Merge branch 'develop' of upstream into ctc_edit_distance_dev

.clang-format

@@ -25,4 +25,3 @@ AllowAllParametersOfDeclarationOnNextLine: true
 BinPackParameters: false
 BinPackArguments: false
 ...
-

.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;

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

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

cmake/external/cares.cmake

+# 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)

cmake/external/grpc.cmake

+# 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)

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()

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)

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()

doc/api/v2/config/layer.rst

@@ -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
 ==========================

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:
 
 <img src="src/compiler.png"/>
 
-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:
 
 <img src="src/paddle-compile.png"/>
 
-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:
 
 <img src="src/distributed_architecture.png"/>
 
-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.
 
-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.
+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.
 
 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:
 
 <img src="src/local_architecture.png"/>
 
 
 ### 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:

doc/getstarted/build_and_install/build_from_source_cn.rst

-从源码编译PaddlePaddle
+从源码编译
 ======================
 
 .. _build_step:
@@ -7,8 +7,11 @@
 ----------------
 
 PaddlePaddle主要使用 `CMake <https://cmake.org>`_ 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 <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ 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
@@ -23,7 +26,6 @@ PaddlePaddle主要使用 `CMake <https://cmake.org>`_ 以及GCC, G++作为编译
    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：
 
 .. code-block:: bash
@@ -31,7 +33,33 @@ PaddlePaddle主要使用 `CMake <https://cmake.org>`_ 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
 ----------------

doc/getstarted/build_and_install/build_from_source_en.rst

-Build PaddlePaddle from Sources
+Build from Sources
 ==========================
 
 .. _build_step:
@@ -9,14 +9,18 @@ How To Build
 PaddlePaddle mainly uses `CMake <https://cmake.org>`_ 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 <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_.
+Docker images `here <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ .
+
+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
 ----------------

doc/getstarted/build_and_install/docker_install_cn.rst

-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。

doc/getstarted/build_and_install/docker_install_en.rst

-PaddlePaddle in Docker Containers
+Run in Docker Containers
 =================================
 
 Run PaddlePaddle in Docker container so that you don't need to care about

doc/getstarted/build_and_install/pip_install_cn.rst

-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

doc/getstarted/build_and_install/pip_install_en.rst

-Install PaddlePaddle Using pip
+Install Using pip
 ================================
 
 You can use current widely used Python package management

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

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
 

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<LoD, std::pair<size_t, size_t>>;
-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<size_t>({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<const char *>(&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<int>(dims.size()), 0);
+    std::copy(dims.begin(), dims.end(), pb_dims->begin());
+    int32_t size = desc.ByteSize();
+    os.write(reinterpret_cast<const char *>(&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<void>();
+    PADDLE_ENFORCE(size < std::numeric_limits<std::streamsize>::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<char[]> buf(new char[kBufSize]);
+      auto &gpu_dev_ctx =
+          static_cast<const platform::CUDADeviceContext &>(dev_ctx);
+      platform::CPUPlace cpu;
+      uintptr_t data = reinterpret_cast<uintptr_t>(data_ptr);
+      while (size != 0) {
+        size_t size_to_write = std::min(kBufSize, static_cast<size_t>(size));
+        memory::Copy(cpu, buf.get(),
+                     boost::get<platform::GPUPlace>(tensor.place()),
+                     reinterpret_cast<const void *>(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<const char *>(data_ptr),
+               static_cast<std::streamsize>(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<const char *>(&size), sizeof(size));
+
+    for (auto &each : lod) {
+      size = each.size() * sizeof(framework::LoD::value_type::value_type);
+      os.write(reinterpret_cast<const char *>(&size), sizeof(size));
+      os.write(reinterpret_cast<const char *>(each.data()),
+               static_cast<std::streamsize>(size));
+    }
+  }
+}
+
+void DeserializeFromStream(std::istream &is, LoDTensor *tensor) {
+  uint32_t version;
+  is.read(reinterpret_cast<char *>(&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<char *>(&size), sizeof(size));
+    std::unique_ptr<char[]> buf(new char[size]);
+    is.read(reinterpret_cast<char *>(buf.get()), size);
+    PADDLE_ENFORCE(desc.ParseFromArray(buf.get(), size),
+                   "Cannot parse tensor desc");
+  }
+  {  // read tensor
+    std::vector<int64_t> dims;
+    dims.reserve(static_cast<size_t>(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<float>(cpu);
+        break;
+      case framework::FP64:
+        buf = tensor->mutable_data<double>(cpu);
+        break;
+      case framework::INT32:
+        buf = tensor->mutable_data<int>(cpu);
+        break;
+      case framework::INT64:
+        buf = tensor->mutable_data<int64_t>(cpu);
+        break;
+      default:
+        PADDLE_THROW("DataType %d not supported", desc.data_type());
+    }
+    is.read(static_cast<char *>(buf), tensor->memory_size());
+  }
+  {  // read lod
+    uint64_t lod_level;
+    is.read(reinterpret_cast<char *>(&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<char *>(&size), sizeof(size));
+      std::vector<size_t> tmp(size / sizeof(size_t));
+      is.read(reinterpret_cast<char *>(tmp.data()),
+              static_cast<std::streamsize>(size));
+      lod[i] = tmp;
+    }
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor.h

@@ -189,5 +189,14 @@ std::pair<LoD, std::pair<size_t, size_t>> 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

paddle/gserver/layers/FactorizationMachineLayer.cpp

+/* 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 <algorithm>
+#include <vector>
+#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<Weight>(
+      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<CpuSparseMatrix*>(inputV.get())) {
+    Matrix::resizeOrCreateSparseMatrix(inputSquare_,
+                                       inputV->getHeight(),
+                                       inputV->getWidth(),
+                                       inputV->getElementCnt(),
+                                       inputV->getValueType());
+    inputSquare_->copyFrom(*inputV);
+    (dynamic_cast<CpuSparseMatrix*>(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<CpuSparseMatrix*>(inputV.get())) {
+      Matrix::resizeOrCreateSparseMatrix(tmpInput_,
+                                         inputV->getHeight(),
+                                         inputV->getWidth(),
+                                         inputV->getElementCnt());
+
+      CpuSparseMatrix* sparseInputV =
+          dynamic_cast<CpuSparseMatrix*>(inputV.get());
+      CpuSparseMatrix* sparseInputSquare =
+          dynamic_cast<CpuSparseMatrix*>(inputSquare_.get());
+      CpuSparseMatrix* sparseTmpInput =
+          dynamic_cast<CpuSparseMatrix*>(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

paddle/gserver/layers/FactorizationMachineLayer.h

+/* 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<Weight> 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

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_;
+      }
     }
   }
 }

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

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
+  for (auto useGpu : {false, true}) {
     testLayerGrad(config,
                   "hsigmoid",
                   100,
-                /* trans */ false, /* useGpu */
-                false);
+                  /* 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);

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) {

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);

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)

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<std::string>("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<platform::CPUPlace, T> : public framework::OpKernel<T> {
 
     const auto *x = ctx.Input<Tensor>("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<platform::CPUPlace, T>
     // 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]

paddle/operators/batch_norm_op.cu.cc

@@ -29,6 +29,12 @@ void ExtractNCWHD(const framework::DDim &dims,
                   const TensorFormat &tensor_format, int *N, int *C, int *H,
                   int *W, int *D) {
   *N = dims[0];
+  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
@@ -37,6 +43,7 @@ void ExtractNCWHD(const framework::DDim &dims,
     *D = dims.size() > 4
              ? (tensor_format == TensorFormat::NCHW ? dims[4] : dims[3])
              : 1;
+  }
 }
 
 template <typename T>
@@ -56,8 +63,8 @@ class BatchNormKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
     // NCHW [batch_size, in_channels, in_height, in_width]
     const auto *x = ctx.Input<Tensor>("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<platform::GPUPlace, T>
 
     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);
 

paddle/operators/conv_cudnn_op.cu.cc

@@ -63,7 +63,7 @@ class CudnnConvOpKernel : public framework::OpKernel<T> {
     cudnnConvolutionDescriptor_t cudnn_conv_desc =
         conv_desc.descriptor<T>(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<T> {
     cudnnConvolutionDescriptor_t cudnn_conv_desc =
         conv_desc.descriptor<T>(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.

paddle/operators/conv_op.cc

@@ -97,7 +97,7 @@ Conv2DOpMaker::Conv2DOpMaker(framework::OpProto* proto,
       .SetDefault({0, 0});
   AddAttr<int>(
       "groups",
-      "(int default:1), the group size of convolution operator. "
+      "(int default:1), the groups number of the convolution operator. "
       "According to grouped convolution in Alex Krizhevsky's Deep CNN paper: "
       "when group=2, the first half of the filters is only connected to the "
       "first half of the input channels, while the second half of the filters "
@@ -112,23 +112,29 @@ Conv2DOpMaker::Conv2DOpMaker(framework::OpProto* proto,
 Convolution Operator.
 
 The convolution operation calculates the output based on the input, filter
-and strides, paddings, groups, dilations parameters. The size of each dimension of the
+and strides, paddings, dilations, groups parameters. The size of each dimension of the
 parameters is checked in the infer-shape.
-Input(Input, Filter) and output(Output) are in NCHW format. Where N is batch
+Input(Input) and Output(Output) are in NCHW format. Where N is batch
 size, C is the number of channels, H is the height of the feature, and W is
-the width of the feature. Parameters(ksize, strides, paddings, dilations) are two elements.
-These two elements represent height and width, respectively.
+the width of the feature.
+Filters(Input) is MCHW format. Where M is the number of output image channels, C is
+the number of input image channels, H is the height of the filter, and W
+is the width of the filter.
+Parameters(strides, paddings, dilations) are two elements. These two elements represent
+height and width, respectively.
 The input(X) size and output(Out) size may be different.
 
 Example:
   Input:
-       Input shape: (N, C_in, H_in, W_in)
-       Filter shape: (C_out, C_in, H_f, W_f)
+       Input shape: $(N, C_{in}, H_{in}, W_{in})$
+       Filter shape: $(C_{out}, C_{in}, H_f, W_f)$
   Output:
-       Output shape: (N, C_out, H_out, W_out)
-  where
-       H_out = (H_in + 2 * paddings[0] - (dilations[0]*(filter_size[0] - 1) + 1)) / strides[0] + 1;
-       W_out = (W_in + 2 * paddings[1] - (dilations[1]*(filter_size[1] - 1) + 1)) / strides[1] + 1;
+       Output shape: $(N, C_{out}, H_{out}, W_{out})$
+  Where
+$$
+       H_{out}= \frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]}+ 1 \\
+       W_{out}= \frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]}+ 1
+$$
 )DOC");
 }
 
@@ -165,7 +171,7 @@ Conv3DOpMaker::Conv3DOpMaker(framework::OpProto* proto,
       .SetDefault({0, 0, 0});
   AddAttr<int>(
       "groups",
-      "(int default:1), the group size of convolution operator. "
+      "(int default:1), the groups number of the convolution operator. "
       "According to grouped convolution in Alex Krizhevsky's Deep CNN paper: "
       "when group=2, the first half of the filters is only connected to the "
       "first half of the input channels, while the second half of the filters "
@@ -174,32 +180,37 @@ Conv3DOpMaker::Conv3DOpMaker(framework::OpProto* proto,
   AddAttr<std::vector<int>>("dilations",
                             "(vector<int> default:{1, 1, 1}), the "
                             "dilations(d_dilation, h_dilation, w_dilation) of "
-                            "convolution operator. Currently, conv3d doesn't "
-                            "support dilation.")
+                            "convolution operator.")
       .SetDefault({1, 1, 1});
 
   AddComment(R"DOC(
 Convolution3D Operator.
 
 The convolution operation calculates the output based on the input, filter
-and strides, paddings, groups parameters. The size of each dimension of the
+and strides, paddings, dilations, groups parameters. The size of each dimension of the
 parameters is checked in the infer-shape.
-Input(Input, Filter) and output(Output) are in NCDHW format. Where N is batch
+Input(Input) and output(Output) are in NCDHW format, where N is batch
 size, C is the number of channels,D is the depth of the feature, H is the height of
-the feature, and W is the width of the feature. Parameters(ksize, strides, paddings)
-are three elements. These three elements represent depth, height and width, respectively.
+the feature, and W is the width of the feature.
+Filters(Input) is MCDHW format, where M is the number of output image channels,
+C is the number of input image channels, D is the depth of the filter,
+H is the height of the filter, and W is the width of the filter.
+Parameters(strides, paddings, dilations) are three elements. These three elements
+represent depth, height and width, respectively.
 The input(X) size and output(Out) size may be different.
 
 Example:
   Input:
-       Input shape: (N, C_in, D_in, H_in, W_in)
-       Filter shape: (C_out, C_in, D_f, H_f, W_f)
+       Input shape: $(N, C_{in}, D_{in}, H_{in}, W_{in})$
+       Filter shape: $(C_{out}, C_{in}, D_f, H_f, W_f)$
   Output:
-       Output shape: (N, C_out, D_out, H_out, W_out)
-  where
-       D_out = (D_in - filter_size[0] + 2 * paddings[0]) / strides[0] + 1;
-       H_out = (H_in - filter_size[1] + 2 * paddings[1]) / strides[1] + 1;
-       W_out = (W_in - filter_size[2] + 2 * paddings[2]) / strides[2] + 1;
+       Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$
+  Where
+  $$
+       D_{out}= \frac{(D_{in} + 2 * paddings[0] - (dilations[0] * (D_f - 1) + 1))}{ strides[0]}+ 1 \\
+       H_{out}= \frac{(H_{in} + 2 * paddings[1] - (dilations[1] * (H_f - 1) + 1))}{ strides[1]}+ 1 \\
+       W_{out}= \frac{(W_{in} + 2 * paddings[2] - (dilations[2] * (W_f - 1) + 1))}{ strides[2]}+ 1
+  $$
 )DOC");
 }
 

paddle/operators/conv_transpose_op.cc

@@ -39,7 +39,7 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const {
                  "ConvTransposeOp input dimension and strides dimension should "
                  "be consistent.");
   PADDLE_ENFORCE_EQ(paddings.size(), strides.size(),
-                    "ConvTransposeOp paddings dimension and Conv strides "
+                    "ConvTransposeOp paddings dimension and strides "
                     "dimension should be the same.");
   PADDLE_ENFORCE_EQ(in_dims[1], filter_dims[0],
                     "In ConvTransposeOp, The input channel should be the same "
@@ -62,24 +62,25 @@ Conv2DTransposeOpMaker::Conv2DTransposeOpMaker(
       "The format of input tensor is NCHW. Where N is batch size, C is the "
       "number of input channels, H is the height of the feature, and "
       "W is the width of the feature.");
-  AddInput("Filter",
+  AddInput(
+      "Filter",
       "(Tensor) The filter tensor of convolution transpose operator. "
-           "The format of the filter tensor is CMHW, where C is the number of "
-           "output image channels, M is the number of input image channels, "
+      "The format of the filter tensor is MCHW, where M is the number of "
+      "input feature channels, C is the number of "
+      "output feature channels,"
       "H is the height of the filter, and W is the width of the filter. "
-           "We enforce groups number == 1 and padding == 0 in "
-           "the convolution transpose scenario.");
+      "We enforce groups number == 1 in the convolution transpose scenario.");
   AddOutput("Output",
             "(Tensor) The output tensor of convolution transpose operator. "
             "The format of output tensor is also NCHW.");
   AddAttr<std::vector<int>>(
       "strides",
-      "(vector<int> defalut:{1, 1}), the strides(h_stride, w_stride) of "
+      "(vector<int> default:{1, 1}), the strides(h_stride, w_stride) of "
       "convolution transpose operator.")
       .SetDefault({1, 1});
   AddAttr<std::vector<int>>(
       "paddings",
-      "(vector<int> defalut:{0, 0}), the paddings(h_pad, w_pad) of convolution "
+      "(vector<int> default:{0, 0}), the paddings(h_pad, w_pad) of convolution "
       "transpose operator.")
       .SetDefault({0, 0});
   AddComment(R"DOC(
@@ -88,21 +89,26 @@ Convolution2D Transpose Operator.
 The convolution transpose operation calculates the output based on the input, filter
 and strides, paddings, groups parameters. The size of each dimension of the
 parameters is checked in the infer-shape.
-
-Input(Input, Filter) and output(Output) are in NCHW format. Where N is batch
-size, C is the number of channels, H is the height of the feature, and 
-W is the width of the feature. Parameters(ksize, strides, paddings) are two elements.
-These two elements represent height and width, respectively.
+Input(Input) and output(Output) are in NCHW format. Where N is batchsize, C is the
+number of channels, H is the height of the feature, and W is the width of the feature.
+Filter(Input) is in MCHW format. Where M is the number of input feature channels,
+C is the number of output feature channels, H is the height of the filter,
+and W is the width of the filter.
+Parameters(strides, paddings) are two elements. These two elements represent height
+and width, respectively.
 The input(X) size and output(Out) size may be different.
+
 Example:
   Input:
-       Input shape: (N, C_in, H_in, W_in)
-       Filter shape: (C_in, C_out, H_f, W_f)
+       Input shape: $(N, C_{in}, H_{in}, W_{in})$
+       Filter shape: $(C_{in}, C_{out}, H_f, W_f)$
   Output:
-       Output shape: (N, C_out, H_out, W_out)
-  where
-       H_out = (H_in - 1) * strides[0] - 2 * paddings[0] + filter_size[0];
-       W_out = (W_in - 1) * strides[1] - 2 * paddings[1] + filter_size[1];
+       Output shape: $(N, C_{out}, H_{out}, W_{out})$
+  Where
+  $$
+       H_{out} = (H_{in} - 1) * strides[0] - 2 * paddings[0] + H_f \\
+       W_{out} = (W_{in} - 1) * strides[1] - 2 * paddings[1] + W_f
+  $$
 )DOC");
 }
 
@@ -117,8 +123,9 @@ Conv3DTransposeOpMaker::Conv3DTransposeOpMaker(
            "W is the width of the feature.");
   AddInput("Filter",
            "(Tensor) The filter tensor of convolution transpose operator."
-           "The format of the filter tensor is CMDHW, where C is the number of "
-           "output image channels, M is the number of input image channels, D "
+           "The format of the filter tensor is MCDHW, where M is the number of "
+           "input feature channels, C is the number of "
+           "output feature channels, D "
            "is the depth of the filter, H is the height of the filter, and "
            "W is the width of the filter."
            "We enforce groups number == 1 and padding == 0 in "
@@ -130,12 +137,12 @@ Conv3DTransposeOpMaker::Conv3DTransposeOpMaker(
             "the number of channels, D is the depth of the feature, H is the "
             "height of the feature, and W is the width of the feature.");
   AddAttr<std::vector<int>>("strides",
-                            "(vector<int> defalut:{1, 1, 1}), the "
+                            "(vector<int> default:{1, 1, 1}), the "
                             "strides{d_stride, h_stride, w_stride} of "
                             "convolution transpose operator.")
       .SetDefault({1, 1, 1});
   AddAttr<std::vector<int>>("paddings",
-                            "(vector<int> defalut:{0, 0, 0}), paddings(d_pad, "
+                            "(vector<int> default:{0, 0, 0}), paddings(d_pad, "
                             "h_pad, w_pad) of convolution transpose operator.")
       .SetDefault({0, 0, 0});
   AddComment(R"DOC(
@@ -144,23 +151,28 @@ Convolution3D Transpose Operator.
 The convolution transpose operation calculates the output based on the input, filter
 and strides, paddings, groups parameters. The size of each dimension of the
 parameters is checked in the infer-shape.
-
-Input(Input, Filter) and output(Output) are in NCDHW format. Where N is batch
-size, C is the number of channels, D is the depth of the feature, 
-H is the height of the feature, and W is the width of the feature. 
-Parameters(ksize, strides, paddings) are three elements.
-These three elements represent depth, height and width, respectively.
+Input(Input) and output(Output) are in NCDHW format. Where N is batch size, C is the
+number of channels, D is the depth of the feature, H is the height of the feature,
+and W is the width of the feature.
+Filter(Input) is in MCDHW format. Where M is the number of input feature channels,
+C is the number of output feature channels, D is the depth of the filter,H is the
+height of the filter, and W is the width of the filter.
+Parameters(strides, paddings) are three elements. These three elements represent
+depth, height and width, respectively.
 The input(X) size and output(Out) size may be different.
+
 Example:   
   Input:
-       Input shape: (N, C_in, D_in, H_in, W_in)
-       Filter shape: (C_in, C_out, D_f, H_f, W_f)
+       Input shape: $(N, C_{in}, D_{in}, H_{in}, W_{in})$
+       Filter shape: $(C_{in}, C_{out}, D_f, H_f, W_f)$
   Output:
-       Output shape: (N, C_out, D_out, H_out, W_out)
-  where
-       D_out = (D_in - 1) * strides[0] - 2 * paddings[0] + filter_size[0];
-       H_out = (H_in - 1) * strides[1] - 2 * paddings[1] + filter_size[1];
-       W_out = (W_in - 1) * strides[2] - 2 * paddings[2] + filter_size[2];
+       Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$
+  Where
+  $$
+       D_{out} = (D_{in} - 1) * strides[0] - 2 * paddings[0] + D_f \\
+       H_{out} = (H_{in} - 1) * strides[1] - 2 * paddings[1] + H_f \\
+       W_{out} = (W_{in} - 1) * strides[2] - 2 * paddings[2] + W_f
+  $$
 )DOC");
 }
 

paddle/operators/conv_transpose_op.h

@@ -63,7 +63,6 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
 
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
-    // TODO(Zhuoyuan): Paddings can be added in future.
     // groups will alway be disabled in conv2dtranspose.
 
     const int batch_size = static_cast<int>(input->dims()[0]);

paddle/operators/detail/CMakeLists.txt

+grpc_library(sendrecvop_grpc SRCS recv_impl.cc send_impl.cc PROTO send_recv.proto DEPS lod_tensor selected_rows)

paddle/operators/detail/recv_impl.cc

+/* 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

paddle/operators/detail/send_impl.cc

+/* 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<framework::LoDTensor>(),
+                 "Only support LoDTensor, %s has wrong type", inname);
+  const framework::LoDTensor& tensor = var->Get<framework::LoDTensor>();
+  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<framework::LoDTensor>();
+  // FIXME(typhoonzero): do not copy.
+  framework::CopyFrom(ret_tensor, ctx.GetPlace(), ctx, out_tensor);
+  return true;
+}
+
+}  // namespace detail
+}  // namespace operators
+}  // namespace paddle

paddle/operators/detail/send_recv.proto

+/* 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

paddle/operators/detail/send_recv_impl.h

+/* 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 <grpc++/channel.h>
+// #include <grpc++/client_context.h>
+// #include <grpc++/create_channel.h>
+// #include <grpc++/security/credentials.h>
+#include "paddle/operators/detail/send_recv.grpc.pb.h"
+#include "paddle/operators/detail/send_recv.pb.h"
+
+#include <grpc++/grpc++.h>
+
+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<framework::LoDTensor> lodtensor_queue_;
+  SimpleBlockQueue<framework::LoDTensor> lodtensor_return_queue_;
+  SimpleBlockQueue<framework::SelectedRows> selected_rows_queue_;
+  SimpleBlockQueue<framework::SelectedRows> 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> channel)
+      : stub_(SendRecvService::NewStub(channel)) {}
+
+  bool SendVariable(const framework::Scope &scope, const std::string &inname,
+                    const std::string &outname);
+
+ private:
+  std::unique_ptr<SendRecvService::Stub> stub_;
+};
+
+}  // namespace detail
+}  // namespace operators
+}  // namespace paddle

paddle/operators/detail/simple_block_queue.h

+/* 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 <condition_variable>
+#include <deque>
+#include <mutex>
+
+namespace paddle {
+namespace operators {
+namespace detail {
+
+template <typename T>
+class SimpleBlockQueue {
+ private:
+  std::mutex mutex_;
+  std::condition_variable condition_;
+  std::deque<T> queue_;
+
+ public:
+  void Push(T const& value) {
+    {
+      std::unique_lock<std::mutex> lock(this->mutex_);
+      queue_.push_front(value);
+    }
+    this->condition_.notify_one();
+  }
+
+  T Pop() {
+    std::unique_lock<std::mutex> 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

paddle/operators/gru_op.h

@@ -71,8 +71,8 @@ class GRUKernel : public framework::OpKernel<T> {
 
     int frame_size = hidden_dims[1];
     math::hl_gru_value<T> gru_value;
-    gru_value.gateWeight = const_cast<T*>(weight_data);
-    gru_value.stateWeight =
+    gru_value.gate_weight = const_cast<T*>(weight_data);
+    gru_value.state_weight =
         const_cast<T*>(weight_data + 2 * frame_size * frame_size);
     Tensor ordered_h0;
     const size_t* order = batch_gate->lod()[2].data();
@@ -82,9 +82,9 @@ class GRUKernel : public framework::OpKernel<T> {
       // to reorder.
       ReorderInitState<Place, T>(context.device_context(), *h0, order,
                                  &ordered_h0, true);
-      gru_value.prevOutValue = ordered_h0.data<T>();
+      gru_value.prev_out_value = ordered_h0.data<T>();
     } else {
-      gru_value.prevOutValue = nullptr;
+      gru_value.prev_out_value = nullptr;
     }
     auto batch_starts = batch_gate->lod()[0];
     size_t num_batch = batch_starts.size() - 1;
@@ -96,14 +96,14 @@ class GRUKernel : public framework::OpKernel<T> {
       Tensor gate_t = batch_gate->Slice(bstart, bend);
       Tensor reset_hidden_prev_t = batch_reset_hidden_prev->Slice(bstart, bend);
       Tensor hidden_t = batch_hidden->Slice(bstart, bend);
-      gru_value.outputValue = hidden_t.data<T>();
-      gru_value.gateValue = gate_t.data<T>();
-      gru_value.resetOutputValue = reset_hidden_prev_t.data<T>();
+      gru_value.output_value = hidden_t.data<T>();
+      gru_value.gate_value = gate_t.data<T>();
+      gru_value.reset_output_value = reset_hidden_prev_t.data<T>();
       math::GRUUnitFunctor<Place, T>::compute(
           dev_ctx, gru_value, frame_size, cur_batch_size,
           math::ActiveType(context.Attr<std::string>("activation")),
           math::ActiveType(context.Attr<std::string>("gate_activation")));
-      gru_value.prevOutValue = gru_value.outputValue;
+      gru_value.prev_out_value = gru_value.output_value;
     }
 
     math::Batch2LoDTensorFunctor<Place, T> to_seq;
@@ -169,20 +169,20 @@ class GRUGradKernel : public framework::OpKernel<T> {
     to_batch(dev_ctx, *hidden_grad, batch_hidden_grad, false, is_reverse);
 
     math::hl_gru_value<T> gru_value;
-    gru_value.gateWeight = const_cast<T*>(weight_data);
-    gru_value.stateWeight =
+    gru_value.gate_weight = const_cast<T*>(weight_data);
+    gru_value.state_weight =
         const_cast<T*>(weight_data + 2 * frame_size * frame_size);
 
     math::hl_gru_grad<T> gru_grad;
     if (weight_grad) {
-      gru_grad.gateWeightGrad =
+      gru_grad.gate_weight_grad =
           weight_grad->mutable_data<T>(context.GetPlace());
       zero(dev_ctx, weight_grad, static_cast<T>(0.0));
-      gru_grad.stateWeightGrad =
+      gru_grad.state_weight_grad =
           weight_grad->data<T>() + 2 * frame_size * frame_size;
     } else {
-      gru_grad.gateWeightGrad = nullptr;
-      gru_grad.stateWeightGrad = nullptr;
+      gru_grad.gate_weight_grad = nullptr;
+      gru_grad.state_weight_grad = nullptr;
     }
 
     auto batch_starts = batch_hidden_grad.lod()[0];
@@ -193,27 +193,27 @@ class GRUGradKernel : public framework::OpKernel<T> {
       int cur_batch_size = bend - bstart;
 
       Tensor gate_t = batch_gate->Slice(bstart, bend);
-      gru_value.gateValue = gate_t.data<T>();
+      gru_value.gate_value = gate_t.data<T>();
       Tensor reset_hidden_prev_t = batch_reset_hidden_prev->Slice(bstart, bend);
-      gru_value.resetOutputValue = reset_hidden_prev_t.data<T>();
+      gru_value.reset_output_value = reset_hidden_prev_t.data<T>();
 
       Tensor hidden_grad_t = batch_hidden_grad.Slice(bstart, bend);
-      gru_grad.outputGrad = hidden_grad_t.data<T>();
+      gru_grad.output_grad = hidden_grad_t.data<T>();
       Tensor gate_grad_t = batch_gate_grad.Slice(bstart, bend);
-      gru_grad.gateGrad = gate_grad_t.data<T>();
+      gru_grad.gate_grad = gate_grad_t.data<T>();
       Tensor reset_hidden_prev_grad_t =
           batch_reset_hidden_prev_grad.Slice(bstart, bend);
-      gru_grad.resetOutputGrad = reset_hidden_prev_grad_t.data<T>();
+      gru_grad.reset_output_grad = reset_hidden_prev_grad_t.data<T>();
       if (n == 0) {
-        gru_value.prevOutValue = h0 ? ordered_h0.data<T>() : nullptr;
-        gru_grad.prevOutGrad =
+        gru_value.prev_out_value = h0 ? ordered_h0.data<T>() : nullptr;
+        gru_grad.prev_out_grad =
             h0 && h0_grad ? ordered_h0_grad.data<T>() : nullptr;
       } else {
         int bstart_pre = static_cast<int>(batch_starts[n - 1]);
         Tensor hidden_prev_t = batch_hidden->Slice(bstart_pre, bstart);
-        gru_value.prevOutValue = hidden_prev_t.data<T>();
+        gru_value.prev_out_value = hidden_prev_t.data<T>();
         Tensor hidden_prev_grad_t = batch_hidden_grad.Slice(bstart_pre, bstart);
-        gru_grad.prevOutGrad = hidden_prev_grad_t.data<T>();
+        gru_grad.prev_out_grad = hidden_prev_grad_t.data<T>();
       }
 
       math::GRUUnitGradFunctor<Place, T>::compute(

paddle/operators/load_op.cc

@@ -38,61 +38,7 @@ class LoadOp : public framework::OperatorBase {
                    out_var_name);
 
     auto *tensor = out_var->GetMutable<framework::LoDTensor>();
-
-    uint32_t version;
-    fin.read(reinterpret_cast<char *>(&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<char *>(&size), sizeof(size));
-      std::unique_ptr<char[]> buf(new char[size]);
-      fin.read(reinterpret_cast<char *>(buf.get()), size);
-      PADDLE_ENFORCE(desc.ParseFromArray(buf.get(), size),
-                     "Cannot parse tensor desc");
-    }
-    {  // read tensor
-      std::vector<int64_t> dims;
-      dims.reserve(static_cast<size_t>(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<float>(cpu);
-          break;
-        case framework::FP64:
-          buf = tensor->mutable_data<double>(cpu);
-          break;
-        case framework::INT32:
-          buf = tensor->mutable_data<int>(cpu);
-          break;
-        case framework::INT64:
-          buf = tensor->mutable_data<int64_t>(cpu);
-          break;
-        default:
-          PADDLE_THROW("DataType %d not supported", desc.data_type());
-      }
-      fin.read(static_cast<char *>(buf), tensor->memory_size());
-    }
-    {  // read lod
-      uint64_t lod_level;
-      fin.read(reinterpret_cast<char *>(&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<char *>(&size), sizeof(size));
-        std::vector<size_t> tmp(size / sizeof(size_t));
-        fin.read(reinterpret_cast<char *>(tmp.data()),
-                 static_cast<std::streamsize>(size));
-        lod[i] = tmp;
-      }
-    }
+    framework::DeserializeFromStream(fin, tensor);
 
     auto place = dev_ctx.GetPlace();
     if (platform::is_gpu_place(place)) {

paddle/operators/log_loss_op.cc

+/* 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 <typename AttrType>
+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<AttrType>("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<float>, log_loss_grad,
+            ops::LogLossGradOp);
+REGISTER_OP_CPU_KERNEL(log_loss,
+                       ops::LogLossKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    log_loss_grad, ops::LogLossGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/log_loss_op.cu

+/* 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<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    log_loss_grad, ops::LogLossGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/log_loss_op.h

+/* 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 <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+
+template <typename Place, typename T, typename AttrType = T>
+class LogLossKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* loss_out = ctx.Output<Tensor>("Loss");
+
+    loss_out->mutable_data<T>(ctx.GetPlace());
+
+    auto epsilon = static_cast<T>(ctx.Attr<AttrType>("epsilon"));
+
+    auto prediction = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Predicted"));
+    auto label = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Labels"));
+
+    auto loss = EigenVector<T>::Flatten(*loss_out);
+    auto place = ctx.GetEigenDevice<Place>();
+
+    loss.device(place) = (-(label * (prediction + epsilon).log()) -
+                          ((static_cast<T>(1) - label) *
+                           (static_cast<T>(1) - prediction + epsilon).log()));
+  }
+};
+
+template <typename Place, typename T, typename AttrType = T>
+class LogLossGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto epsilon = static_cast<T>(ctx.Attr<AttrType>("epsilon"));
+
+    auto prediction = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Predicted"));
+    auto label = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Labels"));
+
+    auto* dloss = ctx.Input<Tensor>(framework::GradVarName("Loss"));
+    auto* dpred = ctx.Output<Tensor>(framework::GradVarName("Predicted"));
+
+    auto dl = EigenVector<T>::Flatten(*dloss);
+    auto place = ctx.GetEigenDevice<Place>();
+
+    if (dpred) {
+      dpred->mutable_data<T>(ctx.GetPlace());
+      auto dx = framework::EigenVector<T>::Flatten(*dpred);
+      dx.device(place) = dl * (-(label / (prediction + epsilon)) +
+                               ((static_cast<T>(1) - label) /
+                                (static_cast<T>(1) - prediction + epsilon)));
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/detail/gru_gpu_kernel.h

@@ -27,174 +27,174 @@ namespace math {
 namespace detail {
 
 /*
- * threads(framePerBlock, batchPerBlock)
- * grid(frameBlocks, batchBlocks)
+ * threads(frame_per_block, batch_per_block)
+ * grid(frame_blocks, batch_blocks)
  */
-template <class OpResetOutput, bool isBatch, typename T>
-__global__ void KeGruForwardResetOutput(OpResetOutput opResetOutput,
-                                        T *gateValue, T *resetOutputValue,
-                                        T *prevOutputValue, int frameSize,
-                                        int batchSize,
+template <class OpResetOutput, bool is_batch, typename T>
+__global__ void KeGruForwardResetOutput(OpResetOutput op_reset_output,
+                                        T *gate_value, T *reset_output_value,
+                                        T *prev_output_value, int frame_size,
+                                        int batch_size,
                                         activation_mode_t active_gate) {
-  const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x;
-  if (frameIdx >= frameSize) return;
-
-  int batchIdx = 0;
-  if (isBatch) {
-    batchIdx = blockIdx.y * blockDim.y + threadIdx.y;
-    if (batchIdx >= batchSize) return;
-    gateValue += batchIdx * 3 * frameSize;
-    resetOutputValue += batchIdx * frameSize;
+  const int frame_idx = blockIdx.x * blockDim.x + threadIdx.x;
+  if (frame_idx >= frame_size) return;
+
+  int batch_idx = 0;
+  if (is_batch) {
+    batch_idx = blockIdx.y * blockDim.y + threadIdx.y;
+    if (batch_idx >= batch_size) return;
+    gate_value += batch_idx * 3 * frame_size;
+    reset_output_value += batch_idx * frame_size;
   }
 
-  T rPrevOut = 0;
-  T rValueResetOutput;
-  T rValueUpdateGate = gateValue[frameIdx + frameSize * 0];
-  T rValueResetGate = gateValue[frameIdx + frameSize * 1];
+  T r_prev_out = 0;
+  T r_value_reset_output;
+  T r_value_update_gate = gate_value[frame_idx + frame_size * 0];
+  T r_value_reset_gate = gate_value[frame_idx + frame_size * 1];
 
-  if (prevOutputValue) {
-    if (isBatch) prevOutputValue += batchIdx * frameSize;
-    rPrevOut = prevOutputValue[frameIdx];
+  if (prev_output_value) {
+    if (is_batch) prev_output_value += batch_idx * frame_size;
+    r_prev_out = prev_output_value[frame_idx];
   }
 
-  opResetOutput(rValueUpdateGate, rValueResetGate, rPrevOut, rValueResetOutput,
-                active_gate);
+  op_reset_output(r_value_update_gate, r_value_reset_gate, r_prev_out,
+                  r_value_reset_output, active_gate);
 
-  gateValue[frameIdx + frameSize * 0] = rValueUpdateGate;
-  gateValue[frameIdx + frameSize * 1] = rValueResetGate;
-  resetOutputValue[frameIdx] = rValueResetOutput;
+  gate_value[frame_idx + frame_size * 0] = r_value_update_gate;
+  gate_value[frame_idx + frame_size * 1] = r_value_reset_gate;
+  reset_output_value[frame_idx] = r_value_reset_output;
 }
 
 /*
- * threads(framePerBlock, batchPerBlock)
- * grid(frameBlocks, batchBlocks)
+ * threads(frame_per_block, batch_per_block)
+ * grid(frame_blocks, batch_blocks)
  */
-template <class OpFinalOutput, bool isBatch, typename T>
-__global__ void KeGruForwardFinalOutput(OpFinalOutput opFinalOutput,
-                                        T *gateValue, T *prevOutputValue,
-                                        T *outputValue, int frameSize,
-                                        int batchSize,
+template <class OpFinalOutput, bool is_batch, typename T>
+__global__ void KeGruForwardFinalOutput(OpFinalOutput op_final_output,
+                                        T *gate_value, T *prev_output_value,
+                                        T *output_value, int frame_size,
+                                        int batch_size,
                                         activation_mode_t active_node) {
-  const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x;
-  if (frameIdx >= frameSize) return;
-  int batchIdx = 0;
-  if (isBatch) {
-    batchIdx = blockIdx.y * blockDim.y + threadIdx.y;
-    if (batchIdx >= batchSize) return;
-    gateValue += batchIdx * 3 * frameSize;
-    outputValue += batchIdx * frameSize;
+  const int frame_idx = blockIdx.x * blockDim.x + threadIdx.x;
+  if (frame_idx >= frame_size) return;
+  int batch_idx = 0;
+  if (is_batch) {
+    batch_idx = blockIdx.y * blockDim.y + threadIdx.y;
+    if (batch_idx >= batch_size) return;
+    gate_value += batch_idx * 3 * frame_size;
+    output_value += batch_idx * frame_size;
   }
 
-  T rOutput;
-  T rPrevOut = 0;
-  T rValueUpdateGate = gateValue[frameIdx + frameSize * 0];
-  T rValueFrameState = gateValue[frameIdx + frameSize * 2];
+  T r_output;
+  T r_prev_out = 0;
+  T r_value_update_gate = gate_value[frame_idx + frame_size * 0];
+  T r_value_frame_state = gate_value[frame_idx + frame_size * 2];
 
-  if (prevOutputValue) {
-    if (isBatch) prevOutputValue += batchIdx * frameSize;
-    rPrevOut = prevOutputValue[frameIdx];
+  if (prev_output_value) {
+    if (is_batch) prev_output_value += batch_idx * frame_size;
+    r_prev_out = prev_output_value[frame_idx];
   }
 
-  opFinalOutput(rValueUpdateGate, rValueFrameState, rPrevOut, rOutput,
-                active_node);
+  op_final_output(r_value_update_gate, r_value_frame_state, r_prev_out,
+                  r_output, active_node);
 
-  gateValue[frameIdx + frameSize * 2] = rValueFrameState;
-  outputValue[frameIdx] = rOutput;
+  gate_value[frame_idx + frame_size * 2] = r_value_frame_state;
+  output_value[frame_idx] = r_output;
 }
 
 /*
- * threads(framePerBlock, batchPerBlock)
- * grid(frameBlocks, batchBlocks)
+ * threads(frame_per_block, batch_per_block)
+ * grid(frame_blocks, batch_blocks)
  */
-template <class OpStateGrad, bool isBatch, typename T>
-__global__ void KeGruBackwardStateGrad(OpStateGrad opStateGrad, T *gateValue,
-                                       T *gateGrad, T *prevOutValue,
-                                       T *prevOutGrad, T *outputGrad,
-                                       int frameSize, int batchSize,
+template <class OpStateGrad, bool is_batch, typename T>
+__global__ void KeGruBackwardStateGrad(OpStateGrad op_state_grad, T *gate_value,
+                                       T *gate_grad, T *prev_out_value,
+                                       T *prev_out_grad, T *output_grad,
+                                       int frame_size, int batch_size,
                                        activation_mode_t active_node) {
-  const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x;
-  if (frameIdx >= frameSize) return;
-  int batchIdx = 0;
-  if (isBatch) {
-    batchIdx = blockIdx.y * blockDim.y + threadIdx.y;
-    if (batchIdx >= batchSize) return;
-    gateValue += batchIdx * 3 * frameSize;
-    gateGrad += batchIdx * 3 * frameSize;
-    outputGrad += batchIdx * frameSize;
+  const int frame_idx = blockIdx.x * blockDim.x + threadIdx.x;
+  if (frame_idx >= frame_size) return;
+  int batch_idx = 0;
+  if (is_batch) {
+    batch_idx = blockIdx.y * blockDim.y + threadIdx.y;
+    if (batch_idx >= batch_size) return;
+    gate_value += batch_idx * 3 * frame_size;
+    gate_grad += batch_idx * 3 * frame_size;
+    output_grad += batch_idx * frame_size;
   }
 
-  T rUpdateGateGrad;
-  T rFrameStateGrad;
-  T rPrevOutValue = 0;
-  T rPrevOutGrad = 0;
-  T rUpdateGateValue = gateValue[frameIdx + frameSize * 0];
-  T rFrameStateValue = gateValue[frameIdx + frameSize * 2];
-  T rOutGrad = outputGrad[frameIdx];
+  T r_update_gate_grad;
+  T r_frame_state_grad;
+  T r_prev_out_value = 0;
+  T r_prev_out_grad = 0;
+  T r_update_gate_value = gate_value[frame_idx + frame_size * 0];
+  T r_frame_state_value = gate_value[frame_idx + frame_size * 2];
+  T r_out_grad = output_grad[frame_idx];
 
-  if (prevOutValue && prevOutGrad) {
-    if (isBatch) prevOutValue += batchIdx * frameSize;
-    rPrevOutValue = prevOutValue[frameIdx];
+  if (prev_out_value && prev_out_grad) {
+    if (is_batch) prev_out_value += batch_idx * frame_size;
+    r_prev_out_value = prev_out_value[frame_idx];
 
-    if (isBatch) prevOutGrad += batchIdx * frameSize;
-    rPrevOutGrad = prevOutGrad[frameIdx];
+    if (is_batch) prev_out_grad += batch_idx * frame_size;
+    r_prev_out_grad = prev_out_grad[frame_idx];
   }
 
-  opStateGrad(rUpdateGateValue, rUpdateGateGrad, rFrameStateValue,
-              rFrameStateGrad, rPrevOutValue, rPrevOutGrad, rOutGrad,
-              active_node);
+  op_state_grad(r_update_gate_value, r_update_gate_grad, r_frame_state_value,
+                r_frame_state_grad, r_prev_out_value, r_prev_out_grad,
+                r_out_grad, active_node);
 
-  gateGrad[frameIdx + frameSize * 0] = rUpdateGateGrad;
-  gateGrad[frameIdx + frameSize * 2] = rFrameStateGrad;
-  if (prevOutGrad) {
-    prevOutGrad[frameIdx] = rPrevOutGrad;
+  gate_grad[frame_idx + frame_size * 0] = r_update_gate_grad;
+  gate_grad[frame_idx + frame_size * 2] = r_frame_state_grad;
+  if (prev_out_grad) {
+    prev_out_grad[frame_idx] = r_prev_out_grad;
   }
 }
 
 /*
- * threads(framePerBlock, batchPerBlock)
- * grid(frameBlocks, batchBlocks)
+ * threads(frame_per_block, batch_per_block)
+ * grid(frame_blocks, batch_blocks)
  */
-template <class OpResetGrad, bool isBatch, typename T>
-__global__ void KeGruBackwardResetGrad(OpResetGrad opResetGrad, T *gateValue,
-                                       T *gateGrad, T *prevOutValue,
-                                       T *prevOutGrad, T *resetOutputGrad,
-                                       int frameSize, int batchSize,
+template <class OpResetGrad, bool is_batch, typename T>
+__global__ void KeGruBackwardResetGrad(OpResetGrad op_reset_grad, T *gate_value,
+                                       T *gate_grad, T *prev_out_value,
+                                       T *prev_out_grad, T *reset_output_grad,
+                                       int frame_size, int batch_size,
                                        activation_mode_t active_gate) {
-  const int frameIdx = blockIdx.x * blockDim.x + threadIdx.x;
-  if (frameIdx >= frameSize) return;
-  int batchIdx = 0;
-  if (isBatch) {
-    batchIdx = blockIdx.y * blockDim.y + threadIdx.y;
-    if (batchIdx >= batchSize) return;
-    gateValue += batchIdx * 3 * frameSize;
-    gateGrad += batchIdx * 3 * frameSize;
-    resetOutputGrad += batchIdx * frameSize;
+  const int frame_idx = blockIdx.x * blockDim.x + threadIdx.x;
+  if (frame_idx >= frame_size) return;
+  int batch_idx = 0;
+  if (is_batch) {
+    batch_idx = blockIdx.y * blockDim.y + threadIdx.y;
+    if (batch_idx >= batch_size) return;
+    gate_value += batch_idx * 3 * frame_size;
+    gate_grad += batch_idx * 3 * frame_size;
+    reset_output_grad += batch_idx * frame_size;
   }
 
-  T rResetGateGrad;
-  T rPrevOutValue = 0;
-  T rPrevOutGrad = 0;
-  T rResetOutputGrad = 0;
-  T rUpdateGateValue = gateValue[frameIdx + frameSize * 0];
-  T rUpdateGateGrad = gateGrad[frameIdx + frameSize * 0];
-  T rResetGateValue = gateValue[frameIdx + frameSize * 1];
-
-  if (prevOutValue && prevOutGrad) {
-    if (isBatch) prevOutValue += batchIdx * frameSize;
-    if (isBatch) prevOutGrad += batchIdx * frameSize;
-    rPrevOutValue = prevOutValue[frameIdx];
-    rPrevOutGrad = prevOutGrad[frameIdx];
-    rResetOutputGrad = resetOutputGrad[frameIdx];
+  T r_reset_gate_grad;
+  T r_prev_out_value = 0;
+  T r_prev_out_grad = 0;
+  T r_reset_output_grad = 0;
+  T r_update_gate_value = gate_value[frame_idx + frame_size * 0];
+  T r_update_gate_grad = gate_grad[frame_idx + frame_size * 0];
+  T r_reset_gate_value = gate_value[frame_idx + frame_size * 1];
+
+  if (prev_out_value && prev_out_grad) {
+    if (is_batch) prev_out_value += batch_idx * frame_size;
+    if (is_batch) prev_out_grad += batch_idx * frame_size;
+    r_prev_out_value = prev_out_value[frame_idx];
+    r_prev_out_grad = prev_out_grad[frame_idx];
+    r_reset_output_grad = reset_output_grad[frame_idx];
   }
 
-  opResetGrad(rUpdateGateValue, rUpdateGateGrad, rResetGateValue,
-              rResetGateGrad, rPrevOutValue, rPrevOutGrad, rResetOutputGrad,
-              active_gate);
+  op_reset_grad(r_update_gate_value, r_update_gate_grad, r_reset_gate_value,
+                r_reset_gate_grad, r_prev_out_value, r_prev_out_grad,
+                r_reset_output_grad, active_gate);
 
-  gateGrad[frameIdx + frameSize * 0] = rUpdateGateGrad;
-  gateGrad[frameIdx + frameSize * 1] = rResetGateGrad;
-  if (prevOutGrad) {
-    prevOutGrad[frameIdx] = rPrevOutGrad;
+  gate_grad[frame_idx + frame_size * 0] = r_update_gate_grad;
+  gate_grad[frame_idx + frame_size * 1] = r_reset_gate_grad;
+  if (prev_out_grad) {
+    prev_out_grad[frame_idx] = r_prev_out_grad;
   }
 }
 }  // namespace detail

paddle/operators/math/detail/gru_kernel.h

@@ -28,23 +28,25 @@ namespace forward {
 template <typename T>
 class gru_resetOutput {
  public:
-  HOSTDEVICE void operator()(T &valueUpdateGate, T &valueResetGate, T &prevOut,
-                             T &valueResetOutput, activation_mode_t actGate) {
-    valueUpdateGate = activation(valueUpdateGate, actGate);
-    valueResetGate = activation(valueResetGate, actGate);
-    valueResetOutput = prevOut * valueResetGate;
+  HOSTDEVICE void operator()(T &value_update_gate, T &value_reset_gate,
+                             T &prev_out, T &value_reset_output,
+                             activation_mode_t act_gate) {
+    value_update_gate = activation(value_update_gate, act_gate);
+    value_reset_gate = activation(value_reset_gate, act_gate);
+    value_reset_output = prev_out * value_reset_gate;
   }
 #ifndef __NVCC__
 #ifndef __AVX__
   static const bool avx = false;
 #else
   static const bool avx = true;
-  HOSTDEVICE void operator()(__m256 &valueUpdateGate, __m256 &valueResetGate,
-                             __m256 &prevOut, __m256 &valueResetOutput,
-                             activation_mode_t actGate) {
-    valueUpdateGate = activation(valueUpdateGate, actGate);
-    valueResetGate = activation(valueResetGate, actGate);
-    valueResetOutput = _mm256_mul_ps(prevOut, valueResetGate);
+  HOSTDEVICE void operator()(__m256 &value_update_gate,
+                             __m256 &value_reset_gate, __m256 &prev_out,
+                             __m256 &value_reset_output,
+                             activation_mode_t act_gate) {
+    value_update_gate = activation(value_update_gate, act_gate);
+    value_reset_gate = activation(value_reset_gate, act_gate);
+    value_reset_output = _mm256_mul_ps(prev_out, value_reset_gate);
   }
 #endif
 #endif
@@ -53,24 +55,26 @@ class gru_resetOutput {
 template <typename T>
 class gru_finalOutput {
  public:
-  HOSTDEVICE void operator()(T &valueUpdateGate, T &valueFrameState, T &prevOut,
-                             T &valueOutput, activation_mode_t actInput) {
-    valueFrameState = activation(valueFrameState, actInput);
-    valueOutput = prevOut - (valueUpdateGate * prevOut) +
-                  (valueUpdateGate * valueFrameState);
+  HOSTDEVICE void operator()(T &value_update_gate, T &value_frame_state,
+                             T &prev_out, T &value_output,
+                             activation_mode_t act_input) {
+    value_frame_state = activation(value_frame_state, act_input);
+    value_output = prev_out - (value_update_gate * prev_out) +
+                   (value_update_gate * value_frame_state);
   }
 #ifndef __NVCC__
 #ifndef __AVX__
   static const bool avx = false;
 #else
   static const bool avx = true;
-  HOSTDEVICE void operator()(__m256 &valueUpdateGate, __m256 &valueFrameState,
-                             __m256 &prevOut, __m256 &valueOutput,
-                             activation_mode_t actInput) {
-    valueFrameState = activation(valueFrameState, actInput);
-    valueOutput = _mm256_add_ps(
-        _mm256_sub_ps(prevOut, _mm256_mul_ps(valueUpdateGate, prevOut)),
-        _mm256_mul_ps(valueUpdateGate, valueFrameState));
+  HOSTDEVICE void operator()(__m256 &value_update_gate,
+                             __m256 &value_frame_state, __m256 &prev_out,
+                             __m256 &value_output,
+                             activation_mode_t act_input) {
+    value_frame_state = activation(value_frame_state, act_input);
+    value_output = _mm256_add_ps(
+        _mm256_sub_ps(prev_out, _mm256_mul_ps(value_update_gate, prev_out)),
+        _mm256_mul_ps(value_update_gate, value_frame_state));
   }
 #endif
 #endif
@@ -82,34 +86,37 @@ namespace backward {
 template <typename T>
 class gru_stateGrad {
  public:
-  HOSTDEVICE void operator()(T &valueUpdateGate, T &gradUpdateGate,
-                             T &valueFrameState, T &gradFrameState,
-                             T &valuePrevOut, T &gradPrevOut, T &gradOutput,
-                             activation_mode_t actInput) {
-    gradUpdateGate = (gradOutput * valueFrameState);
-    gradUpdateGate -= (gradOutput * valuePrevOut);
-    gradPrevOut -= (gradOutput * valueUpdateGate);
-    gradPrevOut += gradOutput;
-    gradFrameState =
-        activation(gradOutput * valueUpdateGate, valueFrameState, actInput);
+  HOSTDEVICE void operator()(T &value_update_gate, T &grad_update_gate,
+                             T &value_frame_state, T &grad_frame_state,
+                             T &value_prev_out, T &grad_prev_out,
+                             T &grad_output, activation_mode_t act_input) {
+    grad_update_gate = (grad_output * value_frame_state);
+    grad_update_gate -= (grad_output * value_prev_out);
+    grad_prev_out -= (grad_output * value_update_gate);
+    grad_prev_out += grad_output;
+    grad_frame_state = activation(grad_output * value_update_gate,
+                                  value_frame_state, act_input);
   }
 #ifndef __NVCC__
 #ifndef __AVX__
   static const bool avx = false;
 #else
   static const bool avx = true;
-  HOSTDEVICE void operator()(__m256 &valueUpdateGate, __m256 &gradUpdateGate,
-                             __m256 &valueFrameState, __m256 &gradFrameState,
-                             __m256 &valuePrevOut, __m256 &gradPrevOut,
-                             __m256 &gradOutput, activation_mode_t actInput) {
-    gradUpdateGate = _mm256_mul_ps(gradOutput, valueFrameState);
-    gradUpdateGate =
-        _mm256_sub_ps(gradUpdateGate, _mm256_mul_ps(gradOutput, valuePrevOut));
-    gradPrevOut = _mm256_add_ps(
-        _mm256_sub_ps(gradPrevOut, _mm256_mul_ps(gradOutput, valueUpdateGate)),
-        gradOutput);
-    gradFrameState = activation(_mm256_mul_ps(gradOutput, valueUpdateGate),
-                                valueFrameState, actInput);
+  HOSTDEVICE void operator()(__m256 &value_update_gate,
+                             __m256 &grad_update_gate,
+                             __m256 &value_frame_state,
+                             __m256 &grad_frame_state, __m256 &value_prev_out,
+                             __m256 &grad_prev_out, __m256 &grad_output,
+                             activation_mode_t act_input) {
+    grad_update_gate = _mm256_mul_ps(grad_output, value_frame_state);
+    grad_update_gate = _mm256_sub_ps(
+        grad_update_gate, _mm256_mul_ps(grad_output, value_prev_out));
+    grad_prev_out = _mm256_add_ps(
+        _mm256_sub_ps(grad_prev_out,
+                      _mm256_mul_ps(grad_output, value_update_gate)),
+        grad_output);
+    grad_frame_state = activation(_mm256_mul_ps(grad_output, value_update_gate),
+                                  value_frame_state, act_input);
   }
 #endif
 #endif
@@ -118,30 +125,32 @@ class gru_stateGrad {
 template <typename T>
 class gru_resetGrad {
  public:
-  HOSTDEVICE void operator()(T &valueUpdateGate, T &gradUpdateGate,
-                             T &valueResetGate, T &gradResetGate,
-                             T &valuePrevOut, T &gradPrevOut,
-                             T &gradResetOutput, activation_mode_t actGate) {
-    gradResetGate = (gradResetOutput * valuePrevOut);
-    gradPrevOut += (gradResetOutput * valueResetGate);
-    gradUpdateGate = activation(gradUpdateGate, valueUpdateGate, actGate);
-    gradResetGate = activation(gradResetGate, valueResetGate, actGate);
+  HOSTDEVICE void operator()(T &value_update_gate, T &grad_update_gate,
+                             T &value_reset_gate, T &grad_reset_gate,
+                             T &value_prev_out, T &grad_prev_out,
+                             T &grad_reset_output, activation_mode_t act_gate) {
+    grad_reset_gate = (grad_reset_output * value_prev_out);
+    grad_prev_out += (grad_reset_output * value_reset_gate);
+    grad_update_gate =
+        activation(grad_update_gate, value_update_gate, act_gate);
+    grad_reset_gate = activation(grad_reset_gate, value_reset_gate, act_gate);
   }
 #ifndef __NVCC__
 #ifndef __AVX__
   static const bool avx = false;
 #else
   static const bool avx = true;
-  HOSTDEVICE void operator()(__m256 &valueUpdateGate, __m256 &gradUpdateGate,
-                             __m256 &valueResetGate, __m256 &gradResetGate,
-                             __m256 &valuePrevOut, __m256 &gradPrevOut,
-                             __m256 &gradResetOutput,
-                             activation_mode_t actGate) {
-    gradResetGate = _mm256_mul_ps(gradResetOutput, valuePrevOut);
-    gradPrevOut = _mm256_add_ps(gradPrevOut,
-                                _mm256_mul_ps(gradResetOutput, valueResetGate));
-    gradUpdateGate = activation(gradUpdateGate, valueUpdateGate, actGate);
-    gradResetGate = activation(gradResetGate, valueResetGate, actGate);
+  HOSTDEVICE void operator()(__m256 &value_update_gate,
+                             __m256 &grad_update_gate, __m256 &value_reset_gate,
+                             __m256 &grad_reset_gate, __m256 &value_prev_out,
+                             __m256 &grad_prev_out, __m256 &grad_reset_output,
+                             activation_mode_t act_gate) {
+    grad_reset_gate = _mm256_mul_ps(grad_reset_output, value_prev_out);
+    grad_prev_out = _mm256_add_ps(
+        grad_prev_out, _mm256_mul_ps(grad_reset_output, value_reset_gate));
+    grad_update_gate =
+        activation(grad_update_gate, value_update_gate, act_gate);
+    grad_reset_gate = activation(grad_reset_gate, value_reset_gate, act_gate);
   }
 #endif
 #endif

paddle/operators/math/gru_compute.cc

@@ -21,29 +21,29 @@ namespace math {
 template <typename T>
 struct GRUUnitFunctor<platform::CPUPlace, T> {
   static void compute(const platform::DeviceContext &context,
-                      hl_gru_value<T> value, int frameSize, int batchSize,
+                      hl_gru_value<T> value, int frame_size, int batch_size,
                       activation_mode_t active_node,
                       activation_mode_t active_gate) {
 #ifndef __NVCC__
-    if (value.prevOutValue) {
+    if (value.prev_out_value) {
       math::gemm<platform::CPUPlace, T>(
-          context, false, false, batchSize, frameSize * 2, frameSize, 1,
-          value.prevOutValue, frameSize, value.gateWeight, frameSize * 2, 1,
-          value.gateValue, frameSize * 3);
+          context, false, false, batch_size, frame_size * 2, frame_size, 1,
+          value.prev_out_value, frame_size, value.gate_weight, frame_size * 2,
+          1, value.gate_value, frame_size * 3);
     }
 
     detail::forward_reset_output(detail::forward::gru_resetOutput<T>(), value,
-                                 frameSize, batchSize, active_gate);
+                                 frame_size, batch_size, active_gate);
 
-    if (value.prevOutValue) {
+    if (value.prev_out_value) {
       math::gemm<platform::CPUPlace, T>(
-          context, false, false, batchSize, frameSize, frameSize, 1,
-          value.resetOutputValue, frameSize, value.stateWeight, frameSize, 1,
-          value.gateValue + frameSize * 2, frameSize * 3);
+          context, false, false, batch_size, frame_size, frame_size, 1,
+          value.reset_output_value, frame_size, value.state_weight, frame_size,
+          1, value.gate_value + frame_size * 2, frame_size * 3);
     }
 
     detail::forward_final_output(detail::forward::gru_finalOutput<T>(), value,
-                                 frameSize, batchSize, active_node);
+                                 frame_size, batch_size, active_node);
 #endif
   }
 };
@@ -51,41 +51,43 @@ struct GRUUnitFunctor<platform::CPUPlace, T> {
 template <typename T>
 struct GRUUnitGradFunctor<platform::CPUPlace, T> {
   static void compute(const platform::DeviceContext &context,
-                      hl_gru_value<T> value, hl_gru_grad<T> grad, int frameSize,
-                      int batchSize, activation_mode_t active_node,
+                      hl_gru_value<T> value, hl_gru_grad<T> grad,
+                      int frame_size, int batch_size,
+                      activation_mode_t active_node,
                       activation_mode_t active_gate) {
 #ifndef __NVCC__
     detail::backward_state_grad(detail::backward::gru_stateGrad<T>(), value,
-                                grad, frameSize, batchSize, active_node);
+                                grad, frame_size, batch_size, active_node);
 
-    if (value.prevOutValue && grad.prevOutGrad) {
+    if (value.prev_out_value && grad.prev_out_grad) {
       math::gemm<platform::CPUPlace, T>(
-          context, false, true, batchSize, frameSize, frameSize, 1,
-          grad.gateGrad + frameSize * 2, frameSize * 3, value.stateWeight,
-          frameSize, 0, grad.resetOutputGrad, frameSize);
+          context, false, true, batch_size, frame_size, frame_size, 1,
+          grad.gate_grad + frame_size * 2, frame_size * 3, value.state_weight,
+          frame_size, 0, grad.reset_output_grad, frame_size);
 
-      if (grad.stateWeightGrad) {
+      if (grad.state_weight_grad) {
         math::gemm<platform::CPUPlace, T>(
-            context, true, false, frameSize, frameSize, batchSize, 1,
-            value.resetOutputValue, frameSize, grad.gateGrad + frameSize * 2,
-            frameSize * 3, 1, grad.stateWeightGrad, frameSize);
+            context, true, false, frame_size, frame_size, batch_size, 1,
+            value.reset_output_value, frame_size,
+            grad.gate_grad + frame_size * 2, frame_size * 3, 1,
+            grad.state_weight_grad, frame_size);
       }
     }
 
     detail::backward_reset_grad(detail::backward::gru_resetGrad<T>(), value,
-                                grad, frameSize, batchSize, active_gate);
+                                grad, frame_size, batch_size, active_gate);
 
-    if (grad.prevOutGrad && value.prevOutValue) {
+    if (grad.prev_out_grad && value.prev_out_value) {
       math::gemm<platform::CPUPlace, T>(
-          context, false, true, batchSize, frameSize, frameSize * 2, 1,
-          grad.gateGrad, frameSize * 3, value.gateWeight, frameSize * 2, 1,
-          grad.prevOutGrad, frameSize);
+          context, false, true, batch_size, frame_size, frame_size * 2, 1,
+          grad.gate_grad, frame_size * 3, value.gate_weight, frame_size * 2, 1,
+          grad.prev_out_grad, frame_size);
 
-      if (grad.gateWeightGrad) {
+      if (grad.gate_weight_grad) {
         math::gemm<platform::CPUPlace, T>(
-            context, true, false, frameSize, frameSize * 2, batchSize, 1,
-            value.prevOutValue, frameSize, grad.gateGrad, frameSize * 3, 1,
-            grad.gateWeightGrad, frameSize * 2);
+            context, true, false, frame_size, frame_size * 2, batch_size, 1,
+            value.prev_out_value, frame_size, grad.gate_grad, frame_size * 3, 1,
+            grad.gate_weight_grad, frame_size * 2);
       }
     }
 #endif

paddle/operators/math/gru_compute.cu

@@ -21,66 +21,66 @@ namespace math {
 template <typename T>
 struct GRUUnitFunctor<platform::GPUPlace, T> {
   static void compute(const platform::DeviceContext &context,
-                      hl_gru_value<T> value, int frameSize, int batchSize,
+                      hl_gru_value<T> value, int frame_size, int batch_size,
                       activation_mode_t active_node,
                       activation_mode_t active_gate) {
     auto stream =
         reinterpret_cast<const platform::CUDADeviceContext &>(context).stream();
     dim3 threads;
     dim3 grid;
-    if (batchSize == 1) {
-      int framePerBlock = frameSize <= 1024 ? frameSize : 1024;
-      int frameBlocks = (frameSize + 1024 - 1) / 1024;
-      threads = dim3(framePerBlock, 1);
-      grid = dim3(frameBlocks, 1);
+    if (batch_size == 1) {
+      int frame_per_block = frame_size <= 1024 ? frame_size : 1024;
+      int frame_blocks = (frame_size + 1024 - 1) / 1024;
+      threads = dim3(frame_per_block, 1);
+      grid = dim3(frame_blocks, 1);
     } else {
       threads = dim3(32, 32);
-      grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32);
+      grid = dim3((frame_size + 32 - 1) / 32, (batch_size + 32 - 1) / 32);
     }
 
-    if (value.prevOutValue) {
+    if (value.prev_out_value) {
       math::gemm<platform::GPUPlace, T>(
-          context, false, false, batchSize, frameSize * 2, frameSize, 1,
-          value.prevOutValue, frameSize, value.gateWeight, frameSize * 2, 1,
-          value.gateValue, frameSize * 3);
+          context, false, false, batch_size, frame_size * 2, frame_size, 1,
+          value.prev_out_value, frame_size, value.gate_weight, frame_size * 2,
+          1, value.gate_value, frame_size * 3);
     }
 
-    if (batchSize == 1) {
+    if (batch_size == 1) {
       detail::KeGruForwardResetOutput<detail::forward::gru_resetOutput<T>,
-                                      /* isBatch= */ false,
+                                      /* is_batch= */ false,
                                       T><<<grid, threads, 0, stream>>>(
-          detail::forward::gru_resetOutput<T>(), value.gateValue,
-          value.resetOutputValue, value.prevOutValue, frameSize, batchSize,
-          active_gate);
+          detail::forward::gru_resetOutput<T>(), value.gate_value,
+          value.reset_output_value, value.prev_out_value, frame_size,
+          batch_size, active_gate);
     } else {
       detail::KeGruForwardResetOutput<detail::forward::gru_resetOutput<T>,
-                                      /* isBatch= */ true,
+                                      /* is_batch= */ true,
                                       T><<<grid, threads, 0, stream>>>(
-          detail::forward::gru_resetOutput<T>(), value.gateValue,
-          value.resetOutputValue, value.prevOutValue, frameSize, batchSize,
-          active_gate);
+          detail::forward::gru_resetOutput<T>(), value.gate_value,
+          value.reset_output_value, value.prev_out_value, frame_size,
+          batch_size, active_gate);
     }
 
-    if (value.prevOutValue) {
+    if (value.prev_out_value) {
       math::gemm<platform::GPUPlace, T>(
-          context, false, false, batchSize, frameSize, frameSize, 1,
-          value.resetOutputValue, frameSize, value.stateWeight, frameSize, 1,
-          value.gateValue + frameSize * 2, frameSize * 3);
+          context, false, false, batch_size, frame_size, frame_size, 1,
+          value.reset_output_value, frame_size, value.state_weight, frame_size,
+          1, value.gate_value + frame_size * 2, frame_size * 3);
     }
 
-    if (batchSize == 1) {
+    if (batch_size == 1) {
       detail::KeGruForwardFinalOutput<detail::forward::gru_finalOutput<T>,
-                                      /* isBatch= */ false,
+                                      /* is_batch= */ false,
                                       T><<<grid, threads, 0, stream>>>(
-          detail::forward::gru_finalOutput<T>(), value.gateValue,
-          value.prevOutValue, value.outputValue, frameSize, batchSize,
+          detail::forward::gru_finalOutput<T>(), value.gate_value,
+          value.prev_out_value, value.output_value, frame_size, batch_size,
           active_node);
     } else {
       detail::KeGruForwardFinalOutput<detail::forward::gru_finalOutput<T>,
-                                      /* isBatch= */ true,
+                                      /* is_batch= */ true,
                                       T><<<grid, threads, 0, stream>>>(
-          detail::forward::gru_finalOutput<T>(), value.gateValue,
-          value.prevOutValue, value.outputValue, frameSize, batchSize,
+          detail::forward::gru_finalOutput<T>(), value.gate_value,
+          value.prev_out_value, value.output_value, frame_size, batch_size,
           active_node);
     }
   }
@@ -89,80 +89,82 @@ struct GRUUnitFunctor<platform::GPUPlace, T> {
 template <typename T>
 struct GRUUnitGradFunctor<platform::GPUPlace, T> {
   static void compute(const platform::DeviceContext &context,
-                      hl_gru_value<T> value, hl_gru_grad<T> grad, int frameSize,
-                      int batchSize, activation_mode_t active_node,
+                      hl_gru_value<T> value, hl_gru_grad<T> grad,
+                      int frame_size, int batch_size,
+                      activation_mode_t active_node,
                       activation_mode_t active_gate) {
     auto stream =
         reinterpret_cast<const platform::CUDADeviceContext &>(context).stream();
     dim3 threads;
     dim3 grid;
-    if (batchSize == 1) {
-      int framePerBlock = frameSize <= 1024 ? frameSize : 1024;
-      int frameBlocks = (frameSize + 1024 - 1) / 1024;
-      threads = dim3(framePerBlock, 1);
-      grid = dim3(frameBlocks, 1);
+    if (batch_size == 1) {
+      int frame_per_block = frame_size <= 1024 ? frame_size : 1024;
+      int frame_blocks = (frame_size + 1024 - 1) / 1024;
+      threads = dim3(frame_per_block, 1);
+      grid = dim3(frame_blocks, 1);
     } else {
       threads = dim3(32, 32);
-      grid = dim3((frameSize + 32 - 1) / 32, (batchSize + 32 - 1) / 32);
+      grid = dim3((frame_size + 32 - 1) / 32, (batch_size + 32 - 1) / 32);
     }
 
-    if (batchSize == 1) {
+    if (batch_size == 1) {
       detail::KeGruBackwardStateGrad<
           detail::backward::gru_stateGrad<T>,
-          /* isBatch= */ false><<<grid, threads, 0, stream>>>(
-          detail::backward::gru_stateGrad<T>(), value.gateValue, grad.gateGrad,
-          value.prevOutValue, grad.prevOutGrad, grad.outputGrad, frameSize,
-          batchSize, active_node);
+          /* is_batch= */ false><<<grid, threads, 0, stream>>>(
+          detail::backward::gru_stateGrad<T>(), value.gate_value,
+          grad.gate_grad, value.prev_out_value, grad.prev_out_grad,
+          grad.output_grad, frame_size, batch_size, active_node);
     } else {
       detail::KeGruBackwardStateGrad<
           detail::backward::gru_stateGrad<T>,
-          /* isBatch= */ true><<<grid, threads, 0, stream>>>(
-          detail::backward::gru_stateGrad<T>(), value.gateValue, grad.gateGrad,
-          value.prevOutValue, grad.prevOutGrad, grad.outputGrad, frameSize,
-          batchSize, active_node);
+          /* is_batch= */ true><<<grid, threads, 0, stream>>>(
+          detail::backward::gru_stateGrad<T>(), value.gate_value,
+          grad.gate_grad, value.prev_out_value, grad.prev_out_grad,
+          grad.output_grad, frame_size, batch_size, active_node);
     }
 
-    if (value.prevOutValue && grad.prevOutGrad) {
+    if (value.prev_out_value && grad.prev_out_grad) {
       math::gemm<platform::GPUPlace, T>(
-          context, false, true, batchSize, frameSize, frameSize, 1,
-          grad.gateGrad + frameSize * 2, frameSize * 3, value.stateWeight,
-          frameSize, 0, grad.resetOutputGrad, frameSize);
+          context, false, true, batch_size, frame_size, frame_size, 1,
+          grad.gate_grad + frame_size * 2, frame_size * 3, value.state_weight,
+          frame_size, 0, grad.reset_output_grad, frame_size);
 
-      if (grad.stateWeightGrad) {
+      if (grad.state_weight_grad) {
         math::gemm<platform::GPUPlace, T>(
-            context, true, false, frameSize, frameSize, batchSize, 1,
-            value.resetOutputValue, frameSize, grad.gateGrad + frameSize * 2,
-            frameSize * 3, 1, grad.stateWeightGrad, frameSize);
+            context, true, false, frame_size, frame_size, batch_size, 1,
+            value.reset_output_value, frame_size,
+            grad.gate_grad + frame_size * 2, frame_size * 3, 1,
+            grad.state_weight_grad, frame_size);
       }
     }
 
-    if (batchSize == 1) {
+    if (batch_size == 1) {
       detail::KeGruBackwardResetGrad<
           detail::backward::gru_resetGrad<T>,
-          /* isBatch= */ false><<<grid, threads, 0, stream>>>(
-          detail::backward::gru_resetGrad<T>(), value.gateValue, grad.gateGrad,
-          value.prevOutValue, grad.prevOutGrad, grad.resetOutputGrad, frameSize,
-          batchSize, active_gate);
+          /* is_batch= */ false><<<grid, threads, 0, stream>>>(
+          detail::backward::gru_resetGrad<T>(), value.gate_value,
+          grad.gate_grad, value.prev_out_value, grad.prev_out_grad,
+          grad.reset_output_grad, frame_size, batch_size, active_gate);
     } else {
       detail::KeGruBackwardResetGrad<
           detail::backward::gru_resetGrad<T>,
-          /* isBatch= */ true><<<grid, threads, 0, stream>>>(
-          detail::backward::gru_resetGrad<T>(), value.gateValue, grad.gateGrad,
-          value.prevOutValue, grad.prevOutGrad, grad.resetOutputGrad, frameSize,
-          batchSize, active_gate);
+          /* is_batch= */ true><<<grid, threads, 0, stream>>>(
+          detail::backward::gru_resetGrad<T>(), value.gate_value,
+          grad.gate_grad, value.prev_out_value, grad.prev_out_grad,
+          grad.reset_output_grad, frame_size, batch_size, active_gate);
     }
 
-    if (grad.prevOutGrad && value.prevOutValue) {
+    if (grad.prev_out_grad && value.prev_out_value) {
       math::gemm<platform::GPUPlace, T>(
-          context, false, true, batchSize, frameSize, frameSize * 2, 1,
-          grad.gateGrad, frameSize * 3, value.gateWeight, frameSize * 2, 1,
-          grad.prevOutGrad, frameSize);
+          context, false, true, batch_size, frame_size, frame_size * 2, 1,
+          grad.gate_grad, frame_size * 3, value.gate_weight, frame_size * 2, 1,
+          grad.prev_out_grad, frame_size);
 
-      if (grad.gateWeightGrad) {
+      if (grad.gate_weight_grad) {
         math::gemm<platform::GPUPlace, T>(
-            context, true, false, frameSize, frameSize * 2, batchSize, 1,
-            value.prevOutValue, frameSize, grad.gateGrad, frameSize * 3, 1,
-            grad.gateWeightGrad, frameSize * 2);
+            context, true, false, frame_size, frame_size * 2, batch_size, 1,
+            value.prev_out_value, frame_size, grad.gate_grad, frame_size * 3, 1,
+            grad.gate_weight_grad, frame_size * 2);
       }
     }
   }

paddle/operators/math/gru_compute.h

@@ -22,28 +22,28 @@ namespace math {
 // TODO(guosheng): refine code style in gru_compute
 template <typename T>
 struct hl_gru_value {
-  T *gateWeight;
-  T *stateWeight;
-  T *gateValue;
-  T *resetOutputValue;
-  T *outputValue;
-  T *prevOutValue;
+  T *gate_weight;
+  T *state_weight;
+  T *gate_value;
+  T *reset_output_value;
+  T *output_value;
+  T *prev_out_value;
 };
 
 template <typename T>
 struct hl_gru_grad {
-  T *gateWeightGrad;
-  T *stateWeightGrad;
-  T *gateGrad;
-  T *resetOutputGrad;
-  T *outputGrad;
-  T *prevOutGrad;
+  T *gate_weight_grad;
+  T *state_weight_grad;
+  T *gate_grad;
+  T *reset_output_grad;
+  T *output_grad;
+  T *prev_out_grad;
 };
 
 template <typename Place, typename T>
 struct GRUUnitFunctor {
   static void compute(const platform::DeviceContext &context,
-                      hl_gru_value<T> value, int frameSize, int batchSize,
+                      hl_gru_value<T> value, int frame_size, int batch_size,
                       activation_mode_t active_node,
                       activation_mode_t active_gate);
 };
@@ -51,8 +51,9 @@ struct GRUUnitFunctor {
 template <typename Place, typename T>
 struct GRUUnitGradFunctor {
   static void compute(const platform::DeviceContext &context,
-                      hl_gru_value<T> value, hl_gru_grad<T> grad, int frameSize,
-                      int batchSize, activation_mode_t active_node,
+                      hl_gru_value<T> value, hl_gru_grad<T> grad,
+                      int frame_size, int batch_size,
+                      activation_mode_t active_node,
                       activation_mode_t active_gate);
 };
 

paddle/operators/pool_op.cc

@@ -105,7 +105,7 @@ Pool2dOpMaker::Pool2dOpMaker(framework::OpProto *proto,
   // TypedAttrChecker don't support vector type.)
   AddAttr<std::vector<int>>(
       "paddings",
-      "(vector<int>, defalut {0,0}), paddings(height, width) of pooling "
+      "(vector<int>, default {0,0}), paddings(height, width) of pooling "
       "operator."
       "If global_pooling = true, paddings and ksize will be ignored.")
       .SetDefault({0, 0});  // TODO(Chengduo): Add checker. (Currently,
@@ -127,10 +127,10 @@ Example:
        X shape: $(N, C, H_{in}, W_{in})$
   Output:
        Out shape: $(N, C, H_{out}, W_{out})$
-  where 
+  Where
        $$ 
-       H_{out} = (H_{in} - ksize[0] + 2 * paddings[0]) / strides[0] + 1 \\
-       W_{out} = (W_{in} - ksize[1] + 2 * paddings[1]) / strides[1] + 1
+       H_{out} = \frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\
+       W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1
        $$
 
 )DOC");
@@ -177,7 +177,7 @@ Pool3dOpMaker::Pool3dOpMaker(framework::OpProto *proto,
                                // TypedAttrChecker don't support vector type.)
   AddAttr<std::vector<int>>(
       "paddings",
-      "(vector<int>, defalut {0,0,0}), paddings(depth, height, "
+      "(vector<int>, default {0,0,0}), paddings(depth, height, "
       "width) of pooling operator. "
       "If global_pooling = true, ksize and paddings will be ignored.")
       .SetDefault({0, 0, 0});  // TODO(Chengduo): Add checker. (Currently,
@@ -199,11 +199,11 @@ Example:
        X shape: $(N, C, D_{in}, H_{in}, W_{in})$
   Output:
        Out shape: $(N, C, D_{out}, H_{out}, W_{out})$
-  where
+  Where
   $$
-       D_{out} = (D_{in} - ksize[0] + 2 * paddings[0]) / strides[0] + 1 \\
-       H_{out} = (H_{in} - ksize[1] + 2 * paddings[1]) / strides[1] + 1 \\
-       W_{out} = (W_{in} - ksize[2] + 2 * paddings[2]) / strides[2] + 1
+       D_{out} = \frac{(D_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\
+       H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 \\
+       W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1
   $$
 
 )DOC");

paddle/operators/pool_with_index_op.cc

@@ -142,7 +142,7 @@ class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
     // TypedAttrChecker don't support vector type.)
     AddAttr<std::vector<int>>(
         "paddings",
-        "(vector<int>, defalut:{0, 0}), paddings(height, width) of pooling "
+        "(vector<int>, default:{0, 0}), paddings(height, width) of pooling "
         "operator. "
         "If global_pooling = true, paddings and will be ignored.")
         .SetDefault({0, 0});  // TODO(Chengduo): Add checker. (Currently,
@@ -166,10 +166,10 @@ Example:
   Output:
        Out shape: $(N, C, H_{out}, W_{out})$
        Mask shape: $(N, C, H_{out}, W_{out})$
-  where
+  Where
        $$
-       H_{out} = (H_{in} - ksize[0] + 2 * paddings[0]) / strides[0] + 1 \\
-       W_{out} = (W_{in} - ksize[1] + 2 * paddings[1]) / strides[1] + 1
+       H_{out} = \frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\
+       W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1
        $$
 
 )DOC");
@@ -220,7 +220,7 @@ class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
     // TypedAttrChecker don't support vector type.)
     AddAttr<std::vector<int>>(
         "paddings",
-        "(vector, defalut {0,0,0}), paddings(depth, "
+        "(vector, default {0,0,0}), paddings(depth, "
         "height, width) of pooling operator. "
         "If global_pooling = true, paddings and ksize will be ignored.")
         .SetDefault({0, 0, 0});  // TODO(Chengduo): Add checker. (Currently,
@@ -244,11 +244,11 @@ Example:
   Output:
        Out shape: $(N, C, D_{out}, H_{out}, W_{out})$
        Mask shape: $(N, C, D_{out}, H_{out}, W_{out})$
-  where
+  Where
        $$
-       D_{out} = (D_{in} - ksize[0] + 2 * paddings[0]) / strides[0] + 1 \\
-       H_{out} = (H_{in} - ksize[1] + 2 * paddings[1]) / strides[1] + 1 \\
-       W_{out} = (W_{in} - ksize[2] + 2 * paddings[2]) / strides[2] + 1
+       D_{out} = \frac{(D_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\
+       H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 \\
+       W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1
        $$
 
 )DOC");

paddle/operators/recv_op.cc

+/* 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 <stdint.h>
+#include <sys/stat.h>
+#include <ostream>
+#include <thread>
+
+#include <unistd.h>
+
+#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<detail::SendRecvServerImpl> service,
+               const std::string &server_address) {
+  ServerBuilder builder;
+  builder.AddListeningPort(server_address, grpc::InsecureServerCredentials());
+  builder.RegisterService(service.get());
+  std::unique_ptr<Server> 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<std::string>("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<framework::LoDTensor>();
+    // FIXME(typhoonzero): do not copy
+    framework::CopyFrom(t, dev_ctx.GetPlace(), dev_ctx, tensor);
+
+    auto *block = Attr<framework::BlockDescBind *>("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<framework::LoDTensor>());
+  }
+
+ 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<detail::SendRecvServerImpl> rpc_service_;
+  std::shared_ptr<std::thread> 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<std::string>("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<framework::BlockDescBind *>("OptimizeBlock", "type BlockDescBind*",
+                                        "optimize network run in server");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(recv, ops::RecvOp, ops::RecvOpMaker);

paddle/operators/roi_pool_op.h

@@ -133,53 +133,47 @@ class CPUROIPoolGradOpKernel : public framework::OpKernel<T> {
     auto* in = ctx.Input<framework::Tensor>("X");
     auto* rois = ctx.Input<framework::Tensor>("ROIs");
     auto* argmax = ctx.Input<framework::Tensor>("Argmax");
-
     auto* out_grad =
         ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
-    auto* x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto* in_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
 
     auto pooled_height = ctx.Attr<int>("pooled_height");
     auto pooled_width = ctx.Attr<int>("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<int64_t>();
-      int rois_num = rois->dims()[0];
-
-      T* x_grad_data = x_grad->mutable_data<T>(ctx.GetPlace());
+      const T* out_grad_data = out_grad->data<T>();
+      const int64_t* argmax_data = argmax->data<int64_t>();
+      T* in_grad_data = in_grad->mutable_data<T>(ctx.GetPlace());
       math::SetConstant<Place, T> set_zero;
-      set_zero(ctx.device_context(), x_grad, static_cast<T>(0));
+      set_zero(ctx.device_context(), in_grad, static_cast<T>(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<T>();
-      size_t pool_channel_offset = pooled_height * pooled_width;
-      const int64_t* argmax_data = argmax->data<int64_t>();
+      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<size_t>(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];
       }
     }
   }

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<framework::LoDTensor>();
-
-    {  // the 1st field, uint32_t version
-      constexpr uint32_t version = 0;
-      fout.write(reinterpret_cast<const char *>(&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<int>(dims.size()), 0);
-      std::copy(dims.begin(), dims.end(), pb_dims->begin());
-      int32_t size = desc.ByteSize();
-      fout.write(reinterpret_cast<const char *>(&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<void>();
-      PADDLE_ENFORCE(size < std::numeric_limits<std::streamsize>::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<char[]> buf(new char[kBufSize]);
-        auto &gpu_dev_ctx =
-            static_cast<const platform::CUDADeviceContext &>(dev_ctx);
-        platform::CPUPlace cpu;
-        uintptr_t data = reinterpret_cast<uintptr_t>(data_ptr);
-        while (size != 0) {
-          size_t size_to_write = std::min(kBufSize, static_cast<size_t>(size));
-          memory::Copy(cpu, buf.get(),
-                       boost::get<platform::GPUPlace>(tensor.place()),
-                       reinterpret_cast<const void *>(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<const char *>(data_ptr),
-                   static_cast<std::streamsize>(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<const char *>(&size), sizeof(size));
-
-      for (auto &each : lod) {
-        size = each.size() * sizeof(framework::LoD::value_type::value_type);
-        fout.write(reinterpret_cast<const char *>(&size), sizeof(size));
-        fout.write(reinterpret_cast<const char *>(each.data()),
-                   static_cast<std::streamsize>(size));
-      }
-    }
+    framework::SerializeToStream(fout, tensor, dev_ctx);
   }
 };
 

paddle/operators/scale_op.cc

@@ -77,4 +77,6 @@ REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker<float>,
                   ops::ScaleGradMaker);
 REGISTER_OP_CPU_KERNEL(scale,
                        ops::ScaleKernel<paddle::platform::CPUPlace, float>,
-                       ops::ScaleKernel<paddle::platform::CPUPlace, double>);
+                       ops::ScaleKernel<paddle::platform::CPUPlace, double>,
+                       ops::ScaleKernel<paddle::platform::CPUPlace, int>,
+                       ops::ScaleKernel<paddle::platform::CPUPlace, int64_t>);

paddle/operators/scale_op.cu

@@ -16,4 +16,6 @@
 
 REGISTER_OP_GPU_KERNEL(
     scale, paddle::operators::ScaleKernel<paddle::platform::GPUPlace, float>,
-    paddle::operators::ScaleKernel<paddle::platform::GPUPlace, double>);
+    paddle::operators::ScaleKernel<paddle::platform::GPUPlace, double>,
+    paddle::operators::ScaleKernel<paddle::platform::GPUPlace, int>,
+    paddle::operators::ScaleKernel<paddle::platform::GPUPlace, int64_t>);

paddle/operators/send_op.cc

+/* 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 <ostream>
+
+#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<std::string>("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<detail::RPCClient> 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<std::string>("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);

paddle/operators/send_recv_op_test.cc

+/* 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 <unistd.h>
+#include <thread>
+
+#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<paddle::framework::OperatorBase> 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<paddle::framework::LoDTensor>();
+  tensor->Resize({10, 10});
+  float *expect = tensor->mutable_data<float>(place);
+  for (int64_t i = 0; i < tensor->numel(); ++i) {
+    expect[i] = static_cast<float>(i);
+  }
+
+  auto out_var = scope.Var("Out");
+  auto out_tensor = out_var->GetMutable<paddle::framework::LoDTensor>();
+  out_tensor->Resize({10, 10});
+  tensor->mutable_data<float>(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<paddle::framework::LoDTensor>();
+  float *expected = tensor->data<float>();
+
+  auto out_var = scope.Var("Out");
+  auto target = out_var->GetMutable<paddle::framework::LoDTensor>();
+  // send fail cause output is none.
+  EXPECT_NE(target->memory_size(), size_t(0));
+  float *actual = target->data<float>();
+  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();
+}

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");
   }

paddle/platform/cuda_profiler.h

+/* 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 <cuda_profiler_api.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+namespace paddle {
+namespace platform {
+
+void CudaProfilerInit(std::string output_file, std::string output_mode,
+                      std::vector<std::string> config_flags) {
+  std::array<char, 128> 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

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

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

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();

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"

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`

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 {

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):

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)

python/paddle/trainer_config_helpers/tests/configs/protostr/test_factorization_machine.protostr

+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
+}
+

python/paddle/trainer_config_helpers/tests/configs/test_factorization_machine.py

+from paddle.trainer_config_helpers import *
+
+data = data_layer(name='data', size=1024)
+
+fm = factorization_machine(input=data, factor_size=10)
+
+outputs(fm)

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:

python/paddle/v2/fluid/__init__.py

@@ -13,13 +13,14 @@ import nets
 import optimizer
 import backward
 import regularizer
+from param_attr import ParamAttr
 
 from core import LoDTensor, CPUPlace, GPUPlace
 
 Tensor = LoDTensor
 __all__ = framework.__all__ + executor.__all__ + [
     'io', 'initializer', 'layers', 'nets', 'optimizer', 'backward',
-    'regularizer', 'LoDTensor', 'CPUPlace', 'GPUPlace', 'Tensor'
+    'regularizer', 'LoDTensor', 'CPUPlace', 'GPUPlace', 'Tensor', 'ParamAttr'
 ]
 
 

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 

python/paddle/v2/fluid/executor.py

 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()

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_

python/paddle/v2/fluid/io.py

 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)

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