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

.travis.yml

@@ -21,7 +21,6 @@ addons:
       - python
       - python-pip
       - python2.7-dev
-      - python-numpy
       - python-wheel
       - libboost-dev
       - curl
@@ -35,8 +34,8 @@ before_install:
   - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
   # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python
   # protobuf version.
-  - pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
-  - pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
+  - sudo pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
+  - sudo pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
   - curl https://glide.sh/get | bash
   - eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
   - go get -u github.com/alecthomas/gometalinter

CMakeLists.txt

@@ -65,8 +65,8 @@ if(NOT CMAKE_BUILD_TYPE)
 endif()
 
 if(ANDROID)
-    if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
-        message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 21")
+    if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
+        message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
     endif()
 
     set(WITH_GPU OFF CACHE STRING

doc/design/functions_operators_layers.md

@@ -86,12 +86,13 @@ def layer.fc(X):
 
 We'd like to have Python bindings to operators in package `paddle.operator`, and Python compositions of operators in package `paddle.layer`.  So we have the following concepts in above illustrative example:
 
-```
+
 | C++ functions/functors | mul          | add          |             |          |
+|------------------------|--------------|--------------|-------------|----------|
 | C++ operator class     | mulOp        | addOp        | FCOp        |          |
 | Python binding         | operator.mul | operator.add | operator.fc |          |
 | Python function        |              |              |             | layer.fc |
-```
+
 
 This is how we differentiate layer and operators in PaddlePaddle:
 

doc/design/ops/dist_train.md

+# Design Doc: Operation Graph Based Parameter Server
+
+## Abstract
+
+We propose an approach to implement the parameter server. In this
+approach, there is no fundamental difference between the trainer and
+the parameter server: they both run subgraphs, but subgraphs of
+different purposes.
+
+## Background
+
+The previous implementations of the parameter server does not run a
+subgraph. parameter initialization, optimizer computation, network
+communication and checkpointing are implemented twice on both the
+trainer and the parameter server.
+
+It would be great if we can write code once and use them on both the
+trainer and the parameter server: reduces code duplication and
+improves extensibility. Given that after the current refactor, we are
+representing everything as a computing graph on the
+trainer. Representing everything as a computing graph on the parameter
+server becomes a natural extension.
+
+## Design
+
+### Graph Converter
+
+The *graph converter* converts the user-defined operation (OP) graph
+into subgraphs to be scheduled on different nodes with the following
+steps:
+
+1. OP placement: the OPs will be placed on different nodes according
+   to heuristic that minimizes estimated total computation
+   time. Currently we will use a simple heuristic that puts parameter
+   varable on parameter server workers and everything else on trainer
+   workers.
+
+1. Add communication OPs to enable the communication between nodes.
+
+We will need these OPs: *Send*, *Recv*, *Enqueue*, *Dequeue*.
+
+Below is an example of converting the user defined graph to the
+subgraphs for the trainer and the parameter server:
+
+<img src="src/local-graph.png" width="300"/>
+
+After converting:
+
+<img src="src/dist-graph.png" width="700"/>
+
+1. The parameter variable W and it's optimizer subgraph are placed on the parameter server.
+1. Operators are added to the subgraphs.
+   - *Send* sends data to the connected *Recv* operator.  The
+	 scheduler on the receive node will only schedule *Recv* operator
+	 to run when the *Send* operator has ran (the *Send* OP will mark
+	 the *Recv* OP runnable automatically).
+   - *Enueue* enqueues the input variable, it can block until space
+     become available in the queue.
+   - *Dequeue* outputs configurable numbers of tensors from the
+     queue. It will block until the queue have the required number of
+     tensors.
+
+
+### Benefits
+
+- Model parallelism become easier to implement: it's an extension to
+  the trainer - parameter server approach. we already have the
+  communication OPs, but need to extend the graph converter's
+  placement functionality.
+
+- User-defined optimizer is easier to add - user can now express it as
+  a subgraph.
+
+- No more duplication logic inside the trainer and the parameter
+  server mentioned in the background section.
+
+### Challenges
+
+- It might be hard for the graph converter to cut a general graph
+  (without any hint for which subgraph is the optimizer). We may need
+  to label which subgraph inside the OP graph is the optimizer.
+
+- It's important to balance the parameter shards of on multiple
+  parameter server. If a single parameter is very big (some
+  word-embedding, fully connected, softmax layer), we need to
+  automatically partition the single parameter onto different
+  parameter servers when possible (only element-wise optimizer depends
+  on the parameter variable).
+
+### Discussion
+
+- In the "Aync SGD" figure, the "W" variable on the parameter server
+  could be read and wrote concurrently, what is our locking strategy?
+  E.g., each variable have a lock cpp method to be invoked by every
+  OP, or, have a lock OP.
+
+- Can the Enqueue OP be implemented under our current tensor design
+  (puts the input tensor into the queue tensor)?
+
+- *Dequeue* OP will have variable numbers of output (depends on the
+  `min_count` attribute), does our current design support it? (similar
+  question for the *Add* OP)
+
+
+### References:
+[1] [TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems](https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45166.pdf)

doc/howto/dev/write_docs_cn.rst

@@ -5,15 +5,13 @@
 PaddlePaddle的文档包括英文文档 ``doc`` 和中文文档 ``doc_cn`` 两个部分。文档都是通过 `cmake`_ 驱动 `sphinx`_ 编译生成，生成后的文档分别存储在编译目录的 ``doc`` 和 ``doc_cn`` 两个子目录下。
 
 
-如何构建PaddlePaddle的文档
-==========================
+如何构建文档
+============
 
-PaddlePaddle的文档构建有直接构建和基于Docker构建两种方式，我们提供了一个构建脚本build_docs.sh来进行构建。
-PaddlePaddle文档需要准备的环境相对较复杂，所以我们推荐使用基于Docker来构建PaddlePaddle的文档。
+PaddlePaddle的文档构建有两种方式。
 
-
-使用Docker构建PaddlePaddle的文档
---------------------------------
+使用Docker构建
+--------------
 
 使用Docker构建PaddlePaddle的文档，需要在系统里先安装好Docker工具包。Docker安装请参考 `Docker的官网 <https://docs.docker.com/>`_ 。安装好Docker之后可以使用源码目录下的脚本构建文档，即
 
@@ -21,58 +19,46 @@ PaddlePaddle文档需要准备的环境相对较复杂，所以我们推荐使
 
     cd TO_YOUR_PADDLE_CLONE_PATH
     cd paddle/scripts/tools/build_docs
-    bash build_docs.sh with_docker
-
-编译完成后，会在当前目录生成两个子目录\:
-
-* doc 英文文档目录
-* doc_cn 中文文档目录
+    sh build_docs.sh
 
+编译完成之后，会在当前目录生成两个子目录\: doc(英文文档目录)和 doc_cn(中文文档目录)。
 打开浏览器访问对应目录下的index.html即可访问本地文档。
 
-
-
-直接构建PaddlePaddle的文档
---------------------------
-
-因为PaddlePaddle的v2 api文档生成过程依赖于py_paddle Python包，用户需要首先确认py_paddle包已经安装。
-
-..  code-block:: bash
-
-    python -c "import py_paddle"
-
-如果提示错误，那么用户需要在本地编译安装PaddlePaddle，请参考 `源码编译文档 <http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/build_from_source_en.html>`_ 。
-注意，用户在首次编译安装PaddlePaddle时，请将WITH_DOC选项关闭。在编译安装正确之后，请再次确认py_paddle包已经安装，即可进行下一步操作。
+直接构建
+--------
 
 如果提示正确，可以执行以下命令编译生成文档，即
 
 ..  code-block:: bash
 
     cd TO_YOUR_PADDLE_CLONE_PATH
-    cd paddle/scripts/tools/build_docs
-    bash build_docs.sh local
-
-编译完成之后，会在当前目录生成两个子目录\:
-
-* doc 英文文档目录
-* doc_cn 中文文档目录
+    mkdir -p build
+    cd build
+    cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKLDNN=OFF -DWITH_MKLML=OFF -DWITH_DOC=ON
+    make gen_proto_py
+    make paddle_docs paddle_docs_cn
 
+编译完成之后，会在当前目录生成两个子目录\: doc(英文文档目录)和 doc_cn(中文文档目录)。
 打开浏览器访问对应目录下的index.html即可访问本地文档。
 
 
-如何书写PaddlePaddle的文档
-==========================
+如何书写文档
+============
 
 PaddlePaddle文档使用 `sphinx`_ 自动生成，用户可以参考sphinx教程进行书写。
 
-如何更新www.paddlepaddle.org文档
-================================
+如何更新文档主题
+================
+
+PaddlePaddle文档主题在 `TO_YOUR_PADDLE_CLONE_PATH/doc_theme` 文件夹下，包含所有和前端网页设计相关的文件。
 
-开发者给PaddlePaddle代码增加的注释以PR的形式提交到github中，提交方式可参见 `贡献文档 <http://doc.paddlepaddle.org/develop/doc_cn/howto/dev/contribute_to_paddle_cn.html>`_ 。
+如何更新doc.paddlepaddle.org
+============================
+
+更新的文档以PR的形式提交到github中，提交方式参见 `贡献文档 <http://doc.paddlepaddle.org/develop/doc_cn/howto/dev/contribute_to_paddle_cn.html>`_ 。
 目前PaddlePaddle的develop分支的文档是自动触发更新的，用户可以分别查看最新的 `中文文档 <http://doc.paddlepaddle.org/develop/doc_cn/>`_ 和
 `英文文档 <http://doc.paddlepaddle.org/develop/doc/>`_ 。
 
 
-
 ..  _cmake: https://cmake.org/
 ..  _sphinx: http://www.sphinx-doc.org/en/1.4.8/

paddle/framework/CMakeLists.txt

@@ -9,6 +9,7 @@ cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
 
 cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor)
 cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor)
+nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
 
 cc_test(variable_test SRCS variable_test.cc)
 

paddle/framework/attribute.h

@@ -45,7 +45,19 @@ class GreaterThanChecker {
  public:
   explicit GreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
   void operator()(T& value) const {
-    PADDLE_ENFORCE(value > lower_bound_, "larger_than check fail");
+    PADDLE_ENFORCE(value > lower_bound_, "larger_than check fails.");
+  }
+
+ private:
+  T lower_bound_;
+};
+
+template <typename T>
+class EqualGreaterThanChecker {
+ public:
+  explicit EqualGreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
+  void operator()(T& value) const {
+    PADDLE_ENFORCE_GE(value, lower_bound_, "equal_larger_than check fails.");
   }
 
  private:
@@ -115,6 +127,11 @@ class TypedAttrChecker {
     return *this;
   }
 
+  TypedAttrChecker& EqualGreaterThan(const T& lower_bound) {
+    value_checkers_.push_back(EqualGreaterThanChecker<T>(lower_bound));
+    return *this;
+  }
+
   // we can add more common limits, like LessThan(), Between()...
 
   TypedAttrChecker& SetDefault(const T& default_value) {

paddle/framework/backward.md

@@ -2,20 +2,20 @@
 
 ## Motivation
 
-In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
-  
+In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the gradient operators/expressions together with the chain rule. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
+
 ## Backward Operator Registry
 
-A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs and output gradients and then calculate its input gradients.
+A backward network is built up with several backward operators. Backward operators take forward operators' inputs outputs, and output gradients and then calculate its input gradients.
 
 |                        | forward operator | backward operator 
 | ---------------------- | ---------------- |------------------------- |		
 | **Operator::inputs_**  | Inputs       | Inputs, Outputs, OutputGradients |	
 | **Operator::outputs_** | Outputs          | InputGradients            |
 
- In most cases, there is a one-to-one correspondence between forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced.
+ In most cases, there is a one-to-one correspondence between the forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced.
 
-For example, we have got a `mul_op`, and we can register it's information and corresponding backward operator by the following macro:
+For example, we have got a `mul_op`, and we can register its information and corresponding backward operator by the following macro:
 
 ```cpp
 REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
@@ -27,17 +27,17 @@ REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
 
 ## Backward Opeartor Creating
 
-Given a certain forward operator, we can get its corresponding backward opeartor by calling:
+Given a certain forward operator, we can get its corresponding backward operator by calling:
 
 ```cpp
 OperatorBase* bwd_op = BuildGradOp(const OperatorBase* fwd_op);
-``` 
+```
 
 The function `BuildGradOp` will sequentially execute following processes:
 
 1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`.
 
-2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these are not necessary for gradient computing.
+2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing.
 
 3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`.
 
@@ -49,31 +49,31 @@ A backward network is a series of backward operators. The main idea of building
 
 In our design, the network itself is also a kind of operator. So the operators contained by a big network may be some small network. 
 
-given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`,`InputGradients`.
+given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`, `InputGradients`.
 
 1. Op 
 
-   when the input forward network is a Op, return its gradient Operator Immediately.
+   when the input forward network is an Op, return its gradient Operator Immediately.
 
 2. NetOp 
 
-   when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to forward NetOp.
+   when the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp.
 
-   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwirte their shared input variable.  
+   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwrite their shared input variable.  
 
    <p align="center">
-   <img src="./images/duplicate_op.png" width="70%" ><br/>
+   <img src="./images/duplicate_op.png" width="50%" ><br/>
 
-   1. shared variable in two operators. 
+   1. Shared variable in operators. 
 
    </p>
 
-   Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator replace the overwirte links. 
+   Share variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator replace the overwrite links. 
 
    <p align="center">
-   <img src="images/duplicate_op2.png" width="90%" ><br/>
+   <img src="images/duplicate_op2.png" width="50%" ><br/>
 
-   2. replace shared variable gradient with `Add` Operator
+   2. Replace shared variable's gradient with `Add` operator.
 
    </p>
 

paddle/framework/ddim.cc

@@ -283,5 +283,14 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
 DDim::DDim(std::initializer_list<int64_t> init_list) {
   *this = make_ddim(init_list);
 }
+
+DDim flatten_to_2d(const DDim& src, int num_col_dims) {
+  int rank = src.size();
+  return make_ddim({product(slice_ddim(src, 0, num_col_dims)),
+                    product(slice_ddim(src, num_col_dims, rank))});
+}
+
+DDim flatten_to_1d(const DDim& src) { return make_ddim({product(src)}); }
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/ddim.h

@@ -115,6 +115,12 @@ int arity(const DDim& ddim);
 
 std::ostream& operator<<(std::ostream&, const DDim&);
 
+// Reshape a tensor to a matrix. The matrix's first dimension(column length)
+// will be the product of tensor's first `num_col_dims` dimensions.
+DDim flatten_to_2d(const DDim& src, int num_col_dims);
+
+DDim flatten_to_1d(const DDim& src);
+
 }  // namespace framework
 }  // namespace paddle
 

paddle/framework/eigen.h

@@ -63,20 +63,35 @@ struct EigenTensor {
 
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
-struct EigenMatrix : public EigenTensor<T, 2, MajorType, IndexType> {};
+struct EigenMatrix : public EigenTensor<T, 2, MajorType, IndexType> {
+  static typename EigenMatrix::Type Reshape(Tensor& tensor, int num_col_dims) {
+    int rank = tensor.dims_.size();
+    PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank,
+                   "`num_col_dims` must be between (0, rank_of_tensor).");
+    return EigenMatrix::From(tensor,
+                             flatten_to_2d(tensor.dims(), num_col_dims));
+  }
+
+  static typename EigenMatrix::ConstType Reshape(const Tensor& tensor,
+                                                 int num_col_dims) {
+    int rank = tensor.dims_.size();
+    PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank,
+                   "`num_col_dims` must be between (0, rank_of_tensor).");
+    return EigenMatrix::From(tensor,
+                             flatten_to_2d(tensor.dims(), num_col_dims));
+  }
+};
 
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 struct EigenVector : public EigenTensor<T, 1, MajorType, IndexType> {
   // Flatten reshapes a Tensor into an EigenVector.
   static typename EigenVector::Type Flatten(Tensor& tensor) {
-    return EigenVector::From(
-        tensor, make_ddim({static_cast<int>(product(tensor.dims_))}));
+    return EigenVector::From(tensor, {product(tensor.dims_)});
   }
 
   static typename EigenVector::ConstType Flatten(const Tensor& tensor) {
-    return EigenVector::From(
-        tensor, make_ddim({static_cast<int>(product(tensor.dims_))}));
+    return EigenVector::From(tensor, {product(tensor.dims_)});
   }
 };
 

paddle/framework/eigen_test.cc

@@ -108,5 +108,24 @@ TEST(Eigen, Matrix) {
   }
 }
 
+TEST(Eigen, MatrixReshape) {
+  Tensor t;
+  float* p = t.mutable_data<float>({2, 3, 6, 4}, platform::CPUPlace());
+  for (int i = 0; i < 2 * 3 * 6 * 4; ++i) {
+    p[i] = static_cast<float>(i);
+  }
+
+  EigenMatrix<float>::Type em = EigenMatrix<float>::Reshape(t, 2);
+
+  ASSERT_EQ(2 * 3, em.dimension(0));
+  ASSERT_EQ(6 * 4, em.dimension(1));
+
+  for (int i = 0; i < 2 * 3; i++) {
+    for (int j = 0; j < 6 * 4; j++) {
+      ASSERT_NEAR(i * 6 * 4 + j, em(i, j), 1e-6f);
+    }
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor.h

@@ -18,8 +18,10 @@
 #ifndef PADDLE_ONLY_CPU
 #include <thrust/device_vector.h>
 #include <thrust/host_vector.h>
+#include <thrust/system/cuda/experimental/pinned_allocator.h>
 #endif
 
+#include <glog/logging.h>
 #include "paddle/framework/ddim.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/enforce.h"
@@ -32,7 +34,8 @@ template <typename T>
 using Vector = std::vector<T>;
 #else
 template <typename T>
-using Vector = thrust::host_vector<T>;
+using Vector = thrust::host_vector<
+    T, thrust::system::cuda::experimental::pinned_allocator<T>>;
 #endif
 
 using LoD = std::vector<Vector<size_t>>;

paddle/framework/lod_tensor_test.cu

+/*
+  Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+  Licensed under the Apache License, Version 2.0 (the "License");
+  you may not use this file except in compliance with the License.
+  You may obtain a copy of the License at
+  http://www.apache.org/licenses/LICENSE-2.0
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+*/
+
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/platform/assert.h"
+
+#include <gtest/gtest.h>
+
+__global__ void test(size_t* a, int size) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < size;
+       i += blockDim.x * gridDim.x) {
+    a[i] *= 2;
+  }
+}
+
+TEST(LoDTensor, LoDInGPU) {
+  paddle::framework::Tensor tensor;
+  paddle::framework::LoDTensor lod_tensor;
+  paddle::platform::GPUPlace place(0);
+
+  paddle::framework::LoD src_lod;
+  src_lod.push_back(std::vector<size_t>{0, 2, 4, 6, 8, 10, 12, 14});
+
+  tensor.Resize({14, 16});
+  tensor.mutable_data<float>(place);
+
+  lod_tensor.set_lod(src_lod);
+  lod_tensor.set_tensor(&tensor);
+  CHECK_EQ(lod_tensor.lod_element(0, 2), 4);
+  CHECK_EQ(lod_tensor.lod_element(0, 4), 8);
+
+  auto lod = lod_tensor.lod();
+
+  test<<<1, 8>>>(lod[0].data(), lod[0].size());
+  cudaDeviceSynchronize();
+
+  for (size_t i = 0; i < src_lod[0].size(); ++i) {
+    CHECK_EQ(lod[0].data()[i], src_lod[0].data()[i] * 2);
+  }
+}

paddle/framework/operator.cc

@@ -123,6 +123,15 @@ OperatorBase::OperatorBase(const std::string& type,
   CheckAllInputOutputSet();
 }
 
+std::vector<std::string> OperatorBase::InputVars() const {
+  std::vector<std::string> ret_val;
+  for (auto& o : outputs_) {
+    ret_val.reserve(ret_val.size() + o.second.size());
+    ret_val.insert(ret_val.end(), o.second.begin(), o.second.end());
+  }
+  return ret_val;
+}
+
 std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
   std::vector<std::string> ret_val;
   if (has_intermediate) {

paddle/framework/operator.h

@@ -94,11 +94,14 @@ class OperatorBase {
 
   const VariableNameMap& Inputs() const { return inputs_; }
   const VariableNameMap& Outputs() const { return outputs_; }
+
   //! Get a input with argument's name described in `op_proto`
   std::string Input(const std::string& name) const;
   //! Get a input which has multiple variables.
   const std::vector<std::string>& Inputs(const std::string& name) const;
 
+  std::vector<std::string> InputVars() const;
+
   //! Get a output with argument's name described in `op_proto`
   std::string Output(const std::string& name) const;
   //! Get an output which has multiple variables.
@@ -311,9 +314,9 @@ class InferShapeContext {
   }
 
   template <typename T>
-  std::vector<const T*> MultiOutput(const std::string& name) const {
+  std::vector<T*> MultiOutput(const std::string& name) const {
     auto names = op_.Outputs(name);
-    std::vector<const T*> res;
+    std::vector<T*> res;
     res.reserve(names.size());
     std::transform(names.begin(), names.end(), std::back_inserter(res),
                    [&](const std::string& sub_name) {

paddle/framework/tensor.h

@@ -43,6 +43,9 @@ class Tensor {
   template <typename T, size_t D, int MajorType, typename IndexType>
   friend struct EigenTensor;
 
+  template <typename T, int MajorType, typename IndexType>
+  friend struct EigenMatrix;
+
   template <typename T, int MajorType, typename IndexType>
   friend struct EigenVector;
 
@@ -81,6 +84,9 @@ class Tensor {
   /*! Return the dimensions of the memory block. */
   inline const DDim& dims() const;
 
+  /*! Return the numel of the memory block. */
+  inline int64_t numel() const;
+
   /*! Resize the dimensions of the memory block. */
   inline Tensor& Resize(const DDim& dims);
 
@@ -168,6 +174,12 @@ class Tensor {
   /*! points to dimensions of memory block. */
   DDim dims_;
 
+  /**
+   * A cache of the number of elements in a tensor.
+   * Would be 0 for an uninitialized tensor.
+   */
+  int64_t numel_;
+
   /**
    * @brief   A PlaceHolder may be shared by more than one tensor.
    *

paddle/framework/tensor_impl.h

@@ -24,7 +24,7 @@ inline void Tensor::check_memory_size() const {
   PADDLE_ENFORCE_NOT_NULL(
       holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");
   PADDLE_ENFORCE_GE(
-      holder_->size(), product(dims_) * sizeof(T) + offset_,
+      holder_->size(), numel() * sizeof(T) + offset_,
       "Tensor's dims_ is out of bound. Call Tensor::mutable_data "
       "first to re-allocate memory.\n"
       "or maybe the required data-type mismatches the data already stored.");
@@ -54,11 +54,11 @@ inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
 template <typename T>
 inline T* Tensor::mutable_data(platform::Place place) {
   static_assert(std::is_pod<T>::value, "T must be POD");
-  PADDLE_ENFORCE_GT(product(dims_), 0,
+  PADDLE_ENFORCE_GT(numel(), 0,
                     "Tensor's numel must be larger than zero to call "
                     "Tensor::mutable_data. Call Tensor::set_dim first.");
   /* some versions of boost::variant don't have operator!= */
-  int64_t size = product(dims_) * sizeof(T);
+  int64_t size = numel() * sizeof(T);
   if (holder_ == nullptr || !(holder_->place() == place) ||
       holder_->size() < size + offset_) {
     if (platform::is_cpu_place(place)) {
@@ -97,7 +97,7 @@ inline void Tensor::CopyFrom(const Tensor& src,
 
   auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
 
-  auto size = product(src.dims_) * sizeof(T);
+  auto size = src.numel() * sizeof(T);
 
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
@@ -131,7 +131,7 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
   PADDLE_ENFORCE_LT(begin_idx, end_idx,
                     "Begin index must be less than end index.");
   PADDLE_ENFORCE_NE(dims_[0], 1, "Can not slice a tensor with dims_[0] = 1.");
-  size_t base = product(dims_) / dims_[0];
+  size_t base = numel() / dims_[0];
   Tensor dst;
   dst.holder_ = holder_;
   DDim dst_dims = dims_;
@@ -143,10 +143,21 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
 
 inline Tensor& Tensor::Resize(const DDim& dims) {
   dims_ = dims;
+  numel_ = product(dims_);
   return *this;
 }
 
 inline const DDim& Tensor::dims() const { return dims_; }
 
+inline int64_t Tensor::numel() const { return numel_; }
+
+template <typename T>
+inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
+  Tensor res;
+  res.ShareDataWith<T>(src);
+  res.Resize(flatten_to_2d(src.dims(), num_col_dims));
+  return res;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/tensor_test.cc

@@ -264,3 +264,16 @@ TEST(Tensor, CopyFrom) {
   }
 #endif
 }
+
+TEST(Tensor, ReshapeToMatrix) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  Tensor src;
+  int* src_ptr = src.mutable_data<int>({2, 3, 4, 9}, CPUPlace());
+  for (int i = 0; i < 2 * 3 * 4 * 9; ++i) {
+    src_ptr[i] = i;
+  }
+  Tensor res = ReshapeToMatrix<int>(src, 2);
+  ASSERT_EQ(res.dims()[0], 2 * 3);
+  ASSERT_EQ(res.dims()[1], 4 * 9);
+}
\ No newline at end of file

paddle/gserver/layers/BatchNormBaseLayer.cpp

@@ -62,14 +62,18 @@ void BatchNormBaseLayer::calFeatureMapSize() {
   const ImageConfig& conf = config_.inputs(0).image_conf();
   imageH_ = inputLayers_[0]->getOutput().getFrameHeight();
   imageW_ = inputLayers_[0]->getOutput().getFrameWidth();
+  imageD_ = inputLayers_[0]->getOutput().getFrameDepth();
+
+  if (0 == imageD_) imageD_ = conf.img_size_z();
   if (imageH_ == 0 && imageW_ == 0) {
     imageH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
     imageW_ = conf.img_size();
   } else {
     getOutput().setFrameHeight(imageH_);
     getOutput().setFrameWidth(imageW_);
+    getOutput().setFrameDepth(imageD_);
   }
-  imgPixels_ = imageH_ * imageW_;
+  imgPixels_ = imageH_ * imageW_ * imageD_;
 }
 
 }  // namespace paddle

paddle/gserver/layers/BatchNormBaseLayer.h

@@ -80,6 +80,7 @@ protected:
 
   /// Height or width of input image feature.
   /// Both of them are 1 if the input is fully-connected layer.
+  int imageD_;
   int imageH_;
   int imageW_;
   /// Height * Width.

paddle/gserver/layers/CudnnBatchNormLayer.cpp

@@ -37,7 +37,7 @@ bool CudnnBatchNormLayer::init(const LayerMap& layerMap,
 }
 
 void CudnnBatchNormLayer::reshape(int batchSize) {
-  hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_, imageW_);
+  hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_ * imageD_, imageW_);
 }
 
 void CudnnBatchNormLayer::forward(PassType passType) {
@@ -104,7 +104,7 @@ void CudnnBatchNormLayer::forward(PassType passType) {
                                    EPS,
                                    batchSize,
                                    channels_,
-                                   imageH_,
+                                   imageH_ * imageD_,
                                    imageW_);
     }
   }

paddle/gserver/layers/DeConv3DLayer.cpp

@@ -53,27 +53,27 @@ bool DeConv3DLayer::init(const LayerMap &layerMap,
 
 size_t DeConv3DLayer::getSize() {
   CHECK_NE(inputLayers_.size(), 0UL);
-  outputH_.clear();
-  outputW_.clear();
-  outputD_.clear();
+  imgSizeW_.clear();
+  imgSizeH_.clear();
+  imgSizeD_.clear();
   N_.clear();
   NOut_.clear();
   size_t layerSize = 0;
   for (size_t i = 0; i < inputLayers_.size(); ++i) {
-    outputW_.push_back(
-        imageSize(imgSizeW_[i], filterSize_[i], padding_[i], stride_[i], true));
-    outputH_.push_back(imageSize(
-        imgSizeH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true));
-    outputD_.push_back(imageSize(
-        imgSizeD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true));
-    NOut_.push_back(outputD_[i] * outputH_[i] * outputW_[i]);
-    N_.push_back(imgSizeD_[i] * imgSizeH_[i] * imgSizeW_[i]);
+    imgSizeW_.push_back(
+        imageSize(outputW_[i], filterSize_[i], padding_[i], stride_[i], true));
+    imgSizeH_.push_back(imageSize(
+        outputH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true));
+    imgSizeD_.push_back(imageSize(
+        outputD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true));
+    NOut_.push_back(imgSizeD_[i] * imgSizeH_[i] * imgSizeW_[i]);
+    N_.push_back(outputD_[i] * outputH_[i] * outputW_[i]);
     CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize);
     layerSize += NOut_[i] * numFilters_;
   }
-  getOutput().setFrameHeight(outputH_[0]);
-  getOutput().setFrameWidth(outputW_[0]);
-  getOutput().setFrameDepth(outputD_[0]);
+  getOutput().setFrameHeight(imgSizeH_[0]);
+  getOutput().setFrameWidth(imgSizeW_[0]);
+  getOutput().setFrameDepth(imgSizeD_[0]);
   return layerSize;
 }
 
@@ -103,9 +103,9 @@ void DeConv3DLayer::forward(PassType passType) {
       }
       colBuf_->col2Vol(outMat->getData() + n * outMat->getStride(),
                        numFilters_,
-                       outputD_[i],
-                       outputH_[i],
-                       outputW_[i],
+                       imgSizeD_[i],
+                       imgSizeH_[i],
+                       imgSizeW_[i],
                        filterSizeZ_[i],
                        filterSizeY_[i],
                        filterSize_[i],
@@ -144,9 +144,9 @@ void DeConv3DLayer::backward(const UpdateCallback &callback) {
         colBuf_->vol2Col(
             getOutputGrad()->getData() + n * getOutputGrad()->getStride(),
             numFilters_,
-            outputD_[i],
-            outputH_[i],
-            outputW_[i],
+            imgSizeD_[i],
+            imgSizeH_[i],
+            imgSizeW_[i],
             filterSizeZ_[i],
             filterSizeY_[i],
             filterSize_[i],

paddle/gserver/layers/DetectionOutputLayer.cpp

@@ -139,7 +139,13 @@ void DetectionOutputLayer::forward(PassType passType) {
                                        allDecodedBBoxes,
                                        &allIndices);
 
-  resetOutput(numKept, 7);
+  if (numKept > 0) {
+    resetOutput(numKept, 7);
+  } else {
+    MatrixPtr outV = getOutputValue();
+    outV = NULL;
+    return;
+  }
   MatrixPtr outV = getOutputValue();
   getDetectionOutput(confBuffer_->getData(),
                      numKept,

paddle/gserver/layers/DetectionUtil.cpp

@@ -469,7 +469,7 @@ size_t getDetectionIndices(
     const size_t numClasses,
     const size_t backgroundId,
     const size_t batchSize,
-    const size_t confThreshold,
+    const real confThreshold,
     const size_t nmsTopK,
     const real nmsThreshold,
     const size_t keepTopK,

paddle/gserver/layers/DetectionUtil.h

@@ -275,7 +275,7 @@ size_t getDetectionIndices(
     const size_t numClasses,
     const size_t backgroundId,
     const size_t batchSize,
-    const size_t confThreshold,
+    const real confThreshold,
     const size_t nmsTopK,
     const real nmsThreshold,
     const size_t keepTopK,

paddle/gserver/layers/Layer.h

@@ -49,6 +49,12 @@ struct LayerState {
 };
 typedef std::shared_ptr<LayerState> LayerStatePtr;
 
+/// Paddle device ID, MKLDNN is -2, CPU is -1
+enum PADDLE_DEVICE_ID {
+  MKLDNN_DEVICE = -2,
+  CPU_DEVICE = -1,
+};
+
 /**
  * @brief Base class for layer.
  * Define necessary variables and functions for every layer.
@@ -59,11 +65,6 @@ protected:
   LayerConfig config_;
   /// whether to use GPU
   bool useGpu_;
-  /// Paddle device ID, MKLDNN is -2, CPU is -1
-  enum PADDLE_DEVICE_ID {
-    MKLDNN_DEVICE = -2,
-    CPU_DEVICE = -1,
-  };
   /// Device Id. MKLDNN is -2, CPU is -1, and GPU is 0, 1, 2 ...
   int deviceId_;
   /// Input layers

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #include "MKLDNNFcLayer.h"
 #include "paddle/utils/Logging.h"
-#include "paddle/utils/Stat.h"
 
 using namespace mkldnn;  // NOLINT
 typedef memory::format format;
@@ -40,6 +39,8 @@ bool MKLDNNFcLayer::init(const LayerMap& layerMap,
   oc_ = getSize();
   oh_ = 1;
   ow_ = 1;
+  ih_ = 1;
+  iw_ = 1;
 
   // input size can not change in FC
   iLayerSize_ = inputLayers_[0]->getSize();
@@ -77,111 +78,86 @@ void MKLDNNFcLayer::convertWeightsToPaddle() {
   wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
 }
 
-void MKLDNNFcLayer::convertOutputToOtherDevice() {
-  copyOutputInfoToOtherDevice();
-  // find other cpu device and reorder output to cpu device
-  int cnt = 0;
-  for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
-    if (outputOtherDevice_[i].deviceId == CPU_DEVICE) {
-      // fc cpu output value do not need convert
-      // just share point
-      outputOtherDevice_[i].value = output_.value;
-      ++cnt;
-    }
-  }
-
-  if (cnt > 1) {
-    LOG(WARNING) << "should not have more than one CPU devie";
-  }
-}
+void MKLDNNFcLayer::reshape(
+    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+  reshapeInput(bs, ih, iw);
 
-void MKLDNNFcLayer::reshape() {
-  const Argument& input = getInput(0, getPrev(0)->getDeviceId());
-  int batchSize = input.getBatchSize();
-  if (bs_ == batchSize) {
-    return;
-  }
-  bs_ = batchSize;
-  ih_ = input.getFrameHeight();
-  iw_ = input.getFrameWidth();
-  if (ih_ == 0) {
-    ih_ = 1;
-  }
-  if (iw_ == 0) {
-    iw_ = 1;
-  }
   CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize());
-  ic_ = iLayerSize_ / (ih_ * iw_);
-  CHECK_EQ(size_t(ic_ * ih_ * iw_), iLayerSize_) << "not divisible";
-  CHECK_EQ(size_t(oc_), getSize());
-  printSizeInfo();
+  ic = iLayerSize_ / (ih * iw);
+  CHECK_EQ(size_t(ic * ih * iw), iLayerSize_) << "not divisible";
+  CHECK_EQ(size_t(oc), getSize());
 
-  // reset output
-  output_.setFrameHeight(oh_);
-  output_.setFrameWidth(ow_);
-  resetOutput(bs_, oc_);
+  reshapeOutput(oh, ow);
+  resizeOutput(bs, oc);
 
-  // reset mkldnn forward
-  resetFwd();
-  needResetBwd_ = true;
-
-  convertWeightsFromPaddle();
+  printSizeInfo();
 }
 
-void MKLDNNFcLayer::resetFwd() {
+void MKLDNNFcLayer::resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                             MKLDNNMatrixPtr& in,
+                             MKLDNNMatrixPtr& wgt,
+                             MKLDNNMatrixPtr& bias,
+                             MKLDNNMatrixPtr& out) {
+  pipeline.clear();
   bool hasBias = biases_ && biases_->getW();
-  const MatrixPtr& wgt = weight_->getW();
-  const MatrixPtr& bias = hasBias ? biases_->getW() : nullptr;
-  const MatrixPtr& out = output_.value;
+  const MatrixPtr& wgtVal = weight_->getW();
+  const MatrixPtr& biasVal = hasBias ? biases_->getW() : nullptr;
+  const MatrixPtr& outVal = output_.value;
 
   if (inputIsOnlyMKLDNN()) {
-    const MatrixPtr& in = getInputValue(0);
-    inVal_ = std::dynamic_pointer_cast<MKLDNNMatrix>(in);
-    CHECK(inVal_) << "Input should be MKLDNNMatrix";
+    const MatrixPtr& inVal = getInputValue(0);
+    in = std::dynamic_pointer_cast<MKLDNNMatrix>(inVal);
+    CHECK(in) << "Input should be MKLDNNMatrix";
   } else {
     CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
-    const MatrixPtr& in = getInputValue(0, CPU_DEVICE);
-    inVal_ = MKLDNNMatrix::create(
-        in, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_);
-  }
-  inVal_->downSpatial();
-  wgtVal_ = MKLDNNMatrix::create(
-      wgt, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_);
-  wgtVal_->downSpatial();
-  biasVal_ =
-      hasBias ? MKLDNNMatrix::create(bias, {oc_}, format::x, engine_) : nullptr;
-  outVal_ = MKLDNNMatrix::create(out, {bs_, oc_}, format::nc, engine_);
+    const MatrixPtr& inVal = getInputValue(0, CPU_DEVICE);
+    in = MKLDNNMatrix::create(
+        inVal, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_);
+  }
+  in->downSpatial();
+  wgt = MKLDNNMatrix::create(
+      wgtVal, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_);
+  wgt->downSpatial();
+  bias = hasBias ? MKLDNNMatrix::create(biasVal, {oc_}, format::x, engine_)
+                 : nullptr;
+  out = MKLDNNMatrix::create(outVal, {bs_, oc_}, format::nc, engine_);
 
   // change original output value to mkldnn output value
-  output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
+  output_.value = std::dynamic_pointer_cast<Matrix>(out);
   if (!outputIsOnlyMKLDNN()) {
-    convertOutputToOtherDevice();
+    // fc cpu output value do not need create convert
+    // just share point
+    getOutput(CPU_DEVICE).value->setData(output_.value->getData());
   }
 
   // create forward handle
   prop_kind pk = prop_kind::forward;
   fc_fwd::desc fwdDesc = hasBias ? fc_fwd::desc(pk,
-                                                inVal_->getMemoryDesc(),
-                                                wgtVal_->getMemoryDesc(),
-                                                biasVal_->getMemoryDesc(),
-                                                outVal_->getMemoryDesc())
+                                                in->getMemoryDesc(),
+                                                wgt->getMemoryDesc(),
+                                                bias->getMemoryDesc(),
+                                                out->getMemoryDesc())
                                  : fc_fwd::desc(pk,
-                                                inVal_->getMemoryDesc(),
-                                                wgtVal_->getMemoryDesc(),
-                                                outVal_->getMemoryDesc());
+                                                in->getMemoryDesc(),
+                                                wgt->getMemoryDesc(),
+                                                out->getMemoryDesc());
   fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
   if (hasBias) {
-    fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *biasVal_, *outVal_));
+    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *bias, *out));
   } else {
-    fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *outVal_));
+    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *out));
   }
   printValueFormatFlow();
 
-  pipelineFwd_.clear();
-  pipelineFwd_.push_back(*fwd_);
+  pipeline.push_back(*fwd_);
 }
 
-void MKLDNNFcLayer::resetBwd() {
+void MKLDNNFcLayer::resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                             MKLDNNMatrixPtr& in,
+                             MKLDNNMatrixPtr& wgt,
+                             MKLDNNMatrixPtr& bias,
+                             MKLDNNMatrixPtr& out) {
+  pipeline.clear();
   if (!needResetBwd_) {
     return;
   }
@@ -190,8 +166,8 @@ void MKLDNNFcLayer::resetBwd() {
 
   /// backward weight
   CHECK(inVal_) << "Should have input value";
-  const MatrixPtr& wgt = weight_->getWGrad();
-  const MatrixPtr& bias = hasBias ? biases_->getWGrad() : nullptr;
+  const MatrixPtr& wgtGrad = weight_->getWGrad();
+  const MatrixPtr& biasGrad = hasBias ? biases_->getWGrad() : nullptr;
 
   // TODO(TJ): merge outgrad
   int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
@@ -202,101 +178,66 @@ void MKLDNNFcLayer::resetBwd() {
   // for CPU device:
   // fc do not need to convert from cpu device since output is always nc format
   // only need create from cpu device
-  const MatrixPtr& out = getOutput(device).grad;
-  outGrad_ = MKLDNNMatrix::create(out, outVal_->getPrimitiveDesc());
-  wgtGrad_ = MKLDNNMatrix::create(wgt, wgtVal_->getPrimitiveDesc());
-  biasGrad_ = hasBias ? MKLDNNMatrix::create(bias, biasVal_->getPrimitiveDesc())
-                      : nullptr;
+  const MatrixPtr& outGrad = getOutput(device).grad;
+  out = MKLDNNMatrix::create(outGrad, outVal_->getPrimitiveDesc());
+  wgt = MKLDNNMatrix::create(wgtGrad, wgtVal_->getPrimitiveDesc());
+  bias = hasBias ? MKLDNNMatrix::create(biasGrad, biasVal_->getPrimitiveDesc())
+                 : nullptr;
 
   // create memory primitive desc
   fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward,
                                       inVal_->getMemoryDesc(),
-                                      wgtGrad_->getMemoryDesc(),
-                                      outGrad_->getMemoryDesc());
+                                      wgt->getMemoryDesc(),
+                                      out->getMemoryDesc());
   fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
   fc_bwdWgt::desc bwdWgtDesc = hasBias
                                    ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgtGrad_->getMemoryDesc(),
-                                                     biasGrad_->getMemoryDesc(),
-                                                     outGrad_->getMemoryDesc())
+                                                     wgt->getMemoryDesc(),
+                                                     bias->getMemoryDesc(),
+                                                     out->getMemoryDesc())
                                    : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgtGrad_->getMemoryDesc(),
-                                                     outGrad_->getMemoryDesc());
+                                                     wgt->getMemoryDesc(),
+                                                     out->getMemoryDesc());
   fc_bwdWgt::primitive_desc bwdWgtPD =
       fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD);
 
   if (hasBias) {
-    bwdWgt_.reset(
-        new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_, *biasGrad_));
+    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt, *bias));
   } else {
-    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_));
+    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt));
   }
-  pipelineBwd_.clear();
-  pipelineBwd_.push_back(*bwdWgt_);
+  pipeline.push_back(*bwdWgt_);
 
   /// backward data
-  device = inputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
-  const MatrixPtr& in = getInputGrad(0, device);
-  if (in == nullptr) {
+  const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
+  if (inGrad == nullptr) {
     return;
   }
-  if (getInput(0, device).getAllCount() > 1) {
-    // TODO(TJ): use outputMaps_ ways when merge outgrad done
+  if (getInput(0, MKLDNN_DEVICE).getAllCount() > 1) {
+    // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
   } else {
-    inGrad_ = MKLDNNMatrix::create(in, inVal_->getPrimitiveDesc());
+    in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
   }
 
-  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(inVal_->getMemoryDesc(),
-                                                  wgtGrad_->getMemoryDesc(),
-                                                  outGrad_->getMemoryDesc());
+  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(
+      inVal_->getMemoryDesc(), wgt->getMemoryDesc(), out->getMemoryDesc());
   fc_bwdData::primitive_desc bwdDataPD =
       fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD);
 
   CHECK(wgtVal_) << "Should have weight memory";
-  bwdData_.reset(new fc_bwdData(bwdDataPD, *outGrad_, *wgtVal_, *inGrad_));
+  bwdData_.reset(new fc_bwdData(bwdDataPD, *out, *wgtVal_, *in));
   printGradFormatFlow();
-  pipelineBwd_.push_back(*bwdData_);
+  pipeline.push_back(*bwdData_);
 }
 
-void MKLDNNFcLayer::forward(PassType passType) {
-  Layer::forward(passType);
-  reshape();
-
-  {
-    REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
-    syncInputValue();
-
-    // just submit forward pipeline
-    stream_->submit(pipelineFwd_);
-  }
-
-  /* activation */ {
-    REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
-    forwardActivation();
-  }
+void MKLDNNFcLayer::updateInputData() {
+  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
 }
 
-void MKLDNNFcLayer::backward(const UpdateCallback& callback) {
-  /* Do derivation */ {
-    REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
-    backwardActivation();
-  }
-
-  {
-    REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
-    resetBwd();
-
-    syncOutputGrad();
-    // just sumbmit backward pipeline
-    stream_->submit(pipelineBwd_);
-  }
-
-  {
-    REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
-    weight_->getParameterPtr()->incUpdate(callback);
-    if (biases_ && biases_->getWGrad()) {
-      biases_->getParameterPtr()->incUpdate(callback);
-    }
+void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
+  weight_->getParameterPtr()->incUpdate(callback);
+  if (biases_ && biases_->getWGrad()) {
+    biases_->getParameterPtr()->incUpdate(callback);
   }
 }
 }  // namespace paddle

paddle/gserver/layers/MKLDNNFcLayer.h

@@ -45,35 +45,28 @@ public:
   bool init(const LayerMap& layerMap,
             const ParameterMap& parameterMap) override;
 
-  void convertWeightsFromPaddle() override;
+  void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
 
-  void convertWeightsToPaddle() override;
+  void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
 
-  void forward(PassType passType) override;
+  void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
 
-  void backward(const UpdateCallback& callback) override;
+  void updateInputData() override;
 
-protected:
-  /**
-   * reshape the input image sizes
-   * and reset output buffer size
-   * and reset mkldnn forward
-   */
-  void reshape();
-
-  /**
-   * reset the forward primitve and memory
-   * only would be called when input size changes
-   */
-  void resetFwd();
-
-  /**
-   * reset the backward primitve and memory for mkldnn fc
-   * only would be called when needed
-   */
-  void resetBwd();
-
-  void convertOutputToOtherDevice() override;
+  void updateWeights(const UpdateCallback& callback) override;
+
+  void convertWeightsFromPaddle() override;
+
+  void convertWeightsToPaddle() override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/MKLDNNLayer.h

@@ -19,6 +19,7 @@ limitations under the License. */
 #include "MKLDNNBase.h"
 #include "mkldnn.hpp"
 #include "paddle/math/MKLDNNMatrix.h"
+#include "paddle/utils/Stat.h"
 
 DECLARE_bool(use_mkldnn);
 
@@ -33,6 +34,8 @@ typedef std::shared_ptr<MKLDNNLayer> MKLDNNLayerPtr;
  */
 class MKLDNNLayer : public Layer {
 protected:
+  // input value element count
+  size_t inputElemenCnt_;
   // batch size
   int bs_;
   // input image channel, height and width
@@ -52,7 +55,7 @@ protected:
   std::vector<mkldnn::primitive> pipelineFwd_;
   std::vector<mkldnn::primitive> pipelineBwd_;
 
-  // MKLDNNMatrixPtr
+  // MKLDNNMatrixPtr with internal format
   MKLDNNMatrixPtr inVal_;
   MKLDNNMatrixPtr inGrad_;
   MKLDNNMatrixPtr outVal_;
@@ -65,6 +68,7 @@ protected:
 public:
   explicit MKLDNNLayer(const LayerConfig& config)
       : Layer(config),
+        inputElemenCnt_(0),
         bs_(0),
         ic_(0),
         ih_(0),
@@ -95,12 +99,104 @@ public:
     if (!Layer::init(layerMap, parameterMap)) {
       return false;
     }
+    checkCPUOutputsNumber();
 
     stream_.reset(new MKLDNNStream());
     engine_ = CPUEngine::Instance().getEngine();
     return true;
   }
 
+  void forward(PassType passType) override {
+    passType_ = passType;
+
+    {
+      REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
+      CHECK(!inputLayers_.empty());
+      copySeqInfoToOutputs();
+      size_t elemenCnt = inputLayers_[0]->getOutput().value->getElementCnt();
+      if (inputElemenCnt_ != elemenCnt) {
+        // reset when input total sizes changed, not only the batchsize
+        inputElemenCnt_ = elemenCnt;
+        reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_);
+        resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_);
+        convertWeightsFromPaddle();
+        needResetBwd_ = true;
+      }
+
+      if (inputLayers_[0]->getType() == "data") {
+        updateInputData();
+      }
+
+      stream_->submit(pipelineFwd_);
+    }
+
+    /* activation */ {
+      REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
+      forwardActivation();
+    }
+  }
+
+  void backward(const UpdateCallback& callback) override {
+    /* Do derivation */ {
+      REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
+      backwardActivation();
+    }
+
+    {
+      REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
+      if (needResetBwd_) {
+        resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_);
+        needResetBwd_ = false;
+      }
+
+      stream_->submit(pipelineBwd_);
+    }
+
+    {
+      REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
+      updateWeights(callback);
+    }
+  }
+
+  /**
+   * reshape the input image sizes
+   * and reset output image and buffer size
+   * output channel can not be changed
+   */
+  virtual void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) = 0;
+
+  /**
+   * reset the mkldnn forward primitve and memory
+   * only would be called when input size changes
+   */
+  virtual void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out) = 0;
+
+  /**
+   * reset the mkldnn backward primitve and memory for mkldnn fc
+   * only would be called when needed
+   */
+  virtual void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out) = 0;
+
+  /**
+   * Update input value data when input layer is "data" type.
+   * Since the input value data address might be changed.
+   */
+  virtual void updateInputData() {}
+
+  /**
+   * Update weights and biases if necessary.
+   */
+  virtual void updateWeights(const UpdateCallback& callback) {}
+
   /**
    * convert weight from paddle format to mkldnn format
    * weight_ will be override
@@ -114,10 +210,38 @@ public:
   virtual void convertWeightsToPaddle() {}
 
   /**
-   * convert MKLDNN output to other device.
-   * only support CPU device yet
+   * add this interface as public for unit test
+   */
+  void addOutputArgument(int deviceId) { Layer::addOutputArgument(deviceId); }
+
+protected:
+  /**
+   * reshape the input image sizes and input batchsize
    */
-  virtual void convertOutputToOtherDevice() {}
+  virtual void reshapeInput(int& batchsize, int& height, int& width) {
+    const Argument& input = inputLayers_[0]->getOutput();
+    batchsize = input.getBatchSize();
+    int h = input.getFrameHeight();
+    int w = input.getFrameWidth();
+    if (h != 0) {
+      height = h;
+    }
+    if (w != 0) {
+      width = w;
+    }
+  }
+
+  /**
+   * reshape output image sizes
+   */
+  virtual void reshapeOutput(size_t height, size_t width) {
+    output_.setFrameHeight(height);
+    output_.setFrameWidth(width);
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      outputOtherDevice_[i].setFrameHeight(height);
+      outputOtherDevice_[i].setFrameWidth(width);
+    }
+  }
 
   /**
    * print info about sizes
@@ -133,8 +257,8 @@ public:
    */
   virtual void printValueFormatFlow() {
     if (inVal_ && outVal_) {
-      VLOG(MKLDNN_FMTS) << "value format flow --- " << inVal_->getFormat()
-                        << " >>> " << outVal_->getFormat();
+      VLOG(MKLDNN_FMTS) << inVal_->getFormat() << " >>> "
+                        << outVal_->getFormat();
     }
   }
 
@@ -143,29 +267,12 @@ public:
    */
   virtual void printGradFormatFlow() {
     if (inGrad_ && outGrad_) {
-      VLOG(MKLDNN_FMTS) << "grad format flow --- " << inGrad_->getFormat()
-                        << " <<< " << outGrad_->getFormat();
+      VLOG(MKLDNN_FMTS) << inGrad_->getFormat() << " <<< "
+                        << outGrad_->getFormat();
     }
   }
 
 protected:
-  /**
-   * copy image size and sequence info to other device
-   * @note: can not directly use Layer::copyOutputToOtherDevice since here only
-   *        copy base info and do not copy data value
-   */
-  void copyOutputInfoToOtherDevice() {
-    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
-      outputOtherDevice_[i].setFrameHeight(output_.getFrameHeight());
-      outputOtherDevice_[i].setFrameWidth(output_.getFrameWidth());
-      outputOtherDevice_[i].sequenceStartPositions =
-          output_.sequenceStartPositions;
-      outputOtherDevice_[i].subSequenceStartPositions =
-          output_.subSequenceStartPositions;
-      outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;
-    }
-  }
-
   /**
    * If input only has MKLDNN device.
    * Otherwise, only support the previous layer using CPU device.
@@ -193,37 +300,12 @@ protected:
     return outputOtherDevice_.size() == 0;
   }
 
-  /**
-   * Sync input value data
-   */
-  void syncInputValue() {
-    if (inputIsOnlyMKLDNN()) {
-      return;
-    }
-    real* iData = getInputValue(0, CPU_DEVICE)->getData();
-    // update input data
-    // since it might be changed if this is after data layer
-    inVal_->updateData(iData);
-  }
-
-  /**
-   * Sync output grad data
-   */
-  void syncOutputGrad() {
-    if (outputIsOnlyMKLDNN()) {
-      return;
-    }
-
-    // update diff
-    real* oDiff = getOutput(CPU_DEVICE).grad->getData();
-    outGrad_->updateData(oDiff);
-  }
-
   /**
    * Set deviceId of this layer.
    */
   void setDevice(int id) { deviceId_ = id; }
 
+private:
   /**
    * Set deviceId of the params used in this layer.
    */
@@ -247,6 +329,42 @@ protected:
       parameter->setDevice(id);
     }
   }
+
+  /**
+   * Check the cpu device number of outputOtherDevice_.
+   * should have only one at most.
+   */
+  void checkCPUOutputsNumber(int max = 1) {
+    int cnt = 0;
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      if (outputOtherDevice_[i].deviceId == CPU_DEVICE) {
+        ++cnt;
+      }
+    }
+    CHECK_LE(cnt, max) << "too much CPU devies";
+  }
+
+  /**
+   * copy SeqInfo from input layer to this output and other output devices.
+   * @note: do not use getInput(0) since it used this deviceId_,
+   *        use "inputLayers_[0]->getOutput()" instead.
+   */
+  void copySeqInfoToOutputs() {
+    if (inputLayers_.empty() || !needSequenceInfo_) {
+      return;
+    }
+    const Argument& input = inputLayers_[0]->getOutput();
+    output_.sequenceStartPositions = input.sequenceStartPositions;
+    output_.subSequenceStartPositions = input.subSequenceStartPositions;
+    output_.cpuSequenceDims = input.cpuSequenceDims;
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      outputOtherDevice_[i].sequenceStartPositions =
+          output_.sequenceStartPositions;
+      outputOtherDevice_[i].subSequenceStartPositions =
+          output_.subSequenceStartPositions;
+      outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;
+    }
+  }
 };
 
 }  // namespace paddle

paddle/gserver/layers/SwitchOrderLayer.cpp

@@ -24,10 +24,12 @@ bool SwitchOrderLayer::init(const LayerMap& layerMap,
   /* Initialize the basic parent class */
   Layer::init(layerMap, parameterMap);
   auto& img_conf = config_.inputs(0).image_conf();
+  size_t inD = img_conf.img_size_z();
   size_t inH =
       img_conf.has_img_size_y() ? img_conf.img_size_y() : img_conf.img_size();
   size_t inW = img_conf.img_size();
   size_t inC = img_conf.channels();
+  inH = inH * inD;
   inDims_ = TensorShape({0, inC, inH, inW});
   outDims_ = TensorShape(4);
 
@@ -64,9 +66,10 @@ void SwitchOrderLayer::setInDims() {
   MatrixPtr input = inputLayers_[0]->getOutputValue();
   size_t batchSize = input->getHeight();
   inDims_.setDim(0, batchSize);
-
+  int d = inputLayers_[0]->getOutput().getFrameDepth();
+  d = (d == 0 ? 1 : d);
   int h = inputLayers_[0]->getOutput().getFrameHeight();
-  if (h != 0) inDims_.setDim(2, h);
+  if (h != 0) inDims_.setDim(2, h * d);
   int w = inputLayers_[0]->getOutput().getFrameWidth();
   if (w != 0) inDims_.setDim(3, w);
   int totalCount = input->getElementCnt();

paddle/gserver/tests/MKLDNNTester.cpp

@@ -63,8 +63,12 @@ void MKLDNNTester::reset(const TestConfig& dnn,
     initTestLayer(
         configs_[i], &(layerMaps_[i]), &(parameters_[i]), &(testLayers_[i]));
   }
-  dnnLayer_ = testLayers_[DNN];
   refLayer_ = testLayers_[REF];
+  dnnLayer_ = std::dynamic_pointer_cast<MKLDNNLayer>(testLayers_[DNN]);
+  CHECK(dnnLayer_);
+  // for comparison with Paddle reference results,
+  // need manually add cpu device output for test
+  dnnLayer_->addOutputArgument(CPU_DEVICE);
   EXPECT_EQ(dataLayers_[DNN].size(), dataLayers_[REF].size());
   EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size());
 
@@ -109,20 +113,22 @@ void MKLDNNTester::randomBotDatas() {
 
 void MKLDNNTester::randomTopDiffs() {
   refLayer_->getOutputGrad()->randomizeUniform();
-  dnnLayer_->getOutputGrad()->copyFrom(*(refLayer_->getOutputGrad()));
-  VLOG(lvl_) << "Random dom Backward Input, TopDiff: ";
+  dnnLayer_->getOutput(CPU_DEVICE)
+      .grad->copyFrom(*(refLayer_->getOutputGrad()));
+  VLOG(lvl_) << "Random Backward Input, TopDiff: ";
   printMatrix(refLayer_->getOutputGrad());
 }
 
 void MKLDNNTester::checkForward() {
-  printTopDatas();
-  double delta = compareMatrix(testLayers_[DNN]->getOutputValue(),
-                               testLayers_[REF]->getOutputValue());
   VLOG(MKLDNN_ALL) << "Check Forward";
+  printTopDatas();
+  double delta = compareMatrix(dnnLayer_->getOutput(-1).value,
+                               refLayer_->getOutputValue());
   EXPECT_LE(fabs(delta), eps_);
 }
 
 void MKLDNNTester::checkBackwardData() {
+  VLOG(MKLDNN_ALL) << "Check Backward Data";
   // TODO(TJ): uncomment me when batch norm ready
   // const bool isBN = dnnLayer_->getType() == "mkldnn_batch_norm";
   for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) {
@@ -144,14 +150,12 @@ void MKLDNNTester::checkBackwardData() {
 }
 
 void MKLDNNTester::checkBackwardWgts() {
+  VLOG(MKLDNN_ALL) << "Check Backward Weight";
   CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size());
   vector<VectorPtr> dnnWgts;  // used to temply save mkldnn weights
   saveWgt(parameters_[DNN], dnnWgts);
 
-  const MKLDNNLayerPtr dnnlayer =
-      std::dynamic_pointer_cast<MKLDNNLayer>(dnnLayer_);
-  CHECK(dnnlayer);
-  dnnlayer->convertWeightsToPaddle();
+  dnnLayer_->convertWeightsToPaddle();
   for (size_t i = 0; i < parameters_[DNN].size(); ++i) {
     const VectorPtr& dnn = parameters_[DNN][i]->getBuf(PARAMETER_VALUE);
     const VectorPtr& ref = parameters_[REF][i]->getBuf(PARAMETER_VALUE);
@@ -189,38 +193,38 @@ void MKLDNNTester::restoreWgt(const vector<VectorPtr>& from,
 }
 
 // clear parameters grad
-void MKLDNNTester::clearWgtDiffs() {
+void MKLDNNTester::clearWgtDiffs(size_t id) {
+  CHECK_LE(id, parameters_.size());
   for (size_t n = 0; n < parameters_.size(); ++n) {
-    for (size_t i = 0; i < parameters_[n].size(); ++i) {
-      const VectorPtr& grad = parameters_[n][i]->getBuf(PARAMETER_GRADIENT);
-      if (grad) {
-        grad->zeroMem();
+    if (id == n || id == parameters_.size()) {
+      for (size_t i = 0; i < parameters_[n].size(); ++i) {
+        const VectorPtr& grad = parameters_[n][i]->getBuf(PARAMETER_GRADIENT);
+        if (grad) {
+          grad->zeroMem();
+        }
       }
     }
   }
 }
 
-void MKLDNNTester::clearBotDiffs() {
-  // dnn and ref
+void MKLDNNTester::clearBotDiffs(size_t id) {
+  CHECK_LE(id, dataLayers_.size());
   for (size_t n = 0; n < dataLayers_.size(); ++n) {
-    // all inputs layers
-    for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
-      dataLayers_[n][i]->getOutputGrad()->zeroMem();
+    if (id == n || id == dataLayers_.size()) {
+      // clear inputs layers of this specific layer
+      for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
+        dataLayers_[n][i]->getOutputGrad()->zeroMem();
+      }
     }
   }
 }
 
-void MKLDNNTester::clearBotDiffs(int n) {
-  CHECK_LT(n, NUM);
-  // all inputs layers
-  for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
-    dataLayers_[n][i]->getOutputGrad()->zeroMem();
-  }
-}
-
-void MKLDNNTester::clearTopDatas() {
+void MKLDNNTester::clearTopDatas(size_t id) {
+  CHECK_LE(id, testLayers_.size());
   for (size_t i = 0; i < testLayers_.size(); ++i) {
-    testLayers_[i]->getOutputValue()->zeroMem();
+    if (id == i || id == testLayers_.size()) {
+      testLayers_[i]->getOutputValue()->zeroMem();
+    }
   }
 }
 
@@ -300,16 +304,24 @@ void MKLDNNTester::runOnce() {
   checkForward();
 
   // test backward
+  // simple updater
+  UpdateCallback updateCallback = [](Parameter* para) {
+    auto& grad = para->getBuf(PARAMETER_GRADIENT);
+    auto& value = para->getBuf(PARAMETER_VALUE);
+    real lr = 1e-3;
+    value->add(*grad, lr);
+  };
   randomTopDiffs();
-  dnnLayer_->backward(nullptr);
-  refLayer_->backward(nullptr);
+  dnnLayer_->backward(updateCallback);
+  refLayer_->backward(updateCallback);
   checkBackwardData();
   checkBackwardWgts();
 
   // clear buffers
   // ref code will addto the diff, dnn code will writeto it
-  // and clearTopDatas() and clearWgtDiffs() should be coverd by test layers
+  // and clearTopDatas(REF) should be coverd by ref layers
   clearBotDiffs(REF);
+  clearWgtDiffs(REF);
 }
 
 void MKLDNNTester::run(const TestConfig& dnn,

paddle/gserver/tests/MKLDNNTester.h

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <vector>
 #include "LayerGradUtil.h"
 #include "paddle/gserver/layers/MKLDNNBase.h"
+#include "paddle/gserver/layers/MKLDNNLayer.h"
 
 namespace paddle {
 
@@ -40,7 +41,8 @@ protected:
   vector<LayerMap> layerMaps_;
   vector<vector<ParameterPtr>> parameters_;
   vector<LayerPtr> testLayers_;
-  LayerPtr dnnLayer_, refLayer_;
+  LayerPtr refLayer_;
+  MKLDNNLayerPtr dnnLayer_;
 
   /// run some iterations, all the result should pass
   size_t iter_;
@@ -88,10 +90,10 @@ private:
   void checkBackwardData();
   void checkBackwardWgts();
 
-  void clearWgtDiffs();
-  void clearBotDiffs();
-  void clearBotDiffs(int n);  // clear specific layer
-  void clearTopDatas();
+  // clear specific layer, clear all when id equals NUM
+  void clearWgtDiffs(size_t id = NUM);
+  void clearBotDiffs(size_t id = NUM);
+  void clearTopDatas(size_t id = NUM);
 
   void printTopDatas();
   void printMatrix(const MatrixPtr& m);

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1703,6 +1703,55 @@ TEST(Layer, BatchNormalizationLayer) {
 #endif
 }
 
+void testBatchNorm3DLayer(const string& type, bool trans, bool useGpu) {
+  TestConfig config;
+  const int CHANNELS = 10;
+  const int IMG_SIZE = 16;
+  const int IMG_SIZE_Y = 8;
+  const int IMG_SIZE_Z = 8;
+  size_t size = CHANNELS * IMG_SIZE * IMG_SIZE_Y * IMG_SIZE_Z;
+  config.layerConfig.set_type(type);
+  config.layerConfig.set_size(size);
+  config.layerConfig.set_active_type("sigmoid");
+  config.biasSize = CHANNELS;
+  config.inputDefs.push_back({INPUT_DATA,
+                              "layer_0",
+                              /* dim= */ size,
+                              /* paraSize= */ CHANNELS});
+
+  config.inputDefs.push_back({INPUT_DATA, "layer_1_running_mean", 1, CHANNELS});
+  config.inputDefs.back().isStatic = true;
+  config.inputDefs.push_back({INPUT_DATA, "layer_2_running_var", 1, CHANNELS});
+  config.inputDefs.back().isStatic = true;
+
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  config.layerConfig.add_inputs();
+  config.layerConfig.add_inputs();
+
+  ImageConfig* img_conf = input->mutable_image_conf();
+  img_conf->set_channels(CHANNELS);
+  img_conf->set_img_size(IMG_SIZE);
+  img_conf->set_img_size_y(IMG_SIZE_Y);
+  img_conf->set_img_size_z(IMG_SIZE_Z);
+
+  testLayerGrad(config,
+                "batch_norm",
+                64,
+                /* trans= */ trans,
+                useGpu,
+                /* useWeight */ true);
+}
+
+TEST(Layer, testBatchNorm3DLayer) {
+  testBatchNorm3DLayer("batch_norm", false, false);
+#ifndef PADDLE_ONLY_CPU
+  testBatchNorm3DLayer("batch_norm", false, true);
+  if (hl_get_cudnn_lib_version() >= int(4000)) {
+    testBatchNorm3DLayer("cudnn_batch_norm", false, true);
+  }
+#endif
+}
+
 void testConvOperator(bool isDeconv) {
   TestConfig config;
   const int NUM_FILTERS = 16;
@@ -2253,26 +2302,27 @@ void test3DDeConvLayer(const string& type, bool trans, bool useGpu) {
   conv->set_stride(2);
   conv->set_stride_y(2);
   conv->set_stride_z(2);
-  conv->set_img_size(IMAGE_SIZE);
-  conv->set_img_size_y(IMAGE_SIZE_Y);
-  conv->set_img_size_z(IMAGE_SIZE_Z);
-  conv->set_output_x(imageSize(conv->img_size(),
+  conv->set_output_x(IMAGE_SIZE);
+  conv->set_output_y(IMAGE_SIZE_Y);
+  conv->set_output_z(IMAGE_SIZE_Z);
+
+  conv->set_img_size(imageSize(conv->output_x(),
                                conv->filter_size(),
                                conv->padding(),
                                conv->stride(),
                                true));
-  conv->set_output_y(imageSize(conv->img_size_y(),
-                               conv->filter_size_y(),
-                               conv->padding_y(),
-                               conv->stride_y(),
-                               true));
-  conv->set_output_z(imageSize(conv->img_size_z(),
-                               conv->filter_size_z(),
-                               conv->padding_z(),
-                               conv->stride_z(),
-                               true));
-  config.layerConfig.set_size(conv->output_x() * conv->output_y() *
-                              conv->output_z() * NUM_FILTERS);
+  conv->set_img_size_y(imageSize(conv->output_y(),
+                                 conv->filter_size_y(),
+                                 conv->padding_y(),
+                                 conv->stride_y(),
+                                 true));
+  conv->set_img_size_z(imageSize(conv->output_z(),
+                                 conv->filter_size_z(),
+                                 conv->padding_z(),
+                                 conv->stride_z(),
+                                 true));
+  config.layerConfig.set_size(conv->img_size() * conv->img_size_y() *
+                              conv->img_size_z() * NUM_FILTERS);
   conv->set_groups(1);
   conv->set_filter_channels(conv->channels() / conv->groups());
   config.inputDefs.push_back(

paddle/math/MKLDNNMatrix.cpp

@@ -33,14 +33,12 @@ MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) {
     size_t width = cnts / dims[0];
     m = Matrix::create(height, width, false, false);
   }
-
   CHECK(m) << " Matrix should not be empty";
+
   CpuMatrixPtr cpuMatrix = std::dynamic_pointer_cast<CpuMatrix>(m);
   CHECK(cpuMatrix) << "Only support create from CPU matrix yet";
-
-  CHECK_EQ(cnts, m->getElementCnt()) << "Count size does not match";
-  return std::make_shared<MKLDNNMatrix>(
-      m->getData(), m->getHeight(), m->getWidth(), pd);
+  CHECK_EQ(cpuMatrix->getElementCnt(), cnts) << "Count size does not match";
+  return std::make_shared<MKLDNNMatrix>(cpuMatrix, pd);
 }
 
 MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m,
@@ -138,7 +136,7 @@ void MKLDNNMatrix::downSpatial() {
       mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr),
       "could not create a memory primitive");
   reset(result);
-  set_data_handle(getData());
+  set_data_handle(data_);
 }
 
 }  // namespace paddle

paddle/math/MKLDNNMatrix.h

@@ -30,11 +30,10 @@ typedef std::shared_ptr<MKLDNNMatrix> MKLDNNMatrixPtr;
  */
 class MKLDNNMatrix : public CpuMatrix, public mkldnn::memory {
 public:
-  MKLDNNMatrix(real* data,
-               size_t height,
-               size_t width,
-               mkldnn::memory::primitive_desc pd)
-      : CpuMatrix(data, height, width, false), mkldnn::memory(pd, data) {}
+  MKLDNNMatrix(CpuMatrixPtr m, mkldnn::memory::primitive_desc pd)
+      : CpuMatrix(m->getData(), m->getHeight(), m->getWidth(), false),
+        mkldnn::memory(pd, m->getData()),
+        m_(m) {}
 
   ~MKLDNNMatrix() {}
 
@@ -81,11 +80,29 @@ public:
   void downSpatial();
 
   /**
-   * Update the memory data handle.
+   * set the memory data handle.
    * Caution: This will not check the buffer size of the data,
    *          it should be coverd by user.
    */
-  void updateData(void* data) { set_data_handle(data); }
+  void setData(real* data) {
+    set_data_handle(data);
+    CpuMatrix::setData(data);
+    m_.reset();
+  }
+
+  /**
+   * override Matrix::getData
+   * check data before return
+   */
+  real* getData() override {
+    CHECK_EQ((void*)data_, get_data_handle());
+    return data_;
+  }
+
+  const real* getData() const override {
+    CHECK_EQ((void*)data_, get_data_handle());
+    return data_;
+  }
 
   /**
    * Get primitive descriptor.
@@ -143,6 +160,10 @@ protected:
                    memory::format srcFmt,
                    memory::format dstFmt,
                    memory::dims dm);
+
+private:
+  // save the CpuMatrixPtr in case the buffer released outside
+  CpuMatrixPtr m_;
 };
 
 }  // namespace paddle

paddle/operators/concat_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/concat_op.h"
+#include <vector>
+
+namespace paddle {
+namespace operators {
+using framework::Tensor;
+
+class ConcatOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto ins = ctx.MultiInput<framework::Tensor>("X");
+    auto *out = ctx.Output<framework::Tensor>("Out");
+    size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
+    size_t n = ins.size();
+
+    PADDLE_ENFORCE_GT(n, 1, "Input tensors count should > 1.");
+
+    auto out_dims = ins[0]->dims();
+    size_t in_zero_dims_size = out_dims.size();
+    for (size_t i = 1; i < n; i++) {
+      for (size_t j = 0; j < in_zero_dims_size; j++) {
+        if (j == axis) {
+          out_dims[axis] += ins[i]->dims()[j];
+          continue;
+        }
+        PADDLE_ENFORCE_EQ(out_dims[j], ins[i]->dims()[j],
+                          "Input tensors should have the same "
+                          "elements except the specify axis.")
+      }
+    }
+    out->Resize(out_dims);
+  }
+};
+
+class ConcatOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ConcatOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "the input tensors of concat operator.").AsDuplicable();
+    AddOutput("Out", "the output tensor of concat operator.");
+    AddComment(R"DOC(
+            Join the input tensors along with the axis.
+            Examples:
+              Input[0] = [[1,2],[3,4]]
+              Input[1] = [[5,6]]
+              axis = 0
+              Output = [[1,2],
+                        [3,4],
+                        [5,6]]
+        )DOC");
+    AddAttr<int>("axis", "The axis which the inputs will be joined with.")
+        .SetDefault(0);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(concat, ops::ConcatOp, ops::ConcatOpMaker)
+REGISTER_OP_CPU_KERNEL(concat,
+                       ops::ConcatKernel<paddle::platform::CPUPlace, float>)

paddle/operators/concat_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/concat_op.h"
+
+namespace ops = paddle::operators;
+// TODO(Yancey1989) Add GPU kernel

paddle/operators/concat_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <vector>
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class ConcatKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto ins = ctx.MultiInput<framework::Tensor>("X");
+    auto* out = ctx.Output<framework::Tensor>("Out");
+    int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
+    size_t n = ins.size();
+    size_t output_axis_dim = 0;
+    size_t before = 1, after = 1;
+    for (size_t i = 0; i < n; i++) {
+      output_axis_dim += ins[i]->dims()[axis];
+    }
+    auto& input_zero = ins[0];
+    for (int64_t i = 0; i < input_zero->dims().size(); i++) {
+      if (i == axis) {
+        continue;
+      }
+      if (i < axis) {
+        before *= input_zero->dims()[i];
+      } else {
+        after *= input_zero->dims()[i];
+      }
+    }
+    size_t output_offset = 0;
+    for (size_t i = 0; i < n; i++) {
+      auto& in = ins[i];
+      auto axis_dim = in->dims()[axis];
+      for (size_t j = 0; j < before; j++) {
+        size_t len = axis_dim * after * sizeof(T);
+        const T* src = in->data<T>() + axis_dim * after * j;
+        T* out_data = out->mutable_data<T>(platform::CPUPlace());
+        T* dest = out_data + output_offset + output_axis_dim * after * j;
+        memcpy(dest, src, len);
+      }
+      output_offset += axis_dim * after;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/cos_sim_op.h

@@ -42,7 +42,7 @@ class CosSimKernel : public framework::OpKernel {
     output_y_norm->mutable_data<T>(context.GetPlace());
 
     auto dims = input_x->dims();
-    int size = static_cast<int>(framework::product(dims));
+    int64_t size = input_x->numel();
     auto new_dims = framework::make_ddim({dims[0], size / dims[0]});
     auto x = EigenMatrix<T>::From(*input_x, new_dims);
     auto y = EigenMatrix<T>::From(*input_y, new_dims);
@@ -72,7 +72,7 @@ class CosSimGradKernel : public framework::OpKernel {
     auto* input_grad_z = context.Input<Tensor>(framework::GradVarName("Out"));
 
     auto dims = input_x->dims();
-    int size = static_cast<int>(framework::product(dims));
+    int64_t size = input_x->numel();
     auto new_dims = framework::make_ddim({dims[0], size / dims[0]});
     auto x = EigenMatrix<T>::From(*input_x, new_dims);
     auto y = EigenMatrix<T>::From(*input_y, new_dims);

paddle/operators/gaussian_random_op.cc

@@ -31,7 +31,7 @@ class CPUGaussianRandomKernel : public framework::OpKernel {
     }
     engine.seed(seed);
     std::normal_distribution<T> dist(mean, std);
-    int64_t size = framework::product(tensor->dims());
+    int64_t size = tensor->numel();
     for (int64_t i = 0; i < size; ++i) {
       data[i] = dist(engine);
     }

paddle/operators/gaussian_random_op.cu

@@ -50,8 +50,8 @@ class GPUGaussianRandomKernel : public framework::OpKernel {
     T mean = static_cast<T>(context.Attr<float>("mean"));
     T std = static_cast<T>(context.Attr<float>("std"));
     thrust::counting_iterator<unsigned int> index_sequence_begin(0);
-    ssize_t N = framework::product(tensor->dims());
-    thrust::transform(index_sequence_begin, index_sequence_begin + N,
+    int64_t size = tensor->numel();
+    thrust::transform(index_sequence_begin, index_sequence_begin + size,
                       thrust::device_ptr<T>(data),
                       GaussianGenerator<T>(mean, std, seed));
   }

paddle/operators/lookup_table_op.cu

@@ -70,7 +70,7 @@ class LookupTableCUDAKernel : public framework::OpKernel {
 
     size_t N = table_t->dims()[0];
     size_t D = table_t->dims()[1];
-    size_t K = product(ids_t->dims());
+    size_t K = ids_t->numel();
     auto ids = ids_t->data<int32_t>();
     auto table = table_t->data<T>();
     auto output = output_t->mutable_data<T>(context.GetPlace());
@@ -91,7 +91,7 @@ class LookupTableGradCUDAKernel : public framework::OpKernel {
 
     int N = d_table_t->dims()[0];
     int D = d_table_t->dims()[1];
-    int K = product(ids_t->dims());
+    int K = ids_t->numel();
     const int32_t* ids = ids_t->data<int32_t>();
     const T* d_output = d_output_t->data<T>();
     T* d_table = d_table_t->mutable_data<T>(context.GetPlace());

paddle/operators/lookup_table_op.h

@@ -35,7 +35,7 @@ class LookupTableKernel : public framework::OpKernel {
     auto ids = ids_t->data<int32_t>();
     auto table = table_t->data<T>();
     auto output = output_t->mutable_data<T>(context.GetPlace());
-    for (ssize_t i = 0; i < product(ids_t->dims()); ++i) {
+    for (int64_t i = 0; i < ids_t->numel(); ++i) {
       PADDLE_ENFORCE_LT(ids[i], N);
       PADDLE_ENFORCE_GE(ids[i], 0);
       memcpy(output + i * D, table + ids[i] * D, D * sizeof(T));
@@ -61,7 +61,7 @@ class LookupTableGradKernel : public framework::OpKernel {
     t.device(context.GetEigenDevice<platform::CPUPlace>()) =
         t.constant(static_cast<T>(0));
 
-    for (ssize_t i = 0; i < product(ids_t->dims()); ++i) {
+    for (int64_t i = 0; i < ids_t->numel(); ++i) {
       PADDLE_ENFORCE_LT(ids[i], N);
       PADDLE_ENFORCE_GE(ids[i], 0);
       for (int j = 0; j < D; ++j) {

paddle/operators/math/im2col_test.cc

@@ -119,4 +119,4 @@ TEST(math, im2col) {
 #ifndef PADDLE_ONLY_CPU
   testIm2col<paddle::platform::GPUPlace>();
 #endif
-}
\ No newline at end of file
+}

paddle/operators/mean_op.h

@@ -49,12 +49,11 @@ class MeanGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto OG = context.Input<Tensor>(framework::GradVarName("Out"));
-    PADDLE_ENFORCE(framework::product(OG->dims()) == 1,
-                   "Mean Gradient should be scalar");
+    PADDLE_ENFORCE(OG->numel() == 1, "Mean Gradient should be scalar");
     auto IG = context.Output<Tensor>(framework::GradVarName("X"));
     IG->mutable_data<T>(context.GetPlace());
 
-    T ig_size = (T)framework::product(IG->dims());
+    T ig_size = static_cast<T>(IG->numel());
     Eigen::DSizes<int, 1> bcast(ig_size);
 
     EigenVector<T>::Flatten(*IG).device(context.GetEigenDevice<Place>()) =

paddle/operators/minus_op.cc

@@ -31,8 +31,7 @@ class MinusOp : public framework::OperatorWithKernel {
     auto *right_tensor = ctx.Input<framework::Tensor>("Y");
 
     PADDLE_ENFORCE_EQ(
-        framework::product(left_tensor->dims()),
-        framework::product(right_tensor->dims()),
+        left_tensor->numel(), right_tensor->numel(),
         "Minus operator must take two tensor with same num of elements");
     ctx.Output<framework::Tensor>("Out")->Resize(left_tensor->dims());
   }

paddle/operators/mul_op.cc

@@ -25,18 +25,27 @@ class MulOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto dim0 = ctx.Input<Tensor>("X")->dims();
-    auto dim1 = ctx.Input<Tensor>("Y")->dims();
-    PADDLE_ENFORCE_EQ(dim0.size(), 2,
-                      "input X(%s) should be a tensor with 2 dims, a matrix",
-                      ctx.op().Input("X"));
-    PADDLE_ENFORCE_EQ(dim1.size(), 2,
-                      "input Y(%s) should be a tensor with 2 dims, a matrix",
-                      ctx.op().Input("Y"));
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+    int x_num_col_dims = Attr<int>("x_num_col_dims");
+    int y_num_col_dims = Attr<int>("y_num_col_dims");
+
+    PADDLE_ENFORCE(x_dims.size() > x_num_col_dims,
+                   "The rank of input tensor X(%s) should be larger than "
+                   "`mul_op`'s `x_num_col_dims`.",
+                   ctx.op().Input("X"));
+    PADDLE_ENFORCE(y_dims.size() > y_num_col_dims,
+                   "The rank of input tensor Y(%s) should be larger than "
+                   "`mul_op`'s `y_num_col_dims`.",
+                   ctx.op().Input("Y"));
+
+    auto x_mat_dims = framework::flatten_to_2d(x_dims, x_num_col_dims);
+    auto y_mat_dims = framework::flatten_to_2d(y_dims, y_num_col_dims);
+
     PADDLE_ENFORCE_EQ(
-        dim0[1], dim1[0],
+        x_mat_dims[1], y_mat_dims[0],
         "First matrix's width must be equal with second matrix's height.");
-    ctx.Output<Tensor>("Out")->Resize({dim0[0], dim1[1]});
+    ctx.Output<Tensor>("Out")->Resize({x_mat_dims[0], y_mat_dims[1]});
   }
 };
 
@@ -47,6 +56,23 @@ class MulOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "The first input of mul op");
     AddInput("Y", "The second input of mul op");
     AddOutput("Out", "The output of mul op");
+    AddAttr<int>(
+        "x_num_col_dims",
+        R"DOC(mul_op can take tensors with more than two dimensions as input `X`, 
+            in that case, tensors will be reshaped to a matrix. The matrix's first 
+            dimension(column length) will be the product of tensor's last 
+            `num_col_dims` dimensions, and the matrix's second dimension(row length)
+            will be the product of tensor's first `rank - num_col_dims` dimensions.
+        )DOC")
+        .SetDefault(1)
+        .EqualGreaterThan(1);
+    AddAttr<int>(
+        "y_num_col_dims",
+        R"DOC(mul_op can take tensors with more than two dimensions as input `Y`,
+             in that case, tensors will be reshaped to a matrix. Just like input `X`.
+        )DOC")
+        .SetDefault(1)
+        .EqualGreaterThan(1);
     AddComment(R"DOC(
 Two Element Mul Operator.
 
@@ -70,10 +96,20 @@ class MulOpGrad : public framework::OperatorWithKernel {
     auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
     auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
-    PADDLE_ENFORCE(x_dims[0] == out_dims[0],
-                   "Out@GRAD M X N must equal to X dims 0, M ");
-    PADDLE_ENFORCE(y_dims[1] == out_dims[1],
-                   "Out@GRAD M X N must equal to Y dims 1, N ");
+
+    auto x_mat_dims =
+        framework::flatten_to_2d(x_dims, Attr<int>("x_num_col_dims"));
+    auto y_mat_dims =
+        framework::flatten_to_2d(y_dims, Attr<int>("y_num_col_dims"));
+
+    PADDLE_ENFORCE_EQ(
+        x_mat_dims[0], out_dims[0],
+        "The first dimension of Out@GRAD must equal to the first dimension of "
+        "the first operand.");
+    PADDLE_ENFORCE_EQ(
+        y_mat_dims[1], out_dims[1],
+        "The second dimension of Out@GRAD must equal to the second "
+        "dimension of the second operand.");
 
     if (x_grad) x_grad->Resize(x_dims);
     if (y_grad) y_grad->Resize(y_dims);

paddle/operators/mul_op.h

 /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
 
    Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
+   You may not use this file except in compliance with the License.
    You may obtain a copy of the License at
 
    http://www.apache.org/licenses/LICENSE-2.0
@@ -31,13 +31,25 @@ template <typename Place, typename T>
 class MulKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* x = context.Input<Tensor>("X");
-    auto* y = context.Input<Tensor>("Y");
-    auto* z = context.Output<Tensor>("Out");
+    const Tensor* x = context.Input<Tensor>("X");
+    const Tensor* y = context.Input<Tensor>("Y");
+    Tensor* z = context.Output<Tensor>("Out");
+    const Tensor x_matrix =
+        x->dims().size() > 2
+            ? framework::ReshapeToMatrix<T>(
+                  *x, context.template Attr<int>("x_num_col_dims"))
+            : *x;
+    const Tensor y_matrix =
+        y->dims().size() > 2
+            ? framework::ReshapeToMatrix<T>(
+                  *y, context.template Attr<int>("y_num_col_dims"))
+            : *y;
+
     z->mutable_data<T>(context.GetPlace());
     auto* device_context =
         const_cast<platform::DeviceContext*>(context.device_context_);
-    math::matmul<Place, T>(*x, false, *y, false, 1, z, 0, device_context);
+    math::matmul<Place, T>(x_matrix, false, y_matrix, false, 1, z, 0,
+                           device_context);
   }
 };
 
@@ -45,23 +57,39 @@ template <typename Place, typename T>
 class MulGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto* x = ctx.Input<Tensor>("X");
-    auto* y = ctx.Input<Tensor>("Y");
-    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    int x_num_col_dims = ctx.template Attr<int>("x_num_col_dims");
+    int y_num_col_dims = ctx.template Attr<int>("y_num_col_dims");
+    const Tensor* x = ctx.Input<Tensor>("X");
+    const Tensor* y = ctx.Input<Tensor>("Y");
+    const Tensor x_matrix =
+        x->dims().size() > 2 ? framework::ReshapeToMatrix<T>(*x, x_num_col_dims)
+                             : *x;
+    const Tensor y_matrix =
+        y->dims().size() > 2 ? framework::ReshapeToMatrix<T>(*y, y_num_col_dims)
+                             : *y;
+    const Tensor* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
 
-    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    Tensor* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    Tensor* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
     auto* device_context =
         const_cast<platform::DeviceContext*>(ctx.device_context_);
     if (dx) {
       dx->mutable_data<T>(ctx.GetPlace());
+      Tensor dx_matrix = dx->dims().size() > 2 ? framework::ReshapeToMatrix<T>(
+                                                     *dx, x_num_col_dims)
+                                               : *dx;
       // dx = dout * y'. dx: M x K, dout : M x N, y : K x N
-      math::matmul<Place, T>(*dout, false, *y, true, 1, dx, 0, device_context);
+      math::matmul<Place, T>(*dout, false, y_matrix, true, 1, &dx_matrix, 0,
+                             device_context);
     }
     if (dy) {
       dy->mutable_data<T>(ctx.GetPlace());
+      Tensor dy_matrix = dy->dims().size() > 2 ? framework::ReshapeToMatrix<T>(
+                                                     *dy, y_num_col_dims)
+                                               : *dy;
       // dy = x' * dout. dy K x N, dout : M x N, x : M x K
-      math::matmul<Place, T>(*x, true, *dout, false, 1, dy, 0, device_context);
+      math::matmul<Place, T>(x_matrix, true, *dout, false, 1, &dy_matrix, 0,
+                             device_context);
     }
   }
 };

paddle/operators/reshape_op.cc

+
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/reshape_op.h"
+
+namespace paddle {
+namespace operators {
+
+class ReshapeOp : public framework::OperatorWithKernel {
+ public:
+  ReshapeOp(const std::string &type, const framework::VariableNameMap &inputs,
+            const framework::VariableNameMap &outputs,
+            const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    // input check
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) shouldn't be null");
+    auto shape = ctx.Attr<std::vector<int>>("shape");
+    PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty.");
+    for (auto dim : shape) {
+      PADDLE_ENFORCE(dim > 0, "Each dimension of shape must be positive.");
+    }
+    // capacity check
+    int64_t capacity =
+        std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>());
+    auto *in = ctx.Input<framework::Tensor>("X");
+    int64_t in_size = framework::product(in->dims());
+    PADDLE_ENFORCE_EQ(capacity, in_size,
+                      "The size of Input(X) mismatches with Attr(shape).");
+    // resize output
+    std::vector<int64_t> shape_int64(shape.size(), 0);
+    std::transform(shape.begin(), shape.end(), shape_int64.begin(),
+                   [](int a) { return static_cast<int64_t>(a); });
+    auto out_dims = framework::make_ddim(shape_int64);
+    ctx.Output<framework::Tensor>("Out")->Resize(out_dims);
+  }
+};
+
+class ReshapeOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ReshapeOpMaker(framework::OpProto *proto,
+                 framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The input tensor of reshape operator.");
+    AddOutput("Out", "The output tensor of reshape operator.");
+    AddAttr<std::vector<int>>("shape", "Target shape of reshape operator.");
+    AddComment(R"DOC(Reshape operator
+
+Reshape Input(X) into the shape specified by Attr(shape).
+
+An example:
+Given a 2-D tensor X with 2 rows and 2 columns
+
+    [[1, 2], [3, 4]]
+
+with target shape = [1, 4], the reshape operator will transform 
+the tensor X into a 1-D tensor:
+
+    [1, 2, 3, 4]
+
+)DOC");
+  }
+};
+
+class ReshapeGradOp : public framework::OperatorWithKernel {
+ public:
+  ReshapeGradOp(const std::string &type,
+                const framework::VariableNameMap &inputs,
+                const framework::VariableNameMap &outputs,
+                const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) shouldn't be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) shouldn't be null.");
+    auto dims = ctx.Input<framework::Tensor>("X")->dims();
+    auto *d_in = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    d_in->Resize(dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+
+REGISTER_OP(reshape, ops::ReshapeOp, ops::ReshapeOpMaker, reshape_grad,
+            ops::ReshapeGradOp);
+REGISTER_OP_CPU_KERNEL(reshape,
+                       ops::ReshapeKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    reshape_grad, ops::ReshapeGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/reshape_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/reshape_op.h"
+
+REGISTER_OP_GPU_KERNEL(
+    reshape,
+    paddle::operators::ReshapeKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    reshape_grad,
+    paddle::operators::ReshapeGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/reshape_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class ReshapeKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const {
+    auto* out = ctx.Output<framework::Tensor>("Out");
+    auto* in = ctx.Input<framework::Tensor>("X");
+    out->mutable_data<T>(ctx.GetPlace());
+
+    auto shape = ctx.Attr<std::vector<int>>("shape");
+    std::vector<int64_t> shape_int64(shape.size(), 0);
+    std::transform(shape.begin(), shape.end(), shape_int64.begin(),
+                   [](int a) { return static_cast<int64_t>(a); });
+    auto out_dims = framework::make_ddim(shape_int64);
+    out->CopyFrom<T>(*in, ctx.GetPlace());
+    out->Resize(out_dims);
+  }
+};
+
+template <typename Place, typename T>
+class ReshapeGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const {
+    auto* d_out = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* d_x = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    d_x->mutable_data<T>(ctx.GetPlace());
+
+    auto in_dims = d_x->dims();
+    d_x->CopyFrom<T>(*d_out, ctx.GetPlace());
+    d_x->Resize(in_dims);
+  }
+};
+}  // namespace operators
+}  // namespace paddle

paddle/operators/rowwise_add_op.cc

@@ -25,14 +25,19 @@ class RowwiseAddOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto dim0 = ctx.Input<Tensor>("X")->dims();
-    auto dim1 = ctx.Input<Tensor>("b")->dims();
-
-    PADDLE_ENFORCE(dim0.size() == 2, "Input 0 must be matrix");
-    PADDLE_ENFORCE(dim1.size() == 1, "The second input must be vector");
-    PADDLE_ENFORCE(dim0[1] == dim1[0], "The width of two input must be same");
-    PADDLE_ENFORCE(ctx.OutputSize("Out") == 1, "The output size must be 1");
-    ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("X")->dims());
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto b_dims = ctx.Input<Tensor>("b")->dims();
+    PADDLE_ENFORCE_GT(
+        x_dims.size(), b_dims.size(),
+        "The rank of input `X` must be larger than the one of input `b`.");
+
+    int num_col_dims = x_dims.size() - b_dims.size();
+
+    PADDLE_ENFORCE_EQ(
+        framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims,
+        "The width of two operands must be same");
+    PADDLE_ENFORCE_EQ(ctx.OutputSize("Out"), 1, "The output size must be 1");
+    ctx.Output<Tensor>("Out")->Resize(x_dims);
   }
 };
 
@@ -61,13 +66,20 @@ class RowwiseAddGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("b"), "b should not be null");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
                             "Input(Out@GRAD) should not be null");
-    auto dims0 = ctx.Input<Tensor>("X")->dims();
-    auto dims1 = ctx.Input<Tensor>("b")->dims();
-    PADDLE_ENFORCE_EQ(1, dims1.size(), "b dims should be 1")
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto b_dims = ctx.Input<Tensor>("b")->dims();
+    PADDLE_ENFORCE_GT(
+        x_dims.size(), b_dims.size(),
+        "The rank of input `X` must be larger than the one of input `b`.");
+
+    int num_col_dims = x_dims.size() - b_dims.size();
+    PADDLE_ENFORCE_EQ(
+        framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims,
+        "The width of two operands must be same");
     auto *dx = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto *db = ctx.Output<Tensor>(framework::GradVarName("b"));
-    if (dx) dx->Resize(dims0);
-    if (db) db->Resize(dims1);
+    if (dx) dx->Resize(x_dims);
+    if (db) db->Resize(b_dims);
   }
 };
 

paddle/operators/rowwise_add_op.h

@@ -33,10 +33,12 @@ class RowwiseAddKernel : public framework::OpKernel {
   void Compute(const framework::ExecutionContext& context) const override {
     auto out = context.Output<Tensor>("Out");
     out->mutable_data<T>(context.GetPlace());
-
-    auto input = EigenMatrix<T>::From(*context.Input<Tensor>("X"));
-    auto bias = EigenVector<T>::From(*context.Input<Tensor>("b"));
-    auto output = EigenMatrix<T>::From(*out);
+    int num_col_dims = context.Input<Tensor>("X")->dims().size() -
+                       context.Input<Tensor>("b")->dims().size();
+    auto input =
+        EigenMatrix<T>::Reshape(*context.Input<Tensor>("X"), num_col_dims);
+    auto bias = EigenVector<T>::Flatten(*context.Input<Tensor>("b"));
+    auto output = EigenMatrix<T>::Reshape(*out, num_col_dims);
 
     const int bias_size = bias.dimension(0);
     const int rest_size = input.size() / bias_size;
@@ -54,12 +56,15 @@ class RowwiseAddGradKernel : public framework::OpKernel {
     auto* dout = context.Input<Tensor>(framework::GradVarName("Out"));
     auto* dx = context.Output<Tensor>(framework::GradVarName("X"));
     auto* db = context.Output<Tensor>(framework::GradVarName("b"));
+    int num_col_dims = context.Input<Tensor>("X")->dims().size() -
+                       context.Input<Tensor>("b")->dims().size();
 
-    auto out_grad = EigenMatrix<T>::From(*dout);
+    auto out_grad = EigenMatrix<T>::Reshape(*dout, num_col_dims);
     auto place = context.GetEigenDevice<Place>();
+
     if (dx) {
       dx->mutable_data<T>(context.GetPlace());
-      EigenMatrix<T>::From(*dx).device(place) = out_grad;
+      EigenMatrix<T>::Reshape(*dx, num_col_dims).device(place) = out_grad;
     }
 
     if (db) {

paddle/operators/squared_l2_distance_op.cc

@@ -41,8 +41,7 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
 
     int rank = framework::arity(x_dims);
     PADDLE_ENFORCE_GE(rank, 2, "Tensor rank should be at least equal to 2.");
-    PADDLE_ENFORCE_EQ(framework::product(x_dims) / x_dims[0],
-                      framework::product(y_dims) / y_dims[0],
+    PADDLE_ENFORCE_EQ(x->numel() / x_dims[0], y->numel() / y_dims[0],
                       "Product of dimensions expcet the first dimension of "
                       "input and target must be equal.");
     PADDLE_ENFORCE(y_dims[0] == 1 || y_dims[0] == x_dims[0],
@@ -50,8 +49,7 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
                    "or to 1.");
 
     ctx.Output<Tensor>("sub_result")
-        ->Resize({static_cast<int>(x_dims[0]),
-                  static_cast<int>(framework::product(x_dims) / x_dims[0])});
+        ->Resize({x_dims[0], x->numel() / x_dims[0]});
     ctx.Output<Tensor>("Out")->Resize({x_dims[0], 1});
   }
 };

paddle/operators/squared_l2_distance_op.h

@@ -39,7 +39,7 @@ class SquaredL2DistanceKernel : public framework::OpKernel {
     auto in0_dims = in0->dims();
     auto in1_dims = in1->dims();
 
-    int cols = framework::product(in0_dims) / in0_dims[0];
+    int cols = in0->numel() / in0_dims[0];
     // reduce dimensions except the first
     auto x =
         EigenMatrix<T>::From(*in0, framework::make_ddim({in0_dims[0], cols}));
@@ -82,7 +82,7 @@ class SquaredL2DistanceGradKernel : public framework::OpKernel {
     auto x_dims = x_g->dims();
     auto y_dims = y_g->dims();
 
-    int cols = framework::product(x_dims) / x_dims[0];
+    int cols = x_g->numel() / x_dims[0];
     // calculate gradient
     auto grad_mat = 2 *
                     (out_grad.broadcast(Eigen::array<int, 2>({{1, cols}}))) *

paddle/operators/sum_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/sum_op.h"
+#include <vector>
+
+namespace paddle {
+namespace operators {
+using framework::Tensor;
+
+class SumOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto ins = ctx.MultiInput<framework::Tensor>("X");
+    auto *out = ctx.Output<framework::Tensor>("Out");
+    int N = ins.size();
+
+    auto in_dim = ins[0]->dims();
+
+    PADDLE_ENFORCE_GT(N, 1, "Input tensors count should > 1.");
+    for (int i = 1; i < N; i++) {
+      auto dim = ins[i]->dims();
+      PADDLE_ENFORCE(in_dim == dim, "Input tensors must have same shape");
+    }
+    out->Resize(in_dim);
+  }
+};
+
+class SumOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SumOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "the input tensors of sum operator.").AsDuplicable();
+    AddOutput("Out", "the output tensor of sum operator.");
+    AddComment(R"DOC(
+            Sum the input tensors.
+        )DOC");
+  }
+};
+
+class SumGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto outputs = ctx.MultiOutput<Tensor>(framework::GradVarName("X"));
+    auto dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+    for (auto output : outputs) {
+      output->Resize(dims);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(sum, ops::SumOp, ops::SumOpMaker, sum_grad, ops::SumGradOp);
+REGISTER_OP_CPU_KERNEL(sum, ops::SumKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(sum_grad,
+                       ops::SumGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/sum_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/sum_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(sum, ops::SumKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(sum_grad,
+                       ops::SumGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/sum_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class SumKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto ins = context.MultiInput<Tensor>("X");
+    auto* out = context.Output<Tensor>("Out");
+    out->mutable_data<T>(context.GetPlace());
+
+    auto place = context.GetEigenDevice<Place>();
+    auto result = EigenVector<T>::Flatten(*out);
+
+    int N = ins.size();
+    auto in = EigenVector<T>::Flatten(*(ins[0]));
+    result.device(place) = in;
+    for (int i = 1; i < N; i++) {
+      auto in = EigenVector<T>::Flatten(*(ins[i]));
+      result.device(place) = result + in;
+    }
+  }
+};
+
+template <typename Place, typename T>
+class SumGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* input = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto outs = context.MultiOutput<Tensor>(framework::GradVarName("X"));
+    for (auto out : outs) {
+      out->mutable_data<T>(context.GetPlace());
+    }
+
+    auto place = context.GetEigenDevice<Place>();
+    auto in = EigenVector<T>::Flatten(*input);
+    for (auto out : outs) {
+      auto result = EigenVector<T>::Flatten(*out);
+      result.device(place) = in;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/uniform_random_op.cc

@@ -35,7 +35,7 @@ class CPUUniformRandomKernel : public framework::OpKernel {
     std::uniform_real_distribution<T> dist(
         static_cast<T>(context.Attr<float>("min")),
         static_cast<T>(context.Attr<float>("max")));
-    int64_t size = framework::product(tensor->dims());
+    int64_t size = tensor->numel();
     for (int64_t i = 0; i < size; ++i) {
       data[i] = dist(engine);
     }

paddle/operators/uniform_random_op.cu

@@ -53,8 +53,8 @@ class GPUUniformRandomKernel : public framework::OpKernel {
     T min = static_cast<T>(context.Attr<float>("min"));
     T max = static_cast<T>(context.Attr<float>("max"));
     thrust::counting_iterator<unsigned int> index_sequence_begin(0);
-    ssize_t N = framework::product(tensor->dims());
-    thrust::transform(index_sequence_begin, index_sequence_begin + N,
+    int64_t size = tensor->numel();
+    thrust::transform(index_sequence_begin, index_sequence_begin + size,
                       thrust::device_ptr<T>(data),
                       UniformGenerator<T>(min, max, seed));
   }

paddle/platform/enforce.h

@@ -25,10 +25,6 @@ limitations under the License. */
 #include "paddle/string/printf.h"
 #include "paddle/string/to_string.h"
 
-#ifdef __GNUC__
-#include <cxxabi.h>  // for __cxa_demangle
-#endif
-
 #ifndef PADDLE_ONLY_CPU
 
 #include "paddle/platform/dynload/cublas.h"
@@ -46,19 +42,6 @@ limitations under the License. */
 namespace paddle {
 namespace platform {
 
-namespace {
-#ifdef __GNUC__
-inline std::string demangle(std::string name) {
-  int status = -4;  // some arbitrary value to eliminate the compiler warning
-  std::unique_ptr<char, void (*)(void*)> res{
-      abi::__cxa_demangle(name.c_str(), NULL, NULL, &status), std::free};
-  return (status == 0) ? res.get() : name;
-}
-#else
-inline std::string demangle(std::string name) { return name; }
-#endif
-}
-
 struct EnforceNotMet : public std::exception {
   std::exception_ptr exp_;
   std::string err_str_;
@@ -79,7 +62,7 @@ struct EnforceNotMet : public std::exception {
       Dl_info info;
       for (int i = 0; i < size; ++i) {
         if (dladdr(call_stack[i], &info)) {
-          auto demangled = demangle(info.dli_sname);
+          auto demangled = info.dli_sname;
           auto addr_offset = static_cast<char*>(call_stack[i]) -
                              static_cast<char*>(info.dli_saddr);
           sout << string::Sprintf("%-3d %*0p %s + %zd\n", i,

paddle/pybind/pybind.cc

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <vector>
 
 #include "paddle/framework/backward.h"
+#include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/net_op.h"
 #include "paddle/operators/recurrent_op.h"
@@ -49,13 +50,18 @@ USE_OP(minus);
 USE_OP(cos_sim);
 USE_CPU_ONLY_OP(gather);
 USE_CPU_ONLY_OP(scatter);
+USE_CPU_ONLY_OP(concat);
 USE_OP(top_k);
 USE_OP(squared_l2_distance);
+USE_OP(sum);
+USE_OP(reshape);
 
 namespace paddle {
 namespace framework {
 
 using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+using LoD = framework::LoD;
 
 static size_t UniqueIntegerGenerator() {
   static std::atomic<size_t> generator;
@@ -115,6 +121,60 @@ PYBIND11_PLUGIN(core) {
         return self.data<float>()[offset];
       });
 
+  py::class_<LoDTensor>(m, "LoDTensor", R"DOC(LoD(Leval of Ddetails) Tensor.
+
+The tensor and LoD info should be created before creating the LoDTensor, then
+call the set_tensor and set_lod functions to set them.
+
+)DOC")
+      .def("__init__",
+           [](LoDTensor &instance,
+              const std::vector<std::vector<size_t>> &lod,
+              Tensor *t) {
+#ifdef PADDLE_ONLY_CPU
+             new (&instance) LoDTensor(lod, t);
+#else
+             paddle::framework::LoD new_lod;
+             new_lod.reserve(lod.size());
+             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
+             new (&instance) LoDTensor(new_lod, t);
+#endif
+           })
+      .def("set_tensor",
+           [](LoDTensor &self, Tensor *tensor) { self.set_tensor(tensor); })
+      .def("set_lod",
+           [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
+#ifdef PADDLE_ONLY_CPU
+             self.set_lod(lod);
+#else
+             paddle::framework::LoD new_lod;
+             new_lod.reserve(lod.size());
+             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
+             self.set_lod(new_lod);
+#endif
+           })
+      .def("tensor",
+           [](LoDTensor &self) -> Tensor & { return self.tensor(); },
+           py::return_value_policy::reference)
+      .def("lod", [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
+#ifdef PADDLE_ONLY_CPU
+        return self.lod();
+#else
+           auto lod = self.lod();
+           std::vector<std::vector<size_t>> new_lod;
+           new_lod.reserve(lod.size());
+           std::transform(lod.begin(), lod.end(), std::back_inserter(new_lod),
+               [](paddle::framework::Vector<size_t> item) ->
+                   std::vector<size_t> {
+                 std::vector<size_t> v;
+                 v.reserve(item.size());
+                 std::copy(item.begin(), item.end(), std::back_inserter(v));
+                 return v;
+               });
+           return new_lod;
+#endif
+      });
+
   py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
 
 All parameter, weight, gradient are variables in Paddle.
@@ -126,6 +186,11 @@ All parameter, weight, gradient are variables in Paddle.
       .def("get_tensor",
            [](Variable &self) -> Tensor * { return self.GetMutable<Tensor>(); },
            py::return_value_policy::reference)
+      .def("get_lod_tensor",
+           [](Variable &self) -> LoDTensor * {
+             return self.GetMutable<LoDTensor>();
+           },
+           py::return_value_policy::reference)
       .def("get_net",
            [](Variable &self) -> operators::NetOp * {
              return self.GetMutable<operators::NetOp>();
@@ -216,7 +281,10 @@ All parameter, weight, gradient are variables in Paddle.
                -> std::map<std::string, std::vector<std::string>> {
                  return op.Outputs();
                })
+      .def("output_vars",
+           [](const OperatorBase &op) { return op.OutputVars(true); })
       .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
+      .def("input_vars", [](const OperatorBase &op) { return op.InputVars(); })
       .def("__str__", &OperatorBase::DebugString)
       .def("no_intermediate_outputs",
            [](const OperatorBase &op) { return op.OutputVars(false); })

paddle/scripts/docker/build.sh

@@ -30,6 +30,8 @@ Configuring cmake in /paddle/build ...
       -DCMAKE_BUILD_TYPE=Release
       -DWITH_DOC=OFF
       -DWITH_GPU=${WITH_GPU:-OFF}
+      -DWITH_MKLDNN=${WITH_MKLDNN:-ON}
+      -DWITH_MKLML=${WITH_MKLML:-ON}
       -DWITH_AVX=${WITH_AVX:-OFF}
       -DWITH_GOLANG=${WITH_GOLANG:-ON}
       -DWITH_SWIG_PY=ON
@@ -50,6 +52,8 @@ cmake .. \
       -DCMAKE_BUILD_TYPE=Release \
       -DWITH_DOC=OFF \
       -DWITH_GPU=${WITH_GPU:-OFF} \
+      -DWITH_MKLDNN=${WITH_MKLDNN:-ON} \
+      -DWITH_MKLML=${WITH_MKLML:-ON} \
       -DWITH_AVX=${WITH_AVX:-OFF} \
       -DWITH_GOLANG=${WITH_GOLANG:-ON} \
       -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} \

paddle/utils/Util.cpp

@@ -320,6 +320,9 @@ void loadFileList(const std::string& fileListFileName,
 }
 
 double getMemoryUsage() {
+#if defined(__ANDROID__)
+  return 0.0;
+#else
   FILE* fp = fopen("/proc/meminfo", "r");
   CHECK(fp) << "failed to fopen /proc/meminfo";
   size_t bufsize = 256 * sizeof(char);
@@ -357,6 +360,7 @@ double getMemoryUsage() {
   delete[] buf;
   double usedMem = 1.0 - 1.0 * (freeMem + bufMem + cacheMem) / totalMem;
   return usedMem;
+#endif
 }
 
 SyncThreadPool* getGlobalSyncThreadPool() {

paddle/utils/Util.h

@@ -33,6 +33,13 @@ limitations under the License. */
 #include "Flags.h"
 #include "hl_gpu.h"
 
+#if defined(__ANDROID__) && (__ANDROID_API__ < 21)
+inline int rand_r(unsigned int* seedp) {
+  (void)seedp;
+  return rand();
+}
+#endif
+
 /**
  * Loop over the elements in a container
  * TODO(yuyang18): It's this foreach useful? Why not use C++ 11 foreach,

proto/ModelConfig.proto

@@ -271,6 +271,7 @@ message ImageConfig {
   // The size of input feature map.
   required uint32 img_size = 8;
   optional uint32 img_size_y = 9;
+  optional uint32 img_size_z = 10 [ default = 1 ];
 }
 
 message PriorBoxConfig {
@@ -519,6 +520,7 @@ message LayerConfig {
   // for HuberRegressionLoss
   optional double delta = 57 [ default = 1.0 ];
 
+  // for 3D data
   optional uint64 depth = 58 [ default = 1 ];
 
   // for switch order layer

python/paddle/trainer/config_parser.py

@@ -1332,6 +1332,12 @@ def parse_image(image, input_layer_name, image_conf):
         get_img_size(input_layer_name, image_conf.channels)
 
 
+def parse_image3d(image, input_layer_name, image_conf):
+    image_conf.channels = image.channels
+    image_conf.img_size, image_conf.img_size_y, image_conf.img_size_z = \
+        get_img3d_size(input_layer_name, image_conf.channels)
+
+
 def parse_norm(norm, input_layer_name, norm_conf):
     norm_conf.norm_type = norm.norm_type
     config_assert(
@@ -2028,6 +2034,7 @@ class ParameterReluLayer(LayerBase):
         config_assert(input_layer.size % partial_sum == 0,
                       "a wrong setting for partial_sum")
         self.set_layer_size(input_layer.size)
+        self.config.partial_sum = partial_sum
         self.create_input_parameter(0, input_layer.size / partial_sum)
 
 
@@ -2365,9 +2372,11 @@ class BatchNormLayer(LayerBase):
                  name,
                  inputs,
                  bias=True,
+                 img3D=False,
                  use_global_stats=True,
                  moving_average_fraction=0.9,
                  batch_norm_type=None,
+                 mean_var_names=None,
                  **xargs):
         if inputs is None:
             inputs = []
@@ -2409,24 +2418,69 @@ class BatchNormLayer(LayerBase):
 
         input_layer = self.get_input_layer(0)
         image_conf = self.config.inputs[0].image_conf
-        parse_image(self.inputs[0].image, input_layer.name, image_conf)
-
-        # Only pass the width and height of input to batch_norm layer
-        # when either of it is non-zero.
-        if input_layer.width != 0 or input_layer.height != 0:
-            self.set_cnn_layer(name, image_conf.img_size_y, image_conf.img_size,
-                               image_conf.channels, False)
+        if img3D:
+            parse_image3d(self.inputs[0].image, input_layer.name, image_conf)
+            # Only pass the width and height of input to batch_norm layer
+            # when either of it is non-zero.
+            if input_layer.width != 0 or input_layer.height != 0:
+                self.set_cnn_layer(
+                    input_layer_name=name,
+                    depth=image_conf.img_size_z,
+                    height=image_conf.img_size_y,
+                    width=image_conf.img_size,
+                    channels=image_conf.channels,
+                    is_print=True)
+            else:
+                self.set_layer_size(input_layer.size)
         else:
-            self.set_layer_size(input_layer.size)
+            parse_image(self.inputs[0].image, input_layer.name, image_conf)
+            # Only pass the width and height of input to batch_norm layer
+            # when either of it is non-zero.
+            if input_layer.width != 0 or input_layer.height != 0:
+                self.set_cnn_layer(
+                    input_layer_name=name,
+                    height=image_conf.img_size_y,
+                    width=image_conf.img_size,
+                    channels=image_conf.channels,
+                    is_print=True)
+            else:
+                self.set_layer_size(input_layer.size)
 
         psize = self.calc_parameter_size(image_conf)
         dims = [1, psize]
+        if mean_var_names is not None:
+            assert len(mean_var_names) == 2
+            self.inputs[1].parameter_name = mean_var_names[0]
+            self.inputs[2].parameter_name = mean_var_names[1]
+
         self.create_input_parameter(0, psize)
         self.create_input_parameter(1, psize, dims)
         self.create_input_parameter(2, psize, dims)
 
         self.create_bias_parameter(bias, psize)
 
+    def set_cnn_layer(self,
+                      input_layer_name,
+                      depth=None,
+                      height=None,
+                      width=None,
+                      channels=None,
+                      is_print=True):
+        depthIsNone = False
+        if depth is None:
+            depth = 1
+            depthIsNone = True
+        size = depth * height * width * channels
+        self.set_layer_size(size)
+        self.set_layer_height_width(height, width)
+        self.set_layer_depth(depth)
+        if is_print and depthIsNone:
+            print("output for %s: c = %d, h = %d, w = %d, size = %d" %
+                  (input_layer_name, channels, height, width, size))
+        elif is_print:
+            print("output for %s: c = %d, d = %d, h = %d, w = %d, size = %d" %
+                  (input_layer_name, channels, depth, height, width, size))
+
     def calc_parameter_size(self, image_conf):
         return image_conf.channels
 
@@ -2688,9 +2742,20 @@ class AddToLayer(LayerBase):
         super(AddToLayer, self).__init__(
             name, 'addto', 0, inputs=inputs, **xargs)
         config_assert(len(inputs) > 0, 'inputs cannot be empty for AddToLayer')
-        for input_index in xrange(len(self.inputs)):
-            input_layer = self.get_input_layer(input_index)
-            self.set_layer_size(input_layer.size)
+
+        if len(self.inputs) > 1:
+            for input_index in xrange(len(self.inputs)):
+                assert self.get_input_layer(0).height == self.get_input_layer(
+                    input_index).height
+                assert self.get_input_layer(0).width == self.get_input_layer(
+                    input_index).width
+                assert self.get_input_layer(0).depth == self.get_input_layer(
+                    input_index).depth
+
+        self.set_layer_size(self.get_input_layer(0).size)
+        self.set_layer_height_width(self.get_input_layer(0).height, \
+                                        self.get_input_layer(0).width)
+        self.set_layer_depth(self.get_input_layer(0).depth)
         self.create_bias_parameter(bias, self.config.size)
 
 
@@ -3370,11 +3435,20 @@ class ConcatenateLayer(LayerBase):
             name, 'concat', 0, inputs=inputs, **xargs)
         size = 0
         for input_index in xrange(len(self.inputs)):
+            assert self.get_input_layer(0).height == self.get_input_layer(
+                input_index).height
+            assert self.get_input_layer(0).width == self.get_input_layer(
+                input_index).width
+            assert self.get_input_layer(0).depth == self.get_input_layer(
+                input_index).depth
             input_layer = self.get_input_layer(input_index)
             input = self.inputs[input_index]
             if self.config.size == 0:
                 size += input_layer.size
 
+        self.set_layer_height_width(self.get_input_layer(0).height, \
+                                    self.get_input_layer(0).width)
+        self.set_layer_depth(self.get_input_layer(0).depth)
         self.set_layer_size(size)
 
 
@@ -3675,8 +3749,8 @@ class SwitchOrderLayer(LayerBase):
     def __init__(self, name, inputs, reshape, **xargs):
         super(SwitchOrderLayer, self).__init__(
             name, 'switch_order', 0, inputs=inputs, **xargs)
-        self.config.reshape_conf.heightAxis.extend(reshape['height'])
-        self.config.reshape_conf.widthAxis.extend(reshape['width'])
+        self.config.reshape_conf.height_axis.extend(reshape['height'])
+        self.config.reshape_conf.width_axis.extend(reshape['width'])
 
 
 # Deprecated, use a new layer specific class instead

python/paddle/trainer_config_helpers/layers.py

@@ -354,6 +354,10 @@ class LayerOutput(object):
     def height(self):
         return cp.g_layer_map[self.full_name].height
 
+    @property
+    def depth(self):
+        return cp.g_layer_map[self.full_name].depth
+
     def set_input(self, input):
         """
         Set the input for a memory layer. Can only be used for memory layer
@@ -943,7 +947,7 @@ def data_layer(name, size, depth=None, height=None, width=None,
     if height is not None and width is not None:
         num_filters = size / (width * height * depth)
         assert num_filters * width * height * depth == size, \
-                "size=%s width=%s height=%s depth=%s"  % (size, width, height, depth)
+                "size=%s width=%s height=%s depth=%s" % (size, width, height, depth)
 
     return LayerOutput(name, LayerType.DATA, size=size, num_filters=num_filters)
 
@@ -1219,7 +1223,8 @@ def detection_output_layer(input_loc,
                            name=None):
     """
     Apply the NMS to the output of network and compute the predict bounding
-    box location.
+    box location. The output of this layer could be None if there is no valid
+    bounding box.
 
     :param name: The Layer Name.
     :type name: basestring
@@ -2953,13 +2958,15 @@ def img_cmrnorm_layer(input,
 def batch_norm_layer(input,
                      act=None,
                      name=None,
+                     img3D=False,
                      num_channels=None,
                      bias_attr=None,
                      param_attr=None,
                      layer_attr=None,
                      batch_norm_type=None,
                      moving_average_fraction=0.9,
-                     use_global_stats=None):
+                     use_global_stats=None,
+                     mean_var_names=None):
     """
     Batch Normalization Layer. The notation of this layer as follow.
 
@@ -3026,6 +3033,8 @@ def batch_norm_layer(input,
                                    :math:`runningMean = newMean*(1-factor)
                                    + runningMean*factor`
     :type moving_average_fraction: float.
+    :param mean_var_names: [mean name, variance name]
+    :type mean_var_names: string list
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -3039,6 +3048,7 @@ def batch_norm_layer(input,
            (batch_norm_type == "cudnn_batch_norm")
     l = Layer(
         name=name,
+        img3D=img3D,
         inputs=Input(
             input.name, image=Image(channels=num_channels), **param_attr.attr),
         active_type=act.name,
@@ -3047,6 +3057,7 @@ def batch_norm_layer(input,
         bias=ParamAttr.to_bias(bias_attr),
         moving_average_fraction=moving_average_fraction,
         use_global_stats=use_global_stats,
+        mean_var_names=mean_var_names,
         **ExtraLayerAttribute.to_kwargs(layer_attr))
 
     return LayerOutput(
@@ -6410,7 +6421,7 @@ def gated_unit_layer(input,
 @wrap_name_default('switch_order')
 def switch_order_layer(input,
                        name=None,
-                       reshape=None,
+                       reshape_axis=None,
                        act=None,
                        layer_attr=None):
     """
@@ -6421,8 +6432,9 @@ def switch_order_layer(input,
     The example usage is:
 
     .. code-block:: python
+       reshape_axis = 3
+       switch = switch_order(input=layer, name='switch', reshape_axis=reshape_axis)
        reshape = {'height':[ 0, 1, 2], 'width':[3]}
-       switch = switch_order(input=layer, name='switch', reshape=reshape)
 
     :param input: The input layer.
     :type input: LayerOutput
@@ -6434,6 +6446,11 @@ def switch_order_layer(input,
     :rtype: LayerOutput
     """
     assert isinstance(input, LayerOutput)
+    assert reshape_axis != None and (reshape_axis > 0 and reshape_axis < 4)
+    height = [ele for ele in xrange(reshape_axis)]
+    width = [ele for ele in range(reshape_axis, 4)]
+    reshape = {'height': height, 'width': width}
+
     l = Layer(
         name=name,
         inputs=input.name,
@@ -6444,6 +6461,7 @@ def switch_order_layer(input,
     return LayerOutput(
         name=name,
         layer_type=LayerType.SWITCH_ORDER_LAYER,
+        activation=act,
         parents=input,
         size=l.config.size)
 

python/paddle/trainer_config_helpers/tests/configs/file_list.sh

@@ -10,6 +10,6 @@ test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_la
 test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer
 test_kmax_seq_socre_layer test_sub_nested_seq_select_layer test_scale_shift_layer
 test_seq_slice_layer test_cross_entropy_over_beam test_pooling3D_layer
-test_conv3d_layer test_deconv3d_layer)
+test_conv3d_layer test_deconv3d_layer test_BatchNorm3D)
 
 export whole_configs=(test_split_datasource)

python/paddle/trainer_config_helpers/tests/configs/protostr/img_layers.protostr

@@ -62,6 +62,7 @@ layers {
   moving_average_fraction: 0.9
   height: 227
   width: 227
+  depth: 1
 }
 layers {
   name: "__crmnorm_0__"

python/paddle/trainer_config_helpers/tests/configs/protostr/img_trans_layers.protostr

@@ -62,6 +62,7 @@ layers {
   moving_average_fraction: 0.9
   height: 256
   width: 256
+  depth: 1
 }
 layers {
   name: "__crmnorm_0__"

python/paddle/trainer_config_helpers/tests/configs/protostr/test_BatchNorm3D.protostr

+type: "nn"
+layers {
+  name: "data3D"
+  type: "data"
+  size: 360
+  active_type: ""
+  height: 6
+  width: 20
+  depth: 3
+}
+layers {
+  name: "__batch_norm_0__"
+  type: "batch_norm"
+  size: 360
+  active_type: "relu"
+  inputs {
+    input_layer_name: "data3D"
+    input_parameter_name: "___batch_norm_0__.w0"
+    image_conf {
+      channels: 1
+      img_size: 20
+      img_size_y: 6
+      img_size_z: 3
+    }
+  }
+  inputs {
+    input_layer_name: "data3D"
+    input_parameter_name: "___batch_norm_0__.w1"
+  }
+  inputs {
+    input_layer_name: "data3D"
+    input_parameter_name: "___batch_norm_0__.w2"
+  }
+  bias_parameter_name: "___batch_norm_0__.wbias"
+  moving_average_fraction: 0.9
+  height: 6
+  width: 20
+  depth: 3
+}
+parameters {
+  name: "___batch_norm_0__.w0"
+  size: 1
+  initial_mean: 1.0
+  initial_std: 0.0
+  initial_strategy: 0
+  initial_smart: false
+}
+parameters {
+  name: "___batch_norm_0__.w1"
+  size: 1
+  initial_mean: 0.0
+  initial_std: 0.0
+  dims: 1
+  dims: 1
+  initial_strategy: 0
+  initial_smart: false
+  is_static: true
+  is_shared: true
+}
+parameters {
+  name: "___batch_norm_0__.w2"
+  size: 1
+  initial_mean: 0.0
+  initial_std: 0.0
+  dims: 1
+  dims: 1
+  initial_strategy: 0
+  initial_smart: false
+  is_static: true
+  is_shared: true
+}
+parameters {
+  name: "___batch_norm_0__.wbias"
+  size: 1
+  initial_mean: 0.0
+  initial_std: 0.0
+  dims: 1
+  dims: 1
+  initial_strategy: 0
+  initial_smart: false
+}
+input_layer_names: "data3D"
+output_layer_names: "__batch_norm_0__"
+sub_models {
+  name: "root"
+  layer_names: "data3D"
+  layer_names: "__batch_norm_0__"
+  input_layer_names: "data3D"
+  output_layer_names: "__batch_norm_0__"
+  is_recurrent_layer_group: false
+}
+

python/paddle/trainer_config_helpers/tests/configs/protostr/test_bi_grumemory.protostr

@@ -74,6 +74,9 @@ layers {
   inputs {
     input_layer_name: "__bidirectional_gru_0___bw"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 parameters {
   name: "___bidirectional_gru_0___fw_transform.w0"

python/paddle/trainer_config_helpers/tests/configs/protostr/test_prelu_layer.protostr

@@ -14,6 +14,29 @@ layers {
     input_layer_name: "input"
     input_parameter_name: "___prelu_layer_0__.w0"
   }
+  partial_sum: 1
+}
+layers {
+  name: "__prelu_layer_1__"
+  type: "prelu"
+  size: 300
+  active_type: ""
+  inputs {
+    input_layer_name: "input"
+    input_parameter_name: "___prelu_layer_1__.w0"
+  }
+  partial_sum: 1
+}
+layers {
+  name: "__prelu_layer_2__"
+  type: "prelu"
+  size: 300
+  active_type: ""
+  inputs {
+    input_layer_name: "input"
+    input_parameter_name: "___prelu_layer_2__.w0"
+  }
+  partial_sum: 5
 }
 parameters {
   name: "___prelu_layer_0__.w0"
@@ -23,14 +46,32 @@ parameters {
   initial_strategy: 0
   initial_smart: true
 }
+parameters {
+  name: "___prelu_layer_1__.w0"
+  size: 300
+  initial_mean: 0.0
+  initial_std: 0.057735026919
+  initial_strategy: 0
+  initial_smart: true
+}
+parameters {
+  name: "___prelu_layer_2__.w0"
+  size: 60
+  initial_mean: 0.0
+  initial_std: 0.129099444874
+  initial_strategy: 0
+  initial_smart: true
+}
 input_layer_names: "input"
-output_layer_names: "__prelu_layer_0__"
+output_layer_names: "__prelu_layer_2__"
 sub_models {
   name: "root"
   layer_names: "input"
   layer_names: "__prelu_layer_0__"
+  layer_names: "__prelu_layer_1__"
+  layer_names: "__prelu_layer_2__"
   input_layer_names: "input"
-  output_layer_names: "__prelu_layer_0__"
+  output_layer_names: "__prelu_layer_2__"
   is_recurrent_layer_group: false
 }
 

python/paddle/trainer_config_helpers/tests/configs/protostr/test_recursive_topology.protostr

@@ -16,6 +16,9 @@ layers {
   inputs {
     input_layer_name: "data"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_1__"
@@ -28,6 +31,9 @@ layers {
   inputs {
     input_layer_name: "__addto_0__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_2__"
@@ -40,6 +46,9 @@ layers {
   inputs {
     input_layer_name: "__addto_1__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_3__"
@@ -52,6 +61,9 @@ layers {
   inputs {
     input_layer_name: "__addto_2__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_4__"
@@ -64,6 +76,9 @@ layers {
   inputs {
     input_layer_name: "__addto_3__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_5__"
@@ -76,6 +91,9 @@ layers {
   inputs {
     input_layer_name: "__addto_4__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_6__"
@@ -88,6 +106,9 @@ layers {
   inputs {
     input_layer_name: "__addto_5__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_7__"
@@ -100,6 +121,9 @@ layers {
   inputs {
     input_layer_name: "__addto_6__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_8__"
@@ -112,6 +136,9 @@ layers {
   inputs {
     input_layer_name: "__addto_7__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_9__"
@@ -124,6 +151,9 @@ layers {
   inputs {
     input_layer_name: "__addto_8__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_10__"
@@ -136,6 +166,9 @@ layers {
   inputs {
     input_layer_name: "__addto_9__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_11__"
@@ -148,6 +181,9 @@ layers {
   inputs {
     input_layer_name: "__addto_10__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_12__"
@@ -160,6 +196,9 @@ layers {
   inputs {
     input_layer_name: "__addto_11__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_13__"
@@ -172,6 +211,9 @@ layers {
   inputs {
     input_layer_name: "__addto_12__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_14__"
@@ -184,6 +226,9 @@ layers {
   inputs {
     input_layer_name: "__addto_13__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_15__"
@@ -196,6 +241,9 @@ layers {
   inputs {
     input_layer_name: "__addto_14__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_16__"
@@ -208,6 +256,9 @@ layers {
   inputs {
     input_layer_name: "__addto_15__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_17__"
@@ -220,6 +271,9 @@ layers {
   inputs {
     input_layer_name: "__addto_16__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_18__"
@@ -232,6 +286,9 @@ layers {
   inputs {
     input_layer_name: "__addto_17__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_19__"
@@ -244,6 +301,9 @@ layers {
   inputs {
     input_layer_name: "__addto_18__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_20__"
@@ -256,6 +316,9 @@ layers {
   inputs {
     input_layer_name: "__addto_19__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_21__"
@@ -268,6 +331,9 @@ layers {
   inputs {
     input_layer_name: "__addto_20__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_22__"
@@ -280,6 +346,9 @@ layers {
   inputs {
     input_layer_name: "__addto_21__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_23__"
@@ -292,6 +361,9 @@ layers {
   inputs {
     input_layer_name: "__addto_22__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_24__"
@@ -304,6 +376,9 @@ layers {
   inputs {
     input_layer_name: "__addto_23__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_25__"
@@ -316,6 +391,9 @@ layers {
   inputs {
     input_layer_name: "__addto_24__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_26__"
@@ -328,6 +406,9 @@ layers {
   inputs {
     input_layer_name: "__addto_25__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_27__"
@@ -340,6 +421,9 @@ layers {
   inputs {
     input_layer_name: "__addto_26__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_28__"
@@ -352,6 +436,9 @@ layers {
   inputs {
     input_layer_name: "__addto_27__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_29__"
@@ -364,6 +451,9 @@ layers {
   inputs {
     input_layer_name: "__addto_28__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_30__"
@@ -376,6 +466,9 @@ layers {
   inputs {
     input_layer_name: "__addto_29__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__addto_31__"
@@ -388,6 +481,9 @@ layers {
   inputs {
     input_layer_name: "__addto_30__"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__fc_layer_0__"

python/paddle/trainer_config_helpers/tests/configs/protostr/util_layers.protostr

@@ -22,6 +22,9 @@ layers {
   inputs {
     input_layer_name: "b"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__concat_0__"
@@ -34,6 +37,9 @@ layers {
   inputs {
     input_layer_name: "b"
   }
+  height: 0
+  width: 0
+  depth: 1
 }
 layers {
   name: "__concat_1__"

python/paddle/trainer_config_helpers/tests/configs/test_BatchNorm3D.py

+from paddle.trainer_config_helpers import *
+
+settings(batch_size=1000, learning_rate=1e-4)
+
+#data = data_layer(name='data', size=180, width=30, height=6)
+#batchNorm = batch_norm_layer(data, num_channels=1)
+#outputs(batchNorm)
+
+data3D = data_layer(name='data3D', size=120 * 3, width=20, height=6, depth=3)
+batchNorm3D = batch_norm_layer(data3D, num_channels=1, img3D=True)
+outputs(batchNorm3D)

python/paddle/trainer_config_helpers/tests/configs/test_prelu_layer.py

@@ -2,5 +2,7 @@ from paddle.trainer_config_helpers import *
 
 data = data_layer(name='input', size=300)
 prelu = prelu_layer(input=data)
+prelu = prelu_layer(input=data, partial_sum=1)
+prelu = prelu_layer(input=data, partial_sum=5)
 
 outputs(prelu)

python/paddle/v2/event.py

@@ -53,10 +53,13 @@ class BeginPass(object):
 class EndPass(WithMetric):
     """
     Event On One Pass Training Complete.
+    To get the output of a specific layer, add "event.gm.getLayerOutputs('predict_layer')"
+    in your event_handler call back
     """
 
-    def __init__(self, pass_id, evaluator):
+    def __init__(self, pass_id, evaluator, gm):
         self.pass_id = pass_id
+        self.gm = gm
         WithMetric.__init__(self, evaluator)
 
 
@@ -73,10 +76,13 @@ class BeginIteration(object):
 class EndIteration(WithMetric):
     """
     Event On One Batch Training Complete.
+    To get the output of a specific layer, add "event.gm.getLayerOutputs('predict_layer')"
+    in your event_handler call back
     """
 
-    def __init__(self, pass_id, batch_id, cost, evaluator):
+    def __init__(self, pass_id, batch_id, cost, evaluator, gm):
         self.pass_id = pass_id
         self.batch_id = batch_id
         self.cost = cost
+        self.gm = gm
         WithMetric.__init__(self, evaluator)

python/paddle/v2/framework/op.py

@@ -43,7 +43,6 @@ class OpDescCreationMethod(object):
         if len(args) != 0:
             raise ValueError("Only keyword arguments are supported.")
         op_desc = framework_pb2.OpDesc()
-
         for input_parameter in self.__op_proto__.inputs:
             input_arguments = kwargs.get(input_parameter.name, [])
             if is_str(input_arguments):
@@ -142,8 +141,8 @@ def create_op_creation_method(op_proto):
     return OpInfo(
         method=__impl__,
         name=op_proto.type,
-        inputs=[var.name for var in op_proto.inputs],
-        outputs=[var.name for var in op_proto.outputs],
+        inputs=[(var.name, var.duplicable) for var in op_proto.inputs],
+        outputs=[(var.name, var.duplicable) for var in op_proto.outputs],
         attrs=[attr.name for attr in op_proto.attrs])
 
 
@@ -180,9 +179,15 @@ class OperatorFactory(object):
         return self.op_methods.get(t)
 
     def get_op_input_names(self, type):
+        return map(lambda x: x[0], self.get_op_info(type).inputs)
+
+    def get_op_inputs(self, type):
         return self.get_op_info(type).inputs
 
     def get_op_output_names(self, type):
+        return map(lambda x: x[0], self.get_op_info(type).outputs)
+
+    def get_op_outputs(self, type):
         return self.get_op_info(type).outputs
 
     def get_op_attr_names(self, type):

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

@@ -19,8 +19,6 @@ py_test(test_scatter_op SRCS test_scatter_op.py)
 py_test(test_fill_zeros_like_op SRCS test_fill_zeros_like_op.py)
 py_test(test_top_k_op SRCS test_top_k_op.py)
 
-py_test(gradient_checker SRCS gradient_checker.py)
-
 py_test(test_rowwise_add_op SRCS test_rowwise_add_op.py)
 
 py_test(test_default_scope_funcs SRCS test_default_scope_funcs.py)
@@ -33,5 +31,8 @@ py_test(test_sgd_op SRCS test_sgd_op.py)
 py_test(test_gradient_checker SRCS test_gradient_checker.py)
 py_test(test_lookup_table SRCS test_lookup_table.py)
 py_test(test_scale_and_identity_op SRCS test_scale_and_identity_op.py)
+py_test(test_sum_op SRCS test_sum_op.py)
 py_test(mnist SRCS mnist.py)
+py_test(test_concat_op SRCS test_concat_op.py)
 py_test(test_squared_l2_distance_op SRCS test_squared_l2_distance_op.py)
+py_test(test_reshape_op SRCS test_reshape_op.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
-
-__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
-    for out_name in Operator.get_op_output_names(op_type):
-        kwargs[out_name] = out_name
-
-    return Operator(op_type, **kwargs)
-
-
-def grad_var_name(var_name):
-    return var_name + "@GRAD"
-
-
-def empty_var_name():
-    return "@EMPTY@"
-
-
-def get_numeric_gradient(op,
-                         input_values,
-                         output_name,
-                         input_to_check,
-                         delta=0.005,
-                         local_scope=None,
-                         in_place=False):
-    """
-    Get Numeric Gradient for an operator's input.
-
-    :param op: C++ operator instance, could be an network
-    :param input_values: The input variables. Should be an dictionary, key is
-    variable name. Value is numpy array.
-    :param output_name: The final output variable name.
-    :param input_to_check: The input variable need to get gradient.
-    :param delta: The perturbation value for numeric gradient method. The
-    smaller delta is, the more accurate result will get. But if that delta is
-     too small, it could occur numerical stability problem.
-    :param local_scope: The local scope used for get_numeric_gradient.
-    :return: The gradient array in numpy format.
-    """
-    if local_scope is None:
-        local_scope = core.Scope()
-
-    # Create all input variable in local_scope
-    for var_name in input_values:
-        var = local_scope.new_var(var_name)
-        tensor = var.get_tensor()
-        tensor.set_dims(input_values[var_name].shape)
-        tensor.alloc_float(core.CPUPlace())
-        tensor.set(input_values[var_name], core.CPUPlace())
-
-    # Create all output variable in local_scope
-    opts = op.outputs()
-    for key in opts:
-        for output in opts[key]:
-            if local_scope.find_var(output) is None:
-                local_scope.new_var(output).get_tensor()
-    op.infer_shape(local_scope)
-
-    # allocate output memory
-    for key in opts:
-        for output in opts[key]:
-            local_scope.find_var(output).get_tensor().alloc_float(core.CPUPlace(
-            ))
-
-    cpu_ctx = core.DeviceContext.create(core.CPUPlace())
-
-    def get_output():
-        op.run(local_scope, cpu_ctx)
-        return numpy.array(local_scope.find_var(output_name).get_tensor()).sum()
-
-    def product(dim):
-        return reduce(lambda a, b: a * b, dim, 1)
-
-    def restore_inputs():
-        for var_name in input_values:
-            tensor_ = local_scope.find_var(var_name).get_tensor()
-            tensor_.set(numpy.copy(input_values[var_name]), core.CPUPlace())
-
-    # get the input tensor that we want to get it's numeric gradient.
-    tensor_to_check = local_scope.find_var(input_to_check).get_tensor()
-    tensor_size = product(tensor_to_check.get_dims())
-    # prepare a numpy array to store the gradient.
-    gradient_flat = numpy.zeros(shape=(tensor_size, ), dtype='float32')
-
-    # we only compute gradient of one element each time.
-    # we use a for loop to compute the gradient of every element.
-    for i in xrange(tensor_size):
-        if in_place:
-            restore_inputs()
-        # get one input element throw it's index i.
-        origin = tensor_to_check.get_float_element(i)
-
-        # add delta to it, run op and then get the sum of the result tensor.
-        x_pos = origin + delta
-        tensor_to_check.set_float_element(i, x_pos)
-        y_pos = get_output()
-
-        # plus delta to this element, run op and get the sum of the result tensor.
-        if in_place:
-            restore_inputs()
-        x_neg = origin - delta
-        tensor_to_check.set_float_element(i, x_neg)
-        y_neg = get_output()
-
-        # restore old value
-        tensor_to_check.set_float_element(i, origin)
-
-        # compute the gradient of this element and store it into a numpy array.
-        gradient_flat[i] = (y_pos - y_neg) / delta / 2
-
-    # reshape the gradient result to the shape of the source tensor.
-    return gradient_flat.reshape(tensor_to_check.get_dims())
-
-
-class GradientChecker(unittest.TestCase):
-    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
-        backward_outs = backward_op.outputs()
-        backward_names = [
-            item for key in backward_outs for item in backward_outs[key]
-        ]
-        for name in backward_names:
-            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, no_grad_set=None):
-        """ 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}
-        :param no_grad_set: the set of variables names without gradients.
-        :type no_grad_set: a set of string
-        :raises: AssertionError, there is different gradient value.
-        """
-        if no_grad_set is None:
-            no_grad_set = set()
-        backward_op = core.Operator.backward(forward_op, no_grad_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]]
-        out_names = filter(lambda x: x != empty_var_name(), out_names)
-        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.
-            abs_a[abs_a < 1e-3] = 1
-
-            diff_mat = numpy.abs(a - b) / abs_a
-            max_diff = numpy.max(diff_mat)
-
-            def err_msg():
-                offset = numpy.argmax(diff_mat > max_relative_error)
-                return "%s Variable %s max gradient diff %f over limit %f, the first " \
-                       "error element is %d" % (
-                       msg_prefix, name, max_diff, max_relative_error, offset)
-
-            self.assertLessEqual(max_diff, max_relative_error, err_msg())
-
-    def check_grad(self,
-                   forward_op,
-                   input_vars,
-                   inputs_to_check,
-                   output_name,
-                   no_grad_set=None,
-                   only_cpu=False,
-                   in_place=False,
-                   max_relative_error=0.005):
-        """
-        :param forward_op: used to create backward_op
-        :param input_vars: numpy value of input variable. The following
-            computation will use these variables.
-        :param inputs_to_check: inputs var names that should check gradient.
-        :param output_name: the output variable name of forward network.
-        :param max_relative_error: The relative tolerance parameter.
-        :param no_grad_set: used when create backward ops
-        :param only_cpu: only compute and check gradient on cpu kernel.
-        :return:
-        """
-        if no_grad_set is None:
-            no_grad_set = set()
-
-        no_tmp_out = forward_op.no_intermediate_outputs()
-        if len(no_tmp_out) != 1:
-            raise ValueError("non temp out_names should be 1")
-
-        inputs = forward_op.inputs()
-        in_names = [item for k in inputs for item in inputs[k]]
-        for no_grad in no_grad_set:
-            if no_grad not in in_names:
-                raise ValueError("no_grad should be in in_names")
-            if no_grad in inputs_to_check:
-                raise ValueError("no_grad should not be in inputs_to_check")
-
-        backward_op = core.Operator.backward(forward_op, no_grad_set)
-
-        places = [core.CPUPlace()]
-        if not only_cpu and core.is_compile_gpu() and backward_op.support_gpu():
-            places.append(core.GPUPlace(0))
-
-        # get numerical gradients
-        numeric_grads = [
-            get_numeric_gradient(
-                forward_op, input_vars, output_name, name, in_place=in_place)
-            for name in inputs_to_check
-        ]
-
-        check_names = [grad_var_name(name) for name in inputs_to_check]
-        for place in places:
-            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/op_test.py

+import unittest
+import numpy as np
+import itertools
+import paddle.v2.framework.core as core
+from paddle.v2.framework.op import Operator
+
+
+def grad_var_name(var_name):
+    return var_name + "@GRAD"
+
+
+def create_op(scope, op_type, inputs, outputs, attrs):
+    kwargs = dict()
+
+    for in_name, in_dup in Operator.get_op_inputs(op_type):
+        if in_name in inputs:
+            kwargs[in_name] = []
+            if in_dup:
+                sub_in = inputs[in_name]
+                for sub_in_name, _ in sub_in:
+                    var = scope.new_var(sub_in_name)
+                    kwargs[in_name].append(sub_in_name)
+            else:
+                var = scope.new_var(in_name)
+                kwargs[in_name].append(in_name)
+
+    for out_name, out_dup in Operator.get_op_outputs(op_type):
+        if out_name in outputs:
+            kwargs[out_name] = []
+            if out_dup:
+                sub_in = outputs[out_name]
+                for sub_in_name, _ in sub_in:
+                    var = scope.new_var(sub_in_name)
+                    kwargs[out_name].append(sub_in_name)
+            else:
+                var = scope.new_var(out_name)
+                kwargs[out_name].append(out_name)
+
+    for attr_name in Operator.get_op_attr_names(op_type):
+        if attr_name in attrs:
+            kwargs[attr_name] = attrs[attr_name]
+    return Operator(op_type, **kwargs)
+
+
+def set_input(scope, op, inputs, place):
+    for in_name, in_dup in Operator.get_op_inputs(op.type()):
+        if in_name in inputs:
+            if in_dup:
+                sub_in = inputs[in_name]
+                for sub_in_name, sub_in_array in sub_in:
+                    var = scope.find_var(sub_in_name)
+                    tensor = var.get_tensor()
+                    tensor.set_dims(sub_in_array.shape)
+                    tensor.set(sub_in_array, place)
+            else:
+                var = scope.find_var(in_name)
+                tensor = var.get_tensor()
+                arr = inputs[in_name]
+                tensor.set_dims(arr.shape)
+                tensor.set(arr, place)
+
+
+def set_output_grad(scope, op, outputs, place):
+    for out_name, out_dup in Operator.get_op_outputs(op.type()):
+        if out_name in outputs:
+            if out_dup:
+                sub_out = outputs[out_name]
+                for sub_out_name, _ in sub_out:
+                    out_tensor = scope.find_var(sub_out_name).get_tensor()
+                    grad_tensor = scope.new_var(grad_var_name(
+                        sub_out_name)).get_tensor()
+                    grad_tensor.set_dims(out_tensor.shape())
+                    data = np.ones(out_tensor.shape(), dtype=np.float32)
+                    grad_tensor.set(data, place)
+            else:
+                out_tensor = scope.find_var(out_name).get_tensor()
+                grad_tensor = scope.new_var(grad_var_name(out_name)).get_tensor(
+                )
+                grad_tensor.set_dims(out_tensor.shape())
+                data = np.ones(out_tensor.shape(), dtype=np.float32)
+                grad_tensor.set(data, place)
+
+
+def get_numeric_gradient(scope,
+                         op,
+                         inputs,
+                         input_to_check,
+                         output_name,
+                         delta=0.005,
+                         in_place=False):
+
+    set_input(scope, op, inputs, core.CPUPlace())
+    op.infer_shape(scope)
+
+    tensor_to_check = scope.find_var(input_to_check).get_tensor()
+
+    def product(dim):
+        return reduce(lambda a, b: a * b, dim, 1)
+
+    ctx = core.DeviceContext.create(core.CPUPlace())
+
+    def get_output():
+        op.run(scope, ctx)
+        return np.array(scope.find_var(output_name).get_tensor()).sum()
+
+    tensor_to_check = scope.find_var(input_to_check).get_tensor()
+    tensor_size = product(tensor_to_check.get_dims())
+    gradient_flat = np.zeros(shape=(tensor_size, ), dtype='float32')
+    # we only compute gradient of one element each time.
+    # we use a for loop to compute the gradient of every element.
+    for i in xrange(tensor_size):
+        if in_place:
+            set_input(scope, op, inputs, core.CPUPlace())
+
+        # get one input element throw it's index i.
+        origin = tensor_to_check.get_float_element(i)
+        # add delta to it, run op and then get the sum of the result tensor.
+        x_pos = origin + delta
+        tensor_to_check.set_float_element(i, x_pos)
+        y_pos = get_output()
+
+        if in_place:
+            set_input(scope, op, inputs, core.CPUPlace())
+
+        x_neg = origin - delta
+        tensor_to_check.set_float_element(i, x_neg)
+        y_neg = get_output()
+
+        tensor_to_check.set_float_element(i, origin)
+        gradient_flat[i] = (y_pos - y_neg) / delta / 2
+
+    return gradient_flat.reshape(tensor_to_check.get_dims())
+
+
+def get_backward_op(scope, op, no_grad_set):
+    backward_op = core.Operator.backward(op, no_grad_set)
+    for input in backward_op.input_vars():
+        var = scope.new_var(input)
+        var.get_tensor()
+    for output in backward_op.output_vars():
+        var = scope.new_var(output)
+        var.get_tensor()
+    return backward_op
+
+
+def get_gradient(scope, op, inputs, outputs, grad_name, place,
+                 no_grad_set=None):
+    ctx = core.DeviceContext.create(place)
+
+    set_input(scope, op, inputs, place)
+
+    op.infer_shape(scope)
+    op.run(scope, ctx)
+
+    if no_grad_set is None:
+        no_grad_set = set()
+
+    backward_op = get_backward_op(scope, op, no_grad_set)
+    set_output_grad(scope, op, outputs, place)
+
+    backward_op.infer_shape(scope)
+    backward_op.run(scope, ctx)
+
+    out = np.array(scope.find_var(grad_name).get_tensor())
+    return out
+
+
+class OpTest(unittest.TestCase):
+    def check_output_with_place(self, place):
+        self.scope = core.Scope()
+        op_inputs = self.inputs if hasattr(self, "inputs") else dict()
+        op_attrs = self.attrs if hasattr(self, "attrs") else dict()
+        self.op = create_op(self.scope, self.op_type, op_inputs, self.outputs,
+                            op_attrs)
+        if isinstance(place, core.GPUPlace) and not self.op.support_gpu():
+            return
+        set_input(self.scope, self.op, self.inputs, place)
+        self.op.infer_shape(self.scope)
+        ctx = core.DeviceContext.create(place)
+        self.op.run(self.scope, ctx)
+
+        for out_name, out_dup in Operator.get_op_outputs(self.op.type()):
+            if out_dup:
+                sub_out = self.outputs[out_name]
+                for sub_out_name in sub_out:
+                    actual = np.array(
+                        self.scope.find_var(sub_out_name).get_tensor())
+                    expect = sub_out[sub_out_name]
+                    self.assertTrue(
+                        np.allclose(
+                            actual, expect, atol=1e-05),
+                        "output name: " + out_name + "has diff")
+            else:
+                actual = np.array(self.scope.find_var(out_name).get_tensor())
+                expect = self.outputs[out_name]
+                self.assertTrue(
+                    np.allclose(
+                        actual, expect, atol=1e-05),
+                    "output name: " + out_name + "has diff")
+
+    def check_output(self):
+        places = [core.CPUPlace()]
+        if core.is_compile_gpu():
+            places.append(core.GPUPlace(0))
+        for place in places:
+            self.check_output_with_place(place)
+
+    def __assert_is_close(self, numeric_grads, analytic_grads, names,
+                          max_relative_error, msg_prefix):
+
+        for a, b, name in itertools.izip(numeric_grads, analytic_grads, names):
+            abs_a = np.abs(a)
+            abs_a[abs_a < 1e-3] = 1
+
+            diff_mat = np.abs(a - b) / abs_a
+            max_diff = np.max(diff_mat)
+
+            def err_msg():
+                offset = np.argmax(diff_mat > max_relative_error)
+                return "%s Variable %s max gradient diff %f over limit %f, the first " \
+                  "error element is %d" % (
+                   msg_prefix, name, max_diff, max_relative_error, offset)
+
+            self.assertLessEqual(max_diff, max_relative_error, err_msg())
+
+    def check_grad(self,
+                   inputs_to_check,
+                   output_name,
+                   no_grad_set=None,
+                   in_place=False,
+                   max_relative_error=0.005):
+        self.scope = core.Scope()
+        op_inputs = self.inputs if hasattr(self, "inputs") else dict()
+        op_attrs = self.attrs if hasattr(self, "attrs") else dict()
+        self.op = create_op(self.scope, self.op_type, op_inputs, self.outputs,
+                            op_attrs)
+        if no_grad_set is None:
+            no_grad_set = set()
+
+        numeric_grads = [
+            get_numeric_gradient(
+                self.scope,
+                self.op,
+                self.inputs,
+                input_to_check,
+                output_name,
+                in_place=in_place) for input_to_check in inputs_to_check
+        ]
+        grad_names = [
+            grad_var_name(input_to_check) for input_to_check in inputs_to_check
+        ]
+
+        cpu_place = core.CPUPlace()
+        cpu_analytic_grads = [
+            get_gradient(self.scope, self.op, self.inputs, self.outputs,
+                         grad_name, cpu_place, no_grad_set)
+            for grad_name in grad_names
+        ]
+
+        self.__assert_is_close(numeric_grads, cpu_analytic_grads, grad_names,
+                               max_relative_error,
+                               "Gradient Check On %s" % str(cpu_place))
+
+        if core.is_compile_gpu() and self.op.support_gpu():
+            gpu_place = core.GPUPlace(0)
+            gpu_analytic_grads = [
+                get_gradient(self.scope, self.op, self.inputs, self.outputs,
+                             grad_name, gpu_place, no_grad_set)
+                for grad_name in grad_names
+            ]
+
+            self.__assert_is_close(numeric_grads, gpu_analytic_grads,
+                                   grad_names, max_relative_error,
+                                   "Gradient Check On %s" % str(gpu_place))
+
+            for c_grad, g_grad, name in itertools.izip(
+                    cpu_analytic_grads, gpu_analytic_grads, grad_names):
+                self.assertTrue(
+                    np.allclose(
+                        c_grad, g_grad, atol=1e-4),
+                    "output name: " + name + " has diff")

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

-import numpy
-import paddle.v2.framework.core as core
-from paddle.v2.framework.op import Operator
-
-
-class OpTestMeta(type):
-    """
-    Operator Test ClassMeta.
-
-    It injects `test_all` method into user's OperatorTest class, to make Python
-    unittest module run that method.
-
-    The `test_all` read what value is stored in `self`. It use self's values to
-    create and run a operator, and check whether that op is OK or not.
-
-    See `test_add_two_op` for example usage.
-    """
-
-    def __new__(cls, name, bases, attrs):
-        obj = super(OpTestMeta, cls).__new__(cls, name, bases, attrs)
-
-        def test_all(self):
-            scope = core.Scope()
-            kwargs = dict()
-            places = [core.CPUPlace()]
-            if core.is_compile_gpu():
-                places.append(core.GPUPlace(0))
-
-            for place in places:
-                for in_name in Operator.get_op_input_names(self.type):
-                    if hasattr(self, "inputs") and in_name in self.inputs:
-                        kwargs[in_name] = in_name
-                        var = scope.new_var(in_name).get_tensor()
-                        arr = self.inputs[in_name]
-                        var.set_dims(arr.shape)
-                        var.set(arr, place)
-                    else:
-                        kwargs[in_name] = "@EMPTY@"
-
-                for out_name in Operator.get_op_output_names(self.type):
-                    if not hasattr(self, "outputs"):
-                        raise ValueError(
-                            "The test op must set self.outputs dict.")
-                    if out_name not in self.outputs:
-                        raise ValueError("The %s is not in self.outputs dict." %
-                                         (out_name))
-                    kwargs[out_name] = out_name
-                    scope.new_var(out_name).get_tensor()
-
-                for attr_name in Operator.get_op_attr_names(self.type):
-                    if hasattr(self, "attrs") and attr_name in self.attrs:
-                        kwargs[attr_name] = self.attrs[attr_name]
-
-                op = Operator(self.type, **kwargs)
-                if isinstance(place, core.GPUPlace) and not op.support_gpu():
-                    return
-
-                op.infer_shape(scope)
-
-                ctx = core.DeviceContext.create(place)
-                op.run(scope, ctx)
-
-                for out_name in Operator.get_op_output_names(self.type):
-                    actual = numpy.array(scope.find_var(out_name).get_tensor())
-                    expect = self.outputs[out_name]
-                    self.assertTrue(
-                        numpy.allclose(
-                            actual, expect, atol=1e-05),
-                        "output name: " + out_name + " has diff")
-
-        obj.test_all = test_all
-        return obj

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

 import unittest
+import numpy as np
+from op_test import OpTest
 
-import numpy
-import paddle.v2.framework.core as core
-from paddle.v2.framework.op import Operator
-
-from op_test_util import OpTestMeta
-
-
-class TestAddOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
 
+class TestAddOp(OpTest):
     def setUp(self):
-        self.type = "add"
+        self.op_type = "add"
         self.inputs = {
-            'X': numpy.random.random((102, 105)).astype("float32"),
-            'Y': numpy.random.random((102, 105)).astype("float32")
+            'X': np.random.random((102, 105)).astype("float32"),
+            'Y': np.random.random((102, 105)).astype("float32")
         }
         self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']}
 
+    def test_check_output(self):
+        self.check_output()
+
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestConcatOp(OpTest):
+    def setUp(self):
+        self.op_type = "concat"
+        x0 = np.random.random((2, 3, 2, 5)).astype('float32')
+        x1 = np.random.random((2, 3, 3, 5)).astype('float32')
+        x2 = np.random.random((2, 3, 4, 5)).astype('float32')
+        axis = 2
+        self.inputs = {'X': [('x0', x0), ('x1', x1), ('x2', x2)]}
+        self.attrs = {'axis': axis}
+        self.outputs = {'Out': np.concatenate((x0, x1, x2), axis=axis)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == '__main__':
+    unittest.main()

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

 import unittest
 import numpy as np
-from gradient_checker import GradientChecker, create_op
-from op_test_util import OpTestMeta
+from op_test import OpTest
 
 
-class TestCosSimOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestCosSimOp(OpTest):
     def setUp(self):
-        self.type = "cos_sim"
+        self.op_type = "cos_sim"
         self.inputs = {
-            'X': np.random.random((32, 64)).astype("float32"),
-            'Y': np.random.random((32, 64)).astype("float32")
+            'X': np.random.random((10, 5)).astype("float32"),
+            'Y': np.random.random((10, 5)).astype("float32")
         }
         expect_x_norm = np.linalg.norm(self.inputs['X'], axis=1)
         expect_y_norm = np.linalg.norm(self.inputs['Y'], axis=1)
@@ -23,38 +20,20 @@ class TestCosSimOp(unittest.TestCase):
             'Out': np.expand_dims(expect_out, 1)
         }
 
+    def test_check_output(self):
+        self.check_output()
 
-class TestCosSimGradOp(GradientChecker):
-    def setUp(self):
-        self.op = create_op("cos_sim")
-        self.inputs = {
-            'X': np.random.random((10, 5)).astype("float32"),
-            'Y': np.random.random((10, 5)).astype("float32")
-        }
-
-    def test_cpu_gpu_compare(self):
-        self.compare_grad(self.op, self.inputs)
-
-    def test_normal(self):
-        self.check_grad(
-            self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.05)
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.05)
 
-    def test_ignore_x(self):
+    def test_check_grad_ingore_x(self):
         self.check_grad(
-            self.op,
-            self.inputs, ["Y"],
-            "Out",
-            max_relative_error=0.05,
-            no_grad_set={"X"})
+            ['Y'], 'Out', max_relative_error=0.05, no_grad_set=set('X'))
 
-    def test_ignore_y(self):
+    def test_check_grad_ignore_y(self):
         self.check_grad(
-            self.op,
-            self.inputs, ["X"],
-            "Out",
-            max_relative_error=0.05,
-            no_grad_set={"Y"})
+            ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y'))
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()

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

 import unittest
 import numpy
-from op_test_util import OpTestMeta
-from gradient_checker import GradientChecker, create_op
+from op_test import OpTest
 
 
-class TestCrossEntropy(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestCrossEntropy(OpTest):
     def setUp(self):
-        self.type = "onehot_cross_entropy"
+        self.op_type = "onehot_cross_entropy"
         batch_size = 30
         class_num = 10
-        X = numpy.random.random((batch_size, class_num)).astype("float32")
-        label = 5 * numpy.ones(batch_size).astype("int32")
+        X = numpy.random.uniform(0.1, 1.0,
+                                 [batch_size, class_num]).astype("float32")
+        label = (class_num / 2) * numpy.ones(batch_size).astype("int32")
         self.inputs = {'X': X, 'label': label}
         Y = []
         for i in range(0, batch_size):
             Y.append(-numpy.log(X[i][label[i]]))
         self.outputs = {'Y': numpy.array(Y).astype("float32")}
 
+    def test_check_output(self):
+        self.check_output()
 
-class CrossEntropyGradOpTest(GradientChecker):
     def test_check_grad(self):
-        op = create_op("onehot_cross_entropy")
-        batch_size = 30
-        class_num = 10
-        inputs = {
-            "X": numpy.random.uniform(
-                0.1, 1.0, [batch_size, class_num]).astype("float32"),
-            "label": (class_num / 2) * numpy.ones(batch_size).astype("int32")
-        }
-        self.check_grad(op, inputs, set("X"), "Y")
+        self.check_grad(['X'], 'Y')
 
 
 if __name__ == "__main__":

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

 import unittest
-from op_test_util import OpTestMeta
-import numpy
+import numpy as np
+from op_test import OpTest
 
 
-class TestFillZerosLikeOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestFillZerosLikeOp(OpTest):
     def setUp(self):
-        self.type = "fill_zeros_like"
-        self.inputs = {'Src': numpy.random.random((219, 232)).astype("float32")}
-        self.outputs = {'Dst': numpy.zeros_like(self.inputs['Src'])}
+        self.op_type = "fill_zeros_like"
+        self.inputs = {'Src': np.random.random((219, 232)).astype("float32")}
+        self.outputs = {'Dst': np.zeros_like(self.inputs["Src"])}
+
+    def test_check_output(self):
+        self.check_output()
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()

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

 import unittest
-from op_test_util import OpTestMeta
-from gradient_checker import GradientChecker, create_op
-import numpy
-import paddle.v2.framework.core as core
-from paddle.v2.framework.op import Operator
+import numpy as np
+from op_test import OpTest
 
 
-class TestGatherOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestGatherOp(OpTest):
     def setUp(self):
-        self.type = "gather"
-        xnp = numpy.random.random((10, 20)).astype("float32")
-        self.inputs = {
-            'X': xnp,
-            'Index': numpy.array([1, 3, 5]).astype("int32")
-        }
-        self.outputs = {'Out': self.inputs['X'][self.inputs['Index']]}
+        self.op_type = "gather"
+        xnp = np.random.random((10, 20)).astype("float32")
+        self.inputs = {'X': xnp, 'Index': np.array([1, 3, 5]).astype("int32")}
+        self.outputs = {'Out': self.inputs["X"][self.inputs["Index"]]}
 
+    def test_check_output(self):
+        self.check_output()
 
-class TestGatherGradOp(GradientChecker):
-    def test_gather_grad(self):
-        op = create_op("gather")
-        xnp = numpy.random.random((10, 20)).astype("float32")
-        inputs = {'X': xnp, 'Index': numpy.array([1, 3, 5]).astype("int32")}
-        self.check_grad(op, inputs, set("X"), "Out")
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
 
 
 if __name__ == "__main__":

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

@@ -14,11 +14,11 @@ class GaussianRandomTest(unittest.TestCase):
 
     def gaussian_random_test(self, place):
         scope = core.Scope()
-        scope.new_var("Out").get_tensor()
+        scope.new_var('Out').get_tensor()
 
         op = Operator(
             "gaussian_random",
-            Out="Out",
+            Out='Out',
             dims=[1000, 784],
             mean=.0,
             std=1.,
@@ -27,10 +27,10 @@ class GaussianRandomTest(unittest.TestCase):
         op.infer_shape(scope)
         context = core.DeviceContext.create(place)
         op.run(scope, context)
-        tensor = numpy.array(scope.find_var("Out").get_tensor())
+        tensor = numpy.array(scope.find_var('Out').get_tensor())
         self.assertAlmostEqual(numpy.mean(tensor), .0, delta=0.1)
         self.assertAlmostEqual(numpy.std(tensor), 1., delta=0.1)
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()

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
+import numpy as np
+import paddle.v2.framework.core as core
+from op_test import get_numeric_gradient
+from op_test import create_op
 
 
 class GetNumericGradientTest(unittest.TestCase):
     def test_add_op(self):
-        add_op = Operator("add", 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")
+        x = np.random.random((10, 1)).astype("float32")
+        y = np.random.random((10, 1)).astype("float32")
+        z = x + y
+        scope = core.Scope()
+        add_op = create_op(scope, "add", {'X': x, 'Y': y}, {'Out': z}, dict())
+        arr = get_numeric_gradient(scope, add_op, {'X': x, 'Y': y}, 'X', 'Out')
         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)
+            shiftx = x - np.max(x)
+            exps = np.exp(shiftx)
+            return exps / np.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, :])
+                d = np.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)
+        X = np.random.random((2, 2)).astype("float32")
+        Y = np.apply_along_axis(stable_softmax, 1, X)
+        dY = np.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)
+        scope = core.Scope()
+        softmax_op = create_op(scope, "softmax", {"X": X}, {"Y": Y}, dict())
+
+        arr = get_numeric_gradient(scope, softmax_op, {"X": X}, "X", "Y")
+        np.testing.assert_almost_equal(arr, dX, decimal=1e-2)
 
 
 if __name__ == "__main__":

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

 import unittest
 import numpy as np
-from op_test_util import OpTestMeta
-from gradient_checker import GradientChecker, create_op
+from op_test import OpTest
 
 
-class TestSigmoidOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
-
+class TestLookupTableOp(OpTest):
     def setUp(self):
-        self.type = 'lookup_table'
-        table = np.random.random((17, 31)).astype('float32')
-        ids = np.random.randint(0, 17, 4).astype('int32')
+        self.op_type = "lookup_table"
+        table = np.random.random((17, 31)).astype("float32")
+        ids = np.random.randint(0, 17, 4).astype("int32")
         self.inputs = {'W': table, 'Ids': ids}
         self.outputs = {'Out': table[ids]}
 
+    def test_check_output(self):
+        self.check_output()
 
-class TestSigmoidGradOp(GradientChecker):
-    def test_grad(self):
-        op = create_op('lookup_table')
-        table = np.random.random((17, 31)).astype('float32')
-        ids = np.random.randint(0, 17, 4).astype('int32')
-        inputs = {'W': table, 'Ids': ids}
-        # comapre gradients 
-        self.compare_grad(op, inputs, set(['Ids']))
-        # check gradients 
-        self.check_grad(op, inputs, set('W'), 'Out')
+    def test_check_grad(self):
+        self.check_grad(['W'], 'Out', no_grad_set=set('Ids'))
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()