diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index eeda759ff18ccb86ce6a585fe41cb972ea3ae295..e718b32cb6c48d11e73600509a17db107f438708 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -23,7 +23,7 @@ repos: - id: clang-format-with-version-check name: clang-format description: Format files with ClangFormat. - entry: bash ./.clang_format.hook -i + entry: bash ./tools/codestyle/clang_format.hook -i language: system files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$ - repo: local @@ -52,7 +52,7 @@ repos: hooks: - id: copyright_checker name: copyright_checker - entry: python ./.copyright.hook + entry: python ./tools/codestyle/copyright.hook language: system files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto|py)$ exclude: (?!.*third_party)^.*$ | (?!.*book)^.*$ diff --git a/Dockerfile b/Dockerfile index 752fea5951bdc8c2cf79a17c960217c88ae62571..fc5069a6c080ed23317695e6822c4c46b5b5c7f9 100644 --- a/Dockerfile +++ b/Dockerfile @@ -76,7 +76,8 @@ RUN easy_install -U pip && \ pip install sphinx-rtd-theme==0.1.9 recommonmark RUN pip install pre-commit 'ipython==5.3.0' && \ - pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' + pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \ + pip install opencv-python #For docstring checker RUN pip install pylint pytest astroid isort diff --git a/benchmark/fluid/Dockerfile b/benchmark/fluid/Dockerfile index b9eaca5ee6b487bb37bb954b3c606c3096d37aeb..707fadb1fae97cefe8a41715cd57d71754abda41 100644 --- a/benchmark/fluid/Dockerfile +++ b/benchmark/fluid/Dockerfile @@ -1,11 +1,18 @@ FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04 + +# Use UBUNTU_MIRROR can speed up apt-get speed. +# ARG UBUNTU_MIRROR +# RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi' + RUN apt-get update && apt-get install -y python python-pip iputils-ping libgtk2.0-dev wget vim net-tools iftop python-opencv RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/lib/libcudnn.so && ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/lib/libnccl.so -RUN pip install -U pip -RUN pip install -U kubernetes paddlepaddle # IMPORTANT: # Add "ENV http_proxy=http://ip:port" if your download is slow, and don't forget to unset it at runtime. +# exmaple: unset http_proxy && unset https_proxy && python fluid_benchmark.py ... + +RUN pip install -U pip +RUN pip install -U kubernetes paddlepaddle RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.cifar.train10()\npaddle.dataset.flowers.fetch()" | python' RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.mnist.train()\npaddle.dataset.mnist.test()\npaddle.dataset.imdb.fetch()" | python' @@ -14,9 +21,11 @@ RUN pip uninstall -y paddlepaddle && mkdir /workspace ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/paddle_k8s /usr/bin ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/k8s_tools.py /root +RUN chmod +x /usr/bin/paddle_k8s ADD *.whl / -RUN pip install /*.whl && rm -f /*.whl && chmod +x /usr/bin/paddle_k8s +RUN pip install /*.whl && rm -f /*.whl ENV LD_LIBRARY_PATH=/usr/local/lib -ADD fluid_benchmark.py recordio_converter.py models/ /workspace/ +ADD fluid_benchmark.py recordio_converter.py args.py recordio_converter.py run.sh run_fluid_benchmark.sh /workspace/ +ADD models/ /workspace/models/ diff --git a/benchmark/fluid/fluid_benchmark.py b/benchmark/fluid/fluid_benchmark.py index aa70783ecd68be543b2d5aabee96a5b09bd72e6a..ece1102dce987cda994ff086b07f756498ce26e6 100644 --- a/benchmark/fluid/fluid_benchmark.py +++ b/benchmark/fluid/fluid_benchmark.py @@ -97,7 +97,7 @@ def dist_transpile(trainer_id, args): return train_program, fluid.default_startup_program() else: raise ValueError( - 'TRAINING_ROLE environment variable must be either TRAINER or PSERVER' + 'PADDLE_TRAINING_ROLE environment variable must be either TRAINER or PSERVER' ) @@ -264,8 +264,6 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader, break else: loss, = exe.run([avg_loss.name], feed=feeder.feed(data)) - if args.update_method == "pserver": - exe.bcast_params() if args.use_reader_op: num_samples += args.batch_size * args.gpus else: @@ -301,9 +299,18 @@ def print_train_time(start_time, end_time, num_samples): (num_samples, train_elapsed, examples_per_sec)) +def print_paddle_envs(): + print('----------- Configuration envs -----------') + for k in os.environ: + if "PADDLE_" in k: + print "ENV %s:%s" % (k, os.environ[k]) + print('------------------------------------------------') + + def main(): args = parse_args() print_arguments(args) + print_paddle_envs() # the unique trainer id, starting from 0, needed by trainer # only diff --git a/benchmark/fluid/kube_gen_job.py b/benchmark/fluid/kube_gen_job.py index 9da8a69af1d7b671b2648b1b3702776c1c0650b0..dfe8b5cdd58456902fa8ec355e9837dface3f7be 100644 --- a/benchmark/fluid/kube_gen_job.py +++ b/benchmark/fluid/kube_gen_job.py @@ -17,6 +17,7 @@ import copy import argparse import random import os +import copy from kube_templates import pserver, trainer, envs @@ -108,10 +109,9 @@ def gen_job(): tn_container["ports"][0]["containerPort"] = spreadport envs.append({"name": "PADDLE_JOB_NAME", "value": args.jobname}) - envs.append({"name": "TRAINERS", "value": str(args.trainers)}) - envs.append({"name": "PSERVERS", "value": str(args.pservers)}) + envs.append({"name": "PADDLE_TRAINERS", "value": str(args.trainers)}) + envs.append({"name": "PADDLE_PSERVERS", "value": str(args.pservers)}) envs.append({"name": "ENTRY", "value": args.entry}) - envs.append({"name": "PADDLE_INIT_PORT", "value": str(args.port)}) envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)}) # NOTE: these directories below are cluster specific, please modify # this settings before you run on your own cluster. @@ -166,17 +166,23 @@ def gen_job(): tn["spec"]["template"]["spec"]["volumes"] = volumes tn_container["volumeMounts"] = volumeMounts - ps_container["env"] = envs - ps_container["env"].append({"name": "TRAINING_ROLE", "value": "PSERVER"}) + ps_container["env"] = copy.deepcopy(envs) + ps_container["env"].append({ + "name": "PADDLE_TRAINING_ROLE", + "value": "PSERVER" + }) tn_container["env"] = envs if args.disttype == "pserver": tn_container["env"].append({ - "name": "TRAINING_ROLE", + "name": "PADDLE_TRAINING_ROLE", "value": "TRAINER" }) elif args.disttype == "nccl2" or args.disttype == "local": # NCCL2 have no training role, set to plain WORKER - tn_container["env"].append({"name": "TRAINING_ROLE", "value": "WORKER"}) + tn_container["env"].append({ + "name": "PADDLE_TRAINING_ROLE", + "value": "WORKER" + }) os.mkdir(args.jobname) if args.disttype == "pserver": diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake index 25c07850dda7b2f69c2207c37b9d2368632104ec..20dda35c5ccd98f5672d867c26ab97a215483543 100644 --- a/cmake/external/mkldnn.cmake +++ b/cmake/external/mkldnn.cmake @@ -45,7 +45,8 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML") ELSE() MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN") ENDIF() -SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-unused-result") +SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result") +SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value") SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}") SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}") ExternalProject_Add( @@ -53,7 +54,7 @@ ExternalProject_Add( ${EXTERNAL_PROJECT_LOG_ARGS} DEPENDS ${MKLDNN_DEPENDS} GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git" - GIT_TAG "db3424ad44901513c03a1ea31ccaacdf633fbe9f" + GIT_TAG "a29d8487a63afca3d5b8c5bbdbb473cf8ccc6e51" PREFIX ${MKLDNN_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR} diff --git a/doc/fluid/api/gen_doc.sh b/doc/fluid/api/gen_doc.sh index acc8b4aa3fb258e5beef2d1e54919d429cf7ea6f..9ce6a9a7c329055a755cdb0a40c8c1c2af09a61c 100755 --- a/doc/fluid/api/gen_doc.sh +++ b/doc/fluid/api/gen_doc.sh @@ -1,7 +1,7 @@ #!/bin/bash python gen_doc.py layers --submodules control_flow device io nn ops tensor detection learning_rate_scheduler metric > layers.rst -for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer +for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer transpiler do python gen_doc.py ${module} > ${module}.rst done diff --git a/doc/fluid/api/transpiler.rst b/doc/fluid/api/transpiler.rst new file mode 100644 index 0000000000000000000000000000000000000000..b3535b449eb0e5ac6563256ddac3bf4a27fd8ce6 --- /dev/null +++ b/doc/fluid/api/transpiler.rst @@ -0,0 +1,46 @@ +.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}` + !DO NOT EDIT THIS FILE MANUALLY! + +========== +transpiler +========== + +DistributeTranspiler +-------------------- + +.. autoclass:: paddle.fluid.transpiler.DistributeTranspiler + :members: + :noindex: + +InferenceTranspiler +------------------- + +.. autoclass:: paddle.fluid.transpiler.InferenceTranspiler + :members: + :noindex: + +memory_optimize +--------------- + +.. autofunction:: paddle.fluid.transpiler.memory_optimize + :noindex: + +release_memory +-------------- + +.. autofunction:: paddle.fluid.transpiler.release_memory + :noindex: + +HashName +-------- + +.. autoclass:: paddle.fluid.transpiler.HashName + :members: + :noindex: + +RoundRobin +---------- + +.. autoclass:: paddle.fluid.transpiler.RoundRobin + :members: + :noindex: diff --git a/doc/fluid/howto/cluster/fluid_cluster_train_cn.md b/doc/fluid/howto/cluster/fluid_cluster_train_cn.md index b99b90056b0a2e51f2668a6d27d94857bdc09c37..55326940ce7c7dbaa5bf19f1950f470527ddf4f0 100644 --- a/doc/fluid/howto/cluster/fluid_cluster_train_cn.md +++ b/doc/fluid/howto/cluster/fluid_cluster_train_cn.md @@ -168,13 +168,13 @@ cd /paddle/python/paddle/fluid/tests/book 第二步，启动Parameter Server： ```bash -PADDLE_INIT_PORT=6174 PADDLE_INIT_PSERVERS=192.168.1.2 TRAINERS=2 POD_IP=192.168.1.2 PADDLE_INIT_TRAINER_ID=1 TRAINING_ROLE=PSERVER python test_fit_a_line.py +PADDLE_PSERVER_PORT=6174 PADDLE_PSERVER_IPS=192.168.1.2 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=192.168.1.2 PADDLE_TRAINER_ID=1 PADDLE_TRAINING_ROLE=PSERVER python test_fit_a_line.py ``` 执行命令后请等待出现提示： ```Server listening on 192.168.1.2:6174 ```, 表示Paramter Server已经正常启动。 第三步，启动Trainer： ```bash -PADDLE_INIT_PORT=6174 PADDLE_INIT_PSERVERS=192.168.1.3 TRAINERS=2 POD_IP=192.168.1.3 PADDLE_INIT_TRAINER_ID=1 TRAINING_ROLE=TRAINER python test_fit_a_line.py +PADDLE_PSERVER_PORT=6174 PADDLE_PSERVER_IPS=192.168.1.3 PADDLE_TRAINERS=2 PADDLE_CURRENT_IPP=192.168.1.3 PADDLE_TRAINER_ID=1 PADDLE_TRAINING_ROLE=TRAINER python test_fit_a_line.py ``` 由于我们定义的Trainer的数量是2个，因此需要在另外一个计算节点上再启动一个Trainer。 diff --git a/doc/fluid/howto/cluster/fluid_recordio.md b/doc/fluid/howto/cluster/fluid_recordio.md index 55ce63ec193948424cd0b87f13d56b9cf6154dfc..92859e8f622d0c155128821c54252113c5016989 100644 --- a/doc/fluid/howto/cluster/fluid_recordio.md +++ b/doc/fluid/howto/cluster/fluid_recordio.md @@ -114,8 +114,8 @@ def gen_train_list(file_pattern, trainers, trainer_id): ret_list.append(f) return ret_list -trainers = int(os.getenv("TRAINERS")) -trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) +trainers = int(os.getenv("PADDLE_TRAINERS")) +trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) data_file = fluid.layers.io.open_files( filenames=gen_train_list("./mnist-[0-9]*.recordio", 2, 0), thread_num=1, diff --git a/doc/fluid/howto/inference/build_and_install_lib_cn.rst b/doc/fluid/howto/inference/build_and_install_lib_cn.rst index c8d9992fcc92c25f8c14f71c79bde9f79fd92b1f..84005b54e07cf810649370d2c1f6b6c522434bf6 100644 --- a/doc/fluid/howto/inference/build_and_install_lib_cn.rst +++ b/doc/fluid/howto/inference/build_and_install_lib_cn.rst @@ -13,6 +13,7 @@ cpu_noavx_openblas `fluid.tgz `_ cuda8.0_cudnn5_avx_mkl `fluid.tgz `_ cuda8.0_cudnn7_avx_mkl `fluid.tgz `_ +cuda9.0_cudnn7_avx_mkl `fluid.tgz `_ ====================== ======================================== 从源码编译 diff --git a/doc/v2/faq/build_and_install/index_cn.rst b/doc/v2/faq/build_and_install/index_cn.rst index f292684fb5fe2df06db5239e7f43fdfa1dd2f2bd..0d644777287aea0a572adb6fa40f498f9c147af7 100644 --- a/doc/v2/faq/build_and_install/index_cn.rst +++ b/doc/v2/faq/build_and_install/index_cn.rst @@ -213,3 +213,12 @@ virtualenv本身也是Python的一个包，可以用pip进行安装： 保存并关闭文件。 这样，每次打开终端时就会自动启动名为‘paddle’的Python环境了。 + +10. 通过pip安装的PaddlePaddle在 :code:`import paddle.fluid` 报找不到 :code:`libmkldnn.so` 或 :code:`libmklml_intel.so` +------------------------------------------------------------------------------------------ +出现这种问题的原因是在导入 :code:`paddle.fluid` 时需要加载 :code:`libmkldnn.so` 和 :code:`libmklml_intel.so`， +但是系统没有找到该文件。一般通过pip安装PaddlePaddle时会将 :code:`libmkldnn.so` 和 :code:`libmklml_intel.so` +拷贝到 :code:`/usr/local/lib` 路径下，所以解决办法是将该路径加到 :code:`LD_LIBRARY_PATH` 环境变量下， +即： :code:`export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH` 。 + +**注意**：如果是在虚拟环境中安装PaddlePaddle， :code:`libmkldnn.so` 和 :code:`libmklml_intel.so` 可能不在 :code:`/usr/local/lib` 路径下。 \ No newline at end of file diff --git a/paddle/contrib/CMakeLists.txt b/paddle/contrib/CMakeLists.txt index 70e3a0583d8ecf9db19a85c0978aae0ce0625570..4b19256ef4533a09162edf907f6cd51146517e46 100644 --- a/paddle/contrib/CMakeLists.txt +++ b/paddle/contrib/CMakeLists.txt @@ -14,4 +14,3 @@ # add_subdirectory(inference) -add_subdirectory(tape) diff --git a/paddle/contrib/inference/demo/simple_on_word2vec.cc b/paddle/contrib/inference/demo/simple_on_word2vec.cc index 192a6414260ce06048b8c765402d89882cabc51b..2a4bfc87069b9fd8ece58dde210a6cb8344da536 100644 --- a/paddle/contrib/inference/demo/simple_on_word2vec.cc +++ b/paddle/contrib/inference/demo/simple_on_word2vec.cc @@ -40,10 +40,9 @@ void Main(bool use_gpu) { //# 2. Prepare input. int64_t data[4] = {1, 2, 3, 4}; - PaddleBuf buf{.data = data, .length = sizeof(data)}; PaddleTensor tensor{.name = "", .shape = std::vector({4, 1}), - .data = buf, + .data = PaddleBuf(data, sizeof(data)), .dtype = PaddleDType::INT64}; // For simplicity, we set all the slots with the same data. @@ -55,14 +54,12 @@ void Main(bool use_gpu) { //# 4. Get output. ASSERT_EQ(outputs.size(), 1UL); - LOG(INFO) << "output buffer size: " << outputs.front().data.length; - const size_t num_elements = outputs.front().data.length / sizeof(float); + LOG(INFO) << "output buffer size: " << outputs.front().data.length(); + const size_t num_elements = outputs.front().data.length() / sizeof(float); // The outputs' buffers are in CPU memory. for (size_t i = 0; i < std::min(5UL, num_elements); i++) { - LOG(INFO) << static_cast(outputs.front().data.data)[i]; + LOG(INFO) << static_cast(outputs.front().data.data())[i]; } - // TODO(Superjomn): this is should be free automatically - free(outputs[0].data.data); } } @@ -86,10 +83,9 @@ void MainThreads(int num_threads, bool use_gpu) { for (int batch_id = 0; batch_id < num_batches; ++batch_id) { // 2. Dummy Input Data int64_t data[4] = {1, 2, 3, 4}; - PaddleBuf buf{.data = data, .length = sizeof(data)}; PaddleTensor tensor{.name = "", .shape = std::vector({4, 1}), - .data = buf, + .data = PaddleBuf(data, sizeof(data)), .dtype = PaddleDType::INT64}; std::vector inputs(4, tensor); std::vector outputs; @@ -99,13 +95,13 @@ void MainThreads(int num_threads, bool use_gpu) { // 4. Get output. ASSERT_EQ(outputs.size(), 1UL); LOG(INFO) << "TID: " << tid << ", " - << "output buffer size: " << outputs.front().data.length; - const size_t num_elements = outputs.front().data.length / sizeof(float); + << "output buffer size: " << outputs.front().data.length(); + const size_t num_elements = + outputs.front().data.length() / sizeof(float); // The outputs' buffers are in CPU memory. for (size_t i = 0; i < std::min(5UL, num_elements); i++) { - LOG(INFO) << static_cast(outputs.front().data.data)[i]; + LOG(INFO) << static_cast(outputs.front().data.data())[i]; } - free(outputs[0].data.data); } }); } diff --git a/paddle/contrib/inference/high_level_api.md b/paddle/contrib/inference/high_level_api.md new file mode 100644 index 0000000000000000000000000000000000000000..563b696143de9cbf67db38048bbd2f7c11b3a66e --- /dev/null +++ b/paddle/contrib/inference/high_level_api.md @@ -0,0 +1,59 @@ +# Inference High-level APIs +This document describes the high-level inference APIs one can use to easily deploy a Paddle model for an application. + +The APIs are described in `paddle_inference_api.h`, just one header file, and two libaries `libpaddle_fluid.so` and `libpaddle_fluid_api.so` are needed. + +## PaddleTensor +We provide the `PaddleTensor` data structure is to give a general tensor interface. + +The definition is + +```c++ +struct PaddleTensor { + std::string name; // variable name. + std::vector shape; + PaddleBuf data; // blob of data. + PaddleDType dtype; +}; +``` + +The data is stored in a continuous memory `PaddleBuf`, and tensor's data type is specified by a `PaddleDType`. +The `name` field is used to specify the name of input variable, +that is important when there are multiple inputs and need to distiuish which variable to set. + +## engine +The inference APIs has two different underlying implementation, currently there are two valid engines: + +- the native engine, which is consists of the native operators and framework, +- the Anakin engine, which is a Anakin library embeded. + +The native engine takes a native Paddle model as input, and supports any model that trained by Paddle, +but the Anakin engine can only take the Anakin model as input(user need to manully transform the format first) and currently not all Paddle models are supported. + +```c++ +enum class PaddleEngineKind { + kNative = 0, // Use the native Fluid facility. + kAnakin, // Use Anakin for inference. +}; +``` + +## PaddlePredictor and how to create one +The main interface is `PaddlePredictor`, there are following methods + +- `bool Run(const std::vector& inputs, std::vector* output_data)` + - take inputs and output `output_data` +- `Clone` to clone a predictor from an existing one, with model parameter shared. + +There is a factory method to help create a predictor, and the user takes the ownership of this object. + +```c++ +template +std::unique_ptr CreatePaddlePredictor(const ConfigT& config); +``` + +By specifying the engine kind and config, one can get an specific implementation. + +## Reference + +- [paddle_inference_api.h](./paddle_inference_api.h) +- [demos](./demo) diff --git a/paddle/contrib/inference/paddle_inference_api.cc b/paddle/contrib/inference/paddle_inference_api.cc index d67e1e7667800d6dd00cb8915b0d6dc7c664970b..dc2842ae0eeb5592b6d4571b70df162886aee7a2 100644 --- a/paddle/contrib/inference/paddle_inference_api.cc +++ b/paddle/contrib/inference/paddle_inference_api.cc @@ -13,3 +13,53 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/contrib/inference/paddle_inference_api.h" + +namespace paddle { + +PaddleBuf::PaddleBuf(PaddleBuf&& other) + : data_(other.data_), + length_(other.length_), + memory_owned_(other.memory_owned_) { + other.memory_owned_ = false; + other.data_ = nullptr; + other.length_ = 0; +} + +PaddleBuf::PaddleBuf(const PaddleBuf& other) { *this = other; } + +PaddleBuf& PaddleBuf::operator=(const PaddleBuf& other) { + // only the buffer with external memory can be copied + assert(!other.memory_owned_); + data_ = other.data_; + length_ = other.length_; + memory_owned_ = other.memory_owned_; + return *this; +} + +void PaddleBuf::Resize(size_t length) { + // Only the owned memory can be reset, the external memory can't be changed. + if (length_ == length) return; + assert(memory_owned_); + Free(); + data_ = new char[length]; + length_ = length; + memory_owned_ = true; +} + +void PaddleBuf::Reset(void* data, size_t length) { + Free(); + memory_owned_ = false; + data_ = data; + length_ = length; +} + +void PaddleBuf::Free() { + if (memory_owned_ && data_) { + assert(length_ > 0); + delete static_cast(data_); + data_ = nullptr; + length_ = 0; + } +} + +} // namespace paddle \ No newline at end of file diff --git a/paddle/contrib/inference/paddle_inference_api.h b/paddle/contrib/inference/paddle_inference_api.h index 77e2d77b6b7fe3eeed865c8de0818d059cfa6c6e..38e3cc21413b9ab715b84f278f00b9df23cb7682 100644 --- a/paddle/contrib/inference/paddle_inference_api.h +++ b/paddle/contrib/inference/paddle_inference_api.h @@ -21,6 +21,7 @@ limitations under the License. */ #pragma once +#include #include #include #include @@ -32,12 +33,38 @@ enum PaddleDType { INT64, }; -struct PaddleBuf { - void* data; // pointer to the data memory. - size_t length; // number of memory bytes. +class PaddleBuf { + public: + PaddleBuf() = default; + PaddleBuf(PaddleBuf&& other); + // Copy only available when memory is managed externally. + explicit PaddleBuf(const PaddleBuf&); + PaddleBuf& operator=(const PaddleBuf&); + // Do not own the memory. + PaddleBuf(void* data, size_t length) + : data_(data), length_(length), memory_owned_{false} {} + // Own memory. + PaddleBuf(size_t length) + : data_(new char[length]), length_(length), memory_owned_(true) {} + // Resize to `length` bytes. + void Resize(size_t length); + // Reset to external memory. + void Reset(void* data, size_t length); + bool empty() const { return length_ == 0; } + void* data() const { return data_; } + size_t length() const { return length_; } + + ~PaddleBuf() { Free(); } + + private: + void Free(); + void* data_{nullptr}; // pointer to the data memory. + size_t length_{0}; // number of memory bytes. + bool memory_owned_{true}; }; struct PaddleTensor { + PaddleTensor() = default; std::string name; // variable name. std::vector shape; // TODO(Superjomn) for LoD support, add a vector> field if needed. @@ -67,8 +94,9 @@ class PaddlePredictor { // Predict an record. // The caller should be responsible for allocating and releasing the memory of - // `inputs`. `inputs` should be alive until Run returns. caller should be - // responsible for releasing the memory of `output_data`. + // `inputs`. `inputs` should be available until Run returns. Caller should be + // responsible for the output tensor's buffer, either allocated or passed from + // outside. virtual bool Run(const std::vector& inputs, std::vector* output_data) = 0; @@ -81,8 +109,7 @@ class PaddlePredictor { // The common configs for all the predictors. struct Config { - std::string model_dir; // path to the model directory. - bool enable_engine{false}; // Enable to execute (part of) the model on + std::string model_dir; // path to the model directory. }; }; diff --git a/paddle/contrib/inference/paddle_inference_api_anakin_engine.cc b/paddle/contrib/inference/paddle_inference_api_anakin_engine.cc index 5bafc58fa53f7d99de571f66b6224f0f2de66e32..ba2d30314715a57c5ab85e5ae1d8ac0512bbc74f 100644 --- a/paddle/contrib/inference/paddle_inference_api_anakin_engine.cc +++ b/paddle/contrib/inference/paddle_inference_api_anakin_engine.cc @@ -48,7 +48,7 @@ bool PaddleInferenceAnakinPredictor::Run( auto d_tensor_in_p = executor_.get_in(input.name); float *d_data_p = d_tensor_in_p->mutable_data(); if (cudaMemcpy(d_data_p, - static_cast(input.data.data), + static_cast(input.data.data()), d_tensor_in_p->valid_size() * sizeof(float), cudaMemcpyHostToDevice) != 0) { LOG(ERROR) << "copy data from CPU to GPU error"; @@ -65,8 +65,11 @@ bool PaddleInferenceAnakinPredictor::Run( for (auto &output : *output_data) { auto *tensor = executor_.get_out(output.name); output.shape = tensor->shape(); + if (output.data.length() < tensor->valid_size() * sizeof(float)) { + output.data.Resize(tensor->valid_size() * sizeof(float)); + } // Copy data from GPU -> CPU - if (cudaMemcpy(output.data.data, + if (cudaMemcpy(output.data.data(), tensor->mutable_data(), tensor->valid_size() * sizeof(float), cudaMemcpyDeviceToHost) != 0) { diff --git a/paddle/contrib/inference/paddle_inference_api_anakin_engine_tester.cc b/paddle/contrib/inference/paddle_inference_api_anakin_engine_tester.cc index 1d41a5c73e75723f8614d810eae09ed8cdc8cf2b..f92e9d4190412f5847e353ef1dc0324cad668c9a 100644 --- a/paddle/contrib/inference/paddle_inference_api_anakin_engine_tester.cc +++ b/paddle/contrib/inference/paddle_inference_api_anakin_engine_tester.cc @@ -37,28 +37,26 @@ TEST(inference, anakin) { float data[1 * 3 * 224 * 224] = {1.0f}; - PaddleBuf buf{.data = data, .length = sizeof(data)}; PaddleTensor tensor{.name = "input_0", .shape = std::vector({1, 3, 224, 224}), - .data = buf, + .data = PaddleBuf(data, sizeof(data)), .dtype = PaddleDType::FLOAT32}; // For simplicity, we set all the slots with the same data. - std::vector paddle_tensor_feeds(1, tensor); + std::vector paddle_tensor_feeds; + paddle_tensor_feeds.emplace_back(std::move(tensor)); - float data_out[1000]; - - PaddleBuf buf_out{.data = data_out, .length = sizeof(data)}; PaddleTensor tensor_out{.name = "prob_out", .shape = std::vector({1000, 1}), - .data = buf_out, + .data = PaddleBuf(), .dtype = PaddleDType::FLOAT32}; - std::vector outputs(1, tensor_out); + std::vector outputs; + outputs.emplace_back(std::move(tensor_out)); ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs)); - float* data_o = static_cast(outputs[0].data.data); + float* data_o = static_cast(outputs[0].data.data()); for (size_t j = 0; j < 1000; ++j) { LOG(INFO) << "output[" << j << "]: " << data_o[j]; } diff --git a/paddle/contrib/inference/paddle_inference_api_impl.cc b/paddle/contrib/inference/paddle_inference_api_impl.cc index bda2981a14482e2c4a29773d37b074506cc344b1..d9129a704bc289ce1d416474537fc9234a07e5b8 100644 --- a/paddle/contrib/inference/paddle_inference_api_impl.cc +++ b/paddle/contrib/inference/paddle_inference_api_impl.cc @@ -178,8 +178,8 @@ bool NativePaddlePredictor::SetFeed(const std::vector &inputs, // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy. std::memcpy(static_cast(input_ptr), - inputs[i].data.data, - inputs[i].data.length); + inputs[i].data.data(), + inputs[i].data.length()); feeds->push_back(input); } return true; @@ -241,10 +241,11 @@ bool NativePaddlePredictor::GetFetch( } outputs->at(i).shape = shape; - outputs->at(i).data.length = sizeof(float) * data.size(); - outputs->at(i).data.data = malloc(outputs->at(i).data.length); - std::memcpy( - outputs->at(i).data.data, data.data(), outputs->at(i).data.length); + auto &buffer = outputs->at(i).data; + if (buffer.empty() || buffer.length() < sizeof(float) * data.size()) { + buffer.Resize(sizeof(float) * data.size()); + } + std::memcpy(buffer.data(), data.data(), buffer.length()); outputs->at(i).dtype = PaddleDType::FLOAT32; // TODO(panyx0718): support other types? fill tensor name? avoid a copy. } diff --git a/paddle/contrib/inference/test_paddle_inference_api_impl.cc b/paddle/contrib/inference/test_paddle_inference_api_impl.cc index 5d843010e02b09087e6b328428e80fb40eb5bb97..88c4e665a3daed0ed34b23b75d360acbd586401f 100644 --- a/paddle/contrib/inference/test_paddle_inference_api_impl.cc +++ b/paddle/contrib/inference/test_paddle_inference_api_impl.cc @@ -27,13 +27,12 @@ namespace paddle { PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) { PaddleTensor pt; - pt.data.data = t->data(); if (t->type() == typeid(int64_t)) { - pt.data.length = t->numel() * sizeof(int64_t); + pt.data.Reset(t->data(), t->numel() * sizeof(int64_t)); pt.dtype = PaddleDType::INT64; } else if (t->type() == typeid(float)) { - pt.data.length = t->numel() * sizeof(float); + pt.data.Reset(t->data(), t->numel() * sizeof(float)); pt.dtype = PaddleDType::FLOAT32; } else { LOG(FATAL) << "unsupported type."; @@ -79,8 +78,8 @@ void MainWord2Vec(bool use_gpu) { std::vector outputs; ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs)); ASSERT_EQ(outputs.size(), 1UL); - size_t len = outputs[0].data.length; - float* data = static_cast(outputs[0].data.data); + size_t len = outputs[0].data.length(); + float* data = static_cast(outputs[0].data.data()); for (size_t j = 0; j < len / sizeof(float); ++j) { ASSERT_LT(data[j], 1.0); ASSERT_GT(data[j], -1.0); @@ -103,8 +102,6 @@ void MainWord2Vec(bool use_gpu) { EXPECT_LT(lod_data[i] - data[i], 1e-3); EXPECT_GT(lod_data[i] - data[i], -1e-3); } - - free(outputs[0].data.data); } void MainImageClassification(bool use_gpu) { @@ -143,13 +140,12 @@ void MainImageClassification(bool use_gpu) { std::vector outputs; ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs)); ASSERT_EQ(outputs.size(), 1UL); - size_t len = outputs[0].data.length; - float* data = static_cast(outputs[0].data.data); + size_t len = outputs[0].data.length(); + float* data = static_cast(outputs[0].data.data()); float* lod_data = output1.data(); for (size_t j = 0; j < len / sizeof(float); ++j) { EXPECT_NEAR(lod_data[j], data[j], 1e-3); } - free(data); } void MainThreadsWord2Vec(bool use_gpu) { @@ -192,8 +188,8 @@ void MainThreadsWord2Vec(bool use_gpu) { // check outputs range ASSERT_EQ(local_outputs.size(), 1UL); - const size_t len = local_outputs[0].data.length; - float* data = static_cast(local_outputs[0].data.data); + const size_t len = local_outputs[0].data.length(); + float* data = static_cast(local_outputs[0].data.data()); for (size_t j = 0; j < len / sizeof(float); ++j) { ASSERT_LT(data[j], 1.0); ASSERT_GT(data[j], -1.0); @@ -205,7 +201,6 @@ void MainThreadsWord2Vec(bool use_gpu) { for (int i = 0; i < refs[tid].numel(); ++i) { EXPECT_NEAR(ref_data[i], data[i], 1e-3); } - free(data); }); } for (int i = 0; i < num_jobs; ++i) { @@ -251,14 +246,13 @@ void MainThreadsImageClassification(bool use_gpu) { // check outputs correctness ASSERT_EQ(local_outputs.size(), 1UL); - const size_t len = local_outputs[0].data.length; - float* data = static_cast(local_outputs[0].data.data); + const size_t len = local_outputs[0].data.length(); + float* data = static_cast(local_outputs[0].data.data()); float* ref_data = refs[tid].data(); EXPECT_EQ(refs[tid].numel(), len / sizeof(float)); for (int i = 0; i < refs[tid].numel(); ++i) { EXPECT_NEAR(ref_data[i], data[i], 1e-3); } - free(data); }); } for (int i = 0; i < num_jobs; ++i) { diff --git a/paddle/contrib/tape/README.md b/paddle/contrib/tape/README.md deleted file mode 100644 index 16c22a45d59664e44c83923371c0f0d957a8ca7f..0000000000000000000000000000000000000000 --- a/paddle/contrib/tape/README.md +++ /dev/null @@ -1,252 +0,0 @@ -# Dynamic Graph on Fluid - -PaddlePaddle Fluid is targeting the autodiff without tape, which, however, is very -challenging and we are still way from there. DyNet and PyTorch provide a good design -idea, the *tape*, that significantly eases the challenge. Also, DyNet provides -a C++ API that is as convenient as Python but with higher efficiency and could -conveniently integrate with industrial/production systems. This package, `tape`, -combines the good of - -1. tape from PyTorch and DyNet -2. C++ API and core from DyNet -3. rich set of operators from PaddlePaddle - -## Overview - -We can implement Dynet-like Tape(See this [survey](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/survey/dynamic_graph.md)) -by wrapping Paddle Fluid's `Operator` and `Variable`. - -The user API is straight forward since - -1. it is imperative. And it uses host language's control flow logic. -1. it avoids extra concepts such as `Scope` and `Executor`. - -All of these benefits come at the cost of just adding one line `reset_global_tape` -at every iteration. - -## Code Structure - -In short, the `Tape` contains a vector of `OpHandle`s. And an `OpHandle` contains its -`type`, the pointers to the `Variable`s, and necessary attributes. - -```c++ -class Variable { -public: - VriableHandle Grad(); // returns its gradient variable -private: - framework::VarDesc desc_; // compile time infershape, necessary for lazy execution - framework::Variable var_; // run time variable, holds data memory -}; - -using VariableHandle = shared_ptr; - -struct OpHandle { - string type_; - map> inputs_; - map> outputs_; - AttributeMap attrs_; -}; - -class Tape { -public: - void AddOp(OpHandle); // add op - void Forward(); // execute the tape_ - void Backward(); // execute the backward of the tape_ -private: - vector tape_; -}; -``` - -We uses `Function` to indicate layers. It takes care of parameter -initialization and `AddOp` to the Tape when it is called. - -```c++ -class Linear { - public: - Linear(int in_dim, int out_dim, const std::string &act) - : w_(new Variable("LinearWeight")), - b_(new Variable("LinearBias")), - act_(act) { - Tape init_tape; - - std::string initializer = "fill_constant"; - framework::AttributeMap attrs; - attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32; - attrs["shape"] = std::vector{in_dim, out_dim}; - attrs["value"] = 1.0f; - init_tape.AddOp(initializer, {}, {{"Out", {w_}}}, attrs); - - attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32; - attrs["shape"] = std::vector{out_dim}; - attrs["value"] = 1.0f; - init_tape.AddOp(initializer, {}, {{"Out", {b_}}}, attrs); - - init_tape.Forward(); - } - - VariableHandle operator()(VariableHandle input) { - VariableHandle pre_bias(new Variable("linear")); - get_global_tape().AddOp("mul", - {{"X", {input}}, {"Y", {w_}}}, - {{"Out", {pre_bias}}}, - {{"x_num_col_dims", 1}, {"y_num_col_dims", 1}}); - VariableHandle pre_act(new Variable("linear")); - get_global_tape().AddOp("elementwise_add", - {{"X", {pre_bias}}, {"Y", {b_}}}, - {{"Out", {pre_act}}}, - {{"axis", 1}}); - VariableHandle post_act(new Variable("linear")); - get_global_tape().AddOp(act_, - {{"X", {pre_act}}}, - {{"Out", {post_act}}}, - {}); - return post_act; - } - - std::vector Params() { return {w_, b_}; } - - private: - VariableHandle w_; - VariableHandle b_; - std::string act_; -}; -``` - -## User API - -```c++ -// Model function -paddle::tape::Linear linear1(3, 3, "relu"); // init weight and bias -paddle::tape::Linear linear2(3, 3, "relu"); // init weight and bias -paddle::tape::Mean mean; - -// Optimizer -paddle::tape::SGD sgd(0.001); - -// Data Feeder -paddle::tape::Fill data_feeder(...); -VariableHandle input(new paddle::tape::Variable("input")); -VariableHandle label(new paddle::tape::Variable("label")); - -for (int i = 0; i < 2; ++i) { - reset_global_tape(); - - data_feeder(input, label); - - auto loss = softmax(linear2(linear1(input)), label); // compile time InferShape & InferVarType - LOG(INFO) << loss.value(); // Run forward up to loss - - // Run backward, store gradient of w at w->Grad() - get_global_tape.Backward(loss); - - // Update w - sgd(linear1.Params()); - sgd(linear2.Params()); -} -``` - -
- -digraph G { - - subgraph cluster_0 { - node [shape=record,style=filled]; - style=filled; - color=lightgrey; - linear1 [label="{type: mul | {input | {X: before_mul1 | Y: weight1}} | {output | Out: before_bias1}}"]; - elementwise_add1 [label="{type: elementwise_add | {input | {X: before_bias1 | Y: bias1}} | {output | Out: before_act1}}"]; - relu1 [label="{type: relu | {input | {X: before_act1 }} | {output | Out: after_act1}}"]; - - linear1 -> elementwise_add1->relu1; - label = "forward tape"; - } - - linear1:before_mul1->before_mul1 - linear1:weight1->weight1 - linear1:before_bias1->before_bias1 - - elementwise_add1:bias1->bias1 - elementwise_add1:before_bias1->before_bias1 - elementwise_add1:before_act1->before_act1 - - relu1:before_act1->before_act1 - relu1:after_act1->after_act1 - - subgraph cluster_1 { - node [shape=record,style=filled]; - style=filled; - color=lightgrey; - linear1_grad [label="{type: mul_grad | {input | {X: before_mul1 | Y: weight1| Out_grad: before_bias1_grad}} | {output |{X_grad: before_mul1_grad | Y_grad: weight1_grad}}}"]; - - elementwise_add1_grad [label="{type: elementwise_add_grad | {input | Out_grad: before_act1_grad} | {output |{X_grad: before_bias1_grad | Y_grad: bias1_grad}}}"]; - - relu1_grad [label="{type: relu_grad | {input | Out_grad: after_act1_grad} | {ouput | {X_grad: before_act1_grad }}}"]; - - linear1_grad -> elementwise_add1_grad ->relu1_grad [dir=back]; - label = "backward tape"; - } - - relu1_grad:after_act1_grad->after_act1_grad - relu1_grad:before_act1_grad->before_act1_grad - - elementwise_add1_grad:before_act1_grad->before_act1_grad - elementwise_add1_grad:before_bias1_grad->before_bias1_grad - elementwise_add1_grad:bias1_grad->bias1_grad - - linear1_grad:before_mul1->before_mul1 - linear1_grad:weight1->weight1 - linear1_grad:before_bias1_grad->before_bias1_grad - linear1_grad:before_mul1_grad->before_mul1_grad - linear1_grad:weight1_grad->weight1_grad - - - subgraph cluster_2 { - node [shape=record]; - label = "Linear1"; - weight1 - bias1 - } - - weight1 -> weight1_grad [ label="Grad()", style="dashed" ]; - bias1 -> bias1_grad [ label="Grad()", style="dashed"]; - - - -} -
- -![Image](https://github.com/tonyyang-svail/Paddle/blob/cpp_tap/paddle/contrib/tape/computation_graph.png) - -## Code Reuse - -We want to stay close to Paddle Fluid as much as possible. - -### Reuse All Operators - -As all Ops are registered at `OpInfoMap`, the effort of adding a new `Function` -is about 10 lines of code, similar to expose an operator to Python. - -### Reuse Compile Time InferShape and InferVarType - -Note that all the symbolic information is stored at `tape::Varaible::desc_`, instead -of `ProgramDesc.block.vars`, we create a temporary `BlockDesc` to do `InferShape` and -`InferVarType` every time we `AddOp` to the tape. - -### Reuse Operator::Run - -We use smart pointer, instead of `Scope`, to manage memory. So we create a temporary -`Scope` for every `Operator::Run()`. - -## Possible Feature - -### Release Memory on Backward - -We can release memory aggressively. During backward, we can delete the OpHandle once -we have finished its backward. Since all the variable is managed by smart pointer, the -memory is automatically released when its `ref_count` goes to 0. - -### Kernel Fusion - -As a symbolic representation of the Tape is constructed first before the actual -execution, it would be possible to perform graph optimization. One use case is kernel -fusion. diff --git a/paddle/contrib/tape/computation_graph.png b/paddle/contrib/tape/computation_graph.png deleted file mode 100644 index 6cf5ead735d5d18b204b079771e53d44483cf016..0000000000000000000000000000000000000000 Binary files a/paddle/contrib/tape/computation_graph.png and /dev/null differ diff --git a/paddle/contrib/tape/function.h b/paddle/contrib/tape/function.h deleted file mode 100644 index 8c9694d9a21b5948361164eab60a663ec4fd3803..0000000000000000000000000000000000000000 --- a/paddle/contrib/tape/function.h +++ /dev/null @@ -1,131 +0,0 @@ -// 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 "paddle/contrib/tape/tape.h" -#include "paddle/contrib/tape/variable.h" -#include "paddle/fluid/framework/type_defs.h" - -namespace paddle { -namespace tape { - -class Function {}; - -class Fill { - public: - Fill(const std::string &initializer, const framework::AttributeMap &attrs) - : initializer_(initializer), attrs_(attrs) {} - - void operator()(VariableHandle var) { - get_global_tape().AddOp(initializer_, {}, {{"Out", {var}}}, attrs_); - } - - private: - const std::string initializer_; - const framework::AttributeMap attrs_; -}; - -class Mean { - public: - VariableHandle operator()(VariableHandle var) { - VariableHandle out(new Variable("mean")); - get_global_tape().AddOp("mean", {{"X", {var}}}, {{"Out", {out}}}, {}); - return out; - } -}; - -class Linear { - public: - Linear(int in_dim, int out_dim, const std::string &act) - : w_(new Variable("LinearWeight")), - b_(new Variable("LinearBias")), - act_(act) { - Tape init_tape; - - std::string initializer = "fill_constant"; - framework::AttributeMap attrs; - attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32; - attrs["shape"] = std::vector{in_dim, out_dim}; - attrs["value"] = 1.0f; - init_tape.AddOp(initializer, {}, {{"Out", {w_}}}, attrs); - - attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32; - attrs["shape"] = std::vector{out_dim}; - attrs["value"] = 1.0f; - init_tape.AddOp(initializer, {}, {{"Out", {b_}}}, attrs); - - init_tape.Forward(); - } - - VariableHandle operator()(VariableHandle input) { - VariableHandle pre_bias(new Variable("linear")); - get_global_tape().AddOp("mul", - {{"X", {input}}, {"Y", {w_}}}, - {{"Out", {pre_bias}}}, - {{"x_num_col_dims", 1}, {"y_num_col_dims", 1}}); - VariableHandle pre_act(new Variable("linear")); - get_global_tape().AddOp("elementwise_add", - {{"X", {pre_bias}}, {"Y", {b_}}}, - {{"Out", {pre_act}}}, - {{"axis", 1}}); - VariableHandle post_act(new Variable("linear")); - get_global_tape().AddOp( - act_, {{"X", {pre_act}}}, {{"Out", {post_act}}}, {}); - return post_act; - } - - std::vector Params() { return {w_, b_}; } - - private: - VariableHandle w_; - VariableHandle b_; - std::string act_; -}; - -class SGD { - public: - SGD(float learning_rate) : learning_rate_(new Variable("sgd")) { - Tape init_tape; - - std::string initializer = "fill_constant"; - framework::AttributeMap attrs; - attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32; - attrs["shape"] = std::vector{1}; - attrs["value"] = learning_rate; - init_tape.AddOp(initializer, {}, {{"Out", {learning_rate_}}}, attrs); - - init_tape.Forward(); - } - - void operator()(VariableHandle input) { - PADDLE_ENFORCE(get_global_tape().HasBeenBackwarded(), - "optimization must happen after the backward"); - Tape temp_tape; - temp_tape.AddOp("sgd", - {{"Param", {input}}, - {"LearningRate", {learning_rate_}}, - {"Grad", {input->Grad()}}}, - {{"ParamOut", {input}}}, - {}); - temp_tape.Forward(); - } - - private: - VariableHandle learning_rate_; -}; -} -} diff --git a/paddle/contrib/tape/tape.cc b/paddle/contrib/tape/tape.cc deleted file mode 100644 index 531499b6fe02abf200b7d4401494fd6350646622..0000000000000000000000000000000000000000 --- a/paddle/contrib/tape/tape.cc +++ /dev/null @@ -1,265 +0,0 @@ -// 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/contrib/tape/tape.h" - -#include -#include -#include -#include -#include - -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/dim.h" -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/operator.h" -#include "paddle/fluid/framework/scope.h" -#include "paddle/fluid/platform/place.h" -#include "paddle/fluid/pybind/pybind.h" - -namespace paddle { -namespace tape { - -// borrowed from -// https://stackoverflow.com/questions/874134/find-if-string-ends-with-another-string-in-c -inline bool ends_with(std::string const &value, std::string const &ending) { - if (ending.size() > value.size()) return false; - return std::equal(ending.rbegin(), ending.rend(), value.rbegin()); -} - -std::ostream &operator<<(std::ostream &os, const framework::VarDesc &var_desc) { - os << var_desc.Name(); - os << "[" << var_desc.GetType() << "]"; - os << "[" << var_desc.GetDataType() << "]"; - os << "{"; - for (auto &i : var_desc.GetShape()) { - os << i << ","; - } - os << "}"; - return os; -} - -std::string to_string(const std::string &type, - const VariableHandleMap &in_vars, - const VariableHandleMap &out_vars, - const framework::AttributeMap &attrs) { - std::stringstream ss; - ss << type << " "; - for (auto ¶m_name : in_vars) { - for (auto &var : param_name.second) { - ss << param_name.first << ":(" << var->Desc() << ") "; - } - } - for (auto ¶m_name : out_vars) { - for (auto &var : param_name.second) { - ss << param_name.first << ":(" << var->Desc() << ") "; - } - } - return ss.str(); -} - -framework::OpDesc CreateOpDesc(const std::string &type, - const VariableHandleMap &in_vars, - const VariableHandleMap &out_vars, - const framework::AttributeMap &attrs) { - framework::VariableNameMap inputs; - for (auto ¶m_name : in_vars) { - for (auto &var : param_name.second) { - inputs[param_name.first].emplace_back(var->Name()); - } - } - framework::VariableNameMap outputs; - for (auto ¶m_name : out_vars) { - for (auto &var : param_name.second) { - outputs[param_name.first].emplace_back(var->Name()); - } - } - return framework::OpDesc(type, inputs, outputs, attrs); -} - -void InferShapeAndVarType(const std::string &type, - const VariableHandleMap &in_vars, - VariableHandleMap *out_vars, - const framework::AttributeMap &attrs) { - framework::OpDesc op_desc = CreateOpDesc(type, in_vars, *out_vars, attrs); - - // Create a temporary block for compile-time - framework::ProgramDesc program_desc; - framework::BlockDesc *block_desc = program_desc.MutableBlock(0); - PADDLE_ENFORCE(block_desc); - - for (auto ¶m_name : in_vars) { - for (auto &var : param_name.second) { - *block_desc->Var(var->Name())->Proto() = *var->MutableDesc()->Proto(); - } - } - for (auto ¶m_name : *out_vars) { - for (auto &var : param_name.second) { - *block_desc->Var(var->Name())->Proto() = *var->MutableDesc()->Proto(); - } - } - - LOG(INFO) << "- " << to_string(type, in_vars, *out_vars, attrs); - op_desc.InferShape(*block_desc); - op_desc.InferVarType(block_desc); - for (auto ¶m_name : *out_vars) { - for (auto &var : param_name.second) { - *var->MutableDesc()->Proto() = *block_desc->Var(var->Name())->Proto(); - } - } - LOG(INFO) << "+ " << to_string(type, in_vars, *out_vars, attrs); -} - -void Tape::AddOp(const std::string &type, - const VariableHandleMap &in_vars, - VariableHandleMap out_vars, - const framework::AttributeMap &attrs) { - InferShapeAndVarType(type, in_vars, &out_vars, attrs); - tape_.emplace_back(type, in_vars, out_vars, attrs); -} - -// Temporary Scope for Operator::Run() -class ScopeWrapper : public framework::Scope { - public: - ScopeWrapper(const VariableHandleMap &in_vars, - const VariableHandleMap &out_vars) { - for (auto &v : in_vars) { - for (auto &vv : v.second) { - if (!vars_.count(vv->Name())) { - vars_[vv->Name()].reset(vv->Var()); - } - } - } - for (auto &v : out_vars) { - for (auto &vv : v.second) { - if (!vars_.count(vv->Name())) { - vars_[vv->Name()].reset(vv->Var()); - } - } - } - } - - ~ScopeWrapper() { - for (auto &pair : vars_) { - pair.second.release(); - } - } -}; - -void Tape::Forward() { - LOG(INFO) << "Starting forward -------------------------"; - PADDLE_ENFORCE(!has_been_backwarded_); - while (current_position_ < tape_.size()) { - OpHandle &op = tape_[current_position_]; - - // Create Output Tensor, this is only necessary for OpWithKernel - for (auto ¶m2var : op.outputs_) { - for (auto &var : param2var.second) { - var->InitializeVariable(); - } - } - - framework::OpDesc op_desc = - CreateOpDesc(op.type_, op.inputs_, op.outputs_, op.attrs_); - ScopeWrapper scope(op.inputs_, op.outputs_); - framework::OpRegistry::CreateOp(op_desc)->Run(scope, platform::CPUPlace()); - current_position_++; - } - - LOG(INFO) << "Finishing forward -------------------------"; -} - -void Tape::Backward(VariableHandle target) { - PADDLE_ENFORCE(!has_been_backwarded_); - - Forward(); - - // TODO(tonyyang-svail): check output of last op is target - backward_tape_.reset(new Tape()); - - framework::AttributeMap attrs; - - // FIXME(tonyyang-svail): Need to infer_data_type - attrs["dtype"] = framework::proto::VarType::Type::VarType_Type_FP32; - attrs["shape"] = std::vector{1}; - attrs["value"] = 1.0f; - backward_tape_->AddOp( - "fill_constant", {}, {{"Out", {target->Grad()}}}, attrs); - - for (auto it = tape_.rbegin(); it != tape_.rend(); ++it) { - framework::OpDesc op_desc = - CreateOpDesc(it->type_, it->inputs_, it->outputs_, it->attrs_); - std::unordered_map grad_to_var; - std::vector> grad_op_descs = - framework::OpInfoMap::Instance() - .Get(op_desc.Type()) - .GradOpMaker()(op_desc, {}, &grad_to_var, {}); - - for (auto &op_desc : grad_op_descs) { - std::unordered_map name2var; - for (auto ¶m2vars : it->inputs_) { - for (auto &a : param2vars.second) { - name2var[a->Name()] = a; - } - } - for (auto ¶m2vars : it->outputs_) { - for (auto &a : param2vars.second) { - name2var[a->Name()] = a; - } - } - - VariableHandleMap in_vars; - VariableHandleMap out_vars; - std::map - loop_over{{&op_desc->Inputs(), &in_vars}, - {&op_desc->Outputs(), &out_vars}}; - for (auto &each : loop_over) { - auto &vmp = *each.first; - auto &vhm = *each.second; - for (auto &p2a : vmp) { - for (auto &argu : p2a.second) { - if (name2var.count(argu)) { - vhm[p2a.first].push_back(name2var[argu]); - } else { - PADDLE_ENFORCE(ends_with(argu, framework::kGradVarSuffix), - argu.c_str()); - std::string name = argu.substr( - 0, argu.size() - std::strlen(framework::kGradVarSuffix)); - PADDLE_ENFORCE(name2var.count(name), name.c_str()); - vhm[p2a.first].push_back(name2var[name]->Grad()); - } - } - } - } - - backward_tape_->AddOp( - op_desc->Type(), in_vars, out_vars, op_desc->GetAttrMap()); - } - - // TODO(tonyyang-svail): how to fill empty grad? - // TODO(tonyyang-svail): Sum var grad is necessary - } - - backward_tape_->Forward(); - has_been_backwarded_ = true; -} - -Tape &get_global_tape() { - static Tape T; - return T; -} - -void reset_global_tape() { get_global_tape() = Tape(); } -} -} diff --git a/paddle/contrib/tape/tape.h b/paddle/contrib/tape/tape.h deleted file mode 100644 index ed79de17a7fca58a2c542831560f0dd5ad34f960..0000000000000000000000000000000000000000 --- a/paddle/contrib/tape/tape.h +++ /dev/null @@ -1,64 +0,0 @@ -// 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 - -#include "paddle/contrib/tape/variable.h" - -namespace paddle { -namespace tape { - -using VariableHandleMap = std::map>; - -struct OpHandle { - OpHandle(const std::string &type, - const VariableHandleMap &in_vars, - const VariableHandleMap &out_vars, - const framework::AttributeMap &attrs) - : type_(type), inputs_(in_vars), outputs_(out_vars), attrs_(attrs) {} - - std::string type_; - VariableHandleMap inputs_; - VariableHandleMap outputs_; - framework::AttributeMap attrs_; -}; - -class Tape { - public: - void AddOp(const std::string &type, - const VariableHandleMap &in_vars, - VariableHandleMap out_vars, - const framework::AttributeMap &attrs); - void Forward(); - void Backward(VariableHandle target); - - bool HasBeenBackwarded() { return has_been_backwarded_; } - - private: - bool has_been_backwarded_ = false; - size_t current_position_ = 0; - - std::vector tape_; - std::shared_ptr backward_tape_; -}; - -Tape &get_global_tape(); - -void reset_global_tape(); -} -} diff --git a/paddle/contrib/tape/test_tape.cc b/paddle/contrib/tape/test_tape.cc deleted file mode 100644 index e9bfd21a7189c5867a52d2b25db09a462d5c7ba7..0000000000000000000000000000000000000000 --- a/paddle/contrib/tape/test_tape.cc +++ /dev/null @@ -1,61 +0,0 @@ -// 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 "gtest/gtest.h" -#include "paddle/contrib/tape/function.h" - -using namespace paddle::tape; - -TEST(Tape, TestMLP) { - LOG(INFO) << "TestMLP"; - Linear linear1(3, 3, "relu"); - Linear linear2(3, 3, "relu"); - Mean mean; - - SGD sgd(0.001); - - std::string initializer = "fill_constant"; - paddle::framework::AttributeMap attrs; - attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32; - attrs["shape"] = std::vector{3, 3}; - attrs["value"] = 1.0f; - Fill filler(initializer, attrs); - - for (int i = 0; i < 2; ++i) { - reset_global_tape(); - - VariableHandle input(new Variable("input")); - filler(input); - - auto loss = mean(linear2(linear1(input))); - - get_global_tape().Backward(loss); - - for (auto w : linear1.Params()) { - sgd(w); - } - for (auto w : linear2.Params()) { - sgd(w); - } - } -} - -int main(int argc, char** argv) { - std::vector places; - places.emplace_back(paddle::platform::CPUPlace()); - paddle::platform::DeviceContextPool::Init(places); - - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/contrib/tape/variable.cc b/paddle/contrib/tape/variable.cc deleted file mode 100644 index 5ec1612909503f666bca0fce3246002879854156..0000000000000000000000000000000000000000 --- a/paddle/contrib/tape/variable.cc +++ /dev/null @@ -1,33 +0,0 @@ -// 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/contrib/tape/variable.h" - -namespace paddle { -namespace tape { - -void Variable::InitializeVariable() { - LOG(INFO) << "Initialzing " << desc_.Name() << " as " << desc_.GetType(); - framework::proto::VarType::Type var_type = desc_.GetType(); - if (var_type == framework::proto::VarType::LOD_TENSOR) { - var_.GetMutable(); - } else if (var_type == framework::proto::VarType::SELECTED_ROWS) { - var_.GetMutable(); - } else { - PADDLE_THROW("Variable type %d is not in [LOD_TENSOR, SELECTED_ROWS]", - var_type); - } -} -} -} diff --git a/paddle/contrib/tape/variable.h b/paddle/contrib/tape/variable.h deleted file mode 100644 index 35c328e69c9ebe25e907a59e4d67b999aff1d876..0000000000000000000000000000000000000000 --- a/paddle/contrib/tape/variable.h +++ /dev/null @@ -1,85 +0,0 @@ -// 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 "paddle/fluid/framework/operator.h" // framework::kGradVarSuffix -#include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/framework/variable.h" - -namespace paddle { -namespace tape { - -class Variable; -using VariableHandle = std::shared_ptr; - -/* - * Combination of - * framework::VarDesc desc_; - * framework::Variable var_; - */ -class Variable { - public: - Variable(const std::string pre_fix) - : desc_(pre_fix + std::to_string(count())) {} - - Variable(const std::string pre_fix, bool is_grad) - : desc_(pre_fix + (is_grad ? framework::kGradVarSuffix - : std::to_string(count()))) {} - - ~Variable() { LOG(INFO) << "Deleting " << Name(); } - - // Instantiate LoDTensor/SelectedRow - void InitializeVariable(); - - VariableHandle Grad() { - if (grad_.expired()) { - VariableHandle new_grad(new Variable(desc_.Name(), true)); - grad_ = new_grad; - return new_grad; - } else { - return VariableHandle(grad_); - } - } - - // Stochastic Gradient Descent with Momentum - // VariableHandle Momentum (); - - // void init(const std::string& initializer, - // const framework::AttributeMap& attrs); - - // void value() {}; - - const framework::VarDesc& Desc() const { return desc_; } - framework::VarDesc* MutableDesc() { return &desc_; } - - // TODO(tonyyang-svail): No need to expose name - std::string Name() const { return desc_.Name(); } - - framework::Variable* Var() { return &var_; } - - private: - int count() { - static int counter = 0; - return counter++; - } - - framework::VarDesc desc_; - framework::Variable var_; - - std::weak_ptr grad_; -}; -} -} diff --git a/paddle/fluid/framework/details/broadcast_op_handle.cc b/paddle/fluid/framework/details/broadcast_op_handle.cc index d5ca061944f33939cea59a5275e691b1966194fa..1d9f1bd6e417e30f0799f0bbed1739cedb4e8fbf 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle.cc @@ -73,6 +73,9 @@ void BroadcastOpHandle::RunImpl() { int root_id = boost::get(in_tensor.place()).device; std::vector> broadcast_calls; + int type = platform::ToNCCLDataType(in_tensor.type()); + size_t numel = static_cast(in_tensor.numel()); + for (auto out_var_handle : out_var_handles) { Variable *out_var = var_scopes.at(out_var_handle->scope_idx_) ->FindVar(out_var_handle->name_); @@ -87,13 +90,11 @@ void BroadcastOpHandle::RunImpl() { send_recv_buffer = const_cast(in_tensor.data()); out_handle = out_var_handle; } else { - send_recv_buffer = - VariableVisitor::GetMutableTensor(out_var).mutable_data( - out_var_handle->place_); + send_recv_buffer = VariableVisitor::GetMutableTensor(out_var) + .Resize(in_tensor.dims()) + .mutable_data(out_var_handle->place_); } - int type = platform::ToNCCLDataType(in_tensor.type()); - size_t numel = static_cast(in_tensor.numel()); broadcast_calls.emplace_back( [send_recv_buffer, numel, type, root_id, &nccl_ctx] { PADDLE_ENFORCE(platform::dynload::ncclBcast( diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.cc b/paddle/fluid/framework/details/multi_devices_graph_builder.cc index 78356cb1be3bd089c26dde663275e2c8109df951..a6fe64fa80d6bf036893d49de56d7274d49a3b30 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.cc @@ -57,6 +57,7 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder( for (auto &p : params) { grad_names_.insert(GradVarName(p)); } + balance_vars_.resize(places_.size(), 0); } void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result, @@ -140,11 +141,30 @@ bool MultiDevSSAGraphBuilder::IsDistTrainOp( checker(op.InputArgumentNames(), recv_vars); } +size_t MultiDevSSAGraphBuilder::GetAppropriateDeviceID( + const std::vector &var_names) const { + int64_t numel_sum = 0; + for (auto var_name : var_names) { + auto var_desc = all_vars_.at(var_name); + PADDLE_ENFORCE_NOT_NULL(var_desc); + auto dim = framework::make_ddim(var_desc->GetShape()); + int64_t numel = framework::product(dim); + PADDLE_ENFORCE_GT(numel, 0); + numel_sum += numel; + } + + auto smallest = + std::min_element(std::begin(balance_vars_), std::end(balance_vars_)); + size_t dev_id = + static_cast(std::distance(std::begin(balance_vars_), smallest)); + balance_vars_[dev_id] += numel_sum; + return dev_id; +} + std::unique_ptr MultiDevSSAGraphBuilder::Build( const ProgramDesc &program) const { - std::unordered_map all_vars; for (auto *var : program.Block(0).AllVars()) { - all_vars[var->Name()] = var; + all_vars_.emplace(var->Name(), var); } auto graph = new SSAGraph(); @@ -161,35 +181,16 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( auto send_vars = FindDistTrainSendVars(program); auto recv_vars = FindDistTrainRecvVars(program); - 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()); size_t cur_device_id = 0; - std::vector balance_grads(places_.size(), 0); - - auto get_appropriate_dev = [&](std::string &g_name) -> size_t { - auto var_desc = all_vars.at(g_name); - PADDLE_ENFORCE_NOT_NULL(var_desc); - auto dim = framework::make_ddim(var_desc->GetShape()); - int64_t numel = framework::product(dim); - PADDLE_ENFORCE_GE(numel, 0); - auto smallest = - std::min_element(std::begin(balance_grads), std::end(balance_grads)); - size_t dev_id = - static_cast(std::distance(std::begin(balance_grads), smallest)); - balance_grads[dev_id] += numel; - return dev_id; - }; - bool is_forwarding = true; + for (auto *op : program.Block(0).AllOps()) { if (boost::get( op->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) == static_cast(OpRole::kRPC)) { - // append rpc op if program is distributed trainer main program. - // always use the first device CreateRPCOp(&result, *op); } else if (IsDistTrainOp(*op, send_vars, recv_vars)) { CreateDistTrainOp(&result, *op); @@ -199,15 +200,19 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( BuildStrategy::GradientScaleStrategy::kCustomized) { CreateScaleLossGradOp(&result); } + // This assumes the backward generating code will ensure IsScaleLossOp + // is true only for the op that scale the final scalar loss. + // It also assumes backward op will always follow the forward op in + // the block. is_forwarding = false; } else { - int op_dev_id = GetOpDeviceID(var_name_on_devices, *op); + int op_dev_id = GetOpDeviceID(*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); + var_name_on_devices_.emplace(var_name, op_dev_id); } } if (!is_forwarding && places_.size() > 1) { @@ -230,19 +235,22 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( switch (strategy_.reduce_) { case BuildStrategy::ReduceStrategy::kReduce: - cur_device_id = get_appropriate_dev(g_name); + cur_device_id = GetAppropriateDeviceID({g_name}); CreateReduceOp(&result, g_name, cur_device_id); - var_name_on_devices[cur_device_id].emplace(g_name); + var_name_on_devices_.emplace(g_name, cur_device_id); bcast_var_name_set[cur_device_id].emplace(p_name); break; case BuildStrategy::ReduceStrategy::kAllReduce: - if (IsSparseGradient(all_vars, g_name)) { + if (IsSparseGradient(g_name)) { CreateReduceOp(&result, g_name, 0); CreateBroadcastOp(&result, g_name, 0); } else { InsertAllReduceOp(&result, g_name); } break; + default: + LOG(FATAL) << "Unknown reduce strategy "; + break; } } } catch (boost::bad_get e) { @@ -261,7 +269,7 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( } /* Dependency graph has been constructed. However, there are still data - harzaeds need to be handled. + hazards need to be handled. */ PolishGraphToSupportDataHazards(&result); @@ -273,11 +281,9 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( return std::unique_ptr(graph); } -bool MultiDevSSAGraphBuilder::IsSparseGradient( - const std::unordered_map &all_vars, - const std::string &og) const { - PADDLE_ENFORCE(all_vars.count(og) != 0); - if (all_vars.at(og)->GetType() == proto::VarType::SELECTED_ROWS) { +bool MultiDevSSAGraphBuilder::IsSparseGradient(const std::string &og) const { + PADDLE_ENFORCE(all_vars_.count(og) != 0); + if (all_vars_.at(og)->GetType() == proto::VarType::SELECTED_ROWS) { return true; } return false; @@ -345,7 +351,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(SSAGraph *result, auto &prev_grad = vars.back(); op_handle->AddInput(prev_grad.get()); - auto var = new VarHandle(vars.size() - 1, i, og, p); + auto var = new VarHandle(vars.size(), i, og, p); vars.emplace_back(var); op_handle->AddOutput(var); } @@ -363,24 +369,23 @@ bool MultiDevSSAGraphBuilder::IsParameterGradientOnce( return is_pg_once; } -int MultiDevSSAGraphBuilder::GetOpDeviceID( - const std::vector> &var_name_on_devices, - const OpDesc &op) const { +int MultiDevSSAGraphBuilder::GetOpDeviceID(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; - } + for (auto &varname : op.InputArgumentNames()) { + int dev_id = GetVarDeviceID(varname); + if (dev_id != -1) { + return dev_id; } } - return var_dev_id; + return -1; +} + +int MultiDevSSAGraphBuilder::GetVarDeviceID(const std::string &varname) const { + auto got = var_name_on_devices_.find(varname); + return got == var_name_on_devices_.end() ? -1 : got->second; } void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(SSAGraph *result) const { @@ -442,13 +447,14 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(SSAGraph *result, op_handle->AddInput(prev_grad.get()); } auto &vars = result->vars_[dst_dev_id][og]; - auto var = - new VarHandle(vars.size() - 1, dst_dev_id, og, places_[dst_dev_id]); + auto var = new VarHandle(vars.size(), dst_dev_id, og, places_[dst_dev_id]); vars.emplace_back(var); op_handle->AddOutput(var); return var; } +// Find the first occurence of `prev_op_name` and make current `op` depend +// on it. void MultiDevSSAGraphBuilder::ConnectOp(SSAGraph *result, OpHandleBase *op, const std::string &prev_op_name) const { for (auto &prev_op : result->ops_) { @@ -463,16 +469,66 @@ void MultiDevSSAGraphBuilder::ConnectOp(SSAGraph *result, OpHandleBase *op, void MultiDevSSAGraphBuilder::CreateDistTrainOp(SSAGraph *result, const OpDesc &op) const { - CreateComputationalOp(result, op, 0); + int op_dev_id = -1; + if (op.Type() == "split_byref") { + op_dev_id = GetVarDeviceID(op.InputArgumentNames()[0]); + if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) { + op_dev_id = GetAppropriateDeviceID(op.InputArgumentNames()); + for (auto &varname : op.InputArgumentNames()) { + var_name_on_devices_.emplace(varname, op_dev_id); + } + } + for (auto &varname : op.OutputArgumentNames()) { + var_name_on_devices_.emplace(varname, op_dev_id); + } + } else if (op.Type() == "concat") { + op_dev_id = GetVarDeviceID(op.InputArgumentNames()[0]); + } else { + PADDLE_ENFORCE( + "the distribute training related op should be in [split_byref, " + "concat]."); + } + + PADDLE_ENFORCE(op_dev_id != -1, + "can not find right place for distributed op: %s", op.Type()); + + CreateComputationalOp(result, op, op_dev_id); if (op.Type() == "concat") { ConnectOp(result, result->ops_.back().get(), "fetch_barrier"); } } +// Create RPC related op handles that connects its in ops and out ops. void MultiDevSSAGraphBuilder::CreateRPCOp(SSAGraph *result, const OpDesc &op) const { - result->ops_.emplace_back( - new RPCOpHandle(op, local_scopes_[0], op.Type(), places_[0])); + int op_dev_id = -1; + if (op.Type() == "send") { + op_dev_id = GetVarDeviceID(op.InputArgumentNames()[0]); + // the variable name which contains .block means it was splited by + // split_byref op + // so that we can balance the variable blocks to all the pserver instances. + if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce && + op.InputArgumentNames()[0].find(".block") == std::string::npos) { + op_dev_id = GetAppropriateDeviceID(op.InputArgumentNames()); + for (auto &varname : op.InputArgumentNames()) { + var_name_on_devices_.emplace(varname, op_dev_id); + } + } + } else if (op.Type() == "recv") { + op_dev_id = GetAppropriateDeviceID(op.OutputArgumentNames()); + for (auto &varname : op.OutputArgumentNames()) { + var_name_on_devices_.emplace(varname, op_dev_id); + } + } else { + // send_barrier and fetch_barrier op can be scheduled on device 0 + op_dev_id = 0; + } + + PADDLE_ENFORCE(op_dev_id != -1, "can not find the right place for rpc op: %s", + op.Type()); + + result->ops_.emplace_back(new RPCOpHandle(op, local_scopes_[op_dev_id], + op.Type(), places_[op_dev_id])); if (op.Type() == "send_barrier") { ConnectOp(result, result->ops_.back().get(), "send"); @@ -488,9 +544,7 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(SSAGraph *result, "send, send_barrier. recv, fetch_barrier]"); } - // TODO(Yancey1989): schedule rpc op on different place may - // increate throughput - CreateOpHandleIOs(result, op, 0); + CreateOpHandleIOs(result, op, op_dev_id); } bool MultiDevSSAGraphBuilder::IsScaleLossOp(const OpDesc &op) const { diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.h b/paddle/fluid/framework/details/multi_devices_graph_builder.h index 78581755fe4890800636944d6cd89875a852cc19..0b6347bf51dc1c347073a0fdcf4ddd91865d846d 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.h +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.h @@ -47,10 +47,11 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder { #endif std::unique_ptr Build(const ProgramDesc &program) const override; + int GetVarDeviceID(const std::string &varname) const override; private: void CreateOpHandleIOs(SSAGraph *result, const OpDesc &op, - size_t place_id) const; + size_t device_id) const; private: std::string loss_var_name_; @@ -96,21 +97,23 @@ 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; + int GetOpDeviceID(const OpDesc &op) const; void InsertAllReduceOp(SSAGraph *result, const std::string &og) const; void CreateBroadcastOp(SSAGraph *result, const std::string &p_name, size_t src_dev_id) const; - bool IsSparseGradient( - const std::unordered_map &all_vars, - const std::string &og) const; + bool IsSparseGradient(const std::string &og) const; + + size_t GetAppropriateDeviceID( + const std::vector &var_names) const; private: BuildStrategy strategy_; + mutable std::unordered_map all_vars_; + mutable std::unordered_map var_name_on_devices_; + mutable std::vector balance_vars_; void SetCommunicationContext(OpHandleBase *op_handle, const platform::Place &p) const; diff --git a/paddle/fluid/framework/details/op_handle_base.cc b/paddle/fluid/framework/details/op_handle_base.cc index f79565fe71c4aef140475c922cbbf5a1e0b7fe03..1f84c3b9e2d7ee9ae51959988fceeb3451b7b3b8 100644 --- a/paddle/fluid/framework/details/op_handle_base.cc +++ b/paddle/fluid/framework/details/op_handle_base.cc @@ -11,8 +11,8 @@ // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. - #include "paddle/fluid/framework/details/op_handle_base.h" +#include namespace paddle { namespace framework { @@ -122,11 +122,16 @@ void OpHandleBase::RunAndRecordEvent(const std::function &callback) { #ifdef PADDLE_WITH_CUDA if (!events_.empty()) { // Use event std::function method = callback; - + // NOTE(zcd): device context must be ordered here because RecordEvent + // will use a mutex to ensure the safe of multi-threads. + std::map ordered_ctxes; for (auto &p : dev_ctxes_) { + ordered_ctxes.emplace(p.second, p.first); + } + for (auto &p : ordered_ctxes) { method = [method, p, this]() { - static_cast(p.second)->RecordEvent( - events_.at(boost::get(p.first).device), + static_cast(p.first)->RecordEvent( + events_.at(boost::get(p.second).device), method); }; } diff --git a/paddle/fluid/framework/details/ssa_graph_builder.h b/paddle/fluid/framework/details/ssa_graph_builder.h index 5fc12a44b51fae26e5a8f5fdba952d3879e82d0f..9eb23c46264f9036f009b0ae9aeeb34ec70c0e53 100644 --- a/paddle/fluid/framework/details/ssa_graph_builder.h +++ b/paddle/fluid/framework/details/ssa_graph_builder.h @@ -30,6 +30,7 @@ class SSAGraphBuilder { SSAGraphBuilder() {} virtual ~SSAGraphBuilder() {} virtual std::unique_ptr Build(const ProgramDesc &program) const = 0; + virtual int GetVarDeviceID(const std::string &var_name) const { return -1; } DISABLE_COPY_AND_ASSIGN(SSAGraphBuilder); diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc index 6c5098ce85b784a3edcf8f48d2cc828aabd8e161..b1706eb12d080364d04108c7ef4da31e1e7c1deb 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -96,6 +96,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( auto cur_ready_vars = ready_vars.PopAll(1, &timeout); if (timeout) { + std::lock_guard l(exception_mu_); if (exception_) { auto exp = *exception_; exception_.reset(); @@ -199,6 +200,7 @@ void ThreadedSSAGraphExecutor::RunOp( ready_var_q->Extend(op->Outputs()); VLOG(10) << op << " " << op->Name() << "Signal posted"; } catch (platform::EnforceNotMet ex) { + std::lock_guard l(exception_mu_); exception_.reset(new platform::EnforceNotMet(ex)); } catch (...) { LOG(FATAL) << "Unknown exception catched"; diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h index 4a2075f1cccb3211316567197da56c01d26f35ce..90430be996758364387b552019762d9c2e9dfe45 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h @@ -56,6 +56,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { std::vector local_scopes_; std::vector places_; platform::DeviceContextPool fetch_ctxs_; + std::mutex exception_mu_; std::unique_ptr exception_; std::atomic running_ops_; diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index 429482bd038a0703d46dcdfd333cccdb58051126..ae98fccc9600a2a75f12fa516c982bec0ef13f9f 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -21,7 +21,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/reader.h" #ifdef PADDLE_WITH_DISTRIBUTE -#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/distributed/grpc_client.h" #endif #include "paddle/fluid/platform/place.h" #include "paddle/fluid/platform/profiler.h" @@ -49,8 +49,8 @@ Executor::Executor(const platform::Place& place) : place_(place) {} #ifdef PADDLE_WITH_DISTRIBUTE void Executor::Complete() { - ::paddle::operators::detail::RPCClient::GetInstance< - ::paddle::operators::detail::GRPCClient>() + ::paddle::operators::distributed::RPCClient::GetInstance< + ::paddle::operators::distributed::GRPCClient>() ->SendComplete(); } #endif @@ -295,13 +295,14 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, std::unique_ptr Executor::Prepare( const ProgramDesc& program, int block_id) { - auto* ctx = new ExecutorPrepareContext(program, block_id); + std::unique_ptr ctx( + new ExecutorPrepareContext(program, block_id)); PADDLE_ENFORCE_LT(static_cast(block_id), program.Size()); auto& block = program.Block(block_id); for (auto& op_desc : block.AllOps()) { ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc)); } - return std::unique_ptr(ctx); + return ctx; } std::vector> Executor::Prepare( @@ -320,7 +321,8 @@ std::vector> Executor::Prepare( } void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, - bool create_local_scope, bool create_vars) { + bool create_local_scope, bool create_vars, + bool keep_kids) { Scope* local_scope = scope; if (create_vars) { if (create_local_scope) { @@ -343,12 +345,20 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, } } platform::DeviceContextPool::Instance().Get(place_)->Wait(); - if (create_vars && create_local_scope) { + if (local_scope != scope) { scope->DeleteScope(local_scope); } else { - // Delete the local scopes created in operators. - scope->DropKids(); + if (!keep_kids) { + // By default, we should delete all kid scopes after run executor because + // some operators may create local scope when running, such as while_op. + // But when while_op also create a local executor to run it's sub block, + // the sub scopes it created should not be dropped immediately, because + // while_grad_op will use some variables created during while_op run, so + // we need to keep the kids and wait for the outer executor to drop them. + scope->DropKids(); + } } + if (FLAGS_benchmark) { VLOG(2) << "-------------------------------------------------------"; VLOG(2) << "Memory used after deleting local scope: " diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index 67a0761dac2a9adcdd0ce2b218c4aa505d688d56..3aa5ffef69cd29681f248e915575c5715ad0d3fa 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -78,7 +78,7 @@ class Executor { void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, bool create_local_scope = true, - bool create_vars = true); + bool create_vars = true, bool keep_kids = false); void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, std::map* feed_targets, diff --git a/paddle/fluid/framework/framework.proto b/paddle/fluid/framework/framework.proto index 68fcc104d48b2b39929ed2198a2dd2eabae10e94..2cf14bd371831ab682166f4256d6966b5ab278c8 100644 --- a/paddle/fluid/framework/framework.proto +++ b/paddle/fluid/framework/framework.proto @@ -27,6 +27,7 @@ enum AttrType { BOOLEANS = 7; BLOCK = 8; LONG = 9; + BLOCKS = 10; } // OpDesc describes an instance of a C++ framework::OperatorBase @@ -46,6 +47,7 @@ message OpDesc { repeated bool bools = 11; optional int32 block_idx = 12; optional int64 l = 13; + repeated int32 blocks_idx = 14; }; message Var { diff --git a/paddle/fluid/framework/lod_tensor.cc b/paddle/fluid/framework/lod_tensor.cc index e331c8128f2e8121dbbfe82b74ea35f2d0d399c0..d29d8ce1c561e45980d10c17c984ca2ed3b453f3 100644 --- a/paddle/fluid/framework/lod_tensor.cc +++ b/paddle/fluid/framework/lod_tensor.cc @@ -51,8 +51,6 @@ std::ostream &operator<<(std::ostream &os, const LoD &lod) { } std::ostream &operator<<(std::ostream &os, const LoDTensor &t) { - PADDLE_ENFORCE(t.type().hash_code() == typeid(float).hash_code()); - if (!platform::is_cpu_place(t.place())) { LoDTensor tt; framework::TensorCopy(t, platform::CPUPlace(), &tt); @@ -70,7 +68,13 @@ std::ostream &operator<<(std::ostream &os, const LoDTensor &t) { // only print first ten elements int64_t size = t.numel() < 10 ? t.numel() : 10; for (int64_t i = 0; i < size; ++i) { - os << t.data()[i] << " "; + if (t.type().hash_code() == typeid(float).hash_code()) { + os << t.data()[i] << " "; + } else if (t.type().hash_code() == typeid(int64_t).hash_code()) { + os << t.data()[i] << " "; + } else { + PADDLE_THROW("LoDTensor data type not in [float, int64_t]"); + } } return os; diff --git a/paddle/fluid/framework/lod_tensor_test.cc b/paddle/fluid/framework/lod_tensor_test.cc index 6dfe7d2d8c1cce3360d99950240bc6de5a063dab..38d3cd96d65f0a54b0ea87b4c677013f3802adfb 100644 --- a/paddle/fluid/framework/lod_tensor_test.cc +++ b/paddle/fluid/framework/lod_tensor_test.cc @@ -26,6 +26,20 @@ namespace paddle { namespace framework { +TEST(LoD, PrintLoDTensor) { + LoDTensor tensor1; + tensor1.mutable_data(platform::CPUPlace()); + tensor1.data()[0] = 0.2; + tensor1.data()[1] = 0.5; + LOG(INFO) << tensor1; + + LoDTensor tensor2; + tensor2.mutable_data(platform::CPUPlace()); + tensor2.data()[0] = 1; + tensor2.data()[1] = 2; + LOG(INFO) << tensor2; +} + TEST(LoD, data) { LoD lod{{0, 1, 2}}; lod.push_back({0, 2, 4, 5}); @@ -37,7 +51,7 @@ TEST(LoD, data) { } } -TEST(LodExpand, test) { +TEST(LoD, ExpandLoD) { LoD lod{{0, 2}}; LoDTensor tensor; tensor.set_lod(lod); diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc index f92769192c218eb7cdc2350ff6e4721b45005806..a190199f1cb1361f67f20c755b8e7ef52c284adc 100644 --- a/paddle/fluid/framework/op_desc.cc +++ b/paddle/fluid/framework/op_desc.cc @@ -211,6 +211,12 @@ void OpDesc::SetBlockAttr(const std::string &name, BlockDesc *block) { need_update_ = true; } +void OpDesc::SetBlocksAttr(const std::string &name, + std::vector blocks) { + this->attrs_[name] = blocks; + need_update_ = true; +} + void OpDesc::SetAttrMap( const std::unordered_map &attr_map) { attrs_ = attr_map; @@ -305,6 +311,13 @@ struct SetAttrDescVisitor : public boost::static_visitor { void operator()(const std::vector &v) const { VectorToRepeated(v, attr_->mutable_bools()); } + void operator()(const std::vector &v) const { + std::vector blocks_idx; + for (auto blk : v) { + blocks_idx.push_back(blk->ID()); + } + VectorToRepeated(blocks_idx, attr_->mutable_blocks_idx()); + } void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->ID()); } void operator()(int64_t v) const { attr_->set_l(v); } void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); } diff --git a/paddle/fluid/framework/op_desc.h b/paddle/fluid/framework/op_desc.h index a02d3e269129596f65a2fb346e76c1af7fbead95..74dd8ec002005dd080424b48b5db1a2574a6974f 100644 --- a/paddle/fluid/framework/op_desc.h +++ b/paddle/fluid/framework/op_desc.h @@ -77,6 +77,8 @@ class OpDesc { void SetBlockAttr(const std::string &name, BlockDesc *block); + void SetBlocksAttr(const std::string &name, std::vector blocks); + Attribute GetAttr(const std::string &name) const; Attribute GetNullableAttr(const std::string &name) const; diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index 9406c6155da860c90739bddac1e81403b094e619..a6788cb6d5d6acb57998fb9b06dfaaf417912dde 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -110,7 +110,6 @@ ParallelExecutor::ParallelExecutor( // Step 3. Convert main_program to SSA form and dependency graph. Also, insert // ncclOp - details::SSAGraphBuilderFactory builder_factory( member_->places_, loss_var_name, params, member_->local_scopes_, build_strategy); @@ -122,9 +121,10 @@ ParallelExecutor::ParallelExecutor( #endif } + builder_ = builder_factory.Create(); member_->executor_.reset(new details::ThreadedSSAGraphExecutor( exec_strategy, member_->local_scopes_, places, - builder_factory.Create()->Build(main_program))); + builder_->Build(main_program))); member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor( exec_strategy, member_->local_scopes_, std::move(var_infos), @@ -133,10 +133,22 @@ ParallelExecutor::ParallelExecutor( void ParallelExecutor::BCastParamsToGPUs( const std::unordered_set &vars) const { - auto *main_scope = member_->local_scopes_[0]; + // the the initialize bcast, all vars would be bcast from device(0), otherwise + // bcast from the specified device. + bool initialize = builder_.get() == nullptr ? true : false; for (auto &var : vars) { - auto *main_var = main_scope->FindVar(var); + int var_dev_id = + builder_.get() == nullptr ? -1 : builder_->GetVarDeviceID(var); + if (!initialize && var_dev_id == -1) continue; + + framework::Variable *main_var = nullptr; + if (initialize) { + main_var = member_->local_scopes_[0]->FindVar(var); + } else { + main_var = member_->local_scopes_[var_dev_id]->FindVar(var); + } + if (main_var == nullptr || !main_var->IsType()) { continue; } @@ -151,7 +163,8 @@ void ParallelExecutor::BCastParamsToGPUs( for (size_t i = 0; i < member_->places_.size(); ++i) { auto place = member_->places_[i]; void *buffer; - if (i == 0) { + + if ((initialize && i == 0) || (!initialize && i == var_dev_id)) { buffer = const_cast(main_tensor.data()); } else { auto local_scope = member_->local_scopes_[i]; diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h index 5247e790649e76567f4527d54499d6e95dac5c27..058f83f07c26224e3180d140630c08a24c40cd80 100644 --- a/paddle/fluid/framework/parallel_executor.h +++ b/paddle/fluid/framework/parallel_executor.h @@ -19,12 +19,14 @@ limitations under the License. */ #include #include #include "paddle/fluid/framework/details/execution_strategy.h" +#include "paddle/fluid/framework/details/multi_devices_graph_builder.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 { @@ -68,6 +70,7 @@ class ParallelExecutor { private: ParallelExecutorPrivate *member_; + std::unique_ptr builder_; }; } // namespace framework diff --git a/paddle/fluid/framework/type_defs.h b/paddle/fluid/framework/type_defs.h index 4879209ece9fdfea91e484a4118c00a2a2a2b4f7..e099e40f121ff13657e563eb608feecbca0551be 100644 --- a/paddle/fluid/framework/type_defs.h +++ b/paddle/fluid/framework/type_defs.h @@ -35,7 +35,8 @@ using VariableNameMap = std::map>; using Attribute = boost::variant, std::vector, std::vector, bool, - std::vector, BlockDesc*, int64_t>; + std::vector, BlockDesc*, int64_t, + std::vector>; using AttributeMap = std::unordered_map; diff --git a/paddle/fluid/inference/analysis/argument.h b/paddle/fluid/inference/analysis/argument.h index 7d7131ed7a188c199eac6fcef5227b278ea97fa3..f7f4e03968a723df1718bd3752bdd1c3430d02be 100644 --- a/paddle/fluid/inference/analysis/argument.h +++ b/paddle/fluid/inference/analysis/argument.h @@ -21,6 +21,8 @@ * big. */ +#pragma once + #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/inference/analysis/data_flow_graph.h" @@ -43,7 +45,7 @@ struct Argument { #define UNLIKELY(condition) __builtin_expect(static_cast(condition), 0) #define ANALYSIS_ARGUMENT_CHECK_FIELD(field__) \ - if (!UNLIKELY(field__)) { \ + if (UNLIKELY(!(field__))) { \ LOG(ERROR) << "field " << #field__ << " should be set."; \ return false; \ } diff --git a/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc b/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc index b75df33b71311acd0e626e5a13c18469b19ef136..c7f40d43c922a328febd343cea7240fcb09f3d02 100644 --- a/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc +++ b/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc @@ -27,7 +27,7 @@ void TensorRTSubGraphPass::Run(DataFlowGraph *graph) { SubGraphFuse(graph, node_inside_subgraph_teller_); } -} // analysis -} // inference +} // namespace analysis +} // namespace inference -} // paddle +} // namespace paddle diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index d6a36eff09c7f70803d3be619b26d16660da1ec2..d3988ae16d7d4ceccaf01503c6200066f2fa4073 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -184,9 +184,9 @@ else() set(DEPS_OPS ${DEPS_OPS} nccl_op) endif() -add_subdirectory(detail) if(WITH_DISTRIBUTE) - + add_subdirectory(distributed) + set(DISTRIBUTE_DEPS "") if(WITH_GRPC) set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf) @@ -195,18 +195,11 @@ if(WITH_DISTRIBUTE) endif() set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") - op_library(prefetch_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(prefetch_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - op_library(recv_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(recv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - op_library(listen_and_serv_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(listen_and_serv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - op_library(send_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(send_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - op_library(send_barrier_op DEPS ${DISTRIBUTE_DEPS}) - op_library(fetch_barrier_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(send_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - set_source_files_properties(fetch_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + foreach(dist_op "prefetch_op" "listen_and_serv_op" "send_op" "recv_op" "send_barrier_op" "fetch_barrier_op") + op_library(${dist_op} DEPS ${DISTRIBUTE_DEPS}) + set_source_files_properties(${dist_op}.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + endforeach() + #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 SERIAL) diff --git a/paddle/fluid/operators/activation_mkldnn_op.cc b/paddle/fluid/operators/activation_mkldnn_op.cc index 46ed99bcf2234f7621d9f00eb48c846d8a355795..137bca5e2b8e2754aed274970e08b03ee816a7f2 100644 --- a/paddle/fluid/operators/activation_mkldnn_op.cc +++ b/paddle/fluid/operators/activation_mkldnn_op.cc @@ -12,16 +12,20 @@ See the License for the specific language governing permissions and limitations under the License. */ -#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 { -using paddle::framework::Tensor; -using paddle::platform::MKLDNNDeviceContext; +using framework::DataLayout; +using framework::Tensor; +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::stream; +using platform::GetMKLDNNFormat; +using platform::MKLDNNDeviceContext; +using platform::to_void_cast; namespace { std::string gethash(const mkldnn::memory::dims &operand_dims, @@ -35,188 +39,260 @@ std::string gethash(const mkldnn::memory::dims &operand_dims, }; return dim2str(operand_dims) + std::to_string(algorithm); } +} // namespace + +template +class MKLDNNActivationKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + const auto *x = ctx.Input("X"); + PADDLE_ENFORCE(x->layout() == DataLayout::kMKLDNN && + x->format() != memory::format::format_undef, + "Wrong layout/format set for Input x tensor"); + + Functor functor; + + auto attrs = functor.GetAttrs(); + for (auto &attr : attrs) { + *attr.second = ctx.Attr(attr.first); + } + functor(ctx); + } +}; -template -void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm, - const T alpha = 0, const T beta = 0) { +template +class MKLDNNActivationGradKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + const auto *diff_y = ctx.Input(framework::GradVarName("Out")); + PADDLE_ENFORCE(diff_y->layout() == DataLayout::kMKLDNN && + diff_y->format() != memory::format::format_undef, + "Wrong layout/format set for Input OutGrad tensor"); + + Functor functor; + + auto attrs = functor.GetAttrs(); + for (auto &attr : attrs) { + *attr.second = ctx.Attr(attr.first); + } + functor(ctx); + } +}; + +template +void eltwise_forward(const framework::ExecutionContext &ctx, + mkldnn::algorithm algorithm, const T alpha = 0, + const T beta = 0) { PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), "It must use CPUPlace."); - auto &dev_ctx = ctx.template device_context(); const auto &mkldnn_engine = dev_ctx.GetEngine(); - // get buffers - const auto *src = ctx.template Input("X"); - const auto *src_data = src->template data(); + const auto *x = ctx.Input("X"); + auto *y = ctx.Output("Out"); - auto *dst = ctx.template Output("Out"); - T *dst_data = dst->template mutable_data(ctx.GetPlace()); + const T *x_data = x->data(); + T *y_data = y->mutable_data(ctx.GetPlace()); - // get memory dim - PADDLE_ENFORCE(src->dims().size() == 2 || src->dims().size() == 4, + PADDLE_ENFORCE(x->dims().size() == 2 || x->dims().size() == 4, "Input dim must be with 2 or 4"); - std::vector src_tz = framework::vectorize2int(src->dims()); + + std::vector src_tz = framework::vectorize2int(x->dims()); + + auto src_format = + src_tz.size() == 2 ? mkldnn::memory::format::nc : x->format(); 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"; + const std::string key_src_layout = + key + ctx.op().Output("Out") + "@eltwise_fwd_src_layout"; + const std::string key_with_layout = key + std::to_string(src_format); + const std::string key_src_mem = key_with_layout + "@eltwise_fwd_src_mem"; + const std::string key_dst_mem = key_with_layout + "@eltwise_fwd_dst_mem"; + const std::string key_fwd = key_with_layout + "@eltwise_fwd"; + const std::string key_fwd_pd = key_with_layout + "@eltwise_fwd_pd"; + + // save input data and layout to be referred in backward path + auto p_src_data = std::make_shared(x_data); + dev_ctx.SetBlob(key_src_data, p_src_data); + auto p_src_layout = std::make_shared(src_format); + dev_ctx.SetBlob(key_src_layout, p_src_layout); 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); + std::shared_ptr dst_memory; 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())); + // create mkldnn memory for input X + auto src_md = platform::MKLDNNMemDesc( + src_tz, platform::MKLDNNGetDataType(), src_format); + auto src_memory = std::shared_ptr( + new memory({src_md, mkldnn_engine}, to_void_cast(x_data))); + // save src_memory to be referred in backward path + dev_ctx.SetBlob(key_src_mem, src_memory); + + // create primitive descriptor for activation forward and save it + auto forward_desc = mkldnn::eltwise_forward::desc( + mkldnn::prop_kind::forward_training, algorithm, + src_memory->get_primitive_desc().desc(), alpha, beta); + auto forward_pd = std::make_shared( + forward_desc, mkldnn_engine); + + // save prim desc into global device context to be referred in backward path + dev_ctx.SetBlob(key_fwd_pd, forward_pd); + + // create mkldnn memory for output y + dst_memory = + std::make_shared(forward_pd->dst_primitive_desc(), y_data); + + dev_ctx.SetBlob(key_dst_mem, dst_memory); + + // create activation primitive + p_fwd = std::make_shared(*forward_pd, *src_memory, + *dst_memory); dev_ctx.SetBlob(key_fwd, p_fwd); } else { // primitives already exist - auto p_src_mem = + auto src_memory = 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 = + PADDLE_ENFORCE(src_memory != nullptr, + "Fail to find eltwise src_memory in device context."); + dst_memory = 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."); + PADDLE_ENFORCE(dst_memory != nullptr, + "Fail to find eltwise dst_memory in device context."); - p_src_mem->set_data_handle(platform::to_void_reinterpret_cast(src_data)); - p_dst_mem->set_data_handle(dst_data); + src_memory->set_data_handle(platform::to_void_cast(x_data)); + dst_memory->set_data_handle(y_data); } // push primitive to stream and wait until it's executed - std::vector pipeline = {*(p_fwd.get())}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + std::vector pipeline; + pipeline.push_back(*p_fwd); + stream(stream::kind::eager).submit(pipeline).wait(); + + y->set_layout(DataLayout::kMKLDNN); + y->set_format(GetMKLDNNFormat(*dst_memory)); } -template -void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm, - const T alpha = 0, const T beta = 0) { +template +void eltwise_grad(const framework::ExecutionContext &ctx, + mkldnn::algorithm algorithm, const T alpha = 0, + const T beta = 0) { auto &dev_ctx = ctx.template device_context(); const auto &mkldnn_engine = dev_ctx.GetEngine(); - // get buffers - const auto *out = ctx.template Input("Out"); - - auto *dout = ctx.template Input(framework::GradVarName("Out")); - const auto *diff_dst = dout->template data(); + const auto *diff_y = ctx.Input(framework::GradVarName("Out")); + auto *diff_x = ctx.Output(framework::GradVarName("X")); - auto *dx = - ctx.template Output(framework::GradVarName("X")); - const T *diff_src = dx->template mutable_data(ctx.GetPlace()); + const T *diff_y_data = diff_y->data(); + T *diff_x_data = diff_x->mutable_data(ctx.GetPlace()); - // get memory dim - std::vector src_tz = framework::vectorize2int(out->dims()); + std::vector diff_dst_tz = framework::vectorize2int(diff_y->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"; + auto diff_y_format = + diff_dst_tz.size() == 2 ? mkldnn::memory::format::nc : diff_y->format(); + const std::string key = gethash(diff_dst_tz, algorithm); const std::string key_src_data = key + ctx.op().Input("Out") + "@eltwise_fwd_src_data"; + const std::string key_src_layout = + key + ctx.op().Input("Out") + "@eltwise_fwd_src_layout"; + const auto p_src_layout = + std::static_pointer_cast(dev_ctx.GetBlob(key_src_layout)); + const std::string key_src_mem = + key + std::to_string(*p_src_layout) + "@eltwise_fwd_src_mem"; + const std::string key_fwd_pd = + key + std::to_string(*p_src_layout) + "@eltwise_fwd_pd"; + const std::string key_with_layouts = + key + std::to_string(*p_src_layout) + "-" + std::to_string(diff_y_format); + const std::string key_diff_src_mem = + key_with_layouts + "@eltwise_diff_src_mem"; + const std::string key_diff_dst_mem = + key_with_layouts + "@eltwise_diff_dst_mem"; + const std::string key_grad = key_with_layouts + "@eltwise_grad"; + 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 = + auto src_memory = std::static_pointer_cast(dev_ctx.GetBlob(key_src_mem)); - p_src_mem->set_data_handle(*p_src_data.get()); + PADDLE_ENFORCE(src_memory != nullptr, + "Fail to find src_memory in device context"); + src_memory->set_data_handle(*p_src_data.get()); + + std::shared_ptr diff_src_memory; - auto p_grad = std::static_pointer_cast( + 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); - + // create mkldnn memory for input diff_y + auto diff_dst_md = platform::MKLDNNMemDesc( + diff_dst_tz, platform::MKLDNNGetDataType(), diff_y_format); + auto diff_dst_memory = std::shared_ptr( + new memory({diff_dst_md, mkldnn_engine}, to_void_cast(diff_y_data))); + dev_ctx.SetBlob(key_diff_dst_mem, diff_dst_memory); + + // retrieve eltwise primitive desc from device context + auto forward_pd = + std::static_pointer_cast( + dev_ctx.GetBlob(key_fwd_pd)); + PADDLE_ENFORCE(forward_pd != nullptr, + "Fail to find eltwise_fwd_pd in device context"); + + // ceate primitive descriptor for activation backward + auto backward_desc = mkldnn::eltwise_backward::desc( + algorithm, diff_dst_memory->get_primitive_desc().desc(), + src_memory->get_primitive_desc().desc(), alpha, beta); + auto backward_pd = mkldnn::eltwise_backward::primitive_desc( + backward_desc, mkldnn_engine, *forward_pd); + + // create mkldnn memory for output diff_src + diff_src_memory = std::make_shared( + backward_pd.diff_src_primitive_desc(), diff_x_data); + dev_ctx.SetBlob(key_diff_src_mem, diff_src_memory); + + // create activation backward primitive 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()))); + backward_pd, *src_memory, *diff_dst_memory, *diff_src_memory); + dev_ctx.SetBlob(key_grad, p_grad); } else { // primitives already exist - auto p_diff_src_mem = std::static_pointer_cast( + diff_src_memory = std::static_pointer_cast( dev_ctx.GetBlob(key_diff_src_mem)); - auto p_diff_dst_mem = std::static_pointer_cast( + auto diff_dst_memory = 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)); + diff_src_memory->set_data_handle( + platform::to_void_reinterpret_cast(diff_x_data)); + diff_dst_memory->set_data_handle( + platform::to_void_reinterpret_cast(diff_y_data)); } // push primitive to stream and wait until it's executed - std::vector pipeline = {*(p_grad.get())}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + std::vector pipeline; + pipeline.push_back(*p_grad); + stream(stream::kind::eager).submit(pipeline).wait(); + + diff_x->set_layout(DataLayout::kMKLDNN); + diff_x->set_format(GetMKLDNNFormat(*diff_src_memory)); } -} // anonymous namespace template struct MKLDNNActivationFunc : public BaseActivationFunctor { - template - void operator()(const ExecContext &ctx) const { + void operator()(const framework::ExecutionContext &ctx) const { eltwise_forward(ctx, algorithm); } }; template struct MKLDNNActivationGradFunc : public BaseActivationFunctor { - template - void operator()(const ExecContext &ctx) const { + void operator()(const framework::ExecutionContext &ctx) const { eltwise_grad(ctx, algorithm); } }; diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index a06ca7952f8556671fa0662329be4eb7dfefc984..286b03d7b7d11a50f33f0190c1a5b9097ed0f4a2 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -19,18 +19,20 @@ 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: \ - void Make() override { \ - AddInput("X", "Input of " #OP_NAME " operator"); \ - AddOutput("Out", "Output of " #OP_NAME " operator").Reuse("X"); \ - AddAttr("use_mkldnn", \ - "(default false) Only used in mkldnn kernel") \ - .SetDefault(false); \ - AddComment(OP_COMMENT); \ - } \ +using paddle::framework::Tensor; + +#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").Reuse("X"); \ + 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) \ @@ -58,7 +60,6 @@ framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx, const framework::OperatorWithKernel& oper, const std::string& name) { framework::LibraryType library{framework::LibraryType::kPlain}; - framework::DataLayout layout = framework::DataLayout::kAnyLayout; #ifdef PADDLE_WITH_MKLDNN auto it = oper.Attrs().find("use_mkldnn"); @@ -82,6 +83,7 @@ class ActivationOp : public framework::OperatorWithKernel { ctx->ShareLoD("X", /*->*/ "Out"); } + protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return GetKernelType(ctx, *this, "X"); @@ -96,6 +98,7 @@ class ActivationOpGrad : public framework::OperatorWithKernel { ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out")); } + protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return GetKernelType(ctx, *this, "Out"); @@ -140,7 +143,7 @@ $$out = \\frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$ __attribute__((unused)) constexpr char TanhShrinkDoc[] = R"DOC( TanhShrink Activation Operator. -$$out = x - \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$ +$$out = x - \\frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$ )DOC"; @@ -382,7 +385,7 @@ class STanhOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( STanh Activation Operator. -$$out = b * \frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$ +$$out = b * \\frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$ )DOC"); } diff --git a/paddle/fluid/operators/batch_norm_mkldnn_op.cc b/paddle/fluid/operators/batch_norm_mkldnn_op.cc index 8206cc9890160da756efb13c991020f09b20126a..cc158e57f7140c84f02bc7e091d8eac0d2b672e1 100644 --- a/paddle/fluid/operators/batch_norm_mkldnn_op.cc +++ b/paddle/fluid/operators/batch_norm_mkldnn_op.cc @@ -21,8 +21,6 @@ namespace operators { using batch_norm_bwd = mkldnn::batch_normalization_backward; using batch_norm_fwd = mkldnn::batch_normalization_forward; -using framework::DataLayout; -using framework::Tensor; using mkldnn::memory; using mkldnn::primitive; using mkldnn::reorder; @@ -31,18 +29,6 @@ using paddle::platform::MKLDNNDeviceContext; using paddle::platform::MKLDNNMemDesc; using platform::to_void_cast; -template -using EigenArrayMap = - Eigen::Map>; -template -using ConstEigenArrayMap = - Eigen::Map>; -template -using EigenVectorArrayMap = Eigen::Map>; -template -using ConstEigenVectorArrayMap = - Eigen::Map>; - namespace { template struct bn_type_traits { diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc index 625ca2d7c4c70d1098b0fb28380d8d1eb24cb338..52b0bf85c07fee380f9e7ba1c703b56367628644 100644 --- a/paddle/fluid/operators/batch_norm_op.cc +++ b/paddle/fluid/operators/batch_norm_op.cc @@ -22,22 +22,6 @@ limitations under the License. */ namespace paddle { namespace operators { -using Tensor = framework::Tensor; -using LoDTensor = framework::LoDTensor; -using DataLayout = framework::DataLayout; - -template -using EigenArrayMap = - Eigen::Map>; -template -using ConstEigenArrayMap = - Eigen::Map>; -template -using EigenVectorArrayMap = Eigen::Map>; -template -using ConstEigenVectorArrayMap = - Eigen::Map>; - class BatchNormOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; diff --git a/paddle/fluid/operators/batch_norm_op.h b/paddle/fluid/operators/batch_norm_op.h index 9e5fc41598f29336074335f3624a2300ad018d09..5e3d630d6889e445c5e84fa836d2d81bb7266779 100644 --- a/paddle/fluid/operators/batch_norm_op.h +++ b/paddle/fluid/operators/batch_norm_op.h @@ -19,6 +19,22 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; +using DataLayout = framework::DataLayout; + +template +using EigenArrayMap = + Eigen::Map>; +template +using ConstEigenArrayMap = + Eigen::Map>; +template +using EigenVectorArrayMap = Eigen::Map>; +template +using ConstEigenVectorArrayMap = + Eigen::Map>; + template class BatchNormKernel : public framework::OpKernel { public: diff --git a/paddle/fluid/operators/bilinear_interp_op.cc b/paddle/fluid/operators/bilinear_interp_op.cc index 2572e813d656353a2187c29da89266733a32f3ce..2dc3399da183fbcf7664066f6f7ce12db3dc6d5e 100644 --- a/paddle/fluid/operators/bilinear_interp_op.cc +++ b/paddle/fluid/operators/bilinear_interp_op.cc @@ -110,6 +110,7 @@ REGISTER_OPERATOR(bilinear_interp, ops::BilinearInterpOp, ops::BilinearInterpOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(bilinear_interp_grad, ops::BilinearInterpOpGrad); -REGISTER_OP_CPU_KERNEL(bilinear_interp, ops::BilinearInterpKernel); +REGISTER_OP_CPU_KERNEL(bilinear_interp, ops::BilinearInterpKernel, + ops::BilinearInterpKernel); REGISTER_OP_CPU_KERNEL(bilinear_interp_grad, ops::BilinearInterpGradKernel); diff --git a/paddle/fluid/operators/bilinear_interp_op.h b/paddle/fluid/operators/bilinear_interp_op.h index 8b03cd5a0635584a45782fe5a4823c37fe4fa8e8..70847cb8c1abe2e94bc844ab8117d1f23fea533b 100644 --- a/paddle/fluid/operators/bilinear_interp_op.h +++ b/paddle/fluid/operators/bilinear_interp_op.h @@ -46,8 +46,10 @@ class BilinearInterpKernel : public framework::OpKernel { int in_chw = channels * in_hw; int out_chw = channels * out_hw; - T ratio_h = (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; - T ratio_w = (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + float ratio_h = + (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; + float ratio_w = + (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; if (in_h == out_h && in_w == out_w) { memcpy(output, input, input_t->numel() * sizeof(T)); @@ -56,24 +58,24 @@ class BilinearInterpKernel : public framework::OpKernel { for (int i = 0; i < out_h; ++i) { // loop for images int h = ratio_h * i; int hid = (h < in_h - 1) ? 1 : 0; - T h1lambda = ratio_h * i - h; - T h2lambda = 1 - h1lambda; + float h1lambda = ratio_h * i - h; + float h2lambda = 1.f - h1lambda; for (int j = 0; j < out_w; ++j) { int w = ratio_w * j; int wid = (w < in_w - 1) ? 1 : 0; - T w1lambda = ratio_w * j - w; - T w2lambda = 1 - w1lambda; + float w1lambda = ratio_w * j - w; + float w2lambda = 1.f - w1lambda; // calculate four position for bilinear interpolation const T* in_pos = &input[k * in_chw + h * in_w + w]; T* out_pos = &output[k * out_chw + i * out_w + j]; for (int c = 0; c < channels; ++c) { // loop for channels // bilinear interpolation - out_pos[0] = + out_pos[0] = static_cast( h2lambda * (w2lambda * in_pos[0] + w1lambda * in_pos[wid]) + h1lambda * (w2lambda * in_pos[hid * in_w] + - w1lambda * in_pos[hid * in_w + wid]); + w1lambda * in_pos[hid * in_w + wid])); in_pos += in_hw; out_pos += out_hw; } @@ -117,8 +119,10 @@ class BilinearInterpGradKernel : public framework::OpKernel { int in_chw = channels * in_hw; int out_chw = channels * out_hw; - T ratio_h = (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; - T ratio_w = (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + float ratio_h = + (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; + float ratio_w = + (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; if (in_h == out_h && in_w == out_w) { memcpy(d_input, d_output, d_input_t->numel() * sizeof(T)); @@ -127,22 +131,24 @@ class BilinearInterpGradKernel : public framework::OpKernel { for (int i = 0; i < out_h; ++i) { // loop for images int h = ratio_h * i; int hid = (h < in_h - 1) ? 1 : 0; - T h1lambda = ratio_h * i - h; - T h2lambda = 1 - h1lambda; + float h1lambda = ratio_h * i - h; + float h2lambda = 1 - h1lambda; for (int j = 0; j < out_w; ++j) { int w = ratio_w * j; int wid = (w < in_w - 1) ? 1 : 0; - T w1lambda = ratio_w * j - w; - T w2lambda = 1 - w1lambda; + float w1lambda = ratio_w * j - w; + float w2lambda = 1 - w1lambda; T* in_pos = &d_input[k * in_chw + h * in_w + w]; const T* out_pos = &d_output[k * out_chw + i * out_w + j]; for (int c = 0; c < channels; ++c) { // loop for channels - in_pos[0] += h2lambda * w2lambda * out_pos[0]; - in_pos[wid] += h2lambda * w1lambda * out_pos[0]; - in_pos[hid * in_w] += h1lambda * w2lambda * out_pos[0]; - in_pos[hid * in_w + wid] += h1lambda * w1lambda * out_pos[0]; + in_pos[0] += static_cast(h2lambda * w2lambda * out_pos[0]); + in_pos[wid] += static_cast(h2lambda * w1lambda * out_pos[0]); + in_pos[hid * in_w] += + static_cast(h1lambda * w2lambda * out_pos[0]); + in_pos[hid * in_w + wid] += + static_cast(h1lambda * w1lambda * out_pos[0]); in_pos += in_hw; out_pos += out_hw; } diff --git a/paddle/fluid/operators/concat_op.h b/paddle/fluid/operators/concat_op.h index 1b1b8bf5ed959dd9c2ce8c9f5c905a75b81865fd..a496301526f58875ff51aeaa5b2094c3c656531c 100644 --- a/paddle/fluid/operators/concat_op.h +++ b/paddle/fluid/operators/concat_op.h @@ -60,34 +60,45 @@ template class ConcatGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const { - auto* in = ctx.Input(framework::GradVarName("Out")); + auto* out_grad = + ctx.Input(framework::GradVarName("Out")); + auto ins = ctx.MultiInput("X"); + auto out_var_names = ctx.Outputs(framework::GradVarName("X")); auto outs = ctx.MultiOutput(framework::GradVarName("X")); int64_t axis = static_cast(ctx.Attr("axis")); + // get output tensor that the name is not kEmptyVarName + std::vector outputs; + for (size_t j = 0; j < outs.size(); ++j) { + if (out_var_names[j] != framework::kEmptyVarName) { + outs[j]->mutable_data(ctx.GetPlace()); + outputs.push_back(outs[j]); + } else { + outputs.push_back(nullptr); + } + } + // Sometimes direct copies will be faster, this maybe need deeply analysis. if (axis == 0 && outs.size() < 10) { size_t input_offset = 0; - auto in_stride = framework::stride_numel(in->dims()); + const auto in_stride = framework::stride_numel(out_grad->dims()); - for (auto& out : outs) { - out->mutable_data(ctx.GetPlace()); - auto out_stride = framework::stride_numel(out->dims()); - StridedNumelCopyWithAxis(ctx.device_context(), axis, out->data(), - out_stride, in->data() + input_offset, - in_stride, out_stride[axis]); + for (size_t i = 0; i < outs.size(); ++i) { + auto out_stride = framework::stride_numel(ins[i]->dims()); + auto* out = outputs[i]; + if (out != nullptr) { + StridedNumelCopyWithAxis( + ctx.device_context(), axis, out->data(), out_stride, + out_grad->data() + input_offset, in_stride, out_stride[axis]); + } input_offset += out_stride[axis]; } } else { - std::vector outputs(outs.size()); - for (size_t j = 0; j < outs.size(); ++j) { - outs[j]->mutable_data(ctx.GetPlace()); - outputs[j] = *outs[j]; - } - auto& dev_ctx = ctx.template device_context(); paddle::operators::math::ConcatGradFunctor concat_grad_functor; - concat_grad_functor(dev_ctx, *in, static_cast(axis), &outputs); + concat_grad_functor(dev_ctx, *out_grad, ins, static_cast(axis), + &outputs); } } }; diff --git a/paddle/fluid/operators/detail/macros.h b/paddle/fluid/operators/detail/macros.h index da1de72dad00db3ffe609e17bd198ef0a56bbfcd..b9e385994efcea0388756e8bd780ebfc719ed08d 100644 --- a/paddle/fluid/operators/detail/macros.h +++ b/paddle/fluid/operators/detail/macros.h @@ -15,13 +15,13 @@ #pragma once #ifdef PADDLE_WITH_GRPC -#include "paddle/fluid/operators/detail/grpc_client.h" -#include "paddle/fluid/operators/detail/grpc_server.h" -#define RPCSERVER_T detail::AsyncGRPCServer -#define RPCCLIENT_T detail::GRPCClient +#include "paddle/fluid/operators/distributed/grpc_client.h" +#include "paddle/fluid/operators/distributed/grpc_server.h" +#define RPCSERVER_T distributed::AsyncGRPCServer +#define RPCCLIENT_T distributed::GRPCClient #else -#include "paddle/fluid/operators/detail/brpc_client.h" -#include "paddle/fluid/operators/detail/brpc_server.h" -#define RPCSERVER_T detail::AsyncBRPCServer -#define RPCCLIENT_T detail::BRPCClient +#include "paddle/fluid/operators/distributed/brpc_client.h" +#include "paddle/fluid/operators/distributed/brpc_server.h" +#define RPCSERVER_T distributed::AsyncBRPCServer +#define RPCCLIENT_T distributed::BRPCClient #endif diff --git a/paddle/fluid/operators/detection_map_op.cc b/paddle/fluid/operators/detection_map_op.cc index 716c8625d35308f98582e6802e90d99d643e188b..d7f49a9590e4ef4ca4d2ad5a92572c70e6bfb6ac 100644 --- a/paddle/fluid/operators/detection_map_op.cc +++ b/paddle/fluid/operators/detection_map_op.cc @@ -175,12 +175,12 @@ class DetectionMAPOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( Detection mAP evaluate operator. The general steps are as follows. First, calculate the true positive and - false positive according to the input of detection and labels, then - calculate the mAP evaluate value. - Supporting '11 point' and 'integral' mAP algorithm. Please get more information - from the following articles: - https://sanchom.wordpress.com/tag/average-precision/ - https://arxiv.org/abs/1512.02325 +false positive according to the input of detection and labels, then +calculate the mAP evaluate value. +Supporting '11 point' and 'integral' mAP algorithm. Please get more information +from the following articles: +https://sanchom.wordpress.com/tag/average-precision/ +https://arxiv.org/abs/1512.02325 )DOC"); } diff --git a/paddle/fluid/operators/detail/CMakeLists.txt b/paddle/fluid/operators/distributed/CMakeLists.txt similarity index 97% rename from paddle/fluid/operators/detail/CMakeLists.txt rename to paddle/fluid/operators/distributed/CMakeLists.txt index abc5aad0430e71928a441c9488dda16dfdd63b9c..312f80e09077f21a47985c1c936c2ac41c292ead 100644 --- a/paddle/fluid/operators/detail/CMakeLists.txt +++ b/paddle/fluid/operators/distributed/CMakeLists.txt @@ -1,8 +1,3 @@ -if(NOT WITH_DISTRIBUTE) - return() -endif() - - if(WITH_GRPC) grpc_library(sendrecvop_grpc SRCS bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc request_handler_impl.cc rpc_client.cc rpc_server.cc grpc_server.cc variable_response.cc PROTO send_recv.proto DEPS lod_tensor diff --git a/paddle/fluid/operators/detail/brpc_client.cc b/paddle/fluid/operators/distributed/brpc_client.cc similarity index 98% rename from paddle/fluid/operators/detail/brpc_client.cc rename to paddle/fluid/operators/distributed/brpc_client.cc index 9a4e410f1d83e93883438fae116c38eb60787673..b394c678fb6503eb73a1e11e6feb814251e9e940 100644 --- a/paddle/fluid/operators/detail/brpc_client.cc +++ b/paddle/fluid/operators/distributed/brpc_client.cc @@ -12,12 +12,12 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/operators/detail/brpc_client.h" +#include "paddle/fluid/operators/distributed/brpc_client.h" #include "paddle/fluid/framework/threadpool.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { DEFINE_int32(brpc_channel_num, 24, "Number of channels to send requests connected to one server"); @@ -175,6 +175,6 @@ ChannelQueuePtr BRPCClient::GetChannel(const std::string& ep) { return q; } -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/brpc_client.h b/paddle/fluid/operators/distributed/brpc_client.h similarity index 94% rename from paddle/fluid/operators/detail/brpc_client.h rename to paddle/fluid/operators/distributed/brpc_client.h index 1e953ea431d51a9586bfd0b352c7f27d079ff1a8..34f140687f91d866536f5e2b647c7445a6624736 100644 --- a/paddle/fluid/operators/detail/brpc_client.h +++ b/paddle/fluid/operators/distributed/brpc_client.h @@ -31,13 +31,13 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/operators/detail/rpc_client.h" -#include "paddle/fluid/operators/detail/send_recv.pb.h" +#include "paddle/fluid/operators/distributed/rpc_client.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" #include "paddle/fluid/platform/macros.h" // for DISABLE_COPY_AND_ASSIGN namespace paddle { namespace operators { -namespace detail { +namespace distributed { struct ChannelContext { brpc::Channel channel; @@ -95,6 +95,6 @@ class BRPCClient : public RPCClient { DISABLE_COPY_AND_ASSIGN(BRPCClient); }; -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/brpc_server.cc b/paddle/fluid/operators/distributed/brpc_server.cc similarity index 86% rename from paddle/fluid/operators/detail/brpc_server.cc rename to paddle/fluid/operators/distributed/brpc_server.cc index 2170abe679f9ededff3b53e3139e56f8aad227cb..862167f02084cfe81db1c0936bbfb0415fa85721 100644 --- a/paddle/fluid/operators/detail/brpc_server.cc +++ b/paddle/fluid/operators/distributed/brpc_server.cc @@ -12,13 +12,13 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/operators/detail/brpc_server.h" -#include "paddle/fluid/operators/detail/request_handler.h" +#include "paddle/fluid/operators/distributed/brpc_server.h" +#include "paddle/fluid/operators/distributed/request_handler.h" namespace sendrecv { typedef std::unordered_map + paddle::operators::distributed::RequestHandler*> HandlerMap; class BRPCServiceImpl : public SendRecvService { @@ -27,17 +27,17 @@ class BRPCServiceImpl : public SendRecvService { : request_send_h_(nullptr), request_get_h_(nullptr), request_prefetch_h_(nullptr) { - auto it = rpc_call_map.find(paddle::operators::detail::kRequestSend); + auto it = rpc_call_map.find(paddle::operators::distributed::kRequestSend); if (it != rpc_call_map.end()) { request_send_h_ = it->second; } - it = rpc_call_map.find(paddle::operators::detail::kRequestSend); + it = rpc_call_map.find(paddle::operators::distributed::kRequestSend); if (it != rpc_call_map.end()) { request_get_h_ = it->second; } - it = rpc_call_map.find(paddle::operators::detail::kRequestPrefetch); + it = rpc_call_map.find(paddle::operators::distributed::kRequestPrefetch); if (it != rpc_call_map.end()) { request_prefetch_h_ = it->second; } @@ -88,15 +88,15 @@ class BRPCServiceImpl : public SendRecvService { } private: - paddle::operators::detail::RequestHandler* request_send_h_; - paddle::operators::detail::RequestHandler* request_get_h_; - paddle::operators::detail::RequestHandler* request_prefetch_h_; + paddle::operators::distributed::RequestHandler* request_send_h_; + paddle::operators::distributed::RequestHandler* request_get_h_; + paddle::operators::distributed::RequestHandler* request_prefetch_h_; }; } // namespace sendrecv namespace paddle { namespace operators { -namespace detail { +namespace distributed { void AsyncBRPCServer::StartServer() { // Instance of your service. @@ -139,6 +139,6 @@ void AsyncBRPCServer::WaitServerReady() { VLOG(3) << "AsyncGRPCServer WaitSeverReady"; } -}; // namespace detail +}; // namespace distributed }; // namespace operators }; // namespace paddle diff --git a/paddle/fluid/operators/detail/brpc_server.h b/paddle/fluid/operators/distributed/brpc_server.h similarity index 88% rename from paddle/fluid/operators/detail/brpc_server.h rename to paddle/fluid/operators/distributed/brpc_server.h index 0105c8074a46849031d8fa9c21a5507a982ec3c3..85a7ad0dfe843dad483d43631b69a79d75211ce9 100644 --- a/paddle/fluid/operators/detail/brpc_server.h +++ b/paddle/fluid/operators/distributed/brpc_server.h @@ -19,12 +19,12 @@ limitations under the License. */ #include #include "brpc/server.h" -#include "paddle/fluid/operators/detail/rpc_server.h" -#include "paddle/fluid/operators/detail/send_recv.pb.h" +#include "paddle/fluid/operators/distributed/rpc_server.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { class AsyncBRPCServer final : public RPCServer { public: @@ -48,6 +48,6 @@ class AsyncBRPCServer final : public RPCServer { int ready_; }; -}; // namespace detail +}; // namespace distributed }; // namespace operators }; // namespace paddle diff --git a/paddle/fluid/operators/detail/bytebuffer_stream.cc b/paddle/fluid/operators/distributed/bytebuffer_stream.cc similarity index 94% rename from paddle/fluid/operators/detail/bytebuffer_stream.cc rename to paddle/fluid/operators/distributed/bytebuffer_stream.cc index a14171563edb0ac9a22b7ae493c965de3efb7823..6e91b447db838c9095432eda22e9e1171e938d31 100644 --- a/paddle/fluid/operators/detail/bytebuffer_stream.cc +++ b/paddle/fluid/operators/distributed/bytebuffer_stream.cc @@ -17,11 +17,11 @@ limitations under the License. */ // file and did some modifications so that we can send gRPC // requests without too much copying of the tensor data. -#include "paddle/fluid/operators/detail/bytebuffer_stream.h" +#include "paddle/fluid/operators/distributed/bytebuffer_stream.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { GrpcByteBufferSource::GrpcByteBufferSource() {} @@ -83,6 +83,6 @@ google::protobuf::int64 GrpcByteBufferSource::ByteCount() const { return byte_count_; } -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/bytebuffer_stream.h b/paddle/fluid/operators/distributed/bytebuffer_stream.h similarity index 99% rename from paddle/fluid/operators/detail/bytebuffer_stream.h rename to paddle/fluid/operators/distributed/bytebuffer_stream.h index 054dd4ff294414cca55d7e033f2c5403bbb85526..e7de172c79c30761483b5d96f5bad19860208832 100644 --- a/paddle/fluid/operators/detail/bytebuffer_stream.h +++ b/paddle/fluid/operators/distributed/bytebuffer_stream.h @@ -106,7 +106,7 @@ class GrpcBufferReader final namespace paddle { namespace operators { -namespace detail { +namespace distributed { // Source provides a way for a particular RPC implementation to provide // received data to ParseFrom. class Source { @@ -183,6 +183,6 @@ class GrpcByteSource : public Source { char space_[sizeof(Reader)]; }; -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_client.cc b/paddle/fluid/operators/distributed/grpc_client.cc similarity index 93% rename from paddle/fluid/operators/detail/grpc_client.cc rename to paddle/fluid/operators/distributed/grpc_client.cc index ea004f7cd340030e61571825941a50e89735ef05..52f931188dc790682626b14da83d0835cad4f1a6 100644 --- a/paddle/fluid/operators/detail/grpc_client.cc +++ b/paddle/fluid/operators/distributed/grpc_client.cc @@ -12,19 +12,20 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/distributed/grpc_client.h" #include #include +#include "glog/logging.h" // For VLOG #include "paddle/fluid/framework/threadpool.h" -#include "paddle/fluid/operators/detail/request_handler.h" +#include "paddle/fluid/operators/distributed/request_handler.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { void GRPCClient::InitImpl() { InitEventLoop(); } @@ -75,6 +76,9 @@ bool GRPCClient::AsyncSendVar(const std::string& ep, var_h.scope = p_scope; var_h.name = var_name_val; var_h.ctx = p_ctx; + var_h.method = "Send"; + + VLOG(3) << var_h.String() << " begin"; // stub context SendProcessor* s = new SendProcessor(ch); @@ -129,6 +133,9 @@ bool GRPCClient::AsyncGetVar(const std::string& ep, var_h.scope = p_scope; var_h.name = var_name_val; var_h.ctx = p_ctx; + var_h.method = "Get"; + + VLOG(3) << var_h.String() << " begin"; // stub context GetProcessor* s = new GetProcessor(ch); @@ -172,6 +179,9 @@ bool GRPCClient::AsyncPrefetchVar(const std::string& ep, var_h.scope = p_scope; var_h.name = out_var_name_val; var_h.ctx = p_ctx; + var_h.method = "Prefetch"; + + VLOG(3) << var_h.String() << " begin"; // stub context GetProcessor* s = new GetProcessor(ch); @@ -243,10 +253,11 @@ void GRPCClient::Proceed() { GPR_ASSERT(ok); PADDLE_ENFORCE(c); if (c->status_.ok()) { + VLOG(3) << c->var_h_.String() << " process"; c->Process(); } else { - LOG(FATAL) << "var: " << c->var_h_.String() - << " grpc error:" << c->status_.error_message(); + LOG(FATAL) << c->var_h_.String() + << " meets grpc error:" << c->status_.error_message(); } delete c; { @@ -276,6 +287,6 @@ std::shared_ptr GRPCClient::GetChannel(const std::string& ep) { return ch; } -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_client.h b/paddle/fluid/operators/distributed/grpc_client.h similarity index 95% rename from paddle/fluid/operators/detail/grpc_client.h rename to paddle/fluid/operators/distributed/grpc_client.h index 44000c028b499d9ad1a0e0dd40a5e287cd61d143..7875939ff510e7e41a2a11ca965b52eedff3d05c 100644 --- a/paddle/fluid/operators/detail/grpc_client.h +++ b/paddle/fluid/operators/distributed/grpc_client.h @@ -38,23 +38,27 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/operators/detail/rpc_client.h" -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#include "paddle/fluid/operators/distributed/rpc_client.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" #include "paddle/fluid/platform/macros.h" // for DISABLE_COPY_AND_ASSIGN namespace paddle { namespace operators { -namespace detail { +namespace distributed { struct VarHandle { + // RPC endpoint. std::string ep; const platform::DeviceContext* ctx; const framework::Scope* scope; + // Variable name. std::string name; + // RPC method name. + std::string method; std::string String() const { std::ostringstream s; - s << "name:[" << name << "] ep:[" << ep << "]"; + s << method << " name:[" << name << "], ep:[" << ep << "]"; return s.str(); } }; @@ -226,6 +230,6 @@ class GRPCClient : public RPCClient { DISABLE_COPY_AND_ASSIGN(GRPCClient); }; -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_serde_test.cc b/paddle/fluid/operators/distributed/grpc_serde_test.cc similarity index 93% rename from paddle/fluid/operators/detail/grpc_serde_test.cc rename to paddle/fluid/operators/distributed/grpc_serde_test.cc index 15892295e6901fe649788c9e34604008fc8cbdfa..3d107b533bcb7bfef3f9b13ec99afbd579a62e52 100644 --- a/paddle/fluid/operators/detail/grpc_serde_test.cc +++ b/paddle/fluid/operators/distributed/grpc_serde_test.cc @@ -21,8 +21,8 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/framework/variable.h" -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" -#include "paddle/fluid/operators/detail/variable_response.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" +#include "paddle/fluid/operators/distributed/variable_response.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/platform/place.h" #include "paddle/fluid/string/printf.h" @@ -50,7 +50,7 @@ void RunSerdeTestSelectedRows(platform::Place place) { for (int i = 0; i < 564; ++i) rows->push_back(i); ::grpc::ByteBuffer msg; - operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg); + operators::distributed::SerializeToByteBuffer("myvar", &var, ctx, &msg); EXPECT_GT(msg.Length(), static_cast(0)); // deserialize @@ -81,10 +81,10 @@ void RunSerdeTestSelectedRows(platform::Place place) { // deserialize zero-copy // framework::Variable var2; - // operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2); + // operators::distributed::DeserializeFromByteBuffer(msg, ctx, &var2); framework::Scope scope; scope.Var("myvar"); - operators::detail::VariableResponse resp(&scope, &ctx); + operators::distributed::VariableResponse resp(&scope, &ctx); EXPECT_EQ(resp.Parse(msg), 0); framework::Variable* var2 = resp.GetVar(); @@ -128,7 +128,7 @@ void RunTestLodTensor(platform::Place place, int from_type = 0) { math::set_constant(ctx, tensor, 31.9); ::grpc::ByteBuffer msg; - operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg); + operators::distributed::SerializeToByteBuffer("myvar", &var, ctx, &msg); EXPECT_GT(msg.Length(), static_cast(0)); // deserialize @@ -171,7 +171,7 @@ void RunTestLodTensor(platform::Place place, int from_type = 0) { // deserialize zero-copy framework::Scope scope; scope.Var("myvar"); - operators::detail::VariableResponse resp(&scope, &ctx); + operators::distributed::VariableResponse resp(&scope, &ctx); if (from_type == 0) { EXPECT_EQ(resp.Parse(msg), 0); } else { diff --git a/paddle/fluid/operators/detail/grpc_server.cc b/paddle/fluid/operators/distributed/grpc_server.cc similarity index 91% rename from paddle/fluid/operators/detail/grpc_server.cc rename to paddle/fluid/operators/distributed/grpc_server.cc index 5a87258901c6563fe793d4041f344011a56d9a01..b9a9b12cecdada570da5af173e394999554e9cb8 100644 --- a/paddle/fluid/operators/detail/grpc_server.cc +++ b/paddle/fluid/operators/distributed/grpc_server.cc @@ -15,13 +15,13 @@ limitations under the License. */ #include #include -#include "paddle/fluid/operators/detail/grpc_server.h" +#include "paddle/fluid/operators/distributed/grpc_server.h" using ::grpc::ServerAsyncResponseWriter; namespace paddle { namespace operators { -namespace detail { +namespace distributed { enum CallStatus { PROCESS = 0, FINISH }; // reference: @@ -41,6 +41,19 @@ class RequestBase { virtual ~RequestBase() {} virtual void Process() = 0; + std::string Status2String(const std::string& method) { + std::string status = "Process"; + if (status_ == FINISH) { + status = "Finish"; + } + + std::ostringstream s; + s << method << " name:[" << GetReqName() << "]" + << ", ep:[" << ctx_.peer() << "]" + << " " << status << " using req_id:" << req_id_; + return s.str(); + } + CallStatus Status() const { std::lock_guard l(status_mu_); return status_; @@ -74,7 +87,7 @@ class RequestSend final : public RequestBase { request_.reset(new VariableResponse(request_handler->scope(), request_handler->dev_ctx(), !request_handler->sync_mode())); - int method_id = static_cast(detail::GrpcMethod::kSendVariable); + int method_id = static_cast(distributed::GrpcMethod::kSendVariable); service_->RequestAsyncUnary( method_id, &ctx_, request_.get(), &responder_, cq_, cq_, reinterpret_cast(static_cast(req_id))); @@ -106,7 +119,7 @@ class RequestGet final : public RequestBase { ::grpc::ServerCompletionQueue* cq, RequestHandler* request_handler, int req_id) : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_) { - auto method_id = static_cast(detail::GrpcMethod::kGetVariable); + auto method_id = static_cast(distributed::GrpcMethod::kGetVariable); service_->RequestAsyncUnary( method_id, &ctx_, &request_, &responder_, cq_, cq_, reinterpret_cast(static_cast(req_id))); @@ -150,7 +163,8 @@ class RequestPrefetch final : public RequestBase { local_scope_(nullptr) { request_.reset(new VariableResponse(request_handler->scope(), request_handler->dev_ctx(), true)); - int method_id = static_cast(detail::GrpcMethod::kPrefetchVariable); + int method_id = + static_cast(distributed::GrpcMethod::kPrefetchVariable); service_->RequestAsyncUnary( method_id, &ctx_, request_.get(), &responder_, cq_, cq_, reinterpret_cast(static_cast(req_id))); @@ -271,7 +285,7 @@ void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name, int req_id) { std::unique_lock lock(cq_mutex_); if (is_shut_down_) { - VLOG(3) << "shutdown, do not TryToRegisterNewSendOne"; + LOG(WARNING) << "shutdown, do not TryToRegisterNewSendOne"; return; } @@ -305,14 +319,14 @@ void AsyncGRPCServer::HandleRequest( bool ok = false; while (true) { - VLOG(3) << "HandleRequest " << rpc_name << " wait next"; + VLOG(4) << "HandleRequest " << rpc_name << " wait next"; if (!cq->Next(&tag, &ok)) { LOG(INFO) << "CompletionQueue " << rpc_name << " shutdown!"; break; } int req_id = static_cast(reinterpret_cast(tag)); - VLOG(3) << "HandleRequest " << rpc_name << ", req_id:" << req_id + VLOG(4) << "HandleRequest " << rpc_name << ", req_id:" << req_id << " get next"; auto& reqs = rpc_reqs_[rpc_name]; @@ -323,22 +337,21 @@ void AsyncGRPCServer::HandleRequest( base = reqs[req_id]; } + VLOG(3) << base->Status2String(rpc_name); + // reference: // https://github.com/tensorflow/tensorflow/issues/5596 // https://groups.google.com/forum/#!topic/grpc-io/xftlRy-IQwM // https://groups.google.com/forum/#!topic/grpc-io/ywATt88Ef_I if (!ok) { LOG(WARNING) << "completion queue:" << rpc_name - << " recv no regular event:argument name[" - << base->GetReqName() << "]"; + << " recv no regular event" + << " context:" << base->Status2String(rpc_name); TryToRegisterNewOne(rpc_name, req_id); delete base; continue; } - VLOG(3) << "queue id:" << rpc_name << ", req_id:" << req_id - << ", status:" << base->Status(); - switch (base->Status()) { case PROCESS: { base->Process(); @@ -354,6 +367,6 @@ void AsyncGRPCServer::HandleRequest( } } -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_server.h b/paddle/fluid/operators/distributed/grpc_server.h similarity index 85% rename from paddle/fluid/operators/detail/grpc_server.h rename to paddle/fluid/operators/distributed/grpc_server.h index f1db7590f6f14d5d44acc12453861a446e278cd2..d2524f5e65db6dedab78f45e17380359b58a3d11 100644 --- a/paddle/fluid/operators/detail/grpc_server.h +++ b/paddle/fluid/operators/distributed/grpc_server.h @@ -29,17 +29,17 @@ limitations under the License. */ #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/framework/var_type.h" -#include "paddle/fluid/operators/detail/grpc_service.h" -#include "paddle/fluid/operators/detail/request_handler.h" -#include "paddle/fluid/operators/detail/rpc_server.h" -#include "paddle/fluid/operators/detail/send_recv.grpc.pb.h" -#include "paddle/fluid/operators/detail/send_recv.pb.h" -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#include "paddle/fluid/operators/distributed/grpc_service.h" +#include "paddle/fluid/operators/distributed/request_handler.h" +#include "paddle/fluid/operators/distributed/rpc_server.h" +#include "paddle/fluid/operators/distributed/send_recv.grpc.pb.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { class RequestBase; @@ -84,6 +84,6 @@ class AsyncGRPCServer final : public RPCServer { std::map> rpc_reqs_; }; -}; // namespace detail +}; // namespace distributed }; // namespace operators }; // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_service.h b/paddle/fluid/operators/distributed/grpc_service.h similarity index 87% rename from paddle/fluid/operators/detail/grpc_service.h rename to paddle/fluid/operators/distributed/grpc_service.h index e0505c2b9d0903837713d7e0032b01ab091c2e04..141be3e68012743a32e4df5de148a55717f8e9a2 100644 --- a/paddle/fluid/operators/detail/grpc_service.h +++ b/paddle/fluid/operators/distributed/grpc_service.h @@ -23,7 +23,7 @@ #include #include #include -#include "paddle/fluid/operators/detail/variable_response.h" +#include "paddle/fluid/operators/distributed/variable_response.h" #include "paddle/fluid/platform/profiler.h" @@ -42,24 +42,25 @@ class ServerContext; // Support parsing/unparsing of tensorflow::VariableResponse. // Wire-format is identical to RecvVariableResponse. template <> -class SerializationTraits { +class SerializationTraits { public: static Status Serialize( - const paddle::operators::detail::VariableResponse& msg, + const paddle::operators::distributed::VariableResponse& msg, grpc_byte_buffer** bp, bool* own_buffer) { PADDLE_ENFORCE(false, "SerializationTraits::Serialize not implemented!"); return Status(); } - static Status Deserialize(grpc_byte_buffer* buffer, - paddle::operators::detail::VariableResponse* msg, - int max_message_size = INT_MAX) { + static Status Deserialize( + grpc_byte_buffer* buffer, + paddle::operators::distributed::VariableResponse* msg, + int max_message_size = INT_MAX) { if (buffer == nullptr) { return Status(StatusCode::INTERNAL, "No payload"); } Status result = g_core_codegen_interface->ok(); if (result.ok()) { - paddle::operators::detail::GrpcByteSource source(buffer); + paddle::operators::distributed::GrpcByteSource source(buffer); int ret = msg->Parse(&source); if (ret != 0) { result = Status(StatusCode::INTERNAL, "VariableResponse parse error"); @@ -73,7 +74,7 @@ class SerializationTraits { namespace paddle { namespace operators { -namespace detail { +namespace distributed { enum class GrpcMethod { kSendVariable, @@ -118,6 +119,6 @@ class GrpcService final { }; }; -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/proto_encoder_helper.h b/paddle/fluid/operators/distributed/proto_encoder_helper.h similarity index 98% rename from paddle/fluid/operators/detail/proto_encoder_helper.h rename to paddle/fluid/operators/distributed/proto_encoder_helper.h index d91d054b2507f32d1e948dde33da06a70cabe775..2fab02e32fe18ee04f86a69bb5bae1cbe7c6762c 100644 --- a/paddle/fluid/operators/detail/proto_encoder_helper.h +++ b/paddle/fluid/operators/distributed/proto_encoder_helper.h @@ -26,7 +26,7 @@ limitations under the License. */ namespace paddle { namespace operators { -namespace detail { +namespace distributed { char* EncodeVarint32(char* dst, uint32_t v) { // Operate on characters as unsigneds @@ -144,6 +144,6 @@ class ProtoEncodeHelper { char* limit_; // Just for CHECKs }; -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/request_handler.h b/paddle/fluid/operators/distributed/request_handler.h similarity index 98% rename from paddle/fluid/operators/detail/request_handler.h rename to paddle/fluid/operators/distributed/request_handler.h index a2d08747d59220d30a5b8fd56074fd2739ae3bab..cf106656aa56c2130d8be8dbe7478c3397f9b9ad 100644 --- a/paddle/fluid/operators/detail/request_handler.h +++ b/paddle/fluid/operators/distributed/request_handler.h @@ -31,7 +31,7 @@ namespace paddle { namespace operators { -namespace detail { +namespace distributed { constexpr char kRequestSend[] = "RequestSend"; constexpr char kRequestGet[] = "RequestGet"; @@ -124,6 +124,6 @@ class RequestHandler { RPCServer* rpc_server_; }; -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/request_handler_impl.cc b/paddle/fluid/operators/distributed/request_handler_impl.cc similarity index 95% rename from paddle/fluid/operators/detail/request_handler_impl.cc rename to paddle/fluid/operators/distributed/request_handler_impl.cc index 7425bee798cd9ba0af8cd777a6db63862c8a4031..cb78c15c01e8e7f47ec759a75090f9a6b880b493 100644 --- a/paddle/fluid/operators/detail/request_handler_impl.cc +++ b/paddle/fluid/operators/distributed/request_handler_impl.cc @@ -20,12 +20,12 @@ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/operators/detail/request_handler_impl.h" -#include "paddle/fluid/operators/detail/rpc_server.h" +#include "paddle/fluid/operators/distributed/request_handler_impl.h" +#include "paddle/fluid/operators/distributed/rpc_server.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { bool RequestSendHandler::Handle(const std::string& varname, framework::Scope* scope, @@ -119,6 +119,6 @@ bool RequestPrefetchHandler::Handle(const std::string& varname, return true; } -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/request_handler_impl.h b/paddle/fluid/operators/distributed/request_handler_impl.h similarity index 95% rename from paddle/fluid/operators/detail/request_handler_impl.h rename to paddle/fluid/operators/distributed/request_handler_impl.h index 3f77c09a9598b431d747f1b824615e49d939098e..abbe8778911a21ece3090bc9790d51a3cb31b6d7 100644 --- a/paddle/fluid/operators/detail/request_handler_impl.h +++ b/paddle/fluid/operators/distributed/request_handler_impl.h @@ -28,11 +28,11 @@ #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/framework/var_type.h" -#include "paddle/fluid/operators/detail/request_handler.h" +#include "paddle/fluid/operators/distributed/request_handler.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { class RequestSendHandler final : public RequestHandler { public: @@ -66,6 +66,6 @@ class RequestPrefetchHandler final : public RequestHandler { const std::string& out_var_name = "") override; }; -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_client.cc b/paddle/fluid/operators/distributed/rpc_client.cc similarity index 88% rename from paddle/fluid/operators/detail/rpc_client.cc rename to paddle/fluid/operators/distributed/rpc_client.cc index 9a791403e3d6b99c5d4de5183e83e1af655d7d4c..c71edf977c18e554c502732e9bf4bb4ea99f8f99 100644 --- a/paddle/fluid/operators/detail/rpc_client.cc +++ b/paddle/fluid/operators/distributed/rpc_client.cc @@ -12,15 +12,15 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/operators/detail/rpc_client.h" +#include "paddle/fluid/operators/distributed/rpc_client.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { std::once_flag RPCClient::init_flag_; std::unique_ptr RPCClient::rpc_client_(nullptr); -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_client.h b/paddle/fluid/operators/distributed/rpc_client.h similarity index 98% rename from paddle/fluid/operators/detail/rpc_client.h rename to paddle/fluid/operators/distributed/rpc_client.h index 47c6ffb4fd7a002fc0bd8053fb3314a2fbf18fd3..72fa6d940886bc676e9d03d13f12d07772f5f5a7 100644 --- a/paddle/fluid/operators/detail/rpc_client.h +++ b/paddle/fluid/operators/distributed/rpc_client.h @@ -22,7 +22,7 @@ namespace paddle { namespace operators { -namespace detail { +namespace distributed { class RPCClient { public: @@ -84,6 +84,6 @@ class RPCClient { static std::once_flag init_flag_; static std::unique_ptr rpc_client_; }; -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_server.cc b/paddle/fluid/operators/distributed/rpc_server.cc similarity index 96% rename from paddle/fluid/operators/detail/rpc_server.cc rename to paddle/fluid/operators/distributed/rpc_server.cc index cd0fe96e2301ee3304fe9a2967df58b9f7072d8d..fa0cb71b3056de92f65139c5402132fc8cbb7a87 100644 --- a/paddle/fluid/operators/detail/rpc_server.cc +++ b/paddle/fluid/operators/distributed/rpc_server.cc @@ -17,11 +17,11 @@ #include #include -#include "paddle/fluid/operators/detail/rpc_server.h" +#include "paddle/fluid/operators/distributed/rpc_server.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { void RPCServer::ShutDown() { LOG(INFO) << "RPCServer ShutDown "; @@ -112,6 +112,6 @@ void RPCServer::WaitCond(const std::string& rpc_name) { lock, [=] { return (cur_cond_.load() == cond || exit_flag_.load()); }); } -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_server.h b/paddle/fluid/operators/distributed/rpc_server.h similarity index 95% rename from paddle/fluid/operators/detail/rpc_server.h rename to paddle/fluid/operators/distributed/rpc_server.h index 2e3342428cb56c34abaca655d5906668cda8f140..cf25e78435bb470b25a46db647ca818571cc83a5 100644 --- a/paddle/fluid/operators/detail/rpc_server.h +++ b/paddle/fluid/operators/distributed/rpc_server.h @@ -19,11 +19,11 @@ #include // NOLINT #include #include -#include "paddle/fluid/operators/detail/request_handler.h" +#include "paddle/fluid/operators/distributed/request_handler.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { class RPCServer { public: @@ -86,6 +86,6 @@ class RPCServer { friend class RequestHandler; }; -}; // namespace detail +}; // namespace distributed }; // namespace operators }; // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_server_test.cc b/paddle/fluid/operators/distributed/rpc_server_test.cc similarity index 87% rename from paddle/fluid/operators/detail/rpc_server_test.cc rename to paddle/fluid/operators/distributed/rpc_server_test.cc index 463a7b80cfac280de5afe91ee85caaaf074cef32..a0693cffabcc561b0adfafc2c49027a890dd5efc 100644 --- a/paddle/fluid/operators/detail/rpc_server_test.cc +++ b/paddle/fluid/operators/distributed/rpc_server_test.cc @@ -22,18 +22,18 @@ limitations under the License. */ #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/operators/detail/macros.h" -#include "paddle/fluid/operators/detail/request_handler_impl.h" -#include "paddle/fluid/operators/detail/rpc_client.h" -#include "paddle/fluid/operators/detail/rpc_server.h" +#include "paddle/fluid/operators/distributed/request_handler_impl.h" +#include "paddle/fluid/operators/distributed/rpc_client.h" +#include "paddle/fluid/operators/distributed/rpc_server.h" namespace framework = paddle::framework; namespace platform = paddle::platform; -namespace detail = paddle::operators::detail; +namespace distributed = paddle::operators::distributed; USE_OP(lookup_table); -std::unique_ptr g_rpc_service; -std::unique_ptr g_req_handler; +std::unique_ptr g_rpc_service; +std::unique_ptr g_req_handler; framework::BlockDesc* AppendPrefetchBlcok(framework::ProgramDesc* program) { auto root_block = program->MutableBlock(0); @@ -113,19 +113,21 @@ void StartServer() { g_req_handler->SetScope(&scope); g_req_handler->SetExecutor(&exe); - g_rpc_service->RegisterRPC(detail::kRequestPrefetch, g_req_handler.get()); + g_rpc_service->RegisterRPC(distributed::kRequestPrefetch, + g_req_handler.get()); g_req_handler->SetRPCServer(g_rpc_service.get()); std::thread server_thread( - std::bind(&detail::RPCServer::StartServer, g_rpc_service.get())); + std::bind(&distributed::RPCServer::StartServer, g_rpc_service.get())); server_thread.join(); } TEST(PREFETCH, CPU) { - g_req_handler.reset(new detail::RequestPrefetchHandler(true)); + g_req_handler.reset(new distributed::RequestPrefetchHandler(true)); g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 1)); - detail::RPCClient* client = detail::RPCClient::GetInstance(); + distributed::RPCClient* client = + distributed::RPCClient::GetInstance(); std::thread server_thread(StartServer); g_rpc_service->WaitServerReady(); diff --git a/paddle/fluid/operators/detail/send_recv.proto b/paddle/fluid/operators/distributed/send_recv.proto similarity index 100% rename from paddle/fluid/operators/detail/send_recv.proto rename to paddle/fluid/operators/distributed/send_recv.proto diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.cc b/paddle/fluid/operators/distributed/sendrecvop_utils.cc similarity index 95% rename from paddle/fluid/operators/detail/sendrecvop_utils.cc rename to paddle/fluid/operators/distributed/sendrecvop_utils.cc index 507b465435609a91ebca97dd70b176c3b79bee02..98129d9f1014c39347e3409533f2bc10092611d2 100644 --- a/paddle/fluid/operators/detail/sendrecvop_utils.cc +++ b/paddle/fluid/operators/distributed/sendrecvop_utils.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/detail/sendrecvop_utils.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" #ifdef PADDLE_WITH_CUDA #include @@ -23,14 +23,14 @@ limitations under the License. */ #include "google/protobuf/io/coded_stream.h" #include "google/protobuf/io/zero_copy_stream.h" #include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/operators/detail/bytebuffer_stream.h" -#include "paddle/fluid/operators/detail/proto_encoder_helper.h" -#include "paddle/fluid/operators/detail/variable_response.h" +#include "paddle/fluid/operators/distributed/bytebuffer_stream.h" +#include "paddle/fluid/operators/distributed/proto_encoder_helper.h" +#include "paddle/fluid/operators/distributed/variable_response.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { using VarMsg = sendrecv::VariableMessage; @@ -222,11 +222,11 @@ void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, const platform::DeviceContext& ctx, const framework::Scope* scope, framework::Variable** var) { - operators::detail::VariableResponse resp(scope, &ctx); + operators::distributed::VariableResponse resp(scope, &ctx); PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!"); *var = resp.GetVar(); } -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.h b/paddle/fluid/operators/distributed/sendrecvop_utils.h similarity index 92% rename from paddle/fluid/operators/detail/sendrecvop_utils.h rename to paddle/fluid/operators/distributed/sendrecvop_utils.h index bd16bf1dab8d933ffd18b6d6d9e3ce1c7d73029b..fe25e73fa608727ba0bb912a82776b330ec8d83a 100644 --- a/paddle/fluid/operators/detail/sendrecvop_utils.h +++ b/paddle/fluid/operators/distributed/sendrecvop_utils.h @@ -25,12 +25,12 @@ limitations under the License. */ #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/framework/var_type.h" -#include "paddle/fluid/operators/detail/send_recv.grpc.pb.h" -#include "paddle/fluid/operators/detail/send_recv.pb.h" +#include "paddle/fluid/operators/distributed/send_recv.grpc.pb.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { typedef void (*DestroyCallback)(void*); @@ -61,6 +61,6 @@ inline std::type_index ToTypeIndex(sendrecv::VariableMessage::Type type) { } } -} // namespace detail +} // namespace distributed } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/variable_response.cc b/paddle/fluid/operators/distributed/variable_response.cc similarity index 94% rename from paddle/fluid/operators/detail/variable_response.cc rename to paddle/fluid/operators/distributed/variable_response.cc index 24cb91a3bb820a0e5d51aaa49154434919080f69..45832c60bf9172497afabac927ba39a7cbfb9a52 100644 --- a/paddle/fluid/operators/detail/variable_response.cc +++ b/paddle/fluid/operators/distributed/variable_response.cc @@ -12,7 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/operators/detail/variable_response.h" +#include "paddle/fluid/operators/distributed/variable_response.h" #include #include @@ -22,12 +22,12 @@ #endif #include "paddle/fluid/platform/profiler.h" -#include "paddle/fluid/operators/detail/send_recv.pb.h" -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" +#include "paddle/fluid/operators/distributed/sendrecvop_utils.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { enum WireType { WIRETYPE_VARINT = 0, @@ -76,6 +76,8 @@ bool ReadRaw(::google::protobuf::io::CodedInputStream* input, if (total_written + size_to_write > length) { size_to_write = length - total_written; } + // This log is useful to see how long a internal block size is of rpc. + VLOG(7) << "copy " << size_to_write << " data to CUDAPlace"; memory::Copy(boost::get(place), reinterpret_cast(p), cpu, data, size_to_write, gpu_dev_ctx.stream()); @@ -103,6 +105,8 @@ bool ReadRaw(::google::protobuf::io::CodedInputStream* input, } // TODO(gongwb): can we avoid copy? platform::CPUPlace cpu; + // This log is useful to see how long a internal block size is of rpc. + VLOG(7) << "copy " << size_to_write << " data to CPUPlace"; memory::Copy(cpu, reinterpret_cast(p), cpu, data, size_to_write); p += size_to_write; @@ -158,13 +162,13 @@ bool VariableResponse::CopySelectRowsTensorData( slr->set_height(meta_.slr_height()); auto* tensor = slr->mutable_value(); tensor->Resize(dims); - PADDLE_ENFORCE_EQ( - static_cast(tensor->numel()), - length / framework::SizeOfType( - paddle::operators::detail::ToTypeIndex(meta_.data_type()))); + PADDLE_ENFORCE_EQ(static_cast(tensor->numel()), + length / framework::SizeOfType( + paddle::operators::distributed::ToTypeIndex( + meta_.data_type()))); void* tensor_data = tensor->mutable_data( ctx.GetPlace(), - paddle::operators::detail::ToTypeIndex(meta_.data_type())); + paddle::operators::distributed::ToTypeIndex(meta_.data_type())); if (!ReadRaw(input, ctx, tensor->place(), tensor_data, length)) { return false; @@ -480,6 +484,6 @@ int VariableResponse::Parse(Source* source) { return 0; } -}; // namespace detail +}; // namespace distributed }; // namespace operators }; // namespace paddle diff --git a/paddle/fluid/operators/detail/variable_response.h b/paddle/fluid/operators/distributed/variable_response.h similarity index 92% rename from paddle/fluid/operators/detail/variable_response.h rename to paddle/fluid/operators/distributed/variable_response.h index 69cfd784f8dd4f129f50c6882061e53e8535b949..1db4a0a522654ff2497b8bd9ee1381b5ab64067a 100644 --- a/paddle/fluid/operators/detail/variable_response.h +++ b/paddle/fluid/operators/distributed/variable_response.h @@ -22,17 +22,17 @@ #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/framework/var_type.h" -#include "paddle/fluid/operators/detail/send_recv.grpc.pb.h" -#include "paddle/fluid/operators/detail/send_recv.pb.h" +#include "paddle/fluid/operators/distributed/send_recv.grpc.pb.h" +#include "paddle/fluid/operators/distributed/send_recv.pb.h" #include "google/protobuf/io/coded_stream.h" #include "google/protobuf/io/zero_copy_stream.h" #include "paddle/fluid/framework/tensor.h" -#include "paddle/fluid/operators/detail/bytebuffer_stream.h" +#include "paddle/fluid/operators/distributed/bytebuffer_stream.h" namespace paddle { namespace operators { -namespace detail { +namespace distributed { class VariableResponse { public: @@ -99,6 +99,6 @@ class VariableResponse { sendrecv::VariableMessage meta_; }; -}; // namespace detail +}; // namespace distributed }; // namespace operators }; // namespace paddle diff --git a/paddle/fluid/operators/fetch_barrier_op.cc b/paddle/fluid/operators/fetch_barrier_op.cc index 98b051afb551f373009d2bd3df1a8daa64b7e6c7..02beb80fc8a9f451393dcdd54492c4f88f908497 100644 --- a/paddle/fluid/operators/fetch_barrier_op.cc +++ b/paddle/fluid/operators/fetch_barrier_op.cc @@ -42,8 +42,8 @@ class FetchBarrierOp : public framework::OperatorBase { // For profiling platform::RecordEvent record_event(Type(), &ctx); - detail::RPCClient* rpc_client = - detail::RPCClient::GetInstance(); + distributed::RPCClient* rpc_client = + distributed::RPCClient::GetInstance(); rpc_client->Wait(); diff --git a/paddle/fluid/operators/gaussian_random_mkldnn_op.cc b/paddle/fluid/operators/gaussian_random_mkldnn_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..76b00b396c1349eff5db1059268e7cf280a8fc64 --- /dev/null +++ b/paddle/fluid/operators/gaussian_random_mkldnn_op.cc @@ -0,0 +1,55 @@ +/* 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 "paddle/fluid/operators/mean_op.h" + +namespace paddle { +namespace operators { + +using framework::DataLayout; +template +class GaussianMKLDNNKernel : public paddle::framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + float mean = context.Attr("mean"); + float std = context.Attr("std"); + auto* tensor = context.Output("Out"); + T* data = tensor->mutable_data(context.GetPlace()); + + unsigned int seed = static_cast(context.Attr("seed")); + std::minstd_rand engine; + if (seed == 0) { + seed = std::random_device()(); + } + engine.seed(seed); + std::normal_distribution dist(mean, std); + int64_t size = tensor->numel(); + for (int64_t i = 0; i < size; ++i) { + data[i] = dist(engine); + } + + // The format of output is set as the mkldnn's format + // TODO(@mozga-intel) The format of matrix sets inside the another layers. + tensor->set_layout(DataLayout::kMKLDNN); + tensor->set_format(mkldnn::memory::format::oihw); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_KERNEL(gaussian_random, MKLDNN, ::paddle::platform::CPUPlace, + ops::GaussianMKLDNNKernel); diff --git a/paddle/fluid/operators/gaussian_random_op.cc b/paddle/fluid/operators/gaussian_random_op.cc index 815c1bb50988be49ca9996e368a59344c6583d58..1488aab1926b5b4ba7bceed582700f5a11fc6c93 100644 --- a/paddle/fluid/operators/gaussian_random_op.cc +++ b/paddle/fluid/operators/gaussian_random_op.cc @@ -15,6 +15,10 @@ limitations under the License. */ #include #include "paddle/fluid/framework/op_registry.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif + namespace paddle { namespace operators { @@ -62,9 +66,20 @@ class GaussianRandomOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { + framework::LibraryType library{framework::LibraryType::kPlain}; + framework::DataLayout layout{framework::DataLayout::kAnyLayout}; + +#ifdef PADDLE_WITH_MKLDNN + if (library == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; + } +#endif + return framework::OpKernelType( static_cast(ctx.Attr("dtype")), - ctx.device_context()); + ctx.device_context(), layout, library); } }; @@ -95,7 +110,9 @@ class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker { "(int, default 5(FP32)) " "Output data type.") .SetDefault(framework::proto::VarType::FP32); - + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddComment(R"DOC( GaussianRandom Operator. diff --git a/paddle/fluid/operators/gen_nccl_id_op.cc b/paddle/fluid/operators/gen_nccl_id_op.cc index f824eee4e7d1ef19c9a38fd5d3369265f9c549a0..697c239e59d158428ae9ba9f7feded19637dff28 100644 --- a/paddle/fluid/operators/gen_nccl_id_op.cc +++ b/paddle/fluid/operators/gen_nccl_id_op.cc @@ -22,7 +22,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/threadpool.h" #include "paddle/fluid/operators/detail/macros.h" -#include "paddle/fluid/operators/detail/request_handler_impl.h" +#include "paddle/fluid/operators/distributed/request_handler_impl.h" #include "paddle/fluid/platform/nccl_helper.h" namespace paddle { @@ -60,7 +60,8 @@ class GenNCCLIdOp : public framework::OperatorBase { std::vector endpoint_list = Attr>("endpoint_list"); - detail::RPCClient* client = detail::RPCClient::GetInstance(); + distributed::RPCClient* client = + distributed::RPCClient::GetInstance(); for (auto& ep : endpoint_list) { VLOG(3) << "sending nccl id to " << ep; @@ -80,11 +81,11 @@ class GenNCCLIdOp : public framework::OperatorBase { // 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::RequestSendHandler rpc_h(true); - std::unique_ptr rpc_service( + distributed::RequestSendHandler rpc_h(true); + std::unique_ptr rpc_service( new RPCSERVER_T(endpoint, 1)); - rpc_service->RegisterRPC(detail::kRequestSend, &rpc_h); + rpc_service->RegisterRPC(distributed::kRequestSend, &rpc_h); rpc_h.SetRPCServer(rpc_service.get()); framework::ProgramDesc empty_program; @@ -95,11 +96,11 @@ class GenNCCLIdOp : public framework::OperatorBase { rpc_h.SetExecutor(&executor); std::thread server_thread( - std::bind(&detail::RPCServer::StartServer, rpc_service.get())); + std::bind(&distributed::RPCServer::StartServer, rpc_service.get())); - rpc_service->SetCond(detail::kRequestSend); + rpc_service->SetCond(distributed::kRequestSend); VLOG(3) << "start getting nccl id from trainer 0..."; - rpc_service->WaitBarrier(detail::kRequestSend); + rpc_service->WaitBarrier(distributed::kRequestSend); VLOG(3) << "got nccl id and stop server..."; rpc_service->ShutDown(); VLOG(3) << "rpc server stopped"; diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index 57c2ce457791d830e4230aa25e1c5b358f476782..d98bf807a9464c1c2294aa0601386a940ddc00f8 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -21,14 +21,14 @@ limitations under the License. */ #include "paddle/fluid/operators/detail/macros.h" -#include "paddle/fluid/operators/detail/request_handler_impl.h" +#include "paddle/fluid/operators/distributed/request_handler_impl.h" #include "paddle/fluid/operators/listen_and_serv_op.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { -void RunServer(std::shared_ptr service) { +void RunServer(std::shared_ptr service) { service->StartServer(); VLOG(4) << "RunServer thread end"; } @@ -101,17 +101,16 @@ void ListenAndServOp::RunSyncLoop( framework::Scope *recv_scope, const std::vector &prefetch_block_id_list) const { size_t num_blocks = program->Size(); + auto optimize_blocks = + Attr>(kOptimizeBlocks); PADDLE_ENFORCE_GE(num_blocks, 2, "server program should have at least 2 blocks"); - std::vector optimize_block_id_list; - for (int blkid = 1; blkid < num_blocks; ++blkid) { - if (std::find(prefetch_block_id_list.begin(), prefetch_block_id_list.end(), - blkid) == prefetch_block_id_list.end()) { - optimize_block_id_list.push_back(blkid); - } + std::vector optimize_blocks_idx; + for (auto blk : optimize_blocks) { + optimize_blocks_idx.push_back(blk->ID()); } - auto optimize_prepared = executor->Prepare(*program, optimize_block_id_list); + auto optimize_prepared = executor->Prepare(*program, optimize_blocks_idx); // Insert placeholder for block0 which holds current op itself. optimize_prepared.insert( optimize_prepared.begin(), @@ -121,12 +120,12 @@ void ListenAndServOp::RunSyncLoop( while (true) { // Get from multiple trainers, we don't care about the order in which // the gradients arrives, just add suffix 0~n and merge the gradient. - rpc_service_->SetCond(detail::kRequestSend); - rpc_service_->WaitBarrier(detail::kRequestSend); + rpc_service_->SetCond(distributed::kRequestSend); + rpc_service_->WaitBarrier(distributed::kRequestSend); if (rpc_service_->IsExit()) { LOG(WARNING) << "get exit!rpc_processor break!"; - rpc_service_->SetCond(detail::kRequestGet); + rpc_service_->SetCond(distributed::kRequestGet); break; } @@ -134,14 +133,14 @@ void ListenAndServOp::RunSyncLoop( // and this will still work. // The optimize blocks which have the same parent ID would run parallel // TODO(Yancey1989): need to use ParallelExecutor for future - int32_t last_parent_blkid = program->Block(1).Parent(); + int32_t last_parent_blkid = optimize_blocks[0]->Parent(); std::vector parallel_blkids; - parallel_blkids.push_back(1); + parallel_blkids.push_back(optimize_blocks[0]->ID()); double ts = GetTimestamp(); - for (size_t i = 1; i < optimize_block_id_list.size(); ++i) { + for (size_t i = 1; i < optimize_blocks.size(); ++i) { // skip the first optimize block because it is already in the // parallel_blkids. - int blkid = optimize_block_id_list[i]; + int blkid = optimize_blocks[i]->ID(); if (program->Block(blkid).Parent() != last_parent_blkid) { ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared, program, recv_scope); @@ -154,11 +153,11 @@ void ListenAndServOp::RunSyncLoop( recv_scope); VLOG(2) << "run all blocks spent " << GetTimestamp() - ts << "(ms)"; - rpc_service_->SetCond(detail::kRequestGet); - rpc_service_->WaitBarrier(detail::kRequestGet); + rpc_service_->SetCond(distributed::kRequestGet); + rpc_service_->WaitBarrier(distributed::kRequestGet); rpc_service_->ResetBarrierCounter(); // reset received sparse vars to avoid reuse it in the next mini-batch - dynamic_cast(request_send_handler_.get()) + dynamic_cast(request_send_handler_.get()) ->ResetSparseVarRecorder(); } // while(true) } @@ -215,13 +214,13 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, } static void FillRequestCtx( - detail::RequestHandler *h, framework::Scope *scope, + distributed::RequestHandler *h, framework::Scope *scope, platform::DeviceContext *dev_ctx, framework::Executor *executor, framework::ProgramDesc *program, std::unordered_map> *prefetch_ctx, - detail::RPCServer *rpc_server) { + distributed::RPCServer *rpc_server) { h->SetScope(scope); h->SetDevCtx(dev_ctx); h->SetExecutor(executor); @@ -249,18 +248,23 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, rpc_service_.reset(new RPCSERVER_T(endpoint, fan_in)); - request_send_handler_.reset(new detail::RequestSendHandler(sync_mode)); - request_get_handler_.reset(new detail::RequestGetHandler(sync_mode)); + request_send_handler_.reset(new distributed::RequestSendHandler(sync_mode)); + request_get_handler_.reset(new distributed::RequestGetHandler(sync_mode)); request_prefetch_handler_.reset( - new detail::RequestPrefetchHandler(sync_mode)); + new distributed::RequestPrefetchHandler(sync_mode)); - rpc_service_->RegisterRPC(detail::kRequestSend, request_send_handler_.get()); - rpc_service_->RegisterRPC(detail::kRequestGet, request_get_handler_.get()); - rpc_service_->RegisterRPC(detail::kRequestPrefetch, + rpc_service_->RegisterRPC(distributed::kRequestSend, + request_send_handler_.get()); + rpc_service_->RegisterRPC(distributed::kRequestGet, + request_get_handler_.get()); + rpc_service_->RegisterRPC(distributed::kRequestPrefetch, request_prefetch_handler_.get()); - auto *optimize_block = Attr(kOptimizeBlock); - auto *program = optimize_block->Program(); + auto optimize_blocks = + Attr>(kOptimizeBlocks); + PADDLE_ENFORCE(optimize_blocks.size() >= 1, + "optimize blocks should be 1 at least on the pserver side."); + auto *program = optimize_blocks[0]->Program(); framework::Executor executor(dev_place); // prepare for prefetch @@ -337,8 +341,9 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker { "a map from grad name to it's optimize block id") .SetDefault({}); AddAttr("sync_mode", "if works at sync_mode or not").SetDefault(true); - AddAttr(kOptimizeBlock, - "BlockID to run on server side."); + AddAttr>( + kOptimizeBlocks, "Optimize blocks to run on server side.") + .SetDefault({}); AddAttr>(kPrefetchVarNameToBlockId, "prefetch blocks to run on server side.") .SetDefault({}); diff --git a/paddle/fluid/operators/listen_and_serv_op.h b/paddle/fluid/operators/listen_and_serv_op.h index 46c3a19e20b3f2dd970a672bb99f98e83d3e25bf..634c1b4f4b541be9f4950a9ef48f944863486705 100644 --- a/paddle/fluid/operators/listen_and_serv_op.h +++ b/paddle/fluid/operators/listen_and_serv_op.h @@ -24,16 +24,16 @@ limitations under the License. */ #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/request_handler.h" -#include "paddle/fluid/operators/detail/rpc_server.h" +#include "paddle/fluid/operators/distributed/request_handler.h" +#include "paddle/fluid/operators/distributed/rpc_server.h" namespace paddle { namespace operators { -constexpr char kOptimizeBlock[] = "OptimizeBlock"; +constexpr char kOptimizeBlocks[] = "optimize_blocks"; constexpr char kPrefetchVarNameToBlockId[] = "prefetch_var_name_to_block_id"; -void RunServer(std::shared_ptr service); +void RunServer(std::shared_ptr service); class ListenAndServOp : public framework::OperatorBase { public: @@ -62,10 +62,11 @@ class ListenAndServOp : public framework::OperatorBase { const platform::Place& dev_place) const override; protected: - mutable std::shared_ptr rpc_service_; - mutable std::shared_ptr request_send_handler_; - mutable std::shared_ptr request_get_handler_; - mutable std::shared_ptr request_prefetch_handler_; + mutable std::shared_ptr rpc_service_; + mutable std::shared_ptr request_send_handler_; + mutable std::shared_ptr request_get_handler_; + mutable std::shared_ptr + request_prefetch_handler_; mutable std::shared_ptr server_thread_; }; diff --git a/paddle/fluid/operators/logical_op.cc b/paddle/fluid/operators/logical_op.cc index db109f5cd053d84718ac85bd4693ecece12ce172..26970db8d2af62bb06fce4eb1a1f21fd41617bd1 100644 --- a/paddle/fluid/operators/logical_op.cc +++ b/paddle/fluid/operators/logical_op.cc @@ -146,6 +146,6 @@ REGISTER_UNARY_LOGICAL_OP(logical_not, "$$Out = !X$$"); REGISTER_UNARY_LOGICAL_KERNEL(logical_not, CPU, paddle::operators::LogicalNotFunctor); REGISTER_BINARY_LOGICAL_OP(logical_xor, - "$$Out = (X || Y) \\, \\&\\& \\, !(X \\&\\& Y)$$"); + "$$Out = (X || Y) \\&\\& !(X \\&\\& Y)$$"); REGISTER_BINARY_LOGICAL_KERNEL(logical_xor, CPU, paddle::operators::LogicalXorFunctor); diff --git a/paddle/fluid/operators/math/concat.cc b/paddle/fluid/operators/math/concat.cc index cc69212466b72f3fa82e8f5f58b4f3229dab28ec..55c8a472aca7fe700ef6a3f96bed1496d7b12b80 100644 --- a/paddle/fluid/operators/math/concat.cc +++ b/paddle/fluid/operators/math/concat.cc @@ -70,21 +70,23 @@ template class ConcatGradFunctor { public: void operator()(const platform::CPUDeviceContext& context, - const framework::Tensor& input, const int axis, - std::vector* outputs) { + const framework::Tensor& input, + const std::vector& ref_inputs, + const int axis, std::vector* outputs) { // TODO(zcd): Add input data validity checking - int num = outputs->size(); + size_t num = outputs->size(); int input_rows = 1; - auto dim_0 = outputs->at(0).dims(); + auto dim_0 = ref_inputs[0]->dims(); for (int i = 0; i < axis; ++i) { input_rows *= dim_0[i]; } + int input_cols = 0; std::vector output_cols(outputs->size()); - for (int i = 0; i < num; ++i) { - int t_cols = outputs->at(i).numel() / input_rows; + for (size_t i = 0; i < num; ++i) { + int t_cols = ref_inputs[i]->numel() / input_rows; input_cols += t_cols; output_cols[i] = t_cols; } @@ -94,11 +96,14 @@ class ConcatGradFunctor { for (int k = 0; k < input_rows; ++k) { const T* src_ptr = input.data() + k * input_cols; int col_idx = 0; - for (int j = 0; j < num; ++j) { + for (size_t j = 0; j < num; ++j) { int col_len = output_cols[j]; - T* dst_ptr = outputs->at(j).data() + k * col_len; - memory::Copy(cpu_place, dst_ptr, cpu_place, src_ptr + col_idx, - sizeof(T) * col_len); + auto* out_tensor = outputs->at(j); + if (out_tensor != nullptr) { + T* dst_ptr = out_tensor->data() + k * col_len; + memory::Copy(cpu_place, dst_ptr, cpu_place, src_ptr + col_idx, + sizeof(T) * col_len); + } col_idx += col_len; } } diff --git a/paddle/fluid/operators/math/concat.cu b/paddle/fluid/operators/math/concat.cu index 4285d38dcd6a4124543cdd2246c82a8203f5a281..5863d74fca21de8b77bc208fb95d8fd52562f7a7 100644 --- a/paddle/fluid/operators/math/concat.cu +++ b/paddle/fluid/operators/math/concat.cu @@ -22,43 +22,24 @@ namespace paddle { namespace operators { namespace math { -template -__device__ T upper_bound(const T* first, T count, T val) { - const T* orig = first; - const T* it = nullptr; - T step = 0; - while (count > 0) { - it = first; - step = count / 2; - it += step; - if (!(val < *it)) { - first = ++it; - count -= step + 1; - } else { - count = step; - } - } - return first - orig; -} - template __global__ void KernelConcat(T** inputs, const int* input_cols, int col_size, const int output_rows, const int output_cols, T* output) { int tid_x = blockIdx.x * blockDim.x + threadIdx.x; - int segment = upper_bound(input_cols, col_size, tid_x) - 1; - - int curr_offset = input_cols[segment]; - int curr_segment = segment; + int curr_segment = 0; + int curr_offset = input_cols[0]; for (; tid_x < output_cols; tid_x += blockDim.x * gridDim.x) { - T curr_col_offset; - while ((curr_col_offset = input_cols[curr_segment + 1]) <= tid_x) { + int curr_col_offset = input_cols[curr_segment + 1]; + while (curr_col_offset <= tid_x) { curr_offset = curr_col_offset; ++curr_segment; + curr_col_offset = input_cols[curr_segment + 1]; } int local_col = tid_x - curr_offset; int segment_width = curr_col_offset - curr_offset; + T* input_ptr = inputs[curr_segment]; int tid_y = blockIdx.y * blockDim.y + threadIdx.y; for (; tid_y < output_rows; tid_y += blockDim.y * gridDim.y) @@ -89,23 +70,25 @@ __global__ void KernelConcatGrad(const T* input_data, const int in_row, const int in_col, const int* out_cols, int out_cols_size, T** outputs_data) { int tid_x = blockIdx.x * blockDim.x + threadIdx.x; - int segment = upper_bound(out_cols, out_cols_size, tid_x) - 1; - int curr_offset = out_cols[segment]; - int curr_segment = segment; + int curr_segment = 0; + int curr_offset = out_cols[0]; for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) { - T curr_col_offset; - while ((curr_col_offset = out_cols[curr_segment + 1]) <= tid_x) { + int curr_col_offset = out_cols[curr_segment + 1]; + while (curr_col_offset <= tid_x) { curr_offset = curr_col_offset; ++curr_segment; + curr_col_offset = out_cols[curr_segment + 1]; } int local_col = tid_x - curr_offset; int segment_width = curr_col_offset - curr_offset; T* output_ptr = outputs_data[curr_segment]; - int tid_y = blockIdx.y * blockDim.y + threadIdx.y; - for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y) - output_ptr[tid_y * segment_width + local_col] = - input_data[tid_y * in_col + tid_x]; + if (output_ptr != nullptr) { + int tid_y = blockIdx.y * blockDim.y + threadIdx.y; + for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y) + output_ptr[tid_y * segment_width + local_col] = + input_data[tid_y * in_col + tid_x]; + } } } @@ -118,10 +101,12 @@ __global__ void KernelConcatGrad(const T* input_data, const int in_row, int split = tid_x / fixed_out_col; int in_offset = tid_x - split * fixed_out_col; T* output_ptr = outputs_data[split]; - int tid_y = blockIdx.y * blockDim.y + threadIdx.y; - for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y) - output_ptr[tid_y * fixed_out_col + in_offset] = - input_data[tid_y * in_col + tid_x]; + if (output_ptr != nullptr) { + int tid_y = blockIdx.y * blockDim.y + threadIdx.y; + for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y) + output_ptr[tid_y * fixed_out_col + in_offset] = + input_data[tid_y * in_col + tid_x]; + } } } @@ -203,17 +188,18 @@ template class ConcatGradFunctor { public: void operator()(const platform::CUDADeviceContext& context, - const framework::Tensor& input, const int axis, - std::vector* outputs) { + const framework::Tensor& input, + const std::vector& ref_inputs, + const int axis, std::vector* outputs) { // TODO(zcd): Add input data validity checking int o_num = outputs->size(); int out_row = 1; - auto dim_0 = outputs->at(0).dims(); + auto dim_0 = ref_inputs[0]->dims(); for (int i = 0; i < axis; ++i) { out_row *= dim_0[i]; } - int out_col = outputs->at(0).numel() / out_row; + int out0_col = ref_inputs[0]->numel() / out_row; int in_col = 0, in_row = out_row; bool sameShape = true; @@ -223,13 +209,17 @@ class ConcatGradFunctor { outputs_cols[0] = 0; for (int i = 0; i < o_num; ++i) { - int t_col = outputs->at(i).numel() / out_row; + int t_col = ref_inputs.at(i)->numel() / out_row; if (sameShape) { - if (t_col != out_col) sameShape = false; + if (t_col != out0_col) sameShape = false; } in_col += t_col; outputs_cols[i + 1] = in_col; - outputs_ptr[i] = outputs->at(i).data(); + if (outputs->at(i) != nullptr) { + outputs_ptr[i] = outputs->at(i)->data(); + } else { + outputs_ptr[i] = nullptr; + } } T** dev_out_gpu_data = @@ -255,7 +245,7 @@ class ConcatGradFunctor { if (sameShape) { KernelConcatGrad<<>>( - input.data(), in_row, in_col, out_col, dev_out_gpu_data); + input.data(), in_row, in_col, out0_col, dev_out_gpu_data); } else { const int* dev_outs_col_data = outputs_cols.CUDAData(context.GetPlace()); KernelConcatGrad<<>>( diff --git a/paddle/fluid/operators/math/concat.h b/paddle/fluid/operators/math/concat.h index 041ce8bf8a2e9528a004c076ead4471a3837c1a6..9e080f2e8be23768dcea47b577043beef37b2eaf 100644 --- a/paddle/fluid/operators/math/concat.h +++ b/paddle/fluid/operators/math/concat.h @@ -57,7 +57,8 @@ template class ConcatGradFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& input, - const int axis, std::vector* outputs); + const std::vector& ref_inputs, + const int axis, std::vector* outputs); }; } // namespace math diff --git a/paddle/fluid/operators/math/math_function.cc b/paddle/fluid/operators/math/math_function.cc index d39154c6f88d6d17c1719eb9a5b048211f4bb52b..c3387be6daa3bd34a6e3410ced23fce5d65f2cf7 100644 --- a/paddle/fluid/operators/math/math_function.cc +++ b/paddle/fluid/operators/math/math_function.cc @@ -30,6 +30,7 @@ template struct SetConstant; template struct SetConstant; template struct SetConstant; template struct SetConstant; +template struct SetConstant; #define DEFINE_CPU_TRANS(RANK) \ template struct TransposeDebugStringEx(sub_scopes[0]); sum_op->Run(*sub_scopes[0], places[0]); WaitOnPlace(places[0]); diff --git a/paddle/fluid/operators/prefetch_op.cc b/paddle/fluid/operators/prefetch_op.cc index f71ba84b318c1f8b0604310f3db8a0826124e207..8734282fe496b8e90af19abd5549566d62316fc3 100644 --- a/paddle/fluid/operators/prefetch_op.cc +++ b/paddle/fluid/operators/prefetch_op.cc @@ -41,8 +41,8 @@ class PrefetchOp : public framework::OperatorBase { platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); - detail::RPCClient* rpc_client = - detail::RPCClient::GetInstance(); + distributed::RPCClient* rpc_client = + distributed::RPCClient::GetInstance(); for (size_t i = 0; i < ins.size(); i++) { if (NeedSend(scope, ins[i])) { diff --git a/paddle/fluid/operators/recurrent_op.cc b/paddle/fluid/operators/recurrent_op.cc index 9c1cee7022a9b9a98f026f7602f0f7badc44a49b..162bfcbb0844d29385d0f8ad5d25a3f8de6bd41b 100644 --- a/paddle/fluid/operators/recurrent_op.cc +++ b/paddle/fluid/operators/recurrent_op.cc @@ -429,7 +429,8 @@ class RecurrentGradOp : public RecurrentBase { auto sum_op = framework::OpRegistry::CreateOp( "sum", {{"X", {pg_names[param_id], new_inside_name}}}, - {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{}); + {{"Out", {pg_names[param_id]}}}, + framework::AttributeMap{{"use_mkldnn", {false}}}); sum_op->Run(cur_scope, place); cur_scope.Rename(new_inside_name, inside_grad_name); diff --git a/paddle/fluid/operators/recv_op.cc b/paddle/fluid/operators/recv_op.cc index 15dfb5469bf51330b98d6699fb3ce708222212ed..9854a31f5b10f5ecd940c0d41c2c3e468fc17bad 100644 --- a/paddle/fluid/operators/recv_op.cc +++ b/paddle/fluid/operators/recv_op.cc @@ -43,8 +43,8 @@ class RecvOp : public framework::OperatorBase { // For profiling platform::RecordEvent record_event(Type(), &ctx); - detail::RPCClient* rpc_client = - detail::RPCClient::GetInstance(); + distributed::RPCClient* rpc_client = + distributed::RPCClient::GetInstance(); for (size_t i = 0; i < outs.size(); i++) { VLOG(3) << "getting " << outs[i] << " from " << epmap[i]; diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc index c6c975a23ce846464388c72af5d8902144ceb16a..6b4572dcccc21e783f1df0b9bcde11d532ff4ba8 100644 --- a/paddle/fluid/operators/send_barrier_op.cc +++ b/paddle/fluid/operators/send_barrier_op.cc @@ -44,8 +44,8 @@ class SendBarrierOp : public framework::OperatorBase { // For profiling platform::RecordEvent record_event(Type(), &ctx); - detail::RPCClient* rpc_client = - detail::RPCClient::GetInstance(); + distributed::RPCClient* rpc_client = + distributed::RPCClient::GetInstance(); VLOG(3) << "SendBarrierOp sync_mode:" << sync_mode; diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc index 84ec36625314572d16e5c537884b6efec420cc60..0cac329aafa8c4c67cae48ba62a48575f5edba92 100644 --- a/paddle/fluid/operators/send_op.cc +++ b/paddle/fluid/operators/send_op.cc @@ -45,8 +45,8 @@ class SendOp : public framework::OperatorBase { // For profiling platform::RecordEvent record_event(Type(), &ctx); - detail::RPCClient* rpc_client = - detail::RPCClient::GetInstance(); + distributed::RPCClient* rpc_client = + distributed::RPCClient::GetInstance(); for (size_t i = 0; i < ins.size(); i++) { if (NeedSend(scope, ins[i])) { diff --git a/paddle/fluid/operators/send_recv_op_test.cc b/paddle/fluid/operators/send_recv_op_test.cc index e550552b195b768d68ec64e9c3b5889b56ca719f..aee6180add5708d31f7ce927b37c4524a291fe3c 100644 --- a/paddle/fluid/operators/send_recv_op_test.cc +++ b/paddle/fluid/operators/send_recv_op_test.cc @@ -129,7 +129,10 @@ void StartServerNet(bool is_sparse, std::atomic *initialized) { // sub program run in listen_and_serv_op, for simple test we use sum f::ProgramDesc program; const auto &root_block = program.Block(0); + std::vector optimize_blocks; auto *optimize_block = program.AppendBlock(root_block); + optimize_blocks.push_back(optimize_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, @@ -139,7 +142,7 @@ void StartServerNet(bool is_sparse, std::atomic *initialized) { attrs.insert({"Fanin", 1}); attrs.insert({"ParamList", std::vector({"Out"})}); attrs.insert({"GradList", std::vector({"x1"})}); - attrs.insert({"OptimizeBlock", optimize_block}); + attrs.insert({"optimize_blocks", optimize_blocks}); attrs.insert({"PrefetchBlock", prefetch_block}); attrs.insert({"grad_to_block_id", std::vector({""})}); attrs.insert({"sync_mode", true}); diff --git a/paddle/fluid/operators/softmax_mkldnn_op.cc b/paddle/fluid/operators/softmax_mkldnn_op.cc index 14b57b11fefb2b726531cb164dbf479f8df26b24..6668e6b9e917eea7ba4a80ac78917b73eb827208 100644 --- a/paddle/fluid/operators/softmax_mkldnn_op.cc +++ b/paddle/fluid/operators/softmax_mkldnn_op.cc @@ -27,8 +27,81 @@ using paddle::platform::MKLDNNMemDesc; using mkldnn::memory; // Note: paddle has also "memory" namespace using mkldnn::primitive; using mkldnn::softmax_forward; +using mkldnn::softmax_backward; using mkldnn::prop_kind; using mkldnn::stream; +using platform::to_void_cast; + +class SoftmaxMKLDNNHandler : public platform::MKLDNNHandler { + public: + SoftmaxMKLDNNHandler( + std::shared_ptr softmax_pd, + const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine, + const std::string& base_key) + : platform::MKLDNNHandler(dev_ctx, engine, base_key), + softmax_pd_(softmax_pd) {} + + SoftmaxMKLDNNHandler( + std::shared_ptr softmax_pd, + std::shared_ptr softmax_bwd_pd, + const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine, + const std::string& base_key) + : platform::MKLDNNHandler(dev_ctx, engine, base_key), + softmax_pd_(softmax_pd), + softmax_bwd_pd_(softmax_bwd_pd) { + // If we are in Grad operatgor then update a key with BWD suffix to + // distinguish from FWD memory primitives + key_ += "-BWD"; + } + + std::shared_ptr AcquireSoftmax( + std::shared_ptr dst_memory_p, + std::shared_ptr src_memory_p) { + /*Generate key*/ + auto prim_key = key_ + "@softmax_p"; + + auto softmax_p = std::static_pointer_cast( + dev_ctx_.GetBlob(prim_key)); + PADDLE_ENFORCE((softmax_p != nullptr) || (is_reusing_ == false), + "Fail to find softmax primitive in device context"); + if (softmax_p == nullptr) { + softmax_p = std::make_shared( + *(softmax_pd_.get()), + *(static_cast(src_memory_p.get())), + *(static_cast(dst_memory_p.get()))); + dev_ctx_.SetBlob(prim_key, softmax_p); + } else { + is_reusing_ = true; + } + + return softmax_p; + } + + std::shared_ptr AcquireSoftmaxBackward( + std::shared_ptr dst_memory_p, + std::shared_ptr diff_dst_memory_p, + std::shared_ptr diff_src_memory_p) { + auto prim_key = key_ + "@softmax_bwd_p"; + auto softmax_bwd_p = std::static_pointer_cast( + dev_ctx_.GetBlob(prim_key)); + PADDLE_ENFORCE((softmax_bwd_p != nullptr) || (is_reusing_ == false), + "Fail to find softmax backward primitive in device context"); + if (softmax_bwd_p == nullptr) { + softmax_bwd_p = std::make_shared( + *softmax_bwd_pd_, *(dst_memory_p.get()), *(diff_dst_memory_p.get()), + *(diff_src_memory_p.get())); + dev_ctx_.SetBlob(prim_key, softmax_bwd_p); + } else { + is_reusing_ = true; + } + + return softmax_bwd_p; + } + + private: + std::shared_ptr softmax_pd_; + std::shared_ptr softmax_bwd_pd_; +}; template class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel { @@ -54,56 +127,27 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel { // Same memory descriptor to be used for input and output memory::dims softmax_tz = {src_tz[0], src_tz[1]}; // 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); - } + const std::string key = + platform::MKLDNNHandler::GetHash(softmax_tz, ctx.op().Output("Out")); + const std::string key_softmax_pd = key + "@softmax_pd"; + + // Currently only NC data format is supported + auto softmax_md = MKLDNNMemDesc( + {softmax_tz}, platform::MKLDNNGetDataType(), 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*/); + auto softmax_pd = std::make_shared( + softmax_desc, mkldnn_engine); + dev_ctx.SetBlob(key_softmax_pd, softmax_pd); + + SoftmaxMKLDNNHandler handler(softmax_pd, dev_ctx, mkldnn_engine, key); + auto softmax_src_memory_p = + handler.AcquireSrcMemory(softmax_md, to_void_cast(input_data)); + auto softmax_dst_memory_p = + handler.AcquireDstMemory(softmax_md, to_void_cast(output_data)); + auto softmax_p = + handler.AcquireSoftmax(softmax_dst_memory_p, softmax_src_memory_p); std::vector pipeline{ *(static_cast(softmax_p.get()))}; @@ -120,6 +164,77 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel { } }; +template +class SoftmaxMKLDNNGradKernel : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + + auto& dev_ctx = ctx.template device_context(); + auto mkldnn_engine = dev_ctx.GetEngine(); + const Tensor* output = ctx.Input("Out"); + const T* dst_data = output->data(); + + auto* dout = ctx.template Input(framework::GradVarName("Out")); + const auto* diff_dst_ptr = dout->template data(); + + auto* dx = + ctx.template Output(framework::GradVarName("X")); + T* diff_src_ptr = dx->template mutable_data(ctx.GetPlace()); + + std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); + std::vector src_tz(dst_tz); + PADDLE_ENFORCE(output->dims().size() == 2UL, + "The input of softmax op must be a 2D matrix."); + // MKL-DNN does support softmax over selected axis. Having 2D Tensor, + // we will make normalization after final eg. axis: 1 + PADDLE_ENFORCE(((src_tz[0] == dst_tz[0]) && (src_tz[1] == dst_tz[1])), + "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 + // retrieve eltwise primitive desc from device context + const std::string key = + platform::MKLDNNHandler::GetHash(softmax_tz, ctx.op().Input("Out")); + const std::string key_softmax_pd = key + "@softmax_pd"; + + auto softmax_pd = + std::static_pointer_cast( + dev_ctx.GetBlob(key_softmax_pd)); + PADDLE_ENFORCE(softmax_pd != nullptr, + "Fail to find softmax_pd in device context"); + + // TODO(jczaja): Add layouts support when there is a need to do so + // Two dimensional softmax does support NC format + auto data_softmax_md = MKLDNNMemDesc( + {softmax_tz}, platform::MKLDNNGetDataType(), memory::format::nc); + auto diff_softmax_md = MKLDNNMemDesc( + {softmax_tz}, platform::MKLDNNGetDataType(), memory::format::nc); + // Normalization is made after innermost dimension eg. C out of NC + auto softmax_bwd_desc = + softmax_backward::desc(diff_softmax_md, data_softmax_md, 1 /* dim: C*/); + auto softmax_bwd_pd = + std::make_shared( + softmax_bwd_desc, mkldnn_engine, *softmax_pd); + + SoftmaxMKLDNNHandler handler(softmax_pd, softmax_bwd_pd, dev_ctx, + mkldnn_engine, key); + auto dst_memory_p = + handler.AcquireDstMemory(data_softmax_md, to_void_cast(dst_data)); + auto diff_dst_memory_p = handler.AcquireDiffDstMemory( + diff_softmax_md, to_void_cast(diff_dst_ptr)); + auto diff_src_memory_p = handler.AcquireDiffSrcMemory( + diff_softmax_md, to_void_cast(diff_src_ptr)); + + // Get primitve from device context + auto softmax_bwd_p = handler.AcquireSoftmaxBackward( + dst_memory_p, diff_dst_memory_p, diff_src_memory_p); + + std::vector pipeline{*softmax_bwd_p}; + stream(stream::kind::eager).submit(pipeline).wait(); + } +}; } // namespace operators } // namespace paddle @@ -127,3 +242,5 @@ namespace ops = paddle::operators; REGISTER_OP_KERNEL(softmax, MKLDNN, ::paddle::platform::CPUPlace, ops::SoftmaxMKLDNNKernel); +REGISTER_OP_KERNEL(softmax_grad, MKLDNN, ::paddle::platform::CPUPlace, + ops::SoftmaxMKLDNNGradKernel); diff --git a/paddle/fluid/operators/softmax_op.cc b/paddle/fluid/operators/softmax_op.cc index 847b3cbd1bd416ae1326211c98ba9d145c103298..31a7458f637921c290fc71ac748143867b4aae19 100644 --- a/paddle/fluid/operators/softmax_op.cc +++ b/paddle/fluid/operators/softmax_op.cc @@ -145,16 +145,30 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel { const framework::ExecutionContext& ctx) const override { // choose cudnn kernel if the runtime supported. framework::LibraryType library_{framework::LibraryType::kPlain}; + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library_ = framework::LibraryType::kCUDNN; } #endif - std::string data_format = ctx.Attr("data_format"); - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), - framework::StringToDataLayout(data_format), library_); +#ifdef PADDLE_WITH_MKLDNN + if (library_ == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; + } +#endif + auto input_data_type = + framework::ToDataType(ctx.Input("X")->type()); + if (input_data_type == framework::proto::VarType::FP16) { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "float16 can only be used on GPU place"); + } + + return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout_, + library_); } }; diff --git a/paddle/fluid/operators/sum_mkldnn_op.cc b/paddle/fluid/operators/sum_mkldnn_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..f78d977760f18c9eb1270e515e68acb208a7c9a4 --- /dev/null +++ b/paddle/fluid/operators/sum_mkldnn_op.cc @@ -0,0 +1,240 @@ +// 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. + +/*Licensed under the Apache License, Version 2.0(the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT 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 "mkldnn.hpp" +#include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" +#include "paddle/fluid/operators/sum_op.h" +#include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/mkldnn_helper.h" + +namespace paddle { +namespace operators { + +using paddle::framework::Tensor; +using paddle::platform::MKLDNNDeviceContext; +using paddle::platform::CPUDeviceContext; +using framework::DataLayout; +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::stream; +using mkldnn::sum; +using mkldnn::reorder; +using platform::to_void_cast; + +template +class SumMKLDNNOpKernel : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + auto& dev_ctx = ctx.template device_context(); + const auto& mkldnn_engine = dev_ctx.GetEngine(); + auto in_vars = ctx.MultiInputVar("X"); + + const int N = in_vars.size(); + auto out_var = ctx.OutputVar("Out"); + bool in_place = out_var == in_vars[0]; + + if (out_var->IsType()) { + LoDTensor* output = ctx.Output("Out"); + T* output_data = output->mutable_data(ctx.GetPlace()); + + std::vector dst_tz = framework::vectorize2int(output->dims()); + auto src_tz = dst_tz; + memory::format output_format{memory::format::format_undef}; + std::vector scales; + std::vector srcs_mpd; + std::vector srcs_mem; + + PADDLE_ENFORCE(in_vars[0]->IsType(), + "Input[0] must be LoDTensors"); + auto& input0 = in_vars[0]->Get(); + PADDLE_ENFORCE(input0.layout() == DataLayout::kMKLDNN && + input0.format() != memory::format::format_undef, + "Wrong layout/format for inputs[0]"); + + memory::format input_format = input0.format(); + + if (src_tz.size() == 1 && (input_format == memory::format::nchw || + input_format == memory::format::nhwc)) { + input_format = memory::format::x; + } + if (src_tz.size() == 2 && (input_format == memory::format::nchw || + input_format == memory::format::nhwc)) { + input_format = memory::format::nc; + } + + for (int i = in_place ? 1 : 0; i < N; i++) { + PADDLE_ENFORCE(in_vars[i]->IsType(), + "all inputs must be all LoDTensors"); + auto& input = in_vars[i]->Get(); + PADDLE_ENFORCE(input.layout() == DataLayout::kMKLDNN && + input.format() != memory::format::format_undef, + "Wrong layout/format for inputs"); + + if (input.numel() == 0) { + continue; + } + + const T* input_data = input.data(); + + auto src_md = + memory::desc(src_tz, memory::data_type::f32, input_format); + auto src_mpd = memory::primitive_desc(src_md, mkldnn_engine); + auto src_mem = memory(src_mpd, to_void_cast(input_data)); + srcs_mpd.push_back(src_mpd); + srcs_mem.push_back(src_mem); + scales.push_back(1.0); + } + + auto dst_md = + memory::desc(dst_tz, memory::data_type::f32, memory::format::any); + + auto sum_pd = sum::primitive_desc(dst_md, scales, srcs_mpd); + + std::shared_ptr dst_mem; + if (in_place) { + dst_mem.reset(new memory(sum_pd.dst_primitive_desc())); + } else { + dst_mem.reset(new memory(sum_pd.dst_primitive_desc(), output_data)); + } + std::vector inputs; + for (size_t i = 0; i < srcs_mem.size(); ++i) { + inputs.push_back(srcs_mem[i]); + } + + auto sum_prim = mkldnn::sum(sum_pd, inputs, *dst_mem); + output_format = (memory::format)platform::GetMKLDNNFormat(sum_pd); + + primitive reorder_prim; + std::shared_ptr target_mem; + if (in_place) { + output_format = input_format; + target_mem.reset(new memory( + {{{src_tz}, memory::data_type::f32, output_format}, mkldnn_engine}, + output_data)); + reorder_prim = reorder(*dst_mem, *target_mem); + } + + std::vector pipeline; + pipeline.push_back(sum_prim); + if (in_place) pipeline.push_back(reorder_prim); + stream(stream::kind::eager).submit(pipeline).wait(); + + output->set_layout(DataLayout::kMKLDNN); + output->set_format(output_format); + } else if (out_var->IsType()) { + // TODO(@mozga-intel) Add MKLDNN SelectedRows support + std::unique_ptr in0; + if (in_place) { + // If is in_place, we store the input[0] to in0 + auto& in_sel0 = in_vars[0]->Get(); + auto& rows = in_sel0.rows(); + in0.reset(new framework::SelectedRows(rows, in_sel0.height())); + in0->mutable_value()->ShareDataWith(in_sel0.value()); + } + + auto get_selected_row = [&](size_t i) -> const SelectedRows& { + if (i == 0 && in0) { + return *in0.get(); + } else { + return in_vars[i]->Get(); + } + }; + auto* out = ctx.Output("Out"); + out->mutable_rows()->clear(); + auto* out_value = out->mutable_value(); + + // Runtime InferShape + size_t first_dim = 0; + for (int i = 0; i < N; i++) { + auto& sel_row = get_selected_row(i); + first_dim += sel_row.rows().size(); + } + auto in_dim = + framework::vectorize(get_selected_row(N - 1).value().dims()); + in_dim[0] = static_cast(first_dim); + + out_value->Resize(framework::make_ddim(in_dim)); + + // if all the input sparse vars are empty, no need to + // merge these vars. + if (first_dim == 0UL) { + return; + } + out_value->mutable_data(ctx.GetPlace()); + math::SelectedRowsAddTo functor; + int64_t offset = 0; + for (int i = 0; i < N; i++) { + auto& sel_row = get_selected_row(i); + if (sel_row.rows().size() == 0) { + continue; + } + PADDLE_ENFORCE_EQ(out->height(), sel_row.height()); + functor(ctx.template device_context(), sel_row, + offset, out); + offset += sel_row.value().numel(); + } + } else if (out_var->IsType()) { + // TODO(@mozga-intel) Add MKLDNN LoDTensorArray support + auto& out_array = *out_var->GetMutable(); + for (size_t i = in_place ? 1 : 0; i < in_vars.size(); ++i) { + PADDLE_ENFORCE(in_vars[i]->IsType(), + "Only support all inputs are TensorArray"); + auto& in_array = in_vars[i]->Get(); + + for (size_t i = 0; i < in_array.size(); ++i) { + if (in_array[i].numel() != 0) { + if (i >= out_array.size()) { + out_array.resize(i + 1); + } + if (out_array[i].numel() == 0) { + framework::TensorCopy(in_array[i], in_array[i].place(), + ctx.device_context(), &out_array[i]); + out_array[i].set_lod(in_array[i].lod()); + } else { + PADDLE_ENFORCE(out_array[i].lod() == in_array[i].lod()); + auto in = EigenVector::Flatten(in_array[i]); + auto result = EigenVector::Flatten(out_array[i]); + result.device(*ctx.template device_context() + .eigen_device()) = result + in; + } + } + } + } + } else { + PADDLE_THROW("Unexpected branch, output variable type is %s", + out_var->Type().name()); + } + } +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OP_KERNEL(sum, MKLDNN, ::paddle::platform::CPUPlace, + paddle::operators::SumMKLDNNOpKernel); diff --git a/paddle/fluid/operators/sum_op.cc b/paddle/fluid/operators/sum_op.cc index 863baba9ea7663d0b21875e0b423dc4a6ce2d59a..fe7c7039c7dec714e265ede1b7167fd800ddc2f7 100644 --- a/paddle/fluid/operators/sum_op.cc +++ b/paddle/fluid/operators/sum_op.cc @@ -18,6 +18,10 @@ limitations under the License. */ #include "paddle/fluid/framework/var_type_inference.h" #include "paddle/fluid/operators/detail/safe_ref.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif + namespace paddle { namespace operators { using framework::Tensor; @@ -63,6 +67,18 @@ class SumOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { auto x_vars = ctx.MultiInputVar("X"); + + framework::LibraryType library{framework::LibraryType::kPlain}; + framework::DataLayout layout{framework::DataLayout::kAnyLayout}; + +#ifdef PADDLE_WITH_MKLDNN + if (library == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; + } +#endif + if (x_vars[0]->IsType()) { int dtype = -1; for (auto& x_var : x_vars) { @@ -80,26 +96,27 @@ class SumOp : public framework::OperatorWithKernel { "Sum operator should have at least one tensor"); return framework::OpKernelType( - static_cast(dtype), - ctx.device_context()); + static_cast(dtype), ctx.GetPlace(), + layout, library); } else if (x_vars[0]->IsType()) { for (auto& var : x_vars) { auto& value = var->Get().value(); if (value.IsInitialized()) { return framework::OpKernelType(framework::ToDataType(value.type()), - ctx.device_context()); + ctx.device_context(), layout, library); } } // if input sparse vars are not initialized, use an default kernel type. return framework::OpKernelType(framework::proto::VarType::FP32, - ctx.device_context()); + ctx.device_context(), layout, library); } else if (x_vars[0]->IsType()) { for (auto& x_var : x_vars) { auto& array = x_var->Get(); for (auto& each : array) { if (each.numel() != 0) { return framework::OpKernelType(framework::ToDataType(each.type()), - ctx.device_context()); + ctx.device_context(), layout, + library); } } } @@ -116,6 +133,9 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("X", "(vector) The input tensors of sum operator.") .AsDuplicable(); AddOutput("Out", "(Tensor) The output tensor of sum operator.").Reuse("X"); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddComment(R"DOC( Sum operator. @@ -132,7 +152,6 @@ class SumOpVarTypeInference : public framework::VarTypeInference { framework::BlockDesc* block) const override { auto& inputs = op_desc.Input("X"); auto var_type = framework::proto::VarType::SELECTED_ROWS; - for (auto& name : op_desc.Input("X")) { VLOG(10) << name << " " << block->FindRecursiveOrCreateVar(name).GetType(); @@ -206,6 +225,7 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker, ops::SumOpVarTypeInference); + REGISTER_OP_CPU_KERNEL( sum, ops::SumKernel, ops::SumKernel, diff --git a/paddle/fluid/operators/tensorrt_engine_op.cc b/paddle/fluid/operators/tensorrt_engine_op.cc index 0ea273af9d5a5c8f1ae112232a9187675031b360..647cfc0a0af2be85e2868c6f68cab962c6631a8d 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.cc +++ b/paddle/fluid/operators/tensorrt_engine_op.cc @@ -14,11 +14,14 @@ #ifdef PADDLE_WITH_CUDA -#include "paddle/fluid/operators/tensorrt_engine_op.h" +#include +#include + #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #include "paddle/fluid/inference/tensorrt/engine.h" #include "paddle/fluid/inference/utils/singleton.h" +#include "paddle/fluid/operators/tensorrt_engine_op.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/tensorrt_engine_op.h b/paddle/fluid/operators/tensorrt_engine_op.h index 8455d24ddf47382b235edda10cb9b2e8934c5f06..295d6ba0395b68cabab3bd4117cedd912df48f5d 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.h +++ b/paddle/fluid/operators/tensorrt_engine_op.h @@ -16,10 +16,12 @@ #ifdef PADDLE_WITH_CUDA +#include +#include + #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/inference/analysis/helper.h" #include "paddle/fluid/inference/tensorrt/engine.h" -#include "paddle/fluid/inference/tensorrt/engine.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/tensorrt_engine_op_test.cc b/paddle/fluid/operators/tensorrt_engine_op_test.cc index 3a2fef48052ae3943abad14bf87c14ca79251c94..358e2d151bb8f990503ea8a51ba5f81e0a1dc816 100644 --- a/paddle/fluid/operators/tensorrt_engine_op_test.cc +++ b/paddle/fluid/operators/tensorrt_engine_op_test.cc @@ -179,7 +179,6 @@ void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) { const std::string& z_name, bool x_created, const shape_t& x_shape, const shape_t& y_shape, const shape_t& z_shape) { - LOG(INFO) << "create fc op"; auto* fc = block_desc.AppendOp(); fc->SetType("mul"); diff --git a/paddle/fluid/operators/test_send_nccl_id.cc b/paddle/fluid/operators/test_send_nccl_id.cc index 5015b1005569ba70b147ebb795243e24ab81ea5c..e2b7b6b8e447381229e4ad594b7974bc0aa159d5 100644 --- a/paddle/fluid/operators/test_send_nccl_id.cc +++ b/paddle/fluid/operators/test_send_nccl_id.cc @@ -21,7 +21,7 @@ limitations under the License. */ #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/operators/detail/macros.h" -#include "paddle/fluid/operators/detail/request_handler_impl.h" +#include "paddle/fluid/operators/distributed/request_handler_impl.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" @@ -37,11 +37,11 @@ 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 distributed = paddle::operators::distributed; namespace string = paddle::string; -std::unique_ptr g_rpc_service; -std::unique_ptr g_req_handler; +std::unique_ptr g_rpc_service; +std::unique_ptr g_req_handler; void StartServer() { f::Scope scope; @@ -57,14 +57,14 @@ void StartServer() { g_req_handler->SetProgram(&empty_program); g_req_handler->SetExecutor(&executor); - g_rpc_service->RegisterRPC(detail::kRequestSend, g_req_handler.get()); + g_rpc_service->RegisterRPC(distributed::kRequestSend, g_req_handler.get()); g_req_handler->SetRPCServer(g_rpc_service.get()); std::thread server_thread( - std::bind(&detail::RPCServer::StartServer, g_rpc_service.get())); + std::bind(&distributed::RPCServer::StartServer, g_rpc_service.get())); - g_rpc_service->SetCond(detail::kRequestSend); - g_rpc_service->WaitBarrier(detail::kRequestSend); + g_rpc_service->SetCond(distributed::kRequestSend); + g_rpc_service->WaitBarrier(distributed::kRequestSend); LOG(INFO) << "got nccl id and stop server..."; g_rpc_service->ShutDown(); @@ -72,7 +72,7 @@ void StartServer() { } TEST(SendNcclId, RPCServer) { - g_req_handler.reset(new detail::RequestSendHandler(true)); + g_req_handler.reset(new distributed::RequestSendHandler(true)); g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 1)); std::thread server_thread(StartServer); @@ -91,7 +91,8 @@ TEST(SendNcclId, RPCServer) { std::string ep = string::Sprintf("127.0.0.1:%d", port); - detail::RPCClient* client = detail::RPCClient::GetInstance(); + distributed::RPCClient* client = + distributed::RPCClient::GetInstance(); LOG(INFO) << "connect to server" << ep; client->AsyncSendVar(ep, dev_ctx, scope, NCCL_ID_VARNAME); diff --git a/paddle/fluid/operators/while_op.cc b/paddle/fluid/operators/while_op.cc index 175c3ac5d79f24e47d21417df8e3eaeb4d5b2335..f440058e8db2024f5c8a0129db3af87a80d6e551 100644 --- a/paddle/fluid/operators/while_op.cc +++ b/paddle/fluid/operators/while_op.cc @@ -203,11 +203,11 @@ class WhileGradOp : public framework::OperatorBase { ->set_lod(inside_tensor.lod()); } } - auto new_inside_name = cur_scope.Rename(inside_grad_name); auto sum_op = framework::OpRegistry::CreateOp( "sum", {{"X", {pg_names[param_id], new_inside_name}}}, - {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{}); + {{"Out", {pg_names[param_id]}}}, + framework::AttributeMap{{"use_mkldnn", {false}}}); sum_op->Run(cur_scope, dev_place); cur_scope.Rename(new_inside_name, inside_grad_name); } diff --git a/paddle/fluid/platform/cpu_info.cc b/paddle/fluid/platform/cpu_info.cc index 40dc7c9a0b6a40f2419ace3ce7e0e5e82bc95c1a..f832d72b53e8d06a32d5c0ac2ecf7130aa28a666 100644 --- a/paddle/fluid/platform/cpu_info.cc +++ b/paddle/fluid/platform/cpu_info.cc @@ -28,9 +28,15 @@ DEFINE_double(fraction_of_cpu_memory_to_use, 1, "Default use 100% of CPU memory for PaddlePaddle," "reserve the rest for page tables, etc"); -DEFINE_uint64( - initial_cpu_memory_in_mb, 500, - "Default initial 500MB of CPU memory for PaddlePaddle, in MD unit."); +DEFINE_uint64(initial_cpu_memory_in_mb, +#ifdef PADDLE_WITH_MKLDNN + /* Aligned with mozga-intel, MKLDNN need at least 5000 MB + * to obtain the best performance*/ + 5000, +#else + 500, +#endif + "Initial CPU memory for PaddlePaddle, in MD unit."); DEFINE_double( fraction_of_cuda_pinned_memory_to_use, 0.5, @@ -59,10 +65,7 @@ inline size_t CpuTotalPhysicalMemory() { size_t CpuMaxAllocSize() { // For distributed systems, it requires configuring and limiting // the fraction of memory to use. - return std::min( - static_cast(FLAGS_fraction_of_cpu_memory_to_use * - CpuTotalPhysicalMemory()), - static_cast(FLAGS_initial_cpu_memory_in_mb * 1 << 20)); + return FLAGS_fraction_of_cpu_memory_to_use * CpuTotalPhysicalMemory(); } size_t CpuMinChunkSize() { @@ -71,8 +74,11 @@ size_t CpuMinChunkSize() { } size_t CpuMaxChunkSize() { - // Allow to allocate the maximum chunk size is roughly 3% of CPU memory. - return CpuMaxAllocSize() / 32; + // Allow to allocate the maximum chunk size is roughly 3% of CPU memory, + // or the initial_cpu_memory_in_mb. + return std::min( + static_cast(CpuMaxAllocSize() / 32), + static_cast(FLAGS_initial_cpu_memory_in_mb * 1 << 20)); } size_t CUDAPinnedMaxAllocSize() { diff --git a/paddle/fluid/platform/dynload/CMakeLists.txt b/paddle/fluid/platform/dynload/CMakeLists.txt index 364c4901b297dbd647faae85b01f682a1daace9c..6dd19aaeffef8aa8a7d1997915908af04273d50c 100644 --- a/paddle/fluid/platform/dynload/CMakeLists.txt +++ b/paddle/fluid/platform/dynload/CMakeLists.txt @@ -1,11 +1,16 @@ cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags enforce) -list(APPEND CUDA_SRCS cublas.cc cudnn.cc curand.cc nccl.cc) +list(APPEND CUDA_SRCS cublas.cc cudnn.cc curand.cc) + +# There is no macOS version of NCCL. +if (NOT APPLE) + list(APPEND CUDA_SRCS nccl.cc) +endif() + if (TENSORRT_FOUND) list(APPEND CUDA_SRCS tensorrt.cc) endif() - configure_file(cupti_lib_path.h.in ${CMAKE_CURRENT_BINARY_DIR}/cupti_lib_path.h) if (CUPTI_FOUND) list(APPEND CUDA_SRCS cupti.cc) diff --git a/paddle/fluid/platform/enforce.h b/paddle/fluid/platform/enforce.h index 7b8c29e1e642ec6bb4023afd8c083311b8b31812..a34e4371cccfd1be0d173fa11595e4368eb65b85 100644 --- a/paddle/fluid/platform/enforce.h +++ b/paddle/fluid/platform/enforce.h @@ -44,8 +44,10 @@ limitations under the License. */ #include "paddle/fluid/platform/dynload/cublas.h" #include "paddle/fluid/platform/dynload/cudnn.h" #include "paddle/fluid/platform/dynload/curand.h" +#ifndef __APPLE__ #include "paddle/fluid/platform/dynload/nccl.h" -#endif +#endif // __APPLE__ +#endif // PADDLE_WITH_CUDA namespace paddle { namespace platform { @@ -174,6 +176,7 @@ inline typename std::enable_if::type throw_on_error( throw std::runtime_error(err + string::Sprintf(args...)); } +#ifndef __APPLE__ template inline typename std::enable_if::type throw_on_error( ncclResult_t stat, const Args&... args) { @@ -184,7 +187,7 @@ inline typename std::enable_if::type throw_on_error( string::Sprintf(args...)); } } - +#endif // __APPLE__ #endif // PADDLE_WITH_CUDA template diff --git a/paddle/fluid/platform/mkldnn_helper.h b/paddle/fluid/platform/mkldnn_helper.h index de711b7d23ef01d57a62087c552ea090f01f0386..ed99932546446eb877c9701de15e2d37d29b5f88 100644 --- a/paddle/fluid/platform/mkldnn_helper.h +++ b/paddle/fluid/platform/mkldnn_helper.h @@ -99,5 +99,143 @@ inline mkldnn::memory::format GetMKLDNNFormat(const mkldnn::memory memory) { memory.get_primitive_desc().desc().data.format); } +inline mkldnn::memory::format GetMKLDNNFormat( + const mkldnn::sum::primitive_desc& memory) { + return static_cast( + memory.dst_primitive_desc().desc().data.format); +} + +class MKLDNNHandler { + public: + MKLDNNHandler(const MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine, + const std::string& base_key) + : dev_ctx_(dev_ctx), + engine_(engine), + key_(base_key), + is_reusing_(false) {} + + std::shared_ptr AcquireSrcMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_src_mem_p"); + } + + std::shared_ptr AcquireWeightsMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_weights_mem_p"); + } + + std::shared_ptr AcquireDstMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_dst_mem_p"); + } + + std::shared_ptr AcquireDiffDstMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_diff_dst_mem_p"); + } + + std::shared_ptr AcquireDiffSrcMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_diff_src_mem_p"); + } + + std::shared_ptr AcquireMemoryFromPrimitive( + mkldnn::memory::primitive_desc mdp, void* ptr, + const std::string& suffix) { + auto local_key = key_ + suffix; + auto mem_p = + std::static_pointer_cast(dev_ctx_.GetBlob(local_key)); + PADDLE_ENFORCE((mem_p != nullptr) || (is_reusing_ == false), + "Fail to find mem primitive in device context"); + if (mem_p == nullptr) { + mem_p = std::make_shared(mdp, ptr); + dev_ctx_.SetBlob(local_key, mem_p); + } else { + mem_p->set_data_handle(ptr); + // Mark that reusing happenned. All primitives from operator instance + // should be reused or none of them. So we check consistency + is_reusing_ = true; + } + return mem_p; + } + + std::shared_ptr AcquireMemory(const mkldnn::memory::desc& md, + void* ptr, + const std::string& suffix) { + /*Generate key*/ + auto local_key = key_ + suffix; + auto mem_p = + std::static_pointer_cast(dev_ctx_.GetBlob(local_key)); + PADDLE_ENFORCE((mem_p != nullptr) || (is_reusing_ == false), + "Fail to find mem primitive in device context"); + if (mem_p == nullptr) { + mem_p = std::make_shared( + mkldnn::memory::primitive_desc{md, engine_}, ptr); + dev_ctx_.SetBlob(local_key, mem_p); + } else { + mem_p->set_data_handle(ptr); + // Mark that reusing happenned. All primitives from operator instance + // should be reused or none of them. So we check consistency + is_reusing_ = true; + } + return mem_p; + } + + std::shared_ptr AcquireMemory( + mkldnn::memory::primitive_desc& mpd, + mkldnn::memory::primitive_desc& user_mpd, + const std::shared_ptr user_memory_p, + const std::string& suffix, std::vector& pipeline) { + // create reorder primitive if the input format is not the preferred one + auto local_key = key_ + suffix; + auto key_reorder_p = key_ + suffix + "reorder_p"; + + auto target_memory_p = + std::static_pointer_cast(dev_ctx_.GetBlob(local_key)); + PADDLE_ENFORCE((target_memory_p != nullptr) || (is_reusing_ == false), + "Fail to find mem primitive in device context"); + if (target_memory_p == nullptr) { + target_memory_p = user_memory_p; + std::shared_ptr reorder_p; + if (mpd != user_mpd) { + target_memory_p = std::make_shared(mpd); + + auto reorder_p = + std::make_shared(*user_memory_p, *target_memory_p); + dev_ctx_.SetBlob(key_reorder_p, reorder_p); + pipeline.push_back(*reorder_p); + } + dev_ctx_.SetBlob(local_key, target_memory_p); + } else { + // Make reorder if needed + auto reorder_p = std::static_pointer_cast( + dev_ctx_.GetBlob(key_reorder_p)); + if (reorder_p != nullptr) { + pipeline.push_back(*reorder_p); + } + is_reusing_ = true; + } + return target_memory_p; + } + + static std::string GetHash(mkldnn::memory::dims& operand_dims, + const std::string& suffix) { + auto dims2str = [](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 dims2str(operand_dims) + suffix; + }; + + protected: + const MKLDNNDeviceContext& dev_ctx_; + mkldnn::engine engine_; + std::string key_; + bool is_reusing_; +}; + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc index bcf6d4dd3087060c016e53722cde80704ef2e834..fcd3356d44ee592233c3883d439d0677714900b8 100644 --- a/paddle/fluid/pybind/protobuf.cc +++ b/paddle/fluid/pybind/protobuf.cc @@ -268,7 +268,8 @@ void BindOpDesc(pybind11::module *m) { .value("STRINGS", pd::proto::AttrType::STRINGS) .value("BOOL", pd::proto::AttrType::BOOLEAN) .value("BOOLS", pd::proto::AttrType::BOOLEANS) - .value("BLOCK", pd::proto::AttrType::BLOCK); + .value("BLOCK", pd::proto::AttrType::BLOCK) + .value("BLOCKS", pd::proto::AttrType::BLOCKS); pybind11::class_ op_desc(*m, "OpDesc", ""); op_desc @@ -293,6 +294,7 @@ void BindOpDesc(pybind11::module *m) { .def("set_attr", &pd::OpDesc::SetAttr) .def("attr", &pd::OpDesc::GetAttr) .def("set_block_attr", &pd::OpDesc::SetBlockAttr) + .def("set_blocks_attr", &pd::OpDesc::SetBlocksAttr) .def("set_serialized_attr", [](pd::OpDesc &self, const std::string &name, const pybind11::bytes &seriralized) { diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index 74036bcb3114df8fc4613bd9f4dc327463397dba..5a45e431df993febab676f22da7116d84e441548 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -159,12 +159,14 @@ PYBIND11_PLUGIN(core) { new (&instance) LoDTensor(new_offset_lod); }) .def("__init__", [](LoDTensor &instance) { new (&instance) LoDTensor(); }) + // We implement offset based LOD in C++ while we use length based with + // Python API. So we changed set_lod to set_recursive_sequence_lengths to + // avoid misuse. + // The discussion is here: + // https://github.com/PaddlePaddle/Paddle/issues/10855 .def("set_lod", [](LoDTensor &self, const std::vector> &lod) { // the input lod is offset-based level-of-detail info - LOG(WARNING) - << "set_lod is deprecated and will be removed by 9.2018, " - "please switch to set_recursive_sequence_lengths."; LoD new_lod; new_lod.reserve(lod.size()); std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod)); @@ -191,14 +193,13 @@ PYBIND11_PLUGIN(core) { .def("lod", [](LoDTensor &self) -> std::vector> { // output the offset-based lod info - LOG(WARNING) << "lod is deprecated and will be removed by 9.2018, " - "please switch to recursive_sequence_lengths."; LoD lod = self.lod(); std::vector> new_lod; new_lod.reserve(lod.size()); std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod)); return new_lod; }) + // Set above comments of set_lod. .def("recursive_sequence_lengths", [](LoDTensor &self) -> std::vector> { // output the length-based lod info diff --git a/paddle/fluid/pybind/tensor_py.h b/paddle/fluid/pybind/tensor_py.h index 93b09ed6922b32a5531224acc470daf0d97f95bd..6da3846ac69980daac4f0fb7401b2573c21c89bf 100644 --- a/paddle/fluid/pybind/tensor_py.h +++ b/paddle/fluid/pybind/tensor_py.h @@ -97,7 +97,7 @@ struct CastToPyBufferImpl { inline pybind11::buffer_info CastToPyBuffer(const framework::Tensor &tensor) { auto buffer_info = details::CastToPyBufferImpl()(tensor); + uint8_t, platform::float16>()(tensor); return buffer_info; } diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh index e8b305326702cf04b752bb2eb413f848daa5ec7b..037688bde9122c1d999e90f2438977b46c1eb531 100755 --- a/paddle/scripts/paddle_build.sh +++ b/paddle/scripts/paddle_build.sh @@ -22,7 +22,7 @@ function print_usage() { echo -e "\n${RED}Usage${NONE}: ${BOLD}${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 @@ -133,7 +133,7 @@ EOF -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \ -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \ -DWITH_CONTRIB=${WITH_CONTRIB:-ON} \ - -DWITH_ANAKIN=ON + -DWITH_ANAKIN=${WITH_ANAKIN:-ON} } function abort(){ @@ -198,7 +198,7 @@ function build_android() { fi ANDROID_STANDALONE_TOOLCHAIN=$ANDROID_TOOLCHAINS_DIR/$ANDROID_ARCH-android-$ANDROID_API - + cat <(new_argv.size()); char** new_argv_address = new_argv.data(); diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py index bd985ad733aa8eece2f8374d033f452a0175a011..45af83708ea63fc1b6aa86f1e8423bb44b7388a6 100644 --- a/python/paddle/fluid/__init__.py +++ b/python/paddle/fluid/__init__.py @@ -44,7 +44,7 @@ import metrics import transpiler from param_attr import ParamAttr, WeightNormParamAttr from data_feeder import DataFeeder -from core import LoDTensor, CPUPlace, CUDAPlace, CUDAPinnedPlace +from core import LoDTensor, CPUPlace, CUDAPlace, CUDAPinnedPlace, Scope from transpiler import DistributeTranspiler, InferenceTranspiler, \ memory_optimize, release_memory from concurrency import (Go, make_channel, channel_send, channel_recv, @@ -83,6 +83,7 @@ __all__ = framework.__all__ + executor.__all__ + concurrency.__all__ + \ 'profiler', 'unique_name', 'recordio_writer', + 'Scope', ] @@ -117,7 +118,7 @@ def __bootstrap__(): read_env_flags = [ 'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir', - 'eager_delete_scope', 'use_mkldnn' + 'eager_delete_scope', 'use_mkldnn', 'initial_cpu_memory_in_mb' ] if core.is_compiled_with_cuda(): read_env_flags += [ diff --git a/python/paddle/fluid/average.py b/python/paddle/fluid/average.py index 6abe8233b07c484494848c566e9898600a7d8f5c..358e24df31bb517604481bb48b9180e579f8460d 100644 --- a/python/paddle/fluid/average.py +++ b/python/paddle/fluid/average.py @@ -36,6 +36,25 @@ def _is_number_or_matrix_(var): class WeightedAverage(object): + """ + Calculate weighted average. + + The average calculating is accomplished via Python totally. + They do not change Paddle's Program, nor do anything to + modify NN model's configuration. They are completely + wrappers of Python functions. + + Examples: + .. code-block:: python + avg = fluid.average.WeightedAverage() + avg.add(value=2.0, weight=1) + avg.add(value=4.0, weight=2) + avg.eval() + + # The result is 3.333333333. + # For (2.0 * 1 + 4.0 * 2) / (1 + 2) = 3.333333333 + """ + def __init__(self): warnings.warn( "The %s is deprecated, please use fluid.metrics.Accuracy instead." % diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py index 4f9622d04dc98f41b503ceb780802d2a4e4c58a0..4faa06303170488d0de2fda4c1461cfe2d623d35 100644 --- a/python/paddle/fluid/backward.py +++ b/python/paddle/fluid/backward.py @@ -132,9 +132,9 @@ def _addup_repetitive_outputs_(op_descs): for idx, op_desc in enumerate(op_descs): for var_name in op_desc.input_arg_names(): if len(renamed_vars[var_name]) > 1: - pending_sum_ops.append( - (_create_op_desc_("sum", {"X": renamed_vars[var_name]}, - {"Out": [var_name]}, {}), idx)) + pending_sum_ops.append((_create_op_desc_( + "sum", {"X": renamed_vars[var_name]}, {"Out": [var_name]}, + {"use_mkldnn": False}), idx)) renamed_vars[var_name] = [var_name] for var_name in op_desc.output_arg_names(): if var_name == core.empty_var_name( @@ -147,7 +147,7 @@ def _addup_repetitive_outputs_(op_descs): else: if len(renamed_vars[var_name]) == 1: new_name = var_name + "@RENAME@" + \ - str(var_rename_count[var_name]) + str(var_rename_count[var_name]) var_rename_count[var_name] += 1 # rename original var_name renamed_vars[var_name][0] = new_name @@ -155,14 +155,15 @@ def _addup_repetitive_outputs_(op_descs): _rename_arg_(pending_sum_ops, var_name, new_name) new_name = var_name + "@RENAME@" + \ - str(var_rename_count[var_name]) + str(var_rename_count[var_name]) var_rename_count[var_name] += 1 op_desc.rename_output(var_name, new_name) renamed_vars[var_name].append(new_name) for var_name, inputs in renamed_vars.iteritems(): if len(inputs) > 1: - pending_sum_ops.append((_create_op_desc_( - "sum", {"X": inputs}, {"Out": [var_name]}, {}), len(op_descs))) + pending_sum_ops.append( + (_create_op_desc_("sum", {"X": inputs}, {"Out": [var_name]}, + {"use_mkldnn": False}), len(op_descs))) # sum_op descs are sorted according to their insert position for p in reversed(pending_sum_ops): op_descs.insert(p[1], p[0]) @@ -434,18 +435,65 @@ def _get_stop_gradients_(program): def append_backward(loss, parameter_list=None, no_grad_set=None, callbacks=None): """ - Append backward part to main_program + Append backward part to main_program. - Args: - loss(Variable): The variable generated by cost function. - parameter_list(list[string]): Parameters that need to be updated by - optimizer. If None, it means all parameters need to be updated. - no_grad_set(set): Variables that have no gradients in Block 0. - All variables with `step_gradient=True` from all blocks will be - automatically added. + A complete neural network training is made up of forward and backward + propagation. However, when we configure a network, we only need to + specify its forwrd part. The backward part is generated automatically + according to the forward part by this function. - Return: - (list[(Variable,Variable)]): list of (parameter, gradient) pair. + In most cases, users do not need to invoke this function manually. It + will be automatically invoked by the optimizer's `minimize` function. + + Args: + loss(Variable): The loss variable of the network. + parameter_list(list[string]|None): Names of parameters that need + to be updated by optimizers. + If it is None, all parameters + will be updated. + Default: None + no_grad_set(set|None): Variables in the Block 0 whose gradients + should be ignored. All variables with + `step_gradient=True` from all blocks will + be automatically added into this set. + Default: None + callbacks(list[callable object]|None): The callbacks are used for + doing some custom jobs during + backward part building. All + callable objects in it will + be invoked once each time a + new gradient operator is added + into the program. The callable + object must has two input + parameters: 'block' and 'context'. + The 'block' is the block which + the new gradient operator will + be added to. The 'context' is a + map, whose keys are gradient + variable names and values are + corresponding original variables. + In addition to this, the 'context' + has another special key-value pair: + the key is string '__current_op_desc__' + and the value is the op_desc of the + gradient operator who has just + triggered the callable object. + + Returns: + list[(Variable,Variable)]: Pairs of parameter and its + corresponding gradients. The key is the parameter and the + value is gradient variable. + + Raises: + AssertionError: If `loss` is not an instance of Variable. + + Examples: + .. code-block:: python + + # network configuration code + # ... + avg_loss = fluid.layers.mean(loss) + param_grad_list = fluid.backward.append_backward(loss=avg_loss) """ assert isinstance(loss, framework.Variable) diff --git a/python/paddle/fluid/clip.py b/python/paddle/fluid/clip.py index 66c3fc6b66d61bc9578f84594409ad0f24c99910..18e2f3045e272fb4712391f87bffd3f367c1c744 100644 --- a/python/paddle/fluid/clip.py +++ b/python/paddle/fluid/clip.py @@ -24,8 +24,6 @@ __all__ = [ 'GradientClipByValue', 'GradientClipByNorm', 'GradientClipByGlobalNorm', - 'append_gradient_clip_ops', - 'error_clip_callback', ] @@ -38,6 +36,25 @@ class BaseErrorClipAttr(object): class ErrorClipByValue(BaseErrorClipAttr): + """ + Clips tensor values to the range [min, max]. + + Given a tensor t, this operation clips its value to min and max inplace. + + - Any values less than min are set to min. + - Any values greater than max are set to max. + + Args: + max (float): The maximum value to clip by. + min (float, optional): The minimum value to clip by. if not set by user, \ + will be set to -max by framework. + + Examples: + .. code-block:: python + + var = fluid.framework.Variable(..., error_clip=ErrorClipByValue(max=5.0), ...) + """ + def __init__(self, max, min=None): max = float(max) if min is None: @@ -99,6 +116,31 @@ class NullGradientClipAttr(BaseGradientClipAttr): class GradientClipByValue(BaseGradientClipAttr): + """ + Clips gradient values to the range [min, max]. + + Given a tensor t, this operation clips its value to min and max inplace. + + - Any values less than min are set to min. + - Any values greater than max are set to max. + + Args: + max (float): The maximum value to clip by. + min (float, optional): The minimum value to clip by. if not set by user, \ + will be set to -max by framework. + + Examples: + .. code-block:: python + + w_param_attrs = ParamAttr(name=None, + initializer=UniformInitializer(low=-1.0, high=1.0, seed=0), + learning_rate=1.0, + regularizer=L1Decay(1.0), + trainable=True, + clip=GradientClipByValue(-1.0, 1.0)) + y_predict = fluid.layers.fc(input=x, size=1, param_attr=w_param_attrs) + """ + def __init__(self, max, min=None): max = float(max) if min is None: @@ -120,6 +162,37 @@ class GradientClipByValue(BaseGradientClipAttr): class GradientClipByNorm(BaseGradientClipAttr): + """ + Clips tensor values to a maximum L2-norm. + + This operator limits the L2 norm of the input :math:`X` within :math:`max\_norm`. + If the L2 norm of :math:`X` is less than or equal to :math:`max\_norm`, :math:`Out` + will be the same as :math:`X`. If the L2 norm of :math:`X` is greater than + :math:`max\_norm`, :math:`X` will be linearly scaled to make the L2 norm of + :math:`Out` equal to :math:`max\_norm`, as shown in the following formula: + + .. math:: + + Out = \\frac{max\_norm * X}{norm(X)}, + + where :math:`norm(X)` represents the L2 norm of :math:`X`. + + Args: + clip_norm (float): The maximum norm value + + Examples: + .. code-block:: python + + w_param_attrs = ParamAttr(name=None, + initializer=UniformInitializer(low=-1.0, high=1.0, seed=0), + learning_rate=1.0, + regularizer=L1Decay(1.0), + trainable=True, + clip=GradientClipByNorm(clip_norm=2.0)) + y_predict = fluid.layers.fc(input=x, size=1, param_attr=w_param_attrs) + + """ + def __init__(self, clip_norm): self.clip_norm = clip_norm @@ -135,6 +208,44 @@ class GradientClipByNorm(BaseGradientClipAttr): class GradientClipByGlobalNorm(BaseGradientClipAttr): + """ + Clips values of multiple tensors by the ratio of the sum of their norms. + + Given a list of tensors t_list, and a clipping ratio clip_norm, this + operation returns a list of clipped tensors list_clipped and the global + norm (global_norm) of all tensors in t_list. + + To perform the clipping, the values :math:`t\_list[i]` are set to: + + .. math:: + + t\_list[i] = t\_list[i] * \\frac{clip\_norm}{\max(global\_norm, clip\_norm)} + + where: + + .. math:: + + global\_norm = \sqrt{\sum_{i=0}^{N-1}(l2norm(t\_list[i]))^2} + + If :math:`clip\_norm > global\_norm` then the entries in t_list remain as they are, + otherwise they're all shrunk by the global ratio. + + Args: + clip_norm (float): The maximum norm value + group_name (str, optional): The group name for this clip. + + Examples: + .. code-block:: python + + p_g_clip = fluid.backward.append_backward(loss=avg_cost_clip) + + with fluid.program_guard(main_program=prog_clip): + fluid.clip.set_gradient_clip( + fluid.clip.GradientClipByGlobalNorm(clip_norm=2.0)) + p_g_clip = fluid.clip.append_gradient_clip_ops(p_g_clip) + + """ + def __init__(self, clip_norm, group_name="default_group"): if not isinstance(group_name, basestring): raise TypeError("'group_name' must be a basestring.") @@ -183,15 +294,16 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr): def set_gradient_clip(clip, param_list=None, program=None): """ - To specify parameters that require gradient clip. - Args: - clip(BaseGradientClipAttr): An instance of some derived class of BaseGradientClipAttr, - which describes the type and detailed attributes of required gradient clip. - param_list(list, None by default): Parameters that require gradient clip. - It can be a list of parameter or a list of parameter's name. - When it's None, all parameters in the program will be included. - program(Program, None by default): The program where parameters are. - Will be the default main program when assigned with None. + To specify parameters that require gradient clip. + + Args: + clip(BaseGradientClipAttr): An instance of some derived class of BaseGradientClipAttr, + which describes the type and detailed attributes of required gradient clip. + param_list(list(Variable)): Parameters that require gradient clip. + It can be a list of parameter or a list of parameter's name. + When it's None, all parameters in the program will be included. + program(Program): The program where parameters are. + Will be the default main program when assigned with None. """ if not isinstance(clip, BaseGradientClipAttr): raise TypeError( diff --git a/python/paddle/fluid/data_feeder.py b/python/paddle/fluid/data_feeder.py index ac396002018d5952bee4aa79ff4aaa5463e2e9e1..c859778b3757f638ac531620f241e684522add57 100644 --- a/python/paddle/fluid/data_feeder.py +++ b/python/paddle/fluid/data_feeder.py @@ -29,6 +29,13 @@ class DataToLoDTensorConverter(object): self.place = place self.lod_level = lod_level self.shape = shape + negtive_count = 0 + for s in self.shape: + if s < 0: + negtive_count += 1 + if negtive_count > 1: + self.shape = None + break if dtype == core.VarDesc.VarType.FP32: self.dtype = 'float32' elif dtype == core.VarDesc.VarType.INT64: @@ -61,7 +68,9 @@ class DataToLoDTensorConverter(object): self._feed_impl_(each_data, lod[1:], lod_level - 1) def done(self): - arr = numpy.array(self.data, dtype=self.dtype).reshape(self.shape) + arr = numpy.array(self.data, dtype=self.dtype) + if self.shape: + arr = arr.reshape(self.shape) t = core.LoDTensor() t.set(arr, self.place) if self.lod_level > 0: @@ -70,6 +79,61 @@ class DataToLoDTensorConverter(object): class DataFeeder(object): + """ + DataFeeder converts the data that returned by a reader into a data + structure that can feed into Executor and ParallelExecutor. The reader + usually returns a list of mini-batch data entries. Each data entry in + the list is one sample. Each sample is a list or a tuple with one + feature or multiple features. + + The simple usage shows below: + + .. code-block:: python + + place = fluid.CPUPlace() + 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])]) + + + If you want to feed data into GPU side separately in advance when you + use multi-GPU to train a model, you can use `decorate_reader` function. + + .. code-block:: python + + place=fluid.CUDAPlace(0) + feeder = fluid.DataFeeder(place=place, feed_list=[data, label]) + reader = feeder.decorate_reader( + paddle.batch(flowers.train(), batch_size=16)) + + Args: + feed_list(list): The Variables or Variables'name that will + feed into model. + place(Place): place indicates feed data into CPU or GPU, if you want to + feed data into GPU, please using `fluid.CUDAPlace(i)` (`i` represents + the GPU id), or if you want to feed data into CPU, please using + `fluid.CPUPlace()`. + program(Program): The Program that will feed data into, if program + is None, it will use default_main_program(). Default None. + + Raises: + ValueError: If some Variable is not in this Program. + + Examples: + .. code-block:: python + + # ... + place = fluid.CPUPlace() + feed_list = [ + main_program.global_block().var(var_name) for var_name in feed_vars_name + ] # feed_vars_name is a list of variables' name. + feeder = fluid.DataFeeder(feed_list, place) + for data in reader(): + outs = exe.run(program=main_program, + feed=feeder.feed(data)) + """ + def __init__(self, feed_list, place, program=None): self.feed_dtypes = [] self.feed_names = [] @@ -99,6 +163,16 @@ class DataFeeder(object): self.place = place def feed(self, iterable): + """ + According to feed_list and iterable, converters the input into + a data structure that can feed into Executor and ParallelExecutor. + + Args: + iterable(list|tuple): the input data. + + Returns: + dict: the result of conversion. + """ converter = [] for lod_level, shape, dtype in six.zip( self.feed_lod_level, self.feed_shapes, self.feed_dtypes): @@ -121,6 +195,20 @@ class DataFeeder(object): return ret_dict def feed_parallel(self, iterable, num_places=None): + """ + Takes multiple mini-batches. Each mini-batch will be feed on each + device in advance. + + Args: + iterable(list|tuple): the input data. + num_places(int): the number of devices. Default None. + + Returns: + dict: the result of conversion. + + Notes: + The number of devices and number of mini-batches must be same. + """ if isinstance(self.place, core.CUDAPlace): places = [ core.CUDAPlace(i) @@ -159,6 +247,24 @@ class DataFeeder(object): multi_devices, num_places=None, drop_last=True): + """ + Converter the input data into a data that returned by reader into + multiple mini-batches. Each mini-batch will be feed on each device. + + Args: + reader(fun): the input data. + multi_devices(bool): the number of places. Default None. + num_places(int): the number of places. Default None. + drop_last(bool): the number of places. Default None. + + Returns: + dict: the result of conversion. + + Raises: + ValueError: If drop_last is False and the data batch which cannot + fit for devices. + """ + def __reader_creator__(): if not multi_devices: for item in reader(): diff --git a/python/paddle/fluid/evaluator.py b/python/paddle/fluid/evaluator.py index 7c6ad6f27dcfd7040f79c72c01413c8cc84a28ba..00ba1a0457583d1cc1fa7136ebd51e9ced167832 100644 --- a/python/paddle/fluid/evaluator.py +++ b/python/paddle/fluid/evaluator.py @@ -41,7 +41,12 @@ def _clone_var_(block, var): class Evaluator(object): """ - Base Class for all evaluators + Warning: better to use the fluid.metrics.* things, more + flexible support via pure Python and Operator, and decoupled + with executor. Short doc are intended to urge new user + start from Metrics. + + Base Class for all evaluators. Args: name(str): The name of evaluator. such as, "accuracy". Used for generate @@ -69,6 +74,10 @@ class Evaluator(object): def reset(self, executor, reset_program=None): """ reset metric states at the begin of each pass/user specified batch + + Args: + executor(Executor|ParallelExecutor): a executor for executing the reset_program + reset_program(Program): a single Program for reset process """ if reset_program is None: reset_program = Program() @@ -85,15 +94,16 @@ class Evaluator(object): def eval(self, executor, eval_program=None): """ Evaluate the statistics merged by multiple mini-batches. + Args: + executor(Executor|ParallelExecutor): a executor for executing the eval_program + eval_program(Program): a single Program for eval process """ raise NotImplementedError() - def create_state(self, suffix, dtype, shape): + def _create_state(self, suffix, dtype, shape): """ Create state variable. - NOTE: It is not a public API. - Args: suffix(str): the state suffix. dtype(str|core.VarDesc.VarType): the state data type @@ -113,9 +123,35 @@ class Evaluator(object): class ChunkEvaluator(Evaluator): """ + Warning: This would be deprecated in the future. Please use fluid.metrics.ChunkEvaluator + instead. + Accumulate counter numbers output by chunk_eval from mini-batches and compute the precision recall and F1-score using the accumulated counter numbers. + For some basics of chunking, please refer to + 'Chunking with Support Vector Machines '. + + Args: + input (Variable): prediction output of the network. + label (Variable): label of the test data set. + chunk_scheme (str): can be IOB/IOE/IOBES and IO. See the chunk_eval op for details. + num_chunk_types (int): the number of chunk type. + excluded_chunk_types (list): A list including chunk type ids, indicating chunk types that are not counted. + + Returns: + tuple: tuple containing: precision, recall, f1_score + + Examples: + .. code-block:: python + + exe = fluid.executor(place) + evaluator = fluid.Evaluator.ChunkEvaluator(input, label) + for epoch in PASS_NUM: + evaluator.reset(exe) + for data in batches: + loss = exe.run(fetch_list=[cost]) + distance, instance_error = distance_evaluator.eval(exe) """ def __init__( @@ -130,11 +166,11 @@ class ChunkEvaluator(Evaluator): if main_program.current_block().idx != 0: raise ValueError("You can only invoke Evaluator in root block") - self.num_infer_chunks = self.create_state( + self.num_infer_chunks = self._create_state( dtype='int64', shape=[1], suffix='num_infer_chunks') - self.num_label_chunks = self.create_state( + self.num_label_chunks = self._create_state( dtype='int64', shape=[1], suffix='num_label_chunks') - self.num_correct_chunks = self.create_state( + self.num_correct_chunks = self._create_state( dtype='int64', shape=[1], suffix='num_correct_chunks') precision, recall, f1_score, num_infer_chunks, num_label_chunks, num_correct_chunks = layers.chunk_eval( input=input, @@ -178,6 +214,8 @@ class ChunkEvaluator(Evaluator): class EditDistance(Evaluator): """ + Warning: This would be deprecated in the future. Please use fluid.metrics.EditDistance + instead. Accumulate edit distance sum and sequence number from mini-batches and compute the average edit_distance and instance error of all batches. @@ -188,15 +226,16 @@ class EditDistance(Evaluator): ignored_tokens(list of int): Tokens that should be removed before calculating edit distance. - Example: + Examples: + .. code-block:: python - exe = fluid.executor(place) - distance_evaluator = fluid.Evaluator.EditDistance(input, label) - for epoch in PASS_NUM: - distance_evaluator.reset(exe) - for data in batches: - loss = exe.run(fetch_list=[cost]) - distance, instance_error = distance_evaluator.eval(exe) + exe = fluid.executor(place) + distance_evaluator = fluid.Evaluator.EditDistance(input, label) + for epoch in PASS_NUM: + distance_evaluator.reset(exe) + for data in batches: + loss = exe.run(fetch_list=[cost]) + distance, instance_error = distance_evaluator.eval(exe) In the above example: 'distance' is the average of the edit distance in a pass. @@ -210,11 +249,11 @@ class EditDistance(Evaluator): if main_program.current_block().idx != 0: raise ValueError("You can only invoke Evaluator in root block") - self.total_distance = self.create_state( + self.total_distance = self._create_state( dtype='float32', shape=[1], suffix='total_distance') - self.seq_num = self.create_state( + self.seq_num = self._create_state( dtype='int64', shape=[1], suffix='seq_num') - self.instance_error = self.create_state( + self.instance_error = self._create_state( dtype='int64', shape=[1], suffix='instance_error') distances, seq_num = layers.edit_distance( input=input, label=label, ignored_tokens=ignored_tokens) @@ -256,9 +295,10 @@ class EditDistance(Evaluator): class DetectionMAP(Evaluator): """ + Warning: This would be deprecated in the future. Please use fluid.metrics.DetectionMAP + instead. Calculate the detection mean average precision (mAP). - TODO (Dang Qingqing): update the following doc. The general steps are as follows: 1. calculate the true positive and false positive according to the input of detection and labels. @@ -293,17 +333,18 @@ class DetectionMAP(Evaluator): - 11point: the 11-point interpolated average precision. - integral: the natural integral of the precision-recall curve. - Example: + Examples: + .. code-block:: python - exe = fluid.executor(place) - map_evaluator = fluid.Evaluator.DetectionMAP(input, - gt_label, gt_box, gt_difficult) - cur_map, accum_map = map_evaluator.get_map_var() - fetch = [cost, cur_map, accum_map] - for epoch in PASS_NUM: - map_evaluator.reset(exe) - for data in batches: - loss, cur_map_v, accum_map_v = exe.run(fetch_list=fetch) + exe = fluid.executor(place) + map_evaluator = fluid.Evaluator.DetectionMAP(input, + gt_label, gt_box, gt_difficult) + cur_map, accum_map = map_evaluator.get_map_var() + fetch = [cost, cur_map, accum_map] + for epoch in PASS_NUM: + map_evaluator.reset(exe) + for data in batches: + loss, cur_map_v, accum_map_v = exe.run(fetch_list=fetch) In the above example: @@ -340,9 +381,10 @@ class DetectionMAP(Evaluator): evaluate_difficult=evaluate_difficult, ap_version=ap_version) - self.create_state(dtype='int32', shape=None, suffix='accum_pos_count') - self.create_state(dtype='float32', shape=None, suffix='accum_true_pos') - self.create_state(dtype='float32', shape=None, suffix='accum_false_pos') + self._create_state(dtype='int32', shape=None, suffix='accum_pos_count') + self._create_state(dtype='float32', shape=None, suffix='accum_true_pos') + self._create_state( + dtype='float32', shape=None, suffix='accum_false_pos') self.has_state = None var = self.helper.create_variable( diff --git a/python/paddle/fluid/executor.py b/python/paddle/fluid/executor.py index 33d8f709412b25d29c6618272500dd7b953d6645..dc275674618ee147dad2e32c7db29132ab55eb29 100644 --- a/python/paddle/fluid/executor.py +++ b/python/paddle/fluid/executor.py @@ -18,17 +18,24 @@ from framework import Program, default_main_program, Variable from . import core __all__ = [ - 'Executor', 'global_scope', 'scope_guard', 'switch_scope', 'fetch_var' + 'Executor', 'global_scope', 'scope_guard', '_switch_scope', 'fetch_var' ] g_scope = core.Scope() def global_scope(): + """ + Get the global/default scope instance. There are a lot of APIs use + :code:`global_scope` as its default value, e.g., :code:`Executor.run` + + Returns: + Scope: The global/default scope instance. + """ return g_scope -def switch_scope(scope): +def _switch_scope(scope): global g_scope ex = g_scope g_scope = scope @@ -37,12 +44,40 @@ def switch_scope(scope): @contextlib.contextmanager def scope_guard(scope): - ex = switch_scope(scope) + """ + Change the global/default scope instance by Python `with` statement. All + variable in runtime will assigned to the new scope. + + Examples: + >>> import paddle.fluid as fluid + >>> new_scope = fluid.Scope() + >>> with fluid.scope_guard(new_scope): + >>> ... + + Args: + scope: The new global/default scope. + """ + ex = _switch_scope(scope) yield - switch_scope(ex) + _switch_scope(ex) def as_numpy(tensor): + """ + Convert a Tensor to a numpy.ndarray, its only support Tensor without LoD information. + For higher dimensional sequence data, please use LoDTensor directly. + Examples: + >>> import paddle.fluid as fluid + >>> outs = executor.run(...) + >>> np_outs = map(lambda x: as_numpy(x), outs) + >>> ... + + Args: + tensor(Variable): a instance of Tensor + + Returns: + numpy.ndarray + """ if isinstance(tensor, list): return [as_numpy(t) for t in tensor] assert isinstance(tensor, core.LoDTensor) @@ -135,14 +170,18 @@ def has_fetch_operators(block, fetch_targets, fetch_holder_name): def fetch_var(name, scope=None, return_numpy=True): """ - Fetch the value of the variable with the given name from the given scope + Fetch the value of the variable with the given name from the + given scope. + Args: name(str): name of the variable. Typically, only persistable variables can be found in the scope used for running the program. scope(core.Scope|None): scope object. It should be the scope where you pass to Executor.run() when running your program. - If None, global_scope() will be used. - return_numpy(bool): whether convert the tensor to numpy.ndarray + If None, global_scope() will be used. Default None. + return_numpy(bool): whether convert the tensor to numpy.ndarray. + Default True. + Returns: LodTensor|numpy.ndarray """ @@ -162,7 +201,7 @@ def fetch_var(name, scope=None, return_numpy=True): return tensor -def get_program_cache_key(feed, fetch_list): +def _get_program_cache_key(feed, fetch_list): feed_var_names = feed.keys() def to_name_str(var): @@ -181,6 +220,25 @@ def get_program_cache_key(feed, fetch_list): class Executor(object): + """ + An Executor in Python, only support the single-GPU running. For multi-cards, please refer to + ParallelExecutor. + Python executor takes a program, add feed operators and fetch operators to this program according + to feed map and fetch_list. Feed map provides input data for the program. fetch_list provides + the variables(or names) that user want to get after program run. Note: the executor will run all + operators in the program but not only the operators dependent by the fetch_list. + It store the global variables into the global scope, and create a local scope for the temporary + variables. The local scope contents will be discarded after every minibatch forward/backward finished. + But the global scope variables will be persistent through different runs. + All of ops in program will be running in sequence. + + Args: + place(core.CPUPlace|core.CUDAPlace(n)): indicate the executor run on which device + + Note: For debugging complicated network in parallel-GPUs, you can test it on the executor. + They has the exactly same arguments, and expected the same results. + """ + def __init__(self, place): self.place = place p = core.Place() @@ -189,6 +247,23 @@ class Executor(object): self.program_caches = dict() def as_lodtensor(self, data): + """ + Convert numpy.ndarray to Tensor, its only support Tensor without LoD information. + For higher dimensional sequence data, please use LoDTensor directly. + + Examples: + >>> import paddle.fluid as fluid + >>> exe = fluid.executor(fluid.CPUPlace()) + >>> data = np.array(size=(100, 200, 300)) + >>> np_outs = map(lambda x: exe.as_lodtensor(x), data) + >>> ... + + Args: + data(numpy.ndarray): a instance of array + + Returns: + LoDTensor + """ if isinstance(data, list): raise RuntimeError("Some of your feed data hold LoD information. \ They can not be completely cast from a list of Python \ @@ -280,23 +355,47 @@ class Executor(object): scope=None, return_numpy=True, use_program_cache=False): - """ Run program by this Executor. Feed data by feed map, fetch result by fetch_list. - + """ + Run program by this Executor. Feed data by feed map, fetch result by fetch_list. Python executor takes a program, add feed operators and fetch operators to this program according to feed map and fetch_list. Feed map provides input data for the program. fetch_list provides - the variables(or names) that user want to get after program run. Note: the executor will run all + the variables(or names) that user want to get after program run. + + Note: the executor will run all operators in the program but not only the operators dependent by the fetch_list - :param program: the program that need to run, if not provied, then default_main_program will be used. - :param feed: feed variable map, e.g. {"image": ImageData, "label": LableData} - :param fetch_list: a list of variable or variable names that user want to get, run will return them according - to this list. - :param feed_var_name: the name for the input variable of feed Operator. - :param fetch_var_name: the name for the output variable of feed Operator. - :param scope: the scope used to run this program, you can switch it to different scope. default is global_scope - :param return_numpy: if convert the fetched tensor to numpy - :param use_program_cache: set use_program_cache to true if program not changed compare to the last step. - :return: result according to fetch_list. + Args: + program(Program): the program that need to run, if not provied, then default_main_program will be used. + feed(dict): feed variable map, e.g. {"image": ImageData, "label": LableData} + fetch_list(list): a list of variable or variable names that user want to get, run will return them according to this list. + feed_var_name(str): the name for the input variable of feed Operator. + fetch_var_name(str): the name for the output variable of fetch Operator. + scope(Scope): the scope used to run this program, you can switch it to different scope. default is global_scope + return_numpy(bool): if convert the fetched tensor to numpy + use_program_cache(bool): set use_program_cache to true if program not changed compare to the last step. + + Returns: + + list(numpy.array): fetch result according to fetch_list. + + + Examples: + + >>> data = layers.data(name='X', shape=[1], dtype='float32') + >>> hidden = layers.fc(input=data, size=10) + >>> layers.assign(hidden, out) + >>> loss = layers.mean(out) + >>> adam = fluid.optimizer.Adam() + >>> adam.minimize(loss) + + >>> cpu = core.CPUPlace() + >>> exe = Executor(cpu) + >>> exe.run(default_startup_program()) + + >>> x = numpy.random.random(size=(10, 1)).astype('float32') + >>> outs = exe.run( + >>> feed={'X': x}, + >>> fetch_list=[loss.name]) """ if feed is None: feed = {} @@ -317,7 +416,7 @@ class Executor(object): if scope is None: scope = global_scope() - cache_key = get_program_cache_key(feed, fetch_list) + cache_key = _get_program_cache_key(feed, fetch_list) if use_program_cache: cached_program = self._get_program_cache(cache_key) if cached_program is None: diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index df0625649d2cf897e103131739aaa4d48f8a097c..4c1c8443a641cde40c392f1c647bc78d6cd3c13c 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -30,8 +30,6 @@ __all__ = [ 'default_startup_program', 'default_main_program', 'program_guard', - 'switch_startup_program', - 'switch_main_program', 'get_var', ] @@ -43,7 +41,8 @@ ZERO_VAR_SUFFIX = core.kZeroVarSuffix() def grad_var_name(var_name): """ - return gradient name for a certain var name + Returns: + str: gradient name for a certain var name """ return var_name + GRAD_VAR_SUFFIX @@ -51,10 +50,12 @@ def grad_var_name(var_name): def convert_np_dtype_to_dtype_(np_dtype): """ Convert the data type in numpy to the data type in Paddle + Args: - np_dtype(np.dtype): the data type in numpy + np_dtype(np.dtype): the data type in numpy. - Returns(core.VarDesc.VarType): the data type in Paddle + Returns: + core.VarDesc.VarType: the data type in Paddle. """ dtype = np.dtype(np_dtype) @@ -120,37 +121,53 @@ def _debug_string_(proto, throw_on_error=True): class Variable(object): """ - Python variable. Every input and output of an operator is a variable. Every - variable belongs to a block. The variable has a name and two variables in - different blocks could have the same name. + In Fluid, every input and output of an operator is a variable. In most + cases, variables are used for holding different kinds of data or training + labels. A variable belongs to a block. All variable has its own name and + two variables in different blocks could have the same name. - There are many kinds of variables. Please reference the framework.proto for - details. + There are many kinds of variables. Each kind of them has its own attributes + and usages. Please reference the framework.proto for details. - Notes: The constructor of Variable should not be invoked directly. Please - use `Block.create_var` to create a variable. - - >>> cur_program = Program() - >>> cur_block = cur_program.current_block() - >>> new_variable = cur_block.create_var( - >>> name="X", shape=[-1, 23, 48], dtype='float32') + Most of a Variable's member variables can be setted to be None. It mean + it is not available or will be specified later. Args: - block(Block): The associated block. It will be passed by - `Block.create_var` automatically. + block(Block): The block that the variable belongs to. type(core.VarDesc.VarType): Variable type. Please reference the framework.proto for details. - shape(tuple|list|None): The shape of variable. -1 means the batch size. + name(str|None): The name of the variable. If setted None, it will be + generated automatically. Default: None + shape(tuple|list|None): The shape of the variable. -1 means the batch size. Some kinds of variable do not contain shape, just set it to None. - dtype(np.dtype|core.VarDesc.VarType|str): The data type of variable. - lod_level(int): The level of lod tensor. 0 means it is not a time + Default: None + dtype(np.dtype|core.VarDesc.VarType|str|None): The data type of variable. + Default: None + lod_level (int|None): The level of lod tensor. 0 means it is not a time series data. - capacity(int): The capacity of Channel variable. Ignored - for other types. - persistable(bool): True if the variable should be saved as check point. - Defaults to False. - stop_gradient(bool): True if the variable will stop to calculate - gradients when backward. Defaults to False. + Default: None + capacity (int|None): The capacity of Channel variable. Ignored for other + types. Default: None + persistable (bool|None): True if the variable is persistable. A persistable + variable will not be deleted after an iteration ending. Defaults: None. + error_clip (BaseErrorClipAttr|None): The error clip attributes of the + corresponding gradient variable. Default: None + stop_gradient (bool): True if the variable will stop to calculate its + gradients when backward. Default: False. + is_data (bool): True if the variable is an input data. Default: False + + Notes: + The constructor of Variable should not be invoked directly. Please + use `Block.create_var` to create a variable. + + Examples: + .. code-block:: python + + cur_program = Program() + cur_block = cur_program.current_block() + new_variable = cur_block.create_var(name="X", + shape=[-1, 23, 48], + dtype='float32') """ def __init__(self, @@ -253,13 +270,14 @@ class Variable(object): Get debug string. Args: - throw_on_error(bool): True if raise an exception when self is not - intialized. + throw_on_error(bool): True if raise an exception when self is + not initialized. with_details(bool): more details about variables and parameters - (e.g. trainable, optimize_attr, ...) will be printed when with_details is True - - Returns(str): The debug string. + (e.g. trainable, optimize_attr, ...) will be printed when + with_details is True. Default False; + Returns: + str: The debug string. """ assert isinstance(throw_on_error, bool) and isinstance(with_details, bool) @@ -276,6 +294,15 @@ class Variable(object): __repr__ = __str__ def set_desc(self, input): + """ + Set the variable description. + + Args: + input(core.VarDesc): The new VarDesc. + + Returns: + None + """ self.desc = input @property @@ -312,6 +339,15 @@ class Variable(object): return self.desc.type() def set_error_clip(self, error_clip): + """ + Set the error_clip. + + Args: + error_clip(BaseErrorClipAttr) : The new error_clip. + + Returns: + None + """ self.error_clip = error_clip @@ -319,8 +355,8 @@ def get_all_op_protos(): """ Get all registered op proto from PaddlePaddle C++ end. - Returns(list): list of OpProto - + Returns: + list: list of OpProto. """ protostrs = core.get_all_op_protos() ret_values = [] @@ -373,9 +409,45 @@ class OpProtoHolder(object): class Operator(object): """ - Python Operator class. The operator represents the build in instructions in a - Block. Users can use the build in instructions to describe their neural - network. + In Fluid, all the operation are represented by Operator, and Operator + is regarded as a build in an instruction of a Block. Users can use the + build in instructions to describe their neural network. + + Args: + block(Block): The block has the current operator. + desc(core.OpDesc): The protobuf description of Operator. + type(str): The type of operator. Default None. + inputs(dict): The input of this Operator. it is a dictionary, for every + element, key is the input parameter name, and value is a list of + variables. Default None. + outputs(dict): The output of this Operator. it is a dictionary, for + every element, key is the input parameter name, and value is a list + of variables. Default None. + attrs(dict): The attributes of this Operator. it is a dictionary, for + every element, key is attribute name, and value is the attribute value. + The attribute type should be as same as the type registered in C++ side. + Default None. + + Returns: + Operator: The initialized Operator. + + Raises: + ValueError: If the passed input, output and attrs doesn't match the + initializing Operator's that registered in C++ side. + + Notes: + The constructor of operator should not be invoked directly. Use + Block.append_op or Block.prepend_op instead. + + Examples: + .. code-block:: python + + cur_program = Program() + cur_block = cur_program.current_block() + # var1 += var2 + var3 + cur_block.append_op(type="sum", + inputs={"X": [var1, var2, var3]}, + outputs={"Out": [var1]}) """ OP_WITHOUT_KERNEL_SET = { 'feed', 'fetch', 'save', 'load', 'recurrent', 'go', @@ -392,31 +464,7 @@ class Operator(object): inputs=None, outputs=None, attrs=None): - """ - Constructor. - Notes: The constructor of operator should not be invoked directly. Use - Block.append_op or Block.prepend_op instead. - - >>> cur_program = Program() - >>> cur_block = cur_program.current_block() - >>> # var1 += var2 + var3 - >>> cur_block.append_op(type="sum", - >>> inputs={"X": [var1, var2, var3]}, - >>> outputs={"Out": [var1]}) - - Args: - block(Block): The block has the current operator. - desc(core.OpDesc): The protobuf description. - type(str): The type of operator. - inputs(dict): The input dictionary. Key is the input parameter name. - Value is a list of variables. - outputs(dict): The output dictionary which has the same format with - inputs. - attrs(dict): The attributes dictionary. Key is attribute name. Value - is the attribute value. The attribute type should be as same as - the type registered in C++ - """ self.block = block self.desc = desc self.attrs = attrs @@ -510,15 +558,20 @@ class Operator(object): if (attr_name not in self.attrs) or ( self.attrs[attr_name] is None): continue - if isinstance(self.attrs[attr_name], Block): + attr_val = self.attrs[attr_name] + if isinstance(attr_val, Block): self.desc.set_block_attr(attr_name, self.attrs[attr_name].desc) - elif isinstance(self.attrs[attr_name], core.BlockDesc) or \ - isinstance(self.attrs[attr_name], core.ProgramDesc): + elif isinstance(attr_val, list) and attr_val and \ + all(isinstance(v, Block) for v in attr_val): + self.desc.set_blocks_attr(attr_name, + [v.desc for v in attr_val]) + elif isinstance(attr_val, core.BlockDesc) or \ + isinstance(attr_val, core.ProgramDesc): self.desc.set_serialized_attr( - attr_name, self.attrs[attr_name].serialize_to_string()) + attr_name, attr_val.serialize_to_string()) else: - self.desc.set_attr(attr_name, self.attrs[attr_name]) + self.desc.set_attr(attr_name, attr_val) self.desc.check_attrs() if self.has_kernel(type): self.desc.infer_var_type(self.block.desc) @@ -529,12 +582,14 @@ class Operator(object): def to_string(self, throw_on_error): """ - To debug string. + Get debug string. + Args: - throw_on_error(bool): raise exception when self is not initialized - when throw_on_error is True + throw_on_error(bool): Whether to raise exception if self is not + initialized. - Returns(str): The debug string. + Returns: + str: The debug string. """ protostr = self.desc.serialize_to_string() @@ -552,29 +607,45 @@ class Operator(object): def input(self, name): """ - Get input arguments by the input parameter name - Args: - name(str): The input parameter name + Get the input arguments according to the input parameter name. - Returns(list): return the list of argument names associated with the - specific parameter name. + Args: + name(str): The input parameter name. + Returns: + list: return the list of argument names that associated with \ + the specific parameter name. """ return self.desc.input(name) def rename_input(self, old_name, new_name): + """ + Rename the `old_name` to `new_name`. + + Args: + old_name(str): The old name of the Operator's input. + new_name(str): The new name of the Operator's input. + + Returns: + None + """ self.desc.rename_input(old_name, new_name) def rename_output(self, old_name, new_name): + """ + Rename the `old_name` to `new_name`. + + Args: + old_name(str): The old name of the Operator's output. + new_name(str): The new name of the Operator's output. + + Returns: + None + """ self.desc.rename_output(old_name, new_name) @property def input_names(self): - """ - Get all input parameter names - Returns(list): return a list of input parameter names - - """ return self.desc.input_names() @property @@ -587,33 +658,23 @@ class Operator(object): def output(self, name): """ - Get output arguments by the output parameter name - Args: - name(str): The output parameter name + Get output arguments by the output parameter name. - Returns(list): return the list of argument names associated with the - specific parameter name. + Args: + name(str): The output parameter name. + Returns: + list: return the list of argument names associated with \ + the specific parameter name. """ return self.desc.output(name) @property def output_names(self): - """ - Get all output parameter names - Returns(list): return a list of output parameter names - - """ return self.desc.output_names() @property def idx(self): - """ - Return the array index of current operator. - Returns(int): The array index in block.ops array - Raises: - ValueError: when the operator is not found. - """ for i, op in enumerate(self.block.ops): if op == self: return i @@ -622,66 +683,87 @@ class Operator(object): def has_attr(self, name): """ - operator has the attribute with name or not. + Whether this Operator has the attribute with name or not. + Args: - name(str): the attribute name + name(str): the attribute name. - Returns(bool): True if has this attribute. + Returns: + bool: True if has this attribute. """ return self.desc.has_attr(name) def attr_type(self, name): """ - Get the type of attribute by attribute name - Args: - name(str): the attribute name + Get the type of attribute by attribute's name. - Returns(core.AttrType): the attribute type + Args: + name(str): the attribute name. + Returns: + core.AttrType: the attribute type. """ return self.desc.attr_type(name) def set_attr(self, name, val): + """ + Set the value of attribute by attribute's name. + + Args: + name(str): the attribute name. + val(bool|int|str|float|list): the value of the attribute. + + Raises: + ValueError: If the type of value doesn't match with desc.attr_type(name). + """ self.attrs[name] = val - self.desc.set_attr(name, val) + if isinstance(val, Block): + self.desc.set_block_attr(name, val.desc) + elif isinstance(val, list) and val and all( + isinstance(v, Block) for v in val): + self.desc.set_blocks_attr(name, [v.desc for v in val]) + elif isinstance(val, core.BlockDesc) or \ + isinstance(val, core.ProgramDesc): + self.desc.set_serialized_attr(name, val.serialize_to_string()) + else: + self.desc.set_attr(name, val) @property def attr_names(self): - """ - Get all attribute names - Returns(list): The list of attribute name - - """ return self.desc.attr_names() def attr(self, name): """ - Get attribute by name + Get the attribute by name. + Args: - name(str): the attribute name + name(str): the attribute name. - Returns(bool|int|str|float|list): The attribute value. The return value + Returns: + bool|int|str|float|list: The attribute value. The return value can be any valid attribute type. - """ return self.desc.attr(name) def block_attr(self, name): """ - Get the block attribute by name - Args: - name(str): the attribute name + Get the block attribute by name. - Returns(int): the block index + Args: + name(str): the attribute name. + Returns: + int: the block index. """ return self.desc.block_attr(name) def all_attrs(self): """ - Get the attribute dict - Returns(dict): The Operator's attribute dict + Get the attribute dict. + + Returns: + dict: The Operator's attribute dict. """ attr_names = self.attr_names attr_map = {} @@ -694,6 +776,35 @@ class Operator(object): class Block(object): + """ + In Fluid, a Program is consistence of multi-Block, and Block stores + VarDesc and OpDesc. In a specific Block, a VarDesc have a unique name. + One block could have some child blocks, and child block's name scopes + should inherit the parent's so that OpDesc in child block can reference + a VarDesc that is stored in the parent block. + Please reference the framework.proto for details. + + Args: + program(Program): The Program that the Block belongs to. + idx(int): The block's id in the Program. + + Notes: + The constructor of Block should not be invoked directly. Please + use `Program.create_block()` to create a block. + + Examples: + .. code-block:: python + + cur_program = Program() + cur_block = cur_program.current_block() + var = cur_block.create_var(name="X", + shape=[-1, 23, 48], + dtype='float32') + cur_block.append_op(type="abs", + inputs={"X": [var]}, + outputs={"Out": [var]}) + """ + def __init__(self, program, idx): self.desc = program.desc.block(idx) self.vars = collections.OrderedDict() # var_name --> var @@ -706,15 +817,17 @@ class Block(object): def to_string(self, throw_on_error, with_details=False): """ - To debug string. + Get debug string. + Args: throw_on_error(bool): raise exception when self is not initialized - when throw_on_error is True + when throw_on_error is True. with_details(bool): more details about variables and parameters - (e.g. trainable, optimize_attr, ...) will be printed when with_details is True - - Returns(str): The debug string. + (e.g. trainable, optimize_attr, ...) will be printed when + with_details is True. Default False. + Returns: + str: The debug string. """ assert isinstance(throw_on_error, bool) and isinstance(with_details, bool) @@ -746,6 +859,15 @@ class Block(object): return self.desc.get_forward_block_idx() def set_forward_block_idx(self, idx): + """ + Set the forward block Idx. + + Args: + idx(int): the block index. + + Returns: + None + """ self.desc.set_forward_block_idx(idx) @property @@ -753,6 +875,19 @@ class Block(object): return self.desc.id def var(self, name): + """ + Get a Variable by name from this block. + + Args: + name(str): the Variable's name. + + Raises: + ValueError: The If input's type is not str, or this block + doesn't have a Variable with the giving name. + + Returns: + Variable: the Variable with the giving name. + """ if not isinstance(name, basestring): raise TypeError( "var require string as parameter, but get %s instead." % @@ -763,6 +898,19 @@ class Block(object): return v def var_recursive(self, name): + """ + Get a Variable by name from this block recursively. + + Args: + name(str): the Variable's name. + + Raises: + ValueError: this block and this parent block doesn't + have a Variable with the giving name. + + Returns: + Variable: the Variable with the giving name. + """ frontier = list() visited = set() @@ -809,6 +957,18 @@ class Block(object): def rename_var(self, name, new_name): """ Rename variable in vars and ops' inputs and outputs + + Args: + name(str): the name that need to be renamed. + new_name(str): the name that need to rename to. + + Raises: + ValueError: If this block doesn't have this the giving name, + or the type of the var with the giving name is not Parameter + or Variable. + + Returns: + Variable: the Variable with the giving name. """ if not self.has_var(name): raise ValueError("var %s is not in current block" % name) @@ -872,12 +1032,27 @@ class Block(object): return param def append_op(self, *args, **kwargs): + """ + Appends a new Operator according to the giving arguments. + + Returns: + Operator: the append Operator. + """ op_desc = self.desc.append_op() op = Operator(block=self, desc=op_desc, *args, **kwargs) self.ops.append(op) return op def insert_op(self, index, *args, **kwargs): + """ + Insert a Operator according to the giving arguments. + + Args: + index(int): the place that the operator to insert. + + Returns: + Operator: the insert Operator. + """ self.sync_with_cpp() op_desc = self.desc.insert_op(index) op = Operator(block=self, desc=op_desc, *args, **kwargs) @@ -885,11 +1060,30 @@ class Block(object): return op def remove_op(self, index): + """ + Remove the specific position operator. + + Args: + index(int): the position that the operator to insert. + + Returns: + None + """ self.sync_with_cpp() self.desc.remove_op(index, index + 1) del self.ops[index] def slice_ops(self, start, end): + """ + Return the Operator between start and end. + + Args: + start(int): the start position. + end(int): the end position. + + Returns: + list: the Operators between start and end. + """ return self.ops[start:end] def prepend_op(self, *args, **kwargs): @@ -900,9 +1094,8 @@ class Block(object): def sync_with_cpp(self): """ - Sync from the desc on the c++ end. - - This method is used to synchronize the c++ desc instance generated by backward. + Sync from the desc on the c++ end. This method is used to synchronize + the c++ desc instance generated by backward. """ # sync variables from cpp for var in self.desc.all_vars(): @@ -967,9 +1160,14 @@ class Block(object): def copy_param_info_from(self, other): """ - Copy the information of parameters from the other block + Copy the information of parameters from the other block. + Args: - other(Block): the other block + other(Block): the other block. + + Raises: + ValueError: If type of input is not Block, or the `other` and this + block is not in the same topology. Returns: None @@ -1001,11 +1199,12 @@ class Block(object): def clone_variable(self, var): """ Clone a variable into current block. + Args: var: the variable to be cloned. Returns: - The new variable cloned from 'var' in current block. + Variable: the new variable cloned from 'var' in current block. """ assert isinstance(var, Variable) ret_var = None @@ -1045,23 +1244,18 @@ class Program(object): Notes: we have default_startup_program and default_main_program by default, a pair of them will shared the parameters. The default_startup_program only run once to initialize parameters, - default_main_program run in every minibatch and adjust the weights. - - Args: - None + default_main_program run in every mini batch and adjust the weights. Returns: - Python Program + A empty program. Examples: - .. code-block:: python - - main_program = Program() - startup_program = Program() - with fluid.program_guard(main_program=main_program, startup_program=startup_program): - fluid.layers.data(name="x", shape=[-1, 784], dtype='float32') - fluid.layers.data(name="y", shape=[-1, 1], dtype='int32') - fluid.layers.fc(name="fc", shape=[10], dtype='float32', act="relu") + >>> main_program = fluid.Program() + >>> startup_program = fluid.Program() + >>> with fluid.program_guard(main_program=main_program, startup_program=startup_program): + >>> fluid.layers.data(name="x", shape=[-1, 784], dtype='float32') + >>> fluid.layers.data(name="y", shape=[-1, 1], dtype='int32') + >>> fluid.layers.fc(name="fc", shape=[10], dtype='float32', act="relu") """ @@ -1075,6 +1269,19 @@ class Program(object): @property def op_role(self): + """ + The operator role. In a enum {Forward, Backward, Optimize}. + + Notes: this is a low level API. It is used only for ParallelExecutor to + duplicate or schedule operator to devices. + + For example, the forward operator should be executed on every device. + The backward operator should be executed on every device and the + parameter gradient of backward (use :code:`op_role_var` to get this + variable) operator should be merged to one device. The optimization + operators should be executed on only one device and broadcast the + optimization result, i.e., the new parameter, to every other device. + """ return self._current_role @op_role.setter @@ -1083,6 +1290,13 @@ class Program(object): @property def op_role_var(self): + """ + The auxiliary variables for :code:`op_role` property. + + See Also: :code:`Program.op_role`'s documentation for details. + + Notes: This is a very low-level API. Users should not use it directly. + """ return self._op_role_var @op_role_var.setter @@ -1091,6 +1305,21 @@ class Program(object): @contextlib.contextmanager def optimized_guard(self, var): + """ + A with guard to set :code:`Optimization` :code:`OpRole` and + :code:`OpRoleVar` automatically. + + Notes: This is a very low level API. Users should not use it directly. + + Args: + var(Variable|str): The variable (name) to be optimized. + + Examples: + + >>> p, g = backward(...) + >>> with program.optimized_guard(p): + >>> p = p - 0.001 * g + """ OpRole = core.op_proto_and_checker_maker.OpRole self._current_role = OpRole.Optimize self._op_role_var = [var.name if isinstance(var, Variable) else var] @@ -1099,18 +1328,35 @@ class Program(object): self._current_role = OpRole.Forward def __str__(self): + """ + Get the protobuf debug string of this Program. + + Returns: + (str): The protobuf debug string. + + Raises: + ValueError: If any of required fields is not set. + """ return self.to_string(True) def to_string(self, throw_on_error, with_details=False): """ To debug string. + Args: - throw_on_error(bool): raise exception when self is not initialized - when throw_on_error is True - with_details(bool): more details about variables and parameters - (e.g. trainable, optimize_attr, ...) will be printed when with_details is True + throw_on_error(bool): raise Value error when any of required fields + is not set. - Returns(str): The debug string. + with_details(bool): True if more details about variables and + parameters, e.g., :code:`trainable`, :code:`optimize_attr`, need + to print. + + Returns + (str): The debug string. + + Raises: + ValueError: If any of required fields is not set and throw_on_error is + True. """ assert isinstance(throw_on_error, bool) and isinstance(with_details, @@ -1126,25 +1372,93 @@ class Program(object): return res_str def get_desc(self): + """ + Get the C++ side of `ProgramDesc` object pointer. The C++ object is + exposed by :code:`pybind`. + + Notes: This is a very low level API. Users should not use this API + directly. + """ return self.desc def clone(self, for_test=False): - """Clone the Program object - Args: - for_test(bool): indicate whether clone for test. + """ + Create a new, duplicated program. + + + Some operators, e.g., :code:`batch_norm`, behave differently between + training and testing. They have an attribute, :code:`is_test`, to + control this behaviour. This method will change the :code:`is_test` + attribute of them to :code:`True` when :code:`for_test=True`. - Set for_test to False when we want to clone the program for training. - Set for_test to True when we want to clone the program for testing. + * Set for_test to False when we want to clone the program for training. + * Set for_test to True when we want to clone the program for testing. + + Notes: This API DOES NOT prune any operator. Use + :code:`clone(for_test=True)` before backward and optimization please. e.g. + + >>> test_program = fluid.default_main_program().clone(for_test=True) + >>> optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9) + >>> optimizer.minimize() Args: - for_test(bool): Some operators, such as batch_norm and drop_out ops, - behave differently in training and testing. If for_test is True, - the is_test attributes in these operators will be set to True for - testing purposes, otherwise, they remain unchanged. + for_test(bool): True if change the :code:`is_test` attribute of + operators to :code:`True`. Returns: - Program: The cloned Program object. - + Program: The new, duplicated Program object. + + Examples: + + 1. To clone a test program, the sample code is: + + >>> import paddle.fluid as fluid + >>> train_program = fluid.Program() + >>> startup_program = fluid.Program() + >>> with fluid.program_guard(train_program, startup_program): + >>> img = fluid.layers.data(name='image', shape=[784]) + >>> hidden = fluid.layers.fc(input=img, size=200, act='relu') + >>> hidden = fluid.layers.dropout(hidden, dropout_prob=0.5) + >>> loss = fluid.layers.cross_entropy( + >>> input=fluid.layers.fc(hidden, size=10, act='softmax'), + >>> label=fluid.layers.data(name='label', shape=[1], dtype='int64')) + >>> + >>> test_program = train_program.clone(for_test=True) + >>> + >>> sgd = fluid.optimizer.SGD(learning_rate=1e-3) + >>> with fluid.program_guard(train_program, startup_program): + >>> sgd.minimize(loss) + + 2. The :code:`clone` method can be avoid if you create program for + training and program for testing individually. + + >>> import paddle.fluid as fluid + >>> + >>> def network(is_test): + >>> img = fluid.layers.data(name='image', shape=[784]) + >>> hidden = fluid.layers.fc(input=img, size=200, act='relu') + >>> hidden = fluid.layers.dropout(hidden, dropout_prob=0.5, is_test=is_test) + >>> loss = fluid.layers.cross_entropy( + >>> input=fluid.layers.fc(hidden, size=10, act='softmax'), + >>> label=fluid.layers.data(name='label', shape=[1], dtype='int64')) + >>> return loss + >>> + >>> train_program = fluid.Program() + >>> startup_program = fluid.Program() + >>> test_program = fluid.Program() + >>> + >>> with fluid.program_guard(train_program, startup_program): + >>> with fluid.unique_name.guard(): + >>> loss = network(is_test=False) + >>> sgd = fluid.optimizer.SGD(learning_rate=1e-3) + >>> sgd.minimize(loss) + >>> + >>> # the test startup program is not used. + >>> with fluid.program_guard(test_program, fluid.Program()): + >>> with fluid.unique_name.guard(): + >>> loss = network(is_test=True) + + The two code snippets above will generate same programs. """ if for_test: p = self.inference_optimize() @@ -1159,6 +1473,21 @@ class Program(object): return p def prune(self, targets): + """ + Prune operators and variables which are not needed to generate + :code:`targets`. + + Notes: This is a very low level API. Users should not use this API + directly. This API is in flux and not stable. + + Args: + targets(list|Variable|Operator): A list of variables or operators + need to be pruned + + Returns: + Program: A new, pruned program. + + """ if not isinstance(targets, list): targets = [targets] targets_idx = [] @@ -1193,6 +1522,17 @@ class Program(object): return res def inference_optimize(self): + """ + This method will create a new program and change the :code:`is_test` + attribute of operators to :code:`True`. All the :code:`Parameter` + information will be lost. + + Notes: This API is a very low level API. Use + :code:`Program.clone(for_test=True)` instead. + + Returns: + Program: The new program. + """ # this is an alternative implement before # core.inference_optimize being fixed. res = Program() @@ -1209,6 +1549,18 @@ class Program(object): @staticmethod def parse_from_string(binary_str): + """ + Deserialize a program desc from protobuf binary string. + + Notes: All information about parameters will be lost after serialization + and deserialization. + + Args: + binary_str(str): The binary prootbuf string. + + Returns: + Program: A deserialized program desc. + """ p = Program() p.desc = core.ProgramDesc(binary_str) p.blocks = [Block(p, i) for i in xrange(p.desc.num_blocks())] @@ -1217,10 +1569,19 @@ class Program(object): @property def random_seed(self): + """ + The default random seed for random operators in Program. Zero means get + the random seed from random device. + + Notes: It must be set before the operators have been added. + """ return self._seed @property def num_blocks(self): + """ + The number of blocks in this program. + """ return self.desc.num_blocks() @random_seed.setter @@ -1233,15 +1594,40 @@ class Program(object): return str(self) def global_block(self): + """ + Get the first block of this program. + """ return self.blocks[0] def block(self, index): + """ + Get the :code:`index` block of this program + Args: + index(int): The index of block to get + + Returns: + Block: The :code:`index` block + """ return self.blocks[index] def current_block(self): + """ + Get the current block. The :code:`current` block is the block to append + operators. + """ return self.blocks[self.current_block_idx] def create_block(self, parent_idx=None): + """ + Create a new block with the :code:`parent_idx` and change the current block + to new block. + + Args: + parent_idx(int): The parent block index. + + Returns: + Block: The new block. + """ new_block_idx = len(self.blocks) parent = self.current_block() if parent_idx is None else self.block( parent_idx) @@ -1251,9 +1637,24 @@ class Program(object): return self.current_block() def rollback(self): + """ + Exit a code block, i.e., roll back to the parent block. + Returns: + None + """ self.current_block_idx = self.current_block().parent_idx def sync_with_cpp(self): + """ + Synchronize Python instance to its binding C++ object instance. + If the program is modified in C++ space, this method should be invoked. + + Notes: This is a very low level API. Users should not invoke it + directly. + + Returns: + None + """ for block_idx in range(len(self.blocks), self.desc.num_blocks()): self.blocks.append(Block(self, block_idx)) for block in self.blocks: @@ -1263,6 +1664,9 @@ class Program(object): """ Copy the information of parameters from other program. + Notes: This is a very low level API. Users should not invoke it + directly. + Args: other(Program): Other program @@ -1282,6 +1686,9 @@ class Program(object): """ Copy the information of data variables from other program. + Notes: This is a very low level API. Users should not invoke it + directly. + Args: other(Program): Other program @@ -1300,12 +1707,41 @@ class Program(object): self.global_block().var(var.name).is_data = True def list_vars(self): + """ + Get all variables from this Program. A iterable object is returned. + + Returns: + iterable: The generator will yield every variable in this program. + """ for each_block in self.blocks: for each_var in each_block.vars.itervalues(): yield each_var class Parameter(Variable): + """ + Parameter is derived from Variable. A parameter is a persistable + Variable, and will be updated by optimizers after each iteration. + The training of a neural network is essentially the updating of + its parameters. + + Relative to a general Variable, a Parameter has several its own + member variables: + + Args: + trainable(bool): True if the parameter need to be updated after + iterations. + optimize_attr(map): Parameter attributes related with optimizing. + Currently, it only contains 'learning_rate'. + Default: {'learning_rate': 1.0} + regularizer(WeightDecayRegularizer): The Regularizer which will + be applied on the parameter. Default: None + gradient_clip_attr(BaseGradientClipAttr): The gradint clip strategy + which will be applied on the parameter. Default: None + do_model_average(bool): True if the model average strategy will + be applied on this parameter. + """ + def __init__(self, block, shape, dtype, **kwargs): if shape is None or dtype is None: raise ValueError("Parameter must set shape and dtype") @@ -1368,8 +1804,15 @@ _startup_program_ = Program() def default_startup_program(): """ - Get default startup program. In startup program, Paddle will initialize - parameters, initialize nccl handle, etc. + Get default/global startup program. + + The layer function in :code:`fluid.layers` will create parameters, readers, + NCCL handles as global variables. The :code:`startup_program` will + initialize them by the operators in startup program. The layer function will + append these initialization operators into startup program. + + This method will return the :code:`default` or the :code:`current` startup + program. Users can use :code:`fluid.program_guard` to switch program. Returns: Program: startup program @@ -1379,7 +1822,15 @@ def default_startup_program(): def default_main_program(): """ - Get default main program. The main program is used for training or testing. + Get default/global main program. The main program is used for training or + testing. + + All layer function in :code:`fluid.layers` will append operators and + variables to the :code:`default_main_program`. + + The :code:`default_main_program` is the default program in a lot of APIs. + For example, the :code:`Executor.run()` will execute the + :code:`default_main_program` when the program is not specified. Returns: Program: main program @@ -1421,20 +1872,34 @@ def switch_startup_program(program): @contextlib.contextmanager def program_guard(main_program, startup_program=None): """ - Switch program with `with` statement + Change the global main program and startup program with `with` statement. + Layer functions in the Python `with` block will append operators and + variables to the new main programs. Examples: - >>> with program_guard(Program()): - >>> data = fluid.layers.data(...) - >>> hidden = fluid.layers.fc(...) + + >>> import paddle.fluid as fluid + >>> main_program = fluid.Program() + >>> startup_program = fluid.Program() + >>> with fluid.program_guard(main_program, startup_program): + >>> data = fluid.layers.data(...) + >>> hidden = fluid.layers.fc(...) + + Notes: The temporary :code:`Program` can be used if the user does not need + to construct either of startup program or main program. + + Examples: + + >>> import paddle.fluid as fluid + >>> main_program = fluid.Program() + >>> # does not care about startup program. Just pass a temporary value. + >>> with fluid.program_guard(main_program, fluid.Program()): + >>> data = ... Args: - main_program(Program): New main program inside `with` statement + main_program(Program): New main program inside `with` statement. startup_program(Program): New startup program inside `with` statement. None means do not change startup program. - - Returns: - None """ if not isinstance(main_program, Program): raise TypeError("main_program should be Program") @@ -1451,7 +1916,8 @@ def program_guard(main_program, startup_program=None): def get_var(name, program=None): """ - Get a variable by name from the global block of a program + Get a variable by name from the global block of a program. + Args: name(str): name of the variable program(Program|None): program object. diff --git a/python/paddle/fluid/inferencer.py b/python/paddle/fluid/inferencer.py index 6baac00905713594acd59bb3819038576fab0674..a81e39695b78f235d6ae896d90117dd392692634 100644 --- a/python/paddle/fluid/inferencer.py +++ b/python/paddle/fluid/inferencer.py @@ -27,13 +27,30 @@ __all__ = ['Inferencer', ] class Inferencer(object): + """ + Inferencer High Level API. + + Args: + infer_func (Python func): Infer function that will return predict Variable + param_path (str): The path where the inference model is saved by fluid.io.save_params + place (Place): place to do the inference + parallel (bool): use parallel_executor to run the inference, it will use multi CPU/GPU. + + Examples: + .. code-block:: python + + 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 + + place = fluid.CPUPlace() + inferencer = fluid.Inferencer( + infer_func=inference_program, param_path="/tmp/model", place=place) + + """ + def __init__(self, infer_func, param_path, place=None, parallel=False): - """ - :param infer_func: a function that will return predict Variable - :param param_path: the path where the inference model is saved by fluid.io.save_params - :param place: place to do the inference - :param parallel: use parallel_executor to run the inference, it will use multi CPU/GPU. - """ self.param_path = param_path self.scope = core.Scope() self.parallel = parallel @@ -60,9 +77,20 @@ class Inferencer(object): def infer(self, inputs, return_numpy=True): """ - :param inputs: a map of {"input_name": input_var} that will be feed into the inference program - to get the predict value - :return: the predict value of the inference model + Do Inference for Inputs + + Args: + inputs (map): a map of {"input_name": input_var} that will be feed into the inference program + return_numpy (bool): transform return value into numpy or not + + Returns: + Tensor or Numpy: the predict value of the inference model for the inputs + + Examples: + .. code-block:: python + + tensor_x = numpy.random.uniform(0, 10, [batch_size, 13]).astype("float32") + results = inferencer.infer({'x': tensor_x}) """ if not isinstance(inputs, dict): raise ValueError( diff --git a/python/paddle/fluid/initializer.py b/python/paddle/fluid/initializer.py index c36ad324e70ccf0c7ca40c6921fcc650e97e8b87..373e9c060de1ee27c165ccd2380cd8c38612c4d9 100644 --- a/python/paddle/fluid/initializer.py +++ b/python/paddle/fluid/initializer.py @@ -19,26 +19,39 @@ from framework import convert_np_dtype_to_dtype_ from core import VarDesc __all__ = [ - 'Constant', 'Uniform', 'Normal', 'Xavier', 'Bilinear', 'force_init_on_cpu', - 'init_on_cpu', 'ConstantInitializer', 'UniformInitializer', - 'NormalInitializer', 'XavierInitializer', 'BilinearInitializer' + 'Constant', 'Uniform', 'Normal', 'Xavier', 'Bilinear', 'MSRA', + 'force_init_on_cpu', 'init_on_cpu', 'ConstantInitializer', + 'UniformInitializer', 'NormalInitializer', 'XavierInitializer', + 'BilinearInitializer', 'MSRAInitializer' ] _force_init_on_cpu_ = False def force_init_on_cpu(): + """ + The flag of whether force to init variables on CPU. + + Examples: + .. code-block:: python + + if force_init_on_cpu(): + pass + + """ return _force_init_on_cpu_ @contextlib.contextmanager def init_on_cpu(): """ - Switch program with `with` statement + Force the variable to be inited on CPU. Examples: - >>> with init_on_cpu(): - >>> step = layers.create_global_var() + .. code-block:: python + + with init_on_cpu(): + step = layers.create_global_var() """ global _force_init_on_cpu_ @@ -104,14 +117,18 @@ class Initializer(object): class ConstantInitializer(Initializer): """Implements the constant initializer + + Args: + value (float): constant value to initialize the variable + + Examples: + .. code-block:: python + + fc = fluid.layers.fc(input=x, size=10, + param_attr=fluid.initializer.Constant(value=2.0)) """ def __init__(self, value=0.0, force_cpu=False): - """Constructor for ConstantInitializer - - Args: - value: constant value to initialize the variable - """ assert value is not None super(ConstantInitializer, self).__init__() self._value = value @@ -146,16 +163,20 @@ class ConstantInitializer(Initializer): class UniformInitializer(Initializer): """Implements the random uniform distribution initializer + + Args: + low (float): lower boundary of the uniform distribution + high (float): upper boundary of the uniform distribution + seed (int): random seed + + Examples: + .. code-block:: python + + fc = fluid.layers.fc(input=x, size=10, + param_attr=fluid.initializer.Uniform(low=-0.5, high=0.5)) """ def __init__(self, low=-1.0, high=1.0, seed=0): - """Constructor for UniformInitializer - - Args: - low: lower boundary of the uniform distribution - high: upper boundary of the uniform distribution - seed: random seed - """ assert low is not None assert high is not None assert high >= low @@ -196,17 +217,21 @@ class UniformInitializer(Initializer): class NormalInitializer(Initializer): - """Implements the random Normal(Gaussian) distribution initializer + """Implements the Random Normal(Gaussian) distribution initializer + + Args: + loc (float): mean of the normal distribution + scale (float): standard deviation of the normal distribution + seed (int): random seed + + Examples: + .. code-block:: python + + fc = fluid.layers.fc(input=x, size=10, + param_attr=fluid.initializer.Normal(loc=0.0, scale=2.0)) """ def __init__(self, loc=0.0, scale=1.0, seed=0): - """Constructor for NormalInitializer - - Args: - loc: mean of the normal distribution - scale: standard deviation of the normal distribution - seed: random seed - """ assert loc is not None assert scale is not None assert seed is not None @@ -246,39 +271,49 @@ class NormalInitializer(Initializer): class XavierInitializer(Initializer): - """Implements the Xavier initializer - + """ This class implements the Xavier weight initializer from the paper - Understanding the difficulty of training deep feedforward neural - networks[1] by Xavier Glorot and Yoshua Bengio. + `Understanding the difficulty of training deep feedforward neural + networks `_ + by Xavier Glorot and Yoshua Bengio. This initializer is designed to keep the scale of the gradients approximately same in all the layers. In case of Uniform distribution, - the range is [-x, x], where x = sqrt(6 / (fan_in + fan_out)). + the range is [-x, x], where + + .. math:: + + x = \sqrt{\\frac{6.0}{fan\_in + fan\_out}} + In case of Normal distribution, the mean is 0 and the standard deviation - is sqrt(2/ (fan_in + fan_out)). + is + + .. math:: + + \sqrt{\\frac{2.0}{fan\_in + fan\_out}} + + + Args: + uniform (bool): whether to use uniform or normal distribution + fan_in (float): fan_in for Xavier initialization. If None, it is + inferred from the variable. + fan_out (float): fan_out for Xavier initialization. If None, it is + inferred from the variable. + seed (int): random seed + + Note: + It is recommended to set fan_in and fan_out to None for most cases. + + Examples: + .. code-block:: python + + fc = fluid.layers.fc( + input=queries, size=10, + param_attr=fluid.initializer.Xavier(uniform=False)) - References: - [1] Understanding the difficulty of training deep feedforward neural - networks. International conference on artificial intelligence and - statistics. - (http://proceedings.mlr.press/v9/glorot10a.html) """ def __init__(self, uniform=True, fan_in=None, fan_out=None, seed=0): - """Constructor for XavierInitializer - - Args: - uniform: whether to use uniform or normal distribution - fan_in: fan_in for Xavier initialization. If None, it is - inferred from the variable. - fan_out: fan_out for Xavier initialization. If None, it is - inferred from the variable. - seed: random seed - - Note: It is recommended to set fan_in and fan_out to None for - most cases. - """ assert uniform is not None assert seed is not None super(XavierInitializer, self).__init__() @@ -342,30 +377,42 @@ class MSRAInitializer(Initializer): """Implements the MSRA initializer a.k.a. Kaiming Initializer This class implements the weight initialization from the paper - Delving Deep into Rectifiers: Surpassing Human-Level Performance on - ImageNet Classification[1] by Kaiming He, Xiangyu Zhang, Shaoqing Ren - and Jian Sun. This is a robust initialization method that particularly - considers the rectifier nonlinearities. In case of Uniform distribution, - the range is [-x, x], where x = sqrt(6 / fan_in). In case of Normal - distribution, the mean is 0 and the standard deviation - is sqrt(2/ fan_in). - - References: - [1] Delving Deep into Rectifiers: Surpassing Human-Level Performance - on ImageNet Classification - (https://arxiv.org/abs/1502.01852) + `Delving Deep into Rectifiers: Surpassing Human-Level Performance on + ImageNet Classification `_ + by Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun. This is a + robust initialization method that particularly considers the rectifier + nonlinearities. In case of Uniform distribution, the range is [-x, x], where + + .. math:: + + x = \sqrt{\\frac{6.0}{fan\_in}} + + In case of Normal distribution, the mean is 0 and the standard deviation + is + + .. math:: + + \sqrt{\\frac{2.0}{fan\_in}} + + Args: + uniform (bool): whether to use uniform or normal distribution + fan_in (float): fan_in for MSRAInitializer. If None, it is\ + inferred from the variable. + seed (int): random seed + + Note: + It is recommended to set fan_in to None for most cases. + + Examples: + .. code-block:: python + + fc = fluid.layers.fc( + input=queries, size=10, + param_attr=fluid.initializer.MSRA(uniform=False)) """ def __init__(self, uniform=True, fan_in=None, seed=0): """Constructor for MSRAInitializer - - Args: - uniform: whether to use uniform or normal distribution - fan_in: fan_in for MSRAInitializer. If None, it is - inferred from the variable. - seed: random seed - - Note: It is recommended to set fan_in to None for most cases. """ assert uniform is not None assert seed is not None @@ -425,34 +472,37 @@ class MSRAInitializer(Initializer): class BilinearInitializer(Initializer): - """Implements the bilinear initializer. - + """ This initializer can be used in transposed convolution operator to act as upsampling. Users can upsample a feature map with shape of (B, C, H, W) by any integer factor. The usage is: - - >>> factor = 2 - >>> w_attr = ParamAttr(learning_rate=0., regularizer=L2Decay(0.), - >>> initializer=Bilinear()) - >>> conv_up = fluid.layers.conv2d_transpose( - >>> input, - >>> num_filters=C, - >>> output_size=None, - >>> filter_size=2 * factor - factor % 2, - >>> padding=ceil((factor - 1) / 2.), - >>> stride=factor, - >>> groups=C, - >>> param_attr=w_attr, - >>> bias_attr=False) - - - Where, `num_filters=C` and `groups=C` means this is channel-wise tranposed + + Examples: + + .. code-block:: python + + factor = 2 + w_attr = ParamAttr(learning_rate=0., regularizer=L2Decay(0.), + initializer=Bilinear()) + conv_up = fluid.layers.conv2d_transpose( + input, + num_filters=C, + output_size=None, + filter_size=2 * factor - factor % 2, + padding=ceil((factor - 1) / 2.), + stride=factor, + groups=C, + param_attr=w_attr, + bias_attr=False) + + Where, `num_filters=C` and `groups=C` means this is channel-wise transposed convolution. The filter shape will be (C, 1, K, K) where K is `filer_size`, This initializer will set a (K, K) interpolation kernel for every channel of the filter identically. The resulting shape of the output feature map will be (B, C, factor * H, factor * W). Note that the learning rate and the weight decay are set to 0 in order to keep coefficient values of bilinear - interpolation unchanged during training. + interpolation unchanged during training. + """ def __init__(self): @@ -469,7 +519,7 @@ class BilinearInitializer(Initializer): be added. Returns: - the initialization op + Operator: the initialization op Raises: ValueError: If type of `var` and `block` is not right. diff --git a/python/paddle/fluid/io.py b/python/paddle/fluid/io.py index 6323c9899e0080b436a52f852c647466b8f94bc1..6e527572f1ca77be9fe069654db00d16ad5c21ef 100644 --- a/python/paddle/fluid/io.py +++ b/python/paddle/fluid/io.py @@ -30,20 +30,42 @@ __all__ = [ def is_parameter(var): - """Check whether the variable is a Parameter. - - This function checks whether the input variable is a Parameter. + """ + Check whether the given variable is an instance of Parameter. Args: - var : The input variable. + var(Variable): The variable to be checked. Returns: - boolean result whether the variable is a Parameter. + bool: True if the given `var` is an instance of Parameter, + False if not. + + Examples: + .. code-block:: python + + param = fluid.default_main_program().global_block().var('fc.w') + res = fluid.io.is_parameter(param) """ return isinstance(var, Parameter) def is_persistable(var): + """ + Check whether the given variable is persistable. + + Args: + var(Variable): The variable to be checked. + + Returns: + bool: True if the given `var` is persistable + False if not. + + Examples: + .. code-block:: python + + param = fluid.default_main_program().global_block().var('fc.w') + res = fluid.io.is_persistable(param) + """ if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \ var.desc.type() == core.VarDesc.VarType.FETCH_LIST: return False @@ -68,20 +90,69 @@ def save_vars(executor, predicate=None, filename=None): """ - Save variables to directory by executor. + Save variables to the given directory by executor. + + There are two ways to specify variables to be saved: The first way, list + variables in a list and assign it to the `vars`. The second way, assign the + `main_program` with an existing program, then all variables in the program + will be saved. The first way has a higher priority. In other words, if `vars` + are assigned, the `main_program` and the `predicate` will be ignored. - :param executor: executor that save variable - :param dirname: directory path - :param main_program: program. If vars is None, then filter all variables in this - program which fit `predicate`. Default default_main_program. - :param predicate: The Predicate describes a callable that returns a variable - as a bool. If it returns true, the corresponding input variable will be saved. - :param vars: variables need to be saved. If vars is specified, program & predicate - will be ignored - :param filename: The name of a single file that all vars are saved to. - If it is None, save variables to separate files. + The `dirname` are used to specify the folder where to save variables. + If you prefer to save variables in separate files in the folder `dirname`, + set `filename` None; if you prefer to save all variables in a single file, + use `filename` to specify it. - :return: None + Args: + executor(Executor): The executor to run for saving variables. + dirname(str): The directory path. + main_program(Program|None): The program whose variables will be saved. + If it is None, the default main program will + be used automatically. + Default: None + vars(list[Variable]|None): The list that contains all variables to save. + It has a higher priority than the `main_program`. + Default: None + predicate(function|None): If it is not None, only variables in the + `main_program` that makes predicate(variable)==True + will be saved. It only works when we are using the + `main_program` to specify variables (In other words + `vars` is None). + Default: None + filename(str|None): The file which to save all variables. If you prefer to save + variables separately, set it to None. + Default: None + + Returns: + None + + Raises: + TypeError: If `main_program` is not an instance of Program nor None. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + + # The first usage: using `main_program` to specify variables + def name_has_fc(var): + res = "fc" in var.name + return res + + prog = fluid.default_main_program() + fluid.io.save_vars(executor=exe, dirname=path, main_program=prog, + vars=None) + # All variables in `main_program` whose name includes "fc" will be saved. + # And variables are going to be saved separately. + + + # The second usage: using `vars` to specify variables + var_list = [var_a, var_b, var_c] + fluid.io.save_vars(executor=exe, dirname=path, vars=var_list, + filename="vars_file") + # var_a, var_b and var_c will be saved. And they are going to be + # saved in the same file named 'var_file' in the path "./my_paddle_model". """ if vars is None: if main_program is None: @@ -129,7 +200,42 @@ def save_vars(executor, def save_params(executor, dirname, main_program=None, filename=None): """ - Save all parameters to directory with executor. + This function filters out all parameters from the give `main_program` + and then save them to the folder `dirname` or the file `filename`. + + Use the `dirname` to specify the saving folder. If you would like to + save parameters in separate files, set `filename` None; if you would + like to save all parameters in a single file, use `filename` to specify + the file name. + + NOTICE: Some variables are not Parameter while they are necessary for + training. So you can NOT save and continue your training just by + `save_params()` and `load_params()`. Please use `save_persistables()` + and `load_persistables()` instead. + + Args: + executor(Executor): The executor to run for saving parameters. + dirname(str): The saving directory path. + main_program(Program|None): The program whose parameters will be + saved. If it is None, the default + main program will be used automatically. + Default: None + filename(str|None): The file to save all parameters. If you prefer + to save parameters in differnet files, set it + to None. + Default: None + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + fluid.io.save_params(executor=exe, dirname=param_path, + main_program=None) """ save_vars( executor, @@ -142,7 +248,37 @@ def save_params(executor, dirname, main_program=None, filename=None): def save_persistables(executor, dirname, main_program=None, filename=None): """ - Save all persistables to directory with executor. + This function filters out all variables with `persistable==True` from the + give `main_program` and then saves these variables to the folder `dirname` + or file `filename`. + + The `dirname` is used to specify the folder where persistable variables + are going to be saved. If you would like to save variables in separate + files, set `filename` None; if you would like to save all variables in a + single file, use `filename` to specify the file name. + + Args: + executor(Executor): The executor to run for saving persistable variables. + dirname(str): The directory path. + main_program(Program|None): The program whose persistbale variables will + be saved. If it is None, the default main + program will be used automatically. + Default: None + filename(str|None): The file to saved all variables. If you prefer to + save variables in differnet files, set it to None. + Default: None + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + fluid.io.save_persistables(executor=exe, dirname=param_path, + main_program=None) """ save_vars( executor, @@ -160,20 +296,69 @@ def load_vars(executor, predicate=None, filename=None): """ - Load variables from directory by executor. + Load variables from the given directory by executor. + + There are two ways to specify variables to be loaded: The first way, list + variables in a list and assign it to the `vars`. The second way, assign the + `main_program` with an existing program, then all variables in the program + will be loaded. The first way has a higher priority. In other words if `vars` + are assigned, the `main_program` and the `predicate` will be ignored. + + The `dirname` are used to specify the folder where to load variables. + If variables were saved in separate files in the folder `dirname`, + set `filename` None; if all variables were saved in a single file, + use `filename` to specify it. - :param executor: executor that load variable - :param dirname: directory path - :param main_program: program. If vars is None, then filter all variables in this - program which fit `predicate`. Default default_main_program(). - :param predicate: The Predicate describes a callable that returns a variable - as a bool. If it returns true, the corresponding input variable will be loaded. - :param vars: variables need to be loaded. If vars is specified, program & - predicate will be ignored - :param filename: The name of the single file that all vars are loaded from. - If it is None, load variables from separate files. + Args: + executor(Executor): The executor to run for loading variables. + dirname(str): The directory path. + main_program(Program|None): The program whose variables will be loaded. + If it is None, the default main program will + be used automatically. + Default: None + vars(list[Variable]|None): The list that contains all variables to load. + It has a higher priority than the `main_program`. + Default: None + predicate(function|None): If it is not None, only variables in the + `main_program` that makes predicate(variable)==True + will be loaded. It only works when we are using the + `main_program` to specify variables (In other words + `vars` is None). + Default: None + filename(str|None): The file which saved all required variables. If variables + were saved in differnet files, set it to None. + Default: None + + Returns: + None + + Raises: + TypeError: If `main_program` is not an instance of Program nor None. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + + # The first usage: using `main_program` to specify variables + def name_has_fc(var): + res = "fc" in var.name + return res - :return: None + prog = fluid.default_main_program() + fluid.io.load_vars(executor=exe, dirname=path, main_program=prog, + vars=None) + # All variables in `main_program` whose name includes "fc" will be loaded. + # And all the variables are supposed to have been saved in differnet files. + + + # The second usage: using `vars` to specify variables + var_list = [var_a, var_b, var_c] + fluid.io.load_vars(executor=exe, dirname=path, vars=var_list, + filename="vars_file") + # var_a, var_b and var_c will be loaded. And they are supposed to haven + # been saved in the same file named 'var_file' in the path "./my_paddle_model". """ if vars is None: if main_program is None: @@ -221,7 +406,42 @@ def load_vars(executor, def load_params(executor, dirname, main_program=None, filename=None): """ - load all parameters from directory by executor. + This function filters out all parameters from the give `main_program` + and then trys to load these parameters from the folder `dirname` or + the file `filename`. + + Use the `dirname` to specify the folder where parameters were saved. If + parameters were saved in separate files in the folder `dirname`, set + `filename` None; if all parameters were saved in a single file, use + `filename` to specify the file name. + + NOTICE: Some variables are not Parameter while they are necessary for + training. So you can NOT save and continue your training just by + `save_params()` and `load_params()`. Please use `save_persistables()` + and `load_persistables()` instead. + + Args: + executor(Executor): The executor to run for loading parameters. + dirname(str): The directory path. + main_program(Program|None): The program whose parameters will be + loaded. If it is None, the default + main program will be used automatically. + Default: None + filename(str|None): The file which saved all parameters. If parameters + were saved in differnet files, set it to None. + Default: None + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + fluid.io.load_params(executor=exe, dirname=param_path, + main_program=None) """ load_vars( executor, @@ -233,7 +453,37 @@ def load_params(executor, dirname, main_program=None, filename=None): def load_persistables(executor, dirname, main_program=None, filename=None): """ - load all persistables from directory by executor. + This function filters out all variables with `persistable==True` from the + give `main_program` and then trys to load these variables from the folder + `dirname` or the file `filename`. + + Use the `dirname` to specify the folder where persistable variables were + saved. If variables were saved in separate files, set `filename` None; + if all variables were saved in a single file, use `filename` to specify + the file name. + + Args: + executor(Executor): The executor to run for loading persistable variables. + dirname(str): The directory path. + main_program(Program|None): The program whose persistbale variables will + be loaded. If it is None, the default main + program will be used automatically. + Default: None + filename(str|None): The file which saved all variables. If variables were + saved in differnet files, set it to None. + Default: None + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + fluid.io.load_persistables(executor=exe, dirname=param_path, + main_program=None) """ load_vars( executor, @@ -306,22 +556,48 @@ def save_inference_model(dirname, model_filename=None, params_filename=None): """ - Build a model especially for inference, - and save it to directory by the executor. + Prune the given `main_program` to build a new program especially for inference, + and then save it and all related parameters to given `dirname` by the `executor`. + + Args: + dirname(str): The directory path to save the inference model. + feeded_var_names(list[str]): Names of variables that need to be feeded data + during inference. + target_vars(list[Variable]): Variables from which we can get inference + results. + executor(Executor): The executor that saves the inference model. + main_program(Program|None): The original program, which will be pruned to + build the inference model. If is setted None, + the default main program will be used. + Default: None. + model_filename(str|None): The name of file to save the inference program + itself. If is setted None, a default filename + `__model__` will be used. + params_filename(str|None): The name of file to save all related parameters. + If it is setted None, parameters will be saved + in separate files . - :param dirname: directory path - :param feeded_var_names: Names of variables that need to be feeded data during inference - :param target_vars: Variables from which we can get inference results. - :param executor: executor that save inference model - :param main_program: original program, which will be pruned to build the inference model. - Default default_main_program(). - :param model_filename: The name of file to save inference program. - If not specified, default filename `__model__` will be used. - :param params_filename: The name of file to save parameters. - It is used for the case that all parameters are saved in a single binary file. - If not specified, parameters are considered saved in separate files. + Returns: + None + + Raises: + ValueError: If `feed_var_names` is not a list of basestring. + ValueError: If `target_vars` is not a list of Variable. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + path = "./infer_model" + fluid.io.save_inference_model(dirname=path, feeded_var_names=['img'], + target_vars=[predict_var], executor=exe) + + # In this exsample, the function will prune the default main program + # to make it suitable for infering the `predict_var`. The pruned + # inference program is going to be saved in the "./infer_model/__model__" + # and parameters are going to be saved in separate files under folder + # "./infer_model". - :return: None """ if isinstance(feeded_var_names, basestring): feeded_var_names = [feeded_var_names] @@ -382,18 +658,49 @@ def load_inference_model(dirname, """ Load inference model from a directory - :param dirname: directory path - :param executor: executor that load inference model - :param model_filename: The name of file to load inference program. - If not specified, default filename `__model__` will be used. - :param params_filename: The name of file to load parameters. - It is used for the case that all parameters are saved in a single binary file. - If not specified, parameters are considered saved in separate files. + Args: + dirname(str): The directory path + executor(Executor): The executor to run for loading inference model. + model_filename(str|None): The name of file to load inference program. + If it is None, the default filename + '__model__' will be used. + Default: None + params_filename(str|None): The name of file to load all parameters. + It is only used for the case that all + parameters were saved in a single binary + file. If parameters were saved in separate + files, set it as 'None'. + + Returns: + tuple: The return of this function is a tuple with three elements: + (program, feed_target_names, fetch_targets). The `program` is a + Program, it's the program for inference. The `feed_target_names` is + a list of str, it contains Names of variables that need to feed + data in the inference program. The `fetch_targets` is a list of + Variable. It contains variables from which we can get inference + results. + + Raises: + ValueError: If `dirname` is not a existing directory. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + path = "./infer_model" + [inference_program, feed_target_names, fetch_targets] = + fluid.io.load_inference_model(dirname=path, executor=exe) + results = exe.run(inference_program, + feed={feed_target_names[0]: tensor_img}, + fetch_list=fetch_targets) + + # In this exsample, the inference program was saved in the + # "./infer_model/__model__" and parameters were saved in + # separate files in ""./infer_model". + # After getting inference program, feed target names and + # fetch targets, we can use an Executor to run the inference + # program to get the inference result. - :return: [program, feed_target_names, fetch_targets] - program: program especially for inference. - feed_target_names: Names of variables that need to feed data - fetch_targets: Variables from which we can get inference results. """ if not os.path.isdir(dirname): raise ValueError("There is no directory named '%s'", dirname) @@ -424,12 +731,25 @@ def load_inference_model(dirname, def get_parameter_value(para, executor): """ - Get the LoDTensor for the parameter + Get the LoDTensor value of the given parameter. + + Args: + para(Parameter): The parameter to get value from. + executor(Executor): The executor to run for retrieving the value. + + Returns: + numpy.array: The given parameter's values. + + Raises: + AssertionError: If the `para` is not an instance of Parameter. - :param executor: executor for retrieving the value - :param para: the given parameter + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param = fluid.default_main_program().global_block().var('fc.w') + p = fluid.io.get_parameter_value(param, exe) - :return: the LoDTensor for the parameter """ assert is_parameter(para) @@ -441,14 +761,30 @@ def get_parameter_value(para, executor): def get_parameter_value_by_name(name, executor, program=None): """ - Get the LoDTensor for paramter with the given name + Get the LoDTensor value of a certain parameter by its name. + + Args: + name(str): The parameter's name. + executor(Executor): The executor to run for retrieving the value. + program(Program | None): The program where to find the parameter. + If it's set to be None, the function will + try to find the parameter in the default + main program. - :param executor: executor for retrieving the value - :param name: the name of the parameter - :param program: the program where the variable is found - Default default_main_program(). + Returns: + numpy.array: The parameter's values. - :return: the LoDTensor for the variable + Raises: + TypeError: If given `name` is not an instance of basestring. + TypeError: If the parameter with the given name doesn't exist. + AssertionError: If there is a varibale named `name` in the + given program but it is not a Parameter. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + p = fluid.io.get_parameter_value('fc.w', exe) """ if program is None: program = default_main_program() @@ -470,16 +806,58 @@ def save_checkpoint(executor, main_program=None, max_num_checkpoints=3): """ - Save Checkpoint will save persistable LodTensor variables from main_program in checkpoint directory, - the directory named by serial number from 0 to (n -1), save_checkpoint use LRU strategy - to keep numbers of checkpoint directory, the numbers of checkpoint directory are max_num_checkpoints at most, - The interval between two saved checkpoints must greater than save_interval_secs. + This function filters out all checkpoint variables from the give + main_program and then saves these variables to the `checkpoint_dir` + directory. + + In the training precess, we generally save a checkpoint in each + iteration. So there might be a lot of checkpoints in the + `checkpoint_dir`. To avoid them taking too much disk space, the + `max_num_checkpoints` are introduced to limit the total number of + checkpoints. If the number of existing checkpints is greater than + the `max_num_checkpoints`, oldest ones will be scroll deleted. + + A variable is a checkpoint variable and will be saved if it meets + all following conditions: + 1. It's persistable. + 2. It's type is not FEED_MINIBATCH nor FETCH_LIST nor RAW. + 3. It's name contains no "@GRAD" nor ".trainer_" nor ".block". - :param executor executor for save the value - :param checkpoint_dir the checkpoint directory - :param trainer_id currect trainer id, if id is equal to 0, the trainer is chief - :param main_program will save all variables in program - :param max_num_checkpoints will keep numbers of checkpoint serials not bigger than max_num_checkpoints + Args: + executor(Executor): The executor to run for save checkpoint. + checkpoint_dir(str): The folder where to save checkpoints. + trainer_id(int): currect trainer id, if id is equal to 0, the trainer + is chief. + trainer_args(dict|None): Current training arguments. Such as 'epoch_id' + and 'step_id'. + Defaut: None + main_program(Program|None): The program whose checkpoint variables will + be saved. If it is None, the default main program will be used. + max_num_checkpoints(int): The max number of total number of existing + checkpoints. + Default: 3 + + Returns: + None + + Raises: + ValueError: If `checkpoint_dir` is None. + AssertionError: If `trainer_args` is not a dict. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + path = "./checkpoints" + prog = fluid.default_main_program() + trainer_args = {"epoch_id": 200, + "step_id": 20} # just an example + fluid.io.save_checkpoint(executor=exe, + checkpoint_dir=path, + trainer_id=0, + trainer_args=trainer_args, + main_program=prog, + max_num_checkpoints=3) """ if checkpoint_dir is None: raise ValueError("'checkpoint_dir' should not be None") @@ -503,13 +881,50 @@ def save_checkpoint(executor, def load_checkpoint(executor, checkpoint_dir, serial, main_program): """ - Load checkpoint from a directory by executor, - it will find the most recent saved checkpoint file and load it auto. + This function filters out all checkpoint variables from the give + main_program and then try to load these variables from the + `checkpoint_dir` directory. + + In the training precess, we generally save a checkpoint in each + iteration. So there are more than one checkpoint in the + `checkpoint_dir` (each checkpoint has its own sub folder), use + `serial` to specify which serial of checkpoint you would like to + load. + + A variable is a checkpoint variable and will be loaded if it meets + all following conditions: + 1. It's persistable. + 2. It's type is not FEED_MINIBATCH nor FETCH_LIST nor RAW. + 3. It's name contains no "@GRAD" nor ".trainer_" nor ".block". + + Args: + executor(Executor): The executor to run for loading checkpoint. + checkpoint_dir(str): The folder where all checkpoints are. + serial(int): The serial of checkpoint you would like to load. + main_program(Program): The program whose checkpoint variables will + be loaded. - :param executor executor for load the value - :param checkpoint_dir the checkpoint directory - :param serial the serial folder in checkpoint directory will be load - :param main_program will load all variables in program + Returns: + None + + Raises: + ValueError: If `checkpoint_dir` is None. + ValueError: If `serial` is None or `serial` is less than 0. + ValueError: If `main_program` is None. + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + path = "./checkpoints" + prog = fluid.default_main_program() + fluid.io.load_checkpoint(executor=exe, checkpoint_dir=path, + serial=9, main_program=prog) + + # In this example, `load_checkpoint` function + # will first filters out all checkpoint variables in the default + # main program, and then try to load these variables form the + # folder "./checkpoints/checkpoint_9/__model__". """ if checkpoint_dir is None: @@ -528,10 +943,10 @@ def load_checkpoint(executor, checkpoint_dir, serial, main_program): def clean_checkpoint(checkpoint_dir, delete_dir=False): """ clean the checkpoint dir, when the train exits normally, the trainer will call clean_checkpoint to delete checkpoint directory saved before. - delete_dir only works when the directory is empty, otherwise, OSError is raised. + delete_dir only works when the directory is empty, otherwise, OSError is raised. - :param checkpoint_dir - :param delete_dir + : param checkpoint_dir + : param delete_dir """ if checkpoint_dir is None: @@ -547,13 +962,40 @@ def load_persist_vars_without_grad(executor, program, has_model_dir=False): """ - load_persist_vars_without_grad will load variables from a directory by an executor, - the variable named end with "@GRAD" will not be loaded. + This function filters out all checkpoint variables from the give + program and then trys to load these variables from the given directory. + + A variable is a checkpoint variable if it meets all following + conditions: + 1. It's persistable. + 2. It's type is not FEED_MINIBATCH nor FETCH_LIST nor RAW. + 3. It's name contains no "@GRAD" nor ".trainer_" nor ".block". - :param executor executor for load the value - :param dirname the checkpoint directory - :param program will load all variables in program - :param has_model_dir if has_model_dir is True, will load variables from sub directory named __model__ + Args: + executor(Executor): The executor to run for loading variables. + dirname(str): The directory path. + program(Program): The program whose checkpoint variables will + be loaded. + has_model_dir(bool): if True, the function loads variables + from a sub directory named '__model__'. + Default: False + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + fluid.io.load_persist_vars_without_grad(executor=exe, + dirname=param_path, program=prog, has_model_dir=True) + + # In this example, `load_persist_vars_without_grad` function + # will first filters out all checkpoint variables in the default + # main program, and then trys to load these variables form the + # folder "./my_paddle_model/__model__". """ if has_model_dir: @@ -569,12 +1011,38 @@ def load_persist_vars_without_grad(executor, def save_persist_vars_without_grad(executor, dirname, program): """ - save_persist_vars_without_grad will save variables to a directory by an executor, - the variable named end with "@GRAD" will not be saved. + This function filters out all checkpoint variables from the give + program and then save these variables to a sub-folder '__model__' of + the given directory. + + A variable is a checkpoint variable if it meets all following + conditions: + 1. It's persistable. + 2. It's type is not FEED_MINIBATCH nor FETCH_LIST nor RAW. + 3. It's name contains no "@GRAD" nor ".trainer_" nor ".block". + + Args: + executor(Executor): The executor to run for saving variables. + dirname(str): The directory path. + program(Program): The program whose checkpoint variables will + be saved. + + Returns: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + prog = fluid.default_main_program() + fluid.io.save_persist_vars_without_grad(executor=exe, + dirname=param_path, program=prog) - :param executor executor for load the value - :param dirname the checkpoint directory - :param program will load all variables in program + # In this example, `save_persist_vars_without_grad` function + # will first filters out all checkpoint variables in the default + # main program, and then saves these variables to the folder + # "./my_paddle_model/__model__". """ cur_dir = _get_model_dir(dirname) save_vars( @@ -620,7 +1088,7 @@ def _is_checkpoint_var(var): the checkpoint will not save or load all the variables. var type is FEED_MINIBATCH/FETCH_LIST/RAW or var name ends with @GRAD are discarded. - :param var + : param var """ if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \ var.desc.type() == core.VarDesc.VarType.FETCH_LIST or \ @@ -701,7 +1169,7 @@ def _write_success(dirname): """ write an empty file named "_SUCCESS" in checkpoint dir, indicate this checkpoint is correct. - :param dirname + : param dirname """ success_file = os.path.join(dirname, SUCCESS_MARK_FILENAME) with open(success_file, 'a') as f: @@ -713,7 +1181,7 @@ def get_latest_checkpoint_serial(checkpoint_dir): """ get the latest file in checkpoint directory, the _SUCCESS file must exist in the directory - :param checkpoint_dir + : param checkpoint_dir """ if not checkpoint_dir: return -1 diff --git a/python/paddle/fluid/layers/__init__.py b/python/paddle/fluid/layers/__init__.py index a568f61dcb2da976baa7847ae26281a34d6f88dd..cd1492da24d5e9d09a9eaac0b1b9c7aaffac6250 100644 --- a/python/paddle/fluid/layers/__init__.py +++ b/python/paddle/fluid/layers/__init__.py @@ -28,8 +28,8 @@ import math_op_patch from math_op_patch import * import detection from detection import * -import metric -from metric import * +import metric_op +from metric_op import * from learning_rate_scheduler import * __all__ = [] @@ -41,5 +41,5 @@ __all__ += control_flow.__all__ __all__ += ops.__all__ __all__ += device.__all__ __all__ += detection.__all__ -__all__ += metric.__all__ +__all__ += metric_op.__all__ __all__ += learning_rate_scheduler.__all__ diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 581770feea98230ce6161bd11dc43f79cecd0048..849474dc58461ac3772f439da7bf5d57592daa8c 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -185,12 +185,14 @@ def Print(input, Returns: Variable: Output tensor, same data with input tensor. + Examples: + .. code-block:: python - value = some_layer(...) - Print(value, summarize=10, - message="The content of some_layer: ") + value = some_layer(...) + Print(value, summarize=10, + message="The content of some_layer: ") ''' helper = LayerHelper('print', **locals()) out = helper.create_tmp_variable(dtype=helper.input_dtype()) @@ -1201,6 +1203,31 @@ class ConditionalBlockGuard(BlockGuard): class ConditionalBlock(object): + ''' + **ConditionalBlock** + + ConditionalBlock is an operator that bind a block to a specific condition, + if the condition matches, the corresponding block will be executed. + + Args: + inputs (Variable): bool conditions. + is_scalar_condition (bool): whether the branch is controled by a scalar. + name(str): name of this ConditionalBlock. + + Examples: + .. code-block:: python + + cond = layers.less_than(x=label, y=limit) + true_image, false_image = layers.split_lod_tensor( + input=image, mask=cond) + true_cond = layers.ConditionalBlock([true_image]) + + with true_cond.block(): + ... + with false_cond.block(): + ... + ''' + def __init__(self, inputs, is_scalar_condition=False, name=None): for each_input in inputs: if not isinstance(each_input, Variable): diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index d5471d182bf19015995aeec2a81ec5a772765712..200db87f1793a41e8327b59677252c19eab567de 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -16,7 +16,7 @@ All layers just related to the detection neural network. """ from layer_function_generator import generate_layer_fn -from layer_function_generator import autodoc +from layer_function_generator import autodoc, templatedoc from ..layer_helper import LayerHelper import tensor import nn @@ -155,7 +155,7 @@ def detection_output(loc, return nmsed_outs -@autodoc() +@templatedoc() def detection_map(detect_res, label, class_num, @@ -166,6 +166,47 @@ def detection_map(detect_res, input_states=None, out_states=None, ap_version='integral'): + """ + ${comment} + + Args: + detect_res: ${detect_res_comment} + label: ${label_comment} + class_num: ${class_num_comment} + background_label: ${background_label_comment} + overlap_threshold: ${overlap_threshold_comment} + evaluate_difficult: ${evaluate_difficult_comment} + has_state: ${has_state_comment} + input_states: If not None, It contains 3 elements: + 1. pos_count ${pos_count_comment}. + 2. true_pos ${true_pos_comment}. + 3. false_pos ${false_pos_comment}. + out_states: If not None, it contains 3 elements. + 1. accum_pos_count ${accum_pos_count_comment}. + 2. accum_true_pos ${accum_true_pos_comment}. + 3. accum_false_pos ${accum_false_pos_comment}. + ap_version: ${ap_type_comment} + + Returns: + ${map_comment} + + + Examples: + .. code-block:: python + + detect_res = fluid.layers.data( + name='detect_res', + shape=[10, 6], + append_batch_size=False, + dtype='float32') + label = fluid.layers.data( + name='label', + shape=[10, 6], + append_batch_size=False, + dtype='float32') + + map_out = fluid.layers.detection_map(detect_res, label, 21) + """ helper = LayerHelper("detection_map", **locals()) def __create_var(type): diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index 8d153b75cd49953770cfa89348914a375be82a82..f3ab47c96b1caa2facfd6d191af014b4c7380cbc 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -186,7 +186,6 @@ class ListenAndServ(object): main_program = self.helper.main_program current_block = main_program.current_block() parent_block = self.parent_block() - empty_block = Program().global_block() parent_block.append_op( type='listen_and_serv', @@ -195,8 +194,9 @@ class ListenAndServ(object): attrs={ 'endpoint': self.endpoint, 'Fanin': self.fan_in, - 'OptimizeBlock': current_block, - 'PrefetchBlock': empty_block, + 'optimize_blocks': [ + current_block + ], # did not support multiple optimize blocks in layers 'sync_mode': True, # did not support async now in layers 'grad_to_block_id': [""] }) diff --git a/python/paddle/fluid/layers/metric.py b/python/paddle/fluid/layers/metric_op.py similarity index 99% rename from python/paddle/fluid/layers/metric.py rename to python/paddle/fluid/layers/metric_op.py index 58de1b6b9fe17a24203e93de6780190b9fc6b3e7..99e82fdd04282177fae63f1fb94b5e32d41c612e 100644 --- a/python/paddle/fluid/layers/metric.py +++ b/python/paddle/fluid/layers/metric_op.py @@ -126,7 +126,7 @@ def auc(input, label, curve='ROC', num_thresholds=200): topk_out, topk_indices = nn.topk(input, k=k) auc_out = helper.create_tmp_variable(dtype="float32") helper.append_op( - type="accuracy", + type="auc", inputs={ "Out": [topk_out], "Indices": [topk_indices], diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index f6f188df0d6a9a33f4ad858f00c1ba0fd36661b9..be22bde4608807aff12ae8fa4b4c723211ffecce 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. """ -All layers just related to the neural network. +All layers just related to the neural network. """ from ..layer_helper import LayerHelper @@ -93,6 +93,7 @@ __all__ = [ 'mean_iou', 'relu', 'log', + 'crop', ] @@ -108,14 +109,14 @@ def fc(input, """ **Fully Connected Layer** - This function creates a fully connected layer in the network. It can take - multiple tensors as its inputs. It creates a variable called weights for - each input tensor, which represents a fully connected weight matrix from - each input unit to each output unit. The fully connected layer multiplies - each input tensor with its coresponding weight to produce an output Tensor. - If multiple input tensors are given, the results of multiple multiplications - will be sumed up. If bias_attr is not None, a bias variable will be created - and added to the output. Finally, if activation is not None, it will be applied + This function creates a fully connected layer in the network. It can take + multiple tensors as its inputs. It creates a variable called weights for + each input tensor, which represents a fully connected weight matrix from + each input unit to each output unit. The fully connected layer multiplies + each input tensor with its coresponding weight to produce an output Tensor. + If multiple input tensors are given, the results of multiple multiplications + will be sumed up. If bias_attr is not None, a bias variable will be created + and added to the output. Finally, if activation is not None, it will be applied to the output as well. This process can be formulated as follows: @@ -197,7 +198,10 @@ def fc(input, else: pre_bias = helper.create_tmp_variable(dtype) helper.append_op( - type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias}) + type="sum", + inputs={"X": mul_results}, + outputs={"Out": pre_bias}, + attrs={"use_mkldnn": use_mkldnn}) # add bias pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims) # add activation @@ -846,7 +850,7 @@ def crf_decoding(input, param_attr, label=None): Returns: Variable: ${viterbi_path_comment} - + Examples: .. code-block:: python @@ -1084,7 +1088,7 @@ def chunk_eval(input, Here is a NER example of labeling for these tagging schemes: .. code-block:: python - + ====== ====== ====== ===== == ============ ===== ===== ===== == ========= Li Ming works at Agricultural Bank of China in Beijing. ====== ====== ====== ===== == ============ ===== ===== ===== == ========= @@ -1110,7 +1114,7 @@ def chunk_eval(input, is the num of chunk types, and `tag_type` get its value from the following table. .. code-block:: python - + Scheme Begin Inside End Single plain 0 - - - IOB 0 1 - - @@ -1146,7 +1150,7 @@ def chunk_eval(input, tuple: tuple containing: precision, recall, f1_score, num_infer_chunks, num_label_chunks, num_correct_chunks - + Examples: .. code-block:: python @@ -1246,7 +1250,7 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=True): """ This function computes the softmax activation among all time-steps for each sequence. The dimension of each time-step should be 1. Thus, the shape of - input Tensor can be either :math:`[N, 1]` or :math:`[N]`, where :math:`N` + input Tensor can be either :math:`[N, 1]` or :math:`[N]`, where :math:`N` is the sum of the length of all sequences. For i-th sequence in a mini-batch: @@ -1266,7 +1270,7 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=True): param_attr (ParamAttr|None): attributes for parameter use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \ library is installed. Default: True - + Returns: Variable: output of sequence_softmax @@ -1827,11 +1831,11 @@ def pool2d(input, ${comment} Args: - input (Variable): The input tensor of pooling operator. The format of - input tensor is NCHW, where N is batch size, C is - the number of channels, H is the height of the + input (Variable): The input tensor of pooling operator. The format of + input tensor is NCHW, where N is batch size, C is + the number of channels, H is the height of the feature, and W is the width of the feature. - pool_size (int): The side length of pooling windows. All pooling + pool_size (int): The side length of pooling windows. All pooling windows are squares with pool_size on a side. pool_type: ${pooling_type_comment} pool_stride (int): stride of the pooling layer. @@ -1840,7 +1844,7 @@ def pool2d(input, use_cudnn: ${use_cudnn_comment} ceil_mode: ${ceil_mode_comment} use_mkldnn: ${use_mkldnn_comment} - name (str|None): A name for this layer(optional). If set None, the + name (str|None): A name for this layer(optional). If set None, the layer will be named automatically. Returns: @@ -1858,10 +1862,10 @@ def pool2d(input, data = fluid.layers.data( name='data', shape=[3, 32, 32], dtype='float32') conv2d = fluid.layers.pool2d( - input=data, - pool_size=2, - pool_type='max', - pool_stride=1, + input=data, + pool_size=2, + pool_type='max', + pool_stride=1, global_pooling=False) """ if pool_type not in ["max", "avg"]: @@ -2226,14 +2230,14 @@ def beam_search_decode(ids, scores, name=None): This layers is to pack the output of beam search layer into sentences and associated scores. It is usually called after the beam search layer. Typically, the output of beam search layer is a tensor of selected ids, with - a tensor of the score of each id. Beam search layer's output ids, however, - are generated directly during the tree search, and they are stacked by each - level of the search tree. Thus we need to reorganize them into sentences, + a tensor of the score of each id. Beam search layer's output ids, however, + are generated directly during the tree search, and they are stacked by each + level of the search tree. Thus we need to reorganize them into sentences, based on the score of each id. This layer takes the output of beam search layer as input and repack them into sentences. Args: - ids (Variable): The selected ids, output of beam search layer. + ids (Variable): The selected ids, output of beam search layer. scores (Variable): The associated scores of the ids, out put of beam search layer. name (str): The name of this layer. It is optional. @@ -2241,7 +2245,7 @@ def beam_search_decode(ids, scores, name=None): Returns: tuple(Variable): a tuple of two output tensors: sentence_ids, sentence_scores. sentence_ids is a tensor with shape [size, length], where size is the - beam size of beam search, and length is the length of each sentence. + beam size of beam search, and length is the length of each sentence. Note that the length of sentences may vary. sentence_scores is a tensor with the same shape as sentence_ids. @@ -2674,18 +2678,35 @@ def sequence_expand(x, y, ref_level=-1, name=None): def beam_search(pre_ids, ids, scores, beam_size, end_id, level=0): ''' + **beam search** + This function implements the beam search algorithm. + Beam search is a classical algorithm for selecting candidate words + in a machine translation task. + + Refer to `Beam search `_ + for more details. + Args: - pre_ids (Variable): ${pre_ids_comment} - ids (Variable): ${ids_comment} - scores (Variable): ${scores_comment} - beam_size (int): ${beam_size_comment} - end_id (int): ${end_id_comment} - level (int): ${level_comment} + pre_ids (Variable): ids in previous step. + ids (Variable): a LoDTensor of shape of [None,k] + scores (Variable): a LoDTensor that has the same shape and LoD with `ids` + beam_size (int): beam size for beam search + end_id (int): the token id which indicates the end of a sequence + level (int): the level of LoDTensor Returns: - tuple: a tuple of beam_search output variables: selected_ids, selected_scores + tuple: a tuple of beam_search output variables: `selected_ids`, `selected_scores` + + Examples: + .. code-block:: python + + # current_score is a Tensor of shape (num_batch_size, embed_size), which + # consists score of each candidate word. + topk_scores, topk_indices = pd.topk(current_score, k=50) + selected_ids, selected_scores = pd.beam_search( + pre_ids, topk_indices, topk_scores, beam_size, end_id=10, level=0) ''' helper = LayerHelper('beam_search', **locals()) score_type = scores.dtype @@ -2901,7 +2922,7 @@ def reduce_mean(input, dim=None, keep_dim=False, name=None): `None`, compute the mean over all elements of :attr:`input` and return a variable with a single element, otherwise it must be in the range :math:`[-rank(input), rank(input))`. If - :math:`dim[i] < 0`, the dimension to reduce is + :math:`dim[i] < 0`, the dimension to reduce is :math:`rank(input) + dim[i]`. keep_dim (bool): Whether to reserve the reduced dimension in the output Tensor. The result tensor will have one fewer dimension @@ -3372,16 +3393,16 @@ def topk(input, k, name=None): Args: input(Variable): The input variable which can be a vector or Tensor with higher rank. - k(int): The number of top elements to look for along the last dimension + k(int): The number of top elements to look for along the last dimension of input. name(str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + will be named automatically. Default: None Returns: - Tuple[Variable]: A tuple with two elements. Each element is a Variable. - The first one is k largest elements along each last - dimensional slice. The second one is indices of values + Tuple[Variable]: A tuple with two elements. Each element is a Variable. + The first one is k largest elements along each last + dimensional slice. The second one is indices of values within the last dimension of input. Raises: @@ -3576,15 +3597,15 @@ def warpctc(input, label, blank=0, norm_by_times=False): It's shape is [Lp, num_classes + 1], where Lp is the sum of all input sequences' length and num_classes is the true number of classes. (not including the blank label). - label (Variable): The ground truth of variable-length sequence, + label (Variable): The ground truth of variable-length sequence, which is a 2-D Tensor with LoD information. It is of the shape [Lg, 1], where Lg is th sum of all labels' length. blank (int, default 0): The blank label index of Connectionist Temporal Classification (CTC) loss, which is in the half-opened interval [0, num_classes + 1). - norm_by_times(bool, default false): Whether to normalize the gradients - by the number of time-step, which is also the sequence's length. - There is no need to normalize the gradients if warpctc layer was + norm_by_times(bool, default false): Whether to normalize the gradients + by the number of time-step, which is also the sequence's length. + There is no need to normalize the gradients if warpctc layer was follewed by a mean_op. Returns: @@ -3690,8 +3711,8 @@ def nce(input, input (Variable): input variable. label (Variable): label. num_total_classes (int):${num_total_classes_comment} - sample_weight (Variable|None): A Variable of shape [batch_size, 1] - storing a weight for each sample. The default weight for each + sample_weight (Variable|None): A Variable of shape [batch_size, 1] + storing a weight for each sample. The default weight for each sample is 1.0. param_attr (ParamAttr|None): attributes for parameter bias_attr (ParamAttr|None): attributes for bias @@ -4081,7 +4102,7 @@ def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None): This layer computes the smooth L1 loss for Variable :attr:`x` and :attr:`y`. It takes the first dimension of :attr:`x` and :attr:`y` as batch size. For each instance, it computes the smooth L1 loss element by element first - and then sums all the losses. So the shape of ouput Variable is + and then sums all the losses. So the shape of ouput Variable is [batch_size, 1]. Args: @@ -4090,14 +4111,14 @@ def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None): y (Variable): A tensor with rank at least 2. The target value of smooth L1 loss op with same shape as :attr:`x`. inside_weight (Variable|None): A tensor with rank at least 2. This - input is optional and should have same shape with :attr:`x`. If - provided, the result of (:attr:`x` - :attr:`y`) will be multiplied + input is optional and should have same shape with :attr:`x`. If + provided, the result of (:attr:`x` - :attr:`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 :attr:`x`. If - provided, the out smooth L1 loss will be multiplied by this tensor + input is optional and should have same shape with :attr:`x`. If + provided, the out smooth L1 loss will be multiplied by this tensor element by element. - sigma (float|None): Hyper parameter of smooth L1 loss layer. A float + sigma (float|None): Hyper parameter of smooth L1 loss layer. A float scalar with default value 1.0. Returns: @@ -4143,7 +4164,7 @@ def one_hot(input, depth): Examples: .. code-block:: python - + label = layers.data(name="label", shape=[1], dtype="float32") one_hot_label = layers.one_hot(input=label, depth=10) """ @@ -4297,10 +4318,10 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None): def lod_reset(x, y=None, target_lod=None): """ Set LoD of :attr:`x` to a new one specified by :attr:`y` or - :attr:`target_lod`. When :attr:`y` provided, :attr:`y.lod` would be - considered as target LoD first, otherwise :attr:`y.data` would be - considered as target LoD. If :attr:`y` is not provided, target LoD should - be specified by :attr:`target_lod`. If target LoD is specified by + :attr:`target_lod`. When :attr:`y` provided, :attr:`y.lod` would be + considered as target LoD first, otherwise :attr:`y.data` would be + considered as target LoD. If :attr:`y` is not provided, target LoD should + be specified by :attr:`target_lod`. If target LoD is specified by :attr:`Y.data` or :attr:`target_lod`, only one level LoD is supported. .. code-block:: text @@ -4354,7 +4375,7 @@ def lod_reset(x, y=None, target_lod=None): Args: x (Variable): Input variable which could be a Tensor or LodTensor. - y (Variable|None): If provided, output's LoD would be derived + y (Variable|None): If provided, output's LoD would be derived from :attr:`y`. target_lod (list|tuple|None): One level LoD which should be considered as target LoD when :attr:`y` not provided. @@ -4670,7 +4691,7 @@ def image_resize(input, """ **Resize a Batch of Images** - The input must be a tensor of the shape (num_batches, channels, in_h, in_w), + The input must be a tensor of the shape (num_batches, channels, in_h, in_w), and the resizing only applies on the last two dimensions(hight and width). Supporting resample methods: @@ -4766,9 +4787,9 @@ def resize_bilinear(input, out_shape=None, scale=None, name=None): def image_resize_short(input, out_short_len, resample='BILINEAR'): """ - Resize a batch of images. The short edge of input images will be - resized to the given 'out_short_len'. The long edge of input images - will be resized proportionately to make images' length-width ratio + Resize a batch of images. The short edge of input images will be + resized to the given 'out_short_len'. The long edge of input images + will be resized proportionately to make images' length-width ratio constant. Args: @@ -4801,7 +4822,7 @@ def gather(input, index): """ **Gather Layer** - Output is obtained by gathering entries of the outer-most dimension + Output is obtained by gathering entries of the outer-most dimension of X indexed by `index` and concatenate them together. .. math:: @@ -4826,7 +4847,7 @@ def gather(input, index): [5, 6]] Args: - input (Variable): The source input with rank>=1. + input (Variable): The source input with rank>=1. index (Variable): The index input with rank=1. Returns: @@ -4862,7 +4883,7 @@ def random_crop(x, shape, seed=None): Returns: ${out_comment} - + Examples: >>> img = fluid.layers.data("img", [3, 256, 256]) >>> cropped_img = fluid.layers.random_crop(img, shape=[3, 224, 224]) @@ -4899,16 +4920,16 @@ def random_crop(x, shape, seed=None): return out -def log(x): +def log(input): """ Calculates the natural log of the given input tensor, element-wise. .. math:: - Out = \\ln(x) + Out = \\ln(input) Args: - x (Variable): Input tensor. + input (Variable): Input tensor. Returns: Variable: The natural log of the input tensor computed element-wise. @@ -4917,27 +4938,27 @@ def log(x): .. code-block:: python - output = fluid.layers.log(x) + output = fluid.layers.log(input) """ helper = LayerHelper('log', **locals()) - dtype = helper.input_dtype() + dtype = helper.input_dtype(input_param_name='x') out = helper.create_tmp_variable(dtype) - helper.append_op(type="log", inputs={"X": input}, outputs={"Out": out}) + helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out}) return out -def relu(x): +def relu(input): """ Relu takes one input data (Tensor) and produces one output data (Tensor) - where the rectified linear function, y = max(0, x), is applied to + where the rectified linear function, y = max(0, input), is applied to the tensor elementwise. .. math:: - Out = \\max(0, x) + Out = \\max(0, input) Args: - x (Variable): The input tensor. + input (Variable): The input tensor. Returns: Variable: The output tensor with the same shape as input. @@ -4946,27 +4967,27 @@ def relu(x): .. code-block:: python - output = fluid.layers.relu(x) + output = fluid.layers.relu(input) """ helper = LayerHelper('relu', **locals()) - dtype = helper.input_dtype() + dtype = helper.input_dtype(input_param_name='x') out = helper.create_tmp_variable(dtype) - helper.append_op(type="relu", inputs={"X": input}, outputs={"Out": out}) + helper.append_op(type="relu", inputs={"X": x}, outputs={"Out": out}) return out def mean_iou(input, label, num_classes): """ Mean Intersection-Over-Union is a common evaluation metric for - semantic image segmentation, which first computes the IOU for each - semantic class and then computes the average over classes. - IOU is defined as follows: - + semantic image segmentation, which first computes the IOU for each + semantic class and then computes the average over classes. + IOU is defined as follows: + .. math:: IOU = \\frac{true\_positiv}{(true\_positive + false\_positive + false\_negative)}. - The predictions are accumulated in a confusion matrix and mean-IOU + The predictions are accumulated in a confusion matrix and mean-IOU is then calculated from it. @@ -4979,12 +5000,12 @@ def mean_iou(input, label, num_classes): Returns: mean_iou (Variable): A Tensor representing the mean intersection-over-union with shape [1]. out_wrong(Variable): A Tensor with shape [num_classes]. The wrong numbers of each class. - out_correct(Variable): A Tensor with shape [num_classes]. The correct numbers of each class. + out_correct(Variable): A Tensor with shape [num_classes]. The correct numbers of each class. Examples: .. code-block:: python - + iou, wrongs, corrects = fluid.layers.mean_iou(predict, label, num_classes) """ helper = LayerHelper('mean_iou', **locals()) @@ -5003,3 +5024,101 @@ def mean_iou(input, label, num_classes): }, attrs={"num_classes": num_classes}) return out_mean_iou, out_wrong, out_correct + + +def crop(x, shape=None, offsets=None, name=None): + """ + Crop input into output, as specified by offsets and shape. + + .. code-block:: text + + * Case 1: + Given + X = [[0, 1, 2, 0, 0] + [0, 3, 4, 0, 0] + [0, 0, 0, 0, 0]], + and + shape = [2, 2], + offsets = [0, 1], + output is: + Out = [[1, 2], + [3, 4]]. + * Case 2: + Given + X = [[0, 1, 2, 5, 0] + [0, 3, 4, 6, 0] + [0, 0, 0, 0, 0]], + and shape is tensor + shape = [[0, 0, 0] + [0, 0, 0]] + and + offsets = [0, 1], + + output is: + Out = [[1, 2, 5], + [3, 4, 6]]. + + Args: + x (Variable): The input tensor variable. + shape (Variable|list/tuple of integer): The output shape is specified + by `shape`, which can a Variable or a list/tupe of integer. + If a tensor Variable, it's rank must be the same as `x`. This way + is suitable for the case that the output shape may be changed each + iteration. If a list/tupe of integer, it's length must be the same + as the rank of `x` + offsets (Variable|list/tuple of integer|None): Specifies the copping + offsets at each dimension. It can be a Variable or or a list/tupe + of integer. If a tensor Variable, it's rank must be the same as `x`. + This way is suitable for the case that the offsets may be changed + each iteration. If a list/tupe of integer, it's length must be the + same as the rank of `x`. If None, the offsets are 0 at each + dimension. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The cropped tensor variable. + + Raises: + ValueError: If shape is not a list, tuple or Variable. + + Examples: + + .. code-block:: python + + x = fluid.layers.data(name="x", shape=[3, 5], dtype="float32") + y = fluid.layers.data(name="y", shape=[2, 3], dtype="float32") + crop = fluid.layers.crop(x, shape=y) + + # or + z = fluid.layers.data(name="z", shape=[3, 5], dtype="float32") + crop = fluid.layers.crop(z, shape=[2, 3]) + + """ + helper = LayerHelper('crop', **locals()) + + if not (isinstance(shape, list) or isinstance(shape, tuple) or \ + isinstance(shape, Variable)): + raise ValueError("The shape should be a list, tuple or Variable.") + + if offsets is None: + offsets = [0] * len(x.shape) + + out = helper.create_tmp_variable(x.dtype) + ipts = {'X': x} + attrs = {} + if isinstance(shape, Variable): + ipts['Y'] = shape + else: + attrs['shape'] = shape + if isinstance(offsets, Variable): + ipts['Offsets'] = offsets + else: + attrs['offsets'] = offsets + + helper.append_op( + type='crop', + inputs=ipts, + outputs={'Out': out}, + attrs=None if len(attrs) == 0 else attrs) + return out diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index 656bd5bb1d7d8e76209331e2a4912738a1cd6bfa..109a2694ad0d23ee35fb48810aba94842718fd6b 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -231,7 +231,11 @@ def sums(input, out=None): helper = LayerHelper('sum', **locals()) if out is None: out = helper.create_tmp_variable(dtype=helper.input_dtype()) - helper.append_op(type='sum', inputs={'X': input}, outputs={'Out': out}) + helper.append_op( + type='sum', + inputs={'X': input}, + outputs={'Out': out}, + attrs={'use_mkldnn': False}) return out @@ -381,7 +385,7 @@ def argmin(x, axis=0): """ **argmin** - This function computes the indices of the min elements + This function computes the indices of the min elements of the input tensor's element along the provided axis. Args: @@ -396,7 +400,7 @@ def argmin(x, axis=0): .. code-block:: python out = fluid.layers.argmin(x=in, axis=0) - out = fluid.layers.argmin(x=in, axis=-1) + out = fluid.layers.argmin(x=in, axis=-1) """ helper = LayerHelper("arg_min", **locals()) out = helper.create_tmp_variable(VarDesc.VarType.INT64) @@ -412,7 +416,7 @@ def argmax(x, axis=0): """ **argmax** - This function computes the indices of the max elements + This function computes the indices of the max elements of the input tensor's element along the provided axis. Args: @@ -427,7 +431,7 @@ def argmax(x, axis=0): .. code-block:: python out = fluid.layers.argmax(x=in, axis=0) - out = fluid.layers.argmax(x=in, axis=-1) + out = fluid.layers.argmax(x=in, axis=-1) """ helper = LayerHelper("arg_max", **locals()) out = helper.create_tmp_variable(VarDesc.VarType.INT64) @@ -546,9 +550,9 @@ def reverse(x, axis): Args: x(Vairbale): the input to be reversed. - axis(int|tuple|list): Axis that along which order of elements - is reversed. If it is a tuple or a list, reversing - will be apply on each axis in the tuple or list. + axis(int|tuple|list): Axis that along which order of elements + is reversed. If it is a tuple or a list, reversing + will be apply on each axis in the tuple or list. Returns: Variable: The reversed tensor. @@ -579,9 +583,9 @@ def save(x, file_path, overwrite=True): Args: x(variable): The Tensor/LoDTensor to be saved. file_path(str): The file path where the variable will be saved. - overwrite(bool): Whether or not cover the given file when it has already - existed. If it's set 'False' and the file is existed, a runtime - error will be thrown. + overwrite(bool): Whether or not cover the given file when it has already + existed. If it's set 'False' and the file is existed, a runtime + error will be thrown. """ helper = LayerHelper("save", **locals()) helper.append_op( @@ -601,8 +605,8 @@ def save_combine(x, file_path, overwrite=True): a single file. file_path(str): The file path where variables will be saved. overwrite(bool): Whether or not cover the given file when it has already - existed. If it's set 'False' and the file is existed, a runtime - error will be thrown. + existed. If it's set 'False' and the file is existed, a runtime + error will be thrown. Returns: There is no return value. diff --git a/python/paddle/fluid/lod_tensor.py b/python/paddle/fluid/lod_tensor.py index 61be39c25912604f842ef8a9a6ec5f0d1cf70257..c417ab393fca88d476d2f1fe83d12f99271d6883 100644 --- a/python/paddle/fluid/lod_tensor.py +++ b/python/paddle/fluid/lod_tensor.py @@ -19,33 +19,41 @@ __all__ = ['create_lod_tensor', 'create_random_int_lodtensor'] def create_lod_tensor(data, lod, place): - """Create a lod tensor from a numpy array, a list, or an existing lod tensor. + """ + Create a lod tensor from a numpy array, a list, or an existing lod tensor. Create a lod tensor by doing the following: + 1. Check that the length-based input lod is valid. + 2. Convert the length-based lod to a offset-based LoD. - 3. Copy the data from a numpy array, a list or a existing lod tensor to + + 3. Copy the data from a numpy array, a list or a existing lod tensor to CPU or GPU device (based on input place). + 4. Set the level of detail (LoD) using the offset-based LoD. - Use example: - Suppose we want LoDTensor to hold data for sequences of word, where each word is - represented by an integer. If we want to create a LoDTensor to represent two - sentences, one of 2 words, and one of 3 words. + Examples: - Then 'data' can be a numpy array of integers with shape (5, 1). - 'lod' will be [[2, 3]], indicating the length(# of words) in each sentence. - This length-based input lod [[2, 3]] will be converted to offset-based lod [[0, 2, 5]] - inside the function call. + Suppose we want LoDTensor to hold data for sequences of word, where each + word is represented by an integer. If we want to create a LoDTensor to + represent two sentences, one of 2 words, and one of 3 words. - Please refer to - github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/lod_tensor.md - for more details regarding LoD. + Then :code:`data` can be a numpy array of integers with shape (5, 1). + :code:`lod` will be [[2, 3]], indicating the length(# of words) in each + sentence. This length-based input lod [[2, 3]] will be converted to + offset-based lod [[0, 2, 5]] inside the function call. + + Please reference :ref:`api_guide_low_level_lod_tensor` for more details + regarding LoD. Args: - data: a numpy array or a LoDTensor or a list holding the data to be copied. - lod: a list of lists indicating the length-based LoD info specified by the user. - place: CPU or GPU place indicating where the data in the new LoDTensor will be stored. + data(numpy.ndarray|list|LoDTensor): a numpy array or a LoDTensor or a + list holding the data to be copied. + lod(list): a list of lists indicating the length-based LoD info + specified by the user. + place(Place): CPU or GPU place indicating where the data in the new + LoDTensor will be stored. Returns: A fluid LoDTensor object with tensor data and lod info. @@ -77,31 +85,38 @@ def create_lod_tensor(data, lod, place): def create_random_int_lodtensor(lod, base_shape, place, low, high): - """Create a LoDTensor containing random integers. + """ + Create a LoDTensor containing random integers. - This function is frequently used in the book examples. So we revised it based on - the new create_lod_tensor API and put it here in the lod_tensor module to simplify - the code. + This function is frequently used in the book examples. So we revised it + based on the new create_lod_tensor API and put it here in the lod_tensor + module to simplify the code. The function does the following: - 1. Calculate the overall shape of the LoDTensor based on the length-based 'lod' input - and the shape of the basic element in 'base_shape'. + + 1. Calculate the overall shape of the LoDTensor based on the length-based + :code:`lod` input and the shape of the basic element in + :code:`base_shape`. + 2. Create a numpy array of this shape. + 3. Create the LoDTensor using create_lod_tensor API. - Suppose we want LoDTensor to hold data for sequences of word, where each word is - represented by an integer. If we want to create a LoDTensor to represent two - sentences, one of 2 words, and one of 3 words. Then 'base_shape' is [1], input - length-based 'lod' is [[2, 3]]. Then the overall shape of the LoDTensor would be - [5, 1], holding 5 words for two sentences. + Suppose we want LoDTensor to hold data for sequences of word, where each + word is represented by an integer. If we want to create a LoDTensor to + represent two sentences, one of 2 words, and one of 3 words. Then + 'base_shape' is [1], input length-based 'lod' is [[2, 3]]. Then the overall + shape of the LoDTensor would be [5, 1], holding 5 words for two sentences. Args: - data: a numpy array or a LoDTensor holding the data to be copied. - lod: a list of lists indicating the length-based LoD info specified by the user. - base_shape: the shape of the basic element to be held by the LoDTensor. - place: CPU or GPU place indicating where the data in the new LoDTensor will be stored. - low: the lower bound of the random integers. - high: the upper bound of the random integers. + lod(list): a list of lists indicating the length-based LoD info + specified by the user. + base_shape(list): the shape of the basic element to be held by the + LoDTensor. + place(Place): CPU or GPU place indicating where the data in the new + LoDTensor will be stored. + low(int): the lower bound of the random integers. + high(int): the upper bound of the random integers. Returns: A fluid LoDTensor object with tensor data and lod info. diff --git a/python/paddle/fluid/metrics.py b/python/paddle/fluid/metrics.py index bb9c6fdc60089fc2b43573a6421a6f9781d2d4a8..c9cd881979a4ea4b14299ce219be4b5bd1f153fc 100644 --- a/python/paddle/fluid/metrics.py +++ b/python/paddle/fluid/metrics.py @@ -23,6 +23,8 @@ import warnings __all__ = [ 'MetricBase', 'CompositeMetric', + 'Precision', + 'Recall', 'Accuracy', 'ChunkEvaluator', 'EditDistance', @@ -46,33 +48,34 @@ def _is_number_or_matrix_(var): class MetricBase(object): """ - Base Class for all evaluators + Base Class for all Metrics. + MetricBase define a group of interfaces for the + model evaluation methods. Metrics accumulate metric states between + consecutive minibatches, at every minibatch, use update + interface to add current minibatch value to global states. + Use eval to compute accumative metric value from last reset() + or from scratch on. + If you need to custom a new metric, please inherit from MetricBase and + custom implementation. Args: - name(str): The name of evaluator. such as, "accuracy". Used for generate - temporary variable name. - Interface: - Note(*) : the states is the attributes who not has _ prefix. - - get_config(): print current states and configuration - reset(): clear the states. If the Metrics states type is not (int, float, np.ndarray), - Please override this method. - update(): update states at every minibatch - eval(): get metric evaluation in numpy type. + name(str): The name of metric instance. such as, "accuracy". + It needed if you want to distinct different metrics in a model. + """ - def __init__(self, name, **kwargs): + def __init__(self, name): self._name = str(name) if name != None else self.__class__.__name__ - self._kwargs = kwargs if kwargs != None else dict() - self.reset() def __str__(self): return self._name def reset(self): """ - states is the attributes who not has _ prefix. - reset the states of metrics. + reset clear the states of metrics. By default, the states + are the members who do not has _ prefix, reset set them to inital states. + If you violate the implicit name rule, please also custom the reset + interface. """ states = { attr: value @@ -90,61 +93,231 @@ class MetricBase(object): setattr(self, attr, None) def get_config(self): + """ + Get the metric and current states. + The states are the members who do not has "_" prefix. + + Args: + None + + Returns: + dict: a dict of metric and states + """ states = { attr: value for attr, value in self.__dict__.iteritems() if not attr.startswith("_") } - config = copy.deepcopy(self._kwargs) + config = {} config.update({"name": self._name, "states": copy.deepcopy(states)}) return config - def update(self): - raise NotImplementedError() + def update(self, preds, labels): + """ + Updates the metric states at every minibatch. + One user can compute the minibatch metric via pure Python, or + via a c++ operator. + + Args: + preds(numpy.array): the predictions of current minibatch + labels(numpy.array): the labels of current minibatch, if the label is one-hot + or soft-label, should custom the corresponding update rule. + """ + raise NotImplementedError( + "Should not use it directly, please extend it.") def eval(self): - raise NotImplementedError() + """ + Evalute the current metrics based the accumulated states. + + Returns: + float|list(float)|numpy.array: the metrics via Python. + """ + raise NotImplementedError( + "Should not use it directly, please extend it.") class CompositeMetric(MetricBase): """ - Compute multiple metrics in each minibatch. + Composite multiple metrics in one instance. for example, merge F1, accuracy, recall into one Metric. + + Examples: + .. code-block:: python + + labels = fluid.layers.data(name="data", shape=[1], dtype="int32") + data = fluid.layers.data(name="data", shape=[32, 32], dtype="int32") + pred = fluid.layers.fc(input=data, size=1000, act="tanh") + comp = fluid.metrics.CompositeMetric() + acc = fluid.metrics.Precision() + recall = fluid.metrics.Recall() + comp.add_metric(acc) + comp.add_metric(recall) + for pass in range(PASSES): + comp.reset() + for data in train_reader(): + loss, preds, labels = exe.run(fetch_list=[cost, preds, labels]) + comp.update(preds=preds, labels=labels) + numpy_acc, numpy_recall = comp.eval() """ - def __init__(self, name=None, **kwargs): - super(CompositeMetric, self).__init__(name, kwargs) + def __init__(self, name=None): + super(CompositeMetric, self).__init__(name) self._metrics = [] def add_metric(self, metric): + """ + add one metric instance to CompositeMetric. + + Args: + metric: a instance of MetricBase. + """ if not isinstance(metric, MetricBase): raise ValueError("SubMetric should be inherit from MetricBase.") self._metrics.append(metric) + def update(self, preds, labels): + """ + Update every metrics in sequence. + + Args: + preds(numpy.array): the predictions of current minibatch + labels(numpy.array): the labels of current minibatch, if the label is one-hot + or soft-label, should custom the corresponding update rule. + """ + for m in self._metrics: + ans.append(m.update(preds, labels)) + def eval(self): + """ + Evaluate every metrics in sequence. + + Returns: + list(float|numpy.array): a list of metrics value in Python. + """ ans = [] for m in self._metrics: ans.append(m.eval()) return ans +class Precision(MetricBase): + """ + Precision (also called positive predictive value) is the fraction of + relevant instances among the retrieved instances. + https://en.wikipedia.org/wiki/Evaluation_of_binary_classifiers + + Note Precision is different with Accuracy in binary classifiers. + accuracy = true positive / total instances + precision = true positive / all positive instance + + Examples: + .. code-block:: python + + metric = fluid.metrics.Precision() + for pass in range(PASSES): + metric.reset() + for data in train_reader(): + loss, preds, labels = exe.run(fetch_list=[cost, preds, labels]) + metric.update(preds=preds, labels=labels) + numpy_precision = metric.eval() + """ + + def __init__(self, name=None): + super(Precision, self).__init__(name) + self.tp = 0 # true positive + self.fp = 0 # false positive + + def update(self, preds, labels): + if not _is_numpy_(preds): + raise ValueError("The 'preds' must be a numpy ndarray.") + if not _is_numpy_(labels): + raise ValueError("The 'labels' must be a numpy ndarray.") + sample_num = labels[0] + for i in range(sample_num): + pred = preds[i].astype("int32") + label = labels[i] + if label == 1: + if pred == label: + self.tp += 1 + else: + self.fp += 1 + + def eval(self): + ap = self.tp + self.fp + return float(self.tp) / ap if ap != 0 else .0 + + +class Recall(MetricBase): + """ + Recall (also known as sensitivity) is the fraction of + relevant instances that have been retrieved over the + total amount of relevant instances + + https://en.wikipedia.org/wiki/Precision_and_recall + + Examples: + .. code-block:: python + + metric = fluid.metrics.Recall() + for pass in range(PASSES): + metric.reset() + for data in train_reader(): + loss, preds, labels = exe.run(fetch_list=[cost, preds, labels]) + metric.update(preds=preds, labels=labels) + numpy_recall = metric.eval() + """ + + def __init__(self, name=None): + super(Recall, self).__init__(name) + self.tp = 0 # true positive + self.fn = 0 # false negtive + + def update(self, preds, labels): + if not _is_numpy_(preds): + raise ValueError("The 'preds' must be a numpy ndarray.") + if not _is_numpy_(labels): + raise ValueError("The 'labels' must be a numpy ndarray.") + sample_num = labels[0] + for i in range(sample_num): + pred = preds[i].astype("int32") + label = labels[i] + if label == 1: + if pred == label: + self.tp += 1 + else: + if pred != label: + self.fn += 1 + + def eval(self): + recall = self.tp + self.fn + return float(self.tp) / recall if recall != 0 else .0 + + class Accuracy(MetricBase): """ Accumulate the accuracy from minibatches and compute the average accuracy for every pass. + https://en.wikipedia.org/wiki/Accuracy_and_precision Args: name: the metrics name - Example: - minibatch_accuracy = fluid.layers.accuracy(pred, label) - accuracy_evaluator = fluid.metrics.Accuracy() - for epoch in PASS_NUM: - accuracy_evaluator.reset() - for data in batches: - loss = exe.run(fetch_list=[cost, minibatch_accuracy]) - accuracy_evaluator.update(value=minibatch_accuracy, weight=batches) - accuracy = accuracy_evaluator.eval() + Examples: + .. code-block:: python + + labels = fluid.layers.data(name="data", shape=[1], dtype="int32") + data = fluid.layers.data(name="data", shape=[32, 32], dtype="int32") + pred = fluid.layers.fc(input=data, size=1000, act="tanh") + minibatch_accuracy = fluid.layers.accuracy(pred, label) + accuracy_evaluator = fluid.metrics.Accuracy() + for pass in range(PASSES): + accuracy_evaluator.reset() + for data in train_reader(): + batch_size = data[0] + loss = exe.run(fetch_list=[cost, minibatch_accuracy]) + accuracy_evaluator.update(value=minibatch_accuracy, weight=batch_size) + numpy_acc = accuracy_evaluator.eval() """ def __init__(self, name=None): @@ -153,6 +326,13 @@ class Accuracy(MetricBase): self.weight = .0 def update(self, value, weight): + """ + Update minibatch states. + + Args: + value(float|numpy.array): accuracy of one minibatch. + weight(int|float): batch size. + """ if not _is_number_or_matrix_(value): raise ValueError( "The 'value' must be a number(int, float) or a numpy ndarray.") @@ -163,9 +343,8 @@ class Accuracy(MetricBase): def eval(self): if self.weight == 0: - raise ValueError( - "There is no data in Accuracy Metrics. Please check layers.accuracy output has added to Accuracy." - ) + raise ValueError("There is no data in Accuracy Metrics. \ + Please check layers.accuracy output has added to Accuracy.") return self.value / self.weight @@ -174,6 +353,25 @@ class ChunkEvaluator(MetricBase): Accumulate counter numbers output by chunk_eval from mini-batches and compute the precision recall and F1-score using the accumulated counter numbers. + For some basics of chunking, please refer to + 'Chunking with Support Vector Machines '. + ChunkEvalEvaluator computes the precision, recall, and F1-score of chunk detection, + and supports IOB, IOE, IOBES and IO (also known as plain) tagging schemes. + + Examples: + .. code-block:: python + + labels = fluid.layers.data(name="data", shape=[1], dtype="int32") + data = fluid.layers.data(name="data", shape=[32, 32], dtype="int32") + pred = fluid.layers.fc(input=data, size=1000, act="tanh") + precision, recall, f1_score, num_infer_chunks, num_label_chunks, num_correct_chunks = layers.chunk_eval( + input=pred, + label=label) + metric = fluid.metrics.ChunkEvaluator() + for data in train_reader(): + loss, preds, labels = exe.run(fetch_list=[cost, preds, labels]) + metric.update(num_infer_chunks, num_label_chunks, num_correct_chunks) + numpy_precision, numpy_recall, numpy_f1 = metric.eval() """ def __init__(self, name=None): @@ -183,9 +381,17 @@ class ChunkEvaluator(MetricBase): self.num_correct_chunks = 0 def update(self, num_infer_chunks, num_label_chunks, num_correct_chunks): + """ + Update the states based on the layers.chunk_eval() ouputs. + Args: + num_infer_chunks(int|numpy.array): The number of chunks in Inference on the given minibatch. + num_label_chunks(int|numpy.array): The number of chunks in Label on the given mini-batch. + num_correct_chunks(int|float|numpy.array): The number of chunks both in Inference and Label on the + given mini-batch. + """ if not _is_number_or_matrix_(num_infer_chunks): raise ValueError( - "The 'num_infer_chunks' must be a number(int, float) or a numpy ndarray." + "The 'num_infer_chunks' must be a number(int) or a numpy ndarray." ) if not _is_number_or_matrix_(num_label_chunks): raise ValueError( @@ -212,21 +418,28 @@ class ChunkEvaluator(MetricBase): class EditDistance(MetricBase): """ + Edit distance is a way of quantifying how dissimilar two strings + (e.g., words) are to one another by counting the minimum number + of operations required to transform one string into the other. + Refer to https://en.wikipedia.org/wiki/Edit_distance + Accumulate edit distance sum and sequence number from mini-batches and compute the average edit_distance and instance error of all batches. Args: name: the metrics name - Example: - edit_distance_metrics = fluid.layers.edit_distance(input, label) - distance_evaluator = fluid.metrics.EditDistance() - for epoch in PASS_NUM: - distance_evaluator.reset() - for data in batches: - loss = exe.run(fetch_list=[cost] + list(edit_distance_metrics)) - distance_evaluator.update(*edit_distance_metrics) - distance, instance_error = distance_evaluator.eval() + Examples: + .. code-block:: python + + distances, seq_num = fluid.layers.edit_distance(input, label) + distance_evaluator = fluid.metrics.EditDistance() + for epoch in PASS_NUM: + distance_evaluator.reset() + for data in batches: + loss = exe.run(fetch_list=[cost] + list(edit_distance_metrics)) + distance_evaluator.update(distances, seq_num) + distance, instance_error = distance_evaluator.eval() In the above example: 'distance' is the average of the edit distance in a pass. @@ -264,16 +477,38 @@ class EditDistance(MetricBase): class DetectionMAP(MetricBase): """ Calculate the detection mean average precision (mAP). - - TODO (Dang Qingqing): update the following doc. - The general steps are as follows: - 1. calculate the true positive and false positive according to the input - of detection and labels. - 2. calculate mAP value, support two versions: '11 point' and 'integral'. - + mAP is the metric to measure the accuracy of object detectors + like Faster R-CNN, SSD, etc. + It is the average of the maximum precisions at different recall values. Please get more information from the following articles: https://sanchom.wordpress.com/tag/average-precision/ + https://arxiv.org/abs/1512.02325 + + The general steps are as follows: + + 1. calculate the true positive and false positive according to the input + of detection and labels. + 2. calculate mAP value, support two versions: '11 point' and 'integral'. + + Examples: + .. code-block:: python + + pred = fluid.layers.fc(input=data, size=1000, act="tanh") + batch_map = layers.detection_map( + input, + label, + class_num, + background_label, + overlap_threshold=overlap_threshold, + evaluate_difficult=evaluate_difficult, + ap_version=ap_version) + metric = fluid.metrics.DetectionMAP() + for data in train_reader(): + loss, preds, labels = exe.run(fetch_list=[cost, batch_map]) + batch_size = data[0] + metric.update(value=batch_map, weight=batch_size) + numpy_map = metric.eval() """ def __init__(self, name=None): @@ -302,17 +537,18 @@ class DetectionMAP(MetricBase): class Auc(MetricBase): """ - Auc Metrics which adapts to binary classification. - Need to note that auc metrics compute the value via Python natively. + Auc metric adapts to the binary classification. + Refer to https://en.wikipedia.org/wiki/Receiver_operating_characteristic#Area_under_the_curve + Need to note that auc metric compute the value via Python natively. If you concern the speed, please use the fluid.layers.auc instead. The `auc` function creates four local variables, `true_positives`, - `true_negatives`, `false_positives` and `false_negatives` that are used to - compute the AUC. To discretize the AUC curve, a linearly spaced set of - thresholds is used to compute pairs of recall and precision values. The area - under the ROC-curve is therefore computed using the height of the recall - values by the false positive rate, while the area under the PR-curve is the - computed using the height of the precision values by the recall. + `true_negatives`, `false_positives` and `false_negatives` that are used to + compute the AUC. To discretize the AUC curve, a linearly spaced set of + thresholds is used to compute pairs of recall and precision values. The area + under the ROC-curve is therefore computed using the height of the recall + values by the false positive rate, while the area under the PR-curve is the + computed using the height of the precision values by the recall. Args: name: metric name @@ -322,22 +558,32 @@ class Auc(MetricBase): curve. "NOTE: only implement the ROC curve type via Python now." + + Examples: + .. code-block:: python + + pred = fluid.layers.fc(input=data, size=1000, act="tanh") + metric = fluid.metrics.Auc() + for data in train_reader(): + loss, preds, labels = exe.run(fetch_list=[cost, preds, labels]) + metric.update(preds, labels) + numpy_auc = metric.eval() """ def __init__(self, name, curve='ROC', num_thresholds=200): - super(MetricBase, self).__init__(name, curve, num_thresholds) + super(Auc, self).__init__(name=name) self._curve = curve self._num_thresholds = num_thresholds self._epsilon = 1e-6 - self.tp_list = np.ndarray((num_thresholds, )) - self.fn_list = np.ndarray((num_thresholds, )) - self.tn_list = np.ndarray((num_thresholds, )) - self.fp_list = np.ndarray((num_thresholds, )) + self.tp_list = np.zeros((num_thresholds, )) + self.fn_list = np.zeros((num_thresholds, )) + self.tn_list = np.zeros((num_thresholds, )) + self.fp_list = np.zeros((num_thresholds, )) - def update(self, labels, predictions, axis=1): + def update(self, preds, labels): if not _is_numpy_(labels): raise ValueError("The 'labels' must be a numpy ndarray.") - if not _is_numpy_(predictions): + if not _is_numpy_(preds): raise ValueError("The 'predictions' must be a numpy ndarray.") kepsilon = 1e-7 # to account for floating point imprecisions @@ -350,12 +596,12 @@ class Auc(MetricBase): tp, fn, tn, fp = 0, 0, 0, 0 for i, lbl in enumerate(labels): if lbl: - if predictions[i, 0] >= thresh: + if predictions[i, 1] >= thresh: tp += 1 else: fn += 1 else: - if predictions[i, 0] >= thresh: + if predictions[i, 1] >= thresh: fp += 1 else: tn += 1 diff --git a/python/paddle/fluid/nets.py b/python/paddle/fluid/nets.py index bbedf6fde0872fd32d81c103bf5fe61449b7f57b..9b3f2aebee73e56ee820dc8ff4c9cfabd1456aaa 100644 --- a/python/paddle/fluid/nets.py +++ b/python/paddle/fluid/nets.py @@ -26,16 +26,87 @@ def simple_img_conv_pool(input, filter_size, pool_size, pool_stride, - act, - param_attr=None, + pool_padding=0, pool_type='max', + global_pooling=False, + conv_stride=1, + conv_padding=0, + conv_dilation=1, + conv_groups=1, + param_attr=None, + bias_attr=None, + act=None, use_cudnn=True, use_mkldnn=False): + """ + The simple_img_conv_pool is composed with one Convolution2d and one Pool2d. + + Args: + input (Variable): The input image with [N, C, H, W] format. + num_filters(int): The number of filter. It is as same as the output + feature channel. + filter_size (int|list|tuple): The filter size. If filter_size is a list or + tuple, it must contain two integers, (filter_size_H, filter_size_W). Otherwise, + the filter_size_H = filter_size_W = filter_size. + pool_size (int|list|tuple): The pooling size of Pool2d layer. If pool_size + is a list or tuple, it must contain two integers, (pool_size_H, pool_size_W). + Otherwise, the pool_size_H = pool_size_W = pool_size. + pool_stride (int|list|tuple): The pooling stride of Pool2d layer. If pool_stride + is a list or tuple, it must contain two integers, (pooling_stride_H, pooling_stride_W). + Otherwise, the pooling_stride_H = pooling_stride_W = pool_stride. + pool_padding (int|list|tuple): The padding of Pool2d layer. If pool_padding is a list or + tuple, it must contain two integers, (pool_padding_H, pool_padding_W). + Otherwise, the pool_padding_H = pool_padding_W = pool_padding. Default 0. + pool_type (str): Pooling type can be :math:`max` for max-pooling and :math:`avg` for + average-pooling. Default :math:`max`. + global_pooling (bool): Whether to use the global pooling. If global_pooling = true, + pool_size and pool_padding while be ignored. Default False + conv_stride (int|list|tuple): The stride size of the Conv2d Layer. If stride is a + list or tuple, it must contain two integers, (conv_stride_H, conv_stride_W). Otherwise, + the conv_stride_H = conv_stride_W = conv_stride. Default: conv_stride = 1. + conv_padding (int|list|tuple): The padding size of the Conv2d Layer. If padding is + a list or tuple, it must contain two integers, (conv_padding_H, conv_padding_W). + Otherwise, the conv_padding_H = conv_padding_W = conv_padding. Default: conv_padding = 0. + conv_dilation (int|list|tuple): The dilation size of the Conv2d Layer. If dilation is + a list or tuple, it must contain two integers, (conv_dilation_H, conv_dilation_W). + Otherwise, the conv_dilation_H = conv_dilation_W = conv_dilation. Default: conv_dilation = 1. + conv_groups (int): The groups number of the Conv2d Layer. According to grouped + convolution in Alex Krizhevsky's Deep CNN paper: when group=2, + the first half of the filters is only connected to the first half + of the input channels, while the second half of the filters is only + connected to the second half of the input channels. Default: groups=1 + param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None + bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None + act (str): Activation type for Conv2d. Default: None + use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + use_mkldnn (bool): Use mkldnn kernels or not, it is valid only when compiled + with mkldnn library. Default: False + + Return: + Variable: The result of input after Convolution2d and Pool2d. + + Examples: + .. code-block:: python + + img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') + conv_pool = fluid.nets.simple_img_conv_pool(input=img, + filter_size=5, + num_filters=20, + pool_size=2, + pool_stride=2, + act="relu") + """ conv_out = layers.conv2d( input=input, num_filters=num_filters, filter_size=filter_size, + stride=conv_stride, + padding=conv_padding, + dilation=conv_dilation, + groups=conv_groups, param_attr=param_attr, + bias_attr=bias_attr, act=act, use_cudnn=use_cudnn, use_mkldnn=use_mkldnn) @@ -45,6 +116,8 @@ def simple_img_conv_pool(input, pool_size=pool_size, pool_type=pool_type, pool_stride=pool_stride, + pool_padding=pool_padding, + global_pooling=global_pooling, use_cudnn=use_cudnn, use_mkldnn=use_mkldnn) return pool_out @@ -60,11 +133,65 @@ def img_conv_group(input, conv_with_batchnorm=False, conv_batchnorm_drop_rate=0.0, pool_stride=1, - pool_type=None, + pool_type="max", use_cudnn=True, use_mkldnn=False): """ - Image Convolution Group, Used for vgg net. + The Image Convolution Group is composed of Convolution2d, BatchNorm, DropOut, + and Pool2d. According to the input arguments, img_conv_group will do serials of + computation for Input using Convolution2d, BatchNorm, DropOut, and pass the last + result to Pool2d. + + Args: + input (Variable): The input image with [N, C, H, W] format. + conv_num_filter(list|tuple): Indicates the numbers of filter of this group. + pool_size (int|list|tuple): The pooling size of Pool2d Layer. If pool_size + is a list or tuple, it must contain two integers, (pool_size_H, pool_size_W). + Otherwise, the pool_size_H = pool_size_W = pool_size. + conv_padding (int|list|tuple): The padding size of the Conv2d Layer. If padding is + a list or tuple, its length must be equal to the length of conv_num_filter. + Otherwise the conv_padding of all Conv2d Layers are the same. Default 1. + conv_filter_size (int|list|tuple): The filter size. If filter_size is a list or + tuple, its length must be equal to the length of conv_num_filter. + Otherwise the conv_filter_size of all Conv2d Layers are the same. Default 3. + conv_act (str): Activation type for Conv2d Layer that is not followed by BatchNorm. + Default: None. + param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None + conv_with_batchnorm (bool|list): Indicates whether to use BatchNorm after Conv2d Layer. + If conv_with_batchnorm is a list, its length must be equal to the length of + conv_num_filter. Otherwise, conv_with_batchnorm indicates whether all the + Conv2d Layer follows a BatchNorm. Default False. + conv_batchnorm_drop_rate (float|list): Indicates the drop_rate of Dropout Layer + after BatchNorm. If conv_batchnorm_drop_rate is a list, its length must be + equal to the length of conv_num_filter. Otherwise, drop_rate of all Dropout + Layers is conv_batchnorm_drop_rate. Default 0.0. + pool_stride (int|list|tuple): The pooling stride of Pool2d layer. If pool_stride + is a list or tuple, it must contain two integers, (pooling_stride_H, + pooling_stride_W). Otherwise, the pooling_stride_H = pooling_stride_W = pool_stride. + Default 1. + pool_type (str): Pooling type can be :math:`max` for max-pooling and :math:`avg` for + average-pooling. Default :math:`max`. + use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + use_mkldnn (bool): Use mkldnn kernels or not, it is valid only when compiled + with mkldnn library. Default: False + + Return: + Variable: The final result after serial computation using Convolution2d, + BatchNorm, DropOut, and Pool2d. + + Examples: + .. code-block:: python + + img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32') + conv_pool = fluid.nets.img_conv_group(input=img, + num_channels=3, + conv_padding=1, + conv_num_filter=[3, 3], + conv_filter_size=3, + conv_act="relu", + pool_size=2, + pool_stride=2) """ tmp = input assert isinstance(conv_num_filter, list) or \ @@ -74,6 +201,7 @@ def img_conv_group(input, if not hasattr(obj, '__len__'): return [obj] * len(conv_num_filter) else: + assert len(obj) == len(conv_num_filter) return obj conv_padding = __extend_list__(conv_padding) @@ -119,6 +247,39 @@ def sequence_conv_pool(input, param_attr=None, act="sigmoid", pool_type="max"): + """ + The sequence_conv_pool is composed with Sequence Convolution and Pooling. + + Args: + input (Variable): The input of sequence_conv, which supports variable-time + length input sequence. The underlying of input is a matrix with shape + (T, N), where T is the total time steps in this mini-batch and N is + the input_hidden_size + num_filters(int): The number of filter. + filter_size (int): The filter size. + param_attr (ParamAttr): The parameters to the Sequence_conv Layer. Default: None. + act (str): Activation type for Sequence_conv Layer. Default: "sigmoid". + pool_type (str): Pooling type can be :math:`max` for max-pooling, :math:`average` for + average-pooling, :math:`sum` for sum-pooling, :math:`sqrt` for sqrt-pooling. + Default :math:`max`. + + Return: + Variable: The final result after Sequence Convolution and Pooling. + + Examples: + .. code-block:: python + + input_dim = len(word_dict) + emb_dim = 128 + hid_dim = 512 + data = fluid.layers.data( ame="words", shape=[1], dtype="int64", lod_level=1) + emb = fluid.layers.embedding(input=data, size=[input_dim, emb_dim], is_sparse=True) + seq_conv = fluid.nets.sequence_conv_pool(input=emb, + num_filters=hid_dim, + filter_size=3, + act="tanh", + pool_type="sqrt") + """ conv_out = layers.sequence_conv( input=input, num_filters=num_filters, @@ -132,9 +293,9 @@ def sequence_conv_pool(input, def glu(input, dim=-1): """ - The gated linear unit composed by split, sigmoid activation and elementwise - multiplication. Specifically, Split the input into two equal sized parts - :math:`a` and :math:`b` along the given dimension and then compute as + The Gated Linear Units(GLU) composed by split, sigmoid activation and element-wise + multiplication. Specifically, Split the input into two equal sized parts, + :math:`a` and :math:`b`, along the given dimension and then compute as following: .. math:: @@ -147,16 +308,16 @@ def glu(input, dim=-1): Args: input (Variable): The input variable which is a Tensor or LoDTensor. dim (int): The dimension along which to split. If :math:`dim < 0`, the - dimension to split along is :math:`rank(input) + dim`. + dimension to split along is :math:`rank(input) + dim`. Default -1. Returns: - Variable: The Tensor variable with half the size of input. + Variable: Variable with half the size of input. Examples: .. code-block:: python - # x is a Tensor variable with shape [3, 6, 9] - fluid.nets.glu(input=x, dim=1) # shape of output: [3, 3, 9] + data = fluid.layers.data(name="words", shape=[3, 6, 9], dtype="float32") + output = fluid.nets.glu(input=data, dim=1) # shape of output: [3, 3, 9] """ a, b = layers.split(input, num_or_sections=2, dim=dim) @@ -189,40 +350,48 @@ def scaled_dot_product_attention(queries, `_. Args: - queries (Variable): The input variable which should be a 3-D Tensor. keys (Variable): The input variable which should be a 3-D Tensor. values (Variable): The input variable which should be a 3-D Tensor. num_heads (int): Head number to compute the scaled dot product - attention. Default value is 1. + attention. Default: 1. dropout_rate (float): The dropout rate to drop the attention weight. - Default value is 0. + Default: 0.0. Returns: - - Variable: A 3-D Tensor computed by multi-head scaled dot product \ - attention. + Variable: A 3-D Tensor computed by multi-head scaled dot product\ + attention. Raises: - ValueError: If input queries, keys, values are not 3-D Tensors. - NOTE: + NOTES: 1. When num_heads > 1, three linear projections are learned respectively - to map input queries, keys and values into queries', keys' and values'. - queries', keys' and values' have the same shapes with queries, keys - and values. - - 1. When num_heads == 1, scaled_dot_product_attention has no learnable - parameters. + to map input queries, keys and values into queries', keys' and values'. + queries', keys' and values' have the same shapes with queries, keys + and values. + 2. When num_heads == 1, scaled_dot_product_attention has no learnable + parameters. Examples: .. code-block:: python - # Suppose q, k, v are Tensors with the following shape: - # q: [3, 5, 9], k: [3, 6, 9], v: [3, 6, 10] - - contexts = fluid.nets.scaled_dot_product_attention(q, k, v) + queries = fluid.layers.data(name="queries", + shape=[3, 5, 9], + dtype="float32", + append_batch_size=False) + queries.stop_gradient = False + keys = fluid.layers.data(name="keys", + shape=[3, 6, 9], + dtype="float32", + append_batch_size=False) + keys.stop_gradient = False + values = fluid.layers.data(name="values", + shape=[3, 6, 10], + dtype="float32", + append_batch_size=False) + values.stop_gradient = False + contexts = fluid.nets.scaled_dot_product_attention(queries, keys, values) contexts.shape # [3, 5, 10] """ if not (len(queries.shape) == len(keys.shape) == len(values.shape) == 3): diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index 54fe9356275c313cd18fbb12edc9d35f38bda772..607a68e2565a247612f0e7b307088f85be91825c 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -26,10 +26,10 @@ from clip import append_gradient_clip_ops, error_clip_callback from contextlib import contextmanager __all__ = [ - 'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', + 'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'Ftrl', 'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer', 'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'RMSPropOptimizer', - 'Adadelta', 'ModelAverage', 'Optimizer' + 'FtrlOptimizer', 'Adadelta', 'ModelAverage', 'Optimizer', 'RMSPropOptimizer' ] @@ -192,15 +192,15 @@ class Optimizer(object): """Add optimization operators to update gradients to variables. Args: - loss: the target that this optimization is for. - parameters_and_grads: a list of (variable, gradient) pair to update. + loss(Variable): the target that this optimization is for. + parameters_and_grads(list(tuple(Variable, Variable))): + a list of (variable, gradient) pair to update. Returns: return_op_list: a list of operators that will complete one step of optimization. This will include parameter update ops, global step update ops and any other custom ops required by subclasses to manage their internal state. - :param startup_program: """ # This is a default implementation of create_optimization_pass that # can be shared by most optimizers. This implementation assumes that @@ -268,7 +268,22 @@ class Optimizer(object): class SGDOptimizer(Optimizer): - """ Simple SGD optimizer without any state. + """ + Optimizer of the stochastic gradient descent algorithm. + + .. math:: + + param\_out = param - learning\_rate * grad + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + + Examples: + .. code-block:: python + + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.2) + sgd_optimizer.minimize(cost) """ def __init__(self, learning_rate, **kwargs): @@ -294,7 +309,37 @@ class SGDOptimizer(Optimizer): class MomentumOptimizer(Optimizer): - """Simple Momentum optimizer with velocity state + """ + + Simple Momentum optimizer with velocity state + + This optimizer has a flag for Nestrov Momentum. + + The update equations are as follows: + + .. math:: + + & velocity = mu * velocity + gradient + + & if (use\_nesterov): + + &\quad param = param - gradient * learning\_rate + mu * velocity * learning\_rate + + & else: + + &\quad param = param - learning\_rate * velocity + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + momentum (float): momentum factor + use_nesterov (bool): enables Nesterov momentum + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.Momentum(learning_rate=0.2, momentum=0.1) + optimizer.minimize(cost) """ _velocity_acc_str = "velocity" @@ -338,7 +383,32 @@ class MomentumOptimizer(Optimizer): class AdagradOptimizer(Optimizer): - """Simple Adagrad optimizer with moment state + """ + **Adaptive Gradient Algorithm (Adagrad)** + + The update is done as follows: + + .. math:: + + moment\_out &= moment + grad * grad + + param\_out &= param - \\frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon} + + The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) + does not have the epsilon attribute. It is added here in our implementation + as also proposed here: http://cs231n.github.io/neural-networks-3/#ada + for numerical stability to avoid the division by zero error. + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + epsilon (float): a small float value for numerical stability. + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.Adagrad(learning_rate=0.2) + optimizer.minimize(cost) """ _moment_acc_str = "moment" @@ -379,7 +449,40 @@ class AdagradOptimizer(Optimizer): class AdamOptimizer(Optimizer): - """Implements the Adam Optimizer + """ + This implements the Adam optimizer from Section 2 of the Adam + paper : https://arxiv.org/abs/1412.6980. + Adam is a first-order gradient-based optimization method based on + adaptive estimates of lower-order moments. + + Adam updates: + + .. math:: + + t & = t + 1 + + moment\_1\_out & = {\\beta}_1 * moment\_1 + (1 - {\\beta}_1) * grad + + moment\_2\_out & = {\\beta}_2 * moment\_2 + (1 - {\\beta}_2) * grad * grad + + learning\_rate & = learning\_rate * \\ + \\frac{\sqrt{1 - {\\beta}_2^t}}{1 - {\\beta}_1^t} + + param\_out & = param - learning\_rate * \\frac{moment\_1}{\sqrt{moment\_2} + \epsilon} + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + beta1 (float): The exponential decay rate for the 1st moment estimates. + beta2 (float): The exponential decay rate for the 2nd moment estimates. + epsilon (float): a small float value for numerical stability. + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.Adam(learning_rate=0.2) + optimizer.minimize(cost) + """ _moment1_acc_str = "moment1" _moment2_acc_str = "moment2" @@ -484,7 +587,42 @@ class AdamOptimizer(Optimizer): class AdamaxOptimizer(Optimizer): - """Implements the Adamax Optimizer + """ + We implement the Adamax optimizer from Section 7 of the Adam + paper: https://arxiv.org/abs/1412.6980. Adamax is a variant of the + Adam algorithm based on the infinity norm. + + Adamax updates: + + .. math:: + + t & = t + 1 + + moment\_out & = {\\beta}_1 * moment + (1 - {\\beta}_1) * grad + + inf\_norm\_out & = max({\\beta}_2 * inf\_norm + \epsilon, |grad|) + + learning\_rate & = \\frac{learning\_rate}{1 - {\\beta}_1^t} + + param\_out & = param - learning\_rate * \\frac{moment\_out}{inf\_norm\_out} + + + The original paper does not have an epsilon attribute. + However, it is added here for numerical stability to prevent the + division by 0 error. + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + beta1 (float): The exponential decay rate for the 1st moment estimates. + beta2 (float): The exponential decay rate for the 2nd moment estimates. + epsilon (float): a small float value for numerical stability. + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.Adamax(learning_rate=0.2) + optimizer.minimize(cost) """ _moment_acc_str = "moment" _inf_norm_acc_str = "inf_norm" @@ -568,7 +706,34 @@ class AdamaxOptimizer(Optimizer): class DecayedAdagradOptimizer(Optimizer): - """Simple Decayed Adagrad optimizer with moment state + """ + **Decayed Adagrad Optimizer** + + The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) + + The update is done as follows: + + .. math:: + + moment\_out & = decay * moment + (1 - decay) * grad * grad + + param\_out & = param - \\frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon} + + The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) + does not have an epsilon attribute. It is added here for numerical + stability to avoid the division by zero error. + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + decay (float): decay rate. + epsilon (float): a small float value for numerical stability. + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.DecayedAdagrad(learning_rate=0.2) + optimizer.minimize(cost) """ _moment_acc_str = "moment" @@ -614,6 +779,7 @@ class DecayedAdagradOptimizer(Optimizer): class AdadeltaOptimizer(Optimizer): """ **Adadelta Optimizer** + Simple Adadelta optimizer with average squared grad state and average squared update state. The details of adadelta please refer to this @@ -628,7 +794,7 @@ class AdadeltaOptimizer(Optimizer): E(dx_t^2) &= \\rho * E(dx_{t-1}^2) + (1-\\rho) * (-g*learning\\_rate)^2 Args: - learning_rate(float): global leraning rate + learning_rate(float): global learning rate rho(float): rho in equation epsilon(float): epsilon in equation @@ -703,37 +869,37 @@ class RMSPropOptimizer(Optimizer): .. math:: - r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 \\\\ + r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 w & = w - \\frac{\\eta} {\\sqrt{r(w,t) + \\epsilon}} \\nabla Q_{i}(w) The first equation calculates moving average of the squared gradient for - each weight. Then dividing the gradient by :math: `sqrt{v(w,t)}`. + each weight. Then dividing the gradient by :math:`sqrt{v(w,t)}`. In some cases, adding a momentum term :math: `\\beta` is beneficial. In our implementation, Nesterov momentum is used: .. math:: - r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 \\\\ + r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 v(w, t) & = \\beta v(w, t-1) + \\frac{\\eta} {\\sqrt{v(w,t) + \\epsilon}} \\nabla Q_{i}(w) w & = w - v(w, t) - where, :math: `\\rho` is a hyperparameter and typical values are 0.9, 0.95 + where, :math:`\\rho` is a hyperparameter and typical values are 0.9, 0.95 and so on. :math: `beta` is the momentum term. :math: `\\epsilon` is a smoothing term to avoid division by zero, usually set somewhere in range from 1e-4 to 1e-8. Args: - learning_rate(float): global leraning rate. + learning_rate(float): global learning rate. rho(float): rho is :math: `\\rho` in equation, set 0.95 by default. epsilon(float): :math: `\\epsilon` in equation is smoothing term to avoid division by zero, set 1e-6 by default. - momentum(float): :math: `\\beta` in equation is the momentum term, + momentum(float): :math:`\\beta` in equation is the momentum term, set 0.0 by default. Raises: @@ -810,6 +976,113 @@ class RMSPropOptimizer(Optimizer): return rmsprop_op +class FtrlOptimizer(Optimizer): + """ + FTRL (Follow The Regularized Leader) Optimizer. + + The paper that proposed Follow The Regularized Leader (FTRL): + (https://www.eecs.tufts.edu/~dsculley/papers/ad-click-prediction.pdf) + + .. math:: + + &new\_accum = squared\_accum + grad^2 + + &if (lr\_power == -0.5): + + &\quad linear\_accum += grad - \\frac{\\sqrt{new\_accum} - \\sqrt{squared\_accum}}{learning\_rate * param} + + &else: + + &\quad linear\_accum += grad - \\frac{new\_accum^{-lr\_power} - accum^{-lr\_power}}{learning\_rate * param} + + + &x = l1 * sign(linear\_accum) - linear\_accum + + &if (lr\_power == -0.5): + + &\quad y = \\frac{\\sqrt{new\_accum}}{learning\_rate} + (2 * l2) + + &\quad pre\_shrink = \\frac{x}{y} + + &\quad param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0) + + &else: + + &\quad y = \\frac{new\_accum^{-lr\_power}}{learning\_rate} + (2 * l2) + + &\quad pre\_shrink = \\frac{x}{y} + + &\quad param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0) + + &squared\_accum += grad^2 + + Args: + learning_rate (float|Variable): global learning rate. + l1 (float): + l2 (float): + lr_power (float): + + Raises: + ValueError: If learning_rate, rho, epsilon, momentum are None. + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.Ftrl(0.0001) + _, params_grads = optimizer.minimize(cost) + """ + + _squared_acc_str = "squared" + _linear_acc_str = "linear" + + def __init__(self, learning_rate, l1=0.0, l2=0.0, lr_power=-0.5, **kwargs): + super(FtrlOptimizer, self).__init__( + learning_rate=learning_rate, **kwargs) + if learning_rate is None: + raise ValueError("learning_rate is not set.") + + self.type = "ftrl" + self._l1 = l1 + self._l2 = l2 + self._lr_power = lr_power + + def _create_accumulators(self, block, parameters): + if not isinstance(block, framework.Block): + raise TypeError("block is not instance of framework.Block.") + + for p in parameters: + self._add_accumulator(self._squared_acc_str, p) + self._add_accumulator(self._linear_acc_str, p) + + def _append_optimize_op(self, block, param_and_grad): + if not isinstance(block, framework.Block): + raise TypeError("block is not instance of framework.Block.") + + squared_acc = self._get_accumulator(self._squared_acc_str, + param_and_grad[0]) + linear_acc = self._get_accumulator(self._linear_acc_str, + param_and_grad[0]) + ftrl_op = block.append_op( + type=self.type, + inputs={ + "Param": param_and_grad[0], + "Grad": param_and_grad[1], + "SquaredAccumulator": squared_acc, + "LinearAccumulator": linear_acc, + "LearningRate": self._create_param_lr(param_and_grad), + }, + outputs={ + "ParamOut": param_and_grad[0], + "SquaredAccumOut": squared_acc, + "LinearAccumOut": linear_acc + }, + attrs={"l1": self._l1, + "l2": self._l1, + "lr_power": self._lr_power}) + + return ftrl_op + + # We short the class name, since users will use the optimizer with the package # name. The sample code: # @@ -826,6 +1099,7 @@ Adamax = AdamaxOptimizer DecayedAdagrad = DecayedAdagradOptimizer Adadelta = AdadeltaOptimizer RMSProp = RMSPropOptimizer +Ftrl = FtrlOptimizer class ModelAverage(Optimizer): @@ -844,7 +1118,9 @@ class ModelAverage(Optimizer): max_average_window: The maximum size of average window. Examples: - ... + + .. code-block:: python + optimizer = fluid.optimizer.Momentum() _, params_grads = optimizer.minimize(cost) model_average = fluid.optimizer.ModelAverage(params_grads, 0.15, diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index 0fdc9a035292b3390cece6c5821a60b1b281e54d..25cc1355d5a53e44b7f45c1f7d80673abcf567ec 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -27,6 +27,40 @@ BuildStrategy = core.ParallelExecutor.BuildStrategy class ParallelExecutor(object): + """ + ParallelExecutor can run program in parallel. + + Args: + use_cuda (bool): Whether to use CUDA or not. + loss_name (str): The loss name must set in training. Default None. + main_program (Program): The program that need to run, if not provided, + then default_main_program will be used. Default None. + share_vars_from(ParallelExecutor): If provied, it will share variables + from the specified ParallelExecutor. Default None. + num_trainers(int): If greater than 1, NCCL will be initialized with + multiple rank of nodes, each node should have same number of GPUs. + Distributed training will be enabled then. Default 1. + trainer_id(int: Must use together with num_trainers. trainer_id is the + "rank" of current node starts from 0. Default 0. + + Returns: + ParallelExecutor: The initialized ParallelExecutor object. + + Raises: + TypeError: If share_vars_from is provided, but not ParallelExecutor object. + + Examples: + .. code-block:: python + + train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=loss.name) + test_exe = fluid.ParallelExecutor(use_cuda=True, + main_program=test_program, + share_vars_from=train_exe) + + train_loss, = train_exe.run([loss.name], feed=feed_dict) + test_loss, = test_exe.run([loss.name], feed=feed_dict) + """ + def __init__(self, use_cuda, loss_name=None, @@ -37,42 +71,6 @@ class ParallelExecutor(object): num_trainers=1, trainer_id=0, **kwargs): - """ - ParallelExecutor can run program in parallel. - - Args: - use_cuda(bool): Whether to use CUDA or not. - 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. - share_vars_from(ParallelExecutor, default None): If provied, - it will share variables from the specified ParallelExecutor. - 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. - - Raises: - TypeError: If share_vars_from is provided, but not ParallelExecutor - object. - - Examples: - .. code-block:: python - - train_exe = fluid.ParallelExecutor( - use_cuda=True, loss_name=loss.name) - test_exe = fluid.ParallelExecutor( - use_cuda=True, - main_program=test_program, - share_vars_from=train_exe) - - 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: @@ -131,10 +129,16 @@ class ParallelExecutor(object): main = main_program main = main if main else framework.default_main_program() scope = executor.global_scope() + # FIXME(Yancey1989): it's a temporary approach to determinate the distribute + # train program, call self.bcast_param() at the end of each mini-batch. + self.is_dist = True if "recv" in [ + op.type for op in main.global_block().ops + ] else False if share_vars_from and not isinstance(share_vars_from, ParallelExecutor): raise TypeError("share_vars_from must be ParallelExecutor.") + local_scopes = share_vars_from.executor.local_scopes( ) if share_vars_from else [] @@ -166,12 +170,14 @@ class ParallelExecutor(object): element in the list will be copied to each device directly. For example, if the feed is a dict: + >>> exe = ParallelExecutor() >>> # the image will be splitted into devices. If there is two devices >>> # each device will process an image with shape (24, 1, 28, 28) >>> exe.run(feed={'image': numpy.random.random(size=(48, 1, 28, 28))}) For example, if the feed is a list: + >>> exe = ParallelExecutor() >>> # each device will process each element in the list. >>> # the 1st device will process an image with shape (48, 1, 28, 28) @@ -182,18 +188,40 @@ class ParallelExecutor(object): >>> {"image": numpy.random.random(size=(32, 1, 28, 28))}, >>> ]) - Args: fetch_list(list): The fetched variable names feed(list|dict|None): The feed variables. If the feed is a dict, tensors in that dict will be splitted into each devices. If the feed is a list, each element of the list will be copied - to each device. + to each device. Default None. feed_dict: Alias for feed parameter, for backward compatibility. - This parameter is deprecated. + This parameter has been deprecated. Default None. + + Returns: + List: The fetched result list. + + Raises: + ValueError: If the feed is a list, but its length is not equal the + length of active places, or its element's is not dict. + + NOTES: + 1. If the feed's type is dict, the number of data that feeds to + ParallelExecutor must be bigger than active places. Otherwise, + it will throw exception from C++ side. Special attention should be + paid to check whether the last batch of the dataset is bigger + than active places. + 2. If active places are more than one, the fetch results for each + variable is a list, and each element of this list is the variable of + respective active place. - Returns: fetched result list. + Examples: + .. code-block:: python + pe = fluid.ParallelExecutor(use_cuda=use_cuda, + loss_name=avg_cost.name, + main_program=fluid.default_main_program()) + loss = pe.run(feed=feeder.feed(cur_batch), + fetch_list=[avg_cost.name])) """ if feed is None and feed_dict is not None: feed = feed_dict @@ -238,9 +266,17 @@ class ParallelExecutor(object): fetch_var_name = '@FETCHED_VAR_NAME@' self.executor.run(fetch_list, fetch_var_name) arr = self.scope.find_var(fetch_var_name).get_lod_tensor_array() + + if self.is_dist: + self.bcast_params() + return [arr[i] for i in range(len(arr))] def bcast_params(self): + """ + Broadcast the parameters to other devices. It is used during + distributed training. + """ self.executor.bcast_params(set(self.persistable_vars)) @property diff --git a/python/paddle/fluid/param_attr.py b/python/paddle/fluid/param_attr.py index 1c6970441bccdc1c1221503256c30c83502bd123..0a42b9fca8dba7a11b414990be6c04c93158864f 100644 --- a/python/paddle/fluid/param_attr.py +++ b/python/paddle/fluid/param_attr.py @@ -22,6 +22,35 @@ __all__ = [ class ParamAttr(object): + """ + Parameter attributes object. To fine-tuning network training process, user + can set parameter's attributes to control training details. Such as learning rate, + regularization, trainable, do_model_average and the method to initialize param. + + + Args: + name(str): The parameter's name. Default None. + initializer(Initializer): The method to initial this parameter. Default None. + learning_rate(float): The parameter's learning rate. The learning rate when + optimize is :math:`global\_lr * parameter\_lr * scheduler\_factor`. + Default 1.0. + regularizer(WeightDecayRegularizer): Regularization factor. Default None. + trainable(bool): Whether this parameter is trainable. Default True. + gradient_clip(BaseGradientClipAttr): The method to clip this parameter's + gradient. Default None. + do_model_average(bool): Whether this parameter should do model average. + Default False. + + Examples: + .. code-block:: python + + w_param_attrs = fluid.ParamAttr(name="fc_weight", + learning_rate=0.5, + regularizer=fluid.L2Decay(1.0), + trainable=True) + y_predict = fluid.layers.fc(input=x, size=10, param_attr=w_param_attrs) + """ + def __init__(self, name=None, initializer=None, @@ -29,7 +58,7 @@ class ParamAttr(object): regularizer=None, trainable=True, gradient_clip=None, - do_model_average=None): + do_model_average=False): self.name = name self.initializer = initializer self.learning_rate = learning_rate @@ -39,6 +68,16 @@ class ParamAttr(object): self.model_average = do_model_average def set_default_initializer(self, initializer): + """ + Set the default initializer, the initializer should be Constant, + Uniform, Normal, Xavier, MSRA. + + Args: + initializer(Initializer): the initializer to set. + + Returns: + None + """ if initializer is None: if self.initializer is None: raise ValueError("ParamAttr.initializer is not set") @@ -50,13 +89,45 @@ class ParamAttr(object): self.initializer = initializer def set_default_param_initializer(self): + """ + Set the default initializer for the parameter with Xavier. + + Args: + None. + + Returns: + None. + """ self.set_default_initializer(Xavier()) def set_default_bias_initializer(self): + """ + Set the default initializer for the bias with Constant(0.0). + + Args: + None. + + Returns: + None. + """ self.set_default_initializer(Constant(0.0)) @staticmethod def to_attr(arg): + """ + Create ParamAttr[s]. + + Args: + arg: Arguments to initialize ParamAttr[s]. arg's type can be + str, Initializer, float, WeightDecayRegularizer, BaseGradientClipAttr, + bool, ParamAttr, or a list of above type. + + Returns: + ParamAttr[s]: ParamAttr[s] initialized with arg. + + Raises: + arg can not initialize a ParamAttr. + """ if arg is None: return ParamAttr() elif isinstance(arg, list) or isinstance(arg, tuple): @@ -75,6 +146,15 @@ class ParamAttr(object): raise TypeError("{0} cast to ParamAttr".format(type(arg))) def to_kwargs(self, with_initializer=False): + """ + Returns the attributes of this parameter. + + Args: + with_initializer(bool): Whether to add initializer attr. + + Returns: + Parameter attributes(map): The attributes of this parameter. + """ kwargs = { 'name': self.name, 'optimize_attr': { @@ -92,9 +172,27 @@ class ParamAttr(object): class WeightNormParamAttr(ParamAttr): """ - Used for weight normalization. Any field in ParamAttr can also be set here. - Besides, an extra field dim can be set to indicate the dimension except - which to normalize. + Used for weight Norm. Weight Norm is a reparameterization of the weight vectors + in a neural network that decouples the length of those weight vectors from + their direction. Weight Norm has been implemented as discussed in this + paper: `Weight Normalization: A Simple Reparameterization to Accelerate + Training of Deep Neural Networks + `_. + + Args: + dim(list): The parameter's name. Default None. + kwargs: Any field in ParamAttr. Default None. + + Examples: + .. code-block:: python + + data = fluid.layers.data(name="data", shape=[3, 32, 32], dtype="float32") + fc = fluid.layers.fc(input=data, + size=1000, + param_attr=WeightNormParamAttr( + dim=None, + name='weight_norm_param')) + """ # List to record the parameters reparameterized by weight normalization. # If these parameters are treated as Variable rather than Parameter, diff --git a/python/paddle/fluid/profiler.py b/python/paddle/fluid/profiler.py index e2bd1d4c9a1ea5ddc0dfd19c769dcb40bfd6d04c..6a321ae024dcb50452bc4d96d7e7e70f590a42c6 100644 --- a/python/paddle/fluid/profiler.py +++ b/python/paddle/fluid/profiler.py @@ -42,6 +42,9 @@ def cuda_profiler(output_file, output_mode=None, config=None): counters/options for profiling by `config` argument. The default config is ['gpustarttimestamp', 'gpustarttimestamp', 'gridsize3d', 'threadblocksize', 'streamid', 'enableonstart 0', 'conckerneltrace']. + Then users can use NVIDIA Visual Profiler + (https://developer.nvidia.com/nvidia-visual-profiler) tools to load this + this output file to visualize results. Args: output_file (string) : The output file name, the result will be @@ -50,6 +53,33 @@ def cuda_profiler(output_file, output_mode=None, config=None): Comma separated values format. It should be 'kvp' or 'csv'. config (list of string) : The profiler options and counters can refer to "Compute Command Line Profiler User Guide". + + Raises: + ValueError: If `output_mode` is not in ['kvp', 'csv']. + + Examples: + + .. code-block:: python + + import paddle.fluid as fluid + import paddle.fluid.profiler as profiler + + epoc = 8 + dshape = [4, 3, 28, 28] + data = fluid.layers.data(name='data', shape=[3, 28, 28], dtype='float32') + conv = fluid.layers.conv2d(data, 20, 3, stride=[1, 1], padding=[1, 1]) + + place = fluid.CUDAPlace(0) + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + + output_file = 'cuda_profiler.txt' + with profiler.cuda_profiler(output_file, 'csv') as nvprof: + for i in range(epoc): + input = np.random.random(dshape).astype('float32') + exe.run(fluid.default_main_program(), feed={'data': input}) + # then use NVIDIA Visual Profiler (nvvp) to load this output file + # to visualize results. """ if output_mode is None: output_mode = 'csv' @@ -69,19 +99,52 @@ def cuda_profiler(output_file, output_mode=None, config=None): def reset_profiler(): - """The profiler clear interface. - reset_profiler will clear the previous time record. + """ + Clear the previous time record. This interface does not work for + `fluid.profiler.cuda_profiler`, it only works for + `fluid.profiler.start_profiler`, `fluid.profiler.stop_profiler`, + and `fluid.profiler.profiler`. + + Examples: + + .. code-block:: python + + import paddle.fluid.profiler as profiler + with profiler.profiler(state, 'total', '/tmp/profile'): + for iter in range(10): + if iter == 2: + profiler.reset_profiler() + # ... """ core.reset_profiler() def start_profiler(state): - """Enable the profiler. + """ + Enable the profiler. Uers can use `fluid.profiler.start_profiler` and + `fluid.profiler.stop_profiler` to insert the code, except the usage of + `fluid.profiler.profiler` interface. Args: state (string) : The profiling state, which should be 'CPU', 'GPU' or 'All'. 'CPU' means only profile CPU. 'GPU' means profiling GPU as well. 'All' also generates timeline. + + Raises: + ValueError: If `state` is not in ['CPU', 'GPU', 'All']. + + Examples: + + .. code-block:: python + + import paddle.fluid.profiler as profiler + + profiler.start_profiler('GPU') + for iter in range(10): + if iter == 2: + profiler.reset_profiler() + # except each iteration + profiler.stop_profiler('total', '/tmp/profile') """ if core.is_profiler_enabled(): return @@ -97,7 +160,10 @@ def start_profiler(state): def stop_profiler(sorted_key=None, profile_path='/tmp/profile'): - """Stop the profiler. + """ + Stop the profiler. Uers can use `fluid.profiler.start_profiler` and + `fluid.profiler.stop_profiler` to insert the code, except the usage of + `fluid.profiler.profiler` interface. Args: sorted_key (string) : If None, the profiling results will be printed @@ -111,6 +177,23 @@ def stop_profiler(sorted_key=None, profile_path='/tmp/profile'): The `ave` means sorting by the average execution time. profile_path (string) : If state == 'All', it will write a profile proto output file. + + Raises: + ValueError: If `sorted_key` is not in + ['calls', 'total', 'max', 'min', 'ave']. + + Examples: + + .. code-block:: python + + import paddle.fluid.profiler as profiler + + profiler.start_profiler('GPU') + for iter in range(10): + if iter == 2: + profiler.reset_profiler() + # except each iteration + profiler.stop_profiler('total', '/tmp/profile') """ if not core.is_profiler_enabled(): return @@ -137,7 +220,12 @@ def profiler(state, sorted_key=None, profile_path='/tmp/profile'): Different from cuda_profiler, this profiler can be used to profile both CPU and GPU program. By defalut, it records the CPU and GPU operator kernels, if you want to profile other program, you can refer the profiling tutorial - to add more records. + to add more records in C++ code. + + If the state == 'All', a profile proto file will be written to + `profile_path`. This file records timeline information during the execution. + Then users can visualize this file to see the timeline, please refer + https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/howto/optimization/timeline.md Args: state (string) : The profiling state, which should be 'CPU' or 'GPU', @@ -156,6 +244,25 @@ def profiler(state, sorted_key=None, profile_path='/tmp/profile'): The `ave` means sorting by the average execution time. profile_path (string) : If state == 'All', it will write a profile proto output file. + + Raises: + ValueError: If `state` is not in ['CPU', 'GPU', 'All']. If `sorted_key` is + not in ['calls', 'total', 'max', 'min', 'ave']. + + Examples: + + .. code-block:: python + + import paddle.fluid.profiler as profiler + + with profiler.profiler('All', 'total', '/tmp/profile') as prof: + for pass_id in range(pass_num): + for batch_id, data in enumerate(train_reader()): + exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[], + use_program_cache=True) + # ... """ start_profiler(state) yield diff --git a/python/paddle/fluid/recordio_writer.py b/python/paddle/fluid/recordio_writer.py index 8d48e9abef0fb9861284c6302b30efb0e3994989..bd57772713057f12b876942de58ee43527e94834 100644 --- a/python/paddle/fluid/recordio_writer.py +++ b/python/paddle/fluid/recordio_writer.py @@ -36,6 +36,45 @@ def convert_reader_to_recordio_file( compressor=core.RecordIOWriter.Compressor.Snappy, max_num_records=1000, feed_order=None): + """ + Convert a Python Reader to a recordio file. + + Please see :ref:`api_guide_python_reader` and :ref:`api_guide_reader_op` for + details. + + Examples: + + >>> import paddle.fluid as fluid + >>> import paddle.dataset.mnist as mnist + >>> import paddle + >>> + >>> tmp_program = fluid.Program() + >>> with fluid.program_guard(tmp_program): + >>> img = fluid.layers.data(name='img', shape=[784]) + >>> label = fluid.layers.data(name='label', shape=[1], dtype='int64') + >>> feeder = fluid.DataFeeder(feed_list=[img, label], place=fluid.CPUPlace()) + >>> # mnist.recordio will be generated in current directory + >>> fluid.recordio_writer.convert_reader_to_recordio_file( + >>> filename="mnist.recordio", + >>> reader_creator=paddle.batch(mnist.train(), batch_size=32), + >>> feeder=feeder) + + Args: + filename(str): The recordio filename. + reader_creator(callable): The Python Reader Creator. See + :ref:`api_guide_python_reader`. + feeder(DataFeeder): The DataFeeder instance. Used to convert + :code:`reader_creator` to :code: `lod_tensor` + compressor: Must in fluid.core.RecordIOWriter.Compressor.Snappy or + fluid.core.RecordIOWriter.Compressor.NoCompress. Use :code:`Snappy` + by default. + max_num_records(int): Maximum number of records in one chuck. Each record + is each return value from reader function + feed_order(list): The order of variable names that the reader returns + + Returns: + int: the number of record that saved. + """ if feed_order is None: feed_order = feeder.feed_names counter = 0 @@ -58,6 +97,17 @@ def convert_reader_to_recordio_files( compressor=core.RecordIOWriter.Compressor.Snappy, max_num_records=1000, feed_order=None): + """ + convert a python reader to many recordio files. + + This API is basically same as :code:`convert_reader_to_recordio_file`, + instead of it will create many recordio files. Each file contains at + most :code:`batch_per_file` records. + + Please reference + :ref:`api_fluid_recordio_writer_convert_reader_to_recordio_file` for more + details. + """ if feed_order is None: feed_order = feeder.feed_names f_name, f_ext = os.path.splitext(filename) diff --git a/python/paddle/fluid/regularizer.py b/python/paddle/fluid/regularizer.py index c4d6829599616cb3ea7791a189e7070974de6ae3..dac474d5ee76590a75311d6bf2c4cb2fe85b6c40 100644 --- a/python/paddle/fluid/regularizer.py +++ b/python/paddle/fluid/regularizer.py @@ -16,8 +16,8 @@ import framework from . import core __all__ = [ - 'append_regularization_ops', 'WeightDecayRegularizer', 'L1Decay', 'L2Decay', - 'L1DecayRegularizer', 'L2DecayRegularizer' + 'append_regularization_ops', 'L1Decay', 'L2Decay', 'L1DecayRegularizer', + 'L2DecayRegularizer' ] @@ -36,7 +36,8 @@ def append_regularization_ops(parameters_and_grads, regularization=None): set. It will be applied with regularizer. Returns: - list of (parameters, gradients) pair with the regularized gradient + list[(Variable, Variable)]: list of (parameters, gradients) \ + pair with the regularized gradient Raises: Exception: Unknown regularization type @@ -100,6 +101,24 @@ class WeightDecayRegularizer(object): class L2DecayRegularizer(WeightDecayRegularizer): """Implements the L2 Weight Decay Regularization + + Small values of L2 can help prevent over fitting the training data. + + .. math:: + + L2WeightDecay = reg\_coeff * parameter + + Args: + regularization_coeff(float): regularization coeff + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.Adagrad( + learning_rate=1e-4, + regularization=fluid.regularizer.L2DecayRegularizer( + regularization_coeff=0.1)) + optimizer.minimize(avg_cost) """ def __init__(self, regularization_coeff=0.0): @@ -154,6 +173,27 @@ class L2DecayRegularizer(WeightDecayRegularizer): class L1DecayRegularizer(WeightDecayRegularizer): """Implements the L1 Weight Decay Regularization + + L1 regularization encourages sparsity. + + .. math:: + + L1WeightDecay = reg\_coeff * sign(parameter) + + Args: + regularization_coeff(float): regularization coeff + + Examples: + .. code-block:: python + + program = fluid.framework.Program() + block = program.global_block() + mul_x = block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="mul.x", + regularizer=fluid.regularizer.L1DecayRegularizer(0.5)) """ def __init__(self, regularization_coeff=0.0): diff --git a/python/paddle/fluid/tests/book/notest_understand_sentiment.py b/python/paddle/fluid/tests/book/notest_understand_sentiment.py index c6687e8ad7fcc45c82d6dcb2256e9055a81cc61c..5d9a47c9ba3db07f240b42732536f1ea37627a11 100644 --- a/python/paddle/fluid/tests/book/notest_understand_sentiment.py +++ b/python/paddle/fluid/tests/book/notest_understand_sentiment.py @@ -194,16 +194,16 @@ def train(word_dict, if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": 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 b1a6b524d33cae97c8982ffb8f780b1b07761a09..74f96f456a8dc917b715d0f4308bb5ea41947f0b 100644 --- a/python/paddle/fluid/tests/book/test_fit_a_line.py +++ b/python/paddle/fluid/tests/book/test_fit_a_line.py @@ -69,16 +69,16 @@ def train(use_cuda, save_dirname, is_local): if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": diff --git a/python/paddle/fluid/tests/book/test_image_classification.py b/python/paddle/fluid/tests/book/test_image_classification.py index 0f3a4c9242a81a3c1fb90268245715a8e59a207a..a2fb186b86c9706ac1aff0de49defbfb06e2eb0f 100644 --- a/python/paddle/fluid/tests/book/test_image_classification.py +++ b/python/paddle/fluid/tests/book/test_image_classification.py @@ -178,16 +178,16 @@ def train(net_type, use_cuda, save_dirname, is_local): if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": 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 99d51ae0076178aca50e36c2c187257a8ba1cbf2..e214ced0b5593c60ebd4a69edff1e961bcb4a72a 100644 --- a/python/paddle/fluid/tests/book/test_label_semantic_roles.py +++ b/python/paddle/fluid/tests/book/test_label_semantic_roles.py @@ -209,16 +209,16 @@ def train(use_cuda, save_dirname=None, is_local=True): if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": diff --git a/python/paddle/fluid/tests/book/test_machine_translation.py b/python/paddle/fluid/tests/book/test_machine_translation.py index 23e5900f127a7a3253c551f8f7fbceba08382209..372d6ec8223f69b69663137a646ba591108c40b7 100644 --- a/python/paddle/fluid/tests/book/test_machine_translation.py +++ b/python/paddle/fluid/tests/book/test_machine_translation.py @@ -200,16 +200,16 @@ def train_main(use_cuda, is_sparse, is_local=True): if is_local: train_loop(framework.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": diff --git a/python/paddle/fluid/tests/book/test_recognize_digits.py b/python/paddle/fluid/tests/book/test_recognize_digits.py index 25bcb8a64103b845adbe2017120ce8d945faf6dd..5f5c8544bbdb87421f129b201a0ebaf4cb8602a1 100644 --- a/python/paddle/fluid/tests/book/test_recognize_digits.py +++ b/python/paddle/fluid/tests/book/test_recognize_digits.py @@ -151,16 +151,16 @@ def train(nn_type, if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": diff --git a/python/paddle/fluid/tests/book/test_recommender_system.py b/python/paddle/fluid/tests/book/test_recommender_system.py index 65d6552acc9b3d31a97a45290e4613a633fffa3c..937d8dd5b065f0c1fdfc052b0342b572e3fbd7ac 100644 --- a/python/paddle/fluid/tests/book/test_recommender_system.py +++ b/python/paddle/fluid/tests/book/test_recommender_system.py @@ -220,16 +220,16 @@ def train(use_cuda, save_dirname, is_local=True): if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": diff --git a/python/paddle/fluid/tests/book/test_word2vec.py b/python/paddle/fluid/tests/book/test_word2vec.py index 3118d88701e5f64ae50f7ee774ea8174aa7758eb..75bed06bd7a9b311ff9466589d6ecab2c37471ce 100644 --- a/python/paddle/fluid/tests/book/test_word2vec.py +++ b/python/paddle/fluid/tests/book/test_word2vec.py @@ -125,16 +125,16 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True): if is_local: train_loop(fluid.default_main_program()) else: - port = os.getenv("PADDLE_INIT_PORT", "6174") - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip... + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... eplist = [] for ip in pserver_ips.split(","): eplist.append(':'.join([ip, port])) pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS")) + trainers = int(os.getenv("PADDLE_TRAINERS")) current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) - training_role = os.getenv("TRAINING_ROLE", "TRAINER") + trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) + training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") t = fluid.DistributeTranspiler() t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt index 21182393bd68db4a379fc3ecf83fc85d27ca9490..219ab9bc2cc74a3c16f7bda69d4d782283574d7e 100644 --- a/python/paddle/fluid/tests/unittests/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt @@ -15,7 +15,7 @@ if(NOT WITH_DISTRIBUTE) endif(NOT WITH_DISTRIBUTE) list(REMOVE_ITEM TEST_OPS test_seq_concat_op) # FIXME(helin): https://github.com/PaddlePaddle/Paddle/issues/8290 -list(REMOVE_ITEM TEST_OPS test_modified_huber_loss_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5184 +list(REMOVE_ITEM TEST_OPS test_modified_huber_loss_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5184 list(REMOVE_ITEM TEST_OPS test_lstm_unit_op) # # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5185 list(REMOVE_ITEM TEST_OPS test_nce) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/7778 list(REMOVE_ITEM TEST_OPS test_recurrent_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/6152 @@ -43,8 +43,6 @@ list(REMOVE_ITEM TEST_OPS test_warpctc_op) list(REMOVE_ITEM TEST_OPS test_dist_train) list(REMOVE_ITEM TEST_OPS test_parallel_executor_crf) list(REMOVE_ITEM TEST_OPS test_parallel_executor_fetch_feed) -# TODO(wuyi): this test hungs on CI, will add it back later -list(REMOVE_ITEM TEST_OPS test_listen_and_serv_op) foreach(TEST_OP ${TEST_OPS}) py_test_modules(${TEST_OP} MODULES ${TEST_OP}) endforeach(TEST_OP) @@ -52,3 +50,4 @@ py_test_modules(test_warpctc_op MODULES test_warpctc_op ENVS FLAGS_warpctc_dir=$ py_test_modules(test_dist_train MODULES test_dist_train SERIAL) py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SERIAL) py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL) +set_tests_properties(test_listen_and_serv_op PROPERTIES TIMEOUT 20) diff --git a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py index 87c11e7880e73b911f21dda77c1cc2b4850b3591..b04f25ef874cc6204211a4f5f5991a0ec8c473dd 100644 --- a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py +++ b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py @@ -15,6 +15,7 @@ import unittest import numpy as np from op_test import OpTest +import paddle.fluid.core as core def bilinear_interp_np(input, out_h, out_w, out_size): @@ -45,9 +46,9 @@ def bilinear_interp_np(input, out_h, out_w, out_size): out[:, :, i, j] = h2lambda*(w2lambda*input[:, :, h, w] + w1lambda*input[:, :, h, w+wid]) + \ - h1lambda*(w2lambda*input[:, :, h+hid, w] + - w1lambda*input[:, :, h+hid, w+wid]) - return out.astype("float32") + h1lambda*(w2lambda*input[:, :, h+hid, w] + + w1lambda*input[:, :, h+hid, w+wid]) + return out.astype(input.dtype) class TestBilinearInterpOp(OpTest): @@ -122,5 +123,44 @@ class TestCase6(TestBilinearInterpOp): self.out_size = np.array([65, 129]).astype("int32") +class TestBilinearInterpOpUint8(OpTest): + def setUp(self): + self.out_size = None + self.init_test_case() + self.op_type = "bilinear_interp" + input_np = np.random.randint( + low=0, high=256, size=self.input_shape).astype("uint8") + output_np = bilinear_interp_np(input_np, self.out_h, self.out_w, + self.out_size) + self.inputs = {'X': input_np} + if self.out_size is not None: + self.inputs['OutSize'] = self.out_size + self.attrs = {'out_h': self.out_h, 'out_w': self.out_w} + self.outputs = {'Out': output_np} + + def test_check_output(self): + self.check_output_with_place(place=core.CPUPlace(), atol=1) + + def init_test_case(self): + self.input_shape = [1, 3, 9, 6] + self.out_h = 10 + self.out_w = 9 + + +class TestCase1Uint8(TestBilinearInterpOpUint8): + def init_test_case(self): + self.input_shape = [2, 3, 128, 64] + self.out_h = 120 + self.out_w = 50 + + +class TestCase2Uint8(TestBilinearInterpOpUint8): + def init_test_case(self): + self.input_shape = [4, 1, 7, 8] + self.out_h = 5 + self.out_w = 13 + self.out_size = np.array([6, 15]).astype("int32") + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_concat_op.py b/python/paddle/fluid/tests/unittests/test_concat_op.py index 1e00d67d5480bfa77a60e1aed52cafac6e8242ca..e9f3c45dc40b3333fe7304f8e4313d156bd5374c 100644 --- a/python/paddle/fluid/tests/unittests/test_concat_op.py +++ b/python/paddle/fluid/tests/unittests/test_concat_op.py @@ -43,7 +43,7 @@ class TestConcatOp(OpTest): self.axis = 1 -class TestConcatOp2(OpTest): +class TestConcatOp2(TestConcatOp): def init_test_data(self): self.x0 = np.random.random((2, 3, 4, 5)).astype('float32') self.x1 = np.random.random((2, 3, 4, 5)).astype('float32') @@ -51,5 +51,16 @@ class TestConcatOp2(OpTest): self.axis = 1 +class TestConcatOp3(TestConcatOp): + def init_test_data(self): + self.x0 = np.random.random((1, 256, 170, 256)).astype('float32') + self.x1 = np.random.random((1, 128, 170, 256)).astype('float32') + self.x2 = np.random.random((1, 128, 170, 256)).astype('float32') + self.axis = 1 + + def test_check_grad(self): + pass + + if __name__ == '__main__': unittest.main() diff --git a/paddle/contrib/tape/CMakeLists.txt b/python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py similarity index 50% rename from paddle/contrib/tape/CMakeLists.txt rename to python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py index 5450359d859de93ca19c56422f1243c7f445aff7..3ae877a60818744f852d3af9a02ffebf5e2affc8 100644 --- a/paddle/contrib/tape/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py @@ -1,25 +1,26 @@ -# 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. -# -if(APPLE) - set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move") -endif(APPLE) +import unittest + +from test_gaussian_random_op import TestGaussianRandomOp + + +class TestMKLDNN(TestGaussianRandomOp): + def init_kernel_type(self): + self.use_mkldnn = True -cc_library(tape_variable SRCS variable.cc DEPS ${FLUID_CORE_MODULES} device_context framework_proto proto_desc operator) -cc_library(tape SRCS tape.cc DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB} tape_variable) -cc_test(test_tape - SRCS test_tape.cc - DEPS tape tape_variable) +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py b/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py index 272caceaf38699438ccae41691bf26b2eb4d2a22..8481500fd78f0ccf34f09c66bec27e195b9aada3 100644 --- a/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py +++ b/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py @@ -25,7 +25,15 @@ class TestGaussianRandomOp(unittest.TestCase): def setUp(self): self.op_type = "gaussian_random" self.inputs = {} - self.attrs = {"shape": [1000, 784], "mean": .0, "std": 1., "seed": 10} + self.use_mkldnn = False + self.init_kernel_type() + self.attrs = { + "shape": [1000, 784], + "mean": .0, + "std": 1., + "seed": 10, + "use_mkldnn": self.use_mkldnn + } self.outputs = ["Out"] @@ -58,6 +66,9 @@ class TestGaussianRandomOp(unittest.TestCase): self.assertAlmostEqual(numpy.mean(tensor), .0, delta=0.1) self.assertAlmostEqual(numpy.std(tensor), 1., delta=0.1) + def init_kernel_type(self): + pass + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index f8cf6f4e2d25c0c03a3a73dca8e6bc1990b3b78b..82074955fae7514d556ba9319c11beb250c4de11 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -401,6 +401,15 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(output) print(str(program)) + def test_maxout(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[3, 5], dtype="float32") + y = layers.data(name='y', shape=[2, 3], dtype="float32") + output = layers.crop(x, shape=y) + self.assertIsNotNone(output) + print(str(program)) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py index 9dec2acb1d7101f8f00565c56e0469edb143d0c6..1cdc69501043d120b9e3cc8ccda3a1212d205886 100644 --- a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py +++ b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py @@ -94,7 +94,7 @@ class TestListenAndServOp(OpTest): self._wait_ps_ready(p1.pid) # raise SIGTERM to pserver - os.kill(p1.pid, signal.SIGKILL) + os.kill(p1.pid, signal.SIGINT) p1.join() # run pserver on CPU in async mode @@ -102,7 +102,7 @@ class TestListenAndServOp(OpTest): self._wait_ps_ready(p2.pid) # raise SIGTERM to pserver - os.kill(p2.pid, signal.SIGKILL) + os.kill(p2.pid, signal.SIGTERM) p2.join() diff --git a/python/paddle/fluid/tests/unittests/test_optimizer.py b/python/paddle/fluid/tests/unittests/test_optimizer.py index e775db1d10f4561b6fb90631757a25c9f74cb777..7286c7c450108c4b5ad7136041bc4e989894a2ba 100644 --- a/python/paddle/fluid/tests/unittests/test_optimizer.py +++ b/python/paddle/fluid/tests/unittests/test_optimizer.py @@ -434,5 +434,71 @@ class TestDecayedAdagradOptimizer(unittest.TestCase): self.assertAlmostEqual(init_ops[1].attr('value'), 0.0) +class TestFtrlOptimizer(unittest.TestCase): + class MockFtrl(optimizer.FtrlOptimizer): + def get_accumulators(self): + return self._accumulators + + def get_squared_str(self): + return self._squared_acc_str + + def get_linear_str(self): + return self._linear_acc_str + + def test_ftrl_optimizer(self): + init_program = framework.Program() + program = framework.Program() + block = program.global_block() + mul_x = block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="mul.x", + optimize_attr={'learning_rate': 1.1}) + mul_y = block.create_var( + dtype="float32", shape=[10, 8], lod_level=0, name="mul.y") + mul_out = block.create_var( + dtype="float32", shape=[5, 8], lod_level=0, name="mul.out") + block.append_op( + type="mul", + inputs={"X": mul_x, + "Y": mul_y}, + outputs={"Out": mul_out}, + attrs={"x_num_col_dims": 1}) + mean_out = block.create_var( + dtype="float32", shape=[1], lod_level=0, name="mean.out") + block.append_op( + type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out}) + learning_rate = 0.01 + ftrl_optimizer = self.MockFtrl( + learning_rate=learning_rate, l1=0.0, l2=0.0, lr_power=-0.5) + params_grads = append_backward(mean_out) + self.assertEqual(len(params_grads), 1) + self.assertEqual(len(ftrl_optimizer.get_accumulators()), 0) + opts = ftrl_optimizer.create_optimization_pass(params_grads, mul_out, + init_program) + self.assertEqual(len(opts), 3) + self.assertEqual([op.type for op in opts], + ["fill_constant", "elementwise_mul", "ftrl"]) + + # Check accumulators + accumulators = ftrl_optimizer.get_accumulators() + self.assertEqual(len(accumulators), 2) + self.assertTrue(ftrl_optimizer.get_squared_str() in accumulators) + self.assertTrue(ftrl_optimizer.get_linear_str() in accumulators) + squared_acc = accumulators[ftrl_optimizer.get_squared_str()] + linear_acc = accumulators[ftrl_optimizer.get_linear_str()] + self.assertEqual(len(squared_acc), 1) + self.assertEqual(len(linear_acc), 1) + self.assertTrue(mul_x.name in squared_acc) + self.assertTrue(mul_x.name in linear_acc) + + # Check init_program + init_ops = init_program.global_block().ops + self.assertEqual(len(init_ops), 3) + self.assertEqual(init_ops[0].type, "fill_constant") + self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate) + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py new file mode 100644 index 0000000000000000000000000000000000000000..7956897d68a3fb49d62ba696d0b6400b4f909989 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py @@ -0,0 +1,26 @@ +# 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 + +from test_sum_op import TestSumOp + + +class TestMKLDNN(TestSumOp): + def init_kernel_type(self): + self.use_mkldnn = True + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_sum_op.py b/python/paddle/fluid/tests/unittests/test_sum_op.py index 2faf5b10647a1fa1d44e4847f017db177ee8808a..1d90414e137a70e6265042e24e106fe565802778 100644 --- a/python/paddle/fluid/tests/unittests/test_sum_op.py +++ b/python/paddle/fluid/tests/unittests/test_sum_op.py @@ -20,12 +20,15 @@ from op_test import OpTest class TestSumOp(OpTest): def setUp(self): self.op_type = "sum" + self.use_mkldnn = False + self.init_kernel_type() x0 = np.random.random((3, 4)).astype('float32') x1 = np.random.random((3, 4)).astype('float32') x2 = np.random.random((3, 4)).astype('float32') self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]} y = x0 + x1 + x2 self.outputs = {'Out': y} + self.attrs = {'use_mkldnn': self.use_mkldnn} def test_check_output(self): self.check_output() @@ -33,6 +36,9 @@ class TestSumOp(OpTest): def test_check_grad(self): self.check_grad(['x0'], 'Out') + def init_kernel_type(self): + pass + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/trainer.py b/python/paddle/fluid/trainer.py index efc28d899304b01a3085891f3ae9396d57c589a1..45ab889beaa1355d0e1e2922aedf0340f70809ba 100644 --- a/python/paddle/fluid/trainer.py +++ b/python/paddle/fluid/trainer.py @@ -33,23 +33,59 @@ __all__ = [ class BeginEpochEvent(object): + """ + The begin of a training epoch. + + Args: + epoch_id(int): The current epoch ID. + """ + def __init__(self, epoch_id): self.epoch = epoch_id class EndEpochEvent(object): + """ + The end of a training epoch. + + Args: + epoch_id(int): The current epoch ID. + """ + def __init__(self, epoch_id): self.epoch = epoch_id class BeginStepEvent(object): + """ + The begin of a training epoch. + + Args: + epoch_id(int): The current epoch ID. + step_id(int): The current step ID. + """ + def __init__(self, epoch_id, step_id): self.epoch = epoch_id self.step = step_id self.fetch_metrics = True + """ + If fetch_metrics is true, the metrics will be fetched at the + EndStepEvent. Default is True. + """ class EndStepEvent(object): + """ + The end of a training step. + + Args: + epoch_id(int): The current epoch ID. + step_id(int): The current step ID. + metrics(list): A list of fetched tensor. The order of this list is same + as the :code:`train_func` returns. + """ + def __init__(self, epoch_id, step_id, metrics): self.epoch = epoch_id self.step = step_id @@ -57,6 +93,27 @@ class EndStepEvent(object): class CheckpointConfig(object): + """ + Parameter object for :code:`fluid.io.save_checkpoint` and + :code:`fluid.Trainer`. Used to configuration how to save checkpoint. + + Args: + checkpoint_dir(str): Directory path to save check point. Default is the + current directory. + + max_num_checkpoints(int): The max number of local check points. + epoch_interval(int): Every number of epoch to save check point. + step_interval(int): Every number of step to save check point. + + Examples: + >>> config = fluid.CheckpointConfig("./checkpoints") + >>> trainer = fluid.Trainer(train_func=train_program, + >>> place=place, + >>> optimizer_func=optimizer_func, + >>> checkpoint_config=config) + >>> trainer.train(...) + """ + def __init__(self, checkpoint_dir=None, max_num_checkpoints=3, @@ -113,11 +170,62 @@ def check_and_get_place(place): class Trainer(object): """ + A trainer wraps MultiGPU/MultiNode training loops and can be used to train a + simple neural network easily. + + This API takes a :code:`train_func`. A :code:`train_func` is a function that + return loss as it first return value. The reset value can be fetched by + EndStepEvent.metrics + + This API also takes a :code:`optimizer_func` that will return an optimizer + instance. + + For example, to train a MLP for MNIST dataset, the sample program is + + >>> import paddle.fluid as fluid + >>> + >>> def mlp(image, layer_sizes=[200, 100], activation="relu", num_classes=10): + >>> hidden = image + >>> for layer_size in layer_sizes: + >>> hidden = fluid.layers.fc(input=hidden, size=layer_size, act=activation) + >>> return fluid.layers.fc(input=hidden, size=num_classes, act="softmax") + >>> + >>> def train_mnist_mlp(): + >>> img = fluid.layers.data(name='image', shape=[784]) + >>> label = fluid.layers.data(name='label', shape=[1], dtype='int64') + >>> prediction = mlp(img) + >>> return fluid.layers.mean(fluid.layers.cross_entropy(prediction, label)) + >>> + >>> def optimizer(): + >>> return fluid.optimizer.Adam() + >>> + >>> trainer = Trainer(train_func=train_mnist_mlp, + >>> optimizer_func=optimizer, + >>> place=fluid.CUDAPlace(0), + >>> parallel=True) + >>> + >>> def train_callback(event): + >>> if isinstance(event, fluid.EndStepEvent): + >>> print "Epoch ID", event.epoch, "Step ID",\ + >>> event.step, "AvgLoss", event.metrics[0] + >>> elif isinstance(event, fluid.EndEpochEvent): + >>> trainer.save_params("./model_{0}".format(event.epoch)) + >>> + >>> trainer.train(num_epochs=100, event_handler=train_callback) + + For more example, please see :ref:`api_guide_high_level_api`. + Args: - train_func(callable): A function which will return loss. The loss must be a scalar. + train_func(callable): A function which will return loss. The loss must be + a scalar tensor. optimizer_func(callable): A function that returns an Optimizer object. - place: The device place of this trainer. + place(CUDAPlace|CPUPlace): The device place of this trainer. If + :code:`parallel=True,` all CUDA Places will be used if :code:`place` + is a :code:`CUDAPlace`. + parallel(bool): True if use multiple devices. + checkpoint_config(CheckpointConfig): Configuration about how to save + checkpoints. """ def __init__(self, @@ -129,9 +237,6 @@ class Trainer(object): checkpoint_config=None): 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 # config for checkpoint # only chief worker will save variables @@ -145,6 +250,10 @@ class Trainer(object): self.scope = core.Scope() + # 1. we need to generate a framework.Program by calling + # program_func. Reference: fluid.program_guard in + # test_word2vec.py + self.startup_program = framework.Program() self.train_program = framework.Program() @@ -277,17 +386,18 @@ class Trainer(object): def train(self, num_epochs, event_handler, reader=None, feed_order=None): """ - Train the model. + Start the train loop to train the model. Args: - 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: - feed_order: Feeding order of reader. None will following the defining + num_epochs(int): The number of epoch. An epoch will process all data in reader + event_handler(callable): The event handler. A function with type (ev:Event)->void + reader(callable): A reader creator object. See also + :ref:`api_guide_python_reader` . + feed_order(list): Feeding order of reader. None will following the defining order in program Returns: - + None """ training_role = os.getenv("PADDLE_TRAINING_ROLE", "") if training_role == "PSERVER": @@ -307,16 +417,24 @@ class Trainer(object): Test the model on given test data Args: - reader: The reader that yields test data. - feed_order: Feeding order of reader. None will following the defining - order in program + reader(callable): The reader that yields test data. + feed_order(list): 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 + """ + Save all parameters into :code:`param_path`. + + Args: + param_path(str): The path to save parameters. + + Returns: + None + """ with self._prog_and_scope_guard(): exe = executor.Executor(self.place) io.save_persistables(exe, dirname=param_path) diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index 9c604170b8b53c9cbcf39b4978ae60ccad84648c..bb61f82a9cf7f837f0403082165a2375d18b574e 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -12,19 +12,11 @@ # See the License for the specific language governing permissions and # limitations under the License. """ -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 +4. append send_op to send splited variables to server and 5. add recv_op to fetch params(splited blocks or origin param) from server. 6. append concat_op to merge splited blocks to update local weights. @@ -44,7 +36,7 @@ import numpy as np from ps_dispatcher import RoundRobin, HashName, PSDispatcher from .. import core, framework from ..framework import Program, default_main_program, \ - default_startup_program, \ + default_startup_program, Block, \ Variable, Parameter, grad_var_name from details import * @@ -117,129 +109,41 @@ def slice_variable(var_list, slice_count, min_block_size=8192): return blocks -class DistributeTranspiler: - def _has_distributed_lookup_table(self): - # process lookup_table_op - # 1. check all lookup_table_op is distributed - # 2. check all lookup_table_op share the same table. - distributed_lookup_table_ops = [] - # support only one distributed_lookup_table now - self.table_name = None - for op in self.origin_program.global_block().ops: - if op.type == LOOKUP_TABLE_TYPE: - if op.attrs['is_distributed'] is True: - if self.table_name is None: - self.table_name = op.input("W")[0] - if self.table_name != op.input("W")[0]: - raise RuntimeError("all distributed lookup_table_ops" - " should have only one table") - distributed_lookup_table_ops.append(op) - else: - if self.table_name is not None: - assert op.input("W")[0] != self.table_name - - return len(distributed_lookup_table_ops) > 0 - - def _update_dist_lookup_table_vars(self, param_list, grad_list, - params_grads): - # TODO(wuyi): put find a way to put dist lookup table stuff all together. - # update self.table_param_grad and self.trainer_side_table_grad_list - program = self.origin_program - if self.has_distributed_lookup_table: - param_list = [ - param for param in param_list if param.name != self.table_name - ] - grad_list = [ - grad for grad in grad_list - if grad.name != grad_var_name(self.table_name) - ] - self.table_param_grad = [ - param_grad for param_grad in params_grads - if param_grad[0].name == self.table_name - ][0] - table_grad_var = self.table_param_grad[1] - if self.sync_mode: - self.trainer_side_table_grad_list = [ - program.global_block().create_var( - name="%s.trainer_%d.pserver_%d" % - (table_grad_var.name, self.trainer_id, index), - type=table_grad_var.type, - shape=table_grad_var.shape, - dtype=table_grad_var.dtype) - for index in range(len(self.pserver_endpoints)) - ] - else: - self.trainer_side_table_grad_list = [ - program.global_block().create_var( - name="%s.pserver_%d" % (table_grad_var.name, index), - type=table_grad_var.type, - shape=table_grad_var.shape, - dtype=table_grad_var.dtype) - for index in range(len(self.pserver_endpoints)) - ] - return param_list, grad_list - - def _init_splited_vars(self, slice_var_up): - # update these mappings for further transpile: - # 1. param_var_mapping: param var name -> [splited params vars] - # 2. grad_var_mapping: grad var name -> [splited grads vars] - # 3. grad_param_mapping: grad.blockx -> param.blockx - # 4. param_grad_ep_mapping: ep -> {"params": [], "grads": []} - - param_list = [] - grad_list = [] - param_grad_set = set() - for p, g in self.params_grads: - # skip parameter marked not trainable - if type(p) == Parameter and p.trainable == False: - continue - if p.name not in param_grad_set: - param_list.append(p) - param_grad_set.add(p.name) - if g.name not in param_grad_set: - grad_list.append(g) - param_grad_set.add(g.name) - - param_list, grad_list = self._update_dist_lookup_table_vars( - param_list, grad_list, self.params_grads) - - if slice_var_up: - # when we slice var up into blocks, we will slice the var according to - # pserver services' count. A pserver may have two or more listening ports. - grad_blocks = slice_variable(grad_list, len(self.pserver_endpoints)) - param_blocks = slice_variable(param_list, - len(self.pserver_endpoints)) - else: - # when we do NOT slice var up into blocks, we will always slice params - # grads into one block. - grad_blocks = slice_variable(grad_list, 1) - param_blocks = slice_variable(param_list, 1) - assert (len(grad_blocks) == len(param_blocks)) - - # origin_varname -> [splited_var] - self.param_var_mapping = self._create_vars_from_blocklist( - self.origin_program, param_blocks) - self.grad_var_mapping = self._create_vars_from_blocklist( - self.origin_program, - grad_blocks, - add_trainer_suffix=self.trainer_num > 1) - self.grad_param_mapping = dict() - for g, p in zip(grad_blocks, param_blocks): - g_name, g_bid, _ = g.split(":") - p_name, p_bid, _ = p.split(":") - self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] = \ - self.param_var_mapping[p_name][int(p_bid)] - - # create mapping of endpoint -> split var to create pserver side program - self.param_grad_ep_mapping = dict() - [ - self.param_grad_ep_mapping.update({ - ep: { - "params": [], - "grads": [] - } - }) for ep in self.pserver_endpoints - ] +class DistributeTranspiler(object): + """ + **DistributeTranspiler** + + Convert the fluid 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. + + Examples: + .. code-block:: python + + # Define your model before these codes. + port = os.getenv("PADDLE_PSERVER_PORT", "6174") + pserver_ips = os.getenv("PADDLE_PSERVER_IPS", "") + eplist = [] + for ip in pserver_ips.split(","): + eplist.append(':'.join([ip, port])) + pserver_endpoints = ",".join(eplist) + trainers = int(os.getenv("PADDLE_TRAINERS")) + current_endpoint = os.getenv("PADDLE_CURRENT_IP", "") + ":" + port + trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0")) + role = os.getenv("PADDLE_TRAINING_ROLE") + + t = distribute_transpiler.DistributeTranspiler() + t.transpile( + trainer_id, pservers=pserver_endpoints, trainers=trainers) + if role == "PSERVER": + pserver_program = t.get_pserver_program(current_endpoint) + pserver_startup_program = t.get_startup_program(current_endpoint, + pserver_program) + elif role == "TRAINER": + trainer_program = t.get_trainer_program() + """ def transpile(self, trainer_id, @@ -250,20 +154,20 @@ class DistributeTranspiler: split_method=RoundRobin, sync_mode=True): """ - :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. + Run the transpiler. + + Args: + trainer_id (int): id for current trainer worker, if you have + n workers, the id may range from 0 ~ n-1 + program (Program|None): program to transpile, + default is fluid.default_main_program(). + pservers (str): comma separated ip:port string for the pserver + list. + trainers (int): number of trainers in the distributed job. + slice_var_up (bool): Do Tensor slice for pservers, default is True. + split_method (PSDispatcher): RoundRobin or HashName can be used + try to choose the best method to balance loads for pservers. + sync_mode (bool): Do sync training or not, default is True. """ assert (split_method.__bases__[0] == PSDispatcher) if program is None: @@ -390,6 +294,12 @@ class DistributeTranspiler: self._split_table_grad_and_add_send_vars(program, pserver_endpoints) def get_trainer_program(self): + """ + Get transpiled trainer side program. + + Returns: + Program: trainer side program. + """ # remove optimize ops and add a send op to main_program delete_ops(self.origin_program.global_block(), self.optimize_ops) # FIXME(typhoonzero): serialize once will fix error occurs when clone. @@ -398,12 +308,19 @@ class DistributeTranspiler: def get_pserver_program(self, endpoint): """ - Get pserver side program using the endpoint. - TODO(panyx0718): Revisit this assumption. what if #blocks > #pservers. - NOTE: assume blocks of the same variable is not distributed - on the same pserver, only change param/grad varnames for - trainers to fetch. + Get parameter server side program. + + Args: + endpoint (str): current parameter server endpoint. + + Returns: + Program: the program for current parameter server to run. """ + # TODO(panyx0718): Revisit this assumption. what if #blocks > #pservers. + # NOTE: assume blocks of the same variable is not distributed + # on the same pserver, only change param/grad varnames for + # trainers to fetch. + # step1 pserver_program = Program() # step2: Create vars to receive vars at parameter servers. @@ -471,7 +388,7 @@ class DistributeTranspiler: self._append_pserver_ops(block, op, endpoint, grad_to_block_id, self.origin_program, merged_var) else: - self._append_pserver_non_opt_ops(block, op, endpoint) + self._append_pserver_non_opt_ops(block, op) def __op_have_grad_input__(op): for varname in op.input_arg_names: @@ -479,19 +396,50 @@ class DistributeTranspiler: return varname return "" + def __clone_lr_op_sub_block__(op, program, lr_block): + if not op.has_attr('sub_block'): + return + + origin_block_desc = op.attr('sub_block') + origin_block = self.origin_program.block(origin_block_desc.id) + assert isinstance(origin_block, Block) + # we put the new sub block to new block to follow the block + # hierarchy of the original blocks + new_sub_block = program.create_block(lr_block.idx) + + # clone vars + for var in origin_block.vars: + new_sub_block.clone_variable(var) + + # clone ops + for origin_op in origin_block.ops: + cloned_op = self._clone_lr_op(program, new_sub_block, origin_op) + # clone sub_block of op + __clone_lr_op_sub_block__(cloned_op, program, new_sub_block) + + # reset the block of op + op.set_attr('sub_block', new_sub_block) + # append lr decay ops to the child block if exists lr_ops = self._get_lr_ops() + # record optimize blocks and we can run them on pserver parallel + optimize_blocks = [] if len(lr_ops) > 0: lr_decay_block = pserver_program.create_block( pserver_program.num_blocks - 1) + optimize_blocks.append(lr_decay_block) for _, op in enumerate(lr_ops): - self._append_pserver_non_opt_ops(lr_decay_block, op, endpoint) + cloned_op = self._append_pserver_non_opt_ops(lr_decay_block, op) + # append sub blocks to pserver_program in lr_decay_op + __clone_lr_op_sub_block__(cloned_op, pserver_program, + lr_decay_block) # append op to the current block grad_to_block_id = [] pre_block_idx = pserver_program.num_blocks - 1 for idx, opt_op in enumerate(opt_op_on_pserver): per_opt_block = pserver_program.create_block(pre_block_idx) + optimize_blocks.append(per_opt_block) # append grad merging ops before clip and weight decay for _, op in enumerate(self.optimize_ops): # find the origin @GRAD var before clipping @@ -510,6 +458,7 @@ class DistributeTranspiler: if global_ops: opt_state_block = pserver_program.create_block( pserver_program.num_blocks - 1) + optimize_blocks.append(opt_state_block) for glb_op in global_ops: __append_optimize_op__(glb_op, opt_state_block, grad_to_block_id, None) @@ -531,11 +480,11 @@ class DistributeTranspiler: assert len(prefetch_var_name_to_block_id) == 0 attrs = { - "OptimizeBlock": pserver_program.block(1), + "optimize_blocks": optimize_blocks, "endpoint": endpoint, "Fanin": self.trainer_num, "sync_mode": self.sync_mode, - "grad_to_block_id": grad_to_block_id + "grad_to_block_id": grad_to_block_id, } if len(prefetch_var_name_to_block_id) > 0: attrs['prefetch_var_name_to_block_id'] \ @@ -556,6 +505,14 @@ class DistributeTranspiler: Get startup program for current parameter server. Modify operator input variables if there are variables that were split to several blocks. + + Args: + endpoint (str): current pserver endpoint. + pserver_program (Program): call get_pserver_program first and + pass the result here. + + Returns: + Program: parameter server side startup program. """ s_prog = Program() orig_s_prog = default_startup_program() @@ -607,6 +564,129 @@ class DistributeTranspiler: # ====================== private transpiler functions ===================== + def _has_distributed_lookup_table(self): + # process lookup_table_op + # 1. check all lookup_table_op is distributed + # 2. check all lookup_table_op share the same table. + distributed_lookup_table_ops = [] + # support only one distributed_lookup_table now + self.table_name = None + for op in self.origin_program.global_block().ops: + if op.type == LOOKUP_TABLE_TYPE: + if op.attrs['is_distributed'] is True: + if self.table_name is None: + self.table_name = op.input("W")[0] + if self.table_name != op.input("W")[0]: + raise RuntimeError("all distributed lookup_table_ops" + " should have only one table") + distributed_lookup_table_ops.append(op) + else: + if self.table_name is not None: + assert op.input("W")[0] != self.table_name + + return len(distributed_lookup_table_ops) > 0 + + def _update_dist_lookup_table_vars(self, param_list, grad_list, + params_grads): + # TODO(wuyi): put find a way to put dist lookup table stuff all together. + # update self.table_param_grad and self.trainer_side_table_grad_list + program = self.origin_program + if self.has_distributed_lookup_table: + param_list = [ + param for param in param_list if param.name != self.table_name + ] + grad_list = [ + grad for grad in grad_list + if grad.name != grad_var_name(self.table_name) + ] + self.table_param_grad = [ + param_grad for param_grad in params_grads + if param_grad[0].name == self.table_name + ][0] + table_grad_var = self.table_param_grad[1] + if self.sync_mode: + self.trainer_side_table_grad_list = [ + program.global_block().create_var( + name="%s.trainer_%d.pserver_%d" % + (table_grad_var.name, self.trainer_id, index), + type=table_grad_var.type, + shape=table_grad_var.shape, + dtype=table_grad_var.dtype) + for index in range(len(self.pserver_endpoints)) + ] + else: + self.trainer_side_table_grad_list = [ + program.global_block().create_var( + name="%s.pserver_%d" % (table_grad_var.name, index), + type=table_grad_var.type, + shape=table_grad_var.shape, + dtype=table_grad_var.dtype) + for index in range(len(self.pserver_endpoints)) + ] + return param_list, grad_list + + def _init_splited_vars(self, slice_var_up): + # update these mappings for further transpile: + # 1. param_var_mapping: param var name -> [splited params vars] + # 2. grad_var_mapping: grad var name -> [splited grads vars] + # 3. grad_param_mapping: grad.blockx -> param.blockx + # 4. param_grad_ep_mapping: ep -> {"params": [], "grads": []} + + param_list = [] + grad_list = [] + param_grad_set = set() + for p, g in self.params_grads: + # skip parameter marked not trainable + if type(p) == Parameter and p.trainable == False: + continue + if p.name not in param_grad_set: + param_list.append(p) + param_grad_set.add(p.name) + if g.name not in param_grad_set: + grad_list.append(g) + param_grad_set.add(g.name) + + param_list, grad_list = self._update_dist_lookup_table_vars( + param_list, grad_list, self.params_grads) + + if slice_var_up: + # when we slice var up into blocks, we will slice the var according to + # pserver services' count. A pserver may have two or more listening ports. + grad_blocks = slice_variable(grad_list, len(self.pserver_endpoints)) + param_blocks = slice_variable(param_list, + len(self.pserver_endpoints)) + else: + # when we do NOT slice var up into blocks, we will always slice params + # grads into one block. + grad_blocks = slice_variable(grad_list, 1) + param_blocks = slice_variable(param_list, 1) + assert (len(grad_blocks) == len(param_blocks)) + + # origin_varname -> [splited_var] + self.param_var_mapping = self._create_vars_from_blocklist( + self.origin_program, param_blocks) + self.grad_var_mapping = self._create_vars_from_blocklist( + self.origin_program, + grad_blocks, + add_trainer_suffix=self.trainer_num > 1) + self.grad_param_mapping = dict() + for g, p in zip(grad_blocks, param_blocks): + g_name, g_bid, _ = g.split(":") + p_name, p_bid, _ = p.split(":") + self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] = \ + self.param_var_mapping[p_name][int(p_bid)] + + # create mapping of endpoint -> split var to create pserver side program + self.param_grad_ep_mapping = dict() + [ + self.param_grad_ep_mapping.update({ + ep: { + "params": [], + "grads": [] + } + }) for ep in self.pserver_endpoints + ] + # transpiler function for dis lookup_table def _replace_lookup_table_op_with_prefetch(self, program, pserver_endpoints): @@ -798,7 +878,8 @@ class DistributeTranspiler: table_opt_block.append_op( type="sum", inputs={"X": pserver_side_table_grad_list}, - outputs={"Out": [grad_var]}) + outputs={"Out": [grad_var]}, + attrs={"use_mkldnn": False}) else: # in async_mode, for table gradient, it also need to be splited to each parameter server origin_grad_name = grad_var.name @@ -1030,7 +1111,8 @@ class DistributeTranspiler: optimize_block.append_op( type="sum", inputs={"X": vars2merge}, - outputs={"Out": merged_var}) + outputs={"Out": merged_var}, + attrs={"use_mkldnn": False}) # TODO(panyx0718): What if it's SELECTED_ROWS. if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS: optimize_block.append_op( @@ -1116,7 +1198,29 @@ class DistributeTranspiler: break return grad_block - def _append_pserver_non_opt_ops(self, optimize_block, opt_op, endpoint): + def _clone_lr_op(self, program, block, op): + inputs = self._get_input_map_from_op( + self.origin_program.global_block().vars, op) + for key, varlist in inputs.iteritems(): + if not isinstance(varlist, list): + varlist = [varlist] + for var in varlist: + if var not in program.global_block().vars: + block.clone_variable(var) + + outputs = self._get_output_map_from_op( + self.origin_program.global_block().vars, op) + for key, varlist in outputs.iteritems(): + if not isinstance(varlist, list): + varlist = [varlist] + for var in varlist: + if var not in program.global_block().vars: + block.clone_variable(var) + + return block.append_op( + type=op.type, inputs=inputs, outputs=outputs, attrs=op.attrs) + + def _append_pserver_non_opt_ops(self, optimize_block, opt_op): program = optimize_block.program # Append the ops for parameters that do not need to be optimized/updated inputs = self._get_input_map_from_op( @@ -1151,7 +1255,7 @@ class DistributeTranspiler: elif not program.global_block().vars.has_key(var.name): program.global_block().clone_variable(var) - optimize_block.append_op( + return optimize_block.append_op( type=opt_op.type, inputs=inputs, outputs=outputs, diff --git a/python/paddle/fluid/transpiler/inference_transpiler.py b/python/paddle/fluid/transpiler/inference_transpiler.py index 202aa76084432b4b2378470919b2e924301f2130..0629f2916b339a6cd19ccadf435a67a17d6da4cc 100644 --- a/python/paddle/fluid/transpiler/inference_transpiler.py +++ b/python/paddle/fluid/transpiler/inference_transpiler.py @@ -19,16 +19,30 @@ from ..executor import global_scope class InferenceTranspiler: + ''' + Convert the fluid program to optimized inference program. + + There are several optimizations, only fuse batch normalization is supported now. + + Examples: + + .. code-block:: python + + # As InferenceTranspiler will modify the original program, + # please clone before use it. + inference_transpiler_program = program.clone() + t = fluid.InferenceTranspiler() + t.transpile(inference_transpiler_program, place) + ''' + def transpile(self, program, place, scope=None): ''' - Transpile the program. Support only fuse batch normalization now. - - :param program: program to transpile - :type program: Program - :param place: inference place - :type place: Place - :param scope: inference scope - :type scope: Scope or None + Run the transpiler. + + Args: + program (Program): program to transpile + place (Place): inference place + scope (Scope|None): inference Scope ''' if not isinstance(program, Program): raise TypeError("program should be as Program type") @@ -49,36 +63,43 @@ class InferenceTranspiler: can be integrated with them. Doing so will give us a forward acceleration, especially in environments like mobile or embedded. - For input X: - - Conv process: X = input * W + bias - - Batch norm process: X' = (X - mean) / std - - Scale Process: Y = a * X' + b + For input :math:`X`: + + - Conv process: :math:`X = input * W + bias` + - Batch norm process: :math:`X' = (X - mean) / std` + - Scale Process: :math:`Y = a * X' + b` After fuse into one operation: - Y = (input * W + bias - mean) / std * a + b - = input * a * W / std + ((bias - mean) / std * a + b) + .. math:: + + Y &= (input * W + bias - mean) / std * a + b \\\\ + &= input * a * W / std + ((bias - mean) / std * a + b) The operator transformation is: + - before: + - conv->batch_norm->any_other_op (bias == 0) - conv->elementwise_add->batch_norm->any_other_op (bias != 0) + - after: + - conv->elementwise_add->any_other_op The transpile stages are: + 1. insert elementwise_add op when bias == 0. 2. fuse the batch_norm's parameters to conv and elementwise_add operators. 3. remove batch_norm ops which are not used in any other ops. 4. adjust the input of any_other_op to be the output of elementwise_add operator. 5. remove unused variables. - :param program: program to transpile - :type program: Program - :param place: inference place - :type place: Place - :param scope: inference scope - :type scope: Scope + Args: + program (Program): program to transpile + place (Place): inference place + scope (Scope): inference Scope + ''' self.scope = scope self.place = place diff --git a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py index 8bfb554845d9b128f000d6c90cf626416a198eef..999ef43ca0feacbddff5f9db59589ce7097fe77e 100644 --- a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py +++ b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py @@ -383,6 +383,16 @@ def memory_optimize(input_program, skip_opt_set=None, print_log=False, level=0): def release_memory(input_program, skip_opt_set=None): + """ + Modify the input program and insert :code:`delete_op` to early drop not used + variables. The modification will be performed inplace. + + Notes: This is an experimental API and could be removed in next few + releases. Users should not use this API. + + Args: + input_program(Program): The program will be inserted :code:`delete_op`. + """ cfgs = _get_cfgs(input_program) for cfg in cfgs: cfg.release_memory(skip_opt_set=skip_opt_set) diff --git a/python/paddle/fluid/transpiler/ps_dispatcher.py b/python/paddle/fluid/transpiler/ps_dispatcher.py index d6a68677527deb09ace0e3a23cbc093d6d7b4349..dcffadd531719431f27feb464ed58a65c04770ee 100644 --- a/python/paddle/fluid/transpiler/ps_dispatcher.py +++ b/python/paddle/fluid/transpiler/ps_dispatcher.py @@ -33,15 +33,21 @@ class PSDispatcher(object): def dispatch(self, varlist): """ - :param varlist: a list of Variables - :return: a map of pserver endpoint -> varname + Args: + varlist(list): a list of Variables + Returns: + a map of pserver endpoint -> varname """ AssertionError("Interface has not been implemented.") class HashName(PSDispatcher): """ - Hash variable names to several endpoints + Hash variable names to several endpoints using python + "hash()" function. + + Args: + pserver_endpoints (list): list of endpoint(ip:port). """ def __init__(self, pserver_endpoints): @@ -61,7 +67,11 @@ class HashName(PSDispatcher): class RoundRobin(PSDispatcher): """ - Distribute variables to serveral endpoints. + Distribute variables to serveral endpoints using + RondRobin method. + + Args: + pserver_endpoints (list): list of endpoint(ip:port). """ def __init__(self, pserver_endpoints): diff --git a/python/paddle/fluid/unique_name.py b/python/paddle/fluid/unique_name.py index 33c53113ae7e8ed9aeada31f2aed6990b6fea110..776619cd36722e338a9fdd5e13bceeaf3724de2c 100644 --- a/python/paddle/fluid/unique_name.py +++ b/python/paddle/fluid/unique_name.py @@ -16,7 +16,7 @@ import collections import contextlib import sys -__all__ = ['generate', 'switch', 'guard', 'UniqueNameGenerator'] +__all__ = ['generate', 'switch', 'guard'] class UniqueNameGenerator(object): diff --git a/python/paddle/reader/decorator.py b/python/paddle/reader/decorator.py index 44a6e344630bb35d28ee29078bf8727053a24bef..1f83cabb8481451736944823be45185deea4f43b 100644 --- a/python/paddle/reader/decorator.py +++ b/python/paddle/reader/decorator.py @@ -336,7 +336,7 @@ def _buf2lines(buf, line_break="\n"): class PipeReader: """ - PipeReader read data by stream from a command, take it's + PipeReader read data by stream from a command, take it's stdout into a pipe buffer and redirect it to the parser to parse, then yield data as your desired format. @@ -352,7 +352,7 @@ class PipeReader: An example: .. code-block:: python - + def example_reader(): for f in myfiles: pr = PipeReader("cat %s"%f) diff --git a/python/paddle/v2/dataset/cifar.py b/python/paddle/v2/dataset/cifar.py index 0a2a1ced11ee5cb2fb407b229ce810d553c2fa46..662655c836dbc54bd6187dcd3dac7354d6c8ecd1 100644 --- a/python/paddle/v2/dataset/cifar.py +++ b/python/paddle/v2/dataset/cifar.py @@ -43,7 +43,7 @@ CIFAR100_URL = URL_PREFIX + 'cifar-100-python.tar.gz' CIFAR100_MD5 = 'eb9058c3a382ffc7106e4002c42a8d85' -def reader_creator(filename, sub_name): +def reader_creator(filename, sub_name, cycle=False): def read_batch(batch): data = batch['data'] labels = batch.get('labels', batch.get('fine_labels', None)) @@ -56,10 +56,13 @@ def reader_creator(filename, sub_name): names = (each_item.name for each_item in f if sub_name in each_item.name) - for name in names: - batch = cPickle.load(f.extractfile(name)) - for item in read_batch(batch): - yield item + while True: + for name in names: + batch = cPickle.load(f.extractfile(name)) + for item in read_batch(batch): + yield item + if not cycle: + break return reader @@ -94,34 +97,40 @@ def test100(): 'test') -def train10(): +def train10(cycle=False): """ 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]. + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: Training reader creator :rtype: callable """ return reader_creator( paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5), - 'data_batch') + 'data_batch', + cycle=cycle) -def test10(): +def test10(cycle=False): """ CIFAR-10 test set creator. It returns a reader creator, each sample in the reader is image pixels in [0, 1] and label in [0, 9]. + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: Test reader creator. :rtype: callable """ return reader_creator( paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5), - 'test_batch') + 'test_batch', + cycle=cycle) def fetch(): diff --git a/python/paddle/v2/dataset/flowers.py b/python/paddle/v2/dataset/flowers.py index 357a4e9b000ea81afe291ff39dde2bed5c67e619..db12076d54064781bd1060947497622b14783768 100644 --- a/python/paddle/v2/dataset/flowers.py +++ b/python/paddle/v2/dataset/flowers.py @@ -76,7 +76,8 @@ def reader_creator(data_file, dataset_name, mapper, buffered_size=1024, - use_xmap=True): + use_xmap=True, + cycle=False): ''' 1. read images from tar file and merge images into batch files in 102flowers.tgz_batch/ @@ -96,6 +97,8 @@ def reader_creator(data_file, :type mapper: callable :param buffered_size: the size of buffer used to process images :type buffered_size: int + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: data reader :rtype: callable ''' @@ -108,15 +111,18 @@ def reader_creator(data_file, file_list = batch_images_from_tar(data_file, dataset_name, img2label) def reader(): - for file in open(file_list): - file = file.strip() - batch = None - with open(file, 'r') as f: - batch = cPickle.load(f) - data = batch['data'] - labels = batch['label'] - for sample, label in itertools.izip(data, batch['label']): - yield sample, int(label) - 1 + while True: + for file in open(file_list): + file = file.strip() + batch = None + with open(file, 'r') as f: + batch = cPickle.load(f) + data = batch['data'] + labels = batch['label'] + for sample, label in itertools.izip(data, batch['label']): + yield sample, int(label) - 1 + if not cycle: + break if use_xmap: cpu_num = int(os.environ.get('CPU_NUM', cpu_count())) @@ -125,7 +131,7 @@ def reader_creator(data_file, return map_readers(mapper, reader) -def train(mapper=train_mapper, buffered_size=1024, use_xmap=True): +def train(mapper=train_mapper, buffered_size=1024, use_xmap=True, cycle=False): ''' Create flowers training set reader. It returns a reader, each sample in the reader is @@ -138,17 +144,23 @@ def train(mapper=train_mapper, buffered_size=1024, use_xmap=True): :type mapper: callable :param buffered_size: the size of buffer used to process images :type buffered_size: int + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: train data reader :rtype: callable ''' return reader_creator( download(DATA_URL, 'flowers', DATA_MD5), download(LABEL_URL, 'flowers', LABEL_MD5), - download(SETID_URL, 'flowers', SETID_MD5), TRAIN_FLAG, mapper, - buffered_size, use_xmap) + download(SETID_URL, 'flowers', SETID_MD5), + TRAIN_FLAG, + mapper, + buffered_size, + use_xmap, + cycle=cycle) -def test(mapper=test_mapper, buffered_size=1024, use_xmap=True): +def test(mapper=test_mapper, buffered_size=1024, use_xmap=True, cycle=False): ''' Create flowers test set reader. It returns a reader, each sample in the reader is @@ -161,14 +173,20 @@ def test(mapper=test_mapper, buffered_size=1024, use_xmap=True): :type mapper: callable :param buffered_size: the size of buffer used to process images :type buffered_size: int + :param cycle: whether to cycle through the dataset + :type cycle: bool :return: test data reader :rtype: callable ''' return reader_creator( download(DATA_URL, 'flowers', DATA_MD5), download(LABEL_URL, 'flowers', LABEL_MD5), - download(SETID_URL, 'flowers', SETID_MD5), TEST_FLAG, mapper, - buffered_size, use_xmap) + download(SETID_URL, 'flowers', SETID_MD5), + TEST_FLAG, + mapper, + buffered_size, + use_xmap, + cycle=cycle) def valid(mapper=test_mapper, buffered_size=1024, use_xmap=True): diff --git a/tools/check_ctest_hung.py b/tools/check_ctest_hung.py new file mode 100644 index 0000000000000000000000000000000000000000..7de76c381b29a1ff8dcf2167f0e861dc261aa47b --- /dev/null +++ b/tools/check_ctest_hung.py @@ -0,0 +1,53 @@ +# 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 sys +import re + + +def escape(input): + o = input.replace("\n", "") + o = o.replace("\r", "") + return o + + +def main(): + usage = """Usage: +1. Download the Paddle_PR_CI_*.log from TeamCity +2. run: python check_ctest_hung.py Paddle_PR_CI_*.log +3. If there is hung ctest, the result likes: +Diff: set(['test_parallel_executor_crf']) + """ + if len(sys.argv) < 2: + print(usage) + exit(0) + + logfile = sys.argv[1] + started = set() + passed = set() + with open(logfile, "r") as fn: + for l in fn.readlines(): + if l.find("Test ") != -1 and \ + l.find("Passed") != -1: + m = re.search("Test\s+#[0-9]*\:\s([a-z0-9_]+)", escape(l)) + passed.add(m.group(1)) + if l.find("Start ") != -1: + start_parts = escape(l).split(" ") + m = re.search("Start\s+[0-9]+\:\s([a-z0-9_]+)", escape(l)) + started.add(m.group(1)) + print "Diff: ", started - passed + + +if __name__ == "__main__": + main() diff --git a/.clang_format.hook b/tools/codestyle/clang_format.hook similarity index 100% rename from .clang_format.hook rename to tools/codestyle/clang_format.hook diff --git a/.copyright.hook b/tools/codestyle/copyright.hook similarity index 100% rename from .copyright.hook rename to tools/codestyle/copyright.hook diff --git a/tools/codestyle/docstring_checker.py b/tools/codestyle/docstring_checker.py index 54a690462699651d3e14f9b24383df01a9740336..8d4b24a0cf6b743b72dca58fd885f927560964bf 100644 --- a/tools/codestyle/docstring_checker.py +++ b/tools/codestyle/docstring_checker.py @@ -291,6 +291,8 @@ class DocstringChecker(BaseChecker): True if successful otherwise False. """ + if node.name.startswith("__") or node.name.startswith("_"): + return True find = False for t in node.body: if not isinstance(t, astroid.Return): @@ -316,6 +318,8 @@ class DocstringChecker(BaseChecker): Returns: True if successful otherwise False. """ + if node.name.startswith("__") or node.name.startswith("_"): + return True args = [] for arg in node.args.get_children(): if (not isinstance(arg, astroid.AssignName)) \ diff --git a/tools/print_signatures.py b/tools/print_signatures.py new file mode 100644 index 0000000000000000000000000000000000000000..5e7ffd44c7b0ba2270069bc4467dc377a58b2417 --- /dev/null +++ b/tools/print_signatures.py @@ -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. +""" +Print all signature of a python module in alphabet order. + +Usage: + ./print_signature "paddle.fluid" > signature.txt +""" +import importlib +import inspect +import collections +import sys +import pydoc + +member_dict = collections.OrderedDict() + + +def visit_member(parent_name, member): + cur_name = ".".join([parent_name, member.__name__]) + if inspect.isclass(member): + for name, value in inspect.getmembers(member): + if hasattr(value, '__name__') and (not name.startswith("_") or + name == "__init__"): + visit_member(cur_name, value) + elif callable(member): + try: + member_dict[cur_name] = inspect.getargspec(member) + except TypeError: # special for PyBind method + member_dict[cur_name] = " ".join([ + line.strip() for line in pydoc.render_doc(member).split('\n') + if "->" in line + ]) + + else: + raise RuntimeError("Unsupported generate signature of member, type {0}". + format(str(type(member)))) + + +def visit_all_module(mod): + for member_name in ( + name + for name in (mod.__all__ if hasattr(mod, "__all__") else dir(mod)) + if not name.startswith("_")): + instance = getattr(mod, member_name, None) + if instance is None: + continue + if inspect.ismodule(instance): + visit_all_module(instance) + else: + visit_member(mod.__name__, instance) + + +visit_all_module(importlib.import_module(sys.argv[1])) + +for name in member_dict: + print name, member_dict[name]