_merge_model ${@:2} ;; "pserver") - ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_pserver_main ${@:2} + ${DEBUGGER} $PADDLE_BIN_PATH/paddle_pserver_main ${@:2} ;; "dump_config") python -m paddle.utils.dump_config ${@:2} @@ -127,7 +110,7 @@ case "$1" in python -m paddle.utils.make_model_diagram ${@:2} ;; "usage") - $MYDIR/../opt/paddle/bin/paddle_usage ${@:2} + $PADDLE_BIN_PATH/paddle_usage ${@:2} ;; "version") version diff --git a/paddle/string/CMakeLists.txt b/paddle/string/CMakeLists.txt index 5becf62672..60667b7287 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/to_string.h b/paddle/string/to_string.h new file mode 100644 index 0000000000..4f478b6a36 --- /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 0000000000..5ff1b007f1 --- /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 af1dceed02..35dcb235e7 100644 --- a/paddle/trainer/NewRemoteParameterUpdater.cpp +++ b/paddle/trainer/NewRemoteParameterUpdater.cpp @@ -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 133e2be104..a0a365aa0b 100644 --- a/paddle/trainer/TrainerConfigHelper.cpp +++ b/paddle/trainer/TrainerConfigHelper.cpp @@ -28,6 +28,7 @@ DECLARE_bool(with_cost); DECLARE_bool(with_gpu); DECLARE_bool(parallel_nn); DECLARE_string(config_args); +DECLARE_bool(use_mkldnn); const char *kConfigParserModuleName = "paddle.trainer.config_parser"; const char *kConfigParserFuncName = "parse_config_and_serialize"; @@ -44,6 +45,7 @@ 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 << ",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 08b2d8a38e..f01ad4142d 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/utils/Flags.cpp b/paddle/utils/Flags.cpp index 320f671ed9..ab1c181c62 100644 --- a/paddle/utils/Flags.cpp +++ b/paddle/utils/Flags.cpp @@ -20,6 +20,13 @@ 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(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 dc4faef833..1832bb515e 100644 --- a/paddle/utils/Flags.h +++ b/paddle/utils/Flags.h @@ -40,3 +40,4 @@ DECLARE_bool(show_layer_stat); DECLARE_string(predict_file); DECLARE_bool(prev_batch_state); DECLARE_string(init_model_path); +DECLARE_bool(use_mkldnn); diff --git a/paddle/utils/tests/CMakeLists.txt b/paddle/utils/tests/CMakeLists.txt index aa923b3553..c770ce1698 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 e1cea8bd0d..6212c2e60a 100644 --- a/proto/CMakeLists.txt +++ b/proto/CMakeLists.txt @@ -9,13 +9,13 @@ 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 ${ABS_FIL} protoc) endforeach() diff --git a/proto/ModelConfig.proto b/proto/ModelConfig.proto index ef2b076c33..95c236ad88 100644 --- a/proto/ModelConfig.proto +++ b/proto/ModelConfig.proto @@ -83,11 +83,14 @@ message ConvConfig { // if not set, use img_size optional uint32 img_size_y = 14; - optional uint32 filter_size_z = 15 [ default = 1 ]; - optional uint32 padding_z = 16 [ default = 1 ]; - optional uint32 stride_z = 17 [ default = 1 ]; - optional uint32 output_z = 18 [ default = 1 ]; - optional uint32 img_size_z = 19 [ default = 1 ]; + optional uint32 dilation = 15 [ default = 1 ]; + optional uint32 dilation_y = 16 [ default = 1 ]; + + optional uint32 filter_size_z = 17 [ default = 1 ]; + optional uint32 padding_z = 18 [ default = 1 ]; + optional uint32 stride_z = 19 [ default = 1 ]; + optional uint32 output_z = 20 [ default = 1 ]; + optional uint32 img_size_z = 21 [ default = 1 ]; } message PoolConfig { @@ -509,6 +512,9 @@ message LayerConfig { optional int32 axis = 54 [ default = 2 ]; repeated uint32 offset = 55; repeated uint32 shape = 56; + + // for HuberRegressionLoss + optional double delta = 57 [ default = 1.0 ]; } message EvaluatorConfig { diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt index b5030da8e7..7bd6d59b00 100644 --- a/python/CMakeLists.txt +++ b/python/CMakeLists.txt @@ -1,5 +1,3 @@ -set(OUTPUT_DIR - "${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 paddle_pserver_main paddle_trainer paddle_merge_model python_api_wheel) +set(paddle_python_deps ${PADDLE_PYTHON_BUILD_DIR}/.timestamp paddle_pserver_main paddle_trainer paddle_merge_model ${MKL_DEPENDS}) +if(WITH_SWIG_PY) + list(APPEND paddle_python_deps python_api_wheel) +endif() +add_custom_target(paddle_python ALL DEPENDS ${paddle_python_deps}) set(PADDLE_PYTHON_PACKAGE_DIR ${CMAKE_CURRENT_BINARY_DIR}/dist/) diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py index b7b696ef0c..c11037c3c8 100644 --- a/python/paddle/trainer/config_parser.py +++ b/python/paddle/trainer/config_parser.py @@ -338,7 +338,8 @@ def RecurrentLayerGroupWithoutOutLinksBegin(name, in_links_count += 1 layer_name = MakeLayerNameInParentSubmodel(name) layer = g_layer_map[layer_name] - ScatterAgentLayer(name=name, size=layer.size) + ScatterAgentLayer( + name=name, size=layer.size, width=layer.width, height=layer.height) pair = g_current_submodel.in_links.add() pair.layer_name = layer_name @@ -869,12 +870,16 @@ class Conv(Cfg): caffe_mode=True, filter_size_y=None, padding_y=None, - stride_y=None): + stride_y=None, + dilation=None, + dilation_y=None): self.add_keys(locals()) if filter_size_y is None: self.filter_size_y = filter_size if padding_y is None: self.padding_y = padding + if dilation_y is None: + self.dilation_y = dilation if stride_y is None: self.stride_y = stride if output_x is not None: @@ -1604,6 +1609,8 @@ class MultiClassCrossEntropySelfNormCostLayer(LayerBase): @config_layer('fc') class FCLayer(LayerBase): + layer_type = 'fc' + def __init__(self, name, size, @@ -1611,14 +1618,27 @@ 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 +1651,11 @@ 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, @@ -2177,8 +2202,8 @@ class MaxOutLayer(LayerBase): maxout_conf = self.config.inputs[0].maxout_conf parse_maxout(self.inputs[0].maxout, input_layer.name, maxout_conf) out_channels = maxout_conf.image_conf.channels / maxout_conf.groups - self.set_cnn_layer(name, g_layer_map[input_layer.name].height, - g_layer_map[input_layer.name].width, out_channels) + self.set_cnn_layer(name, maxout_conf.image_conf.img_size_y, + maxout_conf.image_conf.img_size, out_channels) @config_layer('row_conv') @@ -2212,6 +2237,20 @@ class ClipLayer(LayerBase): self.config.inputs[0].clip_conf.max = max +@config_layer('scale_shift') +class ScaleShiftLayer(LayerBase): + def __init__(self, name, inputs, bias=True, **xargs): + super(ScaleShiftLayer, self).__init__( + name, 'scale_shift', 0, inputs=inputs, **xargs) + config_assert( + len(self.inputs) == 1, + 'ScaleShiftLayer must have one and only one input.') + input_layer = self.get_input_layer(0) + self.set_layer_size(input_layer.size) + self.create_input_parameter(0, 1, [1, 1]) + self.create_bias_parameter(bias, 1) + + # key: cost type # value: cost class g_cost_map = {} @@ -2235,7 +2274,7 @@ define_cost('PnpairValidation', 'pnpair-validation') define_cost('SumOfSquaresCostLayer', 'square_error') define_cost('MultiBinaryLabelCrossEntropy', 'multi_binary_label_cross_entropy') define_cost('SoftBinaryClassCrossEntropy', 'soft_binary_class_cross_entropy') -define_cost('HuberTwoClass', 'huber') +define_cost('HuberTwoClassification', 'huber_classification') define_cost('SumCost', 'sum_cost') define_cost('SmoothL1Cost', 'smooth_l1') @@ -2297,6 +2336,17 @@ class LambdaCost(LayerBase): self.config.max_sort_size = max_sort_size +@config_layer('huber_regression') +class HuberRegressionLoss(LayerBase): + def __init__(self, name, inputs, delta=1., coeff=1., device=None): + super(HuberRegressionLoss, self).__init__( + name, 'huber_regression', 1, inputs=inputs, device=device) + config_assert( + len(self.inputs) == 2, 'HuberRegression must have 2 inputs') + self.config.delta = delta + self.config.coeff = coeff + + @config_layer('nce') class NCELayer(LayerBase): def __init__(self, @@ -2371,9 +2421,11 @@ class GatherAgentLayer(LayerBase): @config_layer('scatter_agent') class ScatterAgentLayer(LayerBase): - def __init__(self, name, size, device=None): + def __init__(self, name, size, width=None, height=None, device=None): super(ScatterAgentLayer, self).__init__( name, 'scatter_agent', size, inputs=[], device=device) + if height and width: + self.set_layer_height_width(height, width) @config_layer('multiplex') @@ -2657,6 +2709,49 @@ class SubSequenceLayer(LayerBase): self.create_bias_parameter(bias, size) +@config_layer('seq_slice') +class SeqSliceLayer(LayerBase): + def __init__(self, name, inputs, starts, ends, bias=False, **xargs): + if isinstance(inputs, list): + assert len(inputs) == 1, ('the first input of sequence slice layer ' + 'is a single sequence input.') + else: + inputs = [inputs] + + if starts is not None: + if isinstance(starts, list): + assert len(starts) == 1, ( + 'the start indices for sequence slice layer cannot ' + 'be a list having more than one element.') + starts = starts[0] + inputs.append(starts) + + if ends is not None: + if isinstance(ends, list): + assert len(ends) == 1, ( + 'the end indices for sequence slice layer cannot ' + 'be a list having more than one element.') + ends = ends[0] + inputs.append(ends) + assert len(inputs) >= 2, ( + 'the sequence slice layer has at least two inputs.') + + super(SeqSliceLayer, self).__init__( + name, 'seq_slice', 0, inputs=inputs, **xargs) + + input_layer0 = self.get_input_layer(0) + size = input_layer0.size + self.set_layer_size(size) + + if len(inputs) == 3: + assert ( + self.get_input_layer(1).size == self.get_input_layer(2).size), ( + 'If start and end indices are both given to' + 'sequence slice layer, they should have the same width.') + elif len(inputs) == 2: + self.config.select_first = (starts is not None) + + @config_layer('sub_nested_seq') class SubNestedSequenceLayer(LayerBase): def __init__(self, name, inputs, selected_indices, bias=False, **xargs): diff --git a/python/paddle/trainer_config_helpers/evaluators.py b/python/paddle/trainer_config_helpers/evaluators.py index 44d52edfa7..57979db4de 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 1bc55c8696..a525ce71d0 100755 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -16,11 +16,13 @@ import functools import collections import inspect +import paddle.trainer.config_parser as cp from paddle.trainer.config_parser import * from .activations import LinearActivation, SigmoidActivation, TanhActivation, \ ReluActivation, IdentityActivation, SoftmaxActivation, BaseActivation from .evaluators import * -from .poolings import MaxPooling, AvgPooling, BasePoolingType +from .poolings import MaxPooling, AvgPooling, BasePoolingType, \ + CudnnAvgPooling, CudnnMaxPooling from .attrs import * from .default_decorators import * @@ -108,7 +110,8 @@ __all__ = [ 'sum_cost', 'rank_cost', 'lambda_cost', - 'huber_cost', + 'huber_regression_cost', + 'huber_classification_cost', 'block_expand_layer', 'maxout_layer', 'out_prod_layer', @@ -132,7 +135,9 @@ __all__ = [ 'sub_nested_seq_layer', 'clip_layer', 'slice_projection', + 'seq_slice_layer', 'kmax_sequence_score_layer', + 'scale_shift_layer', ] @@ -216,7 +221,8 @@ class LayerType(object): RANK_COST = 'rank-cost' LAMBDA_COST = 'lambda_cost' - HUBER = 'huber' + HUBER_REGRESSION = 'huber_regression' + HUBER_CLASSIFICATION = 'huber_classification' CROSS_ENTROPY = 'multi-class-cross-entropy' CROSS_ENTROPY_WITH_SELFNORM = 'multi_class_cross_entropy_with_selfnorm' SOFT_BIN_CLASS_CROSS_ENTROPY = 'soft_binary_class_cross_entropy' @@ -228,8 +234,10 @@ class LayerType(object): CROP_LAYER = 'crop' SUB_NESTED_SEQ = 'sub_nested_seq' CLIP_LAYER = 'clip' + SEQ_SLICE = 'seq_slice' KMAX_SEQ_SCORE = 'kmax_seq_score' + SCALE_SHIFT_LAYER = 'scale_shift' @staticmethod def is_layer_type(type_name): @@ -328,6 +336,14 @@ class LayerOutput(object): self.outputs = outputs self.reverse = reverse + @property + def width(self): + return cp.g_layer_map[self.full_name].width + + @property + def height(self): + return cp.g_layer_map[self.full_name].height + def set_input(self, input): """ Set the input for a memory layer. Can only be used for memory layer @@ -909,7 +925,13 @@ def data_layer(name, size, height=None, width=None, layer_attr=None): width=width, **ExtraLayerAttribute.to_kwargs(layer_attr)) - return LayerOutput(name, LayerType.DATA, size=size) + num_filters = None + if height is not None and width is not None: + num_filters = size / (width * height) + assert num_filters * width * height == size, \ + "size=%s width=%s height=%s" % (size, width, height) + + return LayerOutput(name, LayerType.DATA, size=size, num_filters=num_filters) @wrap_name_default("embedding") @@ -2322,6 +2344,7 @@ def img_conv_layer(input, groups=1, stride=1, padding=0, + dilation=1, bias_attr=None, param_attr=None, shared_biases=True, @@ -2329,6 +2352,7 @@ def img_conv_layer(input, filter_size_y=None, stride_y=None, padding_y=None, + dilation_y=None, trans=False, layer_type=None): """ @@ -2393,6 +2417,11 @@ def img_conv_layer(input, :type padding: int|tuple|list :param padding_y: The y dimension of the padding. :type padding_y: int + :param dilation: The x dimension of the dilation. Or input a tuple for two + image dimension + :type dilation: int|tuple|list + :param dilation_y: The y dimension of the dilation. + :type dilation_y: int :param bias_attr: Convolution bias attribute. None means default bias. False means no bias. :type bias_attr: ParameterAttribute|False @@ -2440,6 +2469,13 @@ def img_conv_layer(input, else: padding_y = padding + if dilation_y is None: + if isinstance(dilation, collections.Sequence): + assert len(dilation) == 2 + dilation, dilation_y = dilation + else: + dilation_y = dilation + if param_attr.attr.get('initial_smart'): # special initial for conv layers. init_w = (2.0 / (filter_size**2 * num_channels))**0.5 @@ -2449,6 +2485,8 @@ def img_conv_layer(input, param_attr.attr["initial_smart"] = False if layer_type: + if dilation > 1 or dilation_y > 1: + assert layer_type in ["cudnn_conv", "cudnn_convt"] if trans: assert layer_type in ["exconvt", "cudnn_convt"] else: @@ -2464,11 +2502,13 @@ def img_conv_layer(input, conv=Conv( filter_size=filter_size, padding=padding, + dilation=dilation, stride=stride, channels=num_channels, groups=groups, filter_size_y=filter_size_y, padding_y=padding_y, + dilation_y=dilation_y, stride_y=stride_y), **param_attr.attr), active_type=act.name, @@ -2574,11 +2614,14 @@ def img_pool_layer(input, elif isinstance(pool_type, AvgPooling): pool_type.name = 'avg' + assert type(pool_type) in [AvgPooling, MaxPooling, CudnnAvgPooling, + CudnnMaxPooling], \ + "only (Cudnn)AvgPooling, (Cudnn)MaxPooling are supported" + type_name = pool_type.name + '-projection' \ if ( isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \ else pool_type.name - pool_size_y = pool_size if pool_size_y is None else pool_size_y stride_y = stride if stride_y is None else stride_y padding_y = padding if padding_y is None else padding_y @@ -4202,8 +4245,7 @@ def conv_operator(img, num_channels = img.num_filters assert isinstance(filter, LayerOutput) - if filter.size is not None: - filter.size = filter_size * filter_size_y * num_filters * num_channels + assert filter.size is not None opCls = ConvTransOperator if trans else ConvOperator @@ -4914,7 +4956,6 @@ def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None): :return: LayerOutput object. :rtype: LayerOutput """ - assert input.layer_type == LayerType.CONV_LAYER assert isinstance(input.activation, LinearActivation) assert groups > 1 if num_channels is None: @@ -5605,16 +5646,77 @@ def sum_cost(input, name=None, layer_attr=None): @wrap_name_default() @layer_support() -def huber_cost(input, label, name=None, coeff=1.0, layer_attr=None): +def huber_regression_cost(input, + label, + name=None, + delta=1.0, + coeff=1.0, + layer_attr=None): """ - A loss layer for huber loss. + In statistics, the Huber loss is a loss function used in robust regression, + that is less sensitive to outliers in data than the squared error loss. + Given a prediction f(x), a label y and :math:`\delta`, the loss function + is defined as: + + .. math: + loss = 0.5*\left ( y-f(x) \right )^2, \left | y-f(x) \right |\leq \delta + loss = \delta \left | y-f(x) \right |-0.5\delta ^2, otherwise The example usage is: .. code-block:: python - cost = huber_cost(input=input_layer, - label=label_layer) + cost = huber_regression_cost(input=input_layer, label=label_layer) + + :param input: The first input layer. + :type input: LayerOutput. + :param label: The input label. + :type input: LayerOutput. + :param name: The name of this layers. It is not necessary. + :type name: None|basestring. + :param delta: The difference between the observed and predicted values. + :type delta: float. + :param coeff: The coefficient affects the gradient in the backward. + :type coeff: float. + :param layer_attr: Extra Layer Attribute. + :type layer_attr: ExtraLayerAttribute + :return: LayerOutput object. + :rtype: LayerOutput. + """ + assert isinstance(input, LayerOutput) + Layer( + name=name, + type=LayerType.HUBER_REGRESSION, + inputs=[input.name, label.name], + delta=delta, + coeff=coeff, + **ExtraLayerAttribute.to_kwargs(layer_attr)) + return LayerOutput( + name, LayerType.HUBER_REGRESSION, parents=[input, label], size=1) + + +@wrap_name_default() +@layer_support() +def huber_classification_cost(input, + label, + name=None, + coeff=1.0, + layer_attr=None): + """ + For classification purposes, a variant of the Huber loss called modified Huber + is sometimes used. Given a prediction f(x) (a real-valued classifier score) and + a true binary class label :math:`y\in \left \{-1, 1 \right \}`, the modified Huber + loss is defined as: + + .. math: + loss = \max \left ( 0, 1-yf(x) \right )^2, yf(x)\geq 1 + loss = -4yf(x), \text{otherwise} + + The example usage is: + + .. code-block:: python + + cost = huber_classification_cost(input=input_layer, label=label_layer) :param input: The first input layer. :type input: LayerOutput. @@ -5634,11 +5736,12 @@ def huber_cost(input, label, name=None, coeff=1.0, layer_attr=None): assert input.size == 1 Layer( name=name, - type=LayerType.HUBER, + type=LayerType.HUBER_CLASSIFICATION, inputs=[input.name, label.name], coeff=coeff, **ExtraLayerAttribute.to_kwargs(layer_attr)) - return LayerOutput(name, LayerType.HUBER, parents=[input, label], size=1) + return LayerOutput( + name, LayerType.HUBER_CLASSIFICATION, parents=[input, label], size=1) @wrap_name_default() @@ -6174,6 +6277,72 @@ def clip_layer(input, min, max, name=None): name, LayerType.CLIP_LAYER, parents=[input], size=input.size) +@wrap_name_default() +def seq_slice_layer(input, starts, ends, name=None): + """ + seq_slice_layer will return one or several sub-sequences from the + input sequence layer given start and end indices. + + - If only start indices are given, and end indices are set to None, + this layer slices the input sequence from the given start indices + to its end. + - If only end indices are given, and start indices are set to None, + this layer slices the input sequence from its beginning to the + given end indices. + - If start and end indices are both given, they should have the same + number of elements. + + If start or end indices contains more than one elements, the input sequence + will be sliced for multiple times. + + + .. code-block:: python + + seq_silce = seq_slice_layer(input=input_seq, + starts=start_pos, ends=end_pos) + + :param name: name of this layer. + :type name: basestring + :param input: input for this layer, it should be a sequence. + :type input: LayerOutput + :param starts: start indices to slice the input sequence. + :type starts: LayerOutput|None + :param ends: end indices to slice the input sequence. + :type ends: LayerOutput|None + :return: LayerOutput object. + :rtype: LayerOutput + + """ + + assert isinstance(input, LayerOutput), ( + 'The first input of seq_slice layer must be a PaddlePaddle layer.') + + if starts is not None: + assert isinstance(starts, LayerOutput), ( + 'The start indices for seq_slice layer ' + 'must be a PaddlePaddle layer.') + if ends is not None: + assert isinstance(ends, LayerOutput), ( + 'The end indices for seq_slice layer must be a PaddlePaddle layer.') + assert starts is not None or ends is not None, ( + 'start and end indices ' + 'cannot be set to None at the same time, at least one of ' + 'them should be given.') + if starts is not None and ends is not None: + assert starts.size == ends.size, ( + 'If start and end indices are both given to seq_slice_layer, ' + 'they should have the same width.') + + Layer( + name=name, + type=LayerType.SEQ_SLICE, + inputs=input.name, + starts=starts.name if starts is not None else None, + ends=ends.name if ends is not None else None) + return LayerOutput( + name, LayerType.SEQ_SLICE, parents=[input], size=input.size) + + @wrap_name_default() @layer_support() def kmax_sequence_score_layer(input, name=None, beam_size=1): @@ -6210,3 +6379,43 @@ def kmax_sequence_score_layer(input, name=None, beam_size=1): return LayerOutput( name, LayerType.KMAX_SEQ_SCORE, parents=[input], size=input.size) + + +@wrap_name_default("scale_shift") +@wrap_param_attr_default() +@wrap_bias_attr_default() +def scale_shift_layer(input, name=None, param_attr=None, bias_attr=None): + """ + A layer applies a linear transformation to each element in each row of + the input matrix. For each element, the layer first re-scale it and then + adds a bias to it. + + This layer is very like the SlopeInterceptLayer, except the scale and + bias are trainable. + + .. math:: + + y = w * x + b + + .. code-block:: python + + scale_shift = scale_shift_layer(input=input_layer, bias_attr=False) + + :param name: The Layer Name. + :type name: basestring + :param input: The input layer. + :type input: LayerOutput. + :param param_attr: The parameter attribute of scaling. + :type param_attr: ParameterAttribute + :param bias_attr: The parameter attribute of shifting. + :type bias_attr: ParameterAttribute + :return: LayerOutput object. + :rtype: LayerOutput + """ + Layer( + name=name, + type=LayerType.SCALE_SHIFT_LAYER, + inputs=Input(input.name, **param_attr.attr), + bias=ParamAttr.to_bias(bias_attr)) + return LayerOutput( + name, LayerType.SCALE_SHIFT_LAYER, parents=[input], size=input.size) diff --git a/python/paddle/trainer_config_helpers/tests/CMakeLists.txt b/python/paddle/trainer_config_helpers/tests/CMakeLists.txt index 6c860fd497..580aef935b 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 a61beb871a..1ca5c8a07e 100755 --- a/python/paddle/trainer_config_helpers/tests/configs/file_list.sh +++ b/python/paddle/trainer_config_helpers/tests/configs/file_list.sh @@ -8,6 +8,7 @@ test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops test_seq_concat_reshape test_pad test_smooth_l1 test_multiplex_layer test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_layer test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer -test_kmax_seq_socre_layer test_seq_select_layers) +test_kmax_seq_socre_layer test_seq_select_layers test_scale_shift_layer +test_seq_slice_layer) export whole_configs=(test_split_datasource) diff --git a/python/paddle/trainer_config_helpers/tests/configs/img_layers.py b/python/paddle/trainer_config_helpers/tests/configs/img_layers.py index 9fda16a540..01d31ef3fa 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/img_layers.py +++ b/python/paddle/trainer_config_helpers/tests/configs/img_layers.py @@ -12,6 +12,7 @@ img_conv = img_conv_layer( num_filters=64, filter_size=(32, 32), padding=(1, 1), + dilation=(1, 1), stride=(1, 1), act=LinearActivation()) img_bn = batch_norm_layer(input=img_conv, act=ReluActivation()) diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cost_layers.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cost_layers.protostr index 05847344be..55ab464ddf 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cost_layers.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_cost_layers.protostr @@ -167,6 +167,20 @@ layers { softmax_selfnorm_alpha: 0.1 coeff: 1.0 } +layers { + name: "__huber_regression_cost_0__" + type: "huber_regression" + size: 1 + active_type: "" + inputs { + input_layer_name: "input" + } + inputs { + input_layer_name: "labels" + } + coeff: 1.0 + delta: 1.0 +} layers { name: "huber_probs" type: "data" @@ -180,8 +194,8 @@ layers { active_type: "" } layers { - name: "__huber_cost_0__" - type: "huber" + name: "__huber_classification_cost_0__" + type: "huber_classification" size: 1 active_type: "" inputs { @@ -300,7 +314,8 @@ output_layer_names: "__rank_cost_0__" output_layer_names: "__lambda_cost_0__" output_layer_names: "__cross_entropy_0__" output_layer_names: "__cross_entropy_with_selfnorm_0__" -output_layer_names: "__huber_cost_0__" +output_layer_names: "__huber_regression_cost_0__" +output_layer_names: "__huber_classification_cost_0__" output_layer_names: "__multi_binary_label_cross_entropy_0__" output_layer_names: "__sum_cost_0__" output_layer_names: "__nce_layer_0__" @@ -324,9 +339,10 @@ sub_models { layer_names: "__lambda_cost_0__" layer_names: "__cross_entropy_0__" layer_names: "__cross_entropy_with_selfnorm_0__" + layer_names: "__huber_regression_cost_0__" layer_names: "huber_probs" layer_names: "huber_label" - layer_names: "__huber_cost_0__" + layer_names: "__huber_classification_cost_0__" layer_names: "__multi_binary_label_cross_entropy_0__" layer_names: "__sum_cost_0__" layer_names: "__nce_layer_0__" @@ -349,7 +365,8 @@ sub_models { output_layer_names: "__lambda_cost_0__" output_layer_names: "__cross_entropy_0__" output_layer_names: "__cross_entropy_with_selfnorm_0__" - output_layer_names: "__huber_cost_0__" + output_layer_names: "__huber_regression_cost_0__" + output_layer_names: "__huber_classification_cost_0__" output_layer_names: "__multi_binary_label_cross_entropy_0__" output_layer_names: "__sum_cost_0__" output_layer_names: "__nce_layer_0__" diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_kmax_seq_socre_layer.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_kmax_seq_socre_layer.protostr index 81bd71f68e..3d32220bfb 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_kmax_seq_socre_layer.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_kmax_seq_socre_layer.protostr @@ -1,12 +1,6 @@ type: "nn" layers { - name: "input" - type: "data" - size: 300 - active_type: "" -} -layers { - name: "data" + name: "input_seq" type: "data" size: 128 active_type: "" @@ -17,7 +11,7 @@ layers { size: 1 active_type: "exponential" inputs { - input_layer_name: "data" + input_layer_name: "input_seq" input_parameter_name: "___fc_layer_0__.w0" } bias_parameter_name: "___fc_layer_0__.wbias" @@ -51,15 +45,14 @@ parameters { initial_strategy: 0 initial_smart: false } -input_layer_names: "data" +input_layer_names: "input_seq" output_layer_names: "__kmax_sequence_score_layer_0__" sub_models { name: "root" - layer_names: "input" - layer_names: "data" + layer_names: "input_seq" layer_names: "__fc_layer_0__" layer_names: "__kmax_sequence_score_layer_0__" - input_layer_names: "data" + input_layer_names: "input_seq" 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_scale_shift_layer.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_scale_shift_layer.protostr new file mode 100644 index 0000000000..35ade126a2 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_scale_shift_layer.protostr @@ -0,0 +1,72 @@ +type: "nn" +layers { + name: "data" + type: "data" + size: 100 + active_type: "" +} +layers { + name: "__scale_shift_0__" + type: "scale_shift" + size: 100 + active_type: "" + inputs { + input_layer_name: "data" + input_parameter_name: "___scale_shift_0__.w0" + } +} +layers { + name: "__scale_shift_1__" + type: "scale_shift" + size: 100 + active_type: "" + inputs { + input_layer_name: "data" + input_parameter_name: "___scale_shift_1__.w0" + } + bias_parameter_name: "___scale_shift_1__.wbias" +} +parameters { + name: "___scale_shift_0__.w0" + size: 1 + initial_mean: 0.0 + initial_std: 1.0 + dims: 1 + dims: 1 + initial_strategy: 0 + initial_smart: true +} +parameters { + name: "___scale_shift_1__.w0" + size: 1 + initial_mean: 0.0 + initial_std: 1.0 + dims: 1 + dims: 1 + initial_strategy: 0 + initial_smart: true +} +parameters { + name: "___scale_shift_1__.wbias" + size: 1 + initial_mean: 0.0 + initial_std: 0.0 + dims: 1 + dims: 1 + initial_strategy: 0 + initial_smart: false +} +input_layer_names: "data" +output_layer_names: "__scale_shift_0__" +output_layer_names: "__scale_shift_1__" +sub_models { + name: "root" + layer_names: "data" + layer_names: "__scale_shift_0__" + layer_names: "__scale_shift_1__" + input_layer_names: "data" + output_layer_names: "__scale_shift_0__" + output_layer_names: "__scale_shift_1__" + is_recurrent_layer_group: false +} + diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_slice_layer.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_slice_layer.protostr new file mode 100644 index 0000000000..5b73d614fe --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_slice_layer.protostr @@ -0,0 +1,79 @@ +type: "nn" +layers { + name: "word" + type: "data" + size: 128 + active_type: "" +} +layers { + name: "starts" + type: "data" + size: 5 + active_type: "" +} +layers { + name: "ends" + type: "data" + size: 5 + active_type: "" +} +layers { + name: "__seq_slice_layer_0__" + type: "seq_slice" + size: 128 + active_type: "" + inputs { + input_layer_name: "word" + } + inputs { + input_layer_name: "starts" + } + inputs { + input_layer_name: "ends" + } +} +layers { + name: "__seq_slice_layer_1__" + type: "seq_slice" + size: 128 + active_type: "" + inputs { + input_layer_name: "word" + } + inputs { + input_layer_name: "starts" + } + select_first: true +} +layers { + name: "__seq_slice_layer_2__" + type: "seq_slice" + size: 128 + active_type: "" + inputs { + input_layer_name: "word" + } + inputs { + input_layer_name: "ends" + } + select_first: false +} +input_layer_names: "word" +output_layer_names: "__seq_slice_layer_0__" +output_layer_names: "__seq_slice_layer_1__" +output_layer_names: "__seq_slice_layer_2__" +sub_models { + name: "root" + layer_names: "word" + layer_names: "starts" + layer_names: "ends" + layer_names: "__seq_slice_layer_0__" + layer_names: "__seq_slice_layer_1__" + layer_names: "__seq_slice_layer_2__" + input_layer_names: "word" + output_layer_names: "__seq_slice_layer_0__" + output_layer_names: "__seq_slice_layer_1__" + output_layer_names: "__seq_slice_layer_2__" + is_recurrent_layer_group: false +} + diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_cost_layers.py b/python/paddle/trainer_config_helpers/tests/configs/test_cost_layers.py index d2a3b702a1..7ce375c708 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/test_cost_layers.py +++ b/python/paddle/trainer_config_helpers/tests/configs/test_cost_layers.py @@ -33,7 +33,9 @@ outputs( input=probs, label=xe_label), cross_entropy_with_selfnorm( input=probs, label=xe_label), - huber_cost( + huber_regression_cost( + input=seq_in, label=labels), + huber_classification_cost( input=data_layer( name='huber_probs', size=1), label=data_layer( diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_kmax_seq_socre_layer.py b/python/paddle/trainer_config_helpers/tests/configs/test_kmax_seq_socre_layer.py index d245c5a41c..48d0cd55da 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/test_kmax_seq_socre_layer.py +++ b/python/paddle/trainer_config_helpers/tests/configs/test_kmax_seq_socre_layer.py @@ -2,9 +2,7 @@ #coding=utf-8 from paddle.trainer_config_helpers import * -data = data_layer(name='input', size=300) - -data = data_layer(name="data", size=128) +data = data_layer(name="input_seq", size=128) scores = fc_layer(input=data, size=1, act=ExpActivation()) kmax_seq_id = kmax_sequence_score_layer(input=scores, beam_size=5) diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_scale_shift_layer.py b/python/paddle/trainer_config_helpers/tests/configs/test_scale_shift_layer.py new file mode 100644 index 0000000000..dd589116fa --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_scale_shift_layer.py @@ -0,0 +1,9 @@ +from paddle.trainer_config_helpers import * + +data = data_layer(name='data', size=100) + +scale = scale_shift_layer(input=data, bias_attr=False) + +scale_shift = scale_shift_layer(input=data) + +outputs(scale, scale_shift) diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_seq_slice_layer.py b/python/paddle/trainer_config_helpers/tests/configs/test_seq_slice_layer.py new file mode 100644 index 0000000000..510ad32208 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_seq_slice_layer.py @@ -0,0 +1,13 @@ +#!/usr/bin/env python +#coding=utf-8 +from paddle.trainer_config_helpers import * + +input_seq = data_layer("word", size=128) +starts = data_layer("starts", size=5) +ends = data_layer("ends", size=5) + +seq_slice1 = seq_slice_layer(input=input_seq, starts=starts, ends=ends) +seq_slice2 = seq_slice_layer(input=input_seq, starts=starts, ends=None) +seq_slice3 = seq_slice_layer(input=input_seq, starts=None, ends=ends) + +outputs(seq_slice1, seq_slice2, seq_slice3) diff --git a/python/paddle/v2/framework/.gitignore b/python/paddle/v2/framework/.gitignore new file mode 100644 index 0000000000..2ff540d576 --- /dev/null +++ b/python/paddle/v2/framework/.gitignore @@ -0,0 +1 @@ +proto diff --git a/python/paddle/v2/framework/op.py b/python/paddle/v2/framework/op.py index 7fd8b55a5d..6ac656321e 100644 --- a/python/paddle/v2/framework/op.py +++ b/python/paddle/v2/framework/op.py @@ -1,7 +1,5 @@ import paddle.v2.framework.core as core -import paddle.v2.framework.proto.op_proto_pb2 as op_proto_pb2 -import paddle.v2.framework.proto.op_desc_pb2 as op_desc_pb2 -import paddle.v2.framework.proto.attribute_pb2 as attribute_pb2 +import paddle.v2.framework.proto.framework_pb2 as framework_pb2 def get_all_op_protos(): @@ -12,22 +10,26 @@ def get_all_op_protos(): protostrs = core.get_all_op_protos() ret_values = [] for pbstr in protostrs: - op_proto = op_proto_pb2.OpProto.FromString(str(pbstr)) + op_proto = framework_pb2.OpProto.FromString(str(pbstr)) ret_values.append(op_proto) return ret_values +def is_str(s): + return isinstance(s, str) or isinstance(s, unicode) + + class OpDescCreationMethod(object): """ A Functor object to convert user input(use key word args) to OpDesc based on OpProto. - + :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): + if not isinstance(op_proto, framework_pb2.OpProto): raise TypeError("Argument should be OpProto") self.__op_proto__ = op_proto @@ -39,26 +41,34 @@ class OpDescCreationMethod(object): """ if len(args) != 0: raise ValueError("Only keyword arguments is supported by Paddle") - op_desc = op_desc_pb2.OpDesc() - - # Inputs - ipts, ipt_format, _ = OpDescCreationMethod.extract_input_or_output( - "input", kwargs, self.__op_proto__.inputs) - op_desc.inputs.extend(ipts) - if ipt_format is not None: - op_desc.attrs.extend([ipt_format]) - - # Outputs - outs, out_format, tmp_index = OpDescCreationMethod.extract_input_or_output( - "output", kwargs, self.__op_proto__.outputs) - op_desc.outputs.extend(outs) - if out_format is not None: - op_desc.attrs.extend([out_format]) - if len(tmp_index) != 0: - tmp_index_attr = op_desc.attrs.add() - tmp_index_attr.type = attribute_pb2.INTS - tmp_index_attr.name = "temporary_index" - tmp_index_attr.ints.extend(tmp_index) + 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 @@ -72,17 +82,17 @@ class OpDescCreationMethod(object): new_attr = op_desc.attrs.add() new_attr.name = attr.name new_attr.type = attr.type - if attr.type == attribute_pb2.INT: + if attr.type == framework_pb2.INT: new_attr.i = user_defined_attr - elif attr.type == attribute_pb2.FLOAT: + elif attr.type == framework_pb2.FLOAT: new_attr.f = user_defined_attr - elif attr.type == attribute_pb2.STRING: + elif attr.type == framework_pb2.STRING: new_attr.s = user_defined_attr - elif attr.type == attribute_pb2.INTS: + elif attr.type == framework_pb2.INTS: new_attr.ints.extend(user_defined_attr) - elif attr.type == attribute_pb2.FLOATS: + elif attr.type == framework_pb2.FLOATS: new_attr.floats.extend(user_defined_attr) - elif attr.type == attribute_pb2.STRINGS: + elif attr.type == framework_pb2.STRINGS: new_attr.strings.extend(user_defined_attr) else: raise NotImplementedError("Not support attribute type " + @@ -90,50 +100,6 @@ class OpDescCreationMethod(object): return op_desc - @staticmethod - def extract_input_or_output(in_out, kwargs, meta): - """ - Extract input variable names or output variable names from key-word - arguments, which base on VarProtos. - - :param in_out: "input" or "output" - :param kwargs: key-word arguments that user inputted. - :param meta: a list of VarProto - :return: The three object will be return. The variable names. The - input_format or output_format attribute(None if the input or output is - not multiple). The temporary variable index list. - """ - multiple = OpDescCreationMethod.any_is_true((m.multiple for m in meta)) - tmp_index = [] - retv = [] - if multiple: - var_format = op_desc_pb2.AttrDesc() - var_format.type = attribute_pb2.INTS - var_format.name = "%s_format" % in_out - var_format.ints.append(0) - - for var in meta: - var_name = var.name - - if var.temporary: - var_name = [core.var_names.temp()] - tmp_index.append(len(retv)) - else: - var_name = kwargs.get(var_name, []) - if not isinstance(var_name, list): - var_name = [var_name] - retv.extend(var_name) - var_format.ints.append(len(var_name) + var_format.ints[-1]) - return retv, var_format, tmp_index - else: - for var in meta: - if var.temporary: - retv.append(kwargs.get(var.name, core.var_names.temp())) - tmp_index.append(len(retv)) - else: - retv.append(kwargs.get(var.name, core.var_names.empty())) - return retv, None, tmp_index - @staticmethod def any_is_true(generator): """ @@ -146,13 +112,12 @@ class OpDescCreationMethod(object): class OpInfo(object): - def __init__(self, name, method, inputs, outputs, attrs, no_temp_outputs): + def __init__(self, name, method, inputs, outputs, attrs): self.name = name self.method = method self.inputs = inputs self.outputs = outputs self.attrs = attrs - self.no_temp_outputs = no_temp_outputs def create_op_creation_method(op_proto): @@ -170,10 +135,7 @@ def create_op_creation_method(op_proto): name=op_proto.type, inputs=[var.name for var in op_proto.inputs], outputs=[var.name for var in op_proto.outputs], - attrs=[attr.name for attr in op_proto.attrs], - no_temp_outputs=[ - var.name for var in op_proto.outputs if not var.temporary - ]) + attrs=[attr.name for attr in op_proto.attrs]) class OperatorFactory(object): @@ -214,8 +176,27 @@ class OperatorFactory(object): def get_op_attr_names(self, type): return self.get_op_info(type).attrs - def get_op_no_temp_output_names(self, type): - return self.get_op_info(type).no_temp_outputs + +class __RecurrentOp__(object): + __proto__ = None + type = 'recurrent_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 10659caa88..661ebd8964 100644 --- a/python/paddle/v2/framework/tests/CMakeLists.txt +++ b/python/paddle/v2/framework/tests/CMakeLists.txt @@ -1,6 +1,5 @@ py_test(test_net SRCS test_net.py) -py_test(test_fc_op SRCS test_fc_op.py) py_test(test_scope SRCS test_scope.py) py_test(test_tensor SRCS test_tensor.py) @@ -14,6 +13,8 @@ py_test(test_add_two_op SRCS test_add_two_op.py) py_test(test_sigmoid_op SRCS test_sigmoid_op.py) py_test(test_softmax_op SRCS test_softmax_op.py) py_test(test_cross_entropy_op SRCS test_cross_entropy_op.py) +py_test(test_gather_op SRCS test_gather_op.py) +py_test(test_scatter_op SRCS test_scatter_op.py) py_test(test_fill_zeros_like_op SRCS test_fill_zeros_like_op.py) py_test(gradient_checker SRCS gradient_checker.py) @@ -21,5 +22,13 @@ 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) +py_test(test_sgd_op SRCS test_sgd_op.py) +py_test(test_gradient_checker SRCS test_gradient_checker.py) +py_test(test_lookup_table SRCS test_lookup_table.py) +py_test(test_scale_and_identity_op SRCS test_scale_and_identity_op.py) +py_test(mnist SRCS mnist.py) diff --git a/python/paddle/v2/framework/tests/gradient_checker.py b/python/paddle/v2/framework/tests/gradient_checker.py index b73c4869d1..518f828bac 100644 --- a/python/paddle/v2/framework/tests/gradient_checker.py +++ b/python/paddle/v2/framework/tests/gradient_checker.py @@ -1,6 +1,7 @@ import unittest import numpy +import itertools import paddle.v2.framework.core as core from paddle.v2.framework.op import Operator @@ -8,6 +9,7 @@ __all__ = ['get_numeric_gradient'] def create_op(op_type): + # TODO need to set attrs kwargs = dict() for in_name in Operator.get_op_input_names(op_type): kwargs[in_name] = in_name @@ -21,12 +23,17 @@ def grad_var_name(var_name): return var_name + "@GRAD" +def empty_var_name(): + return "@EMPTY@" + + def get_numeric_gradient(op, input_values, output_name, input_to_check, delta=0.005, - local_scope=None): + local_scope=None, + in_place=False): """ Get Numeric Gradient for an operator's input. @@ -53,17 +60,19 @@ def get_numeric_gradient(op, tensor.set(input_values[var_name], core.CPUPlace()) # Create all output variable in local_scope - for output in op.outputs(): - if local_scope.find_var(output) is None: - local_scope.new_var(output).get_tensor() - + 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 output in op.outputs(): - local_scope.find_var(output).get_tensor().alloc_float(core.CPUPlace()) + 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(): @@ -73,34 +82,177 @@ def get_numeric_gradient(op, def product(dim): return reduce(lambda a, b: a * b, dim, 1) + def restore_inputs(): + for var_name in input_values: + tensor_ = local_scope.find_var(var_name).get_tensor() + tensor_.set(numpy.copy(input_values[var_name]), core.CPUPlace()) + + # 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): + if in_place: + restore_inputs() + # 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. + if in_place: + restore_inputs() x_neg = origin - delta tensor_to_check.set_float_element(i, x_neg) y_neg = get_output() - tensor_to_check.set_float_element(i, origin) # restore old value + # 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 __is_close(self, numeric_grads, scope, max_relative_error): - for name in numeric_grads: - op_grad = numpy.array( - scope.find_var(grad_var_name(name)).get_tensor()) - is_close = numpy.allclose( - numeric_grads[name], op_grad, rtol=max_relative_error, atol=100) - if not is_close: - return False - return True + def __get_gradient(self, forward_op, backward_op, input_value, grad_names, + place): + """Get the input gradients after running forward and backward operators + on the given places. + + :param forward_op: forward operator + :type forward_op: Operator + :param backward_op: backward operator + :type backward_op: Operator + :param input_value: input values. + :type input_value: dict{string:numpy.array} + :param grad_names: the names of returned input gradients. + :type input_value: a list of string + :param place: the device type. + :type place: CPUPlace or GPUPlace + :return: the input grdients of given grad_names. + :rtype: a list of numpy.array + """ + scope = core.Scope() + ctx = core.DeviceContext.create(place) + + 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]] + + # create input var and set value + for name, value in input_value.iteritems(): + if name not in in_names: + raise ValueError(name + "does not exist in Op's inputs.") + var = scope.new_var(name).get_tensor() + var.set_dims(value.shape) + var.set(value, place) + + # run forward op + for out_name in out_names: + scope.new_var(out_name) + forward_op.infer_shape(scope) + forward_op.run(scope, ctx) + + # set output var's shape + # set output 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 = numpy.ones(out_tensor.shape(), dtype=numpy.float32) + grad_tensor.set(data, place) + + # run backward op + backward_outs = backward_op.outputs() + backward_names = [ + item for key in backward_outs for item in backward_outs[key] + ] + for name in backward_names: + scope.new_var(name) + + backward_op.infer_shape(scope) + backward_op.run(scope, ctx) + + outs = [ + numpy.array(scope.find_var(name).get_tensor()) + for name in grad_names + ] + return outs + + def compare_grad(self, forward_op, input_value, no_grad_set=None): + """ Compare the input gradients between CPU and GPU for the given forward + operator. + + :param forward_op: forward operator + :type forward_op: Operator + :param input_value: input values. + :type input_value: dict{string:numpy.array} + :param no_grad_set: the set of variables names without gradients. + :type no_grad_set: a set of string + :raises: AssertionError, there is different gradient value. + """ + if no_grad_set is None: + no_grad_set = set() + backward_op = core.Operator.backward(forward_op, no_grad_set) + # return if not compile with GPU or not implementing GPU kernel + if not (core.is_compile_gpu() and backward_op.support_gpu()): + return + + outputs = backward_op.outputs() + out_names = [item for k in outputs for item in outputs[k]] + out_names = filter(lambda x: x != empty_var_name(), out_names) + cpu_grads = self.__get_gradient(forward_op, backward_op, input_value, + out_names, core.CPUPlace()) + gpu_grads = self.__get_gradient(forward_op, backward_op, input_value, + out_names, core.GPUPlace(0)) + + for c_grad, g_grad, name in itertools.izip(cpu_grads, gpu_grads, + out_names): + self.assertTrue( + numpy.allclose( + c_grad, g_grad, atol=1e-4), + "output name: " + name + " has diff") + + def __assert_is_close(self, numeric_grads, analytic_grads, names, + max_relative_error, msg_prefix): + """Use relative error for the comparison. + + :param numeric_grads: the numerical graidents. + :type numeric_grads: a list of numpy.array + :param analytic_grads: the analytical graidents. + :type analytic_grads: a list of numpy.array + :param name: the names of gradients, used to print for debug. + :type names: a list of string + :param msg_prefix: string info, used to print for debug. + :type msf_prefix: string + """ + for a, b, name in itertools.izip(numeric_grads, analytic_grads, names): + 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, @@ -109,13 +261,14 @@ class GradientChecker(unittest.TestCase): output_name, no_grad_set=None, only_cpu=False, + in_place=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 output_name: the output variable name of forward network. :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. @@ -124,113 +277,33 @@ class GradientChecker(unittest.TestCase): if no_grad_set is None: no_grad_set = set() - tmp_outs = forward_op.temp_outputs() - no_tmp_out = filter(lambda name: name not in tmp_outs, - forward_op.outputs()) + no_tmp_out = forward_op.no_intermediate_outputs() if len(no_tmp_out) != 1: raise ValueError("non temp out_names should be 1") - in_names = forward_op.inputs() + inputs = forward_op.inputs() + in_names = [item for k in inputs for item in inputs[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) 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 numerical gradients + numeric_grads = [ + get_numeric_gradient( + forward_op, input_vars, output_name, name, in_place=in_place) + for name in inputs_to_check + ] - # get operator gradient according to different device + check_names = [grad_var_name(name) for name in inputs_to_check] for place in places: - scope = core.Scope() - ctx = core.DeviceContext.create(place) - - # create input var and set value - for name, value in input_vars.iteritems(): - if name not in in_names: - raise ValueError(name + " not in op.inputs_") - var = scope.new_var(name).get_tensor() - var.set_dims(value.shape) - var.set(value, place) - - # create output var - for out_name in forward_op.outputs(): - scope.new_var(out_name).get_tensor() - - # infer the shape of output var and compute/set value of output var - forward_op.infer_shape(scope) - forward_op.run(scope, ctx) - - # create output grad var - # set shape as the output var - # set value of this grad to ones - for name in forward_op.outputs(): - out_tensor = scope.find_var(name).get_tensor() - grad_tensor = scope.new_var(grad_var_name(name)).get_tensor() - grad_tensor.set_dims(out_tensor.shape()) - data = 1.0 * numpy.ones(out_tensor.shape()) - grad_tensor.set(data, place) - - # create input grad var - for name in backward_op.outputs(): - scope.new_var(name).get_tensor() - - # infer the shape of input gradient var and compute/set it's value - # with backward op - backward_op.infer_shape(scope) - backward_op.run(scope, ctx) - - if isinstance(place, core.CPUPlace): - msg = "CPU kernel gradient is not close to numeric gradient" - else: - if isinstance(place, core.GPUPlace): - msg = "GPU kernel gradient is not close to numeric gradient" - else: - raise ValueError("unknown place " + type(place)) - self.assertTrue( - self.__is_close(numeric_grad, scope, max_relative_error), msg) - - -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() + # get analytical gradients according to different device + analytic_grads = self.__get_gradient(forward_op, backward_op, + input_vars, check_names, place) + self.__assert_is_close(numeric_grads, analytic_grads, check_names, + max_relative_error, + "Gradient Check On %s" % str(place)) diff --git a/python/paddle/v2/framework/tests/mnist.py b/python/paddle/v2/framework/tests/mnist.py new file mode 100644 index 0000000000..a68f302f9c --- /dev/null +++ b/python/paddle/v2/framework/tests/mnist.py @@ -0,0 +1,249 @@ +import paddle.v2.framework.core as core +from paddle.v2.framework.op import Operator +import numpy +import paddle.v2 as paddle + +BATCH_SIZE = 100 + +scope = core.Scope() +place = core.CPUPlace() +# if you want to test GPU training, you can use gpu place +# place = core.GPUPlace(0) +dev_ctx = core.DeviceContext.create(place) + +init_net = core.Net.create() +forward_net = core.Net.create() +backward_net = None +optimize_net = core.Net.create() + + +def atomic_id(): + id = 0 + while True: + yield id + id += 1 + + +uniq_id = atomic_id().next + + +def data_layer(name, dims): + var = scope.new_var(name) + tensor = var.get_tensor() + tensor.set_dims(dims) # 1 is batch size holder. + return name + + +def feed_data(name, data): + assert isinstance(data, numpy.ndarray) + tensor = scope.find_var(name).get_tensor() + tensor.set_dims(data.shape) + if data.dtype == numpy.dtype('int32'): + tensor.alloc_int(place) + elif data.dtype == numpy.dtype('float32'): + tensor.alloc_float(place) + else: + raise ValueError("data type not supported") + tensor.set(data, place) + + +def grad_var_name(var_name): + return var_name + "@GRAD" + + +def sgd_optimizer(net, param_name, learning_rate=0.005): + grad_name = grad_var_name(param_name) + optimize_op = Operator( + "sgd", + param=param_name, + grad=grad_name, + param_out=param_name, + learning_rate=learning_rate) + net.append_op(optimize_op) + + +# should use operator and add these to the init_network +def init_param(net, param_name, dims): + scope.new_var(param_name) + op = Operator( + "uniform_random", Out=param_name, dims=dims, min=-0.5, max=0.5, seed=10) + op.infer_shape(scope) + net.append_op(op) + + +# fc_layer +def fc_layer(net, input, size, act="softmax", bias=True, param=None, name=None): + """ + Add a fc layer to net + + :param input: input variable name. + :type input: str + :param size: fully connected layer size. + :param act: activation name + :param param: parameter attribute, used for initialize parameters. + :param bias: bias attribute. False will not have a bias. + :param name: the name of fc layer. If not set, model will generate a + readable name + :return: output variable name. + """ + if name is None: + name = 'fc_%d' % uniq_id() + if not isinstance(name, str): + raise ValueError("name should be string") + + input_dims = scope.find_var(input).get_tensor().get_dims() + + w_name = param or name + ".w" + init_param(net=init_net, param_name=w_name, dims=[input_dims[1], size]) + sgd_optimizer(net=optimize_net, param_name=w_name, learning_rate=0.01) + + pre_activation = name + ".mul.out" + scope.new_var(pre_activation) + mul_op = Operator("mul", X=input, Y=w_name, Out=pre_activation) + net.append_op(mul_op) + + # create bias variable if needed + if bias: + bias_name = name + ".b" + init_param(net=init_net, param_name=bias_name, dims=[size]) + sgd_optimizer( + net=optimize_net, param_name=bias_name, learning_rate=0.001) + bias_out = name + ".rowwise_add.out" + scope.new_var(bias_out) + rowwise_append_op = Operator( + "rowwise_add", X=pre_activation, b=bias_name, Out=bias_out) + net.append_op(rowwise_append_op) + pre_activation = bias_out + + activation_op = Operator(act, X=pre_activation, Y=name) + net.append_op(activation_op) + scope.new_var(name) + net.infer_shape(scope) + return name + + +def cross_entropy_layer(net, input, label): + cost_name = 'cross_entropy_%d' % uniq_id() + cross_entropy_op = Operator( + "onehot_cross_entropy", X=input, label=label, Y=cost_name) + net.append_op(cross_entropy_op) + scope.new_var(cost_name) + net.infer_shape(scope) + return cost_name + + +def create_backward_net(forward_net): + net = core.Operator.backward(forward_net, set()) + for input in net.inputs()["all"]: + var = scope.new_var(input) + var.get_tensor() + for output in net.outputs()["all"]: + var = scope.new_var(output) + var.get_tensor() + return net + + +def debug_print_op(op): + print("===============" + op.type() + "==============") + print("***inputs:***") + for input in op.inputs()["all"]: + print input, scope.find_var(input).get_tensor().get_dims() + print("\n***outputs:***") + for output in op.outputs()["all"]: + print output, scope.find_var(output).get_tensor().get_dims() + print("") + print("") + + +def set_cost(cost): + cost_shape = numpy.array(scope.find_var(cost).get_tensor()).shape + cost_grad = \ + scope.find_var(grad_var_name(cost)).get_tensor() + cost_grad.set_dims(cost_shape) + cost_grad.alloc_float(place) + cost_grad.set(numpy.ones(cost_shape).astype("float32"), place) + + +def get_cost_mean(cost): + cost_data = numpy.array(scope.find_var(cost).get_tensor()) + return cost_data.sum() / len(cost_data) + + +def error_rate(predict, label): + predict_var = numpy.array(scope.find_var(predict).get_tensor()).argmax( + axis=1) + label = numpy.array(scope.find_var(label).get_tensor()) + error_num = numpy.sum(predict_var != label) + return error_num / float(len(label)) + + +images = data_layer(name='pixel', dims=[BATCH_SIZE, 784]) +labels = data_layer(name='label', dims=[BATCH_SIZE]) +fc1 = fc_layer(net=forward_net, input=images, size=100, act="sigmoid") +fc2 = fc_layer(net=forward_net, input=fc1, size=100, act="sigmoid") +predict = fc_layer(net=forward_net, input=fc2, size=10, act="softmax") +cost = cross_entropy_layer(net=forward_net, input=predict, label=labels) + +init_net.complete_add_op(True) +forward_net.complete_add_op(True) +backward_net = create_backward_net(forward_net) +optimize_net.complete_add_op(True) + +print(init_net) +print(forward_net) +print(backward_net) +print(optimize_net) + +debug_print_op(forward_net) +debug_print_op(backward_net) +debug_print_op(optimize_net) + +train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=8192), + batch_size=BATCH_SIZE) + + +def test(cost_name): + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=BATCH_SIZE) + cost = [] + error = [] + for data in test_reader(): + image_data = numpy.array(map(lambda x: x[0], data)).astype("float32") + label_data = numpy.array(map(lambda x: x[1], data)).astype("int32") + feed_data(images, image_data) + feed_data(labels, label_data) + + forward_net.infer_shape(scope) + forward_net.run(scope, dev_ctx) + cost.append(get_cost_mean(cost_name)) + error.append(error_rate(predict, "label")) + print("cost=" + str(sum(cost) / float(len(cost))) + " error_rate=" + str( + sum(error) / float(len(error)))) + + +PASS_NUM = 1 + +init_net.run(scope, dev_ctx) +for pass_id in range(PASS_NUM): + batch_id = 0 + + for data in train_reader(): + image_data = numpy.array(map(lambda x: x[0], data)).astype("float32") + label_data = numpy.array(map(lambda x: x[1], data)).astype("int32") + feed_data(images, image_data) + feed_data(labels, label_data) + + forward_net.infer_shape(scope) + forward_net.run(scope, dev_ctx) + set_cost(cost) + backward_net.infer_shape(scope) + backward_net.run(scope, dev_ctx) + + optimize_net.run(scope, dev_ctx) + if batch_id % 100 == 0: + print("pass[" + str(pass_id) + "] batch_id[" + str(batch_id) + "]") + test(cost) + + batch_id = batch_id + 1 diff --git a/python/paddle/v2/framework/tests/op_test_util.py b/python/paddle/v2/framework/tests/op_test_util.py index dd65e0f2dc..3bc05a0fec 100644 --- a/python/paddle/v2/framework/tests/op_test_util.py +++ b/python/paddle/v2/framework/tests/op_test_util.py @@ -64,7 +64,8 @@ class OpTestMeta(type): actual = numpy.array(scope.find_var(out_name).get_tensor()) expect = self.outputs[out_name] self.assertTrue( - numpy.allclose(actual, expect), + numpy.allclose( + actual, expect, atol=1e-05), "output name: " + out_name + "has diff") obj.test_all = test_all 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 c023783064..0def484edd 100644 --- a/python/paddle/v2/framework/tests/test_add_two_op.py +++ b/python/paddle/v2/framework/tests/test_add_two_op.py @@ -19,14 +19,5 @@ class TestAddOp(unittest.TestCase): self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']} -class TestAddGradOp(unittest.TestCase): - def test_add_grad(self): - op = Operator('add_two', X="X", Y="Y", Out="Out") - backward_op = core.Operator.backward(op, set()) - self.assertEqual(backward_op.type(), "add_two_grad") - expected = '''Op(add_two_grad), inputs:(X, Y, Out, Out@GRAD), outputs:(X@GRAD, Y@GRAD).''' - self.assertEqual(expected, str(backward_op)) - - if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/framework/tests/test_cross_entropy_op.py b/python/paddle/v2/framework/tests/test_cross_entropy_op.py index 4815192e25..d4277f2a42 100644 --- a/python/paddle/v2/framework/tests/test_cross_entropy_op.py +++ b/python/paddle/v2/framework/tests/test_cross_entropy_op.py @@ -8,9 +8,8 @@ class TestCrossEntropy(unittest.TestCase): __metaclass__ = OpTestMeta def setUp(self): - # TODO this unit test is not passed self.type = "onehot_cross_entropy" - batch_size = 100 + batch_size = 30 class_num = 10 X = numpy.random.random((batch_size, class_num)).astype("float32") label = 5 * numpy.ones(batch_size).astype("int32") @@ -22,9 +21,9 @@ class TestCrossEntropy(unittest.TestCase): class CrossEntropyGradOpTest(GradientChecker): - def test_softmax_grad(self): + def test_check_grad(self): op = create_op("onehot_cross_entropy") - batch_size = 100 + batch_size = 30 class_num = 10 inputs = { "X": numpy.random.uniform( 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 e24435839d..0000000000 --- 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 -from paddle.v2.framework.op import Operator - - -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 = Operator("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_gather_op.py b/python/paddle/v2/framework/tests/test_gather_op.py new file mode 100644 index 0000000000..e3de3fd0a1 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_gather_op.py @@ -0,0 +1,31 @@ +import unittest +from op_test_util import OpTestMeta +from gradient_checker import GradientChecker, create_op +import numpy +import paddle.v2.framework.core as core +from paddle.v2.framework.op import Operator + + +class TestGatherOp(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "gather" + xnp = numpy.random.random((10, 20)).astype("float32") + self.inputs = { + 'X': xnp, + 'Index': numpy.array([1, 3, 5]).astype("int32") + } + self.outputs = {'Out': self.inputs['X'][self.inputs['Index']]} + + +class TestGatherGradOp(GradientChecker): + def test_gather_grad(self): + op = create_op("gather") + xnp = numpy.random.random((10, 20)).astype("float32") + inputs = {'X': xnp, 'Index': numpy.array([1, 3, 5]).astype("int32")} + self.check_grad(op, inputs, set("X"), "Out") + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_gaussian_random_op.py b/python/paddle/v2/framework/tests/test_gaussian_random_op.py new file mode 100644 index 0000000000..f95ed70b58 --- /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_gradient_checker.py b/python/paddle/v2/framework/tests/test_gradient_checker.py new file mode 100644 index 0000000000..e0b3151208 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_gradient_checker.py @@ -0,0 +1,43 @@ +import unittest +import numpy +from paddle.v2.framework.op import Operator +from gradient_checker import GradientChecker +from gradient_checker import get_numeric_gradient + + +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-4) + + 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) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_lookup_table.py b/python/paddle/v2/framework/tests/test_lookup_table.py new file mode 100644 index 0000000000..19eb464baa --- /dev/null +++ b/python/paddle/v2/framework/tests/test_lookup_table.py @@ -0,0 +1,31 @@ +import unittest +import numpy as np +from op_test_util import OpTestMeta +from gradient_checker import GradientChecker, create_op + + +class TestSigmoidOp(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = 'lookup_table' + table = np.random.random((17, 31)).astype('float32') + ids = np.random.randint(0, 17, 4).astype('int32') + self.inputs = {'W': table, 'Ids': ids} + self.outputs = {'Out': table[ids]} + + +class TestSigmoidGradOp(GradientChecker): + def test_grad(self): + op = create_op('lookup_table') + table = np.random.random((17, 31)).astype('float32') + ids = np.random.randint(0, 17, 4).astype('int32') + inputs = {'W': table, 'Ids': ids} + # comapre gradients + self.compare_grad(op, inputs, set(['Ids'])) + # check gradients + self.check_grad(op, inputs, set('W'), 'Out') + + +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 b5d52b9056..f32b3160d6 100644 --- a/python/paddle/v2/framework/tests/test_mean_op.py +++ b/python/paddle/v2/framework/tests/test_mean_op.py @@ -1,5 +1,6 @@ import unittest from op_test_util import OpTestMeta +from gradient_checker import GradientChecker, create_op import numpy as np @@ -12,5 +13,12 @@ class TestMeanOp(unittest.TestCase): self.outputs = {'Out': np.mean(self.inputs['X'])} +class MeanGradOpTest(GradientChecker): + def test_normal(self): + op = create_op("mean") + inputs = {"X": np.random.random((10, 10)).astype("float32")} + self.check_grad(op, inputs, set("X"), "Out") + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/framework/tests/test_minus_op.py b/python/paddle/v2/framework/tests/test_minus_op.py new file mode 100644 index 0000000000..5abdd4a69b --- /dev/null +++ b/python/paddle/v2/framework/tests/test_minus_op.py @@ -0,0 +1,30 @@ +import unittest +import numpy as np +from gradient_checker import GradientChecker, create_op +from op_test_util import OpTestMeta + + +class MinusOpTest(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "minus" + self.inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'Y': np.random.random((32, 84)).astype("float32") + } + self.outputs = {'Out': (self.inputs['X'] - self.inputs['Y'])} + + +class MinusGradTest(GradientChecker): + def test_left(self): + op = create_op("minus") + inputs = { + "X": np.random.random((10, 10)).astype("float32"), + "Y": np.random.random((10, 10)).astype("float32") + } + self.check_grad(op, inputs, ["X", 'Y'], "Out") + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_mul_op.py b/python/paddle/v2/framework/tests/test_mul_op.py index ec0ac99156..ee0d81a64e 100644 --- a/python/paddle/v2/framework/tests/test_mul_op.py +++ b/python/paddle/v2/framework/tests/test_mul_op.py @@ -1,6 +1,7 @@ import unittest -from op_test_util import OpTestMeta import numpy as np +from gradient_checker import GradientChecker, create_op +from op_test_util import OpTestMeta class TestMulOp(unittest.TestCase): @@ -15,5 +16,19 @@ class TestMulOp(unittest.TestCase): self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} +class MulGradOpTest(GradientChecker): + def test_mul(self): + op = create_op("mul") + inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'Y': np.random.random((84, 100)).astype("float32") + } + # mul op will enlarge the relative error + self.check_grad( + op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.5) + + +# TODO(dzh,qijun) : mulgrad test case need transpose feature of blas library + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/framework/tests/test_net.py b/python/paddle/v2/framework/tests/test_net.py index b30896553d..9339cf28da 100644 --- a/python/paddle/v2/framework/tests/test_net.py +++ b/python/paddle/v2/framework/tests/test_net.py @@ -3,25 +3,34 @@ from paddle.v2.framework.op import Operator import unittest +def fc(X, W, Y): + ret_v = core.Net.create() + + ret_v.append_op(Operator("mul", X="X", Y="W", Out="pre_activation")) + ret_v.append_op(Operator("sigmoid", X="pre_activation", Y=Y)) + ret_v.complete_add_op(True) + return ret_v + + class TestNet(unittest.TestCase): def test_net_all(self): net = core.Net.create() op1 = Operator("add_two", X="X", Y="Y", Out="Out") - net.add_op(op1) + net.append_op(op1) net2 = core.Net.create() - net2.add_op(Operator("fc", X="X", W="w", Y="fc.out")) + net2.append_op(fc(X="X", W="w", Y="fc.out")) net2.complete_add_op(True) - net.add_op(net2) + net.append_op(net2) net.complete_add_op(True) expected = ''' -Op(plain_net), inputs:(@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_operator.py b/python/paddle/v2/framework/tests/test_operator.py index 4f164e1a69..1abc4eeb57 100644 --- a/python/paddle/v2/framework/tests/test_operator.py +++ b/python/paddle/v2/framework/tests/test_operator.py @@ -1,9 +1,7 @@ import unittest import paddle.v2.framework.op as op import paddle.v2.framework.core as core -import paddle.v2.framework.proto.op_proto_pb2 as op_proto_pb2 -import paddle.v2.framework.proto.op_desc_pb2 as op_desc_pb2 -import paddle.v2.framework.proto.attribute_pb2 as attribute_pb2 +import paddle.v2.framework.proto.framework_pb2 as framework_pb2 class TestGetAllProtos(unittest.TestCase): @@ -17,7 +15,7 @@ class TestGetAllProtos(unittest.TestCase): class TestOpDescCreationMethod(unittest.TestCase): def test_plain_input_output(self): - op_proto = op_proto_pb2.OpProto() + op_proto = framework_pb2.OpProto() op_proto.type = "test" ipt = op_proto.inputs.add() ipt.name = "X" @@ -37,25 +35,32 @@ class TestOpDescCreationMethod(unittest.TestCase): method = op.OpDescCreationMethod(op_proto) output = method(X="a", Y="b", Z="c") - - expected = op_desc_pb2.OpDesc() + expected = framework_pb2.OpDesc() expected.type = "test" - expected.inputs.extend(["a", "b"]) - expected.outputs.append("c") + ipt_0 = expected.inputs.add() + ipt_0.parameter = "X" + ipt_0.arguments.extend(["a"]) + ipt_1 = expected.inputs.add() + ipt_1.parameter = 'Y' + ipt_1.arguments.extend(['b']) + opt = expected.outputs.add() + opt.parameter = "Z" + opt.arguments.extend(["c"]) + self.assertEqual(expected, output) def test_multiple_input_plain_output(self): - op_proto = op_proto_pb2.OpProto() + op_proto = framework_pb2.OpProto() op_proto.type = "fc" ipt = op_proto.inputs.add() ipt.name = "X" ipt.comment = "" - ipt.multiple = True + ipt.duplicable = True ipt = op_proto.inputs.add() ipt.name = "W" ipt.comment = "" - ipt.multiple = True + ipt.duplicable = True ipt = op_proto.inputs.add() ipt.name = "b" @@ -70,30 +75,50 @@ class TestOpDescCreationMethod(unittest.TestCase): method = op.OpDescCreationMethod(op_proto) generated1 = method(X="x", W="w", b="b", Y="y") - expected1 = op_desc_pb2.OpDesc() - expected1.inputs.extend(['x', 'w', 'b']) - expected1.outputs.extend(['y']) + expected1 = framework_pb2.OpDesc() + tmp = expected1.inputs.add() + tmp.parameter = "X" + tmp.arguments.extend(['x']) + + tmp = expected1.inputs.add() + tmp.parameter = 'W' + tmp.arguments.extend(['w']) + + tmp = expected1.inputs.add() + tmp.parameter = 'b' + tmp.arguments.extend(['b']) + + tmp = expected1.outputs.add() + tmp.parameter = 'Y' + tmp.arguments.extend(['y']) expected1.type = 'fc' - attr = expected1.attrs.add() - attr.name = 'input_format' - attr.type = attribute_pb2.INTS - attr.ints.extend([0, 1, 2, 3]) self.assertEqual(expected1, generated1) generated2 = method( X=['x1', 'x2', 'x3'], b='b', W=['w1', 'w2', 'w3'], Y='y') - expected2 = op_desc_pb2.OpDesc() - expected2.inputs.extend(['x1', 'x2', 'x3', 'w1', 'w2', 'w3', 'b']) - expected2.outputs.extend(['y']) + expected2 = framework_pb2.OpDesc() + + tmp = expected2.inputs.add() + tmp.parameter = "X" + tmp.arguments.extend(['x1', 'x2', 'x3']) + + tmp = expected2.inputs.add() + tmp.parameter = 'W' + tmp.arguments.extend(['w1', 'w2', 'w3']) + + tmp = expected2.inputs.add() + tmp.parameter = 'b' + tmp.arguments.extend(['b']) + + tmp = expected2.outputs.add() + tmp.parameter = 'Y' + tmp.arguments.extend(['y']) + expected2.type = 'fc' - attr = expected2.attrs.add() - attr.name = 'input_format' - attr.type = attribute_pb2.INTS - attr.ints.extend([0, 3, 6, 7]) self.assertEqual(expected2, generated2) def test_attrs(self): - op_proto = op_proto_pb2.OpProto() + op_proto = framework_pb2.OpProto() op_proto.type = "test" ipt = op_proto.inputs.add() ipt.name = 'X' @@ -105,12 +130,12 @@ class TestOpDescCreationMethod(unittest.TestCase): attr.comment = "" attr.type = type - __add_attr__("int_attr", attribute_pb2.INT) - __add_attr__("float_attr", attribute_pb2.FLOAT) - __add_attr__("string_attr", attribute_pb2.STRING) - __add_attr__("ints_attr", attribute_pb2.INTS) - __add_attr__("floats_attr", attribute_pb2.FLOATS) - __add_attr__("strings_attr", attribute_pb2.STRINGS) + __add_attr__("int_attr", framework_pb2.INT) + __add_attr__("float_attr", framework_pb2.FLOAT) + __add_attr__("string_attr", framework_pb2.STRING) + __add_attr__("ints_attr", framework_pb2.INTS) + __add_attr__("floats_attr", framework_pb2.FLOATS) + __add_attr__("strings_attr", framework_pb2.STRINGS) op_proto.comment = "" self.assertTrue(op_proto.IsInitialized()) @@ -126,76 +151,52 @@ class TestOpDescCreationMethod(unittest.TestCase): floats_attr=[0.2, 3.2, 4.5], strings_attr=["a", "b", "c"]) - expected = op_desc_pb2.OpDesc() + expected = framework_pb2.OpDesc() expected.type = "test" - expected.inputs.extend(['a']) + + ipt = expected.inputs.add() + ipt.parameter = "X" + ipt.arguments.extend(['a']) + attr = expected.attrs.add() attr.name = "int_attr" - attr.type = attribute_pb2.INT + attr.type = framework_pb2.INT attr.i = 10 attr = expected.attrs.add() attr.name = "float_attr" - attr.type = attribute_pb2.FLOAT + attr.type = framework_pb2.FLOAT attr.f = 3.2 attr = expected.attrs.add() attr.name = "string_attr" - attr.type = attribute_pb2.STRING + attr.type = framework_pb2.STRING attr.s = "test_str" attr = expected.attrs.add() attr.name = "ints_attr" - attr.type = attribute_pb2.INTS + attr.type = framework_pb2.INTS attr.ints.extend([0, 1, 2, 3, 4]) attr = expected.attrs.add() attr.name = 'floats_attr' - attr.type = attribute_pb2.FLOATS + attr.type = framework_pb2.FLOATS attr.floats.extend([0.2, 3.2, 4.5]) attr = expected.attrs.add() attr.name = 'strings_attr' - attr.type = attribute_pb2.STRINGS + attr.type = framework_pb2.STRINGS attr.strings.extend(['a', 'b', 'c']) self.assertEqual(expected, generated) - def test_input_temporary_output(self): - op_proto = op_proto_pb2.OpProto() - op_proto.type = "test" - out = op_proto.outputs.add() - out.name = "OUT" - out.comment = "" - - out = op_proto.outputs.add() - out.name = "TMP" - out.comment = "" - out.temporary = True - - out = op_proto.outputs.add() - out.name = "OUT2" - out.comment = "" - op_proto.comment = "" - - method = op.OpDescCreationMethod(op_proto) - generated = method(OUT="a", OUT2="b") - desc = op_desc_pb2.OpDesc() - desc.outputs.extend(["a", core.var_names.temp(), "b"]) - desc.type = "test" - attr = desc.attrs.add() - attr.name = "temporary_index" - attr.type = attribute_pb2.INTS - attr.ints.append(2) - self.assertEqual(generated, desc) - class TestOpCreations(unittest.TestCase): def test_all(self): add_op = op.Operator("add_two", X="a", Y="b", Out="z") self.assertIsNotNone(add_op) # Invoke C++ DebugString() - self.assertEqual('Op(add_two), inputs:(a, b), outputs:(z).', + self.assertEqual('Op(add_two), inputs:{X[a], Y[b]}, outputs:{Out[z]}.', str(add_op)) diff --git a/python/paddle/v2/framework/tests/test_protobuf.py b/python/paddle/v2/framework/tests/test_protobuf.py index 69e98e2f25..848a396b3b 100644 --- a/python/paddle/v2/framework/tests/test_protobuf.py +++ b/python/paddle/v2/framework/tests/test_protobuf.py @@ -1,11 +1,10 @@ -import paddle.v2.framework.proto.op_proto_pb2 as op_proto_lib -import paddle.v2.framework.proto.attribute_pb2 as attr_type_lib +import paddle.v2.framework.proto.framework_pb2 as framework_pb2 import unittest class TestFrameworkProto(unittest.TestCase): def test_all(self): - op_proto = op_proto_lib.OpProto() + op_proto = framework_pb2.OpProto() ipt0 = op_proto.inputs.add() ipt0.name = "a" ipt0.comment = "the input of cosine op" @@ -19,7 +18,7 @@ class TestFrameworkProto(unittest.TestCase): attr = op_proto.attrs.add() attr.name = "scale" attr.comment = "scale of cosine op" - attr.type = attr_type_lib.FLOAT + attr.type = framework_pb2.FLOAT op_proto.type = "cos" self.assertTrue(op_proto.IsInitialized()) diff --git a/python/paddle/v2/framework/tests/test_recurrent_op.py b/python/paddle/v2/framework/tests/test_recurrent_op.py index 5c77c477b3..d6000ab9f9 100644 --- a/python/paddle/v2/framework/tests/test_recurrent_op.py +++ b/python/paddle/v2/framework/tests/test_recurrent_op.py @@ -2,19 +2,74 @@ 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): + +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.set(np.random.random(shape), core.CPUPlace()) + tensor.set(np_data, core.CPUPlace()) return tensor -class TestRNN(unittest.TestCase): +class TestRecurrentOp(unittest.TestCase): ''' Test RNNOp @@ -28,7 +83,7 @@ class TestRNN(unittest.TestCase): memories: - h outputs: - - h + - h ''' input_dim = 30 @@ -36,33 +91,44 @@ class TestRNN(unittest.TestCase): weight_dim = 15 sent_len = 11 - def init(self): + 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() - rnn_op = self.create_rnn_op() ctx = core.DeviceContext.create(core.CPUPlace()) - print 'infer_shape' - rnn_op.infer_shape(self.scope) - - rnn_op.run(self.scope, ctx) + self.rnnop.infer_shape(self.scope) + self.rnnop.run(self.scope, ctx) + return np.array(self.scope.find_var("h").get_tensor()) def create_global_variables(self): # create inlink + x_np_data = self.py_rnn.x create_tensor(self.scope, "x", - [self.sent_len, self.batch_size, self.input_dim]) - create_tensor(self.scope, "W", [self.input_dim, self.input_dim]) - create_tensor(self.scope, "U", [self.input_dim, self.input_dim]) - create_tensor(self.scope, "h_boot", [self.batch_size, self.input_dim]) + [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"], @@ -75,23 +141,27 @@ class TestRNN(unittest.TestCase): outlink_alias=["h@alias"], pre_memories=["h@pre"], memories=["h@alias"]) - return rnnop def create_step_net(self): - var = self.scope.new_var("stepnet") - stepnet = var.get_net() - - x_fc_op = ops.fc(X="x@alias", W="W", Y="Wx") - h_fc_op = ops.fc(X="h@pre", W="U", Y="Uh") - sum_op = ops.add_two(X="Wx", Y="Uh", Out="sum") - sig_op = ops.sigmoid(X="sum", Y="h@alias") + 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.append_op(op) stepnet.complete_add_op(True) - - def test_recurrent(self): - self.init() + 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 f8521eb517..45d569da29 100644 --- a/python/paddle/v2/framework/tests/test_rowwise_add_op.py +++ b/python/paddle/v2/framework/tests/test_rowwise_add_op.py @@ -1,6 +1,7 @@ import unittest -from op_test_util import OpTestMeta import numpy as np +from op_test_util import OpTestMeta +from gradient_checker import GradientChecker, create_op class TestRowwiseAddOp(unittest.TestCase): @@ -15,5 +16,15 @@ class TestRowwiseAddOp(unittest.TestCase): self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['b'])} +class RowwiseAddGradOpTest(GradientChecker): + def test_rowwise_add(self): + op = create_op("rowwise_add") + inputs = { + "X": np.random.uniform(0.1, 1, [5, 10]).astype("float32"), + "b": np.random.uniform(0.1, 1, [10]).astype("float32") + } + self.check_grad(op, inputs, set(["X", "b"]), "Out") + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/framework/tests/test_scale_and_identity_op.py b/python/paddle/v2/framework/tests/test_scale_and_identity_op.py new file mode 100644 index 0000000000..69b301c376 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_scale_and_identity_op.py @@ -0,0 +1,43 @@ +import unittest +from op_test_util import OpTestMeta +from gradient_checker import GradientChecker, create_op +import numpy as np +from paddle.v2.framework.op import Operator + + +class IdentityTest(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "identity" + self.inputs = {'X': np.random.random((32, 784)).astype("float32")} + self.outputs = {'Out': self.inputs['X']} + + +class IdentityGradOpTest(GradientChecker): + def test_normal(self): + op = create_op("identity") + inputs = {"X": np.random.random((10, 10)).astype("float32")} + self.check_grad(op, inputs, set("X"), "Out") + + +class ScaleTest(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "scale" + self.inputs = {'X': np.random.random((32, 784)).astype("float32")} + self.attrs = {'scale': -2.3} + self.outputs = {'Out': self.inputs['X'] * self.attrs['scale']} + + +class ScaleGradTest(GradientChecker): + def test_normal(self): + op = Operator("scale", X="X", Out="Out", scale=3.2) + self.check_grad(op, + {"X": np.random.random((10, 10)).astype("float32")}, + set("X"), "Out") + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_scatter_op.py b/python/paddle/v2/framework/tests/test_scatter_op.py new file mode 100644 index 0000000000..c1f9444889 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_scatter_op.py @@ -0,0 +1,38 @@ +import unittest +from op_test_util import OpTestMeta +from gradient_checker import GradientChecker, create_op +import numpy +import paddle.v2.framework.core as core +from paddle.v2.framework.op import Operator + + +class TestScatterOp(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "scatter" + ref_np = numpy.ones((3, 3)).astype("float32") + index_np = numpy.array([1, 2]).astype("int32") + updates_np = numpy.random.random((2, 3)).astype("float32") + output_np = numpy.copy(ref_np) + output_np[index_np] += updates_np + self.inputs = {'Ref': ref_np, 'Index': index_np, 'Updates': updates_np} + self.outputs = {'Out': output_np} + + +class TestScatterGradOp(GradientChecker): + def test_scatter_grad(self): + op = create_op("scatter") + # test data setup + ref_np = numpy.ones((3, 10)).astype("float32") + index_np = numpy.array([1, 2]).astype("int32") + updates_np = numpy.random.random((2, 10)).astype("float32") + output_np = numpy.copy(ref_np) + output_np[index_np] += updates_np + inputs = {'Ref': ref_np, 'Index': index_np, 'Updates': updates_np} + self.check_grad( + op, inputs, set(["Updates", "Ref"]), "Out", in_place=True) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_sigmoid_op.py b/python/paddle/v2/framework/tests/test_sigmoid_op.py index 2a57a41ed8..273c2e5ab1 100644 --- a/python/paddle/v2/framework/tests/test_sigmoid_op.py +++ b/python/paddle/v2/framework/tests/test_sigmoid_op.py @@ -1,6 +1,7 @@ import unittest -from op_test_util import OpTestMeta import numpy as np +from op_test_util import OpTestMeta +from gradient_checker import GradientChecker, create_op class TestSigmoidOp(unittest.TestCase): @@ -8,12 +9,20 @@ class TestSigmoidOp(unittest.TestCase): def setUp(self): self.type = "sigmoid" - self.inputs = {'X': np.random.random((32, 100)).astype("float32")} + self.inputs = {'X': np.random.random((15, 31)).astype("float32")} self.outputs = {'Y': 1 / (1 + np.exp(-self.inputs['X']))} -#class TestSigmoidGradOp(unittest.TestCase): -#TODO(qingqing) add unit test +class TestSigmoidGradOp(GradientChecker): + def test_grad(self): + op = create_op("sigmoid") + inputs = {"X": np.random.uniform(0.1, 1, [11, 17]).astype("float32")} + # compare gpu and cpu results for backward op. + # this test will be skiped if only compiling CPU version. + self.compare_grad(op, inputs) + # check gradients + self.check_grad(op, inputs, set("X"), "Y", max_relative_error=0.007) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/optimizer.py b/python/paddle/v2/optimizer.py index ba58198033..29f0945eb4 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 a9cba8ca0b..4cfd91882e 100644 --- a/python/paddle/v2/parameters.py +++ b/python/paddle/v2/parameters.py @@ -1,5 +1,20 @@ +# 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 +from collections import OrderedDict import paddle.trainer.config_parser as cp import struct import tarfile @@ -28,9 +43,25 @@ def create(layers): class Parameters(object): """ - Parameters is a dictionary contains Paddle's parameter. The key of - Parameters is the name of parameter. The value of Parameters is a plain - :code:`numpy.ndarry` . + `Parameters` manages all the learnable parameters in a neural network. + It stores parameters' information in an OrderedDict. The key is + the name of a parameter, and value is a parameter's configuration(in + protobuf format), such as initialization mean and std, its size, whether it + is a static parameter, and so on. + + :param __param_conf__: store the configurations of learnable parameters in + the network in an OrderedDict. Parameter is added one by one into the + dict by following their created order in the network: parameters of + the previous layers in a network are careted first. You can visit the + parameters from bottom to top by iterating over this dict. + :type __param_conf__: OrderedDict + :param __gradient_machines__: all of the parameters in a neural network are + appended to a PaddlePaddle gradient machine, which is used internally to + copy parameter values between C++ and Python end. + :type __gradient_machines__: list + :param __tmp_params__: a dict to store dummy parameters if no + __gradient_machines__ is appended to `Parameters`. + :type __tmp_params__: dict Basically usage is @@ -48,7 +79,7 @@ class Parameters(object): """ def __init__(self): - self.__param_conf__ = dict() + self.__param_conf__ = OrderedDict() self.__gradient_machines__ = [] self.__tmp_params__ = dict() @@ -113,16 +144,7 @@ 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 +160,7 @@ 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 +168,19 @@ 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 +237,22 @@ 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 + if self.__param_conf__[key].is_static: + return np.zeros(self.__param_conf__[key].size, dtype=np.float32) + + return self.__getter_inner(key, api.PARAMETER_GRADIENT) + def set(self, parameter_name, value): """ Set parameter by parameter name & matrix. @@ -219,7 +270,7 @@ class Parameters(object): append gradient machine to parameters. This method is used internally in Trainer.train. - :param gradient_machine: Paddle C++ GradientMachine object. + :param gradient_machine: PaddlePaddle C++ GradientMachine object. :type gradient_machine: api.GradientMachine :return: """ @@ -250,7 +301,13 @@ 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/reader/creator.py b/python/paddle/v2/reader/creator.py index d0f18e4b66..97e844b92c 100644 --- a/python/paddle/v2/reader/creator.py +++ b/python/paddle/v2/reader/creator.py @@ -57,7 +57,7 @@ def text_file(path): return reader -def recordio_local(paths, buf_size=100): +def recordio(paths, buf_size=100): """ Creates a data reader from given RecordIO file paths separated by ",", glob pattern is supported. @@ -67,15 +67,19 @@ def recordio_local(paths, buf_size=100): import recordio as rec import paddle.v2.reader.decorator as dec + import cPickle as pickle def reader(): - a = ','.join(paths) - f = rec.reader(a) + if isinstance(paths, basestring): + path = paths + else: + path = ",".join(paths) + f = rec.reader(path) while True: r = f.read() if r is None: break - yield r + yield pickle.loads(r) f.close() return dec.buffered(reader, buf_size) diff --git a/python/paddle/v2/reader/tests/creator_test.py b/python/paddle/v2/reader/tests/creator_test.py index 359f3eeefb..cf190aa664 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 0000000000..a99a35bb82 Binary files /dev/null and b/python/paddle/v2/reader/tests/test_reader_recordio.dat differ diff --git a/python/paddle/v2/tests/test_layer.py b/python/paddle/v2/tests/test_layer.py index f2097e195f..783a0ca85d 100644 --- a/python/paddle/v2/tests/test_layer.py +++ b/python/paddle/v2/tests/test_layer.py @@ -141,12 +141,13 @@ class CostLayerTest(unittest.TestCase): cost8 = layer.rank_cost(left=score, right=score, label=score) cost9 = layer.lambda_cost(input=inference, score=score) cost10 = layer.sum_cost(input=inference) - cost11 = layer.huber_cost(input=score, label=label) + cost11 = layer.huber_regression_cost(input=score, label=label) + cost12 = layer.huber_classification_cost(input=score, label=label) print layer.parse_network([cost1, cost2]) print layer.parse_network([cost3, cost4]) print layer.parse_network([cost5, cost6]) - print layer.parse_network([cost7, cost8, cost9, cost10, cost11]) + print layer.parse_network([cost7, cost8, cost9, cost10, cost11, cost12]) crf = layer.crf(input=inference, label=label) crf_decoding = layer.crf_decoding(input=inference, size=3) diff --git a/python/paddle/v2/trainer.py b/python/paddle/v2/trainer.py index 76bae0bb12..0654a30104 100644 --- a/python/paddle/v2/trainer.py +++ b/python/paddle/v2/trainer.py @@ -27,16 +27,24 @@ class SGD(object): SGD Trainer combines data reader, network topolopy and update_equation together to train/test a neural network. - :param update_equation: The optimizer object. - :type update_equation: paddle.v2.optimizer.Optimizer :param cost: Target cost that neural network should be optimized. :type cost: paddle.v2.config_base.Layer :param parameters: The parameters dictionary. :type parameters: paddle.v2.parameters.Parameters + :param update_equation: The optimizer object. + :type update_equation: paddle.v2.optimizer.Optimizer :param extra_layers: Some layers in the neural network graph are not in the path of cost layer. - :param pserver_spec: pserver location, eg: localhost:3000 :type extra_layers: paddle.v2.config_base.Layer + :param is_local: Whether trainning locally + :type is_local: bool + :param pserver_spec: comma string for pserver location, + eg:127.10.0.10:3000,127.10.0.11:3000, + and this parameter is only used for fault + tolerant mode cluster training. + :type pserver_spec: string + :param use_etcd: Whether using etcd pserver. + :param use_etcd: bool """ def __init__(self, @@ -161,14 +169,14 @@ 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 0000000000..e19453c25d --- /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 38f0a503be..87b3823e52 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -1,5 +1,4 @@ from setuptools import setup, Distribution - class BinaryDistribution(Distribution): def has_ext_modules(foo): return True @@ -18,19 +17,25 @@ packages=['paddle', 'paddle.v2.framework.proto', 'py_paddle'] -setup_requires=["requests", - "numpy>=1.12", - "protobuf==3.1", - "recordio", - "matplotlib", - "rarfile", - "scipy>=0.19.0", - "Pillow", - "nltk>=3.2.2"] +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"] +# the prefix is sys.prefix which should always be usr +paddle_bin_dir = 'opt/paddle/bin' +paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/scripts/paddle_usage', + '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer', + '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model', + '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main', + '${PADDLE_BINARY_DIR}/paddle/scripts/paddle'] + +paddle_rt_lib_dir = 'lib' +paddle_rt_libs = ['${WARPCTC_LIBRARIES}'] +if '${MKL_SHARED_LIBS}'!= '': + paddle_rt_libs += '${MKL_SHARED_LIBS}'.split(';') + setup(name='paddlepaddle', version='${PADDLE_VERSION}', description='Parallel Distributed Deep Learning', @@ -45,14 +50,10 @@ setup(name='paddlepaddle', '': '${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': '${PROJ_BINARY_ROOT}/paddle/framework', - 'py_paddle': '${PROJ_ROOT}/paddle/py_paddle' + 'paddle.v2.framework.proto': '${PADDLE_BINARY_DIR}/paddle/framework', + 'py_paddle': '${PADDLE_SOURCE_DIR}/paddle/py_paddle' }, - scripts=['${PROJ_BINARY_ROOT}/paddle/scripts/paddle'], + scripts=paddle_bins, distclass=BinaryDistribution, - data_files=[('/usr/local/opt/paddle/bin', - ['${PROJ_BINARY_ROOT}/paddle/scripts/paddle_usage', - '${PROJ_BINARY_ROOT}/paddle/trainer/paddle_trainer', - '${PROJ_BINARY_ROOT}/paddle/trainer/paddle_merge_model', - '${PROJ_BINARY_ROOT}/paddle/pserver/paddle_pserver_main'])] + data_files=[(paddle_rt_lib_dir, paddle_rt_libs)] )