update code and fix conflicts.

CMakeLists.txt

@@ -36,8 +36,7 @@ include(simd)
 ################################ Configurations #######################################
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
-option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    ${AVX_FOUND})
-option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      ${AVX_FOUND})
+option(WITH_MKL         "Compile PaddlePaddle with MKL support."        ${AVX_FOUND})
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        ON)
 option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)
@@ -82,10 +81,8 @@ if(ANDROID OR IOS)
         "Disable PYTHON when cross-compiling for Android and iOS" FORCE)
     set(WITH_RDMA OFF CACHE STRING
         "Disable RDMA when cross-compiling for Android and iOS" FORCE)
-    set(WITH_MKLDNN OFF CACHE STRING
-        "Disable MKLDNN when cross-compiling for Android and iOS" FORCE)
-    set(WITH_MKLML OFF CACHE STRING
-        "Disable MKLML package when cross-compiling for Android and iOS" FORCE)
+    set(WITH_MKL OFF CACHE STRING
+        "Disable MKL when cross-compiling for Android and iOS" FORCE)
 
     # Compile PaddlePaddle mobile inference library
     if (NOT WITH_C_API)
@@ -111,6 +108,17 @@ else()
     set(THIRD_PARTY_BUILD_TYPE Release)
 endif()
 
+if(WITH_MKL)
+    set(WITH_MKLML ON)
+    set(WITH_MKLDNN ${AVX2_FOUND})
+    if(NOT WITH_MKLDNN)
+        message(WARNING "Do not have AVX2 intrinsics and disabled MKL-DNN")
+    endif()
+else()
+    set(WITH_MKLML OFF)
+    set(WITH_MKLDNN OFF)
+endif()
+
 ########################################################################################
 
 include(external/mklml)     # download mklml package
@@ -161,8 +169,12 @@ if(WITH_GPU)
   include(cuda)
 endif(WITH_GPU)
 
+if(WITH_MKLML)
+    list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB})
+endif()
+
 if(WITH_MKLDNN)
-    list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB} ${MKLDNN_IOMP_LIB})
+    list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB})
 endif()
 
 if(USE_NNPACK)

cmake/configure.cmake

@@ -76,27 +76,14 @@ else()
     include_directories(${CUDA_TOOLKIT_INCLUDE})
 endif(NOT WITH_GPU)
 
-if(WITH_MKLDNN)
-    add_definitions(-DPADDLE_USE_MKLDNN)
-    if (WITH_MKLML AND MKLDNN_IOMP_DIR)
-        message(STATUS "Enable Intel OpenMP at ${MKLDNN_IOMP_DIR}")
+if (WITH_MKLML AND MKLML_IOMP_LIB)
+    message(STATUS "Enable Intel OpenMP with ${MKLML_IOMP_LIB}")
     set(OPENMP_FLAGS "-fopenmp")
     set(CMAKE_C_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
     set(CMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OPENMP_FLAGS}")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OPENMP_FLAGS}")
-    else()
-        find_package(OpenMP)
-        if(OPENMP_FOUND)
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
-        else()
-            message(WARNING "Can not find OpenMP."
-                 "Some performance features in MKLDNN may not be available")
-        endif()
-    endif()
-
-endif(WITH_MKLDNN)
+endif()
 
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SIMD_FLAG}")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SIMD_FLAG}")

cmake/cross_compiling/ios.cmake

@@ -76,11 +76,9 @@ set(IOS_PLATFORM ${IOS_PLATFORM} CACHE STRING "Type of iOS Platform")
 # Set the architecture for iOS
 if(NOT DEFINED IOS_ARCH)
   if(IOS_PLATFORM STREQUAL "OS")
-    # FIXME(liuyiqun): support "armv7;armv7s;arm64" future
-    set(IOS_ARCH "arm64")
+    set(IOS_ARCH "armv7;armv7s;arm64")
   elseif(IOS_PLATFORM STREQUAL "SIMULATOR")
-    # FIXME(liuyiqun): support "i386;x86_64" future
-    set(IOS_ARCH "x86_64")
+    set(IOS_ARCH "i386;x86_64")
   endif()
 endif()
 set(CMAKE_OSX_ARCHITECTURES ${IOS_ARCH} CACHE string  "Build architecture for iOS")
@@ -248,7 +246,7 @@ set(IOS_COMPILER_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} ${XCODE_IOS_BITCODE_
 
 # Hidden visibilty is required for cxx on iOS 
 set(CMAKE_C_FLAGS "${IOS_COMPILER_FLAGS} ${CMAKE_C_FLAGS}" CACHE STRING "C flags")
-set(CMAKE_CXX_FLAGS "${IOS_COMPILER_FLAGS} -fvisibility-inlines-hidden ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags")
+set(CMAKE_CXX_FLAGS "${IOS_COMPILER_FLAGS} -fvisibility=hidden -fvisibility-inlines-hidden ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags")
 
 set(IOS_LINK_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} -Wl,-search_paths_first")
 

cmake/cuda.cmake

@@ -63,7 +63,6 @@ function(select_nvcc_arch_flags out_variable)
   set(archs_name_default "All")
   if(NOT CMAKE_CROSSCOMPILING)
     list(APPEND archs_names "Auto")
-    set(archs_name_default "Auto")
   endif()
 
   # set CUDA_ARCH_NAME strings (so it will be seen as dropbox in CMake-Gui)

cmake/external/mkldnn.cmake

@@ -40,10 +40,9 @@ INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR})
 
 IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
     SET(MKLDNN_DEPENDS   ${MKLML_PROJECT})
-    SET(MKLDNN_MKLROOT   ${MKLML_ROOT})
-    SET(MKLDNN_IOMP_LIB  ${MKLML_IOMP_LIB})
-    SET(MKLDNN_IOMP_DIR  ${MKLML_LIB_DIR})
-    MESSAGE(STATUS "Build MKLDNN with ${MKLDNN_MKLROOT}")
+    MESSAGE(STATUS "Build MKLDNN with MKLML ${MKLML_ROOT}")
+ELSE()
+    MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
 ENDIF()
 
 SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow")
@@ -57,15 +56,16 @@ ExternalProject_Add(
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
-    CMAKE_ARGS          -DMKLROOT=${MKLDNN_MKLROOT}
+    CMAKE_ARGS          -DMKLROOT=${MKLML_ROOT}
     CMAKE_ARGS          -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
     CMAKE_ARGS          -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
     CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
-                        -DMKLROOT:PATH=${MKLDNN_MKLROOT}
+                        -DMKLROOT:PATH=${MKLML_ROOT}
 )
 
 ADD_LIBRARY(mkldnn SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
 ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
-MESSAGE(STATUS "Mkldnn library: ${MKLDNN_LIB}")
+MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
+add_definitions(-DPADDLE_USE_MKLDNN)
 LIST(APPEND external_project_dependencies mkldnn)

cmake/external/openblas.cmake

@@ -29,7 +29,7 @@ IF(NOT ${CBLAS_FOUND})
         "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
         CACHE FILEPATH "openblas library." FORCE)
 
-    SET(OPENBLAS_CC "${CMAKE_C_COMPILER}")
+    SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable")
 
     IF(CMAKE_CROSSCOMPILING)
         SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER})
@@ -45,15 +45,14 @@ IF(NOT ${CBLAS_FOUND})
                 SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0)
             ENDIF()
         ELSEIF(IOS)
-            # FIXME(liuyiqun): support multiple architectures
+            IF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
                 SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5")
                 SET(OPENBLAS_CC "${OPENBLAS_CC} ${CMAKE_C_FLAGS} -isysroot ${CMAKE_OSX_SYSROOT}")
-            IF(CMAKE_OSX_ARCHITECTURES MATCHES "armv7")
-                SET(OPENBLAS_CC "${OPENBLAS_CC} -arch armv7")
-                SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0)
-            ELSEIF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
                 SET(OPENBLAS_CC "${OPENBLAS_CC} -arch arm64")
                 SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0 CROSS_SUFFIX=${CROSS_SUFFIX})
+            ELSE()
+                MESSAGE(FATAL_ERROR "OpenBLAS only support arm64 architectures on iOS. "
+                       "You can set IOS_USE_VECLIB_FOR_BLAS=ON or USE_EIGEN_FOR_BLAS=ON to use other blas library instead.")
             ENDIF()
         ELSEIF(RPI)
             # use hardfp

cmake/external/warpctc.cmake

@@ -12,6 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+IF(MOBILE_INFERENCE)
+    return()
+ENDIF()
+
 INCLUDE(ExternalProject)
 
 SET(WARPCTC_SOURCES_DIR ${THIRD_PARTY_PATH}/warpctc)

cmake/util.cmake

@@ -115,8 +115,8 @@ function(link_paddle_exe TARGET_NAME)
         target_link_libraries(${TARGET_NAME} log)
     endif(ANDROID)
 
-    if(WITH_MKLDNN AND WITH_MKLML AND MKLDNN_IOMP_DIR)
-      target_link_libraries(${TARGET_NAME} "-L${MKLDNN_IOMP_DIR} -liomp5 -Wl,--as-needed")
+    if(WITH_MKLML AND MKLML_LIB_DIR AND MKLML_IOMP_LIB)
+      target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
     endif()
 
     add_dependencies(${TARGET_NAME} ${external_project_dependencies})

doc/design/mkldnn/README.MD

@@ -36,13 +36,13 @@ Figure 1. PaddlePaddle on IA.
 我们把集成方案大致分为了如下几个方面。
 
 ### CMake
-我们会在`CMakeLists.txt`中会添加`WITH_MKLDNN`的选项，当设置这个值为`ON`的时候会启用编译MKL-DNN功能。同时会自动开启OpenMP用于提高MKL-DNN的性能。
+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。
 
-同时，我们会引入`WITH_MKLML`选项，用于选择是否使用MKL-DNN自带的MKLML安装包。这个安装包可以独立于MKL-DNN使用，但是建议在开启MKL-DNN的同时也打开MKLML的开关，这样才能发挥最好的性能。
+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。
 
-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。
+当关闭`WITH_MKL`时，MKLML和MKL-DNN功能会同时关闭。
 
-**备注**：当`WITH_MKLML=ON`的时候，会优先使用这个包作为PaddlePaddle的CBLAS和LAPACK库，所以会稍微改动`cmake/cblas.cmake`中的逻辑。
+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。
 
 ### Layers
 所有MKL-DNN相关的C++ layers，都会按照PaddlePaddle的目录结构存放在

doc/howto/dev/write_docs_cn.rst

@@ -34,7 +34,7 @@ PaddlePaddle的文档构建有两种方式。
     cd TO_YOUR_PADDLE_CLONE_PATH
     mkdir -p build
     cd build
-    cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKLDNN=OFF -DWITH_MKLML=OFF -DWITH_DOC=ON
+    cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON
     make gen_proto_py
     make paddle_docs paddle_docs_cn
 

doc/mobile/cross_compiling_for_android_cn.md

-# 构建Android平台上的PaddlePaddle库
+# Android平台编译指南
 
 用户可通过如下两种方式，交叉编译Android平台上适用的PaddlePaddle库：
 - 基于Docker容器的编译方式

doc/mobile/cross_compiling_for_ios_cn.md

-# 构建iOS平台上的PaddlePaddle库
+# iOS平台编译指南
 交叉编译iOS平台上适用的PaddlePaddle库，需要在MacOS系统上进行。本文的将介绍在MacOS上，从源码交叉编译iOS平台上适用的PaddlePaddle库。
 
 ## 准备交叉编译环境
@@ -25,7 +25,7 @@ iOS平台可选配置参数：
 - `IOS_PLATFORM`，可设置为`OS/SIMULATOR`，默认值为`OS`。
   - `OS`，构建目标为`arm`架构的iPhone或者iPad等物理设备。
   - `SIMULATOR`，构建目标为`x86`架构的模拟器平台。
-- `IOS_ARCH`，目标架构。针对不同的`IOS_PLATFORM`，可设置的目标架构如下表所示：
+- `IOS_ARCH`，目标架构。针对不同的`IOS_PLATFORM`，可设置的目标架构如下表所示，默认编译所有架构：
 
     <table class="docutils">
     <colgroup>
@@ -41,11 +41,11 @@ iOS平台可选配置参数：
     <tbody valign="top">
       <tr class="row-even">
       <td>OS</td>
-      <td>armv7, armv7s, arm64 (默认)</td>
+      <td>armv7, armv7s, arm64 </td>
     </tr>
     <tr class="row-odd">
       <td>SIMULATOR</td>
-      <td>i386, x86_64 (默认)</td>
+      <td>i386, x86_64 </td>
     </tr>
     </tbody>
     </table>
@@ -66,7 +66,7 @@ iOS平台可选配置参数：
 ```bash
 cmake -DCMAKE_SYSTEM_NAME=iOS \
       -DIOS_PLATFORM=OS \
-      -DIOS_ARCH="arm64" \
+      -DIOS_ARCH="armv7;arm64" \
       -DIOS_ENABLE_BITCODE=ON \
       -DIOS_USE_VECLIB_FOR_BLAS=ON \
       -DCMAKE_INSTALL_PREFIX=your/path/to/install \
@@ -112,6 +112,6 @@ $ make install
 - `lib`目录，其中包含PaddlePaddle的C-API静态库
 - `third_party`目录，其中包含所依赖的所有第三方库
 
-注意，不同架构的PaddlePaddle库建议安装到不同的目录下，然后使用`lipo`工具将多个静态库合并成一个支持多个架构的fat库。
+注意，如果PaddlePaddle库需要同时支持真机和模拟器，则需要分别编译真机和模拟器版本，然后使用`lipo`工具合并fat库。
 
 自此，PaddlePaddle库已经安装完成，用户可将合成的fat库用于深度学习相关的iOS App中，调用方法见C-API文档。

doc/mobile/cross_compiling_for_raspberry_cn.md

-# 构建Raspberry Pi平台上的PaddlePaddle库
+# Raspberry Pi平台编译指南
 
 通常有两个方法来构建基于 Rasspberry Pi 的版本：
 

paddle/cuda/include/hl_gpu.h

@@ -25,7 +25,9 @@ limitations under the License. */
 #include "hl_matrix.h"
 #include "hl_sequence.h"
 #include "hl_sparse.h"
+#ifndef PADDLE_MOBILE_INFERENCE
 #include "hl_warpctc_wrap.h"
+#endif
 
 #ifdef HPPL_STUB_FUNC
 #include "stub/hl_aggregate_stub.h"

paddle/framework/backward.cc

@@ -270,6 +270,19 @@ static bool AllGradInSet(const std::vector<std::string>& names,
       return false;
     }
   }
+  if (VLOG_IS_ON(10)) {
+    std::ostringstream sout;
+    sout << "All input {";
+    for (auto& name : names) {
+      sout << name << ",";
+    }
+    sout << "} is in {";
+    for (auto& name : set) {
+      sout << name << ",";
+    }
+    sout << "}";
+    VLOG(10) << sout.str();
+  }
   return true;
 }
 
@@ -290,14 +303,12 @@ static void CreateGradVarInBlock(
   auto ops = block_desc->AllOps();
   for (size_t op_index = grad_op_start_index; op_index < ops.size();
        ++op_index) {
-    bool need_infer_shape = false;
     std::unordered_set<std::string> new_vars;
     ForEachVarName(ops[op_index]->Outputs(),
                    [&](const std::string& grad_var_name) {
                      if (block_desc->HasVar(grad_var_name)) {
                        return false;
                      }
-                     need_infer_shape = true;
                      auto var = block_desc->Var(grad_var_name);
                      new_vars.insert(var->Name());
                      auto it = param_name_map.find(grad_var_name);
@@ -311,7 +322,6 @@ static void CreateGradVarInBlock(
                      grad_record.op_idx_ = static_cast<int>(op_index);
                      return false; /* not break */
                    });
-    if (need_infer_shape) {
     ops[op_index]->InferVarType(block_desc);
     for (auto& arg : ops[op_index]->OutputArgumentNames()) {
       if (new_vars.find(arg) == new_vars.end()) {
@@ -328,7 +338,6 @@ static void CreateGradVarInBlock(
     }
     ops[op_index]->InferShape(*block_desc);
   }
-  }
 }
 
 std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
@@ -387,6 +396,7 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
     ProgramDescBind& program_desc, int block_idx,
     std::unordered_set<std::string>* no_grad_vars,
     std::unordered_map<std::string, std::string>* grad_to_var) {
+  VLOG(5) << "MakeBlockBackward";
   BlockDescBind* cur_block = program_desc.MutableBlock(block_idx);
   std::vector<OpDescBind*> op_descs = cur_block->AllOps();
   std::unordered_map<std::string, std::vector<size_t>> dup_out_ops;
@@ -394,9 +404,10 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
   std::vector<std::unique_ptr<OpDescBind>> backward_descs;
 
   for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) {
+    VLOG(5) << "Making backward " << (*it)->Type() << " op";
     std::vector<std::unique_ptr<OpDescBind>> op_grads;
 
-    if ((*it)->Type() == "recurrent") {
+    if ((*it)->Type() == "recurrent" || (*it)->Type() == "while") {
       int step_block_idx = (*it)->GetBlockAttr("step_block");
       BlockDescBind* backward_block = CreateStepBlock(
           program_desc, no_grad_vars, grad_to_var, step_block_idx);
@@ -410,6 +421,15 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
       op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var);
     }
 
+    if (VLOG_IS_ON(10)) {
+      std::ostringstream sout;
+      sout << "Made ";
+      for (auto& op_grad : op_grads) {
+        sout << op_grad->Type() << " ";
+      }
+      VLOG(10) << sout.str();
+    }
+
     for (const auto& desc : op_grads) {
       for (const std::string& out_name : desc->OutputArgumentNames()) {
         if (out_name.find("@GRAD") == std::string::npos) {
@@ -425,6 +445,8 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
         op_grads.begin(), op_grads.end(), std::back_inserter(backward_descs),
         [](std::unique_ptr<OpDescBind>& ptr) { return std::move(ptr); });
   }
+
+  VLOG(5) << "Appending Sums";
   // Check whether some variables are written more than once
   std::list<std::pair<size_t, std::unique_ptr<OpDescBind>>> pending_sum_ops;
   for (const auto& dup : dup_out_ops) {
@@ -432,16 +454,22 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
     const std::vector<size_t> dup_op = dup.second;
     if (out_name != kEmptyVarName && dup_op.size() > 1) {
       std::vector<std::string> sum_op_inputs;
+      std::string next_g_name = out_name;
       for (size_t i = 0; i < dup_op.size(); ++i) {
+        VLOG(10) << backward_descs[dup_op[i]]->Type() << " has " << out_name
+                 << " duplicated";
         std::string new_name = out_name + "@RENAME@" + std::to_string(i);
-        backward_descs[dup_op[i]]->Rename(out_name, new_name);
+        backward_descs[dup_op[i]]->RenameOutput(out_name, new_name);
+        backward_descs[dup_op[i]]->RenameInput(out_name, next_g_name);
         sum_op_inputs.emplace_back(new_name);
+        next_g_name = sum_op_inputs.back();
       }
       std::unique_ptr<OpDescBind> sum_op(new OpDescBind(
           "sum", {{"X", sum_op_inputs}}, {{"Out", {out_name}}}, {}));
       pending_sum_ops.push_back({dup_op.back(), std::move(sum_op)});
     }
   }
+
   pending_sum_ops.sort(
       [](const std::pair<size_t, std::unique_ptr<OpDescBind>>& a,
          const std::pair<size_t, std::unique_ptr<OpDescBind>>& b) {
@@ -452,6 +480,8 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
                           std::move(p.second));
   }
 
+  VLOG(5) << "MakeBlockBackward Finished";
+
   return backward_descs;
 }
 

paddle/framework/data_type.h

@@ -29,6 +29,8 @@ inline DataType ToDataType(std::type_index type) {
     return DataType::INT32;
   } else if (typeid(int64_t).hash_code() == type.hash_code()) {
     return DataType::INT64;
+  } else if (typeid(bool).hash_code() == type.hash_code()) {
+    return DataType::BOOL;
   } else {
     PADDLE_THROW("Not supported");
   }

paddle/framework/ddim.cc

@@ -60,8 +60,7 @@ void make_ddim(DDim& ddim, const int64_t* dims, int n) {
       ddim = make_dim<9>(dims);
       break;
     default:
-      throw std::invalid_argument(
-          "Dynamic dimensions must have between [1, 9] dimensions.");
+      PADDLE_THROW("Dynamic dimensions must have between [1, 9] dimensions.");
   }
 }
 

paddle/framework/executor.cc

@@ -120,6 +120,7 @@ void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id,
 
   for (auto& op_desc : block.AllOps()) {
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
+    VLOG(10) << op->DebugString();
     op->Run(*local_scope, *device);
   }
   if (create_local_scope) {

paddle/framework/op_desc.cc

@@ -235,6 +235,23 @@ void OpDescBind::Rename(const std::string &old_name,
   need_update_ = true;
 }
 
+void OpDescBind::RenameOutput(const std::string &old_name,
+                              const std::string &new_name) {
+  for (auto &output : outputs_) {
+    std::replace(output.second.begin(), output.second.end(), old_name,
+                 new_name);
+  }
+  need_update_ = true;
+}
+
+void OpDescBind::RenameInput(const std::string &old_name,
+                             const std::string &new_name) {
+  for (auto &input : inputs_) {
+    std::replace(input.second.begin(), input.second.end(), old_name, new_name);
+  }
+  need_update_ = true;
+}
+
 struct SetAttrDescVisitor : public boost::static_visitor<void> {
   explicit SetAttrDescVisitor(OpDesc::Attr *attr) : attr_(attr) {}
   mutable OpDesc::Attr *attr_;
@@ -448,7 +465,12 @@ const std::vector<std::string> &CompileTimeInferShapeContext::Outputs(
 DDim CompileTimeInferShapeContext::GetDim(const std::string &name) const {
   auto var = block_.FindVarRecursive(name);
   PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name);
+  try {
     return framework::make_ddim(var->Shape());
+  } catch (...) {
+    VLOG(5) << "GetDim of variable " << name << " error";
+    std::rethrow_exception(std::current_exception());
+  }
 }
 
 void CompileTimeInferShapeContext::SetDim(const std::string &name,

paddle/framework/op_desc.h

@@ -73,6 +73,10 @@ class OpDescBind {
 
   void Rename(const std::string &old_name, const std::string &new_name);
 
+  void RenameOutput(const std::string &old_name, const std::string &new_name);
+
+  void RenameInput(const std::string &old_name, const std::string &new_name);
+
   // Only be used in C++
   const AttributeMap &GetAttrMap() const;
 

paddle/framework/operator.cc

@@ -403,19 +403,6 @@ class RuntimeInferShapeContext : public InferShapeContext {
 
 void OperatorWithKernel::Run(const Scope& scope,
                              const platform::DeviceContext& dev_ctx) const {
-  if (VLOG_IS_ON(1)) {
-    auto inputs = this->InputVars();
-    auto outputs = this->OutputVars(true);
-    std::ostringstream sout;
-    sout << "Run operator " << this->Type() << " From [";
-    std::ostream_iterator<std::string> out_it(sout, ",");
-    std::copy(inputs.begin(), inputs.end(), out_it);
-    sout << "] to [";
-    std::copy(outputs.begin(), outputs.end(), out_it);
-    sout << "]";
-    VLOG(1) << sout.str();
-  }
-
   RuntimeInferShapeContext infer_shape_ctx(*this, scope);
   this->InferShape(&infer_shape_ctx);
 

paddle/framework/scope.cc

@@ -38,11 +38,12 @@ Scope& Scope::NewScope() const {
 Variable* Scope::Var(const std::string& name) {
   auto iter = vars_.find(name);
   if (iter != vars_.end()) {
+    VLOG(3) << "Get existing variable " << name;
     return iter->second;
   }
   Variable* v = new Variable();
   vars_[name] = v;
-  VLOG(3) << "Create variable " << name << " on scope";
+  VLOG(3) << "Create variable " << name;
   v->name_ = &(vars_.find(name)->first);
   return v;
 }

paddle/framework/shape_inference.h

@@ -53,6 +53,10 @@ class InferShapeContext {
 
   virtual bool IsRuntime() const = 0;
 
+  // Note: In while op, we need this to be public
+  void SetDims(const std::vector<std::string> &names,
+               const std::vector<framework::DDim> &dims);
+
  protected:
   virtual framework::DDim GetDim(const std::string &name) const = 0;
   virtual void SetDim(const std::string &name, const framework::DDim &dim) = 0;
@@ -60,9 +64,6 @@ class InferShapeContext {
   std::vector<framework::DDim> GetDims(
       const std::vector<std::string> &names) const;
 
-  void SetDims(const std::vector<std::string> &names,
-               const std::vector<framework::DDim> &dims);
-
   std::vector<VarDesc::VarType> GetVarTypes(
       const std::vector<std::string> &names) const;
 

paddle/gserver/layers/MKLDNNLayer.cpp

@@ -22,7 +22,7 @@ namespace paddle {
 bool MKLDNNLayer::init(const LayerMap& layerMap,
                        const ParameterMap& parameterMap) {
   CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
-                          << "Please set WITH_MKLDNN=ON "
+                          << "Please set WITH_MKL=ON "
                           << "and set use_mkldnn=True";
   CHECK(!useGpu_) << "Do not support GPU yet";
 

paddle/math/Storage.cpp

@@ -17,9 +17,13 @@ limitations under the License. */
 #include "paddle/utils/StringUtil.h"
 #include "paddle/utils/Util.h"
 
+#ifndef PADDLE_MOBILE_INFERENCE
 DEFINE_int32(pool_limit_size,
              536870912,
              "maximum memory size managed by a memory pool, default is 512M");
+#else
+DEFINE_int32(pool_limit_size, 0, "default is 0");
+#endif
 
 namespace paddle {
 

paddle/operators/array_operator.h

@@ -42,6 +42,7 @@ class ArrayOp : public framework::OperatorBase {
     } else {
       offset = static_cast<size_t>(*i_tensor.data<int64_t>());
     }
+    VLOG(10) << " Offset = " << offset;
     return offset;
   }
 };

paddle/operators/bilinear_tensor_product_op.h

@@ -174,7 +174,7 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
     // Caculate the gradient of Input(Bias).
     if (d_bias) {
       d_bias->mutable_data<T>(ctx.GetPlace());
-      auto d_bias_mat = EigenMatrix<T>::From(*d_bias);
+      auto d_bias_mat = framework::EigenVector<T>::Flatten(*d_bias);
       d_bias_mat.device(place) = d_out_mat.sum(Eigen::DSizes<int, 1>(0));
     }
   }

paddle/operators/conv_transpose_op.cc

@@ -30,11 +30,6 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const {
   std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
   std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
 
-  for (size_t i = 0; i < paddings.size(); ++i) {
-    PADDLE_ENFORCE_EQ(paddings[i], 0,
-                      "No Padding allowed in conv transpose op.");
-  }
-
   PADDLE_ENFORCE(in_dims.size() == 4 || in_dims.size() == 5,
                  "ConvTransposeOp intput should be 4-D or 5-D tensor.");
   PADDLE_ENFORCE_EQ(in_dims.size(), filter_dims.size(),
@@ -52,7 +47,7 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const {
 
   std::vector<int64_t> output_shape({in_dims[0], filter_dims[1]});
   for (size_t i = 0; i < strides.size(); ++i) {
-    output_shape.push_back((in_dims[i + 2] - 1) * strides[i] +
+    output_shape.push_back((in_dims[i + 2] - 1) * strides[i] - 2 * paddings[i] +
                            filter_dims[i + 2]);
   }
   ctx->SetOutputDim("Output", framework::make_ddim(output_shape));

paddle/operators/conv_transpose_op.h

@@ -62,7 +62,6 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
     Tensor* output = context.Output<Tensor>("Output");
 
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
-    // Actually, no paddings and groups allowed in conv transpose.
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
     // TODO(Zhuoyuan): Paddings can be added in future.
     // groups will alway be disabled in conv2dtranspose.
@@ -148,8 +147,8 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
       } else if (filter_shape_vec.size() == 3) {
         // col2vol: col_matrix -> dy
         // from (c * k_d * k_h * k_w, d * h * w) to (c, o_d, o_h, o_w)
-        col2vol(context.device_context(), col, dilations, strides,
-                std::vector<int>{0, 0, 0}, &output_batch);
+        col2vol(context.device_context(), col, dilations, strides, paddings,
+                &output_batch);
       }
     }
   }
@@ -173,7 +172,6 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
     if ((!input_grad) && (!filter_grad)) return;
 
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
-    // Actually, no paddings and groups allowed in conv transpose.
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
 
     const int batch_size = static_cast<int>(input->dims()[0]);

paddle/operators/cos_sim_op.h

@@ -132,7 +132,7 @@ class CosSimGradKernel : public framework::OpKernel<T> {
       // compute dy
       if (out_grad_y) {
         out_grad_y->mutable_data<T>(context.GetPlace());
-        auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1);
+        auto dy = EigenVector<T>::Flatten(*out_grad_y);
         auto grad = x / norm_prod_bcast - z_bcast * y_bcast / y_snorm_bcast;
         dy.device(place) = (dz_bcast * grad).sum(Eigen::array<int, 1>({{0}}));
       }

paddle/operators/detail/safe_ref.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
+
+namespace paddle {
+namespace operators {
+namespace detail {
+/**
+ * Get Reference From Pointer with check. The error message is printf format,
+ * and passed by `args`
+ */
+template <typename T, typename... ARGS>
+inline T &Ref(T *ptr, ARGS &&... args) {
+  PADDLE_ENFORCE(ptr != nullptr, args...);
+  return *ptr;
+}
+}  // namespace detail
+}  // namespace operators
+}  // namespace paddle

paddle/operators/fill_constant_batch_size_like_op.cc

@@ -101,4 +101,7 @@ REGISTER_OPERATOR(fill_constant_batch_size_like,
 REGISTER_OP_CPU_KERNEL(
     fill_constant_batch_size_like,
     ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace, float>,
-    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace, double>);
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace, double>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace, int>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace,
+                                           int64_t>);

paddle/operators/fill_constant_batch_size_like_op.cu.cc

@@ -19,4 +19,7 @@ namespace ops = paddle::operators;
 REGISTER_OP_GPU_KERNEL(
     fill_constant_batch_size_like,
     ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace, float>,
-    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace, double>);
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace, double>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace, int>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace,
+                                           int64_t>);

paddle/operators/fill_zeros_like_op.cc

@@ -54,5 +54,8 @@ namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, ops::FillZerosLikeOp,
                              ops::FillZerosLikeOpMaker);
 REGISTER_OP_CPU_KERNEL(
-    fill_zeros_like,
-    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, float>);
+    fill_zeros_like, ops::FillZerosLikeKernel<paddle::platform::CPUPlace, int>,
+    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, int64_t>,
+    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, float>,
+    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, double>,
+    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, bool>);

paddle/operators/fill_zeros_like_op.cu.cc

@@ -17,5 +17,8 @@
 
 namespace ops = paddle::operators;
 REGISTER_OP_GPU_KERNEL(
-    fill_zeros_like,
-    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, float>);
+    fill_zeros_like, ops::FillZerosLikeKernel<paddle::platform::GPUPlace, int>,
+    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, int64_t>,
+    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, float>,
+    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, double>,
+    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, bool>);

paddle/operators/is_empty_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/framework/op_registry.h"
+#include "paddle/framework/operator.h"
+
+namespace paddle {
+namespace operators {
+
+constexpr char kInput[] = "X";
+constexpr char kOutput[] = "Out";
+
+class IsEmptyOp : public framework::OperatorBase {
+ public:
+  IsEmptyOp(const std::string &type, const framework::VariableNameMap &inputs,
+            const framework::VariableNameMap &outputs,
+            const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    // get input
+    auto *var = scope.FindVar(Input(kInput));
+    PADDLE_ENFORCE_NOT_NULL(var);
+    auto &tensor = var->Get<framework::LoDTensor>();
+    // get output
+    auto *out = scope.FindVar(Output(kOutput));
+    PADDLE_ENFORCE_NOT_NULL(out);
+    auto *out_tensor = out->GetMutable<framework::LoDTensor>();
+
+    out_tensor->Resize({1});
+    out_tensor->mutable_data<bool>(platform::CPUPlace())[0] =
+        framework::product(tensor.dims()) == 0;
+  }
+};
+
+class IsEmptyOpProtoMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  IsEmptyOpProtoMaker(framework::OpProto *proto,
+                      framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput(kInput, "(Tensor) Tensor which is to be checked.");
+    AddOutput(kOutput, "(Tensor) a boolean Tensor that indicate empty or not.");
+    AddComment(R"DOC(
+IsEmpty Operator which checks whether a tensor is empty.
+
+It will just return product(tensor.ddims()) > 0;
+              )DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_WITHOUT_GRADIENT(is_empty, paddle::operators::IsEmptyOp,
+                             paddle::operators::IsEmptyOpProtoMaker);

paddle/operators/math/CMakeLists.txt

 add_subdirectory(detail)
 
 if(WITH_GPU)
-    nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc im2col.cu DEPS cblas device_context)
+    nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc im2col.cu DEPS cblas device_context framework_proto)
     nv_test(math_function_gpu_test SRCS math_function_test.cu DEPS math_function tensor)
     nv_library(selected_rows_functor SRCS selected_rows_functor.cc selected_rows_functor.cu DEPS selected_rows math_function)
     nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor)
@@ -15,7 +15,7 @@ if(WITH_GPU)
     nv_library(lstm_compute SRCS lstm_compute.cc lstm_compute.cu DEPS device_context activation_functions)
     nv_library(gru_compute SRCS gru_compute.cc gru_compute.cu DEPS device_context activation_functions math_function)
 else()
-    cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context)
+    cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context framework_proto)
     cc_library(selected_rows_functor SRCS selected_rows_functor.cc DEPS selected_rows math_function)
     cc_library(softmax SRCS softmax.cc DEPS device_context)
     cc_library(cross_entropy SRCS cross_entropy.cc DEPS device_context)

paddle/operators/math/im2col.cu

@@ -119,8 +119,8 @@ __global__ void col2im(int n, const T* data_col, int im_height, int im_width,
 
   if (index < n) {
     T val = 0;
-    int w = index % im_width;
-    int h = (index / im_width) % im_height;
+    int w = index % im_width + padding_width;
+    int h = (index / im_width) % im_height + padding_height;
     int c = index / (im_width * im_height);
 
     // compute the start and end of the output

paddle/operators/math/math_function.cc

@@ -250,6 +250,8 @@ void axpy<platform::CPUPlace, double>(const platform::DeviceContext& context,
 template struct SetConstant<platform::CPUPlace, float>;
 template struct SetConstant<platform::CPUPlace, double>;
 template struct SetConstant<platform::CPUPlace, int>;
+template struct SetConstant<platform::CPUPlace, int64_t>;
+template struct SetConstant<platform::CPUPlace, bool>;
 
 #define DEFINE_CPU_TRANS(RANK)                                \
   template struct Transpose<platform::CPUPlace, float, RANK>; \

paddle/operators/math/math_function.cu

@@ -256,6 +256,8 @@ void axpy<platform::GPUPlace, double>(const platform::DeviceContext& context,
 template struct SetConstant<platform::GPUPlace, float>;
 template struct SetConstant<platform::GPUPlace, double>;
 template struct SetConstant<platform::GPUPlace, int>;
+template struct SetConstant<platform::GPUPlace, int64_t>;
+template struct SetConstant<platform::GPUPlace, bool>;
 
 #define DEFINE_GPU_TRANS(RANK)                                \
   template struct Transpose<platform::GPUPlace, float, RANK>; \

paddle/operators/sum_op.cc

@@ -12,6 +12,7 @@ limitations under the License. */
 #include "paddle/operators/sum_op.h"
 #include <vector>
 #include "paddle/framework/var_type_inference.h"
+#include "paddle/operators/detail/safe_ref.h"
 
 namespace paddle {
 namespace operators {
@@ -59,7 +60,8 @@ class SumOp : public framework::OperatorWithKernel {
               x_vars[0]->Get<framework::SelectedRows>().value().type()),
           ctx.device_context());
     } else if (x_vars[0]->IsType<framework::LoDTensorArray>()) {
-      auto& array = x_vars[0]->Get<framework::LoDTensorArray>();
+      for (auto& x_var : x_vars) {
+        auto& array = x_var->Get<framework::LoDTensorArray>();
         for (auto& each : array) {
           if (each.numel() != 0) {
             return framework::OpKernelType(framework::ToDataType(each.type()),
@@ -67,6 +69,8 @@ class SumOp : public framework::OperatorWithKernel {
           }
         }
       }
+      PADDLE_THROW("Cannot find the input data type by all input data");
+    }
     PADDLE_THROW("Unexpected branch. Input type is %s",
                  x_vars[0]->Type().name());
   }
@@ -96,6 +100,11 @@ class SumOpVarTypeInference : public framework::VarTypeInference {
     auto& inputs = op_desc.Input("X");
     auto var_type = framework::VarDesc::SELECTED_ROWS;
 
+    for (auto& name : op_desc.Input("X")) {
+      VLOG(10) << name << " "
+               << block->FindRecursiveOrCreateVar(name)->GetType();
+    }
+
     bool any_input_is_lod_tensor = std::any_of(
         inputs.begin(), inputs.end(), [block](const std::string& name) {
           return block->FindRecursiveOrCreateVar(name)->GetType() ==
@@ -103,7 +112,7 @@ class SumOpVarTypeInference : public framework::VarTypeInference {
         });
 
     auto is_tensor_array = [block](const std::string& name) {
-      return block->FindRecursiveOrCreateVar(name)->GetType() ==
+      return detail::Ref(block->FindRecursiveOrCreateVar(name)).GetType() ==
              framework::VarDesc::LOD_TENSOR_ARRAY;
     };
 
@@ -113,14 +122,26 @@ class SumOpVarTypeInference : public framework::VarTypeInference {
         std::all_of(inputs.begin(), inputs.end(), is_tensor_array);
 
     if (any_input_is_tensor_array) {
-      PADDLE_ENFORCE(all_inputs_are_tensor_array);
+      if (!all_inputs_are_tensor_array) {
+        std::ostringstream os;
+        for (auto& each : inputs) {
+          os << "    " << each << " type is "
+             << detail::Ref(block->FindRecursiveOrCreateVar(each)).GetType()
+             << "\n";
+        }
+        PADDLE_ENFORCE(all_inputs_are_tensor_array,
+                       "Not all inputs are tensor array:\n%s", os.str());
+      }
       var_type = framework::VarDesc::LOD_TENSOR_ARRAY;
     } else if (any_input_is_lod_tensor) {
       var_type = framework::VarDesc::LOD_TENSOR;
     }
 
     auto out_var_name = op_desc.Output("Out").front();
-    block->FindRecursiveOrCreateVar(out_var_name)->SetType(var_type);
+    auto& out_var = detail::Ref(block->FindRecursiveOrCreateVar(out_var_name));
+    out_var.SetType(var_type);
+    auto& in_var = detail::Ref(block->FindVarRecursive(inputs.front()));
+    out_var.SetDataType(in_var.GetDataType());
   }
 };
 

paddle/operators/tensor_array_read_write_op.cc

@@ -12,7 +12,7 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 #include "paddle/operators/array_operator.h"
-
+#include "paddle/operators/detail/safe_ref.h"
 namespace paddle {
 namespace operators {
 
@@ -33,6 +33,8 @@ class WriteToArrayOp : public ArrayOp {
     auto *out =
         scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensorArray>();
     if (offset >= out->size()) {
+      VLOG(10) << "Resize " << Output("Out") << " from " << out->size()
+               << " to " << offset + 1;
       out->resize(offset + 1);
     }
     auto *out_tensor = &out->at(offset);
@@ -85,11 +87,15 @@ class WriteToArrayInferVarType : public framework::VarTypeInference {
  public:
   void operator()(const framework::OpDescBind &op_desc,
                   framework::BlockDescBind *block) const override {
-    for (auto &out_var : op_desc.OutputArgumentNames()) {
-      VLOG(10) << "Set Variable " << out_var << " as LOD_TENSOR_ARRAY";
-      block->FindRecursiveOrCreateVar(out_var)->SetType(
-          framework::VarDesc::LOD_TENSOR_ARRAY);
-    }
+    auto x_name = op_desc.Input("X")[0];
+    auto out_name = op_desc.Output("Out")[0];
+    VLOG(10) << "Set Variable " << out_name << " as LOD_TENSOR_ARRAY";
+    auto &out = detail::Ref(block->FindRecursiveOrCreateVar(out_name),
+                            "Cannot found %s", out_name);
+    out.SetType(framework::VarDesc::LOD_TENSOR_ARRAY);
+    auto &x =
+        detail::Ref(block->FindVarRecursive(x_name), "Cannot found %s", x_name);
+    out.SetDataType(x.GetDataType());
   }
 };
 
@@ -107,11 +113,11 @@ class ReadFromArrayOp : public ArrayOp {
     auto &x_array = x->Get<framework::LoDTensorArray>();
     auto *out = scope.FindVar(Output("Out"));
     PADDLE_ENFORCE(out != nullptr, "Out must be set");
-    auto *out_tesnor = out->GetMutable<framework::LoDTensor>();
+    auto *out_tensor = out->GetMutable<framework::LoDTensor>();
     size_t offset = GetOffset(scope, dev_ctx);
     PADDLE_ENFORCE_LT(offset, x_array.size());
-    out_tesnor->CopyFrom(x_array[offset], dev_ctx.GetPlace(), dev_ctx);
-    out_tesnor->set_lod(x_array[offset].lod());
+    out_tensor->CopyFrom(x_array[offset], dev_ctx.GetPlace(), dev_ctx);
+    out_tensor->set_lod(x_array[offset].lod());
   }
 };
 

paddle/operators/while_op.cc

@@ -14,8 +14,10 @@
 
 #include <vector>
 #include "paddle/framework/executor.h"
+#include "paddle/framework/lod_tensor_array.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
+#include "paddle/operators/detail/safe_ref.h"
 
 namespace paddle {
 namespace operators {
@@ -26,8 +28,9 @@ using LoDTensor = framework::LoDTensor;
 constexpr char kStepBlock[] = "step_block";
 constexpr char kCondition[] = "Condition";
 constexpr char kStepScopes[] = "StepScopes";
-constexpr char kParamGrads[] = "X@Grad";
 constexpr char kParameters[] = "X";
+constexpr char kParamGrads[] = "X@GRAD";
+constexpr char kOutputs[] = "Out";
 
 class WhileOp : public framework::OperatorBase {
  public:
@@ -71,9 +74,9 @@ class WhileOpMaker : public framework::OpProtoAndCheckerMaker {
         kCondition,
         "(Bool) An scalar. When it's False, the While Op will be terminated.")
         .AsDuplicable();
-    AddOutput("Out",
+    AddOutput(kOutputs,
               "A set of variables, which will be assigned with values "
-              "generated by perators inside the block of While Op.")
+              "generated by the operators inside the block of While Op.")
         .AsDuplicable();
     AddOutput(kStepScopes,
               "(StepScopeVar) A vector of local scope, which size equals the "
@@ -104,17 +107,64 @@ class WhileGradOp : public framework::OperatorBase {
     auto *step_scopes =
         scope.FindVar(Input(kStepScopes))->GetMutable<StepScopeVar>();
 
+    auto outside_og_names = Inputs(framework::GradVarName(kOutputs));
+    auto inside_og_names =
+        Attr<std::vector<std::string>>("original_output_grad");
+
+    PADDLE_ENFORCE_EQ(outside_og_names.size(), inside_og_names.size());
+
     for (auto cur_scope_iter = step_scopes->rbegin();
          cur_scope_iter != step_scopes->rend(); ++cur_scope_iter) {
+      VLOG(3) << "Start backward at time_step "
+              << cur_scope_iter - step_scopes->rbegin();
+      framework::Scope &cur_scope = **cur_scope_iter;
+      // Link OG from outside to inside
+      for (size_t i = 0; i < outside_og_names.size(); ++i) {
+        auto outside_og_name = outside_og_names[i];
+        auto inside_og_name = inside_og_names[i];
+        VLOG(10) << "Linking outside " << outside_og_name << " --> inside "
+                 << inside_og_name;
+        auto &og_outside = detail::Ref(scope.FindVar(outside_og_name));
+        auto &og_inside = detail::Ref(cur_scope.Var(inside_og_name));
+        if (og_outside.Type().hash_code() ==
+            typeid(framework::LoDTensor).hash_code()) {
+          auto &outside_tensor = og_outside.Get<framework::LoDTensor>();
+          auto &inside_tensor =
+              detail::Ref(og_inside.GetMutable<framework::LoDTensor>());
+          inside_tensor.set_lod(outside_tensor.lod());
+          inside_tensor.ShareDataWith(outside_tensor);
+        } else if (og_outside.Type().hash_code() ==
+                   typeid(framework::LoDTensorArray).hash_code()) {
+          auto &outside_array = og_outside.Get<framework::LoDTensorArray>();
+          auto &inside_array =
+              detail::Ref(og_inside.GetMutable<framework::LoDTensorArray>());
+          VLOG(10) << outside_og_name << " size = " << outside_array.size();
+          inside_array.resize(outside_array.size());
+
+          for (size_t j = 0; j < inside_array.size(); ++j) {
+            VLOG(10) << j << " " << outside_array[j].numel();
+            if (outside_array[j].numel() != 0) {
+              inside_array[j].set_lod(outside_array[j].lod());
+              inside_array[j].ShareDataWith(outside_array[j]);
+            } else {
+              PADDLE_ENFORCE_EQ(inside_array[j].numel(), 0);
+            }
+          }
+        }
+      }
+
       executor.Run(*program, *cur_scope_iter, block->ID(), false);
 
       auto &pg_names = Outputs(kParamGrads);
       auto &p_names = Inputs(kParameters);
       PADDLE_ENFORCE_EQ(pg_names.size(), p_names.size());
-      for (size_t prog_id = 0; prog_id < pg_names.size(); ++prog_id) {
-        auto inside_grad_name = framework::GradVarName(p_names[prog_id]);
+      for (size_t param_id = 0; param_id < pg_names.size(); ++param_id) {
+        if (pg_names[param_id] == framework::kEmptyVarName) {
+          continue;  // iterator doesn't have gradient
+        }
+        auto inside_grad_name = framework::GradVarName(p_names[param_id]);
 
-        //  // TODO(tonyyang-savil: Not sure we need the following
+        //  // TODO(tonyyang-svail): Not sure we need the following
         //  // If does not compute gradient of that variable inside rnn,
         //  just
         //  // continue
@@ -126,7 +176,7 @@ class WhileGradOp : public framework::OperatorBase {
         // zero gradient variable in step 0
         if (cur_scope_iter == step_scopes->rbegin()) {
           auto *var = (*cur_scope_iter)->FindVar(inside_grad_name);
-          PADDLE_ENFORCE_NOT_NULL(var);
+          PADDLE_ENFORCE_NOT_NULL(var, "Can not find var %s", inside_grad_name);
           if (var->IsType<LoDTensor>()) {
             auto &inside_tensor = var->Get<framework::LoDTensor>();
             framework::AttributeMap attrs;
@@ -135,27 +185,18 @@ class WhileGradOp : public framework::OperatorBase {
             attrs["value"] = 0.0f;
 
             auto zero_op = framework::OpRegistry::CreateOp(
-                "fill_constant", {}, {{"Out", {pg_names[prog_id]}}}, attrs);
+                "fill_constant", {}, {{"Out", {pg_names[param_id]}}}, attrs);
             zero_op->Run(scope, dev_ctx);
           }
         }
 
         // sum gradient
-        auto *outside_var = scope.FindVar(pg_names[prog_id]);
-        PADDLE_ENFORCE_NOT_NULL(outside_var);
-        auto &outside_tensor = *outside_var->GetMutable<framework::LoDTensor>();
-
-        std::string result_var_name;
-        auto *local_result_var = (*cur_scope_iter)->Var(&result_var_name);
-        auto &local_result_tensor =
-            *local_result_var->GetMutable<framework::LoDTensor>();
-
-        local_result_tensor.ShareDataWith(outside_tensor);
-
+        auto new_inside_name = cur_scope.Rename(inside_grad_name);
         auto sum_op = framework::OpRegistry::CreateOp(
-            "sum", {{"X", {result_var_name, inside_grad_name}}},
-            {{"Out", {result_var_name}}}, {});
-        sum_op->Run(**cur_scope_iter, dev_ctx);
+            "sum", {{"X", {pg_names[param_id], new_inside_name}}},
+            {{"Out", {pg_names[param_id]}}}, {});
+        sum_op->Run(cur_scope, dev_ctx);
+        cur_scope.Rename(new_inside_name, inside_grad_name);
       }
     }
   }
@@ -169,29 +210,110 @@ class WhileGradOpDescMaker : public framework::SingleGradOpDescMaker {
   virtual std::unique_ptr<framework::OpDescBind> Apply() const {
     auto *grad = new framework::OpDescBind();
     grad->SetType("while_grad");
-    for (auto &input_param : this->InputNames()) {
-      grad->SetInput(input_param, this->Input(input_param));
-      grad->SetOutput(framework::GradVarName(input_param),
-                      this->InputGrad(input_param));
+    grad->SetInput(kParameters, Input(kParameters));
+    grad->SetOutput(
+        framework::GradVarName(kParameters),
+        InputGrad(kParameters, /*do not drop empty gradient*/ false));
+    grad->SetInput(kOutputs, Output(kOutputs));
+
+    // OG should be re-calculated by step blocks, since many outputs of while op
+    // do not need to calculate gradients.
+    std::unordered_set<std::string> block_ins;
+    {
+      for (auto &p : Input(kParameters)) {
+        block_ins.insert(p);
+      }
+      for (auto &o : Output(kOutputs)) {
+        block_ins.insert(o);
+      }
+    }
+    std::unordered_set<std::string> extra_inputs;
+    for (size_t i = 0; i < grad_block_[0]->OpSize(); ++i) {
+      for (auto &input_name : grad_block_[0]->Op(i)->InputArgumentNames()) {
+        if (block_ins.find(input_name) != block_ins.end()) {
+          continue;
+        }
+        extra_inputs.insert(input_name);
       }
 
-    for (auto &output_param : this->OutputNames()) {
-      grad->SetInput(output_param, this->Output(output_param));
-      if (output_param != kStepScopes) {
-        grad->SetInput(framework::GradVarName(output_param),
-                       this->OutputGrad(output_param));
+      for (auto &output_name : grad_block_[0]->Op(i)->OutputArgumentNames()) {
+        block_ins.insert(output_name);
       }
     }
+
+    std::vector<std::string> extra_inputs_list;
+    extra_inputs_list.resize(extra_inputs.size());
+    std::copy(extra_inputs.begin(), extra_inputs.end(),
+              extra_inputs_list.begin());
+    grad->SetInput(framework::GradVarName(kOutputs), extra_inputs_list);
+    grad->SetInput(kStepScopes, Output(kStepScopes));
     grad->SetAttrMap(this->Attrs());
     grad->SetBlockAttr(kStepBlock, *grad_block_[0]);
+    // record the original output gradient names, since the gradient name of
+    // while operator could be renamed.
+    grad->SetAttr("original_output_grad", extra_inputs_list);
 
     return std::unique_ptr<framework::OpDescBind>(grad);
   }
 };
 
+class WhileGradOpVarTypeInference : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDescBind &op_desc,
+                  framework::BlockDescBind *block) const override {
+    auto p_names = op_desc.Input(kParameters);
+    auto pg_names = op_desc.Output(framework::GradVarName(kParameters));
+
+    for (size_t i = 0; i < p_names.size(); ++i) {
+      auto &p_var = detail::Ref(block->FindVarRecursive(p_names[i]));
+      auto *g_var = block->FindVarRecursive(pg_names[i]);
+      if (g_var != nullptr) {  // Gradient could be @EMPTY@
+        VLOG(5) << "Setting " << pg_names[i] << " following " << p_names[i]
+                << " type: " << p_var.GetType();
+        g_var->SetType(p_var.GetType());
+        g_var->SetDataType(p_var.GetDataType());
+      }
+    }
+  }
+};
+
+class WhileGradOpShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *ctx) const override {
+    ctx->HasInputs(kParameters);
+    ctx->HasOutputs(framework::GradVarName(kParameters));
+    ctx->HasInputs(kOutputs);
+    ctx->HasInputs(framework::GradVarName(kOutputs));
+
+    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;
+    for (size_t i = 0; i < p_names.size(); ++i) {
+      if (pg_names[i] == framework::kEmptyVarName) {
+        continue;
+      }
+      if (var_types[i] == framework::VarDesc::LOD_TENSOR) {
+        names_to_set.push_back(pg_names[i]);
+        dims_to_set.push_back(dims[i]);
+      } 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]);
+      }
+    }
+    ctx->SetDims(names_to_set, dims_to_set);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 REGISTER_OPERATOR(while, paddle::operators::WhileOp,
                   paddle::operators::WhileOpMaker,
                   paddle::operators::WhileGradOpDescMaker);
+REGISTER_OPERATOR(while_grad, paddle::operators::WhileGradOp,
+                  paddle::operators::WhileGradOpShapeInference,
+                  paddle::operators::WhileGradOpVarTypeInference);

paddle/scripts/docker/README.md

@@ -57,8 +57,7 @@ Users can specify the following Docker build arguments with either "ON" or "OFF"
 | `WITH_GPU` | OFF | Generates NVIDIA CUDA GPU code and relies on CUDA libraries. |
 | `WITH_AVX` | OFF | Set to "ON" to enable AVX support. |
 | `WITH_TESTING` | ON | Build unit tests binaries. |
-| `WITH_MKLDNN` | ON | Build with [Intel® MKL DNN](https://github.com/01org/mkl-dnn) support. |
-| `WITH_MKLML` | ON | Build with [Intel® MKL](https://software.intel.com/en-us/mkl) support. |
+| `WITH_MKL` | ON | Build with [Intel® MKL](https://software.intel.com/en-us/mkl) and [Intel® MKL-DNN](https://github.com/01org/mkl-dnn) support. |
 | `WITH_GOLANG` | ON | Build fault-tolerant parameter server written in go. |
 | `WITH_SWIG_PY` | ON | Build with SWIG python API support. |
 | `WITH_C_API` | OFF | Build capi libraries for inference. |

paddle/scripts/docker/build.sh

@@ -34,9 +34,7 @@ function cmake_gen() {
         ${PYTHON_FLAGS}
         -DWITH_DOC=OFF
         -DWITH_GPU=${WITH_GPU:-OFF}
-        -DCUDA_ARCH_NAME=All
-        -DWITH_MKLDNN=${WITH_MKLDNN:-ON}
-        -DWITH_MKLML=${WITH_MKLML:-ON}
+        -DWITH_MKL=${WITH_MKL:-ON}
         -DWITH_AVX=${WITH_AVX:-OFF}
         -DWITH_GOLANG=${WITH_GOLANG:-ON}
         -DWITH_SWIG_PY=ON
@@ -57,9 +55,7 @@ EOF
         ${PYTHON_FLAGS} \
         -DWITH_DOC=OFF \
         -DWITH_GPU=${WITH_GPU:-OFF} \
-        -DCUDA_ARCH_NAME=All \
-        -DWITH_MKLDNN=${WITH_MKLDNN:-ON} \
-        -DWITH_MKLML=${WITH_MKLML:-ON} \
+        -DWITH_MKL=${WITH_MKL:-ON} \
         -DWITH_AVX=${WITH_AVX:-OFF} \
         -DWITH_GOLANG=${WITH_GOLANG:-ON} \
         -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} \

paddle/scripts/submit_local.sh.in

@@ -18,8 +18,8 @@ function version(){
         echo "PaddlePaddle @PADDLE_VERSION@, compiled with"
         echo "    with_avx: @WITH_AVX@"
         echo "    with_gpu: @WITH_GPU@"
+        echo "    with_mkl: @WITH_MKL@"
         echo "    with_mkldnn: @WITH_MKLDNN@"
-        echo "    with_mklml: @WITH_MKLML@"
         echo "    with_double: @WITH_DOUBLE@"
         echo "    with_python: @WITH_PYTHON@"
         echo "    with_rdma: @WITH_RDMA@"
@@ -45,8 +45,8 @@ function ver2num() {
 
 function cpu_config() {
   # auto set KMP_AFFINITY and OMP_DYNAMIC from Hyper Threading Status
-  # only when MKLDNN or MKLML enabled
-  if [ "@WITH_MKLDNN@" == "OFF" ] && [ "@WITH_MKLML@" == "OFF"]; then
+  # only when MKL enabled
+  if [ "@WITH_MKL@" == "OFF" ]; then
     return 0
   fi
   ht=`lscpu |grep "per core"|awk -F':' '{print $2}'|xargs`
@@ -70,8 +70,8 @@ function cpu_config() {
 function threads_config() {
   # auto set OMP_NUM_THREADS and MKL_NUM_THREADS
   # according to trainer_count and total processors
-  # only when MKLDNN or MKLML enabled
-  if [ "@WITH_MKLDNN@" == "OFF" ] && [ "@WITH_MKLML@" == "OFF"]; then
+  # only when MKL enabled
+  if [ "@WITH_MKL@" == "OFF" ]; then
     return 0
   fi
   processors=`grep "processor" /proc/cpuinfo|sort -u|wc -l`

paddle/scripts/travis/build_doc.sh

@@ -6,7 +6,7 @@ mkdir -p $TRAVIS_BUILD_DIR/build
 cd $TRAVIS_BUILD_DIR/build
 
 # Compile Documentation only.
-cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKLDNN=OFF -DWITH_MKLML=OFF -DWITH_DOC=ON
+cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON
 make -j `nproc` gen_proto_py
 make -j `nproc` paddle_docs paddle_docs_cn
 

paddle/trainer/Trainer.cpp

@@ -137,6 +137,10 @@ void Trainer::init(const std::shared_ptr<TrainerConfigHelper>& config,
     }
   }
 
+  if (FLAGS_use_mkldnn) {
+    CHECK_EQ(FLAGS_trainer_count, 1UL) << "MKLDNN only need 1 trainer";
+  }
+
   if (testing) {
     LOG(INFO) << "trainer: in testing mode";
     if (config_->getOptConfig().use_sparse_remote_updater() ||

python/paddle/v2/fluid/framework.py

@@ -12,9 +12,9 @@ def unique_name(prefix):
     return "_".join([prefix, str(uid)])
 
 
-def _debug_string_(proto):
+def _debug_string_(proto, throw_on_error=True):
     error_fields = list()
-    if not proto.IsInitialized(error_fields):
+    if not proto.IsInitialized(error_fields) and throw_on_error:
         raise ValueError("{0} are not initialized\nThe message is {1}".format(
             error_fields, proto))
     return proto.__str__()
@@ -101,9 +101,12 @@ class Variable(object):
         self.stop_gradient = stop_gradient
 
     def __str__(self):
+        return self.to_string(True)
+
+    def to_string(self, throw_on_error):
         protostr = self.desc.serialize_to_string()
         proto = framework_pb2.VarDesc.FromString(str(protostr))
-        return _debug_string_(proto)
+        return _debug_string_(proto, throw_on_error)
 
     __repr__ = __str__
 
@@ -291,10 +294,13 @@ class Operator(object):
             self.desc.infer_var_type(self.block.desc)
             self.desc.infer_shape(self.block.desc)
 
-    def __str__(self):
+    def to_string(self, throw_on_error):
         protostr = self.desc.serialize_to_string()
         proto = framework_pb2.OpDesc.FromString(str(protostr))
-        return _debug_string_(proto)
+        return _debug_string_(proto, throw_on_error)
+
+    def __str__(self):
+        return self.to_string(True)
 
     __repr__ = __str__
 
@@ -349,9 +355,12 @@ class Block(object):
         self.program = program
 
     def __str__(self):
+        return self.to_string(True)
+
+    def to_string(self, throw_on_error):
         protostr = self.desc.serialize_to_string()
         proto = framework_pb2.BlockDesc.FromString(str(protostr))
-        return _debug_string_(proto)
+        return _debug_string_(proto, throw_on_error)
 
     __repr__ = __str__
 
@@ -454,9 +463,12 @@ class Program(object):
         self.current_block_idx = 0
 
     def __str__(self):
+        return self.to_string(True)
+
+    def to_string(self, throw_on_error):
         protostr = self.desc.serialize_to_string()
         proto = framework_pb2.ProgramDesc.FromString(str(protostr))
-        return _debug_string_(proto)
+        return _debug_string_(proto, throw_on_error)
 
     def clone(self):
         p = Program()
@@ -512,7 +524,14 @@ class Program(object):
         assert isinstance(target, Variable)
         if no_grad_set is None:
             no_grad_set = set()
-        param_to_grad_info = self.desc.append_backward(target.desc, no_grad_set)
+        try:
+            param_to_grad_info = self.desc.append_backward(target.desc,
+                                                           no_grad_set)
+        except Exception as e:
+            raise core.EnforceNotMet(
+                str(e) + "\nCurrent protobuf is\n{0}".format(
+                    self.to_string(False)))
+
         self.sync_with_cpp()
         return param_to_grad_info
 

python/paddle/v2/fluid/net_drawer.py

@@ -66,10 +66,13 @@ def parse_graph(program, graph, var_dict, **kwargs):
             if not var_dict.has_key(var):
                 var_dict[var] = "Feed"
 
+    temp_id = 0
     proto = framework_pb2.ProgramDesc.FromString(
         program.desc.serialize_to_string())
     for block in proto.blocks:
         for op in block.ops:
+            op.type = op.type + "_" + str(temp_id)
+            temp_id += 1
             graph.node(**draw_node(op))
             for o in op.outputs:
                 for arg in o.arguments:
@@ -78,6 +81,7 @@ def parse_graph(program, graph, var_dict, **kwargs):
                 for arg in e.arguments:
                     if var_dict.has_key(arg):
                         graph.edge(**draw_edge(var_dict, op, e, arg))
+        break  # only plot the first block
 
 
 def draw_graph(startup_program, main_program, **kwargs):

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

+import numpy as np
 import paddle.v2 as paddle
-import paddle.v2.fluid.layers as layers
 import paddle.v2.fluid.core as core
-import paddle.v2.fluid.optimizer as optimizer
 import paddle.v2.fluid.framework as framework
-from paddle.v2.fluid.io import save_persistables, load_persistables
+import paddle.v2.fluid.layers as layers
 from paddle.v2.fluid.executor import Executor
+from paddle.v2.fluid.io import save_persistables, load_persistables
+from paddle.v2.fluid.optimizer import SGDOptimizer
 
-import numpy as np
-
-x = layers.data(
-    name='x',
-    shape=[13],
-    data_type='float32')
+x = layers.data(name='x', shape=[13], data_type='float32')
 
-y_predict = layers.fc(input=x,
-                      size=1,
-                      act=None)
+y_predict = layers.fc(input=x, size=1, act=None)
 
-y = layers.data(
-    name='y',
-    shape=[1],
-    data_type='float32')
+y = layers.data(name='y', shape=[1], data_type='float32')
 
-cost = layers.square_error_cost(
-    input=y_predict,
-    label=y)
+cost = layers.square_error_cost(input=y_predict, label=y)
 avg_cost = layers.mean(x=cost)
 
-sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.001)
+sgd_optimizer = SGDOptimizer(learning_rate=0.001)
 opts = sgd_optimizer.minimize(avg_cost)
 
 BATCH_SIZE = 20

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

 import numpy as np
 import paddle.v2 as paddle
 import paddle.v2.fluid.core as core
+import paddle.v2.fluid.framework as framework
 import paddle.v2.fluid.layers as layers
 import paddle.v2.fluid.nets as nets
-import paddle.v2.fluid.optimizer as optimizer
 from paddle.v2.fluid.executor import Executor
-import paddle.v2.fluid.framework as framework
 from paddle.v2.fluid.initializer import XavierInitializer
+from paddle.v2.fluid.optimizer import AdamOptimizer
 
 
 def resnet_cifar10(input, depth=32):
-    def conv_bn_layer(input,
-                      ch_out,
-                      filter_size,
-                      stride,
-                      padding,
-                      act='relu'):
+    def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'):
         tmp = layers.conv2d(
             input=input,
             filter_size=filter_size,
@@ -24,9 +19,7 @@ def resnet_cifar10(input, depth=32):
             padding=padding,
             act=None,
             bias_attr=False)
-        return layers.batch_norm(
-            input=tmp,
-            act=act)
+        return layers.batch_norm(input=tmp, act=act)
 
     def shortcut(input, ch_in, ch_out, stride, program, init_program):
         if ch_in != ch_out:
@@ -35,28 +28,11 @@ def resnet_cifar10(input, depth=32):
         else:
             return input
 
-    def basicblock(input,
-                   ch_in,
-                   ch_out,
-                   stride):
-        tmp = conv_bn_layer(
-            input,
-            ch_out,
-            3,
-            stride,
-            1)
-        tmp = conv_bn_layer(
-            tmp,
-            ch_out,
-            3,
-            1,
-            1,
-            act=None)
+    def basicblock(input, ch_in, ch_out, stride):
+        tmp = conv_bn_layer(input, ch_out, 3, stride, 1)
+        tmp = conv_bn_layer(tmp, ch_out, 3, 1, 1, act=None)
         short = shortcut(input, ch_in, ch_out, stride)
-        return layers.elementwise_add(
-            x=tmp,
-            y=short,
-            act='relu')
+        return layers.elementwise_add(x=tmp, y=short, act='relu')
 
     def layer_warp(block_func, input, ch_in, ch_out, count, stride):
         tmp = block_func(input, ch_in, ch_out, stride)
@@ -67,45 +43,17 @@ def resnet_cifar10(input, depth=32):
     assert (depth - 2) % 6 == 0
     n = (depth - 2) / 6
     conv1 = conv_bn_layer(
-        input=input,
-        ch_out=16,
-        filter_size=3,
-        stride=1,
-        padding=1)
-    res1 = layer_warp(
-        basicblock,
-        conv1,
-        16,
-        16,
-        n,
-        1)
-    res2 = layer_warp(
-        basicblock,
-        res1,
-        16,
-        32,
-        n,
-        2)
-    res3 = layer_warp(
-        basicblock,
-        res2,
-        32,
-        64,
-        n,
-        2)
+        input=input, ch_out=16, filter_size=3, stride=1, padding=1)
+    res1 = layer_warp(basicblock, conv1, 16, 16, n, 1)
+    res2 = layer_warp(basicblock, res1, 16, 32, n, 2)
+    res3 = layer_warp(basicblock, res2, 32, 64, n, 2)
     pool = layers.pool2d(
-        input=res3,
-        pool_size=8,
-        pool_type='avg',
-        pool_stride=1)
+        input=res3, pool_size=8, pool_type='avg', pool_stride=1)
     return pool
 
 
 def vgg16_bn_drop(input):
-    def conv_block(input,
-                   num_filter,
-                   groups,
-                   dropouts):
+    def conv_block(input, num_filter, groups, dropouts):
         return nets.img_conv_group(
             input=input,
             pool_size=2,
@@ -123,22 +71,14 @@ def vgg16_bn_drop(input):
     conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0])
     conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0])
 
-    drop = layers.dropout(
-        x=conv5,
-        dropout_prob=0.5)
+    drop = layers.dropout(x=conv5, dropout_prob=0.5)
     fc1 = layers.fc(input=drop,
                     size=512,
                     act=None,
                     param_attr={"initializer": XavierInitializer()})
-    reshape1 = layers.reshape(
-        x=fc1,
-        shape=list(fc1.shape + (1, 1)))
-    bn = layers.batch_norm(
-        input=reshape1,
-        act='relu')
-    drop2 = layers.dropout(
-        x=bn,
-        dropout_prob=0.5)
+    reshape1 = layers.reshape(x=fc1, shape=list(fc1.shape + (1, 1)))
+    bn = layers.batch_norm(input=reshape1, act='relu')
+    drop2 = layers.dropout(x=bn, dropout_prob=0.5)
     fc2 = layers.fc(input=drop2,
                     size=512,
                     act=None,
@@ -165,8 +105,8 @@ cost = layers.cross_entropy(input=predict, label=label)
 avg_cost = layers.mean(x=cost)
 accuracy = layers.accuracy(input=predict, label=label)
 
-# optimizer = optimizer.SGDOptimizer(learning_rate=0.001)
-optimizer = optimizer.AdamOptimizer(learning_rate=0.001)
+# optimizer = SGDOptimizer(learning_rate=0.001)
+optimizer = AdamOptimizer(learning_rate=0.001)
 opts = optimizer.minimize(avg_cost)
 
 BATCH_SIZE = 128

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

+import numpy as np
 import paddle.v2 as paddle
-import paddle.v2.fluid.layers as layers
-import paddle.v2.fluid.nets as nets
 import paddle.v2.fluid.core as core
-import paddle.v2.fluid.optimizer as optimizer
 import paddle.v2.fluid.evaluator as evaluator
 import paddle.v2.fluid.framework as framework
+import paddle.v2.fluid.layers as layers
+import paddle.v2.fluid.nets as nets
 from paddle.v2.fluid.executor import Executor
+from paddle.v2.fluid.optimizer import AdamOptimizer
 
-import numpy as np
-
-images = layers.data(
-    name='pixel',
-    shape=[1, 28, 28],
-    data_type='float32')
-label = layers.data(
-    name='label',
-    shape=[1],
-    data_type='int64')
+images = layers.data(name='pixel', shape=[1, 28, 28], data_type='float32')
+label = layers.data(name='label', shape=[1], data_type='int64')
 conv_pool_1 = nets.simple_img_conv_pool(
     input=images,
     filter_size=5,
@@ -32,17 +25,13 @@ conv_pool_2 = nets.simple_img_conv_pool(
     pool_stride=2,
     act="relu")
 
-predict = layers.fc(input=conv_pool_2,
-                    size=10,
-                    act="softmax")
+predict = layers.fc(input=conv_pool_2, size=10, act="softmax")
 cost = layers.cross_entropy(input=predict, label=label)
 avg_cost = layers.mean(x=cost)
-optimizer = optimizer.AdamOptimizer(learning_rate=0.01, beta1=0.9, beta2=0.999)
+optimizer = AdamOptimizer(learning_rate=0.01, beta1=0.9, beta2=0.999)
 opts = optimizer.minimize(avg_cost)
 
-accuracy, acc_out = evaluator.accuracy(
-    input=predict,
-    label=label)
+accuracy, acc_out = evaluator.accuracy(input=predict, label=label)
 
 BATCH_SIZE = 50
 PASS_NUM = 3

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

+import numpy as np
 import paddle.v2 as paddle
-import paddle.v2.fluid.layers as layers
 import paddle.v2.fluid.core as core
-import paddle.v2.fluid.optimizer as optimizer
 import paddle.v2.fluid.framework as framework
+import paddle.v2.fluid.layers as layers
 from paddle.v2.fluid.executor import Executor
-from paddle.v2.fluid.regularizer import L2DecayRegularizer
 from paddle.v2.fluid.initializer import UniformInitializer
-
-import numpy as np
+from paddle.v2.fluid.optimizer import MomentumOptimizer
+from paddle.v2.fluid.regularizer import L2DecayRegularizer
 
 BATCH_SIZE = 128
-image = layers.data(
-    name='x',
-    shape=[784],
-    data_type='float32')
+image = layers.data(name='x', shape=[784], data_type='float32')
 
 param_attr = {
     'name': None,
@@ -22,32 +18,21 @@ param_attr = {
     'regularization': L2DecayRegularizer(0.0005 * BATCH_SIZE)
 }
 
-hidden1 = layers.fc(input=image,
-                    size=128,
-                    act='relu',
-                    param_attr=param_attr)
-hidden2 = layers.fc(input=hidden1,
-                    size=64,
-                    act='relu',
-                    param_attr=param_attr)
+hidden1 = layers.fc(input=image, size=128, act='relu', param_attr=param_attr)
+hidden2 = layers.fc(input=hidden1, size=64, act='relu', param_attr=param_attr)
 
 predict = layers.fc(input=hidden2,
                     size=10,
                     act='softmax',
                     param_attr=param_attr)
 
-label = layers.data(
-    name='y',
-    shape=[1],
-    data_type='int64')
+label = layers.data(name='y', shape=[1], data_type='int64')
 
 cost = layers.cross_entropy(input=predict, label=label)
 avg_cost = layers.mean(x=cost)
-accuracy = layers.accuracy(
-    input=predict,
-    label=label)
+accuracy = layers.accuracy(input=predict, label=label)
 
-optimizer = optimizer.MomentumOptimizer(learning_rate=0.001, momentum=0.9)
+optimizer = MomentumOptimizer(learning_rate=0.001, momentum=0.9)
 opts = optimizer.minimize(avg_cost)
 
 train_reader = paddle.batch(

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

+import numpy as np
 import paddle.v2 as paddle
-import paddle.v2.fluid.layers as layers
-import paddle.v2.fluid.nets as nets
 import paddle.v2.fluid.core as core
-import paddle.v2.fluid.optimizer as optimizer
 import paddle.v2.fluid.framework as framework
+import paddle.v2.fluid.layers as layers
+import paddle.v2.fluid.nets as nets
 from paddle.v2.fluid.executor import Executor
-
-import numpy as np
+from paddle.v2.fluid.optimizer import SGDOptimizer
 
 IS_SPARSE = True
 USE_GPU = False
@@ -19,10 +18,7 @@ def get_usr_combined_features():
 
     USR_DICT_SIZE = paddle.dataset.movielens.max_user_id() + 1
 
-    uid = layers.data(
-        name='user_id',
-        shape=[1],
-        data_type='int64')
+    uid = layers.data(name='user_id', shape=[1], data_type='int64')
 
     usr_emb = layers.embedding(
         input=uid,
@@ -31,15 +27,11 @@ def get_usr_combined_features():
         param_attr={'name': 'user_table'},
         is_sparse=IS_SPARSE)
 
-    usr_fc = layers.fc(input=usr_emb,
-                       size=32)
+    usr_fc = layers.fc(input=usr_emb, size=32)
 
     USR_GENDER_DICT_SIZE = 2
 
-    usr_gender_id = layers.data(
-        name='gender_id',
-        shape=[1],
-        data_type='int64')
+    usr_gender_id = layers.data(name='gender_id', shape=[1], data_type='int64')
 
     usr_gender_emb = layers.embedding(
         input=usr_gender_id,
@@ -47,14 +39,10 @@ def get_usr_combined_features():
         param_attr={'name': 'gender_table'},
         is_sparse=IS_SPARSE)
 
-    usr_gender_fc = layers.fc(input=usr_gender_emb,
-                              size=16)
+    usr_gender_fc = layers.fc(input=usr_gender_emb, size=16)
 
     USR_AGE_DICT_SIZE = len(paddle.dataset.movielens.age_table)
-    usr_age_id = layers.data(
-        name='age_id',
-        shape=[1],
-        data_type="int64")
+    usr_age_id = layers.data(name='age_id', shape=[1], data_type="int64")
 
     usr_age_emb = layers.embedding(
         input=usr_age_id,
@@ -62,14 +50,10 @@ def get_usr_combined_features():
         is_sparse=IS_SPARSE,
         param_attr={'name': 'age_table'})
 
-    usr_age_fc = layers.fc(input=usr_age_emb,
-                           size=16)
+    usr_age_fc = layers.fc(input=usr_age_emb, size=16)
 
     USR_JOB_DICT_SIZE = paddle.dataset.movielens.max_job_id() + 1
-    usr_job_id = layers.data(
-        name='job_id',
-        shape=[1],
-        data_type="int64")
+    usr_job_id = layers.data(name='job_id', shape=[1], data_type="int64")
 
     usr_job_emb = layers.embedding(
         input=usr_job_id,
@@ -77,16 +61,12 @@ def get_usr_combined_features():
         param_attr={'name': 'job_table'},
         is_sparse=IS_SPARSE)
 
-    usr_job_fc = layers.fc(input=usr_job_emb,
-                           size=16)
+    usr_job_fc = layers.fc(input=usr_job_emb, size=16)
 
     concat_embed = layers.concat(
-        input=[usr_fc, usr_gender_fc, usr_age_fc, usr_job_fc],
-        axis=1)
+        input=[usr_fc, usr_gender_fc, usr_age_fc, usr_job_fc], axis=1)
 
-    usr_combined_features = layers.fc(input=concat_embed,
-                                      size=200,
-                                      act="tanh")
+    usr_combined_features = layers.fc(input=concat_embed, size=200, act="tanh")
 
     return usr_combined_features
 
@@ -95,10 +75,7 @@ def get_mov_combined_features():
 
     MOV_DICT_SIZE = paddle.dataset.movielens.max_movie_id() + 1
 
-    mov_id = layers.data(
-        name='movie_id',
-        shape=[1],
-        data_type='int64')
+    mov_id = layers.data(name='movie_id', shape=[1], data_type='int64')
 
     mov_emb = layers.embedding(
         input=mov_id,
@@ -107,36 +84,24 @@ def get_mov_combined_features():
         param_attr={'name': 'movie_table'},
         is_sparse=IS_SPARSE)
 
-    mov_fc = layers.fc(input=mov_emb,
-                       size=32)
+    mov_fc = layers.fc(input=mov_emb, size=32)
 
     CATEGORY_DICT_SIZE = len(paddle.dataset.movielens.movie_categories())
 
-    category_id = layers.data(
-        name='category_id',
-        shape=[1],
-        data_type='int64')
+    category_id = layers.data(name='category_id', shape=[1], data_type='int64')
 
     mov_categories_emb = layers.embedding(
-        input=category_id,
-        size=[CATEGORY_DICT_SIZE, 32],
-        is_sparse=IS_SPARSE)
+        input=category_id, size=[CATEGORY_DICT_SIZE, 32], is_sparse=IS_SPARSE)
 
     mov_categories_hidden = layers.sequence_pool(
-        input=mov_categories_emb,
-        pool_type="sum")
+        input=mov_categories_emb, pool_type="sum")
 
     MOV_TITLE_DICT_SIZE = len(paddle.dataset.movielens.get_movie_title_dict())
 
-    mov_title_id = layers.data(
-        name='movie_title',
-        shape=[1],
-        data_type='int64')
+    mov_title_id = layers.data(name='movie_title', shape=[1], data_type='int64')
 
     mov_title_emb = layers.embedding(
-        input=mov_title_id,
-        size=[MOV_TITLE_DICT_SIZE, 32],
-        is_sparse=IS_SPARSE)
+        input=mov_title_id, size=[MOV_TITLE_DICT_SIZE, 32], is_sparse=IS_SPARSE)
 
     mov_title_conv = nets.sequence_conv_pool(
         input=mov_title_emb,
@@ -146,13 +111,10 @@ def get_mov_combined_features():
         pool_type="sum")
 
     concat_embed = layers.concat(
-        input=[mov_fc, mov_categories_hidden, mov_title_conv],
-        axis=1)
+        input=[mov_fc, mov_categories_hidden, mov_title_conv], axis=1)
 
     # FIXME(dzh) : need tanh operator
-    mov_combined_features = layers.fc(input=concat_embed,
-                                      size=200,
-                                      act="tanh")
+    mov_combined_features = layers.fc(input=concat_embed, size=200, act="tanh")
 
     return mov_combined_features
 
@@ -162,18 +124,11 @@ def model():
     mov_combined_features = get_mov_combined_features()
 
     # need cos sim
-    inference = layers.cos_sim(
-        X=usr_combined_features,
-        Y=mov_combined_features)
+    inference = layers.cos_sim(X=usr_combined_features, Y=mov_combined_features)
 
-    label = layers.data(
-        name='score',
-        shape=[1],
-        data_type='float32')
+    label = layers.data(name='score', shape=[1], data_type='float32')
 
-    square_cost = layers.square_error_cost(
-        input=inference,
-        label=label)
+    square_cost = layers.square_error_cost(input=inference, label=label)
 
     avg_cost = layers.mean(x=square_cost)
 
@@ -182,7 +137,7 @@ def model():
 
 def main():
     cost = model()
-    sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.2)
+    sgd_optimizer = SGDOptimizer(learning_rate=0.2)
     opts = sgd_optimizer.minimize(cost)
 
     if USE_GPU:

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

+import numpy as np
 import paddle.v2 as paddle
-import paddle.v2.fluid.layers as layers
-import paddle.v2.fluid.nets as nets
 import paddle.v2.fluid.core as core
-import paddle.v2.fluid.optimizer as optimizer
 import paddle.v2.fluid.framework as framework
+import paddle.v2.fluid.layers as layers
+import paddle.v2.fluid.nets as nets
 from paddle.v2.fluid.executor import Executor
-
-import numpy as np
+from paddle.v2.fluid.optimizer import AdamOptimizer
 
 
 def convolution_net(input_dim, class_dim=2, emb_dim=32, hid_dim=32):
@@ -31,7 +30,7 @@ def convolution_net(input_dim, class_dim=2, emb_dim=32, hid_dim=32):
                            act="softmax")
     cost = layers.cross_entropy(input=prediction, label=label)
     avg_cost = layers.mean(x=cost)
-    adam_optimizer = optimizer.AdamOptimizer(learning_rate=0.002)
+    adam_optimizer = AdamOptimizer(learning_rate=0.002)
     opts = adam_optimizer.minimize(avg_cost)
     acc = layers.accuracy(input=prediction, label=label)
     return avg_cost, acc

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

+import numpy as np
 import paddle.v2 as paddle
-import paddle.v2.fluid.layers as layers
-import paddle.v2.fluid.nets as nets
 import paddle.v2.fluid.core as core
-import paddle.v2.fluid.optimizer as optimizer
 import paddle.v2.fluid.framework as framework
+import paddle.v2.fluid.layers as layers
 from paddle.v2.fluid.executor import Executor
-
-import numpy as np
+from paddle.v2.fluid.optimizer import AdamOptimizer
 
 
 def stacked_lstm_net(input_dim,
@@ -41,7 +39,7 @@ def stacked_lstm_net(input_dim,
                            act='softmax')
     cost = layers.cross_entropy(input=prediction, label=label)
     avg_cost = layers.mean(x=cost)
-    adam_optimizer = optimizer.AdamOptimizer(learning_rate=0.002)
+    adam_optimizer = AdamOptimizer(learning_rate=0.002)
     opts = adam_optimizer.minimize(avg_cost)
     acc = layers.accuracy(input=prediction, label=label)
     return avg_cost, acc

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

+import numpy as np
 import paddle.v2 as paddle
-import paddle.v2.fluid.layers as layers
 import paddle.v2.fluid.core as core
-import paddle.v2.fluid.optimizer as optimizer
 import paddle.v2.fluid.framework as framework
+import paddle.v2.fluid.layers as layers
 from paddle.v2.fluid.executor import Executor
-
-import numpy as np
+from paddle.v2.fluid.optimizer import AdamOptimizer
 
 
 def lstm_net(dict_dim, class_dim=2, emb_dim=32, seq_len=80, batch_size=50):
@@ -33,7 +32,7 @@ def lstm_net(dict_dim, class_dim=2, emb_dim=32, seq_len=80, batch_size=50):
     cost = layers.cross_entropy(input=prediction, label=label)
 
     avg_cost = layers.mean(x=cost)
-    adam_optimizer = optimizer.AdamOptimizer(learning_rate=0.002)
+    adam_optimizer = AdamOptimizer(learning_rate=0.002)
     opts = adam_optimizer.minimize(avg_cost)
     acc = layers.accuracy(input=prediction, label=label)
 

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

+import numpy as np
 import paddle.v2 as paddle
-import paddle.v2.fluid.layers as layers
 import paddle.v2.fluid.core as core
-import paddle.v2.fluid.optimizer as optimizer
 import paddle.v2.fluid.framework as framework
+import paddle.v2.fluid.layers as layers
 from paddle.v2.fluid.executor import Executor
-
-import numpy as np
+from paddle.v2.fluid.optimizer import SGDOptimizer
 
 PASS_NUM = 100
 EMBED_SIZE = 32
@@ -17,26 +16,11 @@ IS_SPARSE = True
 word_dict = paddle.dataset.imikolov.build_dict()
 dict_size = len(word_dict)
 
-first_word = layers.data(
-    name='firstw',
-    shape=[1],
-    data_type='int64')
-second_word = layers.data(
-    name='secondw',
-    shape=[1],
-    data_type='int64')
-third_word = layers.data(
-    name='thirdw',
-    shape=[1],
-    data_type='int64')
-forth_word = layers.data(
-    name='forthw',
-    shape=[1],
-    data_type='int64')
-next_word = layers.data(
-    name='nextw',
-    shape=[1],
-    data_type='int64')
+first_word = layers.data(name='firstw', shape=[1], data_type='int64')
+second_word = layers.data(name='secondw', shape=[1], data_type='int64')
+third_word = layers.data(name='thirdw', shape=[1], data_type='int64')
+forth_word = layers.data(name='forthw', shape=[1], data_type='int64')
+next_word = layers.data(name='nextw', shape=[1], data_type='int64')
 
 embed_first = layers.embedding(
     input=first_word,
@@ -64,19 +48,12 @@ embed_forth = layers.embedding(
     param_attr={'name': 'shared_w'})
 
 concat_embed = layers.concat(
-    input=[embed_first, embed_second, embed_third, embed_forth],
-    axis=1)
-hidden1 = layers.fc(input=concat_embed,
-                    size=HIDDEN_SIZE,
-                    act='sigmoid')
-predict_word = layers.fc(input=hidden1,
-                         size=dict_size,
-                         act='softmax')
-cost = layers.cross_entropy(
-    input=predict_word,
-    label=next_word)
+    input=[embed_first, embed_second, embed_third, embed_forth], axis=1)
+hidden1 = layers.fc(input=concat_embed, size=HIDDEN_SIZE, act='sigmoid')
+predict_word = layers.fc(input=hidden1, size=dict_size, act='softmax')
+cost = layers.cross_entropy(input=predict_word, label=next_word)
 avg_cost = layers.mean(x=cost)
-sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.001)
+sgd_optimizer = SGDOptimizer(learning_rate=0.001)
 opts = sgd_optimizer.minimize(avg_cost)
 
 train_reader = paddle.batch(

python/paddle/v2/fluid/tests/test_conv2d_op.py

@@ -110,13 +110,30 @@ class TestConv2dOp(OpTest):
         self.op_type = "conv2d"
 
 
+class TestWithPad(TestConv2dOp):
+    def init_test_case(self):
+        self.pad = [1, 1]
+        self.stride = [1, 1]
+        self.input_size = [2, 3, 5, 5]  # NCHW
+        assert np.mod(self.input_size[1], self.groups) == 0
+        f_c = self.input_size[1] / self.groups
+        self.filter_size = [6, f_c, 3, 3]
+
+
+class TestWithStride(TestConv2dOp):
+    def init_test_case(self):
+        self.pad = [1, 1]
+        self.stride = [2, 2]
+        self.input_size = [2, 3, 6, 6]  # NCHW
+        assert np.mod(self.input_size[1], self.groups) == 0
+        f_c = self.input_size[1] / self.groups
+        self.filter_size = [6, f_c, 3, 3]
+
+
 class TestWithGroup(TestConv2dOp):
     def init_group(self):
         self.groups = 3
 
-    def init_op_type(self):
-        self.op_type = "conv2d"
-
 
 class TestWith1x1(TestConv2dOp):
     def init_test_case(self):
@@ -127,15 +144,9 @@ class TestWith1x1(TestConv2dOp):
         f_c = self.input_size[1] / self.groups
         self.filter_size = [6, f_c, 1, 1]
 
-    def init_dilation(self):
-        self.dilations = [1, 1]
-
     def init_group(self):
         self.groups = 3
 
-    def init_op_type(self):
-        self.op_type = "conv2d"
-
 
 class TestWithDilation(TestConv2dOp):
     def init_test_case(self):
@@ -152,14 +163,19 @@ class TestWithDilation(TestConv2dOp):
     def init_group(self):
         self.groups = 3
 
+
+#----------------Conv2dCudnn----------------
+class TestCudnn(TestConv2dOp):
     def init_op_type(self):
-        self.op_type = "conv2d"
+        self.op_type = "conv_cudnn"
 
 
-#----------------Conv2dCudnn----------------
+class TestCudnnWithPad(TestWithPad):
+    def init_op_type(self):
+        self.op_type = "conv_cudnn"
 
 
-class TestCudnn(TestConv2dOp):
+class TestCudnnWithStride(TestWithStride):
     def init_op_type(self):
         self.op_type = "conv_cudnn"
 

python/paddle/v2/fluid/tests/test_conv2d_transpose_op.py

@@ -4,9 +4,7 @@ from op_test import OpTest
 
 
 def conv2dtranspose_forward_naive(input_, filter_, conv2dtranspose_param):
-    # [2, 3, 5, 5]
     in_n, in_c, in_h, in_w = input_.shape
-    # [3, 6, 3, 3]
     f_c, out_c, f_h, f_w = filter_.shape
     assert in_c == f_c
 
@@ -29,6 +27,7 @@ def conv2dtranspose_forward_naive(input_, filter_, conv2dtranspose_param):
                     j1, j2 = j * stride[0], j * stride[0] + f_w
                     out[n, k, i1:i2, j1:j2] += tmp_out
 
+    out = out[:, :, pad[0]:out_h - pad[0], pad[1]:out_w - pad[1]]
     return out
 
 
@@ -36,8 +35,6 @@ class TestConv2dTransposeOp(OpTest):
     def setUp(self):
         # init as conv transpose
         self.init_op_type()
-
-        # [2, 3, 5, 5] -> kernel [3, 6, 3, 3] -> output [2, 6, 7, 7]
         self.init_test_case()
 
         conv2dtranspose_param = {'stride': self.stride, 'pad': self.pad}
@@ -55,7 +52,6 @@ class TestConv2dTransposeOp(OpTest):
         self.outputs = {'Output': output}
 
     def test_check_output(self):
-        print 'check output here for', self.op_type
         self.check_output()
 
     def test_check_grad_no_input(self):
@@ -88,6 +84,26 @@ class TestConv2dTransposeOp(OpTest):
         self.op_type = "conv2d_transpose"
 
 
+class TestWithPad(TestConv2dTransposeOp):
+    def init_test_case(self):
+        self.pad = [1, 1]
+        self.stride = [1, 1]
+        self.dilations = [1, 1]
+        self.input_size = [2, 3, 5, 5]  # NCHW
+        f_c = self.input_size[1]
+        self.filter_size = [f_c, 6, 3, 3]
+
+
+class TestWithStride(TestConv2dTransposeOp):
+    def init_test_case(self):
+        self.pad = [1, 1]
+        self.stride = [2, 2]
+        self.dilations = [1, 1]
+        self.input_size = [2, 3, 5, 5]  # NCHW
+        f_c = self.input_size[1]
+        self.filter_size = [f_c, 6, 3, 3]
+
+
 # ------------ test_cudnn ------------
 class TestCudnn(TestConv2dTransposeOp):
     def init_op_type(self):

python/paddle/v2/fluid/tests/test_conv3d_transpose_op.py

@@ -4,9 +4,7 @@ from op_test import OpTest
 
 
 def conv3dtranspose_forward_naive(input_, filter_, conv3dtranspose_param):
-    # [2, 3, 5, 5, 5]
     in_n, in_c, in_d, in_h, in_w = input_.shape
-    # [3, 6, 3, 3, 3]
     f_c, out_c, f_d, f_h, f_w = filter_.shape
     assert in_c == f_c
 
@@ -14,7 +12,6 @@ def conv3dtranspose_forward_naive(input_, filter_, conv3dtranspose_param):
     out_d = (in_d - 1) * stride[0] + f_d
     out_h = (in_h - 1) * stride[1] + f_h
     out_w = (in_w - 1) * stride[2] + f_w
-
     out = np.zeros((in_n, out_c, out_d, out_h, out_w))
 
     for n in range(in_n):
@@ -33,6 +30,8 @@ def conv3dtranspose_forward_naive(input_, filter_, conv3dtranspose_param):
                         j1, j2 = j * stride[2], j * stride[2] + f_w
                         out[n, k, d1:d2, i1:i2, j1:j2] += tmp_out
 
+    out = out[:, :, pad[0]:out_d - pad[0], pad[1]:out_h - pad[1], pad[2]:out_w -
+              pad[2]]
     return out
 
 
@@ -40,8 +39,6 @@ class TestConv3dTransposeOp(OpTest):
     def setUp(self):
         # init as conv transpose
         self.init_op_type()
-
-        # [2, 3, 5, 5, 5] -> kernel [3, 6, 3, 3, 3] -> output [2, 6, 7, 7, 7]
         self.init_test_case()
 
         conv3dtranspose_param = {'stride': self.stride, 'pad': self.pad}
@@ -49,7 +46,6 @@ class TestConv3dTransposeOp(OpTest):
         filter_ = np.random.random(self.filter_size).astype("float32")
         output = conv3dtranspose_forward_naive(
             input_, filter_, conv3dtranspose_param).astype("float32")
-        # print 'deconv output py', output, output.shape
 
         self.inputs = {'Input': input_, 'Filter': filter_}
         self.attrs = {
@@ -60,7 +56,6 @@ class TestConv3dTransposeOp(OpTest):
         self.outputs = {'Output': output}
 
     def test_check_output(self):
-        print 'check output here'
         self.check_output()
 
     def test_check_grad(self):
@@ -85,7 +80,7 @@ class TestConv3dTransposeOp(OpTest):
         self.pad = [0, 0, 0]
         self.stride = [1, 1, 1]
         self.dilations = [1, 1, 1]
-        self.input_size = [2, 3, 5, 5, 5]  # NCHW
+        self.input_size = [2, 3, 5, 5, 5]  # NCDHW
         f_c = self.input_size[1]
         self.filter_size = [f_c, 6, 3, 3, 3]
 
@@ -93,5 +88,25 @@ class TestConv3dTransposeOp(OpTest):
         self.op_type = "conv3d_transpose"
 
 
+class TestWithPad(TestConv3dTransposeOp):
+    def init_test_case(self):
+        self.pad = [1, 1, 1]
+        self.stride = [1, 1, 1]
+        self.dilations = [1, 1, 1]
+        self.input_size = [2, 3, 5, 5, 5]  # NCDHW
+        f_c = self.input_size[1]
+        self.filter_size = [f_c, 6, 3, 3, 3]
+
+
+class TestWithStride(TestConv3dTransposeOp):
+    def init_test_case(self):
+        self.pad = [1, 1, 1]
+        self.stride = [2, 2, 2]
+        self.dilations = [1, 1, 1]
+        self.input_size = [2, 3, 5, 5, 5]  # NCDHW
+        f_c = self.input_size[1]
+        self.filter_size = [f_c, 6, 3, 3, 3]
+
+
 if __name__ == '__main__':
     unittest.main()

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

+import unittest
+import numpy as np
+from paddle.v2.fluid.op import Operator
+import paddle.v2.fluid.core as core
+
+
+def create_tensor(scope, name, np_data):
+    tensor = scope.var(name).get_tensor()
+    tensor.set_dims(np_data.shape)
+    tensor.set(np_data, core.CPUPlace())
+    return tensor
+
+
+class TestIsEmptyOp(unittest.TestCase):
+    def setUp(self):
+        self.scope = core.Scope()
+        # create input variables
+        np_data0 = np.array([0, 1, 2])
+        create_tensor(self.scope, "X0", np_data0)
+
+        np_data1 = np.array([1])
+        t = create_tensor(self.scope, "X1", np_data1)
+        t.set_dims([0])
+
+        # create output variables
+        self.scope.var("out")
+
+    def test_no_empty(self):
+        self.one_case("X0", False)
+
+    def test_empty(self):
+        self.one_case("X1", True)
+
+    def one_case(self, input, target):
+        op = Operator(type="is_empty", X=input, Out="out")
+        ctx = core.DeviceContext.create(core.CPUPlace())
+        op.run(self.scope, ctx)
+        out = self.scope.var("out").get_tensor()
+        self.assertEqual(np.array(out)[0], target)
+
+
+if __name__ == "__main__":
+    unittest.main()

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

@@ -2,6 +2,7 @@ import unittest
 import paddle.v2.fluid.layers as layers
 from paddle.v2.fluid.executor import Executor
 import paddle.v2.fluid.core as core
+from paddle.v2.fluid.backward import append_backward_ops
 import numpy
 
 
@@ -16,7 +17,7 @@ class TestWhileOp(unittest.TestCase):
         i = layers.zeros(shape=[1], dtype='int64')
         i.stop_gradient = True
         init = layers.zeros(shape=[10], dtype='float32')
-        mem_array = layers.array_write(init, i=i)
+        mem_array = layers.array_write(x=init, i=i)
         data_array = layers.array_write(x=d0, i=i)
 
         i = layers.increment(i)
@@ -29,17 +30,23 @@ class TestWhileOp(unittest.TestCase):
         i.stop_gradient = True
 
         array_len = layers.fill_constant(shape=[1], dtype='int64', value=3)
+        array_len.stop_gradient = True
         cond = layers.less_than(x=i, y=array_len)
 
         while_op = layers.While(cond=cond)
         with while_op.block():
             d = layers.array_read(array=data_array, i=i)
             prev = layers.array_read(array=mem_array, i=i)
-            i = layers.increment(x=i, in_place=True)
             result = layers.sums(input=[d, prev])
+
+            i = layers.increment(x=i, in_place=True)
             layers.array_write(result, i=i, array=mem_array)
             layers.less_than(x=i, y=array_len, cond=cond)
-        sum_result = layers.array_read(mem_array, i=array_len)
+
+        sum_result = layers.array_read(array=mem_array, i=i)
+        loss = layers.mean(x=sum_result)
+
+        append_backward_ops(loss)
 
         cpu = core.CPUPlace()
         exe = Executor(cpu)