update code

.travis.yml

@@ -25,9 +25,9 @@ addons:
     packages:
       - gcc-4.8
       - g++-4.8
+      - gfortran-4.8
       - git
       - build-essential
-      - libatlas-base-dev
       - python
       - python-pip
       - python2.7-dev

cmake/cblas.cmake

@@ -16,7 +16,7 @@
 set(CBLAS_FOUND OFF)
 
 ## Find MKL First.
-set(MKL_ROOT $ENV{MKL_ROOT} CACHE PATH "Folder contains MKL")
+set(MKL_ROOT $ENV{MKLROOT} CACHE PATH "Folder contains MKL")
 
 find_path(MKL_INCLUDE_DIR mkl.h PATHS
   ${MKL_ROOT}/include)

cmake/external/openblas.cmake

@@ -15,7 +15,6 @@
 INCLUDE(cblas)
 
 IF(NOT ${CBLAS_FOUND})
-    MESSAGE(FATAL_ERROR "Please install OpenBlas, MKL or ATLAS.")
     INCLUDE(ExternalProject)
 
     SET(CBLAS_SOURCES_DIR ${THIRD_PARTY_PATH}/openblas)
@@ -28,20 +27,40 @@ IF(NOT ${CBLAS_FOUND})
         SET(CBLAS_LIBRARIES "${CBLAS_INSTALL_DIR}/lib/libopenblas.a" CACHE FILEPATH "openblas library" FORCE)
     ENDIF(WIN32)
 
+    IF(CMAKE_COMPILER_IS_GNUCC)
+        ENABLE_LANGUAGE(Fortran)
+        LIST(APPEND CBLAS_LIBRARIES gfortran pthread)
+    ENDIF(CMAKE_COMPILER_IS_GNUCC)
+
+    IF(NOT CMAKE_Fortran_COMPILER)
+        MESSAGE(FATAL_ERROR "To build lapack in libopenblas, "
+                "you need to set gfortran compiler: cmake .. -DCMAKE_Fortran_COMPILER=...")
+    ENDIF(NOT CMAKE_Fortran_COMPILER)
+
     ExternalProject_Add(
         openblas
         ${EXTERNAL_PROJECT_LOG_ARGS}
-        URL                 "https://github.com/xianyi/OpenBLAS/archive/v0.2.19.tar.gz"
+        GIT_REPOSITORY      https://github.com/xianyi/OpenBLAS.git
+        GIT_TAG             v0.2.19
         PREFIX              ${CBLAS_SOURCES_DIR}
         INSTALL_DIR         ${CBLAS_INSTALL_DIR}
         BUILD_IN_SOURCE     1
-        CONFIGURE_COMMAND   ""
-        BUILD_COMMAND       make CC=${CMAKE_C_COMPILER} FC=${CMAKE_Fortran_COMPILER}
-        INSTALL_COMMAND     make install PREFIX=<INSTALL_DIR>
+        BUILD_COMMAND       ${CMAKE_MAKE_PROGRAM} FC=${CMAKE_Fortran_COMPILER} CC=${CMAKE_C_COMPILER} HOSTCC=${CMAKE_C_COMPILER} NO_SHARED=1 libs netlib
+        INSTALL_COMMAND     ${CMAKE_MAKE_PROGRAM} install NO_SHARED=1 PREFIX=<INSTALL_DIR>
         UPDATE_COMMAND      ""
+        CONFIGURE_COMMAND   ""
+    )
+
+    ExternalProject_Add_Step(
+        openblas lapacke_install
+        COMMAND ${CMAKE_COMMAND} -E copy "${CBLAS_SOURCES_DIR}/src/openblas/lapack-netlib/LAPACKE/include/lapacke_mangling_with_flags.h" "${CBLAS_INSTALL_DIR}/include/lapacke_mangling.h"
+        COMMAND ${CMAKE_COMMAND} -E copy "${CBLAS_SOURCES_DIR}/src/openblas/lapack-netlib/LAPACKE/include/lapacke.h" "${CBLAS_INSTALL_DIR}/include/lapacke.h"
+        COMMAND ${CMAKE_COMMAND} -E copy "${CBLAS_SOURCES_DIR}/src/openblas/lapack-netlib/LAPACKE/include/lapacke_config.h" "${CBLAS_INSTALL_DIR}/include/lapacke_config.h"
+        COMMAND ${CMAKE_COMMAND} -E copy "${CBLAS_SOURCES_DIR}/src/openblas/lapack-netlib/LAPACKE/include/lapacke_utils.h" "${CBLAS_INSTALL_DIR}/include/lapacke_utils.h"
+        DEPENDEES install
     )
 
     LIST(APPEND external_project_dependencies openblas)
-ENDIF()
+ENDIF(NOT ${CBLAS_FOUND})
 
 INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})

cmake/external/protobuf.cmake

@@ -54,6 +54,7 @@ ExternalProject_Add(
   CONFIGURE_COMMAND
     ${CMAKE_COMMAND} ${PROTOBUF_SOURCES_DIR}/src/protobuf/cmake
     -Dprotobuf_BUILD_TESTS=OFF
+    -DZLIB_ROOT:FILEPATH=${ZLIB_ROOT}
     -DCMAKE_POSITION_INDEPENDENT_CODE=ON
     -DCMAKE_BUILD_TYPE=Release
     -DCMAKE_INSTALL_PREFIX=${PROTOBUF_INSTALL_DIR}

cmake/external/python.cmake

@@ -31,6 +31,7 @@ IF(PYTHONLIBS_FOUND AND PYTHONINTERP_FOUND)
         "please use pip to upgrade protobuf.")
     ENDIF(${PY_GOOGLE.PROTOBUF_VERSION} VERSION_LESS "3.0.0")
 ELSE(PYTHONLIBS_FOUND AND PYTHONINTERP_FOUND)
+    MESSAGE(FATAL_ERROR "Please install python 2.7 before building PaddlePaddle.")
     ##################################### PYTHON ########################################
     SET(PYTHON_SOURCES_DIR ${THIRD_PARTY_PATH}/python)
     SET(PYTHON_INSTALL_DIR ${THIRD_PARTY_PATH}/install/python)

cmake/flags.cmake

@@ -96,6 +96,7 @@ set(COMMON_FLAGS
     -Wno-unused-parameter
     -Wno-unused-function
     -Wno-error=literal-suffix
+    -Wno-error=sign-compare
     -Wno-error=unused-local-typedefs)
 
 set(GPU_COMMON_FLAGS
@@ -105,6 +106,7 @@ set(GPU_COMMON_FLAGS
     -Wdelete-non-virtual-dtor
     -Wno-unused-parameter
     -Wno-unused-function
+    -Wno-error=sign-compare
     -Wno-error=literal-suffix
     -Wno-error=unused-local-typedefs
     -Wno-error=unused-function  # Warnings in Numpy Header.

cmake/system.cmake

@@ -21,6 +21,7 @@ ELSE(WIN32)
         SET(MACOS_VERSION ${VERSION})
         SET(HOST_SYSTEM "macosx")
     ELSE(APPLE)
+
         IF(EXISTS "/etc/issue")
             FILE(READ "/etc/issue" LINUX_ISSUE)
             IF(LINUX_ISSUE MATCHES "CentOS")
@@ -31,6 +32,14 @@ ELSE(WIN32)
                 SET(HOST_SYSTEM "ubuntu")
             ENDIF()
         ENDIF(EXISTS "/etc/issue")
+
+        IF(EXISTS "/etc/redhat-release")
+            FILE(READ "/etc/redhat-release" LINUX_ISSUE)
+            IF(LINUX_ISSUE MATCHES "CentOS")
+                SET(HOST_SYSTEM "centos")
+            ENDIF()
+        ENDIF(EXISTS "/etc/redhat-release")
+
     ENDIF(APPLE)
 ENDIF(WIN32)
 
@@ -47,7 +56,7 @@ SET(EXTERNAL_PROJECT_LOG_ARGS
     LOG_DOWNLOAD    0     # Wrap download in script to log output
     LOG_UPDATE      1     # Wrap update in script to log output
     LOG_CONFIGURE   1     # Wrap configure in script to log output
-    LOG_BUILD       1     # Wrap build in script to log output
+    LOG_BUILD       0     # Wrap build in script to log output
     LOG_TEST        1     # Wrap test in script to log output
-    LOG_INSTALL     1     # Wrap install in script to log output
+    LOG_INSTALL     0     # Wrap install in script to log output
 )

doc/getstarted/build_and_install/build_from_source_en.md

@@ -4,6 +4,8 @@ Installing from Sources
 * [1. Download and Setup](#download)
 * [2. Requirements](#requirements)
 * [3. Build on Ubuntu](#ubuntu)
+* [4. Build on Centos](#centos)
+
 
 ## <span id="download">Download and Setup</span> 
 You can download PaddlePaddle from the [github source](https://github.com/PaddlePaddle/Paddle).
@@ -64,7 +66,8 @@ As a simple example, consider the following:
 
 1. **BLAS Dependencies(optional)**
   
-    Paddle will find BLAS from system's default path. But you can specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`.
+    CMake will search BLAS libraries from system. If not found, OpenBLAS will be downloaded, built and installed automatically.
+    To utilize preinstalled BLAS， you can simply specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`.
 
     ```bash
     # specify MKL
@@ -99,7 +102,7 @@ As a simple example, consider the following:
     ```bash
     # necessary
     sudo apt-get update
-    sudo apt-get install -y g++ make cmake build-essential libatlas-base-dev python python-pip libpython-dev git
+    sudo apt-get install -y g++ make cmake build-essential python python-pip libpython-dev git
     sudo pip install wheel numpy
     sudo pip install 'protobuf>=3.0.0'
     ```
@@ -150,3 +153,64 @@ export PATH=<path to install>/bin:$PATH
 # install PaddlePaddle Python modules.
 sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
 ```
+## <span id="centos">Build on Centos 7</span>
+
+### Install Dependencies
+
+- **CPU Dependencies**
+
+    ```bash
+    # necessary
+    sudo yum update
+    sudo yum install -y epel-release
+    sudo yum install -y make cmake3 python-devel python-pip gcc-gfortran swig git
+    sudo pip install wheel numpy
+    sudo pip install 'protobuf>=3.0.0'
+    ```
+  
+- **GPU Dependencies (optional)**
+
+    To build GPU version, you will need the following installed:
+
+        1. a CUDA-capable GPU
+        2. A supported version of Linux with a gcc compiler and toolchain
+        3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
+        4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
+
+    The CUDA development environment relies on tight integration with the host development environment,
+    including the host compiler and C runtime libraries, and is therefore only supported on
+    distribution versions that have been qualified for this CUDA Toolkit release.
+        
+    After downloading cuDNN library, issue the following commands:
+
+    ```bash
+    sudo tar -xzf cudnn-7.5-linux-x64-v5.1.tgz -C /usr/local
+    sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
+    ```
+    Then you need to set LD\_LIBRARY\_PATH, PATH environment variables in ~/.bashrc.
+
+    ```bash
+    export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
+    export PATH=/usr/local/cuda/bin:$PATH
+    ```
+
+### Build and Install
+
+As usual, the best option is to create build folder under paddle project directory.
+
+```bash
+mkdir build && cd build
+``` 
+
+Finally, you can build and install PaddlePaddle:
+
+```bash
+# you can add build option here, such as:    
+cmake3 .. -DCMAKE_INSTALL_PREFIX=<path to install>
+# please use sudo make install, if you want to install PaddlePaddle into the system
+make -j `nproc` && make install
+# set PaddlePaddle installation path in ~/.bashrc
+export PATH=<path to install>/bin:$PATH
+# install PaddlePaddle Python modules.
+sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
+```

doc/howto/deep_model/rnn/hierarchical_layer_cn.rst

@@ -32,7 +32,7 @@ pooling_layer 的使用示例如下，详细见 :ref:`api_trainer_config_helpers
         
 - `pooling_type` 目前支持两种，分别是：MaxPooling()和AvgPooling()。
 
-- `agg_level=AggregateLevel.TIMESTEP` 时（默认值）：
+- `agg_level=AggregateLevel.EACH_TIMESTEP` 时（默认值）：
 
   - 作用：双层序列经过运算变成一个0层序列，或单层序列经过运算变成一个0层序列
   - 输入：一个双层序列，或一个单层序列
@@ -54,7 +54,7 @@ last_seq 的使用示例如下（ :ref:`api_trainer_config_helpers_layers_first_
         last = last_seq(input=layer,
                         agg_level=AggregateLevel.EACH_SEQUENCE)
         
-- `agg_level=AggregateLevel.TIMESTEP` 时（默认值）：
+- `agg_level=AggregateLevel.EACH_TIMESTEP` 时（默认值）：
 
   - 作用：一个双层序列经过运算变成一个0层序列，或一个单层序列经过运算变成一个0层序列
   - 输入：一个双层序列或一个单层序列

paddle/function/BufferArg.cpp

@@ -20,23 +20,27 @@ limitations under the License. */
 namespace paddle {
 
 const SequenceArg& BufferArg::sequence() const {
-  // CHECK_EQ(bufferType_, TENSOR_SEQUENCE_DATA);
+  CHECK_EQ(bufferType_, TENSOR_SEQUENCE_DATA);
   return dynamic_cast<const SequenceArg&>(*this);
 }
 
 const SparseMatrixArg& BufferArg::sparse() const {
-  // CHECK_EQ(bufferType_, TENSOR_SPARSE);
+  CHECK_EQ(bufferType_, TENSOR_SPARSE);
   return dynamic_cast<const SparseMatrixArg&>(*this);
 }
 
 SparseMatrixArg::SparseMatrixArg(const CpuSparseMatrix& sparse, ArgType argType)
     : BufferArg(sparse, argType),
       row_(reinterpret_cast<void*>(sparse.getRows()), VALUE_TYPE_INT32),
-      col_(reinterpret_cast<void*>(sparse.getCols()), VALUE_TYPE_INT32) {}
+      col_(reinterpret_cast<void*>(sparse.getCols()), VALUE_TYPE_INT32) {
+  bufferType_ = TENSOR_SPARSE;
+}
 
 SparseMatrixArg::SparseMatrixArg(const GpuSparseMatrix& sparse, ArgType argType)
     : BufferArg(sparse, argType),
       row_(reinterpret_cast<void*>(sparse.getRows()), VALUE_TYPE_INT32),
-      col_(reinterpret_cast<void*>(sparse.getCols()), VALUE_TYPE_INT32) {}
+      col_(reinterpret_cast<void*>(sparse.getCols()), VALUE_TYPE_INT32) {
+  bufferType_ = TENSOR_SPARSE;
+}
 
 }  // namespace paddle

paddle/function/BufferArg.h

@@ -23,10 +23,11 @@ limitations under the License. */
 namespace paddle {
 
 enum BufferType {
-  TENSOR_NORMAL = 0,
-  TENSOR_SEQUENCE_ID = 1,
-  TENSOR_SEQUENCE_DATA = 2,
-  TENSOR_SPARSE = 3
+  TENSOR_UNKNOWN = 0,
+  TENSOR_NORMAL = 1,
+  TENSOR_SEQUENCE_ID = 2,
+  TENSOR_SEQUENCE_DATA = 3,
+  TENSOR_SPARSE = 4
 };
 
 enum SparseDataType {
@@ -39,7 +40,6 @@ enum SparseDataFormat { SPARSE_CSR_FORMAT = 0, SPARSE_CSC_FORMAT = 1 };
 class BufferArg;
 class SequenceArg;
 class SparseMatrixArg;
-typedef std::shared_ptr<BufferArg> BufferArgPtr;
 
 /**
  * \brief BufferArg used as the argument type of Function.
@@ -50,6 +50,11 @@ typedef std::shared_ptr<BufferArg> BufferArgPtr;
  * 3. SequenceArg for a Buffer of sequence data.
  * 4. SparseMatrixArg for a Buffer of sparse matrix.
  *
+ * Buffer shape
+ * For most buffers, the first dimension `shape()[0]` represents
+ * the size of the mini-batch.
+ *
+ * Buffer argType
  * There is an ArgType property for the BufferArg used as Function Output.
  * Whether the result of the Function calculation is assigned to the
  * output Buffer or added to the output Buffer is determined by the
@@ -71,6 +76,14 @@ public:
   ArgType getArgType() const { return argType_; }
 
 public:
+  BufferArg(ValueType valueType,
+            const TensorShape& shape,
+            ArgType argType = UNSPECIFIED)
+      : buf_(nullptr),
+        valueType_(valueType),
+        shape_(shape),
+        argType_(argType) {}
+
   BufferArg(void* buf,
             ValueType valueType,
             const TensorShape& shape,
@@ -86,6 +99,7 @@ public:
         valueType_(DataType<real>::value),
         shape_(2),
         argType_(argType) {
+    bufferType_ = TENSOR_NORMAL;
     shape_.setDim(0, matrix.getHeight());
     shape_.setDim(1, matrix.getWidth());
   }
@@ -98,6 +112,7 @@ public:
         valueType_(DataType<real>::value),
         shape_(shape),
         argType_(argType) {
+    bufferType_ = TENSOR_NORMAL;
     CHECK_EQ(matrix.getElementCnt(), shape.getElements());
   }
 
@@ -107,6 +122,7 @@ public:
         valueType_(DataType<real>::value),
         shape_(1),
         argType_(argType) {
+    bufferType_ = TENSOR_NORMAL;
     shape_.setDim(0, vector.getSize());
   }
 
@@ -116,6 +132,7 @@ public:
         valueType_(VALUE_TYPE_INT32),
         shape_(1),
         argType_(argType) {
+    bufferType_ = TENSOR_NORMAL;
     shape_.setDim(0, vector.getSize());
   }
 
@@ -150,6 +167,8 @@ public:
   ValueType valueType() const { return valueType_; }
   BufferType bufferType() const { return bufferType_; }
   const TensorShape& shape() const { return shape_; }
+  bool isSparse() const { return (TENSOR_SPARSE == bufferType_); }
+  bool isSequenceArg() const { return TENSOR_SEQUENCE_DATA == bufferType_; }
 
   const SequenceArg& sequence() const;
   const SparseMatrixArg& sparse() const;
@@ -158,8 +177,8 @@ protected:
   void* buf_;
   ValueType valueType_;
   TensorShape shape_;
-  BufferType bufferType_;
-  ArgType argType_ = UNSPECIFIED;
+  BufferType bufferType_{TENSOR_UNKNOWN};
+  ArgType argType_{UNSPECIFIED};
   // leading dimensions. The size is dims_.size()
   // Dims lds_;
 };
@@ -170,15 +189,24 @@ protected:
 // if a < b then value_.buf_[a] < value_.buf_[b]
 class SequenceIdArg : public BufferArg {
 public:
+  SequenceIdArg(const TensorShape& shape, ArgType argType = UNSPECIFIED)
+      : BufferArg(VALUE_TYPE_INT32, shape, argType) {
+    CHECK_EQ(shape_.ndims(), (size_t)1);
+    CHECK_GT(shape_[0], 1);
+    numSeqs_ = shape_[0] - 1;
+  }
+
   SequenceIdArg(void* buf,
                 const TensorShape& shape,
                 ArgType argType = UNSPECIFIED)
       : BufferArg(buf, VALUE_TYPE_INT32, shape, argType) {
+    bufferType_ = TENSOR_SEQUENCE_ID;
     CHECK_EQ(shape_.ndims(), (size_t)1);
     numSeqs_ = shape_[0] - 1;
   }
 
   SequenceIdArg(const IVector& vector) : BufferArg(vector) {
+    bufferType_ = TENSOR_SEQUENCE_ID;
     numSeqs_ = shape_[0] - 1;
   }
 
@@ -190,26 +218,41 @@ private:
   size_t numSeqs_;
 };
 
-// sequence data
+// sequences data
+// For mini-batch calculate,
+// one batch can contain more than one sequence of data.
+// SequenceArg can be used to represent sequences that contain multiple
+// unequal lengths.
 class SequenceArg : public BufferArg {
 public:
+  SequenceArg(ValueType valueType,
+              const TensorShape& shape,
+              ArgType argType = UNSPECIFIED)
+      : BufferArg(valueType, shape, argType), startPositions_(TensorShape()) {}
+
   SequenceArg(void* buf,
               ValueType valueType,
               const TensorShape& shape,
               const SequenceIdArg& startPositions,
               ArgType argType = UNSPECIFIED)
       : BufferArg(buf, valueType, shape, argType),
-        startPositions_(startPositions) {}
+        startPositions_(startPositions) {
+    bufferType_ = TENSOR_SEQUENCE_DATA;
+  }
 
   SequenceArg(const Matrix& matrix,
               const IVector& vector,
               ArgType argType = UNSPECIFIED)
-      : BufferArg(matrix, argType), startPositions_(vector) {}
+      : BufferArg(matrix, argType), startPositions_(vector) {
+    bufferType_ = TENSOR_SEQUENCE_DATA;
+  }
 
   ~SequenceArg() {}
 
   void* getIdBuf() const { return startPositions_.data(); }
   size_t numSeqs() const { return startPositions_.numSeqs(); }
+  SequenceIdArg& getSequenceId() { return startPositions_; }
+  const SequenceIdArg& getSequenceId() const { return startPositions_; }
 
 private:
   SequenceIdArg startPositions_;
@@ -235,6 +278,7 @@ public:
         nnz_(nnz),
         format_(format),
         type_(type) {
+    bufferType_ = TENSOR_SPARSE;
     CHECK((valueType == VALUE_TYPE_FLOAT) || (valueType == VALUE_TYPE_DOUBLE));
     CHECK_EQ(shape_.ndims(), (size_t)2);
     CHECK_EQ(row_.shape().ndims(), (size_t)1);

paddle/function/BufferArgTest.cpp

@@ -14,9 +14,7 @@ limitations under the License. */
 
 #include "BufferArg.h"
 #include <gtest/gtest.h>
-#include "Function.h"
 #include "paddle/math/MemoryHandle.h"
-#include "paddle/math/SparseMatrix.h"
 
 namespace paddle {
 
@@ -37,55 +35,4 @@ TEST(BufferTest, SequenceIdArg) {
   EXPECT_EQ(buffer.numSeqs(), 9);
 }
 
-TEST(BufferTest, asArgument) {
-  MatrixPtr matrix = Matrix::create(100, 200);
-  VectorPtr vector = Vector::create(100, false);
-  CpuSparseMatrix sparse(200, 300, 50);
-
-  // prepare arguments
-  BufferArgs argments;
-  argments.addArg(*matrix);
-  argments.addArg(*vector);
-  argments.addArg(sparse);
-
-  // function
-  auto function = [=](const BufferArgs& inputs) {
-    EXPECT_EQ(inputs.size(), 3);
-
-    // check inputs[0]
-    EXPECT_EQ(inputs[0].shape().ndims(), 2);
-    EXPECT_EQ(inputs[0].shape()[0], 100);
-    EXPECT_EQ(inputs[0].shape()[1], 200);
-    EXPECT_EQ(inputs[0].data(), matrix->getData());
-
-    EXPECT_EQ(inputs[0].matrix<DEVICE_TYPE_CPU>().getHeight(),
-              matrix->getHeight());
-    EXPECT_EQ(inputs[0].matrix<DEVICE_TYPE_CPU>().getWidth(),
-              matrix->getWidth());
-    EXPECT_EQ(inputs[0].matrix<DEVICE_TYPE_CPU>().getData(), matrix->getData());
-
-    // check inputs[1]
-    EXPECT_EQ(inputs[1].shape().ndims(), 1);
-    EXPECT_EQ(inputs[1].shape()[0], 100);
-    EXPECT_EQ(inputs[1].data(), vector->getData());
-    CpuVector inVector = inputs[1].vector<real, DEVICE_TYPE_CPU>();
-    EXPECT_EQ(inVector.getSize(), vector->getSize());
-    EXPECT_EQ(inVector.getData(), vector->getData());
-
-    // check inputs[2]
-    EXPECT_EQ(inputs[2].shape().ndims(), 2);
-    EXPECT_EQ(inputs[2].shape()[0], 200);
-    EXPECT_EQ(inputs[2].shape()[1], 300);
-    EXPECT_EQ(inputs[2].data(), sparse.getData());
-    // CHECK_EQ(inputs[2].sparse().nnz(), 50);
-    // CHECK_EQ(inputs[2].sparse().dataFormat(), SPARSE_CSR_FORMAT);
-    // CHECK_EQ(inputs[2].sparse().dataType(), SPARSE_FLOAT_VALUE);
-    EXPECT_EQ(inputs[2].sparse().getRowBuf(), sparse.getRows());
-    EXPECT_EQ(inputs[2].sparse().getColBuf(), sparse.getCols());
-  };
-
-  // call function
-  function(argments);
-}
-
 }  // namespace paddle

paddle/function/CMakeLists.txt

@@ -19,13 +19,13 @@ if(WITH_TESTING)
     # TODO:
     # file(GLOB test_files . *OpTest.cpp)
     # add_executable(${test_bin} EXCLUDE_FROM_ALL ${test_files})
-    # add_simple_unittest(CrossMapNormalOpTest)
+    add_simple_unittest(CrossMapNormalOpTest)
     add_simple_unittest(TensorShapeTest)
     add_simple_unittest(TensorTypeTest)
     add_simple_unittest(BufferArgTest)
     add_simple_unittest(FunctionTest)
+    add_simple_unittest(ContextProjectionOpTest)
     add_simple_unittest(PadOpTest)
-    # add_simple_unittest(ContextProjectionOpTest)
 endif()
 endif()
 

paddle/function/ContextProjectionOp.cpp

@@ -17,7 +17,10 @@ limitations under the License. */
 #include "paddle/math/Vector.h"
 
 namespace paddle {
-
+/**
+ * Context Projection Forward with CPU Matrix Device.
+ *
+ */
 template <>
 void ContextProjectionForward<DEVICE_TYPE_CPU>(CpuMatrix& out_mat,
                                                const CpuMatrix& input_mat,
@@ -70,10 +73,30 @@ void ContextProjectionForward<DEVICE_TYPE_CPU>(CpuMatrix& out_mat,
 }
 
 /**
- * \param inputs[0] input value.
- * \param inputs[1] input weight.
- * \param inputs[2] input sequence.
- * \param outputs[0] output value.
+ * Paddle Function for Context Projection Forward.
+ * Calculate the output layer value sequence after context projection.
+ *
+ * What is Context Projection for a sequence?
+ * For example, assumed input (x) has 4 words and the dimension of each word
+ * representation is 2. If we use zero to pad instead of learned weight to pad,
+ * and the context_lenth is 3, the output (y) is:
+ *
+ * @code
+ *  x = [a1, a2;
+ *       b1, b2;
+ *       c1, c2;
+ *       d1, d2]
+ *  y = [0,  0,  a1, a2, b1, b2;
+ *       a1, a2, b1, b2, c1, c2;
+ *       b1, b2, c1, c2, d1, d2;
+ *       c1, c2, d1, d2, 0,  0]
+ * @endcode
+ *
+ * \param outputs[0].matrix   output layer value, n * (d * l)
+ * \param outputs[0].vector   start position sequence, n * 1
+ * \param inputs[0].matrix    input layer value, n * d
+ * \param inputs[0].vector    start position sequence, n * 1
+ * \param inputs[1].matrix    input layer weight, pad * d
  */
 template <DeviceType Device>
 class ContextProjectionForwardFunc : public FunctionBase {
@@ -85,28 +108,37 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    CHECK_EQ((size_t)3, inputs.size());
+    CHECK(1 == inputs.size() || 2 == inputs.size());
     CHECK_EQ((size_t)1, outputs.size());
+    CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg())
+        << "SequenceArg required here";
+    const auto val_seqs = dynamic_cast<const SequenceArg&>(inputs[0]);
+    auto out_seq = dynamic_cast<const SequenceArg&>(outputs[0]);
 
-    CHECK(outputs[0].data() && inputs[0].data() && inputs[2].data());
-    CHECK_EQ(outputs[0].shape().ndims(), (size_t)2);
-    CHECK_EQ(inputs[0].shape().ndims(), (size_t)2);
-    CHECK_EQ(inputs[1].shape().ndims(), (size_t)2);
-    CHECK_EQ(inputs[2].shape().ndims(), (size_t)1);
+    CHECK(out_seq.data() && val_seqs.data() && val_seqs.getSequenceId().data());
+    CHECK_EQ(out_seq.shape().ndims(), (size_t)2);
+    CHECK_EQ(val_seqs.shape().ndims(), (size_t)2);
+    CHECK_EQ(val_seqs.getSequenceId().shape().ndims(), (size_t)1);
+    if (2 == inputs.size()) {
+      CHECK_EQ(inputs[1].shape().ndims(), (size_t)2);
+    }
     /// dim of output = dim of input * context_length
-    CHECK_EQ(outputs[0].shape()[1], inputs[0].shape()[1] * context_length_);
-    /// dim of input == dim of weight
-    CHECK_EQ(inputs[0].shape()[1], inputs[1].shape()[1]);
+    CHECK_EQ(out_seq.shape()[1], val_seqs.shape()[1] * context_length_);
     /// input and output has the same batch_size
-    CHECK_EQ(inputs[0].shape()[0], outputs[0].shape()[0]);
+    CHECK_EQ(val_seqs.shape()[0], out_seq.shape()[0]);
+    /// dim of input == dim of weight
+    if (2 == inputs.size()) {
+      CHECK_EQ(val_seqs.shape()[1], inputs[1].shape()[1]);
+    }
 
-    CHECK_EQ(outputs[0].getArgType(), ADD_TO);
-    auto out_mat = outputs[0].matrix<Device>();
-    auto in_mat = inputs[0].matrix<Device>();
-    auto w_mat = !inputs[1].data()
-                     ? typename Tensor<real, Device>::Matrix(nullptr, 0, 0)
-                     : inputs[1].matrix<Device>();
-    auto seq_vec = inputs[2].vector<int, Device>();
+    CHECK_EQ(out_seq.getArgType(), ADD_TO);
+    auto out_mat = out_seq.matrix<Device>();
+    const auto in_mat = val_seqs.matrix<Device>();
+    const auto w_mat =
+        (2 == inputs.size())
+            ? inputs[1].matrix<Device>()
+            : typename Tensor<real, Device>::Matrix(nullptr, 0, 0);
+    const auto seq_vec = val_seqs.getSequenceId().vector<int, Device>();
     ContextProjectionForward<Device>(out_mat,
                                      in_mat,
                                      w_mat,
@@ -122,8 +154,12 @@ private:
   size_t begin_pad_;
 };
 
+/**
+ * Context Projection Backward with CPU Matrix Device.
+ *
+ */
 template <>
-void ContextProjectionBackward<DEVICE_TYPE_CPU>(CpuMatrix& out_grad_mat,
+void ContextProjectionBackward<DEVICE_TYPE_CPU>(const CpuMatrix& out_grad_mat,
                                                 CpuMatrix& in_grad_mat,
                                                 CpuMatrix& w_grad_mat,
                                                 const CpuIVector& seq_vec,
@@ -146,7 +182,8 @@ void ContextProjectionBackward<DEVICE_TYPE_CPU>(CpuMatrix& out_grad_mat,
         int64_t pad_size =
             std::min(starts[i] - begin, starts[i + 1] - starts[i]);
         if (is_padding && w_grad_mat) {
-          MatrixPtr mat = out_grad_mat.subMatrix(starts[i], pad_size);
+          MatrixPtr mat = const_cast<CpuMatrix&>(out_grad_mat)
+                              .subMatrix(starts[i], pad_size);
           MatrixPtr sub = w_grad_mat.subMatrix(j, pad_size);
           sub->addAtOffset(*mat, j * input_dim);
         }
@@ -157,8 +194,8 @@ void ContextProjectionBackward<DEVICE_TYPE_CPU>(CpuMatrix& out_grad_mat,
         int64_t pad_size =
             std::min(end - starts[i + 1], starts[i + 1] - starts[i]);
         if (is_padding && w_grad_mat) {
-          MatrixPtr mat =
-              out_grad_mat.subMatrix(starts[i + 1] - pad_size, pad_size);
+          MatrixPtr mat = const_cast<CpuMatrix&>(out_grad_mat)
+                              .subMatrix(starts[i + 1] - pad_size, pad_size);
           MatrixPtr sub = w_grad_mat.subMatrix(
               begin_pad + context_start + j - pad_size, pad_size);
           sub->addAtOffset(*mat, j * input_dim);
@@ -169,17 +206,22 @@ void ContextProjectionBackward<DEVICE_TYPE_CPU>(CpuMatrix& out_grad_mat,
       if (end <= begin) continue;
       if (!in_grad_mat) continue;
       MatrixPtr src = in_grad_mat.subMatrix(begin, end - begin);
-      MatrixPtr dst = out_grad_mat.subMatrix(dst_begin, dst_end - dst_begin);
+      MatrixPtr dst = const_cast<CpuMatrix&>(out_grad_mat)
+                          .subMatrix(dst_begin, dst_end - dst_begin);
       src->addAtOffset(*dst, j * input_dim);
     }
   }
 }
 
 /**
- * \param inputs[0] input grad.
- * \param inputs[1] weight grad.
- * \param inputs[2] input sequence.
- * \param outputs[0] output value.
+ * Context Projection Backward Function.
+ * Update the weight gradient and input layer gradient with backprop
+ *
+ * \param inputs[0].matrix          output layer grad, n * (d * l)
+ * \param inputs[0].vector          start position sequence, n * 1
+ * \param outputs[0].matrix         input layer grad, n * d
+ * \param outputs[0].vector         start position sequence, n * 1
+ * \param outputs[1]                weight grad, pad * d
  */
 template <DeviceType Device>
 class ContextProjectionBackwardFunc : public FunctionBase {
@@ -193,32 +235,36 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    CHECK_EQ((size_t)3, inputs.size());
-    CHECK_EQ((size_t)1, outputs.size());
+    CHECK_EQ((size_t)1, inputs.size());
+    CHECK_EQ((size_t)2, outputs.size());
+    CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg())
+        << "SequenceArg required here";
+    const auto in_seq = dynamic_cast<const SequenceArg&>(inputs[0]);
+    auto out_seq = dynamic_cast<const SequenceArg&>(outputs[0]);
+    CHECK(in_seq.data() && in_seq.getSequenceId().data());
+    CHECK_EQ(in_seq.shape().ndims(), (size_t)2);
+    CHECK_EQ(in_seq.getSequenceId().shape().ndims(), (size_t)1);
+    CHECK_EQ(out_seq.shape().ndims(), (size_t)2);
+    CHECK_EQ(out_seq.getSequenceId().shape().ndims(), (size_t)1);
+    CHECK_EQ(outputs[1].shape().ndims(), (size_t)2);
 
-    CHECK(outputs[0].data() && inputs[2].data());
-    CHECK_EQ(outputs[0].shape().ndims(), (size_t)2);
-    CHECK_EQ(inputs[0].shape().ndims(), (size_t)2);
-    CHECK_EQ(inputs[1].shape().ndims(), (size_t)2);
-    CHECK_EQ(inputs[2].shape().ndims(), (size_t)1);
+    /// dim of input grad == dim of weight
+    CHECK_EQ(out_seq.shape()[1], outputs[1].shape()[1]);
+    /// input and output grad has the same batch_size
+    CHECK_EQ(out_seq.shape()[0], in_seq.shape()[0]);
+    /// dim of output grad = dim of input grad * context_length
+    CHECK_EQ(in_seq.shape()[1], out_seq.shape()[1] * context_length_);
+    CHECK_EQ(out_seq.getArgType(), ADD_TO);
+    CHECK_EQ(outputs[1].getArgType(), ADD_TO);
 
-    /// dim of input == dim of weight
-    CHECK_EQ(inputs[0].shape()[1], inputs[1].shape()[1]);
-    /// input and output has the same batch_size
-    CHECK_EQ(inputs[0].shape()[0], outputs[0].shape()[0]);
-    /// dim of output = dim of input * context_length
-    CHECK_EQ(outputs[0].shape()[1], inputs[0].shape()[1] * context_length_);
-
-    CHECK_EQ(outputs[0].getArgType(), ADD_TO);
-
-    auto out_grad_mat = outputs[0].matrix<Device>();
+    const auto seq_vec = in_seq.getSequenceId().vector<int, Device>();
+    const auto out_grad_mat = in_seq.matrix<Device>();
     auto in_grad_mat =
-        !inputs[0].data() ? typename Tensor<real, Device>::Matrix(nullptr, 0, 0)
-                          : inputs[0].matrix<Device>();
-    auto w_grad_mat = !inputs[1].data()
+        !out_seq.data() ? typename Tensor<real, Device>::Matrix(nullptr, 0, 0)
+                        : out_seq.matrix<Device>();
+    auto w_grad_mat = !outputs[1].data()
                           ? typename Tensor<real, Device>::Matrix(nullptr, 0, 0)
-                          : inputs[1].matrix<Device>();
-    auto seq_vec = inputs[2].vector<int, Device>();
+                          : outputs[1].matrix<Device>();
     ContextProjectionBackward<Device>(out_grad_mat,
                                       in_grad_mat,
                                       w_grad_mat,
@@ -238,11 +284,16 @@ private:
   size_t total_pad_;
 };
 
-#if 0
 /**
- * \param inputs[0] input grad.
- * \param inputs[1] input sequence.
- * \param outputs[0] output grad.
+ * Context Projection Backward Data Function
+ * Update input layer grad
+ * input:  sequence of output layer grad
+ * output: sequence of input layer grad
+ *
+ * \param outputs[0].matrix              input layer grad, n * d
+ * \param outputs[0].vector              start position sequence, n * 1
+ * \param inputs[0].matrix               output layer grad, n * (d * l)
+ * \param inputs[0].vector               start positon sequence, n * 1
  */
 template <DeviceType Device>
 class ContextProjectionBackwardDataFunc : public FunctionBase {
@@ -252,32 +303,30 @@ public:
     context_start_ = config.get<int>("context_start");
   }
 
-  void calc(const Arguments& inputs,
-            const Arguments& outputs,
-            const Arguments& inouts) override {
-    CHECK_EQ(2, static_cast<int>(inputs.size()));
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(1, static_cast<int>(inputs.size()));
     CHECK_EQ(1, static_cast<int>(outputs.size()));
-    CHECK_EQ(0, static_cast<int>(inouts.size()));
-    CHECK(inputs[0].getData() && outputs[0].getData() && inputs[1].getData());
-    CHECK_EQ(static_cast<int>(outputs[0].dims_.size()), 2);
-    CHECK_EQ(static_cast<int>(inputs[0].dims_.size()), 2);
-    CHECK_EQ(static_cast<int>(inputs[1].dims_.size()), 1);
-    CHECK_EQ(outputs[0].dims_[1], inputs[0].dims_[1] * context_length_);
+    CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg())
+        << "SequenceArg required here";
+    const auto in_seq = dynamic_cast<const SequenceArg&>(inputs[0]);
+    const auto out_seq = dynamic_cast<const SequenceArg&>(outputs[0]);
+
+    CHECK(in_seq.data() && out_seq.data() && in_seq.getSequenceId().data());
+    CHECK_EQ(static_cast<int>(out_seq.shape().ndims()), 2);
+    CHECK_EQ(static_cast<int>(in_seq.shape().ndims()), 2);
+    CHECK_EQ(static_cast<int>(in_seq.getSequenceId().shape().ndims()), 1);
+    /// output layer grad dim == input layer grad dim * context_length_
+    CHECK_EQ(in_seq.shape().ndims(), out_seq.shape().ndims() * context_length_);
     /// input and output has the same batch_size
-    CHECK_EQ(inputs[0].dims_[0], outputs[0].dims_[0]);
+    CHECK_EQ(in_seq.shape()[0], out_seq.shape()[0]);
+    CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO);
 
-    auto out_grad_mat = std::make_shared<typename MatrixT<Device>::type>(
-        outputs[0].getData(), outputs[0].dims_[0], outputs[0].dims_[1]);
-    const auto in_grad_mat = std::make_shared<typename MatrixT<Device>::type>(
-        inputs[0].getData(), inputs[0].dims_[0], inputs[0].dims_[1]);
-    typename SequenceT<Device>::type seq_vec(
-        inputs[1].dims_[0], reinterpret_cast<int*>(inputs[1].getData()));
+    const auto out_grad_mat = in_seq.matrix<Device>();
+    const auto seq_vec = in_seq.getSequenceId().vector<int, Device>();
+    auto in_grad_mat = out_seq.matrix<Device>();
 
-    ContextProjectionBackwardData<Device>(out_grad_mat.get(),
-                                          in_grad_mat.get(),
-                                          seq_vec,
-                                          context_length_,
-                                          context_start_);
+    ContextProjectionBackwardData<Device>(
+        out_grad_mat, in_grad_mat, seq_vec, context_length_, context_start_);
   }
 
 private:
@@ -286,9 +335,14 @@ private:
 };
 
 /**
- * \param inputs[0] weight grad.
- * \param inputs[1] input sequence.
- * \param outputs[0] output grad.
+ * Context Projection Backward Weight Function
+ * Update weight grad by backprop
+ * input:  sequence of output layer grad
+ * output: weight grad
+ *
+ * \param outputs[0]                   weight grad, pad * d
+ * \param inputs[0].matrix             output layer grad, n * (d * l)
+ * \param inputs[0].vecotr             start positon sequence, n * 1
  */
 template <DeviceType Device>
 class ContextProjectionBackwardWeightFunc : public FunctionBase {
@@ -300,28 +354,25 @@ public:
     total_pad_ = config.get<size_t>("total_pad");
   }
 
-  void calc(const Arguments& inputs,
-            const Arguments& outputs,
-            const Arguments& inouts) override {
-    CHECK_EQ(2, static_cast<int>(inputs.size()));
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(1, static_cast<int>(inputs.size()));
     CHECK_EQ(1, static_cast<int>(outputs.size()));
-    CHECK_EQ(0, static_cast<int>(inouts.size()));
-
-    CHECK(inputs[0].getData() && outputs[0].getData() && inputs[1].getData());
-    CHECK_EQ(static_cast<int>(outputs[0].dims_.size()), 2);
-    CHECK_EQ(static_cast<int>(inputs[0].dims_.size()), 2);
-    CHECK_EQ(static_cast<int>(inputs[1].dims_.size()), 1);
-    CHECK_EQ(outputs[0].dims_[1], inputs[0].dims_[1] * context_length_);
-
-    auto out_grad_mat = std::make_shared<typename MatrixT<Device>::type>(
-        outputs[0].getData(), outputs[0].dims_[0], outputs[0].dims_[1]);
-    auto w_grad_mat = std::make_shared<typename MatrixT<Device>::type>(
-        inputs[0].getData(), inputs[0].dims_[0], inputs[0].dims_[1]);
-    typename SequenceT<Device>::type seq_vec(
-        inputs[1].dims_[0], reinterpret_cast<int*>(inputs[1].getData()));
+    CHECK(inputs[0].isSequenceArg()) << "SequenceArg required here";
+    const auto in_seq = dynamic_cast<const SequenceArg&>(inputs[0]);
+    CHECK(in_seq.data() && in_seq.getSequenceId().data() && outputs[0].data());
+    CHECK_EQ(static_cast<int>(outputs[0].shape().ndims()), 2);
+    CHECK_EQ(static_cast<int>(in_seq.shape().ndims()), 2);
+    CHECK_EQ(static_cast<int>(in_seq.getSequenceId().shape().ndims()), 1);
+    CHECK_EQ(in_seq.shape()[0], outputs[0].shape()[0]);
+    /// output layer grad dim == weight dim * context_length_
+    CHECK_EQ(in_seq.shape()[1], outputs[0].shape()[1] * context_length_);
+    CHECK_EQ(outputs[0].getArgType(), ADD_TO);
 
-    ContextProjectionBackwardWeight<Device>(out_grad_mat.get(),
-                                            w_grad_mat.get(),
+    const auto seq_vec = in_seq.getSequenceId().vector<int, Device>();
+    const auto out_grad_mat = in_seq.matrix<Device>();
+    auto w_grad_mat = outputs[0].matrix<Device>();
+    ContextProjectionBackwardWeight<Device>(out_grad_mat,
+                                            w_grad_mat,
                                             seq_vec,
                                             context_length_,
                                             context_start_,
@@ -335,7 +386,6 @@ private:
   size_t begin_pad_;
   size_t total_pad_;
 };
-#endif
 
 REGISTER_TYPED_FUNC(ContextProjectionForward,
                     CPU,
@@ -350,7 +400,6 @@ REGISTER_TYPED_FUNC(ContextProjectionForward,
 REGISTER_TYPED_FUNC(ContextProjectionBackward,
                     GPU,
                     ContextProjectionBackwardFunc);
-#if 0
 REGISTER_TYPED_FUNC(ContextProjectionBackwardData,
                     GPU,
                     ContextProjectionBackwardDataFunc);
@@ -358,5 +407,4 @@ REGISTER_TYPED_FUNC(ContextProjectionBackwardWeight,
                     GPU,
                     ContextProjectionBackwardWeightFunc);
 #endif
-#endif
 }  // namespace paddle

paddle/function/ContextProjectionOp.h

@@ -21,14 +21,14 @@ namespace paddle {
 /**
  * \brief   Context Projection Forward.
  *
- * \param[out]  outputs           output data.
- * \param[in]   input             input data.
- * \param[in]   weight            input weight.
- * \param[in]   sequence          input data.
- * \param[in]   context_length    consecutive rows for concatenation.
- * \param[in]   context_start     context start position.
- * \param[in]   begin_pad         begining pad position.
- * \param[in]   is_padding        whether padding 0 or not.
+ * \param[in/out]  outputs           output data.
+ * \param[in]      input             input data.
+ * \param[in]      weight            input weight.
+ * \param[in]      sequence          input data.
+ * \param[in]      context_length    consecutive rows for concatenation.
+ * \param[in]      context_start     context start position.
+ * \param[in]      begin_pad         begining pad position.
+ * \param[in]      is_padding        whether padding 0 or not.
  *
  */
 template <DeviceType DType>
@@ -56,7 +56,7 @@ void ContextProjectionForward(
  */
 template <DeviceType DType>
 void ContextProjectionBackward(
-    typename Tensor<real, DType>::Matrix& out_grad,
+    const typename Tensor<real, DType>::Matrix& out_grad,
     typename Tensor<real, DType>::Matrix& in_grad,
     typename Tensor<real, DType>::Matrix& w_grad,
     const typename Tensor<int, DType>::Vector& seq_vec,
@@ -68,7 +68,7 @@ void ContextProjectionBackward(
 
 template <DeviceType DType>
 void ContextProjectionBackwardData(
-    typename Tensor<real, DType>::Matrix& out_grad,
+    const typename Tensor<real, DType>::Matrix& out_grad,
     typename Tensor<real, DType>::Matrix& in_grad,
     const typename Tensor<int, DType>::Vector& sequence,
     size_t context_length,
@@ -76,7 +76,7 @@ void ContextProjectionBackwardData(
 
 template <DeviceType DType>
 void ContextProjectionBackwardWeight(
-    typename Tensor<real, DType>::Matrix& out_grad,
+    const typename Tensor<real, DType>::Matrix& out_grad,
     typename Tensor<real, DType>::Matrix& w_grad,
     const typename Tensor<int, DType>::Vector& seq_vec,
     size_t context_length,

paddle/function/ContextProjectionOpGpu.cu

@@ -138,10 +138,10 @@ void ContextProjectionForward<DEVICE_TYPE_GPU>(GpuMatrix& output,
                                 begin_pad);
 }
 
-__global__ void KeContextProjectionBackwardData(real* out_grad,
+__global__ void KeContextProjectionBackwardData(const real* out_grad,
                                                 const int* sequence,
                                                 real* in_grad,
-                                                int input_dim,
+                                                size_t input_dim,
                                                 int context_length,
                                                 int context_start) {
   int idx = threadIdx.x;
@@ -152,7 +152,8 @@ __global__ void KeContextProjectionBackwardData(real* out_grad,
   real value = 0;
 
   int instances = seq_end - seq_start + context_length - 1;
-  out_grad += seq_start * input_dim * context_length;
+  auto out = const_cast<real*>(out_grad);
+  out += seq_start * input_dim * context_length;
   in_grad += seq_start * input_dim;
   for (int k = 0; k <= input_dim / block_size; k++) {
     if (idx < input_dim) {
@@ -169,7 +170,7 @@ __global__ void KeContextProjectionBackwardData(real* out_grad,
         int outx = (i - context_length) < 0 ? i : (context_length - 1);
         int outy = (i - context_length) < 0 ? 0 : (i - (context_length - 1));
         real* output_r =
-          out_grad + outy * input_dim * context_length + outx * input_dim;
+          out + outy * input_dim * context_length + outx * input_dim;
         for (int j = outy; j < seq_end - seq_start; j++) {
           value += output_r[idx];
           if (j - outy == outx) break;
@@ -194,7 +195,7 @@ __global__ void KeContextProjectionBackwardData(real* out_grad,
  * @param[in]   context_start    context start.
  *
  */
-void hl_context_projection_backward_data(real* out_grad,
+void hl_context_projection_backward_data(const real* out_grad,
                                          const int* sequence,
                                          real* input_grad,
                                          size_t num_sequences,
@@ -216,7 +217,7 @@ void hl_context_projection_backward_data(real* out_grad,
 }
 
 template <>
-void ContextProjectionBackwardData<DEVICE_TYPE_GPU>(GpuMatrix& out_grad,
+void ContextProjectionBackwardData<DEVICE_TYPE_GPU>(const GpuMatrix& out_grad,
                                                     GpuMatrix& in_grad,
                                                     const GpuIVector& sequence,
                                                     size_t context_length,
@@ -231,7 +232,7 @@ void ContextProjectionBackwardData<DEVICE_TYPE_GPU>(GpuMatrix& out_grad,
 }
 
 template<int THREADS_X, int THREADS_Y>
-__global__ void KeContextProjectionBackwardWeight(real* out_grad,
+__global__ void KeContextProjectionBackwardWeight(const real* out_grad,
                                                   const int* sequence,
                                                   real* w_grad,
                                                   int num_sequences,
@@ -254,7 +255,8 @@ __global__ void KeContextProjectionBackwardWeight(real* out_grad,
     for (int seqId = idy; seqId < num_sequences; seqId += THREADS_Y) {
       int seq_start = sequence[seqId];
       int seq_end = sequence[seqId+1];
-      output_r = out_grad + seq_start * w_dim * context_length;
+      output_r = const_cast<real*>(out_grad)
+                    + seq_start * w_dim * context_length;
 
       if (context_start < 0) {
         if (padId + context_start < 0) {
@@ -318,7 +320,7 @@ __global__ void KeContextProjectionBackwardWeight(real* out_grad,
  * beginning.
  *
  */
-void hl_context_projection_backward_weight(real* out_grad,
+void hl_context_projection_backward_weight(const real* out_grad,
                                            const int* sequence,
                                            real* w_grad,
                                            size_t num_sequences,
@@ -346,7 +348,7 @@ void hl_context_projection_backward_weight(real* out_grad,
 
 template <>
 void ContextProjectionBackwardWeight<DEVICE_TYPE_GPU>(
-        GpuMatrix& out_grad,
+        const GpuMatrix& out_grad,
         GpuMatrix& w_grad,
         const GpuIVector& seq_vec,
         size_t context_length,
@@ -365,7 +367,7 @@ void ContextProjectionBackwardWeight<DEVICE_TYPE_GPU>(
 }
 
 template <>
-void ContextProjectionBackward<DEVICE_TYPE_GPU>(GpuMatrix& out_grad,
+void ContextProjectionBackward<DEVICE_TYPE_GPU>(const GpuMatrix& out_grad,
                                                 GpuMatrix& in_grad,
                                                 GpuMatrix& w_grad,
                                                 const GpuIVector& sequence,

paddle/function/ContextProjectionOpTest.cpp

@@ -56,22 +56,25 @@ void testMatrixProjectionForward(int context_start,
   cpu_out.randomizeUniform();
   gpu_out.copyFrom(cpu_out);
 
-  compare.getCpuFunction()->calc(
-      {Tensor(cpu_in.getData(), Dims{batch_size, input_dim}),
-       Tensor(cpu_weight ? cpu_weight->getData() : nullptr,
-              Dims{pad, input_dim}),
-       Tensor(reinterpret_cast<real*>(cpu_seq->getData()),
-              Dims{cpu_seq->getSize()})},
-      {Tensor(cpu_out.getData(), Dims{batch_size, input_dim * context_length})},
-      {});
-  compare.getGpuFunction()->calc(
-      {Tensor(gpu_in.getData(), Dims{batch_size, input_dim}),
-       Tensor(gpu_weight ? gpu_weight->getData() : nullptr,
-              Dims{pad, input_dim}),
-       Tensor(reinterpret_cast<real*>(gpu_seq->getData()),
-              Dims{gpu_seq->getSize()})},
-      {Tensor(gpu_out.getData(), Dims{batch_size, input_dim * context_length})},
-      {});
+  BufferArgs cpu_inputs;
+  BufferArgs cpu_outputs;
+  cpu_inputs.addArg(cpu_in, *cpu_seq);
+  if (cpu_weight) {
+    cpu_inputs.addArg(*cpu_weight, *cpu_seq);
+  }
+  cpu_outputs.addArg(cpu_out, *cpu_seq, ADD_TO);
+
+  compare.getCpuFunction()->calc(cpu_inputs, cpu_outputs);
+
+  BufferArgs gpu_inputs;
+  BufferArgs gpu_outputs;
+  gpu_inputs.addArg(gpu_in, *gpu_seq);
+  if (gpu_weight) {
+    gpu_inputs.addArg(*gpu_weight, *gpu_seq);
+  }
+  gpu_outputs.addArg(gpu_out, *gpu_seq, ADD_TO);
+
+  compare.getGpuFunction()->calc(gpu_inputs, gpu_outputs);
 
   autotest::TensorCheckEqual(cpu_out, gpu_out);
 }
@@ -117,25 +120,23 @@ void testMatrixProjectionBackward(int context_start,
     gpu_w_grad->copyFrom(*cpu_w_grad);
   }
 
-  compare.getCpuFunction()->calc(
-      {Tensor(cpu_in_grad.getData(), Dims{batch_size, input_dim}),
-       Tensor(cpu_w_grad ? cpu_w_grad->getData() : nullptr,
-              Dims{pad, input_dim}),
-       Tensor(reinterpret_cast<real*>(cpu_seq->getData()),
-              Dims{cpu_seq->getSize()})},
-      {Tensor(cpu_out_grad.getData(),
-              Dims{batch_size, input_dim * context_length})},
-      {});
-
-  compare.getGpuFunction()->calc(
-      {Tensor(gpu_in_grad.getData(), Dims{batch_size, input_dim}),
-       Tensor(gpu_w_grad ? gpu_w_grad->getData() : nullptr,
-              Dims{pad, input_dim}),
-       Tensor(reinterpret_cast<real*>(gpu_seq->getData()),
-              Dims{gpu_seq->getSize()})},
-      {Tensor(gpu_out_grad.getData(),
-              Dims{batch_size, input_dim * context_length})},
-      {});
+  BufferArgs cpu_inputs;
+  BufferArgs cpu_outputs;
+  cpu_inputs.addArg(cpu_out_grad, *cpu_seq);
+  cpu_outputs.addArg(cpu_in_grad, *cpu_seq, ADD_TO);
+  cpu_outputs.addArg(
+      cpu_w_grad ? *cpu_w_grad : CpuMatrix(nullptr, 0, input_dim), ADD_TO);
+
+  compare.getCpuFunction()->calc(cpu_inputs, cpu_outputs);
+
+  BufferArgs gpu_inputs;
+  BufferArgs gpu_outputs;
+  gpu_inputs.addArg(gpu_out_grad, *gpu_seq);
+  gpu_outputs.addArg(gpu_in_grad, *gpu_seq, ADD_TO);
+  gpu_outputs.addArg(
+      gpu_w_grad ? *gpu_w_grad : GpuMatrix(nullptr, 0, input_dim), ADD_TO);
+
+  compare.getGpuFunction()->calc(gpu_inputs, gpu_outputs);
 
   autotest::TensorCheckErr(cpu_in_grad, gpu_in_grad);
   if (is_padding) {

paddle/function/CrossMapNormalOp.cpp

@@ -188,8 +188,13 @@ public:
     CHECK(inputs[0].shape() == inputs[3].shape());
     CHECK(inputs[0].shape() == outputs[0].shape());
 
-    // TODO(hedaoyuan): need support ASSIGN_TO mode.
-    CHECK_EQ(outputs[0].getArgType(), ADD_TO);
+    if (outputs[0].getArgType() != ADD_TO) {
+      // Currently, some algorithm implementations are ASSIGN_TO mode,
+      // if need to support the ADD_TO calculation, need to clear the output.
+      typename Tensor<real, Device>::Vector tmp(
+          outputs[0].shape().getElements(), outputs[0].data<real>());
+      tmp.zero();
+    }
 
     size_t samples = inputs[0].shape()[0];
     size_t channels = inputs[0].shape()[1];

paddle/function/CrossMapNormalOpTest.cpp

@@ -27,15 +27,19 @@ TEST(CrossMapNormal, real) {
                     << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW
                     << " size=" << size;
 
-            FunctionCompare compare("CrossMapNormal",
-                                    FuncConfig()
-                                        .set("size", size)
-                                        .set("scale", (real)1.5)
-                                        .set("pow", (real)0.5));
-            Dims dims{numSamples, channels, imgSizeH, imgSizeW};
-            compare.cmpWithArg({Tensor(nullptr, dims)},
-                               {Tensor(nullptr, dims), Tensor(nullptr, dims)},
-                               {});
+            // init Test object
+            FunctionCompare test("CrossMapNormal",
+                                 FuncConfig()
+                                     .set("size", size)
+                                     .set("scale", (real)1.5)
+                                     .set("pow", (real)0.5));
+            // prepare input arguments
+            TensorShape shape{numSamples, channels, imgSizeH, imgSizeW};
+            test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
+            test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, shape));
+            test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, shape));
+            // run Function
+            test.run();
           }
         }
       }
@@ -53,18 +57,19 @@ TEST(CrossMapNormalGrad, real) {
                     << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW
                     << " size=" << size;
 
-            FunctionCompare compare("CrossMapNormalGrad",
-                                    FuncConfig()
-                                        .set("size", size)
-                                        .set("scale", (real)1.5)
-                                        .set("pow", (real)0.5));
-            Dims dims{numSamples, channels, imgSizeH, imgSizeW};
-            compare.cmpWithArg({Tensor(nullptr, dims),
-                                Tensor(nullptr, dims),
-                                Tensor(nullptr, dims),
-                                Tensor(nullptr, dims)},
-                               {Tensor(nullptr, dims)},
-                               {});
+            FunctionCompare test("CrossMapNormalGrad",
+                                 FuncConfig()
+                                     .set("size", size)
+                                     .set("scale", (real)1.5)
+                                     .set("pow", (real)0.5));
+            TensorShape shape{numSamples, channels, imgSizeH, imgSizeW};
+            test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
+            test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
+            test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
+            test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
+            test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, shape));
+            // run Function
+            test.run();
           }
         }
       }

paddle/function/Function.cpp

@@ -79,15 +79,25 @@ FuncConfig& FuncConfig::set<bool>(const std::string& key, bool v) {
 void BufferArgs::addArg(const Matrix& arg,
                         const TensorShape& shape,
                         ArgType argType) {
-  args_.push_back(std::make_shared<BufferArg>(arg, shape, argType));
+  _args_.push_back(new BufferArg(arg, shape, argType));
+  addArg(*_args_.back());
 }
 
 void BufferArgs::addArg(const CpuSparseMatrix& arg, ArgType argType) {
-  args_.push_back(std::make_shared<SparseMatrixArg>(arg, argType));
+  _args_.push_back(new SparseMatrixArg(arg, argType));
+  addArg(*_args_.back());
 }
 
 void BufferArgs::addArg(const GpuSparseMatrix& arg, ArgType argType) {
-  args_.push_back(std::make_shared<SparseMatrixArg>(arg, argType));
+  _args_.push_back(new SparseMatrixArg(arg, argType));
+  addArg(*_args_.back());
+}
+
+void BufferArgs::addArg(const Matrix& matrix,
+                        const IVector& vector,
+                        ArgType argType) {
+  _args_.push_back(new SequenceArg(matrix, vector, argType));
+  addArg(*_args_.back());
 }
 
 ClassRegistrar<FunctionBase> FunctionBase::funcRegistrar_;

paddle/function/Function.h

@@ -50,19 +50,44 @@ protected:
  * Argument type for Function::calc().
  * A BufferArgs contains a set of BufferArg,
  * because Function can have multiple inputs and outputs.
+ *
+ * addArg() with Matix object used to adapt Layer Argument.
+ * Will create a BufferArg object in addArg(),
+ * and free in destructor of BufferArgs.
+ *
+ * addArg() with BufferArg object, just save BufferArg object address,
+ * and the caller needs to guarantee the validity of the BufferArg object
+ * in the BufferArgs life time.
  */
 class BufferArgs {
 public:
   BufferArgs() {}
+
+  ~BufferArgs() {
+    for (auto arg : _args_) {
+      delete arg;
+    }
+  }
+
   size_t size() const { return args_.size(); }
 
   // add argument into BufferArgs
   // Tensor can be Matrix, Vector, IVector.
   // For inputs, do not need argType.
   // For outputs, the argType needs to be specified as ASSIGN_TO or ADD_TO.
-  template <typename Tensor>
-  void addArg(const Tensor& arg, ArgType argType = UNSPECIFIED) {
-    args_.push_back(std::make_shared<BufferArg>(arg, argType));
+  void addArg(const Matrix& arg, ArgType argType = UNSPECIFIED) {
+    _args_.push_back(new BufferArg(arg, argType));
+    addArg(*_args_.back());
+  }
+
+  void addArg(const Vector& arg, ArgType argType = UNSPECIFIED) {
+    _args_.push_back(new BufferArg(arg, argType));
+    addArg(*_args_.back());
+  }
+
+  void addArg(const IVector& arg, ArgType argType = UNSPECIFIED) {
+    _args_.push_back(new BufferArg(arg, argType));
+    addArg(*_args_.back());
   }
 
   // Add arg into BufferArgs and reshape the arg.
@@ -77,20 +102,37 @@ public:
   void addArg(const CpuSparseMatrix& arg, ArgType argType = UNSPECIFIED);
   void addArg(const GpuSparseMatrix& arg, ArgType argType = UNSPECIFIED);
 
+  void addArg(const Matrix& matrix,
+              const IVector& vector,
+              ArgType argType = UNSPECIFIED);
+
   // get argument
   const BufferArg& operator[](size_t num) const {
     CHECK_LT(num, args_.size());
     return *args_[num];
   }
 
+  void addArg(BufferArg& arg) { args_.push_back(&arg); }
+
+  void addArg(SequenceIdArg& arg) { args_.push_back(&arg); }
+
+  void addArg(SequenceArg& arg) { args_.push_back(&arg); }
+
+  void addArg(SparseMatrixArg& arg) { args_.push_back(&arg); }
+
 private:
-  std::vector<BufferArgPtr> args_;
+  std::vector<BufferArg*> args_;
+  // The BufferArg object is constructed and freed by BufferArgs.
+  std::vector<BufferArg*> _args_;
 };
 
 /**
  * \brief Base class for Function.
  * The basic Function implementation requires override init and calc interfaces.
  *
+ * The caller needs to ensure the validity of the arguments
+ * during Function execution.
+ *
  * Function inputs are readonly, Function outputs have two modes: ASSIGN_TO
  * and ADD_TO.
  * If output.getArgType() == ASSIGN_TO, this is assign mode, and the calculation

paddle/function/FunctionTest.cpp

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "Function.h"
 #include <gtest/gtest.h>
+#include "paddle/math/SparseMatrix.h"
 
 namespace paddle {
 
@@ -56,4 +57,110 @@ TEST(Function, BufferArgs) {
   Function<DEVICE_TYPE_GPU>(gpuArgments);
 }
 
+/**
+ * Some tests case are used to check the consistency between the BufferArg type
+ * argument received by Function and the original type argument.
+ *
+ * Use Case:
+ *  TEST() {
+ *    Matrix matrix(...);
+ *    CheckBufferArg lambda = [=](const BufferArg& arg) {
+ *      // check matrix and arg are equivalent
+ *      EXPECT_EQ(matrix, arg);
+ *    }
+ *
+ *   BufferArgs argments{matrix...};
+ *   std::vector<CheckBufferArg> checkFunc{lambda...};
+ *   testBufferArgs(argments, checkFunc);
+ *  }
+ */
+typedef std::function<void(const BufferArg&)> CheckBufferArg;
+
+void testBufferArgs(const BufferArgs& inputs,
+                    const std::vector<CheckBufferArg>& check) {
+  EXPECT_EQ(inputs.size(), check.size());
+  for (size_t i = 0; i < inputs.size(); i++) {
+    check[i](inputs[i]);
+  }
+}
+
+void testBufferArgs(const BufferArgs& inputs, const CheckBufferArg& check) {
+  EXPECT_EQ(inputs.size(), 1);
+  check(inputs[0]);
+}
+
+TEST(Arguments, Matrix) {
+  MatrixPtr matrix = Matrix::create(100, 200);
+  CheckBufferArg check = [=](const BufferArg& arg) {
+    EXPECT_EQ(arg.shape().ndims(), 2);
+    EXPECT_EQ(arg.shape()[0], 100);
+    EXPECT_EQ(arg.shape()[1], 200);
+    EXPECT_EQ(arg.data(), matrix->getData());
+
+    EXPECT_EQ(arg.matrix<DEVICE_TYPE_CPU>().getHeight(), matrix->getHeight());
+    EXPECT_EQ(arg.matrix<DEVICE_TYPE_CPU>().getWidth(), matrix->getWidth());
+    EXPECT_EQ(arg.matrix<DEVICE_TYPE_CPU>().getData(), matrix->getData());
+  };
+
+  BufferArgs argments;
+  argments.addArg(*matrix);
+  std::vector<CheckBufferArg> checkFunc;
+  checkFunc.push_back(check);
+  testBufferArgs(argments, checkFunc);
+}
+
+TEST(Arguments, Vector) {
+  VectorPtr vector = Vector::create(100, false);
+  CheckBufferArg check = [=](const BufferArg& arg) {
+    EXPECT_EQ(arg.shape().ndims(), 1);
+    EXPECT_EQ(arg.shape()[0], 100);
+    EXPECT_EQ(arg.data(), vector->getData());
+
+    CpuVector inVector = arg.vector<real, DEVICE_TYPE_CPU>();
+    EXPECT_EQ(inVector.getSize(), vector->getSize());
+    EXPECT_EQ(inVector.getData(), vector->getData());
+  };
+
+  BufferArgs argments;
+  argments.addArg(*vector);
+  std::vector<CheckBufferArg> checkFunc;
+  checkFunc.push_back(check);
+  testBufferArgs(argments, checkFunc);
+}
+
+TEST(Arguments, CpuSparseMatrix) {
+  CpuSparseMatrix sparse(200, 300, 50);
+  CheckBufferArg check = [=](const BufferArg& arg) {
+    EXPECT_EQ(arg.shape().ndims(), 2);
+    EXPECT_EQ(arg.shape()[0], 200);
+    EXPECT_EQ(arg.shape()[1], 300);
+    EXPECT_EQ(arg.data(), sparse.getData());
+    // CHECK_EQ(arg.sparse().nnz(), 50);
+    // CHECK_EQ(arg.sparse().dataFormat(), SPARSE_CSR_FORMAT);
+    // CHECK_EQ(arg.sparse().dataType(), SPARSE_FLOAT_VALUE);
+    EXPECT_EQ(arg.sparse().getRowBuf(), sparse.getRows());
+    EXPECT_EQ(arg.sparse().getColBuf(), sparse.getCols());
+  };
+
+  BufferArgs argments;
+  argments.addArg(sparse);
+  std::vector<CheckBufferArg> checkFunc;
+  checkFunc.push_back(check);
+  testBufferArgs(argments, checkFunc);
+}
+
+TEST(Arguments, BufferArg) {
+  BufferArg arg(nullptr, VALUE_TYPE_FLOAT, {1, 2, 3});
+  CheckBufferArg check = [=](const BufferArg& arg) {
+    EXPECT_EQ(arg.shape().ndims(), 3);
+    EXPECT_EQ(arg.shape()[0], 1);
+    EXPECT_EQ(arg.shape()[1], 2);
+    EXPECT_EQ(arg.shape()[2], 3);
+  };
+
+  BufferArgs argments;
+  argments.addArg(arg);
+  testBufferArgs(argments, check);
+}
+
 }  // namespace paddle

paddle/function/FunctionTest.h

@@ -15,95 +15,186 @@ limitations under the License. */
 #include "Function.h"
 #include "paddle/math/Vector.h"
 #include "paddle/math/tests/TensorCheck.h"
+#include "paddle/testing/TestUtil.h"
 
 namespace paddle {
 
+typedef std::shared_ptr<BufferArg> BufferArgPtr;
+
+/**
+ * \brief A class for comparing CPU and GPU implementations of Function.
+ *
+ *
+ * Use case:
+ *  // Initializes a test object, the corresponding cpu and gpu Function
+ *  // are constructed according to FunctionName and FuncConfig.
+ *  FunctionCompare test(FunctionName, FuncConfig);
+ *  // Prepare inputs and outputs arguments.
+ *  // Here the input and output can not contain real data,
+ *  // only contains the argument type and shape.
+ *  test.addInputs(input1);
+ *  test.addInputs(input2);
+ *  test.addOutputs(output1);
+ *  test.addOutputs(output2);
+ *  // Run.
+ *  // Will according to the type and shape of arguments(inputs_/outputs_),
+ *  // automatic initialization cpu and gpu function required arguments
+ *  // (cpuInputs_/cpuOutputs_/gpuInputs_/gpuOutputs_).
+ *  // Call the CPU and GPU Function calculation results.
+ *  // Compares CPU and GPU calculation results for consistency.
+ *  test.run();
+ */
 class FunctionCompare {
 public:
   FunctionCompare(const std::string& name, const FuncConfig& config)
-      : cpu(FunctionBase::funcRegistrar_.createByType(name + "-CPU")),
-        gpu(FunctionBase::funcRegistrar_.createByType(name + "-GPU")) {
-    cpu->init(config);
-    gpu->init(config);
+      : cpuFunc_(FunctionBase::funcRegistrar_.createByType(name + "-CPU")),
+        gpuFunc_(FunctionBase::funcRegistrar_.createByType(name + "-GPU")) {
+    cpuFunc_->init(config);
+    gpuFunc_->init(config);
+  }
+
+  ~FunctionCompare() {}
+
+  // input need only contains shape, do not contains data.
+  void addInputs(const BufferArg& input) {
+    size_t size =
+        input.shape().getElements() * sizeOfValuType(input.valueType());
+    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
+    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));
+
+    cpuInputs_.emplace_back(std::make_shared<BufferArg>(
+        cpuMemory_.back()->getBuf(), input.valueType(), input.shape()));
+    gpuInputs_.emplace_back(std::make_shared<BufferArg>(
+        gpuMemory_.back()->getBuf(), input.valueType(), input.shape()));
+  }
+
+  // output need only contains shape, do not contains data.
+  void addOutputs(const BufferArg& output) {
+    size_t size =
+        output.shape().getElements() * sizeOfValuType(output.valueType());
+    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
+    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));
+
+    cpuOutputs_.emplace_back(
+        std::make_shared<BufferArg>(cpuMemory_.back()->getBuf(),
+                                    output.valueType(),
+                                    output.shape(),
+                                    ASSIGN_TO));
+    gpuOutputs_.emplace_back(
+        std::make_shared<BufferArg>(gpuMemory_.back()->getBuf(),
+                                    output.valueType(),
+                                    output.shape(),
+                                    ASSIGN_TO));
   }
 
-  void cmpWithArg(const Arguments& inputs,
-                  const Arguments& outputs,
-                  const Arguments& inouts) {
-    // init cpu and gpu arguments
-    auto initArgs = [=](
-        Arguments& cpuArgs, Arguments& gpuArgs, const Arguments& inArgs) {
-      for (const auto arg : inArgs) {
-        size_t size = sizeof(real);
-        for (const auto dim : arg.dims_) {
-          size *= dim;
-        }
-        if (arg.getData()) {
-          // todo(tianbing), waste unnecessary mem here
-          cpuMemory.emplace_back(std::make_shared<CpuMemoryHandle>(size));
-          gpuMemory.emplace_back(std::make_shared<GpuMemoryHandle>(size));
-          cpuArgs.emplace_back(Tensor((real*)arg.getData(), arg.dims_));
-          gpuArgs.emplace_back(Tensor((real*)arg.getData(), arg.dims_));
-          // already init outside
-        } else {
-          cpuMemory.emplace_back(std::make_shared<CpuMemoryHandle>(size));
-          gpuMemory.emplace_back(std::make_shared<GpuMemoryHandle>(size));
-          cpuArgs.emplace_back(
-              Tensor((real*)cpuMemory.back()->getBuf(), arg.dims_));
-          gpuArgs.emplace_back(
-              Tensor((real*)gpuMemory.back()->getBuf(), arg.dims_));
-          // will use an api to refactor this code.
-          CpuVector cpuVector(size / sizeof(real),
-                              (real*)cpuArgs.back().getData());
-          GpuVector gpuVector(size / sizeof(real),
-                              (real*)gpuArgs.back().getData());
-          cpuVector.uniform(0.001, 1);
-          gpuVector.copyFrom(cpuVector);
-        }
+  void addInputs(const SequenceArg& input) {
+    size_t batchSize = input.shape()[0];
+    size_t numSeqs = batchSize / 10 + 1;
+
+    size_t sizeId = (numSeqs + 1) * sizeOfValuType(VALUE_TYPE_INT32);
+    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(sizeId));
+    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(sizeId));
+
+    TensorShape seqsId({numSeqs + 1});
+    // void* cpuBuffer = cpuMemory_.back()->getBuf();
+    // void* gpuBuffer = gpuMemory_.back()->getBuf();
+
+    size_t size =
+        input.shape().getElements() * sizeOfValuType(input.valueType());
+    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
+    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));
+
+    // TODO: need be implemented.
+  }
+
+  void run() {
+    // prepare cpu/gpu arguments
+    initInputs();
+
+    // function calculate
+    auto callFunction = [](FunctionBase* function,
+                           std::vector<BufferArgPtr>& inputs,
+                           std::vector<BufferArgPtr>& outputs) {
+      BufferArgs inArgs;
+      BufferArgs outArgs;
+      for (auto arg : inputs) {
+        inArgs.addArg(*arg);
       }
+      for (auto arg : outputs) {
+        outArgs.addArg(*arg);
+      }
+      function->calc(inArgs, outArgs);
     };
-    initArgs(cpuInputs, gpuInputs, inputs);
-    initArgs(cpuOutputs, gpuOutputs, outputs);
-    initArgs(cpuInouts, gpuInouts, inouts);
 
-    // function calculate
-    cpu->calc(cpuInputs, cpuOutputs, cpuInouts);
-    gpu->calc(gpuInputs, gpuOutputs, gpuInouts);
+    callFunction(cpuFunc_.get(), cpuInputs_, cpuOutputs_);
+    callFunction(gpuFunc_.get(), gpuInputs_, gpuOutputs_);
 
     // check outputs and inouts
-    auto checkArgs = [=](const Arguments& cpuArgs, const Arguments& gpuArgs) {
-      for (size_t i = 0; i < cpuArgs.size(); i++) {
-        auto cpu = cpuArgs[i];
-        auto gpu = gpuArgs[i];
-        size_t size = 1;
-        for (auto dim : cpu.dims_) {
-          size *= dim;
-        }
-        CpuVector cpuVector(size, (real*)cpu.getData());
-        GpuVector gpuVector(size, (real*)gpu.getData());
-
-        autotest::TensorCheckErr(cpuVector, gpuVector);
-      }
-    };
-    checkArgs(cpuOutputs, gpuOutputs);
-    checkArgs(cpuInouts, gpuInouts);
+    compareOutputs();
+  }
+
+  std::shared_ptr<FunctionBase> getCpuFunction() const { return cpuFunc_; }
+
+  std::shared_ptr<FunctionBase> getGpuFunction() const { return gpuFunc_; }
+
+protected:
+  void initInputs() {
+    for (size_t i = 0; i < cpuInputs_.size(); i++) {
+      initArg(*cpuInputs_[i]);
+
+      // TODO: Need a BufferCopy used to copy from one BufferArg to another.
+      CpuVector cpuVector(cpuInputs_[i]->shape().getElements(),
+                          (real*)cpuInputs_[i]->data());
+      GpuVector gpuVector(gpuInputs_[i]->shape().getElements(),
+                          (real*)gpuInputs_[i]->data());
+
+      gpuVector.copyFrom(cpuVector);
+    }
+  }
+
+  void compareOutputs() {
+    for (size_t i = 0; i < cpuOutputs_.size(); i++) {
+      // TODO, Need a BufferCheck used to compare the two buffers.
+      auto cpu = cpuOutputs_[i];
+      auto gpu = gpuOutputs_[i];
+      CpuVector cpuVector(cpu->shape().getElements(), (real*)cpu->data());
+      GpuVector gpuVector(cpu->shape().getElements(), (real*)gpu->data());
+
+      autotest::TensorCheckErr(cpuVector, gpuVector);
+    }
   }
 
-  std::shared_ptr<FunctionBase> getCpuFunction() const { return cpu; }
+  // only init cpu argument, gpu argument copy from cpu argument.
+  void initArg(BufferArg& arg) {
+    CpuVector vector(arg.shape().getElements(), (real*)arg.data());
+    vector.uniform(0.001, 1);
+  }
 
-  std::shared_ptr<FunctionBase> getGpuFunction() const { return gpu; }
+  void initArg(SequenceIdArg& arg, size_t batchSize) {
+    size_t numSeqs = arg.numSeqs();
+    int* buf = reinterpret_cast<int*>(arg.data());
+    int pos = 0;
+    size_t maxLen = 2 * batchSize / numSeqs;
+    for (int i = 0; i < (int)numSeqs; ++i) {
+      int len = uniformRandom(
+                    std::min<int64_t>(maxLen, batchSize - pos - numSeqs + i)) +
+                1;
+      buf[i] = pos;
+      pos += len;
+      VLOG(1) << " len=" << len;
+    }
+    buf[numSeqs] = batchSize;
+  }
 
 protected:
-  std::shared_ptr<FunctionBase> cpu;
-  std::shared_ptr<FunctionBase> gpu;
-  std::vector<CpuMemHandlePtr> cpuMemory;
-  std::vector<GpuMemHandlePtr> gpuMemory;
-  Arguments cpuInputs;
-  Arguments cpuOutputs;
-  Arguments cpuInouts;
-  Arguments gpuInputs;
-  Arguments gpuOutputs;
-  Arguments gpuInouts;
+  std::shared_ptr<FunctionBase> cpuFunc_;
+  std::shared_ptr<FunctionBase> gpuFunc_;
+  std::vector<CpuMemHandlePtr> cpuMemory_;
+  std::vector<GpuMemHandlePtr> gpuMemory_;
+  std::vector<BufferArgPtr> cpuInputs_;
+  std::vector<BufferArgPtr> cpuOutputs_;
+  std::vector<BufferArgPtr> gpuInputs_;
+  std::vector<BufferArgPtr> gpuOutputs_;
 };
 
 }  // namespace paddle

paddle/function/PadOp.cpp

@@ -89,20 +89,21 @@ public:
    * \param inputs[0] input value.
    * \param outputs[0] output value.
    */
-  void calc(const Arguments& inputs,
-            const Arguments& outputs,
-            const Arguments& inouts) override {
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
     CHECK_EQ(1UL, inputs.size());
     CHECK_EQ(1UL, outputs.size());
-    CHECK_EQ(0UL, inouts.size());
-
-    size_t num = inputs[0].dims_[0];
-    size_t inC = inputs[0].dims_[1];
-    size_t inH = inputs[0].dims_[2];
-    size_t inW = inputs[0].dims_[3];
-
-    Pad<Device>(outputs[0].getData(),
-                inputs[0].getData(),
+    CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO);
+
+    size_t num = inputs[0].shape()[0];
+    size_t inC = inputs[0].shape()[1];
+    size_t inH = inputs[0].shape()[2];
+    size_t inW = inputs[0].shape()[3];
+    typename Tensor<real, Device>::Vector vec(outputs[0].shape().getElements(),
+                                              outputs[0].data<real>());
+    vec.zero();
+
+    Pad<Device>(outputs[0].data<real>(),
+                inputs[0].data<real>(),
                 num,
                 inC,
                 inH,
@@ -140,21 +141,25 @@ public:
    * \param inputs[0] output grad.
    * \param inouts[0] input grad.
    */
-  void calc(const Arguments& inputs,
-            const Arguments& outputs,
-            const Arguments& inouts) override {
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
     CHECK_EQ(1UL, inputs.size());
-    CHECK_EQ(0UL, outputs.size());
-    CHECK_EQ(1UL, inouts.size());
+    CHECK_EQ(1UL, outputs.size());
 
-    size_t n = inouts[0].dims_[0];
-    size_t inC = inouts[0].dims_[1];
-    size_t inH = inouts[0].dims_[2];
-    size_t inW = inouts[0].dims_[3];
+    size_t num = outputs[0].shape()[0];
+    size_t inC = outputs[0].shape()[1];
+    size_t inH = outputs[0].shape()[2];
+    size_t inW = outputs[0].shape()[3];
+
+    if (outputs[0].getArgType() != ADD_TO) {
+      // for unit test
+      typename Tensor<real, Device>::Vector tmp(
+          outputs[0].shape().getElements(), outputs[0].data<real>());
+      tmp.zero();
+    }
 
-    PadGrad<Device>(inouts[0].getData(),
-                    inputs[0].getData(),
-                    n,
+    PadGrad<Device>(outputs[0].data<real>(),
+                    inputs[0].data<real>(),
+                    num,
                     inC,
                     inH,
                     inW,

paddle/function/PadOpTest.cpp

@@ -33,10 +33,12 @@ TEST(Pad, real) {
                                       .set("padh1", 2)
                                       .set("padw0", 3)
                                       .set("padw1", 2));
-          Dims inDims{numSamples, channels, imgSizeH, imgSizeW};
-          Dims outDims{numSamples, channels + 5, imgSizeH + 3, imgSizeW + 5};
-          compare.cmpWithArg(
-              {Tensor(nullptr, inDims)}, {Tensor(nullptr, outDims)}, {});
+          TensorShape inDims{numSamples, channels, imgSizeH, imgSizeW};
+          TensorShape outDims{
+              numSamples, channels + 5, imgSizeH + 3, imgSizeW + 5};
+          compare.addInputs(BufferArg(VALUE_TYPE_FLOAT, inDims));
+          compare.addOutputs(BufferArg(VALUE_TYPE_FLOAT, outDims, ASSIGN_TO));
+          compare.run();
         }
       }
     }
@@ -50,7 +52,6 @@ TEST(PadGrad, real) {
         for (size_t imgSizeW : {5, 32, 96}) {
           VLOG(3) << " numSamples=" << numSamples << " channels=" << channels
                   << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW;
-
           FunctionCompare compare("PadGrad",
                                   FuncConfig()
                                       .set("padc0", 2)
@@ -59,10 +60,12 @@ TEST(PadGrad, real) {
                                       .set("padh1", 2)
                                       .set("padw0", 3)
                                       .set("padw1", 2));
-          Dims inDims{numSamples, channels, imgSizeH, imgSizeW};
-          Dims outDims{numSamples, channels + 5, imgSizeH + 3, imgSizeW + 5};
-          compare.cmpWithArg(
-              {Tensor(nullptr, outDims)}, {}, {Tensor(nullptr, inDims)});
+          TensorShape inDims{numSamples, channels, imgSizeH, imgSizeW};
+          TensorShape outDims{
+              numSamples, channels + 5, imgSizeH + 3, imgSizeW + 5};
+          compare.addInputs(BufferArg(VALUE_TYPE_FLOAT, outDims));
+          compare.addOutputs(BufferArg(VALUE_TYPE_FLOAT, inDims, ASSIGN_TO));
+          compare.run();
         }
       }
     }

paddle/function/TensorShape.h

@@ -55,6 +55,15 @@ public:
     numElements();
   }
 
+  void reshape(std::initializer_list<size_t> dims) {
+    ndims_ = dims.size();
+    if (ndims_ > kMinDims) {
+      dims_.resize(ndims_);
+    }
+    dims_.assign(dims);
+    numElements();
+  }
+
   // number of dimensions of the tensor
   size_t ndims() const { return ndims_; }
 
@@ -82,7 +91,7 @@ private:
 
   // init dims_
   void initDims(size_t ndims) {
-    size_t count = ndims < 4 ? 4 : ndims;
+    size_t count = ndims < kMinDims ? kMinDims : ndims;
     dims_.assign(count, 1);
   }
 
@@ -92,6 +101,7 @@ private:
   // number of elements
   size_t nelements_;
   std::vector<size_t> dims_;
+  static const size_t kMinDims = 4;
 };
 
 }  // namespace paddle

paddle/gserver/layers/ContextProjection.cpp

@@ -118,16 +118,15 @@ void ContextProjection::forward() {
   /// first use state_, otherwise use weight_(padding false === w nullptr)
   auto w_ptr =
       state_ ? state_.get() : is_padding ? weight_->getW().get() : nullptr;
-  auto start_pos = in_->sequenceStartPositions;
-
+  const auto start_pos = in_->sequenceStartPositions->getVector(useGpu_);
   BufferArgs inputs;
   BufferArgs outputs;
-  inputs.addArg(*in_->value);
-  inputs.addArg(CpuMatrix(w_ptr ? w_ptr->getData() : nullptr,
-                          w_ptr ? w_ptr->getHeight() : 0,
-                          input_dim));
-  inputs.addArg(*in_->sequenceStartPositions->getVector(useGpu_));
-  outputs.addArg(*out_->value, ADD_TO);
+  inputs.addArg(*in_->value, *start_pos);
+  if (w_ptr) {
+    inputs.addArg(CpuMatrix(w_ptr->getData(), w_ptr->getHeight(), input_dim),
+                  *start_pos);
+  }
+  outputs.addArg(*out_->value, *start_pos, ADD_TO);
   forward_[0]->calc(inputs, outputs);
 
   if (state_ && config_.context_start() < 0) {
@@ -166,13 +165,16 @@ void ContextProjection::backward(const UpdateCallback& callback) {
 
   BufferArgs inputs;
   BufferArgs outputs;
-  inputs.addArg(CpuMatrix(
-      in_->grad ? in_->grad->getData() : nullptr, batch_size, input_dim));
-  inputs.addArg(CpuMatrix(w_ptr ? w_ptr->getData() : nullptr,
-                          w_ptr ? w_ptr->getHeight() : 0,
-                          input_dim));
-  inputs.addArg(*in_->sequenceStartPositions->getVector(useGpu_));
-  outputs.addArg(*out_->grad, ADD_TO);
+  inputs.addArg(*out_->grad, *in_->sequenceStartPositions->getVector(useGpu_));
+  outputs.addArg(
+      CpuMatrix(
+          in_->grad ? in_->grad->getData() : nullptr, batch_size, input_dim),
+      *in_->sequenceStartPositions->getVector(useGpu_),
+      ADD_TO);
+  outputs.addArg(CpuMatrix(w_ptr ? w_ptr->getData() : nullptr,
+                           w_ptr ? w_ptr->getHeight() : 0,
+                           input_dim),
+                 ADD_TO);
   backward_[0]->calc(inputs, outputs);
 
   if (config_.trainable_padding()) {

paddle/gserver/layers/PadLayer.cpp

@@ -27,11 +27,11 @@ bool PadLayer::init(const LayerMap& layerMap,
   auto& pad_conf = config_.inputs(0).pad_conf();
   auto& img_conf = pad_conf.image_conf();
   CHECK_EQ(config_.inputs_size(), 1);
-  inDims_.push_back(0);
-  inDims_.push_back(img_conf.channels());
-  inDims_.push_back(img_conf.has_img_size_y() ? img_conf.img_size_y()
-                                              : img_conf.img_size());
-  inDims_.push_back(img_conf.img_size());
+  inDims_ = TensorShape(
+      {0,
+       img_conf.channels(),
+       img_conf.has_img_size_y() ? img_conf.img_size_y() : img_conf.img_size(),
+       img_conf.img_size()});
 
   CHECK_EQ(2, pad_conf.pad_c_size());
   CHECK_EQ(2, pad_conf.pad_h_size());
@@ -43,7 +43,7 @@ bool PadLayer::init(const LayerMap& layerMap,
   padw_.push_back(pad_conf.pad_w(0));
   padw_.push_back(pad_conf.pad_w(1));
 
-  outDims_.resize(4);
+  outDims_ = TensorShape(4);
   setOutDims(0);
 
   createFunction(forward_,
@@ -68,20 +68,20 @@ bool PadLayer::init(const LayerMap& layerMap,
   return true;
 }
 
-void PadLayer::setOutDims(int batchSize) {
-  outDims_[0] = batchSize;
-  outDims_[1] = inDims_[1] + padc_[0] + padc_[1];
-  outDims_[2] = inDims_[2] + padh_[0] + padh_[1];
-  outDims_[3] = inDims_[3] + padw_[0] + padw_[1];
+void PadLayer::setOutDims(const size_t batchSize) {
+  outDims_.reshape({batchSize,
+                    inDims_[1] + padc_[0] + padc_[1],
+                    inDims_[2] + padh_[0] + padh_[1],
+                    inDims_[3] + padw_[0] + padw_[1]});
 }
 
-void PadLayer::setTensorDim(int batchSize) {
+void PadLayer::setTensorDim(const size_t batchSize) {
   CHECK_EQ(static_cast<int>(inputLayers_.size()), 1);
-  inDims_[0] = batchSize;
+  inDims_.setDim(0, batchSize);
   int h = inputLayers_[0]->getOutput().getFrameHeight();
-  if (h != 0) inDims_[2];
+  if (h != 0) inDims_.setDim(2, h);
   int w = inputLayers_[0]->getOutput().getFrameWidth();
-  if (w != 0) inDims_[3];
+  if (w != 0) inDims_.setDim(3, w);
   setOutDims(batchSize);
 }
 
@@ -94,22 +94,22 @@ void PadLayer::forward(PassType passType) {
   resetOutput(batchSize, size);
   MatrixPtr outV = getOutputValue();
   REGISTER_TIMER_INFO("PadForward", getName().c_str());
-  forward_[0]->calc({Tensor(input->getData(), inDims_)},
-                    {Tensor(outV->getData(), outDims_)},
-                    {});
+
+  BufferArgs inputs;
+  BufferArgs outputs;
+  inputs.addArg(*getInputValue(0), inDims_);
+  outputs.addArg(*getOutputValue(), outDims_, ASSIGN_TO);
+  forward_[0]->calc(inputs, outputs);
 }
 
 void PadLayer::backward(const UpdateCallback& callback) {
   (void)callback;
-
-  MatrixPtr preGrad = inputLayers_[0]->getOutputGrad();
-  if (NULL == preGrad) {
-    return;
-  }
-  MatrixPtr outGrad = getOutputGrad();
   REGISTER_TIMER_INFO("PadBackward", getName().c_str());
-  backward_[0]->calc({Tensor(outGrad->getData(), outDims_)},
-                     {},
-                     {Tensor(preGrad->getData(), inDims_)});
+
+  BufferArgs inputs;
+  BufferArgs outputs;
+  inputs.addArg(*getOutputGrad(), outDims_);
+  outputs.addArg(*getInputGrad(0), inDims_, ADD_TO);
+  backward_[0]->calc(inputs, outputs);
 }
 }  // namespace paddle

paddle/gserver/layers/PadLayer.h

@@ -33,13 +33,13 @@ public:
   void backward(const UpdateCallback& callback = nullptr);
 
 protected:
-  void setOutDims(int batchSize);
-  void setTensorDim(int batchSize);
+  void setOutDims(const size_t batchSize);
+  void setTensorDim(const size_t batchSize);
 
   std::vector<int> padc_;
   std::vector<int> padh_;
   std::vector<int> padw_;
-  Dims inDims_;
-  Dims outDims_;
+  TensorShape inDims_;
+  TensorShape outDims_;
 };
 }  // namespace paddle

paddle/py_paddle/dataprovider_converter.py

@@ -34,6 +34,10 @@ class IScanner(object):
 
 
 class DenseScanner(IScanner):
+    """
+    :type __mat__: numpy.ndarray
+    """
+
     def __init__(self, input_type, pos):
         IScanner.__init__(self, input_type, pos)
         self.__mat__ = None
@@ -47,6 +51,8 @@ class DenseScanner(IScanner):
     def finish_scan(self, argument):
         assert isinstance(argument, swig_paddle.Arguments)
         assert isinstance(self.input_type, dp2.InputType)
+        if self.__mat__.dtype != numpy.float32:
+            self.__mat__ = self.__mat__.astype(numpy.float32)
         m = swig_paddle.Matrix.createDenseFromNumpy(self.__mat__, True, False)
         argument.setSlotValue(self.pos, m)
 

paddle/scripts/travis/before_install.osx.sh

 #!/bin/bash
 brew update
 brew tap homebrew/science
-brew install python
-sudo pip install --upgrade protobuf
-brew install swig openblas md5sha1sum protobuf
+brew install openblas swig md5sha1sum

paddle/scripts/travis/build_and_test.sh

@@ -6,7 +6,7 @@ if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then
   export PYTHONPATH=/opt/python/2.7.12/lib/python2.7/site-packages
   export PYTHONHOME=/opt/python/2.7.12
   export PATH=/opt/python/2.7.12/bin:${PATH}
-  cmake .. -DON_TRAVIS=ON -DON_COVERALLS=ON -DCOVERALLS_UPLOAD=ON ${EXTRA_CMAKE_OPTS}
+  cmake .. -DCMAKE_Fortran_COMPILER=/usr/bin/gfortran-4.8 -DON_TRAVIS=ON -DON_COVERALLS=ON -DCOVERALLS_UPLOAD=ON ${EXTRA_CMAKE_OPTS}
   NRPOC=`nproc`
   make -j $NPROC
   make coveralls

paddle/scripts/travis/docs.sh

@@ -4,7 +4,7 @@
 source ./common.sh
 
 # Compile Documentation only.
-cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_DOC=ON ${EXTRA_CMAKE_OPTS}
+cmake .. -DCMAKE_BUILD_TYPE=Debug -DCMAKE_Fortran_COMPILER=/usr/bin/gfortran-4.8 -DWITH_GPU=OFF -DWITH_DOC=ON ${EXTRA_CMAKE_OPTS}
 make paddle_docs paddle_docs_cn
 
 # check websites for broken links

python/paddle/trainer/config_parser.py

@@ -2650,7 +2650,7 @@ class AverageLayer(LayerBase):
 
 @config_layer('cos')
 class CosSimLayer(LayerBase):
-    def __init__(self, name, inputs, cos_scale=5, device=None):
+    def __init__(self, name, inputs, cos_scale=1, device=None):
         super(CosSimLayer, self).__init__(
             name, 'cos', 1, inputs=inputs, device=device)
         config_assert(len(self.inputs) == 2, 'CosSimLayer must have 2 inputs')

python/paddle/trainer_config_helpers/layers.py

@@ -1674,7 +1674,7 @@ def trans_layer(input, name=None, layer_attr=None):
 
 @wrap_name_default()
 @layer_support()
-def cos_sim(a, b, scale=5, size=1, name=None, layer_attr=None):
+def cos_sim(a, b, scale=1, size=1, name=None, layer_attr=None):
     """
     Cosine Similarity Layer. The cosine similarity equation is here.
 

python/paddle/trainer_config_helpers/tests/CMakeLists.txt

@@ -9,17 +9,10 @@ add_test(NAME test_reset_hook
         ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/python/paddle/trainer_config_helpers/tests/test_reset_hook.py
     WORKING_DIRECTORY ${PROJ_ROOT}/python/paddle)
 
-if (PROTOBUF_3)
-  add_paddle_exe(protobuf_equal
-    ProtobufEqualMain.cpp)
-  add_test(NAME test_layerHelpers
-    COMMAND
-    ${PROJ_ROOT}/python/paddle/trainer_config_helpers/tests/configs/run_tests.sh ${PYTHON_EXECUTABLE}
-    ${CMAKE_CURRENT_BINARY_DIR}/protobuf_equal
-  )
-else()
-  add_test(NAME test_layerHelpers
-    COMMAND
-    ${PROJ_ROOT}/python/paddle/trainer_config_helpers/tests/configs/run_tests.sh ${PYTHON_EXECUTABLE}
-  )
-endif()
+add_paddle_exe(protobuf_equal
+  ProtobufEqualMain.cpp)
+add_test(NAME test_layerHelpers
+  COMMAND
+  ${PROJ_ROOT}/python/paddle/trainer_config_helpers/tests/configs/run_tests.sh ${PYTHON_EXECUTABLE}
+  ${CMAKE_CURRENT_BINARY_DIR}/protobuf_equal
+)

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

@@ -79,7 +79,7 @@ layers {
   inputs {
     input_layer_name: "b"
   }
-  cos_scale: 5
+  cos_scale: 1
 }
 layers {
   name: "__cos_sim_1__"
@@ -92,7 +92,7 @@ layers {
   inputs {
     input_layer_name: "c"
   }
-  cos_scale: 5
+  cos_scale: 1
 }
 layers {
   name: "__sum_to_one_norm_layer_0__"

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

@@ -2,16 +2,18 @@
 cd `dirname $0`
 
 set -e
+PYTHON_EXEC=$1
+COMPARE_PROTO_UTIL=$2
 
 protostr=`dirname $0`/protostr
 
 files=`ls $protostr | grep -v "unittest"`
 
-./generate_protostr.sh $1
+./generate_protostr.sh ${PYTHON_EXEC}
 
 . ./file_list.sh
 
-if [ -z $1 ]; then
+if [ -z ${COMPARE_PROTO_UTIL} ]; then
   for file in $files
   do
       base_protostr=$protostr/$file
@@ -22,20 +24,20 @@ if [ -z $1 ]; then
 else
   for file in ${configs[*]}
   do
-    if ! $1 $protostr/$file.protostr $protostr/$file.protostr.unittest; then
+    if ! ${COMPARE_PROTO_UTIL} $protostr/$file.protostr $protostr/$file.protostr.unittest; then
       diff $protostr/$file.protostr $protostr/$file.protostr.unittest -u
     fi
-    if ! $1 $protostr/$file.protostr $protostr/$file.protostr.non_file_config.unittest; then
+    if ! ${COMPARE_PROTO_UTIL} $protostr/$file.protostr $protostr/$file.protostr.non_file_config.unittest; then
       diff $protostr/$file.protostr $protostr/$file.protostr.non_file_config.unittest -u
     fi
   done
 
   for file in ${whole_configs[*]}
   do
-    if ! $1 $protostr/$file.protostr $protostr/$file.protostr.unittest --whole; then
+    if ! ${COMPARE_PROTO_UTIL} $protostr/$file.protostr $protostr/$file.protostr.unittest --whole; then
       diff $protostr/$file.protostr $protostr/$file.protostr.unittest -u
     fi
-    if ! $1 $protostr/$file.protostr $protostr/$file.protostr.non_file_config.unittest --whole; then
+    if ! ${COMPARE_PROTO_UTIL} $protostr/$file.protostr $protostr/$file.protostr.non_file_config.unittest --whole; then
       diff $protostr/$file.protostr $protostr/$file.protostr.non_file_config.unittest -u
     fi
   done