Merge remote-tracking branch 'origin/develop' into mul_op

.clang_format.hook

+#!/bin/bash
+set -e
+
+readonly VERSION="3.8"
+
+version=$(clang-format -version)
+
+if ! [[ $version == *"$VERSION"* ]]; then
+    echo "clang-format version check failed."
+    echo "a version contains '$VERSION' is needed, but get '$version'"
+    echo "you can install the right version, and make an soft-link to '\$PATH' env"
+    exit -1
+fi
+
+clang-format $@

.pre-commit-config.yaml

@@ -19,10 +19,10 @@
     -   id: end-of-file-fixer
 -   repo: local
     hooks:
-    -   id: clang-format
+    -   id: clang-format-with-version-check
         name: clang-format
         description: Format files with ClangFormat.
-        entry: clang-format -i
+        entry: ./.clang_format.hook -i
         language: system
         files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$
 -   repo: https://github.com/PaddlePaddle/pre-commit-golang

Dockerfile

@@ -71,20 +71,6 @@ RUN pip install -r /root/requirements.txt
 RUN apt-get install -y libssl-dev libffi-dev
 RUN pip install certifi urllib3[secure]
 
-# TODO(qijun) The template library Eigen doesn't work well with GCC 5 
-# coming with the default Docker image, so we switch to use GCC 4.8 
-# by default. And I will check Eigen library later.
-
-RUN ln -sf gcc-4.8 /usr/bin/gcc && \
-    ln -sf gcc-ar-4.8 /usr/bin/gcc-ar && \
-    ln -sf gcc-nm-4.8 /usr/bin/gcc-nm && \
-    ln -sf gcc-ranlib-4.8 /usr/bin/gcc-ranlib && \
-    ln -sf gcc-4.8 /usr/bin/x86_64-linux-gnu-gcc && \
-    ln -sf gcc-ar-4.8 /usr/bin/x86_64-linux-gnu-gcc-ar && \
-    ln -sf gcc-nm-4.8 /usr/bin/x86_64-linux-gnu-gcc-nm && \
-    ln -sf gcc-ranlib-4.8 /usr/bin/x86_64-linux-gnu-gcc-ranlib && \
-    ln -sf g++-4.8 /usr/bin/g++ && \
-    ln -sf g++-4.8 /usr/bin/x86_64-linux-gnu-g++ 
 
 # Install woboq_codebrowser to /woboq
 RUN git clone https://github.com/woboq/woboq_codebrowser /woboq && \

cmake/flags.cmake

@@ -9,13 +9,6 @@ function(CheckCompilerCXX11Flag)
         if(${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 4.8)
             message(FATAL_ERROR "Unsupported GCC version. GCC >= 4.8 required.")
         endif()
-        if(NOT ANDROID)
-            # TODO(qijun) gcc 4.9 or later versions raise SEGV due to the optimization problem.
-            # Use Debug mode instead for now.
-            if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 4.9 OR CMAKE_CXX_COMPILER_VERSION VERSION_EQUAL 4.9)
-                set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "" FORCE)
-            endif()
-        endif()
     elseif(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
         # cmake >= 3.0 compiler id "AppleClang" on Mac OS X, otherwise "Clang"
         # Apple Clang is a different compiler than upstream Clang which havs different version numbers.
@@ -160,7 +153,7 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF)
 
 # Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc.
 # So, don't set these flags here.
-LIST(APPEND CUDA_NVCC_FLAGS -std=c++11 --default-stream per-thread)
+LIST(APPEND CUDA_NVCC_FLAGS -std=c++11)
 LIST(APPEND CUDA_NVCC_FLAGS --use_fast_math)
 
 if(CMAKE_BUILD_TYPE  STREQUAL "Debug")

doc/design/cluster_train/large_model_dist_train.md

+# Alalysis of large model distributed training in Paddle
+
+***NOTE: This is only some note for how we implemeted this scheme in V1, not a new design.***
+
+## What is it
+
+We often encounter cases that the embedding layer parameters(sparse) are so large that we can not store it in the trainer's memory when training. So we need to put them to several servers, and fetch them row by row instead of fetch all of the parameters.
+
+## How to use
+
+Specify command-line argument like  `--loadsave_parameters_in_pserver=true --ports_num_for_sparse=1  --use_old_updater=1` when starting the paddle trainer. And also add something like `--ports_num_for_sparse=1 --pserver_num_threads=5` when starting pserver processes.
+
+Accrodingly, configure your embedding layers like:
+
+```python
+SPARSE_REMOTE=True
+
+w1 = data_layer(name="w1", size=dict_size)
+emb1 = embedding_layer(input=w1, size=32, param_attr=ParameterAttribute(sparse_update=SPARSE_REMOTE))
+w2 = data_layer(name="w2", size=dict_size)
+emb2 = embedding_layer(input=w2, size=32, param_attr=ParameterAttribute(sparse_update=SPARSE_REMOTE))
+...
+```
+
+## Implementation details
+
+```c++
+enum MatType {
+  MAT_NORMAL,
+  MAT_NORMAL_SHARED,
+  MAT_VALUE_SHARED,
+  MAT_SPARSE_ROW_IDS,
+  MAT_SPARSE_ROW_AUTO_GROW,
+  MAT_CACHE_ROW,
+  MAT_SPARSE_ROW,
+  MAT_SPARSE_ROW_PREFETCH,
+  MAT_SPARSE_ROW_PREFETCH_FULL_SIZE,
+};
+```
+
+`MAT_SPARSE_ROW_PREFETCH` is what we use when configured to fetch only row of matrix when training.
+
+In `trainer_internal.cpp:L93 trainOneBatch`:
+
+```c++
+  if (config_->getOptConfig().use_sparse_remote_updater()) {
+    REGISTER_TIMER("prefetch");
+    gradientMachine_->prefetch(inArgs);
+    parameterUpdater_->getParametersRemote();
+  }
+```
+
+When doing actual network forward and backward, at the beginning of each batch, the trainer will try to download one row of data from pserver.
+
+In `trainer/RemoteParameterUpdater.cpp`: `parameterUpdater_->getParametersRemote();`:
+
+```c++
+if (fullSize) {
+    ...
+} else {
+getParams = [&] {
+    parameterClient_->getParameterSparse(
+        /* recvParameterType= */ PARAMETER_VALUE, sendBackParameterType);
+};
+applyL1 = [](Parameter& para, real decayRate) {
+    para.getMat(PARAMETER_VALUE)->applyL1(/*lr=*/1.0f, decayRate);
+};
+}
+```
+
+Calling `parameterClient_->getParameterSparse` will do remote call to pserver's `getParameterSparse`:
+
+```c++
+void ParameterServer2::getParameterSparse(const SendParameterRequest& request,
+                                          std::vector<Buffer>& inputBuffers,
+                                          SendParameterResponse* response,
+                                          std::vector<Buffer>* outputBuffers) {
+  (void)inputBuffers;
+  auto& buffer = *readWriteBuffer_;
+  size_t numReals = 0;
+  for (const auto& block : request.blocks()) {
+    numReals += getParameterConfig(block).dims(1);
+  }
+  buffer.resize(numReals);
+
+  VLOG(3) << "pserver: getParameterSparse, numReals=" << numReals;
+
+  ReadLockGuard guard(parameterMutex_);
+  size_t offset = 0;
+  for (const auto& block : request.blocks()) {
+    size_t width = getParameterConfig(block).dims(1);
+    Buffer buf = {buffer.data() + offset, width};
+    int type = request.send_back_parameter_type();
+    sendBackParameterSparse(block, type, response, &buf, width, outputBuffers);
+    offset += width;
+  }
+}
+```
+
+`getParameterConfig(block).dims(1)` returns the width of the current "parameter block"(a shard of parameter object),
+then `getParameterSparse` remote call returns only one row of data to the client.

doc/design/mkldnn/README.MD

@@ -101,6 +101,7 @@ if use_mkldnn
 5. 在**Argument**里添加两个`MkldnnMatrixPtr`，取名为`mkldnnValue`和`mkldnnGrad`，用于存放`MkldnnLayer`会用到的memory buffer。 并且添加函数cvt(会修改为一个更加合适的函数名)，用于处理"CPU device"和"MKL-DNN device"之间memory的相互转化。
 6. 在父类`Layer`中的`getOutput`函数中添加一段逻辑，用于判断`deviceId`，并针对device在MKL-DNN和CPU之间不统一的情况，做一个前期转换。 也就是调用`Argument`的cvt函数把output统一到需要的device上。
 7. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
+8. 关于MKLDNN参数的保存。由于MKLDNN参数的格式与PaddlePaddle原有的格式存在不一样的情况，所以需要在保存参数时同时保存该格式信息。目前准备扩展[Header](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/parameter/Parameter.h#L247)里面的`int32_t version`。这个值不管是在v1还是在v2里面，一直保存的是0，所以可以充分利用这个信息，定义一个枚举处理所有MKLDNN的参数格式，从而`MKLDNNLayer`就可以从输入的参数中获取需要的格式信息。
 
 ## References
 

doc/getstarted/build_and_install/build_from_source_en.md

@@ -68,7 +68,7 @@ As a simple example, consider the following:
 
 1. **BLAS Dependencies(optional)**
   
-    CMake will search BLAS libraries from system. If not found, OpenBLAS will be downloaded, built and installed automatically.
+    CMake will search BLAS libraries from the system. If not found, OpenBLAS will be downloaded, built and installed automatically.
     To utilize preinstalled BLAS， you can simply specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`.
 
     ```bash
@@ -131,9 +131,9 @@ As a simple example, consider the following:
     To build GPU version, you will need the following installed:
 
         1. a CUDA-capable GPU
-        2. A supported version of Linux with a gcc compiler and toolchain
+        2. A supported version of Linux with a GCC compiler and toolchain
         3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
-        4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
+        4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
 
     The CUDA development environment relies on tight integration with the host development environment,
     including the host compiler and C runtime libraries, and is therefore only supported on
@@ -172,6 +172,7 @@ export PATH=<path to install>/bin:$PATH
 # install PaddlePaddle Python modules.
 sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
 ```
+
 ## <span id="centos">Build on Centos 7</span>
 
 ### Install Dependencies
@@ -192,9 +193,9 @@ sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
     To build GPU version, you will need the following installed:
 
         1. a CUDA-capable GPU
-        2. A supported version of Linux with a gcc compiler and toolchain
+        2. A supported version of Linux with a GCC compiler and toolchain
         3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
-        4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
+        4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
 
     The CUDA development environment relies on tight integration with the host development environment,
     including the host compiler and C runtime libraries, and is therefore only supported on
@@ -222,7 +223,7 @@ mkdir build && cd build
 ``` 
 
 Finally, you can build and install PaddlePaddle:
-
+  
 ```bash
 # you can add build option here, such as:    
 cmake3 .. -DCMAKE_INSTALL_PREFIX=<path to install>

paddle/capi/gradient_machine.cpp

@@ -146,3 +146,19 @@ paddle_error paddle_gradient_machine_randomize_param(
   m->machine->randParameters();
   return kPD_NO_ERROR;
 }
+
+paddle_error paddle_gradient_machine_get_layer_output(
+    paddle_gradient_machine machine,
+    const char* layerName,
+    paddle_arguments args) {
+  auto m = cast(machine);
+  auto out = paddle::capi::cast<paddle::capi::CArguments>(args);
+  if (m == nullptr || layerName == nullptr || out == nullptr ||
+      m->machine == nullptr) {
+    return kPD_NULLPTR;
+  }
+
+  auto layerOutput = m->machine->getLayerOutput(layerName);
+  out->args.push_back(layerOutput);
+  return kPD_NO_ERROR;
+}

paddle/capi/gradient_machine.h

@@ -39,7 +39,11 @@ PD_API paddle_error paddle_gradient_machine_create_for_inference(
 /**
  * @brief Create a gradient machine used for model inference, using config with
  *        parameters which is generated by `paddle merge_model`.
- * @param [out] machine that used for model inference.
+ *        Example:
+ *          paddle merge_model \
+ *                 --model_dir="pass-00000" \
+ *                 --model_file="merged_model.paddle"
+ * @param [out] machine that used for model inference
  * @param [in] mergedModel
  * @param [in] size
  * @return paddle_error
@@ -97,6 +101,18 @@ paddle_gradient_machine_randomize_param(paddle_gradient_machine machine);
 PD_API paddle_error
 paddle_gradient_machine_destroy(paddle_gradient_machine machine);
 
+/**
+ * @brief Get the output of the layer named `layerName`.
+ * @param [in] gradient machine that have run a inference
+ * @param [in] layerName name of specified layer
+ * @param [out] args output of the specified layer
+ * @return paddle_error
+ */
+PD_API paddle_error
+paddle_gradient_machine_get_layer_output(paddle_gradient_machine machine,
+                                         const char* layerName,
+                                         paddle_arguments args);
+
 #ifdef __cplusplus
 }
 #endif

paddle/framework/CMakeLists.txt

@@ -38,7 +38,7 @@ add_custom_command(TARGET framework_py_proto POST_BUILD
     WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
 
 cc_library(backward SRCS backward.cc DEPS net_op)
-cc_test(backward_test SRCS backward_test.cc DEPS backward)
+cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context)
 
 if(WITH_PYTHON)
 cc_library(paddle_pybind SHARED

paddle/framework/backward.cc

@@ -15,8 +15,11 @@
 #include "paddle/framework/backward.h"
 
 #include <list>
+#include <memory>
+
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/net_op.h"
+#include "paddle/operators/recurrent_op.h"
 
 namespace paddle {
 namespace framework {
@@ -42,11 +45,11 @@ static bool AllInSet(
   return all_in_set;
 }
 
-static std::shared_ptr<OperatorBase> NOP() {
-  auto net_op = std::make_shared<operators::NetOp>();
+static std::unique_ptr<OperatorBase> NOP() {
+  auto net_op = new operators::NetOp();
   net_op->SetType("@NOP@");
   net_op->CompleteAddOp();
-  return net_op;
+  return std::unique_ptr<OperatorBase>(net_op);
 }
 
 //  Get backward operator from a forward operator, a recursive implementation.
@@ -61,11 +64,7 @@ static std::shared_ptr<OperatorBase> NOP() {
 //  operator, in a complex situation, it maybe a NetOp.
 //
 //  See Backward.h for details
-static std::shared_ptr<OperatorBase> BackwardRecursive(
-    const OperatorBase& forwardOp,
-    std::unordered_set<std::string>& no_grad_names, size_t& uniq_id);
-
-std::shared_ptr<OperatorBase> BackwardRecursive(
+static std::unique_ptr<OperatorBase> BackwardRecursive(
     const OperatorBase& forwardOp,
     std::unordered_set<std::string>& no_grad_names, size_t& uniq_id) {
   //  If all input gradients of forwarding operator do not need to calculate,
@@ -90,7 +89,7 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
   }
 
   // Returned gradient network
-  auto net = std::make_shared<operators::NetOp>();
+  auto net = std::unique_ptr<operators::NetOp>(new operators::NetOp());
 
   if (forwardOp.IsNetOp()) {
     // Because forwardOp is a net op, it can static_cast.
@@ -104,14 +103,14 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
     // reversely travel forwardNet and collect all duplicate outputs.
     for (auto it = forwardNet.ops_.rbegin(); it != forwardNet.ops_.rend();
          ++it, ++local_op_id) {
-      auto fwd = *it;
+      auto& fwd = *it;
       auto bwd = BackwardRecursive(*fwd, no_grad_names, uniq_id);
-      net->AddOp(bwd);
       ForEachVarName(bwd->Outputs(),
                      [&dup_output_ops, local_op_id](const std::string& out) {
                        dup_output_ops[out].emplace_back(local_op_id);
                        return false;
                      });
+      net->AddOp(std::move(bwd));
     }
     // Get unique ID for this method.
     auto uid = uniq_id++;
@@ -121,7 +120,7 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
     // to handle this case. For each duplicate output, rename it to an alias
     // (original name with a offset), append an `add` op for its operator,
     // and finally sum all the alias variable to the final output variable y.
-    using Pos = std::pair<size_t, std::shared_ptr<OperatorBase>>;
+    using Pos = std::pair<size_t, std::unique_ptr<OperatorBase>>;
     std::list<Pos> insert_position;
     for (auto& dup_output_op : dup_output_ops) {
       const std::string& name = dup_output_op.first;
@@ -149,13 +148,13 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
         [](const Pos& l, const Pos& r) { return l.first > r.first; });
 
     for (auto& pos : insert_position) {
-      net->InsertOp(pos.first + 1, pos.second);
+      net->InsertOp(pos.first + 1, std::move(pos.second));
     }
   } else {
-    std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
+    std::unique_ptr<OperatorBase> grad_op(OpRegistry::CreateGradOp(forwardOp));
 
-    ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net,
-                                       grad_op](const std::string& grad_input) {
+    ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op](
+                                          const std::string& grad_input) {
       if (no_grad_names.count(grad_input)) {
         // +1 for \0
         std::string prefix = grad_input.substr(
@@ -178,18 +177,34 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
                      return false;
                    });
 
+    // process recurrent gradient op as a special operator.
+    if (forwardOp.Type() == "recurrent_op") {
+      // NOTE clean up cycle call somewhere (RNN's stepnet constains itself), or
+      // this will result in infinite loop.
+      const auto& rnnop =
+          *static_cast<const operators::RecurrentOp*>(&forwardOp);
+      auto rnn_grad_op =
+          static_cast<operators::RecurrentGradientOp*>(grad_op.get());
+      const auto& stepnet_op =
+          *static_cast<const OperatorBase*>(&rnnop.stepnet());
+      // create stepnet's gradient op
+      rnn_grad_op->set_stepnet(
+          BackwardRecursive(stepnet_op, no_grad_names, uniq_id));
+    }
+
     if (net->ops_.empty()) {  // Current no aux op is added to network
       return grad_op;
     }
-    net->AddOp(grad_op);
+    net->AddOp(std::move(grad_op));
   }
   net->SetType("@GENERATED_BACKWARD@");
   net->CompleteAddOp();
-  return net;
-}  // namespace framework
+  return std::unique_ptr<OperatorBase>(
+      static_cast<OperatorBase*>(net.release()));
+}
 
 // See header for comments
-std::shared_ptr<OperatorBase> Backward(
+std::unique_ptr<OperatorBase> Backward(
     const OperatorBase& forwardOp,
     const std::unordered_set<std::string>& no_grad_vars) {
   std::unordered_set<std::string> no_grad_names;

paddle/framework/backward.h

@@ -20,7 +20,7 @@ namespace framework {
 
 // Create the backward operator from a forward operator.
 // TODO(yuyang18): Add more API reference comment.
-extern std::shared_ptr<OperatorBase> Backward(
+extern std::unique_ptr<OperatorBase> Backward(
     const OperatorBase& forwardOp,
     const std::unordered_set<std::string>& no_grad_vars);
 }  // namespace framework

paddle/framework/backward_test.cc

@@ -180,8 +180,7 @@ TEST(Backward, simple_op_not_need_grad) {
   auto no_input_gop = f::Backward(*fwd, {"x", "b"});
   ASSERT_NE(no_input_gop, nullptr);
   ASSERT_TRUE(no_input_gop->IsNetOp());
-  ASSERT_EQ(0UL,
-            std::static_pointer_cast<ops::NetOp>(no_input_gop)->ops_.size());
+  ASSERT_EQ(0UL, static_cast<ops::NetOp *>(no_input_gop.get())->ops_.size());
 }
 
 TEST(Backward, net_fc_backward_normal) {

paddle/framework/op_registry.cc

@@ -17,5 +17,48 @@ limitations under the License. */
 #include <vector>
 
 namespace paddle {
-namespace framework {}  // namespace framework
+namespace framework {
+
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
+                                                   const VarNameMap& inputs,
+                                                   const VarNameMap& outputs,
+                                                   AttributeMap attrs) {
+  auto it = op_info_map().find(type);
+  PADDLE_ENFORCE(it != op_info_map().end(),
+                 "Operator '%s' has not been registered.", type);
+  it->second.checker_->Check(attrs);
+  auto op = it->second.creator_(type, inputs, outputs, attrs);
+  return std::unique_ptr<OperatorBase>(op);
+}
+
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
+  VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
+  VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
+  AttributeMap attrs;
+  for (auto& attr : op_desc.attrs()) {
+    attrs[attr.name()] = GetAttrValue(attr);
+  }
+
+  return CreateOp(op_desc.type(), inputs, outputs, attrs);
+}
+
+OperatorBase::VarNameMap OpRegistry::ConvertOpDescVarsToVarNameMap(
+    const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars) {
+  VarNameMap ret_val;
+  for (auto& var : op_desc_vars) {
+    auto& var_names = ret_val[var.parameter()];
+    auto& var_names_in_proto = var.arguments();
+    var_names.reserve(static_cast<size_t>(var_names_in_proto.size()));
+    std::copy(var_names_in_proto.begin(), var_names_in_proto.end(),
+              std::back_inserter(var_names));
+  }
+  return ret_val;
+}
+
+std::unique_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
+  PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops");
+  return std::unique_ptr<OperatorBase>(BuildGradOp(&op));
+}
+
+}  // namespace framework
 }  // namespace paddle

paddle/framework/op_registry.h

@@ -29,103 +29,6 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-// this class not only make proto but also init attribute checkers.
-class OpProtoAndCheckerMaker {
- public:
-  OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : proto_(proto), op_checker_(op_checker) {}
-
-  ~OpProtoAndCheckerMaker() {
-    PADDLE_ENFORCE(validated_, "should call Validate after build");
-  }
-
-  void Validate() {
-    validated_ = true;
-    CheckNoDuplicatedInOutAttrs();
-  }
-
- protected:
-  struct VariableBuilder {
-    OpProto::Var* var_;
-
-    VariableBuilder& AsDuplicable() {
-      var_->set_duplicable(true);
-      return *this;
-    }
-
-    VariableBuilder& AsIntermediate() {
-      var_->set_intermediate(true);
-      return *this;
-    }
-
-    // TODO(FengJiayi, yuyang18): `AsNoGradient` is a very bad name, because it
-    // means that input/output is not needed when calculate gradient. It does
-    // not mean no gradient when backward. It should be changed soon.
-    VariableBuilder& AsNoGradient() {
-      var_->set_no_gradient(true);
-      return *this;
-    }
-  };
-
-  VariableBuilder AddInput(const std::string& name,
-                           const std::string& comment) {
-    auto* input = proto_->add_inputs();
-    input->set_name(name);
-    input->set_comment(comment);
-    return VariableBuilder{input};
-  }
-
-  VariableBuilder AddOutput(const std::string& name,
-                            const std::string& comment) {
-    auto* output = proto_->add_outputs();
-    output->set_name(name);
-    output->set_comment(comment);
-    return VariableBuilder{output};
-  }
-
-  template <typename T>
-  TypedAttrChecker<T>& AddAttr(const std::string& name,
-                               const std::string& comment,
-                               bool generated = false) {
-    auto* attr = proto_->add_attrs();
-    attr->set_name(name);
-    attr->set_comment(comment);
-    attr->set_generated(generated);
-    attr->set_type(AttrTypeID<T>());
-    return op_checker_->AddAttrChecker<T>(name);
-  }
-
-  void AddComment(const std::string& comment) { proto_->set_comment(comment); }
-
- private:
-  void CheckNoDuplicatedInOutAttrs() {
-    std::unordered_set<std::string> names;
-    auto checker = [&](const std::string& name) {
-      PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name);
-      names.insert(name);
-    };
-    for (auto& attr : proto_->attrs()) {
-      checker(attr.name());
-    }
-    for (auto& input : proto_->inputs()) {
-      checker(input.name());
-    }
-    for (auto& output : proto_->outputs()) {
-      checker(output.name());
-    }
-  }
-
-  OpProto* proto_;
-  OpAttrChecker* op_checker_;
-  bool validated_{false};
-};
-
-class NOPMaker : public OpProtoAndCheckerMaker {
- public:
-  NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {}
-};
-
 class OpRegistry {
   using VarNameMap = OperatorBase::VarNameMap;
   using OpCreator = std::function<OperatorBase*(
@@ -174,48 +77,17 @@ class OpRegistry {
     }
   }
 
-  static std::shared_ptr<OperatorBase> CreateOp(const std::string& type,
+  static std::unique_ptr<OperatorBase> CreateOp(const std::string& type,
                                                 const VarNameMap& inputs,
                                                 const VarNameMap& outputs,
-                                                AttributeMap attrs) {
-    auto it = op_info_map().find(type);
-    PADDLE_ENFORCE(it != op_info_map().end(),
-                   "Operator '%s' has not been registered.", type);
-    it->second.checker_->Check(attrs);
-    auto op = it->second.creator_(type, inputs, outputs, attrs);
-    return std::shared_ptr<OperatorBase>(op);
-  }
-
-  static VarNameMap ConvertOpDescVarsToVarNameMap(
-      const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars) {
-    VarNameMap ret_val;
-    for (auto& var : op_desc_vars) {
-      auto& var_names = ret_val[var.parameter()];
-      auto& var_names_in_proto = var.arguments();
-      var_names.reserve(static_cast<size_t>(var_names_in_proto.size()));
-      std::copy(var_names_in_proto.begin(), var_names_in_proto.end(),
-                std::back_inserter(var_names));
-    }
-    return ret_val;
-  }
+                                                AttributeMap attrs);
 
-  static std::shared_ptr<OperatorBase> CreateOp(const OpDesc& op_desc) {
-    VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
-    VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
-    AttributeMap attrs;
-    for (auto& attr : op_desc.attrs()) {
-      attrs[attr.name()] = GetAttrValue(attr);
-    }
+  static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
 
-    return CreateOp(op_desc.type(), inputs, outputs, attrs);
-  }
+  static VarNameMap ConvertOpDescVarsToVarNameMap(
+      const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars);
 
-  static std::shared_ptr<OperatorBase> CreateGradOp(const OperatorBase& op) {
-    PADDLE_ENFORCE(!op.IsNetOp(),
-                   "Use framework::Backward to get backward ops");
-    std::shared_ptr<OperatorBase> grad_op(BuildGradOp(&op));
-    return grad_op;
-  }
+  static std::unique_ptr<OperatorBase> CreateGradOp(const OperatorBase& op);
 
   static std::unordered_map<std::string, const OpInfo>& op_info_map() {
     static std::unordered_map<std::string, const OpInfo> op_info_map_;
@@ -272,8 +144,18 @@ class OpKernelRegistrar : public Registrar {
                     grad_op_class)                                            \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                             \
       __reg_op__##op_type, "REGISTER_OP must be called in global namespace"); \
-  static ::paddle::framework::OpRegistrar<op_class, op_maker_class,           \
-                                          grad_op_class>                      \
+  class _OpClass_##op_type##_ : public op_class {                             \
+   public:                                                                    \
+    DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_);                            \
+    DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class);                   \
+  };                                                                          \
+  class _OpGradClass_##op_type##_ : public grad_op_class {                    \
+   public:                                                                    \
+    DEFINE_OP_CLONE_METHOD(_OpGradClass_##op_type##_);                        \
+    DEFINE_OP_CONSTRUCTOR(_OpGradClass_##op_type##_, grad_op_class);          \
+  };                                                                          \
+  static ::paddle::framework::OpRegistrar<                                    \
+      _OpClass_##op_type##_, op_maker_class, _OpGradClass_##op_type##_>       \
       __op_registrar_##op_type##__(#op_type, #grad_op_type);                  \
   int TouchOpRegistrar_##op_type() {                                          \
     __op_registrar_##op_type##__.Touch();                                     \
@@ -304,7 +186,8 @@ class OpKernelRegistrar : public Registrar {
   REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
 
 /**
- * Macro to mark what Operator and Kernel we will use and tell the compiler to
+ * Macro to mark what Operator and Kernel
+ * we will use and tell the compiler to
  * link them into target.
  */
 #define USE_OP_ITSELF(op_type)                                    \
@@ -324,7 +207,8 @@ class OpKernelRegistrar : public Registrar {
       __attribute__((unused)) =                                  \
           TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE()
 
-// TODO(fengjiayi): The following macros seems ugly, do we have better method?
+// TODO(fengjiayi): The following macros
+// seems ugly, do we have better method?
 
 #ifdef PADDLE_ONLY_CPU
 #define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU)

paddle/framework/op_registry_test.cc

@@ -76,8 +76,7 @@ TEST(OpRegistry, CreateOp) {
   attr->set_type(paddle::framework::AttrType::FLOAT);
   attr->set_f(scale);
 
-  std::shared_ptr<paddle::framework::OperatorBase> op =
-      paddle::framework::OpRegistry::CreateOp(op_desc);
+  auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::framework::Scope scope;
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);
@@ -118,8 +117,7 @@ TEST(OpRegistry, DefaultValue) {
 
   ASSERT_TRUE(op_desc.IsInitialized());
 
-  std::shared_ptr<paddle::framework::OperatorBase> op =
-      paddle::framework::OpRegistry::CreateOp(op_desc);
+  auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::framework::Scope scope;
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);

paddle/framework/operator.cc

@@ -164,5 +164,43 @@ std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
   return ret_val;
 }
 
+void OpProtoAndCheckerMaker::Validate() {
+  validated_ = true;
+  CheckNoDuplicatedInOutAttrs();
+}
+
+OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddInput(
+    const std::string& name, const std::string& comment) {
+  auto* input = proto_->add_inputs();
+  input->set_name(name);
+  input->set_comment(comment);
+  return OpProtoAndCheckerMaker::VariableBuilder{input};
+}
+
+OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddOutput(
+    const std::string& name, const std::string& comment) {
+  auto* output = proto_->add_outputs();
+  output->set_name(name);
+  output->set_comment(comment);
+  return OpProtoAndCheckerMaker::VariableBuilder{output};
+}
+
+void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() {
+  std::unordered_set<std::string> names;
+  auto checker = [&](const std::string& name) {
+    PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name);
+    names.insert(name);
+  };
+  for (auto& attr : proto_->attrs()) {
+    checker(attr.name());
+  }
+  for (auto& input : proto_->inputs()) {
+    checker(input.name());
+  }
+  for (auto& output : proto_->outputs()) {
+    checker(output.name());
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/operator.h

@@ -67,10 +67,6 @@ class OperatorBase {
   OperatorBase(const std::string& type, const VarNameMap& inputs,
                const VarNameMap& outputs, const AttributeMap& attrs);
 
-  OperatorBase(const OperatorBase& o) = delete;
-  OperatorBase& operator=(const OperatorBase& o) = delete;
-  OperatorBase(OperatorBase&& o) = delete;
-
   virtual ~OperatorBase() {}
 
   template <typename T>
@@ -116,10 +112,14 @@ class OperatorBase {
   void SetType(const std::string& type) { type_ = type; }
   const AttributeMap& Attrs() const { return attrs_; }
 
+  // Return a new operator instance, which is as same as this.
+  // Use unique_ptr to prevent caller forget to delete this pointer.
+  virtual std::unique_ptr<OperatorBase> Clone() const = 0;
+
  protected:
   std::string type_;
   // NOTE: in case of OpGrad, inputs_ contains:
-  // I (Inputs)
+  // I (Inputs)opear
   // O (Outputs)
   // OG (Output Gradients)
   VarNameMap inputs_;
@@ -130,12 +130,100 @@ class OperatorBase {
   AttributeMap attrs_;
 };
 
+// Macro for define a clone method.
+// If you are writing an kernel operator, `Clone` will be defined when you
+// register it. i.e. `Clone` method is not needed to define by yourself.
+#define DEFINE_OP_CLONE_METHOD(CLS)                       \
+  std::unique_ptr<OperatorBase> Clone() const final {     \
+    return std::unique_ptr<OperatorBase>(new CLS(*this)); \
+  }
+
+// Macro for define a default constructor for Operator.
+// You can also use
+//   using PARENT_CLASS::PARENT_CLASS;
+// to use parent's constructor.
+#define DEFINE_OP_CONSTRUCTOR(CLS, PARENT_CLS)                                 \
+  CLS(const std::string& type, const VarNameMap& inputs,                       \
+      const VarNameMap& outputs, const paddle::framework::AttributeMap& attrs) \
+      : PARENT_CLS(type, inputs, outputs, attrs) {}
+
 class NOP : public OperatorBase {
  public:
   using OperatorBase::OperatorBase;
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {}
+  std::unique_ptr<OperatorBase> Clone() const override {
+    return std::unique_ptr<OperatorBase>(new NOP(*this));
+  }
+};
+
+// this class not only make proto but also init attribute checkers.
+class OpProtoAndCheckerMaker {
+ public:
+  OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : proto_(proto), op_checker_(op_checker) {}
+
+  ~OpProtoAndCheckerMaker() {
+    PADDLE_ENFORCE(validated_, "should call Validate after build");
+  }
+
+  void Validate();
+
+ protected:
+  struct VariableBuilder {
+    OpProto::Var* var_;
+
+    VariableBuilder& AsDuplicable() {
+      var_->set_duplicable(true);
+      return *this;
+    }
+
+    VariableBuilder& AsIntermediate() {
+      var_->set_intermediate(true);
+      return *this;
+    }
+
+    // TODO(FengJiayi, yuyang18): `AsNoGradient` is a very bad name, because it
+    // means that input/output is not needed when calculate gradient. It does
+    // not mean no gradient when backward. It should be changed soon.
+    VariableBuilder& AsNoGradient() {
+      var_->set_no_gradient(true);
+      return *this;
+    }
+  };
+
+  VariableBuilder AddInput(const std::string& name, const std::string& comment);
+
+  VariableBuilder AddOutput(const std::string& name,
+                            const std::string& comment);
+
+  template <typename T>
+  TypedAttrChecker<T>& AddAttr(const std::string& name,
+                               const std::string& comment,
+                               bool generated = false) {
+    auto* attr = proto_->add_attrs();
+    attr->set_name(name);
+    attr->set_comment(comment);
+    attr->set_generated(generated);
+    attr->set_type(AttrTypeID<T>());
+    return op_checker_->AddAttrChecker<T>(name);
+  }
+
+  void AddComment(const std::string& comment) { proto_->set_comment(comment); }
+
+ private:
+  void CheckNoDuplicatedInOutAttrs();
+
+  OpProto* proto_;
+  OpAttrChecker* op_checker_;
+  bool validated_{false};
+};
+
+class NOPMaker : public OpProtoAndCheckerMaker {
+ public:
+  NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {}
 };
 
 class InferShapeContext {

paddle/framework/operator_test.cc

@@ -245,3 +245,21 @@ TEST(OpKernel, multi_inputs) {
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   op->Run(scope, cpu_device_context);
 }
+
+class OperatorClone : public paddle::framework::OperatorBase {
+ public:
+  DEFINE_OP_CLONE_METHOD(OperatorClone);
+  OperatorClone(const std::string& type, const VarNameMap& inputs,
+                const VarNameMap& outputs,
+                const paddle::framework::AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+  void InferShape(const paddle::framework::Scope& scope) const override {}
+  void Run(const paddle::framework::Scope& scope,
+           const paddle::platform::DeviceContext& dev_ctx) const override {}
+};
+
+TEST(Operator, Clone) {
+  OperatorClone a("ABC", {}, {}, {});
+  auto b = a.Clone();
+  ASSERT_EQ(a.Type(), b->Type());
+}
\ No newline at end of file

paddle/framework/pybind.cc

@@ -48,29 +48,6 @@ namespace framework {
 
 using Tensor = framework::Tensor;
 
-template <typename ClassType>
-void ExposeOperator(ClassType &m) {
-  m.def("infer_shape", &ClassType::type::InferShape)
-      .def("run", &ClassType::type::Run)
-      .def("type",
-           [](const typename ClassType::type &op) -> std::string {
-             return op.Type();
-           })
-      .def("outputs",
-           [](const typename ClassType::type &op)
-               -> std::map<std::string, std::vector<std::string>> {
-                 return op.Outputs();
-               })
-      .def("inputs",
-           [](const typename ClassType::type &op) { return op.Inputs(); })
-      .def("__str__", &ClassType::type::DebugString)
-      .def("no_intermediate_outputs",
-           [](const typename ClassType::type &op) {
-             return op.OutputVars(false);
-           })
-      .def("support_gpu", &ClassType::type::SupportGPU);
-}
-
 static size_t UniqueIntegerGenerator() {
   static std::atomic<size_t> generator;
   return generator.fetch_add(1);
@@ -207,75 +184,69 @@ All parameter, weight, gradient are variables in Paddle.
       .def(py::init<>())
       .def("__str__", string::to_string<const platform::CPUPlace &>);
 
-  py::class_<OperatorBase, std::shared_ptr<OperatorBase>> operator_base(
-      m, "Operator");
-
-  operator_base.def_static("create", [](py::bytes protobin) {
-    OpDesc desc;
-    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
-                   "Cannot parse user input to OpDesc");
-    PADDLE_ENFORCE(desc.IsInitialized(),
-                   "User OpDesc is not initialized, reason %s",
-                   desc.InitializationErrorString());
-    return OpRegistry::CreateOp(desc);
-  });
-
-  operator_base.def("backward",
-                    [](const OperatorBase &forwardOp,
-                       const std::unordered_set<std::string> &no_grad_vars) {
-                      return Backward(forwardOp, no_grad_vars);
-                    });
-
-  ExposeOperator(operator_base);
-
-  py::class_<operators::NetOp, std::shared_ptr<operators::NetOp>> net(m, "Net");
-
-  net.def_static("create",
-                 []() -> std::shared_ptr<operators::NetOp> {
-                   auto retv = std::make_shared<operators::NetOp>();
-                   retv->SetType("plain_net");
-                   return retv;
-                 })
-      .def("add_op", &operators::NetOp::AddOp)
-      .def("add_op",
-           [](operators::NetOp &self,
-              const std::shared_ptr<operators::NetOp> &net) -> void {
-             self.AddOp(std::static_pointer_cast<OperatorBase>(net));
-           })
-      .def("add_op",
-           [](operators::NetOp &self,
-              const std::shared_ptr<operators::RecurrentOp> &rnn) -> void {
-             self.AddOp(std::static_pointer_cast<OperatorBase>(rnn));
+  py::class_<OperatorBase>(m, "Operator")
+      .def_static("create",
+                  [](py::bytes protobin) {
+                    OpDesc desc;
+                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                                   "Cannot parse user input to OpDesc");
+                    PADDLE_ENFORCE(desc.IsInitialized(),
+                                   "User OpDesc is not initialized, reason %s",
+                                   desc.InitializationErrorString());
+                    return OpRegistry::CreateOp(desc);
+                  })
+      .def("backward",
+           [](const OperatorBase &forwardOp,
+              const std::unordered_set<std::string> &no_grad_vars) {
+             return Backward(forwardOp, no_grad_vars).release();
            })
+      .def("infer_shape", &OperatorBase::InferShape)
+      .def("run", &OperatorBase::Run)
+      .def("type",
+           [](const OperatorBase &op) -> std::string { return op.Type(); })
+      .def("outputs",
+           [](const OperatorBase &op)
+               -> std::map<std::string, std::vector<std::string>> {
+                 return op.Outputs();
+               })
+      .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
+      .def("__str__", &OperatorBase::DebugString)
+      .def("no_intermediate_outputs",
+           [](const OperatorBase &op) { return op.OutputVars(false); })
+      .def("support_gpu", &OperatorBase::SupportGPU);
+
+  py::class_<operators::NetOp, OperatorBase>(m, "Net")
+      .def_static("create",
+                  []() -> operators::NetOp * {
+                    auto *retv = new operators::NetOp;
+                    retv->SetType("plain_net");
+                    return retv;
+                  })
+      .def("add_op", [](operators::NetOp &self,
+                        const OperatorBase &op) { self.AddOp(op); })
       .def("complete_add_op", &operators::NetOp::CompleteAddOp)
       .def("complete_add_op", [](std::shared_ptr<operators::NetOp> &self) {
         self->CompleteAddOp();
       });
 
-  ExposeOperator(net);
-
   // recurrent_op
-  py::class_<operators::RecurrentOp, std::shared_ptr<operators::RecurrentOp>>
-      rnn(m, "RecurrentOp");
-
-  rnn.def_static(
-         "create",
-         [](py::bytes protobin) -> std::shared_ptr<operators::RecurrentOp> {
-           OpDesc desc;
-           PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
-                          "Cannot parse user input to OpDesc");
-           PADDLE_ENFORCE(desc.IsInitialized(),
-                          "User OpDesc is not initialized, reason %s",
-                          desc.InitializationErrorString());
-           auto rnn_op = OpRegistry::CreateOp(desc);
-           return std::dynamic_pointer_cast<operators::RecurrentOp>(rnn_op);
-         })
-      .def("set_stepnet",
-           [](operators::RecurrentOp &self,
-              const std::shared_ptr<operators::NetOp> &net) -> void {
-             self.set_stepnet(net);
-           });
-  ExposeOperator(rnn);
+  py::class_<operators::RecurrentOp, OperatorBase>(m, "RecurrentOp")
+      .def_static(
+          "create",
+          [](py::bytes protobin) -> operators::RecurrentOp * {
+            OpDesc desc;
+            PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                           "Cannot parse user input to OpDesc");
+            PADDLE_ENFORCE(desc.IsInitialized(),
+                           "User OpDesc is not initialized, reason %s",
+                           desc.InitializationErrorString());
+            auto rnn_op = OpRegistry::CreateOp(desc);
+            return static_cast<operators::RecurrentOp *>(rnn_op.release());
+          })
+      .def("set_stepnet", [](operators::RecurrentOp &self,
+                             const operators::NetOp &net) -> void {
+        self.set_stepnet(net.Clone());
+      });
 
   m.def("unique_integer", UniqueIntegerGenerator);
 

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -57,11 +57,14 @@ bool MKLDNNFcLayer::init(const LayerMap& layerMap,
 }
 
 void MKLDNNFcLayer::convertWeightsFromPaddle() {
-  if (FLAGS_use_mkldnn_wgt) {
+  if (hasInitedWgt_) {
     return;
   }
 
-  if (hasInitedWgt_) {
+  // TODO(TJ): dst format should get from wgtVal_
+  int dstFmt = PARAM_FORMAT_MKLDNN_OI;
+  int srcFmt = weight_->getParameterPtr()->getHeaderFormat();
+  if (srcFmt == dstFmt) {
     return;
   }
 
@@ -78,6 +81,7 @@ void MKLDNNFcLayer::convertWeightsFromPaddle() {
   MatrixPtr paddleWgtT;
   paddleWgt->transpose(paddleWgtT, true);
   weight_->getW()->copyFrom(*paddleWgtT);
+  weight_->getParameterPtr()->setHeaderFormat(dstFmt);
   hasInitedWgt_ = true;
 }
 

paddle/gserver/tests/MKLDNNTester.cpp

@@ -330,9 +330,7 @@ void MKLDNNTester::run(const TestConfig& dnn,
   log_ = log;
   lvl_ = level;
 
-  // Firstly test FLAGS_use_mkldnn_wgt = false
-  FLAGS_use_mkldnn_wgt = false;
-  // reset and run once
+  // Firstly test mkldnn init from PARAM_FORMAT_ORIGINAL weight
   reset(dnn, ref, batchSize);
   randomWgtDatas();
   clearWgtDiffs();
@@ -342,17 +340,32 @@ void MKLDNNTester::run(const TestConfig& dnn,
     runOnce();
   }
 
-  // Then test FLAGS_use_mkldnn_wgt = true
-  FLAGS_use_mkldnn_wgt = true;
-  // after run once the mkldnn weight has been stored in dnnlayer
+  if (parameters_[DNN].empty()) {
+    // has no paramters
+    return;
+  }
+
+  // After run some iterations, the mkldnn weight has been stored in dnnLayer
+  // and we can also get the mkldnn weight parameter header format.
+  // Weight parameter should always be index 0 (and bias index 1).
+  // TODO(TJ): should also consider mean and var format when batchnorm ready
+  int dnnWgtFmt = parameters_[DNN][0]->getHeaderFormat();
+  int refWgtFmt = parameters_[REF][0]->getHeaderFormat();
+  if (dnnWgtFmt == refWgtFmt) {
+    // weight format are equal, so no need check more
+    return;
+  }
+
   // then save the weights and restart again
   vector<VectorPtr> dnnWgts, refWgts;
   CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size());
   saveWgt(parameters_[DNN], dnnWgts);
   saveWgt(parameters_[REF], refWgts);
 
-  // restart again with flag true
+  // restart again with dnn weight format
   reset(dnn, ref, batchSize);
+  // TODO(TJ): should also considerate mean and var format when batchnorm ready
+  parameters_[DNN][0]->setHeaderFormat(dnnWgtFmt);
 
   // restore wgt
   restoreWgt(dnnWgts, parameters_[DNN]);

paddle/gserver/tests/MKLDNNTester.h

@@ -108,7 +108,7 @@ private:
    * if many(>failRate) wrong(abs(dnn-ref)/abs(ref)>thres) points return the
    * max(diff/ref)
    * else return sum(abs(a-b)) / sum(abs(b))
-   * The return value should smaller than eps when passing.
+   * The return value should be smaller than eps when passing.
    */
   double getDelta(const real* d1,
                   const real* d2,

paddle/memory/CMakeLists.txt

 add_subdirectory(detail)
 
 cc_library(memory SRCS memory.cc)
-cc_library(memcpy SRCS memcpy.cc DEPS device_context)
+cc_library(memcpy SRCS memcpy.cc)
 
 cc_library(paddle_memory
     DEPS

paddle/memory/detail/system_allocator.cc

@@ -27,7 +27,7 @@ limitations under the License. */
 // between host and device.  Allocates too much would reduce the amount
 // of memory available to the system for paging.  So, by default, we
 // should set false to use_pinned_memory.
-DEFINE_bool(use_pinned_memory, false, "If set, allocate cpu pinned memory.");
+DEFINE_bool(use_pinned_memory, true, "If set, allocate cpu pinned memory.");
 
 namespace paddle {
 namespace memory {

paddle/memory/memcpy.cc

@@ -16,8 +16,6 @@ limitations under the License. */
 
 #include <cstring>  // for memcpy
 
-#include "paddle/platform/device_context.h"
-
 namespace paddle {
 namespace memory {
 

paddle/memory/memory.cc

@@ -13,22 +13,33 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/memory/memory.h"
+
+#include <algorithm>  // for transform
+#include <cstring>    // for memcpy
+#include <memory>     // for unique_ptr
+#include <mutex>      // for call_once
+
 #include "paddle/memory/detail/buddy_allocator.h"
 #include "paddle/memory/detail/system_allocator.h"
 
-#include <cstring>  // for memcpy
-
 namespace paddle {
 namespace memory {
 
-detail::BuddyAllocator* GetCPUBuddyAllocator() {
-  static detail::BuddyAllocator* a = nullptr;
-  if (a == nullptr) {
-    a = new detail::BuddyAllocator(new detail::CPUAllocator,
-                                   platform::CpuMinChunkSize(),
-                                   platform::CpuMaxChunkSize());
-  }
-  return a;
+using BuddyAllocator = detail::BuddyAllocator;
+
+std::once_flag cpu_allocator_flag;
+std::once_flag gpu_allocator_flag;
+
+BuddyAllocator* GetCPUBuddyAllocator() {
+  static std::unique_ptr<BuddyAllocator> a{nullptr};
+
+  std::call_once(cpu_allocator_flag, [&]() {
+    a.reset(new BuddyAllocator(new detail::CPUAllocator,
+                               platform::CpuMinChunkSize(),
+                               platform::CpuMaxChunkSize()));
+  });
+
+  return a.get();
 }
 
 template <>
@@ -48,20 +59,31 @@ size_t Used<platform::CPUPlace>(platform::CPUPlace place) {
 
 #ifndef PADDLE_ONLY_CPU
 
-detail::BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
-  static detail::BuddyAllocator** as = NULL;
-  if (as == NULL) {
+BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
+  using BuddyAllocVec = std::vector<BuddyAllocator*>;
+  static std::unique_ptr<BuddyAllocVec, void (*)(BuddyAllocVec * p)> as{
+      new BuddyAllocVec, [](BuddyAllocVec* p) {
+        std::for_each(p->begin(), p->end(),
+                      [](BuddyAllocator* p) { delete p; });
+      }};
+
+  // GPU buddy allocators
+  auto& allocators = *as.get();
+
+  // GPU buddy allocator initialization
+  std::call_once(gpu_allocator_flag, [&]() {
     int gpu_num = platform::GetDeviceCount();
-    as = new detail::BuddyAllocator*[gpu_num];
+    allocators.reserve(gpu_num);
     for (int gpu = 0; gpu < gpu_num; gpu++) {
       platform::SetDeviceId(gpu);
-      as[gpu] = new detail::BuddyAllocator(new detail::GPUAllocator,
-                                           platform::GpuMinChunkSize(),
-                                           platform::GpuMaxChunkSize());
+      allocators.emplace_back(new BuddyAllocator(new detail::GPUAllocator,
+                                                 platform::GpuMinChunkSize(),
+                                                 platform::GpuMaxChunkSize()));
     }
-  }
+  });
+
   platform::SetDeviceId(gpu_id);
-  return as[gpu_id];
+  return allocators[gpu_id];
 }
 
 template <>

paddle/operators/gather_test.cc

@@ -45,4 +45,8 @@ TEST(Gather, GatherData) {
 
   for (int i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4);
   for (int i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4);
+
+  delete src;
+  delete index;
+  delete output;
 }

paddle/operators/mean_op.h

@@ -55,9 +55,10 @@ class MeanGradKernel : public framework::OpKernel {
     IG->mutable_data<T>(context.GetPlace());
 
     T ig_size = (T)framework::product(IG->dims());
+    Eigen::DSizes<int, 1> bcast(ig_size);
 
     EigenVector<T>::Flatten(*IG).device(context.GetEigenDevice<Place>()) =
-        EigenScalar<T>::From(*OG) / ig_size;
+        (EigenVector<T>::From(*OG) / ig_size).broadcast(bcast);
   }
 };
 

paddle/operators/net_op.cc

@@ -85,7 +85,14 @@ NetOp::NetOp(const std::string& type,
              const framework::OperatorBase::VarNameMap& inputs,
              const framework::OperatorBase::VarNameMap& outputs,
              const framework::AttributeMap& attrs)
-    : OperatorBase(type, inputs, outputs, attrs) {}
+    : framework::OperatorBase(type, inputs, outputs, attrs) {}
+
+std::unique_ptr<framework::OperatorBase> NetOp::Clone() const {
+  PADDLE_ENFORCE(
+      add_op_done_,
+      "Must clone a sealed NetOp, invoke Net::CompleteAddOp before clone");
+  return std::unique_ptr<OperatorBase>(new NetOp(*this));
+}
 
 }  // namespace operators
 }  // namespace paddle

paddle/operators/net_op.h

@@ -41,6 +41,16 @@ class NetOp : public framework::OperatorBase {
   NetOp(const std::string& type, const VarNameMap& inputs,
         const VarNameMap& outputs, const framework::AttributeMap& attrs);
 
+  NetOp(const NetOp& o) : framework::OperatorBase(o.type_, {}, {}, o.attrs_) {
+    this->ops_.reserve(o.ops_.size());
+    std::transform(
+        o.ops_.begin(), o.ops_.end(), std::back_inserter(this->ops_),
+        [](const std::unique_ptr<framework::OperatorBase>& op) {
+          return std::unique_ptr<framework::OperatorBase>(op->Clone());
+        });
+    this->CompleteAddOp();
+  }
+
   /**
    * Infer all the operators' input and output variables' shapes, will be called
    * before every mini-batch
@@ -74,21 +84,27 @@ class NetOp : public framework::OperatorBase {
     return true;
   }
 
+  void AddOp(const framework::OperatorBase& op) { AddOp(op.Clone()); }
+
   /**
    * @brief Add an operator by ptr
    */
-  void AddOp(const std::shared_ptr<OperatorBase>& op) {
+  void AddOp(std::unique_ptr<framework::OperatorBase> op) {
     PADDLE_ENFORCE(!add_op_done_, "Cannot AddOp when this network is sealed");
     PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op");
-    ops_.push_back(op);
+    ops_.push_back(std::move(op));
   }
 
-  void InsertOp(size_t pos, const std::shared_ptr<OperatorBase>& op) {
+  void InsertOp(size_t pos, std::unique_ptr<framework::OperatorBase> op) {
     PADDLE_ENFORCE(!add_op_done_,
                    "Cannot InsertOp when this network is sealed");
     PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op");
     PADDLE_ENFORCE_LE(pos, ops_.size(), "Out of range");
-    ops_.insert(ops_.begin() + pos, op);
+    ops_.insert(ops_.begin() + pos, std::move(op));
+  }
+
+  void InsertOp(size_t pos, const framework::OperatorBase& op) {
+    InsertOp(pos, op.Clone());
   }
 
   void CompleteAddOp(bool calculate = true);
@@ -98,7 +114,9 @@ class NetOp : public framework::OperatorBase {
   bool IsNetOp() const override;
   std::vector<std::string> OutputVars(bool has_intermediate) const override;
 
-  std::vector<std::shared_ptr<OperatorBase>> ops_;
+  std::unique_ptr<framework::OperatorBase> Clone() const override;
+
+  std::vector<std::unique_ptr<framework::OperatorBase>> ops_;
 
  private:
   bool add_op_done_{false};

paddle/operators/net_op_test.cc

@@ -13,6 +13,7 @@ static int run_cnt = 0;
 class TestOp : public framework::OperatorBase {
  public:
   using framework::OperatorBase::OperatorBase;
+  DEFINE_OP_CLONE_METHOD(TestOp);
   void InferShape(const Scope& scope) const override { ++infer_shape_cnt; }
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {
@@ -37,15 +38,12 @@ TEST(OpKernel, all) {
   auto net = std::make_shared<NetOp>();
   ASSERT_NE(net, nullptr);
 
-  auto op1 = std::shared_ptr<TestOp>(
+  net->AddOp(std::unique_ptr<TestOp>(
       new TestOp("test", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
-                 {{"Out", {"y"}}}, {}));
-  net->AddOp(op1);
-
-  auto op2 = std::shared_ptr<TestOp>(
+                 {{"Out", {"y"}}}, {})));
+  net->AddOp(std::unique_ptr<TestOp>(
       new TestOp("test", {{"X", {"y"}}, {"W", {"w2"}}, {"b", {"b2"}}},
-                 {{"Out", {"z"}}}, {}));
-  net->AddOp(op2);
+                 {{"Out", {"z"}}}, {})));
 
   net->CompleteAddOp();
   AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"},
@@ -60,15 +58,31 @@ TEST(OpKernel, all) {
 
 TEST(NetOp, insert_op) {
   NetOp net;
-  auto op1 = std::shared_ptr<framework::NOP>(
+  auto op1 = std::unique_ptr<framework::NOP>(
       new framework::NOP("empty", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
                          {{"Out", {"y"}}}, {}));
-  net.AddOp(op1);
-  net.InsertOp(0, op1);
+  net.AddOp(*op1);
+  net.InsertOp(0, *op1);
   ASSERT_EQ(2UL, net.ops_.size());
-  net.InsertOp(2, op1);
+  net.InsertOp(2, std::move(op1));
   ASSERT_EQ(3UL, net.ops_.size());
 }
 
+TEST(NetOp, Clone) {
+  NetOp net;
+  net.AddOp(
+      std::unique_ptr<framework::NOP>(new framework::NOP{"empty", {}, {}, {}}));
+  net.AddOp(std::unique_ptr<framework::NOP>(
+      new framework::NOP{"empty2", {}, {}, {}}));
+  net.CompleteAddOp(true);
+  auto new_net_op = net.Clone();
+  ASSERT_NE(new_net_op, nullptr);
+  ASSERT_TRUE(new_net_op->IsNetOp());
+  auto* new_net = static_cast<NetOp*>(new_net_op.get());
+  ASSERT_EQ(2, new_net->ops_.size());
+  ASSERT_EQ(new_net->ops_[0]->Type(), "empty");
+  ASSERT_EQ(new_net->ops_[1]->Type(), "empty2");
+}
+
 }  // namespace operators
 }  // namespace paddle

paddle/operators/recurrent_op.h

@@ -34,7 +34,8 @@ class RecurrentAlgorithm {
   void Run(const framework::Scope& scope,
            const platform::DeviceContext& dev_ctx) const;
 
-  void Init(rnn::Argument* arg, std::shared_ptr<NetOp>* stepnet) {
+  void Init(rnn::Argument* arg,
+            std::unique_ptr<framework::OperatorBase>* stepnet) {
     PADDLE_ENFORCE_NOT_NULL(stepnet, "stepnet should be set before.");
     arg_ = arg;
     stepnet_ = stepnet;
@@ -63,7 +64,7 @@ class RecurrentAlgorithm {
   void InitMemories(framework::Scope* step_scopes, bool infer_shape_mode) const;
 
  private:
-  std::shared_ptr<NetOp>* stepnet_;
+  std::unique_ptr<framework::OperatorBase>* stepnet_;
   rnn::Argument* arg_;
   mutable size_t seq_len_;
 };
@@ -80,7 +81,8 @@ class RecurrentGradientAlgorithm {
    * operator.
    */
  public:
-  void Init(rnn::Argument* arg, std::shared_ptr<NetOp>* stepnet) {
+  void Init(rnn::Argument* arg,
+            std::unique_ptr<framework::OperatorBase>* stepnet) {
     PADDLE_ENFORCE_NOT_NULL(stepnet, "stepnet should be set before.");
     arg_ = std::move(arg);
     stepnet_ = stepnet;
@@ -107,16 +109,23 @@ class RecurrentGradientAlgorithm {
  private:
   rnn::Argument* arg_;
   mutable size_t seq_len_;
-  std::shared_ptr<NetOp>* stepnet_;
+  std::unique_ptr<framework::OperatorBase>* stepnet_;
 };
 
-class RecurrentOp final : public framework::OperatorBase {
+class RecurrentOp : public framework::OperatorBase {
  public:
   RecurrentOp(const std::string& type, const VarNameMap& inputs,
               const VarNameMap& outputs, const framework::AttributeMap& attrs);
+
+  RecurrentOp(const RecurrentOp& o)
+      : framework::OperatorBase(
+            static_cast<const framework::OperatorBase&>(o)) {
+    // TODO(yuyang18): Implement copy ctor well.
+    PADDLE_THROW("Not implemented");
+  }
   /**
-     * InferShape must be called before Run.
-     */
+   * InferShape must be called before Run.
+   */
   void InferShape(const framework::Scope& scope) const override {
     alg_.InferShape(scope);
   }
@@ -126,23 +135,32 @@ class RecurrentOp final : public framework::OperatorBase {
     alg_.Run(scope, dev_ctx);
   }
 
-  void set_stepnet(std::shared_ptr<NetOp> net) { stepnet_ = net; }
-  const NetOp* stepnet() const { return stepnet_.get(); }
+  void set_stepnet(std::unique_ptr<OperatorBase> net) {
+    stepnet_ = std::move(net);
+  }
+  const OperatorBase& stepnet() const { return *stepnet_; }
 
   static const rnn::ArgumentName kArgName;
 
  private:
   RecurrentAlgorithm alg_;
   rnn::Argument arg_;
-  std::shared_ptr<NetOp> stepnet_;
+  std::unique_ptr<OperatorBase> stepnet_;
 };
 
-class RecurrentGradientOp final : public framework::OperatorBase {
+class RecurrentGradientOp : public framework::OperatorBase {
  public:
   RecurrentGradientOp(const std::string& type, const VarNameMap& inputs,
                       const VarNameMap& outputs,
                       const framework::AttributeMap& attrs);
 
+  RecurrentGradientOp(const RecurrentGradientOp& o)
+      : framework::OperatorBase(
+            static_cast<const framework::OperatorBase&>(o)) {
+    // TODO(yuyang18): Implement Copy ctor.
+    PADDLE_THROW("Not Implemented");
+  }
+
   /**
    * InferShape must be called before Run.
    */
@@ -157,12 +175,14 @@ class RecurrentGradientOp final : public framework::OperatorBase {
 
   static const rnn::ArgumentName kArgName;
 
-  void set_stepnet(const std::shared_ptr<NetOp>& net) { stepnet_ = net; }
-  const NetOp* stepnet() const { return stepnet_.get(); }
+  void set_stepnet(std::unique_ptr<OperatorBase> net) {
+    stepnet_ = std::move(net);
+  }
+  const OperatorBase& stepnet() const { return *stepnet_; }
 
  private:
   RecurrentGradientAlgorithm alg_;
-  std::shared_ptr<NetOp> stepnet_;
+  std::unique_ptr<OperatorBase> stepnet_;
   rnn::Argument arg_;
 };
 

paddle/operators/scatter_test.cc

@@ -49,4 +49,8 @@ TEST(scatter, ScatterUpdate) {
     EXPECT_EQ(output->data<float>()[i], float(i - 4));
   for (size_t i = 8; i < 16; ++i) EXPECT_EQ(p_output[i], float(0));
   for (size_t i = 8; i < 16; ++i) EXPECT_EQ(output->data<float>()[i], float(0));
+
+  delete src;
+  delete index;
+  delete output;
 }

paddle/operators/sigmoid_op.cc

@@ -44,7 +44,8 @@ class SigmoidOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    ctx.Output<Tensor>(framework::GradVarName("X"))
+        ->Resize(ctx.Input<Tensor>("Y")->dims());
   }
 };
 

paddle/operators/sigmoid_op.h

@@ -37,7 +37,7 @@ class SigmoidKernel : public framework::OpKernel {
     auto Y = EigenVector<T>::Flatten(*output);
     auto place = context.GetEigenDevice<Place>();
 
-    Y.device(place) = 1.0 / (1.0 + (-1.0 * X).exp());
+    Y.device(place) = 1. / (1. + (-X).exp());
   }
 };
 

paddle/parameter/Parameter.cpp

@@ -48,7 +48,8 @@ Parameter::Parameter(const ParameterConfig& config, bool useGpu, bool doInit)
       deviceId_(-1),
       sharedCount_(0),
       updateCounter_(0),
-      updated_(false) {
+      updated_(false),
+      headerFormat_(PARAM_FORMAT_ORIGINAL) {
   setID(-1); /* capture uninitialized id */
   if (useGpu_ && FLAGS_parallel_nn) {
     /* gpu environment is specified by device property */
@@ -285,7 +286,7 @@ bool Parameter::save(const std::string& filename) const {
 bool Parameter::save(std::ostream& s) const {
   CpuVector vec(*bufs_[PARAMETER_VALUE].get());
   Header header;
-  header.version = kFormatVersion;
+  header.format = headerFormat_;
   header.valueSize = sizeof(real);
   header.size = getSize();
 
@@ -344,8 +345,9 @@ bool Parameter::load(std::istream& s) {
   Header header;
   CHECK(s.read(reinterpret_cast<char*>(&header), sizeof(header)))
       << "Fail to read parameter " << getName();
-  CHECK_EQ(header.version, kFormatVersion) << "Incorrect format version: "
-                                           << header.version;
+  CHECK(isHeaderFormatSupported(header.format)) << "Incorrect format version: "
+                                                << header.format;
+  headerFormat_ = header.format;
   CHECK_EQ(header.size, getSize())
       << "The size (" << header.size << ") in the file does not match the size "
       << "(" << getSize() << ") of the parameter: " << getName();

paddle/parameter/Parameter.h

@@ -34,6 +34,20 @@ limitations under the License. */
 
 namespace paddle {
 
+typedef enum {
+  /// The paddle original basic format
+  PARAM_FORMAT_ORIGINAL = 0,
+
+  /// See mkldnn_memory_format_t in
+  /// https://github.com/01org/mkl-dnn/blob/master/include/mkldnn_types.h
+  /// for a detailed description.
+  /// 2D weights tensor in the format (output channels, input channels).
+  PARAM_FORMAT_MKLDNN_OI,
+
+  /// The total format items numbers
+  PARAM_FORMAT_ITEMS,
+} PARAM_FORMAT;
+
 class SparsePrefetchRowCpuMatrix;
 
 class Parameter;
@@ -242,14 +256,30 @@ public:
   /// Initialize the value to 0
   void zeroMem();
 
-  static const int kFormatVersion = 0;
   /// file header structure
   struct Header {
-    int32_t version;     // = 0, file format version
+    int32_t format;      // = PARAM_FORMAT
     uint32_t valueSize;  // = sizeof(real)
     uint64_t size;       // = getSize()
   };
 
+  /**
+   * @brief Is the header format supported.
+   */
+  static bool isHeaderFormatSupported(int32_t fmt) {
+    return fmt < PARAM_FORMAT_ITEMS;
+  }
+
+  /**
+   * @brief Get the format in header.
+   */
+  int getHeaderFormat() { return headerFormat_; }
+
+  /**
+   * @brief Set the format in header.
+   */
+  void setHeaderFormat(int32_t fmt) { headerFormat_ = fmt; }
+
   /**
    * @brief  Parameter Update Hook.
    *
@@ -321,6 +351,9 @@ protected:
   bool updated_;
   SparseFormat format_;
 
+  /// The header format for saving or loading param
+  int32_t headerFormat_;
+
   std::vector<std::shared_ptr<IParameterUpdaterHook>> updaterHooks_;
 
 public:

paddle/platform/CMakeLists.txt

@@ -16,5 +16,8 @@ ELSE()
     set(GPU_CTX_DEPS)
 ENDIF()
 
-cc_library(device_context SRCS device_context.cc DEPS place eigen3 ${GPU_CTX_DEPS})
+# 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})
 nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info)

paddle/platform/device_context.cc

@@ -10,6 +10,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/platform/device_context.h"
+#include "paddle/memory/memory.h"
 
 namespace paddle {
 namespace platform {
@@ -36,6 +37,59 @@ Place CPUDeviceContext::GetPlace() const { return CPUPlace(); }
 
 #ifndef PADDLE_ONLY_CPU
 
+class EigenCudaStreamDevice : public Eigen::StreamInterface {
+ public:
+  EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) {
+    Eigen::initializeDeviceProp();
+  }
+  ~EigenCudaStreamDevice() override {}
+
+  void Reinitialize(const cudaStream_t* cuda_stream, GPUPlace place) {
+    stream_ = cuda_stream;
+    place_ = place;
+    device_prop_ = &Eigen::m_deviceProperties[place.device];
+  }
+
+  const cudaStream_t& stream() const override { return *stream_; }
+
+  const cudaDeviceProp& deviceProperties() const override {
+    return *device_prop_;
+  }
+
+  void* allocate(size_t num_bytes) const override {
+    return paddle::memory::Alloc(place_, num_bytes);
+  }
+
+  void deallocate(void* buffer) const override {
+    paddle::memory::Free(place_, buffer);
+  }
+
+  void* scratchpad() const override {
+    if (scratch_ == NULL) {
+      scratch_ = allocate(Eigen::kCudaScratchSize + sizeof(unsigned int));
+    }
+    return scratch_;
+  }
+
+  unsigned int* semaphore() const override {
+    if (semaphore_ == NULL) {
+      char* scratch =
+          static_cast<char*>(scratchpad()) + Eigen::kCudaScratchSize;
+      semaphore_ = reinterpret_cast<unsigned int*>(scratch);
+      PADDLE_ENFORCE(
+          cudaMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_));
+    }
+    return semaphore_;
+  }
+
+ private:
+  GPUPlace place_;
+  const cudaStream_t* stream_;         // not owned;
+  const cudaDeviceProp* device_prop_;  // not owned;
+  mutable void* scratch_;
+  mutable unsigned int* semaphore_;
+};
+
 template <>
 Eigen::GpuDevice* DeviceContext::get_eigen_device<Eigen::GpuDevice>() const {
   return reinterpret_cast<const CUDADeviceContext*>(this)->eigen_device();
@@ -43,19 +97,9 @@ Eigen::GpuDevice* DeviceContext::get_eigen_device<Eigen::GpuDevice>() const {
 
 CUDADeviceContext::CUDADeviceContext(GPUPlace place) : place_(place) {
   SetDeviceId(place_.device);
-  // TODO(qijun) Pass a created cuda stream to Eigen::CudaStreamDevice directly
-  // here will cause segment fault. We must implement a class derived from
-  // Eigen::StreamInterface, and reinitialize it with a cuda stream and a gpu id
-  // later. Please refer to the implementation of class EigenCudaStreamDevice
-  // in TensorFlow.
-  //
-  // We find that CUDA 7 introduces a new option, the per-thread default stream,
-  // that has two effects. Please refer to https://devblogs.nvidia.com/
-  // parallelforall/gpu-pro-tip-cuda-7-streams-simplify-concurrency/
-  //
-  // So, we decide to use default stream and add –default-stream per-thread nvcc
-  // flag. Than, two threads with two CUDADeviceContexts will run parallelly.
-  eigen_stream_.reset(new Eigen::CudaStreamDevice());
+  PADDLE_ENFORCE(cudaStreamCreate(&stream_));
+  eigen_stream_.reset(new EigenCudaStreamDevice());
+  eigen_stream_->Reinitialize(&stream_, place);
   eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
 }
 
@@ -75,12 +119,13 @@ CUDADeviceContext::~CUDADeviceContext() {
   }
   eigen_stream_.reset();
   eigen_device_.reset();
+  PADDLE_ENFORCE(cudaStreamDestroy(stream_));
 }
 
 Place CUDADeviceContext::GetPlace() const { return place_; }
 
 void CUDADeviceContext::Wait() const {
-  PADDLE_ENFORCE(cudaStreamSynchronize(0));
+  PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
 }
 
 Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
@@ -91,6 +136,7 @@ cublasHandle_t CUDADeviceContext::cublas_handle() {
   if (!cublas_handle_) {
     SetDeviceId(place_.device);
     PADDLE_ENFORCE(dynload::cublasCreate(&cublas_handle_));
+    PADDLE_ENFORCE(dynload::cublasSetStream(cublas_handle_, stream_));
   }
   return cublas_handle_;
 }
@@ -99,10 +145,13 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() {
   if (!cudnn_handle_) {
     SetDeviceId(place_.device);
     PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
+    PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream_));
   }
   return cudnn_handle_;
 }
 
+cudaStream_t CUDADeviceContext::stream() { return stream_; }
+
 curandGenerator_t CUDADeviceContext::curand_generator() {
   if (!curand_generator_) {
     SetDeviceId(place_.device);
@@ -110,6 +159,8 @@ curandGenerator_t CUDADeviceContext::curand_generator() {
                                                   CURAND_RNG_PSEUDO_DEFAULT));
     PADDLE_ENFORCE(
         dynload::curandSetPseudoRandomGeneratorSeed(curand_generator_, seed_));
+
+    PADDLE_ENFORCE(dynload::curandSetStream(curand_generator_, stream_));
   }
   return curand_generator_;
 }

paddle/platform/device_context.h

@@ -52,6 +52,7 @@ class CPUDeviceContext : public DeviceContext {
 };
 
 #ifndef PADDLE_ONLY_CPU
+class EigenCudaStreamDevice;
 
 class CUDADeviceContext : public DeviceContext {
  public:
@@ -76,6 +77,9 @@ class CUDADeviceContext : public DeviceContext {
 
   /*! \brief  Return curand handle in the device context. */
   curandGenerator_t curand_generator();
+
+  /*! \brief  Return cuda stream in the device context. */
+  cudaStream_t      stream();
   // clang-format on
 
  private:
@@ -83,15 +87,16 @@ class CUDADeviceContext : public DeviceContext {
 
  private:
   std::unique_ptr<Eigen::GpuDevice> eigen_device_;
-  std::unique_ptr<Eigen::CudaStreamDevice> eigen_stream_;
+  std::unique_ptr<EigenCudaStreamDevice> eigen_stream_;
 
  private:
   uint64_t seed_;
 
   // clang-format off
-  cudnnHandle_t     cudnn_handle_     = nullptr;
-  cublasHandle_t    cublas_handle_    = nullptr;
-  curandGenerator_t curand_generator_ = nullptr;
+  cudaStream_t       stream_{nullptr};
+  cudnnHandle_t      cudnn_handle_{nullptr};
+  cublasHandle_t     cublas_handle_{nullptr};
+  curandGenerator_t  curand_generator_{nullptr};
   // clang-format on
 };
 

paddle/platform/device_context_test.cc

@@ -45,6 +45,7 @@ TEST(Device, CUDADeviceContext) {
     ASSERT_NE(nullptr, cublas_handle);
     curandGenerator_t curand_handle = device_context->curand_generator();
     ASSERT_NE(nullptr, curand_handle);
+    ASSERT_NE(nullptr, device_context->stream());
     delete device_context;
   }
 }

paddle/platform/enforce.h

@@ -86,7 +86,7 @@ struct EnforceNotMet : public std::exception {
                                   2 + sizeof(void*) * 2, call_stack[i],
                                   demangled, addr_offset);
         } else {
-          sout << string::Sprintf("%-3d %*0p %s\n", i, 2 + sizeof(void*) * 2,
+          sout << string::Sprintf("%-3d %*0p\n", i, 2 + sizeof(void*) * 2,
                                   call_stack[i]);
         }
       }

paddle/pserver/ParameterServer2.cpp

@@ -1032,8 +1032,8 @@ void ParameterServer2::loadValueVector(const LoadValueRequest& request,
   Parameter::Header header;
   CHECK(fs.read(reinterpret_cast<char*>(&header), sizeof(header)))
       << "Fail to read parameters in pserver";
-  CHECK_EQ(header.version, Parameter::kFormatVersion)
-      << "Incorrect format version: " << header.version;
+  CHECK(Parameter::isHeaderFormatSupported(header.format))
+      << "Incorrect format version: " << header.format;
   CHECK_EQ(header.size, (size_t)size_)
       << "The size (" << header.size << ") in the file does not match the size "
       << "(" << size_ << ") of the pserver: " << serverId_;
@@ -1063,7 +1063,8 @@ void ParameterServer2::saveValueVector(const SaveValueRequest& request,
   CpuVector& vec = vectors_[PARAMETER_APPLY] ? *vectors_[PARAMETER_APPLY]
                                              : *vectors_[PARAMETER_VALUE];
   Parameter::Header header;
-  header.version = Parameter::kFormatVersion;
+  // TODO(TJ): save param headerFormat_
+  header.format = PARAM_FORMAT_ORIGINAL;
   header.valueSize = sizeof(real);
   header.size = size_;
 

paddle/scripts/docker/build.sh

@@ -82,10 +82,6 @@ EOF
 fi
 
 
-# To build documentation, we need to run cmake again after installing
-# PaddlePaddle.  This awkwardness is due to
-# https://github.com/PaddlePaddle/Paddle/issues/1854.  It also
-# describes a solution.
 if [[ ${WITH_DOC:-OFF} == "ON" ]]; then
     cat <<EOF
 ========================================
@@ -93,11 +89,6 @@ Building documentation ...
    In /paddle/build_doc
 ========================================
 EOF
-    # build documentation need install Paddle before
-    make install -j `nproc`
-    pip install /usr/local/opt/paddle/share/wheels/*.whl
-    paddle version
-
     mkdir -p /paddle/build_doc
     pushd /paddle/build_doc
     cmake .. \
@@ -106,7 +97,8 @@ EOF
           -DWITH_AVX=${WITH_AVX:-ON} \
           -DWITH_SWIG_PY=ON \
           -DWITH_STYLE_CHECK=OFF
-    make paddle_docs paddle_docs_cn
+    make -j `nproc` gen_proto_py
+    make -j `nproc` paddle_docs paddle_docs_cn
     popd
 fi
 
@@ -128,25 +120,6 @@ EOF
     /woboq/indexgenerator/codebrowser_indexgenerator $WOBOQ_OUT
 fi
 
-# generate deb package for current build
-# FIXME(typhoonzero): should we remove paddle/scripts/deb ?
-if [[ ${WITH_DEB:-ON} == "ON" ]]; then
-    cat <<EOF
-========================================
-Generating .deb package ...
-========================================
-EOF
-    set +e
-    cpack -D CPACK_GENERATOR='DEB' -j `nproc` ..
-    err_code=$?
-    if [ ${err_code} -ne 0 ]; then
-        # cat error logs if cpack failed.
-        cat /paddle/build/_CPack_Packages/Linux/DEB/PreinstallOutput.log
-        exit ${err_code}
-    fi
-    set -e
-fi
-
 cat <<EOF
 ========================================
 Generate /paddle/build/Dockerfile ...
@@ -166,15 +139,15 @@ EOF
 fi
 
 cat >> /paddle/build/Dockerfile <<EOF
-# Use different deb file when building different type of images
-ADD *.deb /
+ADD python/dist/*.whl /
 # run paddle version to install python packages first
 RUN apt-get update &&\
     apt-get install -y wget python-pip && pip install -U pip && \
-    dpkg -i /*.deb ; apt-get install -f -y && \
+    pip install /*.whl; apt-get install -f -y && \
     apt-get clean -y && \
-    rm -f /*.deb && \
-    paddle version
+    rm -f /*.whl && \
+    paddle version && \
+    ldconfig
 ${DOCKERFILE_CUDNN_DSO}
 ${DOCKERFILE_GPU_ENV}
 ADD go/cmd/pserver/pserver /usr/bin/
@@ -182,3 +155,7 @@ ADD go/cmd/master/master /usr/bin/
 # default command shows the paddle version and exit
 CMD ["paddle", "version"]
 EOF
+
+set +xe
+printf "If you need to install PaddlePaddle in develop docker image,"
+printf "please make install or pip install build/python/dist/*.whl.\n"

paddle/scripts/submit_local.sh.in

@@ -56,8 +56,7 @@ if [ -z "${PADDLE_NO_STAT+x}" ]; then
     fi
 fi
 
-
-MYDIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
+PADDLE_BIN_PATH="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
 
 if [ ! -z "${DEBUGGER}" ]; then
     echo "Using debug command ${DEBUGGER}"
@@ -93,34 +92,16 @@ else:
   sys.exit(0)
 EOF
 
-if [ $? -eq 1 ]; then  # Older version installed, or not installed at all
-    echo "First time run paddle, need to install some python dependencies."
-    # setuptools normalizes package version, so we need to use normalized
-    # package version for paddle python package
-    PYTHON_PADDLE_VERSION=$(python -c 'import packaging.version
-import setuptools
-print str(packaging.version.Version("@PADDLE_VERSION@"))
-' 2>/dev/null)
-    BASEDIR=$(dirname "$0")
-    pip install ${BASEDIR}/../opt/paddle/share/wheels/*-${PYTHON_PADDLE_VERSION}-*.whl
-    if [ $? -ne 0 ]; then
-	echo "pip install wheels failed. "
-	echo "Please use 'sudo paddle' at the first time you use PaddlePaddle"
-	echo "PaddlePaddle will install some python dependencies automatically."
-	exit 1
-    fi
-    echo "Python dependencies are installed."
-fi
 
 case "$1" in
     "train")
-        ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_trainer ${@:2}
+        ${DEBUGGER} $PADDLE_BIN_PATH/paddle_trainer ${@:2}
         ;;
     "merge_model")
-        ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_merge_model ${@:2}
+        ${DEBUGGER} $PADDLE_BIN_PATH/paddle_merge_model ${@:2}
         ;;
     "pserver")
-        ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_pserver_main ${@:2}
+        ${DEBUGGER} $PADDLE_BIN_PATH/paddle_pserver_main ${@:2}
         ;;
     "dump_config")
         python -m paddle.utils.dump_config ${@:2}
@@ -129,7 +110,7 @@ case "$1" in
         python -m paddle.utils.make_model_diagram ${@:2}
         ;;
     "usage")
-        $MYDIR/../opt/paddle/bin/paddle_usage ${@:2}
+        $PADDLE_BIN_PATH/paddle_usage ${@:2}
         ;;
     "version")
         version

paddle/trainer/TrainerConfigHelper.cpp

@@ -29,7 +29,6 @@ DECLARE_bool(with_gpu);
 DECLARE_bool(parallel_nn);
 DECLARE_string(config_args);
 DECLARE_bool(use_mkldnn);
-DECLARE_bool(use_mkldnn_wgt);
 
 const char *kConfigParserModuleName = "paddle.trainer.config_parser";
 const char *kConfigParserFuncName = "parse_config_and_serialize";
@@ -47,7 +46,6 @@ TrainerConfigHelper::TrainerConfigHelper(const std::string &configFilePath)
              << ",with_cost=" << FLAGS_with_cost << ",use_gpu=" << FLAGS_use_gpu
              << ",parallel_nn=" << FLAGS_parallel_nn
              << ",use_mkldnn=" << FLAGS_use_mkldnn
-             << ",use_mkldnn_wgt=" << FLAGS_use_mkldnn_wgt
              << ",cudnn_version=" << hl_get_cudnn_lib_version();
   if (!FLAGS_config_args.empty()) {
     configArgs << "," << FLAGS_config_args;

paddle/utils/Flags.cpp

@@ -27,7 +27,6 @@ DEFINE_bool(use_mkldnn, false, "Default still keep use CPU training");
 DEFINE_bool(use_mkldnn, false, "Only support CPU training");
 #endif
 
-DEFINE_bool(use_mkldnn_wgt, false, "Init weight from CPU weight");
 DEFINE_bool(parallel_nn,
             false,
             "Whether to use multi-threads to calculate one neural network."

paddle/utils/Flags.h

@@ -41,4 +41,3 @@ DECLARE_string(predict_file);
 DECLARE_bool(prev_batch_state);
 DECLARE_string(init_model_path);
 DECLARE_bool(use_mkldnn);
-DECLARE_bool(use_mkldnn_wgt);

python/CMakeLists.txt

@@ -50,8 +50,11 @@ add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
     COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_PYTHON_BUILD_DIR}/lib* ${PADDLE_PYTHON_BUILD_DIR}/lib-python
     DEPENDS gen_proto_py copy_paddle_pybind framework_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
 
-add_custom_target(paddle_python ALL DEPENDS
-    ${PADDLE_PYTHON_BUILD_DIR}/.timestamp paddle_pserver_main paddle_trainer paddle_merge_model python_api_wheel ${MKL_DEPENDS})
+set(paddle_python_deps ${PADDLE_PYTHON_BUILD_DIR}/.timestamp paddle_pserver_main paddle_trainer paddle_merge_model ${MKL_DEPENDS})
+if(WITH_SWIG_PY)
+    list(APPEND paddle_python_deps python_api_wheel)
+endif()
+add_custom_target(paddle_python ALL DEPENDS ${paddle_python_deps})
 
 set(PADDLE_PYTHON_PACKAGE_DIR ${CMAKE_CURRENT_BINARY_DIR}/dist/)
 

python/paddle/v2/framework/tests/CMakeLists.txt

@@ -25,3 +25,4 @@ py_test(test_operator SRCS test_operator.py)
 # py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py)
 py_test(test_uniform_random_op SRCS test_uniform_random_op.py)
 py_test(test_recurrent_op SRCS test_recurrent_op.py)
+py_test(test_gradient_checker SRCS test_gradient_checker.py)

python/paddle/v2/framework/tests/gradient_checker.py

 import unittest
 
 import numpy
+import itertools
 import paddle.v2.framework.core as core
 from paddle.v2.framework.op import Operator
 
@@ -8,6 +9,7 @@ __all__ = ['get_numeric_gradient']
 
 
 def create_op(op_type):
+    # TODO need to set attrs
     kwargs = dict()
     for in_name in Operator.get_op_input_names(op_type):
         kwargs[in_name] = in_name
@@ -66,7 +68,6 @@ def get_numeric_gradient(op,
             local_scope.find_var(output).get_tensor().alloc_float(core.CPUPlace(
             ))
 
-    # TODO(yuyang18): Only CPU is support now.
     cpu_ctx = core.DeviceContext.create(core.CPUPlace())
 
     def get_output():
@@ -109,12 +110,110 @@ def get_numeric_gradient(op,
 
 
 class GradientChecker(unittest.TestCase):
-    def assert_is_close(self, numeric_grads, scope, max_relative_error,
-                        msg_prefix):
-        for name in numeric_grads:
-            b = numpy.array(scope.find_var(grad_var_name(name)).get_tensor())
-            a = numeric_grads[name]
+    def __get_gradient(self, forward_op, backward_op, input_value, grad_names,
+                       place):
+        """Get the input gradients after running forward and backward operators
+        on the given places.
+
+        :param forward_op: forward operator
+        :type forward_op: Operator
+        :param backward_op: backward operator
+        :type backward_op: Operator
+        :param input_value: input values.
+        :type input_value: dict{string:numpy.array}
+        :param grad_names: the names of returned input gradients.
+        :type input_value: a list of string
+        :param place: the device type.
+        :type place: CPUPlace or GPUPlace
+        :return: the input grdients of given grad_names.
+        :rtype: a list of numpy.array
+        """
+        scope = core.Scope()
+        ctx = core.DeviceContext.create(place)
+
+        inputs = forward_op.inputs()
+        in_names = [item for k in inputs for item in inputs[k]]
+        outputs = forward_op.outputs()
+        out_names = [item for k in outputs for item in outputs[k]]
+
+        # create input var and set value
+        for name, value in input_value.iteritems():
+            if name not in in_names:
+                raise ValueError(name + "does not exist in Op's inputs.")
+            var = scope.new_var(name).get_tensor()
+            var.set_dims(value.shape)
+            var.set(value, place)
+
+        # run forward op
+        for out_name in out_names:
+            scope.new_var(out_name)
+        forward_op.infer_shape(scope)
+        forward_op.run(scope, ctx)
+
+        # set output var's shape
+        # set output grad to ones
+        for name in out_names:
+            out_tensor = scope.find_var(name).get_tensor()
+            grad_tensor = scope.new_var(grad_var_name(name)).get_tensor()
+            grad_tensor.set_dims(out_tensor.shape())
+            data = numpy.ones(out_tensor.shape(), dtype=numpy.float32)
+            grad_tensor.set(data, place)
+
+        # run backward op
+        for name in backward_op.outputs():
+            scope.new_var(name)
+        backward_op.infer_shape(scope)
+        backward_op.run(scope, ctx)
+
+        outs = [
+            numpy.array(scope.find_var(name).get_tensor())
+            for name in grad_names
+        ]
+        return outs
+
+    def compare_grad(self, forward_op, input_value):
+        """ Compare the input gradients between CPU and GPU for the given forward
+        operator.
+
+        :param forward_op: forward operator
+        :type forward_op: Operator
+        :param input_value: input values.
+        :type input_value: dict{string:numpy.array}
+        :raises: AssertionError, there is different gradient value.
+        """
+        backward_op = core.Operator.backward(forward_op, set())
+        # return if not compile with GPU or not implementing GPU kernel
+        if not (core.is_compile_gpu() and backward_op.support_gpu()):
+            return
 
+        outputs = backward_op.outputs()
+        out_names = [item for k in outputs for item in outputs[k]]
+        cpu_grads = self.__get_gradient(forward_op, backward_op, input_value,
+                                        out_names, core.CPUPlace())
+        gpu_grads = self.__get_gradient(forward_op, backward_op, input_value,
+                                        out_names, core.GPUPlace(0))
+
+        for c_grad, g_grad, name in itertools.izip(cpu_grads, gpu_grads,
+                                                   out_names):
+            self.assertTrue(
+                numpy.allclose(
+                    c_grad, g_grad, atol=1e-4),
+                "output name: " + name + " has diff")
+
+    def __assert_is_close(self, numeric_grads, analytic_grads, names,
+                          max_relative_error, msg_prefix):
+        """Use relative error for the comparison.
+
+        :param numeric_grads: the numerical graidents.
+        :type numeric_grads: a list of numpy.array 
+        :param analytic_grads: the analytical graidents.
+        :type analytic_grads: a list of numpy.array 
+        :param name: the names of gradients, used to print for debug.
+        :type names: a list of string
+        :param msg_prefix: string info, used to print for debug.
+        :type msf_prefix: string
+        """
+        for a, b, name in itertools.izip(numeric_grads, analytic_grads, names):
             abs_a = numpy.abs(a)
             # if abs_a is nearly zero, then use abs error for a, not relative
             # error.
@@ -159,106 +258,26 @@ class GradientChecker(unittest.TestCase):
 
         inputs = forward_op.inputs()
         in_names = [item for k in inputs for item in inputs[k]]
-        outputs = forward_op.outputs()
-        out_names = [item for k in outputs for item in outputs[k]]
-
         for no_grad in no_grad_set:
             if no_grad not in in_names:
                 raise ValueError("no_grad should be in in_names")
-
         backward_op = core.Operator.backward(forward_op, no_grad_set)
 
-        bwd_outputs = backward_op.outputs()
-        bwd_out_names = [item for k in bwd_outputs for item in bwd_outputs[k]]
-
         places = [core.CPUPlace()]
         if not only_cpu and core.is_compile_gpu() and backward_op.support_gpu():
             places.append(core.GPUPlace(0))
 
-        numeric_grad = dict()
-        # get numeric gradient
-        for check_name in inputs_to_check:
-            numeric_grad[check_name] = \
-                get_numeric_gradient(forward_op, input_vars, output_name,
-                                     check_name)
+        # get numerical gradients
+        numeric_grads = [
+            get_numeric_gradient(forward_op, input_vars, output_name, name)
+            for name in inputs_to_check
+        ]
 
-        # get operator gradient according to different device
+        check_names = [grad_var_name(name) for name in inputs_to_check]
         for place in places:
-            scope = core.Scope()
-            ctx = core.DeviceContext.create(place)
-
-            # create input var and set value
-            for name, value in input_vars.iteritems():
-                if name not in in_names:
-                    raise ValueError(name + " not in op.inputs_")
-                var = scope.new_var(name).get_tensor()
-                var.set_dims(value.shape)
-                var.set(value, place)
-
-            # create output var
-            for out_name in out_names:
-                scope.new_var(out_name).get_tensor()
-
-            # infer the shape of output var and compute/set value of output var
-            forward_op.infer_shape(scope)
-            forward_op.run(scope, ctx)
-
-            # create output grad var
-            # set shape as the output var
-            # set value of this grad to ones
-            for name in out_names:
-                out_tensor = scope.find_var(name).get_tensor()
-                grad_tensor = scope.new_var(grad_var_name(name)).get_tensor()
-                grad_tensor.set_dims(out_tensor.shape())
-                data = 1.0 * numpy.ones(out_tensor.shape())
-                grad_tensor.set(data, place)
-
-            # create input grad var
-            for name in bwd_out_names:
-                scope.new_var(name).get_tensor()
-
-            # infer the shape of input gradient var and compute/set it's value
-            # with backward op
-            backward_op.infer_shape(scope)
-            backward_op.run(scope, ctx)
-
-            self.assert_is_close(numeric_grad, scope, max_relative_error,
-                                 "Gradient Check On %s" % str(place))
-
-
-if __name__ == '__main__':
-
-    class GetNumericGradientTest(unittest.TestCase):
-        def test_add_op(self):
-            add_op = Operator('add_two', X="X", Y="Y", Out="Z")
-            x = numpy.random.random((10, 1)).astype("float32")
-            y = numpy.random.random((10, 1)).astype("float32")
-
-            arr = get_numeric_gradient(add_op, {'X': x, "Y": y}, 'Z', 'X')
-            self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-2)
-
-        def test_softmax_op(self):
-            def stable_softmax(x):
-                """Compute the softmax of vector x in a numerically stable way."""
-                shiftx = x - numpy.max(x)
-                exps = numpy.exp(shiftx)
-                return exps / numpy.sum(exps)
-
-            def label_softmax_grad(Y, dY):
-                dX = Y * 0.0
-                for i in range(Y.shape[0]):
-                    d = numpy.dot(Y[i, :], dY[i, :])
-                    dX[i, :] = Y[i, :] * (dY[i, :] - d)
-                return dX
-
-            softmax_op = Operator("softmax", X="X", Y="Y")
-
-            X = numpy.random.random((2, 2)).astype("float32")
-            Y = numpy.apply_along_axis(stable_softmax, 1, X)
-            dY = numpy.ones(Y.shape)
-            dX = label_softmax_grad(Y, dY)
-
-            arr = get_numeric_gradient(softmax_op, {"X": X}, 'Y', 'X')
-            numpy.testing.assert_almost_equal(arr, dX, decimal=1e-2)
-
-    unittest.main()
+            # get analytical gradients according to different device
+            analytic_grads = self.__get_gradient(forward_op, backward_op,
+                                                 input_vars, check_names, place)
+            self.__assert_is_close(numeric_grads, analytic_grads, check_names,
+                                   max_relative_error,
+                                   "Gradient Check On %s" % str(place))

python/paddle/v2/framework/tests/test_gradient_checker.py

+import unittest
+import numpy
+from paddle.v2.framework.op import Operator
+from gradient_checker import GradientChecker
+from gradient_checker import get_numeric_gradient
+
+
+class GetNumericGradientTest(unittest.TestCase):
+    def test_add_op(self):
+        add_op = Operator('add_two', X="X", Y="Y", Out="Z")
+        x = numpy.random.random((10, 1)).astype("float32")
+        y = numpy.random.random((10, 1)).astype("float32")
+
+        arr = get_numeric_gradient(add_op, {'X': x, "Y": y}, 'Z', 'X')
+        self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-4)
+
+    def test_softmax_op(self):
+        def stable_softmax(x):
+            """Compute the softmax of vector x in a numerically stable way."""
+            shiftx = x - numpy.max(x)
+            exps = numpy.exp(shiftx)
+            return exps / numpy.sum(exps)
+
+        def label_softmax_grad(Y, dY):
+            dX = Y * 0.0
+            for i in range(Y.shape[0]):
+                d = numpy.dot(Y[i, :], dY[i, :])
+                dX[i, :] = Y[i, :] * (dY[i, :] - d)
+            return dX
+
+        softmax_op = Operator("softmax", X="X", Y="Y")
+
+        X = numpy.random.random((2, 2)).astype("float32")
+        Y = numpy.apply_along_axis(stable_softmax, 1, X)
+        dY = numpy.ones(Y.shape)
+        dX = label_softmax_grad(Y, dY)
+
+        arr = get_numeric_gradient(softmax_op, {"X": X}, 'Y', 'X')
+        numpy.testing.assert_almost_equal(arr, dX, decimal=1e-2)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/framework/tests/test_mean_op.py

 import unittest
 from op_test_util import OpTestMeta
+from gradient_checker import GradientChecker, create_op
 import numpy as np
 
 
@@ -12,5 +13,12 @@ class TestMeanOp(unittest.TestCase):
         self.outputs = {'Out': np.mean(self.inputs['X'])}
 
 
+class MeanGradOpTest(GradientChecker):
+    def test_normal(self):
+        op = create_op("mean")
+        inputs = {"X": np.random.random((10, 10)).astype("float32")}
+        self.check_grad(op, inputs, set("X"), "Out")
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/v2/framework/tests/test_sigmoid_op.py

 import unittest
-from op_test_util import OpTestMeta
 import numpy as np
+from op_test_util import OpTestMeta
+from gradient_checker import GradientChecker, create_op
 
 
 class TestSigmoidOp(unittest.TestCase):
@@ -8,12 +9,20 @@ class TestSigmoidOp(unittest.TestCase):
 
     def setUp(self):
         self.type = "sigmoid"
-        self.inputs = {'X': np.random.random((32, 100)).astype("float32")}
+        self.inputs = {'X': np.random.random((15, 31)).astype("float32")}
         self.outputs = {'Y': 1 / (1 + np.exp(-self.inputs['X']))}
 
 
-#class TestSigmoidGradOp(unittest.TestCase):
-#TODO(qingqing) add unit test
+class TestSigmoidGradOp(GradientChecker):
+    def test_grad(self):
+        op = create_op("sigmoid")
+        inputs = {"X": np.random.uniform(0.1, 1, [11, 17]).astype("float32")}
+        # compare gpu and cpu results for backward op.
+        # this test will be skiped if only compiling CPU version.
+        self.compare_grad(op, inputs)
+        # check gradients 
+        self.check_grad(op, inputs, set("X"), "Y", max_relative_error=0.007)
+
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/v2/reader/creator.py

@@ -57,7 +57,7 @@ def text_file(path):
     return reader
 
 
-def recordio_local(paths, buf_size=100):
+def recordio(paths, buf_size=100):
     """
     Creates a data reader from given RecordIO file paths separated by ",",
         glob pattern is supported.
@@ -67,15 +67,19 @@ def recordio_local(paths, buf_size=100):
 
     import recordio as rec
     import paddle.v2.reader.decorator as dec
+    import cPickle as pickle
 
     def reader():
-        a = ','.join(paths)
-        f = rec.reader(a)
+        if isinstance(paths, basestring):
+            path = paths
+        else:
+            path = ",".join(paths)
+        f = rec.reader(path)
         while True:
             r = f.read()
             if r is None:
                 break
-            yield r
+            yield pickle.loads(r)
         f.close()
 
     return dec.buffered(reader, buf_size)

python/paddle/v2/reader/tests/creator_test.py

@@ -34,5 +34,27 @@ class TestTextFile(unittest.TestCase):
             self.assertEqual(e, str(idx * 2) + " " + str(idx * 2 + 1))
 
 
+class TestRecordIO(unittest.TestCase):
+    def do_test(self, path):
+        reader = paddle.v2.reader.creator.recordio(path)
+        idx = 0
+        for e in reader():
+            if idx == 0:
+                self.assertEqual(e, (1, 2, 3))
+            elif idx == 1:
+                self.assertEqual(e, (4, 5, 6))
+            idx += 1
+        self.assertEqual(idx, 2)
+
+    def test_recordIO(self):
+        self.do_test(
+            os.path.join(
+                os.path.dirname(__file__), "test_reader_recordio.dat"))
+        self.do_test([
+            os.path.join(
+                os.path.dirname(__file__), "test_reader_recordio.dat")
+        ])
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/v2/trainer.py

@@ -27,16 +27,24 @@ class SGD(object):
     SGD Trainer combines data reader, network topolopy and update_equation together
     to train/test a neural network.
 
-    :param update_equation: The optimizer object.
-    :type update_equation: paddle.v2.optimizer.Optimizer
     :param cost: Target cost that neural network should be optimized.
     :type cost: paddle.v2.config_base.Layer
     :param parameters: The parameters dictionary.
     :type parameters: paddle.v2.parameters.Parameters
+    :param update_equation: The optimizer object.
+    :type update_equation: paddle.v2.optimizer.Optimizer
     :param extra_layers: Some layers in the neural network graph are not
                          in the path of cost layer.
-    :param pserver_spec: pserver location, eg: localhost:3000
     :type extra_layers: paddle.v2.config_base.Layer
+    :param is_local: Whether trainning locally
+    :type is_local: bool
+    :param pserver_spec: comma string for pserver location,
+                         eg:127.10.0.10:3000,127.10.0.11:3000,
+                         and this parameter is only used for fault
+                         tolerant mode cluster training.
+    :type pserver_spec: string
+    :param use_etcd: Whether using etcd pserver.
+    :param use_etcd: bool
     """
 
     def __init__(self,

python/requirements.txt

 requests==2.9.2
 numpy>=1.12
 protobuf==3.1
-recordio
+recordio>=0.1.0
 matplotlib
 rarfile
 scipy>=0.19.0

python/setup.py.in

@@ -24,14 +24,17 @@ if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']:
     setup_requires+=["opencv-python"]
 
 # the prefix is sys.prefix which should always be usr
-paddle_bin_dir = 'local/opt/paddle/bin'
+paddle_bin_dir = 'opt/paddle/bin'
 paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/scripts/paddle_usage',
                '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer',
                '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model',
-               '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main']
+               '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main',
+               '${PADDLE_BINARY_DIR}/paddle/scripts/paddle']
 
-paddle_rt_lib_dir = 'local/lib'
-paddle_rt_libs = [] if '${MKL_SHARED_LIBS}'== '' else '${MKL_SHARED_LIBS}'.split(';')
+paddle_rt_lib_dir = 'lib'
+paddle_rt_libs = ['${WARPCTC_LIBRARIES}']
+if '${MKL_SHARED_LIBS}'!= '':
+  paddle_rt_libs += '${MKL_SHARED_LIBS}'.split(';')
 
 setup(name='paddlepaddle',
       version='${PADDLE_VERSION}',
@@ -50,8 +53,7 @@ setup(name='paddlepaddle',
           'paddle.v2.framework.proto': '${PADDLE_BINARY_DIR}/paddle/framework',
           'py_paddle': '${PADDLE_SOURCE_DIR}/paddle/py_paddle'
       },
-      scripts=['${PADDLE_BINARY_DIR}/paddle/scripts/paddle'],
+      scripts=paddle_bins,
       distclass=BinaryDistribution,
-      data_files=[(paddle_bin_dir, paddle_bins),
-                  (paddle_rt_lib_dir, paddle_rt_libs)]
+      data_files=[(paddle_rt_lib_dir, paddle_rt_libs)]
 )