Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into hsigmoid_gpu

bb61e90f · peterzhang2029 · 3022a790 · 0a8a86e0 · bb61e90f · bb61e90f
39 changed file
--- a/.clang-format
+++ b/.clang-format
@@ -25,4 +25,3 @@ AllowAllParametersOfDeclarationOnNextLine: true
 BinPackParameters: false
 BinPackArguments: false
 ...
--- a/.travis.yml
+++ b/.travis.yml
@@ -42,7 +42,7 @@ before_install:
 script:
  - |
    timeout 2580 paddle/scripts/travis/${JOB}.sh # 43min timeout
-    RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true; else false; fi;
+    RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true ;else exit 1; fi;
  - |
    if [[ "$JOB" != "build_doc" ]]; then exit 0; fi;
    if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;

--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -133,6 +133,8 @@ include(external/any)       # download libn::any
 include(external/eigen)     # download eigen3
 include(external/pybind11)  # download pybind11
 include(external/nccl)
+include(external/cares)
+include(external/grpc)
 include(cudnn)              # set cudnn libraries, must before configure
 include(configure)          # add paddle env configuration

--- a/Dockerfile
+++ b/Dockerfile
@@ -29,7 +29,7 @@ RUN apt-get update && \
    automake locales clang-format swig doxygen cmake  \
    liblapack-dev liblapacke-dev libboost-dev \
    clang-3.8 llvm-3.8 libclang-3.8-dev \
-    net-tools && \
+    net-tools libtool && \
    apt-get clean -y
 # Install Go and glide

--- a/cmake/external/cares.cmake
+++ b/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)
--- a/cmake/external/grpc.cmake
+++ b/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)
+ExternalProject_Add(
+    extern_grpc
+    DEPENDS protobuf zlib
+    GIT_REPOSITORY "https://github.com/grpc/grpc.git"
+    GIT_TAG "v1.7.x"
+    PREFIX          ${GRPC_SOURCES_DIR}
+    UPDATE_COMMAND  ""
+    CONFIGURE_COMMAND ""
+    BUILD_IN_SOURCE 1
+    BUILD_COMMAND   make
+    INSTALL_COMMAND make prefix=${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)
--- a/cmake/external/zlib.cmake
+++ b/cmake/external/zlib.cmake
@@ -50,6 +50,8 @@ ExternalProject_Add(
 )
 LIST(APPEND external_project_dependencies zlib)
+ADD_LIBRARY(zlib_target STATIC IMPORTED GLOBAL)
+SET_PROPERTY(TARGET zlib_target PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES})
 IF(WITH_C_API)
  INSTALL(DIRECTORY ${ZLIB_INCLUDE_DIR} DESTINATION third_party/zlib)

--- a/cmake/generic.cmake
+++ b/cmake/generic.cmake
@@ -467,3 +467,50 @@ function(py_test TARGET_NAME)
             WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
  endif()
 endfunction()
+# grpc_library generate grpc code using grpc_cpp_plugin and protoc
+# then build the generated protobuf code and grpc code with your
+# implementation source codes together. Use SRCS argument for your
+# implementation source files and PROTO argument for your .proto
+# files.
+#
+# Usage: grpc_library(my_target SRCS my_client.cc PROTO my_target.proto DEPS my_dep)
+function(grpc_library TARGET_NAME)
+  set(oneValueArgs PROTO)
+  set(multiValueArgs SRCS DEPS)
+  set(options "")
+  cmake_parse_arguments(grpc_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+  message(STATUS "generating grpc ${grpc_library_PROTO}")
+  get_filename_component(ABS_PROTO ${grpc_library_PROTO} ABSOLUTE)
+  get_filename_component(PROTO_WE ${grpc_library_PROTO} NAME_WE)
+  get_filename_component(PROTO_PATH ${ABS_PROTO} PATH)
+  protobuf_generate_cpp(grpc_proto_srcs grpc_proto_hdrs "${ABS_PROTO}")
+  set(grpc_grpc_srcs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.cc")
+  set(grpc_grpc_hdrs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.h")
+  cc_library("${TARGET_NAME}_proto" SRCS "${grpc_proto_srcs}")
+  add_custom_command(
+          OUTPUT "${grpc_grpc_srcs}" "${grpc_grpc_hdrs}"
+          COMMAND ${PROTOBUF_PROTOC_EXECUTABLE}
+          ARGS --grpc_out "${CMAKE_CURRENT_BINARY_DIR}" -I "${PROTO_PATH}"
+          --plugin=protoc-gen-grpc="${GRPC_CPP_PLUGIN}" "${ABS_PROTO}"
+          DEPENDS "${ABS_PROTO}" ${PROTOBUF_PROTOC_EXECUTABLE} extern_grpc)
+  # FIXME(typhoonzero): grpc generated code do not generate virtual-dtor, mark it
+  # as compiler warnings instead of error. Should try remove the warnings also.
+  set_source_files_properties(
+    ${grpc_grpc_srcs}
+    PROPERTIES
+    COMPILE_FLAGS  "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
+  cc_library("${TARGET_NAME}_grpc" SRCS "${grpc_grpc_srcs}")
+  set_source_files_properties(
+    ${grpc_library_SRCS}
+    PROPERTIES
+    COMPILE_FLAGS  "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
+  cc_library("${TARGET_NAME}" SRCS "${grpc_library_SRCS}" DEPS "${TARGET_NAME}_grpc" "${TARGET_NAME}_proto" "${grpc_library_DEPS}")
+endfunction()
--- a/doc/api/v2/config/layer.rst
+++ b/doc/api/v2/config/layer.rst
@@ -54,7 +54,7 @@ img_conv
 ..  _api_v2.layer_context_projection:
-context_projection 
+context_projection
 ------------------
 ..  autoclass:: paddle.v2.layer.context_projection
    :noindex:
@@ -70,7 +70,7 @@ Image Pooling Layer
 img_pool
 --------
 ..  autoclass:: paddle.v2.layer.img_pool
-    :noindex:   
+    :noindex:
 spp
 ---
@@ -104,7 +104,7 @@ sum_to_one_norm
 ---------------
 ..  autoclass:: paddle.v2.layer.sum_to_one_norm
    :noindex:
 cross_channel_norm
 ------------------
 ..  autoclass:: paddle.v2.layer.cross_channel_norm
@@ -114,7 +114,7 @@ row_l2_norm
 -----------
 ..  autoclass:: paddle.v2.layer.row_l2_norm
    :noindex:
 Recurrent Layers
 ================
@@ -415,6 +415,13 @@ multiplex
 ..  autoclass:: paddle.v2.layer.multiplex
    :noindex:
+Factorization Machine Layer
+============================
+factorization_machine
+---------------------
+..  autoclass:: paddle.v2.layer.factorization_machine
+    :noindex:
 Slicing and Joining Layers
 ==========================

--- a/paddle/framework/lod_tensor.cc
+++ b/paddle/framework/lod_tensor.cc
@@ -13,6 +13,8 @@
   limitations under the License. */
 #include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/data_type.h"
+#include "paddle/framework/framework.pb.h"
 #include "paddle/memory/memcpy.h"
 #include "paddle/memory/memory.h"
@@ -27,11 +29,11 @@
 namespace paddle {
 namespace framework {
-std::ostream& operator<<(std::ostream& os, const LoD& lod) {
+std::ostream &operator<<(std::ostream &os, const LoD &lod) {
  os << "{";
-  for (auto& v : lod) {
+  for (auto &v : lod) {
    os << "{";
-    for (auto& i : v) {
+    for (auto &i : v) {
      os << i << ",";
    }
    os << "}";
@@ -41,7 +43,7 @@ std::ostream& operator<<(std::ostream& os, const LoD& lod) {
  return os;
 }
-LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) {
+LoD SliceLevels(const LoD &in, size_t level_begin, size_t level_end) {
  LoD new_lod;
  new_lod.reserve(level_end - level_begin);
  for (size_t i = level_begin; i < level_end; i++) {
@@ -53,7 +55,7 @@ LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) {
  return new_lod;
 }
-LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
+LoD SliceInLevel(const LoD &in, size_t level, size_t elem_begin,
                 size_t elem_end) {
  PADDLE_ENFORCE_LT(level, in.size());
  PADDLE_ENFORCE_LT(elem_end, in[level].size());
@@ -64,9 +66,9 @@ LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
  res[0].assign(in[level].begin() + elem_begin,
                in[level].begin() + elem_end + 1);
  for (size_t lvl = 1; lvl < res.size(); lvl++) {
-    const auto& in_level = in[level + lvl];
+    const auto &in_level = in[level + lvl];
-    const auto& above_level = res[lvl - 1];
+    const auto &above_level = res[lvl - 1];
-    auto& out_level = res[lvl];
+    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 &a_level = a[i];
-    const auto& b_level = b[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
--- a/paddle/framework/lod_tensor.h
+++ b/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
--- a/paddle/gserver/layers/FactorizationMachineLayer.cpp
+++ b/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
--- a/paddle/gserver/layers/FactorizationMachineLayer.h
+++ b/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
--- a/paddle/gserver/tests/test_LayerGrad.cpp
+++ b/paddle/gserver/tests/test_LayerGrad.cpp
@@ -2465,6 +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);

--- a/paddle/math/CpuSparseMatrix.cpp
+++ b/paddle/math/CpuSparseMatrix.cpp
@@ -260,6 +260,35 @@ void CpuSparseMatrix::printOneRow(std::ostream& os, size_t idx) const {
  os << ";";
 }
+void CpuSparseMatrix::rowScale(size_t cCol, CpuSparseMatrix& b, Matrix& c) {
+  CHECK(getFormat() != SPARSE_CSC) << "Not supported";
+  CHECK_EQ(height_, b.getHeight());
+  CHECK_EQ(width_, b.getWidth());
+  real* A = getValue();
+  real* B = b.getValue();
+  if (b.getValueType() == FLOAT_VALUE) {
+    for (size_t i = 0; i < height_; i++) {
+      size_t start = getRowStartIdx(i);
+      size_t end = getRowStartIdx(i + 1);
+      CHECK_EQ(start, b.getRowStartIdx(i));
+      CHECK_EQ(end, b.getRowStartIdx(i + 1));
+      for (size_t j = start; j < end; j++) {
+        A[j] = B[j] * c.getElement(i, cCol);
+      }
+    }
+  } else if (b.getValueType() == NO_VALUE) {
+    for (size_t i = 0; i < height_; i++) {
+      size_t start = getRowStartIdx(i);
+      size_t end = getRowStartIdx(i + 1);
+      CHECK_EQ(start, b.getRowStartIdx(i));
+      CHECK_EQ(end, b.getRowStartIdx(i + 1));
+      for (size_t j = start; j < end; j++) {
+        A[j] = c.getElement(i, cCol);
+      }
+    }
+  }
+}
 void CpuSparseMatrix::randomizeUniform() {
  CHECK_LE(elementCnt_, height_ * width_);
  if (valueType_ == FLOAT_VALUE) {

--- a/paddle/math/CpuSparseMatrix.h
+++ b/paddle/math/CpuSparseMatrix.h
@@ -239,6 +239,15 @@ public:
              const unsigned int* cols,
              const real* values);
+  /**
+   * @brief this_row = b_row * c_row[cCol]
+   *
+   * @param[in]  cCol   the column of matrix c used to scale each row of b
+   * @param[in]  b      CpuSparseMatrix
+   * @param[in]  c      Matrix
+   */
+  void rowScale(size_t cCol, CpuSparseMatrix& b, Matrix& c);
  void randomizeUniform();
  void copyFrom(const GpuSparseMatrix& src, hl_stream_t stream);

--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -205,8 +205,24 @@ set(DEPS_OPS
    tensor_array_read_write_op
    gru_op
    adagrad_op
-    sgd_op)
+    sgd_op
+    save_op
+    load_op
+    send_op
+    recv_op)
+add_subdirectory(detail)
+op_library(send_op SRCS send_op.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf)
+set_source_files_properties(
+    send_op.cc
+    PROPERTIES
+    COMPILE_FLAGS  "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
+op_library(recv_op SRCS recv_op.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf)
+set_source_files_properties(
+    recv_op.cc
+    PROPERTIES
+    COMPILE_FLAGS  "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
 op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op)
 op_library(cross_entropy_op DEPS cross_entropy)
@@ -235,6 +251,10 @@ op_library(conv_transpose_op DEPS vol2col)
 op_library(gru_op DEPS sequence2batch gru_compute)
 op_library(recurrent_op SRCS recurrent_op.cc DEPS executor)
+# FIXME(typhoonzero): save/load depends lodtensor serialization functions
+op_library(save_op DEPS lod_tensor)
+op_library(load_op DEPS lod_tensor)
 list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 foreach(src ${GENERAL_OPS})
    op_library(${src})
@@ -242,6 +262,8 @@ endforeach()
 set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
 cc_test(gather_test SRCS gather_test.cc DEPS tensor)
 cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
 cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
@@ -251,3 +273,4 @@ if(WITH_GPU)
  cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
 endif()
 cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)
+cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS send_op recv_op sum_op executor)
--- a/paddle/operators/detail/CMakeLists.txt
+++ b/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)
--- a/paddle/operators/detail/recv_impl.cc
+++ b/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
--- a/paddle/operators/detail/send_impl.cc
+++ b/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
--- a/paddle/operators/detail/send_recv.proto
+++ b/paddle/operators/detail/send_recv.proto
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+   http://www.apache.org/licenses/LICENSE-2.0
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+syntax = "proto3";
+package sendrecv;
+service SendRecvService {
+  // For parameter server round-robin like hashing, do not split tensors. 
+  // Send and recv only one tensor
+  rpc SendVariable(VariableMessage) returns (VariableMessage) {}
+}
+// VariableMessage is serialized paddle variable message.
+// It can be:
+// Tensor
+// LoDTensor
+// SelectedRows
+message VariableMessage {
+  string varname = 1;
+  bytes serialized = 2;
+}
+message VoidMessage {
+}
\ No newline at end of file
--- a/paddle/operators/detail/send_recv_impl.h
+++ b/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
--- a/paddle/operators/detail/simple_block_queue.h
+++ b/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
--- a/paddle/operators/load_op.cc
+++ b/paddle/operators/load_op.cc
@@ -38,61 +38,7 @@ class LoadOp : public framework::OperatorBase {
                   out_var_name);
    auto *tensor = out_var->GetMutable<framework::LoDTensor>();
+    framework::DeserializeFromStream(fin, tensor);
-    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;
-      }
-    }
    auto place = dev_ctx.GetPlace();
    if (platform::is_gpu_place(place)) {

--- a/paddle/operators/recv_op.cc
+++ b/paddle/operators/recv_op.cc
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+   http://www.apache.org/licenses/LICENSE-2.0
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+#include <stdint.h>
+#include <sys/stat.h>
+#include <ostream>
+#include <thread>
+#include <unistd.h>
+#include "paddle/framework/data_type.h"
+#include "paddle/framework/executor.h"
+#include "paddle/framework/framework.pb.h"
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/detail/send_recv_impl.h"
+#include "paddle/operators/detail/simple_block_queue.h"
+namespace paddle {
+namespace operators {
+void RunServer(Server **rpc_server,
+               std::shared_ptr<detail::SendRecvServerImpl> service,
+               const std::string &server_address) {
+  ServerBuilder builder;
+  builder.AddListeningPort(server_address, grpc::InsecureServerCredentials());
+  builder.RegisterService(service.get());
+  std::unique_ptr<Server> server(builder.BuildAndStart());
+  *rpc_server = server.get();
+  LOG(INFO) << "Server listening on " << server_address << std::endl;
+  server->Wait();
+}
+class RecvOp : public framework::OperatorBase {
+ public:
+  RecvOp(const std::string &type, const framework::VariableNameMap &inputs,
+         const framework::VariableNameMap &outputs,
+         const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {
+    if (!rpc_service_) {
+      rpc_service_.reset(new detail::SendRecvServerImpl());
+      std::string endpoint = Attr<std::string>("endpoint");
+      server_thread_.reset(
+          new std::thread(RunServer, &rpc_server_, rpc_service_, endpoint));
+    }
+  }
+  virtual ~RecvOp() {
+    rpc_server_->Shutdown();
+    server_thread_->join();
+  }
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    // blocking get one var from client.
+    const framework::LoDTensor &t = rpc_service_->Get();
+    framework::Scope &recv_scope = scope.NewScope();
+    // set graph input var
+    auto *var = recv_scope.Var(Input("RX"));
+    auto *tensor = var->GetMutable<framework::LoDTensor>();
+    // FIXME(typhoonzero): do not copy
+    framework::CopyFrom(t, dev_ctx.GetPlace(), dev_ctx, tensor);
+    auto *block = Attr<framework::BlockDescBind *>("OptimizeBlock");
+    auto *program = block->Program();
+    framework::Executor executor(dev_ctx);
+    // Run sub graph to get optimized tensor
+    executor.Run(*program, &recv_scope, block->ID(),
+                 false /*create_local_scope*/);
+    auto *out_var = recv_scope.FindVar("Out");
+    // push back
+    rpc_service_->Push(out_var->Get<framework::LoDTensor>());
+  }
+ protected:
+  // grpc server instance to track status and gracefully shutdown.
+  // borrow an pointer from server thread.
+  Server *rpc_server_{nullptr};
+  // grpc send/recv service implement to register.
+  std::shared_ptr<detail::SendRecvServerImpl> rpc_service_;
+  std::shared_ptr<std::thread> server_thread_;
+};
+class RecvOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  RecvOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("RX", "(Tensor) Input tensor to be saved");
+    AddComment(R"DOC(
+Recv operator
+This operator will recv tensor from send_op
+)DOC");
+    AddAttr<std::string>("endpoint",
+                         "(string, default 127.0.0.1:6164)"
+                         "IP address to listen on.")
+        .SetDefault("127.0.0.1:6164")
+        .AddCustomChecker([](const std::string &ip) { return !ip.empty(); });
+    AddAttr<framework::BlockDescBind *>("OptimizeBlock", "type BlockDescBind*",
+                                        "optimize network run in server");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(recv, ops::RecvOp, ops::RecvOpMaker);
--- a/paddle/operators/roi_pool_op.h
+++ b/paddle/operators/roi_pool_op.h
@@ -133,53 +133,47 @@ class CPUROIPoolGradOpKernel : public framework::OpKernel<T> {
    auto* in = ctx.Input<framework::Tensor>("X");
    auto* rois = ctx.Input<framework::Tensor>("ROIs");
    auto* argmax = ctx.Input<framework::Tensor>("Argmax");
    auto* out_grad =
        ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
-    auto* x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto* in_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
    auto pooled_height = ctx.Attr<int>("pooled_height");
    auto pooled_width = ctx.Attr<int>("pooled_width");
-    if (x_grad) {
+    if (in_grad) {
-      int channels = in->dims()[1];
-      auto in_stride = framework::stride(in->dims());
-      auto roi_stride = framework::stride(rois->dims());
      const int64_t* rois_data = rois->data<int64_t>();
-      int rois_num = rois->dims()[0];
+      const T* out_grad_data = out_grad->data<T>();
+      const int64_t* argmax_data = argmax->data<int64_t>();
-      T* x_grad_data = x_grad->mutable_data<T>(ctx.GetPlace());
+      T* in_grad_data = in_grad->mutable_data<T>(ctx.GetPlace());
      math::SetConstant<Place, T> set_zero;
-      set_zero(ctx.device_context(), x_grad, static_cast<T>(0));
+      set_zero(ctx.device_context(), in_grad, static_cast<T>(0));
-      size_t roi_offset = roi_stride[0];
+      auto in_stride = framework::stride(in->dims());
-      size_t batch_offset = in_stride[0];
+      auto argmax_stride = framework::stride(argmax->dims());
-      size_t channel_offset = in_stride[1];
+      auto roi_stride = framework::stride(rois->dims());
+      auto out_stride = framework::stride(out_grad->dims());
-      const T* out_grad_data = out_grad->data<T>();
+      int rois_num = rois->dims()[0];
-      size_t pool_channel_offset = pooled_height * pooled_width;
+      int channels = in->dims()[1];
-      const int64_t* argmax_data = argmax->data<int64_t>();
-      for (size_t n = 0; n < rois_num; ++n) {
+      for (int n = 0; n < rois_num; ++n) {
-        size_t roi_batch_idx = rois_data[0];
+        int roi_batch_idx = rois_data[0];
-        T* batch_grad_data = x_grad_data + batch_offset * roi_batch_idx;
+        T* batch_grad_data = in_grad_data + roi_batch_idx * in_stride[0];
        for (int c = 0; c < channels; ++c) {
          for (int ph = 0; ph < pooled_height; ++ph) {
            for (int pw = 0; pw < pooled_width; ++pw) {
-              size_t pool_index = ph * pooled_width + pw;
+              int pool_index = ph * pooled_width + pw;
              if (argmax_data[pool_index] >= 0) {
-                size_t index = static_cast<size_t>(argmax_data[pool_index]);
+                auto index = argmax_data[pool_index];
                batch_grad_data[index] += out_grad_data[pool_index];
              }
            }
          }
-          batch_grad_data += channel_offset;
+          batch_grad_data += in_stride[1];
-          out_grad_data += pool_channel_offset;
+          out_grad_data += out_stride[1];
-          argmax_data += pool_channel_offset;
+          argmax_data += argmax_stride[1];
        }
-        rois_data += roi_offset;
+        rois_data += roi_stride[0];
      }
    }
  }

--- a/paddle/operators/save_op.cc
+++ b/paddle/operators/save_op.cc
@@ -88,73 +88,7 @@ class SaveOp : public framework::OperatorBase {
                   "SaveOp only support LoDTensor, %s has wrong type", iname);
    auto &tensor = var->Get<framework::LoDTensor>();
+    framework::SerializeToStream(fout, tensor, dev_ctx);
-    {  // the 1st field, uint32_t version
-      constexpr uint32_t version = 0;
-      fout.write(reinterpret_cast<const char *>(&version), sizeof(version));
-    }
-    {  // the 2nd field, tensor description
-       // int32_t  size
-       // void*    protobuf message
-      framework::TensorDesc desc;
-      desc.set_data_type(framework::ToDataType(tensor.type()));
-      auto dims = framework::vectorize(tensor.dims());
-      auto *pb_dims = desc.mutable_dims();
-      pb_dims->Resize(static_cast<int>(dims.size()), 0);
-      std::copy(dims.begin(), dims.end(), pb_dims->begin());
-      int32_t size = desc.ByteSize();
-      fout.write(reinterpret_cast<const char *>(&size), sizeof(size));
-      auto out = desc.SerializeAsString();
-      fout.write(out.data(), size);
-    }
-    {  // the 3rd field, tensor data
-      uint64_t size = tensor.memory_size();
-      auto *data_ptr = tensor.data<void>();
-      PADDLE_ENFORCE(size < std::numeric_limits<std::streamsize>::max(),
-                     "Index overflow when writing tensor");
-      if (platform::is_gpu_place(tensor.place())) {
-#ifdef PADDLE_WITH_CUDA
-        constexpr size_t kBufSize = 1024 * 1024 * 64;  // 64MB
-        std::unique_ptr<char[]> buf(new char[kBufSize]);
-        auto &gpu_dev_ctx =
-            static_cast<const platform::CUDADeviceContext &>(dev_ctx);
-        platform::CPUPlace cpu;
-        uintptr_t data = reinterpret_cast<uintptr_t>(data_ptr);
-        while (size != 0) {
-          size_t size_to_write = std::min(kBufSize, static_cast<size_t>(size));
-          memory::Copy(cpu, buf.get(),
-                       boost::get<platform::GPUPlace>(tensor.place()),
-                       reinterpret_cast<const void *>(data), size_to_write,
-                       gpu_dev_ctx.stream());
-          gpu_dev_ctx.Wait();
-          fout.write(buf.get(), size_to_write);
-          data += size_to_write;
-          size -= size_to_write;
-        }
-#else
-        PADDLE_THROW("Unexpected branch");
-#endif
-      } else {
-        fout.write(static_cast<const char *>(data_ptr),
-                   static_cast<std::streamsize>(size));
-      }
-    }
-    {  // the 4th field, lod information
-       // uint64_t lod_level
-       // uint64_t lod_level_1 size in byte.
-       // int*     lod_level_1 data
-       // ...
-      auto lod = tensor.lod();
-      uint64_t size = lod.size();
-      fout.write(reinterpret_cast<const char *>(&size), sizeof(size));
-      for (auto &each : lod) {
-        size = each.size() * sizeof(framework::LoD::value_type::value_type);
-        fout.write(reinterpret_cast<const char *>(&size), sizeof(size));
-        fout.write(reinterpret_cast<const char *>(each.data()),
-                   static_cast<std::streamsize>(size));
-      }
-    }
  }
 };

--- a/paddle/operators/send_op.cc
+++ b/paddle/operators/send_op.cc
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+   http://www.apache.org/licenses/LICENSE-2.0
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+#include <ostream>
+#include "paddle/framework/data_type.h"
+#include "paddle/framework/framework.pb.h"
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/detail/send_recv_impl.h"
+#include "paddle/operators/detail/simple_block_queue.h"
+namespace paddle {
+namespace operators {
+// TODO(typhoonzero): this is a simple implementation which only send
+// one tensor
+class SendOp : public framework::OperatorBase {
+ public:
+  SendOp(const std::string &type, const framework::VariableNameMap &inputs,
+         const framework::VariableNameMap &outputs,
+         const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {
+    // init client when the operator is created at runtime.
+    if (!client_) {
+      std::string endpoint = Attr<std::string>("endpoint");
+      client_.reset(new detail::RPCClient(
+          grpc::CreateChannel(endpoint, grpc::InsecureChannelCredentials())));
+      // TODO(typhoonzero): how to call InitVariables
+    }
+  }
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    auto iname = Input("X");
+    auto oname = Output("Out");
+    // TODO(typhoonzero): currently it's non-blocking,
+    // should block until server responds.
+    bool ret = client_->SendVariable(scope, iname, oname);
+    if (!ret) {
+      LOG(ERROR) << "send variable error";
+    }
+  }
+ protected:
+  std::shared_ptr<detail::RPCClient> client_{nullptr};
+};
+class SendOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SendOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "(Tensor) Input tensor to be saved");
+    AddOutput("Out", "(Tensor) Output fetched from server");
+    AddComment(R"DOC(
+Recv operator
+This operator will recv tensor from send_op
+)DOC");
+    AddAttr<std::string>("endpoint",
+                         "(string, default 127.0.0.1:6164)"
+                         "IP address to listen on.")
+        .SetDefault("127.0.0.1:6164")
+        .AddCustomChecker([](const std::string &ip) { return !ip.empty(); });
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(send, ops::SendOp, ops::SendOpMaker);
--- a/paddle/operators/send_recv_op_test.cc
+++ b/paddle/operators/send_recv_op_test.cc
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+   http://www.apache.org/licenses/LICENSE-2.0
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+// TODO(typhoonzero): add python bindings for this test as
+// a RemoteOptimizer.
+#include <unistd.h>
+#include <thread>
+#include "gtest/gtest.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/framework/operator.h"
+#include "paddle/framework/program_desc.h"
+USE_NO_KERNEL_OP(send);
+USE_NO_KERNEL_OP(recv);
+USE_OP(sum);
+// global for simplicity.
+std::unique_ptr<paddle::framework::OperatorBase> recv_op;
+void InitTensorsInScope(paddle::framework::Scope &scope,
+                        paddle::platform::CPUPlace &place) {
+  paddle::platform::CPUDeviceContext ctx(place);
+  auto var = scope.Var("X");
+  auto tensor = var->GetMutable<paddle::framework::LoDTensor>();
+  tensor->Resize({10, 10});
+  float *expect = tensor->mutable_data<float>(place);
+  for (int64_t i = 0; i < tensor->numel(); ++i) {
+    expect[i] = static_cast<float>(i);
+  }
+  auto out_var = scope.Var("Out");
+  auto out_tensor = out_var->GetMutable<paddle::framework::LoDTensor>();
+  out_tensor->Resize({10, 10});
+  tensor->mutable_data<float>(place);  // allocate
+}
+void AddOp(const std::string &type,
+           const paddle::framework::VariableNameMap &inputs,
+           const paddle::framework::VariableNameMap &outputs,
+           paddle::framework::AttributeMap attrs,
+           paddle::framework::BlockDescBind *block) {
+  // insert output
+  for (auto kv : outputs) {
+    for (auto v : kv.second) {
+      auto var = block->Var(v);
+      var->SetDataType(paddle::framework::DataType::FP32);
+    }
+  }
+  // insert op
+  auto op = block->AppendOp();
+  op->SetType(type);
+  for (auto &kv : inputs) {
+    op->SetInput(kv.first, kv.second);
+  }
+  for (auto &kv : outputs) {
+    op->SetOutput(kv.first, kv.second);
+  }
+  op->SetAttrMap(attrs);
+}
+void StartServerNet() {
+  paddle::framework::Scope scope;
+  paddle::platform::CPUPlace place;
+  InitTensorsInScope(scope, place);
+  // sub program run in recv_op, for simple test we use sum
+  paddle::framework::ProgramDescBind program;
+  paddle::framework::BlockDescBind *block = program.MutableBlock(0);
+  // X for server side tensors, RX for received tensers, must be of same shape.
+  AddOp("sum", {{"X", {"X", "RX"}}}, {{"Out", {"Out"}}}, {}, block);
+  paddle::framework::AttributeMap attrs;
+  attrs.insert({"endpoint", std::string("127.0.0.1:6174")});
+  attrs.insert({"OptimizeBlock", block});
+  recv_op = paddle::framework::OpRegistry::CreateOp("recv", {{"RX", {"RX"}}},
+                                                    {{"Out", {"Out"}}}, attrs);
+  paddle::platform::CPUDeviceContext ctx(place);
+  recv_op->Run(scope, ctx);
+}
+TEST(SendRecvOp, CPU) {
+  std::thread server_thread(StartServerNet);
+  sleep(5);  // wait server to start
+  // local net
+  paddle::framework::Scope scope;
+  paddle::platform::CPUPlace place;
+  InitTensorsInScope(scope, place);
+  paddle::framework::AttributeMap attrs;
+  attrs.insert({"endpoint", std::string("127.0.0.1:6174")});
+  auto send_op = paddle::framework::OpRegistry::CreateOp(
+      "send", {{"X", {"X"}}}, {{"Out", {"Out"}}}, attrs);
+  paddle::platform::CPUDeviceContext ctx(place);
+  send_op->Run(scope, ctx);
+  auto in_var = scope.Var("X");
+  auto tensor = in_var->GetMutable<paddle::framework::LoDTensor>();
+  float *expected = tensor->data<float>();
+  auto out_var = scope.Var("Out");
+  auto target = out_var->GetMutable<paddle::framework::LoDTensor>();
+  // send fail cause output is none.
+  EXPECT_NE(target->memory_size(), size_t(0));
+  float *actual = target->data<float>();
+  for (int64_t i = 0; i < target->numel(); ++i) {
+    EXPECT_EQ(expected[i] * 2, actual[i]);
+  }
+  recv_op.reset();  // dtor can shutdown and join server thread.
+  server_thread.join();
+}
--- a/paddle/operators/sgd_op.cc
+++ b/paddle/operators/sgd_op.cc
@@ -55,7 +55,7 @@ SGD operator
 This operator implements one step of the stochastic gradient descent algorithm.
-$$param_out = param - learning_rate * grad$$
+$$param\_out = param - learning\_rate * grad$$
 )DOC");
  }

--- a/paddle/scripts/docker/build.sh
+++ b/paddle/scripts/docker/build.sh
@@ -16,11 +16,13 @@ function cmake_gen() {
        echo "using python abi: $1"
        if [ "$1" == "cp27-cp27m" ]; then
            export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs4/lib:}
+            export PATH=/opt/python/cp27-cp27m/bin/:${PATH}
            PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27m/bin/python
        -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27m/include/python2.7
        -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs2/lib/libpython2.7.so"
        elif [ "$1" == "cp27-cp27mu" ]; then
            export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs2/lib:}
+            export PATH=/opt/python/cp27-cp27mu/bin/:${PATH}
            PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27mu/bin/python
        -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27mu/include/python2.7
        -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs4/lib/libpython2.7.so"

--- a/paddle/scripts/travis/build_doc.sh
+++ b/paddle/scripts/travis/build_doc.sh
@@ -11,8 +11,9 @@ make -j `nproc` gen_proto_py
 make -j `nproc` paddle_docs paddle_docs_cn
 # check websites for broken links
-linkchecker doc/en/html/index.html
+# It will be failed now!
-linkchecker doc/cn/html/index.html
+#linkchecker doc/en/html/index.html
+#linkchecker doc/cn/html/index.html
 # Parse Github URL
 REPO=`git config remote.origin.url`

--- a/proto/ModelConfig.proto
+++ b/proto/ModelConfig.proto
@@ -544,6 +544,9 @@ message LayerConfig {
  // for batch normalization layer
  // The small constant added to the variance to improve numeric stability.
  optional double epsilon = 60 [ default = 0.00001 ];
+  // for factorization machine layer
+  optional uint32 factor_size = 61;
 }
 message EvaluatorConfig {

--- a/python/paddle/trainer/config_parser.py
+++ b/python/paddle/trainer/config_parser.py
@@ -3870,6 +3870,21 @@ class ScaleSubRegionLayer(LayerBase):
                           image_conf.channels)
+@config_layer('factorization_machine')
+class FactorizationMachineLayer(LayerBase):
+    def __init__(self, name, inputs, factor_size, **xargs):
+        super(FactorizationMachineLayer, self).__init__(
+            name, 'factorization_machine', size=1, inputs=inputs, **xargs)
+        config_assert(
+            len(self.inputs) == 1,
+            'factorization machine layer must have one and only one input.')
+        self.config.factor_size = factor_size
+        input_layer = self.get_input_layer(0)
+        psize = input_layer.size * factor_size
+        dims = [input_layer.size, factor_size]
+        self.create_input_parameter(0, psize, dims)
 # Deprecated, use a new layer specific class instead
 @config_func
 def Layer(name, type, **xargs):

--- a/python/paddle/trainer_config_helpers/layers.py
+++ b/python/paddle/trainer_config_helpers/layers.py
@@ -148,6 +148,7 @@ __all__ = [
    'resize_layer',
    'sub_seq_layer',
    'scale_sub_region_layer',
+    'factorization_machine',
 ]
@@ -264,6 +265,8 @@ class LayerType(object):
    SCALE_SUB_REGION_LAYER = 'scale_sub_region'
+    FACTORIZATION_MACHINE = 'factorization_machine'
    @staticmethod
    def is_layer_type(type_name):
        """
@@ -2985,8 +2988,8 @@ def spp_layer(input,
    A layer performs spatial pyramid pooling.
    Reference:
-        Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition
+        `Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition
-        https://arxiv.org/abs/1406.4729
+        https://arxiv.org/abs/1406.4729`_
    The example usage is:
@@ -3087,8 +3090,8 @@ def img_cmrnorm_layer(input,
    Response normalization across feature maps.
    Reference:
-        ImageNet Classification with Deep Convolutional Neural Networks
+        `ImageNet Classification with Deep Convolutional Neural Networks
-        http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf
+        http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf`_
    The example usage is:
@@ -3154,9 +3157,9 @@ def batch_norm_layer(input,
        y_i &\\gets \\gamma \\hat{x_i} + \\beta \\qquad &//\ scale\ and\ shift
    Reference:
-        Batch Normalization: Accelerating Deep Network Training by Reducing
+        `Batch Normalization: Accelerating Deep Network Training by Reducing
        Internal Covariate Shift
-        http://arxiv.org/abs/1502.03167
+        http://arxiv.org/abs/1502.03167`_
    The example usage is:
@@ -5413,10 +5416,10 @@ def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None):
    to be devided by groups.
    Reference:
-        Maxout Networks
+        `Maxout Networks
-        http://www.jmlr.org/proceedings/papers/v28/goodfellow13.pdf
+        http://www.jmlr.org/proceedings/papers/v28/goodfellow13.pdf`_
-        Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks
+        `Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks
-        https://arxiv.org/pdf/1312.6082v4.pdf
+        https://arxiv.org/pdf/1312.6082v4.pdf`_
    .. math::
       y_{si+j} = \max_k x_{gsi + sk + j}
@@ -5481,9 +5484,9 @@ def ctc_layer(input,
    alignment between the inputs and the target labels is unknown.
    Reference:
-        Connectionist Temporal Classification: Labelling Unsegmented Sequence Data
+        `Connectionist Temporal Classification: Labelling Unsegmented Sequence Data
        with Recurrent Neural Networks
-        http://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf
+        http://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf`_
    Note:
        Considering the 'blank' label needed by CTC, you need to use (num_classes + 1)
@@ -5555,9 +5558,9 @@ def warp_ctc_layer(input,
    install it to :code:`third_party/install/warpctc` directory.
    Reference:
-        Connectionist Temporal Classification: Labelling Unsegmented Sequence Data
+        `Connectionist Temporal Classification: Labelling Unsegmented Sequence Data
        with Recurrent Neural Networks
-        http://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf
+        http://machinelearning.wustl.edu/mlpapers/paper_files/icml2006_GravesFGS06.pdf`_
    Note:
        - Let num_classes represents the category number. Considering the 'blank'
@@ -5777,8 +5780,8 @@ def nce_layer(input,
    Noise-contrastive estimation.
    Reference:
-        A fast and simple algorithm for training neural probabilistic language
+        `A fast and simple algorithm for training neural probabilistic language
-        models. https://www.cs.toronto.edu/~amnih/papers/ncelm.pdf
+        models. https://www.cs.toronto.edu/~amnih/papers/ncelm.pdf`_
    The example usage is:
@@ -5893,8 +5896,8 @@ def rank_cost(left,
    A cost Layer for learning to rank using gradient descent.
    Reference:
-        Learning to Rank using Gradient Descent
+        `Learning to Rank using Gradient Descent
-        http://research.microsoft.com/en-us/um/people/cburges/papers/ICML_ranking.pdf
+        http://research.microsoft.com/en-us/um/people/cburges/papers/ICML_ranking.pdf`_
    .. math::
@@ -6429,8 +6432,8 @@ def smooth_l1_cost(input, label, name=None, coeff=1.0, layer_attr=None):
        smooth_{L1}(x) = \\begin{cases} 0.5x^2& \\text{if}  \\ |x| < 1 \\\\ |x|-0.5& \\text{otherwise} \end{cases}
    Reference:
-        Fast R-CNN
+        `Fast R-CNN
-        https://arxiv.org/pdf/1504.08083v2.pdf
+        https://arxiv.org/pdf/1504.08083v2.pdf`_
    The example usage is:
@@ -6636,8 +6639,8 @@ def prelu_layer(input,
    The Parametric Relu activation that actives outputs with a learnable weight.
    Reference:
-        Delving Deep into Rectifiers: Surpassing Human-Level Performance on
+        `Delving Deep into Rectifiers: Surpassing Human-Level Performance on
-        ImageNet Classification http://arxiv.org/pdf/1502.01852v1.pdf
+        ImageNet Classification http://arxiv.org/pdf/1502.01852v1.pdf`_
    .. math::
       z_i &\\quad if \\quad z_i > 0 \\\\
@@ -6733,8 +6736,8 @@ def gated_unit_layer(input,
    product between :match:`X'` and :math:`\sigma` is finally returned.
    Reference:
-        Language Modeling with Gated Convolutional Networks
+        `Language Modeling with Gated Convolutional Networks
-        https://arxiv.org/abs/1612.08083
+        https://arxiv.org/abs/1612.08083`_
    .. math::
       y=\\text{act}(X \cdot W + b)\otimes \sigma(X \cdot V + c)
@@ -7403,3 +7406,73 @@ def scale_sub_region_layer(input, indices, value, name=None):
        parents=[input, indices],
        num_filters=input.num_filters,
        size=input.size)
+@wrap_name_default()
+@wrap_act_default(act=LinearActivation())
+@wrap_param_attr_default()
+@layer_support()
+def factorization_machine(input,
+                          factor_size,
+                          act=None,
+                          name=None,
+                          param_attr=None,
+                          layer_attr=None):
+    """
+    The Factorization Machine models pairwise feature interactions as inner
+    product of the learned latent vectors corresponding to each input feature.
+    The Factorization Machine can effectively capture feature interactions
+    especially when the input is sparse.
+    This implementation only consider the 2-order feature interactions using
+    Factorization Machine with the formula:
+    .. math::
+        y = \sum_{i=1}^{n-1}\sum_{j=i+1}^n\langle v_i, v_j \rangle x_i x_j
+    Note:
+        X is the input vector with size n. V is the factor matrix. Each row of V
+        is the latent vector corresponding to each input dimesion. The size of
+        each latent vector is k.
+    For details of Factorization Machine, please refer to the paper:
+    Factorization machines.
+    .. code-block:: python
+        first_order = paddle.layer.fc(input=input,
+                                      size=1,
+                                      act=paddle.activation.Linear())
+        second_order = paddle.layer.factorization_machine(input=input,
+                                                          factor_size=10)
+        fm = paddle.layer.addto(input=[first_order, second_order],
+                                act=paddle.activation.Linear(),
+                                bias_attr=False)
+    :param input: The input layer. Supported input types: all input data types
+                  on CPU, and only dense input types on GPU.
+    :type input: LayerOutput
+    :param factor_size: The hyperparameter that defines the dimensionality of
+                        the latent vector size.
+    :type context_len: int
+    :param act: Activation Type. Default is linear activation.
+    :type act: BaseActivation
+    :param param_attr: The parameter attribute. See ParameterAttribute for
+                       details.
+    :type param_attr: ParameterAttribute
+    :param layer_attr: Extra Layer config.
+    :type layer_attr: ExtraLayerAttribute|None
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+    assert isinstance(input, LayerOutput)
+    assert factor_size > 0, "the factor_size must be greater than 0."
+    Layer(
+        inputs=[Input(input.name, **param_attr.attr)],
+        name=name,
+        factor_size=factor_size,
+        type=LayerType.FACTORIZATION_MACHINE,
+        active_type=act.name,
+        **ExtraLayerAttribute.to_kwargs(layer_attr))
+    return LayerOutput(
+        name, LayerType.FACTORIZATION_MACHINE, input, activation=act, size=1)
--- a/python/paddle/trainer_config_helpers/tests/configs/file_list.sh
+++ b/python/paddle/trainer_config_helpers/tests/configs/file_list.sh
@@ -11,6 +11,7 @@ test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_l
 test_kmax_seq_socre_layer test_sub_nested_seq_select_layer test_scale_shift_layer
 test_seq_slice_layer test_cross_entropy_over_beam test_roi_pool_layer test_pooling3D_layer
 test_conv3d_layer test_deconv3d_layer test_BatchNorm3D test_resize_layer
-test_scale_sub_region_layer test_dot_prod_layer test_l2_distance_layer)
+test_scale_sub_region_layer test_dot_prod_layer test_l2_distance_layer
+test_factorization_machine)
 export whole_configs=(test_split_datasource)
--- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_factorization_machine.protostr
+++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_factorization_machine.protostr
+type: "nn"
+layers {
+  name: "data"
+  type: "data"
+  size: 1024
+  active_type: ""
+}
+layers {
+  name: "__factorization_machine_0__"
+  type: "factorization_machine"
+  size: 1
+  active_type: ""
+  inputs {
+    input_layer_name: "data"
+    input_parameter_name: "___factorization_machine_0__.w0"
+  }
+  factor_size: 10
+}
+parameters {
+  name: "___factorization_machine_0__.w0"
+  size: 10240
+  initial_mean: 0.0
+  initial_std: 0.03125
+  dims: 1024
+  dims: 10
+  initial_strategy: 0
+  initial_smart: true
+}
+input_layer_names: "data"
+output_layer_names: "__factorization_machine_0__"
+sub_models {
+  name: "root"
+  layer_names: "data"
+  layer_names: "__factorization_machine_0__"
+  input_layer_names: "data"
+  output_layer_names: "__factorization_machine_0__"
+  is_recurrent_layer_group: false
+}
--- a/python/paddle/trainer_config_helpers/tests/configs/test_factorization_machine.py
+++ b/python/paddle/trainer_config_helpers/tests/configs/test_factorization_machine.py
+from paddle.trainer_config_helpers import *
+data = data_layer(name='data', size=1024)
+fm = factorization_machine(input=data, factor_size=10)
+outputs(fm)
--- a/python/paddle/v2/__init__.py
+++ b/python/paddle/v2/__init__.py
@@ -83,11 +83,10 @@ def set_omp_mkl_env_vars(trainer_count):
        '''Get the number of physical cores'''
        if platform.system() == "Linux":
            num_sockets = int(
-                os.popen("lscpu |grep \"Socket\" |awk -F':' '{print $2}'|xargs")
+                os.popen("grep 'physical id' /proc/cpuinfo | sort -u | wc -l")
                .read())
            num_cores_per_socket = int(
-                os.popen(
+                os.popen("grep 'core id' /proc/cpuinfo | sort -u | wc -l")
-                    "lscpu |grep \"per socket\" |awk -F':' '{print $2}'|xargs")
                .read())
            return num_sockets * num_cores_per_socket
        else: