Merge branch 'develop' of upstream into profiling_tool

cmake/external/mkldnn.cmake

@@ -63,9 +63,17 @@ ExternalProject_Add(
                         -DMKLROOT:PATH=${MKLML_ROOT}
 )
 
-ADD_LIBRARY(mkldnn SHARED IMPORTED GLOBAL)
-SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
-ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
+ADD_LIBRARY(shared_mkldnn SHARED IMPORTED GLOBAL)
+SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
+ADD_DEPENDENCIES(shared_mkldnn ${MKLDNN_PROJECT})
 MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
 add_definitions(-DPADDLE_WITH_MKLDNN)
-LIST(APPEND external_project_dependencies mkldnn)
+LIST(APPEND external_project_dependencies shared_mkldnn)
+
+# generate a static dummy target to track mkldnn dependencies
+# for cc_library(xxx SRCS xxx.c DEPS mkldnn)
+SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/mkldnn_dummy.c)
+FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
+ADD_LIBRARY(mkldnn STATIC ${dummyfile})
+TARGET_LINK_LIBRARIES(mkldnn ${MKLDNN_LIB} ${MKLML_LIB} ${MKLML_IOMP_LIB})
+ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})

doc/design/releasing_process.md

@@ -7,11 +7,9 @@ PaddlePaddle每次发新的版本，遵循以下流程:
 1. 从`develop`分支派生出新的分支，分支名为`release/版本号`。例如，`release/0.10.0`
 1. 将新分支的版本打上tag，tag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`，第二个为`0.10.0rc2`，依次类推。
 1. 对这个版本的提交，做如下几个操作:
+  * 使用Regression Test List作为检查列表，测试本次release的正确性。
+	  * 如果失败，记录下所有失败的例子，在这个`release/版本号`分支中，修复所有bug后，Patch号加一，到第二步
 	* 修改`python/setup.py.in`中的版本信息,并将`istaged`字段设为`True`。
-	* 编译这个版本的Docker发行镜像，发布到dockerhub。如果失败，修复Docker编译镜像问题，Patch号加一，返回第二步
-	* 编译这个版本的Ubuntu Deb包。如果失败，修复Ubuntu Deb包编译问题，Patch号加一，返回第二步。
-	* 使用Regression Test List作为检查列表，测试Docker镜像/ubuntu安装包的功能正确性
-		* 如果失败，记录下所有失败的例子，在这个`release/版本号`分支中，修复所有bug后，Patch号加一，返回第二步
 	* 编译这个版本的python wheel包，并发布到pypi。
 		* 由于pypi.python.org目前遵循[严格的命名规范PEP 513](https://www.python.org/dev/peps/pep-0513)，在使用twine上传之前，需要重命名wheel包中platform相关的后缀，比如将`linux_x86_64`修改成`manylinux1_x86_64`。
 		* pypi上的package名称为paddlepaddle和paddlepaddle_gpu，如果要上传GPU版本的包，需要修改build/python/setup.py中，name: "paddlepaddle_gpu"并重新打包wheel包：`python setup.py bdist_wheel`。
@@ -21,8 +19,8 @@ PaddlePaddle每次发新的版本，遵循以下流程:
 			pip install twine
 			twine upload dist/[package to upload]
 			```
+		* 编译这个版本的Docker发行镜像，发布到dockerhub。如果失败，修复Docker编译镜像问题，Patch号加一，返回第二步
 1. 第三步完成后，将`release/版本号`分支合入master分支，并删除`release/版本号`分支。将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。最后删除`release/版本号`分支。
-1. 编译master分支的Docker发行镜像，发布到dockerhub。编译ubuntu的deb包，发布到github release页面
 1. 协同完成Release Note的书写
 
 
@@ -31,6 +29,30 @@ PaddlePaddle每次发新的版本，遵循以下流程:
 * `release/版本号`分支一旦建立，一般不允许再从`develop`分支合入`release/版本号`。这样保证`release/版本号`分支功能的封闭，方便测试人员测试PaddlePaddle的行为。
 * 在`release/版本号`分支存在的时候，如果有bugfix的行为，需要将bugfix的分支同时merge到`master`, `develop`和`release/版本号`这三个分支。
 
+## 发布wheel包到pypi
+
+使用[PaddlePaddle CI](https://paddleci.ngrok.io/project.html?projectId=Manylinux1&tab=projectOverview)
+完成自动化二进制编译，参考下图，选择需要发布的版本（通常包含一个CPU版本和一个GPU版本），点击"run"右侧的"..."按钮，可以
+弹出下面的选择框，在第二个tab (Changes)里选择需要发布的分支，这里选择0.11.0，然后点击"Run Build"按钮。等待编译完成后
+可以在此页面的"Artifacts"下拉框中找到生成的3个二进制文件，分别对应CAPI，`cp27m`和`cp27mu`的版本。然后按照上述的方法
+使用`twine`工具上传即可。
+
+<img src="ci_build_whl.png">
+
+* 注：CI环境使用 https://github.com/PaddlePaddle/buildtools 这里的DockerImage作为编译环境以支持更多的Linux
+  发型版，如果需要手动编译，也可以使用这些镜像。这些镜像也可以从 https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/ 下载得到。
+* pypi不支持覆盖上传，所以一个版本号的wheel包发布之后，不可以更改。下一个wheel包需要更新版本号才可以上传。
+
+## 发布Docker镜像
+
+上述PaddlePaddle CI编译wheel完成后会自动将Docker镜像push到DockerHub，所以，发布Docker镜像只需要对自动push的镜像打上
+版本号对应的tag即可：
+
+1. 进入 https://hub.docker.com/r/paddlepaddle/paddle/tags/ 查看latest tag的更新时间是否在上述编译wheel包完成后是否最新。
+1. 执行 `docker pull paddlepaddle/paddle:[latest tag]`，latest tag可以是latest或latest-gpu等。
+1. 执行 `docker tag paddlepaddle/paddle:[latest tag] paddlepaddle/paddle:[version]`
+1. 执行 `docker push paddlepaddle/paddle:[version]`
+
 ## PaddlePaddle 分支规范
 
 PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，并适应github的特性做了一些区别。

paddle/framework/CMakeLists.txt

@@ -41,7 +41,7 @@ device_context)
 cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute)
 cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
-cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute)
+cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
     shape_inference data_transform)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry init)

paddle/framework/executor.cc

@@ -111,7 +111,7 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
 
   for (auto& op_desc : block.AllOps()) {
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
-    VLOG(3) << op->DebugString();
+    VLOG(3) << op->DebugStringEx(local_scope);
     op->Run(*local_scope, place_);
     if (FLAGS_check_nan_inf) {
       for (auto& vname : op->OutputVars(true)) {

paddle/framework/operator.cc

@@ -73,6 +73,17 @@ void UseALL() {
   UseCUDNN();
 }
 
+static DDim GetDims(const Scope& scope, const std::string& name) {
+  Variable* var = scope.FindVar(name);
+  if (var->IsType<LoDTensor>()) {
+    return var->Get<LoDTensor>().dims();
+  } else if (var->IsType<SelectedRows>()) {
+    return var->Get<SelectedRows>().GetCompleteDims();
+  } else {
+    return DDim({-1});
+  }
+}
+
 std::string OperatorBase::Input(const std::string& name) const {
   auto& ins = Inputs(name);
   PADDLE_ENFORCE_LE(ins.size(), 1UL,
@@ -105,7 +116,7 @@ const std::vector<std::string>& OperatorBase::Outputs(
   return it->second;
 }
 
-std::string OperatorBase::DebugString() const {
+std::string OperatorBase::DebugStringEx(const Scope* scope) const {
   std::stringstream ss;
   ss << "Op(" << type_ << "), inputs:{";
   for (auto it = inputs_.begin(); it != inputs_.end();) {
@@ -113,6 +124,9 @@ std::string OperatorBase::DebugString() const {
     ss << input.first << "[";
     for (size_t i = 0; i < input.second.size(); ++i) {
       ss << input.second[i];
+      if (scope) {
+        ss << "(" << GetDims(*scope, input.second[i]) << ")";
+      }
       if (i != input.second.size() - 1) {
         ss << ", ";
       }
@@ -129,6 +143,9 @@ std::string OperatorBase::DebugString() const {
     ss << output.first << "[";
     for (size_t i = 0; i < output.second.size(); ++i) {
       ss << output.second[i];
+      if (scope) {
+        ss << "(" << GetDims(*scope, output.second[i]) << ")";
+      }
       if (i != output.second.size() - 1) {
         ss << ", ";
       }

paddle/framework/operator.h

@@ -108,7 +108,10 @@ class OperatorBase {
     return boost::get<T>(attrs_.at(name));
   }
 
-  virtual std::string DebugString() const;
+  /// if scope is not null, also show dimensions of arguments
+  virtual std::string DebugStringEx(const Scope* scope) const;
+
+  std::string DebugString() const { return DebugStringEx(nullptr); }
 
   /// Net will call this function to Run an op.
   virtual void Run(const Scope& scope, const platform::Place& place) const = 0;

paddle/inference/inference.cc

@@ -38,23 +38,16 @@ void InferenceEngine::LoadInferenceModel(
   LOG(INFO) << "program_desc_str's size: " << program_desc_str.size();
 // PicklingTools cannot parse the vector of strings correctly.
 #else
-  // program_desc_str
-  // the inference.model is stored by following python codes:
-  //  inference_program = fluid.io.get_inference_program(predict)
-  //  model_filename = "recognize_digits_mlp.inference.model/inference.model"
-  //  with open(model_filename, "w") as f:
-  //      program_str = inference_program.desc.serialize_to_string()
-  //          f.write(struct.pack('q', len(program_str)))
-  //          f.write(program_str)
-  std::string model_filename = dirname + "/inference.model";
+  std::string model_filename = dirname + "/__model__.dat";
   LOG(INFO) << "loading model from " << model_filename;
-  std::ifstream fs(model_filename, std::ios_base::binary);
-  int64_t size = 0;
-  fs.read(reinterpret_cast<char*>(&size), sizeof(int64_t));
-  LOG(INFO) << "program_desc_str's size: " << size;
+  std::ifstream inputfs(model_filename, std::ios::in | std::ios::binary);
   std::string program_desc_str;
-  program_desc_str.resize(size);
-  fs.read(&program_desc_str[0], size);
+  inputfs.seekg(0, std::ios::end);
+  program_desc_str.resize(inputfs.tellg());
+  inputfs.seekg(0, std::ios::beg);
+  LOG(INFO) << "program_desc_str's size: " << program_desc_str.size();
+  inputfs.read(&program_desc_str[0], program_desc_str.size());
+  inputfs.close();
 #endif
   program_ = new framework::ProgramDesc(program_desc_str);
   GenerateLoadProgram(dirname);

paddle/operators/net_op.cc

@@ -56,11 +56,11 @@ void NetOp::CompleteAddOp(bool calc) {
   std::copy(output_set.begin(), output_set.end(), std::back_inserter(outputs));
 }
 
-std::string NetOp::DebugString() const {
+std::string NetOp::DebugStringEx(const framework::Scope* scope) const {
   std::ostringstream os;
-  os << OperatorBase::DebugString() << std::endl;
+  os << OperatorBase::DebugStringEx(scope) << std::endl;
   for (auto& op : ops_) {
-    std::istringstream is(op->DebugString());
+    std::istringstream is(op->DebugStringEx(scope));
     for (std::string line; std::getline(is, line);) {
       os << "    " << line << std::endl;
     }

paddle/operators/net_op.h

@@ -106,7 +106,8 @@ class NetOp : public framework::OperatorBase {
 
   void CompleteAddOp(bool calculate = true);
 
-  std::string DebugString() const override;
+  std::string DebugStringEx(
+      const framework::Scope* scope = nullptr) const override;
 
   bool IsNetOp() const override;
   std::vector<std::string> OutputVars(bool has_intermediate) const override;

paddle/platform/CMakeLists.txt

@@ -21,10 +21,16 @@ ELSE()
     set(GPU_CTX_DEPS)
 ENDIF()
 
+IF(WITH_MKLDNN)
+    set(MKLDNN_CTX_DEPS mkldnn)
+ELSE()
+    set(MKLDNN_CTX_DEPS)
+ENDIF()
+
 # memcpy deoends on device_context, here add deps individually for
 # avoiding cycle dependencies
 cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator
-    system_allocator memory_block meta_data meta_cache place eigen3 ${GPU_CTX_DEPS})
+    system_allocator memory_block meta_data meta_cache place eigen3 ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
 nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_info)
 
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)

paddle/platform/device_context.cc

@@ -168,5 +168,69 @@ cudaStream_t CUDADeviceContext::stream() const { return stream_; }
 
 #endif
 
+#ifdef PADDLE_WITH_MKLDNN
+MKLDNNDeviceContext::MKLDNNDeviceContext(CPUPlace place)
+    : CPUDeviceContext(place), ready_(false) {
+  stream_.reset(new mkldnn::stream(mkldnn::stream::kind::eager));
+  engine_.reset(new mkldnn::engine(mkldnn::engine::cpu, 0));
+}
+
+template <typename T>
+void MKLDNNDeviceContext::AddElement(const std::string& op_key,
+                                     const T& value) {
+  if (GetElement<T>(op_key)) {
+    return;
+  }
+  GetElementPool<T>().emplace(op_key, std::move(value));
+}
+
+template <typename T>
+const T& MKLDNNDeviceContext::GetElement(const std::string& op_key) const {
+  auto it = GetElementPool<T>().find(op_key);
+  return it == GetElementPool<T>().end() ? nullptr : it->second;
+}
+
+template <>
+const std::unordered_map<const std::string, const MKLDNNMemoryPtr,
+                         std::hash<std::string>>&
+MKLDNNDeviceContext::GetElementPool<MKLDNNMemoryPtr>() const {
+  return memory_pool_;
+}
+
+template <>
+const std::unordered_map<const std::string, const MKLDNNPrimitivePtr,
+                         std::hash<std::string>>&
+MKLDNNDeviceContext::GetElementPool<MKLDNNPrimitivePtr>() const {
+  return primitive_pool_;
+}
+
+template <>
+const std::unordered_map<const std::string, const MKLDNNPrimitiveDescPtr,
+                         std::hash<std::string>>&
+MKLDNNDeviceContext::GetElementPool<MKLDNNPrimitiveDescPtr>() const {
+  return primitive_desc_pool_;
+}
+
+void MKLDNNDeviceContext::Execute(bool block) {
+  if (pipeline_.empty()) {
+    return;
+  }
+  ResetStream();
+  stream_->submit(pipeline_).wait(block);
+  ready_ = false;
+  pipeline_.clear();
+}
+
+void MKLDNNDeviceContext::ResetStream() {
+  if (ready_) {
+    return;
+  }
+  // TODO(TJ): change me when mkldnn have specific method to reset this state
+  stream_.reset(new mkldnn::stream(mkldnn::stream::kind::eager));
+  ready_ = true;
+}
+
+#endif
+
 }  // namespace platform
 }  // namespace paddle

paddle/platform/device_context.h

@@ -21,6 +21,10 @@ limitations under the License. */
 #define EIGEN_USE_GPU
 #endif
 
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/platform/mkldnn_helper.h"
+#endif
+
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/place.h"
 #include "unsupported/Eigen/CXX11/Tensor"
@@ -105,6 +109,54 @@ struct DefaultDeviceContextType<platform::CUDAPlace> {
 
 #endif
 
+#ifdef PADDLE_WITH_MKLDNN
+class MKLDNNDeviceContext : public CPUDeviceContext {
+ public:
+  explicit MKLDNNDeviceContext(CPUPlace place);
+
+  /* \brief  Add new element: memory, primitive or primitive desc */
+  template <typename T>
+  void AddElement(const std::string& op_key, const T& value);
+
+  /* \brief  Get existed element: memory, primitive or primitive desc */
+  template <typename T>
+  const T& GetElement(const std::string& op_key) const;
+
+  /* \brief  Get element pool: memory, primitive or primitive desc pool */
+  template <typename T>
+  const std::unordered_map<const std::string, const T, std::hash<std::string>>&
+  GetElementPool() const;
+
+  /* \brief  Get the active engine */
+  const MKLDNNEngine& engine() const { return *engine_; }
+
+  /* \brief  Submit primitive to pipeline */
+  void Submit(const MKLDNNPrimitivePtr& p) { pipeline_.push_back(*p); }
+
+  /*! \brief  Execute all submitted primitives in pipeline */
+  void Execute(bool block = true);
+
+ protected:
+  /*! \brief  Reset the stream to prepare next exectue */
+  void ResetStream();
+
+ private:
+  std::unordered_map<const std::string, const MKLDNNMemoryPtr,
+                     std::hash<std::string>>
+      memory_pool_;
+  std::unordered_map<const std::string, const MKLDNNPrimitivePtr,
+                     std::hash<std::string>>
+      primitive_pool_;
+  std::unordered_map<const std::string, const MKLDNNPrimitiveDescPtr,
+                     std::hash<std::string>>
+      primitive_desc_pool_;
+  std::vector<MKLDNNPrimitive> pipeline_;
+  MKLDNNStreamPtr stream_;
+  MKLDNNEnginePtr engine_;
+  bool ready_;
+};
+#endif
+
 /*! \brief device context pool singleton */
 class DeviceContextPool {
  public:

paddle/platform/mkldnn_helper.h

+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <mkldnn.hpp>
+
+namespace paddle {
+namespace platform {
+
+using MKLDNNStream = mkldnn::stream;
+using MKLDNNEngine = mkldnn::engine;
+using MKLDNNMemory = mkldnn::memory;
+using MKLDNNPrimitive = mkldnn::primitive;
+using MKLDNNPrimitiveDesc = mkldnn::handle<mkldnn_primitive_desc_t>;
+
+typedef std::unique_ptr<MKLDNNStream> MKLDNNStreamPtr;
+typedef std::unique_ptr<MKLDNNEngine> MKLDNNEnginePtr;
+typedef std::unique_ptr<MKLDNNMemory> MKLDNNMemoryPtr;
+typedef std::unique_ptr<MKLDNNPrimitive> MKLDNNPrimitivePtr;
+typedef std::unique_ptr<MKLDNNPrimitiveDesc> MKLDNNPrimitiveDescPtr;
+
+}  // namespace platform
+}  // namespace paddle

paddle/scripts/docker/build.sh

@@ -193,6 +193,25 @@ EOF
 EOF
 }
 
+function gen_capi_package() {
+  if [[ ${WITH_C_API} == "ON" ]]; then
+    install_prefix="/paddle/build/capi_output"
+    rm -rf $install_prefix
+
+    make DESTDIR="$install_prefix" install
+
+    if [[ ${WITH_MKL:-OFF} == "ON" ]]; then
+      find ./third_party/install -name 'libmklml_gnu.so' -exec cp {} $install_prefix/usr/local/lib \;
+      find ./third_party/install -name 'libmklml_intel.so' -exec cp {} $install_prefix/usr/local/lib \;
+      cp -P ./third_party/install/mkldnn/lib/* $install_prefix/usr/local/lib/
+    fi
+
+    find ./third_party/install -name 'libiomp5.so' -exec cp {} $install_prefix/usr/local/lib \;
+    cd $install_prefix/usr/local
+    ls | egrep -v "^Found.*item$" | xargs tar /paddle/build/paddle.tgz
+  fi
+}
+
 set -xe
 
 cmake_gen ${PYTHON_ABI:-""}
@@ -200,6 +219,11 @@ run_build
 run_test
 gen_docs
 gen_dockerfile
-
-printf "If you need to install PaddlePaddle in develop docker image,"
-printf "please make install or pip install build/python/dist/*.whl.\n"
+gen_capi_package
+
+if [[ ${WITH_C_API:-OFF} == "ON" ]]; then
+  printf "PaddlePaddle C-API libraries was generated on build/paddle.tgz\n" 
+else
+  printf "If you need to install PaddlePaddle in develop docker image,"
+  printf "please make install or pip install build/python/dist/*.whl.\n"
+fi

python/paddle/trainer_config_helpers/attrs.py

@@ -58,12 +58,12 @@ def is_compatible_with(x, Type):
 
 class HookAttribute(object):
     """
-    Hook Attribute object. As a member of ParameterAttribute class, the hook is an auxiliary operation that occurs 
+    Hook Attribute object. As a member of ParameterAttribute class, the hook is an auxiliary operation that occurs
     during training process of a layer with parameters, such as img_conv layer, fc layer.
 
-    :param  type: Hook type, currently supported types: 
+    :param  type: Hook type, currently supported types:
                         'pruning' :  user specify a sparsity_ratio before training started, and the
-                            network will prune the parameters based on the sparsity_ratio. 
+                            network will prune the parameters based on the sparsity_ratio.
                             eg: The definition of Hook object can be hk = HookAttribute('pruning', 0.6)
                             The specific usage can be paddle.layer.img_conv(input=img, filter_size=3,
                                                                        num_channels=3, num_filters=64,
@@ -71,10 +71,10 @@ class HookAttribute(object):
                             The pruning details can be found https://arxiv.org/pdf/1506.02626.pdf
     :type type: string
 
-    :param sparsity_ratio: Must be specified if hook type is 'pruning', 
+    :param sparsity_ratio: Must be specified if hook type is 'pruning',
                         it represents the ratio of the zero elements to be set by the Parameter.
     :type sparsity_ratio: float or None
-	
+
     """
 
     def __init__(self, type, sparsity_ratio=None):
@@ -130,10 +130,12 @@ class ParameterAttribute(object):
     :param sparse_update: Enable sparse update for this parameter. It will
                           enable both local and remote sparse update.
     :type sparse_update: bool
+    :param update_hooks: A HookAttribute object.
+    :type update_hooks: HookAttribute
     :param initializer: If not None, it should be a callable object which accepts
                         a parameter name and returns numpy array for the initial
                         value of the parameter
-    :param initializer: callable object
+    :type initializer: callable object
     """
 
     def __init__(self,

python/paddle/v2/fluid/backward.py

@@ -7,7 +7,7 @@ __all__ = ['append_backward']
 
 def _rename_arg_(op_descs, old_name, new_name, begin_idx=None, end_idx=None):
     """
-    Traverse all ops in op_descs[begin_idx : end_idx], 
+    Traverse all ops in op_descs[begin_idx : end_idx],
     if any op has inputs/outputs named "old_name", rename it as 'new_name'
     """
     if begin_idx is None:
@@ -162,7 +162,7 @@ def _remove_no_grad_branch_(op_descs, no_grad_set):
             if core.grad_var_suffix() in arg and arg in no_grad_set:
                 to_insert.append((_create_op_desc_("fill_zeros_like", {
                     "X": [_strip_grad_suffix_(arg)]
-                }, {"Y": [arg]}, {}), idx))
+                }, {"Out": [arg]}, {}), idx))
 
     map(lambda p: op_descs.insert(p[1], p[0]), reversed(to_insert))
 
@@ -182,7 +182,7 @@ def _append_backward_ops_(target,
         target(Variable): the target variable of forward pass
         block(Block): the block where forward ops are
         target_block(Block): the block which is going to hold new generated grad ops
-        no_grad_dict(dict): 
+        no_grad_dict(dict):
             key(int)  block index
             val(set) a set of varibale names. These varibales have no gradient
         grad_to_var(dict)(output argument):
@@ -276,8 +276,8 @@ def append_backward(loss, parameter_list=None, no_grad_set=None):
         loss(Variable): The variable generated by cost function.
         parameter_list(list): Parameters that need to be updated by optimizer.
             If None, it means all parameters need to be updated.
-        no_grad_set(set): Variables that have no gradients in Block 0. 
-            If None, the set will be generated inside the function and 
+        no_grad_set(set): Variables that have no gradients in Block 0.
+            If None, the set will be generated inside the function and
             contains all variables with `step_gradient=True` from all blocks.
 
     Return:

python/paddle/v2/fluid/io.py

@@ -212,6 +212,11 @@ def save_inference_model(dirname,
             "fetch_var_names": fetch_var_names
         }, f, -1)
 
+    # Save only programDesc of inference_program in binary format
+    # in another file: __model__.dat
+    with open(model_file_name + ".dat", "wb") as fp:
+        fp.write(inference_program.desc.serialize_to_string())
+
     save_params(executor, dirname, main_program)
 
 

python/paddle/v2/fluid/layers/control_flow.py

@@ -998,7 +998,7 @@ class ConditionalBlock(object):
 
         out_list = [
             parent_block.var(var_name) for var_name in parent_block.vars
-            if var_name not in intermediate
+            if var_name in intermediate
         ]
 
         step_scope = parent_block.create_var(

python/paddle/v2/fluid/layers/nn.py

@@ -64,14 +64,14 @@ def fc(input,
                               is flattened: the first `num_flatten_dims`
                               dimensions will be flatten to form the first
                               dimension of the final matrix (height of the
-                              matrix), and the rest `rank(X) - num_col_dims`
+                              matrix), and the rest `rank(X) - num_flatten_dims`
                               dimensions are flattened to form the second
                               dimension of the final matrix (width of the matrix).
                               For example, suppose `X` is a 6-dimensional tensor
                               with a shape [2, 3, 4, 5, 6], and
-                              `x_num_col_dims` = 3. Then, the flattened matrix
+                              `num_flatten_dims` = 3. Then, the flattened matrix
                               will have a shape [2 x 3 x 4, 5 x 6] = [24, 30].
-                              By default, `x_num_col_dims` is set to 1.
+                              By default, `num_flatten_dims` is set to 1.
        param_attr(ParamAttr|list): The parameter attribute for learnable
                                    parameters/weights of the fully connected
                                    layer.
@@ -243,18 +243,21 @@ def gru_unit(input,
 
             r_t & = actGate(xr_{t} + W_r h_{t-1} + b_r)
 
-            ch_t & = actNode(xc_t + W_c dot(r_t, h_{t-1}) + b_c)
+            m_t & = actNode(xm_t + W_c dot(r_t, h_{t-1}) + b_m)
 
-            h_t & = dot((1-u_t), ch_{t-1}) + dot(u_t, h_t)
+            h_t & = dot((1-u_t), m_t) + dot(u_t, h_{t-1})
 
     The inputs of gru unit includes :math:`z_t`, :math:`h_{t-1}`. In terms
     of the equation above, the :math:`z_t` is split into 3 parts - 
-    :math:`xu_t`, :math:`xr_t` and :math:`xc_t`. This means that in order to 
+    :math:`xu_t`, :math:`xr_t` and :math:`xm_t`. This means that in order to 
     implement a full GRU unit operator for an input, a fully 
     connected layer has to be applied, such that :math:`z_t = W_{fc}x_t`.
 
-    This layer has three outputs :math:`h_t`, :math:`dot(r_t, h_{t - 1})`
-    and concatenation of :math:`u_t`, :math:`r_t` and :math:`ch_t`.
+    The terms :math:`u_t` and :math:`r_t` represent the update and reset gates 
+    of the GRU cell. Unlike LSTM, GRU has one lesser gate. However, there is 
+    an intermediate candidate hidden output, which is denoted by :math:`m_t`.
+    This layer has three outputs :math:`h_t`, :math:`dot(r_t, h_{t-1})`
+    and concatenation of :math:`u_t`, :math:`r_t` and :math:`m_t`.
 
     Args:
         input (Variable): The fc transformed input value of current step.