Merge branch 'develop' into cmake

CMakeLists.txt

@@ -54,6 +54,7 @@ option(WITH_C_API       "Compile PaddlePaddle with C-API(Prediction)"   OFF)
 option(WITH_GOLANG      "Compile PaddlePaddle with GOLANG"              OFF)
 option(GLIDE_INSTALL    "Download and install go dependencies "         ON)
 option(USE_NNPACK       "Compile PaddlePaddle with NNPACK library"      OFF)
+option(WITH_DISTRIBUTE  "Compile with grpc distributed support"         OFF)
 option(USE_EIGEN_FOR_BLAS   "Use matrix multiplication in Eigen"        OFF)
 
 # CMAKE_BUILD_TYPE
@@ -67,9 +68,6 @@ if(ANDROID OR IOS)
     if(ANDROID)
         if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
             message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
-        elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
-            # TODO: support glog for Android api 16 ~ 19 in the future
-            message(WARNING "Using the unofficial git repository <https://github.com/Xreki/glog.git> instead")
         endif()
     endif()
 
@@ -83,6 +81,8 @@ if(ANDROID OR IOS)
         "Disable RDMA when cross-compiling for Android and iOS" FORCE)
     set(WITH_MKL OFF CACHE STRING
         "Disable MKL when cross-compiling for Android and iOS" FORCE)
+    set(WITH_GOLANG OFF CACHE STRING
+        "Disable golang when cross-compiling for Android and iOS" FORCE)
 
     # Compile PaddlePaddle mobile inference library
     if (NOT WITH_C_API)

benchmark/paddle/image/googlenet.py

@@ -6,10 +6,21 @@ width = 224
 num_class = 1000
 batch_size = get_config_arg('batch_size', int, 128)
 use_gpu = get_config_arg('use_gpu', bool, True)
-
-args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
+is_infer = get_config_arg("is_infer", bool, False)
+
+args = {
+    'height': height,
+    'width': width,
+    'color': True,
+    'num_class': num_class,
+    'is_infer': is_infer
+}
 define_py_data_sources2(
-    "train.list", None, module="provider", obj="process", args=args)
+    "train.list" if not is_infer else None,
+    "test.list" if is_infer else None,
+    module="provider",
+    obj="process",
+    args=args)
 
 settings(
     batch_size=batch_size,
@@ -146,7 +157,6 @@ def inception(name, input, channels, \
     return cat
 
 
-lab = data_layer(name="label", size=1000)
 data = data_layer(name="input", size=3 * height * width)
 
 # stage 1
@@ -224,6 +234,10 @@ pool5 = img_pool_layer(
 dropout = dropout_layer(name="dropout", input=pool5, dropout_rate=0.4)
 out3 = fc_layer(
     name="output3", input=dropout, size=1000, act=SoftmaxActivation())
-loss3 = cross_entropy(name='loss3', input=out3, label=lab)
 
-outputs(loss3)
+if is_infer:
+    outputs(out3)
+else:
+    lab = data_layer(name="label", size=num_class)
+    loss3 = cross_entropy(name='loss3', input=out3, label=lab)
+    outputs(loss3)

benchmark/paddle/image/provider.py

@@ -13,14 +13,20 @@ def initHook(settings, height, width, color, num_class, **kwargs):
         settings.data_size = settings.height * settings.width * 3
     else:
         settings.data_size = settings.height * settings.width
-
+    settings.is_infer = kwargs.get('is_infer', False)
+    if settings.is_infer:
+        settings.slots = [dense_vector(settings.data_size)]
+    else:
         settings.slots = [dense_vector(settings.data_size), integer_value(1)]
 
 
 @provider(
     init_hook=initHook, min_pool_size=-1, cache=CacheType.CACHE_PASS_IN_MEM)
 def process(settings, file_list):
-    for i in xrange(1024):
+    for i in xrange(2560 if settings.is_infer else 1024):
         img = np.random.rand(1, settings.data_size).reshape(-1, 1).flatten()
+        if settings.is_infer:
+            yield img.astype('float32')
+        else:
             lab = random.randint(0, settings.num_class - 1)
             yield img.astype('float32'), int(lab)

benchmark/paddle/image/resnet.py

@@ -6,11 +6,21 @@ width = 224
 num_class = 1000
 batch_size = get_config_arg('batch_size', int, 64)
 layer_num = get_config_arg("layer_num", int, 50)
-is_test = get_config_arg("is_test", bool, False)
-
-args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
+is_infer = get_config_arg("is_infer", bool, False)
+
+args = {
+    'height': height,
+    'width': width,
+    'color': True,
+    'num_class': num_class,
+    'is_infer': is_infer
+}
 define_py_data_sources2(
-    "train.list", None, module="provider", obj="process", args=args)
+    "train.list" if not is_infer else None,
+    "test.list" if is_infer else None,
+    module="provider",
+    obj="process",
+    args=args)
 
 settings(
     batch_size=batch_size,
@@ -45,7 +55,10 @@ def conv_bn_layer(name,
         act=LinearActivation(),
         bias_attr=False)
     return batch_norm_layer(
-        name=name + "_bn", input=tmp, act=active_type, use_global_stats=is_test)
+        name=name + "_bn",
+        input=tmp,
+        act=active_type,
+        use_global_stats=is_infer)
 
 
 def bottleneck_block(name, input, num_filters1, num_filters2):
@@ -207,7 +220,9 @@ elif layer_num == 152:
 else:
     print("Wrong layer number.")
 
-lbl = data_layer(name="label", size=num_class)
-loss = cross_entropy(name='loss', input=resnet, label=lbl)
-inputs(img, lbl)
-outputs(loss)
+if is_infer:
+    outputs(resnet)
+else:
+    lbl = data_layer(name="label", size=num_class)
+    loss = cross_entropy(name='loss', input=resnet, label=lbl)
+    outputs(loss)

benchmark/paddle/image/run_mkldnn_infer.sh

+set -e
+
+function clock_to_seconds() {
+  hours=`echo $1 | awk -F ':' '{print $1}'`
+  mins=`echo $1 | awk -F ':' '{print $2}'`
+  secs=`echo $1 | awk -F ':' '{print $3}'`
+  echo `bc -l <<< "$secs + $mins * 60 + $hours * 3600"`
+}
+
+function infer() {
+  unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY
+  topology=$1
+  layer_num=$2
+  bs=$3
+  use_mkldnn=$4
+  if [ $4 == "True" ]; then
+    thread=1
+    log="logs/infer-${topology}-${layer_num}-mkldnn-${bs}.log"
+  elif [ $4 == "False" ]; then
+    thread=`nproc`
+    if [ $thread -gt $bs ]; then
+      thread=$bs
+    fi
+    log="logs/infer-${topology}-${layer_num}-${thread}mklml-${bs}.log"
+  else
+    echo "Wrong input $4, use True or False."
+    exit 0
+  fi
+
+  models_in="models/${topology}-${layer_num}/pass-00000/"
+  if [ ! -d $models_in ]; then
+    echo "Training model ${topology}_${layer_num}"
+    paddle train --job=train \
+      --config="${topology}.py" \
+      --use_mkldnn=True \
+      --use_gpu=False \
+      --trainer_count=1 \
+      --num_passes=1 \
+      --save_dir="models/${topology}-${layer_num}" \
+      --config_args="batch_size=128,layer_num=${layer_num}" \
+      > /dev/null 2>&1
+    echo "Done"
+  fi
+  log_period=$((256 / bs))
+  paddle train --job=test \
+    --config="${topology}.py" \
+    --use_mkldnn=$use_mkldnn \
+    --use_gpu=False \
+    --trainer_count=$thread \
+    --log_period=$log_period \
+    --config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True" \
+    --init_model_path=$models_in \
+    2>&1 | tee ${log}
+
+  # calculate the last 5 logs period time of 1280 samples,
+  # the time before are burning time.
+  start=`tail ${log} -n 7 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
+  end=`tail ${log} -n 2 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
+  start_sec=`clock_to_seconds $start`
+  end_sec=`clock_to_seconds $end`
+  fps=`bc <<< "scale = 2; 1280 / ($end_sec - $start_sec)"`
+  echo "Last 1280 samples start: ${start}(${start_sec} sec), end: ${end}(${end_sec} sec;" >> ${log}
+  echo "FPS: $fps images/sec" >> ${log}
+}
+
+if [ ! -f "train.list" ]; then
+  echo " " > train.list
+fi
+if [ ! -f "test.list" ]; then
+  echo " " > test.list
+fi
+if [ ! -d "logs" ]; then
+  mkdir logs
+fi
+if [ ! -d "models" ]; then
+  mkdir -p models
+fi
+
+# inference benchmark
+for use_mkldnn in True False; do
+  for batchsize in 1 2 4 8 16; do
+    infer googlenet v1 $batchsize $use_mkldnn
+    infer resnet 50 $batchsize $use_mkldnn
+    infer vgg 19 $batchsize $use_mkldnn
+  done
+done

benchmark/paddle/image/run_mkldnn.sh → benchmark/paddle/image/run_mkldnn_train.sh

@@ -8,13 +8,13 @@ function train() {
   use_mkldnn=$4
   if [ $4 == "True" ]; then
     thread=1
-    log="logs/${topology}-${layer_num}-mkldnn-${bs}.log"
+    log="logs/train-${topology}-${layer_num}-mkldnn-${bs}.log"
   elif [ $4 == "False" ]; then
     thread=`nproc`
     # each trainer_count use only 1 core to avoid conflict
-    log="logs/${topology}-${layer_num}-${thread}mklml-${bs}.log"
+    log="logs/train-${topology}-${layer_num}-${thread}mklml-${bs}.log"
   else
-    echo "Wrong input $3, use True or False."
+    echo "Wrong input $4, use True or False."
     exit 0
   fi
   args="batch_size=${bs},layer_num=${layer_num}"
@@ -30,13 +30,14 @@ function train() {
     2>&1 | tee ${log} 
 }
 
-if [ ! -d "train.list" ]; then
+if [ ! -f "train.list" ]; then
   echo " " > train.list
 fi
 if [ ! -d "logs" ]; then
   mkdir logs
 fi
 
+# training benchmark
 for use_mkldnn in True False; do
   for batchsize in 64 128 256; do
     train vgg 19 $batchsize $use_mkldnn

benchmark/paddle/image/vgg.py

@@ -6,10 +6,21 @@ width = 224
 num_class = 1000
 batch_size = get_config_arg('batch_size', int, 64)
 layer_num = get_config_arg('layer_num', int, 19)
+is_infer = get_config_arg("is_infer", bool, False)
 
-args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
+args = {
+    'height': height,
+    'width': width,
+    'color': True,
+    'num_class': num_class,
+    'is_infer': is_infer
+}
 define_py_data_sources2(
-    "train.list", None, module="provider", obj="process", args=args)
+    "train.list" if not is_infer else None,
+    "test.list" if is_infer else None,
+    module="provider",
+    obj="process",
+    args=args)
 
 settings(
     batch_size=batch_size,
@@ -98,6 +109,9 @@ elif layer_num == 19:
 else:
     print("Wrong layer number.")
 
-lab = data_layer('label', num_class)
-loss = cross_entropy(input=vgg, label=lab)
-outputs(loss)
+if is_infer:
+    outputs(vgg)
+else:
+    lab = data_layer('label', num_class)
+    loss = cross_entropy(input=vgg, label=lab)
+    outputs(loss)

cmake/external/cares.cmake

@@ -13,7 +13,7 @@
 # limitations under the License.
 #
 
-IF(MOBILE_INFERENCE)
+IF(MOBILE_INFERENCE OR NOT WITH_DISTRIBUTE)
     return()
 ENDIF()
 

cmake/external/glog.cmake

@@ -26,12 +26,21 @@ ENDIF(WIN32)
 
 INCLUDE_DIRECTORIES(${GLOG_INCLUDE_DIR})
 
+IF(ANDROID AND ${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
+  # Using the unofficial glog for Android API < 21
+  SET(GLOG_REPOSITORY "https://github.com/Xreki/glog.git")
+  SET(GLOG_TAG "8a547150548b284382ccb6582408e9140ff2bea8")
+ELSE()
+  SET(GLOG_REPOSITORY "https://github.com/google/glog.git")
+  SET(GLOG_TAG "v0.3.5")
+ENDIF()
+
 ExternalProject_Add(
     extern_glog
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS gflags
-    GIT_REPOSITORY  "https://github.com/google/glog.git"
-    GIT_TAG         v0.3.5
+    GIT_REPOSITORY  ${GLOG_REPOSITORY}
+    GIT_TAG         ${GLOG_TAG}
     PREFIX          ${GLOG_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}

cmake/external/grpc.cmake

@@ -13,7 +13,7 @@
 # limitations under the License.
 #
 
-IF(MOBILE_INFERENCE)
+IF(MOBILE_INFERENCE OR NOT WITH_DISTRIBUTE)
     return()
 ENDIF()
 

cmake/external/protobuf.cmake

@@ -188,14 +188,26 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
         SET(OPTIONAL_CACHE_ARGS "-DZLIB_ROOT:STRING=${ZLIB_ROOT}")
     ENDIF()
 
+    SET(PROTOBUF_REPO "https://github.com/google/protobuf.git")
+    SET(PROTOBUF_TAG "9f75c5aa851cd877fb0d93ccc31b8567a6706546")
+    IF(MOBILE_INFERENCE)
+        # The reason why the official version is not used is described in
+        # https://github.com/PaddlePaddle/Paddle/issues/6114
+        SET(PROTOBUF_REPO "https://github.com/qingqing01/protobuf.git")
+        SET(PROTOBUF_TAG "v3.2.0")
+        IF(NOT BUILD_FOR_HOST)
+            SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} "-Dprotobuf_BUILD_PROTOC_BINARIES=OFF")
+        ENDIF()
+    ENDIF()
+
     ExternalProject_Add(
         ${TARGET_NAME}
         ${EXTERNAL_PROJECT_LOG_ARGS}
         PREFIX          ${PROTOBUF_SOURCES_DIR}
         UPDATE_COMMAND  ""
         DEPENDS         zlib
-        GIT_REPOSITORY  "https://github.com/google/protobuf.git"
-        GIT_TAG         "9f75c5aa851cd877fb0d93ccc31b8567a6706546"
+        GIT_REPOSITORY  ${PROTOBUF_REPO}
+        GIT_TAG         ${PROTOBUF_TAG}
         CONFIGURE_COMMAND
         ${CMAKE_COMMAND} ${PROTOBUF_SOURCES_DIR}/src/${TARGET_NAME}/cmake
             ${OPTIONAL_ARGS}
@@ -213,7 +225,11 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
     )
 ENDFUNCTION()
 
-SET(PROTOBUF_VERSION 3.1)
+IF(NOT MOBILE_INFERENCE)
+    SET(PROTOBUF_VERSION 3.1)
+ELSE()
+    SET(PROTOBUF_VERSION 3.2)
+ENDIF()
 IF(CMAKE_CROSSCOMPILING)
     build_protobuf(protobuf_host TRUE)
     LIST(APPEND external_project_dependencies protobuf_host)

cmake/flags.cmake

@@ -111,6 +111,8 @@ set(COMMON_FLAGS
     -Wno-error=sign-compare
     -Wno-error=unused-local-typedefs
     -Wno-error=parentheses-equality # Warnings in pybind11
+    -Wno-error=ignored-attributes  # Warnings in Eigen, gcc 6.3
+    -Wno-error=terminate  # Warning in PADDLE_ENFORCE
 )
 
 set(GPU_COMMON_FLAGS

cmake/generic.cmake

@@ -227,8 +227,8 @@ function(cc_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS)
     cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     add_executable(${TARGET_NAME} ${cc_test_SRCS})
-    target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} gtest gtest_main)
-    add_dependencies(${TARGET_NAME} ${cc_test_DEPS} gtest gtest_main)
+    target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
+    add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
     add_test(NAME ${TARGET_NAME} COMMAND ${TARGET_NAME} WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
   endif()
 endfunction(cc_test)
@@ -288,8 +288,8 @@ function(nv_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS)
     cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
-    target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} gtest gtest_main)
-    add_dependencies(${TARGET_NAME} ${nv_test_DEPS} gtest gtest_main)
+    target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
+    add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
     add_test(${TARGET_NAME} ${TARGET_NAME})
   endif()
 endfunction(nv_test)
@@ -505,12 +505,12 @@ function(grpc_library TARGET_NAME)
   set_source_files_properties(
     ${grpc_grpc_srcs}
     PROPERTIES
-    COMPILE_FLAGS  "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
+    COMPILE_FLAGS  "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
   cc_library("${TARGET_NAME}_grpc" SRCS "${grpc_grpc_srcs}")
 
   set_source_files_properties(
     ${grpc_library_SRCS}
     PROPERTIES
-    COMPILE_FLAGS  "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
+    COMPILE_FLAGS  "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
   cc_library("${TARGET_NAME}" SRCS "${grpc_library_SRCS}" DEPS "${TARGET_NAME}_grpc" "${TARGET_NAME}_proto" "${grpc_library_DEPS}")
 endfunction()

doc/design/mkldnn/README.MD

 # Intel® MKL-DNN on PaddlePaddle: Design Doc
 
-我们计划将Intel深度神经网络数学库(**MKL-DNN**\[[1](#references)\])集成到PaddlePaddle，充分展现英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。
+我们计划将英特尔深度神经网络数学库[Intel MKL-DNN](https://github.com/01org/mkl-dnn)
+(Intel Math Kernel Library for Deep Neural Networks)集成到PaddlePaddle，
+充分展现英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。
 
-我们短期内的基本目标是：
+<div align="center">
+<img src="image/overview.png"><br/>
+Figure 1. PaddlePaddle on IA
+</div>
+
+近期目标
 
-- 完成常用layer的MKL-DNN实现。
+- 完成常用Layer的MKL-DNN实现。
 - 完成常见深度神经网络VGG，GoogLeNet 和 ResNet的MKL-DNN实现。
 
+目前的优化，主要针对PaddlePaddle在重构之前的代码框架以及V1的API。
+具体的完成状态可以参见[这里](https://github.com/PaddlePaddle/Paddle/projects/21)。
 
 ## Contents
 
 - [Overview](#overview)
 - [Actions](#actions)
  	- [CMake](#cmake)
+ 	- [Matrix](#matrix)
 	- [Layers](#layers)
 	- [Activations](#activations)
-	- [Weights](#weights)
+	- [Parameters](#parameters)
+	- [Gradients](#gradients)
 	- [Unit Tests](#unit-tests)
-	- [Protobuf Messages](#protobuf-messages)
 	- [Python API](#python-api)
-	- [Demos](#demos)
 	- [Benchmarking](#benchmarking)
 	- [Others](#others)
 - [Design Concerns](#design-concerns)
 
 ## Overview
 
-我们会把MKL-DNN作为第三方库集成进PaddlePaddle，整体框架图
+我们会把MKL-DNN会作为第三方库集成进PaddlePaddle，与其他第三方库一样，会在编译PaddlePaddle的时候下载并编译MKL-DNN。
+
+同时，为了进一步提升PaddlePaddle在基本数学运算的计算速度，我们也将MKLML即(MKL small library\[[1](#references)\])
+作为另一个第三方库集成进PaddlePaddle，它只会包括生成好的动态库和头文件。
+
+MKL，MKLML以及MKL-DNN三者关系如下表：
+
+| Name        |  Open Source     | License     | Descriptions  |
+| :---------- | :--------------- | :---------- | :------------ |
+|   MKL       |     No           | Proprietary | Accelerate math processing routines | 
+|   MKLML     |     No           | Proprietary | Small package of MKL, especially for Machine Learning |
+|   MKL-DNN   |     Yes          | Apache 2.0  | Accelerate primitives processing routines especially for Deep Neural Networks  |
+
+MKLML可以与MKL-DNN共同使用，以此达到最好的性能。
+
 <div align="center">
-<img src="image/overview.png" width=350><br/>
-Figure 1. PaddlePaddle on IA.
+<img src="image/engine.png"><br/>
+Figure 2. PaddlePaddle with MKL Engines
 </div>
 
 ## Actions
-我们把集成方案大致分为了如下几个方面。
+
+添加的相关文件和目录结构如下：
+
+```txt
+PaddlePaddle/Paddle
+├── ...
+├── cmake/
+│   ├── external/
+│   │   ├── ...
+│   │   ├── mkldnn.cmake
+│   │   └── mklml.cmake
+└── paddle/
+    ├── ...
+    ├── math/
+    │   ├── ...
+    │   └── MKLDNNMatrix.*
+    └── gserver/
+        ├── ...
+        ├── layers/
+        │   ├── ...
+        │   └── MKLDNN*Layer.*
+        ├── activations/
+        │   ├── ...
+        │   └── MKLDNNActivations.*
+        └── tests/
+            ├── ...
+            ├── MKLDNNTester.*
+            └── test_MKLDNN.cpp
+```
 
 ### CMake
-我们会在`CMakeLists.txt`中会给用户添加一个`WITH_MKL`的开关，他是负责`WITH_MKLML`和`WITH_MKLDNN`的总开关。
+在`CMakeLists.txt`中提供一个与MKL有关的总开关：`WITH_MKL`，它负责决定编译时是否使用MKLML和MKL-DNN
 
-当打开`WITH_MKL`时，会开启MKLML的功能，作为PaddlePaddle的CBLAS和LAPACK库，同时会开启Intel OpenMP用于提高MKLML的性能。 如果系统支持AVX2指令集及以上，同时会开启MKL-DNN功能。
+- `WITH_MKLML` 控制是否使用MKLML库。 
+当打开`WITH_MKL`时，会自动使用MKLML库作为PaddlePaddle的CBLAS和LAPACK库，同时会开启Intel OpenMP用于提高MKLML的性能。
+编译时会把对应的头文件和库放在`build/third_party/install/mklml/*`目录下对应的地方。
+MKLML的库目前都是动态库，主要包括`libiomp5.so`和`libmklml_intel.so`。
+- `WITH_MKLDNN` 控制是否使用MKL-DNN。
+当开启`WITH_MKL`时，会自动根据硬件配置[[2](#references)]选择是否编译MKL-DNN。
+编译时会把对应的头文件和库放在`build/third_party/install/mkldnn/*`目录下对应的地方。
+MKL-DNN的库目前只有动态库`libmkldnn.so`。
 
-当关闭`WITH_MKL`时，MKLML和MKL-DNN功能会同时关闭。
+### Matrix
+目前在PaddlePaddle中数据都是以`NCHW`的格式存储，但是在MKL-DNN中的排列方式不止这一种。
+所以我们定义了一个`MKLDNNMatrix`用于管理MKL-DNN数据的不同格式以及相互之间的转换。
 
-所以，我们会在`cmake/external`目录新建`mkldnn.cmake`和`mklml.cmake`文件，它们会在编译PaddlePaddle的时候下载对应的软件包，并放到PaddlePaddle的third party目录中。
+<div align="center">
+<img src="image/matrix.png"><br/>
+Figure 3. MKLDNNMatrix
+</div>
 
 ### Layers
-所有MKL-DNN相关的C++ layers，都会按照PaddlePaddle的目录结构存放在
-`paddle/gserver/layers`中，并且文件名都会一以*MKLDNN*开头。
+所有MKL-DNN的Layers都会继承于`MKLDNNLayer`，该类继承于PaddlePaddle的基类`Layer`。
+在`MKLDNNLayer`中会提供一些必要的接口和函数，并且会写好`forward`和`backward`的基本逻辑，
+子类只需要使用定义好的接口，实现具体的函数功能即可。
+
+<div align="center">
+<img src="image/layers.png"><br/>
+Figure 4. MKLDNNLayer
+</div>
+
+每个MKLDNNLayer都包含用于内部存储和外部存储的一系列MKLDNNMatrix：
 
-所有MKL-DNN的layers都会继承于一个叫做`MKLDNNLayer`的父类，该父类继承于PaddlePaddle的基类`Layer`。
+- 内部存储（internel memory）：`inVal_`,`inGrad_`,`outVal_`和`outGrad_`，分别代表输入数据，输入梯度，输出数据和输出梯度。
+- 外部存储（external memory）：都是以ext开头，比如`extInVal_`和`extInGrad_`，它们主要是用于，
+当数据格式与PaddlePaddle默认的`NCHW`格式不匹配时，转换内存的工作。
+需要注意的是，PaddlePaddle的activation会直接使用`output_.value`和`output_.grad`，
+所以`extOutVal_`和`extOutGrad_`必须分别与`output_.value`和`output_.grad`共享内存，
+如果不需要外部存储用于转换，那么对应的内部存储也会与它们共享内存。
+- 转换函数（resetXXX）： 包括`resetInValue`，`resetInGrad`，`resetOutValue`和`resetOutGrad`，
+表示对输入数据，输入梯度，输出数据和输出梯度的转换。
+这些函数会根据输入参数重新设置内部和外部存储，当然这两者也可以相等，即表示不需要转换。
 
-在`MKLDNNLayer`中会提供一些必要的接口和函数，并且会写好`forward`和`backward`的基本逻辑。部分函数定义为纯虚函数，子类只需要实现这些函数即可。
+注意：每个`MKLDNNlayer`的子类只需要使用内部存储就可以了，所有外部的转换工作都会在reset系列函数中都准备好。
 
 ### Activations
-由于在PaddlePaddle中，激活函数是独立于layer概念的，所以会在`paddle/gserver/activations`目录下添加`MKLDNNActivation.h`和`MKLDNNActivation.cpp`文件用于定义和使用MKL-DNN的接口。
+在重构前的PaddlePaddle中，激活函数是独立于`Layer`的概念，并且输入输出都是共用一块内存，
+所以添加了对应的`MKLDNNActivation`来实现，方式类似于`MKLDNNLayer`。
+
+### Parameters
+对于有参数的层，我们会保证`MKLDNNLayer`使用的参数与PaddlePaddle申请的buffer共用一块内存。
+如果存在数据排列格式不一样的情况时，我们会在网络训练之前把格式转换为MKL-DNN希望的格式，
+在训练结束的时候再保存为PaddlePaddle的格式，但是整个训练过程中不需要任何转换。
+这样既使得最终保存的参数格式与PaddlePaddle一致，又可以避免不必要的转换。
+
+### Gradients
+由于MKL-DNN的操作都是直接覆盖的形式，也就是说输出的结果不会在原来的数据上累加，
+这样带来的好处就是不需要一直清空memory，节省了不必要的操作。
+但是注意的是，当网络出现分支且在`backward`的时候，需要累加不同Layer传过来的梯度。
+所以在`MKLDNNlayer`中实现了一个merge的方法，此时每个小分支的`Input Gradient`
+会先临时保存在`MKLDNNMatrix`中，由分支处的Layer负责求和，并把结果放到当前层的`output_.grad`中。
+所以整体上，在实现每个子类的时候就不需要关心分支的事情了。
 
-### Weights
-由于有些layer是含有参数的，我们会尽量让MKL-DNN的参数与PaddlePaddle中`parameter`共享一块内存。
-同时，由于MKL-DNN在训练时使用的参数layout可能与PaddlePaddle默认的`nchw`不一致，我们会在网络训练的开始和结束时分别转换这个layout，使得最终保存的参数格式与PaddlePaddle一致。
+<div align="center">
+<img src="image/gradients.png"><br/>
+Figure 5. Merge Gradients
+</div>
 
 ### Unit Tests
-会在`paddle/gserver/test`目录下添加`test_MKLDNN.cpp`和`MKLDNNTester.*`用于MKL-DNN的测试。
-测试分为每个layer(或activation)的单元测试和简单网络的整体测试。
+我们会添加`test_MKLDNN.cpp`和`MKLDNNTester.*`用于MKL-DNN的测试。
+测试分为每个Layer（或Activation）的单元测试和简单网络的整体测试。
 每个测试会对比PaddlePaddle中CPU算出的结果与MKL-DNN的结果，小于某个比较小的阈值认为通过。
 
-### Protobuf Messages
-根据具体layer的需求可能会在`proto/ModelConfig.proto`里面添加必要的选项。
-
 ### Python API
 目前只考虑**v1 API**。
 
@@ -80,41 +172,40 @@ if use_mkldnn
     self.layer_type = mkldnn_*
 ```
 
-所有MKL-DNN的layer type会以*mkldnn_*开头，以示区分。 
-
-并且可能在`python/paddle/trainer_config_helper`目录下的`activations.py `和`layers.py`里面添加必要的MKL-DNN的接口。
+所有MKL-DNN的`layer_type`会以*mkldnn_*开头，这些会在`MKLDNN*Layer`注册layer的时候保证，以示区分。 
 
-### Demos
-
-会在`v1_api_demo`目录下添加一个`mkldnn`的文件夹，里面放入一些用于MKL-DNN测试的demo脚本。
+同时,会在`paddle/utils.Flags`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
 
 ### Benchmarking
-会添加`benchmark/paddle/image/run_mkldnn.sh`，用于测试使用MKL-DNN之后的性能。
+会添加相应的脚本在[这里](https://github.com/PaddlePaddle/Paddle/tree/develop/benchmark/paddle/image)，用于测试和对比在使用MKL-DNN前后的CNN网络性能。
+测试的性能对比结果会在[IntelOptimizedPaddle.md](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/IntelOptimizedPaddle.md)
 
 ### Others
-1. 如果在使用MKL-DNN的情况下，会把CPU的Buffer对齐为64。
+1. 如果在使用MKL-DNN的情况下，会把CPU的Buffer对齐为4096，具体可以参考MKL-DNN中的[memory](https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp#L673)。
 2. 深入PaddlePaddle，寻找有没有其他可以优化的可能，进一步优化。比如可能会用OpenMP改进SGD的更新性能。
 
 ## Design Concerns
 
-为了更好的符合PaddlePaddle的代码风格\[[2](#references)\]，同时又尽可能少的牺牲MKL-DNN的性能\[[3](#references)\]。
+为了更好的符合PaddlePaddle的代码风格\[[3](#references)\]，同时又尽可能少的牺牲MKL-DNN的性能\[[4](#references)\]。
 
 我们总结出一些特别需要注意的点：
 
-1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数，我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MKLDNNLayer`特有的设备ID。
+1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数，
+我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MKLDNNLayer`特有的设备ID。
 2. 重写父类Layer的**init**函数，修改`deviceId_`为`-2`，代表这个layer是用于跑在MKL-DNN的环境下。
-3. 创建`MKLDNNMatrix`，同时继承`CpuMatrix`和`mkldnn::memory`。用于管理MKL-DNN会用到的相关memory函数、接口以及会用的到格式信息。
-4. 创建`MKLDNNBase`，定义一些除了layer和memory相关的类和函数。包括MKL-DNN会用到`MKLDNNStream`和`CPUEngine`，和未来可能还会用到`FPGAEngine`等。
-5. 每个`MKLDNNlayer`都会有`inVal_`,`inGrad_`,`outVal_`和`outGrad_`，分别代表input value， input gradient，output value和output gradient。他们会存放MKL-DNN用到的internal memory。同时还会定义以*ext*开头的`MKLDNNMatrix`(表示external的memory)，主要是在格式与PaddlePaddle默认的`nchw`格式不匹配时，用于转换内存的工作。必要的转换函数也会在`MKLDNNLayer`中提前定义好，每个子类只需要调用定义好的reset buffer函数即可。
-6. 每个`MKLDNNlayer`的resetbuffer相关的函数（包括reset input、output的Value和grad），他们会根据输入参数reset internal和external的memory，当然这两者也可以相等，即表示不需要转换。只需要把握一个原则，每个`MKLDNNlayer`的子类，只需要使用internal的memory就可以了，所有external的转换工作在父类的reset函数中都提前准备好了。
-7. 一般来说，external的memory会尽量与PaddlePaddle中的`value`和`grad`共享内存。同时每个`MKLDNNLayer`中的external output value和gradient(也就是`extOutVal_`和`extOutGrad_`)必须分别与`output_.value`和`output_.grad`共享内存，因为PaddlePaddle的activation会直接使用`output_.value`和`output_.grad`。如果不需要external的buffer用于转换，那么internal的buffer也会与他们共享内存。
-8. 如果MKL-DNN layer的后面接有cpu device，那么就会使`output_.value`与`extOutVal_`共享内存，同时数据格式就是`nchw`，这样下一个cpu device就能拿到正确的数据。在有cpu device的时候，external的memory的格式始终是`nchw`或者`nc`。
-9. 由于MKL-DNN的输出操作都是覆盖data的，不是在原来的数据上累加，所以当网络出现分支时，在`backward`时会需要merge不同layer的梯度。`MKLDNNlayer`中会实现merge的方法，此时每个小分支的input gradient会先临时保存在一个`MKLDNNMatrix`中，由分支处的layer负责求和，并把结果放到这个layer的`output_.grad`中。所以整体上，每个子类并不会需要关心分支的事情，也是在父类都实现好了。
-10. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
+3. 创建`MKLDNNBase`，定义一些除了layer和memory相关的类和函数。
+包括MKL-DNN会用到`MKLDNNStream`和`CPUEngine`，和未来可能还会用到`FPGAEngine`等。
+4. 如果MKL-DNN layer的后面接有cpu device，那么就会使`output_.value`与`extOutVal_`共享内存，
+同时数据格式就是`NCHW`，这样下一个cpu device就能拿到正确的数据。
+在有普通的CPU layer时， `extOutVal_`和`extOutGrad_`的格式始终是`NCHW`或者`NC`。
 
 ## References
-
-1. [Intel Math Kernel Library for Deep Neural Networks (Intel MKL-DNN)](https://github.com/01org/mkl-dnn "Intel MKL-DNN")
-2. [原来的方案](https://github.com/PaddlePaddle/Paddle/pull/3096)会引入**nextLayer**的信息。但是在PaddlePaddle中，无论是重构前的layer还是重构后的op，都不会想要知道next layer/op的信息。
-3. MKL-DNN的高性能格式与PaddlePaddle原有的`NCHW`不同(PaddlePaddle中的CUDNN部分使用的也是`NCHW`，所以不存在这个问题)，所以需要引入一个转换方法，并且只需要在必要的时候转换这种格式，才能更好的发挥MKL-DNN的性能。
+1. [MKL small library](https://github.com/01org/mkl-dnn#linking-your-application)是[Intel MKL](https://software.intel.com/en-us/mkl)的一个子集。
+主要包括了深度学习相关的数学原语与操作，一般由MKL-DNN在发布[新版本](https://github.com/01org/mkl-dnn/releases)时一起更新。
+2. [MKL-DNN System Requirements](https://github.com/01org/mkl-dnn#system-requirements)。
+目前在PaddlePaddle中，仅会在支持AVX2指令集及以上的机器才使用MKL-DNN。
+3. [原来的方案](https://github.com/PaddlePaddle/Paddle/pull/3096)会引入**nextLayer**的信息。
+但是在PaddlePaddle中，无论是重构前的layer还是重构后的op，都不会想要知道next layer/op的信息。
+4. MKL-DNN的高性能格式与PaddlePaddle原有的`NCHW`不同(PaddlePaddle中的cuDNN部分使用的也是`NCHW`，所以不存在这个问题)。
+所以需要引入一个转换方法，并且只需要在必要的时候转换这种格式，才能更好的发挥MKL-DNN的性能。
 

doc/howto/optimization/cpu_profiling.md

-This tutorial introduces techniques we used to profile and tune the
+This tutorial introduces techniques we use to profile and tune the
 CPU performance of PaddlePaddle.  We will use Python packages
-`cProfile` and `yep`, and Google `perftools`.
+`cProfile` and `yep`, and Google's `perftools`.
 
-Profiling is the process that reveals the performance bottlenecks,
+Profiling is the process that reveals performance bottlenecks,
 which could be very different from what's in the developers' mind.
-Performance tuning is to fix the bottlenecks. Performance optimization
+Performance tuning is done to fix these bottlenecks. Performance optimization
 repeats the steps of profiling and tuning alternatively.
 
-PaddlePaddle users program AI by calling the Python API, which calls
+PaddlePaddle users program AI applications by calling the Python API, which calls
 into `libpaddle.so.` written in C++.  In this tutorial, we focus on
 the profiling and tuning of
 
@@ -82,7 +82,7 @@ focus on. We can sort above profiling file by tottime:
 
 We can see that the most time-consuming function is the `built-in
 method run`, which is a C++ function in `libpaddle.so`.  We will
-explain how to profile C++ code in the next section.  At the right
+explain how to profile C++ code in the next section.  At this 
 moment, let's look into the third function `sync_with_cpp`, which is a
 Python function.  We can click it to understand more about it:
 
@@ -135,8 +135,8 @@ to generate the profiling file.  The default filename is
 `main.py.prof`.
 
 Please be aware of the `-v` command line option, which prints the
-analysis results after generating the profiling file.  By taking a
-glance at the print result, we'd know that if we stripped debug
+analysis results after generating the profiling file.  By examining the
+ the print result, we'd know that if we stripped debug
 information from `libpaddle.so` at build time.  The following hints
 help make sure that the analysis results are readable:
 
@@ -155,9 +155,9 @@ help make sure that the analysis results are readable:
    variable `OMP_NUM_THREADS=1` to prevents OpenMP from automatically
    starting multiple threads.
 
-### Look into the Profiling File
+### Examining the Profiling File
 
-The tool we used to look into the profiling file generated by
+The tool we used to examine the profiling file generated by
 `perftools` is [`pprof`](https://github.com/google/pprof), which
 provides a Web-based GUI like `cprofilev`.
 
@@ -194,4 +194,4 @@ time, and `MomentumOp` takes about 17%. Obviously, we'd want to
 optimize `MomentumOp`.
 
 `pprof` would mark performance critical parts of the program in
-red. It's a good idea to follow the hint.
+red. It's a good idea to follow the hints.

paddle/capi/CMakeLists.txt

@@ -4,6 +4,16 @@ else ()
   set(PADDLE_FLOAT_TYPE float)
 endif()
 
+execute_process(
+  COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
+  WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
+  OUTPUT_VARIABLE PADDLE_GIT_COMMIT
+  RESULT_VARIABLE PADDLE_GIT_COMMIT_RESULT
+  ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+if(NOT PADDLE_GIT_COMMIT)
+  set(PADDLE_GIT_COMMIT "no commit information")
+endif()
+
 # config.h used for C-API. It will store Paddle building configuration as a
 # header. Make user just include PaddleCAPI.h then can get building
 # configuration without explicitly set -DPADDLE_WITH_DOUBLE when building their

paddle/capi/config.h.in

@@ -3,6 +3,9 @@
 
 typedef @PADDLE_FLOAT_TYPE@ paddle_real;
 
+#define __PADDLE_VERSION__  "@PADDLE_VERSION@"
+#define __PADDLE_COMMIT__   "@PADDLE_GIT_COMMIT@"
+
 // Since we only support linux and macos in compile, always use clang or
 // gcc 4.8+. DLL_IMPORT/DLL_EXPORT is as simple as below.
 #define PD_API __attribute__((visibility("default")))

paddle/framework/shape_inference.cc

@@ -22,6 +22,12 @@ std::vector<framework::DDim> InferShapeContext::GetInputsDim(
   return GetDims(names);
 }
 
+DDim InferShapeContext::GetInputsElementDim(const std::string &name,
+                                            int idx) const {
+  const std::vector<std::string> &names = Inputs(name);
+  return this->GetDim(names[idx]);
+}
+
 void InferShapeContext::SetOutputsDim(
     const std::string &name, const std::vector<framework::DDim> &dims) {
   auto &names = Outputs(name);

paddle/framework/shape_inference.h

@@ -37,6 +37,7 @@ class InferShapeContext {
   virtual framework::DDim GetInputDim(const std::string &name) const = 0;
 
   std::vector<framework::DDim> GetInputsDim(const std::string &name) const;
+  DDim GetInputsElementDim(const std::string &name, int idx) const;
 
   virtual void SetOutputDim(const std::string &name, const DDim &dim) = 0;
   void SetOutputsDim(const std::string &name,

paddle/function/EigenGemm.cpp

@@ -21,7 +21,7 @@ template <class T>
 struct EigenBlasGemm {
   typedef Eigen::TensorMap<Eigen::Tensor<T, 2, Eigen::RowMajor, int>,
                            Eigen::Aligned>
-      Matrix;
+      EigenMatrix;
 
   static void compute(const bool transA,
                       const bool transB,
@@ -56,14 +56,13 @@ struct EigenBlasGemm {
       sizeB[1] = N;
       CHECK_EQ(N, ldb);
     }
-    Eigen::array<int, 2> sizeC;
-    sizeC[0] = M;
-    sizeC[1] = N;
-    CHECK_EQ(N, ldc);
+    Eigen::array<int, 2> sizeC = {{M, ldc}};
+    Eigen::array<int, 2> offsetC = {{0, 0}};
+    Eigen::array<int, 2> extentC = {{M, N}};
 
-    const Matrix a(const_cast<T*>(A), sizeA);
-    const Matrix b(const_cast<T*>(B), sizeB);
-    Matrix c(C, sizeC);
+    const EigenMatrix a(const_cast<T*>(A), sizeA);
+    const EigenMatrix b(const_cast<T*>(B), sizeB);
+    EigenMatrix c(C, sizeC);
 
     typedef typename Eigen::Tensor<T, 2>::DimensionPair DimPair;
     Eigen::array<DimPair, 1> dims;
@@ -72,6 +71,7 @@ struct EigenBlasGemm {
     dims[0].second = transB ? 1 : 0;
 
     Eigen::DefaultDevice device;
+    if (N == ldc) {
       if (alpha == T(1) && beta == T(0)) {
         c.device(device) = a.contract(b, dims);
       } else if (alpha == T(1) && beta == T(1)) {
@@ -79,6 +79,16 @@ struct EigenBlasGemm {
       } else {
         c.device(device) = alpha * a.contract(b, dims) + beta * c;
       }
+    } else {
+      if (alpha == T(1) && beta == T(0)) {
+        c.slice(offsetC, extentC).device(device) = a.contract(b, dims);
+      } else if (alpha == T(1) && beta == T(1)) {
+        c.slice(offsetC, extentC).device(device) += a.contract(b, dims);
+      } else {
+        c.slice(offsetC, extentC).device(device) =
+            alpha * a.contract(b, dims) + beta * c.slice(offsetC, extentC);
+      }
+    }
   }
 };
 

paddle/gserver/tests/sequence_rnn_matched_inputs.py

@@ -41,7 +41,7 @@ nonseq = embedding_layer(input=label, size=word_dim)
 
 
 # This hierarchical RNN is designed to be equivalent to the simple RNN in
-# sequence_rnn_multi_unequalength_inputs.conf
+# sequence_rnn_mixed_inputs.conf
 def outer_step(subseq, seq, nonseq, encoding):
     outer_mem = memory(name="outer_rnn_state", size=hidden_dim)
 

paddle/gserver/tests/sequence_rnn_mixed_inputs.py

@@ -37,7 +37,7 @@ encoding = embedding_layer(input=data2, size=word_dim)
 
 
 # This hierarchical RNN is designed to be equivalent to the simple RNN in
-# sequence_rnn_multi_unequalength_inputs.conf
+# sequence_rnn_matched_inputs.conf
 def outer_step(subseq, seq, nonseq, encoding):
     outer_mem = memory(name="outer_rnn_state", size=hidden_dim)
 

paddle/math/CMakeLists.txt

@@ -26,8 +26,6 @@ else()
 endif()
 
 if(MOBILE_INFERENCE)
-    list(REMOVE_ITEM MATH_SOURCES
-         ${CMAKE_CURRENT_SOURCE_DIR}/SIMDFunctions.cpp)
     # Remove sparse
     list(REMOVE_ITEM MATH_HEADERS
          ${CMAKE_CURRENT_SOURCE_DIR}/CpuSparseMatrix.h

paddle/math/SIMDFunctions.h

@@ -116,9 +116,11 @@ inline bool vec_check(size_t len) {
 }
 
 namespace internal {
+#ifdef __SSE3__
 void addToImpl(float* a, const float* b, size_t len);
 void batchAddToImpl(float* a, const float* b[], int batch, size_t len);
 void colMaxImpl(float* result, const float* data, int dim, int numSamples);
+#endif
 #ifdef __AVX__
 void decayL1AvxImpl(float* dst, float* src, float lambda, size_t len);
 void decayL1AvxImpl(

paddle/memory/memory.cc

@@ -81,18 +81,33 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
 }
 
 template <>
-void* Alloc<platform::GPUPlace>(platform::GPUPlace place, size_t size) {
-  return GetGPUBuddyAllocator(place.device)->Alloc(size);
+size_t Used<platform::GPUPlace>(platform::GPUPlace place) {
+  return GetGPUBuddyAllocator(place.device)->Used();
 }
 
 template <>
-void Free<platform::GPUPlace>(platform::GPUPlace place, void* p) {
-  GetGPUBuddyAllocator(place.device)->Free(p);
+void* Alloc<platform::GPUPlace>(platform::GPUPlace place, size_t size) {
+  auto* buddy_allocator = GetGPUBuddyAllocator(place.device);
+  auto* ptr = buddy_allocator->Alloc(size);
+  if (ptr == nullptr) {
+    int cur_dev = platform::GetCurrentDeviceId();
+    platform::SetDeviceId(place.device);
+    size_t avail, total;
+    platform::GpuMemoryUsage(avail, total);
+    LOG(WARNING) << "Cannot allocate " << size << " bytes in GPU "
+                 << place.device << ", available " << avail << " bytes";
+    LOG(WARNING) << "total " << total;
+    LOG(WARNING) << "GpuMinChunkSize " << platform::GpuMinChunkSize();
+    LOG(WARNING) << "GpuMaxChunkSize " << platform::GpuMaxChunkSize();
+    LOG(WARNING) << "GPU memory used: " << Used<platform::GPUPlace>(place);
+    platform::SetDeviceId(cur_dev);
+  }
+  return ptr;
 }
 
 template <>
-size_t Used<platform::GPUPlace>(platform::GPUPlace place) {
-  return GetGPUBuddyAllocator(place.device)->Used();
+void Free<platform::GPUPlace>(platform::GPUPlace place, void* p) {
+  GetGPUBuddyAllocator(place.device)->Free(p);
 }
 
 #endif

paddle/operators/CMakeLists.txt

@@ -212,18 +212,22 @@ set(DEPS_OPS
     send_op
     recv_op)
 
+if(WITH_DISTRIBUTE)
 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")
+    COMPILE_FLAGS  "-Wno-non-virtual-dtor -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")
+    COMPILE_FLAGS  "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
+
+cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS send_op recv_op sum_op executor)
+endif()
 
 op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op)
 op_library(cross_entropy_op DEPS cross_entropy)
@@ -275,4 +279,3 @@ 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/concat_op.cc

@@ -25,7 +25,7 @@ class ConcatOp : public framework::OperatorWithKernel {
 
   void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE_GE(ctx->Inputs("X").size(), 1UL,
-                      "Inputs(X) of ConcatOp should be empty.")
+                      "Inputs(X) of ConcatOp should be empty.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of ConcatOp should not be null.");
 
@@ -45,7 +45,7 @@ class ConcatOp : public framework::OperatorWithKernel {
         }
         PADDLE_ENFORCE_EQ(out_dims[j], ins[i][j],
                           "Input tensors should have the same "
-                          "elements except the specify axis.")
+                          "elements except the specify axis.");
       }
     }
     ctx->SetOutputDim("Out", out_dims);

paddle/operators/elementwise_op.h

@@ -35,7 +35,7 @@ class ElementwiseOp : public framework::OperatorWithKernel {
     auto x_dim = ctx->GetInputDim("X");
     auto y_dim = ctx->GetInputDim("Y");
     PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
-                      "Rank of first input must >= rank of second input.")
+                      "Rank of first input must >= rank of second input.");
     ctx->SetOutputDim("Out", x_dim);
     ctx->ShareLoD("X", /*->*/ "Out");
   }
@@ -120,7 +120,7 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel {
     auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
 
     PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
-                      "Rank of first input must >= rank of second input.")
+                      "Rank of first input must >= rank of second input.");
 
     auto x_grad_name = framework::GradVarName("X");
     auto y_grad_name = framework::GradVarName("Y");

paddle/operators/elementwise_op_function.h

@@ -106,7 +106,7 @@ void ElementwiseCompute(const framework::ExecutionContext& ctx) {
   auto x_dims = x->dims();
   auto y_dims = y->dims();
   PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
-                    "Rank of first input must >= rank of second input.")
+                    "Rank of first input must >= rank of second input.");
 
   if (x_dims == y_dims) {
     functor f;

paddle/operators/nce_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/nce_op.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+
+class NCEOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"));
+    PADDLE_ENFORCE(ctx->HasInput("Label"));
+    PADDLE_ENFORCE(ctx->HasInput("Weight"));
+    PADDLE_ENFORCE(ctx->HasOutput("Cost"));
+    PADDLE_ENFORCE(ctx->HasOutput("SampleLogits"));
+    PADDLE_ENFORCE(ctx->HasOutput("SampleLabels"));
+
+    auto x_dims = ctx->GetInputDim("Input");
+    auto label_dims = ctx->GetInputDim("Label");
+    PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0]);
+    int num_true_classes = label_dims.size() == 2 ? label_dims[1] : 1;
+    if (ctx->HasInput("Bias")) {
+      PADDLE_ENFORCE_EQ(ctx->GetInputDim("Weight")[0],
+                        ctx->GetInputDim("Bias")[0]);
+    }
+    auto num_neg_samples = ctx->Attrs().Get<int>("num_neg_samples");
+    auto num_total_classes = ctx->Attrs().Get<int>("num_total_classes");
+    std::vector<int> custom_neg_classes =
+        ctx->Attrs().Get<std::vector<int>>("custom_neg_classes");
+    PADDLE_ENFORCE_EQ(num_total_classes, ctx->GetInputDim("Weight")[0]);
+    if (custom_neg_classes.size() > 0) {
+      PADDLE_ENFORCE_EQ(custom_neg_classes.size(),
+                        static_cast<size_t>(num_neg_samples));
+    }
+    // set dims of output(Out)
+    std::vector<int64_t> out_dims;
+    out_dims.push_back(x_dims[0]);
+    out_dims.push_back(1);
+    ctx->SetOutputDim("Cost", framework::make_ddim(out_dims));
+
+    // set dims of output(SampleOut)
+    std::vector<int64_t> sample_out_dims;
+    sample_out_dims.push_back(x_dims[0]);
+    sample_out_dims.push_back(num_neg_samples + num_true_classes);
+    ctx->SetOutputDim("SampleLogits", framework::make_ddim(sample_out_dims));
+    ctx->SetOutputDim("SampleLabels", framework::make_ddim(sample_out_dims));
+  }
+
+ protected:
+  framework::OpKernelType GetKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
+        ctx.device_context());
+  }
+};
+
+class NCEOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  NCEOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Input", "(Tensor) A tensor of shape [batch_size, dim].");
+    AddInput(
+        "Label",
+        "(Tensor) A tensor of shape [batch_size, num_true_class]. "
+        "'num_true_class' is the number of target classes in each sample."
+        "The number of target classes per sample should be same. "
+        "If you have a variable number of target classes, "
+        "you can pad them out to a constant number by either repeating them"
+        " or by padding with an otherwise unused class.)");
+    AddInput("Weight",
+             "(Tensor) A tensor of shape [num_class, dim]. 'num_class' is the "
+             "total number of class.");
+    AddInput(
+        "Bias",
+        "(Tensor) A tensor of shape [num_class, 1]. 'num_class' is the total "
+        "number of class. It is a dispensable input.")
+        .AsDispensable();
+    AddInput("SampleWeight",
+             "(Tensor) A tensor of shape [batch_size, 1] storing a weight for "
+             "each sample. And it is a dispensable input. The default value of "
+             "sample is 1.")
+        .AsDispensable();
+    AddOutput("Cost",
+              "(Tensor) A tensor of shape [batch_size, 1]. Cost of samples.");
+    AddOutput("SampleLogits",
+              "An intermediate tensor of shape[batch_size, num_neg_samples + "
+              "num_pos_samples]."
+              "This tensor is output of forward kernel and used in backward "
+              "kernel to compute grads."
+              "Given X is  the dot product of input tensor and sampled labels' "
+              "weights."
+              "Then 'SampleLogits' is sigmoid(X).")
+        .AsIntermediate();
+    AddOutput("SampleLabels",
+              "An intermediate tensor of shape[batch_size, num_neg_samples + "
+              "num_pos_samples]."
+              "This tensor is output of forward kernel and used in backward "
+              "kernel to compute grads."
+              "")
+        .AsIntermediate();
+    AddAttr<int>("num_total_classes",
+                 "Total number of classes in all samples.");
+    AddAttr<int>("num_neg_samples",
+                 "The number of negative classes. The default value is 10.")
+        .SetDefault(10);
+    AddAttr<std::vector<int>>("custom_neg_classes",
+                              "This attribute only be used in unitest. Classes "
+                              "in this list wiil be used as negative classes "
+                              "for every samples. Under normal conditions, "
+                              "user should avoid setting this attribute.");
+    AddComment(R"DOC(
+Compute and return the noise-contrastive estimation training loss.
+See [Noise-contrastive estimation: A new estimation principle for unnormalized statistical models](http://www.jmlr.org/proceedings/papers/v9/gutmann10a/gutmann10a.pdf).
+By default this operator uses a uniform distribution for sampling.
+)DOC");
+  }
+};
+
+class NCEOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"));
+    PADDLE_ENFORCE(ctx->HasInput("Weight"));
+    PADDLE_ENFORCE(ctx->HasInput("Cost"));
+    PADDLE_ENFORCE(ctx->HasInput("SampleLogits"));
+    PADDLE_ENFORCE(ctx->HasInput("SampleLabels"));
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Cost")),
+                   "The input(Out@GRAD) should not be null.");
+
+    auto x_dims = ctx->GetInputDim("Input");
+    auto x_grad_name = framework::GradVarName("Input");
+    if (ctx->HasOutput(x_grad_name)) {
+      ctx->SetOutputDim(x_grad_name, x_dims);
+    }
+
+    auto w_dims = ctx->GetInputDim("Weight");
+    auto w_grad_name = framework::GradVarName("Weight");
+    if (ctx->HasOutput(w_grad_name)) {
+      ctx->SetOutputDim(w_grad_name, w_dims);
+    }
+
+    auto bias_grad_name = framework::GradVarName("Bias");
+    if (ctx->HasOutput(bias_grad_name)) {
+      auto bias_dims = ctx->GetInputDim("Bias");
+      ctx->SetOutputDim(bias_grad_name, bias_dims);
+    }
+  }
+
+ protected:
+  framework::OpKernelType GetKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
+        ctx.device_context());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(nce, ops::NCEOp, ops::NCEOpMaker, nce_grad, ops::NCEOpGrad);
+REGISTER_OP_CPU_KERNEL(nce, ops::NCEKernel<paddle::platform::CPUPlace, float>,
+                       ops::NCEKernel<paddle::platform::CPUPlace, double>);
+REGISTER_OP_CPU_KERNEL(nce_grad,
+                       ops::NCEGradKernel<paddle::platform::CPUPlace, float>,
+                       ops::NCEGradKernel<paddle::platform::CPUPlace, double>);

paddle/operators/nce_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 <math.h>
+#include <random>
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "unsupported/Eigen/CXX11/Tensor"
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+void PrepareSamples(const framework::ExecutionContext& context) {
+  auto label = context.Input<Tensor>("Label");
+  const int64_t* label_data = label->data<int64_t>();
+  auto label_dims = label->dims();
+  int num_total_classes = context.Attr<int>("num_total_classes");
+  // for unitest
+  std::vector<int> custom_neg_classes =
+      context.Attr<std::vector<int>>("custom_neg_classes");
+  // random machine
+  std::random_device rd;
+  std::mt19937 rng(rd());
+  std::uniform_int_distribution<int> rand(0, num_total_classes - 1);
+
+  auto sample_labels = context.Output<Tensor>("SampleLabels");
+  auto sample_labels_dims = sample_labels->dims();
+  int64_t* sample_labels_data =
+      sample_labels->mutable_data<int64_t>(context.GetPlace());
+
+  int num_label = label_dims.size() == 2 ? label_dims[1] : 1;
+  int index = 0;
+  for (size_t i = 0; i < label_dims[0]; ++i) {
+    int j = 0;
+    for (; j < num_label; ++j) {
+      sample_labels_data[index++] = label_data[i * num_label + j];
+    }
+    if (custom_neg_classes.size() > 0) {
+      for (auto label : custom_neg_classes) {
+        sample_labels_data[index++] = label;
+      }
+    } else {
+      for (; j < sample_labels_dims[1]; ++j) {
+        // TODO(wanghaoshuang): support more distribution sampling
+        sample_labels_data[index++] = rand(rng);
+      }
+    }
+  }
+}
+
+template <typename Place, typename T>
+class NCEKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    PrepareSamples<Place, T>(context);
+    auto sample_labels = context.Output<Tensor>("SampleLabels");
+    const int64_t* sample_labels_data = sample_labels->data<int64_t>();
+    auto sample_out = context.Output<Tensor>("SampleLogits");
+    T* sample_out_data = sample_out->mutable_data<T>(context.GetPlace());
+    auto label = context.Input<Tensor>("Label");
+    auto sample_weight = context.Input<Tensor>("SampleWeight");
+    const T* sample_weight_data = nullptr;
+    if (sample_weight != nullptr) {
+      sample_weight_data = sample_weight->data<T>();
+    }
+    auto out = context.Output<Tensor>("Cost");
+    T* out_data = out->mutable_data<T>(context.GetPlace());
+    int num_neg_samples = context.Attr<int>("num_neg_samples");
+    int num_total_classes = context.Attr<int>("num_total_classes");
+    int num_true_class = 1;
+    if (label != nullptr) {
+      num_true_class = label->dims()[1];
+    }
+    T b = 1. / num_total_classes * num_neg_samples;
+    // forward bias
+    auto bias = context.Input<Tensor>("Bias");
+    if (bias != nullptr) {
+      const T* bias_data = bias->data<T>();
+      for (size_t i = 0; i < sample_labels->numel(); ++i) {
+        sample_out_data[i] = bias_data[sample_labels_data[i]];
+      }
+    } else {
+      for (size_t i = 0; i < sample_labels->numel(); ++i) {
+        sample_out_data[i] = 0;
+      }
+    }
+    // forward mul
+    auto input_mat = EigenMatrix<T>::From(*(context.Input<Tensor>("Input")));
+    auto weight_mat = EigenMatrix<T>::From(*(context.Input<Tensor>("Weight")));
+    for (size_t i = 0; i < sample_labels->numel(); ++i) {
+      Eigen::Tensor<T, 0, Eigen::RowMajor, Eigen::DenseIndex> result =
+          (input_mat.chip((int)(i / sample_labels->dims()[1]), 0) *
+           weight_mat.chip(sample_labels_data[i], 0))
+              .sum();
+      sample_out_data[i] += result(0);
+      sample_out_data[i] = (1. / (1. + exp(-sample_out_data[i])));
+    }
+    // forward cost
+    for (size_t i = 0; i < sample_labels->dims()[0]; ++i) {
+      size_t j = 0;
+      out_data[i] = 0;
+      T w = sample_weight == nullptr ? 1. : sample_weight_data[i];
+      // for true classes
+      for (; j < num_true_class; ++j) {
+        T o = sample_out_data[i * sample_out->dims()[1] + j];
+        T cost = -log(o / (o + b));
+        out_data[i] += w * cost;
+      }
+      // for sampled neg classes
+      for (; j < sample_labels->dims()[1]; ++j) {
+        T o = sample_out_data[i * sample_out->dims()[1] + j];
+        T cost = -log(b / (o + b));
+        out_data[i] += w * cost;
+      }
+    }
+  }
+};
+
+template <typename Place, typename T>
+class NCEGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto d_out = context.Input<Tensor>(framework::GradVarName("Cost"));
+    const T* d_out_data = d_out->data<T>();
+    auto label = context.Input<Tensor>("Label");
+    auto sample_out = context.Input<Tensor>("SampleLogits");
+    const T* sample_out_data = sample_out->data<T>();
+    auto sample_labels = context.Input<Tensor>("SampleLabels");
+    const int64_t* sample_labels_data = sample_labels->data<int64_t>();
+    auto sample_weight = context.Input<Tensor>("SampleWeight");
+    const T* sample_weight_data = nullptr;
+    if (sample_weight != nullptr) {
+      sample_weight_data = sample_weight->data<T>();
+    }
+    int num_neg_samples = context.Attr<int>("num_neg_samples");
+    int num_total_classes = context.Attr<int>("num_total_classes");
+    int num_true_class = 1;
+    if (label != nullptr) {
+      num_true_class = label->dims()[1];
+    }
+    T b = 1. / num_total_classes * num_neg_samples;
+    Tensor sample_grad;  // tmp tensor
+    T* sample_grad_data =
+        sample_grad.mutable_data<T>(sample_labels->dims(), context.GetPlace());
+    // backward cost
+    for (size_t i = 0; i < sample_labels->numel(); ++i) {
+      T o = sample_out_data[i];
+      T w = sample_weight == nullptr
+                ? 1
+                : sample_weight_data[i / sample_labels->dims()[1]];
+      sample_grad_data[i] = (i % sample_labels->dims()[1]) < num_true_class
+                                ? w * (b / (o + b)) * (o - 1)
+                                : w * (o * (1 - o) / (o + b));
+      sample_grad_data[i] *= d_out_data[i / sample_labels->dims()[1]];
+    }
+    // get d_bias
+    auto d_bias = context.Output<Tensor>(framework::GradVarName("Bias"));
+    if (d_bias != nullptr) {
+      T* d_bias_data = d_bias->mutable_data<T>(context.GetPlace());
+      std::fill(d_bias_data, d_bias_data + d_bias->numel(), 0.0);
+      for (size_t i = 0; i < sample_labels->numel(); ++i) {
+        d_bias_data[sample_labels_data[i]] += sample_grad_data[i];
+      }
+    }
+    // get d_w
+    auto d_w = context.Output<Tensor>(framework::GradVarName("Weight"));
+    if (d_w != nullptr) {
+      auto d_w_data = d_w->mutable_data<T>(context.GetPlace());
+      std::fill(d_w_data, d_w_data + d_w->numel(), 0.0);
+      auto d_w_matrix = EigenMatrix<T>::From(*d_w);
+      auto x_matrix = EigenMatrix<T>::From(*(context.Input<Tensor>("Input")));
+      for (size_t i = 0; i < sample_labels->numel(); ++i) {
+        d_w_matrix.chip(sample_labels_data[i], 0) +=
+            x_matrix.chip((int)(i / sample_labels->dims()[1]), 0) *
+            sample_grad_data[i];
+      }
+    }
+    // get d_x
+    auto d_x = context.Output<Tensor>(framework::GradVarName("Input"));
+    if (d_x != nullptr) {
+      d_x->mutable_data<T>(context.GetPlace());
+      auto d_x_matrix = EigenMatrix<T>::From(*d_x);
+      auto w_matrix = EigenMatrix<T>::From(*(context.Input<Tensor>("Weight")));
+      for (size_t i = 0; i < sample_labels->numel(); ++i) {
+        d_x_matrix.chip((int)(i / sample_labels->dims()[1]), 0) +=
+            w_matrix.chip(sample_labels_data[i], 0) * sample_grad_data[i];
+      }
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle

paddle/operators/sequence_slice_op.h

@@ -54,10 +54,10 @@ class SequenceSliceOpKernel : public framework::OpKernel<T> {
     PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now.");
     PADDLE_ENFORCE_EQ(
         n, static_cast<size_t>(length->dims()[0]),
-        "The size of input-sequence and length-array should be the same")
+        "The size of input-sequence and length-array should be the same");
     PADDLE_ENFORCE_EQ(
         n, static_cast<size_t>(offset->dims()[0]),
-        "The size of input-sequence and offset-array should be the same")
+        "The size of input-sequence and offset-array should be the same");
 
     const int64_t* offset_data = offset->data<int64_t>();
     const int64_t* length_data = length->data<int64_t>();
@@ -78,11 +78,11 @@ class SequenceSliceOpKernel : public framework::OpKernel<T> {
 
     for (size_t i = 0; i < n; ++i) {
       PADDLE_ENFORCE_LT(0, offset_data[i],
-                        "The offset[%d] must greater than zero.", i)
+                        "The offset[%d] must greater than zero.", i);
       PADDLE_ENFORCE_LT(0, length_data[i],
-                        "The length[%d] must greater than zero.", i)
+                        "The length[%d] must greater than zero.", i);
       PADDLE_ENFORCE_LT(lod[0][i] + offset_data[i] + length_data[i],
-                        lod[0][i + 1], "The target tensor's length overflow.")
+                        lod[0][i + 1], "The target tensor's length overflow.");
     }
 
     out->mutable_data<T>(ctx.GetPlace());

paddle/operators/sigmoid_cross_entropy_with_logits_op.cc

@@ -25,20 +25,19 @@ class SigmoidCrossEntropyWithLogitsOp : public framework::OperatorWithKernel {
 
   void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
-    PADDLE_ENFORCE(ctx->HasInput("Labels"),
-                   "Input(Labels) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should be not null.");
 
     auto x_dims = ctx->GetInputDim("X");
-    auto labels_dims = ctx->GetInputDim("Labels");
+    auto labels_dims = ctx->GetInputDim("Label");
     PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
     PADDLE_ENFORCE_EQ(labels_dims.size(), 2,
-                      "Input(Labels)'s rank should be 2.");
+                      "Input(Label)'s rank should be 2.");
     PADDLE_ENFORCE_EQ(x_dims[0], labels_dims[0],
-                      "The 1st dimension of Input(X) and Input(Labels) should "
+                      "The 1st dimension of Input(X) and Input(Label) should "
                       "be equal.");
     PADDLE_ENFORCE_EQ(x_dims[1], labels_dims[1],
-                      "The 2nd dimension of Input(X) and Input(Labels) should "
+                      "The 2nd dimension of Input(X) and Input(Label) should "
                       "be equal.");
 
     ctx->SetOutputDim("Out", x_dims);
@@ -53,26 +52,25 @@ class SigmoidCrossEntropyWithLogitsGradOp
 
   void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
-    PADDLE_ENFORCE(ctx->HasInput("Labels"),
-                   "Input(Labels) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                    "Input(Out@GRAD) shoudl be not null.");
     PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
                    "Output(X@GRAD) should be not null.");
 
     auto x_dims = ctx->GetInputDim("X");
-    auto labels_dims = ctx->GetInputDim("Labels");
+    auto labels_dims = ctx->GetInputDim("Label");
     auto dout_dims = ctx->GetInputDim(framework::GradVarName("Out"));
     PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
     PADDLE_ENFORCE_EQ(labels_dims.size(), 2,
-                      "Input(Labels)'s rank should be 2.");
+                      "Input(Label)'s rank should be 2.");
     PADDLE_ENFORCE_EQ(dout_dims.size(), 2,
                       "Input(Out@Grad)'s rank should be 2.");
     PADDLE_ENFORCE_EQ(x_dims[0], labels_dims[0],
-                      "The 1st dimension of Input(X) and Input(Labels) should "
+                      "The 1st dimension of Input(X) and Input(Label) should "
                       "be equal.");
     PADDLE_ENFORCE_EQ(x_dims[1], labels_dims[1],
-                      "The 2nd dimension of Input(X) and Input(Labels) should "
+                      "The 2nd dimension of Input(X) and Input(Label) should "
                       "be equal.");
     PADDLE_ENFORCE_EQ(x_dims[0], dout_dims[0],
                       "The 1st dimension of Input(X) and Input(Out@Grad) "
@@ -97,7 +95,7 @@ class SigmoidCrossEntropyWithLogitsOpMaker
              "This input is a tensor of logits computed by the previous "
              " operator. Logits are unscaled log probabilities given as "
              "log(p/(1-p)).");
-    AddInput("Labels",
+    AddInput("Label",
              "(Tensor, default Tensor<float>), a 2-D tensor of the same type "
              "and shape as X. This input is a tensor of probabalistic labels "
              "for each logit");

paddle/operators/sigmoid_cross_entropy_with_logits_op.h

@@ -25,8 +25,7 @@ class SigmoidCrossEntropyWithLogitsKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
     const framework::Tensor *X = context.Input<framework::Tensor>("X");
-    const framework::Tensor *Labels =
-        context.Input<framework::Tensor>("Labels");
+    const framework::Tensor *Labels = context.Input<framework::Tensor>("Label");
     framework::Tensor *Out = context.Output<framework::Tensor>("Out");
     Out->mutable_data<T>(context.GetPlace());
 
@@ -52,8 +51,7 @@ class SigmoidCrossEntropyWithLogitsGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext &context) const override {
     const framework::Tensor *X = context.Input<framework::Tensor>("X");
-    const framework::Tensor *Labels =
-        context.Input<framework::Tensor>("Labels");
+    const framework::Tensor *Labels = context.Input<framework::Tensor>("Label");
     const framework::Tensor *dOut =
         context.Input<framework::Tensor>(framework::GradVarName("Out"));
     framework::Tensor *dX =

paddle/operators/sum_op.h

@@ -84,7 +84,7 @@ class SumKernel : public framework::OpKernel<T> {
       int64_t offset = 0;
       for (int i = 0; i < N; i++) {
         PADDLE_ENFORCE_EQ(out->height(),
-                          in_vars[i]->Get<SelectedRows>().height())
+                          in_vars[i]->Get<SelectedRows>().height());
         functor(context.device_context(), in_vars[i]->Get<SelectedRows>(),
                 offset, out);
         offset += in_vars[i]->Get<SelectedRows>().value().numel();

paddle/operators/tensor_array_read_write_op.cc

@@ -37,9 +37,15 @@ class WriteToArrayOp : public ArrayOp {
                << " to " << offset + 1;
       out->resize(offset + 1);
     }
+    if (x_tensor.memory_size() > 0) {
       auto *out_tensor = &out->at(offset);
       CopyFrom(x_tensor, dev_ctx.GetPlace(), dev_ctx, out_tensor);
       out_tensor->set_lod(x_tensor.lod());
+    } else {
+      VLOG(10) << "WARNING: The input tensor 'x_tensor' holds no memory, so "
+                  "nothing has been written to output array["
+               << offset << "].";
+    }
   }
 };
 

paddle/operators/while_op.cc

@@ -287,7 +287,6 @@ class WhileGradOpShapeInference : public framework::InferShapeBase {
 
     auto p_names = ctx->Inputs(kParameters);
     auto pg_names = ctx->Outputs(kParamGrads);
-    auto dims = ctx->GetInputsDim(kParameters);
     auto var_types = ctx->GetInputsVarType(kParameters);
     std::vector<std::string> names_to_set;
     std::vector<framework::DDim> dims_to_set;
@@ -295,13 +294,14 @@ class WhileGradOpShapeInference : public framework::InferShapeBase {
       if (pg_names[i] == framework::kEmptyVarName) {
         continue;
       }
+      auto dims = ctx->GetInputsElementDim(kParameters, i);
       if (var_types[i] == framework::VarDesc::LOD_TENSOR) {
         names_to_set.push_back(pg_names[i]);
-        dims_to_set.push_back(dims[i]);
+        dims_to_set.push_back(dims);
       } else if (var_types[i] == framework::VarDesc::LOD_TENSOR_ARRAY) {
         // not sure how to set the dim of LOD_TENSOR_ARRAY
         names_to_set.push_back(pg_names[i]);
-        dims_to_set.push_back(dims[i]);
+        dims_to_set.push_back(dims);
       }
     }
     ctx->SetDims(names_to_set, dims_to_set);

paddle/optimizer/parameter_optimizer_test.cc

@@ -127,8 +127,3 @@ TEST_F(OptimizerTest, TestGetWeight) { TestGetWeight(); }
 TEST_F(OptimizerTest, TestUpdate) { TestUpdate(); }
 
 TEST_F(OptimizerTest, TestCheckPoint) { TestCheckPoint(); }
-
-int main(int argc, char** argv) {
-  testing::InitGoogleTest(&argc, argv);
-  return RUN_ALL_TESTS();
-}

paddle/optimizer/serialization_test.cc

@@ -46,8 +46,3 @@ TEST(TensorToProto, Case2) {
     EXPECT_EQ(t1[i], t[i]);
   }
 }
-
-int main(int argc, char** argv) {
-  testing::InitGoogleTest(&argc, argv);
-  return RUN_ALL_TESTS();
-}

paddle/platform/enforce.h

@@ -235,15 +235,23 @@ inline void throw_on_error(T e) {
 #define PADDLE_ENFORCE_LE(__VAL0, __VAL1, ...) \
   __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, <=, >, __VA_ARGS__)
 #define PADDLE_ENFORCE_NOT_NULL(__VAL, ...)                  \
-  PADDLE_ENFORCE(nullptr != (__VAL), #__VAL " should not be null\n%s", \
-                 paddle::string::Sprintf("" __VA_ARGS__));
+  do {                                                       \
+    if (UNLIKELY(nullptr == (__VAL))) {                      \
+      PADDLE_THROW(#__VAL " should not be null\n%s",         \
+                   paddle::string::Sprintf("" __VA_ARGS__)); \
+    }                                                        \
+  } while (0)
 
 #define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...)  \
-  PADDLE_ENFORCE(__VAL0 __CMP __VAL1,                                         \
-                 "enforce %s " #__CMP " %s failed, %s " #__INV_CMP " %s\n%s", \
+  do {                                                                  \
+    if (UNLIKELY(!((__VAL0)__CMP(__VAL1)))) {                           \
+      PADDLE_THROW("enforce %s " #__CMP " %s failed, %s " #__INV_CMP    \
+                   " %s\n%s",                                           \
                    #__VAL0, #__VAL1, paddle::string::to_string(__VAL0), \
                    paddle::string::to_string(__VAL1),                   \
-                 paddle::string::Sprintf("" __VA_ARGS__));
+                   paddle::string::Sprintf("" __VA_ARGS__));            \
+    }                                                                   \
+  } while (0)
 
 }  // namespace platform
 }  // namespace paddle

paddle/platform/gpu_info.cc

@@ -75,15 +75,19 @@ size_t GpuMaxChunkSize() {
   GpuMemoryUsage(available, total);
 
   // Reserving the rest memory for page tables, etc.
-  size_t reserving = (1 - FLAGS_fraction_of_gpu_memory_to_use) * total;
+  size_t reserving = 0.05 * total;
 
   // If available less than minimum chunk size, no usable memory exists.
-  available = std::max(available, GpuMinChunkSize()) - GpuMinChunkSize();
+  available =
+      std::max(std::max(available, GpuMinChunkSize()) - GpuMinChunkSize(),
+               reserving) -
+      reserving;
 
-  // If available less than reserving, no usable memory exists.
-  size_t usable = std::max(available, reserving) - reserving;
+  size_t allocating = FLAGS_fraction_of_gpu_memory_to_use * total;
 
-  return usable;
+  PADDLE_ENFORCE_LT(allocating, available);
+
+  return allocating;
 }
 
 void GpuMemcpyAsync(void *dst, const void *src, size_t count,

paddle/pserver/CMakeLists.txt

@@ -49,7 +49,7 @@ if(WITH_TESTING)
   add_subdirectory(test)
 endif()
 
-if(NOT WITH_C_API)
+if(NOT MOBILE_INFERENCE)
   add_executable(paddle_pserver_main ${PSERVER_MAIN_SOURCES})
   link_paddle_exe(paddle_pserver_main)
 

paddle/scripts/docker/build.sh

@@ -36,6 +36,7 @@ function cmake_gen() {
         ${PYTHON_FLAGS}
         -DWITH_DOC=OFF
         -DWITH_GPU=${WITH_GPU:-OFF}
+        -DWITH_DISTRIBUTE=${WITH_DISTRIBUTE:-OFF}
         -DWITH_MKL=${WITH_MKL:-ON}
         -DWITH_AVX=${WITH_AVX:-OFF}
         -DWITH_GOLANG=${WITH_GOLANG:-ON}
@@ -57,6 +58,7 @@ EOF
         ${PYTHON_FLAGS} \
         -DWITH_DOC=OFF \
         -DWITH_GPU=${WITH_GPU:-OFF} \
+        -DWITH_DISTRIBUTE=${WITH_DISTRIBUTE:-OFF} \
         -DWITH_MKL=${WITH_MKL:-ON} \
         -DWITH_AVX=${WITH_AVX:-OFF} \
         -DWITH_GOLANG=${WITH_GOLANG:-ON} \

paddle/testing/CMakeLists.txt

@@ -5,4 +5,8 @@ if(WITH_TESTING)
   add_dependencies(paddle_test_main paddle_proto ${external_project_dependencies})
   add_library(paddle_test_util STATIC TestUtil.cpp)
   add_dependencies(paddle_test_util paddle_proto ${external_project_dependencies})
+  if(NOT MOBILE_INFERENCE)
+    add_library(paddle_gtest_main STATIC paddle_gtest_main.cc)
+    add_dependencies(paddle_gtest_main paddle_memory gtest gflags)
+  endif()
 endif()

paddle/testing/paddle_gtest_main.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 <cstring>
+#include "gflags/gflags.h"
+#include "gtest/gtest.h"
+#include "paddle/memory/memory.h"
+
+int main(int argc, char** argv) {
+  std::vector<char*> new_argv;
+  std::string gflags_env;
+  new_argv.push_back(argv[0]);
+#ifdef PADDLE_WITH_CUDA
+  new_argv.push_back(
+      strdup("--tryfromenv=fraction_of_gpu_memory_to_use,use_pinned_memory"));
+#else
+  new_argv.push_back(strdup("--tryfromenv=use_pinned_memory"));
+#endif
+  int new_argc = static_cast<int>(new_argv.size());
+  char** new_argv_address = new_argv.data();
+  google::ParseCommandLineFlags(&new_argc, &new_argv_address, false);
+  testing::InitGoogleTest(&argc, argv);
+  paddle::memory::Used(paddle::platform::CPUPlace());
+#ifdef PADDLE_WITH_CUDA
+  paddle::memory::Used(paddle::platform::GPUPlace(0));
+#endif
+  return RUN_ALL_TESTS();
+}

paddle/trainer/CMakeLists.txt

@@ -54,7 +54,7 @@ if(WITH_TESTING)
   add_subdirectory(tests)
 endif()
 
-if(NOT WITH_C_API)
+if(NOT MOBILE_INFERENCE)
   add_paddle_exe(paddle_trainer TrainerMain.cpp)
   add_paddle_exe(paddle_merge_model MergeModel.cpp)
 
@@ -74,7 +74,5 @@ endif()
 if(WITH_GOLANG)
   add_dependencies(paddle_trainer_lib paddle_pserver_cclient)
   target_link_libraries(paddle_trainer_lib paddle_pserver_cclient)
-  if(NOT WITH_C_API)
   target_link_libraries(paddle_trainer paddle_pserver_cclient)
-  endif()
 endif(WITH_GOLANG)

python/paddle/v2/fluid/layers.py

@@ -185,6 +185,7 @@ def data(name,
          shape,
          append_batch_size=True,
          dtype='float32',
+         lod_level=0,
          type=core.VarDesc.VarType.LOD_TENSOR,
          main_program=None,
          startup_program=None,
@@ -198,6 +199,7 @@ def data(name,
        append_batch_size: Whether or not to append the data as a batch.
        dtype: The type of data : float32, float_16, int etc
        type: The output type. By default it is LOD_TENSOR.
+       lod_level(int): The LoD Level. 0 means the input data is not a sequence.
        main_program: Name of the main program that calls this
        startup_program: Name of the startup program
        stop_gradient: A boolean that mentions whether gradient should flow.
@@ -228,7 +230,8 @@ def data(name,
         shape=shape,
         dtype=dtype,
         type=type,
-        stop_gradient=stop_gradient)
+        stop_gradient=stop_gradient,
+        lod_level=lod_level)
 
 
 def create_tensor(dtype, name=None, main_program=None, startup_program=None):
@@ -400,6 +403,7 @@ _create_op_func_('sigmoid')
 _create_op_func_('scale')
 _create_op_func_('reshape')
 _create_op_func_('transpose')
+_create_op_func_('sigmoid_cross_entropy_with_logits')
 
 
 def cast(x, dtype, main_program=None):

python/paddle/v2/fluid/tests/book/test_word2vec.py

@@ -58,10 +58,6 @@ train_reader = paddle.batch(
 place = fluid.CPUPlace()
 exe = fluid.Executor(place)
 
-# fix https://github.com/PaddlePaddle/Paddle/issues/5434 then remove
-# below exit line.
-exit(0)
-
 exe.run(fluid.default_startup_program())
 
 for pass_id in range(PASS_NUM):
@@ -79,6 +75,6 @@ for pass_id in range(PASS_NUM):
                                   'nextw': input_data[4]
                               },
                               fetch_list=[avg_cost])
-        if avg_cost_np[0] < 10.0:
+        if avg_cost_np[0] < 5.0:
             exit(0)  # if avg cost less than 10.0, we think our code is good.
 exit(1)

python/paddle/v2/fluid/tests/demo/fc_gan.py

+import errno
+import math
+import os
+
+import matplotlib
+import numpy
+
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+
+matplotlib.use('Agg')
+import matplotlib.pyplot as plt
+import matplotlib.gridspec as gridspec
+
+NOISE_SIZE = 100
+NUM_PASS = 1000
+NUM_REAL_IMGS_IN_BATCH = 121
+NUM_TRAIN_TIMES_OF_DG = 3
+LEARNING_RATE = 2e-5
+
+
+def D(x):
+    hidden = fluid.layers.fc(input=x,
+                             size=200,
+                             act='relu',
+                             param_attr='D.w1',
+                             bias_attr='D.b1')
+    logits = fluid.layers.fc(input=hidden,
+                             size=1,
+                             act=None,
+                             param_attr='D.w2',
+                             bias_attr='D.b2')
+    return logits
+
+
+def G(x):
+    hidden = fluid.layers.fc(input=x,
+                             size=200,
+                             act='relu',
+                             param_attr='G.w1',
+                             bias_attr='G.b1')
+    img = fluid.layers.fc(input=hidden,
+                          size=28 * 28,
+                          act='tanh',
+                          param_attr='G.w2',
+                          bias_attr='G.b2')
+    return img
+
+
+def plot(gen_data):
+    gen_data.resize(gen_data.shape[0], 28, 28)
+    n = int(math.ceil(math.sqrt(gen_data.shape[0])))
+    fig = plt.figure(figsize=(n, n))
+    gs = gridspec.GridSpec(n, n)
+    gs.update(wspace=0.05, hspace=0.05)
+
+    for i, sample in enumerate(gen_data):
+        ax = plt.subplot(gs[i])
+        plt.axis('off')
+        ax.set_xticklabels([])
+        ax.set_yticklabels([])
+        ax.set_aspect('equal')
+        plt.imshow(sample.reshape(28, 28), cmap='Greys_r')
+
+    return fig
+
+
+def main():
+    try:
+        os.makedirs("./out")
+    except OSError as e:
+        if e.errno != errno.EEXIST:
+            raise
+
+    startup_program = fluid.Program()
+    d_program = fluid.Program()
+    dg_program = fluid.Program()
+
+    with fluid.program_guard(d_program, startup_program):
+        img = fluid.layers.data(name='img', shape=[784], dtype='float32')
+        d_loss = fluid.layers.sigmoid_cross_entropy_with_logits(
+            x=D(img),
+            label=fluid.layers.data(
+                name='label', shape=[1], dtype='float32'))
+        d_loss = fluid.layers.mean(x=d_loss)
+
+    with fluid.program_guard(dg_program, startup_program):
+        noise = fluid.layers.data(
+            name='noise', shape=[NOISE_SIZE], dtype='float32')
+        g_img = G(x=noise)
+        g_program = dg_program.clone()
+        dg_loss = fluid.layers.sigmoid_cross_entropy_with_logits(
+            x=D(g_img),
+            label=fluid.layers.fill_constant_batch_size_like(
+                input=noise, dtype='float32', shape=[-1, 1], value=1.0))
+        dg_loss = fluid.layers.mean(x=dg_loss)
+
+    opt = fluid.optimizer.Adam(learning_rate=LEARNING_RATE)
+
+    opt.minimize(loss=d_loss, startup_program=startup_program)
+    opt.minimize(
+        loss=dg_loss,
+        startup_program=startup_program,
+        parameter_list=[
+            p.name for p in g_program.global_block().all_parameters()
+        ])
+    exe = fluid.Executor(fluid.CPUPlace())
+    exe.run(startup_program)
+
+    num_true = NUM_REAL_IMGS_IN_BATCH
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.mnist.train(), buf_size=60000),
+        batch_size=num_true)
+
+    for pass_id in range(NUM_PASS):
+        for batch_id, data in enumerate(train_reader()):
+            num_true = len(data)
+            n = numpy.random.uniform(
+                low=-1.0, high=1.0,
+                size=[num_true * NOISE_SIZE]).astype('float32').reshape(
+                    [num_true, NOISE_SIZE])
+            generated_img = exe.run(g_program,
+                                    feed={'noise': n},
+                                    fetch_list={g_img})[0]
+            real_data = numpy.array(map(lambda x: x[0], data)).astype('float32')
+            real_data = real_data.reshape(num_true, 784)
+            total_data = numpy.concatenate([real_data, generated_img])
+            total_label = numpy.concatenate([
+                numpy.ones(
+                    shape=[real_data.shape[0], 1], dtype='float32'),
+                numpy.zeros(
+                    shape=[real_data.shape[0], 1], dtype='float32')
+            ])
+            d_loss_np = exe.run(d_program,
+                                feed={'img': total_data,
+                                      'label': total_label},
+                                fetch_list={d_loss})[0]
+            for _ in xrange(NUM_TRAIN_TIMES_OF_DG):
+                n = numpy.random.uniform(
+                    low=-1.0, high=1.0,
+                    size=[2 * num_true * NOISE_SIZE]).astype('float32').reshape(
+                        [2 * num_true, NOISE_SIZE, 1, 1])
+                dg_loss_np = exe.run(dg_program,
+                                     feed={'noise': n},
+                                     fetch_list={dg_loss})[0]
+            print("Pass ID={0}, Batch ID={1}, D-Loss={2}, DG-Loss={3}".format(
+                pass_id, batch_id, d_loss_np, dg_loss_np))
+        # generate image each batch
+        fig = plot(generated_img)
+        plt.savefig(
+            'out/{0}.png'.format(str(pass_id).zfill(3)), bbox_inches='tight')
+        plt.close(fig)
+
+
+if __name__ == '__main__':
+    main()

python/paddle/v2/fluid/tests/test_layers.py

@@ -137,6 +137,16 @@ class TestBook(unittest.TestCase):
 
         print(str(program))
 
+    def test_sigmoid_cross_entropy(self):
+        program = Program()
+        with program_guard(program):
+            dat = layers.data(name='data', shape=[10], dtype='float32')
+            lbl = layers.data(name='label', shape=[10], dtype='float32')
+            self.assertIsNotNone(
+                layers.sigmoid_cross_entropy_with_logits(
+                    x=dat, label=lbl))
+        print(str(program))
+
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/v2/fluid/tests/test_nce.py

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+def nce(input, weight, bias, sample_weight, labels, num_classes,
+        num_sample_class):
+    samples = []
+    sample_labels = []
+    batch_size = input.shape[0]
+    num_true_class = labels.shape[1]
+    for i in range(batch_size):
+        w = 1 if sample_weight is None else sample_weight[i]
+        for label in labels[i]:
+            samples.append((i, label, True, w))
+            sample_labels.append(label)
+        for num in range(num_sample_class):
+            samples.append((i, num, False, w))
+            sample_labels.append(num)
+    # forward bias
+    sample_out = np.zeros(len(samples)).astype(np.float32)
+    if bias is not None:
+        for i in range(len(samples)):
+            sample_out[i] = bias[samples[i][1]]
+    # forward weight
+    for i in range(len(samples)):
+        sample_out[i] += np.dot(input[samples[i][0]], weight[samples[i][1]])
+
+    # forward activation
+    sample_out = 1.0 / (1.0 + np.exp(-sample_out))
+    # forward cost
+    out = np.zeros(batch_size).astype(np.float32)
+    b = 1.0 / num_classes * num_sample_class
+    for i in range(len(samples)):
+        o = sample_out[i]
+        cost = -np.log(o / (o + b)) if samples[i][2] else -np.log(b / (o + b))
+        out[samples[i][0]] += cost * samples[i][3]
+    return (out[:, np.newaxis], np.array(sample_out).reshape(
+        batch_size, num_sample_class + num_true_class),
+            np.array(sample_labels).reshape(batch_size,
+                                            num_sample_class + num_true_class))
+
+
+class TestNCE(OpTest):
+    def generate_data(self, dim, batch_size, num_classes, num_true_class,
+                      num_neg_samples):
+        input = np.random.randn(batch_size, dim).astype(np.float32)
+        weight = np.random.randn(num_classes, dim).astype(np.float32)
+        bias = np.random.randn(num_classes).astype(np.float32)
+        sample_weight = np.random.randn(batch_size).astype(np.float32)
+        labels = np.random.randint(0, num_classes, (batch_size, num_true_class))
+        self.attrs = {
+            'num_total_classes': num_classes,
+            'num_neg_samples': num_neg_samples,
+            'custom_neg_classes': range(num_neg_samples)
+        }
+        self.inputs = {
+            'Input': input,
+            'Label': labels,
+            'Weight': weight,
+            'Bias': bias,
+            'SampleWeight': sample_weight
+        }
+
+    def set_data(self):
+        self.generate_data(5, 5, 4, 1, 2)
+
+    def compute(self):
+        out = nce(self.inputs['Input'], self.inputs['Weight'],
+                  self.inputs['Bias'], self.inputs['SampleWeight'],
+                  self.inputs['Label'], self.attrs['num_total_classes'],
+                  self.attrs['num_neg_samples'])
+        self.outputs = {
+            'Cost': out[0],
+            'SampleLogits': out[1],
+            'SampleLabels': out[2]
+        }
+
+    def setUp(self):
+        self.op_type = 'nce'
+        self.set_data()
+        self.compute()
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(
+            ["Input", "Weight", "Bias"], "Cost", max_relative_error=0.02)
+
+
+class TestNCECase1(TestNCE):
+    def set_data(self):
+        self.generate_data(10, 20, 10, 2, 5)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/fluid/tests/test_recurrent_op.py

@@ -454,4 +454,6 @@ class RecurrentOpNoMemBootTest(RecurrentOpTest1):
 
 
 if __name__ == '__main__':
+    # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/6152
+    exit(0)
     unittest.main()

python/paddle/v2/fluid/tests/test_sigmoid_cross_entropy_with_logits_op.py

@@ -2,11 +2,12 @@ import numpy as np
 from op_test import OpTest
 from scipy.special import logit
 from scipy.special import expit
+import unittest
 
 
 class TestSigmoidCrossEntropyWithLogitsOp1(OpTest):
-    '''Test sigmoid_cross_entropy_with_logit_op with binary labels
-    '''
+    """Test sigmoid_cross_entropy_with_logit_op with binary label
+    """
 
     def setUp(self):
         self.op_type = "sigmoid_cross_entropy_with_logits"
@@ -16,16 +17,16 @@ class TestSigmoidCrossEntropyWithLogitsOp1(OpTest):
             'X': logit(
                 np.random.uniform(0, 1, (batch_size, num_classes))
                 .astype("float32")),
-            'Labels': np.random.randint(0, 2, (batch_size, num_classes))
+            'Label': np.random.randint(0, 2, (batch_size, num_classes))
             .astype("float32")
         }
 
         # Fw Pass is implemented as elementwise sigmoid followed by
         # elementwise logistic loss
-        # Labels * -log(sigmoid(X)) + (1 - labels) * -log(1 - sigmoid(X))
+        # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X))
         sigmoid_X = expit(self.inputs['X'])
-        term1 = self.inputs['Labels'] * np.log(sigmoid_X)
-        term2 = (1 - self.inputs['Labels']) * np.log(1 - sigmoid_X)
+        term1 = self.inputs['Label'] * np.log(sigmoid_X)
+        term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
         self.outputs = {'Out': -term1 - term2}
 
     def test_check_output(self):
@@ -36,8 +37,8 @@ class TestSigmoidCrossEntropyWithLogitsOp1(OpTest):
 
 
 class TestSigmoidCrossEntropyWithLogitsOp2(OpTest):
-    '''Test sigmoid_cross_entropy_with_logit_op with probabalistic labels
-    '''
+    """Test sigmoid_cross_entropy_with_logit_op with probabalistic label
+    """
 
     def setUp(self):
         self.op_type = "sigmoid_cross_entropy_with_logits"
@@ -47,16 +48,16 @@ class TestSigmoidCrossEntropyWithLogitsOp2(OpTest):
             'X': logit(
                 np.random.uniform(0, 1, (batch_size, num_classes))
                 .astype("float32")),
-            'Labels': np.random.uniform(0, 1, (batch_size, num_classes))
+            'Label': np.random.uniform(0, 1, (batch_size, num_classes))
             .astype("float32")
         }
 
         # Fw Pass is implemented as elementwise sigmoid followed by
         # elementwise logistic loss
-        # Labels * -log(sigmoid(X)) + (1 - labels) * -log(1 - sigmoid(X))
+        # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X))
         sigmoid_X = expit(self.inputs['X'])
-        term1 = self.inputs['Labels'] * np.log(sigmoid_X)
-        term2 = (1 - self.inputs['Labels']) * np.log(1 - sigmoid_X)
+        term1 = self.inputs['Label'] * np.log(sigmoid_X)
+        term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
         self.outputs = {'Out': -term1 - term2}
 
     def test_check_output(self):
@@ -64,3 +65,7 @@ class TestSigmoidCrossEntropyWithLogitsOp2(OpTest):
 
     def test_check_grad(self):
         self.check_grad(['X'], 'Out')
+
+
+if __name__ == '__main__':
+    unittest.main()