Merge remote-tracking branch 'upstream/develop' into inference

.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 HAS_SYSTEM_PROTOBUF=false -s -j8 static grpc_cpp_plugin | sed "s/-Werror//g" | sh)
+ELSE()
+  SET(BUILD_CMD make HAS_SYSTEM_PROTOBUF=false -s -j8 static grpc_cpp_plugin)
+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  ${BUILD_CMD}
+    INSTALL_COMMAND make prefix=${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

@@ -459,11 +459,58 @@ function(py_test TARGET_NAME)
   if(WITH_TESTING)
     set(options STATIC static SHARED shared)
     set(oneValueArgs "")
-    set(multiValueArgs SRCS DEPS)
-    cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})  
+    set(multiValueArgs SRCS DEPS ARGS)
+    cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     add_test(NAME ${TARGET_NAME}
              COMMAND env PYTHONPATH=${PADDLE_PYTHON_BUILD_DIR}/lib-python
-             ${PYTHON_EXECUTABLE} ${py_test_SRCS}
+             ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
              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

@@ -54,7 +54,7 @@ img_conv
 
 ..  _api_v2.layer_context_projection:
 
-context_projection 
+context_projection
 ------------------
 ..  autoclass:: paddle.v2.layer.context_projection
     :noindex:
@@ -70,7 +70,7 @@ Image Pooling Layer
 img_pool
 --------
 ..  autoclass:: paddle.v2.layer.img_pool
-    :noindex:   
+    :noindex:
 
 spp
 ---
@@ -104,7 +104,7 @@ sum_to_one_norm
 ---------------
 ..  autoclass:: paddle.v2.layer.sum_to_one_norm
     :noindex:
-    
+
 cross_channel_norm
 ------------------
 ..  autoclass:: paddle.v2.layer.cross_channel_norm
@@ -114,7 +114,7 @@ row_l2_norm
 -----------
 ..  autoclass:: paddle.v2.layer.row_l2_norm
     :noindex:
-    
+
 Recurrent Layers
 ================
 
@@ -415,6 +415,13 @@ multiplex
 ..  autoclass:: paddle.v2.layer.multiplex
     :noindex:
 
+Factorization Machine Layer
+============================
+
+factorization_machine
+---------------------
+..  autoclass:: paddle.v2.layer.factorization_machine
+    :noindex:
 
 Slicing and Joining Layers
 ==========================

doc/design/float16.md

@@ -28,6 +28,51 @@ The goal of float16 is to serve as a key for the executor to find and run the co
 - [Eigen](https://github.com/RLovelett/eigen) >= 3.3 supports float16 calculation on both GPU and CPU using the `Eigen::half` class. It is mostly useful for Nvidia GPUs because of the overloaded arithmetic operators using cuda intrinsics. It falls back to using software emulation on CPU for calculation and there is no special treatment to ARM processors.
 - [ARM compute library](https://github.com/ARM-software/ComputeLibrary) >= 17.02.01 supports NEON FP16 kernels (requires ARMv8.2-A CPU).
 
+### CUDA version issue
+There are currently three versions of CUDA that supports `__half` data type, namely, CUDA 7.5, 8.0, and 9.0. 
+CUDA 7.5 and 8.0 define `__half` as a simple struct that has a `uint16_t` data (see [`cuda_fp16.h`](https://github.com/ptillet/isaac/blob/9212ab5a3ddbe48f30ef373f9c1fb546804c7a8c/include/isaac/external/CUDA/cuda_fp16.h)) as follows:
+```
+typedef struct __align__(2) {
+   unsigned short x;
+} __half;
+
+typedef __half half;
+```
+This struct does not define any overloaded arithmetic operators. So you have to directly use `__hadd` instead of `+` to correctly add two half types:
+```
+__global__ void Add() {
+  half a, b, c;
+  c = __hadd(a, b); // correct
+  c = a + b; // compiler error: no operator "+" matches these operands
+}
+```
+CUDA 9.0 provides a major update to the half data type. The related code can be found in the updated [`cuda_fp16.h`](https://github.com/ptillet/isaac/blob/master/include/isaac/external/CUDA/cuda_fp16.h) and the newly added [`cuda_fp16.hpp`](https://github.com/ptillet/isaac/blob/master/include/isaac/external/CUDA/cuda_fp16.hpp).
+
+Essentially, CUDA 9.0 renames the original `__half` type in 7.5 and 8.0 as `__half_raw`, and defines a new `__half` class type that has constructors, conversion operators, and also provides overloaded arithmetic operators such as follows:
+```
+typedef struct __CUDA_ALIGN__(2) {
+    unsigned short x;
+} __half_raw;
+
+
+struct __CUDA_ALIGN__(2) __half {
+protected:
+    unsigned short __x;
+public:
+    // constructors and conversion operators from/to 
+    // __half_raw and other built-in data types
+}
+
+typedef __half half;
+
+__device__ __forceinline__ 
+__half operator+(const __half &lh, const __half &rh) { 
+    return __hadd(lh, rh); 
+}
+
+// Other overloaded operators
+``` 
+This new design makes `c = a + b` work correctly for CUDA half data type. 
 
 ## Implementation
 The float16 class holds a 16-bit `uint16_t` data internally.

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.
+
+7. PaddlePaddle runtimes with the `fetch` OP reports evaluation results back to the converter, the converter reports the evaluation results back to the PaddlePaddle Python.
 
-1. PaddlePaddle runtimes with the `fetch` OP reports evaluation
-   results back to the converter, the convert reports the evaluation
-   results back to the PaddlePaddle Python.
-   
 The output IRs will be cached to optimize the conversion latency.
 
 
 #### Placement Algorithm
 
-Our first implementation will only support "trainer-parameter server"
-placement: the parameters, initializers, and optimizers are placed on
-the PaddlePaddle runtimes with the parameter server role. And
-everything else will be placed on the PaddlePaddle runtimes with the
-trainer role. This has the same functionality of our
-"trainer-parameter server" architecture of PaddlePaddle v0.10.0, but
-is more general and flexible.
+Our first implementation will only support "trainer-parameter server" placement: the parameters, initializers, and optimizers are all placed on the PaddlePaddle runtimes with the parameter server role. Everything else will be placed on the PaddlePaddle runtimes with the trainer role. This has the same functionality as the "trainer-parameter server" architecture of PaddlePaddle v0.10.0, but is more generic and flexible.
 
-In the future, we will implement the general placement algorithm,
-which makes placements according to the input IR, and a model of
-device computation time and device communication time. Model
-parallelism requires the general placement algorithm.
+In the future, a more general placement algorithm should be implemented, which makes placements according to the input IR, and a model of device computation time and device communication time. Model parallelism requires the generic placement algorithm.
 
 
 ### PaddlePaddle Runtime
 
-The PaddlePaddle runtime owns multiple devices (e.g., CPUs, GPUs) and
-runs the IR. The runtime does not need to do OP placement since it's
-already done by the converter.
+The PaddlePaddle runtime owns multiple devices (e.g., CPUs, GPUs) and runs the IR. The runtime does not need to do OP placement since it is already done by the converter.
 
 
 ### Local Training Architecture
 
-The local training architecture will be the same as the distributed
-training architecture, the differences are everything runs locally,
-and there is just one PaddlePaddle runtime:
+The local training architecture will be the same as the distributed training architecture, the difference is that everything runs locally, and there is just one PaddlePaddle runtime:
 
 <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/basic_usage/index_cn.rst

-经典的线性回归任务
-==================
-
-PaddlePaddle是源于百度的一个深度学习平台。这份简短的介绍将向你展示如何利用PaddlePaddle来解决一个经典的线性回归问题。
-
-任务简介
---------
-
-我们展示如何用PaddlePaddle解决 `单变量的线性回归 <https://www.baidu.com/s?wd=单变量线性回归>`_ 问题。线性回归的输入是一批点 `(x, y)` ，其中 `y = wx + b + ε`， 而 ε 是一个符合高斯分布的随机变量。线性回归的输出是从这批点估计出来的参数 `w` 和 `b` 。
-
-一个例子是房产估值。我们假设房产的价格（y）是其大小（x）的一个线性函数，那么我们可以通过收集市场上房子的大小和价格，用来估计线性函数的参数w 和 b。
-
-准备数据
------------
-
-假设变量 `x` 和 `y` 的真实关系为： `y = 2x + 0.3 + ε`，这里展示如何使用观测数据来拟合这一线性关系。首先，Python代码将随机产生2000个观测点，作为线性回归的输入。下面脚本符合PaddlePaddle期待的读取数据的Python程序的模式。
-
-.. code-block:: python
-
-    # dataprovider.py
-    from paddle.trainer.PyDataProvider2 import *
-    import random
-
-    # 定义输入数据的类型: 2个浮点数
-    @provider(input_types=[dense_vector(1), dense_vector(1)],use_seq=False)
-    def process(settings, input_file):
-        for i in xrange(2000):
-            x = random.random()
-            yield [x], [2*x+0.3]
-
-训练模型
------------
-
-为了还原 `y = 2x + 0.3`，我们先从一条随机的直线 `y' = wx + b` 开始，然后利用观测数据调整 `w` 和 `b` 使得 `y'` 和 `y` 的差距不断减小，最终趋于接近。这个过程就是模型的训练过程，而 `w` 和 `b` 就是模型的参数，即我们的训练目标。
-
-在PaddlePaddle里，该模型的网络配置如下。
-
-.. code-block:: python
-
-    # trainer_config.py
-    from paddle.trainer_config_helpers import *
-
-    # 1. 定义数据来源，调用上面的process函数获得观测数据
-    data_file = 'empty.list'
-    with open(data_file, 'w') as f: f.writelines(' ')
-    define_py_data_sources2(train_list=data_file, test_list=None, 
-                            module='dataprovider', obj='process',args={})
-
-    # 2. 学习算法。控制如何改变模型参数 w 和 b
-    settings(batch_size=12, learning_rate=1e-3, learning_method=MomentumOptimizer())
-
-    # 3. 神经网络配置
-    x = data_layer(name='x', size=1)
-    y = data_layer(name='y', size=1)
-    # 线性计算网络层: ȳ = wx + b
-    ȳ = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b'))
-    # 计算误差函数，即  ȳ 和真实 y 之间的距离
-    cost = square_error_cost(input= ȳ, label=y)
-    outputs(cost)
-
-
-这段简短的配置展示了PaddlePaddle的基本用法：
-
-- 第一部分定义了数据输入。一般情况下，PaddlePaddle先从一个文件列表里获得数据文件地址，然后交给用户自定义的函数（例如上面的 `process`函数）进行读入和预处理从而得到真实输入。本文中由于输入数据是随机生成的不需要读输入文件，所以放一个空列表（`empty.list`）即可。
-
-- 第二部分主要是选择学习算法，它定义了模型参数改变的规则。PaddlePaddle提供了很多优秀的学习算法，这里使用一个基于momentum的随机梯度下降(SGD)算法，该算法每批量(batch)读取12个采样数据进行随机梯度计算来更新更新。
-
-- 最后一部分是神经网络的配置。由于PaddlePaddle已经实现了丰富的网络层，所以很多时候你需要做的只是定义正确的网络层并把它们连接起来。这里使用了三种网络单元：
-    
-    - **数据层**：数据层 `data_layer` 是神经网络的入口，它读入数据并将它们传输到接下来的网络层。这里数据层有两个，分别对应于变量 `x` 和 `y`。
-    - **全连接层**：全连接层 `fc_layer` 是基础的计算单元，这里利用它建模变量之间的线性关系。计算单元是神经网络的核心，PaddlePaddle支持大量的计算单元和任意深度的网络连接，从而可以拟合任意的函数来学习复杂的数据关系。
-    - **回归误差代价层**：回归误差代价层 `square_error_cost` 是众多误差代价函数层的一种，它们在训练过程作为网络的出口，用来计算模型的误差，是模型参数优化的目标函数。
-
-定义了网络结构并保存为 `trainer_config.py` 之后，运行以下训练命令：
-
-.. code-block:: bash
-
-    paddle train --config=trainer_config.py --save_dir=./output --num_passes=30
-
-PaddlePaddle将在观测数据集上迭代训练30轮，并将每轮的模型结果存放在 `./output` 路径下。从输出日志可以看到，随着轮数增加误差代价函数的输出在不断的减小，这意味着模型在训练数据上不断的改进，直到逼近真实解：` y = 2x + 0.3 `
-
-模型检验
------------
-
-训练完成后，我们希望能够检验模型的好坏。一种常用的做法是用学习的模型对另外一组测试数据进行预测，评价预测的效果。在这个例子中，由于已经知道了真实答案，我们可以直接观察模型的参数是否符合预期来进行检验。
-
-PaddlePaddle将每个模型参数作为一个numpy数组单独存为一个文件，所以可以利用如下方法读取模型的参数。
-
-.. code-block:: python
-
-    import numpy as np
-    import os
-
-    def load(file_name):
-        with open(file_name, 'rb') as f:
-            f.read(16) # skip header for float type.
-            return np.fromfile(f, dtype=np.float32)
-        
-    print 'w=%.6f, b=%.6f' % (load('output/pass-00029/w'), load('output/pass-00029/b'))
-    # w=1.999743, b=0.300137
-
-.. image:: ./parameters.png
-     :align: center
-     :scale: 80 %
-
-从图中可以看到，虽然 `w` 和 `b` 都使用随机值初始化，但在起初的几轮训练中它们都在快速逼近真实值，并且后续仍在不断改进，使得最终得到的模型几乎与真实模型一致。
-
-这样，我们用PaddlePaddle解决了单变量线性回归问题， 包括数据输入、模型训练和最后的结果验证。

doc/getstarted/basic_usage/index_en.rst

-Simple Linear Regression
-========================
-
-PaddlePaddle is a deep learning platform open-sourced by Baidu. With PaddlePaddle, you can easily train a classic neural network within a couple lines of configuration, or you can build sophisticated models that provide state-of-the-art performance on difficult learning tasks like sentiment analysis, machine translation, image caption and so on.
-
-Problem Background
-------------------
-
-Now, to give you a hint of what using PaddlePaddle looks like, let's start with a fundamental learning problem - `simple linear regression <https://en.wikipedia.org/wiki/Simple_linear_regression>`_: you have observed a set of two-dimensional data points of ``X`` and ``Y``, where ``X`` is an explanatory variable and ``Y`` is corresponding dependent variable, and you want to recover the underlying correlation between ``X`` and ``Y``. Linear regression can be used in many practical scenarios. For example, ``X`` can be a variable about house size, and ``Y`` a variable about house price. You can build a model that captures relationship between them by observing real estate markets.
-
-Prepare the Data
------------------
-
-Suppose the true relationship can be characterized as ``Y = 2X + 0.3``, let's see how to recover this pattern only from observed data. Here is a piece of python code that feeds synthetic data to PaddlePaddle. The code is pretty self-explanatory, the only extra thing you need to add for PaddlePaddle is a definition of input data types.
-
-    .. code-block:: python
-
-        # dataprovider.py
-        from paddle.trainer.PyDataProvider2 import *
-        import random
-
-        # define data types of input: 2 real numbers
-        @provider(input_types=[dense_vector(1), dense_vector(1)],use_seq=False)
-        def process(settings, input_file):
-            for i in xrange(2000):
-                x = random.random()
-                yield [x], [2*x+0.3]
-
-Train a NeuralNetwork
-----------------------
-
-To recover this relationship between ``X`` and ``Y``, we use a neural network with one layer of linear activation units and a square error cost layer. Don't worry if you are not familiar with these terminologies, it's just saying that we are starting from a random line ``Y' = wX + b`` , then we gradually adapt ``w`` and ``b`` to minimize the difference between ``Y'`` and ``Y``. Here is what it looks like in PaddlePaddle:
-
-    .. code-block:: python
-
-        # trainer_config.py
-        from paddle.trainer_config_helpers import *
-
-        # 1. read data. Suppose you saved above python code as dataprovider.py
-        data_file = 'empty.list'
-        with open(data_file, 'w') as f: f.writelines(' ')
-        define_py_data_sources2(train_list=data_file, test_list=None, 
-                module='dataprovider', obj='process',args={})
-
-        # 2. learning algorithm
-        settings(batch_size=12, learning_rate=1e-3, learning_method=MomentumOptimizer())
-
-        # 3. Network configuration
-        x = data_layer(name='x', size=1)
-        y = data_layer(name='y', size=1)
-        y_predict = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b'))
-        cost = square_error_cost(input=y_predict, label=y)
-        outputs(cost)
-
-Some of the most fundamental usages of PaddlePaddle are demonstrated:
-
--  The first part shows how to feed data into PaddlePaddle. In general cases, PaddlePaddle reads raw data from a list of files, and then do some user-defined process to get real input. In this case, we only need to create a placeholder file since we are generating synthetic data on the fly.
-
--  The second part describes learning algorithm. It defines in what ways adjustments are made to model parameters. PaddlePaddle provides a rich set of optimizers, but a simple momentum based optimizer will suffice here, and it processes 12 data points each time.
-
--  Finally, the network configuration. It usually is as simple as "stacking" layers. Three kinds of layers are used in this configuration:
-	-  **Data Layer**: a network always starts with one or more data layers. They provide input data to the rest of the network. In this problem, two data layers are used respectively for ``X`` and ``Y``.
-	-  **FC Layer**: FC layer is short for Fully Connected Layer, which connects all the input units to current layer and does the actual computation specified as activation function. Computation layers like this are the fundamental building blocks of a deeper model.
-	-  **Cost Layer**: in training phase, cost layers are usually the last layers of the network. They measure the performance of current model, and provide guidence to adjust parameters.
-
-Now that everything is ready, you can train the network with a simple command line call:
-
-    .. code-block:: bash
- 
-        paddle train --config=trainer_config.py --save_dir=./output --num_passes=30
- 
-
-This means that PaddlePaddle will train this network on the synthectic dataset for 30 passes, and save all the models under path ``./output``. You will see from the messages printed out during training phase that the model cost is decreasing as time goes by, which indicates we are getting a closer guess.
-
-
-Evaluate the Model
--------------------
-
-Usually, a different dataset that left out during training phase should be used to evalute the models. However, we are lucky enough to know the real answer: ``w=2, b=0.3``, thus a better option is to check out model parameters directly.
-
-In PaddlePaddle, training is just to get a collection of model parameters, which are ``w`` and ``b`` in this case. Each parameter is saved in an individual file in the popular ``numpy`` array format. Here is the code that reads parameters from last pass.
-
-    .. code-block:: python
-
-        import numpy as np
-        import os
-
-        def load(file_name):
-            with open(file_name, 'rb') as f:
-                f.read(16) # skip header for float type.
-                return np.fromfile(f, dtype=np.float32)
-                
-        print 'w=%.6f, b=%.6f' % (load('output/pass-00029/w'), load('output/pass-00029/b'))
-        # w=1.999743, b=0.300137
-
-    .. image:: parameters.png
-        :align: center
-
-Although starts from a random guess, you can see that value of ``w`` changes quickly towards 2 and ``b`` changes quickly towards 0.3. In the end, the predicted line is almost identical with real answer.
-
-There, you have recovered the underlying pattern between ``X`` and ``Y`` only from observed data.

doc/getstarted/build_and_install/build_from_source_cn.rst

+从源码编译
+======================
+
+.. _build_step:
+
+编译方法
+----------------
+
+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 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
+
+   git clone https://github.com/PaddlePaddle/Paddle.git
+   cd Paddle
+   # 如果使用Docker编译环境，执行下面的命令编译CPU-Only的二进制
+   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
+   # 如果不使用Docker编译环境，执行下面的命令
+   mkdir build
+   cd build
+   cmake -DWITH_GPU=OFF -DWITH_TESTING=OFF ..
+   make
+
+编译完成后会在build/python/dist目录下生成输出的whl包，可以选在在当前机器安装也可以拷贝到目标机器安装：
+
+.. code-block:: bash
+
+   pip install python/dist/*.whl
+
+
+.. _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
+----------------
+
+PaddlePaddle编译需要使用到下面的依赖（包含但不限于），其他的依赖软件，会自动在编译时下载。
+
+.. csv-table:: PaddlePaddle编译依赖
+   :header: "依赖", "版本", "说明"
+   :widths: 10, 15, 30
+
+   "CMake", ">=3.5", ""
+   "GCC", "4.8.2", "推荐使用CentOS的devtools2"
+   "Python", "2.7.x", "依赖libpython2.7.so"
+   "pip", ">=9.0", ""
+   "numpy", "", ""
+   "SWIG", ">=2.0", ""
+   "Go", ">=1.8", "可选"
+
+
+.. _build_options:
+
+编译选项
+----------------
+
+PaddlePaddle的编译选项，包括生成CPU/GPU二进制文件、链接何种BLAS库等。
+用户可在调用cmake的时候设置它们，详细的cmake使用方法可以参考
+`官方文档 <https://cmake.org/cmake-tutorial>`_ 。
+
+在cmake的命令行中，通过使用 ``-D`` 命令设置该类编译选项，例如：
+
+..  code-block:: bash
+
+    cmake .. -DWITH_GPU=OFF
+
+..  csv-table:: 编译选项说明
+    :header: "选项", "说明", "默认值"
+    :widths: 1, 7, 2
+
+    "WITH_GPU", "是否支持GPU", "ON"
+    "WITH_C_API", "是否仅编译CAPI", "OFF"
+    "WITH_DOUBLE", "是否使用双精度浮点数", "OFF"
+    "WITH_DSO", "是否运行时动态加载CUDA动态库，而非静态加载CUDA动态库。", "ON"
+    "WITH_AVX", "是否编译含有AVX指令集的PaddlePaddle二进制文件", "ON"
+    "WITH_PYTHON", "是否内嵌PYTHON解释器", "ON"
+    "WITH_STYLE_CHECK", "是否编译时进行代码风格检查", "ON"
+    "WITH_TESTING", "是否开启单元测试", "ON"
+    "WITH_DOC", "是否编译中英文文档", "OFF"
+    "WITH_SWIG_PY", "是否编译PYTHON的SWIG接口，该接口可用于预测和定制化训练", "Auto"
+    "WITH_GOLANG", "是否编译go语言的可容错parameter server", "ON"
+    "WITH_MKL", "是否使用MKL数学库，如果为否则是用OpenBLAS", "ON"
+
+BLAS
++++++
+
+PaddlePaddle支持 `MKL <https://software.intel.com/en-us/intel-mkl>`_ 和
+`OpenBlAS <http://www.openblas.net/>`_ 两种BLAS库。默认使用MKL。如果使用MKL并且机器含有AVX2指令集，
+还会下载MKL-DNN数学库，详细参考 `这里 <https://github.com/PaddlePaddle/Paddle/tree/develop/doc/design/mkldnn#cmake>`_ 。
+
+如果关闭MKL，则会使用OpenBLAS作为BLAS库。
+
+CUDA/cuDNN
++++++++++++
+
+PaddlePaddle在编译时/运行时会自动找到系统中安装的CUDA和cuDNN库进行编译和执行。
+使用参数 :code:`-DCUDA_ARCH_NAME=Auto` 可以指定开启自动检测SM架构，加速编译。
+
+PaddlePaddle可以使用cuDNN v5.1之后的任何一个版本来编译运行，但尽量请保持编译和运行使用的cuDNN是同一个版本。
+我们推荐使用最新版本的cuDNN。
+
+编译选项的设置
+++++++++++++++
+
+PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/cuDNN库。cmake编译时，首先在系统路径（ :code:`/usr/lib:/usr/local/lib` ）中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如 
+
+..  code-block:: bash
+
+    cmake .. -DWITH_GPU=ON -DWITH_TESTING=OFF -DCUDNN_ROOT=/opt/cudnnv5
+
+**注意：这几个编译选项的设置，只在第一次cmake的时候有效。如果之后想要重新设置，推荐清理整个编译目录（** :code:`rm -rf` ）**后，再指定。**

doc/getstarted/build_and_install/build_from_source_en.md

-Installing from Sources
-==========================
-
-* [1. Download and Setup](#download)
-* [2. Requirements](#requirements)
-* [3. Build on Ubuntu](#ubuntu)
-* [4. Build on Centos](#centos)
-
-
-## <span id="download">Download and Setup</span> 
-You can download PaddlePaddle from the [github source](https://github.com/PaddlePaddle/Paddle).
-
-```bash
-git clone https://github.com/PaddlePaddle/Paddle paddle
-cd paddle
-```
-## <span id="requirements">Requirements</span>
-
-To compile the source code, your computer must be equipped with the following dependencies.
-
-- **Compiler**: GCC >= 4.8 or Clang >= 3.3 (AppleClang >= 5.1) and gfortran compiler
-- **CMake**: CMake >= 3.0 (at least CMake 3.4 on Mac OS X)
-- **BLAS**: MKL, OpenBlas or ATLAS
-- **Python**: only support Python 2.7
-- **Go**
-
-**Note:** For CUDA 7.0 and CUDA 7.5, GCC 5.0 and up are not supported!
-For CUDA 8.0, GCC versions later than 5.3 are not supported!
-
-### Options
-
-PaddlePaddle supports some build options. 
-
-<html>
-<table> 
-<thead>
-<tr>
-<th scope="col" class="left">Optional</th>
-<th scope="col" class="left">Description</th>
-</tr>
-</thead>
-<tbody>
-<tr><td class="left">WITH_GPU</td><td class="left">Compile PaddlePaddle with NVIDIA GPU</td></tr>
-<tr><td class="left">WITH_AVX</td><td class="left">Compile PaddlePaddle with AVX intrinsics</td></tr>
-<tr><td class="left">WITH_DSO</td><td class="left">Compile PaddlePaddle with dynamic linked CUDA</td></tr>
-<tr><td class="left">WITH_TESTING</td><td class="left">Compile PaddlePaddle with unit testing</td></tr>
-<tr><td class="left">WITH_SWIG_PY</td><td class="left">Compile PaddlePaddle with inference api</td></tr>
-<tr><td class="left">WITH_STYLE_CHECK</td><td class="left">Compile PaddlePaddle with style check</td></tr>
-<tr><td class="left">WITH_PYTHON</td><td class="left">Compile PaddlePaddle with python interpreter</td></tr>
-<tr><td class="left">WITH_DOUBLE</td><td class="left">Compile PaddlePaddle with double precision</td></tr>
-<tr><td class="left">WITH_RDMA</td><td class="left">Compile PaddlePaddle with RDMA support</td></tr>
-<tr><td class="left">WITH_TIMER</td><td class="left">Compile PaddlePaddle with stats timer</td></tr>
-<tr><td class="left">WITH_PROFILER</td><td class="left">Compile PaddlePaddle with GPU profiler</td></tr>
-<tr><td class="left">WITH_DOC</td><td class="left">Compile PaddlePaddle with documentation</td></tr>
-<tr><td class="left">WITH_COVERAGE</td><td class="left">Compile PaddlePaddle with code coverage</td></tr>
-<tr><td class="left">COVERALLS_UPLOAD</td><td class="left">Package code coverage data to coveralls</td></tr>
-<tr><td class="left">ON_TRAVIS</td><td class="left">Exclude special unit test on Travis CI</td></tr>
-</tbody>
-</table>
-</html>
-
-**Note:**
-  - The GPU version works best with Cuda Toolkit 8.0 and cuDNN v5.
-  - Other versions like Cuda Toolkit 7.0, 7.5 and cuDNN v3, v4 are also supported.
-  - **To utilize cuDNN v5, Cuda Toolkit 7.5 is prerequisite and vice versa.**
-
-As a simple example, consider the following:  
-
-1. **BLAS Dependencies(optional)**
-  
-    CMake will search BLAS libraries from the system. If not found, OpenBLAS will be downloaded, built and installed automatically.
-    To utilize preinstalled BLAS， you can simply specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`.
-
-    ```bash
-    # specify MKL
-    cmake .. -DMKL_ROOT=<mkl_path>
-    # or specify OpenBLAS
-    cmake .. -DOPENBLAS_ROOT=<openblas_path>
-    ```
-
-2. **Doc Dependencies(optional)**
-
-    To generate PaddlePaddle's documentation, install dependencies and set `-DWITH_DOC=ON` as follows:
-
-    ```bash
-    pip install 'sphinx>=1.4.0'
-    pip install sphinx_rtd_theme recommonmark
-
-    # install doxygen on Ubuntu
-    sudo apt-get install doxygen 
-    # install doxygen on Mac OS X
-    brew install doxygen
-
-    # active docs in cmake
-    cmake .. -DWITH_DOC=ON`
-    ```
-
-## <span id="ubuntu">Build on Ubuntu 14.04</span>
-
-### Install Dependencies
-
-- **Paddle Dependencies**
-
-    ```bash
-    # necessary
-    sudo apt-get update
-    sudo apt-get install -y git curl gcc g++ gfortran make build-essential automake
-    sudo apt-get install -y python python-pip python-numpy libpython-dev bison
-    sudo pip install 'protobuf==3.1.0.post1'
-
-    # Install Go
-    # You can follow https://golang.org/doc/install for a detailed explanation.
-    wget -O go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz && \
-    tar -C $HOME -xzf go.tgz && \
-    mkdir $HOME/gopath && \
-    rm go.tgz
-
-    # Setup environment variables
-    export GOROOT=$HOME/go
-    export GOPATH=$HOME/gopath
-    export PATH=$PATH:$GOROOT/bin
-
-    # install cmake 3.4
-    curl -sSL https://cmake.org/files/v3.4/cmake-3.4.1.tar.gz | tar -xz && \
-        cd cmake-3.4.1 && ./bootstrap && make -j4 && sudo make install && \
-        cd .. && rm -rf cmake-3.4.1
-    ```
-
-- **GPU Dependencies (optional)**
-
-    To build GPU version, you will need the following installed:
-
-        1. a CUDA-capable GPU
-        2. A supported version of Linux with a GCC compiler and toolchain
-        3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
-        4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
-
-    The CUDA development environment relies on tight integration with the host development environment,
-    including the host compiler and C runtime libraries, and is therefore only supported on
-    distribution versions that have been qualified for this CUDA Toolkit release.
-        
-    After downloading cuDNN library, issue the following commands:
-
-    ```bash
-    sudo tar -xzf cudnn-7.5-linux-x64-v5.1.tgz -C /usr/local
-    sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
-    ```
-    Then you need to set LD\_LIBRARY\_PATH, PATH environment variables in ~/.bashrc.
-
-    ```bash
-    export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
-    export PATH=/usr/local/cuda/bin:$PATH
-    ```
-
-### Build and Install
-
-As usual, the best option is to create build folder under paddle project directory.
-
-```bash
-mkdir build && cd build
-``` 
-
-Finally, you can build and install PaddlePaddle:
-
-```bash
-# you can add build option here, such as:    
-cmake .. -DCMAKE_INSTALL_PREFIX=<path to install>
-# please use sudo make install, if you want to install PaddlePaddle into the system
-make -j `nproc` && make install
-# set PaddlePaddle installation path in ~/.bashrc
-export PATH=<path to install>/bin:$PATH
-# install PaddlePaddle Python modules.
-sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
-```
-
-## <span id="centos">Build on Centos 7</span>
-
-### Install Dependencies
-
-- **CPU Dependencies**
-
-    ```bash
-    # necessary
-    sudo yum update
-    sudo yum install -y epel-release
-    sudo yum install -y make cmake3 python-devel python-pip gcc-gfortran swig git
-    sudo pip install wheel numpy
-    sudo pip install 'protobuf>=3.0.0'
-    ```
-  
-- **GPU Dependencies (optional)**
-
-    To build GPU version, you will need the following installed:
-
-        1. a CUDA-capable GPU
-        2. A supported version of Linux with a GCC compiler and toolchain
-        3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
-        4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
-
-    The CUDA development environment relies on tight integration with the host development environment,
-    including the host compiler and C runtime libraries, and is therefore only supported on
-    distribution versions that have been qualified for this CUDA Toolkit release.
-        
-    After downloading cuDNN library, issue the following commands:
-
-    ```bash
-    sudo tar -xzf cudnn-7.5-linux-x64-v5.1.tgz -C /usr/local
-    sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
-    ```
-    Then you need to set LD\_LIBRARY\_PATH, PATH environment variables in ~/.bashrc.
-
-    ```bash
-    export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
-    export PATH=/usr/local/cuda/bin:$PATH
-    ```
-
-### Build and Install
-
-As usual, the best option is to create build folder under paddle project directory.
-
-```bash
-mkdir build && cd build
-``` 
-
-Finally, you can build and install PaddlePaddle:
-  
-```bash
-# you can add build option here, such as:    
-cmake3 .. -DCMAKE_INSTALL_PREFIX=<path to install>
-# please use sudo make install, if you want to install PaddlePaddle into the system
-make -j `nproc` && make install
-# set PaddlePaddle installation path in ~/.bashrc
-export PATH=<path to install>/bin:$PATH
-# install PaddlePaddle Python modules.
-sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
-```

doc/getstarted/build_and_install/build_from_source_en.rst

+Build from Sources
+==========================
+
+.. _build_step:
+
+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/>`_ .
+
+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 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
+   cd build
+   cmake -DWITH_GPU=OFF -DWITH_TESTING=OFF ..
+   make
+
+When the compile finishes, you can get the output whl package under
+build/python/dist, then you can choose to install the whl on local
+machine or copy it to the target machine.
+
+.. code-block:: bash
+
+   pip install python/dist/*.whl
+
+
+.. _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
+----------------
+
+PaddlePaddle need the following dependencies when compiling, other dependencies
+will be downloaded automatically.
+
+.. csv-table:: PaddlePaddle Compile Dependencies
+   :header: "Dependency", "Version", "Description"
+   :widths: 10, 15, 30
+
+   "CMake", ">=3.5", ""
+   "GCC", "4.8.2", "Recommend devtools2 for CentOS"
+   "Python", "2.7.x", "Need libpython2.7.so"
+   "pip", ">=9.0", ""
+   "numpy", "", ""
+   "SWIG", ">=2.0", ""
+   "Go", ">=1.8", "Optional"
+
+
+.. _build_options:
+
+Build Options
+----------------
+
+Build options include whether build binaries for CPU or GPU, which BLAS
+library to use etc. You may pass these settings when running cmake.
+For detailed cmake tutorial please refer to `here <https://cmake.org/cmake-tutorial>`_ 。
+
+.. _build_options_bool:
+
+Bool Type Options
+----------------
+
+You can add :code:`-D` argument to pass such options, like:
+
+..  code-block:: bash
+
+    cmake .. -DWITH_GPU=OFF
+
+..  csv-table:: Bool Type Options
+    :header: "Option", "Description", "Default"
+    :widths: 1, 7, 2
+
+    "WITH_GPU", "Build with GPU support", "ON"
+    "WITH_C_API", "Build only CAPI", "OFF"
+    "WITH_DOUBLE", "Build with double precision", "OFF"
+    "WITH_DSO", "Dynamically load CUDA libraries", "ON"
+    "WITH_AVX", "Build with AVX support", "ON"
+    "WITH_PYTHON", "Build with integrated Python interpreter", "ON"
+    "WITH_STYLE_CHECK", "Check code style when building", "ON"
+    "WITH_TESTING", "Build unit tests", "ON"
+    "WITH_DOC", "Build documentaions", "OFF"
+    "WITH_SWIG_PY", "Build Python SWIG interface for V2 API", "Auto"
+    "WITH_GOLANG", "Build fault-tolerant parameter server written in go", "ON"
+    "WITH_MKL", "Use MKL as BLAS library, else use OpenBLAS", "ON"
+
+
+BLAS
++++++
+
+PaddlePaddle supports `MKL <https://software.intel.com/en-us/intel-mkl>`_ and
+`OpenBlAS <http://www.openblas.net/>`_ as BLAS library。By default it uses MKL.
+If you are using MKL and your machine supports AVX2, MKL-DNN will also be downloaded
+and used, for more `details <https://github.com/PaddlePaddle/Paddle/tree/develop/doc/design/mkldnn#cmake>`_ .
+
+If you choose not to use MKL, then OpenBlAS will be used.
+
+CUDA/cuDNN
++++++++++++
+
+PaddlePaddle will automatically find CUDA and cuDNN when compiling and running.
+parameter :code:`-DCUDA_ARCH_NAME=Auto` can be used to detect SM architecture
+automatically in order to speed up the build.
+
+PaddlePaddle can build with any version later than cuDNN v5.1, and we intend to
+keep on with latest cuDNN versions. Be sure to run with the same version of cuDNN
+you built.
+
+Pass Compile Options
+++++++++++++++
+
+You can pass compile options to use intended BLAS/CUDA/Cudnn libraries.
+When running cmake command, it will search system paths like
+:code:`/usr/lib:/usr/local/lib` and then search paths that you
+passed to cmake, i.e.
+
+..  code-block:: bash
+
+    cmake .. -DWITH_GPU=ON -DWITH_TESTING=OFF -DCUDNN_ROOT=/opt/cudnnv5
+
+**NOTE: These options only take effect when running cmake for the first time, you need to clean the cmake cache or clean the build directory (** :code:`rm -rf` **) if you want to change it.**

doc/getstarted/build_and_install/cmake/build_from_source_cn.rst

-PaddlePaddle的编译选项
-======================
-
-PaddlePaddle的编译选项，包括生成CPU/GPU二进制文件、链接何种BLAS库等。用户可在调用cmake的时候设置它们，详细的cmake使用方法可以参考 `官方文档 <https://cmake.org/cmake-tutorial>`_ 。
-
-Bool型的编译选项
-----------------
-用户可在cmake的命令行中，通过使用 ``-D`` 命令设置该类编译选项，例如
-
-..  code-block:: bash
-
-    cmake .. -DWITH_GPU=OFF
-
-..  csv-table:: Bool型的编译选项
-    :widths: 1, 7, 2
-    :file: compile_options.csv
-
-BLAS/CUDA/Cudnn的编译选项
---------------------------
-BLAS
-+++++
-
-PaddlePaddle支持以下任意一种BLAS库：`MKL <https://software.intel.com/en-us/intel-mkl>`_ ，`ATLAS <http://math-atlas.sourceforge.net/>`_ ，`OpenBlAS <http://www.openblas.net/>`_ 和 `REFERENCE BLAS <http://www.netlib.org/blas/>`_ 。
-
-..  csv-table:: BLAS路径相关的编译选项
-    :widths: 1, 2, 7
-    :file: cblas_settings.csv
-
-CUDA/Cudnn
-+++++++++++
-
-PaddlePaddle可以使用cudnn v2之后的任何一个版本来编译运行，但尽量请保持编译和运行使用的cudnn是同一个版本。 我们推荐使用最新版本的cudnn v5.1。
-
-编译选项的设置
-++++++++++++++
-
-PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/Cudnn库。cmake编译时，首先在系统路径(/usr/lib\:/usr/local/lib)中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如 
-
-..  code-block:: bash
-
-    cmake .. -DMKL_ROOT=/opt/mkl/ -DCUDNN_ROOT=/opt/cudnnv5
-
-注意：这几个编译选项的设置，只在第一次cmake的时候有效。如果之后想要重新设置，推荐清理整个编译目录（``rm -rf``）后，再指定。

doc/getstarted/build_and_install/cmake/cblas_settings.csv

-编译选项,描述,注意
-MKL_ROOT,MKL的路径,${MKL_ROOT}/include下需要包含mkl.h，${MKL_ROOT}/lib目录下需要包含mkl_core，mkl_sequential和mkl_intel_lp64三个库。
-ATLAS_ROOT,ATLAS的路径,${ATLAS_ROOT}/include下需要包含cblas.h，${ATLAS_ROOT}/lib下需要包含cblas和atlas两个库。
-OPENBLAS_ROOT,OpenBLAS的路径,${OPENBLAS_ROOT}/include下需要包含cblas.h，${OPENBLAS_ROOT}/lib下需要包含openblas库。
-REFERENCE_CBLAS_ROOT,REFERENCE BLAS的路径,${REFERENCE_CBLAS_ROOT}/include下需要包含cblas.h，${REFERENCE_CBLAS_ROOT}/lib下需要包含cblas库。
\ No newline at end of file

doc/getstarted/build_and_install/cmake/compile_options.csv

-选项,说明,默认值
-WITH_GPU,是否支持GPU。,取决于是否寻找到CUDA工具链
-WITH_DOUBLE,是否使用双精度浮点数。,否
-WITH_DSO,是否运行时动态加载CUDA动态库，而非静态加载CUDA动态库。,是
-WITH_AVX,是否编译含有AVX指令集的PaddlePaddle二进制文件,是
-WITH_PYTHON,是否内嵌PYTHON解释器。方便今后的嵌入式移植工作。,是
-WITH_STYLE_CHECK,是否编译时进行代码风格检查,是
-WITH_RDMA,是否开启RDMA,否
-WITH_TIMER,是否开启计时功能。如果开启会导致运行略慢，打印的日志变多，但是方便调试和测Benchmark,否
-WITH_TESTING,是否开启单元测试,取决于是否寻找到GTEST
-WITH_DOC,是否编译中英文文档,否
-WITH_SWIG_PY,是否编译PYTHON的SWIG接口，该接口可用于预测和定制化训练,取决于是否寻找到SWIG
\ No newline at end of file

doc/getstarted/build_and_install/docker_install_cn.rst

-PaddlePaddle的Docker容器使用方式
+使用Docker安装运行
 ================================
 
-PaddlePaddle目前唯一官方支持的运行的方式是Docker容器。因为Docker能在所有主要操作系统（包括Linux，Mac OS X和Windows）上运行。 请注意，您需要更改 `Dockers设置 <https://github.com/PaddlePaddle/Paddle/issues/627>`_ 才能充分利用Mac OS X和Windows上的硬件资源。
+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。
+您可以在 `Docker官网 <https://docs.docker.com/get-started/>`_ 获得基本的Docker安装和使用方法。
 
-Docker使用入门
-------------------------------
-
-几个基础的概念帮助理解和使用Docker：
+如果您在使用Windows，可以参考
+`这篇 <https://docs.docker.com/toolbox/toolbox_install_windows/>`_
+教程，完成在Windows上安装和使用Docker。
 
-- *镜像*：一个Docker镜像是一个打包好的软件。它包含了这个软件本身和它所依赖的运行环境。PaddlePaddle的Docker镜像就包含了PaddlePaddle的Python库以及其依赖的多个Python库。这样我们可以直接在Docker中运行需要的程序而不需要安装后在执行。可以执行：
+在了解Docker的基本使用方法之后，即可开始下面的步骤：
 
-  .. code-block:: bash
+.. _docker_pull:
 
-     docker images
+获取PaddlePaddle的Docker镜像
+------------------------------
 
-  来列出当前系统中的所有镜像，同样可以执行：
+执行下面的命令获取最新的PaddlePaddle Docker镜像
 
   .. code-block:: bash
-		  
-     docker pull paddlepaddle/paddle:0.10.0
 
-  来下载Docker镜像，paddlepaddle/paddle是从官方镜像源Dockerhub.com下载的，推荐国内用户使用docker.paddlepaddle.org/paddle下载。
+     docker pull paddlepaddle/paddle
 
-- *容器*： 如果说一个Docker镜像就是一个程序，那容器就是这个程序运行时产生的“进程”。
-  实际上，一个容器就是一个操作系统的进程，但是是运行在独立的进程空间，文件系统以及网络之上。
-  可以执行：
+对于国内用户，我们提供了加速访问的镜像源：
 
   .. code-block:: bash
 
-     docker run paddlepaddle/paddle:0.10.0
+     docker pull docker.paddlepaddle.org/paddle
 
-  来使用一个镜像启动一个容器。
-
-- 默认情况下，Docker容器会运行在独立的文件系统空间之上，我们无法在Docker容器中
-  访问到主机上的文件。可以通过*挂载Volume*的方式，将主机上的文件或目录挂载到
-  Docker容器中。下面的命令把当前目录挂载到了容器中的 /data 目录下，容器使用
-  debian镜像，并且启动后执行 :code:`ls /data`。
+下载GPU版本的Docker镜像：
 
   .. code-block:: bash
 
-     docker run --rm -v $(pwd):/data debian ls /data
-
-PaddlePaddle发布的Docker镜像使用说明
-------------------------------
-
-我们把PaddlePaddle的编译环境打包成一个镜像，称为开发镜像，里面涵盖了
-PaddlePaddle需要的所有编译工具。把编译出来的PaddlePaddle也打包成一个镜
-像，称为生产镜像，里面涵盖了PaddlePaddle运行所需的所有环境。每次
-PaddlePaddle发布新版本的时候都会发布对应版本的生产镜像以及开发镜像。运
-行镜像包括纯CPU版本和GPU版本以及其对应的非AVX版本。我们会在
-`dockerhub.com <https://hub.docker.com/r/paddlepaddle/paddle/tags/>`_ 
-和国内镜像`docker.paddlepaddle.org` 提供最新
-的Docker镜像，可以在"tags"标签下找到最新的Paddle镜像版本。
-
-**注意：为了方便在国内的开发者下载Docker镜像，我们提供了国内的镜像服务器供大家使用。如果您在国内，请把文档里命令中的paddlepaddle/paddle替换成docker.paddlepaddle.org/paddle。**
-
-1. 开发镜像：:code:`paddlepaddle/paddle:0.10.0-dev`
-
-   这个镜像包含了Paddle相关的开发工具以及编译和运行环境。用户可以使用开发镜像代替配置本地环境，完成开发，编译，发布，
-   文档编写等工作。由于不同的Paddle的版本可能需要不同的依赖和工具，所以如果需要自行配置开发环境需要考虑版本的因素。
-   开发镜像包含了以下工具：
-   
-   - gcc/clang
-   - nvcc
-   - Python
-   - sphinx
-   - woboq
-   - sshd
-   很多开发者会使用远程的安装有GPU的服务器工作，用户可以使用ssh登录到这台服务器上并执行 :code:`docker exec`进入开发镜像并开始工作，
-   也可以在开发镜像中启动一个SSHD服务，方便开发者直接登录到镜像中进行开发:
-
-   以交互容器方式运行开发镜像：
-
-   .. code-block:: bash
-
-      docker run -it --rm -v $(pwd):/paddle  paddlepaddle/paddle:0.10.0-dev /bin/bash
-
-   或者，可以以后台进程方式运行容器：
-
-   .. code-block:: bash
-
-      docker run -d -p 2202:22 -p 8888:8888 -v $(pwd):/paddle paddlepaddle/paddle:0.10.0-dev /usr/sbin/sshd -D
-
-   然后用密码 :code:`root` SSH进入容器：
-
-   .. code-block:: bash
-
-      ssh -p 2202 root@localhost
-
-   SSH方式的一个优点是我们可以从多个终端进入容器。比如，一个终端运行vi，另一个终端运行Python。另一个好处是我们可以把PaddlePaddle容器运行在远程服务器上，并在笔记本上通过SSH与其连接。
-
-2. 生产镜像：根据CPU、GPU和非AVX区分了如下4个镜像：
-
-   - GPU/AVX：:code:`paddlepaddle/paddle:<version>-gpu`
-   - GPU/no-AVX：:code:`paddlepaddle/paddle:<version>-gpu-noavx`
-   - CPU/AVX：:code:`paddlepaddle/paddle:<version>`
-   - CPU/no-AVX：:code:`paddlepaddle/paddle:<version>-noavx`
-
-   纯CPU镜像以及GPU镜像都会用到AVX指令集，但是2008年之前生产的旧电脑不支持AVX。以下指令能检查Linux电脑是否支持AVX：
-
-   .. code-block:: bash
-
-      if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi
-
-   如果输出是No，就需要选择使用no-AVX的镜像
-
-   **注：在0.10.0之后的版本，PaddlePaddle都可以自动判断硬件是否支持AVX，所以无需判断AVX即可使用**
+     docker pull paddlepaddle/paddle:latest-gpu
+     docker pull docker.paddlepaddle.org/paddle:latest-gpu
 
-   以上方法在GPU镜像里也能用，只是请不要忘记提前在物理机上安装GPU最新驱动。
-   为了保证GPU驱动能够在镜像里面正常运行，我们推荐使用[nvidia-docker](https://github.com/NVIDIA/nvidia-docker)来运行镜像。
+选择下载使用不同的BLAS库的Docker镜像：
 
-   .. code-block:: bash
-
-      nvidia-docker run -it --rm paddledev/paddle:0.10.0-gpu /bin/bash
+  .. code-block:: bash
 
-   注意: 如果使用nvidia-docker存在问题，你也许可以尝试更老的方法，具体如下，但是我们并不推荐这种方法。：
+     # 默认是使用MKL的镜像
+     docker pull paddlepaddle/paddle
+     # 使用OpenBLAS的镜像
+     docker pull paddlepaddle/paddle:latest-openblas
 
-   .. code-block:: bash
+下载指定版本的Docker镜像，可以从 `DockerHub网站 <https://hub.docker.com/r/paddlepaddle/paddle/tags/>`_ 获取可选的tag，并执行下面的命令：
 
-      export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
-      export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
-      docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:0.10.0-gpu
+  .. code-block:: bash
 
-3. 运行以及发布您的AI程序
+     docker pull paddlepaddle/paddle:[tag]
+     # 比如：
+     docker pull docker.paddlepaddle.org/paddle:0.10.0-gpu
 
-   假设您已经完成了一个AI训练的python程序 :code:`a.py`，这个程序是您在开发机上使用开发镜像完成开发。此时您可以运行这个命令在开发机上进行测试运行：
+.. _docker_run:
 
-   .. code-block:: bash
+在Docker中执行PaddlePaddle训练程序
+------------------------------
 
-      docker run -it -v $PWD:/work paddle /work/a.py
+假设您已经在当前目录（比如在/home/work）编写了一个PaddlePaddle的程序 :code:`train.py` （可以参考
+`PaddlePaddleBook <http://www.paddlepaddle.org/docs/develop/book/01.fit_a_line/index.cn.html>`_ 
+编写），就可以使用下面的命令开始执行训练：
 
-   如果要使用GPU，请运行：
+  .. code-block:: bash
 
-   .. code-block:: bash
+     cd /home/work
+     docker run -it -v $PWD:/work paddlepaddle/paddle /work/train.py
+ 
+上述命令中， :code:`-it` 参数说明容器已交互式运行； :code:`-v $PWD:/work`
+指定将当前路径（Linux中$PWD变量会展开为当前路径的绝对路径）挂载到容器内部的 :code:`/work`
+目录； :code:`paddlepaddle/paddle` 指定需要使用的容器； 最后 :code:`/work/train.py`
+为容器内执行的命令，即运行训练程序。
 
-      nvidia-docker run -it -v $PWD:/work paddle /work/a.py
+当然，您也可以进入到Docker容器中，以交互式的方式执行或调试您的代码：
 
+  .. code-block:: bash
+     docker run -it -v $PWD:/work paddlepaddle/paddle /bin/bash
+     cd /work
+     python train.py
 
-   这里`a.py`包含的所有依赖假设都可以在Paddle的运行容器中。如果需要包含更多的依赖、或者需要发布您的应用的镜像，可以编写`Dockerfile`使用`FROM paddledev/paddle:0.10.0`
-   创建和发布自己的AI程序镜像。
+**注：PaddlePaddle Docker镜像为了减小体积，默认没有安装vim，您可以在容器中执行** :code:`apt-get install -y vim` **安装后，在容器中编辑代码。**
 
-运行PaddlePaddle Book
----------------------
+.. _docker_run_book:
 
-Jupyter Notebook是一个开源的web程序，大家可以通过它制作和分享带有代码、公式、图表、文字的交互式文档。用户可以通过网页浏览文档。
+使用Docker启动PaddlePaddle Book教程
+------------------------------
 
+使用Docker可以快速在本地启动一个包含了PaddlePaddle官方Book教程的Jupyter Notebook，可以通过网页浏览。
 PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Notebook。
 如果您想要更深入了解deep learning，PaddlePaddle Book一定是您最好的选择。
+大家可以通过它阅读教程，或者制作和分享带有代码、公式、图表、文字的交互式文档。
 
 我们提供可以直接运行PaddlePaddle Book的Docker镜像，直接运行：
 
-.. code-block:: bash
+  .. code-block:: bash
 
-    docker run -p 8888:8888 paddlepaddle/book
+     docker run -p 8888:8888 paddlepaddle/book
 
 然后在浏览器中输入以下网址：
 
-.. code-block:: text
+  .. code-block:: text
 
-    http://localhost:8888/
+     http://localhost:8888/
 
 就这么简单，享受您的旅程！
 
-通过Docker容器开发PaddlePaddle
-------------------------------
-
-开发人员可以在Docker开发镜像中开发PaddlePaddle。这样开发人员可以以一致的方式在不同的平台上工作 - Linux，Mac OS X和Windows。
+.. _docker_run_gpu:
 
-1. 制作PaddlePaddle开发镜像
-
-   PaddlePaddle每次发布新版本都会发布对应的开发镜像供开发者直接使用。这里介绍如生成造这个开发镜像。
-   生成Docker镜像的方式有两个，一个是直接把一个容器转换成镜像，另一个是创建Dockerfile并运行docker build指令按照Dockerfile生成镜像。第一个方法的好处是简单快捷，适合自己实验，可以快速迭代。第二个方法的好处是Dockerfile可以把整个生成流程描述很清楚，其他人很容易看懂镜像生成过程，持续集成系统也可以简单地复现这个过程。我们采用第二个方法。Dockerfile位于PaddlePaddle repo的根目录。生成生产镜像只需要运行：
-
-   .. code-block:: bash
-      
-      git clone https://github.com/PaddlePaddle/Paddle.git
-      cd Paddle
-      docker build -t paddle:dev .
-
-   docker build这个命令的-t指定了生成的镜像的名字，这里我们用paddle:dev。到此，PaddlePaddle开发镜像就被构建完毕了。
+使用Docker执行GPU训练
+------------------------------
 
-2. 制作PaddlePaddle生产镜像
+为了保证GPU驱动能够在镜像里面正常运行，我们推荐使用
+`nvidia-docker <https://github.com/NVIDIA/nvidia-docker>`_ 来运行镜像。
+请不要忘记提前在物理机上安装GPU最新驱动。
 
-   生产镜像的生成分为两步，第一步是运行：
+  .. code-block:: bash
 
-   .. code-block:: bash
-      
-      docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=OFF" -e "WITH_TEST=ON" paddle:dev
+     nvidia-docker run -it -v $PWD:/work paddledev/paddle:latest-gpu /bin/bash
 
-   以上命令会编译PaddlePaddle，生成运行程序，以及生成创建生产镜像的Dockerfile。所有生成的的文件都在build目录下。“WITH_GPU”控制生成的生产镜像是否支持GPU，“WITH_AVX”控制生成的生产镜像是否支持AVX，”WITH_TEST“控制是否生成单元测试。
+**注: 如果没有安装nvidia-docker，可以尝试以下的方法，将CUDA库和Linux设备挂载到Docker容器内：**
 
-   第二步是运行：
+  .. code-block:: bash
 
-   .. code-block:: bash
-      
-      docker build -t paddle:prod -f build/Dockerfile ./build
+     export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
+     export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
+     docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:latest-gpu
 
-   以上命令会按照生成的Dockerfile把生成的程序拷贝到生产镜像中并做相应的配置，最终生成名为paddle:prod的生产镜像。
+**关于AVX：**
 
-3. 运行单元测试
+AVX是一种CPU指令集，可以加速PaddlePaddle的计算。最新的PaddlePaddle Docker镜像默认
+是开启AVX编译的，所以，如果您的电脑不支持AVX，需要单独
+`编译 <./build_from_source_cn.rst>`_ PaddlePaddle为no-avx版本。
 
-   运行以下指令：
+以下指令能检查Linux电脑是否支持AVX：
 
    .. code-block:: bash
-      
-      docker run -it -v $(pwd):/paddle paddle:dev bash -c "cd /paddle/build && ctest"
-
-文档
-----
-
-Paddle的Docker开发镜像带有一个通过 `woboq code browser
-<https://github.com/woboq/woboq_codebrowser>`_ 生成的HTML版本的C++源代码，便于用户浏览C++源码。
 
-只要在Docker里启动PaddlePaddle的时候给它一个名字，就可以再运行另一个Nginx Docker镜像来服务HTML代码：
-
-.. code-block:: bash
-
-   docker run -d --name paddle-cpu-doc paddle:0.10.0-dev
-   docker run -d --volumes-from paddle-cpu-doc -p 8088:80 nginx
+      if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi
 
-接着我们就能够打开浏览器在 http://localhost:8088/paddle/ 浏览代码。
+如果输出是No，就需要选择使用no-AVX的镜像

doc/getstarted/build_and_install/docker_install_en.rst

-PaddlePaddle in Docker Containers
+Run in Docker Containers
 =================================
 
-Docker container is currently the only officially-supported way to
-running PaddlePaddle.  This is reasonable as Docker now runs on all
-major operating systems including Linux, Mac OS X, and Windows.
-Please be aware that you will need to change `Dockers settings
-<https://github.com/PaddlePaddle/Paddle/issues/627>`_ to make full use
-of your hardware resource on Mac OS X and Windows.
+Run PaddlePaddle in Docker container so that you don't need to care about
+runtime dependencies, also you can run under Windows system. You can get
+tutorials at `here <https://docs.docker.com/get-started/>`_ .
 
-Working With Docker
--------------------
+If you are using Windows, please refer to
+`this <https://docs.docker.com/toolbox/toolbox_install_windows/>`_
+tutorial to start running docker under windows.
 
-Docker is simple as long as we understand a few basic concepts:
+After you've read above tutorials you may proceed the following steps.
 
-- *image*: A Docker image is a pack of software. It could contain one or more programs and all their dependencies. For example, the PaddlePaddle's Docker image includes pre-built PaddlePaddle and Python and many Python packages. We can run a Docker image directly, other than installing all these software. We can type
+.. _docker_pull:
 
-  .. code-block:: bash
-
-     docker images
+Pull PaddlePaddle Docker Image
+------------------------------
 
-  to list all images in the system. We can also run
+Run the following command to download the latest Docker images:
 
   .. code-block:: bash
-		  
-     docker pull paddlepaddle/paddle:0.10.0
 
-  to download a Docker image, paddlepaddle/paddle in this example,
-  from Dockerhub.com.
+     docker pull paddlepaddle/paddle
 
-- *container*: considering a Docker image a program, a container is a
-  "process" that runs the image. Indeed, a container is exactly an
-  operating system process, but with a virtualized filesystem, network
-  port space, and other virtualized environment. We can type
+For users in China, we provide a faster mirror:
 
   .. code-block:: bash
 
-     docker run paddlepaddle/paddle:0.10.0
+     docker pull docker.paddlepaddle.org/paddle
 
-  to start a container to run a Docker image, paddlepaddle/paddle in this example.
-
-- By default docker container have an isolated file system namespace,
-  we can not see the files in the host file system. By using *volume*,
-  mounted files in host will be visible inside docker container.
-  Following command will mount current dirctory into /data inside
-  docker container, run docker container from debian image with
-  command :code:`ls /data`.
+Download GPU version images:
 
   .. code-block:: bash
 
-     docker run --rm -v $(pwd):/data debian ls /data
-
-Usage of CPU-only and GPU Images
-----------------------------------
-
-We package PaddlePaddle's compile environment into a Docker image,
-called the develop image, it contains all compiling tools that
-PaddlePaddle needs. We package compiled PaddlePaddle program into a
-Docker image as well, called the production image, it contains all
-runtime environment that running PaddlePaddle needs. For each version
-of PaddlePaddle, we release both of them. Production image includes
-CPU-only version and a CUDA GPU version and their no-AVX versions.
-
-We put the docker images on `dockerhub.com
-<https://hub.docker.com/r/paddlepaddle/paddle/tags/>`_. You can find the
-latest versions under "tags" tab at dockerhub.com. 
-
-** NOTE: If you are in China, you can use our Docker image registry mirror to speed up the download process. To use it, please replace all paddlepaddle/paddle in the commands to docker.paddlepaddle.org/paddle.**
-
-
-1. development image :code:`paddlepaddle/paddle:<version>-dev`
-
-   This image has packed related develop tools and runtime
-   environment. Users and developers can use this image instead of
-   their own local computer to accomplish development, build,
-   releasing, document writing etc. While different version of paddle
-   may depends on different version of libraries and tools, if you
-   want to setup a local environment, you must pay attention to the
-   versions.  The development image contains:
-   
-   - gcc/clang
-   - nvcc
-   - Python
-   - sphinx
-   - woboq
-   - sshd
-     
-   Many developers use servers with GPUs, they can use ssh to login to
-   the server and run :code:`docker exec` to enter the docker
-   container and start their work.  Also they can start a development
-   docker image with SSHD service, so they can login to the container
-   and start work.
-
-2. Production images, this image might have multiple variants:
-
-   - GPU/AVX：:code:`paddlepaddle/paddle:<version>-gpu`
-   - GPU/no-AVX：:code:`paddlepaddle/paddle:<version>-gpu-noavx`
-   - CPU/AVX：:code:`paddlepaddle/paddle:<version>`
-   - CPU/no-AVX：:code:`paddlepaddle/paddle:<version>-noavx`
-
-   Please be aware that the CPU-only and the GPU images both use the
-   AVX instruction set, but old computers produced before 2008 do not
-   support AVX.  The following command checks if your Linux computer
-   supports AVX:
-
-   .. code-block:: bash
-
-      if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi
-
-   **NOTE：versions after 0.10.0 will automatically detect system AVX support, so manual detect is not needed in this case.**
-   To run the CPU-only image as an interactive container:
-
-   .. code-block:: bash
-
-      docker run -it --rm paddlepaddle/paddle:0.10.0 /bin/bash
-
-   Above method work with the GPU image too -- the recommended way is
-   using `nvidia-docker <https://github.com/NVIDIA/nvidia-docker>`_.
-
-   Please install nvidia-docker first following this `tutorial
-   <https://github.com/NVIDIA/nvidia-docker#quick-start>`_.
-
-   Now you can run a GPU image:
-
-   .. code-block:: bash
-
-      nvidia-docker run -it --rm paddlepaddle/paddle:0.10.0-gpu /bin/bash
-
-
-Train Model Using Python API
-----------------------------
-
-Our official docker image provides a runtime for PaddlePaddle
-programs. The typical workflow will be as follows:
-
-Create a directory as workspace:
-
-.. code-block:: bash
-
-   mkdir ~/workspace
-
-Edit a PaddlePaddle python program using your favourite editor
-
-.. code-block:: bash
-
-   emacs ~/workspace/example.py
-
-Run the program using docker:
-
-.. code-block:: bash
-
-   docker run --rm -v ~/workspace:/workspace paddlepaddle/paddle:0.10.0 python /workspace/example.py
-
-Or if you are using GPU for training:
+     docker pull paddlepaddle/paddle:latest-gpu
+     docker pull docker.paddlepaddle.org/paddle:latest-gpu
 
-.. code-block:: bash
+Choose between different BLAS version:
 
-   nvidia-docker run --rm -v ~/workspace:/workspace paddlepaddle/paddle:0.10.0-gpu python /workspace/example.py
-
-Above commands will start a docker container by running :code:`python
-/workspace/example.py`. It will stop once :code:`python
-/workspace/example.py` finishes.
-
-Another way is to tell docker to start a :code:`/bin/bash` session and
-run PaddlePaddle program interactively:
-
-.. code-block:: bash
-
-   docker run -it -v ~/workspace:/workspace paddlepaddle/paddle:0.10.0 /bin/bash
-   # now we are inside docker container
-   cd /workspace
-   python example.py
-
-Running with GPU is identical:
-
-.. code-block:: bash
-
-   nvidia-docker run -it -v ~/workspace:/workspace paddlepaddle/paddle:0.10.0-gpu /bin/bash
-   # now we are inside docker container
-   cd /workspace
-   python example.py
-
-
-Develop PaddlePaddle or Train Model Using C++ API
----------------------------------------------------
-
-We will be using PaddlePaddle development image since it contains all
-compiling tools and dependencies.
+  .. code-block:: bash
 
-1. Build PaddlePaddle develop image
+     # image using MKL by default
+     docker pull paddlepaddle/paddle
+     # image using OpenBLAS
+     docker pull paddlepaddle/paddle:latest-openblas
 
-   Use following command to build PaddlePaddle develop image:
 
-   .. code-block:: bash
+If you want to use legacy versions, choose a tag from
+`DockerHub <https://hub.docker.com/r/paddlepaddle/paddle/tags/>`_
+and run:
 
-      git clone https://github.com/PaddlePaddle/Paddle.git && cd Paddle
-      docker build -t paddle:dev .
-
-2. Build PaddlePaddle production image
+  .. code-block:: bash
 
-   There are two steps for building production image, the first step is to run:
+     docker pull paddlepaddle/paddle:[tag]
+     # i.e.
+     docker pull docker.paddlepaddle.org/paddle:0.10.0-gpu
 
-   .. code-block:: bash
+.. _docker_run:
 
-      docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=OFF" -e "WITH_TEST=ON" paddle:dev
+Launch your training program in Docker
+------------------------------
 
-   The above command will compile PaddlePaddle and create a Dockerfile for building production image. All the generated files are in the build directory. "WITH_GPU" controls if the generated production image supports GPU. "WITH_AVX" controls if the generated production image supports AVX. "WITH_TEST" controls if the unit test will be generated.
+Assume that you have already written a PaddlePaddle program
+named :code:`train.py` under directory :code:`/home/work` (refer to 
+`PaddlePaddleBook <http://www.paddlepaddle.org/docs/develop/book/01.fit_a_line/index.cn.html>`_ 
+for more samples), then run the following command:
 
-   The second step is to run:
+  .. code-block:: bash
 
-   .. code-block:: bash
+     cd /home/work
+     docker run -it -v $PWD:/work paddlepaddle/paddle /work/train.py
 
-      docker build -t paddle:prod -f build/Dockerfile ./build
+In the above command, :code:`-it` means run the container interactively;
+:code:`-v $PWD:/work` means mount the current directory ($PWD will expand
+to current absolute path in Linux) under :code:`/work` in the container.
+:code:`paddlepaddle/paddle` to specify image to use; finnally
+:code:`/work/train.py` is the command to run inside docker.
 
-   The above command will generate the production image by copying the compiled PaddlePaddle program into the image.
+Also, you can go into the container shell, run or debug your code
+interactively:
 
-3. Run unit test
+  .. code-block:: bash
+     docker run -it -v $PWD:/work paddlepaddle/paddle /bin/bash
+     cd /work
+     python train.py
 
-   Following command will run unit test:
+**NOTE: We did not install vim in the default docker image to reduce the image size, you can run** :code:`apt-get install -y vim` **to install it if you need to edit python files.**
 
-   .. code-block:: bash
-      
-      docker run -it -v $(pwd):/paddle paddle:dev bash -c "cd /paddle/build && ctest"
+.. _docker_run_book:
 
 PaddlePaddle Book
 ------------------
 
-The Jupyter Notebook is an open-source web application that allows
-you to create and share documents that contain live code, equations,
-visualizations and explanatory text in a single browser.
-
-PaddlePaddle Book is an interactive Jupyter Notebook for users and developers.
-We already exposed port 8888 for this book. If you want to
+You can create a container serving PaddlePaddle Book using Jupyter Notebook in
+one minute using Docker. PaddlePaddle Book is an interactive Jupyter Notebook
+for users and developers.If you want to
 dig deeper into deep learning, PaddlePaddle Book definitely is your best choice.
 
 We provide a packaged book image, simply issue the command:
 
-.. code-block:: bash
+  .. code-block:: bash
 
-    docker run -p 8888:8888 paddlepaddle/book
+     docker run -p 8888:8888 paddlepaddle/book
 
 Then, you would back and paste the address into the local browser:
 
-.. code-block:: text
+  .. code-block:: text
 
-    http://localhost:8888/
+     http://localhost:8888/
 
 That's all. Enjoy your journey!
 
+.. _docker_run_gpu:
 
-Documentation
--------------
+Train with Docker with GPU
+------------------------------
 
-Paddle Docker images include an HTML version of C++ source code
-generated using `woboq code browser
-<https://github.com/woboq/woboq_codebrowser>`_.  This makes it easy
-for users to browse and understand the C++ source code.
+We recommend using
+`nvidia-docker <https://github.com/NVIDIA/nvidia-docker>`_
+to run GPU training jobs. Please ensure you have latest
+GPU driver installed before move on.
 
-As long as we give the Paddle Docker container a name, we can run an
-additional Nginx Docker container to serve the volume from the Paddle
-container:
+  .. code-block:: bash
 
-.. code-block:: bash
+     nvidia-docker run -it -v $PWD:/work paddledev/paddle:latest-gpu /bin/bash
 
-   docker run -d --name paddle-cpu-doc paddle:<version>
-   docker run -d --volumes-from paddle-cpu-doc -p 8088:80 nginx
+**NOTE: If you don't have nvidia-docker installed, try the following method to mount CUDA libs and devices into the container.**
 
+  .. code-block:: bash
 
-Then we can direct our Web browser to the HTML version of source code
-at http://localhost:8088/paddle/
+     export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
+     export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
+     docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:latest-gpu
+
+**About AVX:**
+
+AVX is a kind of CPU instruction can accelerate PaddlePaddle's calculations.
+The latest PaddlePaddle Docker image turns AVX on by default, so, if your
+computer doesn't support AVX, you'll probably need to
+`build <./build_from_source_en.rst>`_ with :code:`WITH_AVX=OFF`.
+
+The following command will tell you whether your computer supports AVX.
+
+   .. code-block:: bash
+
+      if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi

doc/getstarted/build_and_install/index_cn.rst

@@ -6,12 +6,13 @@
 安装流程
 ++++++++
 
-PaddlePaddle提供Docker镜像来部署环境。
+PaddlePaddle提供pip和Docker的安装方式：
 
 .. toctree::
    :maxdepth: 1
-   
-   docker_install_cn.rst 
+
+   pip_install_cn.rst
+   docker_install_cn.rst
 
 
 编译流程
@@ -19,9 +20,14 @@ PaddlePaddle提供Docker镜像来部署环境。
 
 ..  warning::
 
-    编译流程主要推荐高级用户查看，普通用户请走安装流程。
+    建议直接使用上述安装流程，方便快速安装。只有在遇到需要独立定制的二进制时才需要编译。
 
 ..  toctree::
     :maxdepth: 1
 
-    cmake/build_from_source_cn.rst
+    build_from_source_cn.rst
+
+常见问题解答
+++++++++++
+
+`常见问题解答 <http://www.paddlepaddle.org/docs/develop/documentation/zh/faq/build_and_install/index_cn.html>`_

doc/getstarted/build_and_install/index_en.rst

 Install and Build
 =================
 
-Install PaddlePaddle
-----------------------
+.. _install_steps:
 
-..  toctree::
-    :maxdepth: 1
+Install Steps
+++++++++
+
+You can choose either pip or Docker to complete your install:
+
+.. toctree::
+   :maxdepth: 1
+
+   pip_install_en.rst
+   docker_install_en.rst
 
-    docker_install_en.rst
 
 Build from Source
 -----------------
 
 ..  warning::
 
-    Please use :code:`docker` image to install paddle. The building guide is used for hacking or contributing PaddlePaddle source code.
+    We recommend to directly install via above installation steps, you'll only need to build PaddlePaddle from source when you need a modifed binary.
 
 ..  toctree::
     :maxdepth: 1
 
     build_from_source_en.md
+
+FAQ
+++++++++++
+
+`FAQ <http://www.paddlepaddle.org/docs/develop/documentation/zh/faq/build_and_install/index_en.html>`_

doc/getstarted/build_and_install/pip_install_cn.rst

+使用pip安装
+================================
+
+PaddlePaddle可以使用常用的Python包管理工具
+`pip <https://pip.pypa.io/en/stable/installing/>`_
+完成安装，并可以在大多数主流的Linux操作系统以及MacOS上执行。
+
+.. _pip_install:
+
+使用pip安装
+------------------------------
+
+
+执行下面的命令即可在当前机器上安装PaddlePaddle的运行时环境，并自动下载安装依赖软件。
+
+  .. code-block:: bash
+
+     pip install paddlepaddle
+
+
+如果需要安装支持GPU的版本，需要执行：
+
+  .. code-block:: bash
+
+     pip install paddlepaddle-gpu
+
+如果需要获取并安装最新的（开发分支）PaddlePaddle，可以从我们的CI系统中下载最新的whl安装包和c-api开发包并安装，
+您可以从下面的表格中找到需要的版本：
+
+如果在点击下面链接时出现如下登陆界面，点击“Log in as guest”即可开始下载：
+
+.. image:: paddleci.png
+   :scale: 50 %
+   :align: center
+
+..  csv-table:: 各个版本最新的whl包
+    :header: "版本说明", "cp27-cp27mu", "cp27-cp27mu", "C-API"
+    :widths: 1, 3, 3, 3
+
+    "cpu_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cpu_avx_openblas", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "暂无"
+    "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+
+.. _pip_dependency:
+
+运行环境依赖
+------------------------------
+
+PaddlePaddle安装包由于不仅仅包含.py程序，而且包含了C++编写的部分，所以我们确保发布的二进制包可以支持主流的Linux操作系统，比如CentOS 6以上，Ubuntu 14.04以上，MacOS 10.12以上。
+
+PaddlePaddle发布的安装包会尽量对齐 `manylinux1 <https://www.python.org/dev/peps/pep-0513/#the-manylinux1-policy>`_ 标准，通常使用CentOS 5作为编译环境。但由于CUDA库通常需要CentOS 6以上，而且CentOS 5即将停止维护，所以我们默认使用CentOS 6作为标准编译环境。
+
+.. csv-table:: PaddlePaddle环境依赖
+   :header: "依赖", "版本", "说明"
+   :widths: 10, 15, 30
+
+   "操作系统", "Linux, MacOS", "CentOS 6以上，Ubuntu 14.04以上，MacOS 10.12以上"
+   "Python", "2.7.x", "暂时不支持Python3"
+   "libc.so", "GLIBC_2.7", "glibc至少包含GLIBC_2.7以上的符号"
+   "libstdc++.so", "GLIBCXX_3.4.11, CXXABI_1.3.3", "至少包含GLIBCXX_3.4.11, CXXABI_1.3.3以上的符号"
+   "libgcc_s.so", "GCC_3.3", "至少包含GCC_3.3以上的符号"
+
+.. _pip_faq:
+
+安装常见问题和解决方法
+------------------------------
+
+- paddlepaddle*.whl is not a supported wheel on this platform.
+  
+  出现这个问题的主要原因是，没有找到和当前系统匹配的paddlepaddle安装包。请检查Python版本是否为2.7系列。另外最新的pip官方源中的安装包默认是manylinux1标准，需要使用最新的pip (>9.0.0) 才可以安装。可以使用下面的命令更新您的pip：
+
+    .. code-block:: bash
+
+       pip install --upgrade pip
+
+  如果仍然存在问题，可以执行：
+
+      .. code-block:: bash
+
+         python -c "import pip; print(pip.pep425tags.get_supported())"
+
+  获取当前系统支持的安装包格式，并检查和需安装的包是否匹配。pypi安装包可以在 `这个 <https://pypi.python.org/pypi/paddlepaddle/0.10.5>`_ 链接中找到。
+
+  如果系统支持的是 linux_x86_64 而安装包是 manylinux1_x86_64 ，需要升级pip版本到最新； 如果系统支持 manylinux1_x86_64 而安装包（本地）是 linux_x86_64 ，可以重命名这个whl包为 manylinux1_x86_64 再安装。
\ No newline at end of file

doc/getstarted/build_and_install/pip_install_en.rst

+Install Using pip
+================================
+
+You can use current widely used Python package management
+tool `pip <https://pip.pypa.io/en/stable/installing/>`_
+to install PaddlePaddle. This method can be used in
+most of current Linux systems or MacOS.
+
+.. _pip_install:
+
+Install Using pip
+------------------------------
+
+Run the following command to install PaddlePaddle on the current
+machine, it will also download requirements.
+
+  .. code-block:: bash
+
+     pip install paddlepaddle
+
+
+If you wish to install GPU version, just run:
+
+  .. code-block:: bash
+
+     pip install paddlepaddle-gpu
+
+If you wish to install the latest develop branch PaddlePaddle, 
+you can download the latest whl package from our CI system. Access
+the below links, log in as guest, then click at the "Artifact"
+tab, you'll find the download link of whl packages.
+
+If the links below shows up the login form, just click "Log in as guest" to start the download:
+
+.. image:: paddleci.png
+   :scale: 50 %
+   :align: center
+
+..  csv-table:: whl package of each version
+    :header: "version", "cp27-cp27mu", "cp27-cp27mu", "C-API"
+    :widths: 1, 3, 3, 3
+
+    "cpu_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cpu_avx_openblas", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "Not Available"
+    "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+
+.. _pip_dependency:
+
+Runtime Dependency
+------------------------------
+
+PaddlePaddle installation packages (whl) does not only contain .py files,
+but also binaries built from C++ code. We ensure that PaddlePaddle can
+run on current mainline Linux distributions, like CentOS 6, Ubuntu 14.04
+and MacOS 10.12.
+
+PaddlePaddle whl packages are trying to satisfy
+`manylinux1 <https://www.python.org/dev/peps/pep-0513/#the-manylinux1-policy>`_
+standard, which uses CentOS 5 as default build environment. But CUDA libraries
+seems only run on CentOS 6 at least, also, CentOS 5 is about to end its lifetime,
+so we use CentOS 6 as default build environment.
+
+.. csv-table:: PaddlePaddle Runtime Deps
+   :header: "Dependency", "version", "description"
+   :widths: 10, 15, 30
+
+   "OS", "Linux, MacOS", "CentOS 6 or later，Ubuntu 14.04 or later，MacOS 10.12 or later"
+   "Python", "2.7.x", "Currently Python3 is not supported"
+   "libc.so", "GLIBC_2.7", "glibc at least include GLIBC_2.7 symbols"
+   "libstdc++.so", "GLIBCXX_3.4.11, CXXABI_1.3.3", "At least include GLIBCXX_3.4.11, CXXABI_1.3.3 symbols"
+   "libgcc_s.so", "GCC_3.3", "At least include GCC_3.3 symbols"
+
+.. _pip_faq:
+
+FAQ
+------------------------------
+
+- paddlepaddle*.whl is not a supported wheel on this platform.
+  
+  The main cause of this issue is that your current platform is
+  not supported. Please check that you are using Python 2.7 series.
+  Besides, pypi only supports manylinux1 standard, you'll need to
+  upgrade your pip to >9.0.0. Then run the below command:
+
+    .. code-block:: bash
+
+       pip install --upgrade pip
+
+  If the problem still exists, run the following command:
+
+      .. code-block:: bash
+
+         python -c "import pip; print(pip.pep425tags.get_supported())"
+
+  Then you'll get supported package suffixes, then check if it matches
+  the file name of the whl package. You can find default whl package at
+  `here <https://pypi.python.org/pypi/paddlepaddle/0.10.5>`_
+
+  If your system supports linux_x86_64 but the whl package is manylinux1_x86_64,
+  you'll need to update pip to the latest version; If your system supports
+  manylinux1_x86_64 but the whl package is linux_x86_64 you can rename the
+  file to manylinux1_x86_64 suffix and then install.

doc/getstarted/index_cn.rst

 新手入门
 ============
 
+.. _quick_install:
+
+快速安装
+++++++++
+
+PaddlePaddle支持使用pip快速安装，目前支持CentOS 6以上, Ubuntu 14.04以及MacOS 10.12，并安装有Python2.7。
+执行下面的命令完成快速安装：
+
+  .. code-block:: bash
+
+     pip install paddlepaddle
+
+如果需要安装支持GPU的版本，需要执行：
+
+  .. code-block:: bash
+
+     pip install paddlepaddle-gpu
+
+更详细的安装和编译方法参考：
+
 ..  toctree::
   :maxdepth: 1
 
   build_and_install/index_cn.rst
-  concepts/use_concepts_cn.rst
 
-- `深度学习入门课程 <http://book.paddlepaddle.org/index.cn.html>`_
+.. _quick_start:
+
+快速开始
+++++++++
+
+创建一个 housing.py 并粘贴此Python代码：
+
+  .. code-block:: python
+
+     import paddle.v2 as paddle
+
+     # Initialize PaddlePaddle.
+     paddle.init(use_gpu=False, trainer_count=1)
+
+     # Configure the neural network.
+     x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13))
+     y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
+
+     # Infer using provided test data.
+     probs = paddle.infer(
+         output_layer=y_predict,
+         parameters=paddle.dataset.uci_housing.model(),
+         input=[item for item in paddle.dataset.uci_housing.test()()])
+
+     for i in xrange(len(probs)):
+         print 'Predicted price: ${:,.2f}'.format(probs[i][0] * 1000)
+
+执行 :code:`python housing.py` 瞧！ 它应该打印出预测住房数据的清单。
+
+..  toctree::
+  :maxdepth: 1
+
+  concepts/use_concepts_cn.rst

doc/getstarted/index_en.rst

 GET STARTED
 ============
 
+.. _quick_install:
+
+Quick Install
+----------------------
+
+You can use pip to install PaddlePaddle with a single command, supports
+CentOS 6 above, Ubuntu 14.04 above or MacOS 10.12, with Python 2.7 installed.
+Simply run the following command to install:
+
+  .. code-block:: bash
+
+     pip install paddlepaddle
+
+If you need to install GPU version, run:
+
+  .. code-block:: bash
+
+     pip install paddlepaddle-gpu
+
+For more details about installation and build:
+
 ..  toctree::
   :maxdepth: 1
 
   build_and_install/index_en.rst
 
-- `Deep Learning 101 <http://book.paddlepaddle.org/index.html>`_
+
+.. _quick_start:
+
+Quick Start
+++++++++
+
+Create a new file called housing.py, and paste this Python
+code:
+
+
+  .. code-block:: python
+
+     import paddle.v2 as paddle
+
+     # Initialize PaddlePaddle.
+     paddle.init(use_gpu=False, trainer_count=1)
+
+     # Configure the neural network.
+     x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13))
+     y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
+
+     # Infer using provided test data.
+     probs = paddle.infer(
+         output_layer=y_predict,
+         parameters=paddle.dataset.uci_housing.model(),
+         input=[item for item in paddle.dataset.uci_housing.test()()])
+
+     for i in xrange(len(probs)):
+         print 'Predicted price: ${:,.2f}'.format(probs[i][0] * 1000)
+
+Run :code:`python housing.py` and voila! It should print out a list of predictions
+for the test housing data.

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
 

doc/howto/optimization/cpu_profiling.md

@@ -71,7 +71,7 @@ cprofilev -a 0.0.0.0 -p 3214 -f profile.out main.py
 
 ```
 
-可以看到最耗时的函数是C++端的`run`函数。这需要联合我们第二节`Python与C++混合代码的性能分析`来进行调优。而`sync_with_cpp`函数的总共耗时很长，每次调用的耗时也很长。于是我们可以点击`sync_with_cpp`的详细信息，了解其调用关系。
+可以看到最耗时的函数是C++端的`run`函数。这需要联合我们第二节`Python`与`C++`混合代码的性能分析来进行调优。而`sync_with_cpp`函数的总共耗时很长，每次调用的耗时也很长。于是我们可以点击`sync_with_cpp`的详细信息，了解其调用关系。
 
 ```text
 Called By:
@@ -121,7 +121,7 @@ python -m yep -v main.py
 
 1. 编译时指定`-g`生成调试信息。使用cmake的话，可以将CMAKE_BUILD_TYPE指定为`RelWithDebInfo`。
 2. 编译时一定要开启优化。单纯的`Debug`编译性能会和`-O2`或者`-O3`有非常大的差别。`Debug`模式下的性能测试是没有意义的。
-3. 运行性能分析的时候，先从单线程开始，再开启多线程，进而多机。毕竟如果单线程调试更容易。可以设置`OMP_NUM_THREADS=1`这个环境变量关闭openmp优化。
+3. 运行性能分析的时候，先从单线程开始，再开启多线程，进而多机。毕竟单线程调试更容易。可以设置`OMP_NUM_THREADS=1`这个环境变量关闭openmp优化。
 
 ### 查看性能分析文件
 

paddle/capi/Matrix.cpp

@@ -55,7 +55,7 @@ paddle_error paddle_matrix_set_row(paddle_matrix mat,
 }
 
 PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat,
-                                          paddle_real* value) {
+                                            paddle_real* value) {
   if (mat == nullptr || value == nullptr) return kPD_NULLPTR;
   auto ptr = cast(mat);
   if (ptr->mat == nullptr) return kPD_NULLPTR;
@@ -75,7 +75,7 @@ PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat,
 }
 
 PD_API paddle_error paddle_matrix_get_value(paddle_matrix mat,
-                                          paddle_real* result) {
+                                            paddle_real* result) {
   if (mat == nullptr || result == nullptr) return kPD_NULLPTR;
   auto ptr = cast(mat);
   if (ptr->mat == nullptr) return kPD_NULLPTR;

paddle/capi/matrix.h

@@ -79,7 +79,7 @@ PD_API paddle_error paddle_matrix_set_row(paddle_matrix mat,
  * @note  value should contain enough element of data to init the mat
  */
 PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat,
-                                          paddle_real* value);
+                                            paddle_real* value);
 
 /**
  * @brief PDMatGetRow Get raw row buffer from matrix
@@ -93,14 +93,14 @@ PD_API paddle_error paddle_matrix_get_row(paddle_matrix mat,
                                           paddle_real** rawRowBuffer);
 
 /**
- * @brief copy data from the matrix 
+ * @brief copy data from the matrix
  * @param [in] mat Target matrix
- * @param [out] result pointer to store the matrix data 
+ * @param [out] result pointer to store the matrix data
  * @return paddle_error
  * @note the space of the result should allocated before invoke this API
  */
 PD_API paddle_error paddle_matrix_get_value(paddle_matrix mat,
-                                          paddle_real* result);
+                                            paddle_real* result);
 /**
  * @brief PDMatCreateNone Create None Matrix
  * @return

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/framework/op_desc.cc

@@ -65,7 +65,7 @@ class CompileTimeInferShapeContext : public InferShapeContext {
     PADDLE_ENFORCE_EQ(in_var->GetType(), VarDesc::LOD_TENSOR,
                       "The %d-th output of Output(%s) must be LoDTensor.", j,
                       out);
-    in_var->SetLoDLevel(out_var->GetLodLevel());
+    out_var->SetLoDLevel(in_var->GetLodLevel());
   }
   bool IsRuntime() const override;
 

paddle/framework/tensor_util.h

@@ -135,18 +135,17 @@ inline void CopyToVector(const Tensor& src, const platform::DeviceContext& ctx,
   auto dst_ptr = static_cast<void*>(dst->data());
 
   if (platform::is_cpu_place(src.place())) {
-    memory::Copy(dst_place, dst_ptr, boost::get<platform::CPUPlace>(src.place()), 
-		    src_ptr, size);
+    memory::Copy(dst_place, dst_ptr,
+                 boost::get<platform::CPUPlace>(src.place()), src_ptr, size);
   }
 #ifdef PADDLE_WITH_CUDA
   else if (platform::is_gpu_place(src.place())) {  // NOLINT
     memory::Copy(
-        dst_place, dst_ptr, boost::get<platform::GPUPlace>(src.place()), src_ptr,
-        size,
+        dst_place, dst_ptr, boost::get<platform::GPUPlace>(src.place()),
+        src_ptr, size,
         reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
   }
 #endif
-
 }
 
 }  // namespace framework

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/CMakeLists.txt

@@ -62,11 +62,11 @@ if(NOT WITH_DOUBLE AND NOT MOBILE_INFERENCE)
 endif()
 
 if(NOT MOBILE_INFERENCE)
-################## test_Evaluator #######################
+    ################## test_Evaluator #######################
     add_unittest(test_Evaluator
         test_Evaluator.cpp)
       
-############### test_RecurrentGradientMachine ###############
+    ############### test_RecurrentGradientMachine ###############
     # TODO(yuyang18): There is some bug in test_RecurrentGradientMachine
     # I will fix it.
     add_unittest_without_exec(test_RecurrentGradientMachine
@@ -77,7 +77,7 @@ if(NOT MOBILE_INFERENCE)
                 ${CMAKE_CURRENT_BINARY_DIR}/test_RecurrentGradientMachine
         WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
       
-############### test_NetworkCompare ###############
+    ############### test_NetworkCompare ###############
     add_unittest_without_exec(test_NetworkCompare
         test_NetworkCompare.cpp)
     if(WITH_GPU)
@@ -89,34 +89,33 @@ if(NOT MOBILE_INFERENCE)
             COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=false
             WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
     endif()
-endif()
 
+    ################# test_CompareSparse ##################
+    add_unittest_without_exec(test_CompareSparse
+        test_CompareSparse.cpp)
+    if(NOT ON_TRAVIS)
+      add_test(NAME test_CompareSparse
+        COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d
+              ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
+                  ./.set_port.sh -p port -n 6
+                      ${CMAKE_CURRENT_BINARY_DIR}/test_CompareSparse
+        WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
+    endif()
+
+    ################ test_CompareTwoNets ######################
+    add_unittest_without_exec(test_CompareTwoNets
+        test_CompareTwoNets.cpp)
+    add_test(NAME test_CompareTwoNets
+      COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d
+            ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
+            ${CMAKE_CURRENT_BINARY_DIR}/test_CompareTwoNets
+        WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
+endif()
 
+################ test_PyDataProvider2 ######################
 add_unittest_without_exec(test_PyDataProvider2
         test_PyDataProvider2.cpp)
-
 add_test(NAME test_PyDataProvider2
    COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/paddle/gserver/tests:${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider2
         WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle
 )
-
-################# test_CompareSparse ##################
-add_unittest_without_exec(test_CompareSparse
-    test_CompareSparse.cpp)
-if(NOT ON_TRAVIS)
-  add_test(NAME test_CompareSparse
-    COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d
-          ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
-              ./.set_port.sh -p port -n 6
-                  ${CMAKE_CURRENT_BINARY_DIR}/test_CompareSparse
-    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
-endif()
-
-################ test_CompareTwoNets ######################
-add_unittest_without_exec(test_CompareTwoNets
-    test_CompareTwoNets.cpp)
-add_test(NAME test_CompareTwoNets
-  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d
-        ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
-        ${CMAKE_CURRENT_BINARY_DIR}/test_CompareTwoNets
-    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/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
-  testLayerGrad(config,
-                "hsigmoid",
-                100,
-                /* trans */ false, /* useGpu */
-                false);
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config,
+                  "hsigmoid",
+                  100,
+                  /* trans */ false,
+                  /* useGpu */ useGpu);
+  }
 }
 
 TEST(Layer, multi_cross) {
@@ -2464,6 +2465,25 @@ TEST(Layer, L2DistanceLayer) {
   }
 }
 
+void testFactorizationMachineLayer(InputType type, bool useGpu) {
+  const int FACTOR_SIZE = 10;
+  TestConfig config;
+  config.layerConfig.set_type("factorization_machine");
+  config.layerConfig.set_factor_size(FACTOR_SIZE);
+  config.layerConfig.set_size(1);
+  config.biasSize = 0;
+  config.inputDefs.push_back({type, "layer_0", 128, 1280});
+  config.layerConfig.add_inputs();
+  testLayerGrad(config, "factorization_machine", 16, false, useGpu, false);
+}
+
+TEST(Layer, FactorizationMachineLayer) {
+  for (auto useGpu : {false, true}) {
+    testFactorizationMachineLayer(INPUT_DATA, useGpu);
+  }
+  testFactorizationMachineLayer(INPUT_SPARSE_FLOAT_VALUE_DATA, false);
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

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

@@ -191,6 +191,7 @@ set(DEPS_OPS
     sum_op
     pool_op
     maxout_op
+    unpool_op
     pool_with_index_op
     conv_op
     conv_transpose_op
@@ -205,8 +206,24 @@ 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)
@@ -219,6 +236,7 @@ op_library(adagrad_op DEPS selected_rows_functor)
 op_library(conv_op DEPS vol2col)
 op_library(pool_op DEPS pooling)
 op_library(maxout_op DEPS maxouting)
+op_library(unpool_op DEPS unpooling)
 op_library(pool_with_index_op DEPS pooling)
 op_library(lod_rank_table_op SRCS lod_rank_table_op.cc DEPS lod_rank_table)
 op_library(lod_tensor_to_array_op SRCS lod_tensor_to_array_op.cc DEPS lod_rank_table_op)
@@ -235,6 +253,10 @@ 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 +264,8 @@ 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 +275,4 @@ 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

@@ -29,14 +29,21 @@ void ExtractNCWHD(const framework::DDim &dims,
                   const TensorFormat &tensor_format, int *N, int *C, int *H,
                   int *W, int *D) {
   *N = dims[0];
-  *C = tensor_format == TensorFormat::NCHW ? dims[1] : dims[dims.size() - 1];
-  *H = tensor_format == TensorFormat::NCHW ? dims[2] : dims[1];
-  *W = dims.size() > 3
-           ? (tensor_format == TensorFormat::NCHW ? dims[3] : dims[2])
-           : 1;
-  *D = dims.size() > 4
-           ? (tensor_format == TensorFormat::NCHW ? dims[4] : dims[3])
-           : 1;
+  if (dims.size() == 2) {
+    *C = dims[1];
+    *H = 1;
+    *W = 1;
+    *D = 1;
+  } else {
+    *C = tensor_format == TensorFormat::NCHW ? dims[1] : dims[dims.size() - 1];
+    *H = tensor_format == TensorFormat::NCHW ? dims[2] : dims[1];
+    *W = dims.size() > 3
+             ? (tensor_format == TensorFormat::NCHW ? dims[3] : dims[2])
+             : 1;
+    *D = dims.size() > 4
+             ? (tensor_format == TensorFormat::NCHW ? dims[4] : dims[3])
+             : 1;
+  }
 }
 
 template <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",
-           "(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, "
-           "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.");
+  AddInput(
+      "Filter",
+      "(Tensor) The filter tensor of convolution transpose operator. "
+      "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 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:
+
+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 {}

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/CMakeLists.txt

@@ -13,8 +13,9 @@ if(WITH_GPU)
     nv_library(context_project SRCS context_project.cc context_project.cu DEPS device_context math_function)
     nv_library(sequence2batch SRCS sequence2batch.cc sequence2batch.cu DEPS device_context)
     nv_library(lstm_compute SRCS lstm_compute.cc lstm_compute.cu DEPS device_context activation_functions)
-    nv_library(gru_compute SRCS gru_compute.cc gru_compute.cu DEPS device_context activation_functions math_function)
     nv_library(maxouting SRCS maxouting.cc maxouting.cu DEPS device_context)
+    nv_library(unpooling SRCS unpooling.cc unpooling.cu DEPS device_context)
+    nv_library(gru_compute SRCS gru_compute.cc gru_compute.cu DEPS device_context activation_functions math_function)
 else()
     cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context framework_proto)
     cc_library(selected_rows_functor SRCS selected_rows_functor.cc DEPS selected_rows math_function)
@@ -26,8 +27,9 @@ else()
     cc_library(context_project SRCS context_project.cc DEPS device_context math_function)
     cc_library(sequence2batch SRCS sequence2batch.cc DEPS device_context)
     cc_library(lstm_compute SRCS lstm_compute.cc DEPS device_context activation_functions)
-    cc_library(gru_compute SRCS gru_compute.cc DEPS device_context activation_functions math_function)
     cc_library(maxouting SRCS maxouting.cc DEPS device_context)
+    cc_library(unpooling SRCS unpooling.cc DEPS device_context)
+    cc_library(gru_compute SRCS gru_compute.cc DEPS device_context activation_functions math_function)
 endif()
 
 cc_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor)

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