diff --git a/.copyright.hook b/.copyright.hook index 09afff2072df3384a429d01d06188218ae6e85d1..86b16ebdc46047c7cb3d7731a71cbf9647a1f2fe 100644 --- a/.copyright.hook +++ b/.copyright.hook @@ -9,7 +9,7 @@ import subprocess import platform COPYRIGHT = ''' - Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +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. diff --git a/CMakeLists.txt b/CMakeLists.txt index 23bbe829ac16180088bfa37df66e23f19b021ea3..030bd19b3fd2f561a847bbc4613e5d2030812a92 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -25,7 +25,6 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: " message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: " "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}") -find_package(Sphinx) if(NOT CMAKE_CROSSCOMPILING) find_package(CUDA QUIET) endif(NOT CMAKE_CROSSCOMPILING) @@ -226,5 +225,7 @@ if(WITH_PYTHON) endif() if(WITH_DOC) + find_package(Sphinx REQUIRED) + find_python_module(recommonmark REQUIRED) add_subdirectory(doc) endif() diff --git a/Dockerfile b/Dockerfile index 164fe84904947bfc3cf71132b5fba04744460b26..ea39efd00bb5c0a7deb3f6d57083d83a673b883c 100644 --- a/Dockerfile +++ b/Dockerfile @@ -70,7 +70,7 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8 # specify sphinx version as 1.5.6 and remove -U option for [pip install -U # sphinx-rtd-theme] since -U option will cause sphinx being updated to newest # version(1.7.1 for now), which causes building documentation failed. -RUN pip install --upgrade pip==9.0.3 && \ +RUN easy_install -U pip && \ pip install -U wheel && \ pip install -U docopt PyYAML sphinx==1.5.6 && \ pip install sphinx-rtd-theme==0.1.9 recommonmark diff --git a/README.md b/README.md index a3b13fe79cc33927e5cc3e091926b111688a941b..8d89c6b1ec9e4aefbd64328dedb4e8c7cc50c21b 100644 --- a/README.md +++ b/README.md @@ -62,9 +62,9 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl ## Installation It is recommended to check out the -[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/docker_install_en.html) +[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/docker_install_en.html) before looking into the -[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/build_from_source_en.html). +[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/build_from_source_en.html). ## Documentation diff --git a/benchmark/fluid/mnist.py b/benchmark/fluid/mnist.py index 1e2185dfac1072d1f1046f4616a9d53a8fc76061..400200c4745017bd9d160bb9e415fde041c0a6c8 100644 --- a/benchmark/fluid/mnist.py +++ b/benchmark/fluid/mnist.py @@ -159,6 +159,7 @@ def run_benchmark(model, args): paddle.dataset.mnist.train(), batch_size=args.batch_size) accuracy = fluid.metrics.Accuracy() + train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name) iters, num_samples, start_time = 0, 0, time.time() for pass_id in range(args.pass_num): accuracy.reset() @@ -175,17 +176,20 @@ def run_benchmark(model, args): y_data = np.array(map(lambda x: x[1], data)).astype("int64") y_data = y_data.reshape([len(y_data), 1]) - outs = exe.run( - fluid.default_main_program(), + outs = train_exe.run( feed={"pixel": img_data, "label": y_data}, - fetch_list=[avg_cost, batch_acc, batch_size_tensor] + fetch_list=[ + avg_cost.name, batch_acc.name, batch_size_tensor.name + ] ) # The accuracy is the accumulation of batches, but not the current batch. - accuracy.update(value=outs[1], weight=outs[2]) + accuracy.update( + value=np.array(np.mean(outs[1])), + weight=np.mean(np.array(outs[2]))) iters += 1 num_samples += len(y_data) - loss = np.array(outs[0]) - acc = np.array(outs[1]) + loss = np.mean(np.array(outs[0])) + acc = np.mean(np.array(outs[1])) train_losses.append(loss) train_accs.append(acc) print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" % diff --git a/benchmark/fluid/resnet.py b/benchmark/fluid/resnet.py index 831fa2c019fc2868cd85b1ca7b2c8c76a2f1628c..0fd7258a804e7c93b0b03da140140394bf90004a 100644 --- a/benchmark/fluid/resnet.py +++ b/benchmark/fluid/resnet.py @@ -241,6 +241,7 @@ def run_benchmark(model, args): exe = fluid.Executor(place) exe.run(fluid.default_startup_program()) accuracy = fluid.average.WeightedAverage() + train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name) if args.use_fake_data: data = train_reader().next() image = np.array(map(lambda x: x[0].reshape(dshape), data)).astype( @@ -264,14 +265,17 @@ def run_benchmark(model, args): data)).astype('float32') label = np.array(map(lambda x: x[1], data)).astype('int64') label = label.reshape([-1, 1]) - loss, acc, weight = exe.run( - fluid.default_main_program(), + loss, acc, weight = train_exe.run( feed={'data': image, 'label': label}, - fetch_list=[avg_cost, batch_acc, batch_size_tensor]) + fetch_list=[ + avg_cost.name, batch_acc.name, batch_size_tensor.name + ]) iters += 1 num_samples += len(label) - accuracy.add(value=acc, weight=weight) + accuracy.add(value=np.array(np.mean(acc)), weight=np.mean(weight)) + loss = np.mean(np.array(loss)) + acc = np.mean(np.array(acc)) train_losses.append(loss) train_accs.append(acc) print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" % diff --git a/benchmark/fluid/vgg.py b/benchmark/fluid/vgg.py index 53e34e0cbd15914791c305db6797f826ebfae34e..2a9566a45c3804183e05db9298cec4f670225a6f 100644 --- a/benchmark/fluid/vgg.py +++ b/benchmark/fluid/vgg.py @@ -169,6 +169,7 @@ def main(): iters, num_samples, start_time = 0, 0, time.time() accuracy = fluid.average.WeightedAverage() + train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name) for pass_id in range(args.pass_num): accuracy.reset() train_accs = [] @@ -184,14 +185,17 @@ def main(): y_data = np.array(map(lambda x: x[1], data)).astype("int64") y_data = y_data.reshape([-1, 1]) - loss, acc, weight = exe.run( - fluid.default_main_program(), + loss, acc, weight = train_exe.run( feed={"pixel": img_data, "label": y_data}, - fetch_list=[avg_cost, batch_acc, batch_size_tensor]) - accuracy.add(value=acc, weight=weight) + fetch_list=[ + avg_cost.name, batch_acc.name, batch_size_tensor.name + ]) + accuracy.add(value=np.array(np.mean(acc)), weight=np.mean(weight)) iters += 1 num_samples += len(y_data) + loss = np.mean(np.array(loss)) + acc = np.mean(np.array(acc)) print( "Pass = %d, Iter = %d, Loss = %f, Accuracy = %f" % (pass_id, iters, loss, acc) diff --git a/cmake/external/boost.cmake b/cmake/external/boost.cmake index 10662fc96704685f030a5d76c6857d4bc20a63d9..499682f644d60c16c3025870e7dd2a890630a2bb 100644 --- a/cmake/external/boost.cmake +++ b/cmake/external/boost.cmake @@ -24,7 +24,7 @@ set(BOOST_PROJECT "extern_boost") # So we use 1.41.0 here. set(BOOST_VER "1.41.0") set(BOOST_TAR "boost_1_41_0") -set(BOOST_URL "http://paddlepaddledeps.bj.bcebos.com/${BOOST_TAR}.tar.gz") +set(BOOST_URL "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz") set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost) set(BOOST_DOWNLOAD_DIR "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}") set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE) diff --git a/cmake/external/eigen.cmake b/cmake/external/eigen.cmake index edc93c2773f46ec9e0bf898557c55c93274e6a01..e029300eee9b99582f085f6b650e03f7dacc091a 100644 --- a/cmake/external/eigen.cmake +++ b/cmake/external/eigen.cmake @@ -21,11 +21,12 @@ else() ExternalProject_Add( extern_eigen3 ${EXTERNAL_PROJECT_LOG_ARGS} - GIT_REPOSITORY "https://github.com/RLovelett/eigen.git" + GIT_REPOSITORY "https://github.com/eigenteam/eigen-git-mirror" # eigen on cuda9.1 missing header of math_funtions.hpp # https://stackoverflow.com/questions/43113508/math-functions-hpp-not-found-when-using-cuda-with-eigen GIT_TAG 917060c364181f33a735dc023818d5a54f60e54c PREFIX ${EIGEN_SOURCE_DIR} + DOWNLOAD_NAME "eigen" UPDATE_COMMAND "" CONFIGURE_COMMAND "" BUILD_COMMAND "" diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake index 5759e5c489724332793bf103b7aacf7ffb068611..25c07850dda7b2f69c2207c37b9d2368632104ec 100644 --- a/cmake/external/mkldnn.cmake +++ b/cmake/external/mkldnn.cmake @@ -45,15 +45,15 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML") ELSE() MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN") ENDIF() - -SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow") -SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} -Wno-error=strict-overflow") +SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-unused-result") +SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}") +SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}") ExternalProject_Add( ${MKLDNN_PROJECT} ${EXTERNAL_PROJECT_LOG_ARGS} DEPENDS ${MKLDNN_DEPENDS} GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git" - GIT_TAG "v0.11" + GIT_TAG "db3424ad44901513c03a1ea31ccaacdf633fbe9f" PREFIX ${MKLDNN_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR} @@ -61,6 +61,7 @@ ExternalProject_Add( CMAKE_ARGS -DMKLROOT=${MKLML_ROOT} CMAKE_ARGS -DCMAKE_C_FLAGS=${MKLDNN_CFLAG} CMAKE_ARGS -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG} + CMAKE_ARGS -DWITH_TEST=OFF -DWITH_EXAMPLE=OFF CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR} -DMKLROOT:PATH=${MKLML_ROOT} ) diff --git a/cmake/external/mklml.cmake b/cmake/external/mklml.cmake index 796bcf28a1dfb308ccb7a2f839742c5c2fcf2002..e9a37b52e61b2525b047352cc70510df83eccb7f 100644 --- a/cmake/external/mklml.cmake +++ b/cmake/external/mklml.cmake @@ -27,8 +27,8 @@ ENDIF() INCLUDE(ExternalProject) SET(MKLML_PROJECT "extern_mklml") -SET(MKLML_VER "mklml_lnx_2018.0.1.20171007") -SET(MKLML_URL "http://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.tgz") +SET(MKLML_VER "mklml_lnx_2018.0.3.20180406") +SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz") SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml") SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}") SET(MKLML_DST_DIR "mklml") diff --git a/cmake/external/snappy.cmake b/cmake/external/snappy.cmake index 80282329c6ac65fbd1493a6838efca4bd9cadaad..af09ed4d5d6e21cc50aba5198a7e9ea56f49451a 100644 --- a/cmake/external/snappy.cmake +++ b/cmake/external/snappy.cmake @@ -47,8 +47,6 @@ ExternalProject_Add( -DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPY_INSTALL_DIR}/lib -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE} - BUILD_COMMAND make -j8 - INSTALL_COMMAND make install ) add_library(snappy STATIC IMPORTED GLOBAL) diff --git a/cmake/external/snappystream.cmake b/cmake/external/snappystream.cmake index 20a96430823d07a07d4bb4602e7fc0cfe55c3bf2..6df636d7fa8757ade73892bda03a80ba9767472b 100644 --- a/cmake/external/snappystream.cmake +++ b/cmake/external/snappystream.cmake @@ -46,8 +46,6 @@ ExternalProject_Add( -DCMAKE_INSTALL_PREFIX:PATH=${SNAPPYSTREAM_INSTALL_DIR} -DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPYSTREAM_INSTALL_DIR}/lib -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE} - BUILD_COMMAND make -j8 - INSTALL_COMMAND make install DEPENDS snappy ) diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake index cc758019827b9a5416a801e4da43d754d4492a73..7117a3a4f31c88b3c4a81e611146123903659ad5 100644 --- a/cmake/inference_lib.cmake +++ b/cmake/inference_lib.cmake @@ -70,6 +70,12 @@ copy(glog_lib DSTS ${dst_dir} ${dst_dir}/lib ) +set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/boost/") +copy(boost_lib + SRCS ${BOOST_INCLUDE_DIR}/boost + DSTS ${dst_dir} +) + if(NOT PROTOBUF_FOUND) set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/protobuf") copy(protobuf_lib @@ -92,6 +98,14 @@ elseif (WITH_MKLML) ) endif() +if(WITH_MKLDNN) + set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/mkldnn") + copy(mkldnn_lib + SRCS ${MKLDNN_INC_DIR} ${MKLDNN_SHARED_LIB} + DSTS ${dst_dir} ${dst_dir}/lib + ) +endif() + if(NOT MOBILE_INFERENCE AND NOT RPI) set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/snappy") copy(snappy_lib @@ -142,4 +156,30 @@ copy(string_lib DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat ) +set(module "pybind") +copy(pybind_lib + SRCS ${CMAKE_CURRENT_BINARY_DIR}/paddle/fluid/${module}/pybind.h + DSTS ${dst_dir}/${module} +) + +# CMakeCache Info +copy(cmake_cache + SRCS ${CMAKE_CURRENT_BINARY_DIR}/CMakeCache.txt + DSTS ${CMAKE_INSTALL_PREFIX}) + add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep}) + +# paddle fluid version +execute_process( + COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1 + OUTPUT_VARIABLE PADDLE_GIT_COMMIT) +set(version_file ${CMAKE_INSTALL_PREFIX}/version.txt) +file(WRITE ${version_file} + "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n" + "WITH_MKL: ${WITH_MKL}\n" + "WITH_GPU: ${WITH_GPU}\n") +if(WITH_GPU) + file(APPEND ${version_file} + "CUDA version: ${CUDA_VERSION}\n" + "CUDNN version: v${CUDNN_MAJOR_VERSION}\n") +endif() diff --git a/contrib/inference/README.md b/contrib/inference/README.md new file mode 100644 index 0000000000000000000000000000000000000000..20969fac6c8f894ffb4a02b48f795e2a0dcbd096 --- /dev/null +++ b/contrib/inference/README.md @@ -0,0 +1,27 @@ +# Embed Paddle Inference in Your Application + +Paddle inference offers the APIs in `C` and `C++` languages. + +One can easily deploy a model trained by Paddle following the steps as below: + +1. Optimize the native model; +2. Write some codes for deployment. + + +Let's explain the steps in detail. + +## Optimize the native Fluid Model + +The native model that get from the training phase needs to be optimized for that. + +- Clean the noise such as the cost operators that do not need inference; +- Prune unnecessary computation fork that has nothing to do with the output; +- Remove extraneous variables; +- Memory reuse for native Fluid executor; +- Translate the model storage format to some third-party engine's, so that the inference API can utilize the engine for acceleration; + +We have an official tool to do the optimization, call `paddle_inference_optimize --help` for more information. + +## Write some codes + +Read `paddle_inference_api.h` for more information. diff --git a/contrib/inference/paddle_inference_api.h b/contrib/inference/paddle_inference_api.h new file mode 100644 index 0000000000000000000000000000000000000000..dbaa7c95b97e954537707566e5b7458e6afd14c8 --- /dev/null +++ b/contrib/inference/paddle_inference_api.h @@ -0,0 +1,69 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include +#include + +namespace paddle { + +class Predictor { +public: + struct Attr; + Predictor() = default; + + // Build the network before inference. + bool Init(const Attr& attr); + + // Predict an record. + // Arguments: + // inputs: the name of the input variables. + // outputs: the name of the output varaibles. + // input_shapes: the shape of the input variables. + // output_shapes: the shape of the output variables. + // input_data: the data of the input variables. + // output_data: the data of the output variables. + bool Run(const std::vector& inputs, + const std::vector& outputs, + const std::vector>& input_shapes, + const std::vector>& output_shapes, + const std::vector>& input_data, + std::vector>* output_data); + + // Clone a predictor that share the model weights. + Predictor* Clone(); + + // Destroy the Predictor. + ~Predictor(); + + struct Attr { + enum class EngineKind; + + std::string model_dir; // path to the model directory. + bool enable_engine{false}; // Enable to execute (part of) the model on + // third-party engines. + EngineKind engine_kind{Attr::EngineKind::kNone}; + + enum class EngineKind { + kNone = -1, // Use the native Fluid facility. + kAnakin, // Use Anakin for inference. + kTensorRT, // Use TensorRT for inference. + kAutoMixedAnakin, // Automatically mix Fluid with Anakin. + kAutoMixedTensorRT, // Automatically mix Fluid with TensorRT. + }; + }; +}; + +} // namespace paddle diff --git a/doc/fluid/design/concepts/functions_operators_layers.md b/doc/fluid/design/concepts/functions_operators_layers.md index 30bc488a18a28d349645d9d2502aae6691a69931..1f86b99e5197c3e0b85fd76fe704520ef21b06d3 100644 --- a/doc/fluid/design/concepts/functions_operators_layers.md +++ b/doc/fluid/design/concepts/functions_operators_layers.md @@ -40,7 +40,7 @@ template class FCOp : public OperatorBase { public: void Run(...) { - add(mul(Input("X"), Input("W")), Input("b"); + add(mul(Input("X"), Input("W")), Input("b")); } }; REGISTER_OP(FCOp, "fc"); diff --git a/doc/fluid/design/concepts/lod_tensor.md b/doc/fluid/design/concepts/lod_tensor.md index a88292e7888d0ebc64ee89ca315dfea38a12c71d..d606d7a790b4b0dc18553f2220d39cec8aa619ec 100644 --- a/doc/fluid/design/concepts/lod_tensor.md +++ b/doc/fluid/design/concepts/lod_tensor.md @@ -155,7 +155,7 @@ into offsets 3 2+3 4+5 1+9 2+10 3+12 ``` -so we know that the first sentence is from word 0 to word 3, and the second sentence from work 3 to word 5. +so we know that the first sentence is from word 0 to word 3, and the second sentence from word 3 to word 5. Similarly, the lengths in the top level LoD diff --git a/doc/fluid/design/dist_train/async_update.md b/doc/fluid/design/dist_train/async_update.md index 6a0835b761b69030ba30697e6e8863928efbf57f..248d2ec18dafdecac9184527638754b6ba4d85b8 100644 --- a/doc/fluid/design/dist_train/async_update.md +++ b/doc/fluid/design/dist_train/async_update.md @@ -4,34 +4,37 @@ For the typical synchronous distributed training, some significant steps are as follows: -1. A Trainer will compute the gradients and SEND them to the Parameter Server(PServer) nodes. -1. After the PServer node received gradients came from all the Trainers, It will aggregate the +1. A trainer process will compute the gradients and **send** them to the parameter server (PS) nodes. +1. After the PS node received gradients came from all the Trainers, It will aggregate the gradient variables for the same parameter into one gradient variable and then apply the aggregated gradient to the respective parameter, finally using an optimize algorithms(SGD, Monument...) to update the parameters. -1. The Trainer would wait for the PServers finished the optimize stage, and GET the parameters from PServer, +1. The Trainer would wait for the PS finished the optimize stage, and GET the parameters from PS, so all the Trainers would get the same parameters. -In the synchronously distributed training, there should be a `Barrier` to synchronise the -parameters after the optimizing stage. The performance of a distributed training job would -depend on the slowest node if there were hundreds or thousands of training nodes in a -Job, the performance of synchronously distributed training might be very poor because of -the slow node. So this design doc would introduce an approach to implement -*asynchronously* distributed training in PaddlePaddle Fluid. +In Synchronous Distributed Training, there is a **barrier** on each PS to wait until all trainers processes +have completed running current mini-batch. After that, all trainers can continue to run the next +mini-batch. So, we can find that the overall performance of Synchronous Distributed Training depends +on the slowest node. + +In Asynchronous Distributed Training, we don't need to wait for a global mini-bach, the optimizer on +the PS will run immediately when the gradient is uploaded to the PS from one trainer. This mode would +train such models that achieve scaling, better throughput. In this design doc, we will introduce how to +implement the Asynchronous Distributed Training base on PaddlePaddle Fluid. ## Design -As the figure above, we describe a global view of asynchronously update process and use +As the figure above, we describe a global view of the asynchronous update process and use the parameter `w1` as an example to introduce the steps: 1. For each gradient variables, they may distribute on different GPU card and aggregate them while they are all calculated. -1. Split the gradient variable into multiple blocks according to the number of PServer +1. Split the gradient variable into multiple blocks according to the number of PS instances and then send them. -1. PServer would run an `Optimize Block` using a specified optimize algorithm to update +1. PS would run an `Optimize Block` using a specified optimize algorithm to update the specified parameter. -1. The trainer will fetch latest parameter from PServer before running forward Op which depends +1. The trainer will fetch the latest parameter from PS before running forward Op which depends on the specified parameter. 1. Broadcast the received variable into multiple GPU cards and continue to run the next mini-batch. @@ -40,8 +43,8 @@ mini-batch. - For the multiple devices distributed training, we need to aggregate the gradient variables which placed on different devices firstly and then schedule a `SendVars` Operator to -send the gradient variables to the multiple PServer instances. -- Schedule `FetchVars` operator to fetch the latest parameter from PServer before running +send the gradient variables to the multiple PS instances. +- Schedule `FetchVars` operator to fetch the latest parameter from PS before running the forward ops. - There could be a large number of gradient variables to be sent, so we need to use another thread pool(IO Threadpool) whose a number of the schedulable threads is larger than the diff --git a/doc/fluid/design/motivation/api.md b/doc/fluid/design/motivation/api.md index e6a4638d9100d9b07c3ee6b92b530a17eae1c162..bc222564e3ec28e306ca0572b6a23104f6e9cbc5 100644 --- a/doc/fluid/design/motivation/api.md +++ b/doc/fluid/design/motivation/api.md @@ -77,8 +77,7 @@ print "The sematic-vector of testA: ", paddle.infer(fA, parameters, testA) ### Example 2. Sharing Parameters between "Models" -We use [GAN](https://github.com/PaddlePaddle/book/tree/develop/gan) in -this example. In the following example program, `d0` and `d1` +We use GAN in this example. In the following example program, `d0` and `d1` correspond to the two networks in the following figure: diff --git a/doc/fluid/design/multi_devices/operator_kernel_type.md b/doc/fluid/design/multi_devices/operator_kernel_type.md index 8c1bc8f76a337006497e5ab5e5a710f9f49261b8..5e391bd62b4f4e123a9a6f35b7adf5726f205635 100644 --- a/doc/fluid/design/multi_devices/operator_kernel_type.md +++ b/doc/fluid/design/multi_devices/operator_kernel_type.md @@ -75,7 +75,7 @@ Different layout leads to different implementation of the operator kernel. There - The inference of Layout is at run-time, not at compile-time. -- Every operator has to implement different kernels for different layouts. Let's take MKLDNN as an example. If we want to implement an MKLDNN convolution operator, we have to implement all the kernels for different layouts, which are listed [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to register kernels for MKLDNN operators. +- Every operator has to implement different kernels for different layouts. Let's take MKLDNN as an example. If we want to implement an MKLDNN convolution operator, we have to implement all the kernels for different layouts, which are listed [here](http://intel.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to register kernels for MKLDNN operators. `Layout` is also defined as a enum variable: diff --git a/doc/fluid/howto/cluster/nccl2_rdma_training.md b/doc/fluid/howto/cluster/nccl2_rdma_training.md new file mode 100644 index 0000000000000000000000000000000000000000..cecd5c3a7a7339e3be6772543a534728ec132105 --- /dev/null +++ b/doc/fluid/howto/cluster/nccl2_rdma_training.md @@ -0,0 +1,110 @@ +# Distributed Training with NCCL2 and RDMA + +When doing distributed multi-GPU training, network bandwith often becomes the +bottle neck. We introduce a way to use NCCL2 to do such training job to +achieve best performace. + +## Prepare Hardwares with RDMA and Multiple GPUs + +I'm using two Linux servers each of them is installed with 8 GPUs and +one 100Gb RDMA card. +Base environment is: + +* OS: CentOS 7.4 +* RDMA device: "Mellanox Technologies MT27700 Family [ConnectX-4]" +* Kernel version: `4.4.88-1.el7.elrepo.x86_64` +* Docker version: `1.12.6` +* Docker storage driver: `overlay2` +* IP addresses: 192.168.16.30,192.168.16.34 + +In general, the steps including: + +1. Install GPU drivers +1. Install RDMA drivers +1. Install "InfiniBand Support" +1. Use docker to run tests and make sure GPUs and RDMA can work inside + the container. + +I'll ommit section "Install GPU drivers" because we can find it easily +somewhere else. + +### Install RDMA drivers + +For my case, I've got two machines with device +"Mellanox Technologies MT27700 Family [ConnectX-4]" installed. The OS was +"CentOS 7.4" and I updated the kernel to version 4.4 so that docker can +work with latest overlay2 filesystem. + +***NOTE: before you start, make sure you have a way to get a console +of the server other than ssh because we may need to re-configure the +network device.*** + +1. Go to http://www.mellanox.com/page/products_dyn?product_family=26, + download `MLNX_OFED` software in the bottom of the page, and upload it + onto the server. +1. Run `./mlnxofedinstall --add-kernel-support` in the software package. +1. Run `/etc/init.d/openibd restart` to make everything work, note that + this operation may cause the network goes down if you are using this + RDMA device as default network device and use ssh to login the server. +1. Re-configure the network interface, for example: + `ifconfig eth2 192.168.16.30/20 up`, then add routes if needed: + `ip route add default via 192.168.16.1 dev eth2`. +1. Do the same thing on the other node. +1. Use `ping` to test if the two nodes have typical ICMP connection. +1. Use either `udaddy` or `ib_write_bw` to test the network connection is + ready and have the desired bandwith. + +### Prepare Docker Image to Run RDMA Programs + +1. Build a docker image using cuda base image like: `nvidia/cuda:8.0-cudnn5-devel-ubuntu16.04` and install paddlepaddle whl + package in it. +1. Start a docker container and mount GPU driver libs into it (you can + skip this step if you are using nvidia-docker). +1. Mount RDMA dirvers and libs into the docker image (see below section), + also `udaddy` and `ib_write_bw` if needed. +1. Mount GPU devices and RDMA devices into the container using `--device` + or just use privileged mode `--privileged`. +1. Start the container using host network mode: `--net=host` + +### RDMA Library Files Needed + +Usually, `MLNX_OFED` install latest supported libs under +`/usr/lib64/mlnx_ofed/valgrind`. Other libs also needed to run RDMA programs +is listed below. These libs must be mounted into the docker container. + +* Libs under `/usr/lib64/mlnx_ofed/valgrind` + * libibcm.so + * libibverbs.so + * libmlx4.so + * libmlx5.so + * libmlx5-rdmav2.so + * librdmacm.so +* Other libs: + * libnl-3.so.200 + * libnl-route-3.so.200 + * libnuma.so.1 + +## Start to Run the Training Job + +Setting NCCL environment variables to turn NCCL switches on and off: + + +| Env Name | Description | +| --- | --- | +| NCCL_SOCKET_IFNAME | The RDMA device, e.g. eth2 | +| NCCL_P2P_DISABLE | Set to 1 to disable P2P transfer between GPUs | +| NCCL_IB_DISABLE | Set to 1 to disable using RDMA | +| NCCL_IB_CUDA_SUPPORT | Set to 1 to enable GPU Direct if supported | +| NCCL_DEBUG | Set debug level: VERSION, WARN, INFO | + +My two servers are: `192.168.16.30,192.168.16.34`, On node 1, Run : + +```bash +PADDLE_TRAINER_ID=0 PADDLE_PORT=48372 PADDLE_WORKERS=192.168.16.30,192.168.16.34 POD_IP=192.168.16.30 stdbuf -oL python vgg16.py +``` + +On node 2, Run: + +```bash +PADDLE_TRAINER_ID=1 PADDLE_PORT=48372 PADDLE_WORKERS=192.168.16.30,192.168.16.34 POD_IP=192.168.16.34 stdbuf -oL python vgg16.py +``` diff --git a/doc/fluid/howto/optimization/cpu_profiling_cn.md b/doc/fluid/howto/optimization/cpu_profiling_cn.md index 8266dec3c6125a09b90ac0ccd4aa5464f5c7db31..198a05a79e19227e90eaafe116217a164cd51a7d 100644 --- a/doc/fluid/howto/optimization/cpu_profiling_cn.md +++ b/doc/fluid/howto/optimization/cpu_profiling_cn.md @@ -1,3 +1,5 @@ +# CPU性能调优 + 此教程会介绍如何使用Python的cProfile包、Python库yep、Google perftools来进行性能分析 (profiling) 与调优（performance tuning）。 Profling 指发现性能瓶颈。系统中的瓶颈可能和程序员开发过程中想象的瓶颈相去甚远。Tuning 指消除瓶颈。性能优化的过程通常是不断重复地 profiling 和 tuning。 @@ -8,7 +10,7 @@ PaddlePaddle 用户一般通过调用 Python API 编写深度学习程序。大 * Python 与 C++ 混合代码的性能分析 -# Python代码的性能分析 +## Python代码的性能分析 ### 生成性能分析文件 diff --git a/doc/fluid/howto/optimization/cpu_profiling_en.md b/doc/fluid/howto/optimization/cpu_profiling_en.md index e95556dd608b7ff0a3eb18873df0015a2da94e7c..216694965b3c878a8a5f3ccd2a0cba8d21d9ce05 100644 --- a/doc/fluid/howto/optimization/cpu_profiling_en.md +++ b/doc/fluid/howto/optimization/cpu_profiling_en.md @@ -1,3 +1,5 @@ +# Tune CPU performance + This tutorial introduces techniques we use to profile and tune the CPU performance of PaddlePaddle. We will use Python packages `cProfile` and `yep`, and Google's `perftools`. @@ -14,7 +16,7 @@ the profiling and tuning of 1. the Python code and 1. the mixture of Python and C++ code. -# Profiling the Python Code +## Profiling the Python Code ### Generate the Performance Profiling File diff --git a/doc/v2/build_and_install/build_from_source_cn.rst b/doc/v2/build_and_install/build_from_source_cn.rst index 115b92a33888abf1e1be400e1abbb58b632a2976..f846928954dd3a05e11054ce2ff2ff839fbefd4b 100644 --- a/doc/v2/build_and_install/build_from_source_cn.rst +++ b/doc/v2/build_and_install/build_from_source_cn.rst @@ -19,8 +19,9 @@ ---------------- PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安装编译依赖的步骤，可选的不同编译环境Docker镜像 -可以在 `这里 `_ 找到。或者 -参考下述可选步骤，从源码中构建用于编译PaddlePaddle的Docker镜像。 +可以在 `这里 `_ 找到，您也可以 +在 `这里 `_ 找到 paddle_manylinux_devel +镜像的编译以及使用方法。或者参考下述可选步骤，从源码中构建用于编译PaddlePaddle的Docker镜像。 如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 `编译依赖`_ 之后才能开始编译的步骤。 diff --git a/doc/v2/build_and_install/build_from_source_en.rst b/doc/v2/build_and_install/build_from_source_en.rst index 8fef9e7347e8d924026999bfda985381750c6b51..d1b5b88dff81d4c5cee3dd13a7dccbc333ab6a17 100644 --- a/doc/v2/build_and_install/build_from_source_en.rst +++ b/doc/v2/build_and_install/build_from_source_en.rst @@ -22,6 +22,8 @@ How To Build You need to use Docker to build PaddlePaddle to avoid installing dependencies by yourself. We have several pre-built Docker images `here `_ , +you can also find how to build and use paddle_manylinux_devel Docker image from +`here `_ Or you can build your own image from source as the optional step below: .. code-block:: bash diff --git a/doc/v2/build_and_install/pip_install_cn.rst b/doc/v2/build_and_install/pip_install_cn.rst index b3d882743785e8ee301b71b696230531d2b7ba58..9b84bb6425af1eeb94a4f2f5d6c2b1e28c62e3c8 100644 --- a/doc/v2/build_and_install/pip_install_cn.rst +++ b/doc/v2/build_and_install/pip_install_cn.rst @@ -10,20 +10,38 @@ PaddlePaddle可以使用常用的Python包管理工具 使用pip安装 ------------------------------ - -执行下面的命令即可在当前机器上安装PaddlePaddle的运行时环境，并自动下载安装依赖软件，版本为cpu_avx_openblas。 +执行下面的命令即可在当前机器上安装PaddlePaddle的运行时环境，并自动下载安装依赖软件。 .. code-block:: bash pip install paddlepaddle +当前的默认版本为0.12.0，cpu_avx_openblas，您可以通过指定版本号来安装其它版本，例如: + + .. code-block:: bash + + pip install paddlepaddle==0.11.0 + -如果需要安装支持GPU的版本（cuda7.5_cudnn5_avx_openblas），需要执行： +如果需要安装支持GPU的版本（cuda8.0_cudnn5_avx_openblas），需要执行： .. code-block:: bash pip install paddlepaddle-gpu +当前的默认版本也是0.12.0，PaddlePaddle针对不同需求提供了更多版本的安装包，部分列表如下： + +================================= ======================================== +版本号 版本说明 +================================= ======================================== +paddlepaddle-gpu==0.12.0 使用CUDA 8.0和cuDNN 5编译的0.12.0版本 +paddlepaddle-gpu==0.11.0.post87 使用CUDA 8.0和cuDNN 7编译的0.11.0版本 +paddlepaddle-gpu==0.11.0.post8 使用CUDA 8.0和cuDNN 5编译的0.11.0版本 +paddlepaddle-gpu==0.11.0 使用CUDA 7.5和cuDNN 5编译的0.11.0版本 +================================= ======================================== + +您可以在 `Release History `_ 中找到paddlepaddle-gpu的各个发行版本。 + 如果需要获取并安装最新的（开发分支）PaddlePaddle，可以从我们的CI系统中下载最新的whl安装包和c-api开发包并安装， 您可以从下面的表格中找到需要的版本： @@ -37,12 +55,11 @@ PaddlePaddle可以使用常用的Python包管理工具 :header: "版本说明", "cp27-cp27mu", "cp27-cp27m" :widths: 1, 3, 3 - "cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cpu_avx_mkl", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `_" + "cpu_avx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `_" + "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `_" + "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `_" + "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `_" .. _pip_dependency: @@ -69,7 +86,7 @@ PaddlePaddle发布的安装包会尽量对齐 `manylinux1 9.0.0) 才可以安装。可以使用下面的命令更新您的pip： .. code-block:: bash diff --git a/doc/v2/build_and_install/pip_install_en.rst b/doc/v2/build_and_install/pip_install_en.rst index 1e409d86b9775094998f72f92954f4bbc1013ea1..fcac76d6a24eb4905a20f797d614db8f743342d7 100644 --- a/doc/v2/build_and_install/pip_install_en.rst +++ b/doc/v2/build_and_install/pip_install_en.rst @@ -12,20 +12,38 @@ Install using pip ------------------------------ Run the following command to install PaddlePaddle on the current -machine, it will also download requirements, the version is cpu_avx_openblas. +machine, it will also download requirements. .. code-block:: bash pip install paddlepaddle +the default version is 0.12.0, cpu_avx_openblas, you can specify the versions to satisfy your demands, like: -If you wish to install GPU version (cuda7.5_cudnn5_avx_openblas), just run: + .. code-block:: bash + + pip install paddlepaddle==0.11.0 + +If you need to install a GPU-enabled version (cuda8.0_cudnn5_avx_openblas), you need to run: .. code-block:: bash pip install paddlepaddle-gpu -If you wish to install the latest develop branch PaddlePaddle, +The default version is also 0.12.0, PaddlePaddle provides several versions of packages for different needs, as shown in the table: + +================================= ======================================== +版本号 版本说明 +================================= ======================================== +paddlepaddle-gpu==0.12.0 0.12.0 built with CUDA 8.0 and cuDNN 5 +paddlepaddle-gpu==0.11.0.post87 0.11.0 built with CUDA 8.0 and cuDNN 7 +paddlepaddle-gpu==0.11.0.post8 0.11.0 built with CUDA 8.0 and cuDNN 5 +paddlepaddle-gpu==0.11.0 0.11.0 built with CUDA 7.5 and cuDNN 5 +================================= ======================================== + +You can find all versions released of paddlepaddle-gpu in `Release History `_ . + +If you wish to install the latest develop branch PaddlePaddle, you can download the latest whl package from our CI system. Access the below links, log in as guest, then click at the "Artifact" tab, you'll find the download link of whl packages. @@ -40,12 +58,11 @@ If the links below shows up the login form, just click "Log in as guest" to star :header: "version", "cp27-cp27mu", "cp27-cp27m" :widths: 1, 3, 3 - "cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" - "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cpu_avx_mkl", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `_" + "cpu_avx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `_" + "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `_" + "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `_" + "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `_" .. _pip_dependency: @@ -79,7 +96,7 @@ FAQ ------------------------------ - paddlepaddle*.whl is not a supported wheel on this platform. - + The main cause of this issue is that your current platform is not supported. Please check that you are using Python 2.7 series. Besides, pypi only supports manylinux1 standard, you'll need to diff --git a/doc/v2/design/mkl/mkldnn.md b/doc/v2/design/mkl/mkldnn.md index 1bd2e7bc34ee79eb753b3520d97e5e7beca89b0b..bd5bcf6f67168c21cebb046a629b948d1661e75c 100644 --- a/doc/v2/design/mkl/mkldnn.md +++ b/doc/v2/design/mkl/mkldnn.md @@ -5,7 +5,7 @@ 充分展现英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。
-
+
Figure 1. PaddlePaddle on IA
@@ -42,16 +42,43 @@ Figure 1. PaddlePaddle on IA MKL，MKLML以及MKL-DNN三者关系如下表： -| Name | Open Source | License | Descriptions | -| :---------- | :--------------- | :---------- | :------------ | -| MKL | No | Proprietary | Accelerate math processing routines | -| MKLML | No | Proprietary | Small package of MKL, especially for Machine Learning | -| MKL-DNN | Yes | Apache 2.0 | Accelerate primitives processing routines especially for Deep Neural Networks | + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
NameOpen SourceLicenseDescriptions
MKLNoProprietaryAccelerate math processing routines
MKLMLNoProprietarySmall package of MKL, especially for Machine Learning
MKL-DNNYesApache 2.0Accelerate primitives processing routines especially for Deep Neural Networks
MKLML可以与MKL-DNN共同使用，以此达到最好的性能。
-
+
Figure 2. PaddlePaddle with MKL Engines
@@ -103,7 +130,7 @@ MKL-DNN的库目前只有动态库`libmkldnn.so`。 所以我们定义了一个`MKLDNNMatrix`用于管理MKL-DNN数据的不同格式以及相互之间的转换。
-
+
Figure 3. MKLDNNMatrix
@@ -113,7 +140,7 @@ Figure 3. MKLDNNMatrix 子类只需要使用定义好的接口，实现具体的函数功能即可。
-
+
Figure 4. MKLDNNLayer
@@ -150,7 +177,7 @@ Figure 4. MKLDNNLayer 所以整体上，在实现每个子类的时候就不需要关心分支的事情了。
-
+
Figure 5. Merge Gradients
diff --git a/doc/v2/images/FullyConnected.jpg b/doc/v2/images/FullyConnected.jpg new file mode 100644 index 0000000000000000000000000000000000000000..b2241f401434e527f95ee4e0e541a3f2ff78fd1e Binary files /dev/null and b/doc/v2/images/FullyConnected.jpg differ diff --git a/doc/v2/images/add_security_group.png b/doc/v2/images/add_security_group.png new file mode 100644 index 0000000000000000000000000000000000000000..bd34f46c9b0ada7027fd53e553e7d033255d25fc Binary files /dev/null and b/doc/v2/images/add_security_group.png differ diff --git a/doc/v2/images/bi_lstm.jpg b/doc/v2/images/bi_lstm.jpg new file mode 100644 index 0000000000000000000000000000000000000000..adec1606d64d6e35ffe7e62abfa9a09309b05c84 Binary files /dev/null and b/doc/v2/images/bi_lstm.jpg differ diff --git a/doc/v2/images/checkpointing.png b/doc/v2/images/checkpointing.png new file mode 100644 index 0000000000000000000000000000000000000000..c221e8474f90f37e31416cbb19c9452207a0d14c Binary files /dev/null and b/doc/v2/images/checkpointing.png differ diff --git a/doc/v2/images/create_efs.png b/doc/v2/images/create_efs.png new file mode 100644 index 0000000000000000000000000000000000000000..e5f1526033d1daf401700989af1d25919bcb7675 Binary files /dev/null and b/doc/v2/images/create_efs.png differ diff --git a/doc/v2/images/csr.png b/doc/v2/images/csr.png new file mode 100644 index 0000000000000000000000000000000000000000..3dc10b8de4f6d3f517624956b1694b689405a031 Binary files /dev/null and b/doc/v2/images/csr.png differ diff --git a/doc/v2/images/data_dispatch.png b/doc/v2/images/data_dispatch.png new file mode 100644 index 0000000000000000000000000000000000000000..5bdcc24d6a6d193cb014f8c38b362451fded5e54 Binary files /dev/null and b/doc/v2/images/data_dispatch.png differ diff --git a/doc/v2/images/dataset.graffle b/doc/v2/images/dataset.graffle new file mode 100644 index 0000000000000000000000000000000000000000..c10a423ed16a23229a9ee33d11bfc82bb59646c8 Binary files /dev/null and b/doc/v2/images/dataset.graffle differ diff --git a/doc/v2/images/dataset.png b/doc/v2/images/dataset.png new file mode 100644 index 0000000000000000000000000000000000000000..2fb7f1cce3b6dd21489392557826e95a9f207c34 Binary files /dev/null and b/doc/v2/images/dataset.png differ diff --git a/doc/v2/images/doc_en.png b/doc/v2/images/doc_en.png new file mode 100644 index 0000000000000000000000000000000000000000..ed6b9178fba91a3bdf45ae797a9924f84146fbc8 Binary files /dev/null and b/doc/v2/images/doc_en.png differ diff --git a/doc/v2/images/efs_mount.png b/doc/v2/images/efs_mount.png new file mode 100644 index 0000000000000000000000000000000000000000..0f9e3cab98445707e5e9baa18ddabe15cdf04576 Binary files /dev/null and b/doc/v2/images/efs_mount.png differ diff --git a/doc/v2/images/encoder-decoder-attention-model.png b/doc/v2/images/encoder-decoder-attention-model.png new file mode 100644 index 0000000000000000000000000000000000000000..79f911d4ba12ac0c0d1a936c9df639c302786914 Binary files /dev/null and b/doc/v2/images/encoder-decoder-attention-model.png differ diff --git a/doc/v2/images/engine.png b/doc/v2/images/engine.png new file mode 100644 index 0000000000000000000000000000000000000000..1f5f65c2cc765a514a3ba9e7b7f468e1dc4b0c3b Binary files /dev/null and b/doc/v2/images/engine.png differ diff --git a/doc/v2/images/file_storage.graffle b/doc/v2/images/file_storage.graffle new file mode 100644 index 0000000000000000000000000000000000000000..50a17e70fa255495337c529a3bf12a5c0024a5be Binary files /dev/null and b/doc/v2/images/file_storage.graffle differ diff --git a/doc/v2/images/file_storage.png b/doc/v2/images/file_storage.png new file mode 100644 index 0000000000000000000000000000000000000000..fccb4e3e7e738224c7f1584326bd5f351ce799aa Binary files /dev/null and b/doc/v2/images/file_storage.png differ diff --git a/doc/v2/images/glossary_rnn.dot b/doc/v2/images/glossary_rnn.dot new file mode 100644 index 0000000000000000000000000000000000000000..2cd0fb1820c44b0e8e0b869f9d39fcad27efa758 --- /dev/null +++ b/doc/v2/images/glossary_rnn.dot @@ -0,0 +1,42 @@ +digraph G{ + subgraph cluster_timestep0 { + label="recurrent timestep i-1" + bgcolor=lightgray + node [style=filled,color=white] + fc0_0 [label="fc 0"] + fc0_1 [label="fc 1"] + fc0_2 [label="fc 2"] + + fc0_0 -> fc0_1 + fc0_1 -> fc0_2 + } + + subgraph cluster_timestep1 { + label="recurrent timestep i" + node [style=filled]; + fc1_0 [label="fc 0"] + fc1_1 [label="fc 1"] + fc1_2 [label="fc 2"] + color=blue + + fc1_0 -> fc1_1 + fc1_1 -> fc1_2 + } + + subgraph cluster_timestep2 { + label="recurrent timestep i+1" + bgcolor=lightgray + node [style=filled,color=white] + fc2_0 [label="fc 0"] + fc2_1 [label="fc 1"] + fc2_2 [label="fc 2"] + + fc2_0 -> fc2_1 + fc2_1 -> fc2_2 + } + + + fc0_1 -> fc1_1 [style="dotted" constraint=false] + fc1_1 -> fc2_1 [style="dotted" constraint=false] + +} \ No newline at end of file diff --git a/doc/v2/images/glossary_rnn_with_memory.dot b/doc/v2/images/glossary_rnn_with_memory.dot new file mode 100644 index 0000000000000000000000000000000000000000..0f101ec2d8f15aec76c57f328046b6b55cf0c7eb --- /dev/null +++ b/doc/v2/images/glossary_rnn_with_memory.dot @@ -0,0 +1,48 @@ +digraph G{ + subgraph cluster_timestep0 { + label="recurrent timestep i-1" + bgcolor=lightgray + node [style=filled,color=white] + fc0_0 [label="fc 0"] + fc0_1 [label="fc 1"] + fc0_2 [label="fc 2"] + m0 [label="memory"] + fc0_0 -> fc0_1 + fc0_1 -> fc0_2 + fc0_1 -> m0 + m0 -> fc0_1 + } + + subgraph cluster_timestep1 { + label="recurrent timestep i" + node [style=filled]; + fc1_0 [label="fc 0"] + fc1_1 [label="fc 1"] + fc1_2 [label="fc 2"] + m1 [label="memory"] + color=blue + fc1_0 -> fc1_1 + fc1_1 -> fc1_2 + fc1_1 -> m1 + m1 -> fc1_1 + } + + subgraph cluster_timestep2 { + label="recurrent timestep i+1" + bgcolor=lightgray + node [style=filled,color=white] + fc2_0 [label="fc 0"] + fc2_1 [label="fc 1"] + fc2_2 [label="fc 2"] + m2 [label="memory"] + fc2_0 -> fc2_1 + fc2_1 -> fc2_2 + fc2_1 -> m2 + m2 -> fc2_1 + } + + + m0 -> m1 [style="dotted" constraint=false] + m1 -> m2 [style="dotted" constraint=false] + +} \ No newline at end of file diff --git a/doc/v2/images/gradients.png b/doc/v2/images/gradients.png new file mode 100644 index 0000000000000000000000000000000000000000..f031bcf8e4cec14e63075b8b9d2c7bbd9f1b1a3c Binary files /dev/null and b/doc/v2/images/gradients.png differ diff --git a/doc/v2/images/init_lock.graffle b/doc/v2/images/init_lock.graffle new file mode 100644 index 0000000000000000000000000000000000000000..fa9149f21b1311eed48ef72ec55e556559d0fc94 Binary files /dev/null and b/doc/v2/images/init_lock.graffle differ diff --git a/doc/v2/images/init_lock.png b/doc/v2/images/init_lock.png new file mode 100644 index 0000000000000000000000000000000000000000..92404ee6d6c0f9a7727952bae3c869ba338ecd7f Binary files /dev/null and b/doc/v2/images/init_lock.png differ diff --git a/doc/v2/images/k8s-paddle-arch.png b/doc/v2/images/k8s-paddle-arch.png new file mode 100644 index 0000000000000000000000000000000000000000..b3800c4fe81302d35e49f7dbacb9221c4dfa5cde Binary files /dev/null and b/doc/v2/images/k8s-paddle-arch.png differ diff --git a/doc/v2/images/layers.png b/doc/v2/images/layers.png new file mode 100644 index 0000000000000000000000000000000000000000..306f79b7a844610915eb8944128f57d2b7a3065a Binary files /dev/null and b/doc/v2/images/layers.png differ diff --git a/doc/v2/images/managed_policy.png b/doc/v2/images/managed_policy.png new file mode 100644 index 0000000000000000000000000000000000000000..c7ecda555b81d7750e9292a9ab72d2f517f76a2a Binary files /dev/null and b/doc/v2/images/managed_policy.png differ diff --git a/doc/v2/images/matrix.png b/doc/v2/images/matrix.png new file mode 100644 index 0000000000000000000000000000000000000000..c33ce9cf0335e47cc8c1253304d0fe179186e6f2 Binary files /dev/null and b/doc/v2/images/matrix.png differ diff --git a/doc/v2/images/nvvp1.png b/doc/v2/images/nvvp1.png new file mode 100644 index 0000000000000000000000000000000000000000..1af23ac3c52929b2b0645d2f9fa4d4c6db1f6e77 Binary files /dev/null and b/doc/v2/images/nvvp1.png differ diff --git a/doc/v2/images/nvvp2.png b/doc/v2/images/nvvp2.png new file mode 100644 index 0000000000000000000000000000000000000000..177c9db708da6863d1075f3e615f5962dbe18b29 Binary files /dev/null and b/doc/v2/images/nvvp2.png differ diff --git a/doc/v2/images/nvvp3.png b/doc/v2/images/nvvp3.png new file mode 100644 index 0000000000000000000000000000000000000000..d8f393667d6569b6f1e61ffccac43fae5888b6db Binary files /dev/null and b/doc/v2/images/nvvp3.png differ diff --git a/doc/v2/images/nvvp4.png b/doc/v2/images/nvvp4.png new file mode 100644 index 0000000000000000000000000000000000000000..51f2f3e183295de6cf8ddaf2b3b8a0862aa35f01 Binary files /dev/null and b/doc/v2/images/nvvp4.png differ diff --git a/doc/v2/images/overview.png b/doc/v2/images/overview.png new file mode 100644 index 0000000000000000000000000000000000000000..8fb7bbb9dd654bf363d701d0c8cd4a557043d188 Binary files /dev/null and b/doc/v2/images/overview.png differ diff --git a/doc/v2/images/paddle-cloud-in-data-center.png b/doc/v2/images/paddle-cloud-in-data-center.png new file mode 100644 index 0000000000000000000000000000000000000000..da5d1a77562480ad1d886f5f21dbd84001d3d508 Binary files /dev/null and b/doc/v2/images/paddle-cloud-in-data-center.png differ diff --git a/doc/v2/images/paddle-etcd.graffle b/doc/v2/images/paddle-etcd.graffle new file mode 100644 index 0000000000000000000000000000000000000000..f973dc9b9dbf72e9bc31e2d32822916cd281f8d9 Binary files /dev/null and b/doc/v2/images/paddle-etcd.graffle differ diff --git a/doc/v2/images/paddle-etcd.png b/doc/v2/images/paddle-etcd.png new file mode 100644 index 0000000000000000000000000000000000000000..57981ceb4b94f0f7d6dfa63f3d28c0402bf9cc31 Binary files /dev/null and b/doc/v2/images/paddle-etcd.png differ diff --git a/doc/v2/images/paddle-model-sharding.graffle b/doc/v2/images/paddle-model-sharding.graffle new file mode 100644 index 0000000000000000000000000000000000000000..fba30f0ca2b47f0d202a432821d95e55aac37ec8 Binary files /dev/null and b/doc/v2/images/paddle-model-sharding.graffle differ diff --git a/doc/v2/images/paddle-model-sharding.png b/doc/v2/images/paddle-model-sharding.png new file mode 100644 index 0000000000000000000000000000000000000000..8c3f6724ef46c6527e63a4cd8cb0b50fe0167124 Binary files /dev/null and b/doc/v2/images/paddle-model-sharding.png differ diff --git a/doc/v2/images/paddle-ps-0.png b/doc/v2/images/paddle-ps-0.png new file mode 100644 index 0000000000000000000000000000000000000000..47ef32806f182cab003da77f1556823b3f6d1721 Binary files /dev/null and b/doc/v2/images/paddle-ps-0.png differ diff --git a/doc/v2/images/paddle-ps-1.png b/doc/v2/images/paddle-ps-1.png new file mode 100644 index 0000000000000000000000000000000000000000..f3125db73096c52bac6e7c60e1675552857c0774 Binary files /dev/null and b/doc/v2/images/paddle-ps-1.png differ diff --git a/doc/v2/images/paddle-ps.graffle b/doc/v2/images/paddle-ps.graffle new file mode 100644 index 0000000000000000000000000000000000000000..0e536ffdd91cd696008b4c01bad3cb53edebdc16 Binary files /dev/null and b/doc/v2/images/paddle-ps.graffle differ diff --git a/doc/v2/images/paddle-task-queues.graffle b/doc/v2/images/paddle-task-queues.graffle new file mode 100644 index 0000000000000000000000000000000000000000..4263ed8bfd2ef0e55058828bf23f2fac3595e5fd Binary files /dev/null and b/doc/v2/images/paddle-task-queues.graffle differ diff --git a/doc/v2/images/paddle-task-queues.png b/doc/v2/images/paddle-task-queues.png new file mode 100644 index 0000000000000000000000000000000000000000..5f980266795776752cebd0c346b85c4a75a47780 Binary files /dev/null and b/doc/v2/images/paddle-task-queues.png differ diff --git a/doc/v2/images/paddle-task-states.graffle b/doc/v2/images/paddle-task-states.graffle new file mode 100644 index 0000000000000000000000000000000000000000..cf1a0b9246d9386a949d2dbb8c32fe84f72eea83 Binary files /dev/null and b/doc/v2/images/paddle-task-states.graffle differ diff --git a/doc/v2/images/paddle-task-states.png b/doc/v2/images/paddle-task-states.png new file mode 100644 index 0000000000000000000000000000000000000000..4ae43cb66c071aee9eb90d875e2373b29af9c3e0 Binary files /dev/null and b/doc/v2/images/paddle-task-states.png differ diff --git a/doc/v2/images/ps_cn.png b/doc/v2/images/ps_cn.png new file mode 100644 index 0000000000000000000000000000000000000000..f9525739cc8bc6506adde642aafa0a85ae3ebebc Binary files /dev/null and b/doc/v2/images/ps_cn.png differ diff --git a/doc/v2/images/ps_en.png b/doc/v2/images/ps_en.png new file mode 100644 index 0000000000000000000000000000000000000000..6537d3d56589ca9f19a77a50a970e4b5275e6ce0 Binary files /dev/null and b/doc/v2/images/ps_en.png differ diff --git a/doc/v2/images/pserver_and_trainer.png b/doc/v2/images/pserver_and_trainer.png new file mode 100644 index 0000000000000000000000000000000000000000..f41fe48920590333ad332bb51eb18e03dc251541 Binary files /dev/null and b/doc/v2/images/pserver_and_trainer.png differ diff --git a/doc/v2/images/pserver_init.graffle b/doc/v2/images/pserver_init.graffle new file mode 100644 index 0000000000000000000000000000000000000000..5f3f1f52be8aa7f9049a8fcd6b7c93c8560c1676 Binary files /dev/null and b/doc/v2/images/pserver_init.graffle differ diff --git a/doc/v2/images/pserver_init.png b/doc/v2/images/pserver_init.png new file mode 100644 index 0000000000000000000000000000000000000000..dfe491ff98dd7db1c336093c80964a260df2cd90 Binary files /dev/null and b/doc/v2/images/pserver_init.png differ diff --git a/doc/v2/images/route53_create_recordset.png b/doc/v2/images/route53_create_recordset.png new file mode 100644 index 0000000000000000000000000000000000000000..34e476c7beac30fcdde13fccc4cc8d08b4be3d35 Binary files /dev/null and b/doc/v2/images/route53_create_recordset.png differ diff --git a/doc/v2/images/route53_create_zone.png b/doc/v2/images/route53_create_zone.png new file mode 100644 index 0000000000000000000000000000000000000000..25b7ddb831c5cba97f4b2edddd27da3234d621af Binary files /dev/null and b/doc/v2/images/route53_create_zone.png differ diff --git a/doc/v2/images/sequence_data.png b/doc/v2/images/sequence_data.png new file mode 100644 index 0000000000000000000000000000000000000000..6e47a46b8955dfe977e85898fe3c9f33ed28de7e Binary files /dev/null and b/doc/v2/images/sequence_data.png differ diff --git a/doc/v2/images/simple_full_hierarchical_recurrent.dot b/doc/v2/images/simple_full_hierarchical_recurrent.dot new file mode 100644 index 0000000000000000000000000000000000000000..ff278a0323bb2c3ef07bf6f016a3a8df05783581 --- /dev/null +++ b/doc/v2/images/simple_full_hierarchical_recurrent.dot @@ -0,0 +1,30 @@ +digraph G { + rankdir=LR; + + subgraph cluster_t0 { + a [label="4"] + b [label="5"] + c [label="2"] + } + + subgraph cluster_t1 { + d [label="0"] + e [label="9"] + } + + subgraph cluster_t2 { + f [label="8"] + g [label="1"] + h [label="4"] + } + + a -> b; + b -> c; + c -> d [constraint=false]; + + d -> e; + e -> f [constraint=false]; + + f -> g; + g -> h; +} \ No newline at end of file diff --git a/doc/v2/images/simple_full_recurrent.dot b/doc/v2/images/simple_full_recurrent.dot new file mode 100644 index 0000000000000000000000000000000000000000..cee281fbac993afbd0cc3416570f95965cdf0a59 --- /dev/null +++ b/doc/v2/images/simple_full_recurrent.dot @@ -0,0 +1,19 @@ +digraph G { + rankdir=LR; + a [label="4"] + b [label="5"] + c [label="2"] + d [label="0"] + e [label="9"] + f [label="8"] + g [label="1"] + h [label="4"] + + a -> b; + b -> c; + c -> d; + d -> e; + e -> f; + f -> g; + g -> h; +} \ No newline at end of file diff --git a/doc/v2/images/submit-job.graffle b/doc/v2/images/submit-job.graffle new file mode 100644 index 0000000000000000000000000000000000000000..677cdfb6d9a32168bf71729eb841fa1ca0dd31d6 Binary files /dev/null and b/doc/v2/images/submit-job.graffle differ diff --git a/doc/v2/images/submit-job.png b/doc/v2/images/submit-job.png new file mode 100644 index 0000000000000000000000000000000000000000..3046a460a7ba708079e88a560debaa215a694680 Binary files /dev/null and b/doc/v2/images/submit-job.png differ diff --git a/doc/v2/images/trainer.graffle b/doc/v2/images/trainer.graffle new file mode 100644 index 0000000000000000000000000000000000000000..43415ed8cf61a5acfa34f8e56b9577f338dbf254 Binary files /dev/null and b/doc/v2/images/trainer.graffle differ diff --git a/doc/v2/images/trainer.png b/doc/v2/images/trainer.png new file mode 100644 index 0000000000000000000000000000000000000000..6537d3d56589ca9f19a77a50a970e4b5275e6ce0 Binary files /dev/null and b/doc/v2/images/trainer.png differ diff --git a/doc/v2/images/trainer_cn.png b/doc/v2/images/trainer_cn.png new file mode 100644 index 0000000000000000000000000000000000000000..f9525739cc8bc6506adde642aafa0a85ae3ebebc Binary files /dev/null and b/doc/v2/images/trainer_cn.png differ diff --git a/doc/v2/images/worker_security_group.png b/doc/v2/images/worker_security_group.png new file mode 100644 index 0000000000000000000000000000000000000000..57eb0265a34ad4223b69600d2a3dd355482e0bf5 Binary files /dev/null and b/doc/v2/images/worker_security_group.png differ diff --git a/doc/v2/images/workflow_of_CAPI.png b/doc/v2/images/workflow_of_CAPI.png new file mode 100644 index 0000000000000000000000000000000000000000..a4399ade048b3fe10d2d9c714bc34333ca068edb Binary files /dev/null and b/doc/v2/images/workflow_of_CAPI.png differ diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index 8b1ca5e16548334ed0c9a6d31b88e0805304579e..d722eec1892206ac44c49e7a12d92be0c54df8c0 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -24,6 +24,6 @@ if(NOT WITH_FLUID_ONLY) endif() add_subdirectory(testing) -if(NOT MOBILE_INFERENCE AND NOT RPI) +if(NOT MOBILE_INFERENCE AND NOT RPI AND NOT WITH_C_API) add_subdirectory(fluid) endif() diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index 340b891e41671df7e61a4a66ec538d4603bb9842..ed1e70c6460b513c1d2e1add18ac037f71d36944 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -5,11 +5,11 @@ proto_library(framework_proto SRCS framework.proto) cc_library(ddim SRCS ddim.cc DEPS eigen3 boost) cc_test(ddim_test SRCS ddim_test.cc DEPS ddim) nv_test(dim_test SRCS dim_test.cu DEPS ddim) - +cc_library(data_type SRCS data_type.cc DEPS framework_proto ddim device_context) if(WITH_GPU) - nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS ddim place memory device_context framework_proto) + nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type) else() - cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS ddim place memory device_context framework_proto) + cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type) endif() cc_test(tensor_test SRCS tensor_test.cc DEPS tensor) @@ -57,7 +57,7 @@ cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor cc_library(attribute SRCS attribute.cc DEPS framework_proto boost) cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc device_context) -cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute) +cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute glog) cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker) cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto) cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context) diff --git a/paddle/fluid/framework/block_desc.cc b/paddle/fluid/framework/block_desc.cc index 1b6f656a006489485a55b5c13b5e2de93c3da0ed..fd409ed4c0f7a504686765909e9c71692aab8824 100644 --- a/paddle/fluid/framework/block_desc.cc +++ b/paddle/fluid/framework/block_desc.cc @@ -134,6 +134,11 @@ OpDesc *BlockDesc::PrependOp() { return ops_.front().get(); } +void BlockDesc::PrependAllocatedOp(std::unique_ptr &&op_desc) { + need_update_ = true; + ops_.emplace_front(std::move(op_desc)); +} + OpDesc *BlockDesc::InsertOp(size_t index) { need_update_ = true; auto it = ops_.begin() + index; diff --git a/paddle/fluid/framework/block_desc.h b/paddle/fluid/framework/block_desc.h index eef19c4f09c60b9df18f154c85c421f5bff9413f..600601669c5d56a3ffc2fb9c804ffad5fde58f0b 100644 --- a/paddle/fluid/framework/block_desc.h +++ b/paddle/fluid/framework/block_desc.h @@ -88,6 +88,8 @@ class BlockDesc { OpDesc *PrependOp(); + void PrependAllocatedOp(std::unique_ptr &&op_desc); + OpDesc *InsertOp(size_t index); /* diff --git a/paddle/fluid/framework/data_device_transform.cc b/paddle/fluid/framework/data_device_transform.cc index 85dbb39e6fba735471446b5e5e71a612282c498a..a876725ac0f17838458065c4b4753a03e2812801 100644 --- a/paddle/fluid/framework/data_device_transform.cc +++ b/paddle/fluid/framework/data_device_transform.cc @@ -36,9 +36,11 @@ void TransDataDevice(const Tensor& in, const platform::Place& dst_place, VLOG(3) << "DeviceTransform in, src_place " << in.place() << " dst_place: " << dst_place; auto* dev_ctx = GetDeviceContext(in.place(), dst_place); - dev_ctx->Wait(); + TensorCopy(in, dst_place, *dev_ctx, out); - dev_ctx->Wait(); + if (platform::is_gpu_place(in.place()) && platform::is_cpu_place(dst_place)) { + dev_ctx->Wait(); + } } } // namespace framework diff --git a/paddle/fluid/framework/data_device_transform_test.cu b/paddle/fluid/framework/data_device_transform_test.cu index df4caa45eba2470f7528d2fbd99cca39cae0b596..a91fe5c99d397ef1bf04f6d22e988b6d3f33e500 100644 --- a/paddle/fluid/framework/data_device_transform_test.cu +++ b/paddle/fluid/framework/data_device_transform_test.cu @@ -32,8 +32,7 @@ struct AddFunctor { class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: - OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("input", "input1 of test op"); AddOutput("output", "output of test op"); AddAttr("use_gpu", "force to use gpu kernel").SetDefault(false); diff --git a/paddle/fluid/framework/data_type.cc b/paddle/fluid/framework/data_type.cc new file mode 100644 index 0000000000000000000000000000000000000000..b6b93cf422a60c1d8e9cb8b477efd562f9fe4758 --- /dev/null +++ b/paddle/fluid/framework/data_type.cc @@ -0,0 +1,102 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/data_type.h" +#include +#include +#include + +namespace paddle { +namespace framework { + +struct DataTypeMap { + std::unordered_map cpp_to_proto_; + std::unordered_map proto_to_cpp_; + std::unordered_map proto_to_str_; + std::unordered_map cpp_to_size_; +}; + +static DataTypeMap* InitDataTypeMap(); +static DataTypeMap& gDataTypeMap() { + static DataTypeMap* g_data_type_map_ = InitDataTypeMap(); + return *g_data_type_map_; +} + +template +static inline void RegisterType(DataTypeMap* map, + proto::VarType::Type proto_type, + const std::string& name) { + map->proto_to_cpp_.emplace(static_cast(proto_type), typeid(T)); + map->cpp_to_proto_.emplace(typeid(T), proto_type); + map->proto_to_str_.emplace(static_cast(proto_type), name); + map->cpp_to_size_.emplace(typeid(T), sizeof(T)); +} + +static DataTypeMap* InitDataTypeMap() { + auto retv = new DataTypeMap(); + +#define RegType(cc_type, proto_type) \ + RegisterType(retv, proto_type, #cc_type) + + // NOTE: Add your customize type here. + RegType(platform::float16, proto::VarType::FP16); + RegType(float, proto::VarType::FP32); + RegType(double, proto::VarType::FP64); + RegType(int, proto::VarType::INT32); + RegType(int64_t, proto::VarType::INT64); + RegType(bool, proto::VarType::BOOL); + RegType(size_t, proto::VarType::SIZE_T); + RegType(int16_t, proto::VarType::INT16); + RegType(uint8_t, proto::VarType::UINT8); + +#undef RegType + return retv; +} + +proto::VarType::Type ToDataType(std::type_index type) { + auto it = gDataTypeMap().cpp_to_proto_.find(type); + if (it != gDataTypeMap().cpp_to_proto_.end()) { + return it->second; + } + PADDLE_THROW("Not support %s as tensor type", type.name()); +} + +std::type_index ToTypeIndex(proto::VarType::Type type) { + auto it = gDataTypeMap().proto_to_cpp_.find(static_cast(type)); + if (it != gDataTypeMap().proto_to_cpp_.end()) { + return it->second; + } + PADDLE_THROW("Not support proto::VarType::Type(%d) as tensor type", + static_cast(type)); +} + +std::string DataTypeToString(const proto::VarType::Type type) { + auto it = gDataTypeMap().proto_to_str_.find(static_cast(type)); + if (it != gDataTypeMap().proto_to_str_.end()) { + return it->second; + } + PADDLE_THROW("Not support proto::VarType::Type(%d) as tensor type", + static_cast(type)); +} + +size_t SizeOfType(std::type_index type) { + auto it = gDataTypeMap().cpp_to_size_.find(type); + if (it != gDataTypeMap().cpp_to_size_.end()) { + return it->second; + } + PADDLE_THROW("Not support %s as tensor type", type.name()); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/data_type.h b/paddle/fluid/framework/data_type.h index 2a528eb3aa562568c92059250f2c9bc5a75ec103..491413db8c8d66fd907801131e89d9303bdef9f2 100644 --- a/paddle/fluid/framework/data_type.h +++ b/paddle/fluid/framework/data_type.h @@ -17,51 +17,14 @@ limitations under the License. */ #include #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/platform/enforce.h" + #include "paddle/fluid/platform/float16.h" namespace paddle { namespace framework { -inline proto::VarType::Type ToDataType(std::type_index type) { - if (typeid(platform::float16).hash_code() == type.hash_code()) { - return proto::VarType::FP16; - } else if (typeid(const float).hash_code() == type.hash_code()) { - // CPPLint complains Using C-style cast. Use static_cast() instead - // One fix to this is to replace float with const float because - // typeid(T) == typeid(const T) - // http://en.cppreference.com/w/cpp/language/typeid - return proto::VarType::FP32; - } else if (typeid(const double).hash_code() == type.hash_code()) { - return proto::VarType::FP64; - } else if (typeid(const int).hash_code() == type.hash_code()) { - return proto::VarType::INT32; - } else if (typeid(const int64_t).hash_code() == type.hash_code()) { - return proto::VarType::INT64; - } else if (typeid(const bool).hash_code() == type.hash_code()) { - return proto::VarType::BOOL; - } else { - PADDLE_THROW("Not supported"); - } -} - -inline std::type_index ToTypeIndex(proto::VarType::Type type) { - switch (type) { - case proto::VarType::FP16: - return typeid(platform::float16); - case proto::VarType::FP32: - return typeid(float); - case proto::VarType::FP64: - return typeid(double); - case proto::VarType::INT32: - return typeid(int); - case proto::VarType::INT64: - return typeid(int64_t); - case proto::VarType::BOOL: - return typeid(bool); - default: - PADDLE_THROW("Not support type %d", type); - } -} +extern proto::VarType::Type ToDataType(std::type_index type); +extern std::type_index ToTypeIndex(proto::VarType::Type type); template inline void VisitDataType(proto::VarType::Type type, Visitor visitor) { @@ -84,37 +47,23 @@ inline void VisitDataType(proto::VarType::Type type, Visitor visitor) { case proto::VarType::BOOL: visitor.template operator()(); break; - default: - PADDLE_THROW("Not supported"); - } -} - -inline std::string DataTypeToString(const proto::VarType::Type type) { - switch (type) { - case proto::VarType::FP16: - return "float16"; - case proto::VarType::FP32: - return "float32"; - case proto::VarType::FP64: - return "float64"; + case proto::VarType::UINT8: + visitor.template operator()(); + break; case proto::VarType::INT16: - return "int16"; - case proto::VarType::INT32: - return "int32"; - case proto::VarType::INT64: - return "int64"; - case proto::VarType::BOOL: - return "bool"; + visitor.template operator()(); + break; default: - PADDLE_THROW("Not support type %d", type); + PADDLE_THROW("Not supported %d", type); } } +extern std::string DataTypeToString(const proto::VarType::Type type); +extern size_t SizeOfType(std::type_index type); inline std::ostream& operator<<(std::ostream& out, const proto::VarType::Type& type) { out << DataTypeToString(type); return out; } - } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/data_type_transform.cc b/paddle/fluid/framework/data_type_transform.cc index c0523f3c795b103c0c27081ec5dc717f6a0f11e0..5a57ec20585c26dbcd4251464718fc819148a7a5 100644 --- a/paddle/fluid/framework/data_type_transform.cc +++ b/paddle/fluid/framework/data_type_transform.cc @@ -91,6 +91,12 @@ void TransDataType(const OpKernelType& kernel_type_for_var, case proto::VarType::BOOL: framework::VisitDataType(dst_type, CastDataType(in, out, ctx)); break; + case proto::VarType::INT16: + framework::VisitDataType(dst_type, CastDataType(in, out, ctx)); + break; + case proto::VarType::UINT8: + framework::VisitDataType(dst_type, CastDataType(in, out, ctx)); + break; default: PADDLE_THROW("Not support type %d", src_type); } diff --git a/paddle/fluid/framework/details/broadcast_op_handle.cc b/paddle/fluid/framework/details/broadcast_op_handle.cc index 2afa47c81bead6fb104f49886713bf75dc1b4dc0..d5ca061944f33939cea59a5275e691b1966194fa 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle.cc @@ -38,9 +38,7 @@ void BroadcastOpHandle::RunImpl() { out_var_handles.size(), places_.size(), "The number of output should equal to the number of places."); - // Wait input done, this Wait is asynchronous operation platform::Place - // &in_place; - WaitInputVarGenerated(*in_var_handle); + WaitInputVarGenerated(); std::vector var_scopes; for (auto *s : local_scopes_) { @@ -50,29 +48,9 @@ void BroadcastOpHandle::RunImpl() { auto *in_var = var_scopes.at(in_var_handle->scope_idx_)->FindVar(in_var_handle->name_); PADDLE_ENFORCE_NOT_NULL(in_var); - Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var); - // NOTE: The tensors' Place of input and output must be all on GPU or all on - // CPU. - for (auto *out_var_handle : out_var_handles) { - if (out_var_handle->IsTheSameVar(*in_var_handle)) { - continue; - } - auto t_out_p = out_var_handle->place_; - auto *out_var = var_scopes.at(out_var_handle->scope_idx_) - ->FindVar(out_var_handle->name_); - PADDLE_ENFORCE_NOT_NULL(out_var); - if (platform::is_gpu_place(in_tensor.place())) { - PADDLE_ENFORCE(platform::is_gpu_place(t_out_p), - "Places of input and output must be all on GPU."); - } else { - t_out_p = platform::CPUPlace(); - } - VariableVisitor::ShareDimsAndLoD(*in_var, out_var); - VariableVisitor::GetMutableTensor(out_var).mutable_data(t_out_p, - in_tensor.type()); - } + InitOutputValue(*in_var_handle, out_var_handles); if (platform::is_cpu_place(in_tensor.place())) { for (auto *out_var_handle : out_var_handles) { @@ -147,11 +125,37 @@ void BroadcastOpHandle::RunImpl() { } } -void BroadcastOpHandle::WaitInputVarGenerated(const VarHandle &in_var) { - if (in_var.generated_op_) { - for (auto &pair : dev_ctxes_) { - in_var.generated_op_->Wait(pair.second); +void BroadcastOpHandle::InitOutputValue( + const VarHandle &in_var_handle, + const std::vector &out_var_handles) const { + std::vector var_scopes; + for (auto *s : local_scopes_) { + var_scopes.emplace_back(s->FindVar(kLocalExecScopeName)->Get()); + } + auto *in_var = + var_scopes.at(in_var_handle.scope_idx_)->FindVar(in_var_handle.name_); + + Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var); + + // NOTE: The tensors' Place of input and output must be all on GPU or all on + // CPU. + for (auto *out_var_handle : out_var_handles) { + if (out_var_handle->IsTheSameVar(in_var_handle)) { + continue; } + auto t_out_p = out_var_handle->place_; + auto *out_var = var_scopes.at(out_var_handle->scope_idx_) + ->FindVar(out_var_handle->name_); + PADDLE_ENFORCE_NOT_NULL(out_var); + if (is_gpu_place(in_tensor.place())) { + PADDLE_ENFORCE(platform::is_gpu_place(t_out_p), + "Places of input and output must be all on GPU."); + } else { + t_out_p = platform::CPUPlace(); + } + VariableVisitor::ShareDimsAndLoD(*in_var, out_var); + VariableVisitor::GetMutableTensor(out_var).mutable_data(t_out_p, + in_tensor.type()); } } diff --git a/paddle/fluid/framework/details/broadcast_op_handle.h b/paddle/fluid/framework/details/broadcast_op_handle.h index 984a95008c0393eff01c2d419cc98949aed14980..629aa00cb817c4b1446e7b750ca62a7c6b1db670 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.h +++ b/paddle/fluid/framework/details/broadcast_op_handle.h @@ -57,7 +57,6 @@ struct BroadcastOpHandle : public OpHandleBase { protected: void RunImpl() override; - void WaitInputVarGenerated(const VarHandle &in_var); private: const std::vector &local_scopes_; @@ -65,6 +64,9 @@ struct BroadcastOpHandle : public OpHandleBase { #ifdef PADDLE_WITH_CUDA const platform::NCCLContextMap *nccl_ctxs_; #endif + + void InitOutputValue(const VarHandle &in_var_handle, + const std::vector &out_var_handles) const; }; } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/build_strategy.h b/paddle/fluid/framework/details/build_strategy.h new file mode 100644 index 0000000000000000000000000000000000000000..91bdfe6134ffbd1404336c9d6d1222a505084b2b --- /dev/null +++ b/paddle/fluid/framework/details/build_strategy.h @@ -0,0 +1,36 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +namespace paddle { +namespace framework { +namespace details { + +struct BuildStrategy { + enum class ReduceStrategy { kAllReduce = 0, kReduce = 1 }; + + enum class GradientScaleStrategy { + kCoeffNumDevice = 0, + kOne = 1, + kCustomized = 2, + }; + + ReduceStrategy reduce_{ReduceStrategy::kAllReduce}; + GradientScaleStrategy gradient_scale_{GradientScaleStrategy::kCoeffNumDevice}; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/computation_op_handle.cc b/paddle/fluid/framework/details/computation_op_handle.cc index 7ff0efe09387b7e5d7cfe0dfe5e129ca9914d90b..df05bb06333d6b964f2f5434c3d43214e5d2cb7a 100644 --- a/paddle/fluid/framework/details/computation_op_handle.cc +++ b/paddle/fluid/framework/details/computation_op_handle.cc @@ -26,20 +26,20 @@ ComputationOpHandle::ComputationOpHandle(const OpDesc &op_desc, Scope *scope, place_(place) {} void ComputationOpHandle::RunImpl() { - auto *cur_ctx = dev_ctxes_[place_]; - for (auto *in : inputs_) { - bool need_wait = in->generated_op_ && - in->generated_op_->DeviceContext(place_) != cur_ctx; - if (need_wait) { - in->generated_op_->Wait(cur_ctx); - } - } + WaitInputVarGenerated(place_); this->RunAndRecordEvent([this] { op_->Run(*scope_->FindVar(kLocalExecScopeName)->Get(), place_); }); } +bool ComputationOpHandle::NeedWait(VarHandleBase *in_var) { + bool need_wait = + in_var && in_var->generated_op_ && + in_var->generated_op_->DeviceContext(place_) != dev_ctxes_[place_]; + return need_wait; +} + std::string ComputationOpHandle::Name() const { return op_->Type(); } } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/computation_op_handle.h b/paddle/fluid/framework/details/computation_op_handle.h index c363b973d9abbae6bea76c2458fbe82a37a342ca..f048f973fdeb6cf7d1485cda8cea7d530d9ba465 100644 --- a/paddle/fluid/framework/details/computation_op_handle.h +++ b/paddle/fluid/framework/details/computation_op_handle.h @@ -36,6 +36,8 @@ struct ComputationOpHandle : public OpHandleBase { protected: void RunImpl() override; + bool NeedWait(VarHandleBase *in_var) override; + private: std::unique_ptr op_; Scope *scope_; diff --git a/paddle/fluid/framework/details/execution_strategy.h b/paddle/fluid/framework/details/execution_strategy.h new file mode 100644 index 0000000000000000000000000000000000000000..e8d510ec955602b5a3f73ca06caa121886eb150b --- /dev/null +++ b/paddle/fluid/framework/details/execution_strategy.h @@ -0,0 +1,29 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +namespace paddle { +namespace framework { +namespace details { + +struct ExecutionStrategy { + size_t num_threads_{0}; + bool use_event_{true}; + bool allow_op_delay_{false}; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/fetch_op_handle.cc b/paddle/fluid/framework/details/fetch_op_handle.cc index a3cae8c64cdff8594c8971b0458c443f54375f11..224e8e1f6efd7a894591ac51c929517cae7539ce 100644 --- a/paddle/fluid/framework/details/fetch_op_handle.cc +++ b/paddle/fluid/framework/details/fetch_op_handle.cc @@ -31,7 +31,7 @@ FetchOpHandle::~FetchOpHandle() { } } -void FetchOpHandle::Wait(platform::DeviceContext *waited_dev) { +void FetchOpHandle::RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx) { PADDLE_THROW("Nobody should wait FetchOp. Unexpceted Error"); } @@ -45,26 +45,21 @@ void FetchOpHandle::WaitAndMergeCPUTensors() const { } void FetchOpHandle::RunImpl() { - auto cpu_ctx = - platform::DeviceContextPool::Instance().Get(platform::CPUPlace()); - for (auto *input : inputs_) { - auto *var = static_cast(input); - if (var->generated_op_) { - var->generated_op_->Wait(cpu_ctx); - } - } + WaitInputVarGenerated(platform::CPUPlace()); + tensors_.resize(inputs_.size()); - auto *var_handle = static_cast(inputs_[0]); - auto &var_name = var_handle->name_; platform::CPUPlace cpu; auto &scopes = *local_scopes_; - for (size_t i = 0; i < scopes.size(); ++i) { - auto &scope = scopes[i]; - auto *var = - scope->FindVar(kLocalExecScopeName)->Get()->FindVar(var_name); + for (size_t i = 0; i < inputs_.size(); ++i) { + auto *var_handle = static_cast(inputs_[i]); + auto &scope = scopes.at(var_handle->scope_idx_); + auto *var = scope->FindVar(kLocalExecScopeName) + ->Get() + ->FindVar(var_handle->name_); PADDLE_ENFORCE_NOT_NULL(var, "Cannot find variable %s in execution scope", - var_name); + var_handle->name_); + auto &t = var->Get(); if (platform::is_gpu_place(t.place())) { #ifdef PADDLE_WITH_CUDA @@ -79,6 +74,15 @@ void FetchOpHandle::RunImpl() { this->WaitAndMergeCPUTensors(); } +void FetchOpHandle::WaitInputVarGenerated(const platform::Place &place) { + auto cpu_ctx = platform::DeviceContextPool::Instance().Get(place); + for (auto *input : inputs_) { + if (input->generated_op_) { + input->generated_op_->RecordWaitEventOnCtx(cpu_ctx); + } + } +} + std::string FetchOpHandle::Name() const { return "Fetch"; } } // namespace details diff --git a/paddle/fluid/framework/details/fetch_op_handle.h b/paddle/fluid/framework/details/fetch_op_handle.h index b49f3df338dc11310a4a0c27c8aaae3602373fcc..e09bdd1d3338bb175c1ddae35b53f98197b68e9a 100644 --- a/paddle/fluid/framework/details/fetch_op_handle.h +++ b/paddle/fluid/framework/details/fetch_op_handle.h @@ -33,7 +33,7 @@ struct FetchOpHandle : public OpHandleBase { ~FetchOpHandle(); - void Wait(platform::DeviceContext *waited_dev) override; + void RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx) override; void WaitAndMergeCPUTensors() const; @@ -42,6 +42,8 @@ struct FetchOpHandle : public OpHandleBase { protected: void RunImpl() override; + void WaitInputVarGenerated(const platform::Place &place) override; + private: FeedFetchList *data_; size_t offset_; diff --git a/paddle/fluid/framework/details/gather_op_handle.cc b/paddle/fluid/framework/details/gather_op_handle.cc index 3dfc972a44c62bd2adfc1331f29ffb1cca537652..2be02304566cf5dbe348fa01fc4171990eafd158 100644 --- a/paddle/fluid/framework/details/gather_op_handle.cc +++ b/paddle/fluid/framework/details/gather_op_handle.cc @@ -55,7 +55,7 @@ void GatherOpHandle::RunImpl() { "Currently, gather_op only can gather SelectedRows."); // Wait input done, this Wait is asynchronous operation - WaitInputVarGenerated(in_var_handles); + WaitInputVarGenerated(); auto &pre_in_value = pre_in_var->Get(); std::vector out_rows; @@ -111,17 +111,6 @@ void GatherOpHandle::RunImpl() { }); } -void GatherOpHandle::WaitInputVarGenerated( - const std::vector &in_var_handles) { - for (auto *in : in_var_handles) { - if (in->generated_op_) { - for (auto pair : dev_ctxes_) { - in->generated_op_->Wait(pair.second); - } - } - } -} - std::string GatherOpHandle::Name() const { return "gather"; } } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/gather_op_handle.h b/paddle/fluid/framework/details/gather_op_handle.h index c394dd7a14b07cb956aa1aedfc0df4fa25744dd7..d11ef8556aa8840949ca8dc7aa176413f70b9f22 100644 --- a/paddle/fluid/framework/details/gather_op_handle.h +++ b/paddle/fluid/framework/details/gather_op_handle.h @@ -39,7 +39,6 @@ struct GatherOpHandle : public OpHandleBase { protected: void RunImpl() override; - void WaitInputVarGenerated(const std::vector &in_var_handles); private: const std::vector &local_scopes_; diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.cc b/paddle/fluid/framework/details/multi_devices_graph_builder.cc index 21197d587b772aa046d6b3ce4394d3057ed6bf35..6b0c0a6b9fb29e641449f0c21109611cccd4e5a9 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.cc @@ -37,25 +37,26 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder( const std::string &loss_var_name, const std::unordered_set ¶ms, const std::vector &local_scopes, - platform::NCCLContextMap *nccl_ctxs, bool use_default_grad_scale) + platform::NCCLContextMap *nccl_ctxs, const BuildStrategy &strategy) : loss_var_name_(loss_var_name), places_(places), local_scopes_(local_scopes), - nccl_ctxs_(nccl_ctxs) { + nccl_ctxs_(nccl_ctxs), + strategy_(strategy) { #else MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder( const std::vector &places, const std::string &loss_var_name, const std::unordered_set ¶ms, - const std::vector &local_scopes, bool use_default_grad_scale) + const std::vector &local_scopes, const BuildStrategy &strategy) : loss_var_name_(loss_var_name), places_(places), - local_scopes_(local_scopes) { + local_scopes_(local_scopes), + strategy_(strategy) { #endif for (auto &p : params) { grad_names_.insert(GradVarName(p)); } - use_default_grad_scale_ = use_default_grad_scale; } void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result, @@ -97,7 +98,7 @@ bool MultiDevSSAGraphBuilder::IsDistTrainOp(const OpDesc &op, return false; }; - if (op.Type() == "split") { + if (op.Type() == "split" || op.Type() == "split_byref") { return checker(op.OutputArgumentNames(), send_op->InputArgumentNames()); } else if (op.Type() == "concat") { return checker(op.InputArgumentNames(), send_op->OutputArgumentNames()); @@ -124,6 +125,12 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( // Find "send" op first for split is in front of send. OpDesc *send_op = GetSendOpDesc(program); + size_t cur_device_id = 0; + std::vector> var_name_on_devices; + std::vector> bcast_var_name_set; + var_name_on_devices.resize(places_.size()); + bcast_var_name_set.resize(places_.size()); + bool is_forwarding = true; for (auto *op : program.Block(0).AllOps()) { if (op->Type() == "send") { @@ -134,22 +141,42 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( CreateComputationalOps(&result, *op, 1); } else if (IsScaleLossOp(*op)) { // user can customize loss@grad if not use_default_grad_scale_ - if (use_default_grad_scale_) { + if (strategy_.gradient_scale_ != + BuildStrategy::GradientScaleStrategy::kCustomized) { CreateScaleLossGradOp(&result); } is_forwarding = false; } else { - CreateComputationalOps(&result, *op, places_.size()); + int op_dev_id = GetOpDeviceID(var_name_on_devices, *op); + if (op_dev_id == -1) { // var on all device + CreateComputationalOps(&result, *op, places_.size()); + } else { + CreateComputationalOp(&result, *op, op_dev_id); + for (auto &var_name : op->OutputArgumentNames()) { + var_name_on_devices[op_dev_id].emplace(var_name); + } + } if (!is_forwarding && places_.size() > 1) { // Currently, we assume that once gradient is generated, it can be // broadcast, and each gradient is only broadcast once. for (auto &og : op->OutputArgumentNames()) { if (IsParameterGradientOnce(og, &og_has_been_broadcast)) { - if (IsSparseGradient(var_types, og)) { - CreateReduceOp(&result, og, 0); - CreateBroadcastOp(&result, og, 0); - } else { - InsertNCCLAllReduceOp(&result, og); + switch (strategy_.reduce_) { + case BuildStrategy::ReduceStrategy::kReduce: + CreateReduceOp(&result, og, cur_device_id); + var_name_on_devices[cur_device_id].emplace(og); + bcast_var_name_set[cur_device_id].emplace( + og.substr(0, og.size() - strlen(kGradVarSuffix))); + cur_device_id = (cur_device_id + 1) % places_.size(); + break; + case BuildStrategy::ReduceStrategy::kAllReduce: + if (IsSparseGradient(var_types, og)) { + CreateReduceOp(&result, og, 0); + CreateBroadcastOp(&result, og, 0); + } else { + InsertNCCLAllReduceOp(&result, og); + } + break; } } } @@ -157,6 +184,13 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( } } + // Insert BCast Ops + for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) { + auto &to_bcast_set = bcast_var_name_set[dev_id]; + for (auto &bcast_name : to_bcast_set) { + CreateBroadcastOp(&result, bcast_name, dev_id); + } + } /* Dependency graph has been constructed. However, there are still data harzaeds need to be handled. @@ -265,6 +299,26 @@ bool MultiDevSSAGraphBuilder::IsParameterGradientOnce( return is_pg_once; } +int MultiDevSSAGraphBuilder::GetOpDeviceID( + const std::vector> &var_name_on_devices, + const OpDesc &op) const { + if (strategy_.reduce_ != BuildStrategy::ReduceStrategy::kReduce) { + return -1; + } + + int var_dev_id = -1; + for (auto &var_name : op.InputArgumentNames()) { + if (var_dev_id != -1) break; + for (size_t i = 0; i < var_name_on_devices.size(); ++i) { + if (var_name_on_devices[i].count(var_name)) { + var_dev_id = static_cast(i); + break; + } + } + } + return var_dev_id; +} + void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(SSAGraph *result) const { for (size_t i = 0; i < places_.size(); ++i) { // Insert ScaleCost OpHandle diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.h b/paddle/fluid/framework/details/multi_devices_graph_builder.h index 674e2779a112c26d05b84cd54df2c826e9a63373..4f708521884247fc013f0ae336ab683c3fe7ef2f 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.h +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.h @@ -17,6 +17,7 @@ #include #include +#include "paddle/fluid/framework/details/build_strategy.h" #include "paddle/fluid/framework/details/ssa_graph_builder.h" namespace paddle { @@ -36,13 +37,13 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder { const std::unordered_set ¶ms, const std::vector &local_scopes, platform::NCCLContextMap *nccl_ctxs, - bool use_default_grad_scale); + const BuildStrategy &strategy); #else MultiDevSSAGraphBuilder(const std::vector &places, const std::string &loss_var_name, const std::unordered_set ¶ms, const std::vector &local_scopes, - bool use_default_grad_scale); + const BuildStrategy &strategy); #endif std::unique_ptr Build(const ProgramDesc &program) const override; @@ -60,7 +61,6 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder { #ifdef PADDLE_WITH_CUDA platform::NCCLContextMap *nccl_ctxs_; #endif - bool use_default_grad_scale_; bool IsScaleLossOp(const OpDesc &op) const; @@ -84,6 +84,10 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder { const std::string &og, std::unordered_set *og_has_been_broadcast) const; + int GetOpDeviceID( + const std::vector> &var_name_on_devices, + const OpDesc &op) const; + void InsertNCCLAllReduceOp(SSAGraph *result, const std::string &og) const; void CreateBroadcastOp(SSAGraph *result, const std::string &p_name, @@ -98,6 +102,9 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder { bool IsSparseGradient( const std::unordered_map &var_types, const std::string &og) const; + + private: + BuildStrategy strategy_; }; } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc index 16aa5d067ab7a222af8fbb6ca8ec18222ecd799b..95aa599cd3e403e9cc66b2b5ad35d0d214d1ab5b 100644 --- a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc +++ b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc @@ -34,12 +34,7 @@ void NCCLAllReduceOpHandle::RunImpl() { return; // No need to all reduce when GPU count = 1; } else { // Wait input done - for (auto *in : inputs_) { - auto &p = static_cast(in)->place_; - if (in->generated_op_) { - in->generated_op_->Wait(dev_ctxes_[p]); - } - } + WaitInputVarGenerated(); auto &var_name = static_cast(this->inputs_[0])->name_; int dtype = -1; diff --git a/paddle/fluid/framework/details/op_handle_base.cc b/paddle/fluid/framework/details/op_handle_base.cc index 534d77860f87be08c8834efd373d90eb199ed6a2..6b064650b4f09737836bda4a43fa421720077929 100644 --- a/paddle/fluid/framework/details/op_handle_base.cc +++ b/paddle/fluid/framework/details/op_handle_base.cc @@ -56,15 +56,15 @@ void OpHandleBase::Run(bool use_event) { RunImpl(); } -void OpHandleBase::Wait(platform::DeviceContext *waited_dev) { +void OpHandleBase::RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx) { #ifdef PADDLE_WITH_CUDA - if (platform::is_cpu_place(waited_dev->GetPlace()) || events_.empty()) { + if (platform::is_cpu_place(waited_ctx->GetPlace()) || events_.empty()) { for (auto &dev_ctx : dev_ctxes_) { dev_ctx.second->Wait(); } } else { auto stream = - static_cast(waited_dev)->stream(); + static_cast(waited_ctx)->stream(); for (auto &ev : events_) { PADDLE_ENFORCE(cudaStreamWaitEvent(stream, ev.second, 0)); } @@ -86,6 +86,28 @@ void OpHandleBase::AddOutput(VarHandleBase *out) { out->generated_op_ = this; } +void OpHandleBase::WaitInputVarGenerated() { + for (auto in_var : inputs_) { + if (NeedWait(in_var)) { + for (auto &pair : dev_ctxes_) { + in_var->generated_op_->RecordWaitEventOnCtx(pair.second); + } + } + } +} + +void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) { + for (auto *in : inputs_) { + if (NeedWait(in)) { + in->generated_op_->RecordWaitEventOnCtx(dev_ctxes_[place]); + } + } +} + +bool OpHandleBase::NeedWait(VarHandleBase *in_var) { + return in_var && in_var->generated_op_; +} + void OpHandleBase::RunAndRecordEvent(const std::function &callback) { #ifdef PADDLE_WITH_CUDA if (!events_.empty()) { // Use event diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h index 00f213f3ed294adcce7c540e3ff346de8e2be7fb..8f94206a87dbae8a81727ca48718886bbabbe25c 100644 --- a/paddle/fluid/framework/details/op_handle_base.h +++ b/paddle/fluid/framework/details/op_handle_base.h @@ -38,12 +38,24 @@ class OpHandleBase { void Run(bool use_event); - virtual void Wait(platform::DeviceContext *waited_dev); + virtual void RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx); void AddInput(VarHandleBase *in); void AddOutput(VarHandleBase *out); + // This method adds the wait events of all the input on all the device + // context. + // NODE: This Wait is asynchronous operation. + virtual void WaitInputVarGenerated(); + + // This method adds the wait events of all the input on the specified device + // context. + // NODE: This Wait is asynchronous operation. + virtual void WaitInputVarGenerated(const platform::Place &place); + + virtual bool NeedWait(VarHandleBase *in_var); + // If the Op involves data transfer of multiple devices that // will likely block other computations. virtual bool IsMultiDeviceTransfer() { return false; } @@ -58,6 +70,14 @@ class OpHandleBase { const std::vector &Inputs() const { return inputs_; } + size_t NoDupInputSize() const { + std::unordered_set res; + for (auto *var : inputs_) { + res.emplace(var); + } + return res.size(); + } + const std::vector &Outputs() const { return outputs_; } protected: diff --git a/paddle/fluid/framework/details/op_registry.h b/paddle/fluid/framework/details/op_registry.h index 06603db31e0092382c0cc05482a038473d647ef1..1c4b059cd0aeff803ca7436d3f198e97a06cd012 100644 --- a/paddle/fluid/framework/details/op_registry.h +++ b/paddle/fluid/framework/details/op_registry.h @@ -95,7 +95,10 @@ struct OpInfoFiller { void operator()(const char* op_type, OpInfo* info) const { info->proto_ = new proto::OpProto; info->checker_ = new OpAttrChecker(); - auto maker = T(info->proto_, info->checker_); + T maker; + maker.SetProto(info->proto_); + maker.SetChecker(info->checker_); + maker.Make(); maker.Validate(); info->proto_->set_type(op_type); PADDLE_ENFORCE( diff --git a/paddle/fluid/framework/details/reduce_op_handle.cc b/paddle/fluid/framework/details/reduce_op_handle.cc index 1bb04c1dfca107f4b7ce4c599e9aa132de3e5985..7160e346dad0615e2fd32b70c096880af0359e1a 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.cc +++ b/paddle/fluid/framework/details/reduce_op_handle.cc @@ -51,7 +51,7 @@ void ReduceOpHandle::RunImpl() { PADDLE_ENFORCE_NOT_NULL(pre_in_var); // Wait input done, this Wait is asynchronous operation - WaitInputVarGenerated(in_var_handles); + WaitInputVarGenerated(); // NOTE: The Places of all input tensor must be all on CPU or all on GPU. std::vector in_places; // used to get dev_ctx @@ -80,19 +80,21 @@ void ReduceOpHandle::RunImpl() { } if (pre_in_var->IsType()) { - std::vector in_selected_rows = - GetInputValues(in_var_handles, var_scopes); - - GatherSelectedRows(in_selected_rows, in_places, dev_ctxes_, t_out_p, - out_var->GetMutable()); + this->RunAndRecordEvent([&] { + std::vector in_selected_rows = + GetInputValues(in_var_handles, var_scopes); + GatherSelectedRows(in_selected_rows, in_places, dev_ctxes_, t_out_p, + out_var->GetMutable()); + }); } else { std::vector lod_tensors = GetInputValues(in_var_handles, var_scopes); - if (paddle::platform::is_cpu_place(lod_tensors[0]->place())) { - ReduceLoDTensor func(lod_tensors, - out_var->GetMutable()); - VisitDataType(ToDataType(lod_tensors[0]->type()), func); + this->RunAndRecordEvent([&] { + ReduceLoDTensor func(lod_tensors, + out_var->GetMutable()); + VisitDataType(ToDataType(lod_tensors[0]->type()), func); + }); } else if (paddle::platform::is_gpu_place(lod_tensors[0]->place())) { #ifdef PADDLE_WITH_CUDA auto pre_in = pre_in_var->Get(); @@ -157,17 +159,6 @@ std::vector ReduceOpHandle::GetInputValues( return in_selected_rows; } -void ReduceOpHandle::WaitInputVarGenerated( - const std::vector &in_var_handles) { - for (auto *in : in_var_handles) { - if (in->generated_op_) { - for (auto pair : dev_ctxes_) { - in->generated_op_->Wait(pair.second); - } - } - } -} - std::string ReduceOpHandle::Name() const { return "reduce"; } } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/reduce_op_handle.h b/paddle/fluid/framework/details/reduce_op_handle.h index 59731d348d17755fbd8bf3b6fa29b32bdefaf71e..c652a2f4eb0f9b73cb19ebbd9d0809210b280ad3 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.h +++ b/paddle/fluid/framework/details/reduce_op_handle.h @@ -60,8 +60,6 @@ struct ReduceOpHandle : public OpHandleBase { protected: void RunImpl() override; - void WaitInputVarGenerated(const std::vector &in_var_handles); - template std::vector GetInputValues( const std::vector &in_var_handles, diff --git a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc index 1cd3113030086104e7fc5c4ba3364a5ff027632b..d9c387e79dc71288e7330597fed57171d447f31b 100644 --- a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc +++ b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc @@ -29,6 +29,7 @@ ScaleLossGradOpHandle::ScaleLossGradOpHandle(size_t num_dev, Scope *scope, ScaleLossGradOpHandle::~ScaleLossGradOpHandle() {} void ScaleLossGradOpHandle::RunImpl() { + // Doesn't wait any event std::string var_name = static_cast(this->outputs_[0])->name_; auto &local_scope = *scope_->FindVar(kLocalExecScopeName)->Get(); diff --git a/paddle/fluid/framework/details/send_op_handle.cc b/paddle/fluid/framework/details/send_op_handle.cc index bd97c5260dbba935e422793e0aa6aac8b6875627..7109659dd7001f91e7674ac7bebbe3a59794cfc0 100644 --- a/paddle/fluid/framework/details/send_op_handle.cc +++ b/paddle/fluid/framework/details/send_op_handle.cc @@ -26,6 +26,7 @@ SendOpHandle::SendOpHandle(const framework::OpDesc &op_desc, place_(place) {} void SendOpHandle::RunImpl() { + // TODO(wuyi): need further analysis whether wait VarDummyHandle. // Wait input done for (auto *in : inputs_) { auto &p = static_cast(in)->place_; @@ -33,7 +34,7 @@ void SendOpHandle::RunImpl() { continue; } if (in->generated_op_) { - in->generated_op_->Wait(dev_ctxes_[p]); + in->generated_op_->RecordWaitEventOnCtx(dev_ctxes_[p]); } } auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get(); diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc index 5e6ed5cb7cdc534332d402380458f39aecd841b8..815f739371e77d953a28be99b38ec1b8ff26506c 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -14,24 +14,21 @@ #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h" -#include "paddle/fluid/framework/details/fetch_op_handle.h" - namespace paddle { namespace framework { namespace details { ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor( - size_t num_threads, bool use_event, - const std::vector &local_scopes, + const ExecutionStrategy &strategy, const std::vector &local_scopes, const std::vector &places, - std::unique_ptr &&graph, bool allow_op_delay) + std::unique_ptr &&graph) : SSAGraphExecutor(std::move(graph)), - pool_(num_threads >= 2 ? new ::ThreadPool(num_threads) : nullptr), + pool_(strategy.num_threads_ >= 2 ? new ::ThreadPool(strategy.num_threads_) + : nullptr), local_scopes_(local_scopes), places_(places), fetch_ctxs_(places), - use_event_(use_event), running_ops_(0), - allow_op_delay_(allow_op_delay) {} + strategy_(strategy) {} FeedFetchList ThreadedSSAGraphExecutor::Run( const std::vector &fetch_tensors) { @@ -45,73 +42,33 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( // Should revisit it if overlapping is available. std::unordered_set delayed_ops; - auto InsertPendingVar = [&pending_vars, &ready_vars](VarHandleBase &var) { - pending_vars.insert(&var); - if (var.generated_op_ == nullptr) { - ready_vars.Push(&var); - } - }; - - auto InsertPendingOp = [&pending_ops](OpHandleBase &op_instance) { - pending_ops.insert({&op_instance, op_instance.Inputs().size()}); - }; - // Transform SSAGraph to pending_ops & pending_vars for (auto &var_map : graph_->vars_) { for (auto &name_pair : var_map) { for (auto &version_pair : name_pair.second) { - InsertPendingVar(*version_pair); + InsertPendingVar(&pending_vars, &ready_vars, version_pair.get()); } } } for (auto &var : graph_->dep_vars_) { - InsertPendingVar(*var); + InsertPendingVar(&pending_vars, &ready_vars, var.get()); } for (auto &op : graph_->ops_) { if (op->Inputs().empty()) { // Special case, Op has no input. ready_ops.insert(op.get()); } else { - InsertPendingOp(*op); + InsertPendingOp(&pending_ops, op.get()); } } // Step 2. Insert FetchOps std::vector> fetch_ops; - FeedFetchList fetch_data(fetch_tensors.size()); - - std::unordered_map> fetched_vars; - - for (auto &fetch_var_name : fetch_tensors) { - for (auto &var_map : graph_->vars_) { - auto it = var_map.find(fetch_var_name); - if (it != var_map.end()) { - fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get()); - } - } - } - std::unordered_set> fetch_dependencies; - for (size_t i = 0; i < fetch_tensors.size(); ++i) { - auto &var_name = fetch_tensors[i]; - auto &vars = fetched_vars.at(var_name); - auto *op = new FetchOpHandle(&fetch_data, i, &local_scopes_); - fetch_ops.emplace_back(op); - - for (auto &p : places_) { - op->SetDeviceContext(p, fetch_ctxs_.Get(p)); - } - - for (auto *var : vars) { - op->AddInput(var); - } + FeedFetchList fetch_data(fetch_tensors.size()); - auto *fetch_dummy = new DummyVarHandle(); - op->AddOutput(fetch_dummy); - fetch_dependencies.emplace(fetch_dummy); - InsertPendingVar(*fetch_dummy); - InsertPendingOp(*op); - } + InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &pending_ops, + &pending_vars, &ready_vars, &fetch_data); auto run_all_ops = [&](std::unordered_set &set) { for (auto *op : set) { @@ -128,7 +85,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( // // NOTE: DelayedOps have a lower priority. It will be scheduled after all // ready_ops have been performed. - if (ready_ops.empty() && allow_op_delay_ && running_ops_ == 0) { + if (ready_ops.empty() && strategy_.allow_op_delay_ && running_ops_ == 0) { run_all_ops(delayed_ops); } else { run_all_ops(ready_ops); @@ -155,7 +112,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( auto &deps = pending_ops[op]; --deps; if (deps == 0) { - if (op->IsMultiDeviceTransfer() && allow_op_delay_) { + if (op->IsMultiDeviceTransfer() && strategy_.allow_op_delay_) { delayed_ops.insert(op); } else { ready_ops.insert(op); @@ -174,12 +131,66 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( return fetch_data; } +void ThreadedSSAGraphExecutor::InsertFetchOps( + const std::vector &fetch_tensors, + std::vector> *fetch_ops, + std::unordered_set> *fetch_dependencies, + std::unordered_map *pending_ops, + std::unordered_set *pending_vars, + BlockingQueue *ready_vars, FeedFetchList *fetch_data) { + std::unordered_map> fetched_vars; + + for (auto &fetch_var_name : fetch_tensors) { + for (auto &var_map : graph_->vars_) { + auto it = var_map.find(fetch_var_name); + if (it != var_map.end()) { + fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get()); + } + } + } + + for (size_t i = 0; i < fetch_tensors.size(); ++i) { + auto &var_name = fetch_tensors[i]; + auto &vars = fetched_vars.at(var_name); + auto *op = new FetchOpHandle(fetch_data, i, &local_scopes_); + fetch_ops->emplace_back(op); + + for (auto &p : places_) { + op->SetDeviceContext(p, fetch_ctxs_.Get(p)); + } + + for (auto *var : vars) { + op->AddInput(var); + } + + auto *fetch_dummy = new DummyVarHandle(); + op->AddOutput(fetch_dummy); + fetch_dependencies->emplace(fetch_dummy); + this->InsertPendingVar(pending_vars, ready_vars, fetch_dummy); + this->InsertPendingOp(pending_ops, op); + } +} + +void ThreadedSSAGraphExecutor::InsertPendingOp( + std::unordered_map *pending_ops, + OpHandleBase *op_instance) const { + pending_ops->insert({op_instance, op_instance->NoDupInputSize()}); +} + +void ThreadedSSAGraphExecutor::InsertPendingVar( + std::unordered_set *pending_vars, + BlockingQueue *ready_vars, VarHandleBase *var) const { + pending_vars->insert(var); + if (var->generated_op_ == nullptr) { + ready_vars->Push(var); + } +} void ThreadedSSAGraphExecutor::RunOp( BlockingQueue *ready_var_q, details::OpHandleBase *op) { auto op_run = [ready_var_q, op, this] { try { VLOG(10) << op << " " << op->Name() << " : " << op->DebugString(); - op->Run(use_event_); + op->Run(strategy_.use_event_); VLOG(10) << op << " " << op->Name() << " Done "; running_ops_--; ready_var_q->Extend(op->Outputs()); diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h index d089b79d91327e38408439a8019ec5189ff6d189..1f7f88d75218e757e4555ad093f3cd6558f624dd 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h @@ -23,6 +23,8 @@ #include #include "ThreadPool.h" // ThreadPool in thrird party #include "paddle/fluid/framework/blocking_queue.h" +#include "paddle/fluid/framework/details/execution_strategy.h" +#include "paddle/fluid/framework/details/fetch_op_handle.h" #include "paddle/fluid/framework/details/ssa_graph_executor.h" namespace paddle { @@ -33,11 +35,10 @@ namespace details { class ThreadedSSAGraphExecutor : public SSAGraphExecutor { public: - ThreadedSSAGraphExecutor(size_t num_threads, bool use_event, + ThreadedSSAGraphExecutor(const ExecutionStrategy &strategy, const std::vector &local_scopes, const std::vector &places, - std::unique_ptr &&graph, - bool allow_op_delay); + std::unique_ptr &&graph); // Run a SSAGraph by a thread pool // Use topological sort algorithm @@ -54,10 +55,26 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { std::vector local_scopes_; std::vector places_; platform::DeviceContextPool fetch_ctxs_; - const bool use_event_; std::unique_ptr exception_; std::atomic running_ops_; - bool allow_op_delay_; + + void InsertPendingOp(std::unordered_map *pending_ops, + OpHandleBase *op_instance) const; + + void InsertPendingVar(std::unordered_set *pending_vars, + BlockingQueue *ready_vars, + VarHandleBase *var) const; + + void InsertFetchOps( + const std::vector &fetch_tensors, + std::vector> *fetch_ops, + std::unordered_set> *fetch_dependencies, + std::unordered_map *pending_ops, + std::unordered_set *pending_vars, + BlockingQueue *ready_vars, FeedFetchList *fetch_data); + + private: + ExecutionStrategy strategy_; }; } // namespace details diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index ce91d7a82674364560b8065277b28b51ae1b303a..4e431561f81b2a84c06dff9fcb041317ebc84ae3 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -228,7 +228,8 @@ static bool has_fetch_operators( void Executor::Run(const ProgramDesc& program, Scope* scope, std::map* feed_targets, std::map* fetch_targets, - bool create_vars, const std::string& feed_holder_name, + bool create_local_scope, bool create_vars, + const std::string& feed_holder_name, const std::string& fetch_holder_name) { platform::RecordBlock b(kProgramId); bool has_feed_ops = @@ -290,8 +291,9 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, } auto ctx = Prepare(*copy_program, 0); - RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets, create_vars, - feed_holder_name, fetch_holder_name); + RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets, + create_local_scope, create_vars, feed_holder_name, + fetch_holder_name); } std::unique_ptr Executor::Prepare( @@ -366,8 +368,9 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, void Executor::RunPreparedContext( ExecutorPrepareContext* ctx, Scope* scope, std::map* feed_targets, - std::map* fetch_targets, bool create_vars, - const std::string& feed_holder_name, const std::string& fetch_holder_name) { + std::map* fetch_targets, bool create_local_scope, + bool create_vars, const std::string& feed_holder_name, + const std::string& fetch_holder_name) { auto& global_block = ctx->prog_.Block(ctx->block_id_); PADDLE_ENFORCE( @@ -387,7 +390,7 @@ void Executor::RunPreparedContext( } } - RunPreparedContext(ctx, scope, create_vars, create_vars); + RunPreparedContext(ctx, scope, create_local_scope, create_vars); // obtain the data of fetch_targets from fetch_holder for (auto* op : global_block.AllOps()) { diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index 4a3d637e2d79f8cbd83412eea2d73e4b497ef1e7..0c3c23611d95e0da67cabfb8fb2755a4a52c991b 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -57,7 +57,7 @@ class Executor { void Run(const ProgramDesc& program, Scope* scope, std::map* feed_targets, std::map* fetch_targets, - bool create_vars = true, + bool create_local_scope = true, bool create_vars = true, const std::string& feed_holder_name = "feed", const std::string& fetch_holder_name = "fetch"); @@ -76,6 +76,7 @@ class Executor { void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, std::map* feed_targets, std::map* fetch_targets, + bool create_local_scope = true, bool create_vars = true, const std::string& feed_holder_name = "feed", const std::string& fetch_holder_name = "fetch"); diff --git a/paddle/fluid/framework/framework.proto b/paddle/fluid/framework/framework.proto index 96f53dc1bc8747e1b8ea84166614f98ff363ae5e..d35125fe8c3c8018c38650dc87b2b1474ded6058 100644 --- a/paddle/fluid/framework/framework.proto +++ b/paddle/fluid/framework/framework.proto @@ -101,6 +101,9 @@ message VarType { FP16 = 4; FP32 = 5; FP64 = 6; + // Tensor is used in C++. + SIZE_T = 19; + UINT8 = 20; // Other types that may need additional descriptions LOD_TENSOR = 7; diff --git a/paddle/fluid/framework/lod_tensor_test.cc b/paddle/fluid/framework/lod_tensor_test.cc index 77e5ec4c7dd14b7ebb6d606b8c401ee714259d40..2ceffc93319359683e87e7fec2d18784c9bf02f3 100644 --- a/paddle/fluid/framework/lod_tensor_test.cc +++ b/paddle/fluid/framework/lod_tensor_test.cc @@ -228,11 +228,12 @@ TEST(LoD, CheckAbsLoD) { ASSERT_FALSE(CheckAbsLoD(abs_lod0)); } -TEST(LoDTensor, RecordIO) { +template +static void TestRecordIO() { LoDTensor tensor; - int* tmp = tensor.mutable_data(make_ddim({4, 5}), platform::CPUPlace()); + T* tmp = tensor.mutable_data(make_ddim({4, 5}), platform::CPUPlace()); for (int i = 0; i < 20; ++i) { - tmp[i] = i; + tmp[i] = static_cast(i); } std::stringstream* stream = new std::stringstream(); @@ -247,7 +248,7 @@ TEST(LoDTensor, RecordIO) { auto assert_tensor_ok = [](const LoDTensor& tensor) { for (int i = 0; i < 20; ++i) { - ASSERT_EQ(tensor.data()[i], i); + ASSERT_EQ(tensor.data()[i], static_cast(i)); } }; @@ -265,5 +266,13 @@ TEST(LoDTensor, RecordIO) { } } +TEST(LoDTensor, RecordIO) { + TestRecordIO(); + TestRecordIO(); + TestRecordIO(); + TestRecordIO(); + TestRecordIO(); +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/op_kernel_type_test.cc b/paddle/fluid/framework/op_kernel_type_test.cc index d37ce149ce3df63692b41289bb03448d54e392f5..db95861c510b52a5b52229541434e6437d3fb9f4 100644 --- a/paddle/fluid/framework/op_kernel_type_test.cc +++ b/paddle/fluid/framework/op_kernel_type_test.cc @@ -27,7 +27,7 @@ TEST(OpKernelType, ToString) { LibraryType::kCUDNN); ASSERT_EQ(paddle::framework::KernelTypeToString(op_kernel_type), - "data_type[float32]:data_layout[NCHW]:place[CPUPlace]:library_type[" + "data_type[float]:data_layout[NCHW]:place[CPUPlace]:library_type[" "CUDNN]"); } diff --git a/paddle/fluid/framework/op_proto_maker.h b/paddle/fluid/framework/op_proto_maker.h index 0beb57ce1609d2e90c05d3255647bd321bc1f6a9..b01a520bba19c1be32363a1a5c381666c82e6afc 100644 --- a/paddle/fluid/framework/op_proto_maker.h +++ b/paddle/fluid/framework/op_proto_maker.h @@ -14,56 +14,57 @@ limitations under the License. */ #pragma once #include +#include "glog/logging.h" #include "paddle/fluid/framework/attribute.h" #include "paddle/fluid/framework/framework.pb.h" - namespace paddle { namespace framework { // this class not only make proto but also init attribute checkers. class OpProtoAndCheckerMaker { public: - using OpProto = proto::OpProto; - using OpAttrChecker = framework::OpAttrChecker; - OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : proto_(proto), op_checker_(op_checker) {} + virtual void Make() = 0; virtual ~OpProtoAndCheckerMaker() { - PADDLE_ENFORCE(validated_, "should call Validate after build"); + CHECK(validated_) << "should call Validate after build"; } + void SetProto(proto::OpProto *proto) { proto_ = proto; } + + void SetChecker(OpAttrChecker *attr_checker) { op_checker_ = attr_checker; } + void Validate(); protected: struct VariableBuilder { - OpProto::Var* var_; + proto::OpProto::Var *var_; - VariableBuilder& AsDuplicable() { + VariableBuilder &AsDuplicable() { var_->set_duplicable(true); return *this; } - VariableBuilder& AsIntermediate() { + VariableBuilder &AsIntermediate() { var_->set_intermediate(true); return *this; } - VariableBuilder& AsDispensable() { + VariableBuilder &AsDispensable() { var_->set_dispensable(true); return *this; } }; - VariableBuilder AddInput(const std::string& name, const std::string& comment); + VariableBuilder AddInput(const std::string &name, const std::string &comment); - VariableBuilder AddOutput(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, + TypedAttrChecker &AddAttr(const std::string &name, + const std::string &comment, bool generated = false) { - auto* attr = proto_->add_attrs(); + auto *attr = proto_->add_attrs(); attr->set_name(name); attr->set_comment(comment); attr->set_generated(generated); @@ -71,21 +72,14 @@ class OpProtoAndCheckerMaker { return op_checker_->AddAttrChecker(name); } - void AddComment(const std::string& comment) { proto_->set_comment(comment); } + void AddComment(const std::string &comment) { proto_->set_comment(comment); } private: void CheckNoDuplicatedInOutAttrs(); - OpProto* proto_; - OpAttrChecker* op_checker_; + proto::OpProto *proto_; + OpAttrChecker *op_checker_; bool validated_{false}; }; - -class NOPMaker : public OpProtoAndCheckerMaker { - public: - NOPMaker(OpProto* proto, framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) {} -}; - } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/op_proto_maker_test.cc b/paddle/fluid/framework/op_proto_maker_test.cc index a8d8c6386af940d4a14016b30de344e1c7877b22..9b5badbc81f9ddf083c81f57f5355e07a8e5e4a2 100644 --- a/paddle/fluid/framework/op_proto_maker_test.cc +++ b/paddle/fluid/framework/op_proto_maker_test.cc @@ -18,9 +18,7 @@ limitations under the License. */ class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { public: - TestAttrProtoMaker(paddle::framework::proto::OpProto* proto, - paddle::framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddAttr("scale", "scale of test op"); AddAttr("scale", "scale of test op"); } @@ -29,15 +27,16 @@ class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { TEST(ProtoMaker, DuplicatedAttr) { paddle::framework::proto::OpProto op_proto; paddle::framework::OpAttrChecker op_checker; - auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker); + TestAttrProtoMaker proto_maker; + proto_maker.SetProto(&op_proto); + proto_maker.SetChecker(&op_checker); + proto_maker.Make(); ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); } class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { public: - TestInOutProtoMaker(paddle::framework::proto::OpProto* proto, - paddle::framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("input", "input of test op"); AddInput("input", "input of test op"); } @@ -46,6 +45,9 @@ class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { TEST(ProtoMaker, DuplicatedInOut) { paddle::framework::proto::OpProto op_proto; paddle::framework::OpAttrChecker op_checker; - auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker); + TestAttrProtoMaker proto_maker; + proto_maker.SetProto(&op_proto); + proto_maker.SetChecker(&op_checker); + proto_maker.Make(); ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); } diff --git a/paddle/fluid/framework/op_registry_test.cc b/paddle/fluid/framework/op_registry_test.cc index 6dc4cf261bad3c004aa53fba5502fe166e3a47f7..18b1649cc71d5edd5b07740bbad1fe8f81128898 100644 --- a/paddle/fluid/framework/op_registry_test.cc +++ b/paddle/fluid/framework/op_registry_test.cc @@ -33,8 +33,7 @@ class CosineOp : public OperatorBase { class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: - CosineOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("input", "input of cosine op"); AddOutput("output", "output of cosine op"); AddAttr("scale", "scale of cosine op") @@ -55,8 +54,7 @@ class MyTestOp : public OperatorBase { class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: - MyTestOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("input", "input of cosine op").AsDuplicable(); AddOutput("output", "output of cosine op").AsIntermediate(); auto my_checker = [](int i) { @@ -212,10 +210,7 @@ namespace framework { class OpKernelTestMaker : public OpProtoAndCheckerMaker { public: - OpKernelTestMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddComment("NoGradOp, same input output. no Grad"); - } + void Make() { AddComment("NoGradOp, same input output. no Grad"); } }; class OpWithKernelTest : public OperatorWithKernel { @@ -275,9 +270,9 @@ TEST(OperatorRegistrar, CUDA) { static int op_test_value = 0; -using paddle::platform::DeviceContext; using paddle::platform::CPUDeviceContext; using paddle::platform::CUDADeviceContext; +using paddle::platform::DeviceContext; namespace paddle { namespace framework { diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index d373c48b1a75c5f75c7520b56f230bc2c146b174..2f480e00c100d579e100de17d3feb957f5ef6167 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.h @@ -33,7 +33,6 @@ limitations under the License. */ #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/variant.h" -#include "paddle/utils/Error.h" namespace paddle { namespace framework { @@ -192,6 +191,10 @@ class ExecutionContext { return op_.Attr(name); } + bool HasInput(const std::string& name) const { return op_.HasInputs(name); } + + bool HasOutput(const std::string& name) const { return op_.HasOutputs(name); } + size_t InputSize(const std::string& name) const { return op_.Inputs(name).size(); } diff --git a/paddle/fluid/framework/operator_test.cc b/paddle/fluid/framework/operator_test.cc index 1bf8c81469bb4afdd00921cfa0acf6089dedbbaa..74043b5d7990178976baf2fad991ae03f9c8dd25 100644 --- a/paddle/fluid/framework/operator_test.cc +++ b/paddle/fluid/framework/operator_test.cc @@ -46,8 +46,7 @@ class OpWithoutKernelTest : public OperatorBase { class OpWithoutKernelCheckerMaker : public OpProtoAndCheckerMaker { public: - OpWithoutKernelCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("input", "input of test op"); AddOutput("output", "output of test op"); AddAttr("scale", "scale of cosine op"); @@ -98,8 +97,7 @@ namespace framework { class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: - OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("x", "input of test op"); AddOutput("y", "output of test op"); AddAttr("scale", "scale of cosine op") @@ -137,9 +135,7 @@ class CPUKernelTest : public OpKernel { class OpKernelTestMultiInputsProtoAndCheckerMaker : public OpProtoAndCheckerMaker { public: - OpKernelTestMultiInputsProtoAndCheckerMaker(OpProto* proto, - OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("xs", "inputs of test op").AsDuplicable(); AddInput("k", "input of test op"); AddOutput("ys", "outputs of test op").AsDuplicable(); diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index 9eea8d1c1861b8a7f6e49621b27c9871b0c1a590..50c3468d556bfe05d6c41906cf35cb671f711b1e 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -52,13 +52,13 @@ std::vector &ParallelExecutor::GetLocalScopes() { } ParallelExecutor::ParallelExecutor( - size_t num_threads, bool use_event, const std::vector &places, const std::unordered_set ¶ms, const std::unordered_set &bcast_vars, const ProgramDesc &main_program, const std::string &loss_var_name, - Scope *scope, const std::vector &local_scopes, bool allow_op_delay, - bool use_default_grad_scale) + Scope *scope, const std::vector &local_scopes, + const ExecutionStrategy &exec_strategy, const BuildStrategy &build_strategy, + size_t num_trainers, size_t trainer_id) : member_(new ParallelExecutorPrivate(places)) { member_->global_scope_ = scope; @@ -80,7 +80,13 @@ ParallelExecutor::ParallelExecutor( // Bcast Parameters to all GPUs #ifdef PADDLE_WITH_CUDA - member_->nccl_ctxs_.reset(new platform::NCCLContextMap(member_->places_)); + auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME); + ncclUniqueId *nccl_id = nullptr; + if (nccl_id_var != nullptr) { + nccl_id = nccl_id_var->GetMutable(); + } + member_->nccl_ctxs_.reset(new platform::NCCLContextMap( + member_->places_, nccl_id, num_trainers, trainer_id)); #endif if (platform::is_gpu_place(places[0]) && member_->local_scopes_.size() != 1 && local_scopes.empty()) { // Is CUDA @@ -93,17 +99,16 @@ ParallelExecutor::ParallelExecutor( #ifdef PADDLE_WITH_CUDA details::MultiDevSSAGraphBuilder builder( member_->places_, loss_var_name, params, member_->local_scopes_, - member_->nccl_ctxs_.get(), use_default_grad_scale); + member_->nccl_ctxs_.get(), build_strategy); #else details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name, params, member_->local_scopes_, - use_default_grad_scale); + build_strategy); #endif auto graph = builder.Build(main_program); member_->executor_.reset(new details::ThreadedSSAGraphExecutor( - num_threads, use_event, member_->local_scopes_, places, std::move(graph), - allow_op_delay)); + exec_strategy, member_->local_scopes_, places, std::move(graph))); // Step 3. Create vars in each scope; for (auto *var : main_program.Block(0).AllVars()) { diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h index ecd107d81f8f5bf5d8b899d0c07797114a7ab767..5247e790649e76567f4527d54499d6e95dac5c27 100644 --- a/paddle/fluid/framework/parallel_executor.h +++ b/paddle/fluid/framework/parallel_executor.h @@ -14,55 +14,60 @@ limitations under the License. */ #pragma once +#include #include #include #include +#include "paddle/fluid/framework/details/execution_strategy.h" #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/framework/op_info.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/platform/device_context.h" - namespace paddle { namespace framework { class ParallelExecutorPrivate; +using details::BuildStrategy; +using details::ExecutionStrategy; + class ParallelExecutor { DISABLE_COPY_AND_ASSIGN(ParallelExecutor); public: - explicit ParallelExecutor(size_t num_threads, bool use_event, - const std::vector& places, - const std::unordered_set& params, - const std::unordered_set& bcast_vars, - const ProgramDesc& main_program, - const std::string& loss_var_name, Scope* scope, - const std::vector& local_scopes, - bool allow_op_delay, bool use_default_grad_scale); + explicit ParallelExecutor(const std::vector &places, + const std::unordered_set ¶ms, + const std::unordered_set &bcast_vars, + const ProgramDesc &main_program, + const std::string &loss_var_name, Scope *scope, + const std::vector &local_scopes, + const ExecutionStrategy &exec_strategy, + const BuildStrategy &build_strategy, + size_t num_trainers = 1, size_t trainer_id = 0); ~ParallelExecutor(); - std::vector& GetLocalScopes(); + std::vector &GetLocalScopes(); /** * Feed tensors to local scopes. The size of tensors should be equal to the * size of local scopes. */ void FeedTensorsIntoLocalScopes( - const std::vector>& tensors); + const std::vector> &tensors); void FeedAndSplitTensorIntoLocalScopes( - const std::unordered_map& tensors); + const std::unordered_map &tensors); - void Run(const std::vector& fetch_tensors, - const std::string& fetched_var_name); + void Run(const std::vector &fetch_tensors, + const std::string &fetched_var_name); - void BCastParamsToGPUs(const std::unordered_set& vars) const; + void BCastParamsToGPUs(const std::unordered_set &vars) const; private: - ParallelExecutorPrivate* member_; + ParallelExecutorPrivate *member_; }; } // namespace framework diff --git a/paddle/fluid/framework/tensor_impl.h b/paddle/fluid/framework/tensor_impl.h index f49d1a47a325b2aac6185073203df124be18b54d..0a1db7758bd9ec0dac133efcbf495de1d690021d 100644 --- a/paddle/fluid/framework/tensor_impl.h +++ b/paddle/fluid/framework/tensor_impl.h @@ -13,54 +13,14 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/memory/memcpy.h" #include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/float16.h" namespace paddle { namespace framework { - -template -struct SizeOfTypeFunctor; - -template -struct SizeOfTypeFunctor { - size_t operator()(std::type_index type) const { - if (typeid(T).hash_code() == type.hash_code()) { - return sizeof(T); - } else { - return 0UL; - } - } -}; - -template <> -struct SizeOfTypeFunctor<> { - size_t operator()(std::type_index type) const { return 0UL; } -}; - -template -struct SizeOfTypeFunctor { - size_t operator()(std::type_index type) const { - SizeOfTypeFunctor head; - size_t head_size = head(type); - if (head_size != 0) { - return head_size; - } - SizeOfTypeFunctor tail; - return tail(type); - } -}; - -static inline size_t SizeOfType(std::type_index type) { - SizeOfTypeFunctor - functor; - size_t size = functor(type); - PADDLE_ENFORCE(size != 0UL, "Cannot get size of type %s", type.name()); - return size; -} - +extern size_t SizeOfType(std::type_index type); inline void Tensor::check_memory_size() const { PADDLE_ENFORCE_NOT_NULL( holder_, "Tensor holds no memory. Call Tensor::mutable_data first."); diff --git a/paddle/fluid/framework/var_type_inference_test.cc b/paddle/fluid/framework/var_type_inference_test.cc index 9e33003b442762210c990b35f30bc3524963b8b4..14b81ddfecb8c996ae8709910c022a074e91eb3c 100644 --- a/paddle/fluid/framework/var_type_inference_test.cc +++ b/paddle/fluid/framework/var_type_inference_test.cc @@ -24,8 +24,7 @@ namespace framework { class SumOpMaker : public OpProtoAndCheckerMaker { public: - SumOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("X", "").AsDuplicable(); AddOutput("Out", ""); AddComment(""); diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt index 50f635a41a99b2ae292d13afde5637a3bf4e6f8c..b98aeed8a0aaabfd39560fad3c074a6668b4f024 100644 --- a/paddle/fluid/inference/CMakeLists.txt +++ b/paddle/fluid/inference/CMakeLists.txt @@ -20,7 +20,9 @@ if(NOT APPLE) endif() if(WITH_TESTING) + # both tests/book and analysis depends the models that generated by python/paddle/fluid/tests/book add_subdirectory(tests/book) + add_subdirectory(analysis) endif() if (TENSORRT_FOUND) diff --git a/paddle/fluid/inference/analysis/CMakeLists.txt b/paddle/fluid/inference/analysis/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..47929ef7490e5edb246625cb0b3ba507039df27a --- /dev/null +++ b/paddle/fluid/inference/analysis/CMakeLists.txt @@ -0,0 +1,2 @@ +cc_library(analysis SRCS dot.cc node.cc node.h) +cc_test(test_node SRCS node_tester.cc DEPS analysis) diff --git a/paddle/fluid/operators/matmul_op.cu.cc b/paddle/fluid/inference/analysis/device.h similarity index 54% rename from paddle/fluid/operators/matmul_op.cu.cc rename to paddle/fluid/inference/analysis/device.h index e021bbe645399e410cde5c3ff7035d4d68c71744..585c9923291e5f9cb6e50dbc4bcd28c374191048 100644 --- a/paddle/fluid/operators/matmul_op.cu.cc +++ b/paddle/fluid/inference/analysis/device.h @@ -1,22 +1,24 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 +http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT 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/fluid/operators/matmul_op.h" +namespace paddle { +namespace inference { +namespace analysis { -namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL( - matmul, ops::MatMulKernel); -REGISTER_OP_CUDA_KERNEL( - matmul_grad, - ops::MatMulGradKernel); +enum class Device { CPU, GPU }; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/dot.cc b/paddle/fluid/inference/analysis/dot.cc new file mode 100644 index 0000000000000000000000000000000000000000..d5471ffcb594a6915e9e65c0fee5adc5f5bdf40c --- /dev/null +++ b/paddle/fluid/inference/analysis/dot.cc @@ -0,0 +1,23 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/dot.h" + +namespace paddle { +namespace inference { +namespace analysis { +size_t Dot::counter = 0; +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/dot.h b/paddle/fluid/inference/analysis/dot.h new file mode 100644 index 0000000000000000000000000000000000000000..4bf1840fdda8508b52d7274a338c5b1c95baf354 --- /dev/null +++ b/paddle/fluid/inference/analysis/dot.h @@ -0,0 +1,155 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +/* + * This file implements some helper classes and methods for DOT programming + * support. It will give a visualization of the graph and that helps to debug + * the logics of each Pass. + */ +#pragma once + +#include +#include +#include +#include +#include + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * A Dot template that helps to build a DOT graph definition. + */ +class Dot { + public: + static size_t counter; + + struct Attr { + std::string key; + std::string value; + + Attr(const std::string& key, const std::string& value) + : key(key), value(value) {} + + std::string repr() const { + std::stringstream ss; + ss << key << "=" << '"' << value << '"'; + return ss.str(); + } + }; + + struct Node { + std::string name; + std::vector attrs; + + Node(const std::string& name, const std::vector& attrs) + : name(name), + attrs(attrs), + id_("node_" + std::to_string(Dot::counter++)) {} + + std::string id() const { return id_; } + + std::string repr() const { + std::stringstream ss; + CHECK(!name.empty()); + ss << id_; + for (size_t i = 0; i < attrs.size(); i++) { + if (i == 0) { + ss << "[label=" << '"' << name << '"' << " "; + } + ss << attrs[i].repr(); + ss << ((i < attrs.size() - 1) ? " " : "]"); + } + return ss.str(); + } + + private: + std::string id_; + }; + + struct Edge { + std::string source; + std::string target; + std::vector attrs; + + Edge(const std::string& source, const std::string& target, + const std::vector& attrs) + : source(source), target(target), attrs(attrs) {} + + std::string repr() const { + std::stringstream ss; + CHECK(!source.empty()); + CHECK(!target.empty()); + ss << source << "->" << target; + for (size_t i = 0; i < attrs.size(); i++) { + if (i == 0) { + ss << "["; + } + ss << attrs[i].repr(); + ss << ((i < attrs.size() - 1) ? " " : "]"); + } + return ss.str(); + } + }; + + Dot() = default; + + explicit Dot(const std::vector& attrs) : attrs_(attrs) {} + + void AddNode(const std::string& name, const std::vector& attrs) { + CHECK(!nodes_.count(name)) << "duplicate Node '" << name << "'"; + nodes_.emplace(name, Node{name, attrs}); + } + + void AddEdge(const std::string& source, const std::string& target, + const std::vector& attrs) { + CHECK(!source.empty()); + CHECK(!target.empty()); + auto sid = nodes_.at(source).id(); + auto tid = nodes_.at(target).id(); + edges_.emplace_back(sid, tid, attrs); + } + + // Compile to DOT language codes. + std::string Build() const { + std::stringstream ss; + const std::string indent = " "; + ss << "digraph G {" << '\n'; + + // Add graph attrs + for (const auto& attr : attrs_) { + ss << indent << attr.repr() << '\n'; + } + // add nodes + for (auto& item : nodes_) { + ss << indent << item.second.repr() << '\n'; + } + // add edges + for (auto& edge : edges_) { + ss << indent << edge.repr() << '\n'; + } + ss << "} // end G"; + return ss.str(); + } + + private: + std::unordered_map nodes_; + std::vector edges_; + std::vector attrs_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/dot_tester.cc b/paddle/fluid/inference/analysis/dot_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..56ceb9bd5d6f41a601d66f6124fb7b4099c9337e --- /dev/null +++ b/paddle/fluid/inference/analysis/dot_tester.cc @@ -0,0 +1,62 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/dot.h" + +#include +#include +#include "paddle/fluid/inference/analysis/data_flow_graph.h" + +namespace paddle { +namespace inference { +namespace analysis { + +class DotTester : public ::testing::Test { + protected: + void SetUp() override { + std::vector attrs({{"title", "hello"}}); + dot.reset(new Dot(attrs)); + dot->AddNode("a", {Dot::Attr{"shape", "box"}, Dot::Attr("color", "blue")}); + dot->AddNode("b", {}); + dot->AddNode("c", {}); + dot->AddEdge("a", "b", {}); + dot->AddEdge("b", "c", {}); + dot->AddEdge("a", "c", {}); + } + + std::unique_ptr dot; +}; + +TEST_F(DotTester, Build) { + auto codes = dot->Build(); + // Output the DOT language code, the generated codes are too long to compare + // the string. + // + // The output is + // + // digraph G { + // title="hello" + // node_1 + // node_2 + // node_0[label="a" shape="box" color="blue"] + // node_0->node_1 + // node_1->node_2 + // node_0->node_2 + // } // end G + LOG(INFO) << '\n' << codes; +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/helper.h b/paddle/fluid/inference/analysis/helper.h new file mode 100644 index 0000000000000000000000000000000000000000..b2d06c5d63ff139186710cd963e07b4ba245f9f3 --- /dev/null +++ b/paddle/fluid/inference/analysis/helper.h @@ -0,0 +1,74 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include +#include +#include + +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace inference { +namespace analysis { + +template +class iterator_range { + IteratorT begin_, end_; + + public: + template + explicit iterator_range(Container &&c) : begin_(c.begin()), end_(c.end()) {} + + iterator_range(const IteratorT &begin, const IteratorT &end) + : begin_(begin), end_(end) {} + + const IteratorT &begin() const { return begin_; } + const IteratorT &end() const { return end_; } +}; + +/* + * An registry helper class, with its records keeps the order they registers. + */ +template +class OrderedRegistry { + public: + T *Register(const std::string &name, T *x) { + PADDLE_ENFORCE(!dic_.count(name)); + dic_[name] = data_.size(); + data_.emplace_back(std::unique_ptr(x)); + return data_.back().get(); + } + + T *Lookup(const std::string &name) { + auto it = dic_.find(name); + if (it == dic_.end()) return nullptr; + return data_[it->second].get(); + } + + protected: + std::unordered_map dic_; + std::vector> data_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle + +#define PADDLE_DISALLOW_COPY_AND_ASSIGN(type__) \ + \ + type__(const type__ &) = delete; \ + \ + void operator=(const type__ &) = delete; diff --git a/paddle/fluid/inference/analysis/node.cc b/paddle/fluid/inference/analysis/node.cc new file mode 100644 index 0000000000000000000000000000000000000000..fe060526080b1ee01aa98f2ff06fb2191eddf9da --- /dev/null +++ b/paddle/fluid/inference/analysis/node.cc @@ -0,0 +1,67 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/node.h" +#include "glog/logging.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace inference { +namespace analysis { + +std::vector Value::dot_attrs() const { + return std::vector({Dot::Attr("style", "filled,rounded"), + Dot::Attr("shape", "box"), + Dot::Attr("fillcolor", "red")}); +} + +std::vector Function::dot_attrs() const { + return std::vector({Dot::Attr("style", "filled,rounded"), + Dot::Attr("shape", "diamond"), + Dot::Attr("fillcolor", "yellow")}); +} + +Node *NodeMap::Create(Node::Type type) { + switch (type) { + case Node::Type::kFunction: + nodes_.emplace_back(new Function); + break; + case Node::Type::kValue: + nodes_.emplace_back(new Value); + break; + default: + PADDLE_THROW("Not supported node type."); + } + nodes_.back()->id_ = size() - 1; + return nodes_.back().get(); +} + +Node *NodeMap::GetMutable(size_t id) { + PADDLE_ENFORCE_GT(size(), id); + return nodes_[id].get(); +} + +const Node &NodeMap::Get(size_t id) const { + PADDLE_ENFORCE_GT(size(), id); + return *nodes_[id].get(); +} + +void NodeMap::Delete(size_t id) { + PADDLE_ENFORCE_LT(id, size()); + nodes_[id]->SetDeleted(); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/node.h b/paddle/fluid/inference/analysis/node.h new file mode 100644 index 0000000000000000000000000000000000000000..07cb7669f98237399c4165947a03c67ce2a86aa8 --- /dev/null +++ b/paddle/fluid/inference/analysis/node.h @@ -0,0 +1,235 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file defines the Node class and its subclasses. A Node is the basis + * analysis element in a computation graph. + * There are basically two kinds of nodes, the function node and value node. + */ +#pragma once + +#include +#include +#include +#include +#include + +#include "paddle/fluid/inference/analysis/device.h" +#include "paddle/fluid/inference/analysis/dot.h" +#include "paddle/fluid/inference/analysis/helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +class NodeMap; + +/* + * Node Representation. + * + * This is a very important class for analysis. It is the base class of all + * nodes computed by a program that may be used as operands to other nodes. + * Node is the super class of other important classes such as Function and + * Value, some nodes can have a name. + */ +class Node { + public: + // Node type. NOTE the new node types should add here. + enum class Type { kNone = -1, kFunction, kValue, kFunctionBlock }; + + Node() = default; + + struct Attr; + + // Cast to a subclass type, Function for example. + template + Subclass &As() { + return *dynamic_cast(this); + } + + // Formatted representation of this Node. + virtual std::string repr() const { + return name() + "(" + std::to_string(id()) + ")"; + } + + // DOT node representation. One Node type can customize its own node + // representation. + virtual std::vector dot_attrs() const { + return std::vector({Dot::Attr("style", "filled")}); + } + + // Get an additional attribute and convert it to T data type. NOTE this will + // silently create a new attribute if not exists. + Attr &attr(const std::string &name) { return attrs_[name]; } + + int id() const { return id_; } + + bool deleted() const { return deleted_; } + void SetDeleted() { deleted_ = true; } + + void SetName(const std::string &name) { name_ = name; } + const std::string &name() const { return name_; } + + void SetType(Type type) { type_ = type; } + Type type() const { return type_; } + + void *extra_info() const { return extra_info_; } + void SetExtraInfo(void *extra_info) { extra_info_ = extra_info; } + + // Input links. + std::vector inlinks; + // Output links. + std::vector outlinks; + + // A helper class to maintain the status from Pass. + // TODO(superjomn) add a checker here to ensure the T is primary. + struct Attr { + // NOTE T should be a primary type or a struct combined by several primary + // types. + // NOTE the STL containers should not use here. + // Some usages + // Attr attr; + // T data; + // attr.data.assign((char*)data, sizeof(data)); + + bool &Bool() { return As(); } + float &Float() { return As(); } + int32_t &Int32() { return As(); } + int64_t &Int64() { return As(); } + + private: + template + T &As() { + // init storage in the first usage. + if (data_.empty()) { + VLOG(4) << "resize data to " << sizeof(T); + type_hash_ = typeid(T).hash_code(); + data_.resize(sizeof(T)); + } + PADDLE_ENFORCE(type_hash_ == typeid(T).hash_code(), "type not matched"); + PADDLE_ENFORCE_EQ(data_.size(), sizeof(T), "Node attr type recast error"); + return *reinterpret_cast(&data_[0]); + } + + private: + std::string data_; + size_t type_hash_{std::numeric_limits::max()}; + }; + + virtual ~Node() {} + + friend class NodeMap; + + PADDLE_DISALLOW_COPY_AND_ASSIGN(Node); + + protected: + // The id number not the name is a node's unique identifier in the computation + // graph. + int id_{-1}; + std::string name_; + Type type_{Type::kNone}; + // Mark this node is deleted by some pass. + bool deleted_{false}; + + void *extra_info_; + + mutable std::unordered_map attrs_; +}; + +class Function; +/* + * Value represents a value node, it has some attributes including dims, data + * type and so on. + */ +class Value : public Node { + public: + enum class DataType { kInt32, kInt64, kFloat32, kFloat64 }; + using Dims = std::vector; + + void SetDataType(DataType data_type) { data_type_ = data_type; } + DataType data_type() const { return data_type_; } + + void SetDims(const Dims &dims) { dims_ = dims; } + const Dims &dims() const { return dims_; } + + Device device() const { return device_; } + void SetDevice(Device device) { device_ = device; } + + std::vector dot_attrs() const override; + + PADDLE_DISALLOW_COPY_AND_ASSIGN(Value); + + protected: + Value() { SetType(Node::Type::kValue); } + friend class NodeMap; + + private: + DataType data_type_; + Dims dims_; + Device device_; +}; + +/* + * Function represents any kind of executable concepts that takes several Values + * as input, and outputs several Values. + */ +class Function : public Node { + public: + std::vector dot_attrs() const override; + + // Get the operator's type from Desc. + const std::string &func_type() const { return func_type_; } + // Set the operator's type. + void SetFuncType(const std::string &func_type) { func_type_ = func_type; } + + PADDLE_DISALLOW_COPY_AND_ASSIGN(Function); + + protected: + std::string func_type_; + Function() { SetType(Node::Type::kFunction); } + friend class NodeMap; +}; + +/* + * FunctionBlock is a Node that contains a sub-graph multiple Node. + */ +struct FunctionBlock : public Node { + std::string repr() const override { return "block-" + std::to_string(id()); } + std::vector subgraph; +}; + +class NodeMap { + public: + // Create a new node with type. + Node *Create(Node::Type type); + + // Get a node by its id. + Node *GetMutable(size_t id); + + const Node &Get(size_t id) const; + + void Delete(size_t id); + + const std::vector> &nodes() { return nodes_; } + + size_t size() const { return nodes_.size(); } + + private: + std::vector> nodes_; + std::unordered_map map_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/node_tester.cc b/paddle/fluid/inference/analysis/node_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..47fea0fdff808c930ca73edb25f5b16fef397e9a --- /dev/null +++ b/paddle/fluid/inference/analysis/node_tester.cc @@ -0,0 +1,34 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/fluid/inference/analysis/node.h" + +#include + +namespace paddle { +namespace inference { +namespace analysis { + +TEST(Node, Attr) { + // Node is an abstract class, use Value instead for they share the same Attr + // logic. + NodeMap nodes; + auto* node = nodes.Create(Node::Type::kValue); + node->attr("v0").Int32() = 2008; + ASSERT_EQ(node->attr("v0").Int32(), 2008); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/engine.h b/paddle/fluid/inference/engine.h index 6b0ac92fa908427a89a6a5fa74dacc3b24abd1c3..ce2b8161715a3fa2278ce950dbac82c6d0042bef 100644 --- a/paddle/fluid/inference/engine.h +++ b/paddle/fluid/inference/engine.h @@ -14,11 +14,15 @@ limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/framework.pb.h" namespace paddle { namespace inference { +struct Buffer; +enum class DeviceType { UNK = -1, CPU, GPU }; + /* * EngineBase is the base class of all inference engines. An inference engine * takes a paddle program as input, and outputs the result in fluid Tensor @@ -45,8 +49,20 @@ class EngineBase { // Execute the engine, that will run the inference network. virtual void Execute(int batch_size) = 0; + // Return the IO buffer that allocated in engine. One can read/write directly + // on the buffer. If the buffer's buffer is nullptr, one can also allocate + // memory and maintain it outside the engine. + virtual Buffer& buffer(const std::string& name) = 0; + virtual ~EngineBase() {} }; // class EngineBase +struct Buffer { + void* buffer{nullptr}; // buffer should be allocated only once. + size_t max_size; // buffer allocated space. + size_t size; // data size. + DeviceType device{DeviceType::UNK}; // tells which device this buffer is on. +}; + } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/CMakeLists.txt b/paddle/fluid/inference/tensorrt/CMakeLists.txt index 288789d6e484100820c937e6081701f1e9245706..b52d083f280e5e7713600a7b748dedd37aca0a1e 100644 --- a/paddle/fluid/inference/tensorrt/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/CMakeLists.txt @@ -1,4 +1,4 @@ +nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto) nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader) -nv_test(test_tensorrt_engine SRCS test_engine.cc engine.cc DEPS dynload_cuda) -set(ENGINE_FILE ${CMAKE_CURRENT_SOURCE_DIR}/engine.cc) +nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine) add_subdirectory(convert) diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt index 3c5909c0be1c690d5148ecfb32b1b6c2dd6f3211..4fb4511d99179e4ea14cde66feb13bc9e114581a 100644 --- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -1,4 +1,4 @@ nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES}) -nv_test(test_trt_activation_op SRCS test_activation_op.cc ${ENGINE_FILE} activation_op.cc - DEPS ${FLUID_CORE_MODULES} activation_op) +nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc io_converter.cc + DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine) nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor) diff --git a/paddle/fluid/inference/tensorrt/convert/activation_op.cc b/paddle/fluid/inference/tensorrt/convert/activation_op.cc index 543784289cfc51048057e467d36fdd1f334eb903..6297051e5a30f1daa512d25d5aa3ab3b2f79f1d1 100644 --- a/paddle/fluid/inference/tensorrt/convert/activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/activation_op.cc @@ -21,15 +21,18 @@ namespace tensorrt { class ReluOpConverter : public OpConverter { public: ReluOpConverter() {} - void operator()(const framework::OpDesc& op) override { + void operator()(const framework::proto::OpDesc& op) override { + // Here the two nullptr looks strange, that's because the + // framework::OpDesc's constructor is strange. + framework::OpDesc op_desc(op, nullptr, nullptr); LOG(INFO) << "convert a fluid relu op to tensorrt activation layer whose " "type is Relu"; const nvinfer1::ITensor* input_tensor = - engine_->GetITensor(op.Input("X")[0]); + engine_->GetITensor(op_desc.Input("X")[0]); nvinfer1::IActivationLayer* layer = TRT_ENGINE_ADD_LAYER( engine_, Activation, *const_cast(input_tensor), nvinfer1::ActivationType::kRELU); - engine_->SetITensor(op.Output("Out")[0], layer->getOutput(0)); + engine_->SetITensor(op_desc.Output("Out")[0], layer->getOutput(0)); } }; diff --git a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc index 431500b90e144e2a30fe705b72e93452f806ca65..209936c3bafb0d31546856dc36c1b48053a0634b 100644 --- a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc @@ -21,7 +21,7 @@ namespace tensorrt { class Conv2dOpConverter : public OpConverter { public: Conv2dOpConverter() {} - void operator()(const framework::OpDesc& op) override { + void operator()(const framework::proto::OpDesc& op) override { LOG(INFO) << "convert a fluid conv2d op to tensorrt conv layer without bias"; } diff --git a/paddle/fluid/inference/tensorrt/convert/io_converter.cc b/paddle/fluid/inference/tensorrt/convert/io_converter.cc index 32e8631fde3f748669d2008b4a060455a37e154e..854f434d93e81237dc85c5df62debcf3b3824b78 100644 --- a/paddle/fluid/inference/tensorrt/convert/io_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/io_converter.cc @@ -23,26 +23,42 @@ namespace tensorrt { using platform::is_gpu_place; using platform::is_cpu_place; -class DefaultInputConverter : public EngineInputConverter { +class DefaultIOConverter : public EngineIOConverter { public: - DefaultInputConverter() {} + DefaultIOConverter() {} // NOTE out is GPU memory. virtual void operator()(const LoDTensor& in, void* out, size_t max_size) override { PADDLE_ENFORCE(out != nullptr); - PADDLE_ENFORCE_LE(in.memory_size(), max_size); + PADDLE_ENFORCE(stream_ != nullptr); const auto& place = in.place(); + size_t size = in.memory_size(); + PADDLE_ENFORCE_LE(size, max_size); if (is_cpu_place(place)) { - PADDLE_ENFORCE(stream_ != nullptr); - PADDLE_ENFORCE_EQ(0, - cudaMemcpyAsync(out, in.data(), in.memory_size(), - cudaMemcpyHostToDevice, *stream_)); - + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out, in.data(), size, + cudaMemcpyHostToDevice, *stream_)); } else if (is_gpu_place(place)) { - PADDLE_ENFORCE_EQ(0, - cudaMemcpyAsync(out, in.data(), in.memory_size(), - cudaMemcpyHostToHost, *stream_)); - + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out, in.data(), size, + cudaMemcpyDeviceToDevice, *stream_)); + } else { + PADDLE_THROW("Unknown device for converter"); + } + cudaStreamSynchronize(*stream_); + } + // NOTE in is GPU memory. + virtual void operator()(const void* in, LoDTensor* out, + size_t max_size) override { + PADDLE_ENFORCE(in != nullptr); + PADDLE_ENFORCE(stream_ != nullptr); + const auto& place = out->place(); + size_t size = out->memory_size(); + PADDLE_ENFORCE_LE(size, max_size); + if (is_cpu_place(place)) { + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data(), in, size, + cudaMemcpyDeviceToHost, *stream_)); + } else if (is_gpu_place(place)) { + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data(), in, size, + cudaMemcpyDeviceToDevice, *stream_)); } else { PADDLE_THROW("Unknown device for converter"); } @@ -50,7 +66,8 @@ class DefaultInputConverter : public EngineInputConverter { } }; -REGISTER_TENSORRT_INPUT_CONVERTER(default, DefaultInputConverter); +// fluid LodTensor <-> tensorrt ITensor +REGISTER_TENSORRT_IO_CONVERTER(default, DefaultIOConverter); } // namespace tensorrt } // namespace inference diff --git a/paddle/fluid/inference/tensorrt/convert/io_converter.h b/paddle/fluid/inference/tensorrt/convert/io_converter.h index 8972dae92be2c2d261a13c48d98e675f64e51d31..71c48e085d25d2bc6720d93735f661f9e3af7b40 100644 --- a/paddle/fluid/inference/tensorrt/convert/io_converter.h +++ b/paddle/fluid/inference/tensorrt/convert/io_converter.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/inference/utils/singleton.h" @@ -25,43 +26,57 @@ namespace tensorrt { using framework::LoDTensor; /* - * Convert Input from Fluid to an Engine. - * TensorRT's ITensor follows row major, NCHW. Fluid is also row major, so in - * most cases just need to copy the data. + * Convert Input from Fluid to TensorRT Engine. + * Convert Output from TensorRT Engine to Fluid. + * + * Note that TensorRT's ITensor follows row major, NCHW. Fluid is also row + * major, + * so in the default case just need to copy the data. */ -class EngineInputConverter { +class EngineIOConverter { public: - EngineInputConverter() {} + EngineIOConverter() {} virtual void operator()(const LoDTensor& in, void* out, size_t max_size) {} + virtual void operator()(const void* in, LoDTensor* out, size_t max_size) {} void SetStream(cudaStream_t* stream) { stream_ = stream; } - static void Run(const std::string& in_op_type, const LoDTensor& in, void* out, - size_t max_size, cudaStream_t* stream) { + static void ConvertInput(const std::string& op_type, const LoDTensor& in, + void* out, size_t max_size, cudaStream_t* stream) { PADDLE_ENFORCE(stream != nullptr); - auto* converter = Registry::Lookup( - in_op_type, "default" /* default_type */); + auto* converter = Registry::Lookup( + op_type, "default" /* default_type */); PADDLE_ENFORCE_NOT_NULL(converter); converter->SetStream(stream); (*converter)(in, out, max_size); } - virtual ~EngineInputConverter() {} + static void ConvertOutput(const std::string& op_type, const void* in, + LoDTensor* out, size_t max_size, + cudaStream_t* stream) { + PADDLE_ENFORCE(stream != nullptr); + auto* converter = Registry::Lookup( + op_type, "default" /* default_type */); + PADDLE_ENFORCE_NOT_NULL(converter); + converter->SetStream(stream); + (*converter)(in, out, max_size); + } + + virtual ~EngineIOConverter() {} protected: cudaStream_t* stream_{nullptr}; }; +#define REGISTER_TENSORRT_IO_CONVERTER(op_type__, Converter__) \ + struct trt_io_##op_type__##_converter { \ + trt_io_##op_type__##_converter() { \ + Registry::Register(#op_type__); \ + } \ + }; \ + trt_io_##op_type__##_converter trt_io_##op_type__##_converter__; + } // namespace tensorrt } // namespace inference } // namespace paddle - -#define REGISTER_TENSORRT_INPUT_CONVERTER(in_op_type__, Converter__) \ - struct trt_input_##in_op_type__##_converter { \ - trt_input_##in_op_type__##_converter() { \ - ::paddle::inference::Registry::Register< \ - Converter__>(#in_op_type__); \ - } \ - }; \ - trt_input_##in_op_type__##_converter trt_input_##in_op_type__##_converter__; diff --git a/paddle/fluid/inference/tensorrt/convert/mul_op.cc b/paddle/fluid/inference/tensorrt/convert/mul_op.cc index f9834ab156c9dcc11f4e89075b7bf5457cf00268..3ca58b139bd3af1947ae7f063060e11d2ea7d577 100644 --- a/paddle/fluid/inference/tensorrt/convert/mul_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/mul_op.cc @@ -21,7 +21,7 @@ namespace tensorrt { class MulOpConverter : public OpConverter { public: MulOpConverter() {} - void operator()(const framework::OpDesc& op) override { + void operator()(const framework::proto::OpDesc& op) override { LOG(INFO) << "convert a fluid mul op to tensorrt fc layer without bias"; } }; diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h index 77c788550b2c7df1f483b926661789b2a54d8fff..1cd3ed9a00acead2599420f88499bd0d74c2974b 100644 --- a/paddle/fluid/inference/tensorrt/convert/op_converter.h +++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h @@ -31,10 +31,10 @@ namespace tensorrt { class OpConverter { public: OpConverter() {} - virtual void operator()(const framework::OpDesc& op) {} + virtual void operator()(const framework::proto::OpDesc& op) {} - void Run(const framework::OpDesc& op, TensorRTEngine* engine) { - std::string type = op.Type(); + void Run(const framework::proto::OpDesc& op, TensorRTEngine* engine) { + std::string type = op.type(); auto* it = Registry::Lookup(type); PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", type); it->SetEngine(engine); @@ -42,14 +42,16 @@ class OpConverter { } // convert fluid op to tensorrt layer - void ConvertOp(const framework::OpDesc& op, TensorRTEngine* engine) { + void ConvertOp(const framework::proto::OpDesc& op, TensorRTEngine* engine) { OpConverter::Run(op, engine); } // convert fluid block to tensorrt network - void ConvertBlock(const framework::BlockDesc& block, TensorRTEngine* engine) { - for (auto op : block.AllOps()) { - OpConverter::Run(*op, engine); + void ConvertBlock(const framework::proto::BlockDesc& block, + TensorRTEngine* engine) { + for (int i = 0; i < block.ops_size(); i++) { + const auto& op = block.ops(i); + OpConverter::Run(op, engine); } } diff --git a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc index 23e3435c21725328d3765fae0d158a83ac21478b..ec33f97c8240dfc09a203d68599bffe78a4abb12 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc @@ -16,6 +16,7 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/inference/tensorrt/convert/io_converter.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/place.h" @@ -26,7 +27,7 @@ namespace paddle { namespace inference { namespace tensorrt { -void Compare(float input, float expect) { +void Compare(const std::string op_type, float input, float expect) { framework::Scope scope; platform::CUDAPlace place; platform::CUDADeviceContext ctx(place); @@ -35,6 +36,7 @@ void Compare(float input, float expect) { auto x_var = scope.Var("X"); auto x_tensor = x_var->GetMutable(); x_tensor->Resize({1, 1}); + x_tensor->mutable_data(place); std::vector init; init.push_back(input); framework::TensorFromVector(init, ctx, x_tensor); @@ -45,14 +47,15 @@ void Compare(float input, float expect) { out_tensor->mutable_data(place); framework::OpDesc op_desc; - op_desc.SetType("relu"); + op_desc.SetType(op_type); op_desc.SetInput("X", {"X"}); op_desc.SetOutput("Out", {"Out"}); - auto relu_op = framework::OpRegistry::CreateOp(op_desc); + auto op = framework::OpRegistry::CreateOp(*op_desc.Proto()); // run fluid op - relu_op->Run(scope, place); + op->Run(scope, place); + // get fluid output std::vector out1; framework::TensorToVector(*out_tensor, ctx, &out1); @@ -63,21 +66,28 @@ void Compare(float input, float expect) { engine->InitNetwork(); engine->DeclareInput("X", nvinfer1::DataType::kFLOAT, nvinfer1::DimsCHW{1, 1, 1}); - + // convert op OpConverter op_converter; - op_converter.ConvertOp(op_desc, engine); + op_converter.ConvertOp(*op_desc.Proto(), engine); engine->DeclareOutput("Out"); engine->FreezeNetwork(); - engine->SetInputFromCPU("X", &input, 1 * sizeof(float)); - // run tensorrt op + // convert LoDTensor to ITensor + size_t size = x_tensor->memory_size(); + EngineIOConverter::ConvertInput(op_type, *x_tensor, + engine->buffer("X").buffer, size, &stream); + // run tensorrt Outp engine->Execute(1); - - float out2; - engine->GetOutputInCPU("Out", &out2, 1 * sizeof(float)); - - ASSERT_EQ(out1[0], out2); + // convert ITensor to LoDTensor + EngineIOConverter::ConvertOutput(op_type, engine->buffer("Out").buffer, + out_tensor, size, &stream); + // get tensorrt output + std::vector out2; + framework::TensorToVector(*out_tensor, ctx, &out2); + + // compare + ASSERT_EQ(out1[0], out2[0]); ASSERT_EQ(out1[0], expect); delete engine; @@ -85,8 +95,8 @@ void Compare(float input, float expect) { } TEST(OpConverter, ConvertRelu) { - Compare(1, 1); // relu(1) = 1 - Compare(-5, 0); // relu(-5) = 0 + Compare("relu", 1, 1); // relu(1) = 1 + Compare("relu", -5, 0); // relu(-5) = 0 } } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc index afcc516e6b76d58e37ce0e60746704cf3933fac7..8f91309a0a00d5131268f026c319e25ba3cb964a 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_io_converter.cc @@ -12,40 +12,63 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/inference/tensorrt/convert/io_converter.h" -#include - namespace paddle { namespace inference { namespace tensorrt { -class EngineInputConverterTester : public ::testing::Test { - public: - void SetUp() override { tensor.Resize({10, 10}); } +void IOConverterTester(const platform::DeviceContext& ctx) { + cudaStream_t stream; + ASSERT_EQ(0, cudaStreamCreate(&stream)); - framework::LoDTensor tensor; -}; + // init fluid in_tensor + framework::LoDTensor in_tensor; + in_tensor.Resize({10, 10}); + auto place = ctx.GetPlace(); + in_tensor.mutable_data(place); + std::vector init; + for (int64_t i = 0; i < 10 * 10; ++i) { + init.push_back(i); + } + framework::TensorFromVector(init, ctx, &in_tensor); -TEST_F(EngineInputConverterTester, DefaultCPU) { + // init tensorrt buffer void* buffer; - tensor.mutable_data(platform::CPUPlace()); - ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0); + size_t size = in_tensor.memory_size(); + ASSERT_EQ(cudaMalloc(&buffer, size), 0); - cudaStream_t stream; - EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(), - &stream); + // convert fluid in_tensor to tensorrt buffer + EngineIOConverter::ConvertInput("test", in_tensor, buffer, size, &stream); + + // convert tensorrt buffer to fluid out_tensor + framework::LoDTensor out_tensor; + out_tensor.Resize({10, 10}); + out_tensor.mutable_data(place); + EngineIOConverter::ConvertOutput("test", buffer, &out_tensor, size, &stream); + + // compare in_tensor and out_tensor + std::vector result; + framework::TensorToVector(out_tensor, ctx, &result); + EXPECT_EQ(init.size(), result.size()); + for (size_t i = 0; i < init.size(); i++) { + EXPECT_EQ(init[i], result[i]); + } + cudaStreamDestroy(stream); } -TEST_F(EngineInputConverterTester, DefaultGPU) { - void* buffer; - tensor.mutable_data(platform::CUDAPlace()); - ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0); +TEST(EngineIOConverterTester, DefaultCPU) { + platform::CPUPlace place; + platform::CPUDeviceContext ctx(place); + IOConverterTester(ctx); +} - cudaStream_t stream; - EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(), - &stream); +TEST(EngineIOConverterTester, DefaultGPU) { + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); + IOConverterTester(ctx); } } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc index aa5fb726f1129eda65a6f39791330b795aad660d..8d66543eb7637c5a8ae670b89ef5996954ba2e7b 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc @@ -29,7 +29,7 @@ TEST(OpConverter, ConvertBlock) { conv2d_op->SetType("conv2d"); OpConverter converter; - converter.ConvertBlock(*block, nullptr /*TensorRTEngine*/); + converter.ConvertBlock(*block->Proto(), nullptr /*TensorRTEngine*/); } } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/engine.cc b/paddle/fluid/inference/tensorrt/engine.cc index df123a59079acc5f549e733b412ab302aa397a92..1c296e33a610493b889359c43629003fd76b893c 100644 --- a/paddle/fluid/inference/tensorrt/engine.cc +++ b/paddle/fluid/inference/tensorrt/engine.cc @@ -30,16 +30,24 @@ void TensorRTEngine::Build(const DescType& paddle_model) { } void TensorRTEngine::Execute(int batch_size) { - infer_context_->enqueue(batch_size, buffers_.data(), *stream_, nullptr); + std::vector buffers; + for (auto& buf : buffers_) { + PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated"); + PADDLE_ENFORCE_GT(buf.max_size, 0); + PADDLE_ENFORCE(buf.device == DeviceType::GPU); + buffers.push_back(buf.buffer); + } + infer_context_->enqueue(batch_size, buffers.data(), *stream_, nullptr); cudaStreamSynchronize(*stream_); } TensorRTEngine::~TensorRTEngine() { // clean buffer - for (auto& buffer : buffers_) { - if (buffer != nullptr) { - PADDLE_ENFORCE_EQ(0, cudaFree(buffer)); - buffer = nullptr; + for (auto& buf : buffers_) { + if (buf.buffer != nullptr) { + PADDLE_ENFORCE_EQ(0, cudaFree(buf.buffer)); + buf.buffer = nullptr; + buf.max_size = 0; } } } @@ -59,7 +67,7 @@ void TensorRTEngine::FreezeNetwork() { infer_context_.reset(infer_engine_->createExecutionContext()); // allocate GPU buffers. - buffers_.resize(buffer_sizes_.size(), nullptr); + buffers_.resize(buffer_sizes_.size()); for (auto& item : buffer_sizes_) { if (item.second == 0) { auto slot_offset = infer_engine_->getBindingIndex(item.first.c_str()); @@ -67,7 +75,11 @@ void TensorRTEngine::FreezeNetwork() { infer_engine_->getBindingDataType(slot_offset))] * AccumDims(infer_engine_->getBindingDimensions(slot_offset)); } - PADDLE_ENFORCE_EQ(0, cudaMalloc(&buffer(item.first), item.second)); + auto& buf = buffer(item.first); + CHECK(buf.buffer == nullptr); // buffer should be allocated only once. + PADDLE_ENFORCE_EQ(0, cudaMalloc(&buf.buffer, item.second)); + buf.size = buf.max_size = item.second; + buf.device = DeviceType::GPU; } } @@ -113,7 +125,7 @@ void TensorRTEngine::DeclareOutput(const std::string& name) { } void* TensorRTEngine::GetOutputInGPU(const std::string& name) { - return buffer(name); + return buffer(name).buffer; } void TensorRTEngine::GetOutputInCPU(const std::string& name, void* dst, @@ -123,11 +135,13 @@ void TensorRTEngine::GetOutputInCPU(const std::string& name, void* dst, PADDLE_ENFORCE(it != buffer_sizes_.end()); PADDLE_ENFORCE_GT(it->second, 0); PADDLE_ENFORCE_GE(max_size, it->second); - PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buffer(name), it->second, + auto& buf = buffer(name); + PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated before"); + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buf.buffer, it->second, cudaMemcpyDeviceToHost, *stream_)); } -void*& TensorRTEngine::buffer(const std::string& name) { +Buffer& TensorRTEngine::buffer(const std::string& name) { PADDLE_ENFORCE(infer_engine_ != nullptr, "call FreezeNetwork first."); auto it = buffer_sizes_.find(name); PADDLE_ENFORCE(it != buffer_sizes_.end()); @@ -137,10 +151,12 @@ void*& TensorRTEngine::buffer(const std::string& name) { void TensorRTEngine::SetInputFromCPU(const std::string& name, void* data, size_t size) { - void* buf = buffer(name); - cudaMemcpyAsync(buf, data, size, cudaMemcpyHostToDevice, *stream_); - PADDLE_ENFORCE_EQ( - 0, cudaMemcpyAsync(buf, data, size, cudaMemcpyHostToDevice, *stream_)); + auto& buf = buffer(name); + PADDLE_ENFORCE_NOT_NULL(buf.buffer); + PADDLE_ENFORCE_LE(size, buf.max_size, "buffer is too small"); + PADDLE_ENFORCE(buf.device == DeviceType::GPU); + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(buf.buffer, data, size, + cudaMemcpyHostToDevice, *stream_)); } void TensorRTEngine::SetITensor(const std::string& name, diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h index ec919b943d3281dd675b15e2f14adb7b3487f46f..b8298c6059e8644327194a1fcf7a7438cc9a7286 100644 --- a/paddle/fluid/inference/tensorrt/engine.h +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -87,7 +87,9 @@ class TensorRTEngine : public EngineBase { // these memory directly for acceleration, for example, output the converted // data directly to the buffer to save data copy overhead. // NOTE this should be used after calling `FreezeNetwork`. - void*& buffer(const std::string& name); + Buffer& buffer(const std::string& name) override; + + cudaStream_t* stream() { return stream_; } // Fill an input from CPU memory with name and size. void SetInputFromCPU(const std::string& name, void* data, size_t size); @@ -116,7 +118,7 @@ class TensorRTEngine : public EngineBase { cudaStream_t* stream_; nvinfer1::ILogger& logger_; - std::vector buffers_; + std::vector buffers_; // max data size for the buffers. std::unordered_map buffer_sizes_; std::unordered_map diff --git a/paddle/fluid/inference/tensorrt/test_engine.cc b/paddle/fluid/inference/tensorrt/test_engine.cc index a08b78f930d30d674247a713fadd3e42e5ada350..e635f0f87d577a1f1ac74687ee60f762be525418 100644 --- a/paddle/fluid/inference/tensorrt/test_engine.cc +++ b/paddle/fluid/inference/tensorrt/test_engine.cc @@ -77,6 +77,37 @@ TEST_F(TensorRTEngineTest, add_layer) { ASSERT_EQ(y_cpu, x_v * 2 + 3); } +TEST_F(TensorRTEngineTest, add_layer_multi_dim) { + // Weight in CPU memory. + // It seems tensorrt FC use col-major: [[1.0, 3.3], [1.1, 4.4]] + // instead of row-major, which is [[1.0, 1.1], [3.3, 4.4]] + float raw_weight[4] = {1.0, 1.1, 3.3, 4.4}; + float raw_bias[2] = {1.3, 2.4}; + + TensorRTEngine::Weight weight(nvinfer1::DataType::kFLOAT, raw_weight, 4); + TensorRTEngine::Weight bias(nvinfer1::DataType::kFLOAT, raw_bias, 2); + auto* x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT, + nvinfer1::DimsCHW{1, 2, 1}); + auto* fc_layer = TRT_ENGINE_ADD_LAYER(engine_, FullyConnected, *x, 2, + weight.get(), bias.get()); + PADDLE_ENFORCE(fc_layer != nullptr); + + engine_->DeclareOutput(fc_layer, 0, "y"); + engine_->FreezeNetwork(); + ASSERT_EQ(engine_->engine()->getNbBindings(), 2); + + float x_v[2] = {1.0, 2.0}; + engine_->SetInputFromCPU("x", reinterpret_cast(&x_v), + 2 * sizeof(float)); + engine_->Execute(1); + + LOG(INFO) << "to get output"; + float y_cpu[2] = {-1., -1.}; + engine_->GetOutputInCPU("y", &y_cpu[0], sizeof(float) * 2); + ASSERT_EQ(y_cpu[0], 4.5); + ASSERT_EQ(y_cpu[1], 14.5); +} + } // namespace tensorrt } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc index c4fd1e298b0daea85db2a407d04ad2d7bcdee0f0..60c761c5281e2f535aab0200c93fb738addcdb87 100644 --- a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc +++ b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc @@ -16,7 +16,6 @@ limitations under the License. */ #include "gtest/gtest.h" #include "paddle/fluid/inference/tests/test_helper.h" -DEFINE_string(data_set, "cifar10", "Data set to test"); DEFINE_string(dirname, "", "Directory of the inference model."); DEFINE_string(fp16_dirname, "", "Directory of the float16 inference model."); DEFINE_int32(batch_size, 1, "Batch size of input data"); @@ -35,19 +34,19 @@ TEST(inference, image_classification) { // 0. Call `paddle::framework::InitDevices()` initialize all the devices // In unittests, this is done in paddle/testing/paddle_gtest_main.cc + const bool is_combined = false; + std::vector> feed_target_shapes = + GetFeedTargetShapes(dirname, is_combined); + paddle::framework::LoDTensor input; // Use normilized image pixels as input data, // which should be in the range [0.0, 1.0]. - if (FLAGS_data_set == "cifar10") { - SetupTensor(&input, {FLAGS_batch_size, 3, 32, 32}, - static_cast(0), static_cast(1)); - } else if (FLAGS_data_set == "imagenet") { - SetupTensor(&input, {FLAGS_batch_size, 3, 224, 224}, - static_cast(0), static_cast(1)); - } else { - LOG(FATAL) << "Only cifar10 or imagenet is supported."; - } - + feed_target_shapes[0][0] = FLAGS_batch_size; + paddle::framework::DDim input_dims = + paddle::framework::make_ddim(feed_target_shapes[0]); + LOG(INFO) << input_dims; + SetupTensor(&input, input_dims, static_cast(0), + static_cast(1)); std::vector cpu_feeds; cpu_feeds.push_back(&input); @@ -60,7 +59,7 @@ TEST(inference, image_classification) { LOG(INFO) << "--- CPU Runs: ---"; LOG(INFO) << "Batch size is " << FLAGS_batch_size; TestInference( - dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat); + dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined); LOG(INFO) << output1.dims(); } @@ -73,7 +72,7 @@ TEST(inference, image_classification) { LOG(INFO) << "--- GPU Runs: ---"; LOG(INFO) << "Batch size is " << FLAGS_batch_size; TestInference( - dirname, cpu_feeds, cpu_fetchs2, FLAGS_repeat); + dirname, cpu_feeds, cpu_fetchs2, FLAGS_repeat, is_combined); LOG(INFO) << output2.dims(); if (!FLAGS_skip_cpu) { diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h index af2a7a5620487a10c1df6152fc4e4bf67b150752..cc1589514aab3b973b4909159748bc4223cdce46 100644 --- a/paddle/fluid/inference/tests/test_helper.h +++ b/paddle/fluid/inference/tests/test_helper.h @@ -89,6 +89,50 @@ void CheckError(const paddle::framework::LoDTensor& output1, EXPECT_EQ(count, 0U) << "There are " << count << " different elements."; } +std::unique_ptr InitProgram( + paddle::framework::Executor* executor, paddle::framework::Scope* scope, + const std::string& dirname, const bool is_combined = false) { + std::unique_ptr inference_program; + if (is_combined) { + // All parameters are saved in a single file. + // Hard-coding the file names of program and parameters in unittest. + // The file names should be consistent with that used in Python API + // `fluid.io.save_inference_model`. + std::string prog_filename = "__model_combined__"; + std::string param_filename = "__params_combined__"; + inference_program = + paddle::inference::Load(executor, scope, dirname + "/" + prog_filename, + dirname + "/" + param_filename); + } else { + // Parameters are saved in separate files sited in the specified + // `dirname`. + inference_program = paddle::inference::Load(executor, scope, dirname); + } + return inference_program; +} + +std::vector> GetFeedTargetShapes( + const std::string& dirname, const bool is_combined = false) { + auto place = paddle::platform::CPUPlace(); + auto executor = paddle::framework::Executor(place); + auto* scope = new paddle::framework::Scope(); + + auto inference_program = InitProgram(&executor, scope, dirname, is_combined); + auto& global_block = inference_program->Block(0); + + const std::vector& feed_target_names = + inference_program->GetFeedTargetNames(); + std::vector> feed_target_shapes; + for (size_t i = 0; i < feed_target_names.size(); ++i) { + auto* var = global_block.FindVar(feed_target_names[i]); + std::vector var_shape = var->GetShape(); + feed_target_shapes.push_back(var_shape); + } + + delete scope; + return feed_target_shapes; +} + template void TestInference(const std::string& dirname, const std::vector& cpu_feeds, @@ -105,7 +149,7 @@ void TestInference(const std::string& dirname, state = paddle::platform::ProfilerState::kCPU; } else { #ifdef PADDLE_WITH_CUDA - state = paddle::platform::ProfilerState::kCUDA; + state = paddle::platform::ProfilerState::kAll; // The default device_id of paddle::platform::CUDAPlace is 0. // Users can get the device_id using: // int device_id = place.GetDeviceId(); @@ -124,26 +168,11 @@ void TestInference(const std::string& dirname, paddle::platform::RecordEvent record_event( "init_program", paddle::platform::DeviceContextPool::Instance().Get(place)); - - if (is_combined) { - // All parameters are saved in a single file. - // Hard-coding the file names of program and parameters in unittest. - // The file names should be consistent with that used in Python API - // `fluid.io.save_inference_model`. - std::string prog_filename = "__model_combined__"; - std::string param_filename = "__params_combined__"; - inference_program = paddle::inference::Load( - &executor, scope, dirname + "/" + prog_filename, - dirname + "/" + param_filename); - } else { - // Parameters are saved in separate files sited in the specified - // `dirname`. - inference_program = paddle::inference::Load(&executor, scope, dirname); - } + inference_program = InitProgram(&executor, scope, dirname, is_combined); } // Disable the profiler and print the timing information paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault, - "load_program_profiler.txt"); + "load_program_profiler"); paddle::platform::ResetProfiler(); // 3. Get the feed_target_names and fetch_target_names @@ -179,10 +208,10 @@ void TestInference(const std::string& dirname, if (PrepareContext) { ctx = executor.Prepare(*inference_program, 0); executor.RunPreparedContext(ctx.get(), scope, &feed_targets, - &fetch_targets, CreateVars); + &fetch_targets, true, CreateVars); } else { executor.Run(*inference_program, scope, &feed_targets, &fetch_targets, - CreateVars); + true, CreateVars); } // Enable the profiler @@ -207,8 +236,7 @@ void TestInference(const std::string& dirname, // Disable the profiler and print the timing information paddle::platform::DisableProfiler( - paddle::platform::EventSortingKey::kDefault, - "run_inference_profiler.txt"); + paddle::platform::EventSortingKey::kDefault, "run_inference_profiler"); paddle::platform::ResetProfiler(); } diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index 256aded8ca234a24229e11f27b9e3e25728ad293..7fce138e3f47e0eb485afb4d5a665eb41f68e286 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -166,6 +166,8 @@ function(op_library TARGET) # NOTE(*): activation use macro to regist the kernels, set use_op manually. if(${TARGET} STREQUAL "activation") file(APPEND ${pybind_file} "USE_OP(relu);\n") + elseif(${TARGET} STREQUAL "reduce") + file(APPEND ${pybind_file} "USE_OP(reduce_sum);\n") else() file(APPEND ${pybind_file} "USE_OP(${TARGET});\n") endif() @@ -184,6 +186,7 @@ endif() add_subdirectory(detail) if(WITH_DISTRIBUTE) + set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf) set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") op_library(send_op DEPS ${DISTRIBUTE_DEPS}) @@ -200,8 +203,16 @@ if(WITH_DISTRIBUTE) set_source_files_properties(send_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op listen_and_serv_op sum_op executor) + if(WITH_GPU) + set_source_files_properties(test_send_nccl_id.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS send_op listen_and_serv_op executor) + op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_grpc) + set_source_files_properties(gen_nccl_id_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + else() + set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op) + endif() else() - set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op) + set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op gen_nccl_id_op) endif() op_library(cross_entropy_op DEPS cross_entropy) @@ -268,6 +279,11 @@ foreach(src ${READER_LIBRARY}) set(OP_LIBRARY ${src} ${OP_LIBRARY}) endforeach() +add_subdirectory(detection) +foreach(src ${DETECTION_LIBRARY}) + set(OP_LIBRARY ${src} ${OP_LIBRARY}) +endforeach() + set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library") cc_test(gather_test SRCS gather_test.cc DEPS tensor) diff --git a/paddle/fluid/operators/accuracy_op.cc b/paddle/fluid/operators/accuracy_op.cc index ac10d759fecb56635d1303fd383a5f9ea18f0a4d..42fcace17926641b5caf677eb3c8ba5222e37190 100644 --- a/paddle/fluid/operators/accuracy_op.cc +++ b/paddle/fluid/operators/accuracy_op.cc @@ -63,8 +63,7 @@ class AccuracyOp : public framework::OperatorWithKernel { class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker { public: - AccuracyOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { // TODO(typhoonzero): support both inference value and indices. AddInput("Out", "The network output of topk (inferences)"); AddInput("Indices", "The the network output of topk (indices)"); diff --git a/paddle/fluid/operators/activation_mkldnn_op.cc b/paddle/fluid/operators/activation_mkldnn_op.cc index ab7c61227114fe7a0ce2ff2515dd560706058b64..b892ac77d9ed60210ddadaecb1a4f214e5a25180 100644 --- a/paddle/fluid/operators/activation_mkldnn_op.cc +++ b/paddle/fluid/operators/activation_mkldnn_op.cc @@ -15,6 +15,7 @@ #include "mkldnn.hpp" #include "paddle/fluid/operators/activation_op.h" #include "paddle/fluid/operators/mkldnn_activation_op.h" +#include "paddle/fluid/platform/mkldnn_helper.h" namespace paddle { namespace operators { @@ -23,6 +24,18 @@ using paddle::framework::Tensor; using paddle::platform::MKLDNNDeviceContext; namespace { +std::string gethash(const mkldnn::memory::dims &operand_dims, + const mkldnn::algorithm algorithm) { + auto dim2str = [](const mkldnn::memory::dims &operand_dims) { + std::string dstr = ""; + for (size_t i = 0; i < operand_dims.size(); ++i) { + dstr += std::to_string(operand_dims[i]) + "-"; + } + return dstr; + }; + return dim2str(operand_dims) + std::to_string(algorithm); +} + template void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm, const T alpha = 0, const T beta = 0) { @@ -37,42 +50,70 @@ void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm, const auto *src_data = src->template data(); auto *dst = ctx.template Output("Out"); - const T *dst_data = dst->template mutable_data(ctx.GetPlace()); + T *dst_data = dst->template mutable_data(ctx.GetPlace()); // get memory dim PADDLE_ENFORCE(src->dims().size() == 2 || src->dims().size() == 4, "Input dim must be with 2 or 4"); std::vector src_tz = framework::vectorize2int(src->dims()); - // create memory description - auto data_md = src_tz.size() == 2 - ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nc) - : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - - // create memory primitives - auto src_memory = - mkldnn::memory({data_md, mkldnn_engine}, - static_cast(const_cast(src_data))); - auto dst_memory = - mkldnn::memory({data_md, mkldnn_engine}, - static_cast(const_cast(dst_data))); - - auto forward_desc = mkldnn::eltwise_forward::desc( - mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta); - - // save prim desc into global device context to be referred in backward path - const std::string key = ctx.op().Output("Out"); - const std::string key_eltwise_pd = key + "@eltwise_pd"; - auto forward_pd = std::make_shared( - forward_desc, mkldnn_engine); - dev_ctx.SetBlob(key_eltwise_pd, forward_pd); - - auto eltwise = mkldnn::eltwise_forward(*forward_pd, src_memory, dst_memory); + const std::string key = gethash(src_tz, algorithm); + const std::string key_src_data = + key + ctx.op().Output("Out") + "@eltwise_fwd_src_data"; + const std::string key_src_mem = key + "@eltwise_fwd_src_mem"; + const std::string key_dst_mem = key + "@eltwise_fwd_dst_mem"; + const std::string key_fwd = key + "@eltwise_fwd"; + + auto p_fwd = std::static_pointer_cast( + dev_ctx.GetBlob(key_fwd)); + + // save input data to be referred in backward path + auto p_src_data = std::make_shared(src_data); + dev_ctx.SetBlob(key_src_data, p_src_data); + + if (p_fwd == nullptr) { + // create memory description + auto data_md = src_tz.size() == 2 + ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, + mkldnn::memory::format::nc) + : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, + mkldnn::memory::format::nchw); + + // create memory primitives + auto p_src_mem = std::make_shared(mkldnn::memory( + {data_md, mkldnn_engine}, platform::to_void_cast(src_data))); + dev_ctx.SetBlob(key_src_mem, p_src_mem); + + auto p_dst_mem = std::make_shared(mkldnn::memory( + {data_md, mkldnn_engine}, platform::to_void_cast(dst_data))); + dev_ctx.SetBlob(key_dst_mem, p_dst_mem); + + auto fwd_desc = mkldnn::eltwise_forward::desc( + mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta); + auto p_fwd_pd = std::make_shared( + fwd_desc, mkldnn_engine); + const std::string key_fwd_pd = key + "eltwise_fwd_pd"; + dev_ctx.SetBlob(key_fwd_pd, p_fwd_pd); + p_fwd = std::make_shared( + *p_fwd_pd, *(p_src_mem.get()), *(p_dst_mem.get())); + dev_ctx.SetBlob(key_fwd, p_fwd); + } else { + // primitives already exist + auto p_src_mem = + std::static_pointer_cast(dev_ctx.GetBlob(key_src_mem)); + PADDLE_ENFORCE(p_src_mem != nullptr, + "Fail to find eltwise p_src_mem in device context."); + auto p_dst_mem = + std::static_pointer_cast(dev_ctx.GetBlob(key_dst_mem)); + PADDLE_ENFORCE(p_dst_mem != nullptr, + "Fail to find eltwise p_src_mem in device context."); + + p_src_mem->set_data_handle(platform::to_void_reinterpret_cast(src_data)); + p_dst_mem->set_data_handle(dst_data); + } // push primitive to stream and wait until it's executed - std::vector pipeline = {eltwise}; + std::vector pipeline = {*(p_fwd.get())}; mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } @@ -83,8 +124,7 @@ void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm, const auto &mkldnn_engine = dev_ctx.GetEngine(); // get buffers - const auto *x = ctx.template Input("X"); - const auto *src = x->template data(); + const auto *out = ctx.template Input("Out"); auto *dout = ctx.template Input(framework::GradVarName("Out")); const auto *diff_dst = dout->template data(); @@ -94,45 +134,73 @@ void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm, const T *diff_src = dx->template mutable_data(ctx.GetPlace()); // get memory dim - std::vector src_tz = framework::vectorize2int(x->dims()); - - // create memory description - auto data_md = src_tz.size() == 2 - ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nc) - : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - - // create memory primitives - auto src_memory = mkldnn::memory( - {data_md, mkldnn_engine}, static_cast(const_cast(src))); - auto diff_src_memory = - mkldnn::memory({data_md, mkldnn_engine}, - static_cast(const_cast(diff_src))); - auto diff_dst_memory = - mkldnn::memory({data_md, mkldnn_engine}, - static_cast(const_cast(diff_dst))); - - auto backward_desc = - mkldnn::eltwise_backward::desc(algorithm, data_md, data_md, alpha, beta); - - // retrieve eltwise primitive desc from device context - const std::string key = ctx.op().Input("Out"); - const std::string key_eltwise_pd = key + "@eltwise_pd"; - const std::shared_ptr forward_pd = dev_ctx.GetBlob(key_eltwise_pd); - PADDLE_ENFORCE(forward_pd != nullptr, - "Fail to find eltwise_pd in device context"); - auto *p_forward_pd = - static_cast(forward_pd.get()); - - auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc( - backward_desc, mkldnn_engine, *p_forward_pd); - - auto eltwise_bwd = mkldnn::eltwise_backward(eltwise_bwd_prim_desc, src_memory, - diff_dst_memory, diff_src_memory); + std::vector src_tz = framework::vectorize2int(out->dims()); + + const std::string key = gethash(src_tz, algorithm); + const std::string key_diff_src_mem = key + "@eltwise_diff_src_mem"; + const std::string key_diff_dst_mem = key + "@eltwise_diff_dst_mem"; + const std::string key_grad = key + "@eltwise_grad"; + + const std::string key_src_data = + key + ctx.op().Input("Out") + "@eltwise_fwd_src_data"; + const auto p_src_data = + std::static_pointer_cast(dev_ctx.GetBlob(key_src_data)); + + const std::string key_src_mem = key + "@eltwise_fwd_src_mem"; + auto p_src_mem = + std::static_pointer_cast(dev_ctx.GetBlob(key_src_mem)); + p_src_mem->set_data_handle(*p_src_data.get()); + + auto p_grad = std::static_pointer_cast( + dev_ctx.GetBlob(key_grad)); + + if (p_grad == nullptr) { + // create memory description + auto data_md = src_tz.size() == 2 + ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, + mkldnn::memory::format::nc) + : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, + mkldnn::memory::format::nchw); + + // create memory primitives + std::shared_ptr p_diff_src_mem = + std::make_shared(mkldnn::memory( + {data_md, mkldnn_engine}, platform::to_void_cast(diff_src))); + dev_ctx.SetBlob(key_diff_src_mem, p_diff_src_mem); + std::shared_ptr p_diff_dst_mem = + std::make_shared(mkldnn::memory( + {data_md, mkldnn_engine}, platform::to_void_cast(diff_dst))); + dev_ctx.SetBlob(key_diff_dst_mem, p_diff_dst_mem); + + auto bwd_desc = mkldnn::eltwise_backward::desc(algorithm, data_md, data_md, + alpha, beta); + + const std::string key_fwd_pd = key + "eltwise_fwd_pd"; + auto *p_fwd_pd = static_cast( + dev_ctx.GetBlob(key_fwd_pd).get()); + + auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc( + bwd_desc, mkldnn_engine, *p_fwd_pd); + + p_grad = std::make_shared( + eltwise_bwd_prim_desc, *static_cast(p_src_mem.get()), + *(static_cast(p_diff_dst_mem.get())), + *(static_cast(p_diff_src_mem.get()))); + } else { + // primitives already exist + auto p_diff_src_mem = std::static_pointer_cast( + dev_ctx.GetBlob(key_diff_src_mem)); + auto p_diff_dst_mem = std::static_pointer_cast( + dev_ctx.GetBlob(key_diff_dst_mem)); + + p_diff_src_mem->set_data_handle( + platform::to_void_reinterpret_cast(diff_src)); + p_diff_dst_mem->set_data_handle( + platform::to_void_reinterpret_cast(diff_dst)); + } // push primitive to stream and wait until it's executed - std::vector pipeline = {eltwise_bwd}; + std::vector pipeline = {*(p_grad.get())}; mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } } // anonymous namespace diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index 87ef55c50b0be46492a695928625d140345d415d..dd71c66a75a039429f6e4b1771bb31508bb6b56d 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -19,19 +19,18 @@ limitations under the License. */ namespace paddle { namespace operators { -#define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT) \ - class OP_NAME##OpMaker \ - : public ::paddle::framework::OpProtoAndCheckerMaker { \ - public: \ - OP_NAME##OpMaker(OpProto *proto, OpAttrChecker *op_checker) \ - : ::paddle::framework::OpProtoAndCheckerMaker(proto, op_checker) { \ - AddInput("X", "Input of " #OP_NAME "operator"); \ - AddOutput("Out", "Output of" #OP_NAME "operator"); \ - AddAttr("use_mkldnn", \ - "(bool, default false) Only used in mkldnn kernel") \ - .SetDefault(false); \ - AddComment(#OP_COMMENT); \ - } \ +#define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT) \ + class OP_NAME##OpMaker \ + : public ::paddle::framework::OpProtoAndCheckerMaker { \ + public: \ + void Make() override { \ + AddInput("X", "Input of " #OP_NAME "operator"); \ + AddOutput("Out", "Output of" #OP_NAME "operator"); \ + AddAttr("use_mkldnn", \ + "(bool, default false) Only used in mkldnn kernel") \ + .SetDefault(false); \ + AddComment(#OP_COMMENT); \ + } \ } #define REGISTER_ACTIVATION_OP_GRAD_MAKER(OP_NAME, KERNEL_TYPE) \ @@ -42,7 +41,7 @@ namespace operators { \ protected: \ std::unique_ptr<::paddle::framework::OpDesc> Apply() const override { \ - auto *op = new ::paddle::framework::OpDesc(); \ + auto* op = new ::paddle::framework::OpDesc(); \ op->SetType(#KERNEL_TYPE "_grad"); \ op->SetInput("Out", Output("Out")); \ op->SetInput(::paddle::framework::GradVarName("Out"), \ @@ -55,23 +54,50 @@ namespace operators { } \ } +framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx, + const framework::OperatorWithKernel& oper, + const std::string& name) { + framework::LibraryType library{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_MKLDNN + auto it = oper.Attrs().find("use_mkldnn"); + if (library == framework::LibraryType::kPlain && it != oper.Attrs().end() && + platform::CanMKLDNNBeUsed(ctx)) { + library = framework::LibraryType::kMKLDNN; + } +#endif + framework::DataLayout layout = framework::DataLayout::kAnyLayout; + return framework::OpKernelType( + framework::ToDataType(ctx.Input(name)->type()), + ctx.GetPlace(), layout, library); +} + class ActivationOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext *ctx) const override { + void InferShape(framework::InferShapeContext* ctx) const override { ctx->SetOutputDim("Out", ctx->GetInputDim("X")); ctx->ShareLoD("X", /*->*/ "Out"); } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return GetKernelType(ctx, *this, "X"); + } }; class ActivationOpGrad : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext *ctx) const override { + void InferShape(framework::InferShapeContext* ctx) const override { ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out")); } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return GetKernelType(ctx, *this, "Out"); + } }; __attribute__((unused)) constexpr char SigmoidDoc[] = R"DOC( @@ -204,8 +230,7 @@ $$out = \frac{x}{1 + |x|}$$ class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - LeakyReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of LeakyRelu operator"); AddOutput("Out", "Output of LeakyRelu operator"); AddAttr("alpha", "The small negative slope").SetDefault(0.02f); @@ -220,8 +245,7 @@ $out = \max(x, \alpha * x)$ class SoftShrinkOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftShrinkOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of Softshrink operator"); AddOutput("Out", "Output of Softshrink operator"); AddAttr("lambda", "non-negative offset").SetDefault(0.5f); @@ -242,8 +266,7 @@ $$ class HardShrinkOpMaker : public framework::OpProtoAndCheckerMaker { public: - HardShrinkOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of HardShrink operator"); AddOutput("Out", "Output of HardShrink operator"); AddAttr("threshold", "The value of threshold for HardShrink") @@ -265,8 +288,7 @@ $$ class BReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - BReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of BRelu operator"); AddOutput("Out", "Output of BRelu operator"); AddAttr("t_min", "The min marginal value of BRelu") @@ -284,8 +306,7 @@ $out = \max(\min(x, t_{min}), t_{max})$ class SoftReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of SoftRelu operator"); AddOutput("Out", "Output of SoftRelu operator"); AddAttr("threshold", "The threshold value of SoftRelu") @@ -301,8 +322,7 @@ $out = \ln(1 + \exp(\max(\min(x, threshold), threshold))$ class ELUOpMaker : public framework::OpProtoAndCheckerMaker { public: - ELUOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of ELU operator"); AddOutput("Out", "Output of ELU operator"); AddAttr("alpha", "The alpha value of ELU").SetDefault(1.0f); @@ -320,8 +340,7 @@ $out = \max(0, x) + \min(0, \alpha * (e^x - 1))$ class Relu6OpMaker : public framework::OpProtoAndCheckerMaker { public: - Relu6OpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of Relu6 operator"); AddOutput("Out", "Output of Relu6 operator"); AddAttr("threshold", "The threshold value of Relu6") @@ -337,8 +356,7 @@ $out = \min(\max(0, x), 6)$ class PowOpMaker : public framework::OpProtoAndCheckerMaker { public: - PowOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of Pow operator"); AddOutput("Out", "Output of Pow operator"); AddAttr("factor", "The exponential factor of Pow").SetDefault(1.0f); @@ -353,8 +371,7 @@ $out = x^{factor}$ class STanhOpMaker : public framework::OpProtoAndCheckerMaker { public: - STanhOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of STanh operator"); AddOutput("Out", "Output of STanh operator"); AddAttr("scale_a", "The scale parameter of a for the input") @@ -372,8 +389,7 @@ $$out = b * \frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$ class ThresholdedReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - ThresholdedReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of ThresholdedRelu operator"); AddOutput("Out", "Output of ThresholdedRelu operator"); AddAttr("threshold", "The threshold location of activation") @@ -394,8 +410,7 @@ $$ class HardSigmoidOpMaker : public framework::OpProtoAndCheckerMaker { public: - HardSigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of HardSigmoid operator"); AddOutput("Out", "Output of HardSigmoid operator"); AddAttr("slope", "Slope for linear approximation of sigmoid") @@ -420,8 +435,7 @@ It is recommended to use the defaults for this activation. class SwishOpMaker : public framework::OpProtoAndCheckerMaker { public: - SwishOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of Swish operator"); AddOutput("Out", "Output of Swish operator"); AddAttr("beta", "Constant beta of swish operator").SetDefault(1.0f); diff --git a/paddle/fluid/operators/adadelta_op.cc b/paddle/fluid/operators/adadelta_op.cc index 7bdb3f274aa9bacb6b261e0d0cd00b72f1d409ae..d1970515f58969948b1d2db5847e4344112f77f9 100644 --- a/paddle/fluid/operators/adadelta_op.cc +++ b/paddle/fluid/operators/adadelta_op.cc @@ -66,8 +66,7 @@ class AdadeltaOp : public framework::OperatorWithKernel { class AdadeltaOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdadeltaOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); AddInput("AvgSquaredGrad", "(Tensor) Input average of squared gradient"); diff --git a/paddle/fluid/operators/adagrad_op.cc b/paddle/fluid/operators/adagrad_op.cc index 1227129429addb0ed412c7f1755fd39c9ca77157..a3ef9ad9f91f1f626bd33876693ecc17ad76b96b 100644 --- a/paddle/fluid/operators/adagrad_op.cc +++ b/paddle/fluid/operators/adagrad_op.cc @@ -67,8 +67,7 @@ class AdagradOp : public framework::OperatorWithKernel { class AdagradOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdagradOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); AddInput("Moment", "(Tensor) Second moment"); diff --git a/paddle/fluid/operators/adam_op.cc b/paddle/fluid/operators/adam_op.cc index f12f0c6663d1785b8af852244ffe32358fb1b693..99b0239855d6241b064a5883c2be3d58078b3b61 100644 --- a/paddle/fluid/operators/adam_op.cc +++ b/paddle/fluid/operators/adam_op.cc @@ -80,8 +80,7 @@ class AdamOp : public framework::OperatorWithKernel { class AdamOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdamOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); AddInput("LearningRate", "(Tensor) Learning rate"); diff --git a/paddle/fluid/operators/adamax_op.cc b/paddle/fluid/operators/adamax_op.cc index 608b855d58a2f701fbb8631cb5f24768a61f3deb..32062574bcf71ff96e451eaa6865b6bbfc3b1c80 100644 --- a/paddle/fluid/operators/adamax_op.cc +++ b/paddle/fluid/operators/adamax_op.cc @@ -74,8 +74,7 @@ class AdamaxOp : public framework::OperatorWithKernel { class AdamaxOpMaker : public framework::OpProtoAndCheckerMaker { public: - AdamaxOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); AddInput("LearningRate", "(Tensor) Learning rate"); diff --git a/paddle/fluid/operators/array_to_lod_tensor_op.cc b/paddle/fluid/operators/array_to_lod_tensor_op.cc index 5db2e4540ef170079328f24ac8d30f7b1901fa1e..149226e92d4d08a25c211bce686ff03c5d7ddf40 100644 --- a/paddle/fluid/operators/array_to_lod_tensor_op.cc +++ b/paddle/fluid/operators/array_to_lod_tensor_op.cc @@ -123,8 +123,7 @@ class ArrayToLoDTensorOp : public framework::OperatorBase { class ArrayToLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ArrayToLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(std::vector) A vector of tensors that is going to " "be casted to a big LoDTensor."); diff --git a/paddle/fluid/operators/assign_op.cc b/paddle/fluid/operators/assign_op.cc index d372213e1b6008b0c4227103dd40730f86a84301..d9294048a9e89662958fd5c6af4fcbe5da3814c2 100644 --- a/paddle/fluid/operators/assign_op.cc +++ b/paddle/fluid/operators/assign_op.cc @@ -94,8 +94,7 @@ class AssignOp : public framework::OperatorBase { class AssignOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - AssignOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor, SelectedRows or LoDTensorArray) The input variable " "could be LoDTensor, SelectedRows or LoDTensorArray.") diff --git a/paddle/fluid/operators/assign_value_op.cc b/paddle/fluid/operators/assign_value_op.cc index 993610fdedde4bafd99f59a0adeeeef4526eb089..4ad6f3443db33fd14b67091d14fd877b951730ff 100644 --- a/paddle/fluid/operators/assign_value_op.cc +++ b/paddle/fluid/operators/assign_value_op.cc @@ -45,8 +45,7 @@ class AssignValueOp : public framework::OperatorWithKernel { class AssignValueOpMaker : public framework::OpProtoAndCheckerMaker { public: - AssignValueOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput("Out", "(Tensor) Output tensor of assign_value operator."); AddAttr>("shape", "(vector) " diff --git a/paddle/fluid/operators/auc_op.cc b/paddle/fluid/operators/auc_op.cc index a168eaeab56128b75bbe97d7ccf843a081b5dced..c9871a9fe6b3b0d0cf671c2d155715f92c94fd8f 100644 --- a/paddle/fluid/operators/auc_op.cc +++ b/paddle/fluid/operators/auc_op.cc @@ -50,8 +50,7 @@ class AucOp : public framework::OperatorWithKernel { class AucOpMaker : public framework::OpProtoAndCheckerMaker { public: - AucOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Out", "A floating point 2D tensor, values are in the range [0, 1]." "Each row is sorted in descending order. This input should be the" diff --git a/paddle/fluid/operators/average_accumulates_op.cc b/paddle/fluid/operators/average_accumulates_op.cc index b21deaf9258567c05a8816b14ac7d6462964e8ba..25864e95d7e290c7f684501893e99c828c511979 100644 --- a/paddle/fluid/operators/average_accumulates_op.cc +++ b/paddle/fluid/operators/average_accumulates_op.cc @@ -111,8 +111,7 @@ class AverageAccumulatesOp : public framework::OperatorWithKernel { class AverageAccumulatesOpMaker : public framework::OpProtoAndCheckerMaker { public: - AverageAccumulatesOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("param", "(Tensor), The parameter to be accumulated."); AddInput("in_sum_1", "(Tensor), A tensor used to store the parameter " diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc index b4bd40d0311bf10ec1fddabab2ee131fe02baf52..6ec8c9d18b466142acdb46b0f46826a2aca7a47e 100644 --- a/paddle/fluid/operators/batch_norm_op.cc +++ b/paddle/fluid/operators/batch_norm_op.cc @@ -126,8 +126,7 @@ class BatchNormOp : public framework::OperatorWithKernel { class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker { public: - BatchNormOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddAttr("is_test", "").SetDefault(false); AddAttr("momentum", "").SetDefault(0.9); AddAttr("epsilon", "") diff --git a/paddle/fluid/operators/batch_size_like.h b/paddle/fluid/operators/batch_size_like.h index dd51a11fbe6ad5e528197b67536518c4b31fa355..483c9f8c2191fa4eb98b91112f9d6753e2fbddc3 100644 --- a/paddle/fluid/operators/batch_size_like.h +++ b/paddle/fluid/operators/batch_size_like.h @@ -53,8 +53,7 @@ class BatchSizeLikeOp : public framework::OperatorWithKernel { class BatchSizeLikeOpMaker : public framework::OpProtoAndCheckerMaker { public: - BatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() final { AddInput("Input", "(Tensor) Tensor " "whose input_dim_idx'th dimension specifies the batch_size"); @@ -68,7 +67,11 @@ class BatchSizeLikeOpMaker : public framework::OpProtoAndCheckerMaker { AddAttr("output_dim_idx", "(int, default 0) The index of output's batch size dimension") .SetDefault(0); + Apply(); } + + protected: + virtual void Apply() = 0; }; } // namespace operators diff --git a/paddle/fluid/operators/beam_search_decode_op.cc b/paddle/fluid/operators/beam_search_decode_op.cc index 68fb988afd8af4e9ac3acb4506c1c31fcf85e5a3..c3dd22119ddab8ecf9213ee274e4cbd4f05e78fd 100644 --- a/paddle/fluid/operators/beam_search_decode_op.cc +++ b/paddle/fluid/operators/beam_search_decode_op.cc @@ -134,8 +134,7 @@ class BeamSearchDecodeOp : public framework::OperatorBase { class BeamSearchDecodeOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - BeamSearchDecodeOpProtoMaker(OpProto* proto, OpAttrChecker* op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Ids", "(LodTensorArray)" "score of the candidate words in each step"); diff --git a/paddle/fluid/operators/beam_search_op.cc b/paddle/fluid/operators/beam_search_op.cc index cff097cca13f3b92c7efe4b69259fdf7c75b3760..df0b50881f4e3ec6f57bdb2b63033931059c486e 100644 --- a/paddle/fluid/operators/beam_search_op.cc +++ b/paddle/fluid/operators/beam_search_op.cc @@ -197,8 +197,7 @@ std::string ItemToString(const BeamSearch::Item &item) { class BeamSearchOpMaker : public framework::OpProtoAndCheckerMaker { public: - BeamSearchOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { // inputs and outputs stored in proto AddInput("pre_ids", "ids in previous step"); AddInput("ids", "a LoDTensor of shape of [None,k]"); diff --git a/paddle/fluid/operators/beam_search_op.h b/paddle/fluid/operators/beam_search_op.h index 9b51db8a45186c2a90cf8b2eb7966d0aaea04028..46bc4f6f936929050276e8b3b93f1eddd62ac638 100644 --- a/paddle/fluid/operators/beam_search_op.h +++ b/paddle/fluid/operators/beam_search_op.h @@ -14,10 +14,6 @@ limitations under the License. */ #pragma once -#ifdef PADDLE_WITH_TESTING -#include "gtest/gtest.h" -#endif - #include #include #include "paddle/fluid/framework/lod_tensor.h" diff --git a/paddle/fluid/operators/bilinear_interp_op.cc b/paddle/fluid/operators/bilinear_interp_op.cc index 69f79bf93be8ac7df9cab43b84cf755f2f3dfeaa..d46fda54e7a9d5bc737a7ec2116daca33ffa015f 100644 --- a/paddle/fluid/operators/bilinear_interp_op.cc +++ b/paddle/fluid/operators/bilinear_interp_op.cc @@ -41,8 +41,7 @@ class BilinearInterpOp : public framework::OperatorWithKernel { class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker { public: - BilinearInterpOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input tensor of bilinear interpolation, " "This is a 4-D tensor with shape of (N x C x h x w)"); diff --git a/paddle/fluid/operators/bilinear_tensor_product_op.cc b/paddle/fluid/operators/bilinear_tensor_product_op.cc index e910ad92d1051aa89fdb3290a977ff376378a227..8d261a118a75ee16027faf60341cefd30c3cdbba 100644 --- a/paddle/fluid/operators/bilinear_tensor_product_op.cc +++ b/paddle/fluid/operators/bilinear_tensor_product_op.cc @@ -65,8 +65,7 @@ class BilinearTensorProductOp : public framework::OperatorWithKernel { class BilinearTensorProductOpMaker : public framework::OpProtoAndCheckerMaker { public: - BilinearTensorProductOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The first input of bilinear_tensor_product operator."); AddInput("Y", "The second input of bilinear_tensor_product operator."); AddInput("Weight", diff --git a/paddle/fluid/operators/cast_op.cc b/paddle/fluid/operators/cast_op.cc index dd0068d571f72c9c22334e523cd091fe4c8da5a6..84660d042c7b12283fabc316d29609f5eddb825d 100644 --- a/paddle/fluid/operators/cast_op.cc +++ b/paddle/fluid/operators/cast_op.cc @@ -21,8 +21,7 @@ namespace operators { class CastOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - CastOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input tensor of cast op"); AddOutput("Out", "The output tensor of cast op"); AddAttr("out_dtype", "output data type"); diff --git a/paddle/fluid/operators/channel_close_op.cc b/paddle/fluid/operators/channel_close_op.cc index 5892650c49e2e9d7345fb94465d124cff57f0a6f..8e2db250a069c488ee98f618bc03df6485022456 100644 --- a/paddle/fluid/operators/channel_close_op.cc +++ b/paddle/fluid/operators/channel_close_op.cc @@ -50,8 +50,7 @@ class ChannelCloseOpOpInferShape : public framework::InferShapeBase { class ChannelCloseOpMaker : public framework::OpProtoAndCheckerMaker { public: - ChannelCloseOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kChannel, "The Channel Variable that should be closed by" " the ChannelClose Op."); diff --git a/paddle/fluid/operators/channel_create_op.cc b/paddle/fluid/operators/channel_create_op.cc index b2fdfd0e1f24ed071bb57b7de8f99b2d5e1d3196..a7f59e4088e3fb328e5b5a83eed65f0f90edb9f0 100644 --- a/paddle/fluid/operators/channel_create_op.cc +++ b/paddle/fluid/operators/channel_create_op.cc @@ -91,8 +91,7 @@ class ChannelCreateOpOpInferShape : public framework::InferShapeBase { class ChannelCreateOpMaker : public framework::OpProtoAndCheckerMaker { public: - ChannelCreateOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput(kOutput, "The object of a Channel type created by ChannelCreate Op."); AddAttr("capacity", "The size of the buffer of Channel.") diff --git a/paddle/fluid/operators/channel_recv_op.cc b/paddle/fluid/operators/channel_recv_op.cc index 25c5c3c95ef6899589c98570df6ecbf9b3241d89..101015e837e28b504b71d919abd5f908a102c812 100644 --- a/paddle/fluid/operators/channel_recv_op.cc +++ b/paddle/fluid/operators/channel_recv_op.cc @@ -72,8 +72,7 @@ class ChannelRecvOp : public framework::OperatorBase { class ChannelRecvOpMaker : public framework::OpProtoAndCheckerMaker { public: - ChannelRecvOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(Channel, "(Channel) A variable which \"receives\" the a value sent" "to it by a channel_send op.") diff --git a/paddle/fluid/operators/channel_send_op.cc b/paddle/fluid/operators/channel_send_op.cc index 66d33617ede5bef8a95de14f5b447c0910fe3eb4..67d6deb511d883ac69426ddd34be2199367cd4c7 100644 --- a/paddle/fluid/operators/channel_send_op.cc +++ b/paddle/fluid/operators/channel_send_op.cc @@ -57,8 +57,7 @@ class ChannelSendOp : public framework::OperatorBase { class ChannelSendOpMaker : public framework::OpProtoAndCheckerMaker { public: - ChannelSendOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(Channel, "(Channel) A variable which \"sends\" the passed in value to " "a listening receiver.") diff --git a/paddle/fluid/operators/chunk_eval_op.cc b/paddle/fluid/operators/chunk_eval_op.cc index 95440ff89e883e754795c67cd58a08f1131df368..62636bb2f9078768180ab1e0016e3565617d24d2 100644 --- a/paddle/fluid/operators/chunk_eval_op.cc +++ b/paddle/fluid/operators/chunk_eval_op.cc @@ -66,8 +66,7 @@ class ChunkEvalOp : public framework::OperatorWithKernel { class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker { public: - ChunkEvalOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Inference", "(Tensor, default: Tensor). " "Predictions from the network."); diff --git a/paddle/fluid/operators/clip_by_norm_op.cc b/paddle/fluid/operators/clip_by_norm_op.cc index f43726b4793f284f14226f90c94ac6eebf632bd5..c87bded034e382c981d119e8499d6780e288031f 100644 --- a/paddle/fluid/operators/clip_by_norm_op.cc +++ b/paddle/fluid/operators/clip_by_norm_op.cc @@ -37,8 +37,7 @@ class ClipByNormOp : public framework::OperatorWithKernel { class ClipByNormOpMaker : public framework::OpProtoAndCheckerMaker { public: - ClipByNormOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input of clip_by_norm op." "The number of dimensions must be between [1, 9]."); diff --git a/paddle/fluid/operators/clip_op.cc b/paddle/fluid/operators/clip_op.cc index c71139fc7c01a696299296e43d06cf195fb3d03f..a679f7e2536a0a44148193f423f5ffe11b5e35fc 100644 --- a/paddle/fluid/operators/clip_op.cc +++ b/paddle/fluid/operators/clip_op.cc @@ -38,8 +38,7 @@ class ClipOp : public framework::OperatorWithKernel { template class ClipOpMaker : public framework::OpProtoAndCheckerMaker { public: - ClipOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor)The input of clip op." "The number of dimensions must be between [1, 9]."); diff --git a/paddle/fluid/operators/compare_op.cc b/paddle/fluid/operators/compare_op.cc index 3a6a357e81949014a70e5bae1ee0e1c8b9d0c2ce..3a4819f3dec9704a4a7c8910dd22e80fda082335 100644 --- a/paddle/fluid/operators/compare_op.cc +++ b/paddle/fluid/operators/compare_op.cc @@ -21,8 +21,7 @@ namespace operators { template class CompareOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - CompareOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { OpComment comment; AddInput("X", string::Sprintf("(LoDTensor) the left hand operand of %s operator", diff --git a/paddle/fluid/operators/concat_op.cc b/paddle/fluid/operators/concat_op.cc index 3bb3bd4eb15881afb5ae42beb944b76b5e8207cb..38337f9aa52435c445420047957500d21069506a 100644 --- a/paddle/fluid/operators/concat_op.cc +++ b/paddle/fluid/operators/concat_op.cc @@ -63,8 +63,7 @@ class ConcatOp : public framework::OperatorWithKernel { class ConcatOpMaker : public framework::OpProtoAndCheckerMaker { public: - ConcatOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input tensors of concat operator.").AsDuplicable(); AddOutput("Out", "Output tensor of concat operator."); AddAttr("axis", diff --git a/paddle/fluid/operators/conditional_block_op.cc b/paddle/fluid/operators/conditional_block_op.cc index 27f74a789beef02d31ebceb9b909e97ebd68232a..5984f80d04bdeb232f8e24264ae979725af24ef4 100644 --- a/paddle/fluid/operators/conditional_block_op.cc +++ b/paddle/fluid/operators/conditional_block_op.cc @@ -108,8 +108,7 @@ class ConditionalBlockOp : public ConditionalOp { class ConditionalBlockOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ConditionalBlockOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The conditional variable of this operator. If X is empty, the " "whole sub-block will not be executed.") diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc index 92748993c32ffb93ae25db8d9916798e657cc804..697d91484257984b104a13b0572cf19b16f8d37e 100644 --- a/paddle/fluid/operators/conv_op.cc +++ b/paddle/fluid/operators/conv_op.cc @@ -106,8 +106,7 @@ framework::OpKernelType ConvOp::GetExpectedKernelType( library); } -Conv2DOpMaker::Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void Conv2DOpMaker::Make() { AddInput( "Input", "(Tensor) The input tensor of convolution operator. " @@ -200,8 +199,7 @@ $$ )DOC"); } -Conv3DOpMaker::Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void Conv3DOpMaker::Make() { AddInput( "Input", "(Tensor) The input tensor of convolution operator. " diff --git a/paddle/fluid/operators/conv_op.h b/paddle/fluid/operators/conv_op.h index f462f00c0803c12ee2f2b0f94dc90afdca500da3..b3140116dfe6a17a400bb88219ff43b249ecb32a 100644 --- a/paddle/fluid/operators/conv_op.h +++ b/paddle/fluid/operators/conv_op.h @@ -60,12 +60,12 @@ inline bool IsExpand(const std::vector& filter_dim, // operator implementations can reuse the code. class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; class Conv3DOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; class ConvOp : public framework::OperatorWithKernel { diff --git a/paddle/fluid/operators/conv_shift_op.cc b/paddle/fluid/operators/conv_shift_op.cc index 82fdd308207adb159632dbb9decd67fd2d1c4646..f2549e814d6f3b5674fe2eec1139f1c3dc6fa0b4 100644 --- a/paddle/fluid/operators/conv_shift_op.cc +++ b/paddle/fluid/operators/conv_shift_op.cc @@ -75,8 +75,7 @@ class ConvShiftGradOp : public framework::OperatorWithKernel { class ConvShiftOpMaker : public framework::OpProtoAndCheckerMaker { public: - ConvShiftOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor, default Tensor), a 2-D tensor with shape B x M, " "where B is the batch size and M is the data dimension."); diff --git a/paddle/fluid/operators/conv_transpose_op.cc b/paddle/fluid/operators/conv_transpose_op.cc index d699dcafa4e2c7e0a3ffb62ec3985e4961fa2133..c27c8e273168407d3aacb05cd6628887cc5760ad 100644 --- a/paddle/fluid/operators/conv_transpose_op.cc +++ b/paddle/fluid/operators/conv_transpose_op.cc @@ -84,9 +84,7 @@ framework::OpKernelType ConvTransposeOp::GetExpectedKernelType( layout_, library_); } -Conv2DTransposeOpMaker::Conv2DTransposeOpMaker(OpProto* proto, - OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void Conv2DTransposeOpMaker::Make() { AddInput( "Input", "(Tensor) The input tensor of convolution transpose operator. " @@ -168,9 +166,7 @@ Example: )DOC"); } -Conv3DTransposeOpMaker::Conv3DTransposeOpMaker(OpProto* proto, - OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void Conv3DTransposeOpMaker::Make() { AddInput("Input", "(Tensor) The input tensor of convolution transpose operator." "The format of input tensor is NCDHW. Where N is batch size, C is " diff --git a/paddle/fluid/operators/conv_transpose_op.h b/paddle/fluid/operators/conv_transpose_op.h index 898121412b17cd6fbbbeb57e9d63842e592703ac..f9d205a5b5c4cff74d02a6c89b83f7584e4a6824 100644 --- a/paddle/fluid/operators/conv_transpose_op.h +++ b/paddle/fluid/operators/conv_transpose_op.h @@ -30,12 +30,12 @@ using DDim = framework::DDim; // operator implementations can reuse the code. class Conv2DTransposeOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv2DTransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; class Conv3DTransposeOpMaker : public framework::OpProtoAndCheckerMaker { public: - Conv3DTransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; class ConvTransposeOp : public framework::OperatorWithKernel { diff --git a/paddle/fluid/operators/cos_sim_op.cc b/paddle/fluid/operators/cos_sim_op.cc index 04ca878e687f9b8e5239d8c4aad7e5f262fda0fa..046dd11910bb0ff46b567c3b89883582782205d3 100644 --- a/paddle/fluid/operators/cos_sim_op.cc +++ b/paddle/fluid/operators/cos_sim_op.cc @@ -62,8 +62,7 @@ class CosSimOp : public framework::OperatorWithKernel { class CosSimOpMaker : public framework::OpProtoAndCheckerMaker { public: - CosSimOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The 1st input of cos_sim op."); AddInput("Y", "The 2nd input of cos_sim op."); AddOutput("Out", "The output of cos_sim op."); diff --git a/paddle/fluid/operators/crf_decoding_op.cc b/paddle/fluid/operators/crf_decoding_op.cc index a83013c428a77a0ead545d87852e1017bc927edf..40f43936db662f2b18ffa540da4794755b5d6fc7 100644 --- a/paddle/fluid/operators/crf_decoding_op.cc +++ b/paddle/fluid/operators/crf_decoding_op.cc @@ -18,8 +18,7 @@ namespace paddle { namespace operators { class CRFDecodingOpMaker : public framework::OpProtoAndCheckerMaker { public: - CRFDecodingOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Emission", "(LoDTensor, default: LoDTensor). A LoDTensor with shape " "[N x D] where N is the size of the mini-batch and D is the total " diff --git a/paddle/fluid/operators/crop_op.cc b/paddle/fluid/operators/crop_op.cc index a8f1fbd529c71d1915c75fa90b7e4e8239d2fa3f..669b3bbe9df4cae1aa381184092dfa51157ab6a3 100644 --- a/paddle/fluid/operators/crop_op.cc +++ b/paddle/fluid/operators/crop_op.cc @@ -52,8 +52,7 @@ class CropOp : public framework::OperatorWithKernel { class CropOpMaker : public framework::OpProtoAndCheckerMaker { public: - CropOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of pad op. " "The input should be a k-D tensor(k > 0 and k < 7)."); diff --git a/paddle/fluid/operators/cross_entropy_op.cc b/paddle/fluid/operators/cross_entropy_op.cc index 2b2a9dc8319f964875371214168ce04cb67fc818..a3bec3da45136bca5cb2763e7ffd6b67703a1813 100644 --- a/paddle/fluid/operators/cross_entropy_op.cc +++ b/paddle/fluid/operators/cross_entropy_op.cc @@ -111,8 +111,7 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel { class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker { public: - CrossEntropyOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor, default Tensor), a 2-D tensor with shape [N x D]," " where N is the batch size and D is the number of classes. " diff --git a/paddle/fluid/operators/ctc_align_op.cc b/paddle/fluid/operators/ctc_align_op.cc index 19e7649660edd0bc90bc6a9537b1cdbb2e7e8ebc..d2b440d9d2e50340af7a7bb4e76e55beea1bcb46 100644 --- a/paddle/fluid/operators/ctc_align_op.cc +++ b/paddle/fluid/operators/ctc_align_op.cc @@ -44,8 +44,7 @@ class CTCAlignOp : public framework::OperatorWithKernel { class CTCAlignOpMaker : public framework::OpProtoAndCheckerMaker { public: - CTCAlignOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(LodTensor, default: LoDTensor), Its shape is " "[Lp, 1], where Lp is the sum of all input sequences' length."); diff --git a/paddle/fluid/operators/cumsum_op.cc b/paddle/fluid/operators/cumsum_op.cc index f7c516a0ba375a68e3adeb44c99f2808dc0418bb..92bb835e8f18e17ae1355fdec29f43b8ffb70460 100644 --- a/paddle/fluid/operators/cumsum_op.cc +++ b/paddle/fluid/operators/cumsum_op.cc @@ -29,8 +29,7 @@ class CumOp : public framework::OperatorWithKernel { class CumsumOpMaker : public framework::OpProtoAndCheckerMaker { public: - CumsumOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Input of Cumsum operator"); AddOutput("Out", "Output of Cumsum operator"); AddAttr("axis", diff --git a/paddle/fluid/operators/decayed_adagrad_op.cc b/paddle/fluid/operators/decayed_adagrad_op.cc index 5a1315fb2a80bf7f7f57388d0d6832686442c4ff..c0f2b49a04d9e88502c4b63bca493cd2b7ad1c5c 100644 --- a/paddle/fluid/operators/decayed_adagrad_op.cc +++ b/paddle/fluid/operators/decayed_adagrad_op.cc @@ -62,8 +62,7 @@ class DecayedAdagradOp : public framework::OperatorWithKernel { class DecayedAdagradOpMaker : public framework::OpProtoAndCheckerMaker { public: - DecayedAdagradOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor) Input parameter"); AddInput("Grad", "(Tensor) Input gradient"); AddInput("Moment", "(Tensor) Second moment"); diff --git a/paddle/fluid/operators/delete_var_op.cc b/paddle/fluid/operators/delete_var_op.cc index 1fe9404c00335edbe3594486f8c403e69f2ab08f..d7a9bfbc437dbf4c723b9c87ff62ec6b62c38638 100644 --- a/paddle/fluid/operators/delete_var_op.cc +++ b/paddle/fluid/operators/delete_var_op.cc @@ -34,8 +34,7 @@ class DeleteVarOp : public framework::OperatorBase { class DeleteVarOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - DeleteVarOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of delete op").AsDuplicable(); AddComment(R"DOC( Delete Operator. diff --git a/paddle/fluid/operators/detail/grpc_client.cc b/paddle/fluid/operators/detail/grpc_client.cc index 661dfa69fe1580ff3890f12defcd124225be0c06..ae60ab15325ef101feb7270a4f5d840cb2112be0 100644 --- a/paddle/fluid/operators/detail/grpc_client.cc +++ b/paddle/fluid/operators/detail/grpc_client.cc @@ -52,7 +52,7 @@ bool RPCClient::AsyncSendVariable(const std::string& ep, // stub context SendProcessor* s = new SendProcessor(ch); s->Prepare(var_h, time_out); - s->response_call_back_ = NULL; + s->response_call_back_ = nullptr; auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_); diff --git a/paddle/fluid/operators/detail/grpc_client.h b/paddle/fluid/operators/detail/grpc_client.h index f6229b71bc01a6de51f50f5fe880ada6e15e74dd..dabce7414d2f0dca74193f1cd10c341793c10ec9 100644 --- a/paddle/fluid/operators/detail/grpc_client.h +++ b/paddle/fluid/operators/detail/grpc_client.h @@ -57,7 +57,9 @@ void ProcGetResponse(const VarHandle& var_h, const grpc::ByteBuffer& msg); class BaseProcessor { public: - explicit BaseProcessor(std::shared_ptr ch) { context_ = NULL; } + explicit BaseProcessor(std::shared_ptr ch) { + context_ = nullptr; + } virtual ~BaseProcessor() {} @@ -105,7 +107,7 @@ class SendProcessor : public BaseProcessor { ::grpc::GenericStub stub_g_; ::grpc::ByteBuffer reply_; - RequestSendCallBack response_call_back_ = NULL; + RequestSendCallBack response_call_back_ = nullptr; }; typedef std::function diff --git a/paddle/fluid/operators/detail/grpc_server.cc b/paddle/fluid/operators/detail/grpc_server.cc index e6ee28ea8d920ef80fead258a9bd0d5f6762c879..eb114a47d99541402f748bfffcf6b10fde3e78e2 100644 --- a/paddle/fluid/operators/detail/grpc_server.cc +++ b/paddle/fluid/operators/detail/grpc_server.cc @@ -184,7 +184,7 @@ class RequestPrefetch final : public RequestBase { framework::Scope* local_scope = &scope_->NewScope(); auto* var = local_scope->FindVar(var_name); InitializeVariable(var, var_desc->GetType()); - executor_->RunPreparedContext(prefetch_ctx_, scope_, false, false); + executor_->RunPreparedContext(prefetch_ctx_, scope_); SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply); @@ -306,7 +306,7 @@ void AsyncGRPCServer::TryToRegisterNewPrefetchOne() { } RequestPrefetch* prefetch = new RequestPrefetch(&service_, cq_prefetch_.get(), sync_mode_, scope_, - dev_ctx_, executor_, program_, prefetch_ctx_); + dev_ctx_, executor_, program_, prefetch_ctx_.get()); VLOG(4) << "Create RequestPrefetch status:" << prefetch->Status(); } diff --git a/paddle/fluid/operators/detail/grpc_server.h b/paddle/fluid/operators/detail/grpc_server.h index 7f9cae21ccca8dd51f9fbe98148d01a51ac6eb84..238aaa29634a7eff65429c27aa3538a185723eb2 100644 --- a/paddle/fluid/operators/detail/grpc_server.h +++ b/paddle/fluid/operators/detail/grpc_server.h @@ -47,6 +47,7 @@ class AsyncGRPCServer final { explicit AsyncGRPCServer(const std::string &address, bool sync_mode) : address_(address), sync_mode_(sync_mode), ready_(0) {} + ~AsyncGRPCServer() {} void WaitServerReady(); void RunSyncUpdate(); @@ -63,8 +64,9 @@ class AsyncGRPCServer final { void SetExecutor(framework::Executor *executor) { executor_ = executor; } - void SetPrefetchPreparedCtx(framework::ExecutorPrepareContext *prepared) { - prefetch_ctx_ = prepared; + void SetPrefetchPreparedCtx( + std::unique_ptr prepared) { + prefetch_ctx_.reset(prepared.release()); } int GetSelectedPort() const { return selected_port_; } @@ -115,7 +117,7 @@ class AsyncGRPCServer final { std::unique_ptr t_get_; std::unique_ptr t_prefetch_; - framework::ExecutorPrepareContext *prefetch_ctx_; + std::unique_ptr prefetch_ctx_; framework::ProgramDesc *program_; framework::Executor *executor_; int selected_port_; diff --git a/paddle/fluid/operators/detail/grpc_server_test.cc b/paddle/fluid/operators/detail/grpc_server_test.cc index 25b95d608d10d6e456d5f563ce9fbe35d812cb0f..b8db0ad987cdfaec1fc9236c3f26e88891376dce 100644 --- a/paddle/fluid/operators/detail/grpc_server_test.cc +++ b/paddle/fluid/operators/detail/grpc_server_test.cc @@ -100,7 +100,7 @@ void StartServer(const std::string& endpoint) { InitTensorsOnServer(&scope, &place, 10); rpc_service_->SetProgram(&program); - rpc_service_->SetPrefetchPreparedCtx(prepared.get()); + rpc_service_->SetPrefetchPreparedCtx(std::move(prepared)); rpc_service_->SetDevCtx(&ctx); rpc_service_->SetScope(&scope); rpc_service_->SetExecutor(&exe); diff --git a/paddle/fluid/operators/detail/send_recv.proto b/paddle/fluid/operators/detail/send_recv.proto index fffa9ae7a43ea5cd7b2bda6fbbf6ef9f7d23009d..9478c5702bcbf99fc88207b8c4843dbccf8a5925 100644 --- a/paddle/fluid/operators/detail/send_recv.proto +++ b/paddle/fluid/operators/detail/send_recv.proto @@ -32,6 +32,7 @@ service SendRecvService { enum VarType { LOD_TENSOR = 0; SELECTED_ROWS = 1; + NCCL_ID = 2; } // NOTICE(gongwb):don't modify this proto if you are not diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.cc b/paddle/fluid/operators/detail/sendrecvop_utils.cc index d68cf467f7b0c6157dc1f69571e5d0c0b3c70348..07c43554bc6a0d71d688a5a5772d0ab3d2de319a 100644 --- a/paddle/fluid/operators/detail/sendrecvop_utils.cc +++ b/paddle/fluid/operators/detail/sendrecvop_utils.cc @@ -14,6 +14,9 @@ limitations under the License. */ #include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#ifdef PADDLE_WITH_CUDA +#include +#endif #include #include // NOLINT @@ -29,129 +32,149 @@ namespace paddle { namespace operators { namespace detail { +using VarMsg = sendrecv::VariableMessage; + +void GetTensorPayload(framework::Variable* var, + const platform::DeviceContext& ctx, VarMsg* request, + void** payload, size_t* payload_size) { + auto tensor = var->Get(); + // FIXME(wuyi): data types in send_recv.proto is copied from + // framework.proto + request->set_data_type( + static_cast(framework::ToDataType(tensor.type()))); + for (auto& dim : framework::vectorize(tensor.dims())) { + request->add_dims(dim); + } + const framework::LoD lod = tensor.lod(); + if (lod.size() > 0) { + request->set_lod_level(lod.size()); + for (auto& each : lod) { + VarMsg::LodData* lod_inner = request->add_lod(); + for (auto& d : each) { + lod_inner->add_lod_data(d); + } + } + } + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + PADDLE_ENFORCE(platform::is_gpu_place(tensor.place())); + platform::CPUPlace cpu; + auto& gpu_dev_ctx = static_cast(ctx); + auto copy_size = tensor.numel() * framework::SizeOfType(tensor.type()); + *payload = memory::Alloc(cpu, copy_size); + + memory::Copy(cpu, *payload, boost::get(tensor.place()), + reinterpret_cast(tensor.data()), copy_size, + gpu_dev_ctx.stream()); + ctx.Wait(); +#endif + } else { + *payload = tensor.data(); + } + *payload_size = tensor.numel() * framework::SizeOfType(tensor.type()); +} + +void GetSelectedRowsPayload(framework::Variable* var, + const platform::DeviceContext& ctx, VarMsg* request, + void** payload, size_t* payload_size) { + auto* slr = var->GetMutable(); + request->set_data_type( + static_cast(framework::ToDataType(slr->value().type()))); + request->set_lod_level(0); + request->set_slr_height(slr->height()); + + for (auto& dim : framework::vectorize(slr->value().dims())) { + request->add_dims(dim); + } + + auto* tensor = slr->mutable_value(); + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + platform::CPUPlace cpu; + auto& gpu_dev_ctx = static_cast(ctx); + auto copy_size = tensor->numel() * framework::SizeOfType(tensor->type()); + *payload = memory::Alloc(cpu, copy_size); + memory::Copy(cpu, *payload, + boost::get(tensor->place()), + reinterpret_cast(tensor->data()), copy_size, + gpu_dev_ctx.stream()); + ctx.Wait(); +#endif + } else { + *payload = slr->mutable_value()->data(); + } + *payload_size = tensor->numel() * framework::SizeOfType(tensor->type()); +} + void SerializeToByteBuffer(const std::string& name, framework::Variable* var, const platform::DeviceContext& ctx, ::grpc::ByteBuffer* msg, const std::string& out_name) { - using VarMsg = sendrecv::VariableMessage; - // When using GPU, need to free the copied CPU buffer - // when the ByteBuffer destroies - // TODO(typhoonzero): add unref here, if we have dependent - // parallelism execution, need to know when to free the tensor. + // Default DestroyCallback does nothing, When using GPU + // the CPU buffer need to be freed. DestroyCallback destroy_callback = [](void* backing) {}; - - auto buffer = std::unique_ptr(new char[1024]); - void* buf = buffer.get(); - + VarMsg request; void* payload = nullptr; size_t payload_size; - ProtoEncodeHelper e(static_cast(buf), 1024); + + request.set_varname(name); // Note: normally the profiler is enabled in 1 trainer, hence only // 1 trainer returns true for ShouldSendProfileState(). It tells PS // servers the trainer's profiling state so that PS can follow the // trainer. - if (platform::ShouldSendProfileState()) { - e.WriteBool(VarMsg::kProfileFieldNumber, platform::IsProfileEnabled()); + request.set_profile(platform::IsProfileEnabled()); + if (!out_name.empty()) { + request.set_out_varname(out_name); } - e.WriteString(VarMsg::kVarnameFieldNumber, name); if (var->IsType()) { - e.WriteUint64(VarMsg::kTypeFieldNumber, 0); + request.set_type(::sendrecv::LOD_TENSOR); + GetTensorPayload(var, ctx, &request, &payload, &payload_size); } else if (var->IsType()) { - e.WriteUint64(VarMsg::kTypeFieldNumber, 1); + request.set_type(::sendrecv::SELECTED_ROWS); + GetSelectedRowsPayload(var, ctx, &request, &payload, &payload_size); +#ifdef PADDLE_WITH_CUDA + } else if (var->IsType()) { + request.set_type(::sendrecv::NCCL_ID); +#endif + } else { + PADDLE_THROW("Serialize does not support type: %s", + typeid(var->Type()).name()); } - if (!out_name.empty()) { - e.WriteString(VarMsg::kOutVarnameFieldNumber, out_name); + if (platform::is_gpu_place(ctx.GetPlace())) { + // GPU data is copied to CPU buffer when sending, + // free the buffer when possible. + destroy_callback = [](void* backing) { + platform::CPUPlace cpu; + memory::Free(cpu, backing); + }; } - switch (framework::ToVarType(var->Type())) { - case framework::proto::VarType_Type_LOD_TENSOR: { - auto tensor = var->Get(); - e.WriteUint64(VarMsg::kDataTypeFieldNumber, - framework::ToDataType(tensor.type())); - for (auto& dim : framework::vectorize(tensor.dims())) { - e.WriteUint64(VarMsg::kDimsFieldNumber, dim); - } - auto lod = tensor.lod(); // std::vector> - if (lod.size() > 0) { - e.WriteUint64(VarMsg::kLodLevelFieldNumber, lod.size()); - - for (auto& each : lod) { - e.WriteVarlengthBeginning(VarMsg::kLodFieldNumber, - 2 + // tag + varintlength of submessage - 1 + // kLodDataFieldNumber - each.size()); - // auto copied from GPU - for (auto& d : each) { - e.WriteUint64(VarMsg::LodData::kLodDataFieldNumber, d); - } - } - } - if (platform::is_gpu_place(ctx.GetPlace())) { -#ifdef PADDLE_WITH_CUDA - PADDLE_ENFORCE(platform::is_gpu_place(tensor.place())); - platform::CPUPlace cpu; - auto& gpu_dev_ctx = - static_cast(ctx); - auto copy_size = tensor.numel() * framework::SizeOfType(tensor.type()); - payload = memory::Alloc(cpu, copy_size); - - memory::Copy(cpu, payload, - boost::get(tensor.place()), - reinterpret_cast(tensor.data()), - copy_size, gpu_dev_ctx.stream()); - ctx.Wait(); - destroy_callback = [](void* backing) { - platform::CPUPlace cpu; - memory::Free(cpu, backing); - }; -#endif - } else { - payload = tensor.data(); - } - payload_size = tensor.numel() * framework::SizeOfType(tensor.type()); - e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size); - } break; - case framework::proto::VarType_Type_SELECTED_ROWS: { - // TODO(typhoonzero): selectedrows implement should not use unique_ptr - auto* slr = var->GetMutable(); - e.WriteUint64(VarMsg::kDataTypeFieldNumber, - framework::ToDataType(slr->value().type())); - for (auto& dim : framework::vectorize(slr->value().dims())) { - e.WriteUint64(VarMsg::kDimsFieldNumber, dim); - } - e.WriteUint64(VarMsg::kLodLevelFieldNumber, 0); - e.WriteUint64(VarMsg::kSlrHeightFieldNumber, slr->height()); - auto* tensor = slr->mutable_value(); - if (platform::is_gpu_place(ctx.GetPlace())) { + std::string header; + request.AppendToString(&header); + auto buffer = std::unique_ptr(new char[1024]); + void* buf = buffer.get(); + ProtoEncodeHelper e(static_cast(buf), 1024); + e.WriteRawBytes(std::string(header.data(), header.size())); +// NCCLID is copied directly to the message, return bytebuffer +// with only one slice if serializing NCCLID. #ifdef PADDLE_WITH_CUDA - platform::CPUPlace cpu; - auto& gpu_dev_ctx = - static_cast(ctx); - auto copy_size = - tensor->numel() * framework::SizeOfType(tensor->type()); - payload = memory::Alloc(cpu, copy_size); - memory::Copy(cpu, payload, - boost::get(tensor->place()), - reinterpret_cast(tensor->data()), - copy_size, gpu_dev_ctx.stream()); - ctx.Wait(); - destroy_callback = [](void* backing) { - platform::CPUPlace cpu; - memory::Free(cpu, backing); - }; -#endif - } else { - payload = slr->mutable_value()->data(); - } - payload_size = tensor->numel() * framework::SizeOfType(tensor->type()); - e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size); - } break; - default: - PADDLE_THROW("Serialize does not support type: %s", - typeid(var->Type()).name()); - break; + if (var->IsType()) { + e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, + NCCL_UNIQUE_ID_BYTES); + const ncclUniqueId& uid = var->Get(); + e.WriteRawBytes(std::string(uid.internal, NCCL_UNIQUE_ID_BYTES)); + + // for serialize NCCL_ID + ::grpc::Slice slices(e.size()); + memcpy(const_cast(slices.begin()), e.data(), e.size()); + ::grpc::ByteBuffer tmp(&slices, 1); + msg->Swap(&tmp); + return; } +#endif + + e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size); // steal reference of tensor data ::grpc::Slice slices[4]; // metadata, tensor, rows meta, rows int num_slices = 2; // only SelectedRows have rows buffer @@ -162,12 +185,9 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, static_cast(payload)), ::grpc::Slice::STEAL_REF); - if (framework::ToVarType(var->Type()) == - framework::proto::VarType_Type_SELECTED_ROWS) { + if (var->IsType()) { auto* slr = var->GetMutable(); - ProtoEncodeHelper e2(static_cast(buf), 128); - // NOTE: rows is of type int64_t size_t rows_memory_size = slr->rows().size() * framework::SizeOfType(typeid(int64_t)); e2.WriteVarlengthBeginning(VarMsg::kRowsFieldNumber, rows_memory_size); @@ -178,10 +198,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, grpc_slice_new_with_user_data( const_cast( reinterpret_cast(slr->rows().data())), - rows_memory_size, - [](void* backing) { - // TODO(typhoonzero): add unref here, same as above. - }, + rows_memory_size, [](void* backing) {}, const_cast( reinterpret_cast(slr->rows().data()))), ::grpc::Slice::STEAL_REF); diff --git a/paddle/fluid/operators/detail/serde_test.cc b/paddle/fluid/operators/detail/serde_test.cc index e9eaaf1cbcd07ed1c8d6fb0b025bc1f1500718fd..15892295e6901fe649788c9e34604008fc8cbdfa 100644 --- a/paddle/fluid/operators/detail/serde_test.cc +++ b/paddle/fluid/operators/detail/serde_test.cc @@ -117,11 +117,11 @@ void RunTestLodTensor(platform::Place place, int from_type = 0) { // serialize var to ByteBuffer framework::Variable var; auto* tensor = var.GetMutable(); - tensor->Resize(framework::make_ddim({4, 8, 4, 2})); + tensor->Resize(framework::make_ddim({512, 8, 4, 2})); framework::LoD lod; lod.push_back(framework::Vector({1, 3, 8})); tensor->set_lod(lod); - int tensor_numel = 4 * 8 * 4 * 2; + int tensor_numel = 512 * 8 * 4 * 2; platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); tensor->mutable_data(place); @@ -142,7 +142,7 @@ void RunTestLodTensor(platform::Place place, int from_type = 0) { EXPECT_TRUE(varmsg.ParseFromString(tmp)); EXPECT_EQ(varmsg.varname(), "myvar"); EXPECT_EQ(varmsg.type(), 0); - EXPECT_EQ(varmsg.dims()[0], 4); + EXPECT_EQ(varmsg.dims()[0], 512); EXPECT_EQ(varmsg.dims()[1], 8); EXPECT_EQ(varmsg.dims()[2], 4); EXPECT_EQ(varmsg.dims()[3], 2); diff --git a/paddle/fluid/operators/detail/variable_response.cc b/paddle/fluid/operators/detail/variable_response.cc index f4a374d56d28a30201f0d482e97e1a40e7a8bf41..462e303096e609c6797ca8cc16266ec3621623fc 100644 --- a/paddle/fluid/operators/detail/variable_response.cc +++ b/paddle/fluid/operators/detail/variable_response.cc @@ -17,6 +17,9 @@ #include #include #include +#ifdef PADDLE_WITH_CUDA +#include +#endif #include "paddle/fluid/platform/profiler.h" #include "paddle/fluid/operators/detail/send_recv.pb.h" @@ -210,15 +213,15 @@ bool ParseLodData(::google::protobuf::io::CodedInputStream* input, } if (wt == WIRETYPE_LENGTH_DELIMITED) { - int length = 0; - if (!input->ReadVarintSizeAsInt(&length)) { + int num_bytes = 0; + if (!input->ReadVarintSizeAsInt(&num_bytes)) { return tag; } - - for (int i = 0; i < length; i++) { + int start_pos = input->CurrentPosition(); + while (input->CurrentPosition() - start_pos < num_bytes) { uint64_t v; if (!input->ReadVarint64(&v)) { - return false; + return tag; } lod->push_back(v); } @@ -275,8 +278,8 @@ int VariableResponse::Parse(Source* source) { break; } case sendrecv::VariableMessage::kTypeFieldNumber: { - uint64_t v; - if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) { + uint32_t v; + if ((wt != WIRETYPE_VARINT) || !input.ReadVarint32(&v)) { return tag; } @@ -284,8 +287,8 @@ int VariableResponse::Parse(Source* source) { break; } case sendrecv::VariableMessage::kDataTypeFieldNumber: { - uint64_t v = 0; - if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) { + uint32_t v = 0; + if ((wt != WIRETYPE_VARINT) || !input.ReadVarint32(&v)) { return tag; } @@ -305,11 +308,12 @@ int VariableResponse::Parse(Source* source) { // packed if (wt == WIRETYPE_LENGTH_DELIMITED) { - int length = 0; - if (!input.ReadVarintSizeAsInt(&length)) { + int num_bytes = 0; + if (!input.ReadVarintSizeAsInt(&num_bytes)) { return tag; } - for (int i = 0; i < length; i++) { + int start_pos = input.CurrentPosition(); + while (input.CurrentPosition() - start_pos < num_bytes) { uint64_t v; if (!input.ReadVarint64(&v)) { return tag; @@ -318,7 +322,6 @@ int VariableResponse::Parse(Source* source) { } break; } - return tag; } case sendrecv::VariableMessage::kLodLevelFieldNumber: { @@ -368,28 +371,45 @@ int VariableResponse::Parse(Source* source) { } case sendrecv::VariableMessage::kSerializedFieldNumber: { PADDLE_ENFORCE((meta_.type() == sendrecv::SELECTED_ROWS || - meta_.type() == sendrecv::LOD_TENSOR) && + meta_.type() == sendrecv::LOD_TENSOR || + meta_.type() == sendrecv::NCCL_ID) && meta_.varname() != "", "meta info should be got first!"); - int length = 0; + int num_bytes = 0; if (wt != WIRETYPE_LENGTH_DELIMITED || - !ReadVarintSizeAsInt(&input, &length)) { + !ReadVarintSizeAsInt(&input, &num_bytes)) { return tag; } + if (meta_.type() == sendrecv::NCCL_ID) { +#ifdef PADDLE_WITH_CUDA + auto* var = scope_->FindVar(meta_.varname()); + if (var != nullptr) { + ncclUniqueId* id = var->GetMutable(); + if (!ReadRaw(&input, *dev_ctx_, platform::CPUPlace(), id->internal, + num_bytes)) { + return tag; + } + } + break; +#else + PADDLE_THROW("Not compiled with CUDA!"); +#endif + } + framework::DDim dims = GetDims(meta_.dims()); if (meta_.type() == sendrecv::LOD_TENSOR) { PADDLE_ENFORCE(meta_.lod_size() >= 0, "lod info should be got first!"); - if (!CopyLodTensorData(&input, *dev_ctx_, dims, length)) { + if (!CopyLodTensorData(&input, *dev_ctx_, dims, num_bytes)) { return tag; } break; } if (meta_.type() == sendrecv::SELECTED_ROWS) { - if (!CopySelectRowsTensorData(&input, *dev_ctx_, dims, length)) { + if (!CopySelectRowsTensorData(&input, *dev_ctx_, dims, num_bytes)) { return tag; } break; @@ -403,13 +423,13 @@ int VariableResponse::Parse(Source* source) { meta_.varname() != "", "meta info should be got first!"); - int length = 0; + int num_bytes = 0; if (wt != WIRETYPE_LENGTH_DELIMITED || - !ReadVarintSizeAsInt(&input, &length)) { + !ReadVarintSizeAsInt(&input, &num_bytes)) { return tag; } - if (!CopySelectRowsData(&input, *dev_ctx_, length)) { + if (!CopySelectRowsData(&input, *dev_ctx_, num_bytes)) { return tag; } break; diff --git a/paddle/fluid/operators/detection/CMakeLists.txt b/paddle/fluid/operators/detection/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..a5bb58c2f4047a3bf2f8592b605772b4fa166c57 --- /dev/null +++ b/paddle/fluid/operators/detection/CMakeLists.txt @@ -0,0 +1,29 @@ +set(LOCAL_DETECTION_LIBS) + +function(detection_library TARGET_NAME) + set(oneValueArgs "") + set(multiValueArgs SRCS DEPS) + set(options "") + set(common_deps op_registry) + set(pybind_flag 0) + cmake_parse_arguments(detection_library "${options}" "${oneValueArgs}" + "${multiValueArgs}" ${ARGN}) + op_library(${TARGET_NAME} SRCS ${detection_library_SRCS} DEPS ${common_deps} ${detection_library_DEPS}) + set(LOCAL_DETECTION_LIBS + ${TARGET_NAME} + ${LOCAL_DETECTION_LIBS} + PARENT_SCOPE) +endfunction() + +detection_library(bipartite_match_op SRCS bipartite_match_op.cc) +detection_library(box_coder_op SRCS box_coder_op.cc box_coder_op.cu) +detection_library(iou_similarity_op SRCS iou_similarity_op.cc +iou_similarity_op.cu) +detection_library(mine_hard_examples_op SRCS mine_hard_examples_op.cc) +detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc) +detection_library(prior_box_op SRCS prior_box_op.cc prior_box_op.cu) +detection_library(target_assign_op SRCS target_assign_op.cc +target_assign_op.cu) + +# Export local libraries to parent +set(DETECTION_LIBRARY ${LOCAL_DETECTION_LIBS} PARENT_SCOPE) diff --git a/paddle/fluid/operators/bipartite_match_op.cc b/paddle/fluid/operators/detection/bipartite_match_op.cc similarity index 98% rename from paddle/fluid/operators/bipartite_match_op.cc rename to paddle/fluid/operators/detection/bipartite_match_op.cc index 1218d9fdc1e6101d17bc09a4ae769f5fbf8e7b15..d437ad5c19828331c749244404ba80d0f3acda2a 100644 --- a/paddle/fluid/operators/bipartite_match_op.cc +++ b/paddle/fluid/operators/detection/bipartite_match_op.cc @@ -182,8 +182,7 @@ class BipartiteMatchKernel : public framework::OpKernel { class BipartiteMatchOpMaker : public framework::OpProtoAndCheckerMaker { public: - BipartiteMatchOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "DistMat", "(LoDTensor or Tensor) this input is a 2-D LoDTensor with shape " diff --git a/paddle/fluid/operators/box_coder_op.cc b/paddle/fluid/operators/detection/box_coder_op.cc similarity index 97% rename from paddle/fluid/operators/box_coder_op.cc rename to paddle/fluid/operators/detection/box_coder_op.cc index ec416f725e75fae57484751ee8a066c0b9da8a70..74848005d0bea6e5c818ff999727aa2b8ad51d84 100644 --- a/paddle/fluid/operators/box_coder_op.cc +++ b/paddle/fluid/operators/detection/box_coder_op.cc @@ -9,7 +9,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/box_coder_op.h" +#include "paddle/fluid/operators/detection/box_coder_op.h" namespace paddle { namespace operators { @@ -60,8 +60,7 @@ class BoxCoderOp : public framework::OperatorWithKernel { class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker { public: - BoxCoderOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "PriorBox", "(Tensor, default Tensor) " diff --git a/paddle/fluid/operators/box_coder_op.cu b/paddle/fluid/operators/detection/box_coder_op.cu similarity index 99% rename from paddle/fluid/operators/box_coder_op.cu rename to paddle/fluid/operators/detection/box_coder_op.cu index 708c7a5fa96c2f9ce6a2d913ca5f30126bb6192f..8cef8e03439df4ca5b0fa94176a21a36f9eb9f70 100644 --- a/paddle/fluid/operators/box_coder_op.cu +++ b/paddle/fluid/operators/detection/box_coder_op.cu @@ -9,7 +9,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/box_coder_op.h" +#include "paddle/fluid/operators/detection/box_coder_op.h" #include "paddle/fluid/platform/cuda_primitives.h" namespace paddle { diff --git a/paddle/fluid/operators/box_coder_op.h b/paddle/fluid/operators/detection/box_coder_op.h similarity index 100% rename from paddle/fluid/operators/box_coder_op.h rename to paddle/fluid/operators/detection/box_coder_op.h diff --git a/paddle/fluid/operators/iou_similarity_op.cc b/paddle/fluid/operators/detection/iou_similarity_op.cc similarity index 95% rename from paddle/fluid/operators/iou_similarity_op.cc rename to paddle/fluid/operators/detection/iou_similarity_op.cc index 4b78ec510d1fb73592ee8af9a641622f4d713f8d..8e58605fcea04f9ffa97ce8cca53c073e7068aaf 100644 --- a/paddle/fluid/operators/iou_similarity_op.cc +++ b/paddle/fluid/operators/detection/iou_similarity_op.cc @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/iou_similarity_op.h" +#include "paddle/fluid/operators/detection/iou_similarity_op.h" namespace paddle { namespace operators { @@ -42,8 +42,7 @@ class IOUSimilarityOp : public framework::OperatorWithKernel { class IOUSimilarityOpMaker : public framework::OpProtoAndCheckerMaker { public: - IOUSimilarityOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor, default LoDTensor) " "Box list X is a 2-D LoDTensor with shape [N, 4] holds N boxes, " diff --git a/paddle/fluid/operators/iou_similarity_op.cu b/paddle/fluid/operators/detection/iou_similarity_op.cu similarity index 92% rename from paddle/fluid/operators/iou_similarity_op.cu rename to paddle/fluid/operators/detection/iou_similarity_op.cu index f40a388d62e66a110656ebb71094d46b5ac147eb..8342b4138c87e6ea1803146bac6d6954a569ef5f 100644 --- a/paddle/fluid/operators/iou_similarity_op.cu +++ b/paddle/fluid/operators/detection/iou_similarity_op.cu @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/iou_similarity_op.h" +#include "paddle/fluid/operators/detection/iou_similarity_op.h" namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( diff --git a/paddle/fluid/operators/iou_similarity_op.h b/paddle/fluid/operators/detection/iou_similarity_op.h similarity index 100% rename from paddle/fluid/operators/iou_similarity_op.h rename to paddle/fluid/operators/detection/iou_similarity_op.h diff --git a/paddle/fluid/operators/mine_hard_examples_op.cc b/paddle/fluid/operators/detection/mine_hard_examples_op.cc similarity index 99% rename from paddle/fluid/operators/mine_hard_examples_op.cc rename to paddle/fluid/operators/detection/mine_hard_examples_op.cc index 277901cff493445e1e85e92e22ea0ada0e1cba43..d4a09bae3a98e4518f9885c1e9182f7033a0d262 100644 --- a/paddle/fluid/operators/mine_hard_examples_op.cc +++ b/paddle/fluid/operators/detection/mine_hard_examples_op.cc @@ -253,8 +253,7 @@ class MineHardExamplesOp : public framework::OperatorWithKernel { class MineHardExamplesOpMaker : public framework::OpProtoAndCheckerMaker { public: - MineHardExamplesOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "ClsLoss", "(Tensor, default Tensor), The classification loss with shape " diff --git a/paddle/fluid/operators/multiclass_nms_op.cc b/paddle/fluid/operators/detection/multiclass_nms_op.cc similarity index 99% rename from paddle/fluid/operators/multiclass_nms_op.cc rename to paddle/fluid/operators/detection/multiclass_nms_op.cc index a12b975326519c776c9f4a1d9f2894b4028c2440..60b93efdce810f8552374449fe5a6fc79b1a92c1 100644 --- a/paddle/fluid/operators/multiclass_nms_op.cc +++ b/paddle/fluid/operators/detection/multiclass_nms_op.cc @@ -309,8 +309,7 @@ class MultiClassNMSKernel : public framework::OpKernel { class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker { public: - MultiClassNMSOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("BBoxes", "(Tensor) A 3-D Tensor with shape [N, M, 4] represents the " "predicted locations of M bounding bboxes, N is the batch size. " diff --git a/paddle/fluid/operators/prior_box_op.cc b/paddle/fluid/operators/detection/prior_box_op.cc similarity index 97% rename from paddle/fluid/operators/prior_box_op.cc rename to paddle/fluid/operators/detection/prior_box_op.cc index 058b13eeb872aaa77a88da37db64a6d59fbdd1cf..4e35c38e4e03d4d0f00601812fdc4803519b89ae 100644 --- a/paddle/fluid/operators/prior_box_op.cc +++ b/paddle/fluid/operators/detection/prior_box_op.cc @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/prior_box_op.h" +#include "paddle/fluid/operators/detection/prior_box_op.h" namespace paddle { namespace operators { @@ -79,8 +79,7 @@ class PriorBoxOp : public framework::OperatorWithKernel { class PriorBoxOpMaker : public framework::OpProtoAndCheckerMaker { public: - PriorBoxOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(Tensor, default Tensor), " "the input feature data of PriorBoxOp, The layout is NCHW."); diff --git a/paddle/fluid/operators/prior_box_op.cu b/paddle/fluid/operators/detection/prior_box_op.cu similarity index 99% rename from paddle/fluid/operators/prior_box_op.cu rename to paddle/fluid/operators/detection/prior_box_op.cu index 0ea8909296f8f52d252b0ec258666cf32d69a8bb..f67e6ca91c0852b5a3be35d23246884d1157caa4 100644 --- a/paddle/fluid/operators/prior_box_op.cu +++ b/paddle/fluid/operators/detection/prior_box_op.cu @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/prior_box_op.h" +#include "paddle/fluid/operators/detection/prior_box_op.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/prior_box_op.h b/paddle/fluid/operators/detection/prior_box_op.h similarity index 100% rename from paddle/fluid/operators/prior_box_op.h rename to paddle/fluid/operators/detection/prior_box_op.h diff --git a/paddle/fluid/operators/target_assign_op.cc b/paddle/fluid/operators/detection/target_assign_op.cc similarity index 97% rename from paddle/fluid/operators/target_assign_op.cc rename to paddle/fluid/operators/detection/target_assign_op.cc index 33ff967e5e8f5afbaa62ba39ce596687ae0a71cd..367001939251114a9cf442fd85c734958ccb2da8 100644 --- a/paddle/fluid/operators/target_assign_op.cc +++ b/paddle/fluid/operators/detection/target_assign_op.cc @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/target_assign_op.h" +#include "paddle/fluid/operators/detection/target_assign_op.h" namespace paddle { namespace operators { @@ -65,8 +65,7 @@ class TargetAssignOp : public framework::OperatorWithKernel { class TargetAssignOpMaker : public framework::OpProtoAndCheckerMaker { public: - TargetAssignOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor), This input is a 3D LoDTensor with shape [M, P, K]. " "Some elements in X will be assigned to Out based on the " diff --git a/paddle/fluid/operators/target_assign_op.cu b/paddle/fluid/operators/detection/target_assign_op.cu similarity index 97% rename from paddle/fluid/operators/target_assign_op.cu rename to paddle/fluid/operators/detection/target_assign_op.cu index 24664f99b20f92108220d27ec58e8fdf3ba6193c..ddf6889942355457fb281b6c33430ab8337db3ed 100644 --- a/paddle/fluid/operators/target_assign_op.cu +++ b/paddle/fluid/operators/detection/target_assign_op.cu @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/target_assign_op.h" +#include "paddle/fluid/operators/detection/target_assign_op.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/target_assign_op.h b/paddle/fluid/operators/detection/target_assign_op.h similarity index 100% rename from paddle/fluid/operators/target_assign_op.h rename to paddle/fluid/operators/detection/target_assign_op.h diff --git a/paddle/fluid/operators/detection_map_op.cc b/paddle/fluid/operators/detection_map_op.cc index 38f43b6d031372948bd82c686a2d9ce5f8ecd07c..716c8625d35308f98582e6802e90d99d643e188b 100644 --- a/paddle/fluid/operators/detection_map_op.cc +++ b/paddle/fluid/operators/detection_map_op.cc @@ -51,7 +51,8 @@ class DetectionMAPOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_EQ(label_dims.size(), 2, "The rank of Input(Label) must be 2, " "the shape is [N, 6]."); - PADDLE_ENFORCE_EQ(label_dims[1], 6, "The shape is of Input(Label) [N, 6]."); + PADDLE_ENFORCE(label_dims[1] == 6 || label_dims[1] == 5, + "The shape of Input(Label) is [N, 6] or [N, 5]."); if (ctx->HasInput("PosCount")) { PADDLE_ENFORCE(ctx->HasInput("TruePos"), @@ -78,8 +79,7 @@ class DetectionMAPOp : public framework::OperatorWithKernel { class DetectionMAPOpMaker : public framework::OpProtoAndCheckerMaker { public: - DetectionMAPOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("DetectRes", "(LoDTensor) A 2-D LoDTensor with shape [M, 6] represents the " "detections. Each row has 6 values: " @@ -89,9 +89,10 @@ class DetectionMAPOpMaker : public framework::OpProtoAndCheckerMaker { "offset is N + 1, if LoD[i + 1] - LoD[i] == 0, means there is " "no detected data."); AddInput("Label", - "(LoDTensor) A 2-D LoDTensor with shape[N, 6] represents the" + "(LoDTensor) A 2-D LoDTensor represents the" "Labeled ground-truth data. Each row has 6 values: " - "[label, is_difficult, xmin, ymin, xmax, ymax], N is the total " + "[label, xmin, ymin, xmax, ymax, is_difficult] or 5 values: " + "[label, xmin, ymin, xmax, ymax], where N is the total " "number of ground-truth data in this mini-batch. For each " "instance, the offsets in first dimension are called LoD, " "the number of offset is N + 1, if LoD[i + 1] - LoD[i] == 0, " diff --git a/paddle/fluid/operators/detection_map_op.h b/paddle/fluid/operators/detection_map_op.h index 431812e2bfcf926cadf8d7be6a7d1a79e78c7762..dd1ab85fd8d0c8170afcd9dd2a49ee55c41dc8be 100644 --- a/paddle/fluid/operators/detection_map_op.h +++ b/paddle/fluid/operators/detection_map_op.h @@ -72,7 +72,7 @@ class DetectionMAPOpKernel : public framework::OpKernel { auto* out_false_pos = ctx.Output("AccumFalsePos"); float overlap_threshold = ctx.Attr("overlap_threshold"); - float evaluate_difficult = ctx.Attr("evaluate_difficult"); + bool evaluate_difficult = ctx.Attr("evaluate_difficult"); auto ap_type = GetAPType(ctx.Attr("ap_type")); int class_num = ctx.Attr("class_num"); @@ -175,14 +175,20 @@ class DetectionMAPOpKernel : public framework::OpKernel { for (int n = 0; n < batch_size; ++n) { std::map> boxes; for (size_t i = label_index[n]; i < label_index[n + 1]; ++i) { - Box box(labels(i, 2), labels(i, 3), labels(i, 4), labels(i, 5)); int label = labels(i, 0); - auto is_difficult = labels(i, 1); - if (std::abs(is_difficult - 0.0) < 1e-6) - box.is_difficult = false; - else - box.is_difficult = true; - boxes[label].push_back(box); + if (input_label.dims()[1] == 6) { + Box box(labels(i, 2), labels(i, 3), labels(i, 4), labels(i, 5)); + auto is_difficult = labels(i, 1); + if (std::abs(is_difficult - 0.0) < 1e-6) + box.is_difficult = false; + else + box.is_difficult = true; + boxes[label].push_back(box); + } else { + PADDLE_ENFORCE_EQ(input_label.dims()[1], 5); + Box box(labels(i, 1), labels(i, 2), labels(i, 3), labels(i, 4)); + boxes[label].push_back(box); + } } gt_boxes->push_back(boxes); } diff --git a/paddle/fluid/operators/dropout_op.cc b/paddle/fluid/operators/dropout_op.cc index 4ed1b548840fabd2383632beb5f35fa6aa096443..07322e720f26213ea777be3cd22f2fead28507f0 100644 --- a/paddle/fluid/operators/dropout_op.cc +++ b/paddle/fluid/operators/dropout_op.cc @@ -37,8 +37,7 @@ class DropoutOp : public framework::OperatorWithKernel { class DropoutOpMaker : public framework::OpProtoAndCheckerMaker { public: - DropoutOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of dropout op."); AddOutput("Out", "The output of dropout op."); AddOutput("Mask", "The random sampled dropout mask.").AsIntermediate(); diff --git a/paddle/fluid/operators/edit_distance_op.cc b/paddle/fluid/operators/edit_distance_op.cc index c7f037d2df4372d0c4e3a261c0dff1fd6704d182..de25a3dab53492e38a92fbcf07ccbe43f7546950 100644 --- a/paddle/fluid/operators/edit_distance_op.cc +++ b/paddle/fluid/operators/edit_distance_op.cc @@ -49,8 +49,7 @@ class EditDistanceOp : public framework::OperatorWithKernel { class EditDistanceOpMaker : public framework::OpProtoAndCheckerMaker { public: - EditDistanceOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Hyps", "(2-D LoDTensor, 2nd dim. equal to 1) " "The indices for hypothesis strings."); diff --git a/paddle/fluid/operators/elementwise_add_op.cc b/paddle/fluid/operators/elementwise_add_op.cc index 4aab54f60236ecc5fa7f70e22f1553c3bfe68198..d2c20537136fc3ac9d1bece24a2238f26215c922 100644 --- a/paddle/fluid/operators/elementwise_add_op.cc +++ b/paddle/fluid/operators/elementwise_add_op.cc @@ -14,26 +14,8 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_add_op.h" #include "paddle/fluid/operators/elementwise_op.h" - -namespace paddle { -namespace operators { -class ElementwiseAddOpMaker : public ElementwiseOpMaker { - public: - ElementwiseAddOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Add", "Out = X + Y"); - AddComment(comment_); - } -}; -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OPERATOR(elementwise_add, ops::ElementwiseOp, - ops::ElementwiseAddOpMaker, ops::ElementwiseOpInferVarType, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(elementwise_add_grad, ops::ElementwiseOpGrad); - +REGISTER_ELEMWISE_OP(elementwise_add, "Add", "Out = X + Y"); REGISTER_OP_CPU_KERNEL( elementwise_add, ops::ElementwiseAddKernel, diff --git a/paddle/fluid/operators/elementwise_div_op.cc b/paddle/fluid/operators/elementwise_div_op.cc index c7ddafcad1d1f6c14791fde665f43881d6b49836..824b1221e5a77c8799dc34820b7f0db180c2439e 100644 --- a/paddle/fluid/operators/elementwise_div_op.cc +++ b/paddle/fluid/operators/elementwise_div_op.cc @@ -14,26 +14,8 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_div_op.h" #include "paddle/fluid/operators/elementwise_op.h" - -namespace paddle { -namespace operators { -class ElementwiseDivOpMaker : public ElementwiseOpMaker { - public: - ElementwiseDivOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Div", "Out = X / Y"); - AddComment(comment_); - } -}; - -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OPERATOR(elementwise_div, ops::ElementwiseOp, - ops::ElementwiseDivOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(elementwise_div_grad, ops::ElementwiseOpGrad); +REGISTER_ELEMWISE_OP(elementwise_div, "Div", "Out = X / Y"); REGISTER_OP_CPU_KERNEL( elementwise_div, ops::ElementwiseDivKernel, diff --git a/paddle/fluid/operators/elementwise_max_op.cc b/paddle/fluid/operators/elementwise_max_op.cc index a4fe386bb1907bf7c0099d2b1109077b21146948..411671335a19ae2283ca9db8b8f6bcbb6a6b630a 100644 --- a/paddle/fluid/operators/elementwise_max_op.cc +++ b/paddle/fluid/operators/elementwise_max_op.cc @@ -14,25 +14,8 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_max_op.h" #include "paddle/fluid/operators/elementwise_op.h" - -namespace paddle { -namespace operators { -class ElementwiseMaxOpMaker : public ElementwiseOpMaker { - public: - ElementwiseMaxOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Max", "Out = max(X, Y)"); - AddComment(comment_); - } -}; -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OPERATOR(elementwise_max, ops::ElementwiseOp, - ops::ElementwiseMaxOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(elementwise_max_grad, ops::ElementwiseOpGrad); +REGISTER_ELEMWISE_OP(elementwise_max, "Max", "Out = max(X, Y)"); REGISTER_OP_CPU_KERNEL( elementwise_max, ops::ElementwiseMaxKernel, diff --git a/paddle/fluid/operators/elementwise_min_op.cc b/paddle/fluid/operators/elementwise_min_op.cc index 68cd6ddb4a938b2b1c33e3f89c6d1151acb27f48..816192083d2275b26e6dd9afc76f2c021a01cf73 100644 --- a/paddle/fluid/operators/elementwise_min_op.cc +++ b/paddle/fluid/operators/elementwise_min_op.cc @@ -14,25 +14,8 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_min_op.h" #include "paddle/fluid/operators/elementwise_op.h" - -namespace paddle { -namespace operators { -class ElementwiseMinOpMaker : public ElementwiseOpMaker { - public: - ElementwiseMinOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Max", "Out = min(X, Y)"); - AddComment(comment_); - } -}; -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OPERATOR(elementwise_min, ops::ElementwiseOp, - ops::ElementwiseMinOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(elementwise_min_grad, ops::ElementwiseOpGrad); +REGISTER_ELEMWISE_OP(elementwise_min, "Min", "Out = min(X, Y)"); REGISTER_OP_CPU_KERNEL( elementwise_min, ops::ElementwiseMinKernel, diff --git a/paddle/fluid/operators/elementwise_mul_op.cc b/paddle/fluid/operators/elementwise_mul_op.cc index 2dec27136ad57ea032d5abb51799bd04ccc0b2e3..ba343909bb87b4f2efa56c0a4ff664b278e90c60 100644 --- a/paddle/fluid/operators/elementwise_mul_op.cc +++ b/paddle/fluid/operators/elementwise_mul_op.cc @@ -14,27 +14,8 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_mul_op.h" #include "paddle/fluid/operators/elementwise_op.h" - -namespace paddle { -namespace operators { - -class ElementwiseMulOpMaker : public ElementwiseOpMaker { - public: - ElementwiseMulOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Mul", "Out = X \\odot\\ Y"); - AddComment(comment_); - } -}; - -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OPERATOR(elementwise_mul, ops::ElementwiseOp, - ops::ElementwiseMulOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(elementwise_mul_grad, ops::ElementwiseOpGrad); +REGISTER_ELEMWISE_OP(elementwise_mul, "Mul", "Out = X \\odot\\ Y"); REGISTER_OP_CPU_KERNEL( elementwise_mul, ops::ElementwiseMulKernel, diff --git a/paddle/fluid/operators/elementwise_op.h b/paddle/fluid/operators/elementwise_op.h index a33634ab2503f988a8a692682ddb238d4794a3c0..d5b57cc2524efcdee112b2ce41cdcd4697fb79e6 100644 --- a/paddle/fluid/operators/elementwise_op.h +++ b/paddle/fluid/operators/elementwise_op.h @@ -54,8 +54,7 @@ class ElementwiseOpInferVarType : public framework::VarTypeInference { class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker { public: - ElementwiseOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() final { AddInput("X", "(Tensor), The first input tensor of elementwise op."); AddInput("Y", "(Tensor), The second input tensor of elementwise op."); AddOutput("Out", "The output of elementwise op."); @@ -64,12 +63,12 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker { "for broadcasting Y onto X.") .SetDefault(-1) .EqualGreaterThan(-1); - comment_ = R"DOC( -Limited Elementwise {name} Operator. + AddComment(string::Sprintf(R"DOC( +Limited Elementwise %s Operator. The equation is: -$${equation}$$ +$$%s$$ $X$ is a tensor of any dimension and the dimensions of tensor $Y$ must be smaller than or equal to the dimensions of $X$. @@ -100,26 +99,13 @@ For example Either of the inputs $X$ and $Y$ or none can carry the LoD (Level of Details) information. However, the output only shares the LoD information with input $X$. -)DOC"; - AddComment(comment_); +)DOC", + GetName(), GetEquation())); } protected: - std::string comment_; - - void Replace(std::string* src, std::string from, std::string to) { - std::size_t len_from = std::strlen(from.c_str()); - std::size_t len_to = std::strlen(to.c_str()); - for (std::size_t pos = src->find(from); pos != std::string::npos; - pos = src->find(from, pos + len_to)) { - src->replace(pos, len_from, to); - } - } - - void SetComment(std::string name, std::string equation) { - Replace(&comment_, "{name}", name); - Replace(&comment_, "{equation}", equation); - } + virtual std::string GetName() const = 0; + virtual std::string GetEquation() const = 0; }; class ElementwiseOpGrad : public framework::OperatorWithKernel { @@ -152,3 +138,16 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel { }; } // namespace operators } // namespace paddle + +#define REGISTER_ELEMWISE_OP(op_type, op_name, equation) \ + class __ElemwiseOp##op_type##Maker__ \ + : public ::paddle::operators::ElementwiseOpMaker { \ + protected: \ + virtual std::string GetName() const { return op_name; } \ + virtual std::string GetEquation() const { return equation; } \ + }; \ + REGISTER_OPERATOR(op_type, ::paddle::operators::ElementwiseOp, \ + __ElemwiseOp##op_type##Maker__, \ + ::paddle::operators::ElementwiseOpInferVarType, \ + ::paddle::framework::DefaultGradOpDescMaker); \ + REGISTER_OPERATOR(op_type##_grad, ::paddle::operators::ElementwiseOpGrad) diff --git a/paddle/fluid/operators/elementwise_pow_op.cc b/paddle/fluid/operators/elementwise_pow_op.cc index 60302c5e59f8ce595861405713045b05d90002e3..5fd6bde9ba0930e29f2161f1ff23ff9f5e7dc85d 100644 --- a/paddle/fluid/operators/elementwise_pow_op.cc +++ b/paddle/fluid/operators/elementwise_pow_op.cc @@ -13,17 +13,15 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/elementwise_pow_op.h" +#include #include "paddle/fluid/operators/elementwise_op.h" namespace paddle { namespace operators { class ElementwisePowOpMaker : public ElementwiseOpMaker { - public: - ElementwisePowOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Pow", "Out = X ^ Y"); - AddComment(comment_); - } + protected: + std::string GetName() const override { return "Pow"; } + std::string GetEquation() const override { return "Out = X ^ Y"; } }; } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/elementwise_sub_op.cc b/paddle/fluid/operators/elementwise_sub_op.cc index 9d0598fc39a3922fa830f18729d90a7dac6a890b..a7562b166b373ee2a8c9b6f379431d88d3e45fcb 100644 --- a/paddle/fluid/operators/elementwise_sub_op.cc +++ b/paddle/fluid/operators/elementwise_sub_op.cc @@ -14,25 +14,8 @@ limitations under the License. */ #include "paddle/fluid/operators/elementwise_sub_op.h" #include "paddle/fluid/operators/elementwise_op.h" - -namespace paddle { -namespace operators { -class ElementwiseSubOpMaker : public ElementwiseOpMaker { - public: - ElementwiseSubOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : ElementwiseOpMaker(proto, op_checker) { - SetComment("Sub", "Out = X - Y"); - AddComment(comment_); - } -}; -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OPERATOR(elementwise_sub, ops::ElementwiseOp, - ops::ElementwiseSubOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(elementwise_sub_grad, ops::ElementwiseOpGrad); +REGISTER_ELEMWISE_OP(elementwise_sub, "Sub", "Out = X - Y"); REGISTER_OP_CPU_KERNEL( elementwise_sub, ops::ElementwiseSubKernel, diff --git a/paddle/fluid/operators/expand_op.cc b/paddle/fluid/operators/expand_op.cc index 4ae91d074d3df8b910a7f5d816a22b6f1d51dff6..5ad0ec251328cc1ba580026bb47bf05316e7dc77 100644 --- a/paddle/fluid/operators/expand_op.cc +++ b/paddle/fluid/operators/expand_op.cc @@ -56,8 +56,7 @@ class ExpandOp : public framework::OperatorWithKernel { class ExpandOpMaker : public framework::OpProtoAndCheckerMaker { public: - ExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor, default Tensor). A tensor with rank in [1, 6]." "X is the input to be expanded."); diff --git a/paddle/fluid/operators/fc_op.cc b/paddle/fluid/operators/fc_op.cc index 45e4d5b2b863a55ae0aa0414ff8697141fd2aa6f..8843a1c44b7004ba5d7935f75d3c99d9c30fc6c0 100644 --- a/paddle/fluid/operators/fc_op.cc +++ b/paddle/fluid/operators/fc_op.cc @@ -72,8 +72,7 @@ framework::OpKernelType FCOpGrad::GetExpectedKernelType( layout, library); } -FCOpMaker::FCOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void FCOpMaker::Make() { AddInput("Input", "(Tensor) The input tensor of fully connected operator. "); AddInput("W", "(Tensor), The second input tensor of fc op."); AddOutput("Out", "(Tensor) The output tensor of fully connected operator. "); diff --git a/paddle/fluid/operators/fc_op.h b/paddle/fluid/operators/fc_op.h index 70fa96440d344397a7427c1338afee85bde923d4..e1b780fc0c401fbf34a9db03aa31137cbc016939 100644 --- a/paddle/fluid/operators/fc_op.h +++ b/paddle/fluid/operators/fc_op.h @@ -45,7 +45,7 @@ class FCOpGrad : public framework::OperatorWithKernel { class FCOpMaker : public framework::OpProtoAndCheckerMaker { public: - FCOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; } // namespace operators diff --git a/paddle/fluid/operators/feed_op.cc b/paddle/fluid/operators/feed_op.cc index debacf07c360b9aa69000a0d891f04239ed08807..bcb3e63ed7dbc775c1de6c4522f0548ea48a6cf0 100644 --- a/paddle/fluid/operators/feed_op.cc +++ b/paddle/fluid/operators/feed_op.cc @@ -66,8 +66,7 @@ class FeedOp : public framework::OperatorBase { class FeedOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - FeedOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of feed op"); AddOutput("Out", "The output of feed op"); AddAttr("col", "(int) The column of feed"); diff --git a/paddle/fluid/operators/fetch_op.cc b/paddle/fluid/operators/fetch_op.cc index 18deec58137676a0b2c8d559e49d0f7a840cd5ba..1640a2a22c69a0e3ab81a2889d6105b2cf4162b7 100644 --- a/paddle/fluid/operators/fetch_op.cc +++ b/paddle/fluid/operators/fetch_op.cc @@ -66,8 +66,7 @@ class FetchOp : public framework::OperatorBase { class FetchOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - FetchOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of fetch op"); AddOutput("Out", "The output of fetch op"); AddAttr("col", "(int) The column of fetch"); diff --git a/paddle/fluid/operators/fill_constant_batch_size_like_op.cc b/paddle/fluid/operators/fill_constant_batch_size_like_op.cc index 72da80baaf9bb3286f09b7ae5fcf24326b391906..1ae78675a0cac8a72aeaef1227b631a41e4a10b2 100644 --- a/paddle/fluid/operators/fill_constant_batch_size_like_op.cc +++ b/paddle/fluid/operators/fill_constant_batch_size_like_op.cc @@ -30,9 +30,8 @@ class FillConstantBatchSizeLikeOp : public BatchSizeLikeOp { }; class FillConstantBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { - public: - FillConstantBatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : BatchSizeLikeOpMaker(proto, op_checker) { + protected: + void Apply() override { AddAttr("dtype", "(int, default 5 (FP32)) " "Output data type") diff --git a/paddle/fluid/operators/fill_constant_op.cc b/paddle/fluid/operators/fill_constant_op.cc index 07e0a80f8d644d4d011f2821785d49ece6cecfb5..130f18dde4f979a6a9925ede9cbf745fcec14d48 100644 --- a/paddle/fluid/operators/fill_constant_op.cc +++ b/paddle/fluid/operators/fill_constant_op.cc @@ -59,8 +59,7 @@ class FillConstantOp : public framework::OperatorBase { class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker { public: - FillConstantOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddAttr("dtype", "(int, default 5 (FP32)) " "Output data type") diff --git a/paddle/fluid/operators/fill_op.cc b/paddle/fluid/operators/fill_op.cc index ee8a2fc353f86cdabd35459a9195c3aa35f63e31..925dc19061e2196a40411f415eb6e5ad59ab52ff 100644 --- a/paddle/fluid/operators/fill_op.cc +++ b/paddle/fluid/operators/fill_op.cc @@ -82,8 +82,7 @@ class FillOp : public framework::OperatorBase { class FillOpMaker : public framework::OpProtoAndCheckerMaker { public: - FillOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddComment(R"DOC(Fill operator Fill an tensor with `value` and `shape`. The type of the tensor is specify by diff --git a/paddle/fluid/operators/fill_zeros_like_op.cc b/paddle/fluid/operators/fill_zeros_like_op.cc index 58c814ba6413626a48310da595a13238994f5ef1..d67bec36b3248be8602da562a88aeb58f5effe39 100644 --- a/paddle/fluid/operators/fill_zeros_like_op.cc +++ b/paddle/fluid/operators/fill_zeros_like_op.cc @@ -33,8 +33,7 @@ class FillZerosLikeOp : public framework::OperatorWithKernel { class FillZerosLikeOpMaker : public framework::OpProtoAndCheckerMaker { public: - FillZerosLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of fill-zeros-like op."); AddOutput("Out", "The variable will be filled up with zeros."); AddComment(R"DOC( diff --git a/paddle/fluid/operators/ftrl_op.cc b/paddle/fluid/operators/ftrl_op.cc index cbdcce9beb3fafb0775d0b5fc39cb381ad128d0c..70ba25c213046cc934f46be067080d5fdbb42f9e 100644 --- a/paddle/fluid/operators/ftrl_op.cc +++ b/paddle/fluid/operators/ftrl_op.cc @@ -64,8 +64,7 @@ class FTRLOp : public framework::OperatorWithKernel { class FTRLOpMaker : public framework::OpProtoAndCheckerMaker { public: - FTRLOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor, default Tensor) " "Input parameter value that has to be updated."); diff --git a/paddle/fluid/operators/gather_op.cc b/paddle/fluid/operators/gather_op.cc index 4c82f5c429038504d9876ee240a705911feb0b7a..e21b57258928856a10d6e86c3e2c6e81fb241ee3 100644 --- a/paddle/fluid/operators/gather_op.cc +++ b/paddle/fluid/operators/gather_op.cc @@ -67,8 +67,7 @@ class GatherGradOp : public framework::OperatorWithKernel { class GatherOpMaker : public framework::OpProtoAndCheckerMaker { public: - GatherOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The source input of gather op"); AddInput("Index", "The index input of gather op"); AddOutput("Out", "The output of gather op"); diff --git a/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc b/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc index 53c706a83e5bfb9e93d485141314e8b652d73593..8050f61d4546f3351645f23ddcc63b2c49f17929 100644 --- a/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc +++ b/paddle/fluid/operators/gaussian_random_batch_size_like_op.cc @@ -32,9 +32,8 @@ class GaussianRandomBatchSizeLikeOp : public BatchSizeLikeOp { }; class GaussianRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { - public: - GaussianRandomBatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : BatchSizeLikeOpMaker(proto, op_checker) { + protected: + void Apply() override { AddAttr("mean", "(float, default 0.0) " "mean of random tensor.") diff --git a/paddle/fluid/operators/gaussian_random_op.cc b/paddle/fluid/operators/gaussian_random_op.cc index 4d197637b3f49f7e63f5b1a5cba212d1bf774f7e..815c1bb50988be49ca9996e368a59344c6583d58 100644 --- a/paddle/fluid/operators/gaussian_random_op.cc +++ b/paddle/fluid/operators/gaussian_random_op.cc @@ -70,8 +70,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel { class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker { public: - GaussianRandomOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput("Out", "Output matrix of gaussian random op"); AddAttr>("shape", diff --git a/paddle/fluid/operators/gen_nccl_id_op.cc b/paddle/fluid/operators/gen_nccl_id_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a5678f63466d368b3dd59380c18f9625cabd368b --- /dev/null +++ b/paddle/fluid/operators/gen_nccl_id_op.cc @@ -0,0 +1,128 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include + +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/threadpool.h" +#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/grpc_server.h" +#include "paddle/fluid/platform/nccl_helper.h" + +namespace paddle { +namespace operators { + +class GenNCCLIdOp : public framework::OperatorBase { + public: + GenNCCLIdOp(const std::string& type, const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + void RunImpl(const framework::Scope& scope, + const platform::Place& dev_place) const override { + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + // put nccl id in CPUPlace + auto& dev_ctx = *pool.Get(platform::CPUPlace()); + int trainer_id = Attr("trainer_id"); + framework::Scope& local_scope = scope.NewScope(); + + if (trainer_id == 0) { + GenerateAndSend(&local_scope, dev_ctx); + } else { + GetIdByServer(&local_scope, dev_ctx); + } + } + + private: + void GenerateAndSend(framework::Scope* scope, + const platform::DeviceContext& dev_ctx) const { + auto var = scope->FindVar(NCCL_ID_VARNAME); + PADDLE_ENFORCE_NOT_NULL(var); + auto id = var->GetMutable(); + PADDLE_ENFORCE(platform::dynload::ncclGetUniqueId(id)); + + std::vector endpoint_list = + Attr>("endpoint_list"); + detail::RPCClient client; + for (auto& ep : endpoint_list) { + VLOG(3) << "sending nccl id to " << ep; + client.AsyncSendVariable(ep, dev_ctx, *scope, NCCL_ID_VARNAME); + } + client.Wait(); + VLOG(3) << "sending completed..."; + } + + void GetIdByServer(framework::Scope* scope, + const platform::DeviceContext& dev_ctx) const { + std::string endpoint = Attr("endpoint"); + // NOTE: Can not use unique_ptr here because the default + // deleter will call GRPC Server's base class's dtor and + // that will cause a wired crash. + detail::AsyncGRPCServer rpc_service(endpoint, true); + framework::ProgramDesc empty_program; + framework::Executor executor(dev_ctx.GetPlace()); + rpc_service.SetScope(scope); + rpc_service.SetDevCtx(&dev_ctx); + rpc_service.SetProgram(&empty_program); + rpc_service.SetExecutor(&executor); + + std::thread server_thread( + std::bind(&detail::AsyncGRPCServer::RunSyncUpdate, &rpc_service)); + rpc_service.SetCond(0); + VLOG(3) << "start getting nccl id from trainer 0..."; + auto recv = rpc_service.Get(); + VLOG(3) << "got nccl id and stop server..."; + rpc_service.ShutDown(); + VLOG(3) << "rpc server stopped"; + server_thread.join(); + } +}; + +class GenNCCLIdOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddOutput("NCCLID", "Raw variable contains a NCCL UniqueId instaces."); + AddComment(R"DOC( +GenNCCLId operator + +For trainer 0: generate a new UniqueId and send it to all the other trainers. +For trainer 1~n: start a gRPC server to get the UniqueId, once got, stop the server. +)DOC"); + AddAttr("endpoint", + "(string), e.g. 127.0.0.1:6175 " + "current listen endpoint"); + AddAttr>( + "endpoint_list", + "['trainer1_ip:port', 'trainer2_ip:port', ...] " + "list of trainer endpoints start from trainer 1") + .SetDefault({}); + AddAttr("trainer_id", + "(int default 0) " + "The index of the trainer in distributed training.") + .SetDefault(0); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(gen_nccl_id, ops::GenNCCLIdOp, ops::GenNCCLIdOpMaker); diff --git a/paddle/fluid/operators/get_places_op.cc b/paddle/fluid/operators/get_places_op.cc index 0d7219ac5c624236b85916d5faf6810dbed2198a..eafc364a15fa17cc5107bba737b0b44e712b0bef 100644 --- a/paddle/fluid/operators/get_places_op.cc +++ b/paddle/fluid/operators/get_places_op.cc @@ -78,8 +78,7 @@ class GetPlacesOp : public framework::OperatorBase { class GetPlacesOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - GetPlacesOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput("Out", "vector of Place"); AddAttr("device_count", "device count").SetDefault(0); AddAttr("device_type", "device type") diff --git a/paddle/fluid/operators/go_op.cc b/paddle/fluid/operators/go_op.cc index b8e1556c23a3b7357ed56d1b83c09622559040a4..48f9d967adc90838dc4c7a09bfaf5a5a1ac9c99b 100644 --- a/paddle/fluid/operators/go_op.cc +++ b/paddle/fluid/operators/go_op.cc @@ -89,8 +89,7 @@ class GoOp : public framework::OperatorBase { class GoOpMaker : public framework::OpProtoAndCheckerMaker { public: - GoOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kX, "A set of variables, which are required by operators inside the " "block of Go Op.") diff --git a/paddle/fluid/operators/gru_op.cc b/paddle/fluid/operators/gru_op.cc index 0a524c914d305661745c5d85cbbee2edb57c97ba..5c746878823b3dcde2573feec00d3d9dac5ceab8 100644 --- a/paddle/fluid/operators/gru_op.cc +++ b/paddle/fluid/operators/gru_op.cc @@ -71,8 +71,7 @@ class GRUOp : public framework::OperatorWithKernel { class GRUOpMaker : public framework::OpProtoAndCheckerMaker { public: - GRUOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(LoDTensor) The first input is a LodTensor, which supports " "variable-time length input sequence. The underlying tensor in " diff --git a/paddle/fluid/operators/gru_unit_op.cc b/paddle/fluid/operators/gru_unit_op.cc index f8d1d44b5423dd09fe5aad11434911af6f14fe77..82a808b01e99ec33b0ca00a065fb301d3c633b19 100644 --- a/paddle/fluid/operators/gru_unit_op.cc +++ b/paddle/fluid/operators/gru_unit_op.cc @@ -71,8 +71,7 @@ class GRUUnitOp : public framework::OperatorWithKernel { class GRUUnitOpMaker : public framework::OpProtoAndCheckerMaker { public: - GRUUnitOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(Tensor) Matrix with shape [batch_size, frame_size * 3] for the " "input."); diff --git a/paddle/fluid/operators/hinge_loss_op.cc b/paddle/fluid/operators/hinge_loss_op.cc index 086b5a97dec9a3d5b8f91b802b92d64ca73bf57c..69e7fa4490b892373d85898b13b976a474a6096a 100644 --- a/paddle/fluid/operators/hinge_loss_op.cc +++ b/paddle/fluid/operators/hinge_loss_op.cc @@ -46,8 +46,7 @@ class HingeLossOp : public framework::OperatorWithKernel { template class HingeLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - HingeLossOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Logits", "The input value (Logits) of Hinge loss op." "Logits is a 2-D tensor with shape [batch_size, 1]."); diff --git a/paddle/fluid/operators/huber_loss_op.cc b/paddle/fluid/operators/huber_loss_op.cc index 74d8e0e2b76adc7a3e69649f277a8c0df6f38056..4ecd8634ff41ff4eba6b5ed1d0fc78068190dce5 100644 --- a/paddle/fluid/operators/huber_loss_op.cc +++ b/paddle/fluid/operators/huber_loss_op.cc @@ -45,8 +45,7 @@ class HuberLossOp : public framework::OperatorWithKernel { template class HuberLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - HuberLossOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input value of huber loss op." "X is a 2-D tensor with shape [batch_size, 1]."); diff --git a/paddle/fluid/operators/im2sequence_op.cc b/paddle/fluid/operators/im2sequence_op.cc index 8c120eec86601146500721bbb4249bc458190093..0669661d225c664010fce97f0a526b62988b92c5 100644 --- a/paddle/fluid/operators/im2sequence_op.cc +++ b/paddle/fluid/operators/im2sequence_op.cc @@ -54,8 +54,7 @@ class Im2SequenceOp : public framework::OperatorWithKernel { class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker { public: - Im2SequenceOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input tensor has NCHW format." "N: batch size" diff --git a/paddle/fluid/operators/increment_op.cc b/paddle/fluid/operators/increment_op.cc index d8c97b27b328b1470bece4a6c1872b5ccc75115e..f0ffc9706689f5afe4546c3483114b38bc2b7872 100644 --- a/paddle/fluid/operators/increment_op.cc +++ b/paddle/fluid/operators/increment_op.cc @@ -47,8 +47,7 @@ class IncrementOp : public framework::OperatorWithKernel { class IncrementOpMaker : public framework::OpProtoAndCheckerMaker { public: - IncrementOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input tensor of increment operator"); AddOutput("Out", "(Tensor) The output tensor of increment operator."); AddAttr("step", diff --git a/paddle/fluid/operators/is_empty_op.cc b/paddle/fluid/operators/is_empty_op.cc index 2a7be90dab1cc23ffe5e1c296c37a4bbeacb7d8e..29b73951bbddd9bfd73c932d7801797590de5e8e 100644 --- a/paddle/fluid/operators/is_empty_op.cc +++ b/paddle/fluid/operators/is_empty_op.cc @@ -12,46 +12,41 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include "paddle/fluid/operators/is_empty_op.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" namespace paddle { namespace operators { -constexpr char kInput[] = "X"; -constexpr char kOutput[] = "Out"; - -class IsEmptyOp : public framework::OperatorBase { +class IsEmptyOp : public framework::OperatorWithKernel { public: - IsEmptyOp(const std::string &type, const framework::VariableNameMap &inputs, - const framework::VariableNameMap &outputs, - const framework::AttributeMap &attrs) - : OperatorBase(type, inputs, outputs, attrs) {} + using framework::OperatorWithKernel::OperatorWithKernel; - private: - void RunImpl(const framework::Scope &scope, - const platform::Place &place) const override { - // get input - auto *var = scope.FindVar(Input(kInput)); - PADDLE_ENFORCE_NOT_NULL(var); - auto &tensor = var->Get(); - // get output - auto *out = scope.FindVar(Output(kOutput)); - PADDLE_ENFORCE_NOT_NULL(out); - auto *out_tensor = out->GetMutable(); + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of IsEmptyOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of IsEmptyOp should not be null."); + ctx->SetOutputDim("Out", {1}); + } - out_tensor->Resize({1}); - out_tensor->mutable_data(platform::CPUPlace())[0] = - framework::product(tensor.dims()) == 0; + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + framework::OpKernelType kt = framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + platform::CPUPlace()); + return kt; } }; -class IsEmptyOpProtoMaker : public framework::OpProtoAndCheckerMaker { +class IsEmptyOpMaker : public framework::OpProtoAndCheckerMaker { public: - IsEmptyOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput(kInput, "(Tensor) Tensor which is to be checked."); - AddOutput(kOutput, "(Tensor) a boolean Tensor that indicate empty or not."); + void Make() override { + AddInput("X", "(LoDTensor) Tensor which is to be checked."); + AddOutput("Out", + "(LoDTensor) a boolean Tensor that indicate empty or not."); AddComment(R"DOC( IsEmpty Operator which checks whether a tensor is empty. @@ -63,5 +58,12 @@ It will just return product(tensor.ddims()) > 0; } // namespace operators } // namespace paddle -REGISTER_OP_WITHOUT_GRADIENT(is_empty, paddle::operators::IsEmptyOp, - paddle::operators::IsEmptyOpProtoMaker); +namespace ops = paddle::operators; + +REGISTER_OPERATOR(is_empty, ops::IsEmptyOp, ops::IsEmptyOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL( + is_empty, ops::IsEmptyOpKernel, + ops::IsEmptyOpKernel, + ops::IsEmptyOpKernel, + ops::IsEmptyOpKernel); diff --git a/paddle/fluid/operators/is_empty_op.h b/paddle/fluid/operators/is_empty_op.h new file mode 100644 index 0000000000000000000000000000000000000000..3e3af22fa8d842b6a1e67418446f1a40949e046b --- /dev/null +++ b/paddle/fluid/operators/is_empty_op.h @@ -0,0 +1,37 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" + +namespace paddle { +namespace operators { + +template +class IsEmptyOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + // get input + auto* input_tensor = context.Input("X"); + // get output + auto* output_tensor = context.Output("Out"); + + output_tensor->mutable_data(platform::CPUPlace())[0] = + framework::product(input_tensor->dims()) == 0; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/l1_norm_op.cc b/paddle/fluid/operators/l1_norm_op.cc index 0c143b7c8aed13a202e2597632d17d8bccc8b66d..bc115090acb473ac3175999ca96c5e00c0aeaeae 100644 --- a/paddle/fluid/operators/l1_norm_op.cc +++ b/paddle/fluid/operators/l1_norm_op.cc @@ -48,8 +48,7 @@ class L1NormGradOp : public framework::OperatorWithKernel { class L1NormOpMaker : public framework::OpProtoAndCheckerMaker { public: - L1NormOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input of l1_norm op."); AddOutput("Out", "(Scalar) The output of l1_norm op."); AddComment(R"DOC( diff --git a/paddle/fluid/operators/label_smooth_op.cc b/paddle/fluid/operators/label_smooth_op.cc index a73c626032f3bf6e97ac5974424e76bacb9a0799..da59bd53bce010d0d6ad2ab14acaffb9cc2f99e6 100644 --- a/paddle/fluid/operators/label_smooth_op.cc +++ b/paddle/fluid/operators/label_smooth_op.cc @@ -47,8 +47,7 @@ class LabelSmoothOp : public framework::OperatorWithKernel { class LabelSmoothOpMaker : public framework::OpProtoAndCheckerMaker { public: - LabelSmoothOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) The input labels of LabelSmooth operator. This " "input can be batched labels in one-hot encoding or output from " diff --git a/paddle/fluid/operators/layer_norm_op.cc b/paddle/fluid/operators/layer_norm_op.cc index de1056aef7bfa2f53f8a92b262e7d15aa7c2b75c..ab097d31e9ab5eafa788539170e7e405df697625 100644 --- a/paddle/fluid/operators/layer_norm_op.cc +++ b/paddle/fluid/operators/layer_norm_op.cc @@ -61,8 +61,7 @@ class LayerNormOp : public framework::OperatorWithKernel { class LayerNormOpMaker : public framework::OpProtoAndCheckerMaker { public: - LayerNormOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) The input tensor."); AddInput("Scale", "(Tensor, optional) Scale is a 1-dimensional tensor of size " diff --git a/paddle/fluid/operators/linear_chain_crf_op.cc b/paddle/fluid/operators/linear_chain_crf_op.cc index 2f29e377fdada918f2c9dca8c2d94eb06278320d..e38525cd7f44de020f364ffd16e71a439048347f 100644 --- a/paddle/fluid/operators/linear_chain_crf_op.cc +++ b/paddle/fluid/operators/linear_chain_crf_op.cc @@ -19,8 +19,7 @@ namespace operators { class LinearChainCRFOpMaker : public framework::OpProtoAndCheckerMaker { public: - LinearChainCRFOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Emission", "(LoDTensor, default LoDTensor) " "A 2-D LoDTensor with shape [N x D], where N is the size of the " diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index 8acbf820250957163397342c645b333f0da0801c..57eb5d9a0e73a51d9e2cef7ad7539c1b9da2c4ea 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -57,8 +57,7 @@ static void ParallelExecuteBlocks( framework::Async([&executor, &prepared, &program, &scope, idx]() { int run_block = idx; // thread local try { - executor->RunPreparedContext(prepared[run_block].get(), scope, - false, false); + executor->RunPreparedContext(prepared[run_block].get(), scope); } catch (std::exception &e) { LOG(ERROR) << "run sub program error " << e.what(); } @@ -211,8 +210,8 @@ static void AsyncUpdateThread( } auto fs = framework::Async([var_name, &executor, &v, prepared] { try { - executor->RunPreparedContext(prepared, v.second->GetMutableLocalScope(), - false, false); + executor->RunPreparedContext(prepared, + v.second->GetMutableLocalScope()); } catch (std::exception &e) { LOG(ERROR) << "run sub program error " << e.what(); } @@ -322,8 +321,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, // prepare for prefetch VLOG(3) << "prefetch block id is " << prefetch_block->ID(); auto prefetch_prepared = executor.Prepare(*program, prefetch_block->ID()); - rpc_service_->SetPrefetchPreparedCtx(prefetch_prepared.get()); - prefetch_prepared.release(); + rpc_service_->SetPrefetchPreparedCtx(std::move(prefetch_prepared)); // start the server listening after all member initialized. server_thread_.reset(new std::thread(RunServer, rpc_service_)); @@ -343,8 +341,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker { public: - ListenAndServOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("X", "(Tensor) Variables that server recv.").AsDuplicable(); AddComment(R"DOC( ListenAndServ operator diff --git a/paddle/fluid/operators/load_combine_op.cc b/paddle/fluid/operators/load_combine_op.cc index e5353144e91455fc71460459e6e799b54f750f71..0522a94195786c767194ec727d982a60451e7c62 100644 --- a/paddle/fluid/operators/load_combine_op.cc +++ b/paddle/fluid/operators/load_combine_op.cc @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ #include - +#include "paddle/fluid/framework/data_type_transform.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/platform/device_context.h" @@ -31,6 +31,7 @@ class LoadCombineOp : public framework::OperatorBase { void RunImpl(const framework::Scope &scope, const platform::Place &place) const override { auto filename = Attr("file_path"); + auto load_as_fp16 = Attr("load_as_fp16"); std::ifstream fin(filename); PADDLE_ENFORCE(static_cast(fin), @@ -59,17 +60,25 @@ class LoadCombineOp : public framework::OperatorBase { // Get data from fin to tensor DeserializeFromStream(fin, tensor, dev_ctx); - if (platform::is_gpu_place(place)) { - // copy CPU to GPU - framework::LoDTensor cpu_tensor; - cpu_tensor.ShareDataWith(*tensor); - cpu_tensor.set_lod(tensor->lod()); - - // reset tensor + auto in_dtype = framework::ToDataType(tensor->type()); + auto out_dtype = + load_as_fp16 ? framework::proto::VarType::FP16 : in_dtype; + + if (in_dtype != out_dtype) { + // convert to float16 tensor + auto in_kernel_type = framework::OpKernelType(in_dtype, place); + auto out_kernel_type = framework::OpKernelType(out_dtype, place); + framework::LoDTensor fp16_tensor; + // copy LoD info to the new tensor + fp16_tensor.set_lod(tensor->lod()); + framework::TransDataType(in_kernel_type, out_kernel_type, *tensor, + &fp16_tensor); + + // reset output tensor out_var->Clear(); tensor = out_var->GetMutable(); - tensor->set_lod(cpu_tensor.lod()); - TensorCopy(cpu_tensor, place, dev_ctx, tensor); + tensor->set_lod(fp16_tensor.lod()); + tensor->ShareDataWith(fp16_tensor); } } } @@ -77,12 +86,18 @@ class LoadCombineOp : public framework::OperatorBase { class LoadCombineOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoadCombineOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput( "Out", "(vector) The output LoDTensors that will be read from the input file.") .AsDuplicable(); + AddAttr( + "load_as_fp16", + "(boolean, default false)" + "If true, the tensor will be first loaded and then " + "converted to float16 data type. Otherwise, the tensor will be " + "directly loaded without data type conversion.") + .SetDefault(false); AddAttr("file_path", "(string) " "LoDTensors will be loaded from \"file_path\".") diff --git a/paddle/fluid/operators/load_op.cc b/paddle/fluid/operators/load_op.cc index c6bd2bf3dfca77dc078eb04b1d90c7d90883203f..93f45cff8a26201b1fbb1c44141e125a67c44037 100644 --- a/paddle/fluid/operators/load_op.cc +++ b/paddle/fluid/operators/load_op.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include +#include "paddle/fluid/framework/data_type_transform.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/profiler.h" @@ -46,14 +47,41 @@ class LoadOp : public framework::OperatorBase { auto *tensor = out_var->GetMutable(); DeserializeFromStream(fin, tensor, *dev_ctx); + + auto load_as_fp16 = Attr("load_as_fp16"); + auto in_dtype = framework::ToDataType(tensor->type()); + auto out_dtype = load_as_fp16 ? framework::proto::VarType::FP16 : in_dtype; + + if (in_dtype != out_dtype) { + // convert to float16 tensor + auto in_kernel_type = framework::OpKernelType(in_dtype, place); + auto out_kernel_type = framework::OpKernelType(out_dtype, place); + framework::LoDTensor fp16_tensor; + // copy LoD info to the new tensor + fp16_tensor.set_lod(tensor->lod()); + framework::TransDataType(in_kernel_type, out_kernel_type, *tensor, + &fp16_tensor); + + // reset output tensor + out_var->Clear(); + tensor = out_var->GetMutable(); + tensor->set_lod(fp16_tensor.lod()); + tensor->ShareDataWith(fp16_tensor); + } } }; class LoadOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoadOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput("Out", "(Tensor) The tensor need to be loaded"); + AddAttr( + "load_as_fp16", + "(boolean, default false)" + "If true, the tensor will be first loaded and then " + "converted to float16 data type. Otherwise, the tensor will be " + "directly loaded without data type conversion.") + .SetDefault(false); AddAttr("file_path", "(string) " "Variable will be loaded from \"file_path\".") diff --git a/paddle/fluid/operators/lod_array_length_op.cc b/paddle/fluid/operators/lod_array_length_op.cc index e6212405770093455ec89bde9dc0a092b956fc83..e4551b8ba681fe92ac5f21bb0b509f43439f6b66 100644 --- a/paddle/fluid/operators/lod_array_length_op.cc +++ b/paddle/fluid/operators/lod_array_length_op.cc @@ -40,8 +40,7 @@ class LoDArrayLengthOp : public framework::OperatorBase { class LoDArrayLengthProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoDArrayLengthProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensorArray) The input tensor array."); AddOutput("Out", "(Tensor) 1x1 CPU Tensor of length, int64_t"); AddComment(R"DOC( diff --git a/paddle/fluid/operators/lod_rank_table_op.cc b/paddle/fluid/operators/lod_rank_table_op.cc index 590b44e14f518c3c60c141c9a0dfe7f2b96f69c6..166952fe23192799443ef9c9d1f7ba5056d19290 100644 --- a/paddle/fluid/operators/lod_rank_table_op.cc +++ b/paddle/fluid/operators/lod_rank_table_op.cc @@ -38,8 +38,7 @@ class LoDRankTableOp : public framework::OperatorBase { class LoDRankTableOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoDRankTableOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) input lod tensor, must contain lod information."); AddOutput("Out", "(LoDRankTable) The rank table of specific level."); diff --git a/paddle/fluid/operators/lod_reset_op.cc b/paddle/fluid/operators/lod_reset_op.cc index 92ebfc274b84f738f5bd688a9a6d9f437b6318aa..0d4e84e85083399e3803d0648dc7a10aa276d536 100644 --- a/paddle/fluid/operators/lod_reset_op.cc +++ b/paddle/fluid/operators/lod_reset_op.cc @@ -47,8 +47,7 @@ class LoDResetOp : public framework::OperatorWithKernel { class LoDResetOpMaker : public framework::OpProtoAndCheckerMaker { public: - LoDResetOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor, LoDTensor) Input variable of LoDResetOp which " "could be a Tensor or LoDTensor, where the data of output " diff --git a/paddle/fluid/operators/lod_tensor_to_array_op.cc b/paddle/fluid/operators/lod_tensor_to_array_op.cc index 543495ce4e66c0955c9ce1b0db480088069b36db..00ba5ce8ee5e4084c8af204cfc37fe80c437f0d7 100644 --- a/paddle/fluid/operators/lod_tensor_to_array_op.cc +++ b/paddle/fluid/operators/lod_tensor_to_array_op.cc @@ -105,8 +105,7 @@ class LoDTensorToArrayOp : public framework::OperatorBase { class LoDTensorToArrayOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - LoDTensorToArrayOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", ""); AddInput("RankTable", ""); AddOutput("Out", ""); diff --git a/paddle/fluid/operators/log_loss_op.cc b/paddle/fluid/operators/log_loss_op.cc index a8258a1afd70574c174abe8d5630ade5d4ac3de6..9d248e03218b83a65b9786cb317aafbe3dbb67ee 100644 --- a/paddle/fluid/operators/log_loss_op.cc +++ b/paddle/fluid/operators/log_loss_op.cc @@ -46,8 +46,7 @@ class LogLossOp : public framework::OperatorWithKernel { template class LogLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - LogLossOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Predicted", "The input value (Predicted) of Log loss op." "Predicted is a 2-D tensor with shape [batch_size, 1]."); diff --git a/paddle/fluid/operators/logical_op.cc b/paddle/fluid/operators/logical_op.cc index 41aa00ee8ac10e0776c066fc3c37f97b0dd40cc3..db109f5cd053d84718ac85bd4693ecece12ce172 100644 --- a/paddle/fluid/operators/logical_op.cc +++ b/paddle/fluid/operators/logical_op.cc @@ -21,8 +21,7 @@ namespace operators { template class BinaryLogicalOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - BinaryLogicalOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { OpComment comment; AddInput("X", string::Sprintf("(LoDTensor) Left hand operand of %s operator", @@ -45,8 +44,7 @@ Each element of Out is calculated by %s template class UnaryLogicalOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - UnaryLogicalOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { OpComment comment; AddInput("X", string::Sprintf("(LoDTensor) Operand of %s operator", comment.type)); diff --git a/paddle/fluid/operators/lookup_sparse_table_op.cc b/paddle/fluid/operators/lookup_sparse_table_op.cc index 66b626ed792ddec9d57fcf6c81655dffcc23ca99..d07a81968565f095cdb6425d104bc7a11bc9cfad 100644 --- a/paddle/fluid/operators/lookup_sparse_table_op.cc +++ b/paddle/fluid/operators/lookup_sparse_table_op.cc @@ -105,8 +105,7 @@ class LookupSparseTableOp : public framework::OperatorBase { class LookupSparseTableOpMaker : public framework::OpProtoAndCheckerMaker { public: - LookupSparseTableOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("W", "(SelectedRows) The input represents embedding table, " "which is a learnable parameter."); diff --git a/paddle/fluid/operators/lookup_table_op.cc b/paddle/fluid/operators/lookup_table_op.cc index 5e59bd1b178ad1803f6f70c5f3f9fd7af495ac3c..bda499432214b8841c8dfc406ee45ca0367920e7 100644 --- a/paddle/fluid/operators/lookup_table_op.cc +++ b/paddle/fluid/operators/lookup_table_op.cc @@ -58,8 +58,7 @@ class LookupTableOp : public framework::OperatorWithKernel { class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { public: - LookupTableOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("W", "(Tensor) The input represents embedding tensors, " "which is a learnable parameter."); diff --git a/paddle/fluid/operators/lrn_op.cc b/paddle/fluid/operators/lrn_op.cc index f5c0e47fda913b4635833c31496644b60a0a8504..52b9cd7fb7019b738098a8649f23277afd40e938 100644 --- a/paddle/fluid/operators/lrn_op.cc +++ b/paddle/fluid/operators/lrn_op.cc @@ -169,8 +169,7 @@ class LRNOp : public framework::OperatorWithKernel { template class LRNOpMaker : public framework::OpProtoAndCheckerMaker { public: - LRNOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input of LRN operator. " "It must be a 4D tenor with NCHW format."); diff --git a/paddle/fluid/operators/lstm_op.cc b/paddle/fluid/operators/lstm_op.cc index 084ee1cfe602af3622ef2a3f35f2892d5540cec7..4751e3e8025e51a687f8fcfd25e603b61e762f6d 100644 --- a/paddle/fluid/operators/lstm_op.cc +++ b/paddle/fluid/operators/lstm_op.cc @@ -103,8 +103,7 @@ class LSTMOp : public framework::OperatorWithKernel { class LSTMOpMaker : public framework::OpProtoAndCheckerMaker { public: - LSTMOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(LoDTensor) the first input is a LodTensor, which support " "variable-time length input sequence. The underlying tensor in " diff --git a/paddle/fluid/operators/lstm_unit_op.cc b/paddle/fluid/operators/lstm_unit_op.cc index e1157ef6c640be17e7f48abe1ab972cf88504526..0895c58f5f58afd444000ebeac7a92e3eb7778d3 100644 --- a/paddle/fluid/operators/lstm_unit_op.cc +++ b/paddle/fluid/operators/lstm_unit_op.cc @@ -48,8 +48,7 @@ class LstmUnitOp : public framework::OperatorWithKernel { class LstmUnitOpMaker : public framework::OpProtoAndCheckerMaker { public: - LstmUnitOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "Lstm unit only applies non-linear activations, please make sure" "that linear tranformation has already been applied to `X`. " diff --git a/paddle/fluid/operators/lstmp_op.cc b/paddle/fluid/operators/lstmp_op.cc index f9261323f0f50c78b3b4b66a9fa8abcdf5ba27e9..e398b51480f6fc0c6c568770b3b2a9746360744e 100644 --- a/paddle/fluid/operators/lstmp_op.cc +++ b/paddle/fluid/operators/lstmp_op.cc @@ -120,8 +120,7 @@ class LSTMPOp : public framework::OperatorWithKernel { class LSTMPOpMaker : public framework::OpProtoAndCheckerMaker { public: - LSTMPOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(LoDTensor) the input for sequence data, which supports " "variable-time length input sequence. The underlying tensor in " diff --git a/paddle/fluid/operators/margin_rank_loss_op.cc b/paddle/fluid/operators/margin_rank_loss_op.cc index 0b41a3e1ffdb32d248bb55651aba242336307e74..b643ba9d7fa61d758e871ebe7a463c22e937fa2c 100644 --- a/paddle/fluid/operators/margin_rank_loss_op.cc +++ b/paddle/fluid/operators/margin_rank_loss_op.cc @@ -42,8 +42,7 @@ class MarginRankLossOp : public framework::OperatorWithKernel { template class MarginRankLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - MarginRankLossOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X1", "(2-D tensor with shape [batch_size x 1]) The score for " "one item X1 to be ranked, from pairwise ranking model."); diff --git a/paddle/fluid/operators/math/blas.cc b/paddle/fluid/operators/math/blas.cc index 3eeb77546b97a0337b46216d837a4f4cff12c89f..6a143b3c056455595fdedc131b0c5f4ee756e1e0 100644 --- a/paddle/fluid/operators/math/blas.cc +++ b/paddle/fluid/operators/math/blas.cc @@ -13,10 +13,40 @@ // limitations under the License. #include "paddle/fluid/operators/math/blas.h" + +#include namespace paddle { namespace operators { namespace math { -// Do nothing. Blas is a header only library. +MatDescriptor CreateMatrixDescriptor(const framework::DDim &tensor_dim, + int num_flatten_cols, bool trans) { + PADDLE_ENFORCE_GT(tensor_dim.size(), 1); + MatDescriptor retv; + if (num_flatten_cols > 1) { + auto flatten_dim = framework::flatten_to_2d(tensor_dim, num_flatten_cols); + retv.height_ = flatten_dim[0]; + retv.width_ = flatten_dim[1]; + } else { + if (tensor_dim.size() == 2) { + retv.height_ = tensor_dim[0]; + retv.width_ = tensor_dim[1]; + } else { + auto dim_vec = framework::vectorize(tensor_dim); + retv.batch_size_ = 1; + for (size_t i = 0; i < dim_vec.size() - 2; ++i) { + retv.batch_size_ *= dim_vec[i]; + } + retv.height_ = dim_vec[dim_vec.size() - 2]; + retv.width_ = dim_vec[dim_vec.size() - 1]; + retv.stride_ = retv.height_ * retv.width_; + } + } + if (trans) { + std::swap(retv.width_, retv.height_); + } + retv.trans_ = trans; + return retv; +} } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/math/blas.h b/paddle/fluid/operators/math/blas.h index 5cd2f855d1135e6dd8343efdaa9855d2526a3520..dabde43850db770d286b13cacd32bee181328d5c 100644 --- a/paddle/fluid/operators/math/blas.h +++ b/paddle/fluid/operators/math/blas.h @@ -46,6 +46,50 @@ namespace paddle { namespace operators { namespace math { +/** + * Matrix Descriptor of a memory buffer. + * + * It is used for Blas::MatMul. MatMul operator can be batched. + * if Mat A is [BatchSize, H, W], Mat B is [BatchSize, H, W]. It will be a + * `batch_size` times of GEMM. The batched GEMM could be faster base on the + * implementation of the blas library. The batch size could be zero. If any + * matrix of `matmul` has a batch size, the will be a batched GEMM, too. e.g., + * Mat A is [BatchSize, H1, W2], and Mat B [H2, W2], The result matrix wil be + * [BatchSize, H1, W2] + * + * The boolean flag, `trans`, describe the memory is the transpose of matrix or + * not. If the trans is true, the last two dims of matrix are transposed. The + * memory layout of the matrix is [Width, Height] or [BatchSize, Width, Height]. + * + * The MatDescriptor is not only the dimension or shape of a matrix, it also + * contains the layout, stride of matrix. It is clearer to have a structure than + * reuse `DDim`. + */ +struct MatDescriptor { + int64_t height_; + int64_t width_; + int64_t stride_{0}; + int64_t batch_size_{0}; + bool trans_; +}; + +/** + * Create Matrix Descriptor from a tensor dim, num_flatten_cols, and transpose + * flag + * + * @param tensor_dim: The dimension of the tensor. The rank of this dimension + * must larger than 1. + * + * @param num_flatten_cols: Reshape a tensor to a matrix. The matrix's first + * dimension(column length) will be the product of tensor's first `num_col_dims` + * dimensions. If num_flatten_cols is zero, the first N-2 dimension will be the + * batch_size of descriptor. + * + * @param trans: True if the matrix is transposed. + */ +extern MatDescriptor CreateMatrixDescriptor(const framework::DDim& tensor_dim, + int num_flatten_cols, bool trans); + template class Blas { public: @@ -90,6 +134,11 @@ class Blas { int K, T alpha, const T* A, const T* B, T beta, T* C, int batchCount, int64_t strideA, int64_t strideB) const; + template + void MatMul(const framework::Tensor& mat_a, const MatDescriptor& dim_a, + const framework::Tensor& mat_b, const MatDescriptor& dim_b, + T alpha, framework::Tensor* mat_out, T beta) const; + private: const DeviceContext& context_; }; diff --git a/paddle/fluid/operators/math/blas_impl.cu.h b/paddle/fluid/operators/math/blas_impl.cu.h index c76fc17d78cce514b5e35ce8e5ca890d7cec1e98..d84c88cb3bc1a13acb83b3444dbd1bfca3cba503 100644 --- a/paddle/fluid/operators/math/blas_impl.cu.h +++ b/paddle/fluid/operators/math/blas_impl.cu.h @@ -96,10 +96,22 @@ struct CUBlas { reinterpret_cast<__half *>(C), ldc)); } - template - static void GEMM_BATCH(ARGS... args) { + static void GEMM_BATCH(cublasHandle_t handle, cublasOperation_t transa, + cublasOperation_t transb, int m, int n, int k, + const float16 *alpha, const float16 *A, int lda, + long long int strideA, const float16 *B, // NOLINT + int ldb, long long int strideB, // NOLINT + const float16 *beta, float16 *C, int ldc, + long long int strideC, // NOLINT + int batchCount) { #if CUDA_VERSION >= 8000 - PADDLE_ENFORCE(platform::dynload::cublasHgemmStridedBatched(args...)); + PADDLE_ENFORCE(platform::dynload::cublasHgemmStridedBatched( + handle, transa, transb, m, n, k, + reinterpret_cast(alpha), + reinterpret_cast(A), lda, strideA, + reinterpret_cast(B), ldb, strideB, + reinterpret_cast(beta), reinterpret_cast<__half *>(C), + ldc, strideC, batchCount)); #else PADDLE_THROW("HgemmStridedBatched is not supported on cuda <= 7.5"); #endif diff --git a/paddle/fluid/operators/math/blas_impl.h b/paddle/fluid/operators/math/blas_impl.h index 7360cc0a90da499c372c6fb3f8d40a26f9093dd8..14b3624b420cb883b36268c0a5a9e8692dbb5b43 100644 --- a/paddle/fluid/operators/math/blas_impl.h +++ b/paddle/fluid/operators/math/blas_impl.h @@ -172,14 +172,39 @@ void Blas::BatchedGEMM( c_array.data(), &ldc, 1 /* group_count */, &batchCount); #else for (int k = 0; k < batchCount; ++k) { - const float *Ak = &A[k * strideA]; - const float *Bk = &B[k * strideB]; - float *Ck = &C[k * M * N]; + auto *Ak = &A[k * strideA]; + auto *Bk = &B[k * strideB]; + auto *Ck = &C[k * M * N]; this->template GEMM(transA, transB, M, N, K, alpha, Ak, Bk, beta, Ck); } #endif } +template +template +void Blas::MatMul(const framework::Tensor &mat_a, + const MatDescriptor &dim_a, + const framework::Tensor &mat_b, + const MatDescriptor &dim_b, T alpha, + framework::Tensor *mat_out, T beta) const { + PADDLE_ENFORCE_EQ(dim_a.width_, dim_b.height_); + CBLAS_TRANSPOSE transA = !dim_a.trans_ ? CblasNoTrans : CblasTrans; + CBLAS_TRANSPOSE transB = !dim_b.trans_ ? CblasNoTrans : CblasTrans; + if (dim_a.batch_size_ == 0 && dim_b.batch_size_ == 0) { + this->template GEMM(transA, transB, dim_a.height_, dim_b.width_, + dim_a.width_, alpha, mat_a.data(), + mat_b.data(), beta, mat_out->data()); + } else { + PADDLE_ENFORCE(dim_a.batch_size_ == dim_b.batch_size_ || + dim_a.batch_size_ == 0 || dim_b.batch_size_ == 0); + this->template BatchedGEMM( + transA, transB, dim_a.height_, dim_b.width_, dim_a.width_, alpha, + mat_a.data(), mat_b.data(), beta, mat_out->data(), + dim_a.batch_size_ == 0 ? dim_b.batch_size_ : dim_a.batch_size_, + dim_a.stride_, dim_b.stride_); + } +} + } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/math/math_function.cc b/paddle/fluid/operators/math/math_function.cc index d62ea387cc55c7399973b6f35bace491a49666dc..d39154c6f88d6d17c1719eb9a5b048211f4bb52b 100644 --- a/paddle/fluid/operators/math/math_function.cc +++ b/paddle/fluid/operators/math/math_function.cc @@ -38,7 +38,9 @@ template struct SetConstant; template struct Transpose; \ template struct Transpose; \ template struct Transpose; \ - template struct Transpose; + template struct Transpose; \ + template struct Transpose; \ + template struct Transpose; DEFINE_CPU_TRANS(1); DEFINE_CPU_TRANS(2); diff --git a/paddle/fluid/operators/math/math_function.cu b/paddle/fluid/operators/math/math_function.cu index b5bf84e5178c143de35ec6dcb16b1bde5577c166..d5af718723e8d44da0971ea7756b8c36e771cca2 100644 --- a/paddle/fluid/operators/math/math_function.cu +++ b/paddle/fluid/operators/math/math_function.cu @@ -33,9 +33,10 @@ template struct SetConstant; template struct SetConstant; template struct SetConstant; -#define DEFINE_GPU_TRANS(RANK) \ - template struct Transpose; \ - template struct Transpose; +#define DEFINE_GPU_TRANS(RANK) \ + template struct Transpose; \ + template struct Transpose; \ + template struct Transpose; DEFINE_GPU_TRANS(1); DEFINE_GPU_TRANS(2); diff --git a/paddle/fluid/operators/math/matmul.h b/paddle/fluid/operators/math/matmul.h deleted file mode 100644 index 87fd38a324e007bcc939c31b6ae8e5d38c3e658c..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/math/matmul.h +++ /dev/null @@ -1,149 +0,0 @@ -/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include -#include "paddle/fluid/operators/math/blas.h" - -namespace paddle { -namespace operators { -namespace math { - -// Implements the logic of numpy matmul: -// https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.matmul.html -// -// but allowing also for a, b to be transposed -// -// Both a & b can be 1- to 3-dimensional. Higher rank tensors are not supported -// yet. -template -class MatMulFunctor { - public: - void operator()(const DeviceContext& context, const framework::Tensor& a, - bool trans_a, const framework::Tensor& b, bool trans_b, - T alpha, framework::Tensor* out, T beta) { - auto dim_a = a.dims(); - auto dim_b = b.dims(); - - PADDLE_ENFORCE(a.place() == b.place() && b.place() == out->place(), - "Tensors must all be in the same place."); - PADDLE_ENFORCE_GE(dim_a.size(), 1, - "Input tensor a must be at least 1-dimensional."); - PADDLE_ENFORCE_GE(dim_b.size(), 1, - "Input tensor b must be at least 1-dimensional."); - - std::vector out_dim; - int64_t batch_count = 1; - if (dim_a.size() > 3) { - PADDLE_ENFORCE(dim_b.size() == dim_a.size(), - "The dimensions of X and Y must be the same, and both of " - "them should be %d-dimensional.", - dim_b.size()); - // The first rank-2 dimensions are accumulated on the batch_count, and the - // last two dimensions are used for matrix multiplication. - for (int j = 0; j < dim_a.size() - 2; ++j) { - PADDLE_ENFORCE_EQ(dim_b[j], dim_a[j], - "The %d-th dimension of X and Y must be the same.", - j); - out_dim.push_back(dim_a[j]); - batch_count *= dim_a[j]; - } - } - - int M = 0, N = 0, kA = 0, kB = 0, batchCountA = 0, batchCountB = 0, - strideA = 0, strideB = 0; - - switch (dim_a.size()) { - case 1: - // similar to np.matmul: - // prepend dimension 1 (no transpose) or append dimension 1 (transpose) - M = trans_a ? dim_a[0] : 1; - kA = trans_a ? 1 : dim_a[0]; - break; - case 2: - M = trans_a ? dim_a[1] : dim_a[0]; - kA = trans_a ? dim_a[0] : dim_a[1]; - break; - case 3: - batchCountA = dim_a[0]; - M = trans_a ? dim_a[2] : dim_a[1]; - kA = trans_a ? dim_a[1] : dim_a[2]; - strideA = M * kA; - break; - default: - batchCountA = batch_count; - size_t mat_s = dim_a.size() - 2; - M = trans_a ? dim_a[mat_s + 1] : dim_a[mat_s]; - kA = trans_a ? dim_a[mat_s] : dim_a[mat_s + 1]; - strideA = M * kA; - } - - switch (dim_b.size()) { - case 1: - // similar to np.matmul: - // append dimension 1 (no transpose) or prepend dimension 1 (transpose) - kB = trans_b ? 1 : dim_b[0]; - N = trans_b ? dim_b[0] : 1; - break; - case 2: - kB = trans_b ? dim_b[1] : dim_b[0]; - N = trans_b ? dim_b[0] : dim_b[1]; - break; - case 3: - batchCountB = dim_b[0]; - kB = trans_b ? dim_b[2] : dim_b[1]; - N = trans_b ? dim_b[1] : dim_b[2]; - strideB = kB * N; - break; - default: - batchCountB = batch_count; - size_t mat_s = dim_b.size() - 2; - kB = trans_b ? dim_b[mat_s + 1] : dim_b[mat_s]; - N = trans_b ? dim_b[mat_s] : dim_b[mat_s + 1]; - strideB = kB * N; - } - - PADDLE_ENFORCE_EQ( - kA, kB, - "First matrix's width must be equal with second matrix's height."); - if (batchCountA && batchCountB) { - PADDLE_ENFORCE_EQ( - batchCountA, batchCountB, - "When input tensors a and b are both batched, they must have the " - "same batch dimension."); - } - int batchCount = std::max(batchCountA, batchCountB); - - CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; - CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; - - auto blas = GetBlas(context); - - if (!batchCount) { - // regular matrix multiplication - blas.GEMM(transA, transB, M, N, kA, alpha, a.data(), b.data(), beta, - out->data()); - } else { - // batched matrix multiplication - blas.BatchedGEMM(transA, transB, M, N, kA, alpha, a.data(), - b.data(), beta, out->data(), batchCount, strideA, - strideB); - } - } -}; - -} // namespace math -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/math/sequence2batch.h b/paddle/fluid/operators/math/sequence2batch.h index 0abda999a52bcbb94e6503692bd11aff26e849ba..62e6307ae9f4236a38c49daaf09fc05c54268159 100644 --- a/paddle/fluid/operators/math/sequence2batch.h +++ b/paddle/fluid/operators/math/sequence2batch.h @@ -64,18 +64,22 @@ class LoDTensor2BatchFunctor { bool is_reverse = false) const { if (!is_cal_batch_lod) { auto lods = batch->lod(); - PADDLE_ENFORCE_GT(lods.size(), 2UL); - PADDLE_ENFORCE_EQ(lods[1].size(), - static_cast(lod_tensor.dims()[0])); + PADDLE_ENFORCE_GT(lods.size(), 2UL, + "The LoD of LoDTensor should inlcude at least 2-level " + "sequence information."); + PADDLE_ENFORCE_EQ( + lods[1].size(), static_cast(lod_tensor.dims()[0]), + "The LoD information should be consistent with the dims."); CopyMatrixRowsFunctor to_batch; to_batch(context, lod_tensor, lods[1], batch, true); return; } auto lods = lod_tensor.lod(); - auto lod = lods[0]; PADDLE_ENFORCE_EQ(lods.size(), 1UL, "Only support one level sequence now."); + auto lod = lods[0]; + std::vector seq_info; for (size_t seq_id = 0; seq_id < lod.size() - 1; ++seq_id) { int length = lod[seq_id + 1] - lod[seq_id]; @@ -157,9 +161,12 @@ class Batch2LoDTensorFunctor { const framework::LoDTensor& batch, framework::LoDTensor* lod_tensor) const { auto in_lod = batch.lod(); - PADDLE_ENFORCE_GT(in_lod.size(), 2UL); - PADDLE_ENFORCE_EQ(in_lod[1].size(), - static_cast(lod_tensor->dims()[0])); + PADDLE_ENFORCE_GT(in_lod.size(), 2UL, + "The LoD of LoDTensor should inlcude at least 2-level " + "sequence information."); + PADDLE_ENFORCE_EQ( + in_lod[1].size(), static_cast(lod_tensor->dims()[0]), + "The LoD information should be consistent with the dims."); CopyMatrixRowsFunctor to_seq; to_seq(context, batch, in_lod[1], lod_tensor, false); } diff --git a/paddle/fluid/operators/matmul_op.cc b/paddle/fluid/operators/matmul_op.cc index e5d33fbc36438f97ff5b604e4efdbfbfa91fcee4..7182149164854038bb67a9f06cdbec8a4a0f1fb2 100644 --- a/paddle/fluid/operators/matmul_op.cc +++ b/paddle/fluid/operators/matmul_op.cc @@ -12,21 +12,264 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/matmul_op.h" #include +#include #include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detail/safe_ref.h" +#include "paddle/fluid/operators/math/blas.h" namespace paddle { namespace operators { +/** + * Get row matrix shape from a vector shape. If the rank of x_dim > 1, the + * original x_dim is returned. + */ +static framework::DDim RowMatrixFromVector(const framework::DDim &x_dim) { + if (x_dim.size() > 1) { + return x_dim; + } + return framework::make_ddim({1, x_dim[0]}); +} + +/** + * Get column matrix shape from a vector shape. If the ran of y_dim > 1, the + * original y_dim is returned. + */ +static framework::DDim ColumnMatrixFromVector(const framework::DDim &y_dim) { + if (y_dim.size() > 1) { + return y_dim; + } + return framework::make_ddim({y_dim[0], 1}); +} + +template +class MatMulKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + auto &x = + detail::Ref(context.Input("X"), "Cannot find X"); + auto &y = + detail::Ref(context.Input("Y"), "Cannot find Y"); + auto *out = context.Output("Out"); + out->mutable_data(context.GetPlace()); + + auto blas = math::GetBlas(context); + auto mat_dim_a = math::CreateMatrixDescriptor( + RowMatrixFromVector(x.dims()), 0, context.Attr("transpose_X")); + auto mat_dim_b = math::CreateMatrixDescriptor( + ColumnMatrixFromVector(y.dims()), 0, context.Attr("transpose_Y")); + blas.MatMul(x, mat_dim_a, y, mat_dim_b, T(1), out, T(0)); + } +}; + +// Reshape a rank-3 tensor from P x M x N to (P * M) x N. +// Identity op if the tensor is not of rank 3. +static framework::Tensor FoldInitDims(const framework::Tensor &input) { + auto output = input; + auto in_dims = input.dims(); + if (in_dims.size() == 3) { + output.Resize({in_dims[0] * in_dims[1], in_dims[2]}); + } + return output; +} + +// Reshape a rank-3 tensor from P x M x N to M x (P * N). +// (Warning: This requires transposing data and writes into new memory.) +// Identity op if the tensor is not of rank 3. +template +static framework::Tensor FoldHeadAndLastDims(const DeviceContext &context, + const framework::Tensor &input) { + auto in_dims = input.dims(); + if (in_dims.size() != 3) { + return input; + } + framework::Tensor output; + output.Resize({in_dims[1], in_dims[0], in_dims[2]}); + output.mutable_data(context.GetPlace()); + std::vector axis = {1, 0, 2}; + math::Transpose trans; + trans(context, input, &output, axis); + output.Resize({in_dims[1], in_dims[0] * in_dims[2]}); + + return output; +} + +/** + * Reshape a tensor to 3-D or 2-D tensor by matrix descriptor. + * + * The shape would be [BatchSize, H, W] or [H, W]. + * If transposed, `H,W` will be swapped. + */ +static void ReshapeTensorIntoMatrixSequence( + framework::Tensor *x, const math::MatDescriptor &descriptor) { + int64_t h, w; + h = descriptor.height_; + w = descriptor.width_; + if (descriptor.trans_) { + std::swap(w, h); + } + if (descriptor.batch_size_) { + x->Resize({descriptor.batch_size_, h, w}); + } else { + x->Resize({h, w}); + } +} + +/** + * Reshape the x,y,out tensor to 3-D or 2-D tensor by matrix descriptor + * Out = matmul(x, y) + * + * This method will first calculate X,Y matrix sequence, and then calculate + * the out shape. + * + * Assume X = [BatchSize, H1, W1], Y = [BatchSize, H2, W2] + * The out = [BatchSize, H1, W2] + * + * If there is no batch size in `X` and `Y`, the out will be [H1, W2] + * If any of `X` and `Y` has batch size BatchSize, the out will have the + * BatchSize. + */ +static void ReshapeXYOutIntoMatrixSequence(framework::Tensor *x, + framework::Tensor *y, + framework::Tensor *out, bool trans_x, + bool trans_y) { + auto x_dim = RowMatrixFromVector(x->dims()); + auto y_dim = ColumnMatrixFromVector(y->dims()); + auto mat_dim_x = math::CreateMatrixDescriptor(x_dim, 0, trans_x); + auto mat_dim_y = math::CreateMatrixDescriptor(y_dim, 0, trans_y); + if (mat_dim_x.batch_size_ == 0 && mat_dim_y.batch_size_ == 0) { + out->Resize({mat_dim_x.height_, mat_dim_y.width_}); + } else { + out->Resize({std::max(mat_dim_x.batch_size_, mat_dim_y.batch_size_), + mat_dim_x.height_, mat_dim_y.width_}); + } + + ReshapeTensorIntoMatrixSequence(x, mat_dim_x); + ReshapeTensorIntoMatrixSequence(y, mat_dim_y); +} + +// Using dimensional constraints on matrix multiplication, it is +// straight-forward to check the following table for when X and Y +// are both matrices. +// +// transpose_X | False | True | False | True +// transpose_Y | False | False | True | True +// -----------+----------+----------+----------+----------- +// dX = | dOut Y^T | Y dOut^T | dOut Y | Y^T dOut^T +// dY = | X^T dOut | X dOut | dOut^T X | dOut^T X^T +// +// When X is a vector of size K, we treat it instead as a matrix of shape +// (1, K). Similarly, when Y is a vector of size K, we treat it instead as +// a matrix of shape (K, 1). +// +// When X and Y are both 3-dimensional tensors, then the first dimension +// the batch dimension can be ignored and the exact same formulas apply +// as for two matrices. +// +// Finally, when, e.g., X is a 3-dimensional tensor but Y is a matrix, we end +// up with formulas like +// +// dY_{ij} = \sum_{p, m} X_{pmi} dOut_{pmj} +// +// To handle this sort of scenario, we reshape X : P x M x K, dOut: P x M x N +// to X: (P * M) x K, dOut: (P * M) x N. +template +class MatMulGradKernel : public framework::OpKernel { + public: + void MatMul(const framework::ExecutionContext &context, + const framework::Tensor &a, bool trans_a, + const framework::Tensor &b, bool trans_b, + framework::Tensor *out) const { + out->mutable_data(context.GetPlace()); + auto blas = math::GetBlas(context); + auto mat_dim_a = math::CreateMatrixDescriptor(a.dims(), 0, trans_a); + auto mat_dim_b = math::CreateMatrixDescriptor(b.dims(), 0, trans_b); + blas.MatMul(a, mat_dim_a, b, mat_dim_b, T(1), out, T(0)); + } + + void CalcInputGrad(const framework::ExecutionContext &context, + const framework::Tensor &a, bool trans_a, + bool is_fold_init_dims_a, const framework::Tensor &b, + bool trans_b, bool is_fold_init_dims_b, + framework::Tensor *out) const { + if (out == nullptr) return; + bool need_combine = (a.dims().size() == 3 || b.dims().size() == 3) && + out->dims().size() == 2; + if (!need_combine) { + MatMul(context, a, trans_a, b, trans_b, out); + } else { + auto &ctx = context.template device_context(); + MatMul(context, is_fold_init_dims_a + ? FoldInitDims(a) + : FoldHeadAndLastDims(ctx, a), + trans_a, is_fold_init_dims_b + ? FoldInitDims(b) + : FoldHeadAndLastDims(ctx, b), + trans_b, out); + } + } + + void Compute(const framework::ExecutionContext &context) const override { + auto x = *context.Input("X"); + auto y = *context.Input("Y"); + auto dout = + *context.Input(framework::GradVarName("Out")); + auto *dx = context.Output(framework::GradVarName("X")); + auto *dy = context.Output(framework::GradVarName("Y")); + bool transpose_x = context.Attr("transpose_X"); + bool transpose_y = context.Attr("transpose_Y"); + + ReshapeXYOutIntoMatrixSequence(&x, &y, &dout, transpose_x, transpose_y); + framework::DDim dx_dims; + if (dx) { + dx_dims = dx->dims(); + if (dx_dims != x.dims()) { + dx->Resize(x.dims()); + } + } + + framework::DDim dy_dims; + if (dy) { + dy_dims = dy->dims(); + if (dy_dims != y.dims()) { + dy->Resize(y.dims()); + } + } -using framework::Tensor; + if (transpose_x && transpose_y) { + CalcInputGrad(context, y, true, true, dout, true, false, dx); + CalcInputGrad(context, dout, true, true, x, true, false, dy); + } else if (transpose_x) { + CalcInputGrad(context, y, false, false, dout, true, false, dx); + CalcInputGrad(context, x, false, false, dout, false, true, dy); + } else if (transpose_y) { + CalcInputGrad(context, dout, false, false, y, false, true, dx); + CalcInputGrad(context, dout, true, true, x, false, true, dy); + } else { + CalcInputGrad(context, dout, false, false, y, true, false, dx); + CalcInputGrad(context, x, true, true, dout, false, true, dy); + } + + if (dx) { + if (dx_dims != x.dims()) { + dx->Resize(dx_dims); + } + } + if (dy) { + if (dy_dims != y.dims()) { + dy->Resize(dy_dims); + } + } + } +}; class MatMulOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(framework::InferShapeContext* context) const override { + void InferShape(framework::InferShapeContext *context) const override { PADDLE_ENFORCE(context->HasInput("X"), "Input(X) of MatMulOp should not be null."); PADDLE_ENFORCE(context->HasInput("Y"), @@ -36,121 +279,41 @@ class MatMulOp : public framework::OperatorWithKernel { auto dim_x = context->GetInputDim("X"); auto dim_y = context->GetInputDim("Y"); - bool transpose_x = context->Attrs().Get("transpose_X"); - bool transpose_y = context->Attrs().Get("transpose_Y"); - - PADDLE_ENFORCE_GE(dim_x.size(), 1, - "Input tensor X must be at least 1-dimensional."); - PADDLE_ENFORCE_GE(dim_y.size(), 1, - "Input tensor Y must be at least 1-dimensional."); - - std::vector out_dim; - int64_t batch_count = 1; - if (dim_x.size() > 3) { - PADDLE_ENFORCE_EQ( - dim_y.size(), dim_x.size(), - "The dimensions of X and Y must be the same, and both of " - "them should be %d-dimensional.", - dim_x.size()); - - // The first rank-2 dimensions are accumulated on the batch_count, and the - // last two dimensions are used for matrix multiplication. - for (int j = 0; j < dim_x.size() - 2; ++j) { - PADDLE_ENFORCE_EQ(dim_y[j], dim_x[j], - "The %d-th dimension of X and Y must be the same.", - j); - out_dim.push_back(dim_x[j]); - batch_count *= dim_x[j]; - } - } - int M = 0, N = 0, KX = 0, KY = 0, batchCountX = 0, batchCountY = 0; - bool remove_initial_dim = false, remove_final_dim = false; - - switch (dim_x.size()) { - case 1: - if (transpose_x) { - M = dim_x[0]; - KX = 1; - } else { - M = 1; - KX = dim_x[0]; - remove_initial_dim = true; - } - break; - case 2: - M = transpose_x ? dim_x[1] : dim_x[0]; - KX = transpose_x ? dim_x[0] : dim_x[1]; - break; - case 3: - batchCountX = dim_x[0]; - M = transpose_x ? dim_x[2] : dim_x[1]; - KX = transpose_x ? dim_x[1] : dim_x[2]; - break; - default: - batchCountX = batch_count; - size_t mat_s = dim_x.size() - 2; - M = transpose_x ? dim_x[mat_s + 1] : dim_x[mat_s]; - KX = transpose_x ? dim_x[mat_s] : dim_x[mat_s + 1]; - break; - } + auto mat_dim_x = + math::CreateMatrixDescriptor(RowMatrixFromVector(dim_x), 0, + context->Attrs().Get("transpose_X")); + auto mat_dim_y = + math::CreateMatrixDescriptor(ColumnMatrixFromVector(dim_y), 0, + context->Attrs().Get("transpose_Y")); - switch (dim_y.size()) { - case 1: - if (transpose_y) { - N = dim_y[0]; - KY = 1; - } else { - N = 1; - KY = dim_y[0]; - remove_final_dim = true; - } - break; - case 2: - KY = transpose_y ? dim_y[1] : dim_y[0]; - N = transpose_y ? dim_y[0] : dim_y[1]; - break; - case 3: - batchCountY = dim_y[0]; - KY = transpose_y ? dim_y[2] : dim_y[1]; - N = transpose_y ? dim_y[1] : dim_y[2]; - break; - default: - batchCountY = batch_count; - size_t mat_s = dim_y.size() - 2; - KY = transpose_y ? dim_y[mat_s + 1] : dim_y[mat_s]; - N = transpose_y ? dim_y[mat_s] : dim_y[mat_s + 1]; + PADDLE_ENFORCE_EQ(mat_dim_x.width_, mat_dim_y.height_); + PADDLE_ENFORCE(mat_dim_x.batch_size_ == mat_dim_y.batch_size_ || + mat_dim_x.batch_size_ == 0 || mat_dim_y.batch_size_ == 0); + std::vector dim_out; + if (mat_dim_x.batch_size_ != 0) { + dim_out = framework::vectorize(dim_x); + dim_out[dim_out.size() - 2] = mat_dim_x.height_; + dim_out[dim_out.size() - 1] = mat_dim_y.width_; + } else if (mat_dim_y.batch_size_ != 0) { + dim_out = framework::vectorize(dim_y); + dim_out[dim_out.size() - 2] = mat_dim_x.height_; + dim_out[dim_out.size() - 1] = mat_dim_y.width_; + } else { + dim_out = {mat_dim_x.height_, mat_dim_y.width_}; } - PADDLE_ENFORCE_EQ( - KX, KY, - "First matrix's width must be equal with second matrix's height."); - if (batchCountX && batchCountY) { - PADDLE_ENFORCE_EQ( - batchCountX, batchCountY, - "When Input(X) and Input(Y) are both three dimensional, they " - "must have the same batch dimension."); + if (dim_x.size() == 1 && dim_out[dim_out.size() - 2] == 1) { + std::swap(dim_out[dim_out.size() - 2], dim_out[dim_out.size() - 1]); + dim_out.resize(dim_out.size() - 1); } - int batchCount = std::max(batchCountX, batchCountY); - std::vector dim_out; - if (batchCount) { - if (dim_x.size() > 3) { - dim_out.insert(dim_out.begin(), out_dim.begin(), out_dim.end()); - } else { - dim_out.push_back(batchCount); - } + if (dim_y.size() == 1 && dim_out[dim_out.size() - 1] == 1) { + dim_out.resize(dim_out.size() - 1); } - if (!remove_initial_dim) { - dim_out.push_back(M); - } - if (!remove_final_dim) { - dim_out.push_back(N); - } - if (dim_out.size() == 0) { - // We don't support 0-dimensional Tensors (scalars), so instead - // treat the output as a Tensor of shape (1, ) in this case. - dim_out.push_back(1); + + if (dim_out.empty()) { + dim_out = {1}; } context->SetOutputDim("Out", framework::make_ddim(dim_out)); context->ShareLoD("X", /*->*/ "Out"); @@ -159,8 +322,7 @@ class MatMulOp : public framework::OperatorWithKernel { class MatMulOpMaker : public framework::OpProtoAndCheckerMaker { public: - MatMulOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The first input of MatMul op"); AddInput("Y", "The second input of MatMul op"); AddOutput("Out", "The output of MatMul op"); @@ -213,7 +375,7 @@ class MatMulOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; protected: - void InferShape(framework::InferShapeContext* context) const override { + void InferShape(framework::InferShapeContext *context) const override { PADDLE_ENFORCE(context->HasInput("X"), "Input(X) should not be null"); PADDLE_ENFORCE(context->HasInput("Y"), "Input(Y) should not be null"); PADDLE_ENFORCE(context->HasInput(framework::GradVarName("Out")), @@ -233,15 +395,52 @@ class MatMulOpGrad : public framework::OperatorWithKernel { } }; +class MatMulOpGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto *retv = new framework::OpDesc(); + retv->SetType("matmul_grad"); + retv->SetInput("X", Input("X")); + retv->SetInput("Y", Input("Y")); + retv->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); + retv->SetOutput(framework::GradVarName("X"), InputGrad("X")); + retv->SetOutput(framework::GradVarName("Y"), InputGrad("Y")); + retv->SetAttrMap(Attrs()); + return std::unique_ptr(retv); + } +}; } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OPERATOR(matmul, ops::MatMulOp, ops::MatMulOpMaker, - paddle::framework::DefaultGradOpDescMaker); + ops::MatMulOpGradMaker); REGISTER_OPERATOR(matmul_grad, ops::MatMulOpGrad); REGISTER_OP_CPU_KERNEL( - matmul, ops::MatMulKernel); + matmul, ops::MatMulKernel, + ops::MatMulKernel, + ops::MatMulKernel); REGISTER_OP_CPU_KERNEL( matmul_grad, - ops::MatMulGradKernel); + ops::MatMulGradKernel, + ops::MatMulGradKernel, + ops::MatMulGradKernel); + +#ifdef PADDLE_WITH_CUDA +REGISTER_OP_CUDA_KERNEL( + matmul, ops::MatMulKernel, + ops::MatMulKernel, + ops::MatMulKernel); +REGISTER_OP_CUDA_KERNEL( + matmul_grad, + ops::MatMulGradKernel, + ops::MatMulGradKernel, + ops::MatMulGradKernel); +#endif diff --git a/paddle/fluid/operators/matmul_op.h b/paddle/fluid/operators/matmul_op.h deleted file mode 100644 index f2e9cfdcdbf93326ae193776a7d5f6a324373603..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/matmul_op.h +++ /dev/null @@ -1,244 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include -#include -#include -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/math/math_function.h" -#include "paddle/fluid/operators/math/matmul.h" - -namespace paddle { -namespace operators { -namespace matmul_detail { - -using Tensor = framework::Tensor; -using DDim = framework::DDim; -using framework::make_ddim; -using framework::vectorize; - -template -class MatMulKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - const Tensor& x = *context.Input("X"); - const Tensor& y = *context.Input("Y"); - Tensor* out = context.Output("Out"); - out->mutable_data(context.GetPlace()); - bool transpose_x = context.Attr("transpose_X"); - bool transpose_y = context.Attr("transpose_Y"); - - math::MatMulFunctor()( - context.template device_context(), x, transpose_x, y, - transpose_y, T(1), out, T(0)); - } -}; - -template -inline Tensor Reshape(const Tensor& input, const DDim& dims) { - Tensor output; - output.ShareDataWith(input); - output.Resize(dims); - return output; -} - -// Reshape a rank-3 tensor from P x M x N to (P * M) x N. -// Identity op if the tensor is not of rank 3. -template -Tensor CombineBatchAndM(const Tensor& input) { - Tensor output; - output.ShareDataWith(input); - auto in_dims = input.dims(); - if (in_dims.size() == 3) { - std::vector out_dims = {in_dims[0] * in_dims[1], in_dims[2]}; - output.Resize(make_ddim(out_dims)); - } - return output; -} - -// Reshape a rank-3 tensor from P x M x N to M x (P * N). -// (Warning: This requires transposing data and writes into new memory.) -// Identity op if the tensor is not of rank 3. -template -Tensor CombineBatchAndN(const DeviceContext& context, const Tensor& input) { - Tensor output; - auto in_dims = input.dims(); - if (in_dims.size() == 3) { - output.Resize({in_dims[1], in_dims[0], in_dims[2]}); - output.mutable_data(context.GetPlace()); - std::vector axis = {1, 0, 2}; - math::Transpose trans; - trans(context, input, &output, axis); - std::vector out_dims = {in_dims[1], in_dims[0] * in_dims[2]}; - output.Resize({in_dims[1], in_dims[0] * in_dims[2]}); - } else { - output.ShareDataWith(input); - } - return output; -} - -// Using dimensional constraints on matrix multiplication, it is -// straight-forward to check the following table for when X and Y -// are both matrices. -// -// transpose_X | False | True | False | True -// transpose_Y | False | False | True | True -// -----------+----------+----------+----------+----------- -// dX = | dOut Y^T | Y dOut^T | dOut Y | Y^T dOut^T -// dY = | X^T dOut | X dOut | dOut^T X | dOut^T X^T -// -// When X is a vector of size K, we treat it instead as a matrix of shape -// (1, K). Similarly, when Y is a vector of size K, we treat it instead as -// a matrix of shape (K, 1). -// -// When X and Y are both 3-dimensional tensors, then the first dimension -// the batch dimension can be ignored and the exact same formulas apply -// as for two matrices. -// -// Finally, when, e.g., X is a 3-dimensional tensor but Y is a matrix, we end -// up with formulas like -// -// dY_{ij} = \sum_{p, m} X_{pmi} dOut_{pmj} -// -// To handle this sort of scenario, we reshape X : P x M x K, dOut: P x M x N -// to X: (P * M) x K, dOut: (P * M) x N. -template -class MatMulGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - const Tensor& x = *context.Input("X"); - const Tensor& y = *context.Input("Y"); - const Tensor& dout = *context.Input(framework::GradVarName("Out")); - Tensor* dx = context.Output(framework::GradVarName("X")); - Tensor* dy = context.Output(framework::GradVarName("Y")); - bool transpose_x = context.Attr("transpose_X"); - bool transpose_y = context.Attr("transpose_Y"); - - std::vector x_dims = vectorize(x.dims()); - std::vector y_dims = vectorize(y.dims()); - - // If X is a vector, reshape it to a matrix. - if (x_dims.size() == 1) { - x_dims.insert(x_dims.begin(), 1); - } - - // If Y is a vector, reshape it to a matrix. - if (y_dims.size() == 1) { - y_dims.push_back(1); - } - - int batch_count = 0; - // The first rank-2 dimensions are accumulated on the batch_count, and the - // last two dimensions are used for matrix multiplication. - if (x_dims.size() > 3) { - batch_count = accumulate(x_dims.begin(), x_dims.end() - 2, 1, - std::multiplies()); - } - // Fix the dOut dimensions. - int M = 0, N = 0, batchCountX = 0, batchCountY = 0; - - switch (x_dims.size()) { - case 2: - M = transpose_x ? x_dims[1] : x_dims[0]; - break; - case 3: - batchCountX = x_dims[0]; - M = transpose_x ? x_dims[2] : x_dims[1]; - break; - default: - batchCountX = batch_count; - size_t mat_s = x_dims.size() - 2; - M = transpose_x ? x_dims[mat_s + 1] : x_dims[mat_s]; - } - - switch (y_dims.size()) { - case 2: - N = transpose_y ? y_dims[0] : y_dims[1]; - break; - case 3: - batchCountY = y_dims[0]; - N = transpose_y ? y_dims[1] : y_dims[2]; - break; - default: - batchCountY = batch_count; - size_t mat_s = y_dims.size() - 2; - N = transpose_y ? y_dims[mat_s] : y_dims[mat_s + 1]; - } - if (batchCountX && batchCountY) { - PADDLE_ENFORCE_EQ( - batchCountX, batchCountY, - "When Input(X) and Input(Y) are both three dimensional, they " - "must have the same batch dimension."); - } - int batchCount = std::max(batchCountX, batchCountY); - std::vector dout_dims = {M, N}; - if (batchCount) { - if (x_dims.size() > 3) { - dout_dims.insert(dout_dims.begin(), x_dims.begin(), x_dims.end() - 2); - } else { - dout_dims.insert(dout_dims.begin(), batchCount); - } - } - Tensor X = Reshape(x, make_ddim(x_dims)); - Tensor Y = Reshape(y, make_ddim(y_dims)); - Tensor dOut = Reshape(dout, make_ddim(dout_dims)); - - auto& dev_ctx = context.template device_context(); - if (dx) { - dx->mutable_data(context.GetPlace()); - const Tensor& dOut_for_dX = - (x_dims.size() == 2 && y_dims.size() == 3) - ? CombineBatchAndN(dev_ctx, dOut) - : dOut; - if (x_dims.size() == 2 && y_dims.size() == 3) { - Y = transpose_y ? CombineBatchAndM(Y) - : CombineBatchAndN(dev_ctx, Y); - } - if (transpose_x) { - math::MatMulFunctor()( - dev_ctx, Y, transpose_y, dOut_for_dX, transpose_x, T(1), dx, T(0)); - } else { - math::MatMulFunctor()( - dev_ctx, dOut_for_dX, transpose_x, Y, !transpose_y, T(1), dx, T(0)); - } - } - - if (dy) { - dy->mutable_data(context.GetPlace()); - const Tensor& dOut_for_dY = (y_dims.size() == 2 && x_dims.size() == 3) - ? CombineBatchAndM(dOut) - : dOut; - if (y_dims.size() == 2 && x_dims.size() == 3) { - X = transpose_x ? CombineBatchAndN(dev_ctx, X) - : CombineBatchAndM(X); - dOut = CombineBatchAndM(dOut); - } - if (transpose_y) { - math::MatMulFunctor()( - dev_ctx, dOut_for_dY, transpose_y, X, transpose_x, T(1), dy, T(0)); - } else { - math::MatMulFunctor()( - dev_ctx, X, !transpose_x, dOut_for_dY, transpose_y, T(1), dy, T(0)); - } - } - } -}; -} // namespace matmul_detail - -using matmul_detail::MatMulKernel; -using matmul_detail::MatMulGradKernel; - -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/max_sequence_len_op.cc b/paddle/fluid/operators/max_sequence_len_op.cc index 4cd7c89b48a2442ee7a5074abbf0f3dd9ea3bcb4..8e508b68eeab69a4595904dcc3ea0a541d9ab6e6 100644 --- a/paddle/fluid/operators/max_sequence_len_op.cc +++ b/paddle/fluid/operators/max_sequence_len_op.cc @@ -41,8 +41,7 @@ class MaxSeqenceLenOp : public framework::OperatorBase { class MaxSeqenceLenOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - MaxSeqenceLenOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("RankTable", "The lod_rank_table."); AddOutput("Out", "The max sequence length."); AddComment( diff --git a/paddle/fluid/operators/maxout_op.cc b/paddle/fluid/operators/maxout_op.cc index e2bcba5a5e15d4d5f10ae4ae64b5262f750137ab..058115cb624627d81b31d0903f7d615d19708c77 100644 --- a/paddle/fluid/operators/maxout_op.cc +++ b/paddle/fluid/operators/maxout_op.cc @@ -22,8 +22,7 @@ using framework::Tensor; class MaxOutOpMaker : public framework::OpProtoAndCheckerMaker { public: - MaxOutOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(Tensor) The input tensor of maxout operator. " diff --git a/paddle/fluid/operators/mean_op.cc b/paddle/fluid/operators/mean_op.cc index a134796bfcaa9dea2483ace9f5045e257916daba..74477eb439dc202c3f5f17fdf3e1647bc5c23512 100644 --- a/paddle/fluid/operators/mean_op.cc +++ b/paddle/fluid/operators/mean_op.cc @@ -32,8 +32,7 @@ class MeanOp : public framework::OperatorWithKernel { class MeanOpMaker : public framework::OpProtoAndCheckerMaker { public: - MeanOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of mean op"); AddOutput("Out", "The output of mean op"); AddComment(R"DOC( diff --git a/paddle/fluid/operators/merge_lod_tensor_op.cc b/paddle/fluid/operators/merge_lod_tensor_op.cc index 4ebf20cbba69bee09dfddb8e928ddc95665e4731..a16861b3b77fc980ab932b9d88859b38ec36108b 100644 --- a/paddle/fluid/operators/merge_lod_tensor_op.cc +++ b/paddle/fluid/operators/merge_lod_tensor_op.cc @@ -121,8 +121,7 @@ class MergeLoDTensorOp : public framework::OperatorBase { class MergeLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - MergeLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input LoDTensor, contains complete lod information to " "construct the output"); diff --git a/paddle/fluid/operators/minus_op.cc b/paddle/fluid/operators/minus_op.cc index a302b24560e680076d62d02b422c6410467deb1d..34571a38a14795a98ac8454cec606077727b5ffa 100644 --- a/paddle/fluid/operators/minus_op.cc +++ b/paddle/fluid/operators/minus_op.cc @@ -48,8 +48,7 @@ class MinusOp : public framework::OperatorWithKernel { class MinusOpMaker : public framework::OpProtoAndCheckerMaker { public: - MinusOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The left tensor of minus operator."); AddInput("Y", "The right tensor of minus operator."); AddOutput("Out", "The output tensor of minus operator."); diff --git a/paddle/fluid/operators/mkldnn_activation_op.h b/paddle/fluid/operators/mkldnn_activation_op.h index f26a165b5a59f01f864d62bbf798f4cbffa65371..85664623d7330e9473286d995bec67879510dbd7 100644 --- a/paddle/fluid/operators/mkldnn_activation_op.h +++ b/paddle/fluid/operators/mkldnn_activation_op.h @@ -13,6 +13,8 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include + #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/safe_ref.h" @@ -60,52 +62,5 @@ class MKLDNNActivationGradKernel } }; -namespace { // NOLINT -framework::OpKernelType GetKernelType( - const framework::ExecutionContext& ctx, - const framework::OperatorWithKernel& oper) { - framework::LibraryType library{framework::LibraryType::kPlain}; -#ifdef PADDLE_WITH_MKLDNN - if (library == framework::LibraryType::kPlain && - platform::CanMKLDNNBeUsed(ctx)) { - library = framework::LibraryType::kMKLDNN; - } -#endif - framework::DataLayout layout = framework::DataLayout::kAnyLayout; - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), - ctx.GetPlace(), layout, library); -} -} // anonymous namespace - -class ActivationWithMKLDNNOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - void InferShape(framework::InferShapeContext* ctx) const override { - ctx->SetOutputDim("Out", ctx->GetInputDim("X")); - ctx->ShareLoD("X", /*->*/ "Out"); - } - - framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { - return GetKernelType(ctx, *this); - } -}; - -class ActivationWithMKLDNNOpGrad : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - void InferShape(framework::InferShapeContext* ctx) const override { - ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out")); - } - - framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { - return GetKernelType(ctx, *this); - } -}; - } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/modified_huber_loss_op.cc b/paddle/fluid/operators/modified_huber_loss_op.cc index 3a0fc74584391d0441105a8ac7d7ac292e10fb8d..35db4c1ad1f6c6481eca397e99fc8c1f0bc7164c 100644 --- a/paddle/fluid/operators/modified_huber_loss_op.cc +++ b/paddle/fluid/operators/modified_huber_loss_op.cc @@ -39,8 +39,7 @@ class ModifiedHuberLossOp : public framework::OperatorWithKernel { class ModifiedHuberLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - ModifiedHuberLossOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input tensor of modified huber loss op. " "X is 2-D tensor with shape [batch_size, 1]."); diff --git a/paddle/fluid/operators/momentum_op.cc b/paddle/fluid/operators/momentum_op.cc index f13ec53905aa3d5b55b865c3514f36211c06a549..dcd73e3c3e40f80e07b73944d1f0cc57fea010d3 100644 --- a/paddle/fluid/operators/momentum_op.cc +++ b/paddle/fluid/operators/momentum_op.cc @@ -62,8 +62,7 @@ class MomentumOp : public framework::OperatorWithKernel { class MomentumOpMaker : public framework::OpProtoAndCheckerMaker { public: - MomentumOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor, default Tensor) " "Input parameter that has to be updated"); diff --git a/paddle/fluid/operators/mul_op.cc b/paddle/fluid/operators/mul_op.cc index 6903cf83b41a54b54382fac2cf58f7bfe192b55f..a43739463c85b38e1dba04c6ec1bfcf4b6cbfa63 100644 --- a/paddle/fluid/operators/mul_op.cc +++ b/paddle/fluid/operators/mul_op.cc @@ -96,8 +96,7 @@ class MulOp : public framework::OperatorWithKernel { class MulOpMaker : public framework::OpProtoAndCheckerMaker { public: - MulOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor), The first input tensor of mul op."); AddInput("Y", "(Tensor), The second input tensor of mul op."); AddOutput("Out", "(Tensor), The output tensor of mul op."); diff --git a/paddle/fluid/operators/multiplex_op.cc b/paddle/fluid/operators/multiplex_op.cc index b698c1bf8a05e053db07db34712a13c8074ee4d0..a4363fd25d57edb5c2509904a1f55634832613be 100644 --- a/paddle/fluid/operators/multiplex_op.cc +++ b/paddle/fluid/operators/multiplex_op.cc @@ -61,8 +61,7 @@ class MultiplexOp : public framework::OperatorWithKernel { class MultiplexOpMaker : public framework::OpProtoAndCheckerMaker { public: - MultiplexOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Ids", "The index tensor of multiplex operator."); AddInput("X", "The candidate tensors of multiplex operator.") .AsDuplicable(); diff --git a/paddle/fluid/operators/nccl_op.cc b/paddle/fluid/operators/nccl_op.cc index 5e4ed886b10bd48bf991ce84a9099611cf5d1d26..0018139cb06fe0573565c920849843e674df6f4c 100644 --- a/paddle/fluid/operators/nccl_op.cc +++ b/paddle/fluid/operators/nccl_op.cc @@ -76,8 +76,7 @@ class NCCLInitOpShapeInference : public framework::InferShapeBase { class NCCLInitOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCCLInitOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kParallelScopes, "The working place of parallel do."); AddOutput("Communicator", "Create Communicator for communicating between gpus"); @@ -118,8 +117,7 @@ class NCCLAllReduceOp : public framework::OperatorWithKernel { // AllReduceOp class NCCLAllReduceOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCCLAllReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of AllReduce op"); AddInput("Communicator", "Communicator for communicating between gpus"); AddOutput("Out", "The output of AllReduce op"); @@ -165,8 +163,7 @@ class NCCLReduceOp : public framework::OperatorWithKernel { // ReduceOp class NCCLReduceOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCCLReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of Reduce op"); AddInput("Communicator", "Communicator for communicating between gpus"); AddOutput("Out", "The output of Reduce op"); @@ -214,8 +211,7 @@ class NCCLBcastOp : public framework::OperatorWithKernel { // BcastOp class NCCLBcastOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCCLBcastOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of BcastSend op"); AddInput("Communicator", "Communicator for communicating between gpus"); AddOutput("Out", "The output of Bcast"); diff --git a/paddle/fluid/operators/nce_op.cc b/paddle/fluid/operators/nce_op.cc index 192bdf8ea553f3a82066f8562458d286ee15a6ee..06092e680a1efbef379ccf40fdf476769f820429 100644 --- a/paddle/fluid/operators/nce_op.cc +++ b/paddle/fluid/operators/nce_op.cc @@ -75,8 +75,7 @@ class NCEOp : public framework::OperatorWithKernel { class NCEOpMaker : public framework::OpProtoAndCheckerMaker { public: - NCEOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Input", "(Tensor) A tensor of shape [batch_size, dim]."); AddInput( "Label", diff --git a/paddle/fluid/operators/norm_op.cc b/paddle/fluid/operators/norm_op.cc index 30a991224fa184257a8e59af5e6a27a0b0a4da86..cdbc975c02214721ceae3a338741101ef32d7ee9 100644 --- a/paddle/fluid/operators/norm_op.cc +++ b/paddle/fluid/operators/norm_op.cc @@ -19,8 +19,7 @@ namespace operators { template class NormOpMaker : public framework::OpProtoAndCheckerMaker { public: - NormOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(Tensor) The input tensor of norm operator. " diff --git a/paddle/fluid/operators/one_hot_op.cc b/paddle/fluid/operators/one_hot_op.cc index 1d42dfdd765166c9596abc08ce8abd534453bc63..4fcb1d69935175c3f643db7a4da04db34492f8fb 100644 --- a/paddle/fluid/operators/one_hot_op.cc +++ b/paddle/fluid/operators/one_hot_op.cc @@ -46,8 +46,7 @@ class OneHotOp : public framework::OperatorWithKernel { class OneHotOpMaker : public framework::OpProtoAndCheckerMaker { public: - OneHotOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor, LoDTensor) Input variable with rank at least 2. " "The last dimension of X should be 1. Each value of X is an index " diff --git a/paddle/fluid/operators/pad_op.cc b/paddle/fluid/operators/pad_op.cc index d2a0106f80144e3550d73ea22f8e012426eb01ae..d4b631a6f5bf9332f4ed1d1a4bda529fbb6ada0a 100644 --- a/paddle/fluid/operators/pad_op.cc +++ b/paddle/fluid/operators/pad_op.cc @@ -48,8 +48,7 @@ class PadOp : public framework::OperatorWithKernel { class PadOpMaker : public framework::OpProtoAndCheckerMaker { public: - PadOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input of pad op. " "The input should be a k-D tensor(k > 0 and k < 7)"); diff --git a/paddle/fluid/operators/parallel_do_op.cc b/paddle/fluid/operators/parallel_do_op.cc index ae34fe2184b43cc104c14672dec30efd3b0e9f3b..1012640d5e2052e4f347ad458cea9072a004f334 100644 --- a/paddle/fluid/operators/parallel_do_op.cc +++ b/paddle/fluid/operators/parallel_do_op.cc @@ -196,8 +196,7 @@ class ParallelDoOp : public framework::OperatorBase { class ParallelDoOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ParallelDoOpProtoMaker(OpProto *proto, framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kInputs, "").AsDuplicable(); AddInput(kParameters, "").AsDuplicable(); AddInput(kPlaces, ""); diff --git a/paddle/fluid/operators/pool_mkldnn_op.cc b/paddle/fluid/operators/pool_mkldnn_op.cc index 63eaaedcd5fc3df17902511dc02b25bf43ccd241..60e936298defe7c6ce8a33bdc7de05b52eb950e7 100644 --- a/paddle/fluid/operators/pool_mkldnn_op.cc +++ b/paddle/fluid/operators/pool_mkldnn_op.cc @@ -18,6 +18,26 @@ limitations under the License. */ namespace paddle { namespace operators { +using mkldnn::memory; // Note: paddle has also "memory" namespace +using mkldnn::pooling_forward; +using mkldnn::pooling_backward; + +// Generate keys for storing/retriving primitives for this operator +// TODO(jczaja): Make hashing function more optimial +static std::string gethash(memory::dims& input_dims, std::string& pooling_type, + std::vector& ksize, std::vector& strides, + std::vector& paddings, std::string suffix) { + auto dims2str = [](memory::dims& operand_dims) { + std::string dstr = ""; + for (size_t i = 0; i < operand_dims.size(); ++i) { + dstr += std::to_string(operand_dims[i]) + "-"; + } + return dstr; + }; + return dims2str(input_dims) + dims2str(ksize) + dims2str(strides) + + dims2str(paddings) + pooling_type + suffix; +} + template class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { public: @@ -34,10 +54,6 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { // Get an unique name from "argument" name of "Out" variable // This name will be used as key when saving info into device context - const std::string key = ctx.op().Output("Out"); - const std::string key_pool_pd = key + "@pool_pd"; - const std::string key_pool_workspace_memory = - key + "@pool_workspace_memory"; std::string pooling_type = ctx.Attr("pooling_type"); std::vector ksize = ctx.Attr>("ksize"); @@ -63,37 +79,71 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { std::vector src_tz = paddle::framework::vectorize2int(input->dims()); std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); - // TODO(pzelazko-intel): support more formats - auto src_md = platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - auto dst_md = platform::MKLDNNMemDesc(dst_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - - std::shared_ptr pool_pd = - CreatePrimitiveDesc(src_md, dst_md, strides, paddings, ksize, - pooling_type, mkldnn_engine); - - // save pool_pd into global device context to be referred in backward path - dev_ctx.SetBlob(key_pool_pd, pool_pd); - - std::shared_ptr workspace_memory = - CreateWorkspaceMemory(pool_pd, pooling_type, mkldnn_engine); - - // save pool_workspace_memory to be referred in backward path - dev_ctx.SetBlob(key_pool_workspace_memory, workspace_memory); - - auto src_memory = - mkldnn::memory({src_md, mkldnn_engine}, - static_cast(const_cast(input_data))); - auto dst_memory = - mkldnn::memory({dst_md, mkldnn_engine}, - static_cast(const_cast(output_data))); + const std::string key = gethash(src_tz, pooling_type, ksize, strides, + paddings, ctx.op().Output("Out")); + const std::string key_pool_p = key + "@pool_p"; + const std::string key_pool_pd = key + "@pool_pd"; + const std::string key_pool_src_mem_p = key + "@pool_src_mem_p"; + const std::string key_pool_dst_mem_p = key + "@pool_dst_mem_p"; + const std::string key_pool_workspace_memory = + key + "@pool_workspace_memory"; - auto pool_prim = mkldnn::pooling_forward(*pool_pd, src_memory, dst_memory, - *workspace_memory); + auto pool_p = + std::static_pointer_cast(dev_ctx.GetBlob(key_pool_p)); + if (pool_p == nullptr) { + // TODO(pzelazko-intel): support more formats + + auto src_md = + platform::MKLDNNMemDesc(src_tz, platform::MKLDNNGetDataType(), + mkldnn::memory::format::nchw); + auto dst_md = + platform::MKLDNNMemDesc(dst_tz, platform::MKLDNNGetDataType(), + mkldnn::memory::format::nchw); + + std::shared_ptr pool_pd = + CreatePrimitiveDesc(src_md, dst_md, strides, paddings, ksize, + pooling_type, mkldnn_engine); + + // save pool_pd into global device context to be referred in backward path + dev_ctx.SetBlob(key_pool_pd, pool_pd); + + std::shared_ptr workspace_memory = + CreateWorkspaceMemory(pool_pd, pooling_type, mkldnn_engine); + + // save pool_workspace_memory to be referred in backward path + dev_ctx.SetBlob(key_pool_workspace_memory, workspace_memory); + + auto pool_src_memory_p = std::make_shared( + memory::primitive_desc{src_md, mkldnn_engine}, + static_cast(const_cast(input_data))); + dev_ctx.SetBlob(key_pool_src_mem_p, pool_src_memory_p); + + auto pool_dst_memory_p = std::make_shared( + memory::primitive_desc{dst_md, mkldnn_engine}, + static_cast(output_data)); + dev_ctx.SetBlob(key_pool_dst_mem_p, pool_dst_memory_p); + + pool_p = std::make_shared( + *pool_pd, *(pool_src_memory_p.get()), *(pool_dst_memory_p.get()), + *workspace_memory); + dev_ctx.SetBlob(key_pool_p, pool_p); + } else { + // Primitives already exist + auto pool_src_memory_p = + std::static_pointer_cast(dev_ctx.GetBlob(key_pool_src_mem_p)); + PADDLE_ENFORCE(pool_src_memory_p != nullptr, + "Fail to find pooling src mem_p in device context"); + auto pool_dst_memory_p = + std::static_pointer_cast(dev_ctx.GetBlob(key_pool_dst_mem_p)); + PADDLE_ENFORCE(pool_dst_memory_p != nullptr, + "Fail to find pooling dst mem_p in device context"); + pool_src_memory_p->set_data_handle( + reinterpret_cast(const_cast(input_data))); + pool_dst_memory_p->set_data_handle(output_data); + } // push primitive to stream and wait until it's executed - std::vector pipeline{pool_prim}; + std::vector pipeline{*(pool_p.get())}; mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } @@ -120,9 +170,10 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { mkldnn::memory::primitive_desc workspace_md = pooling_type == "max" ? pool_pd->workspace_primitive_desc() - : mkldnn::memory::primitive_desc( - {{}, mkldnn::memory::f32, mkldnn::memory::format::nchw}, - engine); + : mkldnn::memory::primitive_desc({{}, + platform::MKLDNNGetDataType(), + mkldnn::memory::format::nchw}, + engine); auto p_workspace_memory = new mkldnn::memory(workspace_md); return std::unique_ptr(p_workspace_memory); @@ -140,13 +191,6 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { const Tensor* out_grad = ctx.Input(framework::GradVarName("Out")); Tensor* in_x_grad = ctx.Output(framework::GradVarName("X")); - // Get an unique name from "argument" name of "Out" variable - // This name will be used as key when referring info from device context - const std::string key = ctx.op().Input("Out"); - const std::string key_pool_pd = key + "@pool_pd"; - const std::string key_pool_workspace_memory = - key + "@pool_workspace_memory"; - std::string pooling_type = ctx.Attr("pooling_type"); std::vector ksize = ctx.Attr>("ksize"); std::vector strides = ctx.Attr>("strides"); @@ -171,43 +215,76 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { std::vector diff_dst_tz = paddle::framework::vectorize2int(out_grad->dims()); - auto diff_src_md = platform::MKLDNNMemDesc(diff_src_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - auto diff_dst_md = platform::MKLDNNMemDesc(diff_dst_tz, mkldnn::memory::f32, - mkldnn::memory::format::nchw); - - // Retrieve pool_pd/pool_workspace_memory from device context - auto pool_pd = - std::static_pointer_cast( - dev_ctx.GetBlob(key_pool_pd)); - PADDLE_ENFORCE(pool_pd != nullptr, - "Fail to find pool_pd in device context"); - - auto workspace_memory = std::static_pointer_cast( - dev_ctx.GetBlob(key_pool_workspace_memory)); - PADDLE_ENFORCE(workspace_memory != nullptr, - "Fail to find workspace_memory in device context"); - - auto pool_bwd_desc = mkldnn::pooling_backward::desc( - pooling_type == "max" ? mkldnn::algorithm::pooling_max - : mkldnn::algorithm::pooling_avg, - diff_src_md, diff_dst_md, strides, ksize, paddings, paddings, - mkldnn::padding_kind::zero); - auto pool_bwd_pd = mkldnn::pooling_backward::primitive_desc( - pool_bwd_desc, mkldnn_engine, *pool_pd); - - auto diff_src_memory = - mkldnn::memory({diff_src_md, mkldnn_engine}, - static_cast(const_cast(in_x_grad_data))); - auto diff_dst_memory = - mkldnn::memory({diff_dst_md, mkldnn_engine}, - static_cast(const_cast(out_grad_data))); + // Get an unique name from "argument" name of "Out" variable + // This name will be used as key when referring info from device context + const std::string key = gethash(diff_src_tz, pooling_type, ksize, strides, + paddings, ctx.op().Input("Out")); + const std::string key_pool_bwd_p = key + "@pool_bwd_p"; + const std::string key_pool_diff_src_mem_p = key + "@pool_diff_src_mem_p"; + const std::string key_pool_diff_dst_mem_p = key + "@pool_diff_dst_mem_p"; + const std::string key_pool_pd = key + "@pool_pd"; + const std::string key_pool_workspace_memory = + key + "@pool_workspace_memory"; - auto bwd_prim = mkldnn::pooling_backward( - pool_bwd_pd, diff_dst_memory, *workspace_memory, diff_src_memory); + auto pool_bwd_p = std::static_pointer_cast( + dev_ctx.GetBlob(key_pool_bwd_p)); + if (pool_bwd_p == nullptr) { + auto diff_src_md = + platform::MKLDNNMemDesc(diff_src_tz, platform::MKLDNNGetDataType(), + mkldnn::memory::format::nchw); + auto diff_dst_md = + platform::MKLDNNMemDesc(diff_dst_tz, platform::MKLDNNGetDataType(), + mkldnn::memory::format::nchw); + // Retrieve pool_pd/pool_workspace_memory from device context + auto pool_pd = + std::static_pointer_cast( + dev_ctx.GetBlob(key_pool_pd)); + PADDLE_ENFORCE(pool_pd != nullptr, + "Fail to find pool_pd in device context"); + + auto workspace_memory = std::static_pointer_cast( + dev_ctx.GetBlob(key_pool_workspace_memory)); + PADDLE_ENFORCE(workspace_memory != nullptr, + "Fail to find workspace_memory in device context"); + + auto pool_diff_src_memory_p = std::make_shared(memory( + {diff_src_md, mkldnn_engine}, static_cast(in_x_grad_data))); + dev_ctx.SetBlob(key_pool_diff_src_mem_p, pool_diff_src_memory_p); + + auto pool_diff_dst_memory_p = std::make_shared( + memory({diff_dst_md, mkldnn_engine}, + static_cast(const_cast(out_grad_data)))); + dev_ctx.SetBlob(key_pool_diff_dst_mem_p, pool_diff_dst_memory_p); + + auto pool_bwd_desc = mkldnn::pooling_backward::desc( + pooling_type == "max" ? mkldnn::algorithm::pooling_max + : mkldnn::algorithm::pooling_avg, + diff_src_md, diff_dst_md, strides, ksize, paddings, paddings, + mkldnn::padding_kind::zero); + auto pool_bwd_pd = mkldnn::pooling_backward::primitive_desc( + pool_bwd_desc, mkldnn_engine, *pool_pd); + + pool_bwd_p = std::make_shared( + pool_bwd_pd, *(pool_diff_dst_memory_p.get()), *workspace_memory, + *(pool_diff_src_memory_p)); + dev_ctx.SetBlob(key_pool_bwd_p, pool_bwd_p); + } else { + // Primitives already exist + auto pool_diff_src_memory_p = std::static_pointer_cast( + dev_ctx.GetBlob(key_pool_diff_src_mem_p)); + PADDLE_ENFORCE(pool_diff_src_memory_p != nullptr, + "Fail to find pooling src mem_p in device context"); + auto pool_diff_dst_memory_p = std::static_pointer_cast( + dev_ctx.GetBlob(key_pool_diff_dst_mem_p)); + PADDLE_ENFORCE(pool_diff_dst_memory_p != nullptr, + "Fail to find pooling dst mem_p in device context"); + pool_diff_src_memory_p->set_data_handle( + reinterpret_cast(in_x_grad_data)); + pool_diff_dst_memory_p->set_data_handle(const_cast(out_grad_data)); + } // push primitive to stream and wait until it's executed - std::vector pipeline{bwd_prim}; + std::vector pipeline{*(pool_bwd_p.get())}; mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } // Compute() }; diff --git a/paddle/fluid/operators/pool_op.cc b/paddle/fluid/operators/pool_op.cc index f2de075e0d82fc5bd0ac41b481ac80314f3857a3..f4fb2b132fe8d59cb50f5a1f7359240ac50445fe 100644 --- a/paddle/fluid/operators/pool_op.cc +++ b/paddle/fluid/operators/pool_op.cc @@ -135,8 +135,7 @@ framework::OpKernelType PoolOpGrad::GetExpectedKernelType( library_); } -Pool2dOpMaker::Pool2dOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void Pool2dOpMaker::Make() { AddInput( "X", "(Tensor) The input tensor of pooling operator. " @@ -229,8 +228,7 @@ Example: )DOC"); } -Pool3dOpMaker::Pool3dOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { +void Pool3dOpMaker::Make() { AddInput("X", "(Tensor) The input tensor of pooling operator. " "The format of input tensor is NCDHW, where N is batch size, C is " diff --git a/paddle/fluid/operators/pool_op.h b/paddle/fluid/operators/pool_op.h index a48127ea6983d3d4ea12ec4925f30af233002ef2..a63963ca926bb94ff99e5cfe6dbcb2b15075bcb8 100644 --- a/paddle/fluid/operators/pool_op.h +++ b/paddle/fluid/operators/pool_op.h @@ -50,12 +50,12 @@ class PoolOpGrad : public framework::OperatorWithKernel { class Pool2dOpMaker : public framework::OpProtoAndCheckerMaker { public: - Pool2dOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; class Pool3dOpMaker : public framework::OpProtoAndCheckerMaker { public: - Pool3dOpMaker(OpProto* proto, OpAttrChecker* op_checker); + void Make() override; }; template diff --git a/paddle/fluid/operators/pool_with_index_op.cc b/paddle/fluid/operators/pool_with_index_op.cc index 848cd61b23c2389d3fe11f585b256d55c1ff177f..873706593e4c856f0079738654a9e7e59a1c0cd8 100644 --- a/paddle/fluid/operators/pool_with_index_op.cc +++ b/paddle/fluid/operators/pool_with_index_op.cc @@ -100,8 +100,7 @@ class MaxPoolWithIndexOpGrad : public framework::OperatorWithKernel { class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker { public: - MaxPool2dWithIndexOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(Tensor) The input tensor of pooling operator. " @@ -177,8 +176,7 @@ Example: class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker { public: - MaxPool3dWithIndexOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input tensor of pooling operator. " "The format of input tensor is NCDHW, where N is batch size, C is " diff --git a/paddle/fluid/operators/positive_negative_pair_op.cc b/paddle/fluid/operators/positive_negative_pair_op.cc index d237da25a00de13057e009b6705d3241b8b26539..4d865b7f17b050ac6f04addc9949f3f65da06ded 100644 --- a/paddle/fluid/operators/positive_negative_pair_op.cc +++ b/paddle/fluid/operators/positive_negative_pair_op.cc @@ -95,8 +95,7 @@ class PositiveNegativePairOp : public framework::OperatorWithKernel { class PositiveNegativePairOpMaker : public framework::OpProtoAndCheckerMaker { public: - PositiveNegativePairOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Score", "(Tensor, float) Model Score on an item (with " "respect to QueryID). It's a 2-D tensor with shape [batch_size, " diff --git a/paddle/fluid/operators/precision_recall_op.cc b/paddle/fluid/operators/precision_recall_op.cc index c34b0d072bdb2f5b97dd4615ff9338d98f2bfbe5..e7ce16f33fb5052ffb41fc05bd1538e2f0dc35be 100644 --- a/paddle/fluid/operators/precision_recall_op.cc +++ b/paddle/fluid/operators/precision_recall_op.cc @@ -90,8 +90,7 @@ class PrecisionRecallOp : public framework::OperatorWithKernel { class PrecisionRecallOpMaker : public framework::OpProtoAndCheckerMaker { public: - PrecisionRecallOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("MaxProbs", "(Tensor, default Tensor) A 2-D tensor with shape N x 1, " "where N is the batch size. Each row contains the max probability " diff --git a/paddle/fluid/operators/prefetch_op.cc b/paddle/fluid/operators/prefetch_op.cc index f9ae01ab5d2972d2a74b36ae6035985d1d874bb6..4cfea958e8e50156c90af8806414b043e15f8a9c 100644 --- a/paddle/fluid/operators/prefetch_op.cc +++ b/paddle/fluid/operators/prefetch_op.cc @@ -64,8 +64,7 @@ class PrefetchOp : public framework::OperatorBase { class PrefetchOpMaker : public framework::OpProtoAndCheckerMaker { public: - PrefetchOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("X", "(LoDTensor) Input Id variables to be sent").AsDuplicable(); AddOutput("RPCClient", "(RPCClient) The RPC client object which will be" diff --git a/paddle/fluid/operators/prelu_op.cc b/paddle/fluid/operators/prelu_op.cc index a066b3e06e5eca2661827425b5b2d0059d5bcc3c..db040509bc08c3f6ad031c5b97c93574e31337e0 100644 --- a/paddle/fluid/operators/prelu_op.cc +++ b/paddle/fluid/operators/prelu_op.cc @@ -38,8 +38,7 @@ class PReluOp : public framework::OperatorWithKernel { class PReluOpMaker : public framework::OpProtoAndCheckerMaker { public: - PReluOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input tensor of prelu operator."); AddInput("Alpha", "The alpha weight of prelu operator."); AddOutput("Out", "The output tensor of prelu operator."); diff --git a/paddle/fluid/operators/print_op.cc b/paddle/fluid/operators/print_op.cc index fafc7e54d7a44d6bb2dadf67135537dc16430e76..db7634918a5179a61304315ecd08350d23fb4642 100644 --- a/paddle/fluid/operators/print_op.cc +++ b/paddle/fluid/operators/print_op.cc @@ -209,8 +209,7 @@ class TensorPrintOp : public framework::OperatorBase { class PrintOpProtoAndCheckMaker : public framework::OpProtoAndCheckerMaker { public: - PrintOpProtoAndCheckMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("In", "Input tensor to be displayed."); AddAttr("first_n", "Only log `first_n` number of times."); AddAttr("message", "A string message to print as a prefix."); diff --git a/paddle/fluid/operators/proximal_adagrad_op.cc b/paddle/fluid/operators/proximal_adagrad_op.cc index e057244c1e974edea1b9bbc76c0585c295495299..8d8075d76111928ec9855eb0b70fe6dbd90a979b 100644 --- a/paddle/fluid/operators/proximal_adagrad_op.cc +++ b/paddle/fluid/operators/proximal_adagrad_op.cc @@ -66,8 +66,7 @@ class ProximalAdagradOp : public framework::OperatorWithKernel { class ProximalAdagradOpMaker : public framework::OpProtoAndCheckerMaker { public: - ProximalAdagradOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor, default Tensor) " "Input parameter that has to be updated."); diff --git a/paddle/fluid/operators/proximal_gd_op.cc b/paddle/fluid/operators/proximal_gd_op.cc index ed1472631870e5aee6b0e8b8f80bb5e6c84a3851..baf9cbcba2ed89f62afc9816e0ab9e0f112e6008 100644 --- a/paddle/fluid/operators/proximal_gd_op.cc +++ b/paddle/fluid/operators/proximal_gd_op.cc @@ -54,8 +54,7 @@ class ProximalGDOp : public framework::OperatorWithKernel { class ProximalGDOpMaker : public framework::OpProtoAndCheckerMaker { public: - ProximalGDOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor, default Tensor) " "Input parameter value that has to be updated."); diff --git a/paddle/fluid/operators/rank_loss_op.cc b/paddle/fluid/operators/rank_loss_op.cc index eb9ff8de3e4b37ef0bbf7477c1bb62856bdb6310..313cf01541dd88a0f4f8bf54fe4436984c2cbcf8 100644 --- a/paddle/fluid/operators/rank_loss_op.cc +++ b/paddle/fluid/operators/rank_loss_op.cc @@ -46,8 +46,7 @@ class RankLossOp : public framework::OperatorWithKernel { class RankLossOpMaker : public framework::OpProtoAndCheckerMaker { public: - RankLossOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Label", "(2-D Tensor with shape [batch_size x 1]) " "The label indicating A ranked higher than B or not."); diff --git a/paddle/fluid/operators/read_op.cc b/paddle/fluid/operators/read_op.cc index bf02b9958927580608b95d6b8ecfddc7231a02d4..72a27d43584d55cd0859c63577ae85ff0f5fdfa8 100644 --- a/paddle/fluid/operators/read_op.cc +++ b/paddle/fluid/operators/read_op.cc @@ -79,8 +79,7 @@ class ReadOp : public framework::OperatorBase { class ReadOpMaker : public framework::OpProtoAndCheckerMaker { public: - ReadOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(op_proto, op_checker) { + void Make() override { AddInput("Reader", "(ReaderHolder) The executed reader."); AddOutput("Out", "(LoDTensor) The output data.").AsDuplicable(); AddComment(R"DOC( diff --git a/paddle/fluid/operators/reader/create_batch_reader_op.cc b/paddle/fluid/operators/reader/create_batch_reader_op.cc index 04c5872bef4600e30ba572a025cc5f0a5e9839ca..4cc7cbc6e89b0712faf9ad9c51480bce00da15f5 100644 --- a/paddle/fluid/operators/reader/create_batch_reader_op.cc +++ b/paddle/fluid/operators/reader/create_batch_reader_op.cc @@ -52,9 +52,8 @@ class CreateBatchReaderOp : public framework::OperatorBase { }; class CreateBatchReaderOpMaker : public DecoratedReaderMakerBase { - public: - CreateBatchReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : DecoratedReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddAttr("batch_size", "How many instances the batch reader yields each time.") .GreaterThan(0); diff --git a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc index e5efac461512a9a1869318d6547233589ca45a77..bc830a2b72e657f79f4c94e24428d38ff2b7c42e 100644 --- a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc +++ b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc @@ -113,14 +113,13 @@ class CreateDoubleBufferReaderOp : public framework::OperatorBase { }; class CreateDoubleBufferReaderOpMaker : public DecoratedReaderMakerBase { - public: - CreateDoubleBufferReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : DecoratedReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddComment(R"DOC( CreateDoubleBufferReader Operator A double buffer reader takes another reader as its 'underlying reader'. - It launches another thread to execute the 'underlying reader' asynchronously, + It launches another thread to execute the 'underlying reader' asynchronously, which prevents reading process from blocking subsequent training. )DOC"); std::unordered_set enum_range; diff --git a/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc b/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc index 0573345ba502b6a9af35710840d5acf7634f332f..249b0b7c6dbc8b8104bce95562e6e9b2a28c77f8 100644 --- a/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc +++ b/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc @@ -65,20 +65,19 @@ class CreateMultiPassReaderOp : public framework::OperatorBase { }; class CreateMultiPassReaderOpMaker : public DecoratedReaderMakerBase { - public: - CreateMultiPassReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : DecoratedReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddAttr("pass_num", "The number of pass to run.").GreaterThan(0); AddComment(R"DOC( CreateMultiPassReader Operator - This operator creates a multi-pass reader. A multi-pass reader - is used to yield data for several pass training continuously. + This operator creates a multi-pass reader. A multi-pass reader + is used to yield data for several pass training continuously. It takes the number of passes to run as one of its attributes - ('pass_num'), and maintains a pass counter to record how many - passes it has completed. When the underlying reader reaches the - EOF, the multi-pass reader checks whether it has completed training - of the given number of pass. If not, the underlying reader will + ('pass_num'), and maintains a pass counter to record how many + passes it has completed. When the underlying reader reaches the + EOF, the multi-pass reader checks whether it has completed training + of the given number of pass. If not, the underlying reader will be re-initialized and starts a new pass automatically. )DOC"); } diff --git a/paddle/fluid/operators/reader/create_random_data_generator_op.cc b/paddle/fluid/operators/reader/create_random_data_generator_op.cc index d1cb8e47da70cab784858caea7e791151fc104dd..55bb9739e0239d31f63c3d8703bcf1d18bf459dc 100644 --- a/paddle/fluid/operators/reader/create_random_data_generator_op.cc +++ b/paddle/fluid/operators/reader/create_random_data_generator_op.cc @@ -84,9 +84,8 @@ class CreateRandomDataGeneratorOp : public framework::OperatorBase { }; class CreateRandomDataGeneratorOpMaker : public FileReaderMakerBase { - public: - CreateRandomDataGeneratorOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : FileReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddAttr("min", "The lower bound of reader's uniform distribution."); AddAttr("max", "The upper bound of reader's uniform distribution."); AddComment(R"DOC( diff --git a/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc b/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc index 2ae29725561769ebe6428002c9983246b8eec724..282ec3f36b98e7aa62d71fb04f72721a5464e21c 100644 --- a/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc +++ b/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc @@ -76,9 +76,8 @@ class CreateRecordIOReaderOp : public framework::OperatorBase { }; class CreateRecordIOReaderOpMaker : public FileReaderMakerBase { - public: - CreateRecordIOReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : FileReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddAttr("filename", "The filename of record io reader"); AddComment(R"DOC( CreateRecordIOReader Operator diff --git a/paddle/fluid/operators/reader/create_shuffle_reader_op.cc b/paddle/fluid/operators/reader/create_shuffle_reader_op.cc index 13825d65913be95f4f444bd9d5271a036ec8b1e2..fd233be945932eee9f9a3c0c578a43d5b7cc83aa 100644 --- a/paddle/fluid/operators/reader/create_shuffle_reader_op.cc +++ b/paddle/fluid/operators/reader/create_shuffle_reader_op.cc @@ -92,9 +92,8 @@ class CreateShuffleReaderOp : public framework::OperatorBase { }; class CreateShuffleReaderOpMaker : public DecoratedReaderMakerBase { - public: - CreateShuffleReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : DecoratedReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddAttr("buffer_size", "The shuffle buffer size.").GreaterThan(0); AddComment(R"DOC( CreateShuffleReader Operator diff --git a/paddle/fluid/operators/reader/create_threaded_reader_op.cc b/paddle/fluid/operators/reader/create_threaded_reader_op.cc index 1cb9bd36455a2287b8ba4fb4ca14a4c5338da098..1db70f3e9699dba604569c36dc35025dfe2c94fe 100644 --- a/paddle/fluid/operators/reader/create_threaded_reader_op.cc +++ b/paddle/fluid/operators/reader/create_threaded_reader_op.cc @@ -53,17 +53,16 @@ class CreateThreadedReaderOp : public framework::OperatorBase { }; class CreateThreadedReaderOpMaker : public DecoratedReaderMakerBase { - public: - CreateThreadedReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : DecoratedReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddComment(R"DOC( CreateThreadedReader Operator - This operator creates a threaded reader. A threaded reader's - 'ReadNext()' can be invoked by several threads at the same - time. - When the attribute 'safe_mode' is true, the threaded reader's - 'ReInit()' is disabled to avoid unexpected bugs in multi-thread + This operator creates a threaded reader. A threaded reader's + 'ReadNext()' can be invoked by several threads at the same + time. + When the attribute 'safe_mode' is true, the threaded reader's + 'ReInit()' is disabled to avoid unexpected bugs in multi-thread environment. )DOC"); } diff --git a/paddle/fluid/operators/reader/open_files_op.cc b/paddle/fluid/operators/reader/open_files_op.cc index 91ad7d56583446ee4686e74187de166f387125df..8c0dac65dd691954b112bfa61622d399b2b9c3e5 100644 --- a/paddle/fluid/operators/reader/open_files_op.cc +++ b/paddle/fluid/operators/reader/open_files_op.cc @@ -185,9 +185,8 @@ class OpenFilesOp : public framework::OperatorBase { }; class OpenFilesOpMaker : public FileReaderMakerBase { - public: - OpenFilesOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) - : FileReaderMakerBase(op_proto, op_checker) { + protected: + void Apply() override { AddAttr>("file_names", "Files to be read."); AddAttr("thread_num", "The maximal concurrent prefetch thread number.") .GreaterThan(0); @@ -196,7 +195,7 @@ class OpenFilesOpMaker : public FileReaderMakerBase { AddComment(R"DOC( OpenFiles Operator - An OpenFilesOp creates a MultiFileReader, which is able to + An OpenFilesOp creates a MultiFileReader, which is able to read data multi-threaded from multiple files. )DOC"); } diff --git a/paddle/fluid/operators/reader/reader_op_registry.cc b/paddle/fluid/operators/reader/reader_op_registry.cc index 52adc54dc22f60280348060ee535b937a3f58263..612e1f5eca3a4836db1fd167fc6bb63400d20177 100644 --- a/paddle/fluid/operators/reader/reader_op_registry.cc +++ b/paddle/fluid/operators/reader/reader_op_registry.cc @@ -53,10 +53,7 @@ std::unique_ptr CreateReaderByFileName( return std::unique_ptr(reader); } -FileReaderMakerBase::FileReaderMakerBase( - framework::OpProtoAndCheckerMaker::OpProto* op_proto, - framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(op_proto, op_checker) { +void FileReaderMakerBase::Make() { AddOutput("Out", "(ReaderHolder) The created random reader.").AsDuplicable(); AddAttr>("shape_concat", "The concat of all data's shapes."); AddAttr>( @@ -68,6 +65,7 @@ FileReaderMakerBase::FileReaderMakerBase( "It means the reader will generate two data each time," "whose shapes are [2,3,4] and [5,6] respectively."); AddAttr>("lod_levels", "The LoD levels of each data."); + Apply(); } void FileReaderInferShape::operator()(framework::InferShapeContext* ctx) const { @@ -128,13 +126,11 @@ void DecoratedReaderInferVarType::operator()( out_reader->SetDataTypes(in_reader->GetDataTypes()); } -DecoratedReaderMakerBase::DecoratedReaderMakerBase( - framework::OpProtoAndCheckerMaker::OpProto* op_proto, - framework::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(op_proto, op_checker) { +void DecoratedReaderMakerBase::Make() { AddInput("UnderlyingReader", "(ReaderHolder) The underlying reader for creating a batch reader."); AddOutput("Out", "(ReaderHolder) The created batch reader."); + Apply(); } } // namespace reader diff --git a/paddle/fluid/operators/reader/reader_op_registry.h b/paddle/fluid/operators/reader/reader_op_registry.h index ec25f55ef5c3bb691b1213328b996c080656bb7b..244bf15f068a47efc29ee54492cdbdeb10025020 100644 --- a/paddle/fluid/operators/reader/reader_op_registry.h +++ b/paddle/fluid/operators/reader/reader_op_registry.h @@ -47,7 +47,10 @@ extern std::vector RestoreShapes( class FileReaderMakerBase : public framework::OpProtoAndCheckerMaker { public: - FileReaderMakerBase(OpProto* op_proto, OpAttrChecker* op_checker); + void Make() final; + + protected: + virtual void Apply() = 0; }; class FileReaderInferShape : public framework::InferShapeBase { @@ -76,7 +79,10 @@ class DecoratedReaderInferVarType : public framework::VarTypeInference { class DecoratedReaderMakerBase : public framework::OpProtoAndCheckerMaker { public: - DecoratedReaderMakerBase(OpProto* op_proto, OpAttrChecker* op_checker); + void Make() final; + + protected: + virtual void Apply() = 0; }; } // namespace reader diff --git a/paddle/fluid/operators/recurrent_op.cc b/paddle/fluid/operators/recurrent_op.cc index 72c2905872c528a7ed05820744f4031799ad9e46..9c1cee7022a9b9a98f026f7602f0f7badc44a49b 100644 --- a/paddle/fluid/operators/recurrent_op.cc +++ b/paddle/fluid/operators/recurrent_op.cc @@ -508,8 +508,7 @@ class RecurrentGradOp : public RecurrentBase { class RecurrentOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - RecurrentOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kInputs, "rnn inputs").AsDuplicable(); AddInput(kInitialStates, "rnn initial states").AsDuplicable(); AddInput(kParameters, diff --git a/paddle/fluid/operators/recv_op.cc b/paddle/fluid/operators/recv_op.cc index a4dcf704a63ae3bad6567ddb042ea23513bccff7..7148bd0e363a71b58581a6c3c5f245d98d5b9d02 100644 --- a/paddle/fluid/operators/recv_op.cc +++ b/paddle/fluid/operators/recv_op.cc @@ -53,8 +53,7 @@ class RecvOp : public framework::OperatorBase { class RecvOpMaker : public framework::OpProtoAndCheckerMaker { public: - RecvOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddOutput("Out", "(Tensor) Variables to get from server.").AsDuplicable(); AddComment(R"DOC( Recv operator diff --git a/paddle/fluid/operators/reduce_op.cc b/paddle/fluid/operators/reduce_op.cc index 093db966472cf100b2f1e4159ce20399cee1f481..eb8c21179db690e20db29c21892fd6258dd75579 100644 --- a/paddle/fluid/operators/reduce_op.cc +++ b/paddle/fluid/operators/reduce_op.cc @@ -90,8 +90,7 @@ class ReduceGradOp : public framework::OperatorWithKernel { class ReduceOpMaker : public framework::OpProtoAndCheckerMaker { public: - ReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() final { AddInput("X", "(Tensor) The input tensor. Tensors with rank at most 6 are " "supported."); @@ -111,78 +110,20 @@ class ReduceOpMaker : public framework::OpProtoAndCheckerMaker { "(bool, default false) " "If true, output a scalar reduced along all dimensions.") .SetDefault(false); - comment_ = R"DOC( -{ReduceOp} Operator. + AddComment(string::Sprintf(R"DOC( +%s Operator. -This operator computes the {reduce} of input tensor along the given dimension. +This operator computes the %s of input tensor along the given dimension. The result tensor has 1 fewer dimension than the input unless keep_dim is true. If reduce_all is true, just reduce along all dimensions and output a scalar. -)DOC"; - AddComment(comment_); +)DOC", + GetOpType(), GetName())); } protected: - std::string comment_; - - void Replace(std::string *src, std::string from, std::string to) { - std::size_t len_from = std::strlen(from.c_str()); - std::size_t len_to = std::strlen(to.c_str()); - for (std::size_t pos = src->find(from); pos != std::string::npos; - pos = src->find(from, pos + len_to)) { - src->replace(pos, len_from, to); - } - } - - void SetComment(std::string name, std::string op) { - Replace(&comment_, "{ReduceOp}", name); - Replace(&comment_, "{reduce}", op); - } -}; - -class ReduceSumOpMaker : public ReduceOpMaker { - public: - ReduceSumOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : ReduceOpMaker(proto, op_checker) { - SetComment("ReduceSum", "sum"); - AddComment(comment_); - } -}; - -class ReduceMeanOpMaker : public ReduceOpMaker { - public: - ReduceMeanOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : ReduceOpMaker(proto, op_checker) { - SetComment("ReduceMean", "mean"); - AddComment(comment_); - } -}; - -class ReduceMaxOpMaker : public ReduceOpMaker { - public: - ReduceMaxOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : ReduceOpMaker(proto, op_checker) { - SetComment("ReduceMax", "max"); - AddComment(comment_); - } -}; - -class ReduceMinOpMaker : public ReduceOpMaker { - public: - ReduceMinOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : ReduceOpMaker(proto, op_checker) { - SetComment("ReduceMin", "min"); - AddComment(comment_); - } -}; - -class ReduceProdOpMaker : public ReduceOpMaker { - public: - ReduceProdOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : ReduceOpMaker(proto, op_checker) { - SetComment("ReduceProd", "production"); - AddComment(comment_); - } + virtual std::string GetName() const = 0; + virtual std::string GetOpType() const = 0; }; } // namespace operators @@ -190,25 +131,21 @@ class ReduceProdOpMaker : public ReduceOpMaker { namespace ops = paddle::operators; -REGISTER_OPERATOR(reduce_sum, ops::ReduceOp, ops::ReduceSumOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(reduce_sum_grad, ops::ReduceGradOp); - -REGISTER_OPERATOR(reduce_mean, ops::ReduceOp, ops::ReduceMeanOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(reduce_mean_grad, ops::ReduceGradOp); - -REGISTER_OPERATOR(reduce_max, ops::ReduceOp, ops::ReduceMaxOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(reduce_max_grad, ops::ReduceGradOp); - -REGISTER_OPERATOR(reduce_min, ops::ReduceOp, ops::ReduceMinOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(reduce_min_grad, ops::ReduceGradOp); - -REGISTER_OPERATOR(reduce_prod, ops::ReduceOp, ops::ReduceProdOpMaker, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(reduce_prod_grad, ops::ReduceGradOp); +#define REGISTER_REDUCE_OP(op_name) \ + class __##op_name##Maker__ : public ops::ReduceOpMaker { \ + protected: \ + virtual std::string GetName() const { return #op_name; } \ + virtual std::string GetOpType() const { return "Reduce " #op_name; } \ + }; \ + REGISTER_OPERATOR(reduce_##op_name, ops::ReduceOp, __##op_name##Maker__, \ + paddle::framework::DefaultGradOpDescMaker); \ + REGISTER_OPERATOR(reduce_##op_name##_grad, ops::ReduceGradOp) + +REGISTER_REDUCE_OP(sum); +REGISTER_REDUCE_OP(mean); +REGISTER_REDUCE_OP(max); +REGISTER_REDUCE_OP(min); +REGISTER_REDUCE_OP(prod); #define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor) \ REGISTER_OP_CPU_KERNEL(reduce_type, \ diff --git a/paddle/fluid/operators/reorder_lod_tensor_by_rank_op.cc b/paddle/fluid/operators/reorder_lod_tensor_by_rank_op.cc index 5c3e1f5678df0270c837ed407d1e6cc662276880..e4f4fe358e0e8cd2080525227f14a3d40f3c1411 100644 --- a/paddle/fluid/operators/reorder_lod_tensor_by_rank_op.cc +++ b/paddle/fluid/operators/reorder_lod_tensor_by_rank_op.cc @@ -23,9 +23,7 @@ namespace operators { class ReorderLoDTensorByRankTableOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ReorderLoDTensorByRankTableOpProtoMaker(OpProto *proto, - OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor), the input lod tensor to be reordered according to " "Input(RankTable)."); diff --git a/paddle/fluid/operators/reshape_op.cc b/paddle/fluid/operators/reshape_op.cc index 5e5ccc3ded95d57dfed37c1ac9c7eae61d36b8c0..7f743f577fbcdaf6f62e01031e25ef09a842c2e9 100644 --- a/paddle/fluid/operators/reshape_op.cc +++ b/paddle/fluid/operators/reshape_op.cc @@ -22,8 +22,7 @@ namespace operators { class ReshapeOpMaker : public framework::OpProtoAndCheckerMaker { public: - ReshapeOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor). The input tensor of reshape operator."); AddInput("Shape", "(Tensor, optional). If provided, reshape according to " diff --git a/paddle/fluid/operators/reshape_op.h b/paddle/fluid/operators/reshape_op.h index ccd7063fe69e0f21b4d2a821bb70902b39c9b9de..3dd8c7c11eca241e747bfa129962032d882ce44c 100644 --- a/paddle/fluid/operators/reshape_op.h +++ b/paddle/fluid/operators/reshape_op.h @@ -92,14 +92,16 @@ class ReshapeOp : public framework::OperatorWithKernel { } if (unk_dim_idx != -1) { - output_shape[unk_dim_idx] = -in_size / capacity; - // in_size < 0 and is un-determinate in compile time, skip the check, - // for example, in_dims = [-1, 8, 1, 1], shape = [-1, 3, 8], - // capacity = -24, in_size = -8, output_shape[0] = 0 - // the following check will fail. if (in_size > 0) { + // in_size < 0 and is un-determinate in compile time, skip the check, + // for example, in_dims = [-1, 8, 1, 1], shape = [-1, 3, 8], + // capacity = -24, in_size = -8, output_shape[0] = 0 + // the following check will fail. + output_shape[unk_dim_idx] = -in_size / capacity; PADDLE_ENFORCE_EQ(output_shape[unk_dim_idx] * capacity, -in_size, "Invalid shape is given."); + } else { + output_shape[unk_dim_idx] = -1; } } else { PADDLE_ENFORCE_EQ(capacity, in_size, "Invalid shape is given."); @@ -122,7 +124,10 @@ class ReshapeKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext &ctx) const { auto *out = ctx.Output("Out"); auto *in = ctx.Input("X"); - auto *shape_tensor = ctx.Input("Shape"); + + auto *shape_tensor = ctx.HasInput("Shape") + ? ctx.Input("Shape") + : nullptr; framework::DDim out_dims = out->dims(); diff --git a/paddle/fluid/operators/rmsprop_op.cc b/paddle/fluid/operators/rmsprop_op.cc index a8855b3ccd1686c75953e762ce42cc27b26202e6..919ebe48ca38040274bd2052b95ef96eccff4db6 100644 --- a/paddle/fluid/operators/rmsprop_op.cc +++ b/paddle/fluid/operators/rmsprop_op.cc @@ -63,8 +63,7 @@ class RmspropOp : public framework::OperatorWithKernel { class RmspropOpMaker : public framework::OpProtoAndCheckerMaker { public: - RmspropOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor, default Tensor) " "Input parameter value that has to be updated."); diff --git a/paddle/fluid/operators/rnn_memory_helper_op.cc b/paddle/fluid/operators/rnn_memory_helper_op.cc index 70f205d887ef710aeed02905713200ce32988987..23e5fc1112d0b1e634d0ab288721cbba57b3ffe5 100644 --- a/paddle/fluid/operators/rnn_memory_helper_op.cc +++ b/paddle/fluid/operators/rnn_memory_helper_op.cc @@ -59,8 +59,7 @@ class RNNMemoryHelperOpShapeInference : public framework::InferShapeBase { class RNNMemoryHelperOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - RNNMemoryHelperOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", ""); AddOutput("Out", ""); AddAttr("dtype", @@ -117,8 +116,7 @@ class RNNMemoryHelperGradOp : public framework::OperatorBase { class RNNMemoryHelperGradOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: - RNNMemoryHelperGradOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(framework::GradVarName("Out"), ""); AddInput("X", ""); AddInput("Out", ""); diff --git a/paddle/fluid/operators/roi_pool_op.cc b/paddle/fluid/operators/roi_pool_op.cc index 397e49ef20ac45515a852f466d693f358ef5461b..293abb0ea4f1ac03c3889ce2937ef8fa0845db73 100644 --- a/paddle/fluid/operators/roi_pool_op.cc +++ b/paddle/fluid/operators/roi_pool_op.cc @@ -98,8 +98,7 @@ class ROIPoolGradOp : public framework::OperatorWithKernel { class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker { public: - ROIPoolOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor), " "the input of ROIPoolOp. " diff --git a/paddle/fluid/operators/roi_pool_op.cu b/paddle/fluid/operators/roi_pool_op.cu index f905d690f984a20622c5fbcbcc813d888dfb19d9..50450b62f7b1c0b2b5abf01a43581a0e2d2cd01e 100644 --- a/paddle/fluid/operators/roi_pool_op.cu +++ b/paddle/fluid/operators/roi_pool_op.cu @@ -38,10 +38,10 @@ __global__ void GPUROIPoolForward( int index = blockIdx.x * blockDim.x + threadIdx.x; int offset = blockDim.x * gridDim.x; for (size_t i = index; i < nthreads; i += offset) { - int pw = index % pooled_width; - int ph = (index / pooled_width) % pooled_height; - int c = (index / pooled_width / pooled_height) % channels; - int n = index / pooled_width / pooled_height / channels; + int pw = i % pooled_width; + int ph = (i / pooled_width) % pooled_height; + int c = (i / pooled_width / pooled_height) % channels; + int n = i / pooled_width / pooled_height / channels; const int64_t* offset_input_rois = input_rois + n * kROISize; int roi_batch_ind = roi_batch_id_data[n]; @@ -52,14 +52,19 @@ __global__ void GPUROIPoolForward( int roi_width = max(roi_end_w - roi_start_w + 1, 1); int roi_height = max(roi_end_h - roi_start_h + 1, 1); - T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); - T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); - - int hstart = static_cast(floor(static_cast(ph) * bin_size_h)); - int wstart = static_cast(floor(static_cast(pw) * bin_size_w)); - int hend = static_cast(ceil(static_cast(ph + 1) * bin_size_h)); - int wend = static_cast(ceil(static_cast(pw + 1) * bin_size_w)); + int hstart = static_cast(floor(static_cast(ph) * + static_cast(roi_height) / + static_cast(pooled_height))); + int wstart = static_cast(floor(static_cast(pw) * + static_cast(roi_width) / + static_cast(pooled_width))); + int hend = static_cast(ceil(static_cast(ph + 1) * + static_cast(roi_height) / + static_cast(pooled_height))); + int wend = static_cast(ceil(static_cast(pw + 1) * + static_cast(roi_width) / + static_cast(pooled_width))); hstart = min(max(hstart + roi_start_h, 0), height); hend = min(max(hend + roi_start_h, 0), height); wstart = min(max(wstart + roi_start_w, 0), width); @@ -79,9 +84,9 @@ __global__ void GPUROIPoolForward( } } } - output_data[index] = maxval; + output_data[i] = maxval; if (argmax_data) { - argmax_data[index] = maxidx; + argmax_data[i] = maxidx; } } } @@ -96,10 +101,10 @@ __global__ void GPUROIPoolBackward( int index = blockIdx.x * blockDim.x + threadIdx.x; int offset = blockDim.x * gridDim.x; for (int i = index; i < nthreads; i += offset) { - int pw = index % pooled_width; - int ph = (index / pooled_width) % pooled_height; - int c = (index / pooled_width / pooled_height) % channels; - int n = index / pooled_width / pooled_height / channels; + int pw = i % pooled_width; + int ph = (i / pooled_width) % pooled_height; + int c = (i / pooled_width / pooled_height) % channels; + int n = i / pooled_width / pooled_height / channels; int roi_batch_ind = roi_batch_id_data[n]; int input_offset = (roi_batch_ind * channels + c) * height * width; @@ -138,6 +143,7 @@ class GPUROIPoolOpKernel : public framework::OpKernel { int width = in_dims[3]; int rois_num = rois->dims()[0]; + if (rois_num == 0) return; int output_size = out->numel(); diff --git a/paddle/fluid/operators/row_conv_op.cc b/paddle/fluid/operators/row_conv_op.cc index 23f720da0b68cd2fd4c9b51182bf82f72078a906..20f140f962c3aac364a1239a663d5f340bbeb6b2 100644 --- a/paddle/fluid/operators/row_conv_op.cc +++ b/paddle/fluid/operators/row_conv_op.cc @@ -76,8 +76,7 @@ class RowConvGradOp : public framework::OperatorWithKernel { class RowConvOpMaker : public framework::OpProtoAndCheckerMaker { public: - RowConvOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor), the input(X) is a LodTensor, which supports " "variable time-length input sequences. The underlying tensor " diff --git a/paddle/fluid/operators/save_combine_op.cc b/paddle/fluid/operators/save_combine_op.cc index 94703393bfa53124d16e34ae4373773eece5f11f..cfee9207083b46f7c27354f22e82a7d3c38a027c 100644 --- a/paddle/fluid/operators/save_combine_op.cc +++ b/paddle/fluid/operators/save_combine_op.cc @@ -18,6 +18,7 @@ limitations under the License. */ #include #include #include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/data_type_transform.h" #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" @@ -69,6 +70,7 @@ class SaveCombineOp : public framework::OperatorBase { const platform::Place &place) const override { auto filename = Attr("file_path"); auto overwrite = Attr("overwrite"); + auto save_as_fp16 = Attr("save_as_fp16"); bool is_present = FileExists(filename); if (is_present && !overwrite) { @@ -100,8 +102,24 @@ class SaveCombineOp : public framework::OperatorBase { inp_var_names[i]); auto &tensor = var->Get(); - // Serialize tensor - framework::SerializeToStream(fout, tensor, dev_ctx); + // Serialize tensors one by one + + // Check types to see if a fp16 transformation is required + auto in_dtype = framework::ToDataType(tensor.type()); + auto out_dtype = + save_as_fp16 ? framework::proto::VarType::FP16 : in_dtype; + + if (in_dtype != out_dtype) { + auto in_kernel_type = framework::OpKernelType(in_dtype, place); + auto out_kernel_type = framework::OpKernelType(out_dtype, place); + framework::LoDTensor out; + // copy LoD info to the new tensor + out.set_lod(tensor.lod()); + framework::TransDataType(in_kernel_type, out_kernel_type, tensor, &out); + framework::SerializeToStream(fout, out, dev_ctx); + } else { + framework::SerializeToStream(fout, tensor, dev_ctx); + } } fout.close(); } @@ -109,8 +127,7 @@ class SaveCombineOp : public framework::OperatorBase { class SaveCombineOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - SaveCombineOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(vector) Input LoDTensors that need to be saved together in a file.") @@ -125,6 +142,12 @@ to a file on disk. "(boolean, default true)" "Overwrite the output file if it exists.") .SetDefault(true); + AddAttr("save_as_fp16", + "(boolean, default false)" + "If true, the tensor will be converted to float16 data " + "type and then saved. Otherwise, the tensor will be " + "directly saved without data type conversion.") + .SetDefault(false); AddAttr( "file_path", "(string)" diff --git a/paddle/fluid/operators/save_load_combine_op_test.cc b/paddle/fluid/operators/save_load_combine_op_test.cc index 2773c32a0a10269e28c24e12527711e3c5b8f869..4743e0d9499b111d8baa921dbb245431713fd7a8 100644 --- a/paddle/fluid/operators/save_load_combine_op_test.cc +++ b/paddle/fluid/operators/save_load_combine_op_test.cc @@ -17,15 +17,17 @@ limitations under the License. */ #include #include "gtest/gtest.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/platform/float16.h" USE_NO_KERNEL_OP(save_combine); USE_NO_KERNEL_OP(load_combine); -int* CreateForSaveCombineOp(int x, int y, const std::vector& lod_info, - std::string var_name, - const paddle::platform::CPUPlace& place, - paddle::framework::Scope* scope, - paddle::framework::LoD* expect_lod) { +template +T* CreateForSaveCombineOp(int x, int y, const std::vector& lod_info, + std::string var_name, + const paddle::platform::CPUPlace& place, + paddle::framework::Scope* scope, + paddle::framework::LoD* expect_lod) { auto var = scope->Var(var_name); auto tensor = var->GetMutable(); tensor->Resize({x, y}); @@ -34,9 +36,10 @@ int* CreateForSaveCombineOp(int x, int y, const std::vector& lod_info, (*expect_lod)[0].push_back(lod_info[i]); } tensor->set_lod(*expect_lod); - int* expect = tensor->mutable_data(place); + T* expect = tensor->mutable_data(place); for (int64_t i = 0; i < tensor->numel(); ++i) { - expect[i] = static_cast(i); + expect[i] = static_cast( + static_cast(i)); // For FP16, we intend to do float(float16(i)) } return expect; } @@ -48,18 +51,20 @@ paddle::framework::LoDTensor* GeneratePlaceholderBeforeLoad( return target; } -int* GetValuesAfterLoadCombineOp(paddle::framework::LoDTensor* target, - const paddle::framework::Scope& scope, - paddle::framework::LoD* actual_lod) { - int* actual = target->data(); +template +T* GetValuesAfterLoadCombineOp(paddle::framework::LoDTensor* target, + const paddle::framework::Scope& scope, + paddle::framework::LoD* actual_lod) { + T* actual = target->data(); *actual_lod = target->lod(); return actual; } -void CheckValues(int* expect, int* actual, paddle::framework::LoD expect_lod, - paddle::framework::LoD actual_lod, const int& numel) { - for (int64_t i = 0; i < numel; ++i) { - EXPECT_EQ(expect[i], actual[i]); +template +void CheckValues(T* expect, U* actual, const paddle::framework::LoD& expect_lod, + const paddle::framework::LoD& actual_lod, const int& numel) { + for (int i = 0; i < numel; ++i) { + EXPECT_EQ(expect[i], static_cast(actual[i])); } EXPECT_EQ(expect_lod.size(), actual_lod.size()); for (size_t i = 0; i < expect_lod.size(); ++i) { @@ -78,26 +83,26 @@ TEST(SaveLoadCombineOp, CPU) { std::vector lod1 = {0, 1, 2, 3, 10}; int numel1 = 100; paddle::framework::LoD expect_lod1; - int* expect1 = CreateForSaveCombineOp(10, 10, lod1, "test_var1", place, - &scope, &expect_lod1); + int* expect1 = CreateForSaveCombineOp(10, 10, lod1, "test_var1", + place, &scope, &expect_lod1); std::vector lod2 = {0, 2, 5, 10}; int numel2 = 200; paddle::framework::LoD expect_lod2; - int* expect2 = CreateForSaveCombineOp(10, 20, lod2, "test_var2", place, - &scope, &expect_lod2); + int* expect2 = CreateForSaveCombineOp(10, 20, lod2, "test_var2", + place, &scope, &expect_lod2); std::vector lod3 = {0, 2, 3, 20}; int numel3 = 4000; paddle::framework::LoD expect_lod3; - int* expect3 = CreateForSaveCombineOp(20, 200, lod3, "test_var3", place, - &scope, &expect_lod3); + int* expect3 = CreateForSaveCombineOp(20, 200, lod3, "test_var3", + place, &scope, &expect_lod3); std::vector lod4 = {0, 1, 20}; int numel4 = 1000; paddle::framework::LoD expect_lod4; - int* expect4 = CreateForSaveCombineOp(20, 50, lod4, "test_var4", place, - &scope, &expect_lod4); + int* expect4 = CreateForSaveCombineOp(20, 50, lod4, "test_var4", + place, &scope, &expect_lod4); // Set attributes std::string filename = "check_tensor.ls"; @@ -123,15 +128,176 @@ TEST(SaveLoadCombineOp, CPU) { load_combine_op->Run(scope, place); paddle::framework::LoD actual_lod1, actual_lod2, actual_lod3, actual_lod4; - int* actual1 = GetValuesAfterLoadCombineOp(target1, scope, &actual_lod1); - int* actual2 = GetValuesAfterLoadCombineOp(target2, scope, &actual_lod2); - int* actual3 = GetValuesAfterLoadCombineOp(target3, scope, &actual_lod3); - int* actual4 = GetValuesAfterLoadCombineOp(target4, scope, &actual_lod4); - - CheckValues(expect1, actual1, expect_lod1, actual_lod1, numel1); - CheckValues(expect2, actual2, expect_lod2, actual_lod2, numel2); - CheckValues(expect3, actual3, expect_lod3, actual_lod3, numel3); - CheckValues(expect4, actual4, expect_lod4, actual_lod4, numel4); + int* actual1 = GetValuesAfterLoadCombineOp(target1, scope, &actual_lod1); + int* actual2 = GetValuesAfterLoadCombineOp(target2, scope, &actual_lod2); + int* actual3 = GetValuesAfterLoadCombineOp(target3, scope, &actual_lod3); + int* actual4 = GetValuesAfterLoadCombineOp(target4, scope, &actual_lod4); + + CheckValues(expect1, actual1, expect_lod1, actual_lod1, numel1); + CheckValues(expect2, actual2, expect_lod2, actual_lod2, numel2); + CheckValues(expect3, actual3, expect_lod3, actual_lod3, numel3); + CheckValues(expect4, actual4, expect_lod4, actual_lod4, numel4); +} + +// FP16 version of SaveLoadCombineOp Test, only altering the saving aspect +// to save as FP16. +TEST(SaveCombineFP16Op, CPU) { + paddle::framework::Scope scope; + paddle::platform::CPUPlace place; + + std::vector lod1 = {0, 1, 2, 3, 10}; + int numel1 = 100; + paddle::framework::LoD expect_lod1; + float* expect1 = CreateForSaveCombineOp( + 10, 10, lod1, "test_var1", place, &scope, &expect_lod1); + + std::vector lod2 = {0, 2, 5, 10}; + int numel2 = 200; + paddle::framework::LoD expect_lod2; + float* expect2 = CreateForSaveCombineOp( + 10, 20, lod2, "test_var2", place, &scope, &expect_lod2); + + std::vector lod3 = {0, 20}; + int numel3 = 4000; + paddle::framework::LoD expect_lod3; + float* expect3 = CreateForSaveCombineOp( + 20, 200, lod3, "test_var3", place, &scope, &expect_lod3); + + std::vector lod4 = {0, 1, 20}; + int numel4 = 1000; + paddle::framework::LoD expect_lod4; + float* expect4 = CreateForSaveCombineOp( + 20, 50, lod4, "test_var4", place, &scope, &expect_lod4); + + // Set attributes + std::string filename = "check_tensor_fp16_save.ls"; + paddle::framework::AttributeMap attrs; + attrs.insert({"file_path", std::string(filename)}); + attrs.insert({"save_as_fp16", true}); + + // Run the save_combine_op + auto save_combine_op = paddle::framework::OpRegistry::CreateOp( + "save_combine", + {{"X", {"test_var1", "test_var2", "test_var3", "test_var4"}}}, {}, attrs); + save_combine_op->Run(scope, place); + + // Set up output vars + auto target1 = GeneratePlaceholderBeforeLoad("out_var1", &scope); + auto target2 = GeneratePlaceholderBeforeLoad("out_var2", &scope); + auto target3 = GeneratePlaceholderBeforeLoad("out_var3", &scope); + auto target4 = GeneratePlaceholderBeforeLoad("out_var4", &scope); + + // Run the load_combine_op + auto load_combine_op = paddle::framework::OpRegistry::CreateOp( + "load_combine", {}, + {{"Out", {"out_var1", "out_var2", "out_var3", "out_var4"}}}, attrs); + load_combine_op->Run(scope, place); + + paddle::framework::LoD actual_lod1, actual_lod2, actual_lod3, actual_lod4; + paddle::platform::float16* actual1 = + GetValuesAfterLoadCombineOp(target1, scope, + &actual_lod1); + paddle::platform::float16* actual2 = + GetValuesAfterLoadCombineOp(target2, scope, + &actual_lod2); + paddle::platform::float16* actual3 = + GetValuesAfterLoadCombineOp(target3, scope, + &actual_lod3); + paddle::platform::float16* actual4 = + GetValuesAfterLoadCombineOp(target4, scope, + &actual_lod4); + + CheckValues(expect1, actual1, expect_lod1, + actual_lod1, numel1); + CheckValues(expect2, actual2, expect_lod2, + actual_lod2, numel2); + CheckValues(expect3, actual3, expect_lod3, + actual_lod3, numel3); + CheckValues(expect4, actual4, expect_lod4, + actual_lod4, numel4); +} + +// FP16 version of SaveLoadCombineOp Test, only altering the loading aspect +// to load tensors with FP16 precision. +TEST(LoadCombineFP16Op, CPU) { + paddle::framework::Scope scope; + paddle::platform::CPUPlace place; + + std::vector lod1 = {0, 1, 2, 3, 10}; + int numel1 = 100; + paddle::framework::LoD expect_lod1; + float* expect1 = CreateForSaveCombineOp( + 10, 10, lod1, "test_var1", place, &scope, &expect_lod1); + + std::vector lod2 = {0, 2, 5, 10}; + int numel2 = 200; + paddle::framework::LoD expect_lod2; + float* expect2 = CreateForSaveCombineOp( + 10, 20, lod2, "test_var2", place, &scope, &expect_lod2); + + std::vector lod3 = {0, 20}; + int numel3 = 4000; + paddle::framework::LoD expect_lod3; + float* expect3 = CreateForSaveCombineOp( + 20, 200, lod3, "test_var3", place, &scope, &expect_lod3); + + std::vector lod4 = {0, 1, 20}; + int numel4 = 1000; + paddle::framework::LoD expect_lod4; + float* expect4 = CreateForSaveCombineOp( + 20, 50, lod4, "test_var4", place, &scope, &expect_lod4); + + // Set attributes + std::string filename = "check_tensor_fp16_load.ls"; + paddle::framework::AttributeMap attrs; + attrs.insert({"file_path", std::string(filename)}); + + // Run the save_combine_op + auto save_combine_op = paddle::framework::OpRegistry::CreateOp( + "save_combine", + {{"X", {"test_var1", "test_var2", "test_var3", "test_var4"}}}, {}, attrs); + save_combine_op->Run(scope, place); + + // Set up output vars + auto load_var1 = scope.Var("out_var1"); + auto load_var2 = scope.Var("out_var2"); + auto load_var3 = scope.Var("out_var3"); + auto load_var4 = scope.Var("out_var4"); + + attrs.insert({"load_as_fp16", true}); + // Run the load_combine_op + auto load_combine_op = paddle::framework::OpRegistry::CreateOp( + "load_combine", {}, + {{"Out", {"out_var1", "out_var2", "out_var3", "out_var4"}}}, attrs); + load_combine_op->Run(scope, place); + + auto* target1 = load_var1->GetMutable(); + auto* target2 = load_var2->GetMutable(); + auto* target3 = load_var3->GetMutable(); + auto* target4 = load_var4->GetMutable(); + + paddle::framework::LoD actual_lod1, actual_lod2, actual_lod3, actual_lod4; + paddle::platform::float16* actual1 = + GetValuesAfterLoadCombineOp(target1, scope, + &actual_lod1); + paddle::platform::float16* actual2 = + GetValuesAfterLoadCombineOp(target2, scope, + &actual_lod2); + paddle::platform::float16* actual3 = + GetValuesAfterLoadCombineOp(target3, scope, + &actual_lod3); + paddle::platform::float16* actual4 = + GetValuesAfterLoadCombineOp(target4, scope, + &actual_lod4); + + CheckValues(expect1, actual1, expect_lod1, + actual_lod1, numel1); + CheckValues(expect2, actual2, expect_lod2, + actual_lod2, numel2); + CheckValues(expect3, actual3, expect_lod3, + actual_lod3, numel3); + CheckValues(expect4, actual4, expect_lod4, + actual_lod4, numel4); } // Test with original SaveLoadTest @@ -141,7 +307,7 @@ TEST(SaveLoadTestWithCombineOp, CPU) { auto var = scope.Var("test_var"); auto tensor = var->GetMutable(); - tensor->Resize({3, 10}); + tensor->Resize({3, 4000}); paddle::framework::LoD expect_lod; expect_lod.resize(1); expect_lod[0].push_back(0); diff --git a/paddle/fluid/operators/save_load_op_test.cc b/paddle/fluid/operators/save_load_op_test.cc index 74385ee47543e3f4887081c2225212996d3df3f1..c4fcc61af4b75e6dc7d5c31e20c5fff358637af5 100644 --- a/paddle/fluid/operators/save_load_op_test.cc +++ b/paddle/fluid/operators/save_load_op_test.cc @@ -63,14 +63,21 @@ TEST(SaveLoadOp, CPU) { } } -TEST(SaveLoadFP16Op, CPU) { +TEST(SaveFP16Op, CPU) { paddle::framework::Scope scope; paddle::platform::CPUPlace place; auto var = scope.Var("test_var"); auto tensor = var->GetMutable(); tensor->Resize({3, 10}); + paddle::framework::LoD expect_lod; + expect_lod.resize(1); + expect_lod[0].push_back(0); + expect_lod[0].push_back(1); + expect_lod[0].push_back(2); + expect_lod[0].push_back(3); + tensor->set_lod(expect_lod); float* expect = tensor->mutable_data(place); for (int64_t i = 0; i < tensor->numel(); ++i) { expect[i] = static_cast(paddle::platform::float16(i)); @@ -93,4 +100,60 @@ TEST(SaveLoadFP16Op, CPU) { for (int64_t i = 0; i < tensor->numel(); ++i) { EXPECT_EQ(expect[i], static_cast(actual[i])); } + auto& actual_lod = target->lod(); + EXPECT_EQ(expect_lod.size(), actual_lod.size()); + for (size_t i = 0; i < expect_lod.size(); ++i) { + for (size_t j = 0; j < expect_lod[i].size(); ++j) { + EXPECT_EQ(expect_lod[i][j], actual_lod[i][j]); + } + } +} + +TEST(LoadFP16Op, CPU) { + paddle::framework::Scope scope; + paddle::platform::CPUPlace place; + + auto var = scope.Var("test_var"); + auto tensor = var->GetMutable(); + tensor->Resize({3, 10}); + + paddle::framework::LoD expect_lod; + expect_lod.resize(1); + expect_lod[0].push_back(0); + expect_lod[0].push_back(1); + expect_lod[0].push_back(2); + expect_lod[0].push_back(3); + + tensor->set_lod(expect_lod); + float* expect = tensor->mutable_data(place); + for (int64_t i = 0; i < tensor->numel(); ++i) { + expect[i] = static_cast(paddle::platform::float16(i)); + } + + paddle::framework::AttributeMap attrs; + attrs.insert({"file_path", std::string("tensor.save")}); + attrs.insert({"load_as_fp16", true}); + + auto save_op = paddle::framework::OpRegistry::CreateOp( + "save", {{"X", {"test_var"}}}, {}, attrs); + save_op->Run(scope, place); + + auto load_var = scope.Var("out_var"); + auto load_op = paddle::framework::OpRegistry::CreateOp( + "load", {}, {{"Out", {"out_var"}}}, attrs); + load_op->Run(scope, place); + + auto target = load_var->Get(); + paddle::platform::float16* actual = target.data(); + for (int64_t i = 0; i < tensor->numel(); ++i) { + EXPECT_EQ(expect[i], static_cast(actual[i])); + } + + auto& actual_lod = target.lod(); + EXPECT_EQ(expect_lod.size(), actual_lod.size()); + for (size_t i = 0; i < expect_lod.size(); ++i) { + for (size_t j = 0; j < expect_lod[i].size(); ++j) { + EXPECT_EQ(expect_lod[i][j], actual_lod[i][j]); + } + } } diff --git a/paddle/fluid/operators/save_op.cc b/paddle/fluid/operators/save_op.cc index dcc1b9ec204e9e273b8fd2b12f2423fc989ba502..e6d27e2dedd7668b93bd8ddc330a897d1c6fa732 100644 --- a/paddle/fluid/operators/save_op.cc +++ b/paddle/fluid/operators/save_op.cc @@ -117,8 +117,7 @@ class SaveOp : public framework::OperatorBase { class SaveOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - SaveOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor ) Input tensor to be saved"); AddComment(R"DOC( Save operator diff --git a/paddle/fluid/operators/scale_op.cc b/paddle/fluid/operators/scale_op.cc index 7dcf33c989c3bcd905da8017ee36ec8ce8032911..4687e21e7155fc7309fb28c881c0d47152df9ad5 100644 --- a/paddle/fluid/operators/scale_op.cc +++ b/paddle/fluid/operators/scale_op.cc @@ -37,8 +37,7 @@ class ScaleOp : public framework::OperatorWithKernel { class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { public: - ScaleOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) Input tensor of scale operator."); AddOutput("Out", "(Tensor) Output tensor of scale operator."); AddComment(R"DOC( diff --git a/paddle/fluid/operators/scatter_op.cc b/paddle/fluid/operators/scatter_op.cc index 95b12455ea4996f00bab8a353ccd425b2c37aed1..bf5e0d864495ce3a651a31c9d5a7664fe9eb2396 100644 --- a/paddle/fluid/operators/scatter_op.cc +++ b/paddle/fluid/operators/scatter_op.cc @@ -78,8 +78,7 @@ class ScatterGradOp : public framework::OperatorWithKernel { class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { public: - ScatterOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The source input of scatter op"); AddInput("Ids", "The index input of scatter op where X will be updated"); AddInput("Updates", "The updated value of updates op"); diff --git a/paddle/fluid/operators/select_op.cc b/paddle/fluid/operators/select_op.cc index 876d8acf0d880a7ef806514014d297f98e04c53d..e71841d4d1815d50cd9800910c9db34e121beffc 100644 --- a/paddle/fluid/operators/select_op.cc +++ b/paddle/fluid/operators/select_op.cc @@ -380,8 +380,7 @@ class SelectOp : public framework::OperatorBase { class SelectOpMaker : public framework::OpProtoAndCheckerMaker { public: - SelectOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kX, "A set of variables, which are required by operators inside the " "cases of Select Op") diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc index 12b844daaa33162b86b7daffa2e4c49785701662..1ce0907f3a9473e37f53bf7b2d42cddcb629dfa6 100644 --- a/paddle/fluid/operators/send_barrier_op.cc +++ b/paddle/fluid/operators/send_barrier_op.cc @@ -57,8 +57,7 @@ class SendBarrierOp : public framework::OperatorBase { class SendBarrierOpMaker : public framework::OpProtoAndCheckerMaker { public: - SendBarrierOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddOutput("RPCClient", "(RPCClient) The RPC client object which is" "initialized at most once."); diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc index e4386b640a298cd216bb60104653f20c4a96e7dc..95bb1f3c695297e6d8134a647925310207118a9b 100644 --- a/paddle/fluid/operators/send_op.cc +++ b/paddle/fluid/operators/send_op.cc @@ -92,8 +92,7 @@ class SendOp : public framework::OperatorBase { class SendOpMaker : public framework::OpProtoAndCheckerMaker { public: - SendOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("X", "(Tensor) Input tensor to be sent").AsDuplicable(); AddOutput("Out", "(Tensor) Output tensor to be received from server") .AsDuplicable(); diff --git a/paddle/fluid/operators/send_recv_op_test.cc b/paddle/fluid/operators/send_recv_op_test.cc index eb51f301bfe2a97c65dd1fec23ff5a44f3843b05..d5303eaf50722234d205264e56892b1723104d53 100644 --- a/paddle/fluid/operators/send_recv_op_test.cc +++ b/paddle/fluid/operators/send_recv_op_test.cc @@ -92,12 +92,16 @@ void InitSelectedRowsInScope(const p::CPUPlace &place, f::Scope *scope) { void AddOp(const std::string &type, const f::VariableNameMap &inputs, const f::VariableNameMap &outputs, f::AttributeMap attrs, - f::BlockDesc *block) { + f::BlockDesc *block, bool is_sparse) { // insert output for (auto kv : outputs) { for (auto v : kv.second) { auto var = block->Var(v); var->SetDataType(f::proto::VarType::FP32); + var->SetPersistable(true); + if (is_sparse) { + var->SetType(f::proto::VarType::SELECTED_ROWS); + } } } @@ -128,7 +132,8 @@ void StartServerNet(bool is_sparse, std::atomic *initialized) { auto *optimize_block = program.AppendBlock(root_block); auto *prefetch_block = program.AppendBlock(root_block); // X for server side tensors, RX for received tensors, must be of same shape. - AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, optimize_block); + AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, optimize_block, + is_sparse); f::AttributeMap attrs; attrs.insert({"endpoint", std::string("127.0.0.1:0")}); attrs.insert({"Fanin", 1}); diff --git a/paddle/fluid/operators/send_vars_op.cc b/paddle/fluid/operators/send_vars_op.cc index 56b3713d6af28d0787e114a672a503e86cbd85fd..f11e84c176ae97dff0fda560ce3ebe2ab72c7bcc 100644 --- a/paddle/fluid/operators/send_vars_op.cc +++ b/paddle/fluid/operators/send_vars_op.cc @@ -66,8 +66,7 @@ class SendVarsOp : public framework::OperatorBase { class SendVarsOpMaker : public framework::OpProtoAndCheckerMaker { public: - SendVarsOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() { AddInput("X", "(Tensor, SelectedRows) Input variables to be sent") .AsDuplicable(); AddOutput("RPCClient", diff --git a/paddle/fluid/operators/sequence_concat_op.cc b/paddle/fluid/operators/sequence_concat_op.cc index 3c21903e3a08dcfb55c6c07370a117d0ad633e69..077b9a5f7d935a39706ef3c2b710522bf1b713ed 100644 --- a/paddle/fluid/operators/sequence_concat_op.cc +++ b/paddle/fluid/operators/sequence_concat_op.cc @@ -43,8 +43,7 @@ class SequenceConcatOp : public framework::OperatorWithKernel { class SequenceConcatOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceConcatOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LodTensorArray) Input is a vector of LoDTensor, " "each of which is a variable-length sequence or nested sequence.") diff --git a/paddle/fluid/operators/sequence_conv_op.cc b/paddle/fluid/operators/sequence_conv_op.cc index 94f4b49b0018fdbff6e67c3c081aa5706ccb2e66..ec6cb24350ae276724aae339590d40be1e9ea400 100644 --- a/paddle/fluid/operators/sequence_conv_op.cc +++ b/paddle/fluid/operators/sequence_conv_op.cc @@ -102,8 +102,7 @@ class SequenceConvGradOp : public framework::OperatorWithKernel { class SequenceConvOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceConvOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(LoDTensor) the input(X) is a LodTensor, which supports " diff --git a/paddle/fluid/operators/sequence_erase_op.cc b/paddle/fluid/operators/sequence_erase_op.cc index 73c0e89512972cda002bd902ee0c78b4b77d8502..1c86486157a02c3b78ed61e840fd8e452b9cb452 100644 --- a/paddle/fluid/operators/sequence_erase_op.cc +++ b/paddle/fluid/operators/sequence_erase_op.cc @@ -37,8 +37,7 @@ class SequenceEraseOp : public framework::OperatorWithKernel { class SequenceEraseOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceEraseOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(2-D LoDTensor with the 2nd dim. equal to 1) " "Input LoDTensor of SequenceEraseOp."); diff --git a/paddle/fluid/operators/sequence_expand_op.cc b/paddle/fluid/operators/sequence_expand_op.cc index 84a35d7172a567a3f6505559fa45a32290288533..944c7f85e5f43679e1875fcce813382be2ba5526 100644 --- a/paddle/fluid/operators/sequence_expand_op.cc +++ b/paddle/fluid/operators/sequence_expand_op.cc @@ -94,8 +94,7 @@ class SequenceExpandOp : public framework::OperatorWithKernel { class SequenceExpandOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor, default LoDTensor) A 2-D LoDTensor whose lod " "level is at most 1."); diff --git a/paddle/fluid/operators/sequence_pool_op.cc b/paddle/fluid/operators/sequence_pool_op.cc index 933c8c26239d49221819a583f999389ed6fb6cb6..5c6fd13d42e43e3502a1cab85a56e019420c708d 100644 --- a/paddle/fluid/operators/sequence_pool_op.cc +++ b/paddle/fluid/operators/sequence_pool_op.cc @@ -38,8 +38,7 @@ class SequencePoolOp : public framework::OperatorWithKernel { class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequencePoolOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) The variable-length input of SequencePoolOp"); AddOutput("Out", "(Tensor) The output of SequencePoolOp does not contain LoD " diff --git a/paddle/fluid/operators/sequence_reshape_op.cc b/paddle/fluid/operators/sequence_reshape_op.cc index a2999650b8903f9d819a8e8011421349e098b219..ef5e6f3210234d59298fcf04c812390643c693d0 100644 --- a/paddle/fluid/operators/sequence_reshape_op.cc +++ b/paddle/fluid/operators/sequence_reshape_op.cc @@ -42,8 +42,7 @@ class SequenceReshapeOp : public framework::OperatorWithKernel { class SequenceReshapeOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceReshapeOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor, default LoDTensor) A 2-D LoDTensor with shape " "being [N, M]."); diff --git a/paddle/fluid/operators/sequence_slice_op.cc b/paddle/fluid/operators/sequence_slice_op.cc index 7cd620af07fa9b5f8fcee3c0f88207ef2800c4a1..df9243dc04c584d70dfa6ca78d5fac8423796466 100644 --- a/paddle/fluid/operators/sequence_slice_op.cc +++ b/paddle/fluid/operators/sequence_slice_op.cc @@ -79,8 +79,7 @@ class SequenceSliceGradOp : public framework::OperatorWithKernel { class SequenceSliceOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceSliceOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor), " "the input of SequenceSliceOp."); diff --git a/paddle/fluid/operators/sequence_softmax_op.cc b/paddle/fluid/operators/sequence_softmax_op.cc index a0d47c12ba606eb62bbbea4d5ea793ce915e8100..c44f8206eb5079fef969e3e527552512eebd0f1a 100644 --- a/paddle/fluid/operators/sequence_softmax_op.cc +++ b/paddle/fluid/operators/sequence_softmax_op.cc @@ -57,8 +57,7 @@ class SequenceSoftmaxOp : public framework::OperatorWithKernel { class SequenceSoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { public: - SequenceSoftmaxOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) 1-D or 2-D input LoDTensor with the 2-nd dimension " "of length 1."); diff --git a/paddle/fluid/operators/sgd_op.cc b/paddle/fluid/operators/sgd_op.cc index bd04c60ffa5c1e5eb8d2051ce495ab6c685b14b5..7a2bdeac09d61603f437ff10d58d0542bb3c3689 100644 --- a/paddle/fluid/operators/sgd_op.cc +++ b/paddle/fluid/operators/sgd_op.cc @@ -68,8 +68,7 @@ class SGDOpInferVarType : public framework::VarTypeInference { class SGDOpMaker : public framework::OpProtoAndCheckerMaker { public: - SGDOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Param", "(Tensor or SelectedRows) Input parameter"); AddInput("LearningRate", "(Tensor) Learning rate of SGD"); AddInput("Grad", "(Tensor or SelectedRows) Input gradient"); diff --git a/paddle/fluid/operators/shrink_rnn_memory_op.cc b/paddle/fluid/operators/shrink_rnn_memory_op.cc index a1871a8e7fb27d351f9d333966baa63c6f32ae01..8146c5f56104b7dec86b1c4491ed10fc2e94b58b 100644 --- a/paddle/fluid/operators/shrink_rnn_memory_op.cc +++ b/paddle/fluid/operators/shrink_rnn_memory_op.cc @@ -69,8 +69,7 @@ class ShrinkRNNMemoryOp : public ArrayOp { class ShrinkRNNMemoryOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ShrinkRNNMemoryOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) The RNN step memory to be shrinked."); AddInput("RankTable", "(LoDRankTable) The lod_rank_table of dynamic RNN."); AddInput("I", diff --git a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc index 5db77d0493fc0abaa0a696cb559c3ca0534d4101..135e2a6f7f877c9ef159a4542b834d5627649e81 100644 --- a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc +++ b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc @@ -86,9 +86,7 @@ class SigmoidCrossEntropyWithLogitsGradOp class SigmoidCrossEntropyWithLogitsOpMaker : public framework::OpProtoAndCheckerMaker { public: - SigmoidCrossEntropyWithLogitsOpMaker(OpProto* proto, - OpAttrChecker* op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor, default Tensor), a 2-D tensor with shape N x D, " "where N is the batch size and D is the number of classes. " diff --git a/paddle/fluid/operators/sign_op.cc b/paddle/fluid/operators/sign_op.cc index 8f8b7abd03212c12ca351e551621e63b4c7148c2..f3985dcc027f974e0213a73ea9a21e268d77615f 100644 --- a/paddle/fluid/operators/sign_op.cc +++ b/paddle/fluid/operators/sign_op.cc @@ -34,8 +34,7 @@ class SignOp : public framework::OperatorWithKernel { template class SignOpMaker : public framework::OpProtoAndCheckerMaker { public: - SignOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) Input tensor of sign operator."); AddOutput("Out", "(Tensor) Output tensor of sign operator."); AddComment(R"DOC( diff --git a/paddle/fluid/operators/smooth_l1_loss_op.cc b/paddle/fluid/operators/smooth_l1_loss_op.cc index 322581fdef27b12a06704abc9c3b8772adf002f2..622420c1c33a62994c81ad9534c4fa37a4a1fa1a 100644 --- a/paddle/fluid/operators/smooth_l1_loss_op.cc +++ b/paddle/fluid/operators/smooth_l1_loss_op.cc @@ -46,8 +46,7 @@ class SmoothL1LossOp : public framework::OperatorWithKernel { class SmoothL1LossOpMaker : public framework::OpProtoAndCheckerMaker { public: - SmoothL1LossOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor, default Tensor) A tensor with rank at least 2. " "The input value of smooth l1 loss op with shape " @@ -106,7 +105,7 @@ class SmoothL1LossGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { - auto in_dims = ctx->GetInputDim("X"); + auto in_dims = ctx->GetInputDim("Diff"); auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); PADDLE_ENFORCE_GE(out_dims.size(), 2, @@ -128,12 +127,33 @@ class SmoothL1LossGradOp : public framework::OperatorWithKernel { } }; +class SmoothL1LossGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* op = new framework::OpDesc(); + op->SetType("smooth_l1_loss_grad"); + op->SetInput("InsideWeight", Input("InsideWeight")); + op->SetInput("OutsideWeight", Input("OutsideWeight")); + op->SetInput("Diff", Output("Diff")); + op->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); + + op->SetAttrMap(Attrs()); + + op->SetOutput(framework::GradVarName("X"), InputGrad("X")); + op->SetOutput(framework::GradVarName("Y"), InputGrad("Y")); + return std::unique_ptr(op); + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OPERATOR(smooth_l1_loss, ops::SmoothL1LossOp, ops::SmoothL1LossOpMaker, - paddle::framework::DefaultGradOpDescMaker); + ops::SmoothL1LossGradMaker); REGISTER_OPERATOR(smooth_l1_loss_grad, ops::SmoothL1LossGradOp); REGISTER_OP_CPU_KERNEL( smooth_l1_loss, diff --git a/paddle/fluid/operators/softmax_mkldnn_op.cc b/paddle/fluid/operators/softmax_mkldnn_op.cc index 71b541d98f6e0d3e12601c9988ca6ffb8bb7554d..14b57b11fefb2b726531cb164dbf479f8df26b24 100644 --- a/paddle/fluid/operators/softmax_mkldnn_op.cc +++ b/paddle/fluid/operators/softmax_mkldnn_op.cc @@ -53,25 +53,60 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel { "Softmax input and output dimensions should match"); // Same memory descriptor to be used for input and output memory::dims softmax_tz = {src_tz[0], src_tz[1]}; - // Currently only supports NC data format - // TODO(jczaja-intel): support more formats - auto softmax_md = - MKLDNNMemDesc({softmax_tz}, memory::f32, memory::format::nc); - // Normalization is made after innermost dimension eg. C out of NC - auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring, - softmax_md, 1 /*dim: C*/); - // create memory primitives - auto softmax_src_memory = - memory({softmax_md, mkldnn_engine}, - static_cast(const_cast(input_data))); - auto softmax_dst_memory = - memory({softmax_md, mkldnn_engine}, - static_cast(const_cast(output_data))); - auto softmax_prim_desc = - softmax_forward::primitive_desc(softmax_desc, mkldnn_engine); - auto softmax = softmax_forward(softmax_prim_desc, softmax_src_memory, - softmax_dst_memory); - std::vector pipeline{softmax}; + // Generate keys for storing/retriving primitives for this operator + // TODO(jczaja): Each MKLDNN operator may have diffrent hashing function + auto gethash = [](memory::dims& operand_dims) { + return std::string(std::to_string(operand_dims[0]) + "-" + + std::to_string(operand_dims[1])); + }; + const std::string key = gethash(softmax_tz); + const std::string key_softmax_p = key + "@softmax_p"; + const std::string key_softmax_src_mem_p = key + "@softmax_src_mem_p"; + const std::string key_softmax_dst_mem_p = key + "@softmax_dst_mem_p"; + + std::shared_ptr softmax_p = dev_ctx.GetBlob(key_softmax_p); + if (softmax_p == nullptr) { + // Currently only NC data format is supported + auto softmax_md = + MKLDNNMemDesc({softmax_tz}, memory::f32, memory::format::nc); + // Normalization is made after innermost dimension eg. C out of NC + auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring, + softmax_md, 1 /*dim: C*/); + // create memory primitives + auto softmax_src_memory_p = std::make_shared( + memory::primitive_desc{softmax_md, mkldnn_engine}, + static_cast(const_cast(input_data))); + dev_ctx.SetBlob(key_softmax_src_mem_p, softmax_src_memory_p); + auto softmax_dst_memory_p = std::make_shared( + memory::primitive_desc{softmax_md, mkldnn_engine}, + static_cast(output_data)); + dev_ctx.SetBlob(key_softmax_dst_mem_p, softmax_dst_memory_p); + + auto softmax_forward_pd = + std::make_shared(softmax_desc, + mkldnn_engine); + softmax_p = std::make_shared( + *(softmax_forward_pd.get()), + *(static_cast(softmax_src_memory_p.get())), + *(static_cast(softmax_dst_memory_p.get()))); + dev_ctx.SetBlob(key_softmax_p, softmax_p); + } else { + // Primitives already exist + auto src_memory_p = std::static_pointer_cast( + dev_ctx.GetBlob(key_softmax_src_mem_p)); + PADDLE_ENFORCE(src_memory_p != nullptr, + "Fail to find softmax src mem_p in device context"); + auto dst_memory_p = std::static_pointer_cast( + dev_ctx.GetBlob(key_softmax_dst_mem_p)); + PADDLE_ENFORCE(dst_memory_p != nullptr, + "Fail to find softmax dst mem_p in device context"); + src_memory_p->set_data_handle( + reinterpret_cast(const_cast(input_data))); + dst_memory_p->set_data_handle(output_data); + } + + std::vector pipeline{ + *(static_cast(softmax_p.get()))}; stream(stream::kind::eager).submit(pipeline).wait(); const bool is_test = ctx.Attr("is_test"); diff --git a/paddle/fluid/operators/softmax_op.cc b/paddle/fluid/operators/softmax_op.cc index aa7b192e327704c02a26c86cc208ebe8a5cd7ba5..cc256aa627bdda0609f496cab93a2dec7d95f348 100644 --- a/paddle/fluid/operators/softmax_op.cc +++ b/paddle/fluid/operators/softmax_op.cc @@ -77,8 +77,7 @@ class SoftmaxOp : public framework::OperatorWithKernel { class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftmaxOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input tensor of softmax. " "2-D with shape [batch_size, input_feature_dimensions]."); diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.cc b/paddle/fluid/operators/softmax_with_cross_entropy_op.cc index 857e5733573497b56520daa7860f4feb4e01cda7..53cb716a979229c99fcbdc12f1aeab4e21b320f3 100644 --- a/paddle/fluid/operators/softmax_with_cross_entropy_op.cc +++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.cc @@ -20,8 +20,7 @@ namespace operators { class SoftmaxWithCrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker { public: - SoftmaxWithCrossEntropyOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Logits", "(Tensor, default: Tensor), The unscaled log probabilities " "which is a 2-D tensor with shape [N x K]. N is the batch_size, " diff --git a/paddle/fluid/operators/split_byref_op.cc b/paddle/fluid/operators/split_byref_op.cc index 7413ce3e9ce60ed733bb4d27e9ec205e5f0a7e1b..bc998e1abbd7131a7497288cc9d66315a6fedc85 100644 --- a/paddle/fluid/operators/split_byref_op.cc +++ b/paddle/fluid/operators/split_byref_op.cc @@ -64,8 +64,7 @@ class SplitByrefOp : public framework::OperatorWithKernel { class SplitByrefOpMaker : public framework::OpProtoAndCheckerMaker { public: - SplitByrefOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) Input tensor of the split operator."); AddOutput("Out", "(Tensor) Output tensors of the split operator.") .AsDuplicable(); diff --git a/paddle/fluid/operators/split_ids_op.cc b/paddle/fluid/operators/split_ids_op.cc index a53cbc8ac5199061dafdc7f4cf560b9e4fc577ab..c867c46873ae7ddbdbda280351e4ab28235bcc08 100644 --- a/paddle/fluid/operators/split_ids_op.cc +++ b/paddle/fluid/operators/split_ids_op.cc @@ -19,8 +19,7 @@ namespace operators { class SplitIdsOpMaker : public framework::OpProtoAndCheckerMaker { public: - SplitIdsOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}"); AddOutput("Out", "(LoDTensor) The outputs of the input Ids.") .AsDuplicable(); diff --git a/paddle/fluid/operators/split_lod_tensor_op.cc b/paddle/fluid/operators/split_lod_tensor_op.cc index 3222cce239988b170501f2b99e9f1253036b7fbc..767449cde981e5925b7144ff1038560c67651f3e 100644 --- a/paddle/fluid/operators/split_lod_tensor_op.cc +++ b/paddle/fluid/operators/split_lod_tensor_op.cc @@ -125,8 +125,7 @@ class SplitLoDTensorOp : public framework::OperatorBase { class SplitLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - SplitLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input LoDTensor"); AddInput("Mask", "A bool column vector which mask the input"); AddOutput("OutTrue", "True branch of input LoDTensor"); diff --git a/paddle/fluid/operators/split_op.cc b/paddle/fluid/operators/split_op.cc index a4398df36bcc2d3b8bbe8949f27f5d6508861d95..5e2b2a994534c2fb1e053c067b36651d358b9da8 100644 --- a/paddle/fluid/operators/split_op.cc +++ b/paddle/fluid/operators/split_op.cc @@ -70,8 +70,7 @@ class SplitOp : public framework::OperatorWithKernel { class SplitOpMaker : public framework::OpProtoAndCheckerMaker { public: - SplitOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) Input tensor of the split operator."); AddOutput("Out", "(Tensor) Output tensors of the split operator.") .AsDuplicable(); diff --git a/paddle/fluid/operators/split_selected_rows_op.cc b/paddle/fluid/operators/split_selected_rows_op.cc index e1ce3d0c1bf11e9a623e4e9adc8f08f5069f4d94..76615a9405d7a8e3fa9dba8d01a956209e02ae8f 100644 --- a/paddle/fluid/operators/split_selected_rows_op.cc +++ b/paddle/fluid/operators/split_selected_rows_op.cc @@ -19,8 +19,7 @@ namespace operators { class SplitSelectedRowsOpMaker : public framework::OpProtoAndCheckerMaker { public: - SplitSelectedRowsOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "The input SelectedRows."); AddOutput("Out", "The outputs of the input SelectedRows.").AsDuplicable(); AddAttr>("height_sections", diff --git a/paddle/fluid/operators/spp_op.cc b/paddle/fluid/operators/spp_op.cc index 1cada95501a76da27081d533b451ce7f6a384a49..a2a96b72f09df86790ad1f90ead9189ff9bd581c 100644 --- a/paddle/fluid/operators/spp_op.cc +++ b/paddle/fluid/operators/spp_op.cc @@ -20,8 +20,7 @@ namespace operators { class SppOpMaker : public framework::OpProtoAndCheckerMaker { public: - SppOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(Tensor) The input tensor of spp operator. " diff --git a/paddle/fluid/operators/squared_l2_distance_op.cc b/paddle/fluid/operators/squared_l2_distance_op.cc index c32f575b541d6a6441cc1b6e999496eacef421a5..42532a294b2ef9ffdb240fac8596278047daf7fe 100644 --- a/paddle/fluid/operators/squared_l2_distance_op.cc +++ b/paddle/fluid/operators/squared_l2_distance_op.cc @@ -56,8 +56,7 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel { class SquaredL2DistanceOpMaker : public framework::OpProtoAndCheckerMaker { public: - SquaredL2DistanceOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) Input of SquaredL2DistanceOp."); AddInput("Y", "(Tensor) Target of SquaredL2DistanceOp."); AddOutput("sub_result", diff --git a/paddle/fluid/operators/squared_l2_norm_op.cc b/paddle/fluid/operators/squared_l2_norm_op.cc index 4ce51259da3530367d91b5da34f06fbe5d969fce..7bd82e0ce4add6d4434e1defaee43da178a6f309 100644 --- a/paddle/fluid/operators/squared_l2_norm_op.cc +++ b/paddle/fluid/operators/squared_l2_norm_op.cc @@ -48,8 +48,7 @@ class SquaredL2NormGradOp : public framework::OperatorWithKernel { class SquaredL2NormOpMaker : public framework::OpProtoAndCheckerMaker { public: - SquaredL2NormOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input of squared_l2_norm op."); AddOutput("Out", "(Scalar) The output of squared_l2_norm op."); AddComment(R"DOC( diff --git a/paddle/fluid/operators/sum_op.cc b/paddle/fluid/operators/sum_op.cc index 108f26fafe7af76eaa613d77ed77748ee43ea234..bcc5e22d4a77349e7cde9a43b83f23d4c867d994 100644 --- a/paddle/fluid/operators/sum_op.cc +++ b/paddle/fluid/operators/sum_op.cc @@ -112,8 +112,7 @@ class SumOp : public framework::OperatorWithKernel { class SumOpMaker : public framework::OpProtoAndCheckerMaker { public: - SumOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(vector) The input tensors of sum operator.") .AsDuplicable(); AddOutput("Out", "(Tensor) The output tensor of sum operator."); diff --git a/paddle/fluid/operators/tensor_array_read_write_op.cc b/paddle/fluid/operators/tensor_array_read_write_op.cc index 2636812c42985536e7ca3475c03bbd8d1638ece6..c703d11eeccf8418250f00c801f47418ee9c85ae 100644 --- a/paddle/fluid/operators/tensor_array_read_write_op.cc +++ b/paddle/fluid/operators/tensor_array_read_write_op.cc @@ -57,8 +57,7 @@ class WriteToArrayOp : public ArrayOp { class WriteToArrayOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: - WriteToArrayOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(LoDTensor) the tensor will be written to tensor array"); AddInput( "I", @@ -148,8 +147,7 @@ class ReadFromArrayOp : public ArrayOp { class ReadFromArrayProtoMaker : public framework::OpProtoAndCheckerMaker { public: - ReadFromArrayProtoMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(TensorArray) the array will be read from."); AddInput("I", "(Tensor) the subscript index in tensor array. The number of " diff --git a/paddle/fluid/operators/test_send_nccl_id.cc b/paddle/fluid/operators/test_send_nccl_id.cc new file mode 100644 index 0000000000000000000000000000000000000000..bbae1d54aa3524fd45cb8ab13c86df8d54b8e643 --- /dev/null +++ b/paddle/fluid/operators/test_send_nccl_id.cc @@ -0,0 +1,94 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include // NOLINT + +#include "gtest/gtest.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/listen_and_serv_op.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" +#include "paddle/fluid/platform/nccl_helper.h" +#include "paddle/fluid/string/printf.h" + +USE_NO_KERNEL_OP(listen_and_serv); + +namespace f = paddle::framework; +namespace p = paddle::platform; +namespace m = paddle::operators::math; +namespace detail = paddle::operators::detail; +namespace string = paddle::string; + +std::unique_ptr rpc_service; + +void StartServer(std::atomic* initialized) { + f::Scope scope; + p::CPUPlace place; + scope.Var(NCCL_ID_VARNAME); + p::DeviceContextPool& pool = p::DeviceContextPool::Instance(); + auto& dev_ctx = *pool.Get(p::CPUPlace()); + + rpc_service.reset(new detail::AsyncGRPCServer("127.0.0.1:0", true)); + + f::ProgramDesc empty_program; + f::Executor executor(dev_ctx.GetPlace()); + rpc_service->SetScope(&scope); + rpc_service->SetDevCtx(&dev_ctx); + rpc_service->SetProgram(&empty_program); + rpc_service->SetExecutor(&executor); + + std::thread server_thread( + std::bind(&detail::AsyncGRPCServer::RunSyncUpdate, rpc_service.get())); + *initialized = true; + rpc_service->SetCond(0); + auto recv = rpc_service->Get(); + LOG(INFO) << "got nccl id and stop server..."; + rpc_service->ShutDown(); + server_thread.join(); +} + +TEST(SendNcclId, Normal) { + std::atomic initialized{false}; + std::thread server_thread(StartServer, &initialized); + while (!initialized) { + } + // wait server to start + // sleep(2); + rpc_service->WaitServerReady(); + + f::Scope scope; + p::CPUPlace place; + p::DeviceContextPool& pool = p::DeviceContextPool::Instance(); + auto& dev_ctx = *pool.Get(p::CPUPlace()); + + auto var = scope.Var(NCCL_ID_VARNAME); + // var->SetType(f::proto::VarType_Type_RAW); + auto id = var->GetMutable(); + p::dynload::ncclGetUniqueId(id); + + int port = rpc_service->GetSelectedPort(); + std::string ep = string::Sprintf("127.0.0.1:%d", port); + detail::RPCClient client; + + client.AsyncSendVariable(ep, dev_ctx, scope, NCCL_ID_VARNAME); + client.Wait(); + server_thread.join(); + auto* ptr = rpc_service.release(); + delete ptr; +} diff --git a/paddle/fluid/operators/top_k_op.cc b/paddle/fluid/operators/top_k_op.cc index 942a5de3f90f20eabe691924a570b61509eccf76..c17d1afc309c65035063348d4934ea1783b018ed 100644 --- a/paddle/fluid/operators/top_k_op.cc +++ b/paddle/fluid/operators/top_k_op.cc @@ -48,8 +48,7 @@ class TopkOp : public framework::OperatorWithKernel { class TopkOpMaker : public framework::OpProtoAndCheckerMaker { public: - TopkOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("X", "(Tensor) The input of Topk op"); AddOutput("Out", "(Tensor) The output tensor of Topk op"); AddOutput("Indices", "(Tensor) The indices of Topk elements of input"); diff --git a/paddle/fluid/operators/transpose_op.cc b/paddle/fluid/operators/transpose_op.cc index 3555cb68cab97c0cf983f1173c3b4ca9307e4f7d..60556a564c25c08612447ebd47a4b432b8a12d29 100644 --- a/paddle/fluid/operators/transpose_op.cc +++ b/paddle/fluid/operators/transpose_op.cc @@ -56,8 +56,7 @@ class TransposeOp : public framework::OperatorWithKernel { class TransposeOpMaker : public framework::OpProtoAndCheckerMaker { public: - TransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(Tensor) The input tensor, tensors with rank up to 6 are supported."); diff --git a/paddle/fluid/operators/uniform_random_batch_size_like_op.cc b/paddle/fluid/operators/uniform_random_batch_size_like_op.cc index 00f00bb403db5e40939a1502b2219fb4d36d58e5..78fee77df8151221459b0afa0d6789bfe82cfda5 100644 --- a/paddle/fluid/operators/uniform_random_batch_size_like_op.cc +++ b/paddle/fluid/operators/uniform_random_batch_size_like_op.cc @@ -32,9 +32,8 @@ class UniformRandomBatchSizeLikeOp : public BatchSizeLikeOp { }; class UniformRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { - public: - UniformRandomBatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : BatchSizeLikeOpMaker(proto, op_checker) { + protected: + void Apply() override { AddComment(R"DOC( Uniform random operator diff --git a/paddle/fluid/operators/uniform_random_op.cc b/paddle/fluid/operators/uniform_random_op.cc index 3b5cf68dd4f28d23e507058337fe55de9b88d3cd..137ea91caedabc3167146d91b063dbe9e2e2b931 100644 --- a/paddle/fluid/operators/uniform_random_op.cc +++ b/paddle/fluid/operators/uniform_random_op.cc @@ -85,8 +85,7 @@ class UniformRandomOp : public framework::OperatorWithKernel { class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker { public: - UniformRandomOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : framework::OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddOutput("Out", "(Tensor) The output tensor of uniform random op"); AddComment(R"DOC( Uniform random operator. diff --git a/paddle/fluid/operators/unpool_op.cc b/paddle/fluid/operators/unpool_op.cc index b3cd87efa21115565b32659cb35fee4b5bed2d4f..1d441b43b14ea194152095874645f8133c423efd 100644 --- a/paddle/fluid/operators/unpool_op.cc +++ b/paddle/fluid/operators/unpool_op.cc @@ -20,8 +20,7 @@ namespace operators { class Unpool2dOpMaker : public framework::OpProtoAndCheckerMaker { public: - Unpool2dOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput( "X", "(Tensor) The input tensor of unpool operator. " diff --git a/paddle/fluid/operators/warpctc_op.cc b/paddle/fluid/operators/warpctc_op.cc index 6835a5dd6286ece20c4ce6f3e951ed4b0057012c..e06c8c962f45a4e91b7efed7431571f0fc6870a3 100644 --- a/paddle/fluid/operators/warpctc_op.cc +++ b/paddle/fluid/operators/warpctc_op.cc @@ -53,8 +53,7 @@ class WarpCTCOp : public framework::OperatorWithKernel { class WarpCTCOpMaker : public framework::OpProtoAndCheckerMaker { public: - WarpCTCOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput("Logits", "(LodTensor, default: LoDTensor), the unscaled " "probabilities of variable-length sequences, which is a 2-D " diff --git a/paddle/fluid/operators/while_op.cc b/paddle/fluid/operators/while_op.cc index 710cc9fc2e716da2e4fd067562a34d312e48b1a1..175c3ac5d79f24e47d21417df8e3eaeb4d5b2335 100644 --- a/paddle/fluid/operators/while_op.cc +++ b/paddle/fluid/operators/while_op.cc @@ -68,8 +68,7 @@ class WhileOp : public framework::OperatorBase { class WhileOpMaker : public framework::OpProtoAndCheckerMaker { public: - WhileOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { + void Make() override { AddInput(kX, "A set of variables, which are required by operators inside the " "block of While Op.") diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt index 598fd4d419078a973647f2f8f20e8a12c8115a8b..b29035bafd34fa81dc6b59691142fe74439202b8 100644 --- a/paddle/fluid/platform/CMakeLists.txt +++ b/paddle/fluid/platform/CMakeLists.txt @@ -1,4 +1,4 @@ -proto_library(profiler_proto SRCS profiler.proto) +proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto) py_proto_compile(profiler_py_proto SRCS profiler.proto) add_custom_target(profiler_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py) @@ -49,7 +49,7 @@ nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_ nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda) nv_test(transform_test SRCS transform_test.cu DEPS memory place device_context) -cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto ${GPU_CTX_DEPS}) +cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framework_proto ${GPU_CTX_DEPS}) cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer) cc_test(profiler_test SRCS profiler_test.cc DEPS profiler) diff --git a/paddle/fluid/platform/mkldnn_helper.h b/paddle/fluid/platform/mkldnn_helper.h index 23f1d615daab91f0e4b353bc7d9a3ca7f5cec5ae..f1187620d81ff3bc1deef2106edb54d6199fa927 100644 --- a/paddle/fluid/platform/mkldnn_helper.h +++ b/paddle/fluid/platform/mkldnn_helper.h @@ -38,6 +38,11 @@ void* to_void_cast(const Type* t) { return static_cast(const_cast(t)); } +template +void* to_void_reinterpret_cast(const Type* t) { + return reinterpret_cast(const_cast(t)); +} + template using tf_desc = typename Type::desc; @@ -71,5 +76,15 @@ inline bool CanMKLDNNBeUsed(const framework::ExecutionContext& ctx) { return use_mkldnn && platform::is_cpu_place(ctx.GetPlace()); } +template +mkldnn::memory::data_type MKLDNNGetDataType() { + return mkldnn::memory::data_undef; +} + +template <> +inline mkldnn::memory::data_type MKLDNNGetDataType() { + return mkldnn::memory::f32; +} + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/nccl_helper.h b/paddle/fluid/platform/nccl_helper.h index 0013597fd516d15c7d502370eec77e1a6a5dca88..09367889a9517956ad01ad2847c31e2633cc643d 100644 --- a/paddle/fluid/platform/nccl_helper.h +++ b/paddle/fluid/platform/nccl_helper.h @@ -14,12 +14,15 @@ #pragma once +#include #include // NOLINT #include #include #include "paddle/fluid/platform/dynload/nccl.h" #include "paddle/fluid/platform/enforce.h" +#define NCCL_ID_VARNAME "NCCLID" + namespace paddle { namespace platform { @@ -50,7 +53,7 @@ class NCCLGroupGuard { } inline ~NCCLGroupGuard() { - PADDLE_ENFORCE(dynload::ncclGroupEnd()); + CHECK_EQ(dynload::ncclGroupEnd(), ncclSuccess); NCCLMutex().unlock(); } }; @@ -73,7 +76,9 @@ struct NCCLContextMap { std::unordered_map contexts_; std::vector order_; - explicit NCCLContextMap(const std::vector &places) { + explicit NCCLContextMap(const std::vector &places, + ncclUniqueId *nccl_id = nullptr, + size_t num_trainers = 1, size_t trainer_id = 0) { PADDLE_ENFORCE(!places.empty()); order_.reserve(places.size()); for (auto &p : places) { @@ -85,18 +90,34 @@ struct NCCLContextMap { order_.size(), contexts_.size(), "NCCL Context Map does not support contain two or more same device"); - if (places.size() > 1) { - std::unique_ptr comms(new ncclComm_t[order_.size()]); + if (places.size() <= 1) { + return; + } + std::unique_ptr comms(new ncclComm_t[order_.size()]); + // if pass nccl_id here, can assume we are doing multi node training + if (nccl_id == nullptr) { + std::lock_guard guard(NCCLGroupGuard::NCCLMutex()); + PADDLE_ENFORCE(platform::dynload::ncclCommInitAll( + comms.get(), static_cast(order_.size()), order_.data())); + } else { + PADDLE_ENFORCE_GT(num_trainers, 1); + // TODO(wuyi): need to ensure each node have same number of GPUs { - std::lock_guard guard(NCCLGroupGuard::NCCLMutex()); - PADDLE_ENFORCE(platform::dynload::ncclCommInitAll( - comms.get(), static_cast(order_.size()), order_.data())); - } - int i = 0; - for (auto &dev_id : order_) { - contexts_.at(dev_id).comm_ = comms[i++]; + int nranks = num_trainers * order_.size(); + NCCLGroupGuard gurad; + for (auto &gpu_id : order_) { + int rank = trainer_id * order_.size() + gpu_id; + VLOG(3) << "init nccl rank: " << rank << " nranks: " << nranks; + PADDLE_ENFORCE(cudaSetDevice(gpu_id)); + PADDLE_ENFORCE(platform::dynload::ncclCommInitRank( + comms.get() + gpu_id, nranks, *nccl_id, rank)); + } } } + int i = 0; + for (auto &dev_id : order_) { + contexts_.at(dev_id).comm_ = comms[i++]; + } } NCCLContextMap(const NCCLContextMap &other) = delete; diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc index 50bc0aba6aa0f056dc0b2d49f6b3b745433e0756..2fb5c6dc6b8ad25fa1ad5fcf7c2acfedd5be4a83 100644 --- a/paddle/fluid/platform/profiler.cc +++ b/paddle/fluid/platform/profiler.cc @@ -173,8 +173,9 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx) { } RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx) - : start_ns_(PosixInNsec()) { + : is_enabled_(false), start_ns_(PosixInNsec()) { if (g_state == ProfilerState::kDisabled) return; + is_enabled_ = true; dev_ctx_ = dev_ctx; name_ = name; PushEvent(name_, dev_ctx_); @@ -183,7 +184,7 @@ RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx) } RecordEvent::~RecordEvent() { - if (g_state == ProfilerState::kDisabled) return; + if (g_state == ProfilerState::kDisabled || !is_enabled_) return; DeviceTracer* tracer = GetDeviceTracer(); if (tracer) { tracer->AddCPURecords(CurAnnotation(), start_ns_, PosixInNsec(), @@ -193,14 +194,16 @@ RecordEvent::~RecordEvent() { PopEvent(name_, dev_ctx_); } -RecordBlock::RecordBlock(int block_id) : start_ns_(PosixInNsec()) { +RecordBlock::RecordBlock(int block_id) + : is_enabled_(false), start_ns_(PosixInNsec()) { if (g_state == ProfilerState::kDisabled) return; + is_enabled_ = true; SetCurBlock(block_id); name_ = string::Sprintf("block_%d", block_id); } RecordBlock::~RecordBlock() { - if (g_state == ProfilerState::kDisabled) return; + if (g_state == ProfilerState::kDisabled || !is_enabled_) return; DeviceTracer* tracer = GetDeviceTracer(); if (tracer) { // We try to put all blocks at the same nested depth in the diff --git a/paddle/fluid/platform/profiler.h b/paddle/fluid/platform/profiler.h index 61b98143e41abb9e47d2c717c7876f1bab7f5077..643bb6183d144ec11a4890d9ea1ca970acb08b4c 100644 --- a/paddle/fluid/platform/profiler.h +++ b/paddle/fluid/platform/profiler.h @@ -74,6 +74,7 @@ struct RecordEvent { ~RecordEvent(); + bool is_enabled_; uint64_t start_ns_; // The device context is used by Event to get the current cuda stream. const DeviceContext* dev_ctx_; @@ -89,6 +90,7 @@ struct RecordBlock { ~RecordBlock(); private: + bool is_enabled_; std::string name_; uint64_t start_ns_; }; diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc index 6471eb3ab7bf05365c0bb2bf68bb74ef9044c527..bcf6d4dd3087060c016e53722cde80704ef2e834 100644 --- a/paddle/fluid/pybind/protobuf.cc +++ b/paddle/fluid/pybind/protobuf.cc @@ -238,6 +238,7 @@ void BindVarDsec(pybind11::module *m) { pybind11::enum_(var_desc, "VarType", "") .value("BOOL", pd::proto::VarType::BOOL) + .value("UINT8", pd::proto::VarType::UINT8) .value("INT16", pd::proto::VarType::INT16) .value("INT32", pd::proto::VarType::INT32) .value("INT64", pd::proto::VarType::INT64) diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index c925686f8382da1758fb7cdc048253290ef69513..50a1c07251b5bc4e7cc27de63f5457d3f94daef5 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -494,20 +494,61 @@ All parameter, weight, gradient are variables in Paddle. m.def("disable_profiler", platform::DisableProfiler); m.def("reset_profiler", platform::ResetProfiler); - py::class_(m, "ParallelExecutor") - .def("__init__", - [](ParallelExecutor &self, size_t num_threads, bool use_event, - const std::vector &places, - const std::unordered_set ¶ms, - const std::unordered_set &bcast_vars, - const ProgramDesc &main_program, const std::string &loss_var_name, - Scope *scope, std::vector &local_scopes, - bool allow_op_delay, bool use_default_grad_scale) { - new (&self) ParallelExecutor( - num_threads, use_event, places, params, bcast_vars, - main_program, loss_var_name, scope, local_scopes, - allow_op_delay, use_default_grad_scale); - }) + // -- python binds for parallel executor. + py::class_ pe(m, "ParallelExecutor"); + py::class_(pe, "ExecutionStrategy") + .def(py::init()) + .def_property( + "num_threads", + [](const ExecutionStrategy &self) { return self.num_threads_; }, + [](ExecutionStrategy &self, size_t num_threads) { + self.num_threads_ = num_threads; + }) + .def_property( + "use_event", + [](const ExecutionStrategy &self) { return self.use_event_; }, + [](ExecutionStrategy &self, bool use_event) { + self.use_event_ = use_event; + }) + .def_property( + "allow_op_delay", + [](const ExecutionStrategy &self) { return self.allow_op_delay_; }, + [](ExecutionStrategy &self, bool allow_op_delay) { + self.allow_op_delay_ = allow_op_delay; + }); + py::class_ build_strategy(pe, "BuildStrategy"); + + py::enum_(build_strategy, "ReduceStrategy") + .value("Reduce", BuildStrategy::ReduceStrategy::kReduce) + .value("AllReduce", BuildStrategy::ReduceStrategy::kAllReduce); + py::enum_(build_strategy, + "GradientScaleStrategy") + .value("CoeffNumDevice", + BuildStrategy::GradientScaleStrategy::kCoeffNumDevice) + .value("One", BuildStrategy::GradientScaleStrategy::kOne) + .value("Customized", BuildStrategy::GradientScaleStrategy::kCustomized); + + build_strategy.def(py::init()) + .def_property( + "reduce_strategy", + [](const BuildStrategy &self) { return self.reduce_; }, + [](BuildStrategy &self, BuildStrategy::ReduceStrategy strategy) { + self.reduce_ = strategy; + }) + .def_property( + "gradient_scale_strategy", + [](const BuildStrategy &self) { return self.gradient_scale_; }, + [](BuildStrategy &self, + BuildStrategy::GradientScaleStrategy strategy) { + self.gradient_scale_ = strategy; + }); + + pe.def(py::init &, + const std::unordered_set &, + const std::unordered_set &, const ProgramDesc &, + const std::string &, Scope *, std::vector &, + const ExecutionStrategy &, const BuildStrategy &, size_t, + size_t>()) .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs) // NOTE: even we return a vec* to Python use reference policy. // We still cannot get local_scope from this vector, since the element diff --git a/paddle/fluid/train/demo/CMakeLists.txt b/paddle/fluid/train/demo/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..78d6e5ff554b9cd9facae85be166a697e0b75337 --- /dev/null +++ b/paddle/fluid/train/demo/CMakeLists.txt @@ -0,0 +1,66 @@ +cmake_minimum_required(VERSION 3.0) + +project(cpp_train_demo CXX C) + +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11") + +if(NOT DEFINED PADDLE_LIB) + message(FATAL_ERROR "please set PADDLE_LIB with -DPADDLE_LIB=/paddle/lib/dir") +endif() + +option(WITH_MKLDNN "Compile PaddlePaddle with MKLDNN" OFF) +option(WITH_MKL "Compile PaddlePaddle with MKL support, default use openblas." OFF) + +include_directories("${PADDLE_LIB}") +include_directories("${PADDLE_LIB}/third_party/install/protobuf/include") +include_directories("${PADDLE_LIB}/third_party/install/glog/include") +include_directories("${PADDLE_LIB}/third_party/install/gflags/include") +include_directories("${PADDLE_LIB}/third_party/install/snappy/include") +include_directories("${PADDLE_LIB}/third_party/install/snappystream/include") +include_directories("${PADDLE_LIB}/third_party/install/zlib/include") + +include_directories("${PADDLE_LIB}/third_party/boost") +include_directories("${PADDLE_LIB}/third_party/eigen3") + +link_directories("${PADDLE_LIB}/third_party/install/snappy/lib") +link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib") +link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib") +link_directories("${PADDLE_LIB}/third_party/install/glog/lib") +link_directories("${PADDLE_LIB}/third_party/install/gflags/lib") +link_directories("${PADDLE_LIB}/third_party/install/zlib/lib") + +add_executable(demo_trainer demo_trainer.cc) + +if(WITH_MKLDNN) + include_directories("${PADDLE_LIB}/third_party/install/mkldnn/include") + set(MKLDNN_LIB ${PADDLE_LIB}/third_party/install/mkldnn/lib/libmkldnn.so.0) +endif() + +if(WITH_MKL) + include_directories("${PADDLE_LIB}/third_party/install/mklml/include") + set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel.so) +else() + if(APPLE) + set(MATH_LIB cblas) + else(APPLE) + set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas.a) + endif(APPLE) +endif() + +if(APPLE) + set(MACOS_LD_FLAGS "-undefined dynamic_lookup -Wl,-all_load -framework CoreFoundation -framework Security") +else(APPLE) + set(ARCHIVE_START "-Wl,--whole-archive") + set(ARCHIVE_END "-Wl,--no-whole-archive") + set(EXTERNAL_LIB "-lrt -ldl -lpthread") +endif(APPLE) + +target_link_libraries(demo_trainer + ${MACOS_LD_FLAGS} + ${ARCHIVE_START} + ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid.a + ${ARCHIVE_END} + ${MATH_LIB} + ${MKLDNN_LIB} + glog gflags protobuf snappystream snappy z + ${EXTERNAL_LIB}) diff --git a/paddle/fluid/train/demo/README.md b/paddle/fluid/train/demo/README.md new file mode 100644 index 0000000000000000000000000000000000000000..fd80a77b02e60c15ae6c58486ed7cbbb6ffefabc --- /dev/null +++ b/paddle/fluid/train/demo/README.md @@ -0,0 +1,66 @@ + +### step 1. build paddle lib + +``` + +# WITH_MKL=ON|OFF +# WITH_MKLDNN=ON|OFF + +PADDLE_LIB=/paddle/lib/dir +cmake .. -DCMAKE_INSTALL_PREFIX=$PADDLE_LIB \ + -DCMAKE_BUILD_TYPE=Release \ + -DWITH_FLUID_ONLY=ON \ + -DWITH_GPU=OFF \ + -DWITH_STYLE_CHECK=OFF \ + -DWITH_MKL=OFF \ + -DWITH_MKLDNN=OFF +make -j8 +make -j8 inference_lib_dist +``` + +### step 2. generate program desc +``` +# please install paddle before run this scripe +pip install --upgrade paddlepaddle-*.whl +python demo_network.py +``` + +This will generate two program desc files: + - startup_program: used to init all parameters + - main_program: main logic of the network + +### step 3. build demo_trainer and run it. + + +``` +# Make a build dir at the same dir of this README.md document. +# The demo dir can be put anywhere. +mkdir build +cd build + +# WITH_MKL=ON|OFF +# WITH_MKLDNN=ON|OFF +PADDLE_LIB=/paddle/lib/dir + +# PADDLE_LIB is the same with CMAKE_INSTALL_PREFIX when building the lib +cmake .. -DPADDLE_LIB=$PADDLE_LIB \ + -DWITH_MKLDNN=OFF \ + -DWITH_MKL=OFF +make + +# copy startup_program and main_program to this dir +cp ../startup_program . +cp ../main_program . + +# run demo cpp trainer +./demo_trainer + +``` + +The output will be: +``` +step: 0 loss: 1069.02 +step: 1 loss: 1069.02 +step: 2 loss: 1069.02 +.... +``` diff --git a/paddle/fluid/train/demo/demo_network.py b/paddle/fluid/train/demo/demo_network.py new file mode 100644 index 0000000000000000000000000000000000000000..41e98c6a24a750a9300b5c2a6d370303cc0e59c5 --- /dev/null +++ b/paddle/fluid/train/demo/demo_network.py @@ -0,0 +1,47 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle.fluid as fluid +import paddle.fluid.framework as framework + + +def train_network(with_optimize): + x = fluid.layers.data(name='x', shape=[13], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None) + + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + + if with_optimize: + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.00001) + sgd_optimizer.minimize(avg_cost) + else: + fluid.backward.append_backward(avg_cost) + + +def save_program_desc(network_func): + startup_program = framework.Program() + train_program = framework.Program() + + with framework.program_guard(train_program, startup_program): + network_func(with_optimize=False) + + with open("startup_program", "w") as f: + f.write(startup_program.desc.serialize_to_string()) + with open("main_program", "w") as f: + f.write(train_program.desc.serialize_to_string()) + + +save_program_desc(train_network) diff --git a/paddle/fluid/train/demo/demo_trainer.cc b/paddle/fluid/train/demo/demo_trainer.cc new file mode 100644 index 0000000000000000000000000000000000000000..813d8386868558bd62a9d5670d540ddeddb2b77d --- /dev/null +++ b/paddle/fluid/train/demo/demo_trainer.cc @@ -0,0 +1,103 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include + +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/init.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/place.h" + +namespace paddle { +namespace train { + +void ReadBinaryFile(const std::string& filename, std::string* contents) { + std::ifstream fin(filename, std::ios::in | std::ios::binary); + PADDLE_ENFORCE(static_cast(fin), "Cannot open file %s", filename); + fin.seekg(0, std::ios::end); + contents->clear(); + contents->resize(fin.tellg()); + fin.seekg(0, std::ios::beg); + fin.read(&(contents->at(0)), contents->size()); + fin.close(); +} + +std::unique_ptr Load( + paddle::framework::Executor* executor, const std::string& model_filename) { + VLOG(3) << "loading model from " << model_filename; + std::string program_desc_str; + ReadBinaryFile(model_filename, &program_desc_str); + + std::unique_ptr main_program( + new paddle::framework::ProgramDesc(program_desc_str)); + return main_program; +} + +} // namespace train +} // namespace paddle + +int main() { + paddle::framework::InitDevices(false); + + const auto cpu_place = paddle::platform::CPUPlace(); + + paddle::framework::Executor executor(cpu_place); + paddle::framework::Scope scope; + auto startup_program = paddle::train::Load(&executor, "startup_program"); + auto train_program = paddle::train::Load(&executor, "main_program"); + + std::string loss_name = ""; + for (auto op_desc : train_program->Block(0).AllOps()) { + if (op_desc->Type() == "mean") { + loss_name = op_desc->Output("Out")[0]; + break; + } + } + + PADDLE_ENFORCE_NE(loss_name, "", "loss not found"); + + // init all parameters + executor.Run(*startup_program.get(), &scope, 0); + + // prepare data + auto x_var = scope.Var("x"); + auto x_tensor = x_var->GetMutable(); + x_tensor->Resize({2, 13}); + + auto x_data = x_tensor->mutable_data(cpu_place); + for (int i = 0; i < 2 * 13; ++i) { + x_data[i] = static_cast(i); + } + + auto y_var = scope.Var("y"); + auto y_tensor = y_var->GetMutable(); + y_tensor->Resize({2, 1}); + auto y_data = y_tensor->mutable_data(cpu_place); + for (int i = 0; i < 2 * 1; ++i) { + y_data[i] = static_cast(i); + } + + auto loss_var = scope.Var(loss_name); + + for (int i = 0; i < 10; ++i) { + executor.Run(*train_program.get(), &scope, 0, false, true); + std::cout << "step: " << i << " loss: " + << loss_var->Get().data()[0] + << std::endl; + } + return 0; +} diff --git a/paddle/gserver/layers/PriorBox.cpp b/paddle/gserver/layers/PriorBox.cpp index af2cc05a954b3a6857c1015104a57339282840b8..56a4d942f0fdcb981f52f6ce0f644ec57a0e3c9a 100644 --- a/paddle/gserver/layers/PriorBox.cpp +++ b/paddle/gserver/layers/PriorBox.cpp @@ -28,7 +28,7 @@ namespace paddle { */ class PriorBoxLayer : public Layer { -public: +public: // NOLINT explicit PriorBoxLayer(const LayerConfig& config) : Layer(config) {} bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) override; @@ -36,7 +36,7 @@ public: void forward(PassType passType) override; void backward(const UpdateCallback& callback) override {} -protected: +protected: // NOLINT int numPriors_; std::vector minSize_; std::vector maxSize_; @@ -109,11 +109,18 @@ void PriorBoxLayer::forward(PassType passType) { real boxWidth = minSize; real boxHeight = minSize; - // priors with different aspect ratios - for (size_t r = 0; r < aspectRatio_.size(); r++) { - real ar = aspectRatio_[r]; - boxWidth = minSize * sqrt(ar); - boxHeight = minSize / sqrt(ar); + // first prior: aspect_ratio == 1.0, compatible to old logic + tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth; + tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight; + tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth; + tmpPtr[idx++] = (centerY + boxHeight / 2.) / imageHeight; + // set the variance. + for (int t = 0; t < 4; t++) tmpPtr[idx++] = variance_[t]; + + if (maxSize_.size() > 0) { + // square prior with size sqrt(minSize * maxSize) + real maxSize = maxSize_[s]; + boxWidth = boxHeight = sqrt(minSize * maxSize); tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth; tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight; tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth; @@ -122,10 +129,14 @@ void PriorBoxLayer::forward(PassType passType) { for (int t = 0; t < 4; t++) tmpPtr[idx++] = variance_[t]; } - if (maxSize_.size() > 0) { - // square prior with size sqrt(minSize * maxSize) - real maxSize = maxSize_[s]; - boxWidth = boxHeight = sqrt(minSize * maxSize); + // priors with different aspect ratios + for (size_t r = 0; r < aspectRatio_.size(); r++) { + real ar = aspectRatio_[r]; + if (fabs(ar - 1.0) < 1e-6) { + continue; + } + boxWidth = minSize * sqrt(ar); + boxHeight = minSize / sqrt(ar); tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth; tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight; tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth; diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh index 7e00bd38487902227c3b4521db20cdbe314059be..92b8b90880bc78dbc281a959a7472c2822f76fc3 100755 --- a/paddle/scripts/docker/build.sh +++ b/paddle/scripts/docker/build.sh @@ -198,7 +198,7 @@ EOF # run paddle version to install python packages first RUN apt-get update &&\ ${NCCL_DEPS}\ - apt-get install -y wget python-pip dmidecode python-tk && pip install -U pip==9.0.3 && \ + apt-get install -y wget python-pip dmidecode python-tk && easy_install -U pip && \ pip install /*.whl; apt-get install -f -y && \ apt-get clean -y && \ rm -f /*.whl && \ diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh index 1595cc9e8aad4d143ca62f84f812dbc791dc1d26..fbe219a1c9cf85f19ae2ab991ae7e4207858f204 100755 --- a/paddle/scripts/paddle_build.sh +++ b/paddle/scripts/paddle_build.sh @@ -20,19 +20,15 @@ #================================================= function print_usage() { - RED='\033[0;31m' - BLUE='\033[0;34m' - BOLD='\033[1m' - NONE='\033[0m' - echo -e "\n${RED}Usage${NONE}: - ${BOLD}$0${NONE} [OPTION]" + ${BOLD}${SCRIPT_NAME}${NONE} [OPTION]" echo -e "\n${RED}Options${NONE}: ${BLUE}build${NONE}: run build for x86 platform ${BLUE}build_android${NONE}: run build for android platform ${BLUE}build_ios${NONE}: run build for ios platform ${BLUE}test${NONE}: run all unit tests + ${BLUE}single_test${NONE}: run a single unit test ${BLUE}bind_test${NONE}: parallel tests bind to different GPU ${BLUE}doc${NONE}: generate paddle documents ${BLUE}html${NONE}: convert C++ source code into HTML @@ -45,7 +41,15 @@ function print_usage() { } function init() { + RED='\033[0;31m' + BLUE='\033[0;34m' + BOLD='\033[1m' + NONE='\033[0m' + PADDLE_ROOT="$( cd "$( dirname "${BASH_SOURCE[0]}")/../../" && pwd )" + if [ -z "${SCRIPT_NAME}" ]; then + SCRIPT_NAME=$0 + fi } function cmake_gen() { @@ -91,7 +95,6 @@ function cmake_gen() { -DWITH_AVX=${WITH_AVX:-OFF} -DWITH_GOLANG=${WITH_GOLANG:-OFF} -DCUDA_ARCH_NAME=${CUDA_ARCH_NAME:-All} - -DWITH_SWIG_PY=ON -DWITH_C_API=${WITH_C_API:-OFF} -DWITH_PYTHON=${WITH_PYTHON:-ON} -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} @@ -309,6 +312,25 @@ EOF fi } +function single_test() { + TEST_NAME=$1 + if [ -z "${TEST_NAME}" ]; then + echo -e "${RED}Usage:${NONE}" + echo -e "${BOLD}${SCRIPT_NAME}${NONE} ${BLUE}single_test${NONE} [test_name]" + exit 1 + fi + mkdir -p ${PADDLE_ROOT}/build + cd ${PADDLE_ROOT}/build + if [ ${WITH_TESTING:-ON} == "ON" ] ; then + cat <> /paddle/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile < /dev/null - return $? -} - function start_build_docker() { docker pull $IMG - if container_running "${CONTAINER_ID}"; then - docker stop "${CONTAINER_ID}" 1>/dev/null - docker rm -f "${CONTAINER_ID}" 1>/dev/null - fi - apt_mirror='s#http://archive.ubuntu.com/ubuntu#mirror://mirrors.ubuntu.com/mirrors.txt#g' DOCKER_ENV=$(cat < 0: + if not (bool(feeded_var_names) and all( + isinstance(name, basestring) for name in feeded_var_names)): + raise ValueError("'feed_var_names' should be a list of str.") if isinstance(target_vars, Variable): target_vars = [target_vars] diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 4b707973e27391a6bdcba138934f62a255e04bb2..dee41448081cbfcd8224ce2abbf3ba7b7b97eb7c 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -49,6 +49,7 @@ __all__ = [ 'reorder_lod_tensor_by_rank', 'ParallelDo', 'Print', + 'is_empty', ] @@ -1562,3 +1563,40 @@ def reorder_lod_tensor_by_rank(x, rank_table): 'RankTable': [rank_table]}, outputs={'Out': [out]}) return out + + +def is_empty(x, cond=None, **ignored): + """ + **Is Empty** + + This layer returns the truth value of whether the variable is empty. + + Args: + x(Variable): Operand of *is_empty* + cond(Variable|None): Optional output variable to store the result + of *is_empty* + + Returns: + Variable: The tensor variable storing the output of *is_empty*. + + Raises: + TypeError: If input cond is not a variable, or cond's dtype is + not bool + + Examples: + .. code-block:: python + + less = fluid.layers.is_empty(x=input) + """ + helper = LayerHelper("is_empty", **locals()) + if cond is None: + cond = helper.create_tmp_variable(dtype='bool') + cond.stop_gradient = True + elif not isinstance(cond, Variable): + raise TypeError("cond takes a variable") + elif cond.dtype != 'bool': + raise TypeError("The data type of cond must be bool") + + helper.append_op( + type='is_empty', inputs={'X': [x]}, outputs={'Out': [cond]}) + return cond diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index a5938fe494265778ef7032c56a8d6d35acd729c5..3a83db12fd13651578deeac6b562bac2f1e4e4b6 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -23,6 +23,7 @@ import nn import math __all__ = [ + 'prior_box', 'multi_box_head', 'bipartite_match', 'target_assign', @@ -564,6 +565,115 @@ def ssd_loss(location, return loss +def prior_box(input, + image, + min_sizes, + max_sizes=None, + aspect_ratios=[1.], + variance=[0.1, 0.1, 0.2, 0.2], + flip=False, + clip=False, + steps=[0.0, 0.0], + offset=0.5, + name=None): + """ + **Prior box operator** + + Generate prior boxes for SSD(Single Shot MultiBox Detector) algorithm. + Each position of the input produce N prior boxes, N is determined by + the count of min_sizes, max_sizes and aspect_ratios, The size of the + box is in range(min_size, max_size) interval, which is generated in + sequence according to the aspect_ratios. + + Args: + input(Variable): The Input Variables, the format is NCHW. + image(Variable): The input image data of PriorBoxOp, + the layout is NCHW. + min_sizes(list|tuple|float value): min sizes of generated prior boxes. + max_sizes(list|tuple|None): max sizes of generated prior boxes. + Default: None. + aspect_ratios(list|tuple|float value): the aspect ratios of generated + prior boxes. Default: [1.]. + variance(list|tuple): the variances to be encoded in prior boxes. + Default:[0.1, 0.1, 0.2, 0.2]. + flip(bool): Whether to flip aspect ratios. Default:False. + clip(bool): Whether to clip out-of-boundary boxes. Default: False. + step(list|turple): Prior boxes step across width and height, If + step[0] == 0.0/step[1] == 0.0, the prior boxes step across + height/weight of the input will be automatically calculated. + Default: [0., 0.] + offset(float): Prior boxes center offset. Default: 0.5 + name(str): Name of the prior box op. Default: None. + + Returns: + boxes(Variable): the output prior boxes of PriorBox. + The layout is [H, W, num_priors, 4]. + H is the height of input, W is the width of input, + num_priors is the total + box count of each position of input. + Variances(Variable): the expanded variances of PriorBox. + The layout is [H, W, num_priors, 4]. + H is the height of input, W is the width of input + num_priors is the total + box count of each position of input + + + Examples: + .. code-block:: python + box, var = prior_box( + input=conv1, + image=images, + min_sizes=[100.], + flip=True, + clip=True) + """ + helper = LayerHelper("prior_box", **locals()) + dtype = helper.input_dtype() + + def _is_list_or_tuple_(data): + return (isinstance(data, list) or isinstance(data, tuple)) + + if not _is_list_or_tuple_(min_sizes): + min_sizes = [min_sizes] + if not _is_list_or_tuple_(aspect_ratios): + aspect_ratios = [aspect_ratios] + if not (_is_list_or_tuple_(steps) and len(steps) == 2): + raise ValueError('steps should be a list or tuple ', + 'with length 2, (step_width, step_height).') + + min_sizes = list(map(float, min_sizes)) + aspect_ratios = list(map(float, aspect_ratios)) + steps = list(map(float, steps)) + + attrs = { + 'min_sizes': min_sizes, + 'aspect_ratios': aspect_ratios, + 'variances': variance, + 'flip': flip, + 'clip': clip, + 'step_w': steps[0], + 'step_h': steps[1], + 'offset': offset + } + if max_sizes is not None and len(max_sizes) > 0 and max_sizes[0] > 0: + if not _is_list_or_tuple_(max_sizes): + max_sizes = [max_sizes] + attrs['max_sizes'] = max_sizes + + box = helper.create_tmp_variable(dtype) + var = helper.create_tmp_variable(dtype) + helper.append_op( + type="prior_box", + inputs={"Input": input, + "Image": image}, + outputs={"Boxes": box, + "Variances": var}, + attrs=attrs, ) + box.stop_gradient = True + var.stop_gradient = True + return box, var + + def multi_box_head(inputs, image, base_size, @@ -660,47 +770,6 @@ def multi_box_head(inputs, clip=True) """ - def _prior_box_(input, - image, - min_sizes, - max_sizes, - aspect_ratios, - variance, - flip=False, - clip=False, - step_w=0.0, - step_h=0.0, - offset=0.5, - name=None): - helper = LayerHelper("prior_box", **locals()) - dtype = helper.input_dtype() - - attrs = { - 'min_sizes': min_sizes, - 'aspect_ratios': aspect_ratios, - 'variances': variance, - 'flip': flip, - 'clip': clip, - 'step_w': step_w, - 'step_h': step_h, - 'offset': offset - } - if len(max_sizes) > 0 and max_sizes[0] > 0: - attrs['max_sizes'] = max_sizes - - box = helper.create_tmp_variable(dtype) - var = helper.create_tmp_variable(dtype) - helper.append_op( - type="prior_box", - inputs={"Input": input, - "Image": image}, - outputs={"Boxes": box, - "Variances": var}, - attrs=attrs, ) - box.stop_gradient = True - var.stop_gradient = True - return box, var - def _reshape_with_axis_(input, axis=1): if not (axis > 0 and axis < len(input.shape)): raise ValueError("The axis should be smaller than " @@ -777,11 +846,10 @@ def multi_box_head(inputs, aspect_ratio = aspect_ratios[i] if not _is_list_or_tuple_(aspect_ratio): aspect_ratio = [aspect_ratio] + step = [step_w[i] if step_w else 0.0, step_h[i] if step_w else 0.0] - box, var = _prior_box_(input, image, min_size, max_size, aspect_ratio, - variance, flip, clip, step_w[i] - if step_w else 0.0, step_h[i] - if step_w else 0.0, offset) + box, var = prior_box(input, image, min_size, max_size, aspect_ratio, + variance, flip, clip, step, offset) box_results.append(box) var_results.append(var) diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index 0a6befd1485a1f79d63873c47a9fd74ab4214f57..4d6ee3c51b7cccdaa3303b5a4cd8e7219b753ccb 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -78,8 +78,8 @@ def data(name, dtype=dtype, type=type, stop_gradient=stop_gradient, - lod_level=lod_level) - data_var.is_data = True + lod_level=lod_level, + is_data=True) return data_var diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py index 35b01a79914b3427836d4abd51aa2e2eb471d517..295d1b7190ec39bcc6efdf72aebede14a99807aa 100644 --- a/python/paddle/fluid/layers/layer_function_generator.py +++ b/python/paddle/fluid/layers/layer_function_generator.py @@ -113,7 +113,7 @@ def generate_layer_fn(op_type): if len(not_intermediate_outputs) != 1: raise ValueError("Only one non intermediate output operator can be", - "automatically generated.") + "automatically generated. {0}".format(op_type)) if not_intermediate_outputs[0].duplicable: raise ValueError( diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 1786be22fdcd0d074b45bc94b3b0c4e8c41b4e8a..561c8bd42f90911bf5a0c898fe01412d42d2c9b1 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -1329,6 +1329,8 @@ def sequence_pool(input, pool_type): sqrt : out.data = [2.82, 6.93, 4.24], where 2.82=(1+3)/sqrt(2), 6.93=(2+4+6)/sqrt(3), 4.24=(5+1)/sqrt(2) max : out.data = [3, 6, 5], where 3=max(1,3), 6=max(2,4,6), 5=max(5,1) + last : out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1) + first : out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1) Args: input(variable): The input variable which is a LoDTensor. @@ -1348,6 +1350,8 @@ def sequence_pool(input, pool_type): sum_x = fluid.layers.sequence_pool(input=x, pool_type='sum') sqrt_x = fluid.layers.sequence_pool(input=x, pool_type='sqrt') max_x = fluid.layers.sequence_pool(input=x, pool_type='max') + last_x = fluid.layers.sequence_pool(input=x, pool_type='last') + first_x = fluid.layers.sequence_pool(input=x, pool_type='first') """ helper = LayerHelper('sequence_pool', **locals()) dtype = helper.input_dtype() @@ -3263,35 +3267,35 @@ def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None): """ **Smooth L1 Loss Operator. ** - This operator computes the smooth l1 loss for X and Y. + This operator computes the smooth L1 loss for X and Y. The operator takes the first dimension of X and Y as batch size. - For each instance, it computes the smooth l1 loss element by element first + For each instance, it computes the smooth L1 loss element by element first and then sums all the losses. So the shape of Out is [batch_size, 1]. Args: x (Variable): A tensor with rank at least 2. The input value of smooth - l1 loss op with shape [batch_size, dim1, ..., dimN]. + L1 loss op with shape [batch_size, dim1, ..., dimN]. y (Variable): A tensor with rank at least 2. The target value of smooth - l1 loss op with same shape as x. + L1 loss op with same shape as x. inside_weight (Variable|None): A tensor with rank at least 2. This input is optional and should have same shape with x. If provided, the result of (x - y) will be multiplied by this tensor element by element. outside_weight (Variable|None): A tensor with rank at least 2. This input is optional and should have same shape with x. If provided, - the out smooth l1 loss will be multiplied by this tensor element + the out smooth L1 loss will be multiplied by this tensor element by element. - sigma (float|None): Hyper parameter of smooth l1 loss op. A float scalar + sigma (float|None): Hyper parameter of smooth L1 loss op. A float scalar with default value 1.0. Returns: - Variable: A tensor with rank be 2. The output smooth l1 loss with + Variable: A tensor with rank be 2. The output smooth L1 loss with shape [batch_size, 1]. Examples: .. code-block:: python data = fluid.layers.data(name='data', shape=[128], dtype='float32') - label = fluid.layers.data(name='label', shape=[100], dtype='int64') + label = fluid.layers.data(name='label', shape=[100], dtype='float32') fc = fluid.layers.fc(input=data, size=100) out = fluid.layers.smooth_l1(x=fc, y=label) """ @@ -3769,13 +3773,13 @@ def label_smooth(label, def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0): """ - Region of interest pooling (also known as RoI pooling) is to perform + Region of interest pooling (also known as RoI pooling) is to perform is to perform max pooling on inputs of nonuniform sizes to obtain fixed-size feature maps (e.g. 7*7). - The operator has three steps: - 1. Dividing each region proposal into equal-sized sections with - the pooled_width and pooled_height - 2. Finding the largest value in each section + The operator has three steps: + 1. Dividing each region proposal into equal-sized sections with + the pooled_width and pooled_height + 2. Finding the largest value in each section 3. Copying these max values to the output buffer Args: @@ -3783,8 +3787,8 @@ def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0): rois (Variable): ROIs (Regions of Interest) to pool over. It should be a 2-D one level LoTensor of shape [num_rois, 4]. The layout is [x1, y1, x2, y2], where (x1, y1) - is the top left coordinates, and (x2, y2) is the - bottom right coordinates. The num_rois is the + is the top left coordinates, and (x2, y2) is the + bottom right coordinates. The num_rois is the total number of ROIs in this batch data. pooled_height (integer): The pooled output height. Default: 1 pooled_width (integer): The pooled output width. Default: 1 @@ -3793,11 +3797,11 @@ def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0): to the scale used when pooling. Default: 1.0 Returns: - pool_out (Variable): The output is a 4-D tensor of the shape + pool_out (Variable): The output is a 4-D tensor of the shape (num_rois, channels, pooled_h, pooled_w). Examples: - pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0) + pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0) """ helper = LayerHelper('roi_pool', **locals()) dtype = helper.input_dtype() diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index 0a314ddfd7c607a3bc7f7c746c4c4990fc4a52e2..0fc48055220ed84c4ab146ad01b05f393e01078e 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -47,6 +47,8 @@ class Optimizer(object): raise TypeError("learning rate should be float or Variable") self.regularization = regularization self._learning_rate = learning_rate + # the learning rate type should be inferenced from loss + self._dtype = None # each program should have a independent learning rate # program -> Variable(learning_rate) self._learning_rate_map = dict() @@ -77,7 +79,7 @@ class Optimizer(object): name=unique_name.generate("learning_rate"), shape=[1], value=float(self._learning_rate), - dtype='float32', + dtype='float32' if self._dtype == None else self._dtype, persistable=True) def global_learning_rate(self, program=None): @@ -200,6 +202,7 @@ class Optimizer(object): # Create any accumulators program = loss.block.program + self._dtype = loss.dtype with program_guard(program, startup_program): global_block = framework.default_main_program().global_block() start = len(global_block.ops) @@ -391,7 +394,7 @@ class AdamOptimizer(Optimizer): beta_shape = [1] self._beta1_pow_acc = self.helper.create_global_variable( name=unique_name.generate('beta1_pow_acc'), - dtype='float32', + dtype='float32' if self._dtype == None else self._dtype, shape=beta_shape, lod_level=0, persistable=True) @@ -400,7 +403,7 @@ class AdamOptimizer(Optimizer): self._beta2_pow_acc = self.helper.create_global_variable( name=unique_name.generate('beta2_pow_acc'), - dtype='float32', + dtype='float32' if self._dtype == None else self._dtype, shape=beta_shape, lod_level=0, persistable=True) @@ -493,7 +496,7 @@ class AdamaxOptimizer(Optimizer): beta_shape = [1] self._beta1_pow_acc = self.helper.create_global_variable( name=unique_name.generate('beta1_pow_acc'), - dtype='float32', + dtype='float32' if self._dtype == None else self._dtype, shape=beta_shape, lod_level=0, persistable=True) @@ -900,8 +903,10 @@ class ModelAverage(Optimizer): # param = (sum_1 + sum_2 + sum_3) / (num_accumulates + old_num_accumulates) tmp = layers.sum(x=[num_accumulates, old_num_accumulates]) sum = layers.sum(x=[sum_1, sum_2, sum_3]) - tmp = layers.cast(x=tmp, dtype='float32') - sum = layers.cast(x=sum, dtype='float32') + tmp = layers.cast( + x=tmp, dtype='float32' if self._dtype == None else self._dtype) + sum = layers.cast( + x=sum, dtype='float32' if self._dtype == None else self._dtype) layers.elementwise_div(x=sum, y=tmp, out=param) def _add_average_restore_op(self, block, param_grad): diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index 6b80b007e9080922241ee6c66e0577a18b357980..3117dfe00c7a3df1035c439dc31b81e67781d0cc 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -19,7 +19,10 @@ import executor import warnings import sys -__all__ = ['ParallelExecutor'] +__all__ = ['ParallelExecutor', 'ExecutionStrategy', 'BuildStrategy'] + +ExecutionStrategy = core.ParallelExecutor.ExecutionStrategy +BuildStrategy = core.ParallelExecutor.BuildStrategy class ParallelExecutor(object): @@ -27,10 +30,12 @@ class ParallelExecutor(object): use_cuda, loss_name=None, main_program=None, - num_threads=None, - allow_op_delay=False, share_vars_from=None, - use_default_grad_scale=True): + exec_strategy=None, + build_strategy=None, + num_trainers=1, + trainer_id=0, + **kwargs): """ ParallelExecutor can run program in parallel. @@ -39,18 +44,13 @@ class ParallelExecutor(object): loss_name(str, default None): The loss name must set in training. main_program(Program, default None): The program that need to run, if not provided, then default_main_program will be used. - num_threads(int, default None): How many threads are used for - training. - allow_op_delay(bool, default False): Whether to delay and buffer - some operators together for scheduling or not, which may - improve performance in some cases, default False. share_vars_from(ParallelExecutor, default None): If provied, it will share variables from the specified ParallelExecutor. - use_default_grad_scale(bool, default True): If set True, a default - scale value equal to `1./device_count` would be multiplied to - gradients of each device and scaled gradients would be - aggregated. Otherwise, a customized scale value should be fed - to the network. + num_trainers(int, default 1): If greater than 1, NCCL will be + initialized with multpile rank of nodes, each node should have + same number of GPUs. Distributed training will be enabled then. + trainer_id(int, default 0): Must use together with num_trainers. + trainer_id is the "rank" of current node starts from 0. Returns: A ParallelExecutor object. @@ -72,6 +72,25 @@ class ParallelExecutor(object): train_loss, = train_exe.run([loss.name], feed=feed_dict) test_loss, = test_exe.run([loss.name], feed=feed_dict) """ + if len(kwargs) != 0: + err_msg = "" + for key in kwargs: + if key in dir(ExecutionStrategy): + err_msg += \ + "Setting {0} by constructor is deprecated. Use " \ + "strategy=ExecutionStrategy(); strategy.{0}=xxx; " \ + "pe=ParallelExecutor(exec_strategy=strategy) " \ + "instead.\n ".format(key) + elif key in dir(BuildStrategy): + err_msg += \ + "Setting {0} by constructor is deprecated. Use " \ + "strategy=BuildStrategy(); See help(" \ + "paddle.fluid.ParallelExecutor.BuildStrategy) \n".format( + key) + else: + err_msg += "Setting {0} by constructor is deprecated. Use strategy.\n".format( + key) + raise ValueError(err_msg) self._places = [] self._act_places = [] @@ -89,15 +108,25 @@ class ParallelExecutor(object): self._places.append(p) assert self._places, "no place for execution" - if num_threads is None: + if exec_strategy is None: + exec_strategy = ExecutionStrategy() + if use_cuda: + exec_strategy.use_event = True + else: + exec_strategy.use_event = False + + if exec_strategy.num_threads == 0: if use_cuda: # Experiments on se-resnext shows that too many threads hurt # performance. Worth tunning for other models in the future. - num_threads = len(self._places) * 2 + exec_strategy.num_threads = len(self._places) * 2 else: - num_threads = min( + exec_strategy.num_threads = min( len(self._places) * 2, multiprocessing.cpu_count()) + if build_strategy is None: + build_strategy = BuildStrategy() + main = main_program main = main if main else framework.default_main_program() scope = executor.global_scope() @@ -116,21 +145,14 @@ class ParallelExecutor(object): ] self.executor = core.ParallelExecutor( - num_threads, - True if use_cuda else False, # use_event self._places, set([ p.name for p in main.global_block().iter_parameters() if not p.stop_gradient ]), - set(self.persistable_vars), - main.desc, - loss_name if loss_name else '', - scope, - local_scopes, - allow_op_delay, - use_default_grad_scale) - + set(self.persistable_vars), main.desc, loss_name + if loss_name else '', scope, local_scopes, exec_strategy, + build_strategy, num_trainers, trainer_id) self.scope = scope def run(self, fetch_list, feed=None, feed_dict=None): diff --git a/python/paddle/fluid/tests/book/CMakeLists.txt b/python/paddle/fluid/tests/book/CMakeLists.txt index 673c965b662a022739f8d489c331f4de9455a926..ee734f3c782adb5196a03aca5718377009a5b4e7 100644 --- a/python/paddle/fluid/tests/book/CMakeLists.txt +++ b/python/paddle/fluid/tests/book/CMakeLists.txt @@ -5,3 +5,5 @@ string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") foreach(src ${TEST_OPS}) py_test(${src} SRCS ${src}.py) endforeach() + +add_subdirectory(high-level-api) diff --git a/python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..182e30a6a9b4249a895d15cfd65c403bb6813d0d --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt @@ -0,0 +1,12 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() + +add_subdirectory(fit_a_line) +add_subdirectory(recognize_digits) +add_subdirectory(image_classification) +add_subdirectory(understand_sentiment) diff --git a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..673c965b662a022739f8d489c331f4de9455a926 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/CMakeLists.txt @@ -0,0 +1,7 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py new file mode 100644 index 0000000000000000000000000000000000000000..4c8505acf322a8ee33799c009b523cd70bd01db3 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py @@ -0,0 +1,131 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle +import paddle.fluid as fluid +import contextlib +import numpy +import unittest + +# train reader +BATCH_SIZE = 20 + +train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.uci_housing.train(), buf_size=500), + batch_size=BATCH_SIZE) + +test_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.uci_housing.test(), buf_size=500), + batch_size=BATCH_SIZE) + + +def inference_program(): + x = fluid.layers.data(name='x', shape=[13], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None) + return y_predict + + +def linear(): + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + y_predict = inference_program() + + loss = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_loss = fluid.layers.mean(loss) + + return avg_loss + + +def train(use_cuda, train_program, save_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + + trainer = fluid.Trainer( + train_func=train_program, + place=place, + optimizer=fluid.optimizer.SGD(learning_rate=0.001)) + + def event_handler(event): + if isinstance(event, fluid.EndStepEvent): + if event.step == 10: + test_metrics = trainer.test( + reader=test_reader, feed_order=['x', 'y']) + print test_metrics + ''' + ... + ['25.768919467926025'] + ['15.343549569447836'] + ... + ''' + if save_dirname is not None: + trainer.save_params(save_dirname) + trainer.stop() + + trainer.train( + reader=train_reader, + num_epochs=100, + event_handler=event_handler, + feed_order=['x', 'y']) + + +# infer +def infer(use_cuda, inference_program, save_dirname=None): + if save_dirname is None: + return + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + inferencer = fluid.Inferencer( + infer_func=inference_program, param_path=save_dirname, place=place) + + batch_size = 10 + tensor_x = numpy.random.uniform(0, 10, [batch_size, 13]).astype("float32") + + results = inferencer.infer({'x': tensor_x}) + print("infer results: ", numpy.array(results[0])) + + +def main(use_cuda): + if use_cuda and not fluid.core.is_compiled_with_cuda(): + return + + # Directory for saving the trained model + save_dirname = "fit_a_line.inference.model" + + train(use_cuda, linear, save_dirname) + infer(use_cuda, inference_program, save_dirname) + + +class TestFitALine(unittest.TestCase): + def test_cpu(self): + with self.program_scope_guard(): + with fluid.unique_name.guard(): + main(use_cuda=False) + + def test_cuda(self): + with self.program_scope_guard(): + with fluid.unique_name.guard(): + main(use_cuda=True) + + @contextlib.contextmanager + def program_scope_guard(self): + prog = fluid.Program() + startup_prog = fluid.Program() + scope = fluid.core.Scope() + with fluid.scope_guard(scope): + with fluid.program_guard(prog, startup_prog): + yield + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..673c965b662a022739f8d489c331f4de9455a926 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt @@ -0,0 +1,7 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py new file mode 100644 index 0000000000000000000000000000000000000000..7fed6d914f75b690e34411aa154359c93b6ca989 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py @@ -0,0 +1,82 @@ +# 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. +""" +CIFAR dataset. + +This module will download dataset from +https://www.cs.toronto.edu/~kriz/cifar.html and parse train/test set into +paddle reader creators. + +The CIFAR-10 dataset consists of 60000 32x32 colour images in 10 classes, +with 6000 images per class. There are 50000 training images and 10000 test +images. + +The CIFAR-100 dataset is just like the CIFAR-10, except it has 100 classes +containing 600 images each. There are 500 training images and 100 testing +images per class. + +""" + +import cPickle +import itertools +import numpy +import paddle.v2.dataset.common +import tarfile + +__all__ = ['train10'] + +URL_PREFIX = 'https://www.cs.toronto.edu/~kriz/' +CIFAR10_URL = URL_PREFIX + 'cifar-10-python.tar.gz' +CIFAR10_MD5 = 'c58f30108f718f92721af3b95e74349a' + + +def reader_creator(filename, sub_name, batch_size=None): + def read_batch(batch): + data = batch['data'] + labels = batch.get('labels', batch.get('fine_labels', None)) + assert labels is not None + for sample, label in itertools.izip(data, labels): + yield (sample / 255.0).astype(numpy.float32), int(label) + + def reader(): + with tarfile.open(filename, mode='r') as f: + names = (each_item.name for each_item in f + if sub_name in each_item.name) + + batch_count = 0 + for name in names: + batch = cPickle.load(f.extractfile(name)) + for item in read_batch(batch): + if isinstance(batch_size, int) and batch_count > batch_size: + break + batch_count += 1 + yield item + + return reader + + +def train10(batch_size=None): + """ + CIFAR-10 training set creator. + + It returns a reader creator, each sample in the reader is image pixels in + [0, 1] and label in [0, 9]. + + :return: Training reader creator + :rtype: callable + """ + return reader_creator( + paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5), + 'data_batch', + batch_size=batch_size) diff --git a/python/paddle/fluid/tests/book/image_classification/notest_image_classification_resnet.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py similarity index 77% rename from python/paddle/fluid/tests/book/image_classification/notest_image_classification_resnet.py rename to python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py index 17db38797cf19ae387f69f66daa42fc78cfcb7d5..1160e500dbd6db784eeb81b72968386347fec59a 100644 --- a/python/paddle/fluid/tests/book/image_classification/notest_image_classification_resnet.py +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py @@ -17,6 +17,7 @@ from __future__ import print_function import paddle import paddle.fluid as fluid import numpy +import cifar10_small_test_set def resnet_cifar10(input, depth=32): @@ -81,46 +82,50 @@ def train_network(): cost = fluid.layers.cross_entropy(input=predict, label=label) avg_cost = fluid.layers.mean(cost) accuracy = fluid.layers.accuracy(input=predict, label=label) - return avg_cost, accuracy + return [avg_cost, accuracy] -def train(use_cuda, save_path): +def train(use_cuda, train_program, save_dirname): BATCH_SIZE = 128 EPOCH_NUM = 1 train_reader = paddle.batch( paddle.reader.shuffle( - paddle.dataset.cifar.train10(), buf_size=128 * 10), + cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10), batch_size=BATCH_SIZE) test_reader = paddle.batch( paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) def event_handler(event): - if isinstance(event, fluid.EndIteration): - if (event.batch_id % 10) == 0: - avg_cost, accuracy = trainer.test(reader=test_reader) + if isinstance(event, fluid.EndStepEvent): + avg_cost, accuracy = trainer.test( + reader=test_reader, feed_order=['pixel', 'label']) - print('BatchID {1:04}, Loss {2:2.2}, Acc {3:2.2}'.format( - event.batch_id + 1, avg_cost, accuracy)) + print('Loss {0:2.2}, Acc {1:2.2}'.format(avg_cost, accuracy)) - if accuracy > 0.01: # Low threshold for speeding up CI - trainer.params.save(save_path) - return + if accuracy > 0.01: # Low threshold for speeding up CI + if save_dirname is not None: + trainer.save_params(save_dirname) + return place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( - train_network, + train_func=train_program, optimizer=fluid.optimizer.Adam(learning_rate=0.001), - place=place, - event_handler=event_handler) - trainer.train(train_reader, EPOCH_NUM, event_handler=event_handler) + place=place) + trainer.train( + reader=train_reader, + num_epochs=EPOCH_NUM, + event_handler=event_handler, + feed_order=['pixel', 'label']) -def infer(use_cuda, save_path): - params = fluid.Params(save_path) + +def infer(use_cuda, inference_program, save_dirname=None): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - inferencer = fluid.Inferencer(inference_network, params, place=place) + inferencer = fluid.Inferencer( + infer_func=inference_program, param_path=save_dirname, place=place) # The input's dimension of conv should be 4-D or 5-D. # Use normilized image pixels as input data, which should be in the range @@ -135,8 +140,14 @@ def main(use_cuda): if use_cuda and not fluid.core.is_compiled_with_cuda(): return save_path = "image_classification_resnet.inference.model" - train(use_cuda, save_path) - infer(use_cuda, save_path) + + train( + use_cuda=use_cuda, train_program=train_network, save_dirname=save_path) + + infer( + use_cuda=use_cuda, + inference_program=inference_network, + save_dirname=save_path) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/book/image_classification/notest_image_classification_vgg.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py similarity index 72% rename from python/paddle/fluid/tests/book/image_classification/notest_image_classification_vgg.py rename to python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py index e83afeed2f72635a40aa2ac21dc0c8611c309de4..1e3e955ba0299f2cc0fcc02d79ae6fd8ff4c1171 100644 --- a/python/paddle/fluid/tests/book/image_classification/notest_image_classification_vgg.py +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py @@ -17,6 +17,7 @@ from __future__ import print_function import paddle import paddle.fluid as fluid import numpy +import cifar10_small_test_set def vgg16_bn_drop(input): @@ -60,46 +61,48 @@ def train_network(): cost = fluid.layers.cross_entropy(input=predict, label=label) avg_cost = fluid.layers.mean(cost) accuracy = fluid.layers.accuracy(input=predict, label=label) - return avg_cost, accuracy + return [avg_cost, accuracy] -def train(use_cuda, save_path): +def train(use_cuda, train_program, save_dirname): BATCH_SIZE = 128 - EPOCH_NUM = 1 - train_reader = paddle.batch( paddle.reader.shuffle( - paddle.dataset.cifar.train10(), buf_size=128 * 10), + cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10), batch_size=BATCH_SIZE) test_reader = paddle.batch( paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) def event_handler(event): - if isinstance(event, fluid.EndIteration): - if (event.batch_id % 10) == 0: - avg_cost, accuracy = trainer.test(reader=test_reader) + if isinstance(event, fluid.EndStepEvent): + avg_cost, accuracy = trainer.test( + reader=test_reader, feed_order=['pixel', 'label']) - print('BatchID {1:04}, Loss {2:2.2}, Acc {3:2.2}'.format( - event.batch_id + 1, avg_cost, accuracy)) + print('Loss {0:2.2}, Acc {1:2.2}'.format(avg_cost, accuracy)) - if accuracy > 0.01: # Low threshold for speeding up CI - trainer.params.save(save_path) - return + if accuracy > 0.01: # Low threshold for speeding up CI + if save_dirname is not None: + trainer.save_params(save_dirname) + return place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( - train_network, - optimizer=fluid.optimizer.Adam(learning_rate=0.001), + train_func=train_program, place=place, - event_handler=event_handler) - trainer.train(train_reader, EPOCH_NUM, event_handler=event_handler) + optimizer=fluid.optimizer.Adam(learning_rate=0.001)) + + trainer.train( + reader=train_reader, + num_epochs=1, + event_handler=event_handler, + feed_order=['pixel', 'label']) -def infer(use_cuda, save_path): - params = fluid.Params(save_path) +def infer(use_cuda, inference_program, save_dirname=None): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - inferencer = fluid.Inferencer(inference_network, params, place=place) + inferencer = fluid.Inferencer( + infer_func=inference_program, param_path=save_dirname, place=place) # The input's dimension of conv should be 4-D or 5-D. # Use normilized image pixels as input data, which should be in the range @@ -114,8 +117,14 @@ def main(use_cuda): if use_cuda and not fluid.core.is_compiled_with_cuda(): return save_path = "image_classification_vgg.inference.model" - train(use_cuda, save_path) - infer(use_cuda, save_path) + + train( + use_cuda=use_cuda, train_program=train_network, save_dirname=save_path) + + infer( + use_cuda=use_cuda, + inference_program=inference_network, + save_dirname=save_path) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/book/label_semantic_roles/no_test_label_semantic_roles.py b/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/no_test_label_semantic_roles.py similarity index 100% rename from python/paddle/fluid/tests/book/label_semantic_roles/no_test_label_semantic_roles.py rename to python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/no_test_label_semantic_roles.py diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..673c965b662a022739f8d489c331f4de9455a926 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/CMakeLists.txt @@ -0,0 +1,7 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_conv.py b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py similarity index 58% rename from python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_conv.py rename to python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py index a8282c71f895718930ea14a1e0bff777441c4c57..2128d4c5b87434ebe30930dc0e338b3b50d921c2 100644 --- a/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_conv.py +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py @@ -21,7 +21,6 @@ import unittest import math import sys import os -import paddle.v2.dataset as dataset BATCH_SIZE = 64 @@ -55,46 +54,57 @@ def train_program(): cost = fluid.layers.cross_entropy(input=predict, label=label) avg_cost = fluid.layers.mean(cost) acc = fluid.layers.accuracy(input=predict, label=label) - return avg_cost, acc + return [avg_cost, acc] -def train(use_cuda, save_dirname): +def train(use_cuda, train_program, save_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.Adam(learning_rate=0.001) - trainer = fluid.Trainer(train_program, place=place, optimizer=optimizer) + + trainer = fluid.Trainer( + train_func=train_program, + place=place, + optimizer=optimizer, + parallel=True) def event_handler(event): - if isinstance(event, fluid.EndIteration): - avg_cost, acc = event.values + if isinstance(event, fluid.EndEpochEvent): + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=BATCH_SIZE) + avg_cost, acc = trainer.test( + reader=test_reader, feed_order=['img', 'label']) + print("avg_cost: %s" % avg_cost) print("acc : %s" % acc) - if (event.batch_id + 1) % 10 == 0: - test_metrics = trainer.test(reader=dataset.mnist.test()) - avg_cost_set = test_metrics[0] - acc_set = test_metrics[1] - - # get test acc and loss - acc = numpy.array(acc_set).mean() - avg_cost = numpy.array(avg_cost_set).mean() - if float(acc) > 0.2: # Smaller value to increase CI speed - trainer.save_params(save_dirname) - else: - print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format( - event.batch_id + 1, float(avg_cost), float(acc))) - if math.isnan(float(avg_cost)): - sys.exit("got NaN loss, training failed.") + if acc > 0.2: # Smaller value to increase CI speed + trainer.save_params(save_dirname) + else: + print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format( + event.epoch + 1, avg_cost, acc)) + if math.isnan(avg_cost): + sys.exit("got NaN loss, training failed.") + elif isinstance(event, fluid.EndStepEvent): + print("Step {0}, Epoch {1} Metrics {2}".format( + event.step, event.epoch, map(numpy.array, event.metrics))) + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=500), + batch_size=BATCH_SIZE) trainer.train( - reader=dataset.mnist.train(), num_pass=100, event_handler=event_handler) + num_epochs=1, + event_handler=event_handler, + reader=train_reader, + feed_order=['img', 'label']) -def infer(use_cuda, save_dirname=None): +def infer(use_cuda, inference_program, save_dirname=None): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() inferencer = fluid.Inferencer( - inference_program, param_path=save_dirname, place=place) + infer_func=inference_program, param_path=save_dirname, place=place) batch_size = 1 tensor_img = numpy.random.uniform(-1.0, 1.0, @@ -102,17 +112,23 @@ def infer(use_cuda, save_dirname=None): results = inferencer.infer({'img': tensor_img}) - print("infer results: ", results[0]) + print("infer results: ", numpy.array(results[0])) def main(use_cuda): save_dirname = "recognize_digits_conv.inference.model" # call train() with is_local argument to run distributed train - train(use_cuda=use_cuda, save_dirname=save_dirname) - infer(use_cuda=use_cuda, save_dirname=save_dirname) + train( + use_cuda=use_cuda, + train_program=train_program, + save_dirname=save_dirname) + infer( + use_cuda=use_cuda, + inference_program=inference_program, + save_dirname=save_dirname) if __name__ == '__main__': - for use_cuda in (False, True): - main(use_cuda=use_cuda) + # for use_cuda in (False, True): + main(use_cuda=True) diff --git a/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_mlp.py b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py similarity index 58% rename from python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_mlp.py rename to python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py index 3efa931d5886e565d2d876a31309883ee1660389..041c8d778e5c03aa68dad6ef450934f09c8d2a52 100644 --- a/python/paddle/fluid/tests/book/notest_recognize_digits/notest_recognize_digits_mlp.py +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py @@ -21,7 +21,6 @@ import unittest import math import sys import os -import paddle.v2.dataset as dataset BATCH_SIZE = 64 @@ -42,46 +41,51 @@ def train_program(): cost = fluid.layers.cross_entropy(input=predict, label=label) avg_cost = fluid.layers.mean(cost) acc = fluid.layers.accuracy(input=predict, label=label) - return avg_cost, acc + return [avg_cost, acc] -def train(use_cuda, save_dirname): +def train(use_cuda, train_program, save_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.Adam(learning_rate=0.001) - trainer = fluid.Trainer(train_program, place=place, optimizer=optimizer) + + trainer = fluid.Trainer( + train_func=train_program, place=place, optimizer=optimizer) def event_handler(event): - if isinstance(event, fluid.EndIteration): - avg_cost, acc = event.values + if isinstance(event, fluid.EndEpochEvent): + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=BATCH_SIZE) + avg_cost, acc = trainer.test( + reader=test_reader, feed_order=['img', 'label']) + print("avg_cost: %s" % avg_cost) print("acc : %s" % acc) - if (event.batch_id + 1) % 10 == 0: - test_metrics = trainer.test(reader=dataset.mnist.test()) - avg_cost_set = test_metrics[0] - acc_set = test_metrics[1] - - # get test acc and loss - acc = numpy.array(acc_set).mean() - avg_cost = numpy.array(avg_cost_set).mean() - if float(acc) > 0.2: # Smaller value to increase CI speed - trainer.save_params(save_dirname) - else: - print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format( - event.batch_id + 1, float(avg_cost), float(acc))) - if math.isnan(float(avg_cost)): - sys.exit("got NaN loss, training failed.") + if acc > 0.2: # Smaller value to increase CI speed + trainer.save_params(save_dirname) + else: + print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format( + event.epoch + 1, avg_cost, acc)) + if math.isnan(avg_cost): + sys.exit("got NaN loss, training failed.") + + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=500), + batch_size=BATCH_SIZE) trainer.train( - reader=dataset.mnist.train(), num_pass=100, event_handler=event_handler) + num_epochs=1, + event_handler=event_handler, + reader=train_reader, + feed_order=['img', 'label']) -def infer(use_cuda, save_dirname=None): +def infer(use_cuda, inference_program, save_dirname=None): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() inferencer = fluid.Inferencer( - inference_program, param_path=save_dirname, place=place) + infer_func=inference_program, param_path=save_dirname, place=place) batch_size = 1 tensor_img = numpy.random.uniform(-1.0, 1.0, @@ -89,17 +93,23 @@ def infer(use_cuda, save_dirname=None): results = inferencer.infer({'img': tensor_img}) - print("infer results: ", results[0]) + print("infer results: ", numpy.array(results[0])) def main(use_cuda): save_dirname = "recognize_digits_mlp.inference.model" # call train() with is_local argument to run distributed train - train(use_cuda=use_cuda, save_dirname=save_dirname) - infer(use_cuda=use_cuda, save_dirname=save_dirname) + train( + use_cuda=use_cuda, + train_program=train_program, + save_dirname=save_dirname) + infer( + use_cuda=use_cuda, + inference_program=inference_program, + save_dirname=save_dirname) if __name__ == '__main__': - for use_cuda in (False, True): - main(use_cuda=use_cuda) + # for use_cuda in (False, True): + main(use_cuda=False) diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..673c965b662a022739f8d489c331f4de9455a926 --- /dev/null +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/CMakeLists.txt @@ -0,0 +1,7 @@ +file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") +string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") + +# default test +foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) +endforeach() diff --git a/python/paddle/fluid/tests/book/understand_sentiment/notest_understand_sentiment_stacked_lstm.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py similarity index 63% rename from python/paddle/fluid/tests/book/understand_sentiment/notest_understand_sentiment_stacked_lstm.py rename to python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py index 9948e5c0234ed78237c94f9a25d6401619267d0d..0d7cbe3874cbc0c2def9d0032737f81e662296d6 100644 --- a/python/paddle/fluid/tests/book/understand_sentiment/notest_understand_sentiment_stacked_lstm.py +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py @@ -17,11 +17,13 @@ from __future__ import print_function import paddle import paddle.fluid as fluid from functools import partial +import numpy as np CLASS_DIM = 2 EMB_DIM = 128 HID_DIM = 512 STACKED_NUM = 3 +BATCH_SIZE = 128 def stacked_lstm_net(data, input_dim, class_dim, emb_dim, hid_dim, stacked_num): @@ -50,7 +52,7 @@ def stacked_lstm_net(data, input_dim, class_dim, emb_dim, hid_dim, stacked_num): return prediction -def inference_network(word_dict): +def inference_program(word_dict): data = fluid.layers.data( name="words", shape=[1], dtype="int64", lod_level=1) @@ -60,57 +62,71 @@ def inference_network(word_dict): return net -def train_network(word_dict): - prediction = inference_network(word_dict) +def train_program(word_dict): + prediction = inference_program(word_dict) label = fluid.layers.data(name="label", shape=[1], dtype="int64") cost = fluid.layers.cross_entropy(input=prediction, label=label) avg_cost = fluid.layers.mean(cost) accuracy = fluid.layers.accuracy(input=prediction, label=label) - return avg_cost, accuracy + return [avg_cost, accuracy] -def train(use_cuda, save_path): - BATCH_SIZE = 128 - EPOCH_NUM = 5 +def train(use_cuda, train_program, save_dirname): + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + optimizer = fluid.optimizer.Adagrad(learning_rate=0.002) word_dict = paddle.dataset.imdb.word_dict() + trainer = fluid.Trainer( + train_func=partial(train_program, word_dict), + place=place, + optimizer=optimizer) - train_data = paddle.batch( + def event_handler(event): + if isinstance(event, fluid.EndEpochEvent): + test_reader = paddle.batch( + paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE) + avg_cost, acc = trainer.test( + reader=test_reader, feed_order=['words', 'label']) + + print("avg_cost: %s" % avg_cost) + print("acc : %s" % acc) + + if acc > 0.2: # Smaller value to increase CI speed + trainer.save_params(save_dirname) + trainer.stop() + + else: + print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format( + event.epoch + 1, avg_cost, acc)) + if math.isnan(avg_cost): + sys.exit("got NaN loss, training failed.") + elif isinstance(event, fluid.EndStepEvent): + print("Step {0}, Epoch {1} Metrics {2}".format( + event.step, event.epoch, map(np.array, event.metrics))) + if event.step == 1: # Run 2 iterations to speed CI + trainer.save_params(save_dirname) + trainer.stop() + + train_reader = paddle.batch( paddle.reader.shuffle( - paddle.dataset.imdb.train(word_dict), buf_size=1000), + paddle.dataset.imdb.train(word_dict), buf_size=25000), batch_size=BATCH_SIZE) - test_data = paddle.batch( - paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE) - - def event_handler(event): - if isinstance(event, fluid.EndIteration): - if (event.batch_id % 10) == 0: - avg_cost, accuracy = trainer.test(reader=test_data) - - print('BatchID {1:04}, Loss {2:2.2}, Acc {3:2.2}'.format( - event.batch_id + 1, avg_cost, accuracy)) + trainer.train( + num_epochs=1, + event_handler=event_handler, + reader=train_reader, + feed_order=['words', 'label']) - if accuracy > 0.01: # Low threshold for speeding up CI - trainer.params.save(save_path) - return - place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - trainer = fluid.Trainer( - partial(train_network, word_dict), - optimizer=fluid.optimizer.Adagrad(learning_rate=0.002), - place=place, - event_handler=event_handler) - - trainer.train(train_data, EPOCH_NUM, event_handler=event_handler) - - -def infer(use_cuda, save_path): - params = fluid.Params(save_path) +def infer(use_cuda, inference_program, save_dirname=None): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() word_dict = paddle.dataset.imdb.word_dict() + inferencer = fluid.Inferencer( - partial(inference_network, word_dict), params, place=place) + infer_func=partial(inference_program, word_dict), + param_path=save_dirname, + place=place) def create_random_lodtensor(lod, place, low, high): data = np.random.random_integers(low, high, @@ -131,8 +147,8 @@ def main(use_cuda): if use_cuda and not fluid.core.is_compiled_with_cuda(): return save_path = "understand_sentiment_stacked_lstm.inference.model" - train(use_cuda, save_path) - infer(use_cuda, save_path) + train(use_cuda, train_program, save_path) + infer(use_cuda, inference_program, save_path) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/book/word2vec/no_test_word2vec_new_api.py b/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py similarity index 71% rename from python/paddle/fluid/tests/book/word2vec/no_test_word2vec_new_api.py rename to python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py index 35e163dc9df5a35ee5774b6b157366c4eabcb0f7..bf86cd9acf8da940fcc2fb5b594e33f9b6965acb 100644 --- a/python/paddle/fluid/tests/book/word2vec/no_test_word2vec_new_api.py +++ b/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py @@ -80,58 +80,71 @@ def inference_program(is_sparse): def train_program(is_sparse): - next_word = fluid.layers.data(name='nextw', shape=[1], dtype='int64') + # The declaration of 'next_word' must be after the invoking of inference_program, + # or the data input order of train program would be [next_word, firstw, secondw, + # thirdw, forthw], which is not correct. predict_word = inference_program(is_sparse) + next_word = fluid.layers.data(name='nextw', shape=[1], dtype='int64') cost = fluid.layers.cross_entropy(input=predict_word, label=next_word) avg_cost = fluid.layers.mean(cost) return avg_cost -def train(use_cuda, is_sparse, save_path): +def train(use_cuda, train_program, save_dirname): train_reader = paddle.batch( paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE) + test_reader = paddle.batch( + paddle.dataset.imikolov.test(word_dict, N), BATCH_SIZE) place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() def event_handler(event): - print type(event) - if isinstance(event, fluid.EndEpochEvent): - avg_cost = trainer.test(reader=paddle.dataset.imikolov.test( - word_dict, N)) - - if avg_cost < 5.0: - trainer.save_params(save_path) - return + if isinstance(event, fluid.EndStepEvent): + outs = trainer.test( + reader=test_reader, + feed_order=['firstw', 'secondw', 'thirdw', 'forthw', 'nextw']) + avg_cost = outs[0] + print("loss= ", avg_cost) + + if avg_cost < 10.0: + trainer.save_params(save_dirname) + trainer.stop() + if math.isnan(avg_cost): sys.exit("got NaN loss, training failed.") trainer = fluid.Trainer( - partial(train_program, is_sparse), - fluid.optimizer.SGD(learning_rate=0.001), + train_func=train_program, + optimizer=fluid.optimizer.SGD(learning_rate=0.001), place=place) + trainer.train( - reader=train_reader, num_epochs=100, event_handler=event_handler) + reader=train_reader, + num_epochs=1, + event_handler=event_handler, + feed_order=['firstw', 'secondw', 'thirdw', 'forthw', 'nextw']) -def infer(use_cuda, is_sparse, save_path): +def infer(use_cuda, inference_program, save_dirname=None): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() inferencer = fluid.Inferencer( - partial(inference_program, is_sparse), - param_path=save_path, - place=place) + infer_func=inference_program, param_path=save_dirname, place=place) lod = [0, 1] first_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1) second_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1) third_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1) fourth_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1) - result = inferencer.infer({ - 'firstw': first_word, - 'secondw': second_word, - 'thirdw': third_word, - 'forthw': fourth_word - }) - print(result) + + result = inferencer.infer( + { + 'firstw': first_word, + 'secondw': second_word, + 'thirdw': third_word, + 'forthw': fourth_word + }, + return_numpy=False) + print(np.array(result[0])) def main(use_cuda, is_sparse): @@ -139,8 +152,16 @@ def main(use_cuda, is_sparse): return save_path = "word2vec.inference.model" - train(use_cuda, is_sparse, save_path) - infer(use_cuda, is_sparse, save_path) + + train( + use_cuda=use_cuda, + train_program=partial(train_program, is_sparse), + save_dirname=save_path) + + infer( + use_cuda=use_cuda, + inference_program=partial(inference_program, is_sparse), + save_dirname=save_path) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/book/notest_understand_sentiment.py b/python/paddle/fluid/tests/book/notest_understand_sentiment.py index 241778e303036d068dc0a40e4574a02eb97ad134..792ed7368d646cd9dff9255eb402b6a9b84f69a6 100644 --- a/python/paddle/fluid/tests/book/notest_understand_sentiment.py +++ b/python/paddle/fluid/tests/book/notest_understand_sentiment.py @@ -170,7 +170,7 @@ def train(word_dict, assert save_dirname is None adagrad = fluid.optimizer.Adagrad(learning_rate=0.002) - optimize_ops, params_grads = adagrad.minimize(cost) + adagrad.minimize(cost) train_data = paddle.batch( paddle.reader.shuffle( diff --git a/python/paddle/fluid/tests/book/test_fit_a_line.py b/python/paddle/fluid/tests/book/test_fit_a_line.py index ecb34699af0dc14782601702ab8afedbca7e1bfd..b1a6b524d33cae97c8982ffb8f780b1b07761a09 100644 --- a/python/paddle/fluid/tests/book/test_fit_a_line.py +++ b/python/paddle/fluid/tests/book/test_fit_a_line.py @@ -33,7 +33,7 @@ def train(use_cuda, save_dirname, is_local): avg_cost = fluid.layers.mean(cost) sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) - optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) + sgd_optimizer.minimize(avg_cost) BATCH_SIZE = 20 diff --git a/python/paddle/fluid/tests/book/test_image_classification.py b/python/paddle/fluid/tests/book/test_image_classification.py index dbcdb5766e7d20efdb12da0ea4c6f005d903849b..0f3a4c9242a81a3c1fb90268245715a8e59a207a 100644 --- a/python/paddle/fluid/tests/book/test_image_classification.py +++ b/python/paddle/fluid/tests/book/test_image_classification.py @@ -125,7 +125,7 @@ def train(net_type, use_cuda, save_dirname, is_local): test_program = fluid.default_main_program().clone(for_test=True) optimizer = fluid.optimizer.Adam(learning_rate=0.001) - optimize_ops, params_grads = optimizer.minimize(avg_cost) + optimizer.minimize(avg_cost) BATCH_SIZE = 128 PASS_NUM = 1 diff --git a/python/paddle/fluid/tests/book/test_label_semantic_roles.py b/python/paddle/fluid/tests/book/test_label_semantic_roles.py index 50ef29c4572f1b12fe9793bbf037cd7fe71a9e53..f1ee5dfd99e1c8b26280c010c1aaca05a004a5b6 100644 --- a/python/paddle/fluid/tests/book/test_label_semantic_roles.py +++ b/python/paddle/fluid/tests/book/test_label_semantic_roles.py @@ -36,7 +36,7 @@ depth = 8 mix_hidden_lr = 1e-3 IS_SPARSE = True -PASS_NUM = 100 +PASS_NUM = 10 BATCH_SIZE = 10 embedding_name = 'emb' @@ -175,19 +175,13 @@ def train(use_cuda, save_dirname=None, is_local=True): decay_steps=100000, decay_rate=0.5, staircase=True)) - optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) + sgd_optimizer.minimize(avg_cost) # TODO(qiao) # add dependency track and move this config before optimizer crf_decode = fluid.layers.crf_decoding( input=feature_out, param_attr=fluid.ParamAttr(name='crfw')) - chunk_evaluator = fluid.evaluator.ChunkEvaluator( - input=crf_decode, - label=target, - chunk_scheme="IOB", - num_chunk_types=int(math.ceil((label_dict_len - 1) / 2.0))) - train_data = paddle.batch( paddle.reader.shuffle( paddle.dataset.conll05.test(), buf_size=8192), @@ -203,7 +197,6 @@ def train(use_cuda, save_dirname=None, is_local=True): def train_loop(main_program): exe.run(fluid.default_startup_program()) - embedding_param = fluid.global_scope().find_var( embedding_name).get_tensor() embedding_param.set( @@ -213,27 +206,19 @@ def train(use_cuda, save_dirname=None, is_local=True): start_time = time.time() batch_id = 0 for pass_id in xrange(PASS_NUM): - chunk_evaluator.reset(exe) for data in train_data(): - cost, precision, recall, f1_score = exe.run( - main_program, - feed=feeder.feed(data), - fetch_list=[avg_cost] + chunk_evaluator.metrics) - pass_precision, pass_recall, pass_f1_score = chunk_evaluator.eval( - exe) + cost = exe.run(main_program, + feed=feeder.feed(data), + fetch_list=[avg_cost]) + cost = cost[0] if batch_id % 10 == 0: - print("avg_cost:" + str(cost) + " precision:" + str( - precision) + " recall:" + str(recall) + " f1_score:" + - str(f1_score) + " pass_precision:" + str( - pass_precision) + " pass_recall:" + str( - pass_recall) + " pass_f1_score:" + str( - pass_f1_score)) + print("avg_cost:" + str(cost)) if batch_id != 0: print("second per batch: " + str((time.time( ) - start_time) / batch_id)) # Set the threshold low to speed up the CI test - if float(pass_precision) > 0.01: + if float(cost) < 60.0: if save_dirname is not None: # TODO(liuyiqun): Change the target to crf_decode fluid.io.save_inference_model(save_dirname, [ diff --git a/python/paddle/fluid/tests/book/test_machine_translation.py b/python/paddle/fluid/tests/book/test_machine_translation.py index 46c6b9c29a265741a99655d5ac29244798f6fec2..e8a75f473f62df528b7f39bf5f9085076e005c25 100644 --- a/python/paddle/fluid/tests/book/test_machine_translation.py +++ b/python/paddle/fluid/tests/book/test_machine_translation.py @@ -185,7 +185,7 @@ def train_main(use_cuda, is_sparse, is_local=True): learning_rate=1e-4, regularization=fluid.regularizer.L2DecayRegularizer( regularization_coeff=0.1)) - optimize_ops, params_grads = optimizer.minimize(avg_cost) + optimizer.minimize(avg_cost) train_data = paddle.batch( paddle.reader.shuffle( diff --git a/python/paddle/fluid/tests/book/test_recognize_digits.py b/python/paddle/fluid/tests/book/test_recognize_digits.py index c115aa4d7d6b514f9207543730e5e76cb0d2040c..578b1162fbd7e3a1b1c0cc934406818f2e07e019 100644 --- a/python/paddle/fluid/tests/book/test_recognize_digits.py +++ b/python/paddle/fluid/tests/book/test_recognize_digits.py @@ -95,7 +95,7 @@ def train(nn_type, test_program = fluid.default_main_program().clone(for_test=True) optimizer = fluid.optimizer.Adam(learning_rate=0.001) - optimize_ops, params_grads = optimizer.minimize(avg_loss) + optimizer.minimize(avg_loss) place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() diff --git a/python/paddle/fluid/tests/book/test_recommender_system.py b/python/paddle/fluid/tests/book/test_recommender_system.py index d022dedbff805d597b68b5a47f7931f2dd946615..7be924f762ddeb045dda890dbfdcd96a65449553 100644 --- a/python/paddle/fluid/tests/book/test_recommender_system.py +++ b/python/paddle/fluid/tests/book/test_recommender_system.py @@ -160,7 +160,7 @@ def train(use_cuda, save_dirname, is_local=True): test_program = fluid.default_main_program().clone(for_test=True) sgd_optimizer = SGDOptimizer(learning_rate=0.2) - optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) + sgd_optimizer.minimize(avg_cost) place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() diff --git a/python/paddle/fluid/tests/book/test_word2vec.py b/python/paddle/fluid/tests/book/test_word2vec.py index 6dec0f6857e86b4b9c1c67af934aa9bfdb1c3df7..30e1a5040cc92b02bbbf90dac97001812ec90134 100644 --- a/python/paddle/fluid/tests/book/test_word2vec.py +++ b/python/paddle/fluid/tests/book/test_word2vec.py @@ -101,7 +101,7 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True): avg_cost = fluid.layers.mean(pd()) sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) - optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) + sgd_optimizer.minimize(avg_cost) train_reader = paddle.batch( paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE) diff --git a/python/paddle/fluid/tests/test_data_feeder.py b/python/paddle/fluid/tests/test_data_feeder.py index 861dd3174a21d59fe12e0b794ecb2a934946ac71..ce3ba3ebc50d7b015f379b5e80b179463a7b231a 100644 --- a/python/paddle/fluid/tests/test_data_feeder.py +++ b/python/paddle/fluid/tests/test_data_feeder.py @@ -13,15 +13,62 @@ # limitations under the License. import paddle.fluid as fluid +import unittest -def test_converter(): - img = fluid.layers.data(name='image', shape=[1, 28, 28]) - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - feeder = fluid.DataFeeder([img, label], fluid.CPUPlace()) - result = feeder.feed([[[0] * 784, [9]], [[1] * 784, [1]]]) - print(result) +class TestDataFeeder(unittest.TestCase): + def test_lod_level_0_converter(self): + img = fluid.layers.data(name='image', shape=[1, 28, 28]) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + feeder = fluid.DataFeeder([img, label], fluid.CPUPlace()) + result = feeder.feed([([0] * 784, [9]), ([1] * 784, [1])]) + print(result) + + self.assertEqual(result['image'].shape(), [2, 1, 28, 28]) + self.assertEqual(result['label'].shape(), [2, 1]) + self.assertEqual(result['image'].lod(), []) + self.assertEqual(result['label'].lod(), []) + + def test_lod_level_1_converter(self): + # lod_level = 1 + # each sentence has a different number of words + sentences = fluid.layers.data( + name='sentences', shape=[1], dtype='int64', lod_level=1) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + feeder = fluid.DataFeeder([sentences, label], fluid.CPUPlace()) + + # lod = [[0, 3, 5, 9]] + # data = [[1, 2, 3], [4, 5], [6, 7, 8, 9]] + # label = [1] * len(data) + result = feeder.feed( + [([1, 2, 3], [1]), ([4, 5], [1]), ([6, 7, 8, 9], [1])]) + print(result) + + self.assertEqual(result['sentences'].shape(), [9, 1]) + self.assertEqual(result['label'].shape(), [3, 1]) + self.assertEqual(result['sentences'].lod(), [[0, 3, 5, 9]]) + self.assertEqual(result['label'].lod(), []) + + def test_lod_level_2_converter(self): + # lod_level = 2 + # paragraphs -> sentences -> words + paragraphs = fluid.layers.data( + name='paragraphs', shape=[1], dtype='int64', lod_level=2) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + feeder = fluid.DataFeeder([paragraphs, label], fluid.CPUPlace()) + + # lod = [[0, 2, 3], [0, 3, 5, 9]] + # data = [[[1, 2, 3], [4, 5]], [[6, 7, 8, 9]]] + # label = [1] * len(data) + result = feeder.feed( + [([[1, 2, 3], [4, 5]], [1]), ([[6, 7, 8, 9]], [1])]) + print(result) + + self.assertEqual(result['paragraphs'].shape(), [9, 1]) + self.assertEqual(result['label'].shape(), [2, 1]) + self.assertEqual(result['paragraphs'].lod(), [[0, 2, 3], [0, 3, 5, 9]]) + self.assertEqual(result['label'].lod(), []) if __name__ == '__main__': - test_converter() + unittest.main() diff --git a/python/paddle/fluid/tests/test_detection.py b/python/paddle/fluid/tests/test_detection.py index 921260ef3f4b1f9e4c65b3ffb440dc34cb0a9376..8569d838bdd414eb84c6c87674990a25a2fdcdf9 100644 --- a/python/paddle/fluid/tests/test_detection.py +++ b/python/paddle/fluid/tests/test_detection.py @@ -109,6 +109,24 @@ class TestDetection(unittest.TestCase): print(str(program)) +class TestPriorBox(unittest.TestCase): + def test_prior_box(self): + data_shape = [3, 224, 224] + images = fluid.layers.data( + name='pixel', shape=data_shape, dtype='float32') + conv1 = fluid.layers.conv2d(images, 3, 3, 2) + box, var = layers.prior_box( + input=conv1, + image=images, + min_sizes=[100.0], + aspect_ratios=[1.], + flip=True, + clip=True) + assert len(box.shape) == 4 + assert box.shape == var.shape + assert box.shape[3] == 4 + + class TestMultiBoxHead(unittest.TestCase): def test_multi_box_head(self): data_shape = [3, 224, 224] diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt index d9190408e151283ece8460286dd67818dd39da3e..2ae9653953c2f5f6a399243bef2c7fb756f9692f 100644 --- a/python/paddle/fluid/tests/unittests/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt @@ -28,11 +28,11 @@ function(py_test_modules TARGET_NAME) if(WITH_TESTING) set(options "") set(oneValueArgs "") - set(multiValueArgs MODULES DEPS ARGS ENVS) + set(multiValueArgs MODULES DEPS ENVS) cmake_parse_arguments(py_test_modules "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) add_test(NAME ${TARGET_NAME} COMMAND env PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_modules_ENVS} - ${PYTHON_EXECUTABLE} -u -m unittest --verbose ${py_test_modules_MODULES} ${py_test_modules_ARGS} + ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/tools/test_runner.py ${py_test_modules_MODULES} WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) endif() endfunction() @@ -66,6 +66,7 @@ list(REMOVE_ITEM TEST_OPS test_fetch_var) list(REMOVE_ITEM TEST_OPS test_parallel_op) list(REMOVE_ITEM TEST_OPS test_dynrnn_static_input) list(REMOVE_ITEM TEST_OPS test_dist_train) +list(REMOVE_ITEM TEST_OPS test_network_with_dtype) # tests that can be bundled together in one python process for speed. if(WITH_FAST_BUNDLE_TEST) @@ -83,6 +84,7 @@ py_test_modules(test_parallel_executor MODULES test_parallel_executor) py_test_modules(test_warpctc_op MODULES test_warpctc_op ENVS FLAGS_warpctc_dir=${WARPCTC_LIB_DIR}) py_test_modules(test_train_dyn_rnn MODULES test_dyn_rnn) py_test_modules(test_mul_op MODULES test_mul_op) +py_test_modules(test_network_with_dtype MODULES test_network_with_dtype) # tests that need to be run in separate process. py_test_modules(test_multihead_attention MODULES test_multihead_attention) diff --git a/python/paddle/fluid/tests/unittests/test_detection_map_op.py b/python/paddle/fluid/tests/unittests/test_detection_map_op.py index a905a854ad157ffa3d7816dfbd445f3e344a1249..f545ad155ccd28c2d34e424d307eed49b37f20fb 100644 --- a/python/paddle/fluid/tests/unittests/test_detection_map_op.py +++ b/python/paddle/fluid/tests/unittests/test_detection_map_op.py @@ -160,7 +160,9 @@ class TestDetectionMAPOp(OpTest): label_count, true_pos, false_pos = get_input_pos( self.class_pos_count, self.true_pos, self.true_pos_lod, self.false_pos, self.false_pos_lod) - for (label, difficult, xmin, ymin, xmax, ymax) in self.label: + for v in self.label: + label = v[0] + difficult = False if len(v) == 5 else v[1] if self.evaluate_difficult: label_count[label] += 1 elif not difficult: @@ -245,6 +247,15 @@ class TestDetectionMAPOpSkipDiff(TestDetectionMAPOp): [2, 0.8, 0, 1], [2, 0.1, 1, 0], [3, 0.2, 0, 1]] +class TestDetectionMAPOpWithoutDiff(TestDetectionMAPOp): + def init_test_case(self): + super(TestDetectionMAPOpWithoutDiff, self).init_test_case() + + # label xmin ymin xmax ymax + self.label = [[1, 0.1, 0.1, 0.3, 0.3], [1, 0.6, 0.6, 0.8, 0.8], + [2, 0.3, 0.3, 0.6, 0.5], [1, 0.7, 0.1, 0.9, 0.3]] + + class TestDetectionMAPOp11Point(TestDetectionMAPOp): def init_test_case(self): super(TestDetectionMAPOp11Point, self).init_test_case() diff --git a/python/paddle/fluid/tests/unittests/test_dist_train.py b/python/paddle/fluid/tests/unittests/test_dist_train.py index 77e9a8f7e72a9e0790ce1d1f48356abcca8eaccf..c2393a288c6ebb5dd4a12f7b591d12cc94f4ea55 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_train.py +++ b/python/paddle/fluid/tests/unittests/test_dist_train.py @@ -52,15 +52,18 @@ class TestSendOp(unittest.TestCase): serv = layers.ListenAndServ( "127.0.0.1:0", ["X"], optimizer_mode=False) with serv.do(): + out_var = main.global_block().create_var( + name="scale_0.tmp_0", + psersistable=True, + dtype="float32", + shape=[32, 32]) x = layers.data( shape=[32, 32], dtype='float32', name="X", append_batch_size=False) fluid.initializer.Constant(value=1.0)(x, main.global_block()) - o = layers.scale(x=x, scale=10.0) - main.global_block().create_var( - name=o.name, psersistable=False, dtype=o.dtype, shape=o.shape) + layers.scale(x=x, scale=10.0, out=out_var) self.server_exe = fluid.Executor(place) self.server_exe.run(main) diff --git a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py new file mode 100644 index 0000000000000000000000000000000000000000..10f8c4f3f0167632bb4a3d454ab026ba73a8f305 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py @@ -0,0 +1,113 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest + +import paddle.fluid as fluid +import paddle.fluid.core as core +import paddle.fluid.layers as layers +from paddle.fluid.transpiler.distribute_transpiler import delete_ops +import numpy + + +class TestDistTranspiler(unittest.TestCase): + def setUp(self): + self.trainer_id = 0 + self.trainers = 2 + self.pservers = 2 + self.pserver_eps = "127.0.0.1:6174,127.0.0.1:6175" + self.current_pserver_ep = "127.0.0.1:6174" + + def net_conf(self): + x = fluid.layers.data(name='x', shape=[1000], dtype='float32') + + y_predict = fluid.layers.fc(input=x, + size=1000, + act=None, + param_attr=fluid.ParamAttr(name='fc_w')) + + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1) + + optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) + return optimize_ops, params_grads + + def test_transpiler(self): + trainer = self.get_trainer() + pserver, startup = self.get_pserver(self.current_pserver_ep) + + self.assertEqual([op.type for op in trainer.global_block().ops], + self.get_expect_trainer_ops()) + + self.assertEqual(len(pserver.blocks), 3) + # block0: listen_and_serv + self.assertEqual([op.type for op in pserver.blocks[0].ops], + ["listen_and_serv"]) + # block2: optimize pass + self.assertEqual([op.type for op in pserver.blocks[1].ops], + ["sum", "scale", "sgd"]) + + # confirm startup program + + self.assertEqual([op.type for op in startup.global_block().ops], [ + "fill_constant", "fill_constant", "uniform_random", "uniform_random" + ]) + + # the variable #fc_w will be split into two blocks + fc_w_var = startup.global_block().var("fc_w.block1") + self.assertEqual(fc_w_var.shape, (500, 1000)) + + def get_main_program(self): + main = fluid.Program() + + with fluid.program_guard(main): + self.net_conf() + + return main + + def get_expect_trainer_ops(self): + trainer = fluid.Program() + + with fluid.program_guard(trainer): + optimize_ops, params_grads = self.net_conf() + + delete_ops(trainer.global_block(), optimize_ops) + return [op.type for op in trainer.global_block().ops + ] + ["split_byref", "send", "concat"] + + def get_trainer(self): + return self._transpiler_instance().get_trainer_program() + + def get_pserver(self, ep): + t = self._transpiler_instance() + pserver = t.get_pserver_program(ep) + startup = t.get_startup_program(ep, pserver) + return pserver, startup + + def _transpiler_instance(self): + main = self.get_main_program() + t = fluid.DistributeTranspiler() + t.transpile( + self.trainer_id, + program=main, + pservers=self.pserver_eps, + trainers=self.trainers) + return t + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_is_empty_op.py b/python/paddle/fluid/tests/unittests/test_is_empty_op.py index 4d11cf226be2ba4ffbe015198fed3191f1e02f72..11121d9b65351eab639b7618fac0e54714cf4680 100644 --- a/python/paddle/fluid/tests/unittests/test_is_empty_op.py +++ b/python/paddle/fluid/tests/unittests/test_is_empty_op.py @@ -14,42 +14,24 @@ import unittest import numpy as np -from paddle.fluid.op import Operator -import paddle.fluid.core as core +from op_test import OpTest -def create_tensor(scope, name, np_data): - tensor = scope.var(name).get_tensor() - tensor.set_dims(np_data.shape) - tensor.set(np_data, core.CPUPlace()) - return tensor - - -class TestIsEmptyOp(unittest.TestCase): +class TestEmpty(OpTest): def setUp(self): - self.scope = core.Scope() - # create input variables - np_data0 = np.array([0, 1, 2]) - create_tensor(self.scope, "X0", np_data0) - - np_data1 = np.array([1]) - t = create_tensor(self.scope, "X1", np_data1) - t.set_dims([0]) + self.op_type = "is_empty" + self.inputs = {'X': np.array([1, 2, 3])} + self.outputs = {'Out': np.array([False])} - # create output variables - self.scope.var("out") + def test_check_output(self): + self.check_output() - def test_no_empty(self): - self.one_case("X0", False) - def test_empty(self): - self.one_case("X1", True) - - def one_case(self, input, target): - op = Operator(type="is_empty", X=input, Out="out") - op.run(self.scope, core.CPUPlace()) - out = self.scope.var("out").get_tensor() - self.assertEqual(np.array(out)[0], target) +class TestNotEmpty(TestEmpty): + def setUp(self): + self.op_type = "is_empty" + self.inputs = {'X': np.array([])} + self.outputs = {'Out': np.array([True])} if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_matmul_op.py b/python/paddle/fluid/tests/unittests/test_matmul_op.py index 44ac4683891ffd3141a126740f4fddb47550e183..cae2c8fa87d9857de8f26cf4962d9370eca66243 100644 --- a/python/paddle/fluid/tests/unittests/test_matmul_op.py +++ b/python/paddle/fluid/tests/unittests/test_matmul_op.py @@ -111,21 +111,24 @@ class Generator(object): # Generate test cases for all possibilities -for dim_X in [1, 2, 3]: - for dim_Y in [1, 2, 3]: - for transpose_X in [False, True]: - for transpose_Y in [False, True]: - test_name = ( - 'TestMatMulOp_dimX_{}_dim_Y_{}_transX_{}_transY_{}'.format( - dim_X, dim_Y, transpose_X, transpose_Y)) - shape_X, shape_Y = generate_compatible_shapes( - dim_X, dim_Y, transpose_X, transpose_Y) - globals()[test_name] = type(test_name, (Generator, OpTest), { - 'shape_X': shape_X, - 'shape_Y': shape_Y, - 'transpose_X': transpose_X, - 'transpose_Y': transpose_Y, - }) +def inject_test(dim_x, dim_y, trans_x, trans_y): + test_name = ('TestMatMulOp_dimX_{}_dim_Y_{}_transX_{}_transY_{}'.format( + dim_x, dim_y, trans_x, trans_y)) + shape_x, shape_y = generate_compatible_shapes(dim_x, dim_y, trans_x, + trans_y) + globals()[test_name] = type(test_name, (Generator, OpTest), { + 'shape_X': shape_x, + 'shape_Y': shape_y, + 'transpose_X': trans_x, + 'transpose_Y': trans_y, + }) + + +for dim_X in (1, 2, 3): + for dim_Y in (1, 2, 3): + for transose_x in (False, True): + for transose_y in (False, True): + inject_test(dim_X, dim_Y, transose_x, transose_y) # Test case n-dim @@ -149,7 +152,7 @@ def generate_compatible_shapes(dim, transpose_X, transpose_Y): return shape_X, shape_Y -# Test case n-dim +# # Test case n-dim for dim in [4]: for transpose_X in [False, True]: for transpose_Y in [False, True]: diff --git a/python/paddle/fluid/tests/unittests/test_network_with_dtype.py b/python/paddle/fluid/tests/unittests/test_network_with_dtype.py new file mode 100644 index 0000000000000000000000000000000000000000..d4835dd18405fc7a0d508a780a734922e0abd12c --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_network_with_dtype.py @@ -0,0 +1,74 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest + +import numpy as np +import paddle +import paddle.fluid as fluid +import paddle.fluid.core as core +from paddle.fluid.executor import Executor + +BATCH_SIZE = 20 + + +class TestNetWithDtype(unittest.TestCase): + def setUp(self): + self.dtype = "float64" + self.init_dtype() + + def run_net_on_place(self, place): + main = fluid.Program() + startup = fluid.Program() + with fluid.program_guard(main, startup): + x = fluid.layers.data(name='x', shape=[13], dtype=self.dtype) + y = fluid.layers.data(name='y', shape=[1], dtype=self.dtype) + y_predict = fluid.layers.fc(input=x, size=1, act=None) + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) + sgd_optimizer.minimize(avg_cost) + + fetch_list = [avg_cost] + train_reader = paddle.batch( + paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE) + feeder = fluid.DataFeeder(place=place, feed_list=[x, y]) + exe = fluid.Executor(place) + exe.run(startup) + for data in train_reader(): + exe.run(main, feed=feeder.feed(data), fetch_list=fetch_list) + # the main program is runable, the datatype is fully supported + break + + def init_dtype(self): + pass + + def test_cpu(self): + place = fluid.CPUPlace() + self.run_net_on_place(place) + + def test_gpu(self): + if not core.is_compiled_with_cuda(): + return + place = fluid.CUDAPlace(0) + self.run_net_on_place(place) + + +# TODO(dzhwinter): make sure the fp16 is runable +# class TestFloat16(TestNetWithDtype): +# def init_dtype(self): +# self.dtype = "float16" + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor.py b/python/paddle/fluid/tests/unittests/test_parallel_executor.py index 4eb25a6e00b7564ac17db568ec78c1c84933af43..056f9e1781997aa1586d972874b652d5b725fe3f 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor.py @@ -205,7 +205,8 @@ class TestParallelExecutorBase(unittest.TestCase): allow_op_delay=False, feed_dict=None, seed=None, - use_parallel_executor=True): + use_parallel_executor=True, + balance_parameter_opt_between_cards=False): def run_executor(exe, feed, fetch_list, program=None): if isinstance(exe, fluid.ParallelExecutor): res = exe.run(fetch_list=fetch_list, feed=feed) @@ -231,10 +232,18 @@ class TestParallelExecutorBase(unittest.TestCase): place = fluid.CUDAPlace(0) startup_exe = fluid.Executor(place) startup_exe.run(startup) + exec_strategy = fluid.ExecutionStrategy() + exec_strategy.allow_op_delay = allow_op_delay + + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce if balance_parameter_opt_between_cards else fluid.BuildStrategy.ReduceStrategy.AllReduce if use_parallel_executor: exe = fluid.ParallelExecutor( - True, loss_name=loss.name, allow_op_delay=allow_op_delay) + True, + loss_name=loss.name, + exec_strategy=exec_strategy, + build_strategy=build_strategy) else: exe = fluid.Executor(place=place) @@ -280,20 +289,27 @@ class TestMNIST(TestParallelExecutorBase): fluid.recordio_writer.convert_reader_to_recordio_file( './mnist.recordio', reader, feeder) - def check_simple_fc_convergence(self): + def check_simple_fc_convergence(self, balance_parameter_opt_between_cards): self.check_network_convergence(simple_fc_net) self.check_network_convergence(simple_fc_net, allow_op_delay=True) img = np.zeros(shape=[32, 784], dtype='float32') label = np.ones(shape=[32, 1], dtype='int64') self.check_network_convergence( - simple_fc_net, feed_dict={"image": img, - "label": label}) + simple_fc_net, + feed_dict={"image": img, + "label": label}, + balance_parameter_opt_between_cards=balance_parameter_opt_between_cards + ) def test_simple_fc(self): - self.check_simple_fc_convergence() + self.check_simple_fc_convergence(False) + + def test_simple_fc_with_new_strategy(self): + self.check_simple_fc_convergence(True) - def check_simple_fc_parallel_accuracy(self): + def check_simple_fc_parallel_accuracy(self, + balance_parameter_opt_between_cards): img = np.zeros(shape=[32, 784], dtype='float32') label = np.ones(shape=[32, 1], dtype='int64') single_first_loss, single_last_loss = self.check_network_convergence( @@ -307,7 +323,9 @@ class TestMNIST(TestParallelExecutorBase): seed=1000, feed_dict={"image": img, "label": label}, - use_parallel_executor=True) + use_parallel_executor=True, + balance_parameter_opt_between_cards=balance_parameter_opt_between_cards + ) for p_f in parallel_first_loss: self.assertAlmostEquals(p_f, single_first_loss[0], delta=1e-6) @@ -315,18 +333,28 @@ class TestMNIST(TestParallelExecutorBase): self.assertAlmostEquals(p_l, single_last_loss[0], delta=1e-6) def test_simple_fc_parallel_accuracy(self): - self.check_simple_fc_parallel_accuracy() + self.check_simple_fc_parallel_accuracy(False) - def check_batchnorm_fc_convergence(self): + def test_simple_fc_parallel_accuracy_with_new_strategy(self): + self.check_simple_fc_parallel_accuracy(True) + + def check_batchnorm_fc_convergence(self, + balance_parameter_opt_between_cards): self.check_network_convergence(fc_with_batchnorm) img = np.zeros(shape=[32, 784], dtype='float32') label = np.ones(shape=[32, 1], dtype='int64') self.check_network_convergence( - fc_with_batchnorm, feed_dict={"image": img, - "label": label}) + fc_with_batchnorm, + feed_dict={"image": img, + "label": label}, + balance_parameter_opt_between_cards=balance_parameter_opt_between_cards + ) def test_batchnorm_fc(self): - self.check_batchnorm_fc_convergence() + self.check_batchnorm_fc_convergence(False) + + def test_batchnorm_fc_with_new_strategy(self): + self.check_batchnorm_fc_convergence(True) class TestResnet(TestParallelExecutorBase): @@ -348,17 +376,22 @@ class TestResnet(TestParallelExecutorBase): # fluid.recordio_writer.convert_reader_to_recordio_file( # "./flowers.recordio", reader, feeder, compressor=fluid.core.RecordIOWriter.Compressor.NoCompress) - def check_resnet_convergence(self): + def check_resnet_convergence(self, balance_parameter_opt_between_cards): import functools batch_size = 2 self.check_network_convergence( functools.partial( SE_ResNeXt50Small, batch_size=batch_size), iter=20, - batch_size=batch_size) + batch_size=batch_size, + balance_parameter_opt_between_cards=balance_parameter_opt_between_cards + ) def test_resnet(self): - self.check_resnet_convergence() + self.check_resnet_convergence(False) + + def test_resnet_with_new_strategy(self): + self.check_resnet_convergence(True) class ModelHyperParams(object): @@ -519,7 +552,7 @@ class TestTransformer(TestParallelExecutorBase): class ParallelExecutorTestingDuringTraining(unittest.TestCase): - def check_network_convergence(self): + def check_network_convergence(self, build_strategy=None): main = fluid.Program() startup = fluid.Program() with fluid.program_guard(main, startup): @@ -539,12 +572,16 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase): feed_dict = {'image': image, 'label': label} train_exe = fluid.ParallelExecutor( - use_cuda=True, loss_name=loss.name, main_program=main) + use_cuda=True, + loss_name=loss.name, + main_program=main, + build_strategy=build_strategy) test_exe = fluid.ParallelExecutor( use_cuda=True, main_program=test_program, - share_vars_from=train_exe) + share_vars_from=train_exe, + build_strategy=build_strategy) for i in xrange(5): test_loss, = test_exe.run([loss.name], feed=feed_dict) @@ -558,8 +595,15 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase): "Train loss: " + str(train_loss) + "\n Test loss:" + str(test_loss)) - def test_parallel(self): - self.check_network_convergence() + def test_parallel_testing(self): + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce + self.check_network_convergence(build_strategy) + + def test_parallel_testing_with_new_strategy(self): + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce + self.check_network_convergence(build_strategy) import paddle.dataset.conll05 as conll05 @@ -648,7 +692,7 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, class TestCRFModel(unittest.TestCase): - def check_network_convergence(self, is_sparse): + def check_network_convergence(self, is_sparse, build_strategy=None): main = fluid.Program() startup = fluid.Program() with fluid.program_guard(main, startup): @@ -696,7 +740,10 @@ class TestCRFModel(unittest.TestCase): exe = fluid.Executor(place) exe.run(startup) - pe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name) + pe = fluid.ParallelExecutor( + use_cuda=True, + loss_name=avg_cost.name, + build_strategy=build_strategy) feeder = fluid.DataFeeder( feed_list=[ @@ -712,11 +759,29 @@ class TestCRFModel(unittest.TestCase): pe.run(feed=feeder.feed(cur_batch), fetch_list=[avg_cost.name]))[0] - def test_update_sparse_parameter(self): - self.check_network_convergence(is_sparse=True) + def test_update_sparse_parameter_all_reduce(self): + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce + self.check_network_convergence( + is_sparse=True, build_strategy=build_strategy) + + def test_update_dense_parameter_all_reduce(self): + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce + self.check_network_convergence( + is_sparse=False, build_strategy=build_strategy) + + def test_update_sparse_parameter_reduce(self): + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce + self.check_network_convergence( + is_sparse=True, build_strategy=build_strategy) - def test_update_dense_parameter(self): - self.check_network_convergence(is_sparse=False) + def test_update_dense_parameter_reduce(self): + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce + self.check_network_convergence( + is_sparse=False, build_strategy=build_strategy) # test fetch all the variables of global_block @@ -784,7 +849,7 @@ class TestFetchOp(unittest.TestCase): assert not math.isnan(np.sum(ret[i])) and \ not math.isinf(np.sum(ret[i])) - def test_update_sparse_parameter(self): + def test_fetch_op(self): tst_reader = paddle.batch(flowers.test(use_xmap=False), batch_size=16) tst_reader_iter = tst_reader() @@ -796,5 +861,42 @@ class TestFetchOp(unittest.TestCase): self.parallel_exe(train_inputs, seed=1) +class TestFeedParallel(unittest.TestCase): + def test_main(self): + main = fluid.Program() + startup = fluid.Program() + startup.random_seed = 1 + with fluid.scope_guard(fluid.core.Scope()): + with fluid.program_guard(main, startup): + data = fluid.layers.data( + name='image', shape=[3, 224, 224], dtype='float32') + label = fluid.layers.data( + name='label', shape=[1], dtype='int64') + out = Lenet(data, class_dim=102) + loss = fluid.layers.cross_entropy(input=out, label=label) + loss = fluid.layers.mean(loss) + opt = fluid.optimizer.Momentum( + learning_rate=0.1, + momentum=0.9, + regularization=fluid.regularizer.L2Decay(1e-4)) + + opt.minimize(loss) + place = fluid.CUDAPlace(0) + feeder = fluid.DataFeeder(place=place, feed_list=[data, label]) + reader = feeder.decorate_reader( + paddle.batch( + flowers.train(), batch_size=16), multi_devices=True) + exe = fluid.Executor(place) + exe.run(startup) + pe = fluid.ParallelExecutor( + use_cuda=True, loss_name=loss.name, main_program=main) + + for batch_id, data in enumerate(reader()): + loss_np = np.array(pe.run(feed=data, fetch_list=[loss.name])[0]) + print batch_id, loss_np + if batch_id == 2: + break + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_split_var.py b/python/paddle/fluid/tests/unittests/test_split_var.py index 79d387f0066672058d1640f4e5fd28ed8913fe4c..0c5e8901b903375c7d4de32943e657b205d8fae9 100644 --- a/python/paddle/fluid/tests/unittests/test_split_var.py +++ b/python/paddle/fluid/tests/unittests/test_split_var.py @@ -21,15 +21,7 @@ import random class TestSplitVar(unittest.TestCase): - def test_check_output(self): - # split below shapes to 10 servers - shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10]] - expected_sizes = [ - [15], [1024], - [2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 784], - [2040, 2040, 2040, 2040], - [1150, 1150, 1150, 1150, 1150, 1150, 1100] - ] + def check_split_output(self, shapes, expected_sizes, min_size): var_list = [] program = fluid.Program() for shape in shapes: @@ -39,7 +31,7 @@ class TestSplitVar(unittest.TestCase): # dtype=core.VarDesc.VarType.LOD_TENSOR, shape=shape) var_list.append(var) - blocks = split_dense_variable(var_list, 10) + blocks = split_dense_variable(var_list, 10, min_size) all_sizes = [] for s in expected_sizes: for s2 in s: @@ -48,6 +40,25 @@ class TestSplitVar(unittest.TestCase): varname, block_id, size = block_str.split(":") self.assertEqual(int(size), all_sizes[i]) + def test_1k(self): + shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10]] + expected_sizes = [ + [15], [1024], + [2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 784], + [2040, 2040, 2040, 2040], + [1150, 1150, 1150, 1150, 1150, 1150, 1100] + ] + + self.check_split_output(shapes, expected_sizes, 1024) + + def test_check_output_8k(self): + shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10], + [6, 33, 33, 33]] + expected_sizes = [[15], [1024], [10976, 10976], [8160], [8000], + [35937, 35937, 35937, 35937, 35937, 35937]] + + self.check_split_output(shapes, expected_sizes, 8192) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/trainer.py b/python/paddle/fluid/trainer.py index d44cb16bfb1545fc840d1a38155ec407afd4473d..7da123dd92ed9d111d68cd70efb8ce1493452609 100644 --- a/python/paddle/fluid/trainer.py +++ b/python/paddle/fluid/trainer.py @@ -12,17 +12,18 @@ # See the License for the specific language governing permissions and # limitations under the License. +import contextlib import os + import core -import framework -import executor + import data_feeder -import contextlib +import executor +import framework import io -import transpiler - # optimizer is same as the parameter of Trainer.__init__. Rename it to opt_module import optimizer as opt_module +import parallel_executor from transpiler import distribute_transpiler __all__ = [ @@ -48,43 +49,85 @@ class BeginStepEvent(object): def __init__(self, epoch_id, step_id): self.epoch = epoch_id self.step = step_id + self.fetch_metrics = True class EndStepEvent(object): - def __init__(self, epoch_id, step_id): + def __init__(self, epoch_id, step_id, metrics): self.epoch = epoch_id self.step = step_id + self.metrics = metrics + + +def check_and_get_place(place): + """ + Check the type of place or get the default place + Args: + place(None|core.CUDAPlace|core.CPUPlace): the place that trainer will be executed on. + + Raises: + TypeError if the type mismatched. + + Returns: + the original place if it is not None. + if fluid is compiled with CUDA, returns CUDAPlace(0) by default. + Otherwise returns CPUPlace by default. + """ + if place is None: + if core.is_compiled_with_cuda(): + return core.CUDAPlace(0) + else: + return core.CPUPlace() + else: + if not isinstance(place, core.CUDAPlace) and not isinstance( + place, core.CPUPlace): + raise TypeError("Place should be either CUDAPlace or CPUPlace") + return place class Trainer(object): """ Args: - program_func(callable): A function which will return loss. The loss must be a scaler. + train_func(callable): A function which will return loss. The loss must be a scalar. optimizer(optimizer.Optimizer): The optimizer should be an instance of Optimizer place: The device place of this trainer. """ - def __init__(self, program_func, optimizer, param_path=None, place=None): + def __init__(self, + train_func, + optimizer, + param_path=None, + place=None, + parallel=False): + self.__stop = False + self.parallel = parallel # 1. we need to generate a framework.Program by calling # program_func. Reference: fluid.program_guard in # test_word2vec.py + if not isinstance(optimizer, opt_module.Optimizer): + raise TypeError("The optimizer should be an instance of Optimizer") + self.scope = core.Scope() self.startup_program = framework.Program() self.train_program = framework.Program() with framework.program_guard(self.train_program, self.startup_program): - loss = program_func() + program_func_outs = train_func() + self.train_func_outputs = program_func_outs if isinstance( + program_func_outs, list) else [program_func_outs] + self.test_program = self.train_program.clone() if not isinstance(optimizer, opt_module.Optimizer): raise TypeError( "The optimizer should be an instance of Optimizer") - + # The fisrt element of program_func_outs is loss. + loss = self.train_func_outputs[0] optimize_ops, params_grads = optimizer.minimize(loss) - self.place = Trainer._check_and_get_place(place) + self.place = check_and_get_place(place) - self.dist_transpile_if_necessary(optimize_ops, params_grads) + self._dist_transpile_if_necessary(optimize_ops, params_grads) # 2. move the default_main_program to self.program and run the # default_startup program on an empty core.Scope() @@ -97,7 +140,40 @@ class Trainer(object): # load params from param_path into scope io.load_persistables(exe, dirname=param_path) - def dist_transpile_if_necessary(self, optimize_ops, params_grads): + def _transpile_nccl2_dist(self): + # PADDLE_TRAINER_IPS + if "PADDLE_TRAINER_IPS" not in os.environ: + self.nccl_id_var = None + else: + self.trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + port = os.getenv("PADDLE_PSERVER_PORT") + worker_ips = os.getenv("PADDLE_TRAINER_IPS") + worker_endpoints = [] + for ip in worker_ips.split(","): + worker_endpoints.append(':'.join([ip, port])) + self.num_trainers = len(worker_endpoints) + current_endpoint = os.getenv("POD_IP") + ":" + port + worker_endpoints.remove(current_endpoint) + # TODO(wuyi): use self.nccl_id_var, self.num_trainers and self.trainer_id + # in ParallelExecutor to start + # distributed training using NCCL2 + self.nccl_id_var = self.startup_program.global_block().create_var( + name="NCCLID", persistable=True, type=core.VarDesc.VarType.RAW) + self.startup_program.global_block().append_op( + type="gen_nccl_id", + inputs={}, + outputs={"NCCLID": self.nccl_id_var}, + attrs={ + "endpoint": current_endpoint, + "endpoint_list": worker_endpoints, + "trainer_id": self.trainer_id + }) + + def _dist_transpile_if_necessary(self, optimize_ops, params_grads): + self._transpile_nccl2_dist() + if self.nccl_id_var != None: + return + if "PADDLE_TRAINING_ROLE" not in os.environ: return @@ -135,12 +211,13 @@ class Trainer(object): 'TRAINING_ROLE environment variable must be either TRAINER or PSERVER' ) - def train(self, - num_epochs, - event_handler, - reader=None, - parallel=False, - feed_order=None): + def stop(self): + """ + stop training + """ + self.__stop = True + + def train(self, num_epochs, event_handler, reader=None, feed_order=None): """ Train the model. @@ -148,28 +225,37 @@ class Trainer(object): num_epochs: The number of epoch. An epoch will process all data in reader event_handler: The event handler. A function with type (ev:Event)->void reader: - parallel: True if use multi-CPUs or multi-GPUs feed_order: Feeding order of reader. None will following the defining order in program Returns: """ - if parallel: - raise NotImplementedError( - "Parallel Executor version of trainer is not implemented") - training_role = os.getenv("PADDLE_TRAINING_ROLE", "") if training_role == "PSERVER": with self._prog_and_scope_guard(): exe = executor.Executor(self.place) exe.run() return + if self.parallel: + self._train_by_parallel_executor(num_epochs, event_handler, reader, + feed_order) + else: + self._train_by_executor(num_epochs, event_handler, reader, + feed_order) - self._train_by_executor(num_epochs, event_handler, reader, feed_order) + def test(self, reader, feed_order): + """ + Test the model on given test data - def test(self, reader): - pass + Args: + reader: The reader that yields test data. + feed_order: Feeding order of reader. None will following the defining + order in program + """ + + return self._test_by_executor(reader, feed_order, + self.train_func_outputs) def save_params(self, param_path): # reference: save_persistables in io.py @@ -177,32 +263,6 @@ class Trainer(object): exe = executor.Executor(self.place) io.save_persistables(exe, dirname=param_path) - @staticmethod - def _check_and_get_place(place): - """ - Check the type of place or get the default place - Args: - place(None|core.CUDAPlace|core.CPUPlace): the place that trainer will be executed on. - - Raises: - TypeError if the type mismatched. - - Returns: - the original place if it is not None. - if fluid is compiled with CUDA, returns CUDAPlace(0) by default. - Otherwise returns CPUPlace by default. - """ - if place is None: - if core.is_compiled_with_cuda(): - return core.CUDAPlace(0) - else: - return core.CPUPlace() - else: - if not isinstance(place, core.CUDAPlace) and not isinstance( - place, core.CPUPlace): - raise TypeError("Place should be either CUDAPlace or CPUPlace") - return place - @contextlib.contextmanager def _prog_and_scope_guard(self): with framework.program_guard( @@ -225,26 +285,88 @@ class Trainer(object): """ with self._prog_and_scope_guard(): + feed_var_list = build_feed_var_list(self.train_program, feed_order) + feeder = data_feeder.DataFeeder( + feed_list=feed_var_list, place=self.place) exe = executor.Executor(self.place) - if feed_order is None: - feed_var_list = [ - var - for var in self.train_program.global_block( - ).vars.itervalues() - if hasattr(var, 'is_data') and var.is_data - ] - else: - feed_var_list = [ - self.train_program.global_block().var(var_name) - for var_name in feed_order - ] - + reader = feeder.decorate_reader(reader, multi_devices=False) + self._train_by_any_executor(event_handler, exe, num_epochs, reader) + + def _train_by_any_executor(self, event_handler, exe, num_epochs, reader): + for epoch_id in range(num_epochs): + event_handler(BeginEpochEvent(epoch_id)) + for step_id, data in enumerate(reader()): + if self.__stop: + return + begin_event = BeginStepEvent(epoch_id, step_id) + event_handler(begin_event) + if begin_event.fetch_metrics: + metrics = exe.run(feed=data, + fetch_list=[ + var.name + for var in self.train_func_outputs + ]) + else: + metrics = exe.run(feed=data, fetch_list=[]) + event_handler(EndStepEvent(epoch_id, step_id, metrics)) + event_handler(EndEpochEvent(epoch_id)) + + def _test_by_executor(self, reader, feed_order, fetch_list): + with executor.scope_guard(self.scope): + feed_var_list = build_feed_var_list(self.test_program, feed_order) + feeder = data_feeder.DataFeeder( + feed_list=feed_var_list, place=self.place) + exe = executor.Executor(self.place) + accumulated = len(fetch_list) * [0] + count = 0 + for data in reader(): + outs = exe.run(program=self.test_program, + feed=feeder.feed(data), + fetch_list=fetch_list) + accumulated = [x[0] + x[1][0] for x in zip(accumulated, outs)] + count += 1 + + return [x / count for x in accumulated] + + def _train_by_parallel_executor(self, num_epochs, event_handler, reader, + feed_order): + with self._prog_and_scope_guard(): + pe = self._get_or_create_parallel_executor() + feed_var_list = build_feed_var_list(self.train_program, feed_order) feeder = data_feeder.DataFeeder( feed_list=feed_var_list, place=self.place) - for epoch_id in range(num_epochs): - event_handler(BeginEpochEvent(epoch_id)) - for step_id, data in enumerate(reader()): - event_handler(BeginStepEvent(epoch_id, step_id)) - exe.run(feed=feeder.feed(data), fetch_list=[]) - event_handler(EndStepEvent(epoch_id, step_id)) - event_handler(EndEpochEvent(epoch_id)) + reader = feeder.decorate_reader(reader, multi_devices=True) + self._train_by_any_executor(event_handler, pe, num_epochs, reader) + + def _get_parallel_executor(self): + return getattr(self, 'parallel_executor', None) + + def _get_or_create_parallel_executor(self): + if self._get_parallel_executor() is None: + self.parallel_executor = parallel_executor.ParallelExecutor( + use_cuda=isinstance(self.place, core.CUDAPlace), + loss_name=self.train_func_outputs[0].name) + return self._get_parallel_executor() + + +def build_feed_var_list(program, feed_order): + if not isinstance(program, framework.Program): + raise TypeError("The 'program' should be an object of Program") + + if isinstance(feed_order, list): + feed_var_list = [ + program.global_block().var(var_name) for var_name in feed_order + ] + else: + if not isinstance(feed_order, dict): + raise TypeError( + "The 'feed_order' should be either None, list or dict.") + if not sorted(feed_order.values()) == range(len(feed_order)): + raise ValueError( + "The values of 'feed_order' should be a permutation of [0, len(feed_order))" + ) + sorted_pair_list = sorted(feed_order.items(), key=lambda item: item[1]) + feed_var_list = [ + program.global_block().var(pair[0]) for pair in sorted_pair_list + ] + return feed_var_list diff --git a/python/paddle/fluid/transpiler/__init__.py b/python/paddle/fluid/transpiler/__init__.py index 6d3c1b947f4acb1335b25e6eb0099d5d532c895a..413c36c5c41bbe0169f1c050ccdac040202d66df 100644 --- a/python/paddle/fluid/transpiler/__init__.py +++ b/python/paddle/fluid/transpiler/__init__.py @@ -11,6 +11,7 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + from distribute_transpiler import DistributeTranspiler from inference_transpiler import InferenceTranspiler from memory_optimization_transpiler import memory_optimize, release_memory diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index 640ac9f085e6dc83bb04faafadf4846089ad3e29..42ff0a9eb1112ed5709749e3867794c80be8f1d1 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -17,8 +17,10 @@ from __future__ import print_function import math import distributed_splitter as splitter -from .. import core -from ..framework import Program, default_main_program, Variable, Parameter +from .. import core, framework +from ..framework import Program, default_main_program, \ + default_startup_program, \ + Variable, Parameter, grad_var_name LOOKUP_TABLE_TYPE = "lookup_table" LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad" @@ -91,30 +93,33 @@ def same_or_split_var(p_name, var_name): return p_name == var_name or p_name.startswith(var_name + ".block") -def split_dense_variable(var_list, - pserver_count, - min_block_size=1024, - max_block_size=1048576): +def split_dense_variable(var_list, service_count, min_block_size=8192): """ - We may need to split dense tensor to one or more blocks and put - them equally onto parameter server. One block is a sub-tensor - aligned by dim[0] of the tensor. - - We need to have a minimal block size so that the calculations in - the parameter server side can gain better performance. By default - minimum block size is 1024. The max block size is used to prevent - very large blocks that may cause send error. - :return: A list of VarBlocks. Each VarBlock specifies a shard of - the var. + We may need to split dense tensor to one or more blocks and put + them equally onto parameter server. One block is a sub-tensor + aligned by dim[0] of the tensor. + + We need to have a minimal block size so that the calculations in + the parameter server side can gain better performance. By default + minimum block size 8K elements (maybe 16bit or 32bit or 64bit). + + Args: + var_list (list): List of variables. + service_count (int): Numel of pserver services. A pserver may have two + or more listening ports. + min_block_size (int): Minimum splitted block size. + Returns: + blocks (list[(varname, block_id, current_block_size)]): A list + of VarBlocks. Each VarBlock specifies a shard of the var. """ blocks = [] for var in var_list: - split_count = pserver_count + split_count = service_count var_numel = reduce(lambda x, y: x * y, var.shape) max_pserver_count = int(math.floor(var_numel / float(min_block_size))) if max_pserver_count == 0: max_pserver_count = 1 - if max_pserver_count < pserver_count: + if max_pserver_count < service_count: split_count = max_pserver_count block_size = int(math.ceil(var_numel / float(split_count))) @@ -153,43 +158,43 @@ class DistributeTranspiler: split_method=splitter.round_robin, sync_mode=True): """ - Transpile the program to distributed data-parallelism programs. - The main_program will be transformed to use a remote parameter server - to do parameter optimization. And the optimization graph will be put - into a parameter server program. - - Use different methods to split trainable variables to different - parameter servers. - - Steps to transpile trainer: - 1. split variable to multiple blocks, aligned by product(dim[1:]) (width). - 2. rename splited grad variables to add trainer_id suffix ".trainer_%d". - 3. modify trainer program add split_op to each grad variable. - 4. append send_op to send splited variables to server and fetch - params(splited blocks or origin param) from server. - 5. append concat_op to merge splited blocks to update local weights. - - Steps to transpile pserver: - 1. create new program for parameter server. - 2. create params and grad variables that assigned to current server instance. - 3. create a sub-block in the server side program - 4. append ops that should run on current server instance. - 5. add listen_and_serv op - - :param trainer_id: one unique id for each trainer in a job. - :type trainer_id: int - :param program: program to transpile, default is default_main_program - :type program: Program - :param pservers: parameter server endpoints like "m1:6174,m2:6174" - :type pservers: string - :param trainers: total number of workers/trainers in the job - :type trainers: int - :param split_method: A function to determin how to split variables - to different servers equally. - :type split_method: function - :type sync_mode: boolean default True - :param sync_mode: if sync_mode is set True, it means that dist transpiler - will transpile the program into sync_mode pserver and trainer program. + Transpile the program to distributed data-parallelism programs. + The main_program will be transformed to use a remote parameter server + to do parameter optimization. And the optimization graph will be put + into a parameter server program. + + Use different methods to split trainable variables to different + parameter servers. + + Steps to transpile trainer: + 1. split variable to multiple blocks, aligned by product(dim[1:]) (width). + 2. rename splited grad variables to add trainer_id suffix ".trainer_%d". + 3. modify trainer program add split_op to each grad variable. + 4. append send_op to send splited variables to server and fetch + params(splited blocks or origin param) from server. + 5. append concat_op to merge splited blocks to update local weights. + + Steps to transpile pserver: + 1. create new program for parameter server. + 2. create params and grad variables that assigned to current server instance. + 3. create a sub-block in the server side program + 4. append ops that should run on current server instance. + 5. add listen_and_serv op + + :param trainer_id: one unique id for each trainer in a job. + :type trainer_id: int + :param program: program to transpile, default is default_main_program + :type program: Program + :param pservers: parameter server endpoints like "m1:6174,m2:6174" + :type pservers: string + :param trainers: total number of workers/trainers in the job + :type trainers: int + :param split_method: A function to determin how to split variables + to different servers equally. + :type split_method: function + :type sync_mode: boolean default True + :param sync_mode: if sync_mode is set True, it means that dist transpiler + will transpile the program into sync_mode pserver and trainer program. """ assert (callable(split_method)) if program is None: @@ -244,7 +249,7 @@ class DistributeTranspiler: ] grad_list = [ grad for grad in grad_list - if grad.name != framework.grad_var_name(self.table_name) + if grad.name != grad_var_name(self.table_name) ] self.table_param_grad = [ param_grad for param_grad in params_grads @@ -268,6 +273,7 @@ class DistributeTranspiler: grad_var_mapping = self._append_split_op(program, grad_blocks) param_var_mapping = self._create_vars_from_blocklist(program, param_blocks) + # step3: Add gradients as send op inputs and parameters as send # op outputs. send_inputs = [] @@ -275,9 +281,11 @@ class DistributeTranspiler: for b in grad_blocks: # append by order varname, block_id, _ = b.split(":") send_inputs.append(grad_var_mapping[varname][int(block_id)]) + for b in param_blocks: varname, block_id, _ = b.split(":") send_outputs.append(param_var_mapping[varname][int(block_id)]) + # let send_op know which endpoint to send which var to, eplist has the same # order as send_inputs. eplist = split_method(send_inputs, pserver_endpoints) @@ -415,7 +423,7 @@ class DistributeTranspiler: def __append_optimize_op__(op, block, grad_to_block_id): if self._is_opt_op(op): self._append_pserver_ops(block, op, endpoint, grad_to_block_id, - default_main_program()) + self.origin_program) else: self._append_pserver_non_opt_ops(block, op) @@ -494,7 +502,7 @@ class DistributeTranspiler: were split to several blocks. """ s_prog = Program() - orig_s_prog = framework.default_startup_program() + orig_s_prog = default_startup_program() params = self.param_grad_ep_mapping[endpoint]["params"] def _get_splited_name_and_shape(varname): @@ -619,7 +627,7 @@ class DistributeTranspiler: # 2. add split_ids_op and send_vars_op to send gradient to pservers # there should only be one table_name all_ops = program.global_block().ops - table_grad_name = framework.grad_var_name(self.table_name) + table_grad_name = grad_var_name(self.table_name) for op in all_ops: if table_grad_name in op.output_arg_names: op_index = list(all_ops).index(op) @@ -692,7 +700,7 @@ class DistributeTranspiler: persistable=True) grad_var = _clone_var( pserver_program.global_block(), - self.origin_program.global_block().vars[framework.grad_var_name( + self.origin_program.global_block().vars[grad_var_name( self.table_name)], persistable=False) @@ -749,9 +757,18 @@ class DistributeTranspiler: Create vars for each split. NOTE: only grads need to be named for different trainers, use add_trainer_suffix to rename the grad vars. - :return: A dict mapping from original var name to each var split. + Args: + program (ProgramDesc): ProgramDesc which gradients blong. + block_list (list[(varname, block_id, block_size)]): List of gradient blocks. + add_trainer_suffix (Bool): Add trainer suffix to new variable's name if set True. + Returns: + var_mapping (dict(varname->[new_varname_variable])):A dict mapping + from original var name to each var split. """ + + # varname->[(block_id, current_block_size)] block_map = dict() + var_mapping = dict() for block_str in block_list: varname, offset, size = block_str.split(":") @@ -822,7 +839,16 @@ class DistributeTranspiler: persistable=persistable) def _append_split_op(self, program, gradblocks): - # Split variables that need to be split and append respective ops + """ + Split variables that need to be split and append respective ops + Args: + program (ProgramDesc): ProgramDesc that gradients blong. + gradblocks (list[(varname, block_id, block_size)]): List of gradient blocks. + Returns: + var_mapping (dict(varname->[new_splitted_variable])):A dict mapping + from original var name to each var split. + """ + add_suffix = False if self.trainer_num > 1: add_suffix = True @@ -1146,6 +1172,12 @@ class DistributeTranspiler: return lr_ops def _get_optimize_pass(self): + """ + Get optimizer operators, paramters and gradients from origin_program + Returns: + opt_ops (list): optimize operators. + params_grads (dict): paramter->gradient. + """ block = self.origin_program.global_block() opt_ops = [] params_grads = [] diff --git a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py index 49034b47b2d184e4027bcebc29413a163340fdaa..80a8f7c09cfe521f8f94a27e85fc8d86c02b3e97 100644 --- a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py +++ b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py @@ -24,7 +24,8 @@ dtype_to_size = { core.VarDesc.VarType.INT16: 2, core.VarDesc.VarType.INT32: 4, core.VarDesc.VarType.INT64: 8, - core.VarDesc.VarType.BOOL: 1 + core.VarDesc.VarType.BOOL: 1, + core.VarDesc.VarType.UINT8: 1, } SUB_BLOCK_OPS = [ diff --git a/tools/aws_benchmarking/server/cluster_master.py b/tools/aws_benchmarking/server/cluster_master.py index 1333a942bf013a8182585b56e5843803c56945b1..a9b24846544d8aca5e4c7bd5709e70564c088431 100644 --- a/tools/aws_benchmarking/server/cluster_master.py +++ b/tools/aws_benchmarking/server/cluster_master.py @@ -20,6 +20,7 @@ import time import threading import logging import copy +import csv import netaddr import boto3 @@ -136,6 +137,12 @@ parser.add_argument( parser.add_argument( '--master_server_ip', type=str, default="", help="master server private ip") +parser.add_argument( + '--metric_data_identifier', + type=str, + default="**metrics_data: ", + help="key string to identify metrics data") + parser.add_argument( '--no_clean_up', type=str2bool, @@ -155,6 +162,11 @@ logging.basicConfig( log_files = ["master.log"] +metrics = {} + +metrics_csv_file_name = "metrics.csv" +is_metrics_file_created = False + def create_subnet(): # if no vpc id provided, list vpcs @@ -329,12 +341,42 @@ def create_pservers(): cleanup(args.task_name) +def save_metrics_data(str_msg): + #parse msg + logging.info("found metrics data, saving it to csv file") + global is_metrics_file_created + metrics_raw = str_msg.split(",") + with open(args.log_path + metrics_csv_file_name, 'a') as csvfile: + csv_fieldnames = [] + csv_write_data = {} + for metric in metrics_raw: + metric_data = metric.split("=") + metric_key = metric_data[0].strip() + metric_val = float(metric_data[1].strip()) + if not metric_key in metrics: + metrics[metric_key] = [] + metric_repo = metrics[metric_key] + metric_repo.append(metric_val) + csv_fieldnames.append(metric_key) + csv_write_data[metric_key] = metric_val + writer = csv.DictWriter(csvfile, fieldnames=csv_fieldnames) + if not is_metrics_file_created: + writer.writeheader() + is_metrics_file_created = True + writer.writerow(csv_write_data) + logging.info("csv file appended") + + def log_to_file(source, filename): if not filename in log_files: log_files.append(filename) with open(args.log_path + filename, "a") as log_file: for line in iter(source.readline, ""): log_file.write(line) + if (line.startswith(args.metric_data_identifier)): + #found key data, trying to add to csv + line = line.replace(args.metric_data_identifier, "") + save_metrics_data(line) def parse_command(command_raw, defaults={}): diff --git a/tools/manylinux1/README.md b/tools/manylinux1/README.md index 898e00bd37c7b7bcbcb4a56476ff10c87381e47a..0e5905040175047f5b79939d97a3efcf38992944 100644 --- a/tools/manylinux1/README.md +++ b/tools/manylinux1/README.md @@ -28,3 +28,38 @@ git clone https://github.com/paddlepaddle/paddle cd paddle/tools/manylinux1 REPO=[yourrepo] ./build_all.sh ``` + +## Build PaddlePaddle for the different Python ABIs + +Choose one of the following Python ABI and set the correct environment variables. + +- cp27-cp27m + + ```bash + export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs4/lib:} + export PATH=/opt/python/cp27-cp27m/bin/:${PATH} + export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27m/bin/python + -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27m/include/python2.7 + -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs2/lib/libpython2.7.so" + ``` + +- cp27-cp27mu + + ```bash + export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs2/lib:} + export PATH=/opt/python/cp27-cp27mu/bin/:${PATH} + export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27mu/bin/python + -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27mu/include/python2.7 + -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs4/lib/libpython2.7.so" + ``` + +And then add the `PYTHON_FLAGS` as your cmake flags: + +```bash +cmake .. + ${PYTHON_FLAGS} \ + -DWITH_GPU=OFF \ + ... +``` + +You can find more details about cmake flags at [here](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/build_from_source_en.html#appendix-build-options) diff --git a/tools/test_runner.py b/tools/test_runner.py new file mode 100644 index 0000000000000000000000000000000000000000..9dc750b89058cd73355a2f7984d577252c03526d --- /dev/null +++ b/tools/test_runner.py @@ -0,0 +1,48 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import os +import sys +import paddle.fluid as fluid +import importlib +import cStringIO + + +def main(): + sys.path.append(os.getcwd()) + some_test_failed = False + for module_name in sys.argv[1:]: + buffer = cStringIO.StringIO() + main = fluid.Program() + startup = fluid.Program() + scope = fluid.core.Scope() + with fluid.program_guard(main, startup): + with fluid.scope_guard(scope): + with fluid.unique_name.guard(): + test_loader = unittest.TestLoader() + module = importlib.import_module(module_name) + tests = test_loader.loadTestsFromModule(module) + res = unittest.TextTestRunner(stream=buffer).run(tests) + if not res.wasSuccessful(): + some_test_failed = True + print >> sys.stderr, module_name, 'failed\n', buffer.getvalue( + ) + + if some_test_failed: + exit(1) + + +if __name__ == '__main__': + main() diff --git a/tools/timeline.py b/tools/timeline.py index 8cd6353d46f496831cb61c1cdbbd156ca0579fb4..b413bb6fe0505df8fb09fa0759fefb6509b95bc9 100644 --- a/tools/timeline.py +++ b/tools/timeline.py @@ -171,7 +171,7 @@ if args.timeline_path: profile_paths = profile_path.split(',') profile_dict = dict() -if len(profile_path) == 1: +if len(profile_paths) == 1: with open(profile_path, 'r') as f: profile_s = f.read() profile_pb = profiler_pb2.Profile()