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

cmake/cblas.cmake

 # Find the CBlas and lapack libraries
 #
-# It will search MKL, atlas, OpenBlas, reference-cblas in order.
+# It will search MKLML, atlas, OpenBlas, reference-cblas in order.
 #
 # If any cblas implementation found, the following variable will be set.
-#    CBLAS_PROVIDER  # one of MKL, ATLAS, OPENBLAS, REFERENCE
+#    CBLAS_PROVIDER  # one of MKLML, ATLAS, OPENBLAS, REFERENCE
 #    CBLAS_INC_DIR   # the include directory for cblas.
 #    CBLAS_LIBS      # a list of libraries should be linked by paddle.
 #                    # Each library should be full path to object file.
-#
-# User should set one of MKL_ROOT, ATLAS_ROOT, OPENBLAS_ROOT, REFERENCE_CBLAS_ROOT
-# during cmake. If none of them set, it will try to find cblas implementation in
-# system paths.
-#
 
 set(CBLAS_FOUND OFF)
 
@@ -30,44 +25,6 @@ if(WITH_MKLML AND MKLML_INC_DIR AND MKLML_LIB)
   return()
 endif()
 
-## Then find MKL.
-set(INTEL_MKL_ROOT "/opt/intel/mkl" CACHE PATH "Folder contains intel mkl libs")
-set(MKL_ROOT $ENV{MKL_ROOT} CACHE PATH "Folder contains env MKL")
-
-set(MKL_INCLUDE_SEARCH_PATHS
-  ${MKL_ROOT}/include
-  ${INTEL_MKL_ROOT}/include)
-set(MKL_LIB_SEARCH_PATHS
-  ${MKL_ROOT}/lib
-  ${MKL_ROOT}/lib/intel64
-  ${INTEL_MKL_ROOT}/lib
-  ${INTEL_MKL_ROOT}/lib/intel64)
-
-find_path(MKL_INC_DIR mkl.h PATHS
-  ${MKL_INCLUDE_SEARCH_PATHS})
-find_path(MKL_LAPACK_INC_DIR mkl_lapacke.h PATHS
-  ${MKL_INCLUDE_SEARCH_PATHS})
-find_library(MKL_CORE_LIB NAMES mkl_core PATHS
-  ${MKL_LIB_SEARCH_PATHS})
-find_library(MKL_SEQUENTIAL_LIB NAMES mkl_sequential PATHS
-  ${MKL_LIB_SEARCH_PATHS})
-find_library(MKL_INTEL_LP64 NAMES mkl_intel_lp64 PATHS
-  ${MKL_LIB_SEARCH_PATHS})
-
-if(MKL_LAPACK_INC_DIR AND MKL_INC_DIR AND MKL_CORE_LIB AND MKL_SEQUENTIAL_LIB AND MKL_INTEL_LP64)
-  set(CBLAS_FOUND ON)
-  set(CBLAS_PROVIDER MKL)
-  set(CBLAS_INC_DIR ${MKL_INC_DIR} ${MKL_LAPACK_INC_DIR})
-  set(CBLAS_LIBRARIES ${MKL_INTEL_LP64} ${MKL_SEQUENTIAL_LIB} ${MKL_CORE_LIB})
-
-  add_definitions(-DPADDLE_USE_MKL)
-  add_definitions(-DLAPACK_FOUND)
-
-  message(STATUS "Found MKL (include: ${MKL_INC_DIR}, library: ${CBLAS_LIBRARIES})")
-  message(STATUS "Found lapack in MKL (include: ${MKL_LAPACK_INC_DIR})")
-  return()
-endif()
-
 ## Then find atlas.
 set(ATLAS_ROOT $ENV{ATLAS_ROOT} CACHE PATH "Folder contains Atlas")
 set(ATLAS_INCLUDE_SEARCH_PATHS

cmake/external/mkldnn.cmake

@@ -46,16 +46,20 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
     MESSAGE(STATUS "Build MKLDNN with ${MKLDNN_MKLROOT}")
 ENDIF()
 
+SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow")
+SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} -Wno-error=strict-overflow")
 ExternalProject_Add(
     ${MKLDNN_PROJECT}
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
-    GIT_TAG             "v0.10"
+    GIT_TAG             "v0.11"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
     CMAKE_ARGS          -DMKLROOT=${MKLDNN_MKLROOT}
+    CMAKE_ARGS          -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
+    CMAKE_ARGS          -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
     CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
                         -DMKLROOT:PATH=${MKLDNN_MKLROOT}
 )

cmake/external/mklml.cmake

@@ -27,8 +27,8 @@ ENDIF()
 INCLUDE(ExternalProject)
 
 SET(MKLML_PROJECT       "extern_mklml")
-SET(MKLML_VER           "mklml_lnx_2018.0.20170720")
-SET(MKLML_URL           "https://github.com/01org/mkl-dnn/releases/download/v0.10/${MKLML_VER}.tgz")
+SET(MKLML_VER           "mklml_lnx_2018.0.1.20171007")
+SET(MKLML_URL           "https://github.com/01org/mkl-dnn/releases/download/v0.11/${MKLML_VER}.tgz")
 SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
 SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
 SET(MKLML_DST_DIR       "mklml")

cmake/external/openblas.cmake

@@ -86,7 +86,7 @@ IF(NOT ${CBLAS_FOUND})
         UPDATE_COMMAND      ""
         CONFIGURE_COMMAND   ""
     )
-
+    SET(CBLAS_PROVIDER openblas)
     IF(WITH_C_API)
         INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
         # Because libopenblas.a is a symbolic link of another library, thus need to
@@ -115,7 +115,7 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
 # linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
 SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
 FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
-IF(${CBLAS_PROVIDER} MATCHES MKL)
+IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
     ADD_LIBRARY(cblas SHARED ${dummyfile})
 ELSE()
     ADD_LIBRARY(cblas STATIC ${dummyfile})

cmake/generic.cmake

@@ -93,7 +93,7 @@ include_directories(${CMAKE_CURRENT_BINARY_DIR})
 if(NOT APPLE AND NOT ANDROID)
     find_package(Threads REQUIRED)
     link_libraries(${CMAKE_THREAD_LIBS_INIT})
-    set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -ldl -lrt")
+    set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -pthread -ldl -lrt")
 endif(NOT APPLE AND NOT ANDROID)
 
 function(merge_static_libs TARGET_NAME)

doc/api/v2/config/layer.rst

@@ -82,6 +82,11 @@ maxout
 ..  autoclass:: paddle.v2.layer.maxout
     :noindex:
 
+roi_pool
+--------
+..  autoclass:: paddle.v2.layer.roi_pool
+    :noindex:
+
 Norm Layer
 ==========
 

doc/api/v2/data.rst

@@ -2,112 +2,9 @@
 Data Reader Interface and DataSets
 ==================================
 
+..  toctree::
+    :maxdepth: 1
 
-DataTypes
-=========
-
-..  automodule:: paddle.v2.data_type
-    :members:
-    :noindex:
-
-DataFeeder
-==========
-
-..  automodule:: paddle.v2.data_feeder
-    :members:
-    :noindex:
-
-Reader
-======
-
-..  automodule:: paddle.v2.reader
-    :members:
-    :noindex:
-
-..  automodule:: paddle.v2.reader.creator
-    :members:
-    :noindex:
-
-minibatch
-=========
-
-..  automodule:: paddle.v2.minibatch
-    :members:
-    :noindex:
-
-Dataset
-=======
-
-..  automodule:: paddle.v2.dataset
-    :members:
-    :noindex:
-
-mnist
-+++++
-
-..  automodule:: paddle.v2.dataset.mnist
-    :members:
-    :noindex:
-
-cifar
-+++++
-
-..  automodule:: paddle.v2.dataset.cifar
-    :members:
-    :noindex:
-
-conll05
-+++++++
-
-..  automodule:: paddle.v2.dataset.conll05
-    :members: get_dict,get_embedding,test
-    :noindex:
-
-imdb
-++++
-
-..  automodule:: paddle.v2.dataset.imdb
-    :members:
-    :noindex:
-
-imikolov
-++++++++
-
-..  automodule:: paddle.v2.dataset.imikolov
-    :members:
-    :noindex:
-
-movielens
-+++++++++
-
-..  automodule:: paddle.v2.dataset.movielens
-    :members:
-    :noindex:
-
-..  autoclass:: paddle.v2.dataset.movielens.MovieInfo
-    :noindex:
-    
-..  autoclass:: paddle.v2.dataset.movielens.UserInfo
-    :noindex:
-
-sentiment
-+++++++++
-
-..  automodule:: paddle.v2.dataset.sentiment
-    :members:
-    :noindex:
-
-uci_housing
-+++++++++++
-
-..  automodule:: paddle.v2.dataset.uci_housing
-    :members:
-    :noindex:
-
-wmt14
-+++++
-
-..  automodule:: paddle.v2.dataset.wmt14
-    :members:
-    :noindex:
-
+    data/data_reader.rst
+    data/image.rst
+    data/dataset.rst

doc/api/v2/data/data_reader.rst

+=====================
+Data Reader Interface
+=====================
+
+
+DataTypes
+=========
+
+..  automodule:: paddle.v2.data_type
+    :members:
+    :noindex:
+
+DataFeeder
+==========
+
+..  automodule:: paddle.v2.data_feeder
+    :members:
+    :noindex:
+
+Reader
+======
+
+..  automodule:: paddle.v2.reader
+    :members:
+    :noindex:
+
+..  automodule:: paddle.v2.reader.creator
+    :members:
+    :noindex:
+
+minibatch
+=========
+
+..  automodule:: paddle.v2.minibatch
+    :members:
+    :noindex:

doc/api/v2/data/dataset.rst

+Dataset
+=======
+
+..  automodule:: paddle.v2.dataset
+    :members:
+    :noindex:
+
+mnist
++++++
+
+..  automodule:: paddle.v2.dataset.mnist
+    :members:
+    :noindex:
+
+cifar
++++++
+
+..  automodule:: paddle.v2.dataset.cifar
+    :members:
+    :noindex:
+
+conll05
++++++++
+
+..  automodule:: paddle.v2.dataset.conll05
+    :members: get_dict,get_embedding,test
+    :noindex:
+
+imdb
+++++
+
+..  automodule:: paddle.v2.dataset.imdb
+    :members:
+    :noindex:
+
+imikolov
+++++++++
+
+..  automodule:: paddle.v2.dataset.imikolov
+    :members:
+    :noindex:
+
+movielens
++++++++++
+
+..  automodule:: paddle.v2.dataset.movielens
+    :members:
+    :noindex:
+
+..  autoclass:: paddle.v2.dataset.movielens.MovieInfo
+    :noindex:
+    
+..  autoclass:: paddle.v2.dataset.movielens.UserInfo
+    :noindex:
+
+sentiment
++++++++++
+
+..  automodule:: paddle.v2.dataset.sentiment
+    :members:
+    :noindex:
+
+uci_housing
++++++++++++
+
+..  automodule:: paddle.v2.dataset.uci_housing
+    :members:
+    :noindex:
+
+wmt14
++++++
+
+..  automodule:: paddle.v2.dataset.wmt14
+    :members:
+    :noindex:

doc/api/v2/data/image.rst

+Image Interface
+===============
+
+..  automodule:: paddle.v2.image
+    :members:

doc/design/mkldnn/README.MD

@@ -15,6 +15,7 @@
  	- [CMake](#cmake)
 	- [Layers](#layers)
 	- [Activations](#activations)
+	- [Weights](#weights)
 	- [Unit Tests](#unit-tests)
 	- [Protobuf Messages](#protobuf-messages)
 	- [Python API](#python-api)
@@ -45,17 +46,23 @@ Figure 1. PaddlePaddle on IA.
 
 ### Layers
 所有MKL-DNN相关的C++ layers，都会按照PaddlePaddle的目录结构存放在
-`paddle/gserver/layers`中，并且文件名都会一以*Mkldnn*开头。
+`paddle/gserver/layers`中，并且文件名都会一以*MKLDNN*开头。
 
-所有MKL-DNN的layers都会继承于一个叫做`MkldnnLayer`的父类，该父类继承于PaddlePaddle的基类`Layer`。
+所有MKL-DNN的layers都会继承于一个叫做`MKLDNNLayer`的父类，该父类继承于PaddlePaddle的基类`Layer`。
+
+在`MKLDNNLayer`中会提供一些必要的接口和函数，并且会写好`forward`和`backward`的基本逻辑。部分函数定义为纯虚函数，子类只需要实现这些函数即可。
 
 ### Activations
-由于在PaddlePaddle中，激活函数是独立于layer概念的，所以会在`paddle/gserver/activations`目录下添加一个`MkldnnActivation.h`文件定义一些用于MKL-DNN的接口，实现方法还是会在`ActivationFunction.cpp`文件。
+由于在PaddlePaddle中，激活函数是独立于layer概念的，所以会在`paddle/gserver/activations`目录下添加`MKLDNNActivation.h`和`MKLDNNActivation.cpp`文件用于定义和使用MKL-DNN的接口。
 
-### Unit Tests
-会在`paddle/gserver/test`目录下添加`test_Mkldnn.cpp`和`MkldnnTester.*`用于MKL-DNN的测试。
+### Weights
+由于有些layer是含有参数的，我们会尽量让MKL-DNN的参数与PaddlePaddle中`parameter`共享一块内存。
+同时，由于MKL-DNN在训练时使用的参数layout可能与PaddlePaddle默认的`nchw`不一致，我们会在网络训练的开始和结束时分别转换这个layout，使得最终保存的参数格式与PaddlePaddle一致。
 
-Activation的测试，计划在PaddlePaddle原有的测试文件上直接添加新的测试type。
+### Unit Tests
+会在`paddle/gserver/test`目录下添加`test_MKLDNN.cpp`和`MKLDNNTester.*`用于MKL-DNN的测试。
+测试分为每个layer(或activation)的单元测试和简单网络的整体测试。
+每个测试会对比PaddlePaddle中CPU算出的结果与MKL-DNN的结果，小于某个比较小的阈值认为通过。
 
 ### Protobuf Messages
 根据具体layer的需求可能会在`proto/ModelConfig.proto`里面添加必要的选项。
@@ -82,7 +89,7 @@ if use_mkldnn
 会在`v1_api_demo`目录下添加一个`mkldnn`的文件夹，里面放入一些用于MKL-DNN测试的demo脚本。
 
 ### Benchmarking
-会考虑添加部分逻辑在`benchmark/paddle/image/run.sh`，添加使用MKL-DNN的测试。
+会添加`benchmark/paddle/image/run_mkldnn.sh`，用于测试使用MKL-DNN之后的性能。
 
 ### Others
 1. 如果在使用MKL-DNN的情况下，会把CPU的Buffer对齐为64。
@@ -94,14 +101,16 @@ if use_mkldnn
 
 我们总结出一些特别需要注意的点：
 
-1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数，我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MkldnnLayer`特有的设备ID。
+1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数，我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MKLDNNLayer`特有的设备ID。
 2. 重写父类Layer的**init**函数，修改`deviceId_`为`-2`，代表这个layer是用于跑在MKL-DNN的环境下。
-3. 创建`MkldnnMatrix`，用于管理MKL-DNN会用到的相关memory函数、接口以及会用的到格式信息。
-4. 创建`MkldnnBase`，定义一些除了layer和memory相关的类和函数。包括MKL-DNN会用到`MkldnnStream`和`CpuEngine`，和未来可能还会用到`FPGAEngine`等。
-5. 在**Argument**里添加两个`MkldnnMatrixPtr`，取名为`mkldnnValue`和`mkldnnGrad`，用于存放`MkldnnLayer`会用到的memory buffer。 并且添加函数cvt(会修改为一个更加合适的函数名)，用于处理"CPU device"和"MKL-DNN device"之间memory的相互转化。
-6. 在父类`Layer`中的`getOutput`函数中添加一段逻辑，用于判断`deviceId`，并针对device在MKL-DNN和CPU之间不统一的情况，做一个前期转换。 也就是调用`Argument`的cvt函数把output统一到需要的device上。
-7. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
-8. 关于MKLDNN参数的保存。由于MKLDNN参数的格式与PaddlePaddle原有的格式存在不一样的情况，所以需要在保存参数时同时保存该格式信息。目前准备扩展[Header](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/parameter/Parameter.h#L247)里面的`int32_t version`。这个值不管是在v1还是在v2里面，一直保存的是0，所以可以充分利用这个信息，定义一个枚举处理所有MKLDNN的参数格式，从而`MKLDNNLayer`就可以从输入的参数中获取需要的格式信息。
+3. 创建`MKLDNNMatrix`，同时继承`CpuMatrix`和`mkldnn::memory`。用于管理MKL-DNN会用到的相关memory函数、接口以及会用的到格式信息。
+4. 创建`MKLDNNBase`，定义一些除了layer和memory相关的类和函数。包括MKL-DNN会用到`MKLDNNStream`和`CPUEngine`，和未来可能还会用到`FPGAEngine`等。
+5. 每个`MKLDNNlayer`都会有`inVal_`,`inGrad_`,`outVal_`和`outGrad_`，分别代表input value， input gradient，output value和output gradient。他们会存放MKL-DNN用到的internal memory。同时还会定义以*ext*开头的`MKLDNNMatrix`(表示external的memory)，主要是在格式与PaddlePaddle默认的`nchw`格式不匹配时，用于转换内存的工作。必要的转换函数也会在`MKLDNNLayer`中提前定义好，每个子类只需要调用定义好的reset buffer函数即可。
+6. 每个`MKLDNNlayer`的resetbuffer相关的函数（包括reset input、output的Value和grad），他们会根据输入参数reset internal和external的memory，当然这两者也可以相等，即表示不需要转换。只需要把握一个原则，每个`MKLDNNlayer`的子类，只需要使用internal的memory就可以了，所有external的转换工作在父类的reset函数中都提前准备好了。
+7. 一般来说，external的memory会尽量与PaddlePaddle中的`value`和`grad`共享内存。同时每个`MKLDNNLayer`中的external output value和gradient(也就是`extOutVal_`和`extOutGrad_`)必须分别与`output_.value`和`output_.grad`共享内存，因为PaddlePaddle的activation会直接使用`output_.value`和`output_.grad`。如果不需要external的buffer用于转换，那么internal的buffer也会与他们共享内存。
+8. 如果MKL-DNN layer的后面接有cpu device，那么就会使`output_.value`与`extOutVal_`共享内存，同时数据格式就是`nchw`，这样下一个cpu device就能拿到正确的数据。在有cpu device的时候，external的memory的格式始终是`nchw`或者`nc`。
+9. 由于MKL-DNN的输出操作都是覆盖data的，不是在原来的数据上累加，所以当网络出现分支时，在`backward`时会需要merge不同layer的梯度。`MKLDNNlayer`中会实现merge的方法，此时每个小分支的input gradient会先临时保存在一个`MKLDNNMatrix`中，由分支处的layer负责求和，并把结果放到这个layer的`output_.grad`中。所以整体上，每个子类并不会需要关心分支的事情，也是在父类都实现好了。
+10. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
 
 ## References
 

doc/design/ops/sequence_decoder.md

+# Design: Sequence Decoder Generating LoDTensors
+In tasks such as machine translation and image to text, 
+a [sequence decoder](https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/README.md) is necessary to generate sequences.
+
+This documentation describes how to implement the sequence decoder as an operator.
+
+## Beam Search based Decoder
+The [beam search algorithm](https://en.wikipedia.org/wiki/Beam_search) is necessary when generating sequences, 
+it is a heuristic search algorithm that explores the paths by expanding the most promising node in a limited set.
+
+In the old version of PaddlePaddle, a C++ class `RecurrentGradientMachine` implements the general sequence decoder based on beam search, 
+due to the complexity, the implementation relays on a lot of special data structures, 
+quite trivial and hard to be customized by users.
+
+There are a lot of heuristic tricks in the sequence generation tasks, 
+so the flexibility of sequence decoder is very important to users.
+
+During PaddlePaddle's refactoring work,
+some new concept is proposed such as [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) and [TensorArray](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/tensor_array.md) that can better support sequence usage,
+and they can help to make the implementation of beam search based sequence decoder **more transparent and modular** .
+
+For example, the RNN sates, candidates IDs and probabilities of beam search can be represented as `LoDTensors`;
+the selected candidate's IDs in each time step can be stored in a `TensorArray`, and `Packed` to the sentences translated.
+
+## Changing LoD's absolute offset to relative offsets
+The current `LoDTensor` is designed to store levels of variable-length sequences,
+it stores several arrays of integers each represents a level.
+
+The integers in each level represents the begin and end (not inclusive) offset of a sequence **in the underlying tensor**, 
+let's call this format the **absolute-offset LoD** for clear.
+
+The relative-offset LoD can fast retrieve any sequence but fails to represent empty sequences, for example, a two-level LoD is as follows
+```python
+[[0, 3, 9]
+ [0, 2, 3, 3, 3, 9]]
+```
+The first level tells that there are two sequences:
+- the first's offset is `[0, 3)`
+- the second's offset is `[3, 9)`
+
+while on the second level, there are several empty sequences that both begin and end at `3`.
+It is impossible to tell how many empty second-level sequences exist in the first-level sequences.
+
+There are many scenarios that relay on empty sequence representation,
+such as machine translation or image to text, one instance has no translations or the empty candidate set for a prefix.
+
+So let's introduce another format of LoD, 
+it stores **the offsets of the lower level sequences** and is called **relative-offset** LoD.
+
+For example, to represent the same sequences of the above data
+
+```python
+[[0, 3, 6]
+ [0, 2, 3, 3, 3, 9]]
+```
+
+the first level represents that there are two sequences, 
+their offsets in the second-level LoD is `[0, 3)` and `[3, 5)`.
+
+The second level is the same with the relative offset example because the lower level is a tensor.
+It is easy to find out the second sequence in the first-level LoD has two empty sequences.
+
+The following demos are based on relative-offset LoD.
+
+## Usage in a simple machine translation model
+Let's start from a simple machine translation model that is simplified from [machine translation chapter](https://github.com/PaddlePaddle/book/tree/develop/08.machine_translation) to draw a simple blueprint of what a sequence decoder can do and how to use it.
+
+The model has an encoder that learns the semantic vector from a sequence,
+and a decoder which uses the sequence decoder to generate new sentences.
+
+**Encoder**
+```python
+import paddle as pd
+
+dict_size = 8000
+source_dict_size = dict_size
+target_dict_size = dict_size
+word_vector_dim = 128
+encoder_dim = 128
+decoder_dim = 128
+beam_size = 5
+max_length = 120
+
+# encoder
+src_word_id = pd.data(
+    name='source_language_word',
+    type=pd.data.integer_value_sequence(source_dict_dim))
+src_embedding = pd.embedding(size=source_dict_size, size=word_vector_dim)
+
+src_word_vec = pd.lookup(src_embedding, src_word_id)
+
+encoder_out_seq = pd.gru(input=src_word_vec, size=encoder_dim)
+
+encoder_ctx = pd.last_seq(encoder_out_seq)
+# encoder_ctx_proj is the learned semantic vector
+encoder_ctx_proj = pd.fc(
+    encoder_ctx, size=decoder_dim, act=pd.activation.Tanh(), bias=None)
+```
+
+**Decoder**
+
+```python
+def generate():
+    decoder = pd.while_loop()
+    with decoder.step():
+        decoder_mem = decoder.memory(init=encoder_ctx)  # mark the memory
+        generated_ids = decoder.memory() # TODO init to batch_size <s>s
+        generated_scores = decoder.memory() # TODO init to batch_size 1s or 0s
+
+        target_word = pd.lookup(trg_embedding, gendrated_ids)
+        # expand encoder_ctx's batch to fit target_word's lod
+        # for example
+        # decoder_mem.lod is
+        # [[0 1 3],
+        #  [0 1 3 6]]
+        # its tensor content is [a1 a2 a3 a4 a5]
+        # which means there are 2 sentences to translate
+        #   - the first sentence has 1 translation prefixes, the offsets are [0, 1)
+        #   - the second sentence has 2 translation prefixes, the offsets are [1, 3) and [3, 6)
+        # the target_word.lod is 
+        # [[0, 1, 6]
+        #  [0, 2, 4, 7, 9 12]]
+        # which means 2 sentences to translate, each has 1 and 5 prefixes
+        # the first prefix has 2 candidates
+        # the following has 2, 3, 2, 3 candidates
+        # the encoder_ctx_expanded's content will be
+        # [a1 a1 a2 a2 a3 a3 a3 a4 a4 a5 a5 a5]
+        encoder_ctx_expanded = pd.lod_expand(encoder_ctx, target_word)
+        decoder_input = pd.fc(
+            act=pd.activation.Linear(),
+            input=[target_word, encoder_ctx],
+            size=3 * decoder_dim)
+        gru_out, cur_mem = pd.gru_step(
+            decoder_input, mem=decoder_mem, size=decoder_dim)
+        scores = pd.fc(
+            gru_out,
+            size=trg_dic_size,
+            bias=None,
+            act=pd.activation.Softmax())
+        # K is an config
+        topk_scores, topk_ids = pd.top_k(scores, K)
+        topk_generated_scores = pd.add_scalar(topk_scores, generated_scores)
+
+        selected_ids, selected_generation_scores = decoder.beam_search(
+            topk_ids, topk_generated_scores)
+
+        # update the states
+        decoder_mem.update(cur_mem)  # tells how to update state
+        generated_ids.update(selected_ids)
+        generated_scores.update(selected_generation_scores)
+
+        decoder.output(selected_ids)
+        decoder.output(selected_generation_scores)
+
+translation_ids, translation_scores = decoder()
+```
+The `decoder.beam_search` is a operator that given the candidates and the scores of translations including the candidates,
+return the result of the beam search algorithm.
+
+In this way, users can customize anything on the inputs or outputs of beam search, for example, two ways to prune some translation prefixes
+
+1. meke the correspondind elements in `topk_generated_scores` zero or some small values, beam_search will discard this candidate.
+2. remove some specific candidate in `selected_ids`
+3. get the final `translation_ids`, remove the translation sequence in it.
+
+The implementation of sequence decoder can reuse the C++ class [RNNAlgorithm](https://github.com/Superjom/Paddle/blob/68cac3c0f8451fe62a4cdf156747d6dc0ee000b3/paddle/operators/dynamic_recurrent_op.h#L30),
+so the python syntax is quite similar to a [RNN](https://github.com/Superjom/Paddle/blob/68cac3c0f8451fe62a4cdf156747d6dc0ee000b3/doc/design/block.md#blocks-with-for-and-rnnop).
+
+Both of them are two-level `LoDTensors`
+
+- the first level represents `batch_size` of (source) sentences;
+- the second level represents the candidate ID sets for translation prefix.
+
+for example, 3 source sentences to translate, and has 2, 3, 1 candidates.
+
+Unlike an RNN, in sequence decoder, the previous state and the current state have different LoD and shape,
+a `lod_expand` operator is used to expand the LoD of the previous state to fit the current state.
+
+For example, the previous state
+
+* LoD is `[0, 1, 3][0, 2, 5, 6]`
+* content of tensor is `a1 a2 b1 b2 b3 c1`
+
+the current state stored in `encoder_ctx_expanded`
+
+* LoD is `[0, 2, 7][0 3 5 8 9 11 11]`
+* the content is 
+  - a1 a1 a1 (a1 has 3 candidates, so the state should be copied 3 times for each candidates)
+  - a2 a2
+  - b1 b1 b1
+  - b2
+  - b3 b3
+  - None (c1 has 0 candidates, so c1 is dropped)
+
+Benefit from the relative offset LoD, empty candidate set can be represented naturally.
+
+the status in each time step can be stored in `TensorArray`, and `Pack`ed to a final LoDTensor, the corresponding syntax is 
+
+```python
+decoder.output(selected_ids)
+decoder.output(selected_generation_scores)
+```
+
+the `selected_ids` is the candidate ids for the prefixes, 
+it will be `Packed` by `TensorArray` to a two-level `LoDTensor`,
+the first level represents the source sequences,
+the second level represents generated sequences.
+
+Pack the `selected_scores` will get a `LoDTensor` that stores scores of each candidate of translations.
+
+Pack the `selected_generation_scores` will get a `LoDTensor`, and each tail is the probability of the translation.
+
+## LoD and shape changes during decoding
+<p align="center">
+  <img src="./images/LOD-and-shape-changes-during-decoding.jpg"/>
+</p>
+
+According the image above, the only phrase to change LoD is beam search.
+
+## Beam search design
+The beam search algorthm will be implemented as one method of the sequence decoder, it has 3 inputs
+
+1. `topk_ids`, top K candidate ids for each prefix.
+2. `topk_scores`, the corresponding scores for `topk_ids`
+3. `generated_scores`, the score of the prefixes.
+
+All of the are LoDTensors, so that the sequence affilication is clear.
+Beam search will keep a beam for each prefix and select a smaller candidate set for each prefix.
+
+It will return three variables
+
+1. `selected_ids`, the final candidate beam search function selected for the next step.
+2. `selected_scores`, the scores for the candidates.
+3. `generated_scores`, the updated scores for each prefixes (with the new candidates appended).
+
+## Introducing the LoD-based `Pack` and `Unpack` methods in `TensorArray`
+The `selected_ids`, `selected_scores` and `generated_scores` are LoDTensors,
+and they exist in each time step,
+so it is natural to store them in arrays.
+
+Currently, PaddlePaddle has a module called `TensorArray` which can store an array of tensors,
+the results of beam search are better to store in a `TensorArray`.
+
+The `Pack` and `UnPack` in `TensorArray` are used to package tensors in the array to a `LoDTensor` or split the `LoDTensor` to an array of tensors. 
+It needs some extensions to support pack or unpack an array of `LoDTensors`.

doc/faq/local/index_cn.rst

@@ -99,7 +99,7 @@ PaddlePaddle支持Sparse的训练，sparse训练需要训练特征是 :code:`spa
 利用更多的计算资源
 ++++++++++++++++++
 
-利用更多的计算资源可以分为一下几个方式来进行\:
+利用更多的计算资源可以分为以下几个方式来进行\:
 
 * 单机CPU训练
 

doc/howto/dev/new_op_cn.md

@@ -214,7 +214,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 
     ```cpp
     // if use Eigen unsupported module before include head files
-    #define EIGEN_USE_GPU
+    // #define EIGEN_USE_GPU
 
     namespace ops = paddle::operators;
     REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);

paddle/capi/Matrix.cpp

@@ -54,6 +54,46 @@ paddle_error paddle_matrix_set_row(paddle_matrix mat,
   return kPD_NO_ERROR;
 }
 
+PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat,
+                                          paddle_real* value) {
+  if (mat == nullptr || value == nullptr) return kPD_NULLPTR;
+  auto ptr = cast(mat);
+  if (ptr->mat == nullptr) return kPD_NULLPTR;
+  paddle::real* buf = ptr->mat->getRowBuf(0);
+  size_t width = ptr->mat->getWidth();
+  size_t height = ptr->mat->getHeight();
+  if (ptr->mat->useGpu()) {
+#ifdef PADDLE_WITH_CUDA
+    hl_memcpy(buf, value, sizeof(paddle::real) * width * height);
+#else
+    return kPD_NOT_SUPPORTED;
+#endif
+  } else {
+    std::copy(value, value + width * height, buf);
+  }
+  return kPD_NO_ERROR;
+}
+
+PD_API paddle_error paddle_matrix_get_value(paddle_matrix mat,
+                                          paddle_real* result) {
+  if (mat == nullptr || result == nullptr) return kPD_NULLPTR;
+  auto ptr = cast(mat);
+  if (ptr->mat == nullptr) return kPD_NULLPTR;
+  paddle::real* buf = ptr->mat->getRowBuf(0);
+  size_t width = ptr->mat->getWidth();
+  size_t height = ptr->mat->getHeight();
+  if (ptr->mat->useGpu()) {
+#ifdef PADDLE_WITH_CUDA
+    hl_memcpy(result, buf, width * height * sizeof(paddle::real));
+#else
+    return kPD_NOT_SUPPORTED;
+#endif
+  } else {
+    std::copy(buf, buf + width * height, result);
+  }
+  return kPD_NO_ERROR;
+}
+
 paddle_error paddle_matrix_get_row(paddle_matrix mat,
                                    uint64_t rowID,
                                    paddle_real** rawRowBuffer) {

paddle/capi/examples/model_inference/dense/main.c

@@ -27,18 +27,20 @@ int main() {
   CHECK(paddle_arguments_resize(in_args, 1));
 
   // Create input matrix.
-  paddle_matrix mat = paddle_matrix_create(/* sample_num */ 1,
+  paddle_matrix mat = paddle_matrix_create(/* sample_num */ 10,
                                            /* size */ 784,
                                            /* useGPU */ false);
   srand(time(0));
-  paddle_real* array;
 
-  // Get First row.
-  CHECK(paddle_matrix_get_row(mat, 0, &array));
+  std::vector<paddle_real> input;
+  input.resize(784 * 10);
 
-  for (int i = 0; i < 784; ++i) {
-    array[i] = rand() / ((float)RAND_MAX);
+  for (int i = 0; i < input.size(); ++i) {
+    input[i] = rand() / ((float)RAND_MAX);
   }
+  
+  // Set value for the input matrix
+  CHECK(paddle_matrix_set_value(mat, input.data()));
 
   CHECK(paddle_arguments_set_value(in_args, 0, mat));
 
@@ -51,11 +53,17 @@ int main() {
 
   CHECK(paddle_arguments_get_value(out_args, 0, prob));
 
-  CHECK(paddle_matrix_get_row(prob, 0, &array));
+  std::std::vector<paddle_real> result;
+  int height;
+  int width;
+
+  CHECK(paddle_matrix_get_shape(prob, &height, &width);
+  result.resize(height * width);
+  CHECK(paddle_matrix_get_value(prob, result.data()));
 
   printf("Prob: ");
-  for (int i = 0; i < 10; ++i) {
-    printf("%.2f ", array[i]);
+  for (int i = 0; i < height * width; ++i) {
+    printf("%.2f ", result[i]);
   }
   printf("\n");
 

paddle/capi/matrix.h

@@ -70,6 +70,16 @@ PD_API paddle_error paddle_matrix_set_row(paddle_matrix mat,
                                           uint64_t rowID,
                                           paddle_real* rowArray);
 
+/**
+ * @brief paddle_matrix_set_value Set value to matrix.
+ * @param mat Target Matrix
+ * @param value Row data.
+ * @return paddle_error
+ * @note  value should contain enough element of data to init the mat
+ */
+PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat,
+                                          paddle_real* value);
+
 /**
  * @brief PDMatGetRow Get raw row buffer from matrix
  * @param [in] mat Target matrix
@@ -81,6 +91,15 @@ PD_API paddle_error paddle_matrix_get_row(paddle_matrix mat,
                                           uint64_t rowID,
                                           paddle_real** rawRowBuffer);
 
+/**
+ * @brief copy data from the matrix 
+ * @param [in] mat Target matrix
+ * @param [out] result pointer to store the matrix data 
+ * @return paddle_error
+ * @note the space of the result should allocated before invoke this API
+ */
+PD_API paddle_error paddle_matrix_get_value(paddle_matrix mat,
+                                          paddle_real* result);
 /**
  * @brief PDMatCreateNone Create None Matrix
  * @return

paddle/capi/tests/test_Matrix.cpp

@@ -45,3 +45,49 @@ TEST(CAPIMatrix, createNone) {
   paddle_matrix mat = paddle_matrix_create_none();
   ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat));
 }
+
+TEST(CAPIMatrix, cpu_get_set_value) {
+  paddle_matrix mat = paddle_matrix_create(128, 32, false);
+  std::vector<paddle_real> sample;
+  std::vector<paddle_real> result;
+  sample.resize(128 * 32);
+  result.resize(128 * 32);
+  for (size_t i = 0; i < sample.size(); ++i) {
+    sample[i] = 1.0 / (i + 1.0);
+  }
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_set_value(mat, sample.data()));
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_value(mat, result.data()));
+  for (size_t i = 0; i < sample.size(); ++i) {
+    ASSERT_NEAR(sample[i], result[i], 1e-5);
+  }
+
+  uint64_t height, width;
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_shape(mat, &height, &width));
+  ASSERT_EQ(128UL, height);
+  ASSERT_EQ(32UL, width);
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat));
+}
+
+#ifdef PADDLE_WITH_CUDA
+TEST(CAPIMatrix, gpu_get_set_value) {
+  paddle_matrix mat = paddle_matrix_create(128, 32, true);
+  std::vector<paddle_real> sample;
+  std::vector<paddle_real> result;
+  sample.resize(128 * 32);
+  result.resize(128 * 32);
+  for (size_t i = 0; i < sample.size(); ++i) {
+    sample[i] = 1.0 / (i + 1.0);
+  }
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_set_value(mat, sample.data()));
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_value(mat, result.data()));
+  for (size_t i = 0; i < sample.size(); ++i) {
+    ASSERT_NEAR(sample[i], result[i], 1e-5);
+  }
+
+  uint64_t height, width;
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_shape(mat, &height, &width));
+  ASSERT_EQ(128UL, height);
+  ASSERT_EQ(32UL, width);
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat));
+}
+#endif

paddle/framework/backward.cc

@@ -321,8 +321,6 @@ static void CreateGradVarInBlock(
         auto* param = block_desc->FindVarRecursive(pname);
         auto* grad = block_desc->FindVar(arg);
         if (param == nullptr) {
-          LOG(WARNING) << "Cannot find forward variable of " << arg
-                       << ". Set its gradient to FP32";
           grad->SetDataType(DataType::FP32);
         } else {
           grad->SetDataType(param->GetDataType());
@@ -408,6 +406,11 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
 
     for (const auto& desc : op_grads) {
       for (const std::string& out_name : desc->OutputArgumentNames()) {
+        if (out_name.find("@GRAD") == std::string::npos) {
+          // Not all outputs of a backward operator is a gradient. Only gradient
+          // need to be sum. Skip variables are not gradient.
+          continue;
+        }
         dup_out_ops[out_name].emplace_back(grad_desc_idx);
       }
       ++grad_desc_idx;

paddle/framework/backward_test.cc

@@ -21,7 +21,7 @@
 #include "paddle/framework/var_desc.h"
 #include "paddle/operators/net_op.h"
 
-USE_OP(fill_constant);
+USE_NO_KERNEL_OP(fill_constant);
 
 namespace paddle {
 namespace framework {

paddle/framework/block_desc.cc

@@ -50,6 +50,15 @@ VarDescBind *BlockDescBind::FindVarRecursive(const std::string &name) const {
   return it->second.get();
 }
 
+VarDescBind *BlockDescBind::FindRecursiveOrCreateVar(
+    const std::string &name_bytes) {
+  VarDescBind *res = FindVarRecursive(name_bytes);
+  if (res == nullptr) {
+    res = Var(name_bytes);
+  }
+  return res;
+}
+
 bool BlockDescBind::HasVarRecursive(const std::string &name) const {
   return FindVarRecursive(name) != nullptr;
 }

paddle/framework/block_desc.h

@@ -58,6 +58,8 @@ class BlockDescBind {
 
   VarDescBind *FindVarRecursive(const std::string &name_bytes) const;
 
+  VarDescBind *FindRecursiveOrCreateVar(const std::string &name_bytes);
+
   bool HasVarRecursive(const std::string &var_name) const;
 
   std::set<std::string> LocalVarNames() const {

paddle/framework/data_type.h

@@ -34,6 +34,21 @@ inline DataType ToDataType(std::type_index type) {
   }
 }
 
+inline std::type_index ToTypeIndex(DataType type) {
+  switch (type) {
+    case DataType::FP32:
+      return typeid(float);
+    case DataType::FP64:
+      return typeid(double);
+    case DataType::INT32:
+      return typeid(int);
+    case DataType::INT64:
+      return typeid(int64_t);
+    default:
+      PADDLE_THROW("Not support type %d", type);
+  }
+}
+
 template <typename Visitor>
 inline void VisitDataType(DataType type, Visitor visitor) {
   switch (type) {

paddle/framework/ddim.cc

@@ -79,6 +79,13 @@ DDim make_ddim(const std::vector<int64_t>& dims) {
   return result;
 }
 
+DDim make_ddim(const std::vector<int>& dims) {
+  std::vector<int64_t> res(dims.size());
+  std::transform(dims.begin(), dims.end(), res.begin(),
+                 [](int d) { return static_cast<int64_t>(d); });
+  return make_ddim(res);
+}
+
 /// @cond HIDDEN
 // XXX For some reason, putting this in an anonymous namespace causes errors
 class DynamicMutableIndexer : public boost::static_visitor<int64_t&> {

paddle/framework/ddim.h

@@ -81,6 +81,8 @@ struct DDim {
  */
 DDim make_ddim(const std::vector<int64_t>& dims);
 
+DDim make_ddim(const std::vector<int>& dims);
+
 /**
  * \brief Make a DDim from an initializer list
  *

paddle/framework/op_desc.cc

@@ -357,7 +357,8 @@ void OpDescBind::InferVarType(BlockDescBind *block) const {
                 "LOD_TENSOR";
     for (auto &out_pair : this->outputs_) {
       for (auto &out_var_name : out_pair.second) {
-        block->Var(out_var_name)->SetType(VarDesc::LOD_TENSOR);
+        block->FindRecursiveOrCreateVar(out_var_name)
+            ->SetType(VarDesc::LOD_TENSOR);
       }
     }
   }

paddle/framework/scope.cc

@@ -98,5 +98,23 @@ void Scope::DeleteScope(Scope* scope) {
   delete scope;
 }
 
+void Scope::Rename(const std::string& origin_name,
+                   const std::string& new_name) const {
+  auto origin_it = vars_.find(origin_name);
+  PADDLE_ENFORCE(origin_it != vars_.end(),
+                 "Cannot find original variable with name %s", origin_name);
+  auto new_it = vars_.find(new_name);
+  PADDLE_ENFORCE(new_it == vars_.end(),
+                 "The variable with name %s is already in the scope", new_name);
+  vars_[new_name] = origin_it->second;
+  vars_.erase(origin_it);
+}
+
+std::string Scope::Rename(const std::string& origin_name) const {
+  auto var_name = string::Sprintf("%p.%d", this, vars_.size());
+  Rename(origin_name, var_name);
+  return var_name;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/scope.h

@@ -68,11 +68,18 @@ class Scope {
   // enumerate all the variables current contains.
   std::vector<std::string> GetAllNames(bool recursive = false) const;
 
+  // Rename variable to a new name
+  void Rename(const std::string& origin_name,
+              const std::string& new_name) const;
+
+  // Rename variable to a new name and return the new name
+  std::string Rename(const std::string& origin_name) const;
+
  private:
   // Call Scope::NewScope for a sub-scope.
   explicit Scope(Scope const* parent) : parent_(parent) {}
 
-  std::unordered_map<std::string, Variable*> vars_;
+  mutable std::unordered_map<std::string, Variable*> vars_;
   mutable std::list<Scope*> kids_;
   Scope const* parent_{nullptr};
 

paddle/function/CMakeLists.txt

@@ -45,6 +45,7 @@ if(WITH_GPU)
     add_simple_unittest(BlockExpandOpTest)
     add_simple_unittest(CropOpTest)
     add_simple_unittest(SwitchOpTest)
+    add_simple_unittest(ScaleSubRegionOpTest)
 endif()
 
 add_simple_unittest(Im2ColTest)

paddle/function/FunctionTest.h

@@ -110,6 +110,7 @@ public:
         function2_(FunctionBase::funcRegistrar_.createByType(name2)) {
     function1_->init(config);
     function2_->init(config);
+    initArgsCallback_ = nullptr;
   }
 
   ~Compare2Function() {}
@@ -170,6 +171,10 @@ public:
                                       *seq2_));
   }
 
+  void registerInitCallback(std::function<void(BufferArg&, size_t)> callback) {
+    initArgsCallback_ = callback;
+  }
+
   // output need only contains shape, do not contains data.
   void addOutputs(const BufferArg& output, ArgType argType = ASSIGN_TO) {
     size_t size =
@@ -340,6 +345,10 @@ protected:
         initArg(*func1Inputs_[i]);
       }
 
+      if (initArgsCallback_ != nullptr) {
+        initArgsCallback_(*func1Inputs_[i], i);
+      }
+
       copyArg_(*func1Inputs_[i], *func2Inputs_[i]);
     }
   }
@@ -386,6 +395,7 @@ protected:
   std::shared_ptr<SequenceIdArg> seq1_;
   std::shared_ptr<SequenceIdArg> seq2_;
   test::CopyArgument<DType1, DType2> copyArg_;
+  std::function<void(BufferArg&, size_t)> initArgsCallback_;
 };
 
 class CpuGpuFuncCompare

paddle/function/ScaleSubRegionOp.cpp

+/* 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 "ScaleSubRegionOp.h"
+#include "paddle/function/TensorShape.h"
+
+namespace paddle {
+
+template <>
+void ScaleSubRegion<DEVICE_TYPE_CPU>(real* outputs,
+                                     const real* inputs,
+                                     const real* indices,
+                                     const TensorShape shape,
+                                     const FuncConfig& conf) {
+  real value = conf.get<real>("value");
+
+  int number = shape[0];
+  int channel = shape[1];
+  int height = shape[2];
+  int width = shape[3];
+
+  memcpy(outputs, inputs, number * channel * height * width * sizeof(real));
+
+  for (int n = 0; n < number; ++n) {
+    // indices start from 1
+    int offset = n * 6;
+    for (int c = indices[offset] - 1; c < indices[offset + 1]; ++c) {
+      for (int h = indices[offset + 2] - 1; h < indices[offset + 3]; ++h) {
+        for (int w = indices[offset + 4] - 1; w < indices[offset + 5]; ++w) {
+          int idx = ((n * channel + c) * height + h) * width + w;
+          outputs[idx] *= value;
+        }
+      }
+    }
+  }
+}
+
+template <>
+void ScaleSubRegionGrad<DEVICE_TYPE_CPU>(const real* inGrad,
+                                         real* outGrad,
+                                         const real* indices,
+                                         const TensorShape shape,
+                                         const FuncConfig& conf) {
+  real value = conf.get<real>("value");
+
+  int number = shape[0];
+  int channel = shape[1];
+  int height = shape[2];
+  int width = shape[3];
+
+  for (int n = 0; n < number; ++n) {
+    for (int c = 0; c < channel; ++c) {
+      for (int h = 0; h < height; ++h) {
+        for (int w = 0; w < width; ++w) {
+          int idx = ((n * channel + c) * height + h) * width + w;
+          int offset = n * 6;
+          if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) &&
+              h >= (indices[offset + 2] - 1) &&
+              h <= (indices[offset + 3] - 1) &&
+              w >= (indices[offset + 4] - 1) &&
+              w <= (indices[offset + 5] - 1)) {
+            outGrad[idx] += inGrad[idx] * value;
+          } else {
+            outGrad[idx] += inGrad[idx];
+          }
+        }
+      }
+    }
+  }
+}
+
+/**
+ * \brief For each instance, ScaleSubRegion can be used to multiply a value to
+ *        a specified sub continuous region. By providing start index and end
+ *        index for C/H/W, you can specify the location and shape of the region.
+ *
+ * Argument in this Function:
+ * \param inputs    A 4-D tensor with shape [N, C, H, W], only one input.
+ * \param indices   A 2-D tensor with shape [N, 6], indicates the sub region.
+ * \param outputs   A 4-D tensor with same shape as inputs, output value.
+ */
+template <DeviceType Device>
+class ScaleSubRegionFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override { conf_ = config; }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(2UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+    CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO);
+
+    TensorShape shape = inputs[0].shape();
+
+    ScaleSubRegion<Device>(outputs[0].data<real>(),
+                           inputs[0].data<real>(),
+                           inputs[1].data<real>(),
+                           shape,
+                           conf_);
+  }
+
+private:
+  FuncConfig conf_;
+};
+
+/**
+ * \brief The backward propagation of ScaleSubRegion Function.
+ *
+ * Argument in this Function:
+ * \param inputs  A 4-D tensor with shape [N, C, H, W], output gradient.
+ * \param indices A 2-D tensor with shape [N, 6], indicates the sub region.
+ * \param outputs A 4-D tensor with shape [N, C, H, W], gradient of input value.
+ */
+
+template <DeviceType Device>
+class ScaleSubRegionGradFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override { conf_ = config; }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(2UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+    CHECK_EQ(outputs[0].getArgType(), ADD_TO);
+
+    TensorShape shape = inputs[0].shape();
+
+    ScaleSubRegionGrad<Device>(inputs[0].data<real>(),
+                               outputs[0].data<real>(),
+                               inputs[1].data<real>(),
+                               shape,
+                               conf_);
+  }
+
+private:
+  FuncConfig conf_;
+};
+
+REGISTER_TYPED_FUNC(ScaleSubRegion, CPU, ScaleSubRegionFunc);
+REGISTER_TYPED_FUNC(ScaleSubRegionGrad, CPU, ScaleSubRegionGradFunc);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_TYPED_FUNC(ScaleSubRegion, GPU, ScaleSubRegionFunc);
+REGISTER_TYPED_FUNC(ScaleSubRegionGrad, GPU, ScaleSubRegionGradFunc);
+#endif
+
+}  // namespace paddle

paddle/function/ScaleSubRegionOp.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 "Function.h"
+
+namespace paddle {
+
+/**
+ * \brief Function to multiply a value to values in specified sub continuous
+ *        region. Indices must be provided to indcate the location and shape of
+ *        the region and the multiplied value is passed by configure variable.
+ *
+ *
+ * \param[out] outputs  Output value.
+ * \param[in]  inputs   Input data which contains NCHW information.
+ * \param[in]  indices  Indices data to indcate the sub region.
+ * \param[in]  shape    Tensor shape of input value.
+ * \param[in]  conf     Configure variable which contains the multiplied value.
+ */
+template <DeviceType Device>
+void ScaleSubRegion(real* outputs,
+                    const real* inputs,
+                    const real* indices,
+                    const TensorShape shape,
+                    const FuncConfig& conf);
+
+/**
+ * \brief Backward propagation function of ScaleSubRegion.
+ *
+ * \param[out] inGrad   Gradients of previous layer.
+ * \param[in]  outGrad  Output gradient.
+ * \param[in]  indices  Indices data.
+ * \param[in]  shape    The Shape of input tensor.
+ * \param[in]  conf     Configure variable.
+ */
+template <DeviceType Device>
+void ScaleSubRegionGrad(const real* inGrad,
+                        real* outGrad,
+                        const real* indices,
+                        const TensorShape shape,
+                        const FuncConfig& conf);
+}  // namespace paddle

paddle/function/ScaleSubRegionOpGpu.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 "ScaleSubRegionOp.h"
+#include "hl_base.h"
+
+namespace paddle {
+
+__global__ void KeScaleSubRegion(real* outputs,
+                                 const real* inputs,
+                                 const real* indices,
+                                 real value,
+                                 int channel,
+                                 int height,
+                                 int width,
+                                 int nthreads) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int w = idx % width;
+    const int h = (idx / width) % height;
+    const int c = (idx / width / height) % channel;
+    const int n = idx / width / height / channel;
+
+    const int offset = n * 6;
+    if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) &&
+        h >= (indices[offset + 2] - 1) && h <= (indices[offset + 3] - 1) &&
+        w >= (indices[offset + 4] - 1) && w <= (indices[offset + 5] - 1)) {
+      outputs[idx] = inputs[idx] * value;
+    } else {
+      outputs[idx] = inputs[idx];
+    }
+  }
+}
+
+template <>
+void ScaleSubRegion<DEVICE_TYPE_GPU>(real* outputs,
+                                     const real* inputs,
+                                     const real* indices,
+                                     const TensorShape shape,
+                                     const FuncConfig& conf) {
+  real value = conf.get<real>("value");
+
+  int number = shape[0];
+  int channel = shape[1];
+  int height = shape[2];
+  int width = shape[3];
+
+  size_t nth = number * channel * height * width;
+  int blockSize = 1024;
+  int gridSize = (nth + blockSize - 1) / blockSize;
+
+  KeScaleSubRegion<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>(
+      outputs, inputs, indices, value, channel, height, width, nth);
+  CHECK_SYNC("ScaleSubRegion");
+}
+
+__global__ void KeScaleSubRegionDiff(const real* inGrad,
+                                     real* outGrad,
+                                     const real* indices,
+                                     real value,
+                                     int channel,
+                                     int height,
+                                     int width,
+                                     int nthreads) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int w = idx % width;
+    const int h = (idx / width) % height;
+    const int c = (idx / width / height) % channel;
+    const int n = idx / width / height / channel;
+
+    const int offset = n * 6;
+    if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) &&
+        h >= (indices[offset + 2] - 1) && h <= (indices[offset + 3] - 1) &&
+        w >= (indices[offset + 4] - 1) && w <= (indices[offset + 5] - 1)) {
+      outGrad[idx] += inGrad[idx] * value;
+    } else {
+      outGrad[idx] += inGrad[idx];
+    }
+  }
+}
+
+template <>
+void ScaleSubRegionGrad<DEVICE_TYPE_GPU>(const real* inGrad,
+                                         real* outGrad,
+                                         const real* indices,
+                                         const TensorShape shape,
+                                         const FuncConfig& conf) {
+  real value = conf.get<real>("value");
+
+  int number = shape[0];
+  int channel = shape[1];
+  int height = shape[2];
+  int width = shape[3];
+
+  size_t nth = number * channel * height * width;
+  int blockSize = 1024;
+  int gridSize = (nth + blockSize - 1) / blockSize;
+
+  KeScaleSubRegionDiff<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>(
+      inGrad, outGrad, indices, value, channel, height, width, nth);
+  CHECK_SYNC("ScaleSubRegionGrad");
+}
+
+}  // namespace paddle

paddle/function/ScaleSubRegionOpTest.cpp

+/* 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 <gtest/gtest.h>
+#include "FunctionTest.h"
+
+namespace paddle {
+
+TEST(ScaleSubRegion, real) {
+  for (size_t numSamples : {5, 32}) {
+    for (size_t channels : {5, 32}) {
+      for (size_t imgSizeH : {5, 33}) {
+        for (size_t imgSizeW : {5, 32}) {
+          for (real value : {-0.5, 0.0, 0.5}) {
+            for (bool firstHalf : {false, true}) {
+              VLOG(3) << " numSamples=" << numSamples
+                      << " channels=" << channels << " imgSizeH=" << imgSizeH
+                      << " imgSizeW=" << imgSizeW;
+
+              for (bool testGrad : {false, true}) {
+                CpuGpuFuncCompare compare(
+                    testGrad ? "ScaleSubRegionGrad" : "ScaleSubRegion",
+                    FuncConfig().set<real>("value", value));
+
+                TensorShape shape{numSamples, channels, imgSizeH, imgSizeW};
+                TensorShape indicesShape{numSamples, 6};
+
+                compare.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
+                compare.addInputs(BufferArg(VALUE_TYPE_FLOAT, indicesShape));
+
+                compare.registerInitCallback([=](BufferArg& arg, size_t index) {
+                  if (index == 1) {
+                    real* data = (real*)arg.data();
+
+                    for (size_t i = 0; i < numSamples; ++i) {
+                      size_t offset = i * 6;
+                      data[offset] = firstHalf ? 1 : channels / 2;
+                      data[offset + 1] = firstHalf ? channels / 2 : channels;
+                      data[offset + 2] = firstHalf ? 1 : imgSizeH / 2;
+                      data[offset + 3] = firstHalf ? imgSizeH / 2 : imgSizeH;
+                      data[offset + 4] = firstHalf ? 1 : imgSizeW / 2;
+                      data[offset + 5] = firstHalf ? imgSizeW / 2 : imgSizeW;
+                    }
+                  }
+                });
+
+                compare.addOutputs(
+                    BufferArg(
+                        VALUE_TYPE_FLOAT, shape, testGrad ? ADD_TO : ASSIGN_TO),
+                    testGrad ? ADD_TO : ASSIGN_TO);
+                compare.run();
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNAddtoLayer.cpp

@@ -62,16 +62,14 @@ void MKLDNNAddtoLayer::resetFwd(std::vector<primitive>& pipeline,
                                 MKLDNNMatrixPtr& wgt,
                                 MKLDNNMatrixPtr& bias,
                                 MKLDNNMatrixPtr& out) {
-  if (biases_) {
-    LOG(FATAL) << "not implemented yet";
-  }
-  resetFwdBuffers(inVals_, out);
+  resetFwdBuffers(inVals_, bias, out);
   in = inVals_[0];
 
   std::shared_ptr<sum::primitive_desc> fwdPD;
-  resetFwdPD(fwdPD, inVals_, out);
+  std::shared_ptr<sum::primitive_desc> biasPD;
+  resetFwdPD(fwdPD, biasPD, inVals_, bias, out);
 
-  resetFwdPipeline(pipeline, fwdPD, inVals_, out);
+  resetFwdPipeline(pipeline, fwdPD, biasPD, inVals_, bias, out);
 }
 
 void MKLDNNAddtoLayer::resetBwd(std::vector<primitive>& pipeline,
@@ -79,7 +77,7 @@ void MKLDNNAddtoLayer::resetBwd(std::vector<primitive>& pipeline,
                                 MKLDNNMatrixPtr& wgt,
                                 MKLDNNMatrixPtr& bias,
                                 MKLDNNMatrixPtr& out) {
-  resetBwdBuffers(inGrads_, out);
+  resetBwdBuffers(inGrads_, bias, out);
   in = inGrads_[0];
 
   // backward only need share output grad to input grad
@@ -89,6 +87,20 @@ void MKLDNNAddtoLayer::resetBwd(std::vector<primitive>& pipeline,
       inputLayers_[i]->getOutputGrad()->setData(inGrads_[i]->getData());
     }
   }
+
+  // backward bias
+  bwdBias_ = nullptr;
+  if (bias) {
+    std::vector<float> scales(bs_, 1.0);
+    std::vector<memory::primitive_desc> srcPDs(bs_, bias->getPrimitiveDesc());
+    auto biasPD = sum::primitive_desc(bias->getMemoryDesc(), scales, srcPDs);
+    std::vector<primitive::at> srcs;
+    for (size_t i = 0; i < grads_.size(); ++i) {
+      srcs.push_back(*(grads_[i]));
+    }
+    bwdBias_.reset(new sum(biasPD, srcs, *bias));
+    pipeline.push_back(*bwdBias_);
+  }
 }
 
 void MKLDNNAddtoLayer::updateWeights(const UpdateCallback& callback) {
@@ -97,7 +109,25 @@ void MKLDNNAddtoLayer::updateWeights(const UpdateCallback& callback) {
   }
 }
 
+void MKLDNNAddtoLayer::prepareBias(MKLDNNMatrixPtr& bias,
+                                   const MatrixPtr& biasMat,
+                                   const MKLDNNMatrixPtr& out,
+                                   std::vector<MKLDNNMatrixPtr>& outs) {
+  auto pd = MKLDNNMatrix::createPrimitiveDesc(
+      {(int)layerSize_}, memory::format::x, engine_);
+  bias = MKLDNNMatrix::create(pd, biasMat);
+  outs.clear();
+  real* data = out->getData();
+  CHECK_EQ(bs_ * layerSize_, out->getElementCnt());
+  for (int i = 0; i < bs_; ++i) {
+    MatrixPtr tmp =
+        Matrix::create(data + i * layerSize_, 1, layerSize_, false, false);
+    outs.push_back(MKLDNNMatrix::create(bias->getPrimitiveDesc(), tmp));
+  }
+}
+
 void MKLDNNAddtoLayer::resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                                       MKLDNNMatrixPtr& bias,
                                        MKLDNNMatrixPtr& out) {
   inputs.resize(inputLayers_.size());
   for (size_t i = 0; i < inputs.size(); i++) {
@@ -110,12 +140,20 @@ void MKLDNNAddtoLayer::resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
   }
 
   resetOutValue(out, inputs[0]->getPrimitiveDesc());
+
+  if (biases_ && biases_->getW()) {
+    prepareBias(bias, biases_->getW(), out, vals_);
+  } else {
+    bias = nullptr;
+  }
 }
 
 void MKLDNNAddtoLayer::resetFwdPD(std::shared_ptr<sum::primitive_desc>& pd,
+                                  std::shared_ptr<sum::primitive_desc>& biasPD,
                                   std::vector<MKLDNNMatrixPtr>& inputs,
+                                  MKLDNNMatrixPtr bias,
                                   MKLDNNMatrixPtr out) {
-  std::vector<double> scales(inputs.size(), 1.0);
+  std::vector<float> scales(inputs.size(), 1.0);
   std::vector<memory::primitive_desc> srcPDs;
   for (size_t i = 0; i < inputs.size(); i++) {
     srcPDs.push_back(inputs[i]->getPrimitiveDesc());
@@ -123,12 +161,23 @@ void MKLDNNAddtoLayer::resetFwdPD(std::shared_ptr<sum::primitive_desc>& pd,
   CHECK(out);
   pd.reset(new sum::primitive_desc(out->getMemoryDesc(), scales, srcPDs));
   CHECK_PRIMITIVE_DESC_EQ(out, pd->dst_primitive_desc());
+
+  biasPD = nullptr;
+  if (bias) {
+    std::vector<float> scales(2, 1.0);
+    std::vector<memory::primitive_desc> srcPDs(2, bias->getPrimitiveDesc());
+    biasPD.reset(
+        new sum::primitive_desc(bias->getMemoryDesc(), scales, srcPDs));
+    CHECK_PRIMITIVE_DESC_EQ(bias, biasPD->dst_primitive_desc());
+  }
 }
 
 void MKLDNNAddtoLayer::resetFwdPipeline(
     std::vector<primitive>& pipeline,
     std::shared_ptr<sum::primitive_desc>& pd,
+    std::shared_ptr<sum::primitive_desc>& biasPD,
     std::vector<MKLDNNMatrixPtr>& inputs,
+    MKLDNNMatrixPtr& bias,
     MKLDNNMatrixPtr& out) {
   std::vector<primitive::at> srcs;
   for (size_t i = 0; i < inputs.size(); i++) {
@@ -136,9 +185,23 @@ void MKLDNNAddtoLayer::resetFwdPipeline(
   }
   fwd_.reset(new sum(*pd, srcs, *out));
   pipeline.push_back(*fwd_);
+
+  fwdBias_.clear();
+  if (biasPD == nullptr || bias == nullptr) {
+    return;
+  }
+  fwdBias_.resize(vals_.size());
+  for (size_t i = 0; i < vals_.size(); ++i) {
+    std::vector<primitive::at> srcs;
+    srcs.push_back(*(vals_[i]));
+    srcs.push_back(*bias);
+    fwdBias_[i].reset(new sum(*biasPD, srcs, *vals_[i]));
+    pipeline.push_back(*fwdBias_[i]);
+  }
 }
 
 void MKLDNNAddtoLayer::resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                                       MKLDNNMatrixPtr& bias,
                                        MKLDNNMatrixPtr& out) {
   CHECK(outVal_);
   resetOutGrad(out, outVal_->getPrimitiveDesc());
@@ -149,6 +212,12 @@ void MKLDNNAddtoLayer::resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
     resetInGrad(inputs[i], inVal_->getPrimitiveDesc(), i);
     CHECK_PRIMITIVE_DESC_EQ(inputs[i], out->getPrimitiveDesc());
   }
+
+  if (biases_ && biases_->getWGrad()) {
+    prepareBias(bias, biases_->getWGrad(), out, grads_);
+  } else {
+    bias = nullptr;
+  }
 }
 
 }  // namespace paddle

paddle/gserver/layers/MKLDNNAddtoLayer.h

@@ -32,9 +32,15 @@ protected:
   // layer size == ic * ih * iw == oc * oh *ow, and can not be changed
   size_t layerSize_;
 
-  // TODO(TJ): this part has not been optimized by MKL-DNN
   std::unique_ptr<Weight> biases_;
 
+  // buffers for adding bias
+  std::vector<MKLDNNMatrixPtr> vals_;
+  std::vector<MKLDNNMatrixPtr> grads_;
+  // primitives for adding bias
+  std::vector<std::shared_ptr<mkldnn::primitive>> fwdBias_;
+  std::shared_ptr<mkldnn::primitive> bwdBias_;
+
 public:
   explicit MKLDNNAddtoLayer(const LayerConfig& config) : MKLDNNLayer(config) {}
 
@@ -91,20 +97,34 @@ protected:
    *                    reset pipeline.
    */
   void resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                       MKLDNNMatrixPtr& bias,
                        MKLDNNMatrixPtr& out);
   void resetFwdPD(std::shared_ptr<mkldnn::sum::primitive_desc>& pd,
+                  std::shared_ptr<mkldnn::sum::primitive_desc>& biasPD,
                   std::vector<MKLDNNMatrixPtr>& inputs,
+                  MKLDNNMatrixPtr bias,
                   MKLDNNMatrixPtr out);
   void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
                         std::shared_ptr<mkldnn::sum::primitive_desc>& pd,
+                        std::shared_ptr<mkldnn::sum::primitive_desc>& biasPD,
                         std::vector<MKLDNNMatrixPtr>& inputs,
+                        MKLDNNMatrixPtr& bias,
                         MKLDNNMatrixPtr& out);
 
   /**
    * Backward functions: reset buffers(inputs, output, bias)
    */
   void resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                       MKLDNNMatrixPtr& bias,
                        MKLDNNMatrixPtr& out);
+
+  /**
+   * prepare for bias
+   */
+  void prepareBias(MKLDNNMatrixPtr& bias,
+                   const MatrixPtr& biasMat,
+                   const MKLDNNMatrixPtr& out,
+                   std::vector<MKLDNNMatrixPtr>& outs);
 };
 
 }  // namespace paddle

paddle/gserver/layers/MKLDNNBatchNormLayer.cpp

@@ -119,7 +119,7 @@ void MKLDNNBatchNormLayer::reshape(
     int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
   reshapeInput(bs, ih, iw);
   oh = ih;
-  ow = ow;
+  ow = iw;
   // ic_ and oc can not be changed
   CHECK_EQ(inputElemenCnt_ / bs / ih / iw, (size_t)ic)
       << "Input channel can not be changed";

paddle/gserver/layers/MKLDNNLayer.cpp

@@ -287,7 +287,7 @@ void MKLDNNLayer::resetMergeGrad(MKLDNNMatrixPtr& out) {
     return;
   }
   CHECK(out) << "should have reset internal ouput grad";
-  std::vector<double> scales(outputMap_.size(), 1.0);
+  std::vector<float> scales(outputMap_.size(), 1.0);
   std::vector<memory::primitive_desc> srcPDs;
   std::vector<primitive::at> srcs;
   for (auto it = outputMap_.begin(); it != outputMap_.end(); ++it) {

paddle/gserver/layers/ROIPoolLayer.cpp

+/* 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 "ROIPoolLayer.h"
+
+namespace paddle {
+
+REGISTER_LAYER(roi_pool, ROIPoolLayer);
+
+bool ROIPoolLayer::init(const LayerMap& layerMap,
+                        const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+
+  const ROIPoolConfig& layerConf = config_.inputs(0).roi_pool_conf();
+  pooledWidth_ = layerConf.pooled_width();
+  pooledHeight_ = layerConf.pooled_height();
+  spatialScale_ = layerConf.spatial_scale();
+
+  return true;
+}
+
+void ROIPoolLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  const ROIPoolConfig& layerConf = config_.inputs(0).roi_pool_conf();
+  height_ = getInput(0).getFrameHeight();
+  if (!height_) height_ = layerConf.height();
+  width_ = getInput(0).getFrameWidth();
+  if (!width_) width_ = layerConf.width();
+  channels_ = getInputValue(0)->getWidth() / width_ / height_;
+
+  size_t batchSize = getInput(0).getBatchSize();
+  size_t numROIs = getInput(1).getBatchSize();
+
+  MatrixPtr dataValue = getInputValue(0);
+  MatrixPtr roiValue = getInputValue(1);
+  resetOutput(numROIs, channels_ * pooledHeight_ * pooledWidth_);
+  MatrixPtr outputValue = getOutputValue();
+
+  if (useGpu_) {  // TODO(guosheng): implement on GPU later
+    MatrixPtr dataCpuBuffer;
+    Matrix::resizeOrCreate(dataCpuBuffer,
+                           dataValue->getHeight(),
+                           dataValue->getWidth(),
+                           false,
+                           false);
+    MatrixPtr roiCpuBuffer;
+    Matrix::resizeOrCreate(roiCpuBuffer,
+                           roiValue->getHeight(),
+                           roiValue->getWidth(),
+                           false,
+                           false);
+    dataCpuBuffer->copyFrom(*dataValue);
+    roiCpuBuffer->copyFrom(*roiValue);
+    dataValue = dataCpuBuffer;
+    roiValue = roiCpuBuffer;
+    MatrixPtr outputCpuBuffer;
+    Matrix::resizeOrCreate(outputCpuBuffer,
+                           outputValue->getHeight(),
+                           outputValue->getWidth(),
+                           false,
+                           false);
+    outputCpuBuffer->copyFrom(*outputValue);
+    outputValue = outputCpuBuffer;
+  }
+
+  real* bottomData = dataValue->getData();
+  size_t batchOffset = dataValue->getWidth();
+  size_t channelOffset = height_ * width_;
+  real* bottomROIs = roiValue->getData();
+  size_t roiOffset = roiValue->getWidth();
+  size_t poolChannelOffset = pooledHeight_ * pooledWidth_;
+
+  real* outputData = outputValue->getData();
+  Matrix::resizeOrCreate(maxIdxs_,
+                         numROIs,
+                         channels_ * pooledHeight_ * pooledWidth_,
+                         false,
+                         false);
+  real* argmaxData = maxIdxs_->getData();
+
+  for (size_t n = 0; n < numROIs; ++n) {
+    // the first five elememts of each RoI should be:
+    // batch_idx, roi_x_start, roi_y_start, roi_x_end, roi_y_end
+    size_t roiBatchIdx = bottomROIs[0];
+    size_t roiStartW = round(bottomROIs[1] * spatialScale_);
+    size_t roiStartH = round(bottomROIs[2] * spatialScale_);
+    size_t roiEndW = round(bottomROIs[3] * spatialScale_);
+    size_t roiEndH = round(bottomROIs[4] * spatialScale_);
+    CHECK_GE(roiBatchIdx, 0);
+    CHECK_LT(roiBatchIdx, batchSize);
+    size_t roiHeight = std::max(roiEndH - roiStartH + 1, 1UL);
+    size_t roiWidth = std::max(roiEndW - roiStartW + 1, 1UL);
+    real binSizeH =
+        static_cast<real>(roiHeight) / static_cast<real>(pooledHeight_);
+    real binSizeW =
+        static_cast<real>(roiWidth) / static_cast<real>(pooledWidth_);
+    real* batchData = bottomData + batchOffset * roiBatchIdx;
+    for (size_t c = 0; c < channels_; ++c) {
+      for (size_t ph = 0; ph < pooledHeight_; ++ph) {
+        for (size_t pw = 0; pw < pooledWidth_; ++pw) {
+          size_t hstart = static_cast<size_t>(std::floor(ph * binSizeH));
+          size_t wstart = static_cast<size_t>(std::floor(pw * binSizeW));
+          size_t hend = static_cast<size_t>(std::ceil((ph + 1) * binSizeH));
+          size_t wend = static_cast<size_t>(std::ceil((pw + 1) * binSizeW));
+          hstart = std::min(std::max(hstart + roiStartH, 0UL), height_);
+          wstart = std::min(std::max(wstart + roiStartW, 0UL), width_);
+          hend = std::min(std::max(hend + roiStartH, 0UL), height_);
+          wend = std::min(std::max(wend + roiStartW, 0UL), width_);
+
+          bool isEmpty = (hend <= hstart) || (wend <= wstart);
+          size_t poolIndex = ph * pooledWidth_ + pw;
+          if (isEmpty) {
+            outputData[poolIndex] = 0;
+            argmaxData[poolIndex] = -1;
+          }
+
+          for (size_t h = hstart; h < hend; ++h) {
+            for (size_t w = wstart; w < wend; ++w) {
+              size_t index = h * width_ + w;
+              if (batchData[index] > outputData[poolIndex]) {
+                outputData[poolIndex] = batchData[index];
+                argmaxData[poolIndex] = index;
+              }
+            }
+          }
+        }
+      }
+      batchData += channelOffset;
+      outputData += poolChannelOffset;
+      argmaxData += poolChannelOffset;
+    }
+    bottomROIs += roiOffset;
+  }
+  if (useGpu_) {
+    getOutputValue()->copyFrom(*outputValue);
+  }
+}
+
+void ROIPoolLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inGradValue = getInputGrad(0);
+  MatrixPtr outGradValue = getOutputGrad();
+  MatrixPtr roiValue = getInputValue(1);
+
+  if (useGpu_) {
+    MatrixPtr inGradCpuBuffer;
+    Matrix::resizeOrCreate(inGradCpuBuffer,
+                           inGradValue->getHeight(),
+                           inGradValue->getWidth(),
+                           false,
+                           false);
+    MatrixPtr outGradCpuBuffer;
+    Matrix::resizeOrCreate(outGradCpuBuffer,
+                           outGradValue->getHeight(),
+                           outGradValue->getWidth(),
+                           false,
+                           false);
+    MatrixPtr roiCpuBuffer;
+    Matrix::resizeOrCreate(roiCpuBuffer,
+                           roiValue->getHeight(),
+                           roiValue->getWidth(),
+                           false,
+                           false);
+    inGradCpuBuffer->copyFrom(*inGradValue);
+    outGradCpuBuffer->copyFrom(*outGradValue);
+    roiCpuBuffer->copyFrom(*roiValue);
+    inGradValue = inGradCpuBuffer;
+    outGradValue = outGradCpuBuffer;
+    roiValue = roiCpuBuffer;
+  }
+
+  real* bottomROIs = roiValue->getData();
+  size_t numROIs = getInput(1).getBatchSize();
+  size_t roiOffset = getInputValue(1)->getWidth();
+
+  real* inDiffData = inGradValue->getData();
+  size_t batchOffset = getInputValue(0)->getWidth();
+  size_t channelOffset = height_ * width_;
+
+  real* outDiffData = outGradValue->getData();
+  size_t poolChannelOffset = pooledHeight_ * pooledWidth_;
+  real* argmaxData = maxIdxs_->getData();
+
+  for (size_t n = 0; n < numROIs; ++n) {
+    size_t roiBatchIdx = bottomROIs[0];
+    real* batchDiffData = inDiffData + batchOffset * roiBatchIdx;
+    for (size_t c = 0; c < channels_; ++c) {
+      for (size_t ph = 0; ph < pooledHeight_; ++ph) {
+        for (size_t pw = 0; pw < pooledWidth_; ++pw) {
+          size_t poolIndex = ph * pooledWidth_ + pw;
+          if (argmaxData[poolIndex] > 0) {
+            size_t index = static_cast<size_t>(argmaxData[poolIndex]);
+            batchDiffData[index] += outDiffData[poolIndex];
+          }
+        }
+      }
+      batchDiffData += channelOffset;
+      outDiffData += poolChannelOffset;
+      argmaxData += poolChannelOffset;
+    }
+    bottomROIs += roiOffset;
+  }
+
+  if (useGpu_) {
+    getInputGrad(0)->copyFrom(*inGradValue);
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/ROIPoolLayer.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 "Layer.h"
+
+namespace paddle {
+
+/**
+ * A layer used by Fast R-CNN to extract feature maps of ROIs from the last
+ * feature map.
+ * - Input: This layer needs two input layers: The first input layer is a
+ *          convolution layer; The second input layer contains the ROI data
+ *          which is the output of ProposalLayer in Faster R-CNN. layers for
+ *          generating bbox location offset and the classification confidence.
+ * - Output: The ROIs' feature map.
+ * Reference:
+ *    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun.
+ *    Faster R-CNN: Towards Real-Time Object Detection with Region Proposal
+ * Networks
+ */
+
+class ROIPoolLayer : public Layer {
+protected:
+  size_t channels_;
+  size_t width_;
+  size_t height_;
+  size_t pooledWidth_;
+  size_t pooledHeight_;
+  real spatialScale_;
+
+  // Since there is no int matrix, use real maxtrix instead.
+  MatrixPtr maxIdxs_;
+
+public:
+  explicit ROIPoolLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+};
+}  // namespace paddle

paddle/gserver/layers/ScaleSubRegionLayer.cpp

+/* 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 "ScaleSubRegionLayer.h"
+#include "paddle/utils/Stat.h"
+namespace paddle {
+
+REGISTER_LAYER(scale_sub_region, ScaleSubRegionLayer);
+
+bool ScaleSubRegionLayer::init(const LayerMap& layerMap,
+                               const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+  CHECK_EQ(static_cast<int>(inputLayers_.size()), 2);
+  auto& conf = config_.inputs(0).scale_sub_region_conf();
+  value_ = conf.value();
+
+  createFunction(forward_, "ScaleSubRegion", FuncConfig().set("value", value_));
+  createFunction(
+      backward_, "ScaleSubRegionGrad", FuncConfig().set("value", value_));
+
+  return true;
+}
+
+void ScaleSubRegionLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  auto in0 = getInput(0);
+  imgH_ = in0.getFrameHeight();
+  imgW_ = in0.getFrameWidth();
+  if (imgH_ == 0 || imgW_ == 0) {
+    auto& conf = config_.inputs(0).scale_sub_region_conf();
+    imgH_ = conf.image_conf().img_size_y();
+    imgW_ = conf.image_conf().img_size();
+  }
+  MatrixPtr imgV = in0.value;
+  size_t batchSize = imgV->getHeight();
+  size_t spatialSize = imgH_ * imgW_;
+  channelsNum_ = imgV->getWidth() / spatialSize;
+  shape_ = TensorShape({batchSize, channelsNum_, imgH_, imgW_});
+
+  resetOutput(batchSize, imgV->getWidth());
+  auto& out = getOutput();
+  out.setFrameHeight(imgH_);
+  out.setFrameWidth(imgW_);
+
+  MatrixPtr indicesV = getInputValue(1);
+  indicesShape_ = TensorShape({batchSize, 6});
+
+  REGISTER_TIMER_INFO("ScaleSubRegionForward", getName().c_str());
+  BufferArgs inArgs;
+  BufferArgs outArgs;
+  inArgs.addArg(*imgV, shape_);
+  inArgs.addArg(*indicesV, indicesShape_);
+  outArgs.addArg(*out.value, shape_, ASSIGN_TO);
+  forward_[0]->calc(inArgs, outArgs);
+}
+
+void ScaleSubRegionLayer::backward(const UpdateCallback& callback) {
+  REGISTER_TIMER_INFO("ScaleSubRegionBackward", getName().c_str());
+  BufferArgs inArgs;
+  BufferArgs outArgs;
+  inArgs.addArg(*getOutputGrad(), shape_);
+  inArgs.addArg(*getInputValue(1), indicesShape_);
+  outArgs.addArg(*getInputGrad(0), shape_, ADD_TO);
+  backward_[0]->calc(inArgs, outArgs);
+}
+
+}  // namespace paddle

paddle/operators/fill_constant_op.h → paddle/gserver/layers/ScaleSubRegionLayer.h

@@ -13,25 +13,40 @@ 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"
+
+#include "Layer.h"
 
 namespace paddle {
-namespace operators {
-
-template <typename Place, typename T>
-class FillConstantOpKernel : public framework::OpKernel<T> {
- public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    auto* out = ctx.Output<framework::Tensor>("Out");
-    out->mutable_data<T>(ctx.GetPlace());
-    auto value = ctx.Attr<float>("value");
-
-    auto out_eigen = framework::EigenVector<T>::Flatten(*out);
-    auto place = ctx.GetEigenDevice<Place>();
-    out_eigen.device(place) = out_eigen.constant(static_cast<T>(value));
-  }
+
+/**
+ * \brief  For each instance, this layer can be used to multiply a value to a
+ *         specified sub continuous region. By providing start index and end
+ *         index for C/H/W, you can specify the location and shape of the
+ *         region.
+ *
+ *         input_0: Input value.
+ *         input_1: Indices value to specify the location an shape of the
+ *                  region.
+ */
+class ScaleSubRegionLayer : public Layer {
+public:
+  explicit ScaleSubRegionLayer(const LayerConfig& config) : Layer(config) {}
+
+  ~ScaleSubRegionLayer() {}
+
+  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+
+  void forward(PassType passType);
+
+  void backward(const UpdateCallback& callback = nullptr);
+
+protected:
+  TensorShape shape_;
+  TensorShape indicesShape_;
+  size_t imgH_;
+  size_t imgW_;
+  size_t channelsNum_;
+  real value_;
 };
 
-}  // namespace operators
 }  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -53,7 +53,7 @@ TEST(Operator, dot_mul) {
 TEST(Projection, context) {
   for (auto contextStart : {-5, -3, -1, 0, 3}) {
     for (auto contextLength : {1, 2, 5, 7}) {
-      for (auto batchSize : {1, 2, 5, 20, 50}) {
+      for (auto batchSize : {1, 2, 5, 20}) {
         for (auto trainablePadding : {false, true}) {
           LOG(INFO) << " contextStart=" << contextStart
                     << " contextLength=" << contextLength
@@ -585,14 +585,14 @@ TEST(Layer, maxoutLayer) {
 }
 void testFcLayer(string format, size_t nnz) {
   TestConfig config;
-  config.biasSize = 4096;
+  config.biasSize = 1024;
   config.layerConfig.set_type("fc");
-  config.layerConfig.set_size(4096);
+  config.layerConfig.set_size(1024);
   config.layerConfig.set_active_type("sigmoid");
   config.layerConfig.set_drop_rate(0.1);
 
   config.inputDefs.push_back(
-      {INPUT_DATA, "layer_0", 8192, nnz, ParaSparse(format)});
+      {INPUT_DATA, "layer_0", 2048, nnz, ParaSparse(format)});
   config.layerConfig.add_inputs();
 
   LOG(INFO) << config.inputDefs[0].sparse.sparse << " "
@@ -609,9 +609,9 @@ void testFcLayer(string format, size_t nnz) {
 }
 
 TEST(Layer, fcLayer) {
-  testFcLayer("", 4096 * 4096 * 2);
-  testFcLayer("csc", 4096 * 40);
-  testFcLayer("csr", 4096 * 40);
+  testFcLayer("", 1024 * 1024 * 2);
+  testFcLayer("csc", 1024 * 10);
+  testFcLayer("csr", 1024 * 10);
 }
 
 TEST(Layer, SelectiveFullyConnectedLayer) {
@@ -1995,7 +1995,7 @@ TEST(Layer, multibox_loss) {
 TEST(Layer, TransLayer) {
   TestConfig config;
   const int height = 128;
-  const int width = 1028;
+  const int width = 256;
   config.layerConfig.set_type("trans");
   config.layerConfig.set_size(width);
 
@@ -2056,6 +2056,43 @@ TEST(Layer, CropLayer) {
   }
 }
 
+TEST(Layer, roi_pool) {
+  TestConfig config;
+  config.layerConfig.set_type("roi_pool");
+  config.biasSize = 0;
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ROIPoolConfig* roiPoolConf = input->mutable_roi_pool_conf();
+  roiPoolConf->set_pooled_width(7);
+  roiPoolConf->set_pooled_height(7);
+  roiPoolConf->set_spatial_scale(1. / 16);
+  roiPoolConf->set_width(14);
+  roiPoolConf->set_height(14);
+
+  const size_t roiNum = 10;
+  const size_t roiDim = 10;
+  const size_t batchSize = 5;
+  MatrixPtr roiValue = Matrix::create(roiNum, roiDim, false, false);
+  roiValue->zeroMem();
+  real* roiData = roiValue->getData();
+  for (size_t i = 0; i < roiNum; ++i) {
+    roiData[i * roiDim + 0] = std::rand() % batchSize;
+    roiData[i * roiDim + 1] = std::rand() % 224;  // xMin
+    roiData[i * roiDim + 2] = std::rand() % 224;  // yMin
+    size_t xMin = static_cast<size_t>(roiData[i * roiDim + 1]);
+    size_t yMin = static_cast<size_t>(roiData[i * roiDim + 2]);
+    roiData[i * roiDim + 3] = xMin + std::rand() % (224 - xMin);  // xMax
+    roiData[i * roiDim + 4] = yMin + std::rand() % (224 - yMin);  // yMax
+  }
+
+  config.inputDefs.push_back({INPUT_DATA, "input", 3 * 14 * 14, {}});
+  config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "rois", roiValue, {}});
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "roi_pool", batchSize, false, useGpu, false);
+  }
+}
+
 TEST(Layer, SwitchOrderLayer) {
   TestConfig config;
   // config input_0
@@ -2358,6 +2395,38 @@ TEST(Layer, ScaleShiftLayer) {
   }
 }
 
+TEST(Layer, ScaleSubRegionLayer) {
+  const size_t batchSize = 64;
+  const size_t size = 4096;
+  TestConfig config;
+  config.layerConfig.set_type("scale_sub_region");
+  config.inputDefs.push_back({INPUT_DATA, "input", size, 0});
+  MatrixPtr indicesV = Matrix::create(batchSize, 6, false, false);
+  auto* data = indicesV->getData();
+  for (size_t i = 0; i < batchSize; ++i) {
+    data[i * 2] = 2;
+    data[i * 2 + 1] = 4;
+    data[i * 2 + 2] = 16;
+    data[i * 2 + 3] = 32;
+    data[i * 2 + 4] = 16;
+    data[i * 2 + 5] = 32;
+  }
+  config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "indices", indicesV, {}});
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ScaleSubRegionConfig* scaleSubRegionConf =
+      input->mutable_scale_sub_region_conf();
+  ImageConfig* imgConf = scaleSubRegionConf->mutable_image_conf();
+  imgConf->set_img_size(32);
+  imgConf->set_img_size_y(32);
+  imgConf->set_channels(4);
+  scaleSubRegionConf->set_value(2.0);
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "scale_sub_region", batchSize, false, useGpu, false);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/gserver/tests/test_MKLDNN.cpp

@@ -269,6 +269,7 @@ void testBatchNormLayer(const testBatchNormDesc& pm) {
 TEST(MKLDNNLayer, BatchNormLayer) {
   testBatchNormLayer({4, 10, 6, 6});
   testBatchNormLayer({16, 32, 16, 16});
+  testBatchNormLayer({4, 16, 8, 10});
 }
 
 struct testImageDesc {
@@ -300,13 +301,8 @@ void testAddtoLayer(const testImageDesc& pm, const size_t nInputs) {
   TestConfig dnnConfig;
   getAddtoConfig(dnnConfig, pm, nInputs);
   dnnConfig.layerConfig.set_type("mkldnn_addto");
-  // TODO(TJ): test with bias
-  for (auto withBias : {false}) {
-    if (withBias) {
-      dnnConfig.biasSize = pm.ic * pm.ih * pm.iw;
-    } else {
-      dnnConfig.biasSize = 0;
-    }
+  for (auto withBias : {false, true}) {
+    dnnConfig.biasSize = withBias ? pm.ic * pm.ih * pm.iw : 0;
     RUN_MKLDNN_TEST_LAYER(dnnConfig, "addto", pm)
   }
 }

paddle/math/MKLDNNMatrix.cpp

@@ -152,12 +152,7 @@ void MKLDNNMatrix::downSpatial() {
   }
   memory::desc md = memory::desc(dstDims, getDtype(), dstFmt);
   memory::primitive_desc pd = memory::primitive_desc(md, getEngine());
-  mkldnn_primitive_t result;
-  mkldnn::error::wrap_c_api(
-      mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr),
-      "could not create a memory primitive");
-  reset(result);
-  set_data_handle(data_);
+  resetMKLDNNMemory(pd, data_);
 }
 
 }  // namespace paddle

paddle/math/MKLDNNMatrix.h

@@ -145,6 +145,27 @@ public:
     m_.reset();
   }
 
+  /**
+   * override the CpuMatrix::resize
+   */
+  void resize(size_t newHeight, size_t newWidth) override {
+    m_->resize(newHeight, newWidth);
+    if (data_ == m_->getData() && elementCnt_ == newHeight * newWidth) {
+      return;
+    }
+    CpuMatrix::setData(data_);
+    height_ = newHeight;
+    width_ = newWidth;
+    elementCnt_ = newHeight * newWidth;
+    stride_ = width_;
+    auto pd = mkldnn::memory::primitive_desc(
+        mkldnn::memory::desc({(int)newHeight, (int)newWidth},
+                             getDtype(),
+                             mkldnn::memory::format::nc),
+        getEngine());
+    resetMKLDNNMemory(pd, data_);
+  }
+
   /**
    * override Matrix::getData
    * check data before return
@@ -215,6 +236,17 @@ protected:
                    memory::format srcFmt,
                    memory::format dstFmt,
                    memory::dims dm);
+  /**
+   * reset this MKLDNN Memory from primitve desc
+   */
+  void resetMKLDNNMemory(memory::primitive_desc pd, real* data) {
+    mkldnn_primitive_t result;
+    mkldnn::error::wrap_c_api(
+        mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr),
+        "could not create a memory primitive");
+    reset(result);
+    set_data_handle(data);
+  }
 
 private:
   // save the CpuMatrixPtr in case the buffer released outside

paddle/math/MathFunctions.cpp

@@ -206,7 +206,7 @@ double dotProduct<double>(const int n, const double* x, const double* y) {
 }
 #endif
 
-#if defined(PADDLE_USE_MKL) || defined(PADDLE_USE_MKLML)
+#if defined(PADDLE_USE_MKLML)
 
 template <>
 void vExp<float>(const int n, const float* a, float* r) {
@@ -295,38 +295,6 @@ template void vAdd(const int n, const double* a, const double* b, double* r);
 
 #endif
 
-#ifdef PADDLE_USE_MKL
-template <>
-void vInvSqrt<float>(const int n, const float* a, float* r) {
-  vsInvSqrt(n, a, r);
-}
-
-template <>
-void vInvSqrt<double>(const int n, const double* a, double* r) {
-  vdInvSqrt(n, a, r);
-}
-
-template <>
-void vLog1p<float>(const int n, const float* a, float* r) {
-  vsLog1p(n, a, r);
-}
-
-template <>
-void vLog1p<double>(const int n, const double* a, double* r) {
-  vdLog1p(n, a, r);
-}
-
-template <>
-void vTanh<float>(const int n, const float* a, float* r) {
-  vsTanh(n, a, r);
-}
-
-template <>
-void vTanh<double>(const int n, const double* a, double* r) {
-  vdTanh(n, a, r);
-}
-#else
-
 DEFINE_MATRIX_BINARY_OP(vInvSqrt, b = 1.0f / std::sqrt(a));
 template <class T>
 void vInvSqrt(const int n, const T* a, T* r) {
@@ -357,6 +325,4 @@ template void vLog1p(const int n, const double* a, double* r);
 template void vTanh(const int n, const float* a, float* r);
 template void vTanh(const int n, const double* a, double* r);
 
-#endif
-
 }  // namespace paddle

paddle/math/MathFunctions.h

@@ -21,11 +21,6 @@ limitations under the License. */
 #include <mkl_vml_functions.h>
 #endif
 
-#ifdef PADDLE_USE_MKL
-#include <mkl.h>
-#include <mkl_lapacke.h>
-#endif
-
 #if defined(PADDLE_USE_ATLAS) || defined(PADDLE_USE_VECLIB)
 extern "C" {
 #include <cblas.h>

paddle/math/tests/TensorCheck.h

@@ -169,7 +169,7 @@ void TensorCheck(AssertEq compare,
       count++;
     }
   }
-  EXPECT_EQ(count, 0) << "There are " << count << " different element.";
+  EXPECT_EQ(count, 0) << "There are " << count << " different elements.";
 }
 
 template <typename AssertEq, typename Tensor1, typename Tensor2>

paddle/operators/accuracy_op.cu

@@ -65,7 +65,7 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
 
     size_t num_samples = inference->dims()[0];
     size_t infer_width = inference->dims()[1];
-    cudaMemset((void**)&accuracy_data, 0, sizeof(float));
+    PADDLE_ENFORCE(cudaMemset(accuracy_data, 0, sizeof(float)));
 
     if (num_samples == 0) {
       return;

paddle/operators/accuracy_op.h

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #pragma once
 #include <algorithm>
-#include "paddle/framework/eigen.h"
 #include "paddle/framework/op_registry.h"
 
 namespace paddle {
@@ -22,18 +21,6 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
-
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
-
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenScalar = framework::EigenScalar<T, MajorType, IndexType>;
-
 template <typename Place, typename T>
 class AccuracyKernel : public framework::OpKernel<T> {
  public:

paddle/operators/array_operator.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+#include "paddle/framework/lod_tensor_array.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+class ArrayOp : public framework::OperatorBase {
+ public:
+  ArrayOp(const std::string &type, const framework::VariableNameMap &inputs,
+          const framework::VariableNameMap &outputs,
+          const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+
+ protected:
+  size_t GetOffset(const framework::Scope &scope,
+                   const platform::DeviceContext &dev_ctx) const {
+    auto *i = scope.FindVar(Input("I"));
+    PADDLE_ENFORCE(i != nullptr, "I must be set");
+    auto &i_tensor = i->Get<framework::LoDTensor>();
+    PADDLE_ENFORCE_EQ(i_tensor.numel(), 1);
+    size_t offset;
+    if (platform::is_gpu_place(i_tensor.place())) {
+      // FIXME: Avoid copy from GPU to CPU
+      framework::Tensor t;
+      t.CopyFrom(i_tensor, platform::CPUPlace(), dev_ctx);
+      dev_ctx.Wait();
+      offset = static_cast<size_t>(*t.data<int64_t>());
+    } else {
+      offset = static_cast<size_t>(*i_tensor.data<int64_t>());
+    }
+    return offset;
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/array_to_lod_tensor_op.cc

@@ -140,6 +140,23 @@ class ArrayToLoDTensorInferShape : public framework::InferShapeBase {
                    "ArrayToLoDTensorOp must has input X.");
     PADDLE_ENFORCE(context->HasInput("RankTable"),
                    "ArrayToLoDTensorOp must has input RankTable.");
+    context->SetOutputDim("Out", context->GetInputDim("X"));
+  }
+};
+
+class ArrayToLoDTensorGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDescBind> Apply() const override {
+    auto *grad_op = new framework::OpDescBind();
+    grad_op->SetType("lod_tensor_to_array");
+    grad_op->SetInput("X", OutputGrad("Out"));
+    grad_op->SetInput("RankTable", Input("RankTable"));
+    grad_op->SetOutput("Out", InputGrad("X"));
+    grad_op->SetAttrMap(Attrs());
+    return std::unique_ptr<framework::OpDescBind>(grad_op);
   }
 };
 
@@ -149,4 +166,5 @@ class ArrayToLoDTensorInferShape : public framework::InferShapeBase {
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(array_to_lod_tensor, ops::ArrayToLoDTensorOp,
                   ops::ArrayToLoDTensorOpProtoMaker,
-                  ops::ArrayToLoDTensorInferShape);
+                  ops::ArrayToLoDTensorInferShape,
+                  ops::ArrayToLoDTensorGradMaker);

paddle/operators/batch_norm_op.cc

@@ -19,9 +19,6 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
 template <typename T>
 using EigenArrayMap =

paddle/operators/chunk_eval_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/chunk_eval_op.h"
+
+namespace paddle {
+namespace operators {
+
+class ChunkEvalOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Inference"),
+                   "Input(Inference) of ChunkEvalOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Label"),
+                   "Input(Label) of ChunkEvalOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Precision"),
+                   "Output(Precision) of ChunkEvalOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Recall"),
+                   "Output(Recall) of ChunkEvalOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("F1-Score"),
+                   "Output(F1-Score) of ChunkEvalOp should not be null.");
+
+    auto inference_dim = ctx->GetInputDim("Inference");
+    auto label_dim = ctx->GetInputDim("Label");
+
+    PADDLE_ENFORCE(inference_dim == label_dim,
+                   "Inference's shape must be the same as Label's shape.");
+
+    ctx->SetOutputDim("Precision", {1});
+    ctx->SetOutputDim("Recall", {1});
+    ctx->SetOutputDim("F1-Score", {1});
+  }
+
+ protected:
+  framework::OpKernelType GetKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    return framework::OpKernelType(framework::DataType::FP32,
+                                   ctx.device_context());
+  }
+};
+
+class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ChunkEvalOpMaker(framework::OpProto *proto,
+                   framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Inference",
+             "(Tensor, default: Tensor<int>). Predictions from the network.");
+    AddInput("Label",
+             "(Tensor, default: Tensor<int>). The true tag sequences.");
+    AddOutput("Precision",
+              "(float). The evaluated precision (called positive predictive "
+              "value) of chunks on the given mini-batch.");
+    AddOutput("Recall",
+              "(float). The evaluated recall (true positive rate or "
+              "sensitivity) of chunks on the given mini-batch.");
+    AddOutput("F1-Score",
+              "(float). The evaluated F1-Score on the given mini-batch.");
+    AddAttr<int>("num_chunk_types",
+                 "(int). The number of chunk type. See below for details.");
+    AddAttr<std::string>(
+        "chunk_scheme",
+        "(string, default IOB). The labeling scheme indicating "
+        "how to encode the chunks. Must be IOB, IOE, IOBES or plain. See below "
+        "for details.")
+        .SetDefault("IOB");
+    AddAttr<std::vector<int>>("excluded_chunk_types",
+                              "(list<int>) A list including chunk type ids "
+                              "indicating chunk types that are not counted. "
+                              "See below for details.")
+        .SetDefault(std::vector<int>{});
+    AddComment(R"DOC(
+For some basics of chunking, please refer to
+‘Chunking with Support Vector Mechines <https://aclanthology.info/pdf/N/N01/N01-1025.pdf>’.
+
+
+CheckEvalOp computes the precision, recall, and F1-score of chunk detection,
+and supports IOB, IOE, IOBES and IO (also known as plain) tagging schemes.
+Here is a NER example of labeling for these tagging schemes:
+
+ 	     Li     Ming    works  at  Agricultural   Bank   of    China  in  Beijing.
+  IO:    I-PER  I-PER   O      O   I-ORG          I-ORG  I-ORG I-ORG  O   I-LOC
+  IOB:   B-PER  I-PER   O      O   B-ORG          I-ORG  I-ORG I-ORG  O   B-LOC
+  IOE:   I-PER  E-PER   O      O   I-ORG          I-ORG  I-ORG E-ORG  O   E-LOC
+  IOBES: B-PER  E-PER   O      O   I-ORG          I-ORG  I-ORG E-ORG  O   S-LOC
+
+There are three chunk types(named entity types) including PER(person), ORG(orgnazation)
+and LOC(LOCATION), and we can see that the labels have the form <tag type>-<chunk type>.
+
+Since the calculations actually use label ids rather than labels, extra attention
+should be paid when mapping labels to ids to make CheckEvalOp work. The key point
+is that the listed equations are satisfied by ids.
+
+    tag_type = label % num_tag_type
+    chunk_type = label / num_tag_type
+
+where `num_tag_type` is the num of tag types in the tagging scheme, `num_chunk_type`
+is the num of chunk types, and `tag_type` get its value from the following table.
+
+    Scheme Begin Inside End   Single
+     plain   0     -      -     -
+     IOB     0     1      -     -
+     IOE     -     0      1     -
+     IOBES   0     1      2     3
+
+Still use NER as example, assuming the tagging scheme is IOB while chunk types are ORG,
+PER and LOC. To satisfy the above equations, the label map can be like this:
+
+    B-ORG  0
+    I-ORG  1
+    B-PER  2
+    I-PER  3
+    B-LOC  4
+    I-LOC  5
+    O      6
+
+It’s not hard to verify the equations noting that the num of chunk types
+is 3 and the num of tag types in IOB scheme is 2. For example, the label
+id of I-LOC is 5, the tag type id of I-LOC is 1, and the chunk type id of
+I-LOC is 2, which consistent with the results from the equations.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(chunk_eval, ops::ChunkEvalOp,
+                             ops::ChunkEvalOpMaker);
+REGISTER_OP_CPU_KERNEL(chunk_eval,
+                       ops::ChunkEvalKernel<paddle::platform::CPUPlace, float>);

paddle/operators/chunk_eval_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 <set>
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+
+template <typename Place, typename T>
+class ChunkEvalKernel : public framework::OpKernel<T> {
+ public:
+  struct Segment {
+    int begin;
+    int end;
+    int type;
+    bool operator==(const Segment& y) const {
+      return begin == y.begin && end == y.end && type == y.type;
+    }
+  };
+
+  void GetSegments(const int* label, int length, std::vector<Segment>& segments,
+                   int num_chunk_types, int num_tag_types, int other_chunk_type,
+                   int tag_begin, int tag_inside, int tag_end,
+                   int tag_single) const {
+    segments.clear();
+    segments.reserve(length);
+    int chunk_start = 0;
+    bool in_chunk = false;
+    int tag = -1;
+    int type = other_chunk_type;
+    for (int i = 0; i < length; ++i) {
+      int prev_tag = tag;
+      int prev_type = type;
+      PADDLE_ENFORCE_LE(label[i], num_chunk_types * num_tag_types);
+      tag = label[i] % num_tag_types;
+      type = label[i] / num_tag_types;
+      if (in_chunk && ChunkEnd(prev_tag, prev_type, tag, type, other_chunk_type,
+                               tag_begin, tag_inside, tag_end, tag_single)) {
+        Segment segment{
+            chunk_start,  // begin
+            i - 1,        // end
+            prev_type,
+        };
+        segments.push_back(segment);
+        in_chunk = false;
+      }
+      if (ChunkBegin(prev_tag, prev_type, tag, type, other_chunk_type,
+                     tag_begin, tag_inside, tag_end, tag_single)) {
+        chunk_start = i;
+        in_chunk = true;
+      }
+    }
+    if (in_chunk) {
+      Segment segment{
+          chunk_start,  // begin
+          length - 1,   // end
+          type,
+      };
+      segments.push_back(segment);
+    }
+  }
+
+  bool ChunkEnd(int prev_tag, int prev_type, int tag, int type,
+                int other_chunk_type, int tag_begin, int tag_inside,
+                int tag_end, int tag_single) const {
+    if (prev_type == other_chunk_type) return false;
+    if (type == other_chunk_type) return true;
+    if (type != prev_type) return true;
+    if (prev_tag == tag_begin) return tag == tag_begin || tag == tag_single;
+    if (prev_tag == tag_inside) return tag == tag_begin || tag == tag_single;
+    if (prev_tag == tag_end) return true;
+    if (prev_tag == tag_single) return true;
+    return false;
+  }
+
+  bool ChunkBegin(int prev_tag, int prev_type, int tag, int type,
+                  int other_chunk_type, int tag_begin, int tag_inside,
+                  int tag_end, int tag_single) const {
+    if (prev_type == other_chunk_type) return type != other_chunk_type;
+    if (type == other_chunk_type) return false;
+    if (type != prev_type) return true;
+    if (tag == tag_begin) return true;
+    if (tag == tag_inside) return prev_tag == tag_end || prev_tag == tag_single;
+    if (tag == tag_end) return prev_tag == tag_end || prev_tag == tag_single;
+    if (tag == tag_single) return true;
+    return false;
+  }
+
+  void Compute(const framework::ExecutionContext& context) const override {
+    // initialize to parse configurations
+    int num_chunk_types, num_tag_types;
+    int other_chunk_type;
+    int tag_begin, tag_inside, tag_end, tag_single;
+    std::vector<Segment> label_segments;
+    std::vector<Segment> output_segments;
+    std::set<int> excluded_chunk_types;
+    int64_t num_output_segments = 0;
+    int64_t num_label_segments = 0;
+    int64_t num_correct = 0;
+    if (context.Attr<std::string>("chunk_scheme") == "IOB") {
+      num_tag_types = 2;
+      tag_begin = 0;
+      tag_inside = 1;
+      tag_end = -1;
+      tag_single = -1;
+    } else if (context.Attr<std::string>("chunk_scheme") == "IOE") {
+      num_tag_types = 2;
+      tag_begin = -1;
+      tag_inside = 0;
+      tag_end = 1;
+      tag_single = -1;
+    } else if (context.Attr<std::string>("chunk_scheme") == "IOBES") {
+      num_tag_types = 4;
+      tag_begin = 0;
+      tag_inside = 1;
+      tag_end = 2;
+      tag_single = 3;
+    } else if (context.Attr<std::string>("chunk_scheme") == "plain") {
+      num_tag_types = 1;
+      tag_begin = -1;
+      tag_inside = -1;
+      tag_end = -1;
+      tag_single = -1;
+    } else {
+      PADDLE_THROW("Unknown chunk scheme.");
+    }
+    other_chunk_type = num_chunk_types = context.Attr<int>("num_chunk_types");
+    excluded_chunk_types.insert(
+        context.Attr<std::vector<int>>("excluded_chunk_types").begin(),
+        context.Attr<std::vector<int>>("excluded_chunk_types").end());
+
+    auto* inference = context.Input<LoDTensor>("Inference");
+    auto* label = context.Input<LoDTensor>("Label");
+    auto* precision = context.Output<Tensor>("Precision");
+    auto* recall = context.Output<Tensor>("Recall");
+    auto* f1 = context.Output<Tensor>("F1-Score");
+
+    const int* inference_data = inference->data<int>();
+    const int* label_data = label->data<int>();
+    T* precision_data = precision->mutable_data<T>(context.GetPlace());
+    T* racall_data = recall->mutable_data<T>(context.GetPlace());
+    T* f1_data = f1->mutable_data<T>(context.GetPlace());
+
+    auto lod = label->lod();
+    PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now.");
+    PADDLE_ENFORCE(lod == inference->lod(),
+                   "LoD must be same between Inference and Label.");
+    int num_sequences = lod[0].size() - 1;
+    for (int i = 0; i < num_sequences; ++i) {
+      int seq_length = lod[0][i + 1] - lod[0][i];
+      EvalOneSeq(inference_data + lod[0][i], label_data + lod[0][i], seq_length,
+                 output_segments, label_segments, num_output_segments,
+                 num_label_segments, num_correct, num_chunk_types,
+                 num_tag_types, other_chunk_type, tag_begin, tag_inside,
+                 tag_end, tag_single, excluded_chunk_types);
+    }
+    *precision_data = !num_output_segments ? 0 : static_cast<T>(num_correct) /
+                                                     num_output_segments;
+    *racall_data = !num_label_segments ? 0 : static_cast<T>(num_correct) /
+                                                 num_label_segments;
+    *f1_data = !num_correct ? 0 : 2 * (*precision_data) * (*racall_data) /
+                                      ((*precision_data) + (*racall_data));
+  }
+
+  void EvalOneSeq(const int* output, const int* label, int length,
+                  std::vector<Segment>& output_segments,
+                  std::vector<Segment>& label_segments,
+                  int64_t& num_output_segments, int64_t& num_label_segments,
+                  int64_t& num_correct, int num_chunk_types, int num_tag_types,
+                  int other_chunk_type, int tag_begin, int tag_inside,
+                  int tag_end, int tag_single,
+                  const std::set<int>& excluded_chunk_types) const {
+    GetSegments(output, length, output_segments, num_chunk_types, num_tag_types,
+                other_chunk_type, tag_begin, tag_inside, tag_end, tag_single);
+    GetSegments(label, length, label_segments, num_chunk_types, num_tag_types,
+                other_chunk_type, tag_begin, tag_inside, tag_end, tag_single);
+    size_t i = 0, j = 0;
+    while (i < output_segments.size() && j < label_segments.size()) {
+      if (output_segments[i] == label_segments[j] &&
+          excluded_chunk_types.count(output_segments[i].type) != 1) {
+        ++num_correct;
+      }
+      if (output_segments[i].end < label_segments[j].end) {
+        ++i;
+      } else if (output_segments[i].end > label_segments[j].end) {
+        ++j;
+      } else {
+        ++i;
+        ++j;
+      }
+    }
+    for (auto& segment : label_segments) {
+      if (excluded_chunk_types.count(segment.type) != 1) ++num_label_segments;
+    }
+    for (auto& segment : output_segments) {
+      if (excluded_chunk_types.count(segment.type) != 1) ++num_output_segments;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/clip_by_norm_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/clip_by_norm_op.h"
+
+namespace paddle {
+namespace operators {
+
+class ClipByNormOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of ClipByNormOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of ClipByNormOp should not be null.");
+    auto max_norm = ctx->Attrs().Get<float>("max_norm");
+    PADDLE_ENFORCE_GT(max_norm, 0, "max_norm should be greater than 0.");
+    auto x_dims = ctx->GetInputDim("X");
+    ctx->SetOutputDim("Out", x_dims);
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+};
+
+class ClipByNormOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ClipByNormOpMaker(framework::OpProto* proto,
+                    framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(Tensor) The input of clip_by_norm op."
+             "The number of dimensions must be between [1, 9].");
+    AddOutput("Out",
+              "(Tensor) The output of clip_by_norm op with shape as input(X)");
+    AddAttr<float>("max_norm", "(float) The maximum norm value.");
+    AddComment(R"DOC(
+ClipByNorm operator limits the L2 norm of the input 'X' within 'max_norm'. 
+If the L2 norm of 'X' is less than or equal to 'max_norm', 'Out' will be 
+the same as 'X'. If the L2 norm of 'X' is greater than 'max_norm', 'X' will 
+be linearly scaled to make the L2 norm of 'Out' equal to 'max_norm', as 
+shown in the following formula：
+
+'Out' = 'max_norm' * 'X' / norm('X'),
+
+where norm('X') represents the L2 norm of 'X'.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(clip_by_norm, ops::ClipByNormOp,
+                             ops::ClipByNormOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    clip_by_norm, ops::ClipByNormKernel<paddle::platform::CPUPlace, float>);

paddle/operators/increment_op.cu → paddle/operators/clip_by_norm_op.cu

@@ -12,11 +12,8 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#include "paddle/operators/increment_op.h"
+#include "paddle/operators/clip_by_norm_op.h"
 
+namespace ops = paddle::operators;
 REGISTER_OP_GPU_KERNEL(
-    increment,
-    paddle::operators::IncrementKernel<paddle::platform::GPUPlace, float>,
-    paddle::operators::IncrementKernel<paddle::platform::GPUPlace, double>,
-    paddle::operators::IncrementKernel<paddle::platform::GPUPlace, int>,
-    paddle::operators::IncrementKernel<paddle::platform::GPUPlace, int64_t>);
+    clip_by_norm, ops::ClipByNormKernel<paddle::platform::GPUPlace, float>);

paddle/operators/increment_op.h → paddle/operators/clip_by_norm_op.h

@@ -16,23 +16,35 @@
 
 #include "paddle/framework/eigen.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/platform/transform.h"
 
 namespace paddle {
 namespace operators {
-template <typename Place, typename T>
-class IncrementKernel : public framework::OpKernel<T> {
- public:
-  virtual void Compute(const framework::ExecutionContext& context) const {
-    auto* tensor = context.Output<framework::Tensor>("Out");
-    auto* in = context.Input<framework::Tensor>("X");
-    tensor->mutable_data<T>(in->place());
 
-    auto step = static_cast<T>(context.Attr<float>("step"));
+using Tensor = framework::Tensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
-    auto eigen_out = framework::EigenVector<T>::Flatten(*tensor);
-    auto eigen_in = framework::EigenVector<T>::Flatten(*in);
-    auto& place = context.GetEigenDevice<Place>();
-    eigen_out.device(place) = eigen_in + step;
+template <typename Place, typename T>
+class ClipByNormKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto max_norm = context.Attr<T>("max_norm");
+    auto* input = context.Input<Tensor>("X");
+    auto* output = context.Output<Tensor>("Out");
+    output->mutable_data<T>(context.GetPlace());
+
+    auto x = EigenVector<T>::Flatten(*input);
+    auto out = EigenVector<T>::Flatten(*output);
+    auto x_norm = x.square().sum().sqrt();
+    auto place = context.GetEigenDevice<Place>();
+
+    auto temp = (x_norm <= max_norm).template cast<T>().eval();
+    auto scaling = temp + (static_cast<T>(1) - temp) * max_norm / x_norm;
+    Eigen::array<int, 1> one_dim{{1}};
+    Eigen::DSizes<int, 1> m_dsize(input->numel());
+    out.device(place) = x * scaling.reshape(one_dim).broadcast(m_dsize);
   }
 };
 

paddle/operators/compare_op.cc

@@ -14,6 +14,7 @@
 
 #include "paddle/operators/compare_op.h"
 #include "paddle/framework/op_registry.h"
+
 namespace paddle {
 namespace operators {
 template <typename OpComment>
@@ -61,19 +62,34 @@ class CompareOpInferShape : public framework::InferShapeBase {
   }
 };
 
+class CompareOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  framework::OpKernelType GetKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    framework::OpKernelType kt = OperatorWithKernel::GetKernelType(ctx);
+    // CompareOp kernel's device type is decided by input tensor place
+    kt.place_ = ctx.Input<framework::LoDTensor>("X")->place();
+    return kt;
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
-#define REGISTER_LOGICAL_OP(op_type, _equation)                               \
-  struct _##op_type##Comment {                                                \
-    static char type[];                                                       \
-    static char equation[];                                                   \
-  };                                                                          \
-  char _##op_type##Comment::type[]{#op_type};                                 \
-  char _##op_type##Comment::equation[]{_equation};                            \
-  REGISTER_OP_WITH_KERNEL(                                                    \
-      op_type, ::paddle::operators::CompareOpProtoMaker<_##op_type##Comment>, \
-      ::paddle::operators::CompareOpInferShape<_##op_type##Comment>,          \
+#define REGISTER_LOGICAL_OP(op_type, _equation)                      \
+  struct _##op_type##Comment {                                       \
+    static char type[];                                              \
+    static char equation[];                                          \
+  };                                                                 \
+  char _##op_type##Comment::type[]{#op_type};                        \
+  char _##op_type##Comment::equation[]{_equation};                   \
+  REGISTER_OPERATOR(                                                 \
+      op_type, ::paddle::operators::CompareOp,                       \
+      ::paddle::operators::CompareOpProtoMaker<_##op_type##Comment>, \
+      ::paddle::operators::CompareOpInferShape<_##op_type##Comment>, \
       ::paddle::framework::EmptyGradOpMaker);
 
 REGISTER_LOGICAL_OP(less_than, "Out = X < Y");

paddle/operators/expand_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/expand_op.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+
+class ExpandOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null.");
+
+    std::vector<int> expand_times =
+        ctx->Attrs().Get<std::vector<int>>("expand_times");
+    auto x_dims = ctx->GetInputDim("X");
+
+    PADDLE_ENFORCE_EQ(static_cast<size_t>(x_dims.size()), expand_times.size(),
+                      "The number of Attr(expand_times)'s value must be equal "
+                      "to the rank of Input(X).");
+    PADDLE_ENFORCE_LE(x_dims.size(), 6,
+                      "The rank of Input(X) must not be greater than 6.");
+
+    std::vector<int64_t> out_shape(x_dims.size());
+    for (size_t i = 0; i < expand_times.size(); ++i) {
+      PADDLE_ENFORCE_GE(expand_times[i], 1,
+                        "Each value of Attr(expand_times) should not be "
+                        "less than 1.");
+      out_shape[i] = x_dims[i] * expand_times[i];
+    }
+
+    ctx->SetOutputDim("Out", framework::make_ddim(out_shape));
+    if (out_shape[0] == x_dims[0]) {
+      ctx->ShareLoD("X", "Out");
+    }
+  }
+};
+
+class ExpandOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ExpandOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(Tensor, default Tensor<float>) A tensor with rank in [1, 6]."
+             "X is the input tensor to be expanded.");
+    AddOutput("Out",
+              "(Tensor, default Tensor<float>) A tensor with rank in [1, 6]."
+              "The rank of Output(Out) is same as Input(X) except that each "
+              "dimension size of Output(Out) is equal to corresponding "
+              "dimension size of Input(X) multiplying corresponding value of "
+              "Attr(expand_times).");
+    AddAttr<std::vector<int>>("expand_times",
+                              "Expand times number for each dimension.");
+    AddComment(R"DOC(
+Expand operator tiles the input by given times number. You should set times
+number for each dimension by providing attribute 'expand_times'. The rank of X
+should be in [1, 6]. Please notice that size of 'expand_times' must be same with
+X's rank. Following is a using case:
+
+Input(X) is a 3-D tensor with shape [2, 3, 1]:
+
+        [
+           [[1], [2], [3]],
+           [[4], [5], [6]]
+        ]
+
+Attr(expand_times):  [1, 2, 2]
+
+Output(Out) is a 3-D tensor with shape [2, 6, 2]:
+
+        [
+            [[1, 1], [2, 2], [3, 3], [1, 1], [2, 2], [3, 3]],
+            [[4, 4], [5, 5], [6, 6], [4, 4], [5, 5], [6, 6]]
+        ]
+
+)DOC");
+  }
+};
+
+class ExpandGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Input(Out@GRAD) should not be null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    std::vector<int> expand_times =
+        ctx->Attrs().Get<std::vector<int>>("expand_times");
+    auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
+
+    for (size_t i = 0; i < expand_times.size(); ++i) {
+      PADDLE_ENFORCE_EQ(x_dims[i] * expand_times[i], out_dims[i],
+                        "Each dimension size of Input(Out@GRAD) should be "
+                        "equal to multiplication of crroresponding dimension "
+                        "size of Input(X) and Attr(expand_times) value.");
+    }
+
+    auto x_grad_name = framework::GradVarName("X");
+
+    if (ctx->HasOutput(x_grad_name)) {
+      ctx->SetOutputDim(x_grad_name, x_dims);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(expand, ops::ExpandOp, ops::ExpandOpMaker, expand_grad,
+            ops::ExpandGradOp);
+REGISTER_OP_CPU_KERNEL(expand,
+                       ops::ExpandKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    expand_grad, ops::ExpandGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/fill_constant_op.cu → paddle/operators/expand_op.cu

@@ -13,12 +13,11 @@
    limitations under the License. */
 
 #define EIGEN_USE_GPU
-#include "paddle/framework/op_registry.h"
-#include "paddle/operators/fill_constant_op.h"
+
+#include "paddle/operators/expand_op.h"
 
 namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(expand,
+                       ops::ExpandKernel<paddle::platform::GPUPlace, float>);
 REGISTER_OP_GPU_KERNEL(
-    fill_constant, ops::FillConstantOpKernel<paddle::platform::GPUPlace, float>,
-    ops::FillConstantOpKernel<paddle::platform::GPUPlace, double>,
-    ops::FillConstantOpKernel<paddle::platform::GPUPlace, int>,
-    ops::FillConstantOpKernel<paddle::platform::GPUPlace, int64_t>);
+    expand_grad, ops::ExpandGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/expand_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 <boost/preprocessor/arithmetic/div.hpp>
+#include <boost/preprocessor/arithmetic/mod.hpp>
+#include <boost/preprocessor/comparison/greater.hpp>
+#include <boost/preprocessor/comparison/greater_equal.hpp>
+#include <boost/preprocessor/control/if.hpp>
+#include <boost/preprocessor/repetition/repeat.hpp>
+#include <iostream>
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/framework/operator.h"
+
+#define MAX_RANK_SUPPORTED 6
+
+#define EXPAND_TEMPLATE(z, n, data) \
+  case n + 1: {                     \
+    Expand<n + 1>(context);         \
+    break;                          \
+  }
+#define REP_EXPAND_TEMPLATE(n) BOOST_PP_REPEAT(n, EXPAND_TEMPLATE, ~)
+#define COND(n)                                               \
+  BOOST_PP_GREATER_EQUAL(BOOST_PP_DIV(n, MAX_RANK_SUPPORTED), \
+                         BOOST_PP_MOD(n, MAX_RANK_SUPPORTED))
+#define EXPAND_GRAD_CASE(n)                                        \
+  case n: {                                                        \
+    ExpandBackward<n>(context, reshape_dims_vec, reduce_dims_vec); \
+    break;                                                         \
+  }
+#define EXPAND_GRAD_TEMPLATE(z, n, data) \
+  BOOST_PP_IF(COND(n), EXPAND_GRAD_CASE(n), )
+#define REP_EXPAND_GRAD_TEMPLATE(n) BOOST_PP_REPEAT(n, EXPAND_GRAD_TEMPLATE, ~)
+
+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 T, size_t D, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class ExpandKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto rank = context.Input<Tensor>("X")->dims().size();
+    switch (rank) {
+      REP_EXPAND_TEMPLATE(MAX_RANK_SUPPORTED)
+      default:
+        PADDLE_ENFORCE(false,
+                       "Only support tensor with rank being between 1 and 6.");
+    }
+  }
+
+ protected:
+  template <int Rank>
+  void Expand(const framework::ExecutionContext& context) const {
+    auto* in0 = context.Input<Tensor>("X");
+    auto& expand_times = context.Attr<std::vector<int>>("expand_times");
+    auto* out0 = context.Output<Tensor>("Out");
+    Eigen::DSizes<int, Rank> bcast_dims;
+    auto x_dims = in0->dims();
+    for (size_t i = 0; i < expand_times.size(); ++i) {
+      bcast_dims[i] = expand_times[i];
+    }
+    auto x = EigenTensor<T, Rank>::From(*in0);
+    out0->mutable_data<T>(context.GetPlace());
+    auto y = EigenTensor<T, Rank>::From(*out0);
+    auto place = context.GetEigenDevice<Place>();
+    y.device(place) = x.broadcast(bcast_dims);
+  }
+};
+
+template <typename Place, typename T>
+class ExpandGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* in0 = context.Input<Tensor>("X");
+    auto& expand_times = context.Attr<std::vector<int>>("expand_times");
+    auto x_dims = in0->dims();
+    // 1. reshape_dims_vec is the broadcast parameter. For each dimension i,
+    //    if expand_times[i] > 1 and x_dims[i] > 1, i will be splitted to two
+    //    dimensions [expand_times[i], x_dims[i]].
+    // 2. reduce_dims_vec is the dimension parameter to compute gradients. For
+    //    each dimension expanded, the gradients should be summed to original
+    //    size.
+    std::vector<int> reshape_dims_vec;
+    std::vector<int> reduce_dims_vec;
+    for (size_t i = 0; i < expand_times.size(); ++i) {
+      if (expand_times[i] == 1) {
+        reshape_dims_vec.push_back(x_dims[i]);
+      } else {
+        if (x_dims[i] == 1) {
+          reduce_dims_vec.push_back(reshape_dims_vec.size());
+          reshape_dims_vec.push_back(expand_times[i]);
+        } else {
+          reduce_dims_vec.push_back(reshape_dims_vec.size());
+          reshape_dims_vec.push_back(expand_times[i]);
+          reshape_dims_vec.push_back(x_dims[i]);
+        }
+      }
+    }
+
+    int dims = reshape_dims_vec.size() * MAX_RANK_SUPPORTED +
+               reduce_dims_vec.size() - MAX_RANK_SUPPORTED - 1;
+    // no need reduce, just copy
+    if (reduce_dims_vec.size() == 0) {
+      auto* in0 = context.Input<Tensor>(framework::GradVarName("Out"));
+      auto* out0 = context.Output<Tensor>(framework::GradVarName("X"));
+      out0->mutable_data<T>(context.GetPlace());
+      out0->CopyFrom(*in0, context.GetPlace(), context.device_context());
+    } else {
+      switch (dims) {
+        REP_EXPAND_GRAD_TEMPLATE(72)
+        default:
+          PADDLE_ENFORCE(
+              false, "Only support tensor with rank being between 1 and 6.");
+      }
+    }
+  }
+
+ protected:
+  template <int Dims>
+  void ExpandBackward(const framework::ExecutionContext& context,
+                      const std::vector<int>& reshape_dims_vec,
+                      const std::vector<int>& reduce_dims_vec) const {
+    size_t reshape_size = Dims / MAX_RANK_SUPPORTED + 1;
+    size_t reduce_size = Dims % MAX_RANK_SUPPORTED + 1;
+    PADDLE_ENFORCE_EQ(reshape_size, reshape_dims_vec.size(),
+                      "Inconsistent size between template Dims and "
+                      "reshape dimensions.");
+    PADDLE_ENFORCE_EQ(reduce_size, reduce_dims_vec.size(),
+                      "Inconsistent size between template Dims and "
+                      "reduce dimensions.");
+    auto* in0 = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto* out0 = context.Output<Tensor>(framework::GradVarName("X"));
+    auto x = EigenVector<T>::Flatten(*(context.Input<Tensor>("X")));
+    out0->mutable_data<T>(context.GetPlace());
+    auto x_grad = EigenVector<T>::Flatten(*out0);
+    Eigen::DSizes<int, Dims / MAX_RANK_SUPPORTED + 1> reshape_dims;
+    for (size_t i = 0; i < reshape_size; ++i) {
+      reshape_dims[i] = reshape_dims_vec[i];
+    }
+    Eigen::DSizes<int, Dims % MAX_RANK_SUPPORTED + 1> reduce_dims;
+    for (size_t i = 0; i < reduce_size; ++i) {
+      reduce_dims[i] = reduce_dims_vec[i];
+    }
+    auto out_grad = EigenVector<T>::Flatten(*in0);
+    x_grad.device(context.GetEigenDevice<Place>()) =
+        out_grad.reshape(reshape_dims).sum(reduce_dims).reshape(x.dimensions());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/fill_constant_batch_size_like_op.cc

@@ -75,10 +75,10 @@ class FillConstantBatchSizeLikeOpMaker
               "with the specified value");
     AddAttr<std::vector<int>>("shape", "(vector<int>) The shape of the output");
     AddAttr<int>("input_dim_idx",
-                 "(int, default 0) the index of input's batch size dimension")
+                 "(int, default 0) The index of input's batch size dimension")
         .SetDefault(0);
     AddAttr<int>("output_dim_idx",
-                 "(int, default 0) the index of output's batch size dimension")
+                 "(int, default 0) The index of output's batch size dimension")
         .SetDefault(0);
     AddAttr<float>("value", "(float, default 0) The value to be filled")
         .SetDefault(0.0f);

paddle/operators/fill_constant_batch_size_like_op.cu

@@ -12,7 +12,6 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#define EIGEN_USE_GPU
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/fill_constant_batch_size_like_op.h"
 

paddle/operators/fill_constant_batch_size_like_op.h

@@ -13,8 +13,8 @@ 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"
+#include "paddle/operators/math/math_function.h"
 
 namespace paddle {
 namespace operators {
@@ -27,9 +27,8 @@ class FillConstantBatchSizeLikeOpKernel : public framework::OpKernel<T> {
     out->mutable_data<T>(ctx.GetPlace());
     auto value = ctx.Attr<float>("value");
 
-    auto out_eigen = framework::EigenVector<T>::Flatten(*out);
-    auto place = ctx.GetEigenDevice<Place>();
-    out_eigen.device(place) = out_eigen.constant(static_cast<T>(value));
+    math::SetConstant<Place, T> setter;
+    setter(ctx.device_context(), out, static_cast<T>(value));
   }
 };
 

paddle/operators/fill_constant_op.cc

@@ -12,33 +12,41 @@ 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/fill_constant_op.h"
+#include "paddle/framework/data_type.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/math_function.h"
 
 namespace paddle {
 namespace operators {
 
-class FillConstantOp : public framework::OperatorWithKernel {
+class FillConstantInferShape : public framework::InferShapeBase {
  public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
-  void InferShape(framework::InferShapeContext *ctx) const override {
+  void operator()(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of FillConstantOp should not be null.");
     auto &shape = ctx->Attrs().Get<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 dims = framework::make_ddim(shape_int64);
-    ctx->SetOutputDim("Out", dims);
+    ctx->SetOutputDim("Out", framework::make_ddim(shape));
   }
+};
 
- protected:
-  framework::OpKernelType GetKernelType(
-      const framework::ExecutionContext &ctx) const override {
-    int data_type = ctx.Attr<int>("data_type");
-    VLOG(10) << " FillConstant data_type = " << data_type;
-    return framework::OpKernelType(static_cast<framework::DataType>(data_type),
-                                   ctx.device_context());
+class FillConstantOp : public framework::OperatorBase {
+ public:
+  using framework::OperatorBase::OperatorBase;
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    auto data_type = static_cast<framework::DataType>(Attr<int>("data_type"));
+    auto value = Attr<float>("value");
+    auto force_cpu = Attr<bool>("force_cpu");
+    auto &out =
+        *scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();
+    out.Resize(framework::make_ddim(Attr<std::vector<int>>("shape")));
+    if (force_cpu) {
+      auto cpu = platform::CPUPlace();
+      out.mutable_data(cpu, framework::ToTypeIndex(data_type));
+    } else {
+      out.mutable_data(dev_ctx.GetPlace(), framework::ToTypeIndex(data_type));
+    }
+    math::set_constant(dev_ctx, &out, value);
   }
 };
 
@@ -54,6 +62,11 @@ class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<std::vector<int>>("shape", "(vector<int>) The shape of the output");
     AddAttr<float>("value", "(float, default 0) The value to be filled")
         .SetDefault(0.0f);
+    AddAttr<bool>("force_cpu",
+                  "(bool, default false) Force fill output variable to cpu "
+                  "memory. Otherwise, fill output variable to the running "
+                  "device")
+        .SetDefault(false);
     AddOutput("Out",
               "(Tensor) Tensor of specified shape will be filled "
               "with the specified value");
@@ -69,10 +82,6 @@ Fill up a variable with specified constant value.
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(fill_constant, ops::FillConstantOp,
-                             ops::FillConstantOpMaker);
-REGISTER_OP_CPU_KERNEL(
-    fill_constant, ops::FillConstantOpKernel<paddle::platform::CPUPlace, float>,
-    ops::FillConstantOpKernel<paddle::platform::CPUPlace, double>,
-    ops::FillConstantOpKernel<paddle::platform::CPUPlace, int>,
-    ops::FillConstantOpKernel<paddle::platform::CPUPlace, int64_t>);
+REGISTER_OPERATOR(fill_constant, ops::FillConstantOp,
+                  ops::FillConstantInferShape, ops::FillConstantOpMaker,
+                  paddle::framework::EmptyGradOpMaker);

paddle/operators/fill_zeros_like_op.cu

@@ -12,7 +12,6 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#define EIGEN_USE_GPU
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/fill_zeros_like_op.h"
 

paddle/operators/fill_zeros_like_op.h

@@ -13,8 +13,8 @@ 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"
+#include "paddle/operators/math/math_function.h"
 
 namespace paddle {
 namespace operators {
@@ -23,10 +23,11 @@ template <typename Place, typename T>
 class FillZerosLikeKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* output = context.Output<framework::Tensor>("Y");
-    output->mutable_data<T>(context.GetPlace());
-    auto t = framework::EigenVector<T>::Flatten(*output);
-    t.device(context.GetEigenDevice<Place>()) = t.constant(static_cast<T>(0));
+    auto* out = context.Output<framework::Tensor>("Y");
+    out->mutable_data<T>(context.GetPlace());
+
+    math::SetConstant<Place, T> setter;
+    setter(context.device_context(), out, static_cast<T>(0));
   }
 };
 

paddle/operators/increment_op.cc

@@ -12,22 +12,57 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#include "paddle/operators/increment_op.h"
+#include "paddle/framework/op_registry.h"
 
 namespace paddle {
 namespace operators {
 
-class IncrementOp : public framework::OperatorWithKernel {
+class IncrementInferShape : public framework::InferShapeBase {
  public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
-  void InferShape(framework::InferShapeContext *ctx) const override {
+  void operator()(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of IncrementOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of IncrementOp should not be null.");
+    PADDLE_ENFORCE_EQ(1, framework::product(ctx->GetInputDim("X")));
     ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
-    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+};
+
+struct IncrementFunctor {
+  IncrementFunctor(const framework::LoDTensor &x, framework::LoDTensor *out,
+                   float value)
+      : x_(x), out_(out), value_(value) {}
+
+  template <typename T>
+  void operator()() const {
+    *out_->data<T>() = *x_.data<T>() + static_cast<T>(value_);
+  }
+
+  const framework::LoDTensor &x_;
+  framework::LoDTensor *out_;
+  float value_;
+};
+
+class IncrementOp : public framework::OperatorBase {
+ public:
+  IncrementOp(const std::string &type, const framework::VariableNameMap &inputs,
+              const framework::VariableNameMap &outputs,
+              const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    auto &x = scope.FindVar(Input("X"))->Get<framework::LoDTensor>();
+    auto &out =
+        *scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();
+
+    PADDLE_ENFORCE(platform::is_cpu_place(x.place()));
+    out.Resize(x.dims());
+    out.mutable_data(x.place(), x.type());
+    float value = Attr<float>("step");
+    framework::VisitDataType(framework::ToDataType(out.type()),
+                             IncrementFunctor(x, &out, value));
   }
 };
 
@@ -59,10 +94,10 @@ class IncrementGradOpMaker : public framework::SingleGradOpDescMaker {
 
   std::unique_ptr<framework::OpDescBind> Apply() const override {
     auto *grad_op = new framework::OpDescBind();
-    grad_op->SetType("scale");
-    grad_op->SetInput("X", OutputGrad("Out"));
-    grad_op->SetOutput("Out", InputGrad("X"));
-    grad_op->SetAttr("scale", 1.0f);
+    grad_op->SetType("increment");
+    grad_op->SetInput("X", Output("Out"));
+    grad_op->SetOutput("Out", Input("X"));
+    grad_op->SetAttr("step", -boost::get<float>(GetAttr("step")));
     return std::unique_ptr<framework::OpDescBind>(grad_op);
   }
 };
@@ -71,11 +106,5 @@ class IncrementGradOpMaker : public framework::SingleGradOpDescMaker {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-
-REGISTER_OPERATOR(increment, ops::IncrementOp, ops::IncrementOpMaker,
-                  ops::IncrementGradOpMaker);
-REGISTER_OP_CPU_KERNEL(
-    increment, ops::IncrementKernel<paddle::platform::CPUPlace, float>,
-    ops::IncrementKernel<paddle::platform::CPUPlace, double>,
-    ops::IncrementKernel<paddle::platform::CPUPlace, int>,
-    ops::IncrementKernel<paddle::platform::CPUPlace, int64_t>);
+REGISTER_OPERATOR(increment, ops::IncrementOp, ops::IncrementInferShape,
+                  ops::IncrementOpMaker, ops::IncrementGradOpMaker);

paddle/operators/lod_array_length_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/framework/lod_tensor_array.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+class LoDArrayLengthOp : public framework::OperatorBase {
+ public:
+  LoDArrayLengthOp(const std::string &type,
+                   const framework::VariableNameMap &inputs,
+                   const framework::VariableNameMap &outputs,
+                   const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    auto &x = scope.FindVar(Input("X"))->Get<framework::LoDTensorArray>();
+    auto &out =
+        *scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();
+    out.Resize({1});
+    auto cpu = platform::CPUPlace();
+    *out.mutable_data<int64_t>(cpu) = static_cast<int64_t>(x.size());
+  }
+};
+
+class LoDArrayLengthProtoMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  LoDArrayLengthProtoMaker(framework::OpProto *proto,
+                           framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "(LoDTensorArray) The input tensor array.");
+    AddOutput("Out", "(Tensor) 1x1 CPU Tensor of length, int64_t");
+    AddComment(R"DOC(Get the length of lod tensor array
+
+Out = len(X)
+
+NOTE: The output is a CPU Tensor since the control variable should be only in
+CPU and the length of LoDTensorArray should be used as control variables.
+)DOC");
+  }
+};
+
+class LoDArrayLengthInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *context) const override {
+    PADDLE_ENFORCE(context->HasInput("X"));
+    PADDLE_ENFORCE(context->HasOutput("Out"));
+    context->SetOutputDim("Out", {1});
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(lod_array_length, ops::LoDArrayLengthOp,
+                  ops::LoDArrayLengthInferShape, ops::LoDArrayLengthProtoMaker,
+                  paddle::framework::EmptyGradOpMaker);

paddle/operators/lod_rank_table_op.cc

@@ -66,7 +66,8 @@ class LoDRankTableInferVarType : public framework::VarTypeInference {
   void operator()(const framework::OpDescBind &op_desc,
                   framework::BlockDescBind *block) const override {
     for (auto &o : op_desc.Output("Out")) {
-      block->Var(o)->SetType(framework::VarDesc::LOD_RANK_TABLE);
+      block->FindRecursiveOrCreateVar(o)->SetType(
+          framework::VarDesc::LOD_RANK_TABLE);
     }
   }
 };

paddle/operators/lod_tensor_to_array_op.cc

@@ -133,6 +133,22 @@ class LoDTensorToArrayInferVarType : public framework::VarTypeInference {
   }
 };
 
+class LoDTensorToArrayGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDescBind> Apply() const override {
+    auto *grad_op = new framework::OpDescBind();
+    grad_op->SetType("array_to_lod_tensor");
+    grad_op->SetInput("X", OutputGrad("Out"));
+    grad_op->SetInput("RankTable", Input("RankTable"));
+    grad_op->SetOutput("Out", InputGrad("X"));
+    grad_op->SetAttrMap(Attrs());
+    return std::unique_ptr<framework::OpDescBind>(grad_op);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -140,4 +156,5 @@ namespace ops = paddle::operators;
 REGISTER_OPERATOR(lod_tensor_to_array, ops::LoDTensorToArrayOp,
                   ops::LoDTensorToArrayOpProtoMaker,
                   ops::LoDTensorToArrayInferShape,
-                  ops::LoDTensorToArrayInferVarType);
+                  ops::LoDTensorToArrayInferVarType,
+                  ops::LoDTensorToArrayGradMaker);

paddle/operators/lstm_op.cc

@@ -24,6 +24,11 @@ class LSTMOp : public framework::OperatorWithKernel {
   void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Input"),
                    "Input(Input) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Weight"),
+                   "Input(Weight) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Bias"),
+                   "Input(Bias) of LSTM should not be null.");
+
     PADDLE_ENFORCE(ctx->HasOutput("Hidden"),
                    "Output(Hidden) of LSTM should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Cell"),
@@ -59,11 +64,13 @@ class LSTMOp : public framework::OperatorWithKernel {
                       "The second dimension of Input(Weight) "
                       "should be 4 * %d.",
                       frame_size);
+
     auto b_dims = ctx->GetInputDim("Bias");
     PADDLE_ENFORCE_EQ(b_dims.size(), 2, "The rank of Input(Bias) should be 2.");
     PADDLE_ENFORCE_EQ(b_dims[0], 1,
                       "The first dimension of Input(Bias) should be 1.");
-    if (ctx->Attrs().Get<bool>("usePeepholes")) {
+
+    if (ctx->Attrs().Get<bool>("use_peepholes")) {
       PADDLE_ENFORCE_EQ(b_dims[1], 7 * frame_size,
                         "The second dimension of Input(Bias) should be "
                         "7 * %d if enable peepholes connection",
@@ -74,6 +81,7 @@ class LSTMOp : public framework::OperatorWithKernel {
                         "4 * %d if disable peepholes connection",
                         frame_size);
     }
+
     framework::DDim out_dims({in_dims[0], frame_size});
     ctx->SetOutputDim("Hidden", out_dims);
     ctx->SetOutputDim("Cell", out_dims);
@@ -118,14 +126,13 @@ class LSTMOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("Bias",
              "(Tensor) the learnable weights, which contains two parts: "
              "input-hidden bias weight and peephole connections weight if "
-             "setting `usePeepholes` True. "
-             "1. `usePeepholes = False` "
+             "setting `use_peepholes` True. "
+             "1. `use_peepholes = False` "
              " - The shape is (1 x 4D). "
              " - Bias = {b_c, b_i, b_f, b_o}."
-             "2. `usePeepholes = True` "
+             "2. `use_peepholes = True` "
              " - The shape is (1 x 7D). "
-             " - Bias = {b_c, b_i, b_f, b_o, W_ic, W_fc, W_oc}.")
-        .AsDispensable();
+             " - Bias = {b_c, b_i, b_f, b_o, W_ic, W_fc, W_oc}.");
     AddOutput("Hidden",
               "(LoDTensor) the hidden state of LSTM operator. "
               "The shape is (T x D), and lod is the same with the `Input`.");
@@ -145,29 +152,32 @@ class LSTMOpMaker : public framework::OpProtoAndCheckerMaker {
               "(LoDTensor) This LoDTensor is obtained in the forward and used "
               "in the backward.")
         .AsIntermediate();
-    AddAttr<bool>("usePeepholes",
-                  "(bool, default True) "
+    AddAttr<bool>("use_peepholes",
+                  "(bool, defalut: True) "
                   "whether to enable diagonal/peephole connections.")
         .SetDefault(true);
-    AddAttr<bool>("isReverse",
-                  "(bool, default False) "
+    AddAttr<bool>("is_reverse",
+                  "(bool, defalut: False) "
                   "whether to compute reversed LSTM.")
         .SetDefault(false);
     AddAttr<std::string>(
-        "gateActivation",
-        "(string, default sigmoid)"
+        "gate_activation",
+        "(string, default: sigmoid)"
         "The activation for input gate, forget gate and output "
         "gate, `sigmoid` by default.")
-        .SetDefault("sigmoid");
-    AddAttr<std::string>("cellActivation",
-                         "(string, default tanh)"
+        .SetDefault("sigmoid")
+        .InEnum({"sigmoid", "tanh", "relu", "identity"});
+    AddAttr<std::string>("cell_activation",
+                         "(string, default: tanh)"
                          "The activation for cell output, `tanh` by defalut.")
-        .SetDefault("tanh");
-    AddAttr<std::string>("candidateActivation",
-                         "(string, default tanh)"
+        .SetDefault("tanh")
+        .InEnum({"sigmoid", "tanh", "relu", "identity"});
+    AddAttr<std::string>("candidate_activation",
+                         "(string, default: tanh)"
                          "The activation for candidate hidden state, "
                          "`tanh` by default.")
-        .SetDefault("tanh");
+        .SetDefault("tanh")
+        .InEnum({"sigmoid", "tanh", "relu", "identity"});
     AddComment(R"DOC(
 Long-Short Term Memory (LSTM) Operator.
 
@@ -203,7 +213,7 @@ are the cell input and cell output activation functions and `tanh` is usually
 used for them. \f$\tilde{c_t}\f$ is also called candidate hidden state,
 which is computed based on the current input and the previous hidden state.
 
-Set usePeepholes False to disable peephole connection 
+Set `use_peepholes` False to disable peephole connection 
 (http://www.bioinf.jku.at/publications/older/2604.pdf). The formula
 is omitted here.
 
@@ -226,23 +236,27 @@ class LSTMGradOp : public framework::OperatorWithKernel {
                    "Input(Hidden) of LSTM should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Cell"),
                    "Input(Cell) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Weight"),
+                   "Input(Weight) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Bias"),
+                   "Input(Bias) of LSTM should not be null.");
 
     PADDLE_ENFORCE(ctx->HasInput("BatchGate"),
                    "Input(BatchGate) of LSTM should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("BatchCellPreAct"),
                    "Input(BatchGate) of LSTM should not be null.");
 
-    auto in_g_name = framework::GradVarName("Input");
-    if (ctx->HasOutput(in_g_name))
-      ctx->SetOutputDim(in_g_name, ctx->GetInputDim("Input"));
-
-    auto w_g_name = framework::GradVarName("Weight");
-    if (ctx->HasOutput(w_g_name))
-      ctx->SetOutputDim(w_g_name, ctx->GetInputDim("Weight"));
-
-    auto b_g_name = framework::GradVarName("Bias");
-    if (ctx->HasOutput(b_g_name))
-      ctx->SetOutputDim(b_g_name, ctx->GetInputDim("Bias"));
+    auto SetOutGradDim = [&ctx](const std::string& name) {
+      auto g_name = framework::GradVarName(name);
+      if (ctx->HasOutput(g_name))
+        ctx->SetOutputDim(g_name, ctx->GetInputDim(name));
+    };
+
+    SetOutGradDim("Input");
+    SetOutGradDim("Weight");
+    SetOutGradDim("Bias");
+    SetOutGradDim("H0");
+    SetOutGradDim("C0");
   }
 
  protected:

paddle/operators/lstm_op.h

@@ -28,6 +28,15 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
+template <typename Place, typename T>
+inline void ReorderInitState(const platform::DeviceContext& ctx,
+                             const framework::Tensor& src, const size_t* index,
+                             framework::Tensor* dst, bool indexed_src) {
+  math::CopyMatrixRowsFunctor<Place, T> row_shuffle;
+  dst->mutable_data<T>(src.dims(), ctx.GetPlace());
+  row_shuffle(ctx, src, index, *dst, indexed_src);
+}
+
 template <typename Place, typename T>
 class LSTMKernel : public framework::OpKernel<T> {
  public:
@@ -36,6 +45,9 @@ class LSTMKernel : public framework::OpKernel<T> {
     auto* weight = ctx.Input<Tensor>("Weight");
     auto* bias = ctx.Input<Tensor>("Bias");
 
+    auto* hidden_t0 = ctx.Input<Tensor>("H0");
+    auto* cell_t0 = ctx.Input<Tensor>("C0");
+
     auto* batch_gate = ctx.Output<LoDTensor>("BatchGate");
     batch_gate->mutable_data<T>(ctx.GetPlace());
     auto* hidden_out = ctx.Output<LoDTensor>("Hidden");
@@ -43,12 +55,7 @@ class LSTMKernel : public framework::OpKernel<T> {
     auto* cell_out = ctx.Output<LoDTensor>("Cell");
     cell_out->mutable_data<T>(ctx.GetPlace());
 
-    // Now the function ShareLoD in InferShape is not implemented.
-    // So copy LoD here.
-    ctx.ShareLoD("Input", "Hidden");
-    ctx.ShareLoD("Input", "Cell");
-
-    bool is_reverse = ctx.Attr<bool>("isReverse");
+    bool is_reverse = ctx.Attr<bool>("is_reverse");
     math::LoDTensor2BatchFunctor<Place, T> to_batch;
     auto& device_ctx = ctx.device_context();
     to_batch(device_ctx, *input, *batch_gate, true, is_reverse);
@@ -71,7 +78,7 @@ class LSTMKernel : public framework::OpKernel<T> {
     }
 
     math::LstmMetaValue<T> lstm_value;
-    if (bias) {
+    if (bias && ctx.Attr<bool>("use_peepholes")) {
       T* bias_data = const_cast<T*>(bias->data<T>());
       // the code style in LstmMetaValue will be updated later.
 
@@ -84,6 +91,16 @@ class LSTMKernel : public framework::OpKernel<T> {
       lstm_value.checkOg = nullptr;
     }
     lstm_value.prevStateValue = nullptr;
+    Tensor ordered_c0;
+    const size_t* order = batch_gate->lod()[2].data();
+    if (cell_t0) {
+      // Since the batch computing for LSTM reorders the input sequence
+      // according to their length. The initialized cell state also needs
+      // to reorder.
+      ReorderInitState<Place, T>(device_ctx, *cell_t0, order, &ordered_c0,
+                                 true);
+      lstm_value.prevStateValue = ordered_c0.data<T>();
+    }
 
     // Use the local variable as here.
     LoDTensor batch_hidden, batch_cell;
@@ -94,9 +111,9 @@ class LSTMKernel : public framework::OpKernel<T> {
 
     auto batch_starts = batch_gate->lod()[0];
     size_t num_batch = batch_starts.size() - 1;
-    auto gate_act = ctx.Attr<std::string>("gateActivation");
-    auto cell_act = ctx.Attr<std::string>("cellActivation");
-    auto cand_act = ctx.Attr<std::string>("candidateActivation");
+    auto gate_act = ctx.Attr<std::string>("gate_activation");
+    auto cell_act = ctx.Attr<std::string>("cell_activation");
+    auto cand_act = ctx.Attr<std::string>("candidate_activation");
 
     for (size_t n = 0; n < num_batch; n++) {
       int bstart = static_cast<int>(batch_starts[n]);
@@ -109,15 +126,28 @@ class LSTMKernel : public framework::OpKernel<T> {
 
       int cur_batch_size = bend - bstart;
 
-      if (n != 0) {
+      if (n > 0) {
         int pre_h_start = static_cast<int>(batch_starts[n - 1]);
         int pre_h_end = pre_h_start + cur_batch_size;
         auto pre_hidden_t = batch_hidden.Slice(pre_h_start, pre_h_end);
         math::matmul<Place, T>(device_ctx, pre_hidden_t, false, *weight, false,
                                static_cast<T>(1.0), &gate_t,
                                static_cast<T>(1.0));
+      } else if (hidden_t0) {
+        // If n == 0 and there is no initialized hidden state, that is to say
+        // the H0 is zeros, the calculation W_h * H0 will be skiped.
+        // If n == 0 and there is initialized hidden state, calculate W_h * H0.
+
+        // Since the batch computing for LSTM reorders the input sequence
+        // according to their length. The initialized hidden state also needs
+        // to reorder.
+        Tensor ordered_h0;
+        ReorderInitState<Place, T>(device_ctx, *hidden_t0, order, &ordered_h0,
+                                   true);
+        math::matmul<Place, T>(device_ctx, ordered_h0, false, *weight, false,
+                               static_cast<T>(1.0), &gate_t,
+                               static_cast<T>(1.0));
       }
-      // else if : FIXME support the initial hidden and cell
 
       lstm_value.gateValue = gate_t.data<T>();
       lstm_value.outputValue = out_t.data<T>();
@@ -160,6 +190,12 @@ class LSTMGradKernel : public framework::OpKernel<T> {
     auto* weight_g = ctx.Output<Tensor>(framework::GradVarName("Weight"));
     auto* bias_g = ctx.Output<Tensor>(framework::GradVarName("Bias"));
 
+    auto* h0 = ctx.Input<Tensor>("H0");
+    auto* c0 = ctx.Input<Tensor>("C0");
+
+    auto* h0_g = ctx.Output<Tensor>(framework::GradVarName("H0"));
+    auto* c0_g = ctx.Output<Tensor>(framework::GradVarName("C0"));
+
     auto& device_ctx = ctx.device_context();
     math::SetConstant<Place, T> zero;
     if (weight_g) {
@@ -167,13 +203,25 @@ class LSTMGradKernel : public framework::OpKernel<T> {
       zero(device_ctx, weight_g, static_cast<T>(0.0));
     }
 
+    // ordered_h0/c0 is the reordered hidden/cell initialization.
+    // ordered_h0_g/c0_g is the reordered gradient of hidden/cell
+    // initialization.
+    Tensor ordered_h0, ordered_c0, ordered_h0_g, ordered_c0_g;
+    const size_t* order = batch_gate->lod()[2].data();
+    if (c0) {
+      ReorderInitState<Place, T>(device_ctx, *c0, order, &ordered_c0, true);
+    }
+    if (c0 && c0_g) {
+      ordered_c0_g.mutable_data<T>(c0_g->dims(), ctx.GetPlace());
+    }
+
     auto in_dims = input->dims();
     auto out_dims = hidden_g->dims();
     int frame_size = static_cast<int>(in_dims[1] / 4);
     PADDLE_ENFORCE_EQ(frame_size, out_dims[1]);
 
     math::LstmMetaValue<T> lstm_value;
-    if (bias) {
+    if (bias && ctx.Attr<bool>("use_peepholes")) {
       T* bias_data = const_cast<T*>(bias->data<T>());
       lstm_value.checkIg = bias_data + 4 * frame_size;
       lstm_value.checkFg = lstm_value.checkIg + frame_size;
@@ -185,9 +233,13 @@ class LSTMGradKernel : public framework::OpKernel<T> {
     }
 
     math::LstmMetaGrad<T> lstm_grad;
+
     if (bias && bias_g) {
-      T* bias_g_data = const_cast<T*>(bias_g->mutable_data<T>(ctx.GetPlace()));
+      bias_g->mutable_data<T>(ctx.GetPlace());
       zero(device_ctx, bias_g, static_cast<T>(0.0));
+    }
+    if (bias && bias_g && ctx.Attr<bool>("use_peepholes")) {
+      T* bias_g_data = bias_g->data<T>();
       lstm_grad.checkIgGrad = bias_g_data + 4 * frame_size;
       lstm_grad.checkFgGrad = lstm_grad.checkIgGrad + frame_size;
       lstm_grad.checkOgGrad = lstm_grad.checkFgGrad + frame_size;
@@ -199,36 +251,30 @@ class LSTMGradKernel : public framework::OpKernel<T> {
 
     math::LoDTensor2BatchFunctor<Place, T> to_batch;
 
-    // use the local variable as here.
-    LoDTensor batch_hidden;
-    batch_hidden.mutable_data<T>(out_dims, ctx.GetPlace());
-    batch_hidden.set_lod(batch_gate->lod());
-    to_batch(device_ctx, *hidden_out, batch_hidden, false);
+    auto ToBatch = [&batch_gate, &to_batch](
+        const platform::DeviceContext& ctx, const framework::LoDTensor& src,
+        const framework::DDim& dims, framework::LoDTensor& dst) {
+      dst.mutable_data<T>(dims, ctx.GetPlace());
+      dst.set_lod(batch_gate->lod());
+      to_batch(ctx, src, dst, false);
+    };
 
-    LoDTensor batch_hidden_g;
-    batch_hidden_g.mutable_data<T>(out_dims, ctx.GetPlace());
-    batch_hidden_g.set_lod(batch_gate->lod());
-    to_batch(device_ctx, *hidden_g, batch_hidden_g, false);
+    LoDTensor batch_hidden, batch_hidden_g, batch_cell;
+    ToBatch(device_ctx, *hidden_out, out_dims, batch_hidden);
+    ToBatch(device_ctx, *hidden_g, out_dims, batch_hidden_g);
+    ToBatch(device_ctx, *cell_out, out_dims, batch_cell);
 
-    LoDTensor batch_cell;
-    batch_cell.mutable_data<T>(out_dims, ctx.GetPlace());
-    batch_cell.set_lod(batch_gate->lod());
-    to_batch(device_ctx, *cell_out, batch_cell, false);
-
-    LoDTensor batch_cell_g;
+    LoDTensor batch_cell_g, batch_gate_g;
     batch_cell_g.mutable_data<T>(out_dims, ctx.GetPlace());
-    batch_cell_g.set_lod(batch_gate->lod());
     // TODO(qingqing) support the case output cell has gradient.
     // to_batch(device_ctx, *cell_g, batch_cell_g, false);
     zero(device_ctx, &batch_cell_g, static_cast<T>(0.0));
-
-    LoDTensor batch_gate_g;
     batch_gate_g.mutable_data<T>(batch_gate->dims(), ctx.GetPlace());
     batch_gate_g.set_lod(batch_gate->lod());
 
-    auto gate_act = ctx.Attr<std::string>("gateActivation");
-    auto cell_act = ctx.Attr<std::string>("cellActivation");
-    auto cand_act = ctx.Attr<std::string>("candidateActivation");
+    auto gate_act = ctx.Attr<std::string>("gate_activation");
+    auto cell_act = ctx.Attr<std::string>("cell_activation");
+    auto cand_act = ctx.Attr<std::string>("candidate_activation");
 
     auto batch_starts = batch_gate->lod()[0];
     size_t num_batch = batch_starts.size() - 1;
@@ -250,15 +296,15 @@ class LSTMGradKernel : public framework::OpKernel<T> {
       lstm_grad.gateGrad = gate_g.data<T>();
       lstm_grad.outputGrad = out_g.data<T>();
 
-      if (n) {
+      if (n > 0) {
         int bstart_pre = static_cast<int>(batch_starts[n - 1]);
         Tensor cell_pre = batch_cell.Slice(bstart_pre, bstart);
         Tensor cell_pre_g = batch_cell_g.Slice(bstart_pre, bstart);
         lstm_value.prevStateValue = cell_pre.data<T>();
         lstm_grad.prevStateGrad = cell_pre_g.data<T>();
       } else {
-        lstm_value.prevStateValue = nullptr;
-        lstm_grad.prevStateGrad = nullptr;
+        lstm_value.prevStateValue = c0 ? ordered_c0.data<T>() : nullptr;
+        lstm_grad.prevStateGrad = c0_g ? ordered_c0_g.data<T>() : nullptr;
       }
 
       int cur_batch_size = bend - bstart;
@@ -266,7 +312,7 @@ class LSTMGradKernel : public framework::OpKernel<T> {
           device_ctx, lstm_value, lstm_grad, frame_size, cur_batch_size,
           gate_act, cell_act, cand_act);
 
-      if (n != 0) {
+      if (n > 0) {
         int pre_h_start = static_cast<int>(batch_starts[n - 1]);
         int pre_h_end = pre_h_start + cur_batch_size;
         auto pre_hidden_g = batch_hidden_g.Slice(pre_h_start, pre_h_end);
@@ -280,6 +326,19 @@ class LSTMGradKernel : public framework::OpKernel<T> {
                                  static_cast<T>(1.0), weight_g,
                                  static_cast<T>(1.0));
         }
+      } else {
+        if (h0 && weight_g) {
+          ReorderInitState<Place, T>(device_ctx, *h0, order, &ordered_h0, true);
+          math::matmul<Place, T>(device_ctx, ordered_h0, true, gate_g, false,
+                                 static_cast<T>(1.0), weight_g,
+                                 static_cast<T>(1.0));
+        }
+        if (h0 && h0_g) {
+          ordered_h0_g.mutable_data<T>(h0_g->dims(), ctx.GetPlace());
+          math::matmul<Place, T>(device_ctx, gate_g, false, *weight, true,
+                                 static_cast<T>(1.0), &ordered_h0_g,
+                                 static_cast<T>(0.0));
+        }
       }
     }
 
@@ -302,6 +361,13 @@ class LSTMGradKernel : public framework::OpKernel<T> {
       math::gemv<Place, T>(device_ctx, true, m, n, 1., batch_gate_g.data<T>(),
                            ones.data<T>(), 0., bias_g->data<T>());
     }
+
+    if (h0 && h0_g) {
+      ReorderInitState<Place, T>(device_ctx, ordered_h0_g, order, h0_g, false);
+    }
+    if (c0 && c0_g) {
+      ReorderInitState<Place, T>(device_ctx, ordered_c0_g, order, c0_g, false);
+    }
   }
 };
 

paddle/operators/lstm_unit_op.cc

@@ -34,10 +34,10 @@ class LstmUnitOp : public framework::OperatorWithKernel {
     auto c_prev_dims = ctx->GetInputDim("C_prev");
 
     PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank must be 2.");
-    PADDLE_ENFORCE(x_dims[0] == c_prev_dims[0],
-                   "Batch size of inputs and states must be equal");
-    PADDLE_ENFORCE(x_dims[1] == c_prev_dims[1] * 4,
-                   "Dimension of FC should equal to prev state * 4");
+    PADDLE_ENFORCE_EQ(x_dims[0], c_prev_dims[0],
+                      "Batch size of inputs and states must be equal");
+    PADDLE_ENFORCE_EQ(x_dims[1], c_prev_dims[1] * 4,
+                      "Dimension of FC should equal to prev state * 4");
 
     int b_size = c_prev_dims[0];  // batch size
     int s_dim = c_prev_dims[1];   // state dim

paddle/operators/math/CMakeLists.txt

@@ -13,7 +13,7 @@ if(WITH_GPU)
     nv_library(context_project SRCS context_project.cc context_project.cu DEPS device_context)
     nv_library(sequence2batch SRCS sequence2batch.cc sequence2batch.cu DEPS device_context)
     nv_library(lstm_compute SRCS lstm_compute.cc lstm_compute.cu DEPS device_context activation_functions)
-    nv_library(gru_compute SRCS gru_compute.cc gru_compute.cu DEPS device_context activation_functions)
+    nv_library(gru_compute SRCS gru_compute.cc gru_compute.cu DEPS device_context activation_functions math_function)
 else()
     cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context operator)
     cc_library(selected_rows_functor SRCS selected_rows_functor.cc DEPS selected_rows math_function)

paddle/operators/math/detail/lstm_cpu_kernel.h

@@ -52,9 +52,9 @@ void naive_lstm_forward_one_sequence(Op op, LstmMetaValue<T> value,
     rValueIg = valueIg[i];
     rValueFg = valueFg[i];
     rValueOg = valueOg[i];
-    rCheckI = value.checkIg[i];
-    rCheckF = value.checkFg[i];
-    rCheckO = value.checkOg[i];
+    rCheckI = value.checkIg ? value.checkIg[i] : 0;
+    rCheckF = value.checkFg ? value.checkFg[i] : 0;
+    rCheckO = value.checkOg ? value.checkOg[i] : 0;
 
     if (value.prevStateValue) {
       rPrevState = value.prevStateValue[i];
@@ -114,9 +114,9 @@ void naive_lstm_backward_one_sequence(Op op, LstmMetaValue<T> value,
     rValueIg = valueIg[i];
     rValueFg = valueFg[i];
     rValueOg = valueOg[i];
-    rCheckI = value.checkIg[i];
-    rCheckF = value.checkFg[i];
-    rCheckO = value.checkOg[i];
+    rCheckI = value.checkIg ? value.checkIg[i] : 0;
+    rCheckF = value.checkFg ? value.checkFg[i] : 0;
+    rCheckO = value.checkOg ? value.checkOg[i] : 0;
     rState = value.stateValue[i];
     rStateAtv = value.stateActiveValue[i];
     rOutputGrad = grad.outputGrad[i];
@@ -155,9 +155,9 @@ void avx_lstm_forward_one_sequence(Op op, LstmMetaValue<T> value, int frameSize,
   __m256 rValueIg;
   __m256 rValueFg;
   __m256 rValueOg;
-  __m256 rCheckI;
-  __m256 rCheckF;
-  __m256 rCheckO;
+  __m256 rCheckI = _mm256_set1_ps(0.0f);
+  __m256 rCheckF = _mm256_set1_ps(0.0f);
+  __m256 rCheckO = _mm256_set1_ps(0.0f);
   __m256 rState;
   __m256 rPrevState = _mm256_set1_ps(0.0f);
   __m256 rStateAtv;
@@ -173,9 +173,11 @@ void avx_lstm_forward_one_sequence(Op op, LstmMetaValue<T> value, int frameSize,
     rValueIg = valueIg[i];
     rValueFg = valueFg[i];
     rValueOg = valueOg[i];
-    rCheckI = ((__m256 *)value.checkIg)[i];
-    rCheckF = ((__m256 *)value.checkFg)[i];
-    rCheckO = ((__m256 *)value.checkOg)[i];
+    if (value.checkIg) {
+      rCheckI = ((__m256 *)value.checkIg)[i];
+      rCheckF = ((__m256 *)value.checkFg)[i];
+      rCheckO = ((__m256 *)value.checkOg)[i];
+    }
 
     if (value.prevStateValue) {
       rPrevState = ((__m256 *)value.prevStateValue)[i];
@@ -216,9 +218,9 @@ void avx_lstm_backward_one_sequence(Op op, LstmMetaValue<T> value,
   __m256 rState;
   __m256 rStateAtv;
   __m256 rOutputGrad;
-  __m256 rCheckI;
-  __m256 rCheckF;
-  __m256 rCheckO;
+  __m256 rCheckI = _mm256_set1_ps(0.0f);
+  __m256 rCheckF = _mm256_set1_ps(0.0f);
+  __m256 rCheckO = _mm256_set1_ps(0.0f);
   __m256 rCheckIGrad;
   __m256 rCheckFGrad;
   __m256 rCheckOGrad;
@@ -237,9 +239,11 @@ void avx_lstm_backward_one_sequence(Op op, LstmMetaValue<T> value,
     rValueIg = valueIg[i];
     rValueFg = valueFg[i];
     rValueOg = valueOg[i];
-    rCheckI = ((__m256 *)value.checkIg)[i];
-    rCheckF = ((__m256 *)value.checkFg)[i];
-    rCheckO = ((__m256 *)value.checkOg)[i];
+    if (value.checkIg) {
+      rCheckI = ((__m256 *)value.checkIg)[i];
+      rCheckF = ((__m256 *)value.checkFg)[i];
+      rCheckO = ((__m256 *)value.checkOg)[i];
+    }
     rState = ((__m256 *)value.stateValue)[i];
     rStateAtv = ((__m256 *)value.stateActiveValue)[i];
     rOutputGrad = ((__m256 *)grad.outputGrad)[i];

paddle/operators/math/detail/lstm_gpu_kernel.h

@@ -55,9 +55,10 @@ __global__ void KeLstmForward(Op op, LstmMetaValue<T> value, int frameSize,
   T rValueIg;
   T rValueFg;
   T rValueOg;
-  T rCheckI = value.checkIg[frameIdx];
-  T rCheckF = value.checkFg[frameIdx];
-  T rCheckO = value.checkOg[frameIdx];
+
+  T rCheckI = value.checkIg ? value.checkIg[frameIdx] : 0;
+  T rCheckF = value.checkFg ? value.checkFg[frameIdx] : 0;
+  T rCheckO = value.checkOg ? value.checkOg[frameIdx] : 0;
 
   rValueIn = value.gateValue[frameIdx];
   rValueIg = value.gateValue[frameIdx + frameSize];
@@ -121,9 +122,10 @@ __global__ void KeLstmBackward(Op op, LstmMetaValue<T> value,
   T rStateGrad;
   T rStateAtv;
   T rOutputGrad;
-  T rCheckI = value.checkIg[frameIdx];
-  T rCheckF = value.checkFg[frameIdx];
-  T rCheckO = value.checkOg[frameIdx];
+  T rCheckI = value.checkIg ? value.checkIg[frameIdx] : 0;
+  T rCheckF = value.checkFg ? value.checkFg[frameIdx] : 0;
+  T rCheckO = value.checkOg ? value.checkOg[frameIdx] : 0;
+
   T rCheckIGrad;
   T rCheckFGrad;
   T rCheckOGrad;

paddle/operators/math/math_function.cc

@@ -234,8 +234,8 @@ void gemv<platform::CPUPlace, double>(const platform::DeviceContext& context,
 
 template struct SetConstant<platform::CPUPlace, float>;
 
-struct TensorSetConstant {
-  TensorSetConstant(framework::Tensor* tensor, float value)
+struct TensorSetConstantCPU {
+  TensorSetConstantCPU(framework::Tensor* tensor, float value)
       : tensor_(tensor), value_(value) {}
   template <typename T>
   void operator()() const {
@@ -252,7 +252,7 @@ void set_constant_with_place<platform::CPUPlace>(
     const platform::DeviceContext& context, framework::Tensor* tensor,
     float value) {
   framework::VisitDataType(framework::ToDataType(tensor->type()),
-                           TensorSetConstant(tensor, value));
+                           TensorSetConstantCPU(tensor, value));
 }
 
 struct TensorSetConstantWithPlace : public boost::static_visitor<void> {

paddle/operators/math/math_function.cu

@@ -233,8 +233,8 @@ void gemv<platform::GPUPlace, double>(const platform::DeviceContext& context,
 
 template struct SetConstant<platform::GPUPlace, float>;
 
-struct TensorSetConstant {
-  TensorSetConstant(const platform::DeviceContext& context,
+struct TensorSetConstantGPU {
+  TensorSetConstantGPU(const platform::DeviceContext& context,
                     framework::Tensor* tensor, float value)
       : context_(context), tensor_(tensor), value_(value) {}
 
@@ -254,7 +254,7 @@ void set_constant_with_place<platform::GPUPlace>(
     const platform::DeviceContext& context, framework::Tensor* tensor,
     float value) {
   framework::VisitDataType(framework::ToDataType(tensor->type()),
-                           TensorSetConstant(context, tensor, value));
+                           TensorSetConstantGPU(context, tensor, value));
 }
 
 }  // namespace math

paddle/operators/math/math_function.h

@@ -19,11 +19,6 @@ limitations under the License. */
 #include <mkl_vml_functions.h>
 #endif
 
-#ifdef PADDLE_USE_MKL
-#include <mkl.h>
-#include <mkl_lapacke.h>
-#endif
-
 #ifdef PADDLE_USE_ATLAS
 extern "C" {
 #include <cblas.h>

paddle/operators/math/sequence2batch.cc

@@ -22,8 +22,8 @@ template <typename T>
 class CopyMatrixRowsFunctor<platform::CPUPlace, T> {
  public:
   void operator()(const platform::DeviceContext& context,
-                  const framework::LoDTensor& src, const size_t* index,
-                  framework::LoDTensor& dst, bool is_src_index) {
+                  const framework::Tensor& src, const size_t* index,
+                  framework::Tensor& dst, bool is_src_index) {
     auto src_dims = src.dims();
     auto dst_dims = dst.dims();
     PADDLE_ENFORCE_EQ(src_dims.size(), 2UL,

paddle/operators/math/sequence2batch.cu

@@ -41,8 +41,8 @@ template <typename T>
 class CopyMatrixRowsFunctor<platform::GPUPlace, T> {
  public:
   void operator()(const platform::DeviceContext& context,
-                  const framework::LoDTensor& src, const size_t* index,
-                  framework::LoDTensor& dst, bool is_src_index) {
+                  const framework::Tensor& src, const size_t* index,
+                  framework::Tensor& dst, bool is_src_index) {
     auto src_dims = src.dims();
     auto dst_dims = dst.dims();
     PADDLE_ENFORCE_EQ(src_dims.size(), 2,

paddle/operators/math/sequence2batch.h

@@ -30,8 +30,8 @@ class CopyMatrixRowsFunctor {
   // copy the input src to the indexed rows of output dst.
   // The indexed rows are based on the input index.
   void operator()(const platform::DeviceContext& context,
-                  const framework::LoDTensor& src, const size_t* index,
-                  framework::LoDTensor& dst, bool is_src_index);
+                  const framework::Tensor& src, const size_t* index,
+                  framework::Tensor& dst, bool is_src_index);
 };
 
 template <typename Place, typename T>
@@ -57,7 +57,7 @@ class LoDTensor2BatchFunctor {
                   bool is_reverse = false) const {
     if (!is_cal_batch_lod) {
       auto lods = batch.lod();
-      PADDLE_ENFORCE_EQ(lods.size(), 2UL);
+      PADDLE_ENFORCE_GT(lods.size(), 2UL);
       PADDLE_ENFORCE_EQ(lods[1].size(),
                         static_cast<size_t>(lod_tensor.dims()[0]));
       CopyMatrixRowsFunctor<Place, T> to_batch;
@@ -66,8 +66,8 @@ class LoDTensor2BatchFunctor {
     }
 
     auto lods = lod_tensor.lod();
-    PADDLE_ENFORCE_EQ(lods.size(), 1UL, "Only support one level sequence now.");
     auto lod = lods[0];
+    PADDLE_ENFORCE_EQ(lods.size(), 1UL, "Only support one level sequence now.");
 
     std::vector<SeqInfo> seq_info;
     for (size_t seq_id = 0; seq_id < lod.size() - 1; ++seq_id) {
@@ -78,8 +78,7 @@ class LoDTensor2BatchFunctor {
     std::sort(seq_info.begin(), seq_info.end(),
               [](SeqInfo a, SeqInfo b) { return a.length > b.length; });
 
-    // calculate the start position of each batch
-    // (numBatch equal the maxLength of sequences)
+    // Calculate the start position of each batch.
     // example:  sequences = {s0, s1, s2}
     //           s0: 0 0 0 0, s1: 1 1 1 1 1, s2: 2 2 2
     //           num_batch = 5,
@@ -95,19 +94,25 @@ class LoDTensor2BatchFunctor {
     //                                6, 2, 11,
     //                                7, 3,
     //                                8}
-    // The batch number represents batch size after rearranging the
+    //           seq_order = {1, 0, 2}, the sort order.
+    //               where 1 is the second sequence,
+    //                     0 is the first sequence,
+    //                     2 is the third sequence.
+    // The num_batch represents batch size after rearranging the
     // input LodTensor. It is also the maximum length of input sequence.
 
     paddle::framework::LoD batch_lods;
     batch_lods.emplace_back(std::vector<size_t>{0});
     batch_lods.emplace_back(std::vector<size_t>{0});
+    batch_lods.emplace_back(std::vector<size_t>{0});
 
     // batch_lods[0] is the start positions for batch LoDTensor
     int num_batch = seq_info[0].length;
     batch_lods[0].resize(static_cast<size_t>(num_batch + 1));
     // batch_lods[1] is the raw index in the input LoDTensor
-    auto dims = lod_tensor.dims();
-    batch_lods[1].resize(static_cast<size_t>(dims[0]));
+    batch_lods[1].resize(static_cast<size_t>(lod_tensor.dims()[0]));
+    // batch_lods[2] is the sort order for the input LoDTensor.
+    batch_lods[2].resize(seq_info.size());
 
     size_t* batch_starts = batch_lods[0].data();
     size_t* seq2batch_idx = batch_lods[1].data();
@@ -127,6 +132,10 @@ class LoDTensor2BatchFunctor {
       }
       batch_starts[n + 1] = static_cast<size_t>(batch_id);
     }
+    size_t* seq_order = batch_lods[2].data();
+    for (size_t i = 0; i < seq_info.size(); ++i) {
+      seq_order[i] = seq_info[i].seq_idx;
+    }
     batch.set_lod(batch_lods);
 
     CopyMatrixRowsFunctor<Place, T> to_batch;
@@ -141,8 +150,7 @@ class Batch2LoDTensorFunctor {
                   const framework::LoDTensor& batch,
                   framework::LoDTensor& lod_tensor) const {
     auto in_lod = batch.lod();
-    PADDLE_ENFORCE_EQ(in_lod.size(), 2UL,
-                      "The LoD size of input `batch` should be 2.");
+    PADDLE_ENFORCE_GT(in_lod.size(), 2UL);
     PADDLE_ENFORCE_EQ(in_lod[1].size(),
                       static_cast<size_t>(lod_tensor.dims()[0]));
     CopyMatrixRowsFunctor<Place, T> to_seq;

paddle/operators/mean_op.cc

@@ -51,6 +51,7 @@ class MeanGradOp : public framework::OperatorWithKernel {
 
   void InferShape(framework::InferShapeContext* ctx) const override {
     ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
+    ctx->ShareLoD("X", framework::GradVarName("X"));
   }
 };
 

paddle/operators/momentum_op.cc

@@ -75,7 +75,7 @@ class MomentumOpMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("VelocityOut", "(Tensor) Output updated velocity");
 
     AddAttr<float>("mu", "(float) Momentum coefficient");
-    AddAttr<bool>("useNesterov",
+    AddAttr<bool>("use_nesterov",
                   "(bool, default false) "
                   "Use Nesterov Momentum")
         .SetDefault(false);

paddle/operators/momentum_op.h

@@ -34,7 +34,7 @@ class MomentumOpKernel : public framework::OpKernel<T> {
     velocity_out->mutable_data<T>(ctx.GetPlace());
 
     float mu = ctx.Attr<float>("mu");
-    bool use_nesterov = ctx.Attr<bool>("useNesterov");
+    bool use_nesterov = ctx.Attr<bool>("use_nesterov");
 
     auto p_out = framework::EigenVector<T>::Flatten(*param_out);
     auto v_out = framework::EigenVector<T>::Flatten(*velocity_out);

paddle/operators/mul_op.cu

@@ -12,7 +12,6 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#define EIGEN_USE_GPU
 #include "paddle/operators/mul_op.h"
 
 namespace ops = paddle::operators;

paddle/operators/mul_op.h

@@ -16,16 +16,12 @@
 
 #include "paddle/operators/math/math_function.h"
 
-#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 EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
 template <typename Place, typename T>
 class MulKernel : public framework::OpKernel<T> {

paddle/operators/nccl/nccl_gpu_common.h

@@ -35,6 +35,7 @@ constexpr int kInvalidGPUId = -1;
 struct Communicator {
   std::vector<ncclComm_t> comms_;
   std::unordered_map<int, int> comm_id_map_;
+  bool inited_;
 
   Communicator() {}
 
@@ -42,17 +43,21 @@ struct Communicator {
 
   void InitAll(const std::vector<int>& gpus) {
     comms_.resize(gpus.size());
+    inited_ = false;
     for (size_t i = 0; i < gpus.size(); ++i) {
       comm_id_map_[gpus[i]] = i;
     }
     PADDLE_ENFORCE(
         dynload::ncclCommInitAll(comms_.data(), gpus.size(), gpus.data()));
+    inited_ = true;
   }
 
   ~Communicator() {
-    for (size_t i = 0; i < comms_.size(); ++i) {
-      // FIXME(dzh) : PADDLE_ENFORCE return void
-      dynload::ncclCommDestroy(comms_[i]);
+    if (inited_) {
+      for (size_t i = 0; i < comms_.size(); ++i) {
+        // FIXME(dzh) : PADDLE_ENFORCE return void
+        dynload::ncclCommDestroy(comms_[i]);
+      }
     }
   }
 

paddle/operators/nccl_op_test.cu

@@ -26,7 +26,6 @@
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/program_desc.h"
 #include "paddle/framework/var_desc.h"
-#include "paddle/operators/math/math_function.h"
 #include "paddle/operators/nccl/nccl_gpu_common.h"
 #include "paddle/platform/device_context.h"
 #include "paddle/platform/enforce.h"

paddle/operators/pool_cudnn_op.cu

@@ -37,11 +37,11 @@ class PoolCudnnOpKernel : public framework::OpKernel<T> {
     const T *input_data = input->data<T>();
     T *output_data = output->mutable_data<T>(ctx.GetPlace());
 
-    std::string pooling_type = ctx.Attr<std::string>("poolingType");
+    std::string pooling_type = ctx.Attr<std::string>("pooling_type");
     std::vector<int> ksize = ctx.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
-    if (ctx.Attr<bool>("globalPooling")) {
+    if (ctx.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
         ksize[i] = static_cast<int>(input->dims()[i + 2]);
@@ -92,12 +92,12 @@ class PoolCudnnGradOpKernel : public framework::OpKernel<T> {
         ctx.Input<Tensor>(framework::GradVarName("Out"));
     Tensor *input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
 
-    std::string pooling_type = ctx.Attr<std::string>("poolingType");
+    std::string pooling_type = ctx.Attr<std::string>("pooling_type");
     std::vector<int> ksize = ctx.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
 
-    if (ctx.Attr<bool>("globalPooling")) {
+    if (ctx.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
         ksize[i] = static_cast<int>(input->dims()[i + 2]);

paddle/operators/pool_op.cc

@@ -29,7 +29,7 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const {
 
   auto in_x_dims = ctx->GetInputDim("X");
 
-  std::string pooling_type = ctx->Attrs().Get<std::string>("poolingType");
+  std::string pooling_type = ctx->Attrs().Get<std::string>("pooling_type");
   std::vector<int> ksize = ctx->Attrs().Get<std::vector<int>>("ksize");
   std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
   std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
@@ -37,7 +37,7 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const {
   PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5,
                  "Pooling intput should be 4-D or 5-D tensor.");
 
-  if (ctx->Attrs().Get<bool>("globalPooling")) {
+  if (ctx->Attrs().Get<bool>("global_pooling")) {
     ksize.resize(static_cast<size_t>(in_x_dims.size()) - 2);
     for (size_t i = 0; i < ksize.size(); ++i) {
       paddings[i] = 0;
@@ -83,20 +83,20 @@ Pool2dOpMaker::Pool2dOpMaker(framework::OpProto *proto,
             "H is the height of the feature, "
             "and W is the width of the feature.");
 
-  AddAttr<std::string>("poolingType",
+  AddAttr<std::string>("pooling_type",
                        "(string), pooling type, can be \"max\" for max-pooling "
                        "and \"avg\" for average-pooling.")
       .InEnum({"max", "avg"});
   AddAttr<std::vector<int>>("ksize",
                             "(vector<int>) The pooling window "
                             "size(height, width) of the pooling operator. "
-                            "If globalPooling = true, ksize and paddings will "
+                            "If global_pooling = true, ksize and paddings will "
                             "be ignored.");  // TODO(Chengduo): Add checker.
                                              // (Currently,
   // TypedAttrChecker don't support vector type.)
-  AddAttr<bool>("globalPooling",
+  AddAttr<bool>("global_pooling",
                 "(bool, default false) Whether to use the global pooling. "
-                "If globalPooling = true, ksize and paddings will be ignored.")
+                "If global_pooling = true, ksize and paddings will be ignored.")
       .SetDefault(false);
   AddAttr<std::vector<int>>("strides",
                             "(vector<int>, default {1, 1}), strides(height, "
@@ -107,7 +107,7 @@ Pool2dOpMaker::Pool2dOpMaker(framework::OpProto *proto,
       "paddings",
       "(vector<int>, defalut {0,0}), paddings(height, width) of pooling "
       "operator."
-      "If globalPooling = true, paddings and ksize will be ignored.")
+      "If global_pooling = true, paddings and ksize will be ignored.")
       .SetDefault({0, 0});  // TODO(Chengduo): Add checker. (Currently,
   // TypedAttrChecker don't support vector type.)
 
@@ -115,7 +115,7 @@ Pool2dOpMaker::Pool2dOpMaker(framework::OpProto *proto,
 Pool2d Operator.
 
 The pooling2d operation calculates the output based on
-the input, poolingType and ksize, strides, paddings parameters.
+the input, pooling_type and ksize, strides, paddings parameters.
 Input(X) and output(Out) are in NCHW format, where N is batch size, C is the
 number of channels, H is the height of the feature, and W is the width of the feature.
 Parameters(ksize, strides, paddings) are two elements.
@@ -152,7 +152,7 @@ Pool3dOpMaker::Pool3dOpMaker(framework::OpProto *proto,
             "the number of channels, and D, H and W is the depth, height and "
             "width of the feature, respectively.");
 
-  AddAttr<std::string>("poolingType",
+  AddAttr<std::string>("pooling_type",
                        "(string) Pooling type, can be \"max\" for max-pooling "
                        "and \"avg\" for average-pooling.")
       .InEnum({"max", "avg"});
@@ -160,13 +160,14 @@ Pool3dOpMaker::Pool3dOpMaker(framework::OpProto *proto,
       "ksize",
       "(vector<int>) The pooling window size(depth, height, "
       "width) of pooling operator. "
-      "If globalPooling = true, ksize and paddings will "
+      "If global_pooling = true, ksize and paddings will "
       "be ignored.");  // TODO(Chengduo): Add checker.
                        // (Currently,
   // TypedAttrChecker don't support vector type.)
-  AddAttr<bool>("globalPooling",
-                "(bool, default false) Whether to use the global pooling. "
-                "If globalPooling = true, ksize and paddings wille be ignored.")
+  AddAttr<bool>(
+      "global_pooling",
+      "(bool, default false) Whether to use the global pooling. "
+      "If global_pooling = true, ksize and paddings wille be ignored.")
       .SetDefault(false);
   AddAttr<std::vector<int>>(
       "strides",
@@ -178,7 +179,7 @@ Pool3dOpMaker::Pool3dOpMaker(framework::OpProto *proto,
       "paddings",
       "(vector<int>, defalut {0,0,0}), paddings(depth, height, "
       "width) of pooling operator. "
-      "If globalPooling = true, ksize and paddings will be ignored.")
+      "If global_pooling = true, ksize and paddings will be ignored.")
       .SetDefault({0, 0, 0});  // TODO(Chengduo): Add checker. (Currently,
                                // TypedAttrChecker don't support vector type.)
 
@@ -186,7 +187,7 @@ Pool3dOpMaker::Pool3dOpMaker(framework::OpProto *proto,
 Pool3d Operator.
 
 The pooling3d operation calculates the output based on
-the input, poolingType, ksize, strides, and paddings parameters.
+the input, pooling_type, ksize, strides, and paddings parameters.
 Input(X) and output(Out) are in NCDHW format, where N is batch
 size, C is the number of channels, and D, H and W are the depth, height and
 width of the feature, respectively. Parameters(ksize, strides, paddings) 

paddle/operators/pool_op.h

@@ -57,11 +57,11 @@ class PoolKernel : public framework::OpKernel<T> {
     const Tensor* in_x = context.Input<Tensor>("X");
     Tensor* out = context.Output<Tensor>("Out");
 
-    std::string pooling_type = context.Attr<std::string>("poolingType");
+    std::string pooling_type = context.Attr<std::string>("pooling_type");
     std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
-    if (context.Attr<bool>("globalPooling")) {
+    if (context.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
         ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
@@ -119,12 +119,12 @@ class PoolGradKernel : public framework::OpKernel<T> {
         context.Input<Tensor>(framework::GradVarName("Out"));
     Tensor* in_x_grad = context.Output<Tensor>(framework::GradVarName("X"));
 
-    std::string pooling_type = context.Attr<std::string>("poolingType");
+    std::string pooling_type = context.Attr<std::string>("pooling_type");
     std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
 
-    if (context.Attr<bool>("globalPooling")) {
+    if (context.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
         ksize[i] = static_cast<int>(in_x->dims()[i + 2]);

paddle/operators/pool_with_index_op.cc

@@ -44,7 +44,7 @@ class MaxPoolWithIndexOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5,
                    "Pooling intput should be 4-D or 5-D tensor.");
 
-    if (ctx->Attrs().Get<bool>("globalPooling")) {
+    if (ctx->Attrs().Get<bool>("global_pooling")) {
       ksize.resize(static_cast<size_t>(in_x_dims.size()) - 2);
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
@@ -110,14 +110,14 @@ class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<std::vector<int>>("ksize",
                               "(vector<int>) The pooling window size(height, "
                               "width) of pooling operator. "
-                              "If globalPooling = true, ksize and paddings "
+                              "If global_pooling = true, ksize and paddings "
                               "will be ignored.");  // TODO(Chengduo): Add
                                                     // checker. (Currently,
     // TypedAttrChecker don't support vector type.)
     AddAttr<bool>(
-        "globalPooling",
+        "global_pooling",
         "(bool, default false) Whether to use the global pooling. "
-        "If globalPooling = true, ksize and paddings will be ignored.")
+        "If global_pooling = true, ksize and paddings will be ignored.")
         .SetDefault(false);
     AddAttr<std::vector<int>>("strides",
                               "(vector<int>, default {1, 1}), strides(height, "
@@ -128,7 +128,7 @@ class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
         "paddings",
         "(vector<int>, defalut {0, 0}), paddings(height, width) of pooling "
         "operator. "
-        "If globalPooling = true, paddings and will be ignored.")
+        "If global_pooling = true, paddings and will be ignored.")
         .SetDefault({0, 0});  // TODO(Chengduo): Add checker. (Currently,
     // TypedAttrChecker don't support vector type.)
 
@@ -188,14 +188,14 @@ class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<std::vector<int>>("ksize",
                               "(vector<int>) The pooling window size(depth, "
                               "height, width) of pooling operator. "
-                              "If globalPooling = true, ksize and paddings "
+                              "If global_pooling = true, ksize and paddings "
                               "will be ignored.");  // TODO(Chengduo): Add
                                                     // checker. (Currently,
     // TypedAttrChecker don't support vector type.)
     AddAttr<bool>(
-        "globalPooling",
+        "global_pooling",
         "(bool, default false) Whether to use the global pooling. "
-        "If globalPooling = true, ksize and paddings will be ignored.")
+        "If global_pooling = true, ksize and paddings will be ignored.")
         .SetDefault(false);
     AddAttr<std::vector<int>>("strides",
                               "(vector<int>, default {1,1,1}), strides(depth, "
@@ -206,7 +206,7 @@ class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
         "paddings",
         "(vector, defalut {0,0,0}), paddings(depth, "
         "height, width) of pooling operator. "
-        "If globalPooling = true, paddings and ksize will be ignored.")
+        "If global_pooling = true, paddings and ksize will be ignored.")
         .SetDefault({0, 0, 0});  // TODO(Chengduo): Add checker. (Currently,
     // TypedAttrChecker don't support vector type.)
 

paddle/operators/pool_with_index_op.h

@@ -35,7 +35,7 @@ class MaxPoolWithIndexKernel : public framework::OpKernel<T> {
     std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
-    if (context.Attr<bool>("globalPooling")) {
+    if (context.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
         ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
@@ -72,7 +72,7 @@ class MaxPoolWithIndexGradKernel : public framework::OpKernel<T> {
     std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
-    if (context.Attr<bool>("globalPooling")) {
+    if (context.Attr<bool>("global_pooling")) {
       for (size_t i = 0; i < ksize.size(); ++i) {
         paddings[i] = 0;
         ksize[i] = static_cast<int>(in_x_grad->dims()[i + 2]);

paddle/operators/recurrent_op.cc

@@ -387,8 +387,8 @@ class RecurrentGradOp : public RecurrentBase {
         auto &p_names = Inputs(kParameters);
         PADDLE_ENFORCE_EQ(pg_names.size(), p_names.size());
 
-        for (size_t prog_id = 0; prog_id < pg_names.size(); ++prog_id) {
-          auto inside_grad_name = framework::GradVarName(p_names[prog_id]);
+        for (size_t param_id = 0; param_id < pg_names.size(); ++param_id) {
+          auto inside_grad_name = framework::GradVarName(p_names[param_id]);
 
           // If does not compute gradient of that variable inside rnn, just
           // continue
@@ -406,27 +406,19 @@ class RecurrentGradOp : public RecurrentBase {
             attrs["value"] = 0.0f;
 
             auto zero_op = framework::OpRegistry::CreateOp(
-                "fill_constant", {}, {{"Out", {pg_names[prog_id]}}}, attrs);
+                "fill_constant", {}, {{"Out", {pg_names[param_id]}}}, attrs);
             zero_op->Run(scope, dev_ctx);
           }
 
+          auto new_inside_name = cur_scope.Rename(inside_grad_name);
           // sum gradient
-          auto *outside_var = scope.FindVar(pg_names[prog_id]);
-          PADDLE_ENFORCE(outside_var != nullptr);
-          auto &outside_tensor =
-              *outside_var->GetMutable<framework::LoDTensor>();
-
-          std::string result_var_name;
-          auto *local_result_var = cur_scope.Var(&result_var_name);
-          auto &local_result_tensor =
-              *local_result_var->GetMutable<framework::LoDTensor>();
-
-          local_result_tensor.ShareDataWith(outside_tensor);
 
           auto sum_op = framework::OpRegistry::CreateOp(
-              "sum", {{"X", {result_var_name, inside_grad_name}}},
-              {{"Out", {result_var_name}}}, {});
+              "sum", {{"X", {pg_names[param_id], new_inside_name}}},
+              {{"Out", {pg_names[param_id]}}}, {});
           sum_op->Run(cur_scope, dev_ctx);
+
+          cur_scope.Rename(new_inside_name, inside_grad_name);
         }
       }
       VLOG(5) << "Accumulate Parameter finished ";

paddle/operators/sequence_concat_op.cc

@@ -68,38 +68,42 @@ class SequenceConcatOpMaker : public framework::OpProtoAndCheckerMaker {
                  "The level should be less than the level number of inputs.")
         .SetDefault(0);
     AddComment(R"DOC(
-Sequence Concat Operator.
-
-The sequence_concat operator concatenates multiple LoDTensors.
-It supports a sequence (LoD Tensor with level number is 1)
+The sequence_concat operator concatenates multiple LoDTensors. 
+It only supports sequence (LoD Tensor with level number is 1) 
 or a nested sequence (LoD tensor with level number is 2) as its input.
-The following examples explain how the operator works:
 - Case1:
   If the axis is other than 0(here, axis is 1 and level is 1),
-  each input should have the same LoD information and the LoD
+  each input should have the same LoD information and the LoD 
   information of the output keeps the same as the input.
 
-    LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)
-    LoD(x1) = {{0,2,4}, {0,1,2,3,4}}; Dims(x1) = (4,4,4)
-    LoD(Out) = {{0,2,4}, {0,1,2,3,4}}; Dims(Out) = (4,7,4)
+  LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)
+  LoD(x1) = {{0,2,4}, {0,1,2,3,4}}; Dims(x1) = (4,4,4)
+  LoD(Out) = {{0,2,4}, {0,1,2,3,4}}; Dims(Out) = (4,7,4)
 
 - Case2:
-  If the axis is 0(here, leve is 0), the inputs are concatenated along
+  If the axis is 0(here, leve is 0), the inputs are concatenated along 
   time steps, the LoD information of the output need to re-compute.
+  The LoD information of level-1 should be same.
 
-    LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)
-    LoD(x1) = {{0,3,5}, {0,1,2,3,5}}; Dims(x1) = (5,3,4)
-    LoD(Out) = {{0,5,9}, {0,1,2,3,4,5,6,7,9}}; Dims(Out) = (9,3,4)
+  LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)
+  LoD(x1) = {{0,2,4}, {0,1,3,5,7}}; Dims(x1) = (7,3,4)
+  LoD(Out) = {{0,2,4}, {0,2,5,8,11}}; Dims(Out) = (11,3,4)
 
 - Case3:
   If the axis is 0(here, level is 1).
 
-    LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)
-    LoD(x1) = {{0,3,5}, {0,1,3,4,5}}; Dims(x1) = (5,3,4)
-    LoD(Out) = {{0,5,9}, {0,2,5,7,9}}; Dims(Out) = (9,3,4)
+  LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)
+  LoD(x1) = {{0,3,4}, {0,1,3,5,7}}; Dims(x1) = (7,3,4)
+  LoD(Out) = {{0,5,8}, {0,1,2,3,5,7,8,9,11}}; Dims(Out) = (11,3,4)
 
-NOTE: The levels of all the inputs should be the same.
+- Case4:
+  If the LoD number is 1, axis is 0, level is 0
 
+  LoD(x0) = {{0,1,2,3,4}}; Dims(x0) = (4,3,4)
+  LoD(x1) = {{0,1,3,5,7}}; Dims(x1) = (7,3,4)
+  LoD(Out) = {{0,2,5,8,11}}; Dims(Out) = (11,3,4)
+
+NOTE: The levels of all the inputs should be the same.
     )DOC");
   }
 };