diff --git a/.clang-format b/.clang-format index 9ba433b17362424973626470d930356c2173dd84..aff93435f58c522f5ed1090aef2005f76e91cf31 100644 --- a/.clang-format +++ b/.clang-format @@ -25,4 +25,3 @@ AllowAllParametersOfDeclarationOnNextLine: true BinPackParameters: false BinPackArguments: false ... - diff --git a/.travis.yml b/.travis.yml index c51e02eb79a9e53a2b8d1d663e8f0c3e0d8c3a61..e2d49daa1981396628efa5d16459eb70e9e76884 100644 --- a/.travis.yml +++ b/.travis.yml @@ -42,7 +42,7 @@ before_install: script: - | timeout 2580 paddle/scripts/travis/${JOB}.sh # 43min timeout - RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true; else false; fi; + RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true ;else exit 1; fi; - | if [[ "$JOB" != "build_doc" ]]; then exit 0; fi; if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi; diff --git a/CMakeLists.txt b/CMakeLists.txt index 65164b8472b902be8b0b9d5fb99807d012b8a666..e76512166fcaea5daf2a67d1259331b680f15b7c 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -133,6 +133,8 @@ include(external/any) # download libn::any include(external/eigen) # download eigen3 include(external/pybind11) # download pybind11 include(external/nccl) +include(external/cares) +include(external/grpc) include(cudnn) # set cudnn libraries, must before configure include(configure) # add paddle env configuration diff --git a/Dockerfile b/Dockerfile index 150344a8116e2be9b5bab8e5fdcc9c37f4025020..857d3f3e5f64791146741ffb29feabfcb2ecbb84 100644 --- a/Dockerfile +++ b/Dockerfile @@ -29,7 +29,7 @@ RUN apt-get update && \ automake locales clang-format swig doxygen cmake \ liblapack-dev liblapacke-dev libboost-dev \ clang-3.8 llvm-3.8 libclang-3.8-dev \ - net-tools && \ + net-tools libtool && \ apt-get clean -y # Install Go and glide diff --git a/cmake/external/cares.cmake b/cmake/external/cares.cmake new file mode 100644 index 0000000000000000000000000000000000000000..e05111ee18efc906e39bcb56fb1be3b3c3dff5d6 --- /dev/null +++ b/cmake/external/cares.cmake @@ -0,0 +1,45 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# + +IF(MOBILE_INFERENCE) + return() +ENDIF() + +include (ExternalProject) + +# NOTE: c-ares is needed when linking with grpc. + +SET(CARES_SOURCES_DIR ${THIRD_PARTY_PATH}/cares) +SET(CARES_INSTALL_DIR ${THIRD_PARTY_PATH}/install/cares) +SET(CARES_INCLUDE_DIR "${CARES_INSTALL_DIR}/include/" CACHE PATH "cares include directory." FORCE) + +ExternalProject_Add( + extern_cares + GIT_REPOSITORY "https://github.com/c-ares/c-ares.git" + GIT_TAG "cares-1_13_0" + PREFIX ${CARES_SOURCES_DIR} + UPDATE_COMMAND "" + CONFIGURE_COMMAND ./buildconf && ./configure --disable-shared --prefix=${CARES_INSTALL_DIR} + BUILD_IN_SOURCE 1 + BUILD_COMMAND make + INSTALL_COMMAND make install +) + +ADD_LIBRARY(cares STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET cares PROPERTY IMPORTED_LOCATION + "${CARES_INSTALL_DIR}/lib/libcares.a") + +include_directories(${CARES_INCLUDE_DIR}) +ADD_DEPENDENCIES(cares extern_cares) diff --git a/cmake/external/grpc.cmake b/cmake/external/grpc.cmake new file mode 100644 index 0000000000000000000000000000000000000000..219ea1b90881ccdbaf3fd41510fb4f2a8b6ec0f4 --- /dev/null +++ b/cmake/external/grpc.cmake @@ -0,0 +1,66 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# + +IF(MOBILE_INFERENCE) + return() +ENDIF() + +include (ExternalProject) + +SET(GRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/grpc) +SET(GRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/grpc) +SET(GRPC_INCLUDE_DIR "${GRPC_INSTALL_DIR}/include/" CACHE PATH "grpc include directory." FORCE) +SET(GRPC_CPP_PLUGIN "${GRPC_INSTALL_DIR}/bin/grpc_cpp_plugin" CACHE FILEPATH "GRPC_CPP_PLUGIN" FORCE) +IF(APPLE) + SET(BUILD_CMD make -n | sed "s/-Werror//g" | sh) +ELSE() + SET(BUILD_CMD make) +ENDIF() + +ExternalProject_Add( + extern_grpc + DEPENDS protobuf zlib + GIT_REPOSITORY "https://github.com/grpc/grpc.git" + GIT_TAG "v1.7.x" + PREFIX ${GRPC_SOURCES_DIR} + UPDATE_COMMAND "" + CONFIGURE_COMMAND "" + BUILD_IN_SOURCE 1 + # NOTE(yuyang18): + # Disable -Werror, otherwise the compile will fail in MacOS. + # It seems that we cannot configure that by make command. + # Just dry run make command and remove `-Werror`, then use a shell to run make commands + BUILD_COMMAND ${BUILD_CMD} HAS_SYSTEM_PROTOBUF=false -s -j8 static grpc_cpp_plugin + INSTALL_COMMAND make prefix=${GRPC_INSTALL_DIR} install +) + +# FIXME(typhoonzero): hack to get static lib path, try a better way like merge them. +ADD_LIBRARY(grpc++_unsecure STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET grpc++_unsecure PROPERTY IMPORTED_LOCATION + "${GRPC_INSTALL_DIR}/lib/libgrpc++_unsecure.a") + +ADD_LIBRARY(grpc++ STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET grpc++ PROPERTY IMPORTED_LOCATION + "${GRPC_INSTALL_DIR}/lib/libgrpc++.a") +ADD_LIBRARY(gpr STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET gpr PROPERTY IMPORTED_LOCATION + "${GRPC_INSTALL_DIR}/lib/libgpr.a") + +ADD_LIBRARY(grpc_unsecure STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET grpc_unsecure PROPERTY IMPORTED_LOCATION + "${GRPC_INSTALL_DIR}/lib/libgrpc_unsecure.a") + +include_directories(${GRPC_INCLUDE_DIR}) +ADD_DEPENDENCIES(grpc++_unsecure extern_grpc) diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake index be7f6a9465970711170bd15dcecaadeaa8a55f86..7cfe1e68078eed023fd0cc6971c573bb0108b4cc 100644 --- a/cmake/external/protobuf.cmake +++ b/cmake/external/protobuf.cmake @@ -15,7 +15,18 @@ INCLUDE(ExternalProject) # Always invoke `FIND_PACKAGE(Protobuf)` for importing function protobuf_generate_cpp FIND_PACKAGE(Protobuf QUIET) -SET(PROTOBUF_FOUND "OFF") +macro(UNSET_VAR VAR_NAME) + UNSET(${VAR_NAME} CACHE) + UNSET(${VAR_NAME}) +endmacro() +UNSET_VAR(PROTOBUF_INCLUDE_DIR) +UNSET_VAR(PROTOBUF_FOUND) +UNSET_VAR(PROTOBUF_PROTOC_EXECUTABLE) +UNSET_VAR(PROTOBUF_PROTOC_LIBRARY) +UNSET_VAR(PROTOBUF_LITE_LIBRARY) +UNSET_VAR(PROTOBUF_LIBRARY) +UNSET_VAR(PROTOBUF_INCLUDE_DIR) +UNSET_VAR(Protobuf_PROTOC_EXECUTABLE) if(NOT COMMAND protobuf_generate_python) # before cmake 3.4, protobuf_genrerate_python is not defined. function(protobuf_generate_python SRCS) @@ -110,7 +121,6 @@ macro(PROMPT_PROTOBUF_LIB) # FIND_Protobuf.cmake uses `Protobuf_PROTOC_EXECUTABLE`. # make `protobuf_generate_cpp` happy. SET(Protobuf_PROTOC_EXECUTABLE ${PROTOBUF_PROTOC_EXECUTABLE}) - FOREACH(dep ${protobuf_DEPS}) ADD_DEPENDENCIES(protobuf ${dep}) ADD_DEPENDENCIES(protobuf_lite ${dep}) @@ -128,11 +138,11 @@ endmacro() set(PROTOBUF_ROOT "" CACHE PATH "Folder contains protobuf") if (NOT "${PROTOBUF_ROOT}" STREQUAL "") - find_path(PROTOBUF_INCLUDE_DIR google/protobuf/message.h PATHS ${PROTOBUF_ROOT}/include) - find_library(PROTOBUF_LIBRARY protobuf PATHS ${PROTOBUF_ROOT}/lib) - find_library(PROTOBUF_LITE_LIBRARY protobuf-lite PATHS ${PROTOBUF_ROOT}/lib) - find_library(PROTOBUF_PROTOC_LIBRARY protoc PATHS ${PROTOBUF_ROOT}/lib) - find_program(PROTOBUF_PROTOC_EXECUTABLE protoc PATHS ${PROTOBUF_ROOT}/bin) + find_path(PROTOBUF_INCLUDE_DIR google/protobuf/message.h PATHS ${PROTOBUF_ROOT}/include NO_DEFAULT_PATH) + find_library(PROTOBUF_LIBRARY protobuf PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH) + find_library(PROTOBUF_LITE_LIBRARY protobuf-lite PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH) + find_library(PROTOBUF_PROTOC_LIBRARY protoc PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH) + find_program(PROTOBUF_PROTOC_EXECUTABLE protoc PATHS ${PROTOBUF_ROOT}/bin NO_DEFAULT_PATH) if (PROTOBUF_INCLUDE_DIR AND PROTOBUF_LIBRARY AND PROTOBUF_LITE_LIBRARY AND PROTOBUF_PROTOC_LIBRARY AND PROTOBUF_PROTOC_EXECUTABLE) message(STATUS "Using custom protobuf library in ${PROTOBUF_ROOT}.") SET_PROTOBUF_VERSION() diff --git a/cmake/external/zlib.cmake b/cmake/external/zlib.cmake index a98e069b7cd1654ddd5868560d0905eab6d9c692..1638cd8fdfc34575132462859e056a1907f0b2f1 100644 --- a/cmake/external/zlib.cmake +++ b/cmake/external/zlib.cmake @@ -50,6 +50,8 @@ ExternalProject_Add( ) LIST(APPEND external_project_dependencies zlib) +ADD_LIBRARY(zlib_target STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET zlib_target PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES}) IF(WITH_C_API) INSTALL(DIRECTORY ${ZLIB_INCLUDE_DIR} DESTINATION third_party/zlib) diff --git a/cmake/generic.cmake b/cmake/generic.cmake index 7b82d409a3b64a5fc8fdfe526a2e82a4e1c9fa8e..c917ca0ff4e087b7caae8876da127bec6b39b798 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -467,3 +467,50 @@ function(py_test TARGET_NAME) WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) endif() endfunction() + +# grpc_library generate grpc code using grpc_cpp_plugin and protoc +# then build the generated protobuf code and grpc code with your +# implementation source codes together. Use SRCS argument for your +# implementation source files and PROTO argument for your .proto +# files. +# +# Usage: grpc_library(my_target SRCS my_client.cc PROTO my_target.proto DEPS my_dep) + +function(grpc_library TARGET_NAME) + set(oneValueArgs PROTO) + set(multiValueArgs SRCS DEPS) + set(options "") + cmake_parse_arguments(grpc_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + + message(STATUS "generating grpc ${grpc_library_PROTO}") + + get_filename_component(ABS_PROTO ${grpc_library_PROTO} ABSOLUTE) + get_filename_component(PROTO_WE ${grpc_library_PROTO} NAME_WE) + get_filename_component(PROTO_PATH ${ABS_PROTO} PATH) + + protobuf_generate_cpp(grpc_proto_srcs grpc_proto_hdrs "${ABS_PROTO}") + set(grpc_grpc_srcs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.cc") + set(grpc_grpc_hdrs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.h") + cc_library("${TARGET_NAME}_proto" SRCS "${grpc_proto_srcs}") + + add_custom_command( + OUTPUT "${grpc_grpc_srcs}" "${grpc_grpc_hdrs}" + COMMAND ${PROTOBUF_PROTOC_EXECUTABLE} + ARGS --grpc_out "${CMAKE_CURRENT_BINARY_DIR}" -I "${PROTO_PATH}" + --plugin=protoc-gen-grpc="${GRPC_CPP_PLUGIN}" "${ABS_PROTO}" + DEPENDS "${ABS_PROTO}" ${PROTOBUF_PROTOC_EXECUTABLE} extern_grpc) + + # FIXME(typhoonzero): grpc generated code do not generate virtual-dtor, mark it + # as compiler warnings instead of error. Should try remove the warnings also. + set_source_files_properties( + ${grpc_grpc_srcs} + PROPERTIES + COMPILE_FLAGS "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + cc_library("${TARGET_NAME}_grpc" SRCS "${grpc_grpc_srcs}") + + set_source_files_properties( + ${grpc_library_SRCS} + PROPERTIES + COMPILE_FLAGS "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + cc_library("${TARGET_NAME}" SRCS "${grpc_library_SRCS}" DEPS "${TARGET_NAME}_grpc" "${TARGET_NAME}_proto" "${grpc_library_DEPS}") +endfunction() diff --git a/doc/api/v2/config/layer.rst b/doc/api/v2/config/layer.rst index d4d182f6692e09b3e40f3620b77d9a0f20ec5af3..c3f9c18d0663a7a24880b441981875c1e4f015aa 100644 --- a/doc/api/v2/config/layer.rst +++ b/doc/api/v2/config/layer.rst @@ -54,7 +54,7 @@ img_conv .. _api_v2.layer_context_projection: -context_projection +context_projection ------------------ .. autoclass:: paddle.v2.layer.context_projection :noindex: @@ -70,7 +70,7 @@ Image Pooling Layer img_pool -------- .. autoclass:: paddle.v2.layer.img_pool - :noindex: + :noindex: spp --- @@ -104,7 +104,7 @@ sum_to_one_norm --------------- .. autoclass:: paddle.v2.layer.sum_to_one_norm :noindex: - + cross_channel_norm ------------------ .. autoclass:: paddle.v2.layer.cross_channel_norm @@ -114,7 +114,7 @@ row_l2_norm ----------- .. autoclass:: paddle.v2.layer.row_l2_norm :noindex: - + Recurrent Layers ================ @@ -415,6 +415,13 @@ multiplex .. autoclass:: paddle.v2.layer.multiplex :noindex: +Factorization Machine Layer +============================ + +factorization_machine +--------------------- +.. autoclass:: paddle.v2.layer.factorization_machine + :noindex: Slicing and Joining Layers ========================== diff --git a/doc/design/refactor/distributed_architecture.md b/doc/design/refactor/distributed_architecture.md index ac7e98ccf1aadbb973a4801fde842375cf63448c..2b4f921ae93c3b443ed62a28b1fa9fbda14f73ab 100644 --- a/doc/design/refactor/distributed_architecture.md +++ b/doc/design/refactor/distributed_architecture.md @@ -2,106 +2,70 @@ ## Abstract -PaddlePaddle v0.10.0 uses the "trainer-parameter server" -architecture. We run multiple replicated instances of trainers (runs -the same code written by the user) and parameter servers for -distributed training. This architecture served us well, but has some -limitations: +PaddlePaddle version 0.10.0 uses the "trainer-parameter server" architecture. We run multiple instances of trainers (where each trainer runs the same model) and parameter servers for distributed training. This architecture serves well, but has few limitations: -1. Need to write special code to handle tasks which should only be run - by a single trainer. E.g., initializing model and saving model. +1. There is a need to write special code that handles tasks which should only be run on a single trainer. E.g., initializing the model, saving the model etc. -2. Model parallelism is hard: need to write if-else branches conditioned - on the trainer ID to partition model onto each trainer, and manually - write the inter-model-shard communication code. +2. Model parallelism is hard: It would need all the if-else branches conditioned on the trainer ID to partition the model onto the trainers, and eventually manually writing out the inter-model-shard communication code to communicate between different trainers. -3. The user can not directly specify the parameter update rule: need - to modify the parameter server C++ code and compile a new - binary. This adds complication for researchers: A lot of extra - effort is required. Besides, the training job submission program - may not allow running arbitrary binaries. +3. The user can not directly specify the parameter update rule: This would need to modify the parameter server code and compile a new binary. This makes things more complicated for researchers: A lot of extra effort is required to make this work. Besides, the training job submission program may not allow running arbitrary binaries. -This design doc discusses PaddlePaddle's new distributed training -architecture that addresses the above limitations. +This design doc discusses PaddlePaddle's new distributed training architecture that addresses the above mentioned limitations. ## Analysis -We will assume the user writes the trainer program by Python, the same -analysis holds if the trainer program is written in C++. +The assumption is that the user writes the trainer program in either Python or C++. ### Limitation 1 -If we look at the Python code that the user writes, there are two -kinds of functionalities: +There are two basic functionalities in the trainer program: -- The training logic such as load / save model and print log. -- The neural network definition such as the definition of the data - layer, the fully connected layer, the cost function and the +1. The training logic such as loading / saving the model and printing out the logs. +2. The neural network definition such as the definition of the data layer, the fully connected layer, the cost function and the optimizer. -When we training with PaddlePaddle v0.10.0 distributedly, multiple -replicated Python instances are running on different nodes: both the -training logic and the neural network computation is replicated. +When we train using PaddlePaddle v0.10.0 in a distributed fashion, multiple instances of the same Python code are run on different nodes, hence both: the +training logic as well as the neural network computation logic, is replicated. -The tasks that should only run once all belong to the training logic, -if we only replicate the neural network computation, but do **not** -replicate the training logic, the limitation could be solved. +The tasks that only need to be run once belong to the training logic. Hence if we only replicate the neural network computation part, and do **not** +replicate the training logic, the limitation mentioned above can be avoided. ### Limitation 2 -Model parallelism means running a single model on multiple nodes by -partitioning the model onto different nodes and managing the -inter-model-shard communications. +Model parallelism means that a single model is partitioned into different components and each node runs one of the component separately. This comes at the extra cost of managing the +inter-model-shard communication between nodes. -PaddlePaddle should be able to modify the nerual network computation -definition to support model parallelism automatically. However, the -computation is only specified in Python code, and PaddlePaddle can not -modify Python code. +PaddlePaddle should ideally be able to modify the neural network computation and figure out the support for model parallelism automatically. However, the +computation is only specified in Python code which sits outside of PaddlePaddle, hence PaddlePaddle can not support the feature in this setup. -Just like compiler uses a intermediate representation (IR) so that -programmer does not need to manually optimize their code in most of -the cases - the compiler will optimize the IR: +Similar to how a compiler uses an intermediate representation (IR) so that the programmer does not need to manually optimize their code for most of the cases, we can have an intermediate representation in PaddlePaddle as well. The compiler optimizes the IR as follows: -We can have our own IR too: PaddlePaddle can support model parallel by -converting the IR so the user no longer need to manually do it in -Python: +PaddlePaddle can support model parallelism by converting the IR so that the user no longer needs to manually perform the computation and operations in the Python component: -The IR for PaddlePaddle after refactor is called `Block`, it specifies -the computation dependency graph and the variables used in the -computation. +The IR for PaddlePaddle after refactoring is called a `Block`, it specifies the computation dependency graph and the variables used in the computation. ### Limitation 3 -The user can not directly specify the parameter update rule for the -parameter server because the parameter server does not use the same -computation definition as the trainer. Instead, the update rule is -baked in the parameter server. The user can not specify the update -rule in the same way of specifying the trainer computation. +The user can not directly specify the parameter update rule for the parameter server in the Python module, since the parameter server does not use the same computation definition as the trainer. Instead, the update rule is baked inside the parameter server. The user can not specify the update rule explicitly. -This could be fixed by making the parameter server run the same -computation definition as the trainer. For a detailed explanation, -please -see +This could be fixed by making the parameter server run the same computation definition as the trainer (the user's Python module). For a detailed explanation, refer to this document - [Design Doc: Operation Graph Based Parameter Server](./dist_train.md) ## Distributed Training Architecture -The new distributed training architecture can address the above -limitations. Below is the illustration: +The revamped distributed training architecture can address the above discussed limitations. Below is the illustration of how it does so: -The architecture includes major components: *PaddlePaddle Python*, -*PaddlePaddle converter* and *PaddlePaddle runtime*: +The major components in the architecture are: *PaddlePaddle Python*, *PaddlePaddle converter* and *PaddlePaddle runtime*. ### PaddlePaddle Python -PaddlePaddle Python is the Python library that user's Python trainer -invoke to build the neural network topology, start training, etc. +PaddlePaddle Python is the Python library that user's Python code invokes, to read the data. build the neural network topology, start training, etc. ```Python paddle.init() @@ -117,102 +81,60 @@ for i in range(1000): print cost_val ``` -The code above is a typical Python trainer code, the neural network -topology is built using helper functions such as -`paddle.layer.fc`. The training is done by calling `session.eval` -iteratively. +The above code is what a typical Python trainer code is, the neural network topology is built using the helper functions such as `paddle.layer.fc`. Training is done by calling `session.eval` iteratively. #### session.eval -As shown in the graph, `session.eval` sends the IR and the evaluation -inputs/targets to the PaddlePaddle cluster for evaluation. The -targets can be any variable in the computation graph. When the target -is the `optimizer` variable, the neural network will be optimized -once. When the target is the `cost` variable, `session.eval` returns -the cost value. +As shown in the graph, `session.eval` sends the IR and the evaluation inputs or targets to the PaddlePaddle cluster for evaluation. +The targets can be any variable in the computation graph. When the target is say, the `optimizer` variable, the neural network will be optimized once. When the target is the `cost` variable, `session.eval` returns the cost value. Based on what the target is, an appropriate action is taken. -The Python `session` is a wrapper of the C++ `Session` class. For more -information about `Session`, please -see [Design Doc: Session](./session.md). +The Python `session` is a wrapper of the C++ `Session` class. For more information about `Session`, refer to this document - [Design Doc: Session](./session.md). ### PaddlePaddle Converter -PaddlePaddle converter automatically converts the IR in the request -(IR and evaluation inputs/targets) from PaddlePaddle Python to new -partitioned IRs and dispatch the new IRs and evaluation inputs/targets -to different PaddlePaddle runtimes. Below are the steps: +The PaddlePaddle converter automatically converts the IR in the request (IR and evaluation inputs/targets) from PaddlePaddle Python to partitioned IRs and dispatches the new IRs and evaluation inputs/targets to different PaddlePaddle runtimes. Below are the steps that are followed : -1. Add `feed` OP that feeds the eval inputs, and `fetch` OP that - fetches the eval targets to the IR. +1. Add a `feed` OP that feeds the eval inputs, and a `fetch` OP that fetches the eval targets to the IR. -1. Extract a new computation (sub)graph with `feed` and `fetch` OP as - the boundary. The runtime does not need to run the OP that is not - dependent by the `fetch` OP. +2. Extract a new computation (sub)graph with the `feed` and `fetch` OPs as the boundary. The runtime does not need to run the OP that is not dependent on the `fetch` OP. -1. Optimizes the computation graph. +3. Optimize the computation graph. -1. Place the OPs in the graph onto different devices on different - PaddlePaddle runtime according to a placement algorithm and device - constraint specified by the user. +4. Place the OPs in the graph onto different devices on different PaddlePaddle runtime according to a placement algorithm and the device constraints specified by the user. -1. Partition the graph according to runtime boundaries and add `send` / - `recv` OP pair on the runtime boundaries. +5. Partition the graph according to runtime boundaries and add `send` / `recv` OP pair on the runtime boundaries. -1. Dispatch the partitioned graph to different PaddlePaddle runtimes. +6. Dispatch the partitioned graph to different PaddlePaddle runtimes. + +7. PaddlePaddle runtimes with the `fetch` OP reports evaluation results back to the converter, the converter reports the evaluation results back to the PaddlePaddle Python. -1. PaddlePaddle runtimes with the `fetch` OP reports evaluation - results back to the converter, the convert reports the evaluation - results back to the PaddlePaddle Python. - The output IRs will be cached to optimize the conversion latency. #### Placement Algorithm -Our first implementation will only support "trainer-parameter server" -placement: the parameters, initializers, and optimizers are placed on -the PaddlePaddle runtimes with the parameter server role. And -everything else will be placed on the PaddlePaddle runtimes with the -trainer role. This has the same functionality of our -"trainer-parameter server" architecture of PaddlePaddle v0.10.0, but -is more general and flexible. +Our first implementation will only support "trainer-parameter server" placement: the parameters, initializers, and optimizers are all placed on the PaddlePaddle runtimes with the parameter server role. Everything else will be placed on the PaddlePaddle runtimes with the trainer role. This has the same functionality as the "trainer-parameter server" architecture of PaddlePaddle v0.10.0, but is more generic and flexible. -In the future, we will implement the general placement algorithm, -which makes placements according to the input IR, and a model of -device computation time and device communication time. Model -parallelism requires the general placement algorithm. +In the future, a more general placement algorithm should be implemented, which makes placements according to the input IR, and a model of device computation time and device communication time. Model parallelism requires the generic placement algorithm. ### PaddlePaddle Runtime -The PaddlePaddle runtime owns multiple devices (e.g., CPUs, GPUs) and -runs the IR. The runtime does not need to do OP placement since it's -already done by the converter. +The PaddlePaddle runtime owns multiple devices (e.g., CPUs, GPUs) and runs the IR. The runtime does not need to do OP placement since it is already done by the converter. ### Local Training Architecture -The local training architecture will be the same as the distributed -training architecture, the differences are everything runs locally, -and there is just one PaddlePaddle runtime: +The local training architecture will be the same as the distributed training architecture, the difference is that everything runs locally, and there is just one PaddlePaddle runtime: ### Training Data -In PaddlePaddle v0.10.0, training data is typically read -with [data reader](../reader/README.md) from Python. This approach is -no longer efficient when training distributedly since the Python -process no longer runs on the same node with the trainer processes, -the Python reader will need to read from the distributed filesystem -(assuming it has the access) and send to the trainers, doubling the -network traffic. - -When doing distributed training, the user can still use Python data -reader: the training data are sent with `session.eval`. However should -be used for debugging purpose only. The users are encouraged to use -the read data OPs. +In PaddlePaddle v0.10.0, training data is typically read with a [data reader](../reader/README.md) from Python. This approach is no longer efficient when training in a distributed fashion since the Python process no longer runs on the same node with the trainer processes. The Python reader will need to read from the distributed filesystem (assuming it has the required access) and send to the trainers, doubling the network traffic. + +When doing distributed training, the user can still use Python data reader: the training data are sent with `session.eval`. However this should be used for debugging purpose only. The users are encouraged to use the read data OPs. ## References: diff --git a/doc/getstarted/build_and_install/build_from_source_cn.rst b/doc/getstarted/build_and_install/build_from_source_cn.rst index 55665ac8edfcf20290936fba4c3e410b33e1f3d4..3c525bdad6f6118dcd560e2cb7bfaf89737c1362 100644 --- a/doc/getstarted/build_and_install/build_from_source_cn.rst +++ b/doc/getstarted/build_and_install/build_from_source_cn.rst @@ -1,4 +1,4 @@ -从源码编译PaddlePaddle +从源码编译 ====================== .. _build_step: @@ -7,8 +7,11 @@ ---------------- PaddlePaddle主要使用 `CMake `_ 以及GCC, G++作为编译工具。 -我们推荐您使用PaddlePaddle编译环境镜像完成编译，这样可以免去单独安装编译依赖的步骤，可选的不同编译环境 +我们推荐您使用PaddlePaddle Docker编译环境镜像完成编译，这样可以免去单独安装编译依赖的步骤，可选的不同编译环境Docker镜像 可以在 `这里 `_ 找到。 + +如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 `编译依赖`_ 之后才能开始编译的步骤。 + 编译PaddlePaddle，需要执行： .. code-block:: bash @@ -22,7 +25,6 @@ PaddlePaddle主要使用 `CMake `_ 以及GCC, G++作为编译 cd build cmake -DWITH_GPU=OFF -DWITH_TESTING=OFF .. make - 编译完成后会在build/python/dist目录下生成输出的whl包，可以选在在当前机器安装也可以拷贝到目标机器安装： @@ -31,7 +33,33 @@ PaddlePaddle主要使用 `CMake `_ 以及GCC, G++作为编译 pip install python/dist/*.whl -.. _build_step: +.. _run_test: + +执行单元测试 +---------------- + +如果您期望在编译完成后立即执行所有的单元测试，可以按照下面的方法： + +使用Docker的情况下，设置 :code:`RUN_TEST=ON` 和 :code:`WITH_TESTING=ON` 就会在完成编译之后，立即执行单元测试。 +开启 :code:`WITH_GPU=ON` 可以指定同时执行GPU上的单元测试。 + +.. code-block:: bash + + docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x paddle/scripts/docker/build.sh + +如果不使用Docker，可以执行ctest命令即可： + +.. code-block:: bash + + mkdir build + cd build + cmake -DWITH_GPU=OFF -DWITH_TESTING=OFF .. + make + ctest + # 指定执行其中一个单元测试 test_mul_op + ctest -R test_mul_op + +.. _compile_deps: 编译依赖 ---------------- diff --git a/doc/getstarted/build_and_install/build_from_source_en.rst b/doc/getstarted/build_and_install/build_from_source_en.rst index 9a3ed7dd57137ddf3d6213222c17433822b01dbb..76fbc43de2e83580dd79b874507c103533022436 100644 --- a/doc/getstarted/build_and_install/build_from_source_en.rst +++ b/doc/getstarted/build_and_install/build_from_source_en.rst @@ -1,4 +1,4 @@ -Build PaddlePaddle from Sources +Build from Sources ========================== .. _build_step: @@ -9,14 +9,18 @@ How To Build PaddlePaddle mainly uses `CMake `_ and GCC, G++ as compile tools. We recommend you to use our pre-built Docker image to run the build to avoid installing dependencies by yourself. We have several build environment -Docker images `here `_. +Docker images `here `_ . + +If you choose not to use Docker image for your build, you need to install the +below `Compile Dependencies`_ before run the build. + Then run: .. code-block:: bash git clone https://github.com/PaddlePaddle/Paddle.git cd Paddle - # run the following command to build CPU-Only binaries if you are using docker + # run the following command to build a CPU-Only binaries if you are using docker docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x paddle/scripts/docker/build.sh # else run these commands mkdir build @@ -32,7 +36,35 @@ machine or copy it to the target machine. pip install python/dist/*.whl -.. _build_step: + +.. _run_test: + +Run Tests +---------------- + +If you wish to run the tests, you may follow the below steps: + +When using Docker, set :code:`RUN_TEST=ON` and :code:`WITH_TESTING=ON` will run test immediately after the build. +Set :code:`WITH_GPU=ON` Can also run tests on GPU. + +.. code-block:: bash + + docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x paddle/scripts/docker/build.sh + +If you don't use Docker, just run ctest will start the tests: + +.. code-block:: bash + + mkdir build + cd build + cmake -DWITH_GPU=OFF -DWITH_TESTING=ON .. + make + ctest + # run a single test like test_mul_op + ctest -R test_mul_op + + +.. _compile_deps: Compile Dependencies ---------------- diff --git a/doc/getstarted/build_and_install/docker_install_cn.rst b/doc/getstarted/build_and_install/docker_install_cn.rst index 07933b2e0bbca809f6c4e90e7ff8f71d1b3304b2..f78b1fb0e11aa028a4b7abb5270740b97f8039e9 100644 --- a/doc/getstarted/build_and_install/docker_install_cn.rst +++ b/doc/getstarted/build_and_install/docker_install_cn.rst @@ -1,4 +1,4 @@ -使用Docker安装运行PaddlePaddle +使用Docker安装运行 ================================ 使用Docker安装和运行PaddlePaddle可以无需考虑依赖环境即可运行。并且也可以在Windows的docker中运行。 diff --git a/doc/getstarted/build_and_install/docker_install_en.rst b/doc/getstarted/build_and_install/docker_install_en.rst index 9b977c9c72e36b4b47cbf56ae848ab83d9895783..d7acc7aeb744b19d83acb520d07c8551168dd096 100644 --- a/doc/getstarted/build_and_install/docker_install_en.rst +++ b/doc/getstarted/build_and_install/docker_install_en.rst @@ -1,4 +1,4 @@ -PaddlePaddle in Docker Containers +Run in Docker Containers ================================= Run PaddlePaddle in Docker container so that you don't need to care about diff --git a/doc/getstarted/build_and_install/pip_install_cn.rst b/doc/getstarted/build_and_install/pip_install_cn.rst index 41312da48c055826186a560ef9653653e45d1047..b26bf4c95cb18f36408eb75894e8b9b674efc67b 100644 --- a/doc/getstarted/build_and_install/pip_install_cn.rst +++ b/doc/getstarted/build_and_install/pip_install_cn.rst @@ -1,4 +1,4 @@ -使用pip安装PaddlePaddle +使用pip安装 ================================ PaddlePaddle可以使用常用的Python包管理工具 diff --git a/doc/getstarted/build_and_install/pip_install_en.rst b/doc/getstarted/build_and_install/pip_install_en.rst index 4f295e14baa1465a93b8eef1b3f3b6b47eeea905..113790e4e4ca116e91f11f8a233eae874d9d1b7a 100644 --- a/doc/getstarted/build_and_install/pip_install_en.rst +++ b/doc/getstarted/build_and_install/pip_install_en.rst @@ -1,4 +1,4 @@ -Install PaddlePaddle Using pip +Install Using pip ================================ You can use current widely used Python package management diff --git a/doc/howto/index_cn.rst b/doc/howto/index_cn.rst index 76d3e0a0092f89005605a23e14e712530112a5ac..eb95356c67c5df22e4f543f958eb31d79f2c6195 100644 --- a/doc/howto/index_cn.rst +++ b/doc/howto/index_cn.rst @@ -19,7 +19,6 @@ .. toctree:: :maxdepth: 1 - dev/build_cn.rst dev/write_docs_cn.rst 模型配置 diff --git a/doc/howto/index_en.rst b/doc/howto/index_en.rst index 1b6034be4edffd2cbc822018b733b9a3836ea84a..1fbfcd260b912078f00ed5b720ed607db725c4e2 100644 --- a/doc/howto/index_en.rst +++ b/doc/howto/index_en.rst @@ -18,7 +18,6 @@ Development .. toctree:: :maxdepth: 1 - dev/build_en.rst dev/new_layer_en.rst dev/contribute_to_paddle_en.md diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc index a0f2906c749054c1ff9f624e47df432ec2bd6ac8..fdf6de4babff3bb3c253aaf516636882237e6faf 100644 --- a/paddle/framework/lod_tensor.cc +++ b/paddle/framework/lod_tensor.cc @@ -13,6 +13,8 @@ limitations under the License. */ #include "paddle/framework/lod_tensor.h" +#include "paddle/framework/data_type.h" +#include "paddle/framework/framework.pb.h" #include "paddle/memory/memcpy.h" #include "paddle/memory/memory.h" @@ -27,11 +29,11 @@ namespace paddle { namespace framework { -std::ostream& operator<<(std::ostream& os, const LoD& lod) { +std::ostream &operator<<(std::ostream &os, const LoD &lod) { os << "{"; - for (auto& v : lod) { + for (auto &v : lod) { os << "{"; - for (auto& i : v) { + for (auto &i : v) { os << i << ","; } os << "}"; @@ -41,7 +43,7 @@ std::ostream& operator<<(std::ostream& os, const LoD& lod) { return os; } -LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) { +LoD SliceLevels(const LoD &in, size_t level_begin, size_t level_end) { LoD new_lod; new_lod.reserve(level_end - level_begin); for (size_t i = level_begin; i < level_end; i++) { @@ -53,7 +55,7 @@ LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) { return new_lod; } -LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin, +LoD SliceInLevel(const LoD &in, size_t level, size_t elem_begin, size_t elem_end) { PADDLE_ENFORCE_LT(level, in.size()); PADDLE_ENFORCE_LT(elem_end, in[level].size()); @@ -64,9 +66,9 @@ LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin, res[0].assign(in[level].begin() + elem_begin, in[level].begin() + elem_end + 1); for (size_t lvl = 1; lvl < res.size(); lvl++) { - const auto& in_level = in[level + lvl]; - const auto& above_level = res[lvl - 1]; - auto& out_level = res[lvl]; + const auto &in_level = in[level + lvl]; + const auto &above_level = res[lvl - 1]; + auto &out_level = res[lvl]; out_level.assign(in_level.begin() + above_level.front(), in_level.begin() + above_level.back() + 1); } @@ -74,33 +76,33 @@ LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin, // to make the first offset equals 0, all the elements minus the first // element size_t front = res[lvl].front(); - for (auto& ele : res[lvl]) { + for (auto &ele : res[lvl]) { ele -= front; } } return res; } -LoD ToAbsOffset(const LoD& in) { +LoD ToAbsOffset(const LoD &in) { // the lowest level stores relative offsets if (in.empty() || in.size() == 1) return in; LoD result = in; for (int level = result.size() - 2; level >= 0; level--) { - for (auto& ele : result[level]) { + for (auto &ele : result[level]) { ele = result[level + 1][ele]; } } return result; } -bool operator==(const LoD& a, const LoD& b) { +bool operator==(const LoD &a, const LoD &b) { if (a.size() != b.size()) { return false; } for (size_t i = 0; i < a.size(); i++) { - const auto& a_level = a[i]; - const auto& b_level = b[i]; + const auto &a_level = a[i]; + const auto &b_level = b[i]; if (a_level.size() != b_level.size()) { return false; } @@ -151,7 +153,7 @@ void LoDTensor::ShrinkInLevel(size_t level, size_t elem_begin, } using LoDAndOffset = std::pair>; -LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD& lod, size_t start_idx, +LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD &lod, size_t start_idx, size_t end_idx, size_t start_level) { LoD sub_lod; @@ -170,7 +172,7 @@ LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD& lod, size_t start_idx, return LoDAndOffset{sub_lod, {start_idx, end_idx}}; } -void AppendLoD(LoD* lod, const LoD& lod_length) { +void AppendLoD(LoD *lod, const LoD &lod_length) { PADDLE_ENFORCE( lod->empty() || lod->size() == lod_length.size(), "The lod_length should has the same size with the appended lod."); @@ -178,12 +180,139 @@ void AppendLoD(LoD* lod, const LoD& lod_length) { *lod = LoD(lod_length.size(), std::vector({0})); } for (size_t i = 0; i < lod->size(); ++i) { - auto& level = (*lod)[i]; + auto &level = (*lod)[i]; for (size_t len : lod_length[i]) { level.push_back(level.back() + len); } } } +void SerializeToStream(std::ostream &os, const LoDTensor &tensor, + const platform::DeviceContext &dev_ctx) { + // TODO(typhoonzero): serialize to ostream + { // the 1st field, uint32_t version + constexpr uint32_t version = 0; + os.write(reinterpret_cast(&version), sizeof(version)); + } + { // the 2nd field, tensor description + // int32_t size + // void* protobuf message + framework::TensorDesc desc; + desc.set_data_type(framework::ToDataType(tensor.type())); + auto dims = framework::vectorize(tensor.dims()); + auto *pb_dims = desc.mutable_dims(); + pb_dims->Resize(static_cast(dims.size()), 0); + std::copy(dims.begin(), dims.end(), pb_dims->begin()); + int32_t size = desc.ByteSize(); + os.write(reinterpret_cast(&size), sizeof(size)); + auto out = desc.SerializeAsString(); + os.write(out.data(), size); + } + { // the 3rd field, tensor data + uint64_t size = tensor.memory_size(); + auto *data_ptr = tensor.data(); + PADDLE_ENFORCE(size < std::numeric_limits::max(), + "Index overflow when writing tensor"); + if (platform::is_gpu_place(tensor.place())) { +#ifdef PADDLE_WITH_CUDA + constexpr size_t kBufSize = 1024 * 1024 * 64; // 64MB + std::unique_ptr buf(new char[kBufSize]); + auto &gpu_dev_ctx = + static_cast(dev_ctx); + platform::CPUPlace cpu; + uintptr_t data = reinterpret_cast(data_ptr); + while (size != 0) { + size_t size_to_write = std::min(kBufSize, static_cast(size)); + memory::Copy(cpu, buf.get(), + boost::get(tensor.place()), + reinterpret_cast(data), size_to_write, + gpu_dev_ctx.stream()); + gpu_dev_ctx.Wait(); + os.write(buf.get(), size_to_write); + data += size_to_write; + size -= size_to_write; + } +#else + PADDLE_THROW("Unexpected branch"); +#endif + } else { + os.write(static_cast(data_ptr), + static_cast(size)); + } + } + { // the 4th field, lod information + // uint64_t lod_level + // uint64_t lod_level_1 size in byte. + // int* lod_level_1 data + // ... + auto lod = tensor.lod(); + uint64_t size = lod.size(); + os.write(reinterpret_cast(&size), sizeof(size)); + + for (auto &each : lod) { + size = each.size() * sizeof(framework::LoD::value_type::value_type); + os.write(reinterpret_cast(&size), sizeof(size)); + os.write(reinterpret_cast(each.data()), + static_cast(size)); + } + } +} + +void DeserializeFromStream(std::istream &is, LoDTensor *tensor) { + uint32_t version; + is.read(reinterpret_cast(&version), sizeof(version)); + PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported"); + framework::TensorDesc desc; + { // int32_t size + // proto buffer + int32_t size; + is.read(reinterpret_cast(&size), sizeof(size)); + std::unique_ptr buf(new char[size]); + is.read(reinterpret_cast(buf.get()), size); + PADDLE_ENFORCE(desc.ParseFromArray(buf.get(), size), + "Cannot parse tensor desc"); + } + { // read tensor + std::vector dims; + dims.reserve(static_cast(desc.dims().size())); + std::copy(desc.dims().begin(), desc.dims().end(), std::back_inserter(dims)); + tensor->Resize(framework::make_ddim(dims)); + + void *buf; + platform::Place cpu = platform::CPUPlace(); + switch (desc.data_type()) { + case framework::FP32: + buf = tensor->mutable_data(cpu); + break; + case framework::FP64: + buf = tensor->mutable_data(cpu); + break; + case framework::INT32: + buf = tensor->mutable_data(cpu); + break; + case framework::INT64: + buf = tensor->mutable_data(cpu); + break; + default: + PADDLE_THROW("DataType %d not supported", desc.data_type()); + } + is.read(static_cast(buf), tensor->memory_size()); + } + { // read lod + uint64_t lod_level; + is.read(reinterpret_cast(&lod_level), sizeof(lod_level)); + auto &lod = *tensor->mutable_lod(); + lod.resize(lod_level); + for (uint64_t i = 0; i < lod_level; ++i) { + uint64_t size; + is.read(reinterpret_cast(&size), sizeof(size)); + std::vector tmp(size / sizeof(size_t)); + is.read(reinterpret_cast(tmp.data()), + static_cast(size)); + lod[i] = tmp; + } + } +} + } // namespace framework } // namespace paddle diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h index 21bdfca1111f16d5b8ea71be004ddb8da12fd03c..9411c96aea4c10ebf921cc3e3b442769c8acbefa 100644 --- a/paddle/framework/lod_tensor.h +++ b/paddle/framework/lod_tensor.h @@ -189,5 +189,14 @@ std::pair> GetSubLoDAndAbsoluteOffset( void AppendLoD(LoD* lod, const LoD& lod_length); +/* + * Serialize/Desiralize LoDTensor to std::ostream + * You can pass ofstream or ostringstream to serilize to file + * or to a in memory string. GPU tensor will be copied to CPU. + */ +void SerializeToStream(std::ostream& os, const LoDTensor& tensor, + const platform::DeviceContext& dev_ctx); +void DeserializeFromStream(std::istream& is, LoDTensor* tensor); + } // namespace framework } // namespace paddle diff --git a/paddle/gserver/layers/FactorizationMachineLayer.cpp b/paddle/gserver/layers/FactorizationMachineLayer.cpp new file mode 100644 index 0000000000000000000000000000000000000000..be26b9ba88c279036f73b0a0baaff164755fe067 --- /dev/null +++ b/paddle/gserver/layers/FactorizationMachineLayer.cpp @@ -0,0 +1,158 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "FactorizationMachineLayer.h" +#include +#include +#include "paddle/math/SparseMatrix.h" +#include "paddle/utils/Logging.h" +#include "paddle/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(factorization_machine, FactorizationMachineLayer); + +bool FactorizationMachineLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + + factorSize_ = config_.factor_size(); + + /* initialize the latentVectors_ */ + CHECK_EQ(inputLayers_.size(), 1UL); + size_t inputSize = inputLayers_[0]->getSize(); + CHECK_EQ(parameters_[0]->getSize(), inputSize * factorSize_); + latentVectors_ = std::unique_ptr( + new Weight(inputSize, factorSize_, parameters_[0])); + + return true; +} + +void FactorizationMachineLayer::forward(PassType passType) { + Layer::forward(passType); + + const MatrixPtr& inputV = getInputValue(0); + + size_t batchSize = inputV->getHeight(); + size_t outputSize = getSize(); + size_t inputSize = inputLayers_[0]->getSize(); + reserveOutput(batchSize, outputSize); + + MatrixPtr outV = getOutputValue(); + + Matrix::resizeOrCreate( + latentVectorsSquare_, inputSize, factorSize_, false, useGpu_); + Matrix::resizeOrCreate( + inputMulFactor_, batchSize, factorSize_, false, useGpu_); + Matrix::resizeOrCreate(tmpOut_, batchSize, factorSize_, false, useGpu_); + + REGISTER_TIMER_INFO("FmInputMulFactorTimer", getName().c_str()); + inputMulFactor_->mul(*inputV, *latentVectors_->getW()); + inputMulFactor_->square2(*tmpOut_); + outV->sumRows(*tmpOut_, 0.5, 0); + + if (dynamic_cast(inputV.get())) { + Matrix::resizeOrCreateSparseMatrix(inputSquare_, + inputV->getHeight(), + inputV->getWidth(), + inputV->getElementCnt(), + inputV->getValueType()); + inputSquare_->copyFrom(*inputV); + (dynamic_cast(inputSquare_.get()))->square2(); + } else { + Matrix::resizeOrCreate( + inputSquare_, inputV->getHeight(), inputV->getWidth(), false, useGpu_); + inputV->square2(*inputSquare_); + } + latentVectors_->getW()->square2(*latentVectorsSquare_); + tmpOut_->mul(*inputSquare_, *latentVectorsSquare_); + outV->sumRows(*tmpOut_, -0.5, 1.0); + + /* activation */ { + REGISTER_TIMER_INFO("FmFwAtvTimer", getName().c_str()); + forwardActivation(); + } +} + +void FactorizationMachineLayer::backward(const UpdateCallback& callback) { + /* Do derivation */ { backwardActivation(); } + + const MatrixPtr& inputV = getInputValue(0); + const MatrixPtr& oGrad = getOutputGrad(); + + Matrix::resizeOrCreate( + tmpSum_, 1, latentVectors_->getW()->getHeight(), false, useGpu_); + MatrixPtr tmpSumTrans = Matrix::create(tmpSum_->getRowBuf(0), + latentVectors_->getW()->getHeight(), + 1, + false, + useGpu_); + + /* Calculate the gradients of the latentVectors_ matrix */ + if (latentVectors_->getWGrad()) { + if (dynamic_cast(inputV.get())) { + Matrix::resizeOrCreateSparseMatrix(tmpInput_, + inputV->getHeight(), + inputV->getWidth(), + inputV->getElementCnt()); + + CpuSparseMatrix* sparseInputV = + dynamic_cast(inputV.get()); + CpuSparseMatrix* sparseInputSquare = + dynamic_cast(inputSquare_.get()); + CpuSparseMatrix* sparseTmpInput = + dynamic_cast(tmpInput_.get()); + sparseTmpInput->copyFrom(*sparseInputV); + + sparseTmpInput->rowScale(0, *sparseInputV, *oGrad); + latentVectors_->getWGrad()->mul( + *sparseTmpInput->getTranspose(), *inputMulFactor_, 1, 1); + sparseTmpInput->rowScale(0, *sparseInputSquare, *oGrad); + + Matrix::resizeOrCreate(negOnes_, 1, inputV->getHeight(), false, useGpu_); + negOnes_->zeroMem(); + negOnes_->add(-1); + tmpSum_->mul(*negOnes_, *sparseTmpInput, 1, 0); + } else { + Matrix::resizeOrCreate( + tmpInput_, inputV->getHeight(), inputV->getWidth(), false, useGpu_); + + tmpInput_->rowScale(0, *inputV, *oGrad); + latentVectors_->getWGrad()->mul( + *tmpInput_->getTranspose(), *inputMulFactor_, 1, 1); + tmpInput_->rowScale(0, *inputSquare_, *oGrad); + + tmpSum_->sumCols(*tmpInput_, -1, 0); + } + + latentVectors_->getWGrad()->addRowScale( + 0, *latentVectors_->getW(), *tmpSumTrans); + + /* Increasing the number of gradient */ + latentVectors_->getParameterPtr()->incUpdate(callback); + } + + /* Calculate the input layers gradient */ + MatrixPtr inGrad = getInputGrad(0); + if (inGrad != NULL) { + inGrad->mul( + *inputMulFactor_, *latentVectors_->getW()->getTranspose(), 1, 1); + tmpSumTrans->sumRows(*latentVectorsSquare_, -1, 0); + inGrad->addColScale(0, *inputV, *tmpSum_); + inGrad->rowScale(0, *inGrad, *oGrad); + } +} + +} // namespace paddle diff --git a/paddle/gserver/layers/FactorizationMachineLayer.h b/paddle/gserver/layers/FactorizationMachineLayer.h new file mode 100644 index 0000000000000000000000000000000000000000..df20a49934d5dd444f127842c8fdb7c77f4ebeb1 --- /dev/null +++ b/paddle/gserver/layers/FactorizationMachineLayer.h @@ -0,0 +1,80 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "Layer.h" +#include "paddle/math/Matrix.h" +#include "paddle/utils/ThreadLocal.h" + +namespace paddle { +/** + * @brief The Factorization Machine models pairwise (order-2) feature + * interactions as inner product of the learned latent vectors corresponding + * to each input feature. + * + * The Factorization Machine can effectively capture feature interactions + * especially when the input is sparse. While in principle FM can model higher + * order feature interaction, in practice usually only order-2 feature + * interactions are considered. The Factorization Machine Layer here only + * computes the order-2 interations with the formula: + * + * \f[ + * y = \sum_{i=1}^{n-1}\sum_{j=i+1}^n\langle v_i, v_j \rangle x_i x_j + * \f] + * + * The detailed calculation for forward and backward can be found at this paper: + * + * Factorization machines. + * + * The config file api is factorization_machine. + */ + +class FactorizationMachineLayer : public Layer { +protected: + // The latent vectors, shape: (size, factorSize_) + // Each row of the latentVectors_ matrix is the latent vector + // corresponding to one input feature dimension + std::unique_ptr latentVectors_; + // The hyperparameter that defines the dimensionality of the factorization + size_t factorSize_; + +private: + // Store the square values of the letent vectors matrix + MatrixPtr latentVectorsSquare_; + // Store the square values of input matrix + MatrixPtr inputSquare_; + // The result of input matrix * latent vector matrix that will be used in + // both forward and backward step + MatrixPtr inputMulFactor_; + // Store temporary calculation result + MatrixPtr tmpOut_; + MatrixPtr tmpSum_; + MatrixPtr tmpInput_; + // Negative identity matrix + MatrixPtr negOnes_; + +public: + explicit FactorizationMachineLayer(const LayerConfig& config) + : Layer(config) {} + ~FactorizationMachineLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; +}; + +} // namespace paddle diff --git a/paddle/gserver/layers/HierarchicalSigmoidLayer.cpp b/paddle/gserver/layers/HierarchicalSigmoidLayer.cpp index d62a8d846e5b347aa44ce1951c043d5813a5b3ff..236f8096bdb6e024cf3c9c73eba422616a777a23 100644 --- a/paddle/gserver/layers/HierarchicalSigmoidLayer.cpp +++ b/paddle/gserver/layers/HierarchicalSigmoidLayer.cpp @@ -64,49 +64,111 @@ void HierarchicalSigmoidLayer::forward(PassType passType) { batchSize, codeLength_, /* trans */ false, - useGpu(deviceId_)); + false); Matrix::resizeOrCreate(preOutput_.grad, batchSize, codeLength_, /* trans */ false, - useGpu(deviceId_)); - + false); IVectorPtr label = getInput(*getLabelLayer()).ids; - preOutput_.value->zeroMem(); + if (useGpu_) { + Matrix::resizeOrCreate(cpuOutput_, + output_.value->getHeight(), + output_.value->getWidth(), + /* trans */ false, + false); + IVector::resizeOrCreate(cpuLabel_, label->getSize(), false); + cpuLabel_->copyFrom(*label); + cpuOutput_->copyFrom(*output_.value); + } else { + cpuOutput_ = output_.value; + cpuLabel_ = label; + } /* add the bias-vector */ if (biases_.get() != NULL) { - preOutput_.value->addByBitCode(numClasses_, *label, *biases_->getW()); + if (useGpu_) { + Matrix::resizeOrCreate(cpuBias_, + 1, + numClasses_ - 1, + /* trans */ false, + false); + cpuBias_->copyFrom(*biases_->getW()); + } else { + cpuBias_ = biases_->getW(); + } + preOutput_.value->addByBitCode(numClasses_, *cpuLabel_, *cpuBias_); } for (size_t i = 0; i < inputLayers_.size() - 1; ++i) { MatrixPtr input = getInputValue(i); + if (useGpu_) { + Matrix::resizeOrCreate(cpuInput_, + input->getHeight(), + input->getWidth(), + /* trans */ false, + false); + Matrix::resizeOrCreate(cpuWeight_, + weights_[i]->getW()->getHeight(), + weights_[i]->getW()->getWidth(), + /* trans */ false, + false); + cpuInput_->copyFrom(*input); + cpuWeight_->copyFrom(*weights_[i]->getW()); + } else { + cpuInput_ = input; + cpuWeight_ = weights_[i]->getW(); + } preOutput_.value->mulByBitCode( - numClasses_, *label, *weights_[i]->getW(), *input); + numClasses_, *cpuLabel_, *cpuWeight_, *cpuInput_); } // keep consistent with the clipping in the following softrelu preOutput_.value->clip(-40.0, 40.0); preOutput_.value->sumByBitCode(numClasses_, - *label, - *output_.value, + *cpuLabel_, + *cpuOutput_, -1); // scaleSum preOutput_.value->softrelu(*preOutput_.value); - MatrixPtr sum = - Matrix::create(batchSize, 1, /* trans= */ false, useGpu(deviceId_)); + MatrixPtr sum = Matrix::create(batchSize, 1, /* trans= */ false, false); preOutput_.value->rowSum(*sum); - output_.value->add(*sum); + cpuOutput_->add(*sum); + if (useGpu_) { + output_.value->copyFrom(*cpuOutput_); + } else { + output_.value = cpuOutput_; + } } void HierarchicalSigmoidLayer::backward(const UpdateCallback& callback) { IVectorPtr label = getInput(*getLabelLayer()).ids; + if (useGpu_) { + IVector::resizeOrCreate(cpuLabel_, label->getSize(), false); + cpuLabel_->copyFrom(*label); + } else { + cpuLabel_ = label; + } preOutput_.grad->one(); preOutput_.grad->softreluDerivative(*preOutput_.value); - preOutput_.grad->subByBitCode(numClasses_, *label); + preOutput_.grad->subByBitCode(numClasses_, *cpuLabel_); if (biases_ && biases_->getWGrad()) { - preOutput_.grad->addByBitCodeBackward( - numClasses_, *label, *biases_->getWGrad()); - + MatrixPtr biases_grad = biases_->getWGrad(); + if (useGpu_) { + Matrix::resizeOrCreate(cpuBias_, + 1, + numClasses_ - 1, + /* trans */ false, + false); + cpuBias_->copyFrom(*biases_grad); + } else { + cpuBias_ = biases_grad; + } + preOutput_.grad->addByBitCodeBackward(numClasses_, *cpuLabel_, *cpuBias_); + if (useGpu_) { + biases_grad->copyFrom(*cpuBias_); + } else { + biases_grad = cpuBias_; + } /* Increasing the number of gradient */ biases_->getParameterPtr()->incUpdate(callback); } @@ -115,9 +177,31 @@ void HierarchicalSigmoidLayer::backward(const UpdateCallback& callback) { /* Calculate the W-gradient for the current layer */ MatrixPtr input = getInputValue(i); if (weights_[i]->getWGrad()) { + MatrixPtr weights_grad = weights_[i]->getWGrad(); + if (useGpu_) { + Matrix::resizeOrCreate(cpuInput_, + input->getHeight(), + input->getWidth(), + /* trans */ false, + false); + Matrix::resizeOrCreate(cpuWeightGrad_, + weights_grad->getHeight(), + weights_grad->getWidth(), + /* trans */ false, + false); + cpuInput_->copyFrom(*input); + cpuWeightGrad_->copyFrom(*weights_grad); + } else { + cpuInput_ = input; + cpuWeightGrad_ = weights_grad; + } preOutput_.grad->mulByBitCodeBackwardWeight( - numClasses_, *label, *weights_[i]->getWGrad(), *input); - + numClasses_, *cpuLabel_, *cpuWeightGrad_, *cpuInput_); + if (useGpu_) { + weights_grad->copyFrom(*cpuWeightGrad_); + } else { + weights_grad = cpuWeightGrad_; + } /* Increasing the number of gradient */ weights_[i]->getParameterPtr()->incUpdate(callback); } @@ -125,8 +209,30 @@ void HierarchicalSigmoidLayer::backward(const UpdateCallback& callback) { /* Calculate the input layers error */ MatrixPtr inputGrad = getInputGrad(i); if (inputGrad) { + if (useGpu_) { + Matrix::resizeOrCreate(cpuInputGrad_, + inputGrad->getHeight(), + inputGrad->getWidth(), + /* trans */ false, + false); + Matrix::resizeOrCreate(cpuWeight_, + weights_[i]->getW()->getHeight(), + weights_[i]->getW()->getWidth(), + /* trans */ false, + false); + cpuInputGrad_->copyFrom(*inputGrad); + cpuWeight_->copyFrom(*weights_[i]->getW()); + } else { + cpuInputGrad_ = inputGrad; + cpuWeight_ = weights_[i]->getW(); + } preOutput_.grad->mulByBitCodeBackwardError( - numClasses_, *label, *weights_[i]->getW(), *inputGrad); + numClasses_, *cpuLabel_, *cpuWeight_, *cpuInputGrad_); + if (useGpu_) { + inputGrad->copyFrom(*cpuInputGrad_); + } else { + inputGrad = cpuInputGrad_; + } } } } diff --git a/paddle/gserver/layers/HierarchicalSigmoidLayer.h b/paddle/gserver/layers/HierarchicalSigmoidLayer.h index 9afd40b1674680da962d6e51caa56b46279b70de..7f896e61ca26e3e22b99b65b1285384a121f7f02 100644 --- a/paddle/gserver/layers/HierarchicalSigmoidLayer.h +++ b/paddle/gserver/layers/HierarchicalSigmoidLayer.h @@ -80,6 +80,15 @@ protected: int codeLength_; /// temporary result of output_ Argument preOutput_; + + /// The temporary variables in CPU memory. + MatrixPtr cpuWeight_; + MatrixPtr cpuWeightGrad_; + MatrixPtr cpuInput_; + MatrixPtr cpuInputGrad_; + MatrixPtr cpuBias_; + MatrixPtr cpuOutput_; + IVectorPtr cpuLabel_; }; } // namespace paddle diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp index cacf10692942f5eca2f6c498183f4acc00768460..c5359f272b4bed4d4d2483bf19d7ae482b0d33dd 100644 --- a/paddle/gserver/tests/test_LayerGrad.cpp +++ b/paddle/gserver/tests/test_LayerGrad.cpp @@ -681,12 +681,13 @@ TEST(Layer, hsigmoidLayer) { config.layerConfig.add_inputs(); config.layerConfig.add_inputs(); - // Not support GPU now - testLayerGrad(config, - "hsigmoid", - 100, - /* trans */ false, /* useGpu */ - false); + for (auto useGpu : {false, true}) { + testLayerGrad(config, + "hsigmoid", + 100, + /* trans */ false, + /* useGpu */ useGpu); + } } TEST(Layer, multi_cross) { @@ -2464,6 +2465,25 @@ TEST(Layer, L2DistanceLayer) { } } +void testFactorizationMachineLayer(InputType type, bool useGpu) { + const int FACTOR_SIZE = 10; + TestConfig config; + config.layerConfig.set_type("factorization_machine"); + config.layerConfig.set_factor_size(FACTOR_SIZE); + config.layerConfig.set_size(1); + config.biasSize = 0; + config.inputDefs.push_back({type, "layer_0", 128, 1280}); + config.layerConfig.add_inputs(); + testLayerGrad(config, "factorization_machine", 16, false, useGpu, false); +} + +TEST(Layer, FactorizationMachineLayer) { + for (auto useGpu : {false, true}) { + testFactorizationMachineLayer(INPUT_DATA, useGpu); + } + testFactorizationMachineLayer(INPUT_SPARSE_FLOAT_VALUE_DATA, false); +} + int main(int argc, char** argv) { testing::InitGoogleTest(&argc, argv); initMain(argc, argv); diff --git a/paddle/math/CpuSparseMatrix.cpp b/paddle/math/CpuSparseMatrix.cpp index bf62229c03bb1d6e2bdf86d8c56a8157938fb832..dc6979cf5a5229fb09866189f28217889d58c2d0 100644 --- a/paddle/math/CpuSparseMatrix.cpp +++ b/paddle/math/CpuSparseMatrix.cpp @@ -260,6 +260,35 @@ void CpuSparseMatrix::printOneRow(std::ostream& os, size_t idx) const { os << ";"; } +void CpuSparseMatrix::rowScale(size_t cCol, CpuSparseMatrix& b, Matrix& c) { + CHECK(getFormat() != SPARSE_CSC) << "Not supported"; + CHECK_EQ(height_, b.getHeight()); + CHECK_EQ(width_, b.getWidth()); + real* A = getValue(); + real* B = b.getValue(); + if (b.getValueType() == FLOAT_VALUE) { + for (size_t i = 0; i < height_; i++) { + size_t start = getRowStartIdx(i); + size_t end = getRowStartIdx(i + 1); + CHECK_EQ(start, b.getRowStartIdx(i)); + CHECK_EQ(end, b.getRowStartIdx(i + 1)); + for (size_t j = start; j < end; j++) { + A[j] = B[j] * c.getElement(i, cCol); + } + } + } else if (b.getValueType() == NO_VALUE) { + for (size_t i = 0; i < height_; i++) { + size_t start = getRowStartIdx(i); + size_t end = getRowStartIdx(i + 1); + CHECK_EQ(start, b.getRowStartIdx(i)); + CHECK_EQ(end, b.getRowStartIdx(i + 1)); + for (size_t j = start; j < end; j++) { + A[j] = c.getElement(i, cCol); + } + } + } +} + void CpuSparseMatrix::randomizeUniform() { CHECK_LE(elementCnt_, height_ * width_); if (valueType_ == FLOAT_VALUE) { diff --git a/paddle/math/CpuSparseMatrix.h b/paddle/math/CpuSparseMatrix.h index aad1348353d558abca72ed0fa5cf943237e3ac78..522b436a2a69179d3f4f17c919d5ba024102db7b 100644 --- a/paddle/math/CpuSparseMatrix.h +++ b/paddle/math/CpuSparseMatrix.h @@ -239,6 +239,15 @@ public: const unsigned int* cols, const real* values); + /** + * @brief this_row = b_row * c_row[cCol] + * + * @param[in] cCol the column of matrix c used to scale each row of b + * @param[in] b CpuSparseMatrix + * @param[in] c Matrix + */ + void rowScale(size_t cCol, CpuSparseMatrix& b, Matrix& c); + void randomizeUniform(); void copyFrom(const GpuSparseMatrix& src, hl_stream_t stream); diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt index a4c4374cf2f8b4b034d05e3a4c2221300a944214..7e5d4fd640f4399d1a217d1a0be76b3da457c0cc 100644 --- a/paddle/operators/CMakeLists.txt +++ b/paddle/operators/CMakeLists.txt @@ -205,8 +205,24 @@ set(DEPS_OPS tensor_array_read_write_op gru_op adagrad_op - sgd_op) + sgd_op + save_op + load_op + send_op + recv_op) +add_subdirectory(detail) +op_library(send_op SRCS send_op.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf) +set_source_files_properties( + send_op.cc + PROPERTIES + COMPILE_FLAGS "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + +op_library(recv_op SRCS recv_op.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf) +set_source_files_properties( + recv_op.cc + PROPERTIES + COMPILE_FLAGS "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op) op_library(cross_entropy_op DEPS cross_entropy) @@ -235,6 +251,10 @@ op_library(conv_transpose_op DEPS vol2col) op_library(gru_op DEPS sequence2batch gru_compute) op_library(recurrent_op SRCS recurrent_op.cc DEPS executor) +# FIXME(typhoonzero): save/load depends lodtensor serialization functions +op_library(save_op DEPS lod_tensor) +op_library(load_op DEPS lod_tensor) + list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS}) foreach(src ${GENERAL_OPS}) op_library(${src}) @@ -242,6 +262,8 @@ endforeach() set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library") + + cc_test(gather_test SRCS gather_test.cc DEPS tensor) cc_test(net_op_test SRCS net_op_test.cc DEPS net_op) cc_test(scatter_test SRCS scatter_test.cc DEPS tensor) @@ -251,3 +273,4 @@ if(WITH_GPU) cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context) endif() cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op) +cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS send_op recv_op sum_op executor) diff --git a/paddle/operators/batch_norm_op.cc b/paddle/operators/batch_norm_op.cc index f884e6efa917ce3f8554dce0e248f2b29273e3f3..ac97bd83ab7e7838871586cfe5acb832084b6cec 100644 --- a/paddle/operators/batch_norm_op.cc +++ b/paddle/operators/batch_norm_op.cc @@ -62,13 +62,14 @@ class BatchNormOp : public framework::OperatorWithKernel { const auto x_dims = ctx->GetInputDim("X"); const TensorFormat tensor_format = StringToTensorFormat(ctx->Attrs().Get("tensor_format")); + + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "Input X must have 2 to 5 dimensions."); + const int C = (tensor_format == TensorFormat::NCHW ? x_dims[1] : x_dims[x_dims.size() - 1]); - PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5, - "Input X must have 3 to 5 dimensions."); - PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL); PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], C); PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1UL); @@ -146,8 +147,8 @@ class BatchNormKernel : public framework::OpKernel { const auto *x = ctx.Input("X"); const auto &x_dims = x->dims(); - PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5, - "The Input dim size should be between 3 and 5"); + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "The Input dim size should be between 2 and 5"); const int N = x_dims[0]; const int C = (tensor_format == TensorFormat::NCHW ? x_dims[1] @@ -339,8 +340,8 @@ class BatchNormGradKernel // Get the size for each dimension. // NCHW [batch_size, in_channels, in_height, in_width] const auto &x_dims = x->dims(); - PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5, - "The Input dim size should be between 3 and 5"); + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "The Input dim size should be between 2 and 5"); const int N = x_dims[0]; const int C = (tensor_format == TensorFormat::NCHW ? x_dims[1] diff --git a/paddle/operators/batch_norm_op.cu.cc b/paddle/operators/batch_norm_op.cu.cc index 726d1ea1b8d7ced93f94bb0e5bb4df9e43b0ac7b..7b2f3187007fa2491afa75de1cde1910c6ce9bb8 100644 --- a/paddle/operators/batch_norm_op.cu.cc +++ b/paddle/operators/batch_norm_op.cu.cc @@ -29,14 +29,21 @@ void ExtractNCWHD(const framework::DDim &dims, const TensorFormat &tensor_format, int *N, int *C, int *H, int *W, int *D) { *N = dims[0]; - *C = tensor_format == TensorFormat::NCHW ? dims[1] : dims[dims.size() - 1]; - *H = tensor_format == TensorFormat::NCHW ? dims[2] : dims[1]; - *W = dims.size() > 3 - ? (tensor_format == TensorFormat::NCHW ? dims[3] : dims[2]) - : 1; - *D = dims.size() > 4 - ? (tensor_format == TensorFormat::NCHW ? dims[4] : dims[3]) - : 1; + if (dims.size() == 2) { + *C = dims[1]; + *H = 1; + *W = 1; + *D = 1; + } else { + *C = tensor_format == TensorFormat::NCHW ? dims[1] : dims[dims.size() - 1]; + *H = tensor_format == TensorFormat::NCHW ? dims[2] : dims[1]; + *W = dims.size() > 3 + ? (tensor_format == TensorFormat::NCHW ? dims[3] : dims[2]) + : 1; + *D = dims.size() > 4 + ? (tensor_format == TensorFormat::NCHW ? dims[4] : dims[3]) + : 1; + } } template @@ -56,8 +63,8 @@ class BatchNormKernel : public framework::OpKernel { // NCHW [batch_size, in_channels, in_height, in_width] const auto *x = ctx.Input("X"); const auto &x_dims = x->dims(); - PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5, - "The Input dim size should be between 3 and 5"); + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "The Input dim size should be between 2 and 5"); int N, C, H, W, D; ExtractNCWHD(x_dims, tensor_format, &N, &C, &H, &W, &D); @@ -180,8 +187,8 @@ class BatchNormGradKernel const auto &x_dims = x->dims(); - PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5, - "The Input dim size should be between 3 and 5"); + PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5, + "The Input dim size should be between 2 and 5"); int N, C, H, W, D; ExtractNCWHD(x_dims, tensor_format, &N, &C, &H, &W, &D); diff --git a/paddle/operators/conv_cudnn_op.cu.cc b/paddle/operators/conv_cudnn_op.cu.cc index a9763d424801cfced5fe4c4718a335a24b81cfdc..3f97dc7ee0a61944a8a57314b5ec7f33df619bf3 100644 --- a/paddle/operators/conv_cudnn_op.cu.cc +++ b/paddle/operators/conv_cudnn_op.cu.cc @@ -63,7 +63,7 @@ class CudnnConvOpKernel : public framework::OpKernel { cudnnConvolutionDescriptor_t cudnn_conv_desc = conv_desc.descriptor(paddings, strides, dilations); -#if CUDNN_VERSION_MIN(7, 0, 0) +#if CUDNN_VERSION_MIN(7, 0, 1) // cudnn 7 can support groups, no need to do it mannually // FIXME(typhoonzero): find a better way to disable groups // rather than setting it to 1. @@ -180,7 +180,7 @@ class CudnnConvGradOpKernel : public framework::OpKernel { cudnnConvolutionDescriptor_t cudnn_conv_desc = conv_desc.descriptor(paddings, strides, dilations); -#if CUDNN_VERSION_MIN(7, 0, 0) +#if CUDNN_VERSION_MIN(7, 0, 1) // cudnn 7 can support groups, no need to do it mannually // FIXME(typhoonzero): find a better way to disable groups // rather than setting it to 1. diff --git a/paddle/operators/detail/CMakeLists.txt b/paddle/operators/detail/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..f6bdc63cc2cfae526fe911ee4d989675452d5c5d --- /dev/null +++ b/paddle/operators/detail/CMakeLists.txt @@ -0,0 +1 @@ +grpc_library(sendrecvop_grpc SRCS recv_impl.cc send_impl.cc PROTO send_recv.proto DEPS lod_tensor selected_rows) diff --git a/paddle/operators/detail/recv_impl.cc b/paddle/operators/detail/recv_impl.cc new file mode 100644 index 0000000000000000000000000000000000000000..89dc5045221156eed7aa9411bc96ad86f91136d2 --- /dev/null +++ b/paddle/operators/detail/recv_impl.cc @@ -0,0 +1,44 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "send_recv_impl.h" + +namespace paddle { +namespace operators { +namespace detail { + +Status SendRecvServerImpl::SendVariable(ServerContext *context, + const VariableMessage *in_var, + VariableMessage *out_var) { + framework::LoDTensor t; + // TODO(typhoonzero): desirealize in_tensor and run pserver network. + std::istringstream iss(in_var->serialized()); + framework::DeserializeFromStream(iss, &t); + lodtensor_queue_.Push(std::move(t)); + // Block util the sub graph is done. + t = lodtensor_return_queue_.Pop(); + std::ostringstream oss; + // FIXME(typhoonzero): get context from op. + framework::SerializeToStream(oss, t, platform::CPUDeviceContext()); + std::string *varname = out_var->mutable_varname(); + *varname = in_var->varname(); + std::string *serialized = out_var->mutable_serialized(); + *serialized = oss.str(); + + return Status::OK; +} + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/detail/send_impl.cc b/paddle/operators/detail/send_impl.cc new file mode 100644 index 0000000000000000000000000000000000000000..da1ddf75d2afb85670c5ea0c9884376415f28208 --- /dev/null +++ b/paddle/operators/detail/send_impl.cc @@ -0,0 +1,54 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "send_recv_impl.h" + +namespace paddle { +namespace operators { +namespace detail { + +bool RPCClient::SendVariable(const framework::Scope& scope, + const std::string& inname, + const std::string& outname) { + ClientContext context; + VariableMessage msg, out_msg; + // FIXME(typhoonzero): pass device context to here. + auto ctx = platform::CPUDeviceContext(); + auto* var = scope.FindVar(inname); + PADDLE_ENFORCE(var); + // TODO(typhoonzero): support SelectedRows + PADDLE_ENFORCE(var->IsType(), + "Only support LoDTensor, %s has wrong type", inname); + const framework::LoDTensor& tensor = var->Get(); + std::ostringstream oss; + framework::SerializeToStream(oss, tensor, ctx); + msg.set_varname(inname); + msg.set_serialized(oss.str()); + Status status = stub_->SendVariable(&context, msg, &out_msg); + if (!status.ok()) { + return false; + } + std::istringstream iss(out_msg.serialized()); + framework::LoDTensor ret_tensor; + framework::DeserializeFromStream(iss, &ret_tensor); + auto* outvar = scope.FindVar(outname); + framework::LoDTensor* out_tensor = outvar->GetMutable(); + // FIXME(typhoonzero): do not copy. + framework::CopyFrom(ret_tensor, ctx.GetPlace(), ctx, out_tensor); + return true; +} + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/detail/send_recv.proto b/paddle/operators/detail/send_recv.proto new file mode 100644 index 0000000000000000000000000000000000000000..962c7d59819dede022474aec4a2d7f538d28c688 --- /dev/null +++ b/paddle/operators/detail/send_recv.proto @@ -0,0 +1,35 @@ +/* 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. */ + +syntax = "proto3"; + +package sendrecv; + +service SendRecvService { + // For parameter server round-robin like hashing, do not split tensors. + // Send and recv only one tensor + rpc SendVariable(VariableMessage) returns (VariableMessage) {} +} + +// VariableMessage is serialized paddle variable message. +// It can be: +// Tensor +// LoDTensor +// SelectedRows +message VariableMessage { + string varname = 1; + bytes serialized = 2; +} + +message VoidMessage {} \ No newline at end of file diff --git a/paddle/operators/detail/send_recv_impl.h b/paddle/operators/detail/send_recv_impl.h new file mode 100644 index 0000000000000000000000000000000000000000..b9a5340a8636db7b5d6ec7b21368632d3916b4aa --- /dev/null +++ b/paddle/operators/detail/send_recv_impl.h @@ -0,0 +1,87 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/framework/data_type.h" +#include "paddle/framework/lod_tensor.h" +#include "paddle/framework/scope.h" +#include "paddle/framework/selected_rows.h" +#include "paddle/operators/detail/simple_block_queue.h" + +// #include +// #include +// #include +// #include +#include "paddle/operators/detail/send_recv.grpc.pb.h" +#include "paddle/operators/detail/send_recv.pb.h" + +#include + +using grpc::Channel; +using grpc::Server; +using grpc::ServerContext; +using grpc::ServerReader; +using grpc::ServerBuilder; + +using grpc::ClientContext; +using grpc::ClientReader; +using grpc::ClientReaderWriter; +using grpc::ClientWriter; +using grpc::Status; +using sendrecv::SendRecvService; +using sendrecv::VariableMessage; +using sendrecv::VoidMessage; + +namespace paddle { +namespace operators { +namespace detail { + +class SendRecvServerImpl final : public SendRecvService::Service { + public: + explicit SendRecvServerImpl() {} + + Status SendVariable(ServerContext *context, const VariableMessage *in_var, + VariableMessage *out_var) override; + + const framework::LoDTensor Get() { return this->lodtensor_queue_.Pop(); } + + void Push(const framework::LoDTensor &tensor) { + this->lodtensor_return_queue_.Push(tensor); + } + + private: + SimpleBlockQueue lodtensor_queue_; + SimpleBlockQueue lodtensor_return_queue_; + SimpleBlockQueue selected_rows_queue_; + SimpleBlockQueue selected_rows_return_queue_; +}; + +// RPCClient is a class to send tensors to pserver sub-network +// using different hashing methods. +class RPCClient { + public: + RPCClient(std::shared_ptr channel) + : stub_(SendRecvService::NewStub(channel)) {} + + bool SendVariable(const framework::Scope &scope, const std::string &inname, + const std::string &outname); + + private: + std::unique_ptr stub_; +}; + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/detail/simple_block_queue.h b/paddle/operators/detail/simple_block_queue.h new file mode 100644 index 0000000000000000000000000000000000000000..44899217579532af2c1d2e6074ec0e08231e7b86 --- /dev/null +++ b/paddle/operators/detail/simple_block_queue.h @@ -0,0 +1,52 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include +#include +#include + +namespace paddle { +namespace operators { +namespace detail { + +template +class SimpleBlockQueue { + private: + std::mutex mutex_; + std::condition_variable condition_; + std::deque queue_; + + public: + void Push(T const& value) { + { + std::unique_lock lock(this->mutex_); + queue_.push_front(value); + } + this->condition_.notify_one(); + } + + T Pop() { + std::unique_lock lock(this->mutex_); + this->condition_.wait(lock, [=] { return !this->queue_.empty(); }); + T rc(std::move(this->queue_.back())); + this->queue_.pop_back(); + return rc; + } +}; + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/load_op.cc b/paddle/operators/load_op.cc index b0838eed1611c1d51e57fc2300606f753982dc89..4e58b84430f2a8697bbbc1acf971fd063120f563 100644 --- a/paddle/operators/load_op.cc +++ b/paddle/operators/load_op.cc @@ -38,61 +38,7 @@ class LoadOp : public framework::OperatorBase { out_var_name); auto *tensor = out_var->GetMutable(); - - uint32_t version; - fin.read(reinterpret_cast(&version), sizeof(version)); - PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported"); - framework::TensorDesc desc; - { // int32_t size - // proto buffer - int32_t size; - fin.read(reinterpret_cast(&size), sizeof(size)); - std::unique_ptr buf(new char[size]); - fin.read(reinterpret_cast(buf.get()), size); - PADDLE_ENFORCE(desc.ParseFromArray(buf.get(), size), - "Cannot parse tensor desc"); - } - { // read tensor - std::vector dims; - dims.reserve(static_cast(desc.dims().size())); - std::copy(desc.dims().begin(), desc.dims().end(), - std::back_inserter(dims)); - tensor->Resize(framework::make_ddim(dims)); - - void *buf; - platform::Place cpu = platform::CPUPlace(); - switch (desc.data_type()) { - case framework::FP32: - buf = tensor->mutable_data(cpu); - break; - case framework::FP64: - buf = tensor->mutable_data(cpu); - break; - case framework::INT32: - buf = tensor->mutable_data(cpu); - break; - case framework::INT64: - buf = tensor->mutable_data(cpu); - break; - default: - PADDLE_THROW("DataType %d not supported", desc.data_type()); - } - fin.read(static_cast(buf), tensor->memory_size()); - } - { // read lod - uint64_t lod_level; - fin.read(reinterpret_cast(&lod_level), sizeof(lod_level)); - auto &lod = *tensor->mutable_lod(); - lod.resize(lod_level); - for (uint64_t i = 0; i < lod_level; ++i) { - uint64_t size; - fin.read(reinterpret_cast(&size), sizeof(size)); - std::vector tmp(size / sizeof(size_t)); - fin.read(reinterpret_cast(tmp.data()), - static_cast(size)); - lod[i] = tmp; - } - } + framework::DeserializeFromStream(fin, tensor); auto place = dev_ctx.GetPlace(); if (platform::is_gpu_place(place)) { diff --git a/paddle/operators/log_loss_op.cc b/paddle/operators/log_loss_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..257e5c8a49e935dcbdc33e5060118ef1804fa8d7 --- /dev/null +++ b/paddle/operators/log_loss_op.cc @@ -0,0 +1,115 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/log_loss_op.h" + +namespace paddle { +namespace operators { + +class LogLossOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Predicted"), + "Input(Predicted) must be initialized."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input(Labels) must be initialized."); + + auto pred_dims = ctx->GetInputDim("Predicted"); + auto label_dims = ctx->GetInputDim("Labels"); + + PADDLE_ENFORCE_EQ(pred_dims, label_dims); + PADDLE_ENFORCE_EQ(pred_dims.size(), 2, + "The rank of Input(Predicted) must be 2 and the shape is " + "[batch_size, 1]."); + PADDLE_ENFORCE_EQ(pred_dims[1], 1, + "Each row of Input(Predicted) contains a real value, " + "so the 2nd dimension of Input(X) must be 1."); + + ctx->SetOutputDim("Loss", {pred_dims[0], 1}); + ctx->ShareLoD("Predicted", "Loss"); + } +}; + +template +class LogLossOpMaker : public framework::OpProtoAndCheckerMaker { + public: + LogLossOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Predicted", + "The input value (Predicted) of Log loss op." + "Predicted is a 2-D tensor with shape [batch_size, 1]."); + AddInput("Labels", + "The target value (Labels) of Log loss op." + "Labels is a 2-D tensor with shape [batch_size, 1]."); + AddOutput("Loss", + "The output tensor with shape [batch_size, 1] " + "which represents the log loss."); + AddAttr("epsilon", "Epsilon in log loss."); + AddComment(R"DOC( +LogLoss Operator. + +Log loss is a loss function used for binary classification. Log Loss quantifies +the accuracy of a classifier by penalising false classifications. Minimising the +Log Loss is equivalent to maximising the accuracy of the classifier. We define +Predicted as the values predicted by our model and Labels as the target ground +truth value. Log loss can evaluate how close the predicted values are to the +target. The shapes of Predicted and Labels are both [batch_size, 1]. +The equation is: + +$$ +Loss = - Labels * log(Predicted + \epsilon) - + (1 - Labels) * log(1 - Predicted + \epsilon) +$$ + +)DOC"); + } +}; + +class LogLossGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Predicted"), + "Input(Predicted) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input(Labels) should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Loss")), + "Input(Loss@GRAD) should not be null."); + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Predicted")), + "Output(Predicted@GRAD) should not be null."); + + auto pred_dims = ctx->GetInputDim("Predicted"); + auto label_dims = ctx->GetInputDim("Labels"); + auto loss_grad_dims = ctx->GetInputDim(framework::GradVarName("Loss")); + PADDLE_ENFORCE_EQ(loss_grad_dims, pred_dims); + + auto pred_grad_name = framework::GradVarName("Predicted"); + ctx->SetOutputDim(pred_grad_name, pred_dims); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(log_loss, ops::LogLossOp, ops::LogLossOpMaker, log_loss_grad, + ops::LogLossGradOp); +REGISTER_OP_CPU_KERNEL(log_loss, + ops::LogLossKernel); +REGISTER_OP_CPU_KERNEL( + log_loss_grad, ops::LogLossGradKernel); diff --git a/paddle/operators/log_loss_op.cu b/paddle/operators/log_loss_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..6c189ef3412d7a56205502c7913e93218a03b929 --- /dev/null +++ b/paddle/operators/log_loss_op.cu @@ -0,0 +1,22 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/log_loss_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL(log_loss, + ops::LogLossKernel); +REGISTER_OP_GPU_KERNEL( + log_loss_grad, ops::LogLossGradKernel); diff --git a/paddle/operators/log_loss_op.h b/paddle/operators/log_loss_op.h new file mode 100644 index 0000000000000000000000000000000000000000..73404fce9157fa750a51451fa93646bc4059481a --- /dev/null +++ b/paddle/operators/log_loss_op.h @@ -0,0 +1,75 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenVector = framework::EigenVector; + +template +class LogLossKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* loss_out = ctx.Output("Loss"); + + loss_out->mutable_data(ctx.GetPlace()); + + auto epsilon = static_cast(ctx.Attr("epsilon")); + + auto prediction = EigenVector::Flatten(*ctx.Input("Predicted")); + auto label = EigenVector::Flatten(*ctx.Input("Labels")); + + auto loss = EigenVector::Flatten(*loss_out); + auto place = ctx.GetEigenDevice(); + + loss.device(place) = (-(label * (prediction + epsilon).log()) - + ((static_cast(1) - label) * + (static_cast(1) - prediction + epsilon).log())); + } +}; + +template +class LogLossGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto epsilon = static_cast(ctx.Attr("epsilon")); + + auto prediction = EigenVector::Flatten(*ctx.Input("Predicted")); + auto label = EigenVector::Flatten(*ctx.Input("Labels")); + + auto* dloss = ctx.Input(framework::GradVarName("Loss")); + auto* dpred = ctx.Output(framework::GradVarName("Predicted")); + + auto dl = EigenVector::Flatten(*dloss); + auto place = ctx.GetEigenDevice(); + + if (dpred) { + dpred->mutable_data(ctx.GetPlace()); + auto dx = framework::EigenVector::Flatten(*dpred); + dx.device(place) = dl * (-(label / (prediction + epsilon)) + + ((static_cast(1) - label) / + (static_cast(1) - prediction + epsilon))); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/recv_op.cc b/paddle/operators/recv_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..c69e416e10f2a9ced1f1b22c39235e4c9338e77c --- /dev/null +++ b/paddle/operators/recv_op.cc @@ -0,0 +1,121 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include +#include +#include +#include + +#include + +#include "paddle/framework/data_type.h" +#include "paddle/framework/executor.h" +#include "paddle/framework/framework.pb.h" +#include "paddle/framework/lod_tensor.h" +#include "paddle/framework/op_registry.h" +#include "paddle/operators/detail/send_recv_impl.h" +#include "paddle/operators/detail/simple_block_queue.h" + +namespace paddle { +namespace operators { + +void RunServer(Server **rpc_server, + std::shared_ptr service, + const std::string &server_address) { + ServerBuilder builder; + builder.AddListeningPort(server_address, grpc::InsecureServerCredentials()); + builder.RegisterService(service.get()); + std::unique_ptr server(builder.BuildAndStart()); + *rpc_server = server.get(); + LOG(INFO) << "Server listening on " << server_address << std::endl; + server->Wait(); +} + +class RecvOp : public framework::OperatorBase { + public: + RecvOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorBase(type, inputs, outputs, attrs) { + if (!rpc_service_) { + rpc_service_.reset(new detail::SendRecvServerImpl()); + std::string endpoint = Attr("endpoint"); + server_thread_.reset( + new std::thread(RunServer, &rpc_server_, rpc_service_, endpoint)); + } + } + + virtual ~RecvOp() { + rpc_server_->Shutdown(); + server_thread_->join(); + } + + void Run(const framework::Scope &scope, + const platform::DeviceContext &dev_ctx) const override { + // blocking get one var from client. + const framework::LoDTensor &t = rpc_service_->Get(); + framework::Scope &recv_scope = scope.NewScope(); + // set graph input var + auto *var = recv_scope.Var(Input("RX")); + auto *tensor = var->GetMutable(); + // FIXME(typhoonzero): do not copy + framework::CopyFrom(t, dev_ctx.GetPlace(), dev_ctx, tensor); + + auto *block = Attr("OptimizeBlock"); + auto *program = block->Program(); + framework::Executor executor(dev_ctx); + // Run sub graph to get optimized tensor + executor.Run(*program, &recv_scope, block->ID(), + false /*create_local_scope*/); + + auto *out_var = recv_scope.FindVar("Out"); + // push back + rpc_service_->Push(out_var->Get()); + } + + protected: + // grpc server instance to track status and gracefully shutdown. + // borrow an pointer from server thread. + Server *rpc_server_{nullptr}; + // grpc send/recv service implement to register. + std::shared_ptr rpc_service_; + std::shared_ptr server_thread_; +}; + +class RecvOpMaker : public framework::OpProtoAndCheckerMaker { + public: + RecvOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("RX", "(Tensor) Input tensor to be saved"); + AddComment(R"DOC( +Recv operator + +This operator will recv tensor from send_op +)DOC"); + AddAttr("endpoint", + "(string, default 127.0.0.1:6164)" + "IP address to listen on.") + .SetDefault("127.0.0.1:6164") + .AddCustomChecker([](const std::string &ip) { return !ip.empty(); }); + AddAttr("OptimizeBlock", "type BlockDescBind*", + "optimize network run in server"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(recv, ops::RecvOp, ops::RecvOpMaker); diff --git a/paddle/operators/roi_pool_op.h b/paddle/operators/roi_pool_op.h index 1691eb482b03eab9fc793974ba1f39fbf17beafa..3812c66c65457b9d1337690d1a82759aab9a9732 100644 --- a/paddle/operators/roi_pool_op.h +++ b/paddle/operators/roi_pool_op.h @@ -133,53 +133,47 @@ class CPUROIPoolGradOpKernel : public framework::OpKernel { auto* in = ctx.Input("X"); auto* rois = ctx.Input("ROIs"); auto* argmax = ctx.Input("Argmax"); - auto* out_grad = ctx.Input(framework::GradVarName("Out")); - auto* x_grad = ctx.Output(framework::GradVarName("X")); + auto* in_grad = ctx.Output(framework::GradVarName("X")); auto pooled_height = ctx.Attr("pooled_height"); auto pooled_width = ctx.Attr("pooled_width"); - if (x_grad) { - int channels = in->dims()[1]; - auto in_stride = framework::stride(in->dims()); - auto roi_stride = framework::stride(rois->dims()); - + if (in_grad) { const int64_t* rois_data = rois->data(); - int rois_num = rois->dims()[0]; - - T* x_grad_data = x_grad->mutable_data(ctx.GetPlace()); + const T* out_grad_data = out_grad->data(); + const int64_t* argmax_data = argmax->data(); + T* in_grad_data = in_grad->mutable_data(ctx.GetPlace()); math::SetConstant set_zero; - set_zero(ctx.device_context(), x_grad, static_cast(0)); + set_zero(ctx.device_context(), in_grad, static_cast(0)); - size_t roi_offset = roi_stride[0]; - size_t batch_offset = in_stride[0]; - size_t channel_offset = in_stride[1]; + auto in_stride = framework::stride(in->dims()); + auto argmax_stride = framework::stride(argmax->dims()); + auto roi_stride = framework::stride(rois->dims()); + auto out_stride = framework::stride(out_grad->dims()); - const T* out_grad_data = out_grad->data(); - size_t pool_channel_offset = pooled_height * pooled_width; - const int64_t* argmax_data = argmax->data(); + int rois_num = rois->dims()[0]; + int channels = in->dims()[1]; - for (size_t n = 0; n < rois_num; ++n) { - size_t roi_batch_idx = rois_data[0]; - T* batch_grad_data = x_grad_data + batch_offset * roi_batch_idx; + for (int n = 0; n < rois_num; ++n) { + int roi_batch_idx = rois_data[0]; + T* batch_grad_data = in_grad_data + roi_batch_idx * in_stride[0]; for (int c = 0; c < channels; ++c) { for (int ph = 0; ph < pooled_height; ++ph) { for (int pw = 0; pw < pooled_width; ++pw) { - size_t pool_index = ph * pooled_width + pw; - + int pool_index = ph * pooled_width + pw; if (argmax_data[pool_index] >= 0) { - size_t index = static_cast(argmax_data[pool_index]); + auto index = argmax_data[pool_index]; batch_grad_data[index] += out_grad_data[pool_index]; } } } - batch_grad_data += channel_offset; - out_grad_data += pool_channel_offset; - argmax_data += pool_channel_offset; + batch_grad_data += in_stride[1]; + out_grad_data += out_stride[1]; + argmax_data += argmax_stride[1]; } - rois_data += roi_offset; + rois_data += roi_stride[0]; } } } diff --git a/paddle/operators/save_op.cc b/paddle/operators/save_op.cc index 56909fb65f44ad00314103e21bee9535fbd59317..d4921cb80c8d78c52ae1887c36819b52621470eb 100644 --- a/paddle/operators/save_op.cc +++ b/paddle/operators/save_op.cc @@ -88,73 +88,7 @@ class SaveOp : public framework::OperatorBase { "SaveOp only support LoDTensor, %s has wrong type", iname); auto &tensor = var->Get(); - - { // the 1st field, uint32_t version - constexpr uint32_t version = 0; - fout.write(reinterpret_cast(&version), sizeof(version)); - } - { // the 2nd field, tensor description - // int32_t size - // void* protobuf message - framework::TensorDesc desc; - desc.set_data_type(framework::ToDataType(tensor.type())); - auto dims = framework::vectorize(tensor.dims()); - auto *pb_dims = desc.mutable_dims(); - pb_dims->Resize(static_cast(dims.size()), 0); - std::copy(dims.begin(), dims.end(), pb_dims->begin()); - int32_t size = desc.ByteSize(); - fout.write(reinterpret_cast(&size), sizeof(size)); - auto out = desc.SerializeAsString(); - fout.write(out.data(), size); - } - { // the 3rd field, tensor data - uint64_t size = tensor.memory_size(); - auto *data_ptr = tensor.data(); - PADDLE_ENFORCE(size < std::numeric_limits::max(), - "Index overflow when writing tensor"); - if (platform::is_gpu_place(tensor.place())) { -#ifdef PADDLE_WITH_CUDA - constexpr size_t kBufSize = 1024 * 1024 * 64; // 64MB - std::unique_ptr buf(new char[kBufSize]); - auto &gpu_dev_ctx = - static_cast(dev_ctx); - platform::CPUPlace cpu; - uintptr_t data = reinterpret_cast(data_ptr); - while (size != 0) { - size_t size_to_write = std::min(kBufSize, static_cast(size)); - memory::Copy(cpu, buf.get(), - boost::get(tensor.place()), - reinterpret_cast(data), size_to_write, - gpu_dev_ctx.stream()); - gpu_dev_ctx.Wait(); - fout.write(buf.get(), size_to_write); - data += size_to_write; - size -= size_to_write; - } -#else - PADDLE_THROW("Unexpected branch"); -#endif - } else { - fout.write(static_cast(data_ptr), - static_cast(size)); - } - } - { // the 4th field, lod information - // uint64_t lod_level - // uint64_t lod_level_1 size in byte. - // int* lod_level_1 data - // ... - auto lod = tensor.lod(); - uint64_t size = lod.size(); - fout.write(reinterpret_cast(&size), sizeof(size)); - - for (auto &each : lod) { - size = each.size() * sizeof(framework::LoD::value_type::value_type); - fout.write(reinterpret_cast(&size), sizeof(size)); - fout.write(reinterpret_cast(each.data()), - static_cast(size)); - } - } + framework::SerializeToStream(fout, tensor, dev_ctx); } }; diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc index 5745580504fb9bda551f21665bff5c65ae82aeb9..e5c10fec4d840c58a74758a65ddfa93421ab4827 100644 --- a/paddle/operators/scale_op.cc +++ b/paddle/operators/scale_op.cc @@ -77,4 +77,6 @@ REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker, ops::ScaleGradMaker); REGISTER_OP_CPU_KERNEL(scale, ops::ScaleKernel, - ops::ScaleKernel); + ops::ScaleKernel, + ops::ScaleKernel, + ops::ScaleKernel); diff --git a/paddle/operators/scale_op.cu b/paddle/operators/scale_op.cu index 820fd4e6855bb192ec3292ea6983d5ecae73b6e6..0d707751598e65bc56bf73a435c10b4acd6d8ed0 100644 --- a/paddle/operators/scale_op.cu +++ b/paddle/operators/scale_op.cu @@ -16,4 +16,6 @@ REGISTER_OP_GPU_KERNEL( scale, paddle::operators::ScaleKernel, - paddle::operators::ScaleKernel); + paddle::operators::ScaleKernel, + paddle::operators::ScaleKernel, + paddle::operators::ScaleKernel); diff --git a/paddle/operators/send_op.cc b/paddle/operators/send_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a3059847f2d420359b347e3a5d514d8a3829a4e2 --- /dev/null +++ b/paddle/operators/send_op.cc @@ -0,0 +1,84 @@ +/* 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/framework/data_type.h" +#include "paddle/framework/framework.pb.h" +#include "paddle/framework/lod_tensor.h" +#include "paddle/framework/op_registry.h" + +#include "paddle/operators/detail/send_recv_impl.h" +#include "paddle/operators/detail/simple_block_queue.h" + +namespace paddle { +namespace operators { + +// TODO(typhoonzero): this is a simple implementation which only send +// one tensor +class SendOp : public framework::OperatorBase { + public: + SendOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorBase(type, inputs, outputs, attrs) { + // init client when the operator is created at runtime. + if (!client_) { + std::string endpoint = Attr("endpoint"); + client_.reset(new detail::RPCClient( + grpc::CreateChannel(endpoint, grpc::InsecureChannelCredentials()))); + // TODO(typhoonzero): how to call InitVariables + } + } + void Run(const framework::Scope &scope, + const platform::DeviceContext &dev_ctx) const override { + auto iname = Input("X"); + auto oname = Output("Out"); + // TODO(typhoonzero): currently it's non-blocking, + // should block until server responds. + bool ret = client_->SendVariable(scope, iname, oname); + if (!ret) { + LOG(ERROR) << "send variable error"; + } + } + + protected: + std::shared_ptr client_{nullptr}; +}; + +class SendOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SendOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "(Tensor) Input tensor to be saved"); + AddOutput("Out", "(Tensor) Output fetched from server"); + AddComment(R"DOC( +Recv operator + +This operator will recv tensor from send_op +)DOC"); + AddAttr("endpoint", + "(string, default 127.0.0.1:6164)" + "IP address to listen on.") + .SetDefault("127.0.0.1:6164") + .AddCustomChecker([](const std::string &ip) { return !ip.empty(); }); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(send, ops::SendOp, ops::SendOpMaker); diff --git a/paddle/operators/send_recv_op_test.cc b/paddle/operators/send_recv_op_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..ac03eb3752e7cd31dd80f4caa39dc0625f0409d5 --- /dev/null +++ b/paddle/operators/send_recv_op_test.cc @@ -0,0 +1,125 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +// TODO(typhoonzero): add python bindings for this test as +// a RemoteOptimizer. + +#include +#include + +#include "gtest/gtest.h" +#include "paddle/framework/op_registry.h" +#include "paddle/framework/operator.h" +#include "paddle/framework/program_desc.h" + +USE_NO_KERNEL_OP(send); +USE_NO_KERNEL_OP(recv); +USE_OP(sum); + +// global for simplicity. +std::unique_ptr recv_op; + +void InitTensorsInScope(paddle::framework::Scope &scope, + paddle::platform::CPUPlace &place) { + paddle::platform::CPUDeviceContext ctx(place); + auto var = scope.Var("X"); + auto tensor = var->GetMutable(); + tensor->Resize({10, 10}); + float *expect = tensor->mutable_data(place); + for (int64_t i = 0; i < tensor->numel(); ++i) { + expect[i] = static_cast(i); + } + + auto out_var = scope.Var("Out"); + auto out_tensor = out_var->GetMutable(); + out_tensor->Resize({10, 10}); + tensor->mutable_data(place); // allocate +} + +void AddOp(const std::string &type, + const paddle::framework::VariableNameMap &inputs, + const paddle::framework::VariableNameMap &outputs, + paddle::framework::AttributeMap attrs, + paddle::framework::BlockDescBind *block) { + // insert output + for (auto kv : outputs) { + for (auto v : kv.second) { + auto var = block->Var(v); + var->SetDataType(paddle::framework::DataType::FP32); + } + } + + // insert op + auto op = block->AppendOp(); + op->SetType(type); + for (auto &kv : inputs) { + op->SetInput(kv.first, kv.second); + } + for (auto &kv : outputs) { + op->SetOutput(kv.first, kv.second); + } + op->SetAttrMap(attrs); +} + +void StartServerNet() { + paddle::framework::Scope scope; + paddle::platform::CPUPlace place; + InitTensorsInScope(scope, place); + + // sub program run in recv_op, for simple test we use sum + paddle::framework::ProgramDescBind program; + paddle::framework::BlockDescBind *block = program.MutableBlock(0); + // X for server side tensors, RX for received tensers, must be of same shape. + AddOp("sum", {{"X", {"X", "RX"}}}, {{"Out", {"Out"}}}, {}, block); + + paddle::framework::AttributeMap attrs; + attrs.insert({"endpoint", std::string("127.0.0.1:6174")}); + attrs.insert({"OptimizeBlock", block}); + recv_op = paddle::framework::OpRegistry::CreateOp("recv", {{"RX", {"RX"}}}, + {{"Out", {"Out"}}}, attrs); + paddle::platform::CPUDeviceContext ctx(place); + recv_op->Run(scope, ctx); +} + +TEST(SendRecvOp, CPU) { + std::thread server_thread(StartServerNet); + sleep(5); // wait server to start + // local net + paddle::framework::Scope scope; + paddle::platform::CPUPlace place; + InitTensorsInScope(scope, place); + + paddle::framework::AttributeMap attrs; + attrs.insert({"endpoint", std::string("127.0.0.1:6174")}); + + auto send_op = paddle::framework::OpRegistry::CreateOp( + "send", {{"X", {"X"}}}, {{"Out", {"Out"}}}, attrs); + paddle::platform::CPUDeviceContext ctx(place); + send_op->Run(scope, ctx); + + auto in_var = scope.Var("X"); + auto tensor = in_var->GetMutable(); + float *expected = tensor->data(); + + auto out_var = scope.Var("Out"); + auto target = out_var->GetMutable(); + // send fail cause output is none. + EXPECT_NE(target->memory_size(), size_t(0)); + float *actual = target->data(); + for (int64_t i = 0; i < target->numel(); ++i) { + EXPECT_EQ(expected[i] * 2, actual[i]); + } + recv_op.reset(); // dtor can shutdown and join server thread. + server_thread.join(); +} diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc index 72f4e4d5cbcd692423fa2a3e9ec8e7033b552c3c..5576d7b8be060a3c58cb18ed667041562cf853b8 100644 --- a/paddle/operators/sgd_op.cc +++ b/paddle/operators/sgd_op.cc @@ -55,7 +55,7 @@ SGD operator This operator implements one step of the stochastic gradient descent algorithm. -$$param_out = param - learning_rate * grad$$ +$$param\_out = param - learning\_rate * grad$$ )DOC"); } diff --git a/paddle/platform/cuda_profiler.h b/paddle/platform/cuda_profiler.h new file mode 100644 index 0000000000000000000000000000000000000000..b6311cb23d695c3cd851bcca120c24cced7fdd62 --- /dev/null +++ b/paddle/platform/cuda_profiler.h @@ -0,0 +1,53 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include +#include + +namespace paddle { +namespace platform { + +void CudaProfilerInit(std::string output_file, std::string output_mode, + std::vector config_flags) { + std::array buf; + std::string tmpl = "/tmp/cuda_profile_config.XXXXXX"; + PADDLE_ENFORCE_LT(tmpl.size(), buf.size()); + memcpy(buf.data(), tmpl.data(), tmpl.size()); + auto result = mktemp(buf.data()); + PADDLE_ENFORCE(strlen(result) != 0); + std::string config_file = result; + + { + std::ofstream ofs(config_file, std::ios::out | std::ios::trunc); + PADDLE_ENFORCE(ofs.is_open(), "ofstream: ", ofs.rdstate()); + for (const auto& line : config_flags) { + ofs << line << std::endl; + } + } + + PADDLE_ENFORCE(output_mode == "kvp" || output_mode == "csv"); + cudaOutputMode_t mode = output_mode == "csv" ? cudaCSV : cudaKeyValuePair; + PADDLE_ENFORCE( + cudaProfilerInitialize(config_file.c_str(), output_file.c_str(), mode)); +} + +void CudaProfilerStart() { PADDLE_ENFORCE(cudaProfilerStart()); } + +void CudaProfilerStop() { PADDLE_ENFORCE(cudaProfilerStop()); } + +} // namespace platform +} // namespace paddle diff --git a/paddle/platform/dynload/cudnn.cc b/paddle/platform/dynload/cudnn.cc index d3e4cb567d71b987724366b6a0896f5df0eb6055..761d9edd87f428ba140d29a566fc3401199bab15 100644 --- a/paddle/platform/dynload/cudnn.cc +++ b/paddle/platform/dynload/cudnn.cc @@ -37,6 +37,10 @@ CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DEFINE_WRAP); CUDNN_DNN_ROUTINE_EACH_R5(DEFINE_WRAP); #endif +#ifdef CUDNN_DNN_ROUTINE_EACH_R7 +CUDNN_DNN_ROUTINE_EACH_R7(DEFINE_WRAP); +#endif + } // namespace dynload } // namespace platform } // namespace paddle diff --git a/paddle/platform/dynload/cudnn.h b/paddle/platform/dynload/cudnn.h index b2d69da93bcd4a5c8e694a18ca648ddc4bd947af..61caac545014db2a09e2ada0b508419578c49740 100644 --- a/paddle/platform/dynload/cudnn.h +++ b/paddle/platform/dynload/cudnn.h @@ -135,6 +135,12 @@ CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP) CUDNN_DNN_ROUTINE_EACH_R5(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP) #endif +#if CUDNN_VERSION >= 7001 +#define CUDNN_DNN_ROUTINE_EACH_R7(__macro) \ + __macro(cudnnSetConvolutionGroupCount); +CUDNN_DNN_ROUTINE_EACH_R7(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP) +#endif + } // namespace dynload } // namespace platform } // namespace paddle diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc index f55a1edce31ccf2498dcfcf0b30ba1012d7a7d1a..c16d3e0cbe01f90a5aa9a5d7a523cd4e282e4771 100644 --- a/paddle/pybind/pybind.cc +++ b/paddle/pybind/pybind.cc @@ -37,6 +37,7 @@ limitations under the License. */ #ifdef PADDLE_WITH_CUDA #include "paddle/operators/nccl/nccl_gpu_common.h" +#include "paddle/platform/cuda_profiler.h" #include "paddle/platform/gpu_info.h" #endif @@ -460,6 +461,10 @@ All parameter, weight, gradient are variables in Paddle. m.def("op_support_gpu", OpSupportGPU); #ifdef PADDLE_WITH_CUDA m.def("get_cuda_device_count", platform::GetCUDADeviceCount); + + m.def("nvprof_init", platform::CudaProfilerInit); + m.def("nvprof_start", platform::CudaProfilerStart); + m.def("nvprof_stop", platform::CudaProfilerStop); #endif return m.ptr(); diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh index fda2a2f1b764106a7a108e8c56bc90ce3459e9b5..a2fdc5ce69bfdf0fadb808e4b51c8eef4ff7dfd6 100644 --- a/paddle/scripts/docker/build.sh +++ b/paddle/scripts/docker/build.sh @@ -16,11 +16,13 @@ function cmake_gen() { echo "using python abi: $1" if [ "$1" == "cp27-cp27m" ]; then export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs4/lib:} + export PATH=/opt/python/cp27-cp27m/bin/:${PATH} PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27m/bin/python -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27m/include/python2.7 -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs2/lib/libpython2.7.so" elif [ "$1" == "cp27-cp27mu" ]; then export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs2/lib:} + export PATH=/opt/python/cp27-cp27mu/bin/:${PATH} PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27mu/bin/python -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27mu/include/python2.7 -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs4/lib/libpython2.7.so" diff --git a/paddle/scripts/travis/build_doc.sh b/paddle/scripts/travis/build_doc.sh index 28d82343ed32273740d0c52d0451681e43b3675e..7d54f0254c8ea9367a34233602293db5b8593f9a 100755 --- a/paddle/scripts/travis/build_doc.sh +++ b/paddle/scripts/travis/build_doc.sh @@ -11,8 +11,9 @@ make -j `nproc` gen_proto_py make -j `nproc` paddle_docs paddle_docs_cn # check websites for broken links -linkchecker doc/en/html/index.html -linkchecker doc/cn/html/index.html +# It will be failed now! +#linkchecker doc/en/html/index.html +#linkchecker doc/cn/html/index.html # Parse Github URL REPO=`git config remote.origin.url` diff --git a/proto/ModelConfig.proto b/proto/ModelConfig.proto index e2f5592248fd0b6166c2d11af02cef7815673def..2fcdbbc8bd671f8ae911cf82c7a91091f252a82f 100644 --- a/proto/ModelConfig.proto +++ b/proto/ModelConfig.proto @@ -544,6 +544,9 @@ message LayerConfig { // for batch normalization layer // The small constant added to the variance to improve numeric stability. optional double epsilon = 60 [ default = 0.00001 ]; + + // for factorization machine layer + optional uint32 factor_size = 61; } message EvaluatorConfig { diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py index cfe2a34a1f34a9c828486a7a6dbe320f230bb986..5b173694dd0e4a52c0179f12f5edd74e2c41cb8c 100644 --- a/python/paddle/trainer/config_parser.py +++ b/python/paddle/trainer/config_parser.py @@ -2400,6 +2400,14 @@ class CropLayer(LayerBase): image_conf.img_size_y = input_layer.height image_conf.channels = input_layer.size / (input_layer.width * input_layer.height) + # only support for 4-dims inputs and NCHW order + if (len(self.config.inputs) == 2): + self.set_layer_height_width( + self.get_input_layer(1).height, self.get_input_layer(1).width) + self.set_layer_size(self.get_input_layer(1).size) + else: + self.set_layer_height_width(shape[-2], shape[-1]) + self.set_layer_size(reduce(lambda x, y: x * y, shape[1:])) @config_layer('batch_norm') @@ -3849,6 +3857,26 @@ class SwitchOrderLayer(LayerBase): name, 'switch_order', 0, inputs=inputs, **xargs) self.config.reshape_conf.height_axis.extend(reshape['height']) self.config.reshape_conf.width_axis.extend(reshape['width']) + input_layer = self.get_input_layer(0) + if reshape is None: + self.set_layer_size(input_layer.size) + else: + in_h = input_layer.height + in_w = input_layer.width + out_dims = None + if input_layer.has_depth(): + in_d = input_layer.depth + in_c = input_layer.size / in_h / in_w / in_d + # batch_size, depth, height, width, channel + out_dims = [0, in_d, in_h, in_w, in_c] + else: + in_c = input_layer.size / in_h / in_w + # batch_size, height, width, channel + out_dims = [0, in_h, in_w, in_c] + # Because (reshape['width'][0] > 0) always be true. + # So out_dims[0] won't be used. + size = reduce(lambda x, y: x * y, out_dims[reshape['width'][0]:]) + self.set_layer_size(size) @config_layer('scale_sub_region') @@ -3870,6 +3898,21 @@ class ScaleSubRegionLayer(LayerBase): image_conf.channels) +@config_layer('factorization_machine') +class FactorizationMachineLayer(LayerBase): + def __init__(self, name, inputs, factor_size, **xargs): + super(FactorizationMachineLayer, self).__init__( + name, 'factorization_machine', size=1, inputs=inputs, **xargs) + config_assert( + len(self.inputs) == 1, + 'factorization machine layer must have one and only one input.') + self.config.factor_size = factor_size + input_layer = self.get_input_layer(0) + psize = input_layer.size * factor_size + dims = [input_layer.size, factor_size] + self.create_input_parameter(0, psize, dims) + + # Deprecated, use a new layer specific class instead @config_func def Layer(name, type, **xargs): diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py index 469e667e80900b26578db6199e6426be8d0e5945..f6dc58b9c0ed0b14ad9db098892af14274aed0c1 100644 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -148,6 +148,7 @@ __all__ = [ 'resize_layer', 'sub_seq_layer', 'scale_sub_region_layer', + 'factorization_machine', ] @@ -264,6 +265,8 @@ class LayerType(object): SCALE_SUB_REGION_LAYER = 'scale_sub_region' + FACTORIZATION_MACHINE = 'factorization_machine' + @staticmethod def is_layer_type(type_name): """ @@ -2985,8 +2988,8 @@ def spp_layer(input, A layer performs spatial pyramid pooling. Reference: - Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition - https://arxiv.org/abs/1406.4729 + `Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition + https://arxiv.org/abs/1406.4729`_ The example usage is: @@ -3087,8 +3090,8 @@ def img_cmrnorm_layer(input, Response normalization across feature maps. Reference: - ImageNet Classification with Deep Convolutional Neural Networks - http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf + `ImageNet Classification with Deep Convolutional Neural Networks + http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf`_ The example usage is: @@ -3154,9 +3157,9 @@ def batch_norm_layer(input, y_i &\\gets \\gamma \\hat{x_i} + \\beta \\qquad &//\ scale\ and\ shift Reference: - Batch Normalization: Accelerating Deep Network Training by Reducing + `Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift - http://arxiv.org/abs/1502.03167 + http://arxiv.org/abs/1502.03167`_ The example usage is: @@ -5413,10 +5416,10 @@ def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None): to be devided by groups. Reference: - Maxout Networks - http://www.jmlr.org/proceedings/papers/v28/goodfellow13.pdf - Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks - https://arxiv.org/pdf/1312.6082v4.pdf + `Maxout Networks + http://www.jmlr.org/proceedings/papers/v28/goodfellow13.pdf`_ + `Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks + https://arxiv.org/pdf/1312.6082v4.pdf`_ .. math:: y_{si+j} = \max_k x_{gsi + sk + j} @@ -5481,9 +5484,9 @@ def ctc_layer(input, alignment between the inputs and the target labels is unknown. Reference: - Connectionist Temporal Classification: Labelling Unsegmented Sequence Data + `Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks - http://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf + http://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf`_ Note: Considering the 'blank' label needed by CTC, you need to use (num_classes + 1) @@ -5555,9 +5558,9 @@ def warp_ctc_layer(input, install it to :code:`third_party/install/warpctc` directory. Reference: - Connectionist Temporal Classification: Labelling Unsegmented Sequence Data + `Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks - http://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf + http://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf`_ Note: - Let num_classes represents the category number. Considering the 'blank' @@ -5777,8 +5780,8 @@ def nce_layer(input, Noise-contrastive estimation. Reference: - A fast and simple algorithm for training neural probabilistic language - models. https://www.cs.toronto.edu/~amnih/papers/ncelm.pdf + `A fast and simple algorithm for training neural probabilistic language + models. https://www.cs.toronto.edu/~amnih/papers/ncelm.pdf`_ The example usage is: @@ -5893,8 +5896,8 @@ def rank_cost(left, A cost Layer for learning to rank using gradient descent. Reference: - Learning to Rank using Gradient Descent - http://research.microsoft.com/en-us/um/people/cburges/papers/ICML_ranking.pdf + `Learning to Rank using Gradient Descent + http://research.microsoft.com/en-us/um/people/cburges/papers/ICML_ranking.pdf`_ .. math:: @@ -6429,8 +6432,8 @@ def smooth_l1_cost(input, label, name=None, coeff=1.0, layer_attr=None): smooth_{L1}(x) = \\begin{cases} 0.5x^2& \\text{if} \\ |x| < 1 \\\\ |x|-0.5& \\text{otherwise} \end{cases} Reference: - Fast R-CNN - https://arxiv.org/pdf/1504.08083v2.pdf + `Fast R-CNN + https://arxiv.org/pdf/1504.08083v2.pdf`_ The example usage is: @@ -6636,8 +6639,8 @@ def prelu_layer(input, The Parametric Relu activation that actives outputs with a learnable weight. Reference: - Delving Deep into Rectifiers: Surpassing Human-Level Performance on - ImageNet Classification http://arxiv.org/pdf/1502.01852v1.pdf + `Delving Deep into Rectifiers: Surpassing Human-Level Performance on + ImageNet Classification http://arxiv.org/pdf/1502.01852v1.pdf`_ .. math:: z_i &\\quad if \\quad z_i > 0 \\\\ @@ -6733,8 +6736,8 @@ def gated_unit_layer(input, product between :match:`X'` and :math:`\sigma` is finally returned. Reference: - Language Modeling with Gated Convolutional Networks - https://arxiv.org/abs/1612.08083 + `Language Modeling with Gated Convolutional Networks + https://arxiv.org/abs/1612.08083`_ .. math:: y=\\text{act}(X \cdot W + b)\otimes \sigma(X \cdot V + c) @@ -6870,6 +6873,7 @@ def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None): :param input: The input of this layer. If two inputs are given, the second one will be regarded as the reference. + And the input must be 4-dims and in NCHW order. :type input: LayerOutput | Sequence :param offset: The crop offset. :type offset: Sequence @@ -7403,3 +7407,73 @@ def scale_sub_region_layer(input, indices, value, name=None): parents=[input, indices], num_filters=input.num_filters, size=input.size) + + +@wrap_name_default() +@wrap_act_default(act=LinearActivation()) +@wrap_param_attr_default() +@layer_support() +def factorization_machine(input, + factor_size, + act=None, + name=None, + param_attr=None, + layer_attr=None): + """ + The Factorization Machine models pairwise feature interactions as inner + product of the learned latent vectors corresponding to each input feature. + The Factorization Machine can effectively capture feature interactions + especially when the input is sparse. + + This implementation only consider the 2-order feature interactions using + Factorization Machine with the formula: + + .. math:: + y = \sum_{i=1}^{n-1}\sum_{j=i+1}^n\langle v_i, v_j \rangle x_i x_j + + Note: + X is the input vector with size n. V is the factor matrix. Each row of V + is the latent vector corresponding to each input dimesion. The size of + each latent vector is k. + + For details of Factorization Machine, please refer to the paper: + Factorization machines. + + .. code-block:: python + first_order = paddle.layer.fc(input=input, + size=1, + act=paddle.activation.Linear()) + second_order = paddle.layer.factorization_machine(input=input, + factor_size=10) + fm = paddle.layer.addto(input=[first_order, second_order], + act=paddle.activation.Linear(), + bias_attr=False) + + :param input: The input layer. Supported input types: all input data types + on CPU, and only dense input types on GPU. + :type input: LayerOutput + :param factor_size: The hyperparameter that defines the dimensionality of + the latent vector size. + :type context_len: int + :param act: Activation Type. Default is linear activation. + :type act: BaseActivation + :param param_attr: The parameter attribute. See ParameterAttribute for + details. + :type param_attr: ParameterAttribute + :param layer_attr: Extra Layer config. + :type layer_attr: ExtraLayerAttribute|None + :return: LayerOutput object. + :rtype: LayerOutput + """ + assert isinstance(input, LayerOutput) + assert factor_size > 0, "the factor_size must be greater than 0." + + Layer( + inputs=[Input(input.name, **param_attr.attr)], + name=name, + factor_size=factor_size, + type=LayerType.FACTORIZATION_MACHINE, + active_type=act.name, + **ExtraLayerAttribute.to_kwargs(layer_attr)) + return LayerOutput( + name, LayerType.FACTORIZATION_MACHINE, input, activation=act, size=1) diff --git a/python/paddle/trainer_config_helpers/tests/configs/file_list.sh b/python/paddle/trainer_config_helpers/tests/configs/file_list.sh index a21f67a2d99e7eab39708e2a571d30d7e9f20ce6..10c941f707498ec45e79bed9d3f8054eea19887d 100755 --- a/python/paddle/trainer_config_helpers/tests/configs/file_list.sh +++ b/python/paddle/trainer_config_helpers/tests/configs/file_list.sh @@ -11,6 +11,7 @@ test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_l test_kmax_seq_socre_layer test_sub_nested_seq_select_layer test_scale_shift_layer test_seq_slice_layer test_cross_entropy_over_beam test_roi_pool_layer test_pooling3D_layer test_conv3d_layer test_deconv3d_layer test_BatchNorm3D test_resize_layer -test_scale_sub_region_layer test_dot_prod_layer test_l2_distance_layer) +test_scale_sub_region_layer test_dot_prod_layer test_l2_distance_layer +test_factorization_machine) export whole_configs=(test_split_datasource) diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_factorization_machine.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_factorization_machine.protostr new file mode 100644 index 0000000000000000000000000000000000000000..4f3002b19942ed58970bfd64e5978c1601273992 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_factorization_machine.protostr @@ -0,0 +1,39 @@ +type: "nn" +layers { + name: "data" + type: "data" + size: 1024 + active_type: "" +} +layers { + name: "__factorization_machine_0__" + type: "factorization_machine" + size: 1 + active_type: "" + inputs { + input_layer_name: "data" + input_parameter_name: "___factorization_machine_0__.w0" + } + factor_size: 10 +} +parameters { + name: "___factorization_machine_0__.w0" + size: 10240 + initial_mean: 0.0 + initial_std: 0.03125 + dims: 1024 + dims: 10 + initial_strategy: 0 + initial_smart: true +} +input_layer_names: "data" +output_layer_names: "__factorization_machine_0__" +sub_models { + name: "root" + layer_names: "data" + layer_names: "__factorization_machine_0__" + input_layer_names: "data" + output_layer_names: "__factorization_machine_0__" + is_recurrent_layer_group: false +} + diff --git a/python/paddle/trainer_config_helpers/tests/configs/test_factorization_machine.py b/python/paddle/trainer_config_helpers/tests/configs/test_factorization_machine.py new file mode 100644 index 0000000000000000000000000000000000000000..b249de0fee3c8ca4ad0520872fa2497c493d31b5 --- /dev/null +++ b/python/paddle/trainer_config_helpers/tests/configs/test_factorization_machine.py @@ -0,0 +1,7 @@ +from paddle.trainer_config_helpers import * + +data = data_layer(name='data', size=1024) + +fm = factorization_machine(input=data, factor_size=10) + +outputs(fm) diff --git a/python/paddle/v2/__init__.py b/python/paddle/v2/__init__.py index 33a0829ba8d635ebd68b50f3da07da958fb79dcb..70f61e84997efdbe3d6f268d249be8bac15b9ecd 100644 --- a/python/paddle/v2/__init__.py +++ b/python/paddle/v2/__init__.py @@ -83,11 +83,10 @@ def set_omp_mkl_env_vars(trainer_count): '''Get the number of physical cores''' if platform.system() == "Linux": num_sockets = int( - os.popen("lscpu |grep \"Socket\" |awk -F':' '{print $2}'|xargs") + os.popen("grep 'physical id' /proc/cpuinfo | sort -u | wc -l") .read()) num_cores_per_socket = int( - os.popen( - "lscpu |grep \"per socket\" |awk -F':' '{print $2}'|xargs") + os.popen("grep 'core id' /proc/cpuinfo | sort -u | wc -l") .read()) return num_sockets * num_cores_per_socket else: diff --git a/python/paddle/v2/fluid/evaluator.py b/python/paddle/v2/fluid/evaluator.py index bd4a6fda1fd20e68d5a42e76f6ab516bb5c00cff..137c5736226b689340748d5098ca51659d5acff8 100644 --- a/python/paddle/v2/fluid/evaluator.py +++ b/python/paddle/v2/fluid/evaluator.py @@ -26,9 +26,9 @@ class Evaluator(object): name(str): The name of evaluator. such as, "accuracy". Used for generate temporary variable name. main_program(Program, optional): The evaluator should be added to this - main_program. Default g_main_program + main_program. Default default_main_program() startup_program(Program, optional):The parameter should be added to this - startup_program. Default g_startup_program + startup_program. Default default_startup_program() Attributes: states(list): The list of state variables. states will be reset to zero diff --git a/python/paddle/v2/fluid/executor.py b/python/paddle/v2/fluid/executor.py index 3e26d1b983a3c924ce2392c266bcd32e27c7b309..bdc82eede9d93a7cf904999a6b869ce2d23c90dc 100644 --- a/python/paddle/v2/fluid/executor.py +++ b/python/paddle/v2/fluid/executor.py @@ -1,6 +1,6 @@ import numpy as np from . import core -from framework import Program, g_main_program +from framework import Program, default_main_program __all__ = ['Executor', 'g_scope'] @@ -103,7 +103,7 @@ class Executor(object): fetch_list = [] if program is None: - program = g_main_program + program = default_main_program() if not isinstance(program, Program): raise TypeError() diff --git a/python/paddle/v2/fluid/framework.py b/python/paddle/v2/fluid/framework.py index 6d6ea23f55eebc57cb120582a7c82d77eb1df45c..1c42e4d44f5046e0db171fdaeb8e7af38a2cae07 100644 --- a/python/paddle/v2/fluid/framework.py +++ b/python/paddle/v2/fluid/framework.py @@ -6,7 +6,7 @@ import proto.framework_pb2 as framework_pb2 __all__ = [ 'Block', 'Variable', 'Program', 'Operator', 'default_startup_program', - 'default_main_program', 'g_startup_program', 'g_main_program' + 'default_main_program' ] @@ -654,13 +654,13 @@ class Parameter(Variable): # program is a global instance. -g_main_program = Program() -g_startup_program = Program() +_main_program_ = Program() +_startup_program_ = Program() def default_startup_program(): - return g_startup_program + return _startup_program_ def default_main_program(): - return g_main_program + return _main_program_ diff --git a/python/paddle/v2/fluid/io.py b/python/paddle/v2/fluid/io.py index e5b2aa3b919df4cec1091c0bbd39b7e400cc6867..e147ac22ad289eb00c83def66974d875fcdc31f8 100644 --- a/python/paddle/v2/fluid/io.py +++ b/python/paddle/v2/fluid/io.py @@ -1,8 +1,7 @@ import os import cPickle as pickle -from paddle.v2.fluid.framework import Program, Parameter, g_main_program, \ - Variable +from paddle.v2.fluid.framework import Program, Parameter, default_main_program, Variable __all__ = [ 'save_vars', 'save_params', 'save_persistables', 'load_vars', 'load_params', @@ -46,7 +45,7 @@ def save_vars(executor, dirname, main_program=None, vars=None, predicate=None): """ if vars is None: if main_program is None: - main_program = g_main_program + main_program = default_main_program() if not isinstance(main_program, Program): raise TypeError("program should be as Program type or None") @@ -98,7 +97,7 @@ def load_vars(executor, dirname, main_program=None, vars=None, predicate=None): :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 g_program. + 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 variables will be loaded. :param vars: variables need to be loaded. If specify vars, program & @@ -107,7 +106,7 @@ def load_vars(executor, dirname, main_program=None, vars=None, predicate=None): """ if vars is None: if main_program is None: - main_program = g_main_program + main_program = default_main_program() if not isinstance(main_program, Program): raise TypeError("program's type should be Program") @@ -154,7 +153,7 @@ def load_persistables(executor, dirname, main_program=None): def get_inference_program(target_vars, main_program=None): if main_program is None: - main_program = g_main_program + main_program = default_main_program() if not isinstance(target_vars, list): target_vars = [target_vars] @@ -177,12 +176,12 @@ def save_inference_model(dirname, :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 g_main_program. + Default default_main_program(). :return: None """ if main_program is None: - main_program = g_main_program + main_program = default_main_program() if not isinstance(target_vars, list): target_vars = [target_vars] @@ -272,10 +271,10 @@ def get_parameter_value_by_name(name, executor, program=None): :param executor: executor for retrieving the value :param name: the name of the parameter :param program: the program where the variable is found - Default g_main_program. + Default default_main_program(). :return: the LoDTensor for the variable """ if program is None: - program = g_main_program + program = default_main_program() var = program.global_block().var(name) return get_parameter_value(var, executor) diff --git a/python/paddle/v2/fluid/layer_helper.py b/python/paddle/v2/fluid/layer_helper.py index 5f8855551114a9a9b671d1630c9e8a3f0cb5c04b..7762b0d88f3a62c3b919d64a6565bfc3951c4e25 100644 --- a/python/paddle/v2/fluid/layer_helper.py +++ b/python/paddle/v2/fluid/layer_helper.py @@ -1,8 +1,7 @@ import copy import itertools -from framework import Variable, g_main_program, \ - g_startup_program, unique_name, dtype_is_floating +from framework import Variable, default_main_program, default_startup_program, unique_name, dtype_is_floating from paddle.v2.fluid.initializer import Constant, Xavier @@ -22,7 +21,7 @@ class LayerHelper(object): def main_program(self): prog = self.kwargs.get('main_program', None) if prog is None: - return g_main_program + return default_main_program() else: return prog @@ -30,7 +29,7 @@ class LayerHelper(object): def startup_program(self): prog = self.kwargs.get('startup_program', None) if prog is None: - return g_startup_program + return default_startup_program() else: return prog diff --git a/python/paddle/v2/fluid/layers.py b/python/paddle/v2/fluid/layers.py index 28bc3d214b559a089efb2bb736eb49cb1ba4de25..6adfac3a32c7c8ecf035ba3a8f757a0efb6f9b68 100644 --- a/python/paddle/v2/fluid/layers.py +++ b/python/paddle/v2/fluid/layers.py @@ -1,7 +1,7 @@ -from . import core +import core import proto.framework_pb2 as framework_pb2 from framework import OpProtoHolder, Variable, Program, Operator -from initializer import Constant, Normal, Xavier +from initializer import Constant, Normal, Xavier, Initializer from paddle.v2.fluid.layer_helper import LayerHelper, unique_name import re import cStringIO @@ -1587,6 +1587,97 @@ def array_length(array, main_program=None): return tmp +def conv2d_transpose(input, + num_filters, + output_size=None, + filter_size=None, + padding=None, + stride=None, + param_attr=None, + param_initializer=None, + main_program=None, + startup_program=None): + """ + The transpose of conv2d layer. + + This layer is also known as deconvolution layer. + + 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 + image channel. + output_size(int|tuple|None): The output image size. If output size is a + tuple, it must contain two integers, (image_H, image_W). This + parameter only works when filter_size is None. + filter_size(int|tuple|None): The filter size. If filter_size is a tuple, + it must contain two integers, (filter_size_H, filter_size_W). + Otherwise, the filter will be a square. None if use output size to + calculate filter_size + padding(int|tuple): The padding size. If padding is a tuple, it must + contain two integers, (padding_H, padding_W). Otherwise, the + padding_H = padding_W = padding. + stride(int|tuple): The stride size. If stride is a tuple, it must + contain two integers, (stride_H, stride_W). Otherwise, the + stride_H = stride_W = stride. + param_attr: Parameter Attribute. + param_initializer(Initializer): Parameter Initializer. Default is Xavier + main_program(Program): the main program + startup_program(Program): the startup program + + Returns: + Variable: Output image. + """ + helper = LayerHelper("conv2d_transpose", **locals()) + if not isinstance(input, Variable): + raise TypeError("Input of conv2d_transpose must be Variable") + input_channel = input.shape[1] + + op_attr = dict() + + if isinstance(padding, int): + op_attr['paddings'] = [padding, padding] + elif padding is not None: + op_attr['paddings'] = padding + + if isinstance(stride, int): + op_attr['strides'] = stride + elif stride is not None: + op_attr['strides'] = stride + + if filter_size is None: + if output_size is None: + raise ValueError("output_size must be set when filter_size is None") + if isinstance(output_size, int): + output_size = [output_size, output_size] + + padding = op_attr.get('paddings', [0, 0]) + stride = op_attr.get('strides', [1, 1]) + + h_in = input.shape[2] + w_in = input.shape[3] + filter_size_h = output_size[0] - (h_in - 1) * stride[0] + 2 * padding[0] + filter_size_w = output_size[1] - (w_in - 1) * stride[1] + 2 * padding[1] + filter_size = [filter_size_h, filter_size_w] + elif isinstance(filter_size, int): + filter_size = [filter_size, filter_size] + + filter_shape = [input_channel, num_filters] + filter_size + img_filter = helper.create_parameter( + dtype=input.dtype, + shape=filter_shape, + attr=helper.param_attr, + initializer=param_initializer) + + out = helper.create_tmp_variable(dtype=input.dtype) + helper.append_op( + type='conv2d_transpose', + inputs={'Input': [input], + 'Filter': [img_filter]}, + outputs={'Output': out}, + attrs=op_attr) + return out + + class ConditionalBlockGuard(BlockGuard): def __init__(self, block): if not isinstance(block, ConditionalBlock): diff --git a/python/paddle/v2/fluid/profiler.py b/python/paddle/v2/fluid/profiler.py new file mode 100644 index 0000000000000000000000000000000000000000..2dbba72c644d1b9d06e65e19cc7e581b9a668cca --- /dev/null +++ b/python/paddle/v2/fluid/profiler.py @@ -0,0 +1,46 @@ +import paddle.v2.fluid.core as core +from contextlib import contextmanager + +__all__ = ['CudaProfiler'] + +NVPROF_CONFIG = [ + "gpustarttimestamp", + "gpuendtimestamp", + "gridsize3d", + "threadblocksize", + "streamid", + "enableonstart 0", + "conckerneltrace", +] + + +@contextmanager +def cuda_profiler(output_file, output_mode=None, config=None): + """The CUDA profiler. + This fuctions is used to profile CUDA program by CUDA runtime application + programming interface. The profiling result will be written into + `output_file` with Key-Value pair format or Comma separated values format. + The user can set the output mode by `output_mode` argument and set the + counters/options for profiling by `config` argument. The default config + caontains 'gpustarttimestamp', 'gpustarttimestamp', 'gridsize3d', + 'threadblocksize', 'streamid', 'enableonstart 0', 'conckerneltrace'. + + Args: + output_file (string) : The output file name, the result will be + written into this file. + output_mode (string) : The output mode has Key-Value pair format and + Comma separated values format. It should be 'kv' or 'csv'. + config (string) : The profiler options and counters can refer to + "Compute Command Line Profiler User Guide". + """ + if output_mode is None: + output_mode = 'csv' + if output_mode not in ['kv', 'csv']: + raise ValueError("The output mode must be 'key-value' or 'csv'.") + config = NVPROF_CONFIG if config is None else config + core.nvprof_init(output_file, output_mode, config) + # Enables profiler collection by the active CUDA profiling tool. + core.nvprof_start() + yield + # Disables profiler collection. + core.nvprof_stop() diff --git a/python/paddle/v2/fluid/tests/book/test_image_classification_train.py b/python/paddle/v2/fluid/tests/book/test_image_classification_train.py index cc45b10b90868858c61334a3a43acf65c3d4eaf5..0f0cc5b5406ef51ac3504a95ea716056ae8730af 100644 --- a/python/paddle/v2/fluid/tests/book/test_image_classification_train.py +++ b/python/paddle/v2/fluid/tests/book/test_image_classification_train.py @@ -69,8 +69,7 @@ def vgg16_bn_drop(input): drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5) fc1 = fluid.layers.fc(input=drop, size=512, act=None) - reshape1 = fluid.layers.reshape(x=fc1, shape=list(fc1.shape + (1, 1))) - bn = fluid.layers.batch_norm(input=reshape1, act='relu') + bn = fluid.layers.batch_norm(input=fc1, act='relu') drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5) fc2 = fluid.layers.fc(input=drop2, size=512, act=None) return fc2 diff --git a/python/paddle/v2/fluid/tests/test_array_read_write_op.py b/python/paddle/v2/fluid/tests/test_array_read_write_op.py index b7790b01062d480cbd6c9e1a626d318385b4f61e..f6120aedecf1015c279b8f218f5e37f2e598ab91 100644 --- a/python/paddle/v2/fluid/tests/test_array_read_write_op.py +++ b/python/paddle/v2/fluid/tests/test_array_read_write_op.py @@ -3,7 +3,7 @@ import paddle.v2.fluid.core as core import paddle.v2.fluid.layers as layers from paddle.v2.fluid.executor import Executor from paddle.v2.fluid.backward import append_backward_ops -from paddle.v2.fluid.framework import g_main_program +from paddle.v2.fluid.framework import default_main_program import numpy @@ -66,7 +66,7 @@ class TestArrayReadWrite(unittest.TestCase): append_backward_ops(total_sum_scaled) - g_vars = map(g_main_program.global_block().var, + g_vars = map(default_main_program().global_block().var, [each_x.name + "@GRAD" for each_x in x]) g_out = [ item.sum() diff --git a/python/paddle/v2/fluid/tests/test_batch_norm_op.py b/python/paddle/v2/fluid/tests/test_batch_norm_op.py index 71f9599e0de83c86808f7e62547f80d3d50ffc7d..e766a68c0e338b07e47260e40edc544c98555382 100644 --- a/python/paddle/v2/fluid/tests/test_batch_norm_op.py +++ b/python/paddle/v2/fluid/tests/test_batch_norm_op.py @@ -21,6 +21,13 @@ def get_backward_op(scope, op, no_grad_set): def _reference_training(x, scale, offset, epsilon, data_format): + x_shape = x.shape + if len(x_shape) == 2: + if data_format == "NCHW": + x = np.reshape(x, (x.shape[0], x.shape[1], 1, 1)) + else: + x = np.reshape(x, (x.shape[0], 1, 1, x.shape[1])) + if data_format == "NCHW": n, c, h, w = x.shape x_square = x * x @@ -39,6 +46,8 @@ def _reference_training(x, scale, offset, epsilon, data_format): offset_tile = np.reshape(offset, (1, c, 1, 1)) offset_tile = np.reshape(offset_tile, (1, c, 1, 1)) y = normalized * scale_tile + offset_tile + if len(x_shape) == 2: + y = np.reshape(y, (y.shape[0], y.shape[1])) return y, mean, var elif data_format == "NHWC": x_square = x * x @@ -48,7 +57,10 @@ def _reference_training(x, scale, offset, epsilon, data_format): mean = x_sum / element_count var = x_square_sum / element_count - mean * mean normalized = (x - mean) / np.sqrt(var + epsilon) - return (normalized * scale + offset), mean, var + y = normalized * scale + offset + if len(x_shape) == 2: + y = np.reshape(y, x_shape) + return y, mean, var else: raise ValueError("Unknown data order.") @@ -65,6 +77,18 @@ def _reference_grad(x, grad_y, scale, mean, var, epsilon, data_format): # (x - mean) * sum(grad_y * (x - mean)) / (var + epsilon)) # transfer from (N, C, H, W) to (N, H, W, C) to simplify computation + x_shape = x.shape + + if len(x_shape) == 2: + if data_format == "NCHW": + x = np.reshape(x, (x.shape[0], x.shape[1], 1, 1)) + grad_y = np.reshape(grad_y, + (grad_y.shape[0], grad_y.shape[1], 1, 1)) + else: + x = np.reshape(x, (x.shape[0], 1, 1, x.shape[1])) + grad_y = np.reshape(grad_y, + (grad_y.shape[0], 1, 1, grad_y.shape[1])) + if data_format == "NCHW": x = np.transpose(x, (0, 2, 3, 1)) grad_y = np.transpose(grad_y, (0, 2, 3, 1)) @@ -83,6 +107,9 @@ def _reference_grad(x, grad_y, scale, mean, var, epsilon, data_format): grad_x = np.transpose(grad_x, (0, 3, 1, 2)) x = np.transpose(x, (0, 3, 1, 2)) grad_y = np.transpose(grad_y, (0, 3, 1, 2)) + + if len(x_shape) == 2: + grad_x = np.reshape(grad_x, x_shape) return grad_x, grad_scale, grad_offset @@ -127,7 +154,7 @@ class TestBatchNormOp(OpTest): momentum = 0.9 # N, H, W, C: 2, 3, 4, 2 - n, h, w, c = 2, 3, 4, 2 + n, h, w, c = 2, 3, 4, 5 x_shape = [n, h, w, c] scale_shape = [c] @@ -184,20 +211,23 @@ class TestBatchNormOp(OpTest): print 'python: NHWC, NCHW, backward checking passed' def test_forward_backward(self): - def test_with_place(place, tensor_format): + def test_with_place(place, tensor_format, shape): # attr epsilon = 0.00001 momentum = 0.9 - # N, H, W, C: 12, 3, 4, 2 - n, h, w, c = 2, 3, 4, 2 - - if data_format == "NHWC": - x_shape = [n, h, w, c] - elif data_format == "NCHW": - x_shape = [n, c, h, w] + if len(shape) == 2: + x_shape = shape + c = shape[1] else: - raise ValueError("Unknown data type.") + # n, h, w, c = 2, 3, 4, 2 + n, h, w, c = shape[0], shape[1], shape[2], shape[3] + if data_format == "NHWC": + x_shape = [n, h, w, c] + elif data_format == "NCHW": + x_shape = [n, c, h, w] + else: + raise ValueError("Unknown data type.") scale_shape = [c] x_val = np.random.random_sample(x_shape).astype(np.float32) @@ -219,7 +249,10 @@ class TestBatchNormOp(OpTest): # for gradient test # y_grad = np.ones(x_shape).astype(np.float32) y_grad = np.zeros(x_shape).astype(np.float32) - y_grad[0, 0, 0, 0] = 1. + if len(y_grad.shape) == 2: + y_grad[0, 0] = 1. + else: + y_grad[0, 0, 0, 0] = 1. # y_grad = np.random.random_sample(x_shape).astype(np.float32) x_grad_ref, scale_grad_ref, bias_grad_ref = _reference_grad( x_val, y_grad, scale_val, saved_mean, var_ref, epsilon, @@ -313,7 +346,8 @@ class TestBatchNormOp(OpTest): places.append(core.GPUPlace(0)) for place in places: for data_format in ["NCHW", "NHWC"]: - test_with_place(place, data_format) + test_with_place(place, data_format, [2, 3, 4, 5]) + test_with_place(place, data_format, [2, 3]) if __name__ == '__main__': diff --git a/python/paddle/v2/fluid/tests/test_conditional_block.py b/python/paddle/v2/fluid/tests/test_conditional_block.py index d953ee7ddc37d150d87cbd680379410a4d16f6b1..2b9d8f351a2836cd723d629d4790de1e068d0ea3 100644 --- a/python/paddle/v2/fluid/tests/test_conditional_block.py +++ b/python/paddle/v2/fluid/tests/test_conditional_block.py @@ -1,7 +1,7 @@ import unittest import paddle.v2.fluid.layers as layers import paddle.v2.fluid.core as core -from paddle.v2.fluid.framework import g_startup_program, g_main_program +from paddle.v2.fluid.framework import default_startup_program, default_main_program from paddle.v2.fluid.executor import Executor from paddle.v2.fluid.backward import append_backward_ops import numpy @@ -19,7 +19,7 @@ class ConditionalBlock(unittest.TestCase): cpu = core.CPUPlace() exe = Executor(cpu) - exe.run(g_startup_program) + exe.run(default_startup_program()) x = numpy.random.random(size=(10, 1)).astype('float32') @@ -29,7 +29,9 @@ class ConditionalBlock(unittest.TestCase): append_backward_ops(loss=loss) outs = exe.run( feed={'X': x}, - fetch_list=[g_main_program.block(0).var(data.name + "@GRAD")])[0] + fetch_list=[ + default_main_program().block(0).var(data.name + "@GRAD") + ])[0] print outs diff --git a/python/paddle/v2/fluid/tests/test_executor_and_mul.py b/python/paddle/v2/fluid/tests/test_executor_and_mul.py index 558273e30dff7fb74f78751f4fe569f79a453d0d..b1ef87c5cb1711c419b401c5950839816f7f4160 100644 --- a/python/paddle/v2/fluid/tests/test_executor_and_mul.py +++ b/python/paddle/v2/fluid/tests/test_executor_and_mul.py @@ -1,9 +1,10 @@ import unittest -from paddle.v2.fluid.layers import mul, data, sequence_pool + +import numpy import paddle.v2.fluid.core as core + from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.framework import g_main_program -import numpy +from paddle.v2.fluid.layers import mul, data class TestExecutor(unittest.TestCase): @@ -19,10 +20,7 @@ class TestExecutor(unittest.TestCase): a_np = numpy.random.random((100, 784)).astype('float32') b_np = numpy.random.random((784, 100)).astype('float32') exe = Executor(place) - outs = exe.run(g_main_program, - feed={'a': a_np, - 'b': b_np}, - fetch_list=[out]) + outs = exe.run(feed={'a': a_np, 'b': b_np}, fetch_list=[out]) out = outs[0] self.assertEqual((100, 100), out.shape) self.assertTrue(numpy.allclose(out, numpy.dot(a_np, b_np))) diff --git a/python/paddle/v2/fluid/tests/test_image_classification_layer.py b/python/paddle/v2/fluid/tests/test_image_classification_layer.py index 8e8e1b0a8c07a60cb1404462f976d10fe26e87f6..2fd609d4474e97ecd96adcd146f2f550e0772740 100644 --- a/python/paddle/v2/fluid/tests/test_image_classification_layer.py +++ b/python/paddle/v2/fluid/tests/test_image_classification_layer.py @@ -1,6 +1,6 @@ import unittest -import paddle.v2.fluid.layers as layers +import paddle.v2.fluid as fluid import paddle.v2.fluid.nets as nets from paddle.v2.fluid.framework import Program @@ -29,27 +29,35 @@ class TestLayer(unittest.TestCase): def test_batch_norm_layer(self): main_program = Program() startup_program = Program() - images = layers.data( + images = fluid.layers.data( name='pixel', shape=[3, 48, 48], dtype='float32', main_program=main_program) - layers.batch_norm( + hidden1 = fluid.layers.batch_norm( input=images, main_program=main_program, startup_program=startup_program) + hidden2 = fluid.layers.fc(input=hidden1, + size=128, + act='relu', + main_program=main_program) + hidden3 = fluid.layers.batch_norm( + input=hidden2, + main_program=main_program, + startup_program=startup_program) - # print str(main_program) + print str(main_program) def test_dropout_layer(self): main_program = Program() startup_program = Program() - images = layers.data( + images = fluid.layers.data( name='pixel', shape=[3, 48, 48], dtype='float32', main_program=main_program) - layers.dropout( + fluid.layers.dropout( x=images, dropout_prob=0.5, main_program=main_program, @@ -61,7 +69,7 @@ class TestLayer(unittest.TestCase): main_program = Program() startup_program = Program() - images = layers.data( + images = fluid.layers.data( name='pixel', shape=[3, 48, 48], dtype='float32', @@ -77,19 +85,19 @@ class TestLayer(unittest.TestCase): def test_elementwise_add_with_act(self): main_program = Program() startup_program = Program() - image1 = layers.data( + image1 = fluid.layers.data( name='pixel1', shape=[3, 48, 48], dtype='float32', main_program=main_program, startup_program=startup_program) - image2 = layers.data( + image2 = fluid.layers.data( name='pixel2', shape=[3, 48, 48], dtype='float32', main_program=main_program, startup_program=startup_program) - out = layers.elementwise_add( + out = fluid.layers.elementwise_add( x=image1, y=image2, act='relu', diff --git a/python/paddle/v2/fluid/tests/test_layers.py b/python/paddle/v2/fluid/tests/test_layers.py index 87dc6d1a6270e0f8425b56601d04049450c73380..62b2a0f9a11aa20e170fd108083abe04caedc4f3 100644 --- a/python/paddle/v2/fluid/tests/test_layers.py +++ b/python/paddle/v2/fluid/tests/test_layers.py @@ -65,6 +65,15 @@ class TestBook(unittest.TestCase): print str(program) + def test_conv2d_transpose(self): + program = Program() + kwargs = {'main_program': program} + img = layers.data( + name='pixel', shape=[3, 2, 2], dtype='float32', **kwargs) + layers.conv2d_transpose( + input=img, num_filters=10, output_size=28, **kwargs) + print str(program) + def test_recognize_digits_conv(self): program = Program() diff --git a/python/paddle/v2/fluid/tests/test_lod_rank_table.py b/python/paddle/v2/fluid/tests/test_lod_rank_table.py index bbc11930b9e804c2769cc590c298c6e90dc36ca6..30d619fe318517345195281b17f88e9916b6afb3 100644 --- a/python/paddle/v2/fluid/tests/test_lod_rank_table.py +++ b/python/paddle/v2/fluid/tests/test_lod_rank_table.py @@ -1,6 +1,5 @@ from paddle.v2.fluid.layers import lod_rank_table, data from paddle.v2.fluid.executor import Executor -from paddle.v2.fluid.framework import g_main_program import paddle.v2.fluid.core as core import numpy import unittest @@ -18,7 +17,7 @@ class TestLoDRankTable(unittest.TestCase): tensor = core.LoDTensor() tensor.set(numpy.random.random(size=(17, 100)), cpu) tensor.set_lod([[0, 1, 3], [0, 5, 6, 7], [0, 3, 4, 9, 10, 13, 16, 17]]) - exe.run(g_main_program, scope=scope, feed={'x': tensor}) + exe.run(scope=scope, feed={'x': tensor}) var = scope.find_var(rank_table.name) table = var.get_lod_rank_table() self.assertEqual([(0, 5), (1, 1), (2, 1)], table.items()) diff --git a/python/paddle/v2/fluid/tests/test_log_loss_op.py b/python/paddle/v2/fluid/tests/test_log_loss_op.py new file mode 100644 index 0000000000000000000000000000000000000000..2eeaa90758c57ef0d92a8ad7b0a4c1b1f2c38be3 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_log_loss_op.py @@ -0,0 +1,33 @@ +import unittest +import numpy as np +from op_test import OpTest + + +class TestLogLossOp(OpTest): + def setUp(self): + self.op_type = 'log_loss' + samples_num = 32 + + predicted = np.random.uniform(0.1, 1.0, + (samples_num, 1)).astype("float32") + labels = np.random.randint(0, 2, (samples_num, 1)).astype("float32") + epsilon = 1e-4 + self.inputs = { + 'Predicted': predicted, + 'Labels': labels, + } + + self.attrs = {'epsilon': epsilon} + loss = -labels * np.log(predicted + epsilon) - ( + 1 - labels) * np.log(1 - predicted + epsilon) + self.outputs = {'Loss': loss} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['Predicted'], 'Loss', max_relative_error=0.03) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_operator_desc.py b/python/paddle/v2/fluid/tests/test_operator_desc.py index e8362d2e9c6038c04c24dce35de8c53bfde78142..ce34d95ac8cb2644dee9c551cd8e85b33609919a 100644 --- a/python/paddle/v2/fluid/tests/test_operator_desc.py +++ b/python/paddle/v2/fluid/tests/test_operator_desc.py @@ -1,11 +1,15 @@ import unittest -from paddle.v2.fluid.framework import Variable, Program, g_main_program + import paddle.v2.fluid.core as core +from paddle.v2.fluid.framework import Program, default_startup_program + +main_program = default_startup_program() + class TestOperator(unittest.TestCase): def test_error_type(self): - block = g_main_program.create_block() + block = main_program.create_block() try: block.append_op() self.assertFail() diff --git a/python/paddle/v2/fluid/tests/test_parameter.py b/python/paddle/v2/fluid/tests/test_parameter.py index 13f6278ad8b7244e7980b32463f29d7a824b4572..694344acbbd3b7c80cb0ff48ada843f794061282 100644 --- a/python/paddle/v2/fluid/tests/test_parameter.py +++ b/python/paddle/v2/fluid/tests/test_parameter.py @@ -1,17 +1,19 @@ import unittest -from paddle.v2.fluid.framework import g_main_program +from paddle.v2.fluid.framework import default_main_program import paddle.v2.fluid.core as core from paddle.v2.fluid.executor import Executor import paddle.v2.fluid.io as io from paddle.v2.fluid.initializer import ConstantInitializer import numpy as np +main_program = default_main_program() + class TestParameter(unittest.TestCase): def test_param(self): shape = [784, 100] val = 1.0625 - b = g_main_program.global_block() + b = main_program.global_block() param = b.create_parameter( name='fc.w', shape=shape, @@ -23,9 +25,9 @@ class TestParameter(unittest.TestCase): self.assertEqual(core.DataType.FP32, param.dtype) self.assertEqual(0, param.block.idx) exe = Executor(core.CPUPlace()) - p = exe.run(g_main_program, fetch_list=[param])[0] + p = exe.run(main_program, fetch_list=[param])[0] self.assertTrue(np.allclose(p, np.ones(shape) * val)) - p = io.get_parameter_value_by_name('fc.w', exe, g_main_program) + p = io.get_parameter_value_by_name('fc.w', exe, main_program) self.assertTrue(np.allclose(np.array(p), np.ones(shape) * val)) diff --git a/python/paddle/v2/fluid/tests/test_profiler.py b/python/paddle/v2/fluid/tests/test_profiler.py new file mode 100644 index 0000000000000000000000000000000000000000..e8f24251b9daeb60d689835308d0e908aa502d51 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_profiler.py @@ -0,0 +1,28 @@ +import unittest +import numpy as np +import paddle.v2.fluid as fluid +import paddle.v2.fluid.profiler as profiler +import paddle.v2.fluid.layers as layers + + +class TestProfiler(unittest.TestCase): + def test_nvprof(self): + if not fluid.core.is_compile_gpu(): + return + epoc = 8 + dshape = [4, 3, 28, 28] + data = layers.data(name='data', shape=[3, 28, 28], dtype='float32') + conv = layers.conv2d(data, 20, 3, stride=[1, 1], padding=[1, 1]) + + place = fluid.GPUPlace(0) + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + + with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof: + for i in range(epoc): + input = np.random.random(dshape).astype("float32") + exe.run(fluid.default_main_program(), feed={'data': input}) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_program.py b/python/paddle/v2/fluid/tests/test_program.py index 15653a1dbf5b1a66edd3f768bee5a36be1bb7a7a..1a9313c68aab165d85ae29051faeacb4927ac2c9 100644 --- a/python/paddle/v2/fluid/tests/test_program.py +++ b/python/paddle/v2/fluid/tests/test_program.py @@ -1,37 +1,38 @@ from __future__ import print_function import unittest -from paddle.v2.fluid.framework import Program -from paddle.v2.fluid.framework import g_main_program +from paddle.v2.fluid.framework import Program, default_main_program import paddle.v2.fluid.layers as layers +main_program = default_main_program() + class TestProgram(unittest.TestCase): def test_program(self): - b = g_main_program.current_block() + b = main_program.current_block() self.assertEqual(-1, b.parent_idx) self.assertEqual(0, b.idx) - b = g_main_program.create_block() + b = main_program.create_block() self.assertEqual(1, b.idx) self.assertEqual(0, b.parent_idx) - b = g_main_program.create_block() + b = main_program.create_block() self.assertEqual(2, b.idx) self.assertEqual(1, b.parent_idx) - g_main_program.rollback() + main_program.rollback() - b = g_main_program.current_block() + b = main_program.current_block() self.assertEqual(1, b.idx) self.assertEqual(0, b.parent_idx) - b = g_main_program.create_block() + b = main_program.create_block() self.assertEqual(3, b.idx) self.assertEqual(1, b.parent_idx) - g_main_program.rollback() - b = g_main_program.current_block() + main_program.rollback() + b = main_program.current_block() self.assertEqual(1, b.idx) self.assertEqual(0, b.parent_idx) diff --git a/python/paddle/v2/fluid/tests/test_shrink_rnn_memory.py b/python/paddle/v2/fluid/tests/test_shrink_rnn_memory.py index 05f6a560644f18da6ff2e015911901cd73cc36c9..86db4c64b493d94cc675ed4bcee7e2925fef1977 100644 --- a/python/paddle/v2/fluid/tests/test_shrink_rnn_memory.py +++ b/python/paddle/v2/fluid/tests/test_shrink_rnn_memory.py @@ -3,9 +3,11 @@ import paddle.v2.fluid.core as core from paddle.v2.fluid.executor import Executor import paddle.v2.fluid.layers as layers from paddle.v2.fluid.backward import append_backward_ops -from paddle.v2.fluid.framework import g_main_program +from paddle.v2.fluid.framework import default_main_program import numpy +main_program = default_main_program() + class TestShrinkRNNMemory(unittest.TestCase): def test_shrink_rnn_memory(self): @@ -36,7 +38,7 @@ class TestShrinkRNNMemory(unittest.TestCase): append_backward_ops(loss=mem3_mean) x_grad = exe.run( feed={'x': tensor}, - fetch_list=[g_main_program.global_block().var('x@GRAD')])[0] + fetch_list=[main_program.global_block().var('x@GRAD')])[0] self.assertAlmostEqual(1.0, x_grad.sum(), delta=0.1) diff --git a/python/paddle/v2/fluid/tests/test_variable.py b/python/paddle/v2/fluid/tests/test_variable.py index 92ffdceb6c84fb2669f8c1bb556c46fb1c03c411..f1e4c0ba21d5c4f10d2b5011bdb5abaebaec5431 100644 --- a/python/paddle/v2/fluid/tests/test_variable.py +++ b/python/paddle/v2/fluid/tests/test_variable.py @@ -1,5 +1,5 @@ import unittest -from paddle.v2.fluid.framework import g_main_program, Program, convert_np_dtype_to_dtype_ +from paddle.v2.fluid.framework import default_main_program, Program, convert_np_dtype_to_dtype_ import paddle.v2.fluid.core as core import numpy as np @@ -18,7 +18,7 @@ class TestVariable(unittest.TestCase): self.assertRaises(ValueError, lambda: convert("int8")) def test_var(self): - b = g_main_program.current_block() + b = default_main_program().current_block() w = b.create_var( dtype="float64", shape=[784, 100], lod_level=0, name="fc.w") self.assertNotEqual(str(w), "")