diff --git a/CMakeLists.txt b/CMakeLists.txt
index 00996cb7ed5cc573c42b69be6db369c3654d6d1a..7c7eb260aea8478f4833cb79253f4481e10b8685 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -31,9 +31,6 @@ if(NOT CMAKE_CROSSCOMPILING)
 endif(NOT CMAKE_CROSSCOMPILING)
 find_package(Git REQUIRED)
 find_package(Threads REQUIRED)
-if(NOT ANDROID AND NOT IOS)
-    find_package(Boost QUIET)
-endif()
 
 include(simd)
 
@@ -55,6 +52,8 @@ option(WITH_COVERAGE    "Compile PaddlePaddle with code coverage"       OFF)
 option(COVERALLS_UPLOAD "Package code coverage data to coveralls"       OFF)
 option(ON_TRAVIS        "Exclude special unit test on Travis CI"        OFF)
 option(WITH_C_API       "Compile PaddlePaddle with C-API(Prediction)"   OFF)
+# TODO: Only compile PaddlePaddle fluid version by WITH_FLUID option. 
+option(WITH_FLUID       "Compile PaddlePaddle fluid only(TODO)"         ON)
 option(WITH_GOLANG      "Compile PaddlePaddle with GOLANG"              OFF)
 option(GLIDE_INSTALL    "Download and install go dependencies "         ON)
 option(USE_NNPACK       "Compile PaddlePaddle with NNPACK library"      OFF)
@@ -107,6 +106,10 @@ if (WITH_C_API AND WITH_PYTHON)
     "different Python interpreter from compiling.")
 endif()
 
+if (WITH_C_API)
+  set(WITH_FLUID OFF CACHE STRING "Disable install fluid when compile the C_API" FORCE)
+endif()
+
 if(MOBILE_INFERENCE)
     set(THIRD_PARTY_BUILD_TYPE MinSizeRel)
 else()
@@ -134,6 +137,7 @@ include(external/openblas)  # download, build, install openblas
 include(external/mkldnn)    # download, build, install mkldnn
 include(external/swig)      # download, build, install swig
 include(external/warpctc)   # download, build, install warpctc
+include(external/boost)     # download, build, install boost
 include(external/any)       # download libn::any
 include(external/eigen)     # download eigen3
 include(external/pybind11)  # download pybind11
@@ -158,7 +162,6 @@ include_directories("${PADDLE_SOURCE_DIR}")
 include_directories("${PADDLE_SOURCE_DIR}/paddle/cuda/include")
 include_directories("${CMAKE_CURRENT_BINARY_DIR}/proto")
 include_directories("${CMAKE_CURRENT_BINARY_DIR}/go/pserver/client/c")
-include_directories(${Boost_INCLUDE_DIRS})
 
 set(EXTERNAL_LIBS
     ${GFLAGS_LIBRARIES}
diff --git a/Dockerfile b/Dockerfile
index 857d3f3e5f64791146741ffb29feabfcb2ecbb84..6ac9901ac6cea12e97047efdfb6272c957f166ae 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -27,7 +27,7 @@ RUN apt-get update && \
     curl sed grep graphviz libjpeg-dev zlib1g-dev  \
     python-matplotlib gcc-4.8 g++-4.8 \
     automake locales clang-format swig doxygen cmake  \
-    liblapack-dev liblapacke-dev libboost-dev \
+    liblapack-dev liblapacke-dev \
     clang-3.8 llvm-3.8 libclang-3.8-dev \
     net-tools libtool && \
     apt-get clean -y
diff --git a/cmake/external/boost.cmake b/cmake/external/boost.cmake
new file mode 100644
index 0000000000000000000000000000000000000000..137f11da7f2f1c46eebf6590d93402786ef543c9
--- /dev/null
+++ b/cmake/external/boost.cmake
@@ -0,0 +1,51 @@
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+include(ExternalProject)
+
+set(BOOST_PROJECT       "extern_boost")
+set(BOOST_VER           "1.66.0")
+set(BOOST_TAR           "boost_1_66_0")
+set(BOOST_URL           "https://dl.bintray.com/boostorg/release/${BOOST_VER}/source/${BOOST_TAR}.tar.gz")
+set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost)
+set(BOOST_DOWNLOAD_DIR  "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}")
+set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)
+
+include_directories(${BOOST_INCLUDE_DIR})
+
+ExternalProject_Add(
+    ${BOOST_PROJECT}
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    DOWNLOAD_DIR          ${BOOST_DOWNLOAD_DIR}
+    DOWNLOAD_COMMAND      wget --no-check-certificate ${BOOST_URL} -c -q -O ${BOOST_TAR}.tar.gz
+                          && tar zxf ${BOOST_TAR}.tar.gz
+    DOWNLOAD_NO_PROGRESS  1
+    PREFIX                ${BOOST_SOURCES_DIR}
+    CONFIGURE_COMMAND     ""
+    BUILD_COMMAND         ""
+    INSTALL_COMMAND       ""
+    UPDATE_COMMAND        ""
+)
+
+if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
+    set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/boost_dummy.c)
+    file(WRITE ${dummyfile} "const char *dummy = \"${dummyfile}\";")
+    add_library(boost STATIC ${dummyfile})
+else()
+    add_library(boost INTERFACE)
+endif()
+
+add_dependencies(boost ${BOOST_PROJECT})
+list(APPEND external_project_dependencies boost)
+set(Boost_INCLUDE_DIR ${BOOST_INCLUDE_DIR})
diff --git a/cmake/external/eigen.cmake b/cmake/external/eigen.cmake
index c4712f19eb80b34ffbf713d2b13fc0c775312af1..d49c8d601102cf865287c33349bff5eee6a90f6d 100644
--- a/cmake/external/eigen.cmake
+++ b/cmake/external/eigen.cmake
@@ -1,8 +1,8 @@
 INCLUDE(ExternalProject)
 
 SET(EIGEN_SOURCE_DIR ${THIRD_PARTY_PATH}/eigen3)
-
-INCLUDE_DIRECTORIES(${EIGEN_SOURCE_DIR}/src/extern_eigen3)
+SET(EIGEN_INCLUDE_DIR ${EIGEN_SOURCE_DIR}/src/extern_eigen3)
+INCLUDE_DIRECTORIES(${EIGEN_INCLUDE_DIR})
 
 ExternalProject_Add(
     extern_eigen3
@@ -28,3 +28,9 @@ endif()
 add_dependencies(eigen3 extern_eigen3)
 
 LIST(APPEND external_project_dependencies eigen3)
+
+IF(NOT WITH_C_API AND WITH_FLUID)
+    INSTALL(FILES ${EIGEN_INCLUDE_DIR}/Eigen/Core DESTINATION third_party/eigen3/Eigen)
+    INSTALL(DIRECTORY ${EIGEN_INCLUDE_DIR}/Eigen/src DESTINATION third_party/eigen3/Eigen)
+    INSTALL(DIRECTORY ${EIGEN_INCLUDE_DIR}/unsupported/Eigen DESTINATION third_party/eigen3/unsupported)
+ENDIF()
diff --git a/cmake/external/gflags.cmake b/cmake/external/gflags.cmake
index d4f252bb9f64c8db82b841fedf0817f5d8596501..60946304541a20809276c3e665d8524baf209006 100644
--- a/cmake/external/gflags.cmake
+++ b/cmake/external/gflags.cmake
@@ -52,7 +52,7 @@ ADD_DEPENDENCIES(gflags extern_gflags)
 
 LIST(APPEND external_project_dependencies gflags)
 
-IF(WITH_C_API)
+IF(WITH_C_API OR WITH_FLUID)
   INSTALL(DIRECTORY ${GFLAGS_INCLUDE_DIR} DESTINATION third_party/gflags)
   IF(ANDROID)
     INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib/${ANDROID_ABI})
diff --git a/cmake/external/glog.cmake b/cmake/external/glog.cmake
index 0c6b3aafcb4e990b9d4549820137474e5968a7aa..382fbda3b5cfeba893f03871cf65498d20804f36 100644
--- a/cmake/external/glog.cmake
+++ b/cmake/external/glog.cmake
@@ -68,7 +68,7 @@ LINK_LIBRARIES(glog gflags)
 
 LIST(APPEND external_project_dependencies glog)
 
-IF(WITH_C_API)
+IF(WITH_C_API OR WITH_FLUID)
   INSTALL(DIRECTORY ${GLOG_INCLUDE_DIR} DESTINATION third_party/glog)
   IF(ANDROID)
     INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib/${ANDROID_ABI})
diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake
index ff5855052dabaa0b63099cd219f3f04e22f1aa85..365a370a9cfb708379bcff18ae6aa0725d420ae1 100644
--- a/cmake/external/protobuf.cmake
+++ b/cmake/external/protobuf.cmake
@@ -250,7 +250,7 @@ IF(NOT PROTOBUF_FOUND)
     SET(PROTOBUF_PROTOC_LIBRARY ${extern_protobuf_PROTOC_LIBRARY}
         CACHE FILEPATH "protoc library." FORCE)
 
-    IF(WITH_C_API)
+    IF(WITH_C_API OR WITH_FLUID)
         INSTALL(DIRECTORY ${PROTOBUF_INCLUDE_DIR} DESTINATION third_party/protobuf)
         IF(ANDROID)
             INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})
diff --git a/doc/api/v1/data_provider/dataprovider_cn.rst b/doc/api/v1/data_provider/dataprovider_cn.rst
deleted file mode 100644
index d08c6b3efacbc35ae274d5b207fe91e747124e79..0000000000000000000000000000000000000000
--- a/doc/api/v1/data_provider/dataprovider_cn.rst
+++ /dev/null
@@ -1,15 +0,0 @@
-.. _api_dataprovider:
-
-DataProvider的介绍
-==================
-
-DataProvider是PaddlePaddle负责提供数据的模块。其作用是将数据传入内存或显存，让神经网络可以进行训练或预测。用户可以通过简单使用Python接口 :ref:`api_pydataprovider2` ，来自定义传数据的过程。如果有更复杂的使用，或者需要更高的效率，用户也可以在C++端自定义一个 ``DataProvider`` 。
-
-PaddlePaddle需要用户在网络配置（trainer_config.py）中定义使用哪种DataProvider，并且在DataProvider中实现如何访问训练文件列表（train.list）或测试文件列表（test.list）。
-
-- train.list和test.list存放在本地（推荐直接存放到训练目录，以相对路径引用)。一般情况下，两者均为纯文本文件，其中每一行对应一个数据文件地址：
-  
-  - 如果数据文件存于本地磁盘，这个地址则为它的绝对路径或相对路径(相对于PaddlePaddle程序运行时的路径)。
-  - 地址也可以为hdfs文件路径，或者数据库连接路径等。
-  - 由于这个地址会被DataProvider使用，因此，如何解析该地址也是用户自定义DataProvider时需要考虑的地方。
-- 如果没有设置test.list，或设置为None，那么在训练过程中不会执行测试操作；否则，会根据命令行参数指定的测试方式，在训练过程中进行测试，从而防止过拟合。
diff --git a/doc/api/v1/data_provider/dataprovider_en.rst b/doc/api/v1/data_provider/dataprovider_en.rst
deleted file mode 100644
index 96efbb1da959daec561009fdcc95d353b191dec8..0000000000000000000000000000000000000000
--- a/doc/api/v1/data_provider/dataprovider_en.rst
+++ /dev/null
@@ -1,34 +0,0 @@
-Introduction
-==============
-DataProvider is a module that loads training or testing data into cpu or gpu
-memory for the following triaining or testing process.
-
-For simple use, users can use Python :code:`PyDataProvider` to dynamically reads
-the original data in any format or in any form, and then transfer them into a
-data format PaddlePaddle requires. The process is extremly flexible and highly
-customized, with sacrificing the efficiency only a little. This is extremly
-useful when you have to dynamically generate certain kinds of data according to,
-for example, the training performance.
-
-Besides, users also can customize a C++ :code:`DataProvider` for a more
-complex usage, or for a higher efficiency.
-
-The following parameters are required to define in the PaddlePaddle network
-configuration file (trainer_config.py): which DataProvider is chosen to used,
-and specific parameters for DataProvider, including training file list
-(train.list) and testing file list (test.list).
-
-Train.list and test.list are simply two plain text files, which defines path
-of training or testing data. It is recommended that directly placing them into
-the training directory, and reference to them by using a relative path (
-relative to the PaddePaddle program).
-
-Testing or evaluating will not be performed during training if the test.list is
-not set or set to None. Otherwise, PaddlePaddle will evaluate the trained model
-by the specified tesing data while training, every testing period (a user
-defined command line parameter in PaddlePaddle) to prevent over-fitting.
-
-Each line of train.list and test.list is an absolute or relative path (relative
-to the PaddePaddle program runtime) of data file. Fascinatingly more, each line
-can also be a HDFS file path or a SQL connection string. As long as the user
-assures how to access each file in DataProvider.
diff --git a/doc/api/v1/data_provider/pydataprovider2_cn.rst b/doc/api/v1/data_provider/pydataprovider2_cn.rst
deleted file mode 100644
index 8f9db31cfb9946e1d2db3872718bd92787d861f0..0000000000000000000000000000000000000000
--- a/doc/api/v1/data_provider/pydataprovider2_cn.rst
+++ /dev/null
@@ -1,229 +0,0 @@
-..  _api_pydataprovider2:
-
-PyDataProvider2的使用
-=====================
-
-PyDataProvider2是PaddlePaddle使用Python提供数据的推荐接口。该接口使用多线程读取数据，并提供了简单的Cache功能；同时可以使用户只关注如何从文件中读取每一条数据，而不用关心数据如何传输，如何存储等等。
-
-..  contents::
-
-MNIST的使用场景
----------------
-
-我们以MNIST手写识别为例，来说明PyDataProvider2的简单使用场景。
-
-样例数据
-++++++++
-
-MNIST是一个包含有70,000张灰度图片的数字分类数据集。样例数据 ``mnist_train.txt`` 如下：
-
-..  literalinclude:: src/mnist_train.txt
-
-其中每行数据代表一张图片，行内使用 ``;`` 分成两部分。第一部分是图片的标签，为0-9中的一个数字；第二部分是28*28的图片像素灰度值。 对应的 ``train.list`` 即为这个数据文件的名字：
-
-..  literalinclude:: src/train.list
-
-dataprovider的使用
-++++++++++++++++++
-
-..  literalinclude:: src/mnist_provider.dict.py
-
-- 首先，引入PaddlePaddle的PyDataProvider2包。
-- 其次，定义一个Python的 `Decorator <http://www.learnpython.org/en/Decorators>`_ `@provider`_ 。用于将下一行的数据输入函数标记成一个PyDataProvider2，同时设置它的input_types属性。
-  
-  - `input_types`_：设置这个PyDataProvider2返回什么样的数据。本例根据网络配置中 ``data_layer`` 的名字，显式指定返回的是一个28*28维的稠密浮点数向量和一个[0-9]的10维整数标签。
-
-    ..  literalinclude:: src/mnist_config.py
-         :lines: 9-10
-
-  - 注意：如果用户不显示指定返回数据的对应关系，那么PaddlePaddle会根据layer的声明顺序，来确定对应关系。但这个关系可能不正确，所以推荐使用显式指定的方式来设置input_types。
-- 最后，实现数据输入函数（如本例的 ``process`` 函数）。
-
-  - 该函数的功能是：打开文本文件，读取每一行，将行中的数据转换成与input_types一致的格式，然后返回给PaddlePaddle进程。注意，
-    
-    - 返回的顺序需要和input_types中定义的顺序一致。
-    - 返回时，必须使用Python关键词 ``yield`` ，相关概念是 ``generator`` 。
-    - 一次yield调用，返回一条完整的样本。如果想为一个数据文件返回多条样本，只需要在函数中调用多次yield即可（本例中使用for循环进行多次调用）。
-  
-  - 该函数具有两个参数：
-  
-    - settings：在本例中没有使用，具体可以参考 `init_hook`_ 中的说明。
-    - filename：为 ``train.list`` 或 ``test.list`` 中的一行，即若干数据文件路径的某一个。
-
-网络配置中的调用
-++++++++++++++++
-
-在网络配置里，只需要一行代码就可以调用这个PyDataProvider2，如，
-
-..  literalinclude:: src/mnist_config.py
-     :lines: 1-7
-
-训练数据是 ``train.list`` ，没有测试数据，调用的PyDataProvider2是 ``mnist_provider`` 模块中的 ``process`` 函数。
-
-小结
-+++++
-
-至此，简单的PyDataProvider2样例就说明完毕了。对用户来说，仅需要知道如何从 **一个文件** 中读取 **一条样本** ，就可以将数据传送给PaddlePaddle了。而PaddlePaddle则会帮用户做以下工作：
-
-* 将数据组合成Batch进行训练
-* 对训练数据进行Shuffle
-* 多线程的数据读取
-* 缓存训练数据到内存(可选)
-* CPU->GPU双缓存
-
-是不是很简单呢？
-
-时序模型的使用场景
-------------------
-样例数据
-++++++++
-
-时序模型是指数据的某一维度是一个序列形式，即包含时间步信息。所谓时间步信息，不一定和时间有关系，只是说明数据的顺序是重要的。例如，文本信息就是一个序列数据。
-
-本例采用英文情感分类的数据，即将一段英文文本数据，分类成正面情绪和负面情绪两类(用0和1表示)。样例数据 ``sentimental_train.txt`` 如下：
-
-..  literalinclude:: src/sentimental_train.txt
-
-dataprovider的使用
-++++++++++++++++++
-
-相对MNIST而言，这个dataprovider较复杂，主要原因是增加了初始化机制 `init_hook`_。本例的 ``on_init`` 函数就是根据该机制配置的，它会在dataprovider创建的时候执行。
-
-- 其中 ``input_types`` 和在 `@provider`_ 中配置的效果一致。本例中的输入特征是词ID的序列，因此使用 ``integer_value_sequence`` 类型来设置。
-- 将 ``dictionary`` 存入settings对象，在 ``process`` 函数中使用。 dictionary是从网络配置中传入的dict对象，即一个将单词字符串映射到单词ID的字典。
-
-..  literalinclude:: src/sentimental_provider.py
-
-网络配置中的调用
-++++++++++++++++
-
-调用这个PyDataProvider2的方法，基本上和MNIST样例一致，除了
-
-* 在配置中需要读取外部字典。
-* 在声明DataProvider的时候传入dictionary作为参数。
-
-..  literalinclude:: src/sentimental_config.py
-     :emphasize-lines: 12-14
-
-参考(Reference)
----------------
-
-@provider
-+++++++++
-
-``@provider`` 是一个Python的 `Decorator`_ ，可以将某一个函数标记成一个PyDataProvider2。如果不了解 `Decorator`_ 是什么也没关系，只需知道这是一个标记属性的方法就可以了。它包含的属性参数如下:
-
-*  input_types：数据输入格式。具体的格式说明，请参考 `input_types`_ 。
-*  should_shuffle：是不是要对数据做Shuffle。训练时默认shuffle，测试时默认不shuffle。
-*  min_pool_size：设置内存中最小暂存的数据条数，也是PaddlePaddle所能够保证的shuffle粒度。如果为-1，则会预先读取全部数据到内存中。
-*  pool_size： 设置内存中暂存的数据条数。如果为-1（默认），则不在乎内存暂存多少条数据。如果设置，则推荐大于训练时batch size的值，并且在内存足够的情况下越大越好。
-*  can_over_batch_size：是否允许暂存略微多余pool_size的数据。由于这样做可以避免很多死锁问题，一般推荐设置成True。
-*  calc_batch_size：可以传入一个函数，用于自定义每条数据的batch size（默认为1）。
-*  cache： 数据缓存的策略，具体请参考 `cache`_ 。
-*  init_hook：初始化时调用的函数，具体请参考 `init_hook`_ 。
-*  check：如果为true，会根据input_types检查数据的合法性。
-*  check_fail_continue：如果为true，那么当check出数据不合法时，会扔到这条数据，继续训练或预测。（对check=false的情况，没有作用）
-
-input_types
-+++++++++++
-
-PaddlePaddle的数据包括四种主要类型，和三种序列模式。
-
-四种数据类型：
-
-* dense_vector：稠密的浮点数向量。
-* sparse_binary_vector：稀疏的01向量，即大部分值为0，但有值的地方必须为1。
-* sparse_float_vector：稀疏的向量，即大部分值为0，但有值的部分可以是任何浮点数。
-* integer：整数标签。
-
-三种序列模式：
-
-* SequenceType.NO_SEQUENCE：不是一条序列
-* SequenceType.SEQUENCE：是一条时间序列
-* SequenceType.SUB_SEQUENCE： 是一条时间序列，且序列的每一个元素还是一个时间序列。
-
-不同的数据类型和序列模式返回的格式不同，列表如下：
-
-+----------------------+---------------------+-----------------------------------+------------------------------------------------+
-|                      | NO_SEQUENCE         | SEQUENCE                          |  SUB_SEQUENCE                                  |
-+======================+=====================+===================================+================================================+
-| dense_vector         | [f, f, ...]         | [[f, ...], [f, ...], ...]         | [[[f, ...], ...], [[f, ...], ...],...]         |
-+----------------------+---------------------+-----------------------------------+------------------------------------------------+
-| sparse_binary_vector | [i, i, ...]         | [[i, ...], [i, ...], ...]         | [[[i, ...], ...], [[i, ...], ...],...]         |
-+----------------------+---------------------+-----------------------------------+------------------------------------------------+
-| sparse_float_vector  | [(i,f), (i,f), ...] | [[(i,f), ...], [(i,f), ...], ...] | [[[(i,f), ...], ...], [[(i,f), ...], ...],...] |
-+----------------------+---------------------+-----------------------------------+------------------------------------------------+
-| integer_value        |  i                  | [i, i, ...]                       | [[i, ...], [i, ...], ...]                      |
-+----------------------+---------------------+-----------------------------------+------------------------------------------------+
-
-其中，f代表一个浮点数，i代表一个整数。
-
-注意：对sparse_binary_vector和sparse_float_vector，PaddlePaddle存的是有值位置的索引。例如，
-
-- 对一个5维非序列的稀疏01向量 ``[0, 1, 1, 0, 0]`` ，类型是sparse_binary_vector，返回的是 ``[1, 2]`` 。
-- 对一个5维非序列的稀疏浮点向量 ``[0, 0.5, 0.7, 0, 0]`` ，类型是sparse_float_vector，返回的是 ``[(1, 0.5), (2, 0.7)]`` 。
-
-init_hook
-+++++++++
-
-init_hook可以传入一个函数。该函数在初始化的时候会被调用，其参数如下:
-
-* 第一个参数是settings对象，它和数据传入函数的第一个参数（如本例中 ``process`` 函数的 ``settings`` 参数）必须一致。该对象具有以下两个属性：
-    * settings.input_types：数据输入格式，具体请参考 `input_types`_ 。
-    * settings.logger：一个logging对象。
-* 其他参数使用 ``kwargs`` （key word arguments）传入，包括以下两种：
-    * PaddlePaddle定义的参数: 1）is_train：bool型参数，表示用于训练或预测；2）file_list：所有文件列表。
-    * 用户定义的参数：使用args在网络配置中设置。
-
-注意：PaddlePaddle保留添加参数的权力，因此init_hook尽量使用 ``**kwargs`` 来接受不使用的函数以保证兼容性。
-
-cache
-+++++
-
-PyDataProvider2提供了两种简单的Cache策略：
-
-* CacheType.NO_CACHE：不缓存任何数据，每次都会从python端读取数据
-* CacheType.CACHE_PASS_IN_MEM：第一个pass会从python端读取数据，剩下的pass会直接从内存里
-  读取数据。 
-
-
-注意事项
---------
-
-可能的内存泄露问题
-++++++++++++++++++
-
-PaddlePaddle将train.list中的每一行都传递给process函数，从而生成多个generator。当训练数据非常多时，就会生成非常多的generator。
-
-虽然每个generator在没有调用的时候，是几乎不占内存的；但当调用过一次后，generator便会存下当前的上下文(Context)，而这个Context可能会非常大。并且，generator至少需要调用两次才会知道是否停止。所以，即使process函数里面只有一个yield，也需要两次随机选择到相同generator的时候，才会释放该段内存。
-
-..  code-block:: python
-
-    def func():
-        yield 0
-
-    f = func()  # 创建generator
-    tmp = next(f)  # 调用一次，返回0
-    tmp = next(f)  # 调用第二次的时候，才会Stop Iteration
-
-由于顺序调用这些generator不会出现上述问题，因此有两种解决方案：
-
-1. **最佳推荐**：将样本的地址放入另一个文本文件，train.list写入那个文本文件的地址。即不要将每一个样本都放入train.list。
-2. 在generator的上下文中尽量留下非常少的变量引用，例如
-
-..  code-block:: python
-
-    def real_process(fn):
-        # ... read from fn
-        return result   # 当函数返回的时候，python可以解除掉内部变量的引用。
-
-    def process(fn):
-        yield real_process(fn)
-
-注意：这个问题是PyDataProvider读数据时候的逻辑问题，很难整体修正。
-
-内存不够用的情况
-++++++++++++++++
-
-PyDataProvider2会尽可能多的使用内存。因此，对于内存较小的机器，推荐使用 ``pool_size`` 变量来设置内存中暂存的数据条。具体请参考 `@provider`_ 中的说明。
-
diff --git a/doc/api/v1/data_provider/pydataprovider2_en.rst b/doc/api/v1/data_provider/pydataprovider2_en.rst
deleted file mode 100644
index e8fb6292779790765154502bff319ea10ab1e70b..0000000000000000000000000000000000000000
--- a/doc/api/v1/data_provider/pydataprovider2_en.rst
+++ /dev/null
@@ -1,249 +0,0 @@
-..  _api_pydataprovider2:
-
-PyDataProvider2
-===============
-
-We highly recommand users to use PyDataProvider2 to provide training or testing
-data to PaddlePaddle. The user only needs to focus on how to read a single
-sample from the original data file by using PyDataProvider2, leaving all of the
-trivial work, including, transfering data into cpu/gpu memory, shuffle, binary
-serialization to PyDataProvider2. PyDataProvider2 uses multithreading and a
-fanscinating but simple cache strategy to optimize the efficiency of the data
-providing process.
-
-DataProvider for the non-sequential model
------------------------------------------
-
-Here we use the MNIST handwriting recognition data as an example to illustrate
-how to write a simple PyDataProvider.
-
-MNIST is a handwriting classification data set. It contains 70,000 digital
-grayscale images. Labels of the training sample range from 0 to 9. All the
-images have been size-normalized and centered into images with the same size
-of 28 x 28 pixels.
-
-A small part of the original data as an example is shown as below:
-
-.. literalinclude:: src/mnist_train.txt
-
-Each line of the data contains two parts, separated by :code:`;`. The first part is
-label of an image. The second part contains 28x28 pixel float values.
-
-Just write path of the above data into train.list. It looks like this:
-
-.. literalinclude:: src/train.list
-
-The corresponding dataprovider is shown as below:
-
-.. literalinclude:: src/mnist_provider.dict.py
-
-The first line imports PyDataProvider2 package.
-The main function is the process function, that has two parameters.
-The first parameter is the settings, which is not used in this example.
-The second parameter is the filename, that is exactly each line of train.list.
-This parameter is passed to the process function by PaddlePaddle.
-
-:code:`@provider` is a Python
-`Decorator <http://www.learnpython.org/en/Decorators>`_ .
-It sets some properties to DataProvider, and constructs a real PaddlePaddle
-DataProvider from a very simple user implemented python function. It does not
-matter if you are not familiar with `Decorator`_. You can keep it simple by
-just taking :code:`@provider` as a fixed mark above the provider function you
-implemented.
-
-`input_types`_ defines the data format that a DataProvider returns.
-In this example, it is set to a 28x28-dimensional dense vector and an integer
-scalar, whose value ranges from 0 to 9.
-`input_types`_ can be set to several kinds of input formats, please refer to the
-document of `input_types`_ for more details.
-
-
-The process method is the core part to construct a real DataProvider in
-PaddlePaddle. It implements how to open the text file, how to read one sample
-from the original text file, convert them into `input_types`_, and give them
-back to PaddlePaddle process at line 23.
-Note that data yielded by the process function must follow the same order that
-`input_types`_ are defined.
-
-
-With the help of PyDataProvider2, user can focus on how to generate ONE traning
-sample by using keywords :code:`yield`.
-:code:`yield` is a python keyword, and a concept related to it includes
-:code:`generator`.
-
-Only a few lines of codes need to be added into the training configuration file,
-you can take this as an example.
-
-.. literalinclude:: src/mnist_config.py
-
-Here we specify training data by :code:`train.list`, and no testing data is specified.
-The method which actually provide data is :code:`process`.
-
-User also can use another style to provide data, which defines the
-:code:`data_layer`'s name explicitly when `yield`. For example,
-the :code:`dataprovider` is shown as below.
-
-.. literalinclude:: src/mnist_provider.dict.py
-   :linenos:
-
-If user did't give the :code:`data_layer`'s name, PaddlePaddle will use
-the order of :code:`data_layer` definition roughly to determine which feature to
-which :code:`data_layer`. This order may be not correct, so TO DEFINE THE
-:code:`data_layer`'s NAMES EXPLICITLY IS THE RECOMMANDED WAY TO PROVIDER DATA.
-
-Now, this simple example of using PyDataProvider is finished.
-The only thing that the user should know is how to generte **one sample** from
-**one data file**.
-And PaddlePadle will do all of the rest things\:
-
-* Form a training batch
-* Shuffle the training data
-* Read data with multithreading
-* Cache the training data (Optional)
-* CPU-> GPU double buffering.
-
-Is this cool?
-
-..  _api_pydataprovider2_sequential_model:
-
-DataProvider for the sequential model
--------------------------------------
-A sequence model takes sequences as its input. A sequence is made up of several
-timesteps. The so-called timestep, is not necessary to have something to do
-with time. It can also be explained to that the order of data are taken into
-consideration into model design and training.
-For example, the sentence can be interpreted as a kind of sequence data in NLP
-tasks.
-
-Here is an example on data proivider for English sentiment classification data.
-The original input data are simple English text, labeled into positive or
-negative sentiment (marked by 0 and 1 respectively).
-
-A small part of the original data as an example can be found in the path below:
-
-.. literalinclude:: src/sentimental_train.txt
-
-The corresponding data provider can be found in the path below:
-
-.. literalinclude:: src/sentimental_provider.py
-
-This data provider for sequential model is a little more complex than that
-for MINST dataset.
-A new initialization method is introduced here.
-The method :code:`on_init` is configured to DataProvider by :code:`@provider`'s
-:code:`init_hook` parameter, and it will be invoked once DataProvider is
-initialized. The :code:`on_init` function has the following parameters:
-
-* The first parameter is the settings object.
-* The rest parameters are passed by key word arguments. Some of them are passed
-  by PaddlePaddle, see reference for `init_hook`_.
-  The :code:`dictionary` object is a python dict object passed from the trainer
-  configuration file, and it maps word string to word id.
-
-To pass these parameters into DataProvider, the following lines should be added
-into trainer configuration file.
-
-.. literalinclude:: src/sentimental_config.py
-
-The definition is basically same as MNIST example, except:
-* Load dictionary in this configuration
-* Pass it as a parameter to the DataProvider
-
-The `input_types` is configured in method :code:`on_init`. It has the same
-effect to configure them by :code:`@provider`'s :code:`input_types` parameter.
-However, the :code:`input_types` is set at runtime, so we can set it to
-different types according to the input data. Input of the neural network is a
-sequence of word id, so set :code:`seq_type` to :code:`integer_value_sequence`.
-
-Durning :code:`on_init`, we save :code:`dictionary` variable to
-:code:`settings`, and it will be used in :code:`process`. Note the settings
-parameter for the process function and for the on_init's function are a same
-object.
-
-The basic processing logic is the same as MNIST's :code:`process` method. Each
-sample in the data file is given back to PaddlePaddle process.
-
-Thus, the basic usage of PyDataProvider is here.
-Please refer to the following section reference for details.
-
-Reference
----------
-
-@provider
-+++++++++
-
-.. autofunction:: paddle.trainer.PyDataProvider2.provider
-
-input_types
-+++++++++++
-
-PaddlePaddle has four data types, and three sequence types.
-The four data types are:
-
-* :code:`dense_vector`: dense float vector.
-* :code:`sparse_binary_vector`: sparse binary vector, most of the value is 0, and
-  the non zero elements are fixed to 1.
-* :code:`sparse_float_vector`: sparse float vector, most of the value is 0, and some
-  non zero elements can be any float value. They are given by the user.
-* :code:`integer`: an integer scalar, that is especially used for label or word index.
-
-The three sequence types are:
-
-* :code:`SequenceType.NO_SEQUENCE` means the sample is not a sequence.
-* :code:`SequenceType.SEQUENCE` means the sample is a sequence.
-* :code:`SequenceType.SUB_SEQUENCE` means it is a nested sequence, that each timestep of
-  the input sequence is also a sequence.
-
-Different input type has a defferenct input format. Their formats are shown
-in the above table.
-
-+----------------------+---------------------+-----------------------------------+------------------------------------------------+
-|                      | NO_SEQUENCE         | SEQUENCE                          |  SUB_SEQUENCE                                  |
-+======================+=====================+===================================+================================================+
-| dense_vector         | [f, f, ...]         | [[f, ...], [f, ...], ...]         | [[[f, ...], ...], [[f, ...], ...],...]         |
-+----------------------+---------------------+-----------------------------------+------------------------------------------------+
-| sparse_binary_vector | [i, i, ...]         | [[i, ...], [i, ...], ...]         | [[[i, ...], ...], [[i, ...], ...],...]         |
-+----------------------+---------------------+-----------------------------------+------------------------------------------------+
-| sparse_float_vector  | [(i,f), (i,f), ...] | [[(i,f), ...], [(i,f), ...], ...] | [[[(i,f), ...], ...], [[(i,f), ...], ...],...] |
-+----------------------+---------------------+-----------------------------------+------------------------------------------------+
-| integer_value        |  i                  | [i, i, ...]                       | [[i, ...], [i, ...], ...]                      |
-+----------------------+---------------------+-----------------------------------+------------------------------------------------+
-
-where f represents a float value, i represents an integer value.
-
-init_hook
-+++++++++
-
-init_hook is a function that is invoked once the data provoder is initialized.
-Its parameters lists as follows:
-
-* The first parameter is a settings object, which is the same to :code:`settings`
-  in :code:`process` method. The object contains several attributes, including:
-
-  * :code:`settings.input_types`: the input types. Reference `input_types`_.
-  * :code:`settings.logger`: a logging object.
-
-* The rest parameters are the key word arguments. It is made up of PaddpePaddle
-  pre-defined parameters and user defined parameters.
-
-  * PaddlePaddle-defined parameters including:
-
-    * :code:`is_train` is a bool parameter that indicates the DataProvider is used in
-      training or testing.
-    * :code:`file_list` is the list of all files.
-
-  * User-defined parameters args can be set in training configuration.
-
-Note, PaddlePaddle reserves the right to add pre-defined parameter, so please
-use :code:`**kwargs` in init_hook to ensure compatibility by accepting the
-parameters which your init_hook does not use.
-
-cache
-+++++
-DataProvider provides two simple cache strategy. They are:
-
-* :code:`CacheType.NO_CACHE` means do not cache any data, then data is read at runtime by
-  the user implemented python module every pass.
-* :code:`CacheType.CACHE_PASS_IN_MEM` means the first pass reads data by the user
-  implemented python module, and the rest passes will directly read data from
-  memory.
diff --git a/doc/api/v1/data_provider/src/mnist_config.py b/doc/api/v1/data_provider/src/mnist_config.py
deleted file mode 100644
index d2af9d849ec48347d4d56c2b5b6b517671c37518..0000000000000000000000000000000000000000
--- a/doc/api/v1/data_provider/src/mnist_config.py
+++ /dev/null
@@ -1,24 +0,0 @@
-#   Copyright (c) 2018 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.
-
-from paddle.trainer_config_helpers import *
-
-define_py_data_sources2(
-    train_list='train.list',
-    test_list=None,
-    module='mnist_provider',
-    obj='process')
-
-img = data_layer(name='pixel', size=784)
-label = data_layer(name='label', size=10)
diff --git a/doc/api/v1/data_provider/src/mnist_provider.dict.py b/doc/api/v1/data_provider/src/mnist_provider.dict.py
deleted file mode 100644
index 284f7dadb065f8c8a891e949a800280cde9edda9..0000000000000000000000000000000000000000
--- a/doc/api/v1/data_provider/src/mnist_provider.dict.py
+++ /dev/null
@@ -1,38 +0,0 @@
-#   Copyright (c) 2018 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.
-
-from paddle.trainer.PyDataProvider2 import *
-
-
-# Define a py data provider
-@provider(
-    input_types={'pixel': dense_vector(28 * 28),
-                 'label': integer_value(10)})
-def process(settings, filename):  # settings is not used currently.
-    f = open(filename, 'r')  # open one of training file
-
-    for line in f:  # read each line
-        label, pixel = line.split(';')
-
-        # get features and label
-        pixels_str = pixel.split(' ')
-
-        pixels_float = []
-        for each_pixel_str in pixels_str:
-            pixels_float.append(float(each_pixel_str))
-
-        # give data to paddle.
-        yield {"pixel": pixels_float, 'label': int(label)}
-
-    f.close()  # close file
diff --git a/doc/api/v1/data_provider/src/mnist_train.txt b/doc/api/v1/data_provider/src/mnist_train.txt
deleted file mode 100644
index 34be718ad9ea49e7d9aabc34ad70099479df3b31..0000000000000000000000000000000000000000
--- a/doc/api/v1/data_provider/src/mnist_train.txt
+++ /dev/null
@@ -1,3 +0,0 @@
-5;0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.215686 0.533333 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.67451 0.992157 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.070588 0.886275 0.992157 0 0 0 0 0 0 0 0 0 0 0.192157 0.070588 0 0 0 0 0 0 0 0 0 0 0 0 0 0.670588 0.992157 0.992157 0 0 0 0 0 0 0 0 0 0.117647 0.933333 0.858824 0.313725 0 0 0 0 0 0 0 0 0 0 0 0.090196 0.858824 0.992157 0.831373 0 0 0 0 0 0 0 0 0 0.141176 0.992157 0.992157 0.611765 0.054902 0 0 0 0 0 0 0 0 0 0 0.258824 0.992157 0.992157 0.529412 0 0 0 0 0 0 0 0 0 0.368627 0.992157 0.992157 0.419608 0.003922 0 0 0 0 0 0 0 0 0 0.094118 0.835294 0.992157 0.992157 0.517647 0 0 0 0 0 0 0 0 0 0.603922 0.992157 0.992157 0.992157 0.603922 0.545098 0.043137 0 0 0 0 0 0 0 0.447059 0.992157 0.992157 0.956863 0.062745 0 0 0 0 0 0 0 0 0.011765 0.666667 0.992157 0.992157 0.992157 0.992157 0.992157 0.745098 0.137255 0 0 0 0 0 0.152941 0.866667 0.992157 0.992157 0.521569 0 0 0 0 0 0 0 0 0 0.070588 0.992157 0.992157 0.992157 0.803922 0.352941 0.745098 0.992157 0.945098 0.317647 0 0 0 0 0.580392 0.992157 0.992157 0.764706 0.043137 0 0 0 0 0 0 0 0 0 0.070588 0.992157 0.992157 0.776471 0.043137 0 0.007843 0.27451 0.882353 0.941176 0.176471 0 0 0.180392 0.898039 0.992157 0.992157 0.313725 0 0 0 0 0 0 0 0 0 0 0.070588 0.992157 0.992157 0.713725 0 0 0 0 0.627451 0.992157 0.729412 0.062745 0 0.509804 0.992157 0.992157 0.776471 0.035294 0 0 0 0 0 0 0 0 0 0 0.494118 0.992157 0.992157 0.968627 0.168627 0 0 0 0.423529 0.992157 0.992157 0.364706 0 0.717647 0.992157 0.992157 0.317647 0 0 0 0 0 0 0 0 0 0 0 0.533333 0.992157 0.984314 0.945098 0.603922 0 0 0 0.003922 0.466667 0.992157 0.988235 0.976471 0.992157 0.992157 0.788235 0.007843 0 0 0 0 0 0 0 0 0 0 0 0.686275 0.882353 0.364706 0 0 0 0 0 0 0.098039 0.588235 0.992157 0.992157 0.992157 0.980392 0.305882 0 0 0 0 0 0 0 0 0 0 0 0 0.101961 0.67451 0.321569 0 0 0 0 0 0 0 0.105882 0.733333 0.976471 0.811765 0.713725 0 0 0 0 0 0 0 0 0 0 0 0 0 0.65098 0.992157 0.321569 0 0 0 0 0 0 0 0 0 0.25098 0.007843 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0.94902 0.219608 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.968627 0.764706 0.152941 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.498039 0.25098 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0;
-0;0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.298039 0.333333 0.333333 0.333333 0.337255 0.333333 0.333333 0.109804 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.027451 0.223529 0.776471 0.964706 0.988235 0.988235 0.988235 0.992157 0.988235 0.988235 0.780392 0.098039 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.14902 0.698039 0.988235 0.992157 0.988235 0.901961 0.87451 0.568627 0.882353 0.976471 0.988235 0.988235 0.501961 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.188235 0.647059 0.988235 0.988235 0.745098 0.439216 0.098039 0 0 0 0.572549 0.988235 0.988235 0.988235 0 0 0 0 0 0 0 0 0 0 0 0 0 0.2 0.933333 0.992157 0.941176 0.247059 0 0 0 0 0 0 0.188235 0.898039 0.992157 0.992157 0 0 0 0 0 0 0 0 0 0 0 0.039216 0.639216 0.933333 0.988235 0.913725 0.278431 0 0 0 0 0 0 0 0.113725 0.843137 0.988235 0.988235 0 0 0 0 0 0 0 0 0 0 0 0.235294 0.988235 0.992157 0.988235 0.815686 0.07451 0 0 0 0 0 0 0 0.333333 0.988235 0.988235 0.552941 0 0 0 0 0 0 0 0 0 0 0.211765 0.878431 0.988235 0.992157 0.701961 0.329412 0.109804 0 0 0 0 0 0 0 0.698039 0.988235 0.913725 0.145098 0 0 0 0 0 0 0 0 0 0.188235 0.890196 0.988235 0.988235 0.745098 0.047059 0 0 0 0 0 0 0 0 0 0.882353 0.988235 0.568627 0 0 0 0 0 0 0 0 0 0.2 0.933333 0.992157 0.992157 0.992157 0.447059 0.294118 0 0 0 0 0 0 0 0 0.447059 0.992157 0.768627 0 0 0 0 0 0 0 0 0 0 0.623529 0.988235 0.988235 0.988235 0.988235 0.992157 0.47451 0 0 0 0 0 0 0 0.188235 0.933333 0.87451 0.509804 0 0 0 0 0 0 0 0 0 0 0.992157 0.988235 0.937255 0.792157 0.988235 0.894118 0.082353 0 0 0 0 0 0 0.027451 0.647059 0.992157 0.654902 0 0 0 0 0 0 0 0 0 0 0 0.623529 0.988235 0.913725 0.329412 0.376471 0.184314 0 0 0 0 0 0 0.027451 0.513725 0.988235 0.635294 0.219608 0 0 0 0 0 0 0 0 0 0 0 0.196078 0.929412 0.988235 0.988235 0.741176 0.309804 0 0 0 0 0 0 0.529412 0.988235 0.678431 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.223529 0.992157 0.992157 1 0.992157 0.992157 0.992157 0.992157 1 0.992157 0.992157 0.882353 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.023529 0.478431 0.654902 0.658824 0.952941 0.988235 0.988235 0.988235 0.992157 0.988235 0.729412 0.278431 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.196078 0.647059 0.764706 0.764706 0.768627 0.580392 0.047059 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0;
-4;0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.180392 0.470588 0.623529 0.623529 0.623529 0.588235 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.243137 0.494118 0.862745 0.870588 0.960784 0.996078 0.996078 0.996078 0.996078 0.992157 0.466667 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.317647 0.639216 0.639216 0.639216 0.639216 0.639216 0.470588 0.262745 0.333333 0.929412 0.694118 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.811765 0.694118 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.811765 0.694118 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.811765 0.694118 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.184314 0.992157 0.694118 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.192157 0.996078 0.384314 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.454902 0.980392 0.219608 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.564706 0.941176 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.588235 0.776471 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.945098 0.560784 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.054902 0.952941 0.356863 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.337255 0.917647 0.109804 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.698039 0.701961 0.019608 0.4 0.662745 0.662745 0.662745 0.662745 0.662745 0.662745 0.662745 0.376471 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.090196 0.639216 0.972549 0.945098 0.913725 0.996078 0.996078 0.996078 0.996078 1 0.996078 0.996078 1 0.996078 0 0 0 0 0 0 0 0 0 0 0.007843 0.105882 0.717647 0.776471 0.905882 0.996078 0.996078 0.988235 0.980392 0.862745 0.537255 0.223529 0.223529 0.368627 0.376471 0.6 0.6 0.6 0 0 0 0 0 0 0 0 0.262745 0.470588 0.6 0.996078 0.996078 0.996078 0.996078 0.847059 0.356863 0.156863 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.909804 0.705882 0.823529 0.635294 0.490196 0.219608 0.113725 0.062745 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.152941 0.152941 0.156863 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0;
diff --git a/doc/api/v1/data_provider/src/sentimental_config.py b/doc/api/v1/data_provider/src/sentimental_config.py
deleted file mode 100644
index 56adde13b97b30095729da5246805a2902870676..0000000000000000000000000000000000000000
--- a/doc/api/v1/data_provider/src/sentimental_config.py
+++ /dev/null
@@ -1,28 +0,0 @@
-#   Copyright (c) 2018 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.
-
-from paddle.trainer_config_helpers import *
-
-dictionary = dict()
-...  #  read dictionary from outside
-
-define_py_data_sources2(
-    train_list='train.list',
-    test_list=None,
-    module='sentimental_provider',
-    obj='process',
-    # above codes same as mnist sample.
-    args={  # pass to provider.
-        'dictionary': dictionary
-    })
diff --git a/doc/api/v1/data_provider/src/sentimental_provider.py b/doc/api/v1/data_provider/src/sentimental_provider.py
deleted file mode 100644
index 59a2b6f7f5faff847f6fda15827632c4ab236b21..0000000000000000000000000000000000000000
--- a/doc/api/v1/data_provider/src/sentimental_provider.py
+++ /dev/null
@@ -1,57 +0,0 @@
-#   Copyright (c) 2018 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.
-
-from paddle.trainer.PyDataProvider2 import *
-
-
-def on_init(settings, dictionary, **kwargs):
-    # on_init will invoke when data provider is initialized. The dictionary
-    # is passed from trainer_config, and is a dict object with type
-    # (word string => word id).
-
-    # set input types in runtime. It will do the same thing as
-    # @provider(input_types) will do, but it is set dynamically during runtime.
-    settings.input_types = {
-        # The text is a sequence of integer values, and each value is a word id.
-        # The whole sequence is the sentences that we want to predict its
-        # sentimental.
-        'data': integer_value_sequence(len(dictionary)),  # text input
-        'label': integer_value(2)  # label positive/negative
-    }
-
-    # save dictionary as settings.dictionary. 
-    # It will be used in process method.
-    settings.dictionary = dictionary
-
-
-@provider(init_hook=on_init)
-def process(settings, filename):
-    f = open(filename, 'r')
-
-    for line in f:  # read each line of file
-        label, sentence = line.split('\t')  # get label and sentence
-        words = sentence.split(' ')  # get words
-
-        # convert word string to word id
-        # the word not in dictionary will be ignored.
-        word_ids = []
-
-        for each_word in words:
-            if each_word in settings.dictionary:
-                word_ids.append(settings.dictionary[each_word])
-
-        # give data to paddle.
-        yield word_ids, int(label)
-
-    f.close()
diff --git a/doc/api/v1/data_provider/src/sentimental_train.txt b/doc/api/v1/data_provider/src/sentimental_train.txt
deleted file mode 100644
index 0060ac267c4bf884a8e19553bc4bdea48c89c9b2..0000000000000000000000000000000000000000
--- a/doc/api/v1/data_provider/src/sentimental_train.txt
+++ /dev/null
@@ -1,3 +0,0 @@
-0       I saw this movie at the AFI Dallas festival . It all takes place at a lake house and it looks wonderful .
-1       This documentary makes you travel all around the globe . It contains rare and stunning sequels from the wilderness .
-...
diff --git a/doc/api/v1/data_provider/src/train.list b/doc/api/v1/data_provider/src/train.list
deleted file mode 100644
index 92bdc0a8b4c21b3091341d93afc5c536c3cffe47..0000000000000000000000000000000000000000
--- a/doc/api/v1/data_provider/src/train.list
+++ /dev/null
@@ -1 +0,0 @@
-mnist_train.txt
diff --git a/doc/api/v1/index_cn.rst b/doc/api/v1/index_cn.rst
deleted file mode 100644
index cf146dc088e3905a751ff55c26fd82ef0ba02c89..0000000000000000000000000000000000000000
--- a/doc/api/v1/index_cn.rst
+++ /dev/null
@@ -1,37 +0,0 @@
-API中文手册
-============
-
-DataProvider API
-----------------
-
-..  toctree::
-    :maxdepth: 1
-
-    data_provider/dataprovider_cn.rst
-    data_provider/pydataprovider2_cn.rst
-
-..  _api_trainer_config:
-
-Model Config API
-----------------
-
-..  toctree::
-    :maxdepth: 1
-
-    trainer_config_helpers/optimizers.rst
-    trainer_config_helpers/data_sources.rst
-    trainer_config_helpers/layers.rst
-    trainer_config_helpers/activations.rst
-    trainer_config_helpers/poolings.rst
-    trainer_config_helpers/networks.rst
-    trainer_config_helpers/evaluators.rst
-    trainer_config_helpers/attrs.rst
-
-
-Applications API
-----------------
-
-..  toctree::
-    :maxdepth: 1
-
-    predict/swig_py_paddle_cn.rst
diff --git a/doc/api/v1/index_en.rst b/doc/api/v1/index_en.rst
deleted file mode 100644
index 10c297a71d6988c002de868e804ed9ee2345fbd7..0000000000000000000000000000000000000000
--- a/doc/api/v1/index_en.rst
+++ /dev/null
@@ -1,37 +0,0 @@
-API
-===
-
-DataProvider API
-----------------
-
-..  toctree::
-    :maxdepth: 1
-
-    data_provider/dataprovider_en.rst
-    data_provider/pydataprovider2_en.rst
-
-..  _api_trainer_config:
-
-Model Config API
-----------------
-
-..  toctree::
-    :maxdepth: 1
-
-    trainer_config_helpers/optimizers.rst
-    trainer_config_helpers/data_sources.rst
-    trainer_config_helpers/layers.rst
-    trainer_config_helpers/activations.rst 
-    trainer_config_helpers/poolings.rst
-    trainer_config_helpers/networks.rst
-    trainer_config_helpers/evaluators.rst
-    trainer_config_helpers/attrs.rst
-
-
-Applications API
-----------------
-
-..  toctree::
-    :maxdepth: 1
-
-    predict/swig_py_paddle_en.rst
diff --git a/doc/api/v1/predict/src/predict_sample.py b/doc/api/v1/predict/src/predict_sample.py
deleted file mode 100644
index 51349250e80ce11e476d952991f0d046f65286b4..0000000000000000000000000000000000000000
--- a/doc/api/v1/predict/src/predict_sample.py
+++ /dev/null
@@ -1,135 +0,0 @@
-# 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.
-
-from py_paddle import swig_paddle, DataProviderConverter
-from paddle.trainer.PyDataProvider2 import dense_vector
-from paddle.trainer.config_parser import parse_config
-
-TEST_DATA = [[[
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.215686, 0.533333, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.67451, 0.992157, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0.070588, 0.886275, 0.992157, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.192157,
-    0.070588, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.670588, 0.992157,
-    0.992157, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.117647, 0.933333, 0.858824, 0.313725,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.090196, 0.858824, 0.992157, 0.831373, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0.141176, 0.992157, 0.992157, 0.611765, 0.054902, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0.258824, 0.992157, 0.992157, 0.529412, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0.368627, 0.992157, 0.992157, 0.419608, 0.003922, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0.094118, 0.835294, 0.992157, 0.992157, 0.517647, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0.603922, 0.992157, 0.992157, 0.992157, 0.603922,
-    0.545098, 0.043137, 0, 0, 0, 0, 0, 0, 0, 0.447059, 0.992157, 0.992157,
-    0.956863, 0.062745, 0, 0, 0, 0, 0, 0, 0, 0, 0.011765, 0.666667, 0.992157,
-    0.992157, 0.992157, 0.992157, 0.992157, 0.745098, 0.137255, 0, 0, 0, 0, 0,
-    0.152941, 0.866667, 0.992157, 0.992157, 0.521569, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0.070588, 0.992157, 0.992157, 0.992157, 0.803922, 0.352941, 0.745098,
-    0.992157, 0.945098, 0.317647, 0, 0, 0, 0, 0.580392, 0.992157, 0.992157,
-    0.764706, 0.043137, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.070588, 0.992157, 0.992157,
-    0.776471, 0.043137, 0, 0.007843, 0.27451, 0.882353, 0.941176, 0.176471, 0,
-    0, 0.180392, 0.898039, 0.992157, 0.992157, 0.313725, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0.070588, 0.992157, 0.992157, 0.713725, 0, 0, 0, 0, 0.627451,
-    0.992157, 0.729412, 0.062745, 0, 0.509804, 0.992157, 0.992157, 0.776471,
-    0.035294, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.494118, 0.992157, 0.992157,
-    0.968627, 0.168627, 0, 0, 0, 0.423529, 0.992157, 0.992157, 0.364706, 0,
-    0.717647, 0.992157, 0.992157, 0.317647, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0.533333, 0.992157, 0.984314, 0.945098, 0.603922, 0, 0, 0, 0.003922,
-    0.466667, 0.992157, 0.988235, 0.976471, 0.992157, 0.992157, 0.788235,
-    0.007843, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.686275, 0.882353, 0.364706, 0,
-    0, 0, 0, 0, 0, 0.098039, 0.588235, 0.992157, 0.992157, 0.992157, 0.980392,
-    0.305882, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.101961, 0.67451, 0.321569,
-    0, 0, 0, 0, 0, 0, 0, 0.105882, 0.733333, 0.976471, 0.811765, 0.713725, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.65098, 0.992157, 0.321569, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0.25098, 0.007843, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1,
-    0.94902, 0.219608, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0.968627, 0.764706, 0.152941, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.498039, 0.25098, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0
-]], [[
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0.298039, 0.333333, 0.333333, 0.333333, 0.337255,
-    0.333333, 0.333333, 0.109804, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0.027451, 0.223529, 0.776471, 0.964706, 0.988235, 0.988235, 0.988235,
-    0.992157, 0.988235, 0.988235, 0.780392, 0.098039, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0.14902, 0.698039, 0.988235, 0.992157, 0.988235, 0.901961,
-    0.87451, 0.568627, 0.882353, 0.976471, 0.988235, 0.988235, 0.501961, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.188235, 0.647059, 0.988235, 0.988235,
-    0.745098, 0.439216, 0.098039, 0, 0, 0, 0.572549, 0.988235, 0.988235,
-    0.988235, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.2, 0.933333, 0.992157,
-    0.941176, 0.247059, 0, 0, 0, 0, 0, 0, 0.188235, 0.898039, 0.992157,
-    0.992157, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.039216, 0.639216, 0.933333,
-    0.988235, 0.913725, 0.278431, 0, 0, 0, 0, 0, 0, 0, 0.113725, 0.843137,
-    0.988235, 0.988235, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.235294, 0.988235,
-    0.992157, 0.988235, 0.815686, 0.07451, 0, 0, 0, 0, 0, 0, 0, 0.333333,
-    0.988235, 0.988235, 0.552941, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.211765,
-    0.878431, 0.988235, 0.992157, 0.701961, 0.329412, 0.109804, 0, 0, 0, 0, 0,
-    0, 0, 0.698039, 0.988235, 0.913725, 0.145098, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0.188235, 0.890196, 0.988235, 0.988235, 0.745098, 0.047059, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0.882353, 0.988235, 0.568627, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.2,
-    0.933333, 0.992157, 0.992157, 0.992157, 0.447059, 0.294118, 0, 0, 0, 0, 0,
-    0, 0, 0, 0.447059, 0.992157, 0.768627, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0.623529, 0.988235, 0.988235, 0.988235, 0.988235, 0.992157, 0.47451, 0, 0,
-    0, 0, 0, 0, 0, 0.188235, 0.933333, 0.87451, 0.509804, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0.992157, 0.988235, 0.937255, 0.792157, 0.988235, 0.894118,
-    0.082353, 0, 0, 0, 0, 0, 0, 0.027451, 0.647059, 0.992157, 0.654902, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0.623529, 0.988235, 0.913725, 0.329412, 0.376471,
-    0.184314, 0, 0, 0, 0, 0, 0, 0.027451, 0.513725, 0.988235, 0.635294,
-    0.219608, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.196078, 0.929412, 0.988235,
-    0.988235, 0.741176, 0.309804, 0, 0, 0, 0, 0, 0, 0.529412, 0.988235,
-    0.678431, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.223529, 0.992157,
-    0.992157, 1, 0.992157, 0.992157, 0.992157, 0.992157, 1, 0.992157, 0.992157,
-    0.882353, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.023529,
-    0.478431, 0.654902, 0.658824, 0.952941, 0.988235, 0.988235, 0.988235,
-    0.992157, 0.988235, 0.729412, 0.278431, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0.196078, 0.647059, 0.764706, 0.764706, 0.768627,
-    0.580392, 0.047059, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-    0, 0, 0, 0, 0, 0, 0
-]]]
-
-
-def main():
-    conf = parse_config("./mnist_model/trainer_config.py", "")
-    print conf.data_config.load_data_args
-    network = swig_paddle.GradientMachine.createFromConfigProto(
-        conf.model_config)
-    assert isinstance(network, swig_paddle.GradientMachine)  # For code hint.
-    network.loadParameters("./mnist_model/")
-    converter = DataProviderConverter([dense_vector(784)])
-    inArg = converter(TEST_DATA)
-    print network.forwardTest(inArg)
-
-
-if __name__ == '__main__':
-    swig_paddle.initPaddle("--use_gpu=0")
-    main()
diff --git a/doc/api/v1/predict/swig_py_paddle_cn.rst b/doc/api/v1/predict/swig_py_paddle_cn.rst
deleted file mode 100644
index 42f333dba2e996e70572b3cda085b83e402ede8e..0000000000000000000000000000000000000000
--- a/doc/api/v1/predict/swig_py_paddle_cn.rst
+++ /dev/null
@@ -1,58 +0,0 @@
-.. _api_swig_py_paddle:
-
-基于Python的预测
-================
-
-预测流程
---------
-
-PaddlePaddle使用swig对常用的预测接口进行了封装，通过编译会生成py_paddle软件包，安装该软件包就可以在python环境下实现模型预测。可以使用python的 ``help()`` 函数查询软件包相关API说明。
-
-基于Python的模型预测，主要包括以下五个步骤。
-
-1. 初始化PaddlePaddle环境
-
-   在程序开始阶段，通过调用 ``swig_paddle.initPaddle()`` 并传入相应的命令行参数初始化PaddlePaddle。
-
-2. 解析模型配置文件
-   
-   初始化之后，可以通过调用 ``parse_config()`` 解析训练模型时用的配置文件。注意预测数据通常不包含label, 同时预测网络通常直接输出最后一层的结果而不是像训练网络一样再接一层cost layer，所以一般需要对训练用的模型配置文件稍作相应修改才能在预测时使用。
-
-3. 构造paddle.GradientMachine
-  
-   通过调用 ``swig_paddle.GradientMachine.createFromConfigproto()`` 传入上一步解析出来的模型配置就可以创建一个 ``GradientMachine``。
-
-4. 准备预测数据
-  
-   swig_paddle中的预测接口的参数是自定义的C++数据类型，py_paddle里面提供了一个工具类 ``DataProviderConverter`` 可以用于接收和PyDataProvider2一样的输入数据并转换成预测接口所需的数据类型。
-
-5. 模型预测
-  
-   通过调用 ``forwardTest()`` 传入预测数据，直接返回计算结果。
-
-
-预测Demo
---------
-
-如下是一段使用mnist model来实现手写识别的预测代码。完整的代码见 ``src_root/doc/ui/predict/predict_sample.py`` 。mnist model可以通过 ``src_root\demo\mnist`` 目录下的demo训练出来。
-
-..  literalinclude:: src/predict_sample.py
-    :language: python
-    :lines: 15-18,121-136
-
-
-Demo预测输出如下，其中value即为softmax层的输出。由于TEST_DATA包含两条预测数据，所以输出的value包含两个向量 。
-
-..  code-block:: text
-
-    [{'id': None, 'value': array(
-      [[  5.53018653e-09,   1.12194102e-05,   1.96644767e-09,
-          1.43630644e-02,   1.51111044e-13,   9.85625684e-01,
-          2.08823112e-10,   2.32777140e-08,   2.00186201e-09,
-          1.15501715e-08],
-       [  9.99982715e-01,   1.27787406e-10,   1.72296313e-05,
-          1.49316648e-09,   1.36540484e-11,   6.93137714e-10,
-          2.70634608e-08,   3.48565123e-08,   5.25639710e-09,
-          4.48684503e-08]], dtype=float32)}]
-
-
diff --git a/doc/api/v1/predict/swig_py_paddle_en.rst b/doc/api/v1/predict/swig_py_paddle_en.rst
deleted file mode 100644
index 1c628e6971fa5643e6a9ca629488049957686193..0000000000000000000000000000000000000000
--- a/doc/api/v1/predict/swig_py_paddle_en.rst
+++ /dev/null
@@ -1,59 +0,0 @@
-Python Prediction
-==================
-
-PaddlePaddle offers a set of clean prediction interfaces for python with the help of
-SWIG. The main steps of predict values in python are:
-
-* Parse training configurations
-* Construct GradientMachine
-* Prepare data
-* Predict
-
-Here is a sample python script that shows the typical prediction process for the
-MNIST classification problem. A complete sample code could be found at
-:code:`src_root/doc/ui/predict/predict_sample.py`.
-
-..  literalinclude:: src/predict_sample.py
-    :language: python
-    :lines: 15-18,90-100,101-104
-
-The module that does the most of the job is py_paddle.swig_paddle, it's
-generated by SWIG and has complete documents, for more details you can use
-python's :code:`help()` function. Let's walk through the above python script:
-
-* At the beginning, use :code:`swig_paddle.initPaddle()` to initialize
-  PaddlePaddle with command line arguments, for more about command line arguments
-  see :ref:`cmd_detail_introduction` .
-* Parse the configuration file that is used in training with :code:`parse_config()`.
-  Because data to predict with always have no label, and output of prediction work
-  normally is the output layer rather than the cost layer, so you should modify
-  the configuration file accordingly before using it in the prediction work.
-* Create a neural network with
-  :code:`swig_paddle.GradientMachine.createFromConfigproto()`, which takes the
-  parsed configuration :code:`conf.model_config` as argument. Then load the
-  trained parameters from the model with :code:`network.loadParameters()`.
-* Create a data converter object of utility class :code:`DataProviderConverter`.
-    - Note: As swig_paddle can only accept C++ matrices, we offer a utility
-      class DataProviderConverter that can accept the same input data with
-      PyDataProvider2, for more information please refer to document
-      of :ref:`api_pydataprovider2` .
-* Do the prediction with :code:`forwardTest()`, which takes the converted
-  input data and outputs the activations of the output layer.
-
-Here is a typical output:
-
-..  code-block:: text
-
-    [{'id': None, 'value': array([[  5.53018653e-09,   1.12194102e-05,   1.96644767e-09,
-          1.43630644e-02,   1.51111044e-13,   9.85625684e-01,
-          2.08823112e-10,   2.32777140e-08,   2.00186201e-09,
-          1.15501715e-08],
-       [  9.99982715e-01,   1.27787406e-10,   1.72296313e-05,
-          1.49316648e-09,   1.36540484e-11,   6.93137714e-10,
-          2.70634608e-08,   3.48565123e-08,   5.25639710e-09,
-          4.48684503e-08]], dtype=float32)}]
-
-:code:`value` is the output of the output layer, each row represents result of
-the corresponding row in the input data, each element represents activation of
-the corresponding neuron in the output layer.
-
diff --git a/doc/api/v2/fluid/layers.rst b/doc/api/v2/fluid/layers.rst
index 4c4713b17c608a567af519dd249821d6ec2ea8a7..f738bf15641d9fca0bfb0c10821de778ceee0d79 100644
--- a/doc/api/v2/fluid/layers.rst
+++ b/doc/api/v2/fluid/layers.rst
@@ -18,6 +18,11 @@ dynamic_lstm
 ..  autofunction:: paddle.v2.fluid.layers.dynamic_lstm
     :noindex:
 
+dynamic_gru
+-----------
+..  autofunction:: paddle.v2.fluid.layers.dynamic_gru
+    :noindex:
+
 data
 ----
 ..  autofunction:: paddle.v2.fluid.layers.data
@@ -500,6 +505,21 @@ swish
 ..  autofunction:: paddle.v2.fluid.layers.swish
     :noindex:
 
+im2sequence
+------
+..  autofunction:: paddle.v2.fluid.layers.im2sequence
+    :noindex:
+
+edit_distance
+---------------
+..  autofunction:: paddle.v2.fluid.layers.edit_distance_error
+    :noindex:
+
+ctc_greedy_decoder
+---------------
+..  autofunction:: paddle.v2.fluid.layers.ctc_greedy_decoder
+    :noindex:
+
 l2_normalize
 ------------
 ..  autofunction:: paddle.v2.fluid.layers.l2_normalize
@@ -510,6 +530,11 @@ sequence_reshape
 ..  autofunction:: paddle.v2.fluid.layers.sequence_reshape
     :noindex:
 
+row_conv
+--------
+..  autofunction:: paddle.v2.fluid.layers.row_conv
+    :noindex:
+
 multiplex
 ---------
 ..  autofunction:: paddle.v2.fluid.layers.multiplex
diff --git a/doc/design/csp.md b/doc/design/csp.md
new file mode 100644
index 0000000000000000000000000000000000000000..ba9cacfdea7dcf7c6499b562dfc58400d082f2c8
--- /dev/null
+++ b/doc/design/csp.md
@@ -0,0 +1,96 @@
+# Design Doc: CSP in PaddlePaddle Fluid
+
+## Motivation
+
+Concurrent programming is important for deep learning.  Few example applications are:
+
+1.  The main thread keeps reading the next mini-batch while another thread uses the GPU for computing.
+2.  The main thread performs the computation while another thread uploads the local gradients from each trainer to the parameter server.
+
+Most DL systems, including TensorFlow, Caffe2, and MxNet, can asynchronously execute operators in a graph. However, Fluid doesn't have the concept of a graph at all, as the design goal of Fluid is that of a programming language.
+
+## Concurrent Programming Models
+
+There were many concurrent programming models, implemented in various forms:
+
+| concurrent programming model | implementation |
+|-----|-----|
+| mutex | types and functions in standard libraries |
+| semaphore | types and functions in standard libraries |
+| communicating sequential processes (CSP) | Go programming language |
+| actor model | Erlang programming language |
+| message passing | MPI |
+| bulk synchronous parallel (BSP) | Pregel distributed programming framework |
+
+Since Fluid was designed to be a programming language, we would like to implement CSP in Fluid.
+
+### CSP v.s. Actor Model
+
+A well-known implementation of Actor Model is the Erlang programming language.  In Actor Model, *processes* could send messages to another process and receive messages from another process given the process IDs.  We can find the three ingredients, process with ID, send, and recv, in MPI too.  Indeed, we can rewrite Erlang programs in Python + MPI with possibly fewer lines of code.  Our concern with Actor Model is that it doesn't seem reasonable to implement process management in a programming language's runtime library; instead, it should be the operating systems' responsibility to manage processes and libraries like MPI for send/recv.
+
+## CSP in Fluid
+
+Fluid has two fundamental control-flows: *if-else* and *while*.  If we are to implement CSP, we need the following:
+
+1. a new data type: *channel* and operators *send* and *recv*,
+1. *goroutine* or thread, and
+1. a new control-flow: select.
+
+We also need Python wrappers for the above components.
+
+The type *channel* is conceptually the blocking queue.  In Go, its implemented is a [blocking circular queue](https://github.com/golang/go/blob/68ce117cf17b8debf5754bfd476345779b5b6616/src/runtime/chan.go#L31-L50), which supports send and recv.
+
+The `select` operation has been in OS kernels long before Go language.  All Unix kernels implement system calls *poll* and *select*.  They monitor multiple file descriptors to see if I/O is possible on any of them.  This takes O(N) time.  Since Linux 2.6, a new system call, *epoll*, can do the same in O(1) time.  In BSD systems, there is a similar system call *kqueue*.  Go's Linux implementation uses epoll.
+
+It might be a good idea to implement Fluid's select using epoll too.  In this design doc, we start from the O(N) way, so we could focus on Python binding and the syntax.
+
+### Type Channel
+
+Fluid supports many data types:
+
+1. Tensor,
+1. Row-sparse Tensor
+1. LoD Tensor,
+1. Tensor array, etc
+
+Each data type is registered in the [`framework.proto`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L117-L127) as an enum value.  To add a new type channel, we need to add a new type enum.
+
+To expose a C++ type to Python, we need to edit the [`pybind.cc`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc) file.  [Here](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc#L120-L164) is an example how we expose C++ class LoDTensor.
+
+## Syntax Design
+
+### Create Channel
+
+In Go, we create a channel by specifying the element type and buffer size:
+
+```go
+ch  := make(chan int)       // a channel without buffer
+ch1 := make(chan int, 100)  // a channel that can buffer 100 ints.
+```
+
+In Fluid, we should be able to do the same:
+
+```python
+ch  = fluid.make_chan(dtype=INT)
+ch1 = fluid.make_chan(dtype=INT, 100)
+```
+
+In addition to that, we want channels that can hold more complex element types, e.g., Tensors of float16:
+
+```python
+ch = fluid.make_chan(dtype=Tensor, etype=float16)
+```
+
+or Tensors of Tensors of float16 etc.
+
+The point here is that we need a consistent way to compose types, like in C++ we can have `Tensor<Tensor<...<float16>...> >`.
+
+### Send and Recv
+
+### Select
+
+## Example Programs
+
+### 1. RPC between Trainers and Parameter Servers
+
+### 2. Concurrent Minibatch Loading
diff --git a/doc/design/dist_refactor/parameter_server.md b/doc/design/dist_refactor/parameter_server.md
index 1094f06d461275a9ad4034d5e48b39856d967b71..805dd13048d41b995d2a01cda52b2ea33e4bbe1d 100644
--- a/doc/design/dist_refactor/parameter_server.md
+++ b/doc/design/dist_refactor/parameter_server.md
@@ -9,16 +9,16 @@ different purposes.
 
 ## Background
 
-The previous implementations of the parameter server does not run a
+The previous implementations of the parameter server do not run a
 fluid sub-program. Parameter initialization, optimizer computation, network
 communication and checkpointing are implemented twice on both the
-trainer and the parameter server.
+trainer as well as the parameter server.
 
-It would be great if we can write code once and use them on both the
-trainer and the parameter server: reduces code duplication and
-improves extensibility. Given that after the current refactor, we are
-representing everything as a computing graph on the
-trainer. Representing everything as a computing graph on the parameter
+It would be great if we can write code once and use them on both: the
+trainer and the parameter server, since this reduces code duplication and
+improves extensibility. Given that after the current refactoring, we are
+representing everything as a computation graph on the
+trainer. Representing everything as a computation graph on the parameter
 server becomes a natural extension.
 
 ## Design
@@ -30,9 +30,9 @@ into sub-programs to be scheduled on different nodes with the following
 steps:
 
 1. OP placement: the OPs will be placed on different nodes according
-   to heuristic that minimizes estimated total computation
+   to a heuristic that minimizes the estimated total computation
    time. Currently we will use a simple heuristic that puts parameter
-   varable on parameter server workers and everything else on trainer
+   variable on parameter server workers and everything else on trainer
    workers.
 1. Add communication OPs to enable the communication between nodes.
 
@@ -47,22 +47,22 @@ After converting:
 
 <img src="src/dist-graph.png" width="700"/>
 
-1. The parameter variable W and it's optimizer program are placed on the parameter server.
+1. The parameter variable W and its optimizer program are placed on the parameter server.
 1. Operators are added to the program.
    - *Send* sends data to the connected *Recv* operator.  The
 	 scheduler on the receive node will only schedule *Recv* operator
 	 to run when the *Send* operator has ran (the *Send* OP will mark
 	 the *Recv* OP runnable automatically).
-   - *Enueue* enqueues the input variable, it can block until space
+   - *Enqueue* enqueues the input variable, it can block until space
      become available in the queue.
    - *Dequeue* outputs configurable numbers of tensors from the
-     queue. It will block until the queue have the required number of
+     queue. It will block until the queue has the required number of
      tensors.
 
 
 ### Benefits
 
-- Model parallelism become easier to implement: it's an extension to
+- Model parallelism becomes easier to implement: it is an extension to
   the trainer - parameter server approach. We can have several "Transpilers"
   to achieve different goals.
 - User-defined optimizer is easier to add - user can now express it as
@@ -72,22 +72,22 @@ After converting:
 
 ### Challenges
 
-- It's important to balance the parameter shards of on multiple
-  parameter server. If a single parameter is very big (some
+- It is important to balance the parameter shards on multiple
+  parameter servers. If a single parameter is very big (for example: some
   word-embedding, fully connected, softmax layer), we need to
   automatically partition the single parameter onto different
   parameter servers when possible (only element-wise optimizer depends
   on the parameter variable).
-- In the "Aync SGD" figure, the "W" variable on the parameter server
-  could be read and wrote concurrently. See
+- In the "Async SGD" figure, the "W" variable on the parameter server
+  could be read and written concurrently. See
   [here](https://github.com/PaddlePaddle/Paddle/pull/6394) for more
-  details about concurrent program in fluid.
+  details about concurrent program in Fluid.
 
 ### Discussion
 
 - Can the Enqueue OP be implemented under our current tensor design
-  (puts the input tensor into the queue tensor)?
-- *Dequeue* OP will have variable numbers of output (depends on the
+  (put the input tensor into the queue tensor)?
+- *Dequeue* OP will have variable numbers of output (depending on the
   `min_count` attribute), does our current design support it? (similar
   question for the *Add* OP)
 
diff --git a/doc/design/ops/sequence_decoder.md b/doc/design/ops/sequence_decoder.md
index 9db5fb8e9a9f89b004bf71ddc064cd976c0d0bee..c4a9bbeeefca0e05c335dd60233691e8bac33015 100644
--- a/doc/design/ops/sequence_decoder.md
+++ b/doc/design/ops/sequence_decoder.md
@@ -22,7 +22,7 @@ The current `LoDTensor` is designed to store levels of variable-length sequences
 The integers in each level represent the begin and end (not inclusive) offset of a sequence **in the underlying tensor**,
 let's call this format the **absolute-offset LoD** for clarity.
 
-The relative-offset LoD can retrieve any sequence very quickly but fails to represent empty sequences, for example, a two-level LoD is as follows
+The absolute-offset LoD can retrieve any sequence very quickly 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]]
@@ -119,7 +119,7 @@ def generate():
         encoder_ctx_expanded = pd.lod_expand(encoder_ctx, target_word)
         decoder_input = pd.fc(
             act=pd.activation.Linear(),
-            input=[target_word, encoder_ctx],
+            input=[target_word, encoder_ctx_expanded],
             size=3 * decoder_dim)
         gru_out, cur_mem = pd.gru_step(
             decoder_input, mem=decoder_mem, size=decoder_dim)
diff --git a/doc/getstarted/build_and_install/docker_install_cn.rst b/doc/getstarted/build_and_install/docker_install_cn.rst
index bae42593ddc6f7a7eb47d603752ad6efa9820b45..98fada7bdb46f4dd2927d6f93bcbcebbe7d18604 100644
--- a/doc/getstarted/build_and_install/docker_install_cn.rst
+++ b/doc/getstarted/build_and_install/docker_install_cn.rst
@@ -25,14 +25,14 @@
 
   .. code-block:: bash
 
-     docker pull docker.paddlepaddle.org/paddle
+     docker pull docker.paddlepaddlehub.com/paddle
 
 下载GPU版本（cuda8.0_cudnn5_avx_mkl）的Docker镜像：
 
   .. code-block:: bash
 
      docker pull paddlepaddle/paddle:latest-gpu
-     docker pull docker.paddlepaddle.org/paddle:latest-gpu
+     docker pull docker.paddlepaddlehub.com/paddle:latest-gpu
 
 选择下载使用不同的BLAS库的Docker镜像：
 
@@ -49,7 +49,7 @@
 
      docker pull paddlepaddle/paddle:[tag]
      # 比如：
-     docker pull docker.paddlepaddle.org/paddle:0.10.0-gpu
+     docker pull docker.paddlepaddlehub.com/paddle:0.11.0-gpu
 
 .. _docker_run:
 
diff --git a/doc/getstarted/build_and_install/docker_install_en.rst b/doc/getstarted/build_and_install/docker_install_en.rst
index 56a7c68e4d39c45249fa55a964dc48b7081596a6..b1d0890b4cdddb77114a80276130afd07c22d270 100644
--- a/doc/getstarted/build_and_install/docker_install_en.rst
+++ b/doc/getstarted/build_and_install/docker_install_en.rst
@@ -26,14 +26,14 @@ For users in China, we provide a faster mirror:
 
   .. code-block:: bash
 
-     docker pull docker.paddlepaddle.org/paddle
+     docker pull docker.paddlepaddlehub.com/paddle
 
 Download GPU version (cuda8.0_cudnn5_avx_mkl) images:
 
   .. code-block:: bash
 
      docker pull paddlepaddle/paddle:latest-gpu
-     docker pull docker.paddlepaddle.org/paddle:latest-gpu
+     docker pull docker.paddlepaddlehub.com/paddle:latest-gpu
 
 Choose between different BLAS version:
 
@@ -53,7 +53,7 @@ and run:
 
      docker pull paddlepaddle/paddle:[tag]
      # i.e.
-     docker pull docker.paddlepaddle.org/paddle:0.10.0-gpu
+     docker pull docker.paddlepaddlehub.com/paddle:0.11.0-gpu
 
 .. _docker_run:
 
diff --git a/doc/howto/optimization/cpu_profiling.md b/doc/howto/optimization/cpu_profiling.md
index 1775374cf6e518586c28bbd8e04946c74df7e4c5..368af40cc7308cf6f4c609361078fe3ba02213ed 100644
--- a/doc/howto/optimization/cpu_profiling.md
+++ b/doc/howto/optimization/cpu_profiling.md
@@ -60,8 +60,7 @@ each column is as follows:
 | column | meaning |
 | --- | --- |
 | ncalls | the number of calls into a function |
-| tottime | the total execution time of the function, not including the
- execution time of other functions called by the function |
+| tottime | the total execution time of the function, not including the execution time of other functions called by the function |
 | percall | tottime divided by ncalls |
 | cumtime | the total execution time of the function, including the execution time of other functions being called |
 | percall | cumtime divided by ncalls |
diff --git a/doc/howto/usage/cluster/fluid_cluster_train_en.md b/doc/howto/usage/cluster/fluid_cluster_train_en.md
index a64004a7c4ea12bc0d949d7f11f3e26af62bf912..11904a6f71bb6ce37417aeffb8e408ec65961b12 100644
--- a/doc/howto/usage/cluster/fluid_cluster_train_en.md
+++ b/doc/howto/usage/cluster/fluid_cluster_train_en.md
@@ -2,27 +2,27 @@
 
 ## Introduction
 
-In this article, we'll explain how to config and run distributed training jobs with PaddlePaddle Fluid in a bare metal cluster.
+In this article, we'll explain how to configure and run distributed training jobs with PaddlePaddle Fluid in a bare metal cluster.
 
 ## Preparations
 
-### Get your cluster ready
+### Getting the cluster ready
 
-Prepare your computer nodes in the cluster. Nodes in this cluster can be of any specification that runs PaddlePaddle, and with a unique IP address assigned to it. Make sure they can communicate with each other.
+Prepare the compute nodes in the cluster. Nodes in this cluster can be of any specification that runs PaddlePaddle, and with a unique IP address assigned to it. Make sure they can communicate to each other.
 
 ### Have PaddlePaddle installed
 
 PaddlePaddle must be installed on all nodes. If you have GPU cards on your nodes, be sure to properly install drivers and CUDA libraries.
 
-PaddlePaddle build and installation guide can be found from [here](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html).
+PaddlePaddle build and installation guide can be found  [here](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html).
 
-### Update training script
+### Update the training script
 
 #### Non-cluster training script
 
 Let's take [Deep Learning 101](http://www.paddlepaddle.org/docs/develop/book/01.fit_a_line/index.html)'s first chapter: "fit a line" as an example.
 
-This demo's non-cluster version with fluid API is as follows:
+The non-cluster version of this demo with fluid API is as follows:
 
 ``` python
 import paddle.v2 as paddle
@@ -65,25 +65,25 @@ for pass_id in range(PASS_NUM):
 exit(1)
 ```
 
-We created a simple fully connected neural networks training program and handed it to the fluid executor to run for 100 passes.
+We created a simple fully-connected neural network training program and handed it to the fluid executor to run for 100 passes.
 
-Now let's try to convert it to a distributed version to run in a cluster.
+Now let's try to convert it to a distributed version to run on a cluster.
 
 #### Introducing parameter server
 
-As you see from the non-cluster version of training script, there is only one role in it: the trainer, who does the computing as well as holding parameters. In cluster training, since multi-trainers are working on the same task, they need one centralized place to hold and distribute parameters. This centralized place is called the Parameter Server in PaddlePaddle.
+As we can see from the non-cluster version of training script, there is only one role in the script: the trainer, that performs the computing as well as holds the parameters. In cluster training, since multi-trainers are working on the same task, they need one centralized place to hold and distribute parameters. This centralized place is called the Parameter Server in PaddlePaddle.
 
-![parameter server architect](src/trainer.png)
+![parameter server architecture](src/trainer.png)
 
-Parameter Server in fluid does not only hold parameters but is also assigned with a part of the program. Trainers communicate with parameter servers via send/receive OPs. For more tech detail, please refer to this [document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/dist_refactor/distributed_architecture.md).
+Parameter Server in fluid not only holds the parameters but is also assigned with a part of the program. Trainers communicate with parameter servers via send/receive OPs. For more technical details, please refer to  [this document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/dist_refactor/distributed_architecture.md).
 
-Now we need to create program for both trainers and parameter servers, the question is how?
+Now we need to create programs for both: trainers and parameter servers, the question is how?
 
 #### Slice the program
 
-Fluid provides a tool called "Distribute Transpiler" to automatically convert the non-cluster program into cluster program.
+Fluid provides a tool called "Distributed Transpiler" that automatically converts the non-cluster program into cluster program.
 
-The idea behind this tool is to find optimize OPs and gradient parameters, slice the program into 2 pieces and connect them with send/receive OP.
+The idea behind this tool is to find the optimize OPs and gradient parameters, slice the program into 2 pieces and connect them with send/receive OP.
 
 Optimize OPs and gradient parameters can be found from the return values of optimizer's minimize function.
 
@@ -94,9 +94,9 @@ To put them together:
 
 optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) #get optimize OPs and gradient parameters
 
-t = fluid.DistributeTranspiler() # create transpiler instance
+t = fluid.DistributeTranspiler() # create the transpiler instance
 # slice the program into 2 pieces with optimizer_ops and gradient parameters list, as well as pserver_endpoints, which is a comma separated list of [IP:PORT] and number of trainers
-t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2) 
+t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
 
 ... #create executor
 
@@ -119,7 +119,7 @@ for pass_id in range(100):
 
 ### E2E demo
 
-Please find the complete demo from [here](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py). In parameter server node run this in the command line:
+Please find the complete demo from [here](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py). In parameter server node run the following in the command line:
 
 ``` bash
 PSERVERS=192.168.1.2:6174 SERVER_ENDPOINT=192.168.1.2:6174 TRAINING_ROLE=PSERVER python notest_dist_fit_a_line.py
@@ -129,12 +129,12 @@ PSERVERS=192.168.1.2:6174 SERVER_ENDPOINT=192.168.1.2:6174 TRAINING_ROLE=PSERVER
 
 Wait until the prompt `Server listening on 192.168.1.2:6174`
 
-Then in 2 of your trainer node run this:
+Then in 2 of your trainer nodes run this:
 
 ``` bash
 PSERVERS=192.168.1.2:6174 SERVER_ENDPOINT=192.168.1.2:6174 TRAINING_ROLE=TRAINER python notest_dist_fit_a_line.py
 ```
 
-*the reason you need to run this command twice in 2 nodes is: in the script we set the trainer count to be 2. You can change this setting on line 50*
+*the reason you need to run this command twice in 2 nodes is because: in the script we set the trainer count to be 2. You can change this setting on line 50*
 
 Now you have 2 trainers and 1 parameter server up and running.
diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt
index 4a98ede278fad85ff2beef3c8e7dd158912f693a..3f9c132ef6ae03c7614e10484715676c8019821e 100644
--- a/paddle/CMakeLists.txt
+++ b/paddle/CMakeLists.txt
@@ -18,7 +18,7 @@ else()
     add_subdirectory(capi)
   endif()
 
-  if(Boost_FOUND)
+  if(NOT ANDROID AND NOT IOS)
     add_subdirectory(memory)
     add_subdirectory(platform)
     add_subdirectory(framework)
diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt
index afb55bdaaae8e6d46d5791991c92f2b61652ac55..b83007ac3bc6ed8713ca65fddabccfd292a2732f 100644
--- a/paddle/framework/CMakeLists.txt
+++ b/paddle/framework/CMakeLists.txt
@@ -1,7 +1,7 @@
 # ddim lib
 proto_library(framework_proto SRCS framework.proto)
 
-cc_library(ddim SRCS ddim.cc DEPS eigen3)
+cc_library(ddim SRCS ddim.cc DEPS eigen3 boost)
 cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
 nv_test(dim_test SRCS dim_test.cu DEPS ddim)
 
@@ -45,7 +45,7 @@ cc_test(data_layout_transform_test SRCS data_layout_transform_test.cc DEPS data_
 cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor
         framework_proto selected_rows data_device_transform data_type_transform data_layout_transform)
 
-cc_library(attribute SRCS attribute.cc DEPS framework_proto)
+cc_library(attribute SRCS attribute.cc DEPS framework_proto boost)
 cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc
 device_context)
 cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute)
@@ -88,3 +88,10 @@ cc_test(init_test SRCS init_test.cc DEPS init)
 
 cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
 cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
+      
+if(NOT WITH_C_API AND WITH_FLUID)
+  file(GLOB FRAMEWORK_HEADERS *.h)
+  install(FILES ${FRAMEWORK_HEADERS} DESTINATION include/paddle/framework)
+  install(FILES ${CMAKE_CURRENT_BINARY_DIR}/framework.pb.h DESTINATION include/paddle/framework)
+  install(FILES details/cow_ptr.h details/op_registry.h DESTINATION include/paddle/framework/details)
+endif()
diff --git a/paddle/framework/attribute.cc b/paddle/framework/attribute.cc
index b0fd4d2750eb2529706d871947332d39494505cd..5074e8f5a05ed4e824b3db7e506b30eb1b70c3fd 100644
--- a/paddle/framework/attribute.cc
+++ b/paddle/framework/attribute.cc
@@ -61,6 +61,9 @@ Attribute GetAttrValue(const proto::OpDesc::Attr& attr_desc) {
       }
       return val;
     }
+    case proto::AttrType::LONG: {
+      return attr_desc.l();
+    }
     default:
       PADDLE_THROW("Unsupport attr type %d", attr_desc.type());
   }
diff --git a/paddle/framework/attribute.h b/paddle/framework/attribute.h
index c1c63d9cb13acb195b3bc3b30088f5fa7daf2a3d..bcff9bc4c48f8f233b7f811640c2789f9618a972 100644
--- a/paddle/framework/attribute.h
+++ b/paddle/framework/attribute.h
@@ -168,6 +168,32 @@ struct ExtractAttribute<bool> {
   const std::string& attr_name_;
 };
 
+template <>
+struct ExtractAttribute<int64_t> {
+  explicit ExtractAttribute(const std::string& attr_name)
+      : attr_name_(attr_name) {}
+
+  int64_t* operator()(Attribute& attr) const {
+    if (attr.type() == typeid(int)) {  // NOLINT
+      int val = boost::get<int>(attr);
+      attr = static_cast<int64_t>(val);
+    } else if (attr.type() == typeid(float)) {  // NOLINT
+      int val = boost::get<float>(attr);
+      attr = static_cast<int64_t>(val);
+    }
+    int64_t* attr_value = nullptr;
+    try {
+      attr_value = &boost::get<int64_t>(attr);
+    } catch (boost::bad_get& bad_get) {
+      PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
+                   attr_name_, attr.type().name());
+    }
+    return attr_value;
+  }
+
+  const std::string& attr_name_;
+};
+
 // check whether a certain attribute fit its limits
 // an attribute can have more than one limits
 template <typename T>
diff --git a/paddle/framework/block_desc.cc b/paddle/framework/block_desc.cc
index 54498e175dacfa0a220e3d839f4feb02502b2c03..dd2ed87252102aee6d384f37365d19305f19b281 100644
--- a/paddle/framework/block_desc.cc
+++ b/paddle/framework/block_desc.cc
@@ -75,7 +75,7 @@ std::vector<VarDesc *> BlockDesc::AllVars() const {
 
 OpDesc *BlockDesc::AppendOp() {
   need_update_ = true;
-  ops_.emplace_back(new OpDesc());
+  ops_.emplace_back(new OpDesc(this));
   return ops_.back().get();
 }
 
@@ -86,7 +86,7 @@ void BlockDesc::AppendAllocatedOp(std::unique_ptr<OpDesc> &&op_desc) {
 
 OpDesc *BlockDesc::PrependOp() {
   need_update_ = true;
-  ops_.emplace_front(new OpDesc());
+  ops_.emplace_front(new OpDesc(this));
   return ops_.front().get();
 }
 
@@ -153,7 +153,7 @@ BlockDesc::BlockDesc(ProgramDesc *prog, proto::BlockDesc *desc)
     vars_[var_desc.name()].reset(new VarDesc(var_desc));
   }
   for (const proto::OpDesc &op_desc : desc_->ops()) {
-    ops_.emplace_back(new OpDesc(op_desc, prog));
+    ops_.emplace_back(new OpDesc(op_desc, prog, this));
   }
 }
 
@@ -162,7 +162,7 @@ BlockDesc::BlockDesc(const BlockDesc &other, proto::BlockDesc *desc,
     : prog_(prog), desc_(desc) {
   need_update_ = true;
   for (auto &op : other.ops_) {
-    ops_.emplace_back(new OpDesc(*op));
+    ops_.emplace_back(new OpDesc(*op, this));
   }
 
   for (auto &it : other.vars_) {
diff --git a/paddle/framework/executor.cc b/paddle/framework/executor.cc
index 1382bfca19a674a404916a5c709276ce41219d2f..bd58c0a7f8161f6b45f2b500f3685e4028d97e96 100644
--- a/paddle/framework/executor.cc
+++ b/paddle/framework/executor.cc
@@ -116,8 +116,9 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
 
   for (auto& op_desc : block.AllOps()) {
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
-    VLOG(3) << op->DebugStringEx(local_scope);
+    VLOG(4) << op->DebugStringEx(local_scope);
     op->Run(*local_scope, place_);
+    VLOG(3) << op->DebugStringEx(local_scope);
     if (FLAGS_do_memory_benchmark) {
       VLOG(2) << "Memory used after operator " + op->Type() + " running: "
               << memory::memory_usage(place_);
diff --git a/paddle/framework/framework.proto b/paddle/framework/framework.proto
index ea69b87e2ac7dc587333b623c310182bb39eb452..5b6ef03f610926578d2c02dcf06f399f106a30a1 100644
--- a/paddle/framework/framework.proto
+++ b/paddle/framework/framework.proto
@@ -26,6 +26,7 @@ enum AttrType {
   BOOLEAN = 6;
   BOOLEANS = 7;
   BLOCK = 8;
+  LONG = 9;
 }
 
 // OpDesc describes an instance of a C++ framework::OperatorBase
@@ -44,6 +45,7 @@ message OpDesc {
     optional bool b = 10;
     repeated bool bools = 11;
     optional int32 block_idx = 12;
+    optional int64 l = 13;
   };
 
   message Var {
diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc
index b29f528f3f749efa3463125c774c2f4d4ebcbc7c..53b0d0fe083579da4f0bb600f292765aa2aa0d8a 100644
--- a/paddle/framework/lod_tensor.cc
+++ b/paddle/framework/lod_tensor.cc
@@ -107,9 +107,10 @@ LoD ToAbsOffset(const LoD &in) {
   // the lowest level stores relative offsets
   if (in.empty() || in.size() == 1) return in;
   LoD result = in;
-  for (int level = result.size() - 2; level >= 0; level--) {
-    for (auto &ele : result[level]) {
-      ele = result[level + 1][ele];
+  for (auto level = static_cast<int>(in.size() - 2); level >= 0; level--) {
+    for (size_t i = 0; i < in[level].size(); ++i) {
+      size_t index = in[level][i];
+      result[level][i] = result[level + 1][index];
     }
   }
   return result;
diff --git a/paddle/framework/op_desc.cc b/paddle/framework/op_desc.cc
index 1c0372bb16c04e155a68a0411939e4887322107a..f8df2cf97ad532f06cb1393b1a24cd789f8bde29 100644
--- a/paddle/framework/op_desc.cc
+++ b/paddle/framework/op_desc.cc
@@ -97,7 +97,7 @@ void OpDesc::CopyFrom(const OpDesc &op_desc) {
   need_update_ = true;
 }
 
-OpDesc::OpDesc(const proto::OpDesc &desc, ProgramDesc *prog)
+OpDesc::OpDesc(const proto::OpDesc &desc, ProgramDesc *prog, BlockDesc *block)
     : desc_(desc), need_update_(false) {
   // restore inputs_
   int input_size = desc_.inputs_size();
@@ -131,6 +131,7 @@ OpDesc::OpDesc(const proto::OpDesc &desc, ProgramDesc *prog)
       attrs_[attr_name] = prog->MutableBlock(bid);
     }
   }
+  this->block_ = block;
 }
 
 proto::OpDesc *OpDesc::Proto() {
@@ -282,6 +283,7 @@ struct SetAttrDescVisitor : public boost::static_visitor<void> {
     VectorToRepeated(v, attr_->mutable_bools());
   }
   void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->ID()); }
+  void operator()(int64_t v) const { attr_->set_l(v); }
   void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); }
 };
 
diff --git a/paddle/framework/op_desc.h b/paddle/framework/op_desc.h
index a5ffb162928bfd355d35d3f9b63aab59a88dd061..13695cff59f0bfd79c48eb28670ecc67a0309332 100644
--- a/paddle/framework/op_desc.h
+++ b/paddle/framework/op_desc.h
@@ -25,7 +25,6 @@ namespace framework {
 
 class BlockDesc;
 class ProgramDesc;
-
 class OpDesc {
  public:
   OpDesc() {}
@@ -33,7 +32,14 @@ class OpDesc {
   OpDesc(const std::string &type, const VariableNameMap &inputs,
          const VariableNameMap &outputs, const AttributeMap &attrs);
 
-  OpDesc(const proto::OpDesc &desc, ProgramDesc *prog);
+  OpDesc(const proto::OpDesc &desc, ProgramDesc *prog, BlockDesc *block);
+
+  explicit OpDesc(BlockDesc *block) : block_(block) {}
+
+  OpDesc(const OpDesc &other, BlockDesc *block) {
+    *this = other;
+    block_ = block;
+  }
 
   void CopyFrom(const OpDesc &op_desc);
 
@@ -117,6 +123,10 @@ class OpDesc {
 
   void Flush();
 
+  BlockDesc *Block() { return this->block_; }
+
+  void SetBlock(BlockDesc *block) { this->block_ = block; }
+
  private:
   template <typename MapType>
   static std::vector<typename MapType::key_type> MapKeys(const MapType &map) {
@@ -129,6 +139,7 @@ class OpDesc {
   }
 
   proto::OpDesc desc_;
+  BlockDesc *block_;  // not_own
   // input arg name => input variable names
   VariableNameMap inputs_;
   // output arg name => output variable names
diff --git a/paddle/framework/type_defs.h b/paddle/framework/type_defs.h
index d834d343759fa279a1444c6337956ffce1b9061a..1eedbbc419ab660f5ce00aa891ef80ca245bc0a8 100644
--- a/paddle/framework/type_defs.h
+++ b/paddle/framework/type_defs.h
@@ -35,7 +35,7 @@ using VariableNameMap = std::map<std::string, std::vector<std::string>>;
 using Attribute =
     boost::variant<boost::blank, int, float, std::string, std::vector<int>,
                    std::vector<float>, std::vector<std::string>, bool,
-                   std::vector<bool>, BlockDesc*>;
+                   std::vector<bool>, BlockDesc*, int64_t>;
 
 using AttributeMap = std::unordered_map<std::string, Attribute>;
 
diff --git a/paddle/framework/var_desc.h b/paddle/framework/var_desc.h
index fc482c467404a6b9dfed64c43871d91d3d10c766..9316b14bb695c185efd6db4296d422ef0c476d57 100644
--- a/paddle/framework/var_desc.h
+++ b/paddle/framework/var_desc.h
@@ -66,6 +66,8 @@ class VarDesc {
 
   std::string Name() const { return desc_.name(); }
 
+  void SetName(std::string name) { desc_.set_name(name); }
+
   void SetShape(const std::vector<int64_t> &dims);
 
   void SetDataType(proto::DataType data_type);
diff --git a/paddle/framework/variable_test.cc b/paddle/framework/variable_test.cc
index e4732d9718e2b46a068963d44c4c1e04024f2330..e5585c8724d712e273d086001b6cbc3d59c46ebe 100644
--- a/paddle/framework/variable_test.cc
+++ b/paddle/framework/variable_test.cc
@@ -12,19 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-/*
-  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 <memory>
 #include <string>
 
diff --git a/paddle/gserver/layers/PriorBox.cpp b/paddle/gserver/layers/PriorBox.cpp
index 337b9ba7bc0fc4e4bb80ee7b248d934f111379d5..8faf032f550836579522016b4fff3db7e94746e3 100644
--- a/paddle/gserver/layers/PriorBox.cpp
+++ b/paddle/gserver/layers/PriorBox.cpp
@@ -69,7 +69,7 @@ bool PriorBoxLayer::init(const LayerMap& layerMap,
   if (maxSize_.size() > 0) CHECK_EQ(minSize_.size(), maxSize_.size());
 
   // flip aspect ratios
-  for (int index = 0; index < tmp.size(); index++) {
+  for (unsigned index = 0; index < tmp.size(); index++) {
     real ar = tmp[index];
     if (fabs(ar - 1.) < 1e-6) continue;
     aspectRatio_.push_back(ar);
diff --git a/paddle/inference/CMakeLists.txt b/paddle/inference/CMakeLists.txt
index 02ca8a45a851d262eed6962a9a227b5009ef03a5..ae4d3fd2f58daf87a650428e04722581610ed780 100644
--- a/paddle/inference/CMakeLists.txt
+++ b/paddle/inference/CMakeLists.txt
@@ -1,12 +1,28 @@
-set(FLUID_CORE_MODULES
-    backward proto_desc paddle_memory executor prune init ${GLOB_OP_LIB})
+set(FLUID_CORE_MODULES proto_desc paddle_memory executor prune init)
 
 cc_library(paddle_fluid_api
     SRCS inference.cc
-    DEPS ${FLUID_CORE_MODULES})
+    DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
 
-# Merge all modules into a simgle static library
-cc_library(paddle_fluid DEPS paddle_fluid_api ${FLUID_CORE_MODULES})
+# Merge all modules into a single static library
+cc_library(paddle_fluid DEPS paddle_fluid_api ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
+
+# Create shared library
+add_library(paddle_fluid_shared SHARED inference.cc)
+
+target_circle_link_libraries(paddle_fluid_shared
+  ARCHIVE_START
+  ${GLOB_OP_LIB}
+  ARCHIVE_END
+  ${FLUID_CORE_MODULES})
+
+SET_TARGET_PROPERTIES(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
+
+# install library & headers
+if(NOT WITH_C_API AND WITH_FLUID)
+  install(FILES inference.h DESTINATION include/paddle/inference)
+  install(TARGETS paddle_fluid_shared DESTINATION lib)
+endif()
 
 add_executable(example example.cc)
 if(APPLE)
diff --git a/paddle/memory/CMakeLists.txt b/paddle/memory/CMakeLists.txt
index 8841c14ee083fccfd2271efd0c331805919a09d9..496098f80423854be62dc99b8601209ff6a6b182 100644
--- a/paddle/memory/CMakeLists.txt
+++ b/paddle/memory/CMakeLists.txt
@@ -1,7 +1,7 @@
 add_subdirectory(detail)
 
 cc_library(memory SRCS memory.cc DEPS place enforce)
-cc_library(memcpy SRCS memcpy.cc)
+cc_library(memcpy SRCS memcpy.cc DEPS place)
 
 cc_library(paddle_memory
     DEPS
@@ -14,3 +14,10 @@ cc_library(paddle_memory
     system_allocator)
 
 cc_test(memory_test SRCS memory_test.cc DEPS place paddle_memory)
+
+if(NOT WITH_C_API AND WITH_FLUID)
+  file(GLOB MEMORY_HEADERS *.h)
+  file(GLOB MEMORY_DETAIL_HEADERS detail/*.h)
+  install(FILES ${MEMORY_HEADERS} DESTINATION include/paddle/memory)
+  install(FILES ${MEMORY_DETAIL_HEADERS} DESTINATION include/paddle/memory/detail)
+endif()
diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt
index 6745a8da17723d663913a29f28e5ea9eedc0372a..15f7cb6b560590f55e276fde4900d2e3c0045fb8 100644
--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -156,6 +156,7 @@ op_library(parallel_do_op DEPS executor)
 # Regist multiple Kernel to pybind
 if (WITH_GPU)
 op_library(conv_op SRCS conv_op.cc conv_op.cu.cc conv_cudnn_op.cu.cc DEPS vol2col)
+op_library(edit_distance_op SRCS edit_distance_op.cc edit_distance_op.cu DEPS math_function)
 op_library(pool_op SRCS pool_op.cc pool_op.cu.cc pool_cudnn_op.cu.cc DEPS pooling)
 op_library(conv_transpose_op SRCS conv_transpose_op.cc conv_transpose_op.cu.cc
   conv_transpose_cudnn_op.cu.cc DEPS vol2col)
diff --git a/paddle/operators/beam_search_op.cc b/paddle/operators/beam_search_op.cc
index 4c71d66d22899d2cf6418935bf9358a0f73cec27..844ade40eb2a7ae239b079daa609f03b9e7a06df 100644
--- a/paddle/operators/beam_search_op.cc
+++ b/paddle/operators/beam_search_op.cc
@@ -24,8 +24,18 @@ namespace operators {
 void BeamSearch::operator()(const framework::LoDTensor &pre_ids,
                             framework::LoDTensor *selected_ids,
                             framework::LoDTensor *selected_scores) {
+  auto abs_lod = framework::ToAbsOffset(ids_->lod());
+  auto &high_level = abs_lod[lod_level_];
+
   auto items = SelectTopBeamSizeItems();
-  auto selected_items = ToMap(items);
+  auto selected_items = ToMap(items, high_level.back());
+  VLOG(3) << "selected_items:";
+  for (size_t i = 0; i < selected_items.size(); ++i) {
+    VLOG(3) << "offset:" << i;
+    for (auto &item : selected_items[i]) {
+      VLOG(3) << ItemToString(item);
+    }
+  }
   PruneEndidCandidates(pre_ids, &selected_items);
   // calculate the output tensor's height
   size_t num_instances = std::accumulate(
@@ -63,11 +73,12 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids,
   low_level.push_back(low_offset);
 
   // fill lod
-  auto abs_lod = framework::ToAbsOffset(ids_->lod());
-  auto &high_level = abs_lod[lod_level_];
   framework::LoD lod(2);
   lod[0].assign(high_level.begin(), high_level.end());
   lod[1].assign(low_level.begin(), low_level.end());
+  if (!framework::CheckLoD(lod)) {
+    PADDLE_THROW("lod %s is not right", framework::LoDToString(lod));
+  }
   selected_ids->set_lod(lod);
   selected_scores->set_lod(lod);
 }
@@ -90,13 +101,11 @@ int BeamSearch::PruneEndidCandidates(const framework::LoDTensor &pre_ids,
 }
 
 std::vector<std::vector<BeamSearch::Item>> BeamSearch::ToMap(
-    const std::vector<std::vector<Item>> &items) {
+    const std::vector<std::vector<Item>> &items, size_t element_num) {
   std::vector<std::vector<Item>> result;
+  result.resize(element_num);
   for (auto &entries : items) {
     for (const auto &item : entries) {
-      if (item.offset >= result.size()) {
-        result.resize(item.offset + 1);
-      }
       result[item.offset].push_back(item);
     }
   }
@@ -122,6 +131,14 @@ BeamSearch::SelectTopBeamSizeItems() {
     }
     result.emplace_back(items);
   }
+  VLOG(3) << "SelectTopBeamSizeItems result size " << result.size();
+  for (auto &items : result) {
+    VLOG(3) << "item set:";
+    for (auto &item : items) {
+      VLOG(3) << ItemToString(item);
+    }
+  }
+
   return result;
 }
 
@@ -159,6 +176,22 @@ bool BeamSearch::NextItemSet(std::vector<BeamSearch::Item> *items) {
   return true;
 }
 
+std::ostream &operator<<(std::ostream &os, const BeamSearch::Item &item) {
+  os << "{";
+  os << "offset: " << item.offset << ", ";
+  os << "id: " << item.id << ", ";
+  os << "score: " << item.score << "";
+  os << "}";
+
+  return os;
+}
+
+std::string ItemToString(const BeamSearch::Item &item) {
+  std::ostringstream stream;
+  stream << item;
+  return stream.str();
+}
+
 class BeamSearchProtoAndCheckerMaker
     : public framework::OpProtoAndCheckerMaker {
  public:
@@ -186,8 +219,40 @@ class BeamSearchProtoAndCheckerMaker
   }
 };
 
+class BeamSearchInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *context) const override {
+    for (const std::string &arg :
+         std::vector<std::string>({"pre_ids", "ids", "scores"})) {
+      PADDLE_ENFORCE(context->HasInput(arg),
+                     "BeamSearch need input argument '%s'", arg);
+    }
+    for (const std::string &arg :
+         std::vector<std::string>({"selected_ids", "selected_scores"})) {
+      PADDLE_ENFORCE(context->HasOutput(arg),
+                     "BeamSearch need output argument '%s'", arg);
+    }
+  }
+};
+
+class BeamSearchInferVarType : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc &op_desc,
+                  framework::BlockDesc *block) const override {
+    for (auto &o : op_desc.Output("selected_ids")) {
+      block->Var(o)->SetType(framework::proto::VarDesc::LOD_TENSOR);
+    }
+    for (auto &o : op_desc.Output("selected_scores")) {
+      block->Var(o)->SetType(framework::proto::VarDesc::LOD_TENSOR);
+    }
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_WITHOUT_GRADIENT(beam_search, paddle::operators::BeamSearchOp,
-                             paddle::operators::BeamSearchProtoAndCheckerMaker);
+REGISTER_OPERATOR(beam_search, paddle::operators::BeamSearchOp,
+                  paddle::operators::BeamSearchProtoAndCheckerMaker,
+                  paddle::operators::BeamSearchInferShape,
+                  paddle::operators::BeamSearchInferVarType,
+                  paddle::framework::EmptyGradOpMaker);
diff --git a/paddle/operators/beam_search_op.h b/paddle/operators/beam_search_op.h
index 45d14d68fe8d1c4a84aa826e68e76692444765a8..7ad85874fcbd6ea48d688b32f2cc982d6b76d3c4 100644
--- a/paddle/operators/beam_search_op.h
+++ b/paddle/operators/beam_search_op.h
@@ -136,8 +136,6 @@ class BeamSearch {
   void operator()(const framework::LoDTensor& pre_ids,
                   framework::LoDTensor* selected_ids,
                   framework::LoDTensor* selected_scores);
-
- protected:
   /*
    * The basic items help to sort.
    */
@@ -155,6 +153,7 @@ class BeamSearch {
     score_t score;
   };
 
+ protected:
   /*
    * Delete all the records that follows the end token.
    */
@@ -166,7 +165,7 @@ class BeamSearch {
    * NOTE low performance
    */
   std::vector<std::vector<Item>> ToMap(
-      const std::vector<std::vector<Item>>& inputs);
+      const std::vector<std::vector<Item>>& inputs, size_t element_num);
 
   /*
    * For each source, select top beam_size records.
@@ -187,6 +186,10 @@ class BeamSearch {
   int end_id_{0};
 };
 
+std::ostream& operator<<(std::ostream& os, const BeamSearch::Item& item);
+
+std::string ItemToString(const BeamSearch::Item& item);
+
 class BeamSearchOp : public framework::OperatorBase {
  public:
   BeamSearchOp(const std::string& type,
@@ -203,7 +206,6 @@ class BeamSearchOp : public framework::OperatorBase {
 
   void Run(const framework::Scope& scope,
            const platform::Place& dev_place) const override {
-    LOG(INFO) << "run beam search op";
     auto ids_var = scope.FindVar(Input("ids"));
     auto scores_var = scope.FindVar(Input("scores"));
     auto pre_ids_var = scope.FindVar(Input("pre_ids"));
@@ -217,10 +219,8 @@ class BeamSearchOp : public framework::OperatorBase {
     size_t level = Attr<int>("level");
     size_t beam_size = Attr<int>("beam_size");
     int end_id = Attr<int>("end_id");
-    LOG(INFO) << "init beam search";
     BeamSearch alg(ids, scores, level, beam_size, end_id);
 
-    LOG(INFO) << "after beam search";
     auto selected_ids_var = scope.FindVar(Output("selected_ids"));
     auto selected_scores_var = scope.FindVar(Output("selected_scores"));
     PADDLE_ENFORCE_NOT_NULL(selected_ids_var);
@@ -229,9 +229,7 @@ class BeamSearchOp : public framework::OperatorBase {
         *selected_ids_var->GetMutable<framework::LoDTensor>();
     auto& selected_scores_tensor =
         *selected_scores_var->GetMutable<framework::LoDTensor>();
-    LOG(INFO) << "run beam search";
     alg(pre_ids, &selected_ids_tensor, &selected_scores_tensor);
-    LOG(INFO) << "finish beam search";
   }
 };
 
diff --git a/paddle/operators/bipartite_match_op.cc b/paddle/operators/bipartite_match_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..83c8778fe4cec4d9d80de691e117a39fdd92f494
--- /dev/null
+++ b/paddle/operators/bipartite_match_op.cc
@@ -0,0 +1,189 @@
+/* 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/framework/op_registry.h"
+#include "paddle/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+
+class BipartiteMatchOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("DistMat"),
+                   "Input(DistMat) of BipartiteMatch should not be null.");
+
+    auto dims = ctx->GetInputDim("DistMat");
+    PADDLE_ENFORCE_EQ(dims.size(), 2, "The rank of Input(DistMat) must be 2.");
+
+    ctx->SetOutputDim("ColToRowMatchIndices", dims);
+    ctx->SetOutputDim("ColToRowMatchDis", dims);
+  }
+};
+
+template <typename T>
+class BipartiteMatchKernel : public framework::OpKernel<T> {
+ public:
+  // The match_indices must be initialized to -1 at first.
+  // The match_dist must be initialized to 0 at first.
+  void BipartiteMatch(const Tensor& dist, int* match_indices,
+                      T* match_dist) const {
+    constexpr T kEPS = static_cast<T>(1e-6);
+    PADDLE_ENFORCE_EQ(dist.dims().size(), 2, "The rank of dist must be 2.");
+    int64_t row = dist.dims()[0];
+    int64_t col = dist.dims()[1];
+    auto* dist_data = dist.data<T>();
+    std::vector<int> row_pool;
+    for (int i = 0; i < row; ++i) {
+      row_pool.push_back(i);
+    }
+    while (row_pool.size() > 0) {
+      int max_idx = -1;
+      int max_row_idx = -1;
+      T max_dist = -1;
+      for (int64_t j = 0; j < col; ++j) {
+        if (match_indices[j] != -1) {
+          continue;
+        }
+        for (size_t k = 0; k < row_pool.size(); ++k) {
+          int m = row_pool[k];
+          // distance is 0 between m-th row and j-th column
+          if (dist_data[m * col + j] < kEPS) {
+            continue;
+          }
+          if (dist_data[m * col + j] > max_dist) {
+            max_idx = j;
+            max_row_idx = m;
+            max_dist = dist_data[m * col + j];
+          }
+        }
+      }
+      if (max_idx == -1) {
+        // Cannot find good match.
+        break;
+      } else {
+        PADDLE_ENFORCE_EQ(match_indices[max_idx], -1);
+        match_indices[max_idx] = max_row_idx;
+        match_dist[max_idx] = max_dist;
+        // Erase the row index.
+        row_pool.erase(
+            std::find(row_pool.begin(), row_pool.end(), max_row_idx));
+      }
+    }
+  }
+
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* dist_mat = context.Input<LoDTensor>("DistMat");
+    auto* match_indices = context.Output<Tensor>("ColToRowMatchIndices");
+    auto* match_dist = context.Output<Tensor>("ColToRowMatchDis");
+
+    auto& dev_ctx = context.device_context<platform::CPUDeviceContext>();
+
+    auto col = dist_mat->dims()[1];
+
+    int64_t n = dist_mat->lod().size() == 0UL
+                    ? 1
+                    : static_cast<int64_t>(dist_mat->lod().back().size() - 1);
+    if (dist_mat->lod().size()) {
+      PADDLE_ENFORCE_EQ(dist_mat->lod().size(), 1UL,
+                        "Only support 1 level of LoD.");
+    }
+    match_indices->mutable_data<int>({n, col}, context.GetPlace());
+    match_dist->mutable_data<T>({n, col}, context.GetPlace());
+
+    math::SetConstant<platform::CPUDeviceContext, int> iset;
+    iset(dev_ctx, match_indices, static_cast<int>(-1));
+    math::SetConstant<platform::CPUDeviceContext, T> tset;
+    tset(dev_ctx, match_dist, static_cast<T>(0));
+
+    int* indices = match_indices->data<int>();
+    T* dist = match_dist->data<T>();
+    if (n == 1) {
+      BipartiteMatch(*dist_mat, indices, dist);
+    } else {
+      auto lod = dist_mat->lod().back();
+      for (size_t i = 0; i < lod.size() - 1; ++i) {
+        Tensor one_ins = dist_mat->Slice(lod[i], lod[i + 1]);
+        BipartiteMatch(one_ins, indices + i * col, dist + i * col);
+      }
+    }
+  }
+};
+
+class BipartiteMatchOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  BipartiteMatchOpMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput(
+        "DistMat",
+        "(LoDTensor or Tensor) this input is a 2-D LoDTensor with shape "
+        "[K, M]. It is pair-wise distance matrix between the entities "
+        "represented by each row and each column. For example, assumed one "
+        "entity is A with shape [K], another entity is B with shape [M]. The "
+        "DistMat[i][j] is the distance between A[i] and B[j]. The bigger "
+        "the distance is, the better macthing the pairs are. Please note, "
+        "This tensor can contain LoD information to represent a batch of "
+        "inputs. One instance of this batch can contain different numbers of "
+        "entities.");
+    AddOutput("ColToRowMatchIndices",
+              "(Tensor) A 2-D Tensor with shape [N, M] in int type. "
+              "N is the batch size. If ColToRowMatchIndices[i][j] is -1, it "
+              "means B[j] does not match any entity in i-th instance. "
+              "Otherwise, it means B[j] is matched to row "
+              "ColToRowMatchIndices[i][j] in i-th instance. The row number of "
+              "i-th instance is saved in ColToRowMatchIndices[i][j].");
+    AddOutput("ColToRowMatchDis",
+              "(Tensor) A 2-D Tensor with shape [N, M] in float type. "
+              "N is batch size. If ColToRowMatchIndices[i][j] is -1, "
+              "ColToRowMatchDis[i][j] is also -1.0. Otherwise, assumed "
+              "ColToRowMatchIndices[i][j] = d, and the row offsets of each "
+              "instance are called LoD. Then "
+              "ColToRowMatchDis[i][j] = DistMat[d+LoD[i]][j]");
+    AddComment(R"DOC(
+This operator is a greedy bipartite matching algorithm, which is used to
+obtain the matching with the maximum distance based on the input
+distance matrix. For input 2D matrix, the bipartite matching algorithm can
+find the matched column for each row, also can find the matched row for
+each column. And this operator only calculate matched indices from column
+to row. For each instance, the number of matched indices is the number of
+of columns of the input ditance matrix.
+
+There are two outputs to save matched indices and distance.
+A simple description, this algothrim matched the best (maximum distance)
+row entity to the column entity and the matched indices are not duplicated
+in each row of ColToRowMatchIndices. If the column entity is not matched
+any row entity, set -1 in ColToRowMatchIndices.
+
+Please note that the input DistMat can be LoDTensor (with LoD) or Tensor.
+If LoDTensor with LoD, the height of ColToRowMatchIndices is batch size.
+If Tensor, the height of ColToRowMatchIndices is 1.
+
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(bipartite_match, ops::BipartiteMatchOp,
+                  ops::BipartiteMatchOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+REGISTER_OP_CPU_KERNEL(bipartite_match, ops::BipartiteMatchKernel<float>,
+                       ops::BipartiteMatchKernel<double>);
diff --git a/paddle/operators/conv_transpose_op.cc b/paddle/operators/conv_transpose_op.cc
index a2382a7e42eb9c5c6a8f13265b0e6173e6b05f76..089290a506db10f676c8d7eb92663d2cb56892af 100644
--- a/paddle/operators/conv_transpose_op.cc
+++ b/paddle/operators/conv_transpose_op.cc
@@ -160,8 +160,8 @@ Example:
        Output shape: $(N, C_{out}, H_{out}, W_{out})$
   Where
   $$
-       H_{out} = (H_{in} - 1) * strides[0] - 2 * paddings[0] + H_f \\
-       W_{out} = (W_{in} - 1) * strides[1] - 2 * paddings[1] + W_f
+       H_{out} = (H_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (H_f - 1) + 1 \\
+       W_{out} = (W_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (W_f - 1) + 1
   $$
 )DOC");
 }
@@ -249,9 +249,9 @@ Example:
        Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$
   Where
   $$
-       D_{out} = (D_{in} - 1) * strides[0] - 2 * paddings[0] + D_f \\
-       H_{out} = (H_{in} - 1) * strides[1] - 2 * paddings[1] + H_f \\
-       W_{out} = (W_{in} - 1) * strides[2] - 2 * paddings[2] + W_f
+       D_{out} = (D_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (D_f - 1) + 1 \\
+       H_{out} = (H_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (H_f - 1) + 1 \\
+       W_{out} = (W_{in} - 1) * strides[2] - 2 * paddings[2] + dilations[2] * (W_f - 1) + 1
   $$
 )DOC");
 }
diff --git a/paddle/operators/conv_transpose_op.h b/paddle/operators/conv_transpose_op.h
index a42ade41b165d1bfa00d2db0e45d40cf5d7b00bc..8c0d57afcd21d8622fb6316f7b988d79a45b57fe 100644
--- a/paddle/operators/conv_transpose_op.h
+++ b/paddle/operators/conv_transpose_op.h
@@ -141,9 +141,9 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
       if (data_dim == 2U) {
         // col2im: col_matrix -> dy
         // from (c * k_h * k_w, h * w) to (c, o_h, o_w)
-        col2im(dev_ctx, col, std::vector<int>{dilations[0], dilations[1]},
-               strides, std::vector<int>{paddings[0], paddings[1], paddings[0],
-                                         paddings[1]},
+        col2im(dev_ctx, col, dilations, strides,
+               std::vector<int>{paddings[0], paddings[1], paddings[0],
+                                paddings[1]},
                &output_batch);
       } else if (data_dim == 3U) {
         // col2vol: col_matrix -> dy
@@ -247,8 +247,7 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
         if (data_dim == 2U) {
           // im2col: dy -> col matrix
           // from (c, o_h, o_w) to (c * k_h * k_w, h * w)
-          im2col(dev_ctx, output_grad_batch,
-                 std::vector<int>{dilations[0], dilations[1]}, strides,
+          im2col(dev_ctx, output_grad_batch, dilations, strides,
                  std::vector<int>{paddings[0], paddings[1], paddings[0],
                                   paddings[1]},
                  &col);
diff --git a/paddle/operators/ctc_align_op.h b/paddle/operators/ctc_align_op.h
index 589413feb3dcbb7fea1f0a878b35d4bf714b5318..fed89aa1e899a2450b315f352b9695056ed13aec 100644
--- a/paddle/operators/ctc_align_op.h
+++ b/paddle/operators/ctc_align_op.h
@@ -51,7 +51,7 @@ class CTCAlignKernel : public framework::OpKernel<T> {
       T prev_token = -1;
       for (size_t i = input_lod[level][seq_idx];
            i < input_lod[level][seq_idx + 1]; ++i) {
-        if (input_data[i] != blank &&
+        if ((unsigned)input_data[i] != blank &&
             !(merge_repeated && input_data[i] == prev_token)) {
           output_data[output_idx] = input_data[i];
           ++output_idx;
diff --git a/paddle/operators/edit_distance_op.cc b/paddle/operators/edit_distance_op.cc
index 62a1fcebe7b7222ffceafc3ca2bc74e3998225f6..7e7dfc79eba5c9a75366415e5f4b3183653a5cc6 100644
--- a/paddle/operators/edit_distance_op.cc
+++ b/paddle/operators/edit_distance_op.cc
@@ -25,6 +25,8 @@ class EditDistanceOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(ctx->HasInput("Hyps"), "Input(Hyps) shouldn't be null.");
     PADDLE_ENFORCE(ctx->HasInput("Refs"), "Input(Refs) shouldn't be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) shouldn't be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("SequenceNum"),
+                   "Output(SequenceNum) shouldn't be null.");
     auto hyp_dims = ctx->GetInputDim("Hyps");
     auto ref_dims = ctx->GetInputDim("Refs");
     PADDLE_ENFORCE(hyp_dims.size() == 2 && hyp_dims[1] == 1,
@@ -34,6 +36,7 @@ class EditDistanceOp : public framework::OperatorWithKernel {
                    "Input(Refs) must be a 2-D LoDTensor with the 2nd dimension "
                    "equal to 1.");
     ctx->SetOutputDim("Out", ctx->GetInputDim("Refs"));
+    ctx->SetOutputDim("SequenceNum", {1});
   }
 
  protected:
@@ -54,6 +57,7 @@ class EditDistanceOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("Refs",
              "(2-D LoDTensor<int64_t>, 2nd dim. equal to 1) "
              "The indices for reference strings.");
+    AddOutput("SequenceNum", "The sequence count of current batch");
     AddAttr<bool>("normalized",
                   "(bool, default false) Indicated whether to normalize "
                   "the edit distance by the length of reference string.")
diff --git a/paddle/operators/edit_distance_op.cu b/paddle/operators/edit_distance_op.cu
index 338fd79bcc125b86c7764645c2fd8953d4477d2a..c3e116af086627d576c7a788caebd45667d70017 100644
--- a/paddle/operators/edit_distance_op.cu
+++ b/paddle/operators/edit_distance_op.cu
@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include <algorithm>
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/math_function.h"
 #include "paddle/platform/cuda_helper.h"
 #include "paddle/platform/gpu_info.h"
 
@@ -72,6 +73,8 @@ class EditDistanceGPUKernel : public framework::OpKernel<T> {
 
     auto* x1_t = ctx.Input<framework::LoDTensor>("Hyps");
     auto* x2_t = ctx.Input<framework::LoDTensor>("Refs");
+    auto* sequence_num = ctx.Output<framework::Tensor>("SequenceNum");
+    sequence_num->mutable_data<int64_t>(ctx.GetPlace());
 
     auto normalized = ctx.Attr<bool>("normalized");
     auto stream = reinterpret_cast<const platform::CUDADeviceContext&>(
@@ -88,7 +91,11 @@ class EditDistanceGPUKernel : public framework::OpKernel<T> {
                      "Reference string %d is empty.", i);
     }
 
-    auto num_strs = hyp_lod.size() - 1;
+    const size_t num_strs = hyp_lod.size() - 1;
+    math::SetConstant<platform::CUDADeviceContext, int64_t> set_constant;
+    set_constant(ctx.template device_context<platform::CUDADeviceContext>(),
+                 sequence_num, static_cast<int64_t>(num_strs));
+
     out_t->Resize({static_cast<int64_t>(num_strs), 1});
     out_t->mutable_data<T>(ctx.GetPlace());
     auto out = out_t->data<T>();
diff --git a/paddle/operators/edit_distance_op.h b/paddle/operators/edit_distance_op.h
index 4c5a29813ce39e42111c0ee5f3c16d5cefac4651..974299e604d22422e9024382e85c843ad831575d 100644
--- a/paddle/operators/edit_distance_op.h
+++ b/paddle/operators/edit_distance_op.h
@@ -16,7 +16,6 @@ limitations under the License. */
 #include <algorithm>
 #include "paddle/framework/eigen.h"
 #include "paddle/framework/op_registry.h"
-
 namespace paddle {
 namespace operators {
 
@@ -28,6 +27,8 @@ class EditDistanceKernel : public framework::OpKernel<T> {
 
     auto* x1_t = ctx.Input<framework::LoDTensor>("Hyps");
     auto* x2_t = ctx.Input<framework::LoDTensor>("Refs");
+    auto* sequence_num = ctx.Output<framework::Tensor>("SequenceNum");
+    int64_t* seq_num_data = sequence_num->mutable_data<int64_t>(ctx.GetPlace());
 
     auto normalized = ctx.Attr<bool>("normalized");
 
@@ -41,6 +42,7 @@ class EditDistanceKernel : public framework::OpKernel<T> {
                      "Reference string %d is empty.", i);
     }
     auto num_strs = hyp_lod.size() - 1;
+    *seq_num_data = static_cast<int64_t>(num_strs);
 
     out_t->Resize({static_cast<int64_t>(num_strs), 1});
     out_t->mutable_data<float>(ctx.GetPlace());
diff --git a/paddle/operators/im2sequence_op.cc b/paddle/operators/im2sequence_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..31baaedf6914b1a6939fc762491ef35013db4bb6
--- /dev/null
+++ b/paddle/operators/im2sequence_op.cc
@@ -0,0 +1,157 @@
+/* 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/im2sequence_op.h"
+
+namespace paddle {
+namespace operators {
+
+class Im2SequenceOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of Im2SequenceOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of Im2SequenceOp op should not be null.");
+
+    auto in_dim = ctx->GetInputDim("X");
+    PADDLE_ENFORCE_EQ(in_dim.size(), 4,
+                      "Input(X) format must be 4D tensor, eg., NCHW.");
+
+    auto kernels = ctx->Attrs().Get<std::vector<int>>("kernels");
+    auto strides = ctx->Attrs().Get<std::vector<int>>("strides");
+    auto paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
+
+    int batch_size = in_dim[0];
+    int img_channels = in_dim[1];
+    int img_height = in_dim[2];
+    int img_width = in_dim[3];
+
+    int output_height = OutputSize(img_height, kernels[0], paddings[0],
+                                   paddings[2], strides[0]);
+    int output_width =
+        OutputSize(img_width, kernels[1], paddings[1], paddings[3], strides[1]);
+
+    ctx->SetOutputDim("Out", {batch_size * output_height * output_width,
+                              img_channels * kernels[0] * kernels[1]});
+  }
+};
+
+class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  Im2SequenceOpMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(Tensor) The input tensor has NCHW format."
+             "N: batch size"
+             "C: channels"
+             "H: height"
+             "W: width");
+    AddOutput("Out", "(LodTensor) The output data of im2sequence op,");
+    AddAttr<std::vector<int>>("kernels",
+                              "(vector<int>), the "
+                              "kernels(kernel_height, kernel_width)");
+    AddAttr<std::vector<int>>("strides",
+                              "(vector<int> default:{1, 1}), the "
+                              "strides(h_stride, w_stride)")
+        .SetDefault({1, 1});
+    AddAttr<std::vector<int>>("paddings",
+                              "(vector<int> default:{0, 0, 0, 0}), the "
+                              "paddings(up_pad, left_pad, down_pad, right_pad)")
+        .SetDefault({0, 0, 0, 0});
+    AddComment(R"DOC(
+This op uses kernels to scan images and converts these images to sequences.
+After expanding, The number of time steps are output_height * output_width
+and the dimension of each time step is kernel_height * kernel_width * channels,
+in which:
+
+output_height =
+    1 + (padding_height + padding_down + img_height - kernel_height + stride_height - 1) /
+            stride_height;
+output_width =
+    1 + (padding_left + padding+right + img_width - kernel_width + stride_width - 1) /
+            stride_width;
+
+This op can be used after convolution neural network, and before recurrent neural network.
+
+Given:
+
+x = [[[[ 6.  2.  1.]
+       [ 8.  3.  5.]
+       [ 0.  2.  6.]]
+
+      [[ 2.  4.  4.]
+       [ 6.  3.  0.]
+       [ 6.  4.  7.]]]
+
+     [[[ 6.  7.  1.]
+       [ 5.  7.  9.]
+       [ 2.  4.  8.]]
+
+      [[ 1.  2.  1.]
+       [ 1.  3.  5.]
+       [ 9.  0.  8.]]]]
+x.dims = {2, 2, 3, 3}
+
+And:
+
+kernels = [2, 2]
+strides = [1, 1]
+paddings = [0, 0, 0, 0]
+
+Then:
+
+output.data = [[ 6.  2.  8.  3.  2.  4.  6.  3.]
+               [ 2.  1.  3.  5.  4.  4.  3.  0.]
+               [ 8.  3.  0.  2.  6.  3.  6.  4.]
+               [ 3.  5.  2.  6.  3.  0.  4.  7.]
+               [ 6.  7.  5.  7.  1.  2.  1.  3.]
+               [ 7.  1.  7.  9.  2.  1.  3.  5.]
+               [ 5.  7.  2.  4.  1.  3.  9.  0.]
+               [ 7.  9.  4.  8.  3.  5.  0.  8.]]
+output.dims = {8, 9}
+output.lod = [[0, 4, 8]]
+
+)DOC");
+  }
+};
+
+class Im2SequenceGradOp : 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) shouldn't be null.");
+    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(im2sequence, ops::Im2SequenceOp, ops::Im2SequenceOpMaker,
+            im2sequence_grad, ops::Im2SequenceGradOp);
+REGISTER_OP_CPU_KERNEL(
+    im2sequence,
+    ops::Im2SequenceKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CPU_KERNEL(
+    im2sequence_grad,
+    ops::Im2SequenceGradKernel<paddle::platform::CPUDeviceContext, float>);
diff --git a/paddle/operators/im2sequence_op.cu b/paddle/operators/im2sequence_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..9db7529112f2710d6ff4af2b444e304543486de3
--- /dev/null
+++ b/paddle/operators/im2sequence_op.cu
@@ -0,0 +1,25 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/im2sequence_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_CUDA_KERNEL(
+    im2sequence,
+    ops::Im2SequenceKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    im2sequence_grad,
+    ops::Im2SequenceGradKernel<paddle::platform::CUDADeviceContext, float>);
diff --git a/paddle/operators/im2sequence_op.h b/paddle/operators/im2sequence_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..aeb810015134babc132909b3e820fa8391233b1c
--- /dev/null
+++ b/paddle/operators/im2sequence_op.h
@@ -0,0 +1,135 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   You may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+
+#include "paddle/framework/data_layout.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/im2col.h"
+#include "paddle/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+
+inline int OutputSize(int input_size, int filter_size, int padding_0,
+                      int padding_1, int stride) {
+  const int output_size =
+      (input_size + padding_0 + padding_1 - filter_size) / stride + 1;
+  return output_size;
+}
+
+template <typename DeviceContext, typename T>
+class Im2SequenceKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    const Tensor* in = ctx.Input<Tensor>("X");
+    LoDTensor* out = ctx.Output<LoDTensor>("Out");
+    out->mutable_data<T>(ctx.GetPlace());
+    // TODO(wanghaoshuang): Add layout checker after 'set_layout'
+    // being available for python API
+    // PADDLE_ENFORCE_EQ(in->layout(), framework::DataLayout::kNCHW,
+    //                  "Input(X) layout must be NCHW");
+    auto in_dim = in->dims();
+    int batch_size = in_dim[0];
+    int img_channels = in_dim[1];
+    int img_height = in_dim[2];
+    int img_width = in_dim[3];
+
+    auto kernels = ctx.Attr<std::vector<int>>("kernels");
+    auto strides = ctx.Attr<std::vector<int>>("strides");
+    auto paddings = ctx.Attr<std::vector<int>>("paddings");
+    int output_height = OutputSize(img_height, kernels[0], paddings[0],
+                                   paddings[2], strides[0]);
+    int output_width =
+        OutputSize(img_width, kernels[1], paddings[1], paddings[3], strides[1]);
+
+    const std::vector<int> dilations({1, 1});
+
+    auto out_dims = out->dims();
+    out->Resize({batch_size, out->numel() / batch_size});
+    for (int i = 0; i < batch_size; i++) {
+      const Tensor src =
+          in->Slice(i, i + 1).Resize({img_channels, img_height, img_width});
+      Tensor dst = out->Slice(i, i + 1).Resize(
+          {output_height, output_width, img_channels, kernels[0], kernels[1]});
+
+      math::Im2ColFunctor<math::ColFormat::kOCF, DeviceContext, T> f;
+      auto& dev_ctx = ctx.template device_context<DeviceContext>();
+      f(dev_ctx, src, dilations, strides, paddings, &dst);
+    }
+    out->Resize(out_dims);
+
+    // set lod information
+    // TODO(wanghaoshuang): Move this to InferShape
+    framework::LoD lod(1);
+    lod[0].reserve(batch_size + 1);
+    for (int i = 0, offset = 0; i < batch_size + 1; ++i) {
+      lod[0][i] = offset;
+      offset += output_height * output_width;
+    }
+    out->set_lod(lod);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class Im2SequenceGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in = ctx.Input<Tensor>("X");
+    Tensor* d_out =
+        const_cast<Tensor*>(ctx.Input<Tensor>(framework::GradVarName("Out")));
+    auto* d_x = ctx.Output<Tensor>(framework::GradVarName("X"));
+    d_x->mutable_data<T>(ctx.GetPlace());
+
+    auto x_v = framework::EigenVector<T>::Flatten(*d_x);
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
+    x_v.device(place) = x_v.constant(0.0);
+
+    auto in_dim = in->dims();
+    int batch_size = in_dim[0];
+    int img_channels = in_dim[1];
+    int img_height = in_dim[2];
+    int img_width = in_dim[3];
+
+    auto kernels = ctx.Attr<std::vector<int>>("kernels");
+    auto strides = ctx.Attr<std::vector<int>>("strides");
+    auto paddings = ctx.Attr<std::vector<int>>("paddings");
+    int output_height = OutputSize(img_height, kernels[0], paddings[0],
+                                   paddings[2], strides[0]);
+    int output_width =
+        OutputSize(img_width, kernels[1], paddings[1], paddings[3], strides[1]);
+
+    const std::vector<int> dilations({1, 1});
+
+    auto d_out_dims = d_out->dims();
+    d_out->Resize({batch_size, d_out->numel() / batch_size});
+    for (int i = 0; i < batch_size; i++) {
+      Tensor dst =
+          d_x->Slice(i, i + 1).Resize({img_channels, img_height, img_width});
+      const Tensor src = d_out->Slice(i, i + 1).Resize(
+          {output_height, output_width, img_channels, kernels[0], kernels[1]});
+      math::Col2ImFunctor<math::ColFormat::kOCF, DeviceContext, T> f;
+      auto& dev_ctx = ctx.template device_context<DeviceContext>();
+      f(dev_ctx, src, dilations, strides, paddings, &dst);
+    }
+    d_out->Resize(d_out_dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/iou_similarity_op.cc b/paddle/operators/iou_similarity_op.cc
new file mode 100755
index 0000000000000000000000000000000000000000..c520b28b83e66dbf53d2e19985370be4a2f69e23
--- /dev/null
+++ b/paddle/operators/iou_similarity_op.cc
@@ -0,0 +1,96 @@
+/* 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/iou_similarity_op.h"
+
+namespace paddle {
+namespace operators {
+
+class IOUSimilarityOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of IOUSimilarityOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Y"),
+                   "Input(Y) of IOUSimilarityOp should not be null.");
+    auto x_dims = ctx->GetInputDim("X");
+    auto y_dims = ctx->GetInputDim("Y");
+
+    PADDLE_ENFORCE_EQ(x_dims.size(), 2UL, "The rank of Input(X) must be 2.");
+    PADDLE_ENFORCE_EQ(x_dims[1], 4UL, "The shape of X is [N, 4]");
+    PADDLE_ENFORCE_EQ(y_dims.size(), 2UL, "The rank of Input(Y) must be 2.");
+    PADDLE_ENFORCE_EQ(y_dims[1], 4UL, "The shape of Y is [M, 4]");
+
+    ctx->ShareLoD("X", /*->*/ "Out");
+    ctx->SetOutputDim("Out", framework::make_ddim({x_dims[0], y_dims[0]}));
+  }
+};
+
+class IOUSimilarityOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  IOUSimilarityOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(LoDTensor, default LoDTensor<float>) "
+             "Box list X is a 2-D LoDTensor with shape [N, 4] holds N boxes, "
+             "each box is represented as [xmin, ymin, xmax, ymax], "
+             "the shape of X is [N, 4]. [xmin, ymin] is the left top "
+             "coordinate of the box if the input is image feature map, they "
+             "are close to the origin of the coordinate system. "
+             "[xmax, ymax] is the right bottom coordinate of the box. "
+             "This tensor can contain LoD information to represent a batch "
+             "of inputs. One instance of this batch can contain different "
+             "numbers of entities.");
+    AddInput("Y",
+             "(Tensor, default Tensor<float>) "
+             "Box list Y holds M boxes, each box is represented as "
+             "[xmin, ymin, xmax, ymax], the shape of X is [N, 4]. "
+             "[xmin, ymin] is the left top coordinate of the box if the "
+             "input is image feature map, and [xmax, ymax] is the right "
+             "bottom coordinate of the box.");
+
+    AddOutput("Out",
+              "(LoDTensor, the lod is same as input X) The output of "
+              "iou_similarity op, a tensor with shape [N, M] "
+              "representing pairwise iou scores.");
+
+    AddComment(R"DOC(
+IOU Similarity Operator.
+Computes intersection-over-union (IOU) between two box lists.
+ Box list 'X' should be a LoDTensor and 'Y' is a common Tensor,
+ boxes in 'Y' are shared by all instance of the batched inputs of X.
+ Given two boxes A and B, the calculation of IOU is as follows:
+
+$$
+IOU(A, B) = 
+\frac{area(A\cap B)}{area(A)+area(B)-area(A\cap B)}
+$$
+
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(iou_similarity, ops::IOUSimilarityOp,
+                             ops::IOUSimilarityOpMaker);
+
+REGISTER_OP_CPU_KERNEL(
+    iou_similarity,
+    ops::IOUSimilarityKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::IOUSimilarityKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/iou_similarity_op.cu b/paddle/operators/iou_similarity_op.cu
new file mode 100755
index 0000000000000000000000000000000000000000..fa5052624618c35875b241419946f69b776c81d4
--- /dev/null
+++ b/paddle/operators/iou_similarity_op.cu
@@ -0,0 +1,21 @@
+/* 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/iou_similarity_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    iou_similarity,
+    ops::IOUSimilarityKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::IOUSimilarityKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/paddle/operators/iou_similarity_op.h b/paddle/operators/iou_similarity_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..e36177069d7b18ea23759f99c4679218fbfd32b8
--- /dev/null
+++ b/paddle/operators/iou_similarity_op.h
@@ -0,0 +1,90 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/framework/op_registry.h"
+#include "paddle/platform/for_range.h"
+
+template <typename T>
+inline HOSTDEVICE T IOUSimilarity(T xmin1, T ymin1, T xmax1, T ymax1, T xmin2,
+                                  T ymin2, T xmax2, T ymax2) {
+  constexpr T zero = static_cast<T>(0);
+  T area1 = (ymax1 - ymin1) * (xmax1 - xmin1);
+  T area2 = (ymax2 - ymin2) * (xmax2 - xmin2);
+  T inter_xmax = xmax1 > xmax2 ? xmax2 : xmax1;
+  T inter_ymax = ymax1 > ymax2 ? ymax2 : ymax1;
+  T inter_xmin = xmin1 > xmin2 ? xmin1 : xmin2;
+  T inter_ymin = ymin1 > ymin2 ? ymin1 : ymin2;
+  T inter_height = inter_ymax - inter_ymin;
+  T inter_width = inter_xmax - inter_xmin;
+  inter_height = inter_height > zero ? inter_height : zero;
+  inter_width = inter_width > zero ? inter_width : zero;
+  T inter_area = inter_width * inter_height;
+  T union_area = area1 + area2 - inter_area;
+  T sim_score = inter_area / union_area;
+  return sim_score;
+}
+
+template <typename T>
+struct IOUSimilarityFunctor {
+  IOUSimilarityFunctor(const T* x, const T* y, T* z, int cols)
+      : x_(x), y_(y), z_(z), cols_(static_cast<size_t>(cols)) {}
+
+  inline HOSTDEVICE void operator()(size_t row_id) const {
+    T x_min1 = x_[row_id * 4];
+    T y_min1 = x_[row_id * 4 + 1];
+    T x_max1 = x_[row_id * 4 + 2];
+    T y_max1 = x_[row_id * 4 + 3];
+    for (size_t i = 0; i < cols_; ++i) {
+      T x_min2 = y_[i * 4];
+      T y_min2 = y_[i * 4 + 1];
+      T x_max2 = y_[i * 4 + 2];
+      T y_max2 = y_[i * 4 + 3];
+
+      T sim = IOUSimilarity(x_min1, y_min1, x_max1, y_max1, x_min2, y_min2,
+                            x_max2, y_max2);
+
+      z_[row_id * cols_ + i] = sim;
+    }
+  }
+  const T* x_;
+  const T* y_;
+  T* z_;
+  const size_t cols_;
+};
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class IOUSimilarityKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    const framework::LoDTensor* in_x = ctx.Input<framework::LoDTensor>("X");
+    const framework::Tensor* in_y = ctx.Input<framework::Tensor>("Y");
+    framework::LoDTensor* out = ctx.Output<framework::LoDTensor>("Out");
+
+    int x_n = in_x->dims()[0];
+    int y_n = in_y->dims()[0];
+    IOUSimilarityFunctor<T> functor(in_x->data<T>(), in_y->data<T>(),
+                                    out->mutable_data<T>(ctx.GetPlace()), y_n);
+
+    platform::ForRange<DeviceContext> for_range(
+        static_cast<const DeviceContext&>(ctx.device_context()), x_n);
+    for_range(functor);
+  }
+};  // namespace operators
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc
index bb03def4391da80c6219f7863d300fd3c8d8c7ac..2405852f53d46356a474897d3a111d1c94eed081 100644
--- a/paddle/operators/lookup_table_op.cc
+++ b/paddle/operators/lookup_table_op.cc
@@ -66,6 +66,12 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker {
                   "(boolean, default false) "
                   "Sparse update")
         .SetDefault(false);
+    AddAttr<int64_t>("padding_idx",
+                     "(int64, default -1) "
+                     "If the value is -1, it makes no effect to lookup. "
+                     "Otherwise the given value indicates padding the output "
+                     "with zeros whenever lookup encounters it in Ids.")
+        .SetDefault(-1);
     AddComment(R"DOC(
 Lookup Table Operator.
 
diff --git a/paddle/operators/lookup_table_op.cu b/paddle/operators/lookup_table_op.cu
index 261a28da694bf551d8d9e630139680aebc4be51a..d97390fa1c53fa0bdf16ab34cb209b994621f83c 100644
--- a/paddle/operators/lookup_table_op.cu
+++ b/paddle/operators/lookup_table_op.cu
@@ -21,9 +21,11 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-template <typename T, int BlockDimX, int BlockDimY, int GridDimX>
+template <typename T, int BlockDimX, int BlockDimY, int GridDimX,
+          bool PaddingFlag>
 __global__ void LookupTable(T* output, const T* table, const int64_t* ids,
-                            const int64_t N, const int64_t K, const int64_t D) {
+                            const int64_t N, const int64_t K, const int64_t D,
+                            const int64_t padding_idx) {
   int idx = threadIdx.x;
   int idy = blockIdx.x + threadIdx.y * GridDimX;
 
@@ -34,7 +36,14 @@ __global__ void LookupTable(T* output, const T* table, const int64_t* ids,
     T* out = output + idy * D;
     const T* tab = table + id * D;
     for (int i = idx; i < D; i += BlockDimX) {
-      out[i] = tab[i];
+      if (PaddingFlag) {
+        if (id == padding_idx)
+          out[i] = static_cast<T>(0);
+        else
+          out[i] = tab[i];
+      } else {
+        out[i] = tab[i];
+      }
     }
     idy += BlockDimY * GridDimX;
   }
@@ -67,6 +76,7 @@ class LookupTableCUDAKernel : public framework::OpKernel<T> {
     auto* table_t = context.Input<LoDTensor>("W");
     auto* ids_t = context.Input<LoDTensor>("Ids");
     auto* output_t = context.Output<LoDTensor>("Out");
+    int64_t padding_idx = context.Attr<int64_t>("padding_idx");
 
     size_t N = table_t->dims()[0];
     size_t D = table_t->dims()[1];
@@ -77,10 +87,17 @@ class LookupTableCUDAKernel : public framework::OpKernel<T> {
 
     dim3 threads(128, 8);
     dim3 grids(8, 1);
-    LookupTable<
-        T, 128, 8,
-        8><<<grids, threads, 0, context.cuda_device_context().stream()>>>(
-        output, table, ids, N, K, D);
+
+    if (padding_idx == -1)
+      LookupTable<
+          T, 128, 8, 8,
+          false><<<grids, threads, 0, context.cuda_device_context().stream()>>>(
+          output, table, ids, N, K, D, padding_idx);
+    else
+      LookupTable<
+          T, 128, 8, 8,
+          true><<<grids, threads, 0, context.cuda_device_context().stream()>>>(
+          output, table, ids, N, K, D, padding_idx);
   }
 };
 
@@ -91,6 +108,8 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
     auto& dev_ctx =
         context.template device_context<platform::CUDADeviceContext>();
     bool is_sparse = context.Attr<bool>("is_sparse");
+    // Since paddings are not trainable and fixed in forward, the gradient of
+    // paddings makes no sense and we don't deal with it in backward.
     if (is_sparse) {
       auto* ids = context.Input<LoDTensor>("Ids");
       auto* table = context.Input<LoDTensor>("W");
diff --git a/paddle/operators/lookup_table_op.h b/paddle/operators/lookup_table_op.h
index 2fd3335868406455ec01f9ded6bacc7bda5e2a67..0842c422f7bfd3cad9b36dfdbab930f3cc4a8728 100644
--- a/paddle/operators/lookup_table_op.h
+++ b/paddle/operators/lookup_table_op.h
@@ -32,16 +32,30 @@ class LookupTableKernel : public framework::OpKernel<T> {
     auto* table_t = context.Input<LoDTensor>("W");      // float tensor
     auto* ids_t = context.Input<LoDTensor>("Ids");      // int tensor
     auto* output_t = context.Output<LoDTensor>("Out");  // float tensor
+    int64_t padding_idx = context.Attr<int64_t>("padding_idx");
 
     int N = table_t->dims()[0];
     int D = table_t->dims()[1];
     auto* ids = ids_t->data<int64_t>();
     auto* table = table_t->data<T>();
     auto* output = output_t->mutable_data<T>(context.GetPlace());
-    for (int64_t i = 0; i < ids_t->numel(); ++i) {
-      PADDLE_ENFORCE_LT(ids[i], N);
-      PADDLE_ENFORCE_GE(ids[i], 0);
-      memcpy(output + i * D, table + ids[i] * D, D * sizeof(T));
+
+    if (padding_idx == -1) {
+      for (int64_t i = 0; i < ids_t->numel(); ++i) {
+        PADDLE_ENFORCE_LT(ids[i], N);
+        PADDLE_ENFORCE_GE(ids[i], 0);
+        memcpy(output + i * D, table + ids[i] * D, D * sizeof(T));
+      }
+    } else {
+      for (int64_t i = 0; i < ids_t->numel(); ++i) {
+        if (ids[i] == padding_idx) {
+          memset(output + i * D, 0, D * sizeof(T));
+        } else {
+          PADDLE_ENFORCE_LT(ids[i], N);
+          PADDLE_ENFORCE_GE(ids[i], 0);
+          memcpy(output + i * D, table + ids[i] * D, D * sizeof(T));
+        }
+      }
     }
   }
 };
@@ -51,6 +65,8 @@ class LookupTableGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     bool is_sparse = context.Attr<bool>("is_sparse");
+    // Since paddings are not trainable and fixed in forward, the gradient of
+    // paddings makes no sense and we don't deal with it in backward.
     if (is_sparse) {
       auto* ids = context.Input<LoDTensor>("Ids");
       auto* table = context.Input<LoDTensor>("W");
diff --git a/paddle/operators/nce_op.cc b/paddle/operators/nce_op.cc
index 84ba3ead2b52547b989a4541f31ea31ffcce6c63..994ddf717e7a5b883d8071c6a47da0b4b4074f2e 100644
--- a/paddle/operators/nce_op.cc
+++ b/paddle/operators/nce_op.cc
@@ -124,7 +124,8 @@ class NCEOpMaker : public framework::OpProtoAndCheckerMaker {
                               "This attribute only be used in unitest. Classes "
                               "in this list wiil be used as negative classes "
                               "for every samples. Under normal conditions, "
-                              "user should avoid setting this attribute.");
+                              "user should avoid setting this attribute.")
+        .SetDefault({});
     AddComment(R"DOC(
 Compute and return the noise-contrastive estimation training loss.
 See [Noise-contrastive estimation: A new estimation principle for unnormalized statistical models](http://www.jmlr.org/proceedings/papers/v9/gutmann10a/gutmann10a.pdf).
diff --git a/paddle/operators/nce_op.h b/paddle/operators/nce_op.h
index e6b496f7896dcb412be8ff096fdccb2f0b682369..86fa13a649ce7fdcaad64e2609ceea2fb4d7e072 100644
--- a/paddle/operators/nce_op.h
+++ b/paddle/operators/nce_op.h
@@ -197,7 +197,8 @@ class NCEGradKernel : public framework::OpKernel<T> {
     // get d_x
     auto d_x = context.Output<Tensor>(framework::GradVarName("Input"));
     if (d_x != nullptr) {
-      d_x->mutable_data<T>(context.GetPlace());
+      auto* d_x_data = d_x->mutable_data<T>(context.GetPlace());
+      std::fill(d_x_data, d_x_data + d_x->numel(), 0.0);
       auto d_x_matrix = EigenMatrix<T>::From(*d_x);
       auto w_matrix = EigenMatrix<T>::From(*(context.Input<Tensor>("Weight")));
       for (int64_t i = 0; i < sample_labels->numel(); ++i) {
diff --git a/paddle/operators/parallel_do_op.cc b/paddle/operators/parallel_do_op.cc
index a00458ea068dd703d2c7f362511ed08bc212d2a8..09e808902f8fe3a7a07153d3432866c18e81dc7c 100644
--- a/paddle/operators/parallel_do_op.cc
+++ b/paddle/operators/parallel_do_op.cc
@@ -31,6 +31,7 @@ static constexpr char kParallelScopes[] = "parallel_scopes";
 static constexpr char kParallelBlock[] = "sub_block";
 
 using LoDTensor = framework::LoDTensor;
+using SelectedRows = framework::SelectedRows;
 
 static void SplitTensorAndMoveTensorToScopes(
     const framework::Scope &scope, std::vector<framework::Scope *> *sub_scopes,
@@ -64,6 +65,30 @@ static void SplitTensorAndMoveTensorToScopes(
   }
 }
 
+inline void CopyOrShare(const framework::Variable &src,
+                        const platform::Place &dst_place,
+                        framework::Variable *dst) {
+  if (src.IsType<LoDTensor>()) {
+    if (src.Get<LoDTensor>().place() == dst_place) {
+      dst->GetMutable<LoDTensor>()->ShareDataWith(src.Get<LoDTensor>());
+    } else {
+      Copy(src.Get<LoDTensor>(), dst_place, dst->GetMutable<LoDTensor>());
+    }
+  } else if (src.IsType<SelectedRows>()) {
+    auto &src_sr = src.Get<SelectedRows>();
+    auto *dst_sr = dst->GetMutable<SelectedRows>();
+    dst_sr->set_rows(src_sr.rows());
+    dst_sr->set_height(src_sr.height());
+    if (src_sr.value().place() == dst_place) {
+      dst_sr->mutable_value()->ShareDataWith(src_sr.value());
+    } else {
+      Copy(src_sr.value(), dst_place, dst_sr->mutable_value());
+    }
+  } else {
+    PADDLE_THROW("Expect LoDTensor/SelectedRows, get %s", src.Type().name());
+  }
+}
+
 void WaitOnPlace(const platform::Place place) {
   platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
   auto &dev_ctx = *pool.Get(place);
@@ -210,30 +235,30 @@ class ParallelDoGradOp : public framework::OperatorBase {
     }
     WaitOnPlaces(places);
 
-    // merge grad
+    AccumulateGrad(scope, place, sub_scopes, places);
+  }
+
+  void AccumulateGrad(const framework::Scope &scope,
+                      const platform::Place &place,
+                      const std::vector<framework::Scope *> &sub_scopes,
+                      const platform::PlaceList &places) const {
     for (auto &s : Outputs(framework::GradVarName(kParameters))) {
-      auto &result = sub_scopes[0]->FindVar(s)->Get<LoDTensor>();
       std::string tmp_name;
-      auto *tmp = sub_scopes[0]->Var(&tmp_name)->GetMutable<LoDTensor>();
+      auto *tmp = sub_scopes[0]->Var(&tmp_name);
 
       for (size_t i = 1; i < sub_scopes.size(); ++i) {
-        auto &tensor_to_merge = sub_scopes[i]->FindVar(s)->Get<LoDTensor>();
-        if (!(places[i] == places[0])) {
-          framework::Copy(tensor_to_merge, places[0], tmp);
-          WaitOnPlace(places[0]);
-        } else {
-          tmp->ShareDataWith(tensor_to_merge);
-        }
+        CopyOrShare(*sub_scopes[i]->FindVar(s), places[0], tmp);
+        WaitOnPlace(places[0]);
 
         auto sum_op = framework::OpRegistry::CreateOp(
             "sum", {{"X", {s, tmp_name}}}, {{"Out", {s}}},
             framework::AttributeMap{});
+        VLOG(3) << sum_op->DebugStringEx(sub_scopes[0]);
         sum_op->Run(*sub_scopes[0], places[0]);
         WaitOnPlace(places[0]);
       }
 
-      VLOG(3) << result;
-      framework::Copy(result, place, scope.FindVar(s)->GetMutable<LoDTensor>());
+      CopyOrShare(*sub_scopes[0]->FindVar(s), place, scope.FindVar(s));
     }
     WaitOnPlaces(places);
   }
@@ -289,7 +314,7 @@ class ParallelDoGradOpShapeInference : public framework::InferShapeBase {
 
     PADDLE_ENFORCE(ctx->HasInputs(kParameters));
     PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(kParameters)));
-    PADDLE_ENFORCE(ctx->HasInput(kInputs));
+    PADDLE_ENFORCE(ctx->HasInputs(kInputs));
 
     for (auto &s : output) {
       PADDLE_ENFORCE(ctx->HasInputs(s));
diff --git a/paddle/operators/prior_box_op.cc b/paddle/operators/prior_box_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..105ff4ac3e3ba889aad880f4204af15829c6da47
--- /dev/null
+++ b/paddle/operators/prior_box_op.cc
@@ -0,0 +1,154 @@
+/* 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/prior_box_op.h"
+
+namespace paddle {
+namespace operators {
+
+class PriorBoxOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"),
+                   "Input(Input) of PriorBoxOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Image"),
+                   "Input(Image) of PriorBoxOp should not be null.");
+
+    auto image_dims = ctx->GetInputDim("Image");
+    auto input_dims = ctx->GetInputDim("Input");
+    PADDLE_ENFORCE(image_dims.size() == 4, "The layout of image is NCHW.");
+    PADDLE_ENFORCE(input_dims.size() == 4, "The layout of input is NCHW.");
+
+    PADDLE_ENFORCE_LT(input_dims[2], image_dims[2],
+                      "The height of input must smaller than image.");
+
+    PADDLE_ENFORCE_LT(input_dims[3], image_dims[3],
+                      "The width of input must smaller than image.");
+
+    auto min_sizes = ctx->Attrs().Get<std::vector<int>>("min_sizes");
+    auto max_sizes = ctx->Attrs().Get<std::vector<int>>("max_sizes");
+    auto variances = ctx->Attrs().Get<std::vector<float>>("variances");
+    auto aspect_ratios = ctx->Attrs().Get<std::vector<float>>("aspect_ratios");
+    bool flip = ctx->Attrs().Get<bool>("flip");
+
+    PADDLE_ENFORCE_GT(min_sizes.size(), 0,
+                      "Size of min_sizes must be at least 1.");
+    for (size_t i = 0; i < min_sizes.size(); ++i) {
+      PADDLE_ENFORCE_GT(min_sizes[i], 0, "min_sizes[%d] must be positive.", i);
+    }
+
+    std::vector<float> aspect_ratios_vec;
+    ExpandAspectRatios(aspect_ratios, flip, aspect_ratios_vec);
+
+    int num_priors = aspect_ratios_vec.size() * min_sizes.size();
+    if (max_sizes.size() > 0) {
+      PADDLE_ENFORCE_EQ(max_sizes.size(), min_sizes.size(),
+                        "The number of min_size and max_size must be equal.");
+      for (size_t i = 0; i < min_sizes.size(); ++i) {
+        PADDLE_ENFORCE_GT(max_sizes[i], min_sizes[i],
+                          "max_size[%d] must be greater than min_size[%d].", i,
+                          i);
+        num_priors += 1;
+      }
+    }
+
+    PADDLE_ENFORCE_EQ(variances.size(), 4, "Must and only provide 4 variance.");
+    for (size_t i = 0; i < variances.size(); ++i) {
+      PADDLE_ENFORCE_GT(variances[i], 0.0,
+                        "variance[%d] must be greater than 0.", i);
+    }
+
+    const float step_h = ctx->Attrs().Get<float>("step_h");
+    PADDLE_ENFORCE_GT(step_h, 0.0, "step_h should be larger than 0.");
+    const float step_w = ctx->Attrs().Get<float>("step_w");
+    PADDLE_ENFORCE_GT(step_w, 0.0, "step_w should be larger than 0.");
+
+    std::vector<int64_t> dim_vec(4);
+    dim_vec[0] = input_dims[2];
+    dim_vec[1] = input_dims[3];
+    dim_vec[2] = num_priors;
+    dim_vec[3] = 4;
+    ctx->SetOutputDim("Boxes", framework::make_ddim(dim_vec));
+    ctx->SetOutputDim("Variances", framework::make_ddim(dim_vec));
+  }
+};
+
+class PriorBoxOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  PriorBoxOpMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Input",
+             "(Tensor, default Tensor<float>), "
+             "the input feature data of PriorBoxOp, The layout is NCHW.");
+    AddInput("Image",
+             "(Tensor, default Tensor<float>), "
+             "the input image data of PriorBoxOp, The layout is NCHW.");
+    AddOutput("Boxes",
+              "(Tensor, default Tensor<float>), the output prior boxes of "
+              "PriorBoxOp. The layout is [H, W, num_priors, 4]. "
+              "H is the height of input, W is the width of input, num_priors "
+              "is the box count of each position.");
+    AddOutput("Variances",
+              "(Tensor, default Tensor<float>), the expanded variances of "
+              "PriorBoxOp. The layout is [H, W, num_priors, 4]. "
+              "H is the height of input, W is the width of input, num_priors "
+              "is the box count of each position.");
+    AddAttr<std::vector<int>>("min_sizes", "(vector<int>) ",
+                              "List of min sizes of generated prior boxes.");
+    AddAttr<std::vector<int>>("max_sizes", "(vector<int>) ",
+                              "List of max sizes of generated prior boxes.");
+    AddAttr<std::vector<float>>(
+        "aspect_ratios", "(vector<float>) ",
+        "List of aspect ratios of generated prior boxes.");
+    AddAttr<std::vector<float>>(
+        "variances", "(vector<float>) ",
+        "List of variances to be encoded in prior boxes.");
+    AddAttr<bool>("flip", "(bool) ", "Whether to flip aspect ratios.")
+        .SetDefault(true);
+    AddAttr<bool>("clip", "(bool) ", "Whether to clip out-of-boundary boxes.")
+        .SetDefault(true);
+    AddAttr<float>("step_w",
+                   "Prior boxes step across width, 0 for auto calculation.")
+        .SetDefault(0.0);
+    AddAttr<float>("step_h",
+                   "Prior boxes step across height, 0 for auto calculation.")
+        .SetDefault(0.0);
+    AddAttr<float>("offset",
+                   "(float) "
+                   "Prior boxes center offset.")
+        .SetDefault(0.5);
+    AddComment(R"DOC(
+Prior box operator
+Generate prior boxes for SSD(Single Shot MultiBox Detector) algorithm.
+Each position of the input produce N prior boxes, N is determined by
+ the count of min_sizes, max_sizes and aspect_ratios, The size of the
+ box is in range(min_size, max_size) interval, which is generated in
+ sequence according to the aspect_ratios.
+
+Please get more information from the following papers:
+https://arxiv.org/abs/1512.02325.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(prior_box, ops::PriorBoxOp, ops::PriorBoxOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    prior_box, ops::PriorBoxOpKernel<paddle::platform::CPUPlace, float>,
+    ops::PriorBoxOpKernel<paddle::platform::CPUPlace, double>);
diff --git a/paddle/operators/prior_box_op.h b/paddle/operators/prior_box_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..e0a663ace8f38c2d08fd4714c1247d3313ffae3e
--- /dev/null
+++ b/paddle/operators/prior_box_op.h
@@ -0,0 +1,188 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/math_function.h"
+#include "paddle/platform/transform.h"
+
+namespace paddle {
+namespace operators {
+
+inline void ExpandAspectRatios(const std::vector<float>& input_aspect_ratior,
+                               bool flip,
+                               std::vector<float>& output_aspect_ratior) {
+  constexpr float epsilon = 1e-6;
+  output_aspect_ratior.clear();
+  output_aspect_ratior.push_back(1.);
+  for (size_t i = 0; i < input_aspect_ratior.size(); ++i) {
+    float ar = input_aspect_ratior[i];
+    bool already_exist = false;
+    for (size_t j = 0; j < output_aspect_ratior.size(); ++j) {
+      if (fabs(ar - output_aspect_ratior[j]) < epsilon) {
+        already_exist = true;
+        break;
+      }
+    }
+    if (!already_exist) {
+      output_aspect_ratior.push_back(ar);
+      if (flip) {
+        output_aspect_ratior.push_back(1. / ar);
+      }
+    }
+  }
+}
+
+template <typename T>
+struct ClipFunctor {
+  HOSTDEVICE T operator()(T in) const {
+    return std::min<T>(std::max<T>(in, 0.), 1.);
+  }
+};
+
+template <typename Place, typename T>
+class PriorBoxOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* input = ctx.Input<paddle::framework::Tensor>("Input");
+    auto* image = ctx.Input<paddle::framework::Tensor>("Image");
+    auto* boxes = ctx.Output<paddle::framework::Tensor>("Boxes");
+    auto* vars = ctx.Output<paddle::framework::Tensor>("Variances");
+
+    auto min_sizes = ctx.Attr<std::vector<int>>("min_sizes");
+    auto max_sizes = ctx.Attr<std::vector<int>>("max_sizes");
+    auto input_aspect_ratio = ctx.Attr<std::vector<float>>("aspect_ratios");
+    auto variances = ctx.Attr<std::vector<float>>("variances");
+    auto flip = ctx.Attr<bool>("flip");
+    auto clip = ctx.Attr<bool>("clip");
+
+    std::vector<float> aspect_ratios;
+    ExpandAspectRatios(input_aspect_ratio, flip, aspect_ratios);
+
+    T step_w = static_cast<T>(ctx.Attr<float>("step_w"));
+    T step_h = static_cast<T>(ctx.Attr<float>("step_h"));
+    T offset = static_cast<T>(ctx.Attr<float>("offset"));
+
+    auto img_width = image->dims()[3];
+    auto img_height = image->dims()[2];
+
+    auto feature_width = input->dims()[3];
+    auto feature_height = input->dims()[2];
+
+    T step_width, step_height;
+    if (step_w == 0 || step_h == 0) {
+      step_width = static_cast<T>(img_width) / feature_width;
+      step_height = static_cast<T>(img_height) / feature_height;
+    } else {
+      step_width = step_w;
+      step_height = step_h;
+    }
+
+    int num_priors = aspect_ratios.size() * min_sizes.size();
+    if (max_sizes.size() > 0) {
+      num_priors += max_sizes.size();
+    }
+
+    boxes->mutable_data<T>(ctx.GetPlace());
+    vars->mutable_data<T>(ctx.GetPlace());
+
+    auto e_boxes = framework::EigenTensor<T, 4>::From(*boxes);
+    for (int h = 0; h < feature_height; ++h) {
+      for (int w = 0; w < feature_width; ++w) {
+        T center_x = (w + offset) * step_width;
+        T center_y = (h + offset) * step_height;
+        T box_width, box_height;
+        int idx = 0;
+        for (size_t s = 0; s < min_sizes.size(); ++s) {
+          int min_size = min_sizes[s];
+          // first prior: aspect_ratio = 1, size = min_size
+          box_width = box_height = min_size;
+          // xmin
+          e_boxes(h, w, idx, 0) = (center_x - box_width / 2.) / img_width;
+          // ymin
+          e_boxes(h, w, idx, 1) = (center_y - box_height / 2.) / img_height;
+          // xmax
+          e_boxes(h, w, idx, 2) = (center_x + box_width / 2.) / img_width;
+          // ymax
+          e_boxes(h, w, idx, 3) = (center_y + box_height / 2.) / img_height;
+
+          idx++;
+          if (max_sizes.size() > 0) {
+            int max_size = max_sizes[s];
+            // second prior: aspect_ratio = 1,
+            // size = sqrt(min_size * max_size)
+            box_width = box_height = sqrt(min_size * max_size);
+            // xmin
+            e_boxes(h, w, idx, 0) = (center_x - box_width / 2.) / img_width;
+            // ymin
+            e_boxes(h, w, idx, 1) = (center_y - box_height / 2.) / img_height;
+            // xmax
+            e_boxes(h, w, idx, 2) = (center_x + box_width / 2.) / img_width;
+            // ymax
+            e_boxes(h, w, idx, 3) = (center_y + box_height / 2.) / img_height;
+            idx++;
+          }
+
+          // rest of priors
+          for (size_t r = 0; r < aspect_ratios.size(); ++r) {
+            float ar = aspect_ratios[r];
+            if (fabs(ar - 1.) < 1e-6) {
+              continue;
+            }
+            box_width = min_size * sqrt(ar);
+            box_height = min_size / sqrt(ar);
+            // xmin
+            e_boxes(h, w, idx, 0) = (center_x - box_width / 2.) / img_width;
+            // ymin
+            e_boxes(h, w, idx, 1) = (center_y - box_height / 2.) / img_height;
+            // xmax
+            e_boxes(h, w, idx, 2) = (center_x + box_width / 2.) / img_width;
+            // ymax
+            e_boxes(h, w, idx, 3) = (center_y + box_height / 2.) / img_height;
+            idx++;
+          }
+        }
+      }
+    }
+
+    if (clip) {
+      platform::Transform<platform::CPUDeviceContext> trans;
+      ClipFunctor<T> clip_func;
+      trans(ctx.template device_context<platform::CPUDeviceContext>(),
+            boxes->data<T>(), boxes->data<T>() + boxes->numel(),
+            boxes->data<T>(), clip_func);
+    }
+
+    framework::Tensor var_t;
+    var_t.mutable_data<T>(
+        framework::make_ddim({1, static_cast<int>(variances.size())}),
+        ctx.GetPlace());
+    auto var_et = framework::EigenTensor<T, 2>::From(var_t);
+    for (size_t i = 0; i < variances.size(); ++i) {
+      var_et(0, i) = variances[i];
+    }
+
+    int box_num = feature_height * feature_width * num_priors;
+    auto var_dim = vars->dims();
+    vars->Resize({box_num, static_cast<int>(variances.size())});
+
+    auto e_vars = framework::EigenMatrix<T, Eigen::RowMajor>::From(*vars);
+    e_vars = var_et.broadcast(Eigen::DSizes<int, 2>(box_num, 1));
+
+    vars->Resize(var_dim);
+  }
+};  // namespace operators
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/recv_op.cc b/paddle/operators/recv_op.cc
index 60360807351f1cf41c19aa06e30640e5cf473e07..593c35879ae2b3680b93ac5d8443110e61cb99fe 100644
--- a/paddle/operators/recv_op.cc
+++ b/paddle/operators/recv_op.cc
@@ -49,7 +49,7 @@ static void CreateTensorFromMessageType(framework::Variable *var,
     var->GetMutable<framework::SelectedRows>();
   } else {
     PADDLE_THROW(
-        "VraibleMessage type %d is not in "
+        "VariableMessage type %d is not in "
         "[LoDTensor, SelectedRows]",
         var_type);
   }
@@ -121,17 +121,17 @@ class RecvOp : public framework::OperatorBase {
         if (it != grad_list.end()) {
           param_var_name = param_list[it - grad_list.begin()];
         } else {
-          LOG(ERROR) << "grad have no paired param:" << grad_var_name;
+          LOG(ERROR) << "grad has no paired param:" << grad_var_name;
         }
-        VLOG(3) << "recved grad: " << grad_var_name
+        VLOG(3) << "received grad: " << grad_var_name
                 << " updating param: " << param_var_name;
         if (fan_in > 1) {
           grad_var_name = this->GetGradVarNameForTrainer(grad_var_name);
         }
         auto *var = recv_scope.FindVar(grad_var_name);
         if (var == nullptr) {
-          LOG(ERROR) << "can not find server side var: " << grad_var_name;
-          PADDLE_THROW("can not find server side var");
+          LOG(ERROR) << "Can not find server side var: " << grad_var_name;
+          PADDLE_THROW("Can not find server side var");
         }
         detail::DeserializeFromMessage(v.second, dev_ctx, var);
       }
@@ -165,7 +165,7 @@ class RecvOpMaker : public framework::OpProtoAndCheckerMaker {
     AddComment(R"DOC(
 Recv operator
 
-This operator will recv tensor from send_op
+This operator will recieve tensor from send_op
 )DOC");
     AddAttr<std::string>("endpoint",
                          "(string, default 127.0.0.1:6164)"
@@ -176,11 +176,11 @@ This operator will recv tensor from send_op
         kOptimizeBlock, "Serialized ProgramDesc string for recv to run.");
     AddAttr<std::vector<std::string>>(
         "ParamList", "type list of string",
-        "grad->param name mapping to find which param to optimize.")
+        "grad->param name mapping to find which parameters to optimize.")
         .SetDefault({});
     AddAttr<std::vector<std::string>>(
         "GradList", "type list of string",
-        "grad->param name mapping to find which param to optimize.")
+        "grad->param name mapping to find which parameters to optimize.")
         .SetDefault({});
     AddAttr<int>("Fanin", "type int",
                  "Number of trainers in the current cluster job")
diff --git a/paddle/operators/reduce_op.cc b/paddle/operators/reduce_op.cc
index 09b7091358e65221374a604122b742d763cfbafc..4a06babeda00f2420df80f81f876a0047a3285ef 100644
--- a/paddle/operators/reduce_op.cc
+++ b/paddle/operators/reduce_op.cc
@@ -190,10 +190,22 @@ REGISTER_OP(reduce_min, ops::ReduceOp, ops::ReduceMinOpMaker, reduce_min_grad,
 #define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor)         \
   REGISTER_OP_CPU_KERNEL(reduce_type,                                          \
                          ops::ReduceKernel<paddle::platform::CPUDeviceContext, \
-                                           float, ops::functor>);              \
+                                           float, ops::functor>,               \
+                         ops::ReduceKernel<paddle::platform::CPUDeviceContext, \
+                                           double, ops::functor>,              \
+                         ops::ReduceKernel<paddle::platform::CPUDeviceContext, \
+                                           int, ops::functor>,                 \
+                         ops::ReduceKernel<paddle::platform::CPUDeviceContext, \
+                                           int64_t, ops::functor>);            \
   REGISTER_OP_CPU_KERNEL(                                                      \
       reduce_type##_grad,                                                      \
       ops::ReduceGradKernel<paddle::platform::CPUDeviceContext, float,         \
+                            ops::grad_functor>,                                \
+      ops::ReduceGradKernel<paddle::platform::CPUDeviceContext, double,        \
+                            ops::grad_functor>,                                \
+      ops::ReduceGradKernel<paddle::platform::CPUDeviceContext, int,           \
+                            ops::grad_functor>,                                \
+      ops::ReduceGradKernel<paddle::platform::CPUDeviceContext, int64_t,       \
                             ops::grad_functor>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_CPU_KERNEL);
diff --git a/paddle/operators/reduce_op.cu b/paddle/operators/reduce_op.cu
index 1dd948ed8a79cce8468f2fe210b5636e7dd1f99e..4ed1e051db4df579afe1c1ca24a06fa1baf3e13a 100644
--- a/paddle/operators/reduce_op.cu
+++ b/paddle/operators/reduce_op.cu
@@ -20,10 +20,22 @@ namespace ops = paddle::operators;
 #define REGISTER_REDUCE_GPU_KERNEL(reduce_type, functor, grad_functor)    \
   REGISTER_OP_CUDA_KERNEL(                                                \
       reduce_type, ops::ReduceKernel<paddle::platform::CUDADeviceContext, \
-                                     float, ops::functor>);               \
+                                     float, ops::functor>,                \
+      ops::ReduceKernel<paddle::platform::CUDADeviceContext, double,      \
+                        ops::functor>,                                    \
+      ops::ReduceKernel<paddle::platform::CUDADeviceContext, int,         \
+                        ops::functor>,                                    \
+      ops::ReduceKernel<paddle::platform::CUDADeviceContext, int64_t,     \
+                        ops::functor>);                                   \
   REGISTER_OP_CUDA_KERNEL(                                                \
       reduce_type##_grad,                                                 \
       ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, float,   \
+                            ops::grad_functor>,                           \
+      ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, double,  \
+                            ops::grad_functor>,                           \
+      ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int,     \
+                            ops::grad_functor>,                           \
+      ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int64_t, \
                             ops::grad_functor>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_GPU_KERNEL);
diff --git a/paddle/operators/send_op.cc b/paddle/operators/send_op.cc
index 807533a6c6129af073c5efa7817bd5ad637fec23..5aa66c20eaf77959089100f8dcee55f2bc83a71a 100644
--- a/paddle/operators/send_op.cc
+++ b/paddle/operators/send_op.cc
@@ -62,13 +62,13 @@ class SendOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   SendOpMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "(Tensor) Input tensor to be send").AsDuplicable();
-    AddOutput("Out", "(Tensor) Output tensor to get from server")
+    AddInput("X", "(Tensor) Input tensor to be sent").AsDuplicable();
+    AddOutput("Out", "(Tensor) Output tensor to be received from server")
         .AsDuplicable();
     AddComment(R"DOC(
 Send operator
 
-This operator will send tensor to recv_op.
+This operator will send tensor to recv_op at the parameter server.
 )DOC");
     AddAttr<std::vector<std::string>>("endpoints",
                                       "(string vector, default 127.0.0.1:6164)"
diff --git a/paddle/operators/sequence_expand_op.h b/paddle/operators/sequence_expand_op.h
index 2ba628e9c37278025e31779ab0468db46f2ff40a..6021526eee8e0a1f58885f6de38b14048787a828 100644
--- a/paddle/operators/sequence_expand_op.h
+++ b/paddle/operators/sequence_expand_op.h
@@ -32,6 +32,7 @@ class SequenceExpandKernel : public framework::OpKernel<T> {
     const T* x_data = x->data<T>();
     auto x_dims = x->dims();
     auto* y = context.Input<LoDTensor>("Y");
+    PADDLE_ENFORCE(!y->lod().empty(), "y should have lod");
     PADDLE_ENFORCE_EQ(static_cast<size_t>(x_dims[0]),
                       y->lod().back().size() - 1,
                       "The size of last lod level in Input(Y)"
diff --git a/paddle/operators/sequence_reshape_op.h b/paddle/operators/sequence_reshape_op.h
index c6f528ab8a73294bb8ee91425f34e44c66f1932c..aaae7ab29281b72848515b80cc60931c13a294c9 100644
--- a/paddle/operators/sequence_reshape_op.h
+++ b/paddle/operators/sequence_reshape_op.h
@@ -35,7 +35,7 @@ class SequenceReshapeKernel : public framework::OpKernel<T> {
     PADDLE_ENFORCE_EQ(in_lod.size(), 1UL,
                       "Only support one level sequence now.");
     PADDLE_ENFORCE_EQ(
-        in_dims[0], in_lod[0].back(),
+        (uint64_t)in_dims[0], in_lod[0].back(),
         "Inconsistent size between X.shape[0] and X.lod()[0].back().");
 
     auto in_lod_l0 = in_lod[0];
diff --git a/paddle/operators/top_k_op.h b/paddle/operators/top_k_op.h
index e9cd9bbd4d964c28f305fb4ab4c4733ed27ebfff..bf42e15e6b234125d9ec24e8500367b9915213ab 100644
--- a/paddle/operators/top_k_op.h
+++ b/paddle/operators/top_k_op.h
@@ -22,6 +22,7 @@ namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
 
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
@@ -33,9 +34,9 @@ class TopkKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     // Get the top k elements of each row of input tensor
     // FIXME: only deal with matrix(2d tensor).
-    auto* input = ctx.Input<Tensor>("X");
-    auto* output = ctx.Output<Tensor>("Out");
-    auto* indices = ctx.Output<Tensor>("Indices");
+    auto* input = ctx.Input<LoDTensor>("X");
+    auto* output = ctx.Output<LoDTensor>("Out");
+    auto* indices = ctx.Output<LoDTensor>("Indices");
     // k is determined by Attr
     const size_t k = static_cast<int>(ctx.Attr<int>("k"));
 
diff --git a/paddle/operators/transpose_op.cc b/paddle/operators/transpose_op.cc
index 11615d806a61b3525d2ed50f5ea5940e8d61c8f8..c7ae162638ca5e929cca14c841cc3eceeea5f64e 100644
--- a/paddle/operators/transpose_op.cc
+++ b/paddle/operators/transpose_op.cc
@@ -59,44 +59,39 @@ class TransposeOpMaker : public framework::OpProtoAndCheckerMaker {
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput(
         "X",
-        "(Tensor)The input tensor, tensors with rank at most 6 are supported");
-    AddOutput("Out", "(Tensor)The output tensor");
+        "(Tensor) The input tensor, tensors with rank up to 6 are supported.");
+    AddOutput("Out", "(Tensor)The output tensor.");
     AddAttr<std::vector<int>>(
         "axis",
-        "(vector<int>)A list of values, and the size of the list should be "
-        "the same with the input tensor rank, the tensor will "
-        "permute the axes according the the values given");
+        "(vector<int>) A list of values, and the size of the list should be "
+        "the same with the input tensor rank. This operator permutes the input "
+        "tensor's axes according to the values given.");
     AddComment(R"DOC(
 Transpose Operator.
 
-The input tensor will be permuted according to the axis values given.
-The op functions is similar to how numpy.transpose works in python.
+The input tensor will be permuted according to the axes given.
+The behavior of this operator is similar to how `numpy.transpose` works.
 
-For example:
+- suppose the input `X` is a 2-D tensor:
+    $$
+    X = \begin{pmatrix}
+    0 &1 &2 \\
+    3 &4 &5
+    \end{pmatrix}$$
 
-    .. code-block:: text
+    the given `axes` is: $[1, 0]$, and $Y$ = transpose($X$, axis)
 
-      input = numpy.arange(6).reshape((2,3))
+    then the output $Y$ is:
 
-      the input is:
+    $$
+    Y = \begin{pmatrix}
+         0 &3 \\
+         1 &4  \\
+         2 &5
+    \end{pmatrix}$$
 
-      array([[0, 1, 2],
-             [3, 4, 5]])
-
-      given axis is:
-
-      [1, 0]
-
-      output = input.transpose(axis)
-
-      then the output is:
-
-      array([[0, 3],
-             [1, 4],
-             [2, 5]])
-
-So, given a input tensor of shape(N, C, H, W) and the axis is {0, 2, 3, 1},
-the output tensor shape will be (N, H, W, C)
+- Given a input tensor with shape $(N, C, H, W)$ and the `axes` is 
+$[0, 2, 3, 1]$, then shape of the output tensor will be: $(N, H, W, C)$.
 
 )DOC");
   }
diff --git a/paddle/platform/CMakeLists.txt b/paddle/platform/CMakeLists.txt
index 44f6d85cd1510f309595ca711de2e0f767219580..d68caea99719b37816391f9bddcc5cac051025b2 100644
--- a/paddle/platform/CMakeLists.txt
+++ b/paddle/platform/CMakeLists.txt
@@ -10,7 +10,7 @@ cc_test(cpu_info_test SRCS cpu_info_test.cc DEPS cpu_info)
 
 nv_library(gpu_info SRCS gpu_info.cc DEPS gflags glog enforce)
 
-cc_library(place SRCS place.cc DEPS enforce)
+cc_library(place SRCS place.cc DEPS enforce boost)
 cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
 
 add_subdirectory(dynload)
@@ -39,3 +39,11 @@ nv_test(nccl_test SRCS nccl_test.cu DEPS dynload_cuda gpu_info device_context)
 
 cc_library(profiler SRCS profiler.cc DEPS device_context)
 cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
+
+if(NOT WITH_C_API AND WITH_FLUID)
+  file(GLOB PLATFORM_HEADERS *.h)
+  file(GLOB PLATFORM_dynload_HEADERS dynload/*.h)
+  install(FILES ${PLATFORM_HEADERS} DESTINATION include/paddle/platform)
+  install(FILES ${PLATFORM_HEADERS} DESTINATION include/paddle/platform/dynload)
+  install(FILES details/device_ptr_cast.h DESTINATION include/paddle/platform/details)
+endif()
diff --git a/paddle/pybind/print_operators_doc.cc b/paddle/pybind/print_operators_doc.cc
index 99694fa592059d979297b72748125d02b2dd70a3..b55ddee17616ced4de659be8e55acd5e072c66b7 100644
--- a/paddle/pybind/print_operators_doc.cc
+++ b/paddle/pybind/print_operators_doc.cc
@@ -64,6 +64,8 @@ std::string AttrType(paddle::framework::proto::AttrType at) {
       return "bool array";
     case paddle::framework::proto::BLOCK:
       return "block id";
+    case paddle::framework::proto::LONG:
+      return "long";
   }
   return "UNKNOWN";  // not possible
 }
diff --git a/paddle/pybind/protobuf.cc b/paddle/pybind/protobuf.cc
index 4f959481537d29c089be24f9ae306f860c196c0f..371d6119d4ab73e683821d0dc5db5194f44a64ce 100644
--- a/paddle/pybind/protobuf.cc
+++ b/paddle/pybind/protobuf.cc
@@ -212,6 +212,7 @@ void BindVarDsec(py::module &m) {
              return name;
            },
            py::return_value_policy::reference)
+      .def("set_name", &VarDesc::SetName)
       .def("set_shape", &VarDesc::SetShape)
       .def("set_dtype", &VarDesc::SetDataType)
       .def("shape", &VarDesc::Shape, py::return_value_policy::reference)
@@ -280,7 +281,8 @@ void BindOpDesc(py::module &m) {
       .def("check_attrs", &OpDesc::CheckAttrs)
       .def("infer_shape", &OpDesc::InferShape)
       .def("infer_var_type", &OpDesc::InferVarType)
-      .def("serialize_to_string", SerializeMessage<OpDesc>);
+      .def("serialize_to_string", SerializeMessage<OpDesc>)
+      .def("block", &OpDesc::Block, py::return_value_policy::reference);
 }
 
 }  // namespace pybind
diff --git a/paddle/string/CMakeLists.txt b/paddle/string/CMakeLists.txt
index 60667b72873f9422aec1807972a81ab680de2e64..751776dbb5c00972c0b6893fcfb2e710f3f082d7 100644
--- a/paddle/string/CMakeLists.txt
+++ b/paddle/string/CMakeLists.txt
@@ -1,5 +1,10 @@
 cc_library(stringpiece SRCS piece.cc)
 cc_test(stringpiece_test SRCS piece_test.cc DEPS stringpiece glog gflags)
-
 cc_test(stringprintf_test SRCS printf_test.cc DEPS glog gflags)
 cc_test(to_string_test SRCS to_string_test.cc)
+
+if(NOT WITH_C_API AND WITH_FLUID)
+  file(GLOB STRING_HEADERS *.h)
+  install(FILES ${STRING_HEADERS} DESTINATION include/paddle/string)
+  install(FILES tinyformat/tinyformat.h DESTINATION include/paddle/string/tinyformat)
+endif()
diff --git a/python/paddle/v2/dataset/wmt16.py b/python/paddle/v2/dataset/wmt16.py
index bbc28a2da99052308471931122946d0d96b54da5..c8818f715beadd9499ae588f2c19a57fbf26f372 100644
--- a/python/paddle/v2/dataset/wmt16.py
+++ b/python/paddle/v2/dataset/wmt16.py
@@ -171,8 +171,9 @@ def train(src_dict_size, trg_dict_size, src_lang="en"):
         callable: The train reader.
     """
 
-    assert (src_lang in ["en", "de"], ("An error language type.  Only support: "
-                                       "en (for English); de(for Germany)"))
+    if src_lang not in ["en", "de"]:
+        raise ValueError("An error language type.  Only support: "
+                         "en (for English); de(for Germany).")
     src_dict_size, trg_dict_size = __get_dict_size(src_dict_size, trg_dict_size,
                                                    src_lang)
 
@@ -218,9 +219,9 @@ def test(src_dict_size, trg_dict_size, src_lang="en"):
         callable: The test reader.
     """
 
-    assert (src_lang in ["en", "de"],
-            ("An error language type.  "
-             "Only support: en (for English); de(for Germany)"))
+    if src_lang not in ["en", "de"]:
+        raise ValueError("An error language type. "
+                         "Only support: en (for English); de(for Germany).")
 
     src_dict_size, trg_dict_size = __get_dict_size(src_dict_size, trg_dict_size,
                                                    src_lang)
@@ -266,9 +267,9 @@ def validation(src_dict_size, trg_dict_size, src_lang="en"):
     Returns:
         callable: The validation reader.
     """
-    assert (src_lang in ["en", "de"],
-            ("An error language type.  "
-             "Only support: en (for English); de(for Germany)"))
+    if src_lang not in ["en", "de"]:
+        raise ValueError("An error language type. "
+                         "Only support: en (for English); de(for Germany).")
     src_dict_size, trg_dict_size = __get_dict_size(src_dict_size, trg_dict_size,
                                                    src_lang)
 
@@ -304,9 +305,9 @@ def get_dict(lang, dict_size, reverse=False):
 
     dict_path = os.path.join(paddle.v2.dataset.common.DATA_HOME,
                              "wmt16/%s_%d.dict" % (lang, dict_size))
-    assert (os.path.exists(dict_path), "Word dictionary does not exist. "
-            "Please invoke paddle.dataset.wmt16.train/test/validation "
-            "first to build the dictionary.")
+    assert os.path.exists(dict_path), "Word dictionary does not exist. "
+    "Please invoke paddle.dataset.wmt16.train/test/validation first "
+    "to build the dictionary."
     tar_file = os.path.join(paddle.v2.dataset.common.DATA_HOME, "wmt16.tar.gz")
     return __load_dict(tar_file, dict_size, lang, reverse)
 
diff --git a/python/paddle/v2/fluid/clip.py b/python/paddle/v2/fluid/clip.py
index 386df9823de9119287abf87569eab0b283ecc802..3028029e60fde2f481b4348ab1b0a4980ebb2b60 100644
--- a/python/paddle/v2/fluid/clip.py
+++ b/python/paddle/v2/fluid/clip.py
@@ -12,14 +12,18 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import copy
+
 import functools
 import layers
 import framework
 from . import core
 
 __all__ = [
-    'GradientClipByValue',
     'ErrorClipByValue',
+    'GradientClipByValue',
+    'GradientClipByNorm',
+    'GradientClipByGlobalNorm',
     'append_gradient_clip_ops',
     'error_clip_callback',
 ]
@@ -155,10 +159,11 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
         return param, new_grad
 
 
-def gradient_clip_by_global_norm(clip_norm,
-                                 param_list=None,
-                                 group_name="default_group",
-                                 program=None):
+def set_gradient_clip(clip, param_list=None, program=None):
+    if not isinstance(clip, BaseGradientClipAttr):
+        raise TypeError(
+            "'clip' should be an instance of BaseGradientClipAttr's derived class"
+        )
     if program is None:
         program = framework.default_main_program()
     if param_list is None:
@@ -171,8 +176,7 @@ def gradient_clip_by_global_norm(clip_norm,
         )
 
     for param in param_list:
-        param.gradient_clip_attr = GradientClipByGlobalNorm(clip_norm,
-                                                            group_name)
+        param.gradient_clip_attr = copy.deepcopy(clip)
 
 
 def append_gradient_clip_ops(param_grad):
diff --git a/python/paddle/v2/fluid/distribute_transpiler.py b/python/paddle/v2/fluid/distribute_transpiler.py
index 573774a2324791c1786e39700aeb27e64e2e8f9a..abcad899bfac9ba3eff20cde825e136d867a4485 100644
--- a/python/paddle/v2/fluid/distribute_transpiler.py
+++ b/python/paddle/v2/fluid/distribute_transpiler.py
@@ -38,14 +38,14 @@ def split_dense_variable(var_list,
                          min_block_size=1024,
                          max_block_size=1048576):
     """
-        We may need to split dense tensor to one or several blocks and put
+        We may need to split dense tensor to one or more blocks and put
         them equally onto parameter server. One block is a sub-tensor
         aligned by dim[0] of the tensor.
-        
+
         We need to have a minimal block size so that the calculations in
         the parameter server side can gain better performance. By default
-        mininum block size is 1024. The max block size is used to prevent
-        too large block that may causing send error.
+        minimum block size is 1024. The max block size is used to prevent
+        very large blocks that may cause send error.
     """
     blocks = []
     for var in var_list:
@@ -64,7 +64,7 @@ def split_dense_variable(var_list,
             remains = block_size % dim1
             if remains != 0:
                 block_size += dim1 - remains
-        # update split_count after align
+        # update split_count after aligning
         split_count = int(math.ceil(var_numel / float(block_size)))
         for block_id in xrange(split_count):
             curr_block_size = min(block_size, var_numel - (
@@ -83,18 +83,18 @@ class DistributeTranspiler:
                   trainers=1,
                   split_method=round_robin):
         """
-            Transpile the program to a distributed data-parallelism programs.
-            The main_program will be transform to use a remote parameter server
+            Transpile the program to distributed data-parallelism programs.
+            The main_program will be transformed to use a remote parameter server
             to do parameter optimization. And the optimization graph will be put
-            in to a parameter server program.
+            into a parameter server program.
 
-            Use different methods to split trainable varialbles to different
+            Use different methods to split trainable variables to different
             parameter servers.
 
             :param optimize_ops: op list of optimization, should be the
                                  return value of Optimizer.minimize
             :type optimize_ops: list
-            :param program: program to optimize, default default_main_program
+            :param program: program to optimize, default is default_main_program
             :param pservers: parameter server endpoints like "m1:6174,m2:6174"
             :type pservers: string
             :return: return a list of programs
@@ -106,11 +106,11 @@ class DistributeTranspiler:
         self.trainers = trainers
         self.optimize_ops = optimize_ops
         # steps to transpile:
-        # 1. split variable to multiple blocks, align by product(dim[1:]) (width).
+        # 1. split variable to multiple blocks, aligned by product(dim[1:]) (width).
         # 2. modify trainer program add split_op to each Grad.
         # 3. append send_op to trainer.
         # 4. append concat_op to trainer to update local weights.
-        # 5. create new program as parameter server.
+        # 5. create new program for parameter server.
         # 6. create parameter server program by split_method generated endpoint->VarBlock
 
         pserver_endpoints = pservers.split(",")
@@ -136,10 +136,10 @@ class DistributeTranspiler:
         for b in param_blocks:
             varname, block_id, _ = b.split(":")
             send_outputs.append(param_var_mapping[varname][int(block_id)])
-        # let send_op know which endpoint to send which var, eplist is of the same
-        # order of send_inputs.
+        # let send_op know which endpoint to send which var to, eplist has the same
+        # order as send_inputs.
         eplist = split_method(send_inputs, pserver_endpoints)
-        # create mapping of endpoint -> splited var to create pserver side program
+        # create mapping of endpoint -> split var to create pserver side program
         self.param_grad_ep_mapping = dict()
         for i, ep in enumerate(eplist):
             param = send_outputs[i]
@@ -149,6 +149,7 @@ class DistributeTranspiler:
             self.param_grad_ep_mapping[ep]["params"].append(param)
             self.param_grad_ep_mapping[ep]["grads"].append(grad)
 
+        # create send_op
         send_op = program.global_block().append_op(
             type="send",
             inputs={"X": send_inputs},
@@ -167,6 +168,7 @@ class DistributeTranspiler:
                 attrs={"axis": 0})
 
     def _create_vars_from_blocklist(self, program, block_list):
+        # Create respective variables using the block_list
         block_map = dict()
         var_mapping = dict()
         for block_str in block_list:
@@ -207,11 +209,12 @@ class DistributeTranspiler:
             dtype=var.dtype,
             type=var.type,
             lod_level=var.lod_level,
-            # HACK: let all param in pserver persistable so child
+            # HACK: let all param in pserver be persistable so the child
             # program in recv can get them
             persistable=True)
 
     def _append_split_op(self, program, gradblocks):
+        # Split variables that need to be split and append respective ops
         var_mapping = self._create_vars_from_blocklist(program, gradblocks)
         for varname, splited_vars in var_mapping.iteritems():
             # variable that don't need to split have empty splited_vars
@@ -248,6 +251,7 @@ class DistributeTranspiler:
         return self.program
 
     def _create_var_for_trainers(self, block, var, trainers):
+        # For each trainer, create the necessary variables
         var_list = []
         for i in xrange(trainers):
             var_each = block.create_var(
@@ -262,7 +266,7 @@ class DistributeTranspiler:
                                    param_shape):
         """
         Returns the shape for optimizer inputs that need to be reshaped when
-        Param and Grad is splited to multiple servers.
+        Param and Grad is split to multiple servers.
         """
         # HACK(typhoonzero): Should use functions of corresponding optimizer in
         # optimizer.py to get the shape, do not  bind this in the transpiler.
@@ -300,7 +304,7 @@ class DistributeTranspiler:
             else:
                 for n in param_names:
                     if n.startswith(op.inputs["Param"].name+".block") and \
-                        n != op.inputs["Param"].name:
+                       n != op.inputs["Param"].name:
                         return True
                 return False
         else:
@@ -396,7 +400,7 @@ class DistributeTranspiler:
                 dtype=var.dtype,
                 shape=new_shape)
 
-        # change outputs ParamOut variable
+        # change output's ParamOut variable
         opt_op.outputs["ParamOut"] = new_inputs["Param"]
         program.global_block().append_op(
             type=opt_op.type,
@@ -405,6 +409,7 @@ class DistributeTranspiler:
             attrs=opt_op.attrs)
 
     def _append_pserver_non_opt_ops(self, program, pserver_program, opt_op):
+        # Append the ops for parameters that do not need to be optimized/updated
         for _, var in opt_op.inputs.iteritems():
             program.global_block().create_var(
                 name=var.name,
@@ -424,7 +429,7 @@ class DistributeTranspiler:
 
     def get_pserver_program(self, endpoint):
         """
-        get pserver side program by endpoint
+        Get pserver side program using the endpoint
 
         NOTE: assume blocks of the same variable is not distributed
         on the same pserver, only change param/grad varnames for
@@ -450,6 +455,7 @@ class DistributeTranspiler:
                     shape=v.shape)
         # step6
         optimize_sub_program = Program()
+        # Iterate through the ops and append ops as needed
         for idx, opt_op in enumerate(self.optimize_ops):
             is_op_on_pserver = self._is_op_on_pserver(endpoint,
                                                       self.optimize_ops, idx)
@@ -461,6 +467,7 @@ class DistributeTranspiler:
             else:
                 self._append_pserver_non_opt_ops(optimize_sub_program,
                                                  pserver_program, opt_op)
+        # Append the recv op
         pserver_program.global_block().append_op(
             type="recv",
             inputs={"RX": self.param_grad_ep_mapping[endpoint]["grads"]
@@ -486,7 +493,7 @@ class DistributeTranspiler:
         """
         Get startup program for current parameter server.
         Modify operator input variables if there are variables that
-        was splited to several blocks.
+        were split to several blocks.
         """
         s_prog = Program()
         orig_s_prog = framework.default_startup_program()
diff --git a/python/paddle/v2/fluid/evaluator.py b/python/paddle/v2/fluid/evaluator.py
index 396d56fc8b236d95d38517f3513521aa969e47be..2686a5bdfcf0e2d26ce8f58cceff1967b06d835b 100644
--- a/python/paddle/v2/fluid/evaluator.py
+++ b/python/paddle/v2/fluid/evaluator.py
@@ -205,3 +205,63 @@ class ChunkEvaluator(Evaluator):
             [precision], dtype='float32'), np.array(
                 [recall], dtype='float32'), np.array(
                     [f1_score], dtype='float32')
+
+
+class EditDistance(Evaluator):
+    """
+    Accumulate edit distance sum and sequence number from mini-batches and
+    compute the average edit_distance of all batches.
+
+    Args:
+        input: the sequences predicted by network.
+        label: the target sequences which must has same sequence count
+        with input.
+        ignored_tokens(list of int): Tokens that should be removed before
+        calculating edit distance.
+
+    Example:
+
+        exe = fluid.executor(place)
+        distance_evaluator = fluid.Evaluator.EditDistance(input, label)
+        for epoch in PASS_NUM:
+            distance_evaluator.reset(exe)
+            for data in batches:
+                loss, sum_distance = exe.run(fetch_list=[cost] + distance_evaluator.metrics)
+                avg_distance = distance_evaluator.eval(exe)
+            pass_distance = distance_evaluator.eval(exe)
+
+        In the above example:
+        'sum_distance' is the sum of the batch's edit distance.
+        'avg_distance' is the average of edit distance from the firt batch to the current batch.
+        'pass_distance' is the average of edit distance from all the pass.
+
+    """
+
+    def __init__(self, input, label, ignored_tokens=None, **kwargs):
+        super(EditDistance, self).__init__("edit_distance", **kwargs)
+        main_program = self.helper.main_program
+        if main_program.current_block().idx != 0:
+            raise ValueError("You can only invoke Evaluator in root block")
+
+        self.total_error = self.create_state(
+            dtype='float32', shape=[1], suffix='total_error')
+        self.seq_num = self.create_state(
+            dtype='int64', shape=[1], suffix='seq_num')
+        error, seq_num = layers.edit_distance(
+            input=input, label=label, ignored_tokens=ignored_tokens)
+        #error = layers.cast(x=error, dtype='float32')
+        sum_error = layers.reduce_sum(error)
+        layers.sums(input=[self.total_error, sum_error], out=self.total_error)
+        layers.sums(input=[self.seq_num, seq_num], out=self.seq_num)
+        self.metrics.append(sum_error)
+
+    def eval(self, executor, eval_program=None):
+        if eval_program is None:
+            eval_program = Program()
+        block = eval_program.current_block()
+        with program_guard(main_program=eval_program):
+            total_error = _clone_var_(block, self.total_error)
+            seq_num = _clone_var_(block, self.seq_num)
+            seq_num = layers.cast(x=seq_num, dtype='float32')
+            out = layers.elementwise_div(x=total_error, y=seq_num)
+        return np.array(executor.run(eval_program, fetch_list=[out])[0])
diff --git a/python/paddle/v2/fluid/layer_helper.py b/python/paddle/v2/fluid/layer_helper.py
index 0b0064ade90d2b70dd1458cb4d20d741fbf1efcd..8c481444e9bf895d29ed4e4952e825c2eaafc915 100644
--- a/python/paddle/v2/fluid/layer_helper.py
+++ b/python/paddle/v2/fluid/layer_helper.py
@@ -100,7 +100,8 @@ class LayerHelper(object):
             if dtype is None:
                 dtype = each.dtype
             elif dtype != each.dtype:
-                raise ValueError("Data Type mismatch")
+                raise ValueError("Data Type mismatch: %d to %d" %
+                                 (dtype, each.dtype))
         return dtype
 
     def create_parameter(self,
diff --git a/python/paddle/v2/fluid/layers/control_flow.py b/python/paddle/v2/fluid/layers/control_flow.py
index 5f01fdb076d3bf7d060a805d1431f4973993a843..0fcbfe0e2f2f9686366139e84b7fdcc158bf0aa7 100644
--- a/python/paddle/v2/fluid/layers/control_flow.py
+++ b/python/paddle/v2/fluid/layers/control_flow.py
@@ -289,6 +289,7 @@ class ParallelDo(object):
                 for in_var_name in op.input(iname):
                     if in_var_name not in local_inputs:
                         params.append(in_var_name)
+        params = list(set(params))
 
         return [parent_block.var(name) for name in params]
 
@@ -769,7 +770,7 @@ def topk(input, k):
           array = fluid.layers.topk(x, k)
     """
     helper = LayerHelper('topk', **locals())
-    topk_out = helper.create_tmp_variable(dtype=input.data_type)
+    topk_out = helper.create_tmp_variable(dtype=input.dtype)
     topk_indices = helper.create_tmp_variable(dtype='int64')
     helper.append_op(
         type='top_k',
diff --git a/python/paddle/v2/fluid/layers/math_op_patch.py b/python/paddle/v2/fluid/layers/math_op_patch.py
index 11197b70a3d4cae08afbb49ad31013ab40e4dad2..f359e70126f7601b75261e795b5a37bdc241112e 100644
--- a/python/paddle/v2/fluid/layers/math_op_patch.py
+++ b/python/paddle/v2/fluid/layers/math_op_patch.py
@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from ..framework import Variable, unique_name
-from ..registry import OpProtoHolder
+from layer_function_generator import OpProtoHolder
 
 __all__ = ['monkey_patch_variable']
 
diff --git a/python/paddle/v2/fluid/layers/nn.py b/python/paddle/v2/fluid/layers/nn.py
index 26e36ec0bbe684f0eff9698117f12e0c702a0d31..f834eac755fd1c5261acd36610c696988b969f09 100644
--- a/python/paddle/v2/fluid/layers/nn.py
+++ b/python/paddle/v2/fluid/layers/nn.py
@@ -19,16 +19,51 @@ from ..layer_helper import LayerHelper
 from ..initializer import Normal, Constant
 from ..framework import Variable
 from ..param_attr import ParamAttr
+from layer_function_generator import autodoc
 from tensor import concat
 
 __all__ = [
-    'fc', 'embedding', 'dynamic_lstm', 'gru_unit', 'linear_chain_crf',
-    'crf_decoding', 'cos_sim', 'cross_entropy', 'square_error_cost', 'accuracy',
-    'chunk_eval', 'sequence_conv', 'conv2d', 'sequence_pool', 'pool2d',
-    'batch_norm', 'beam_search_decode', 'conv2d_transpose', 'sequence_expand',
-    'lstm_unit', 'reduce_sum', 'reduce_mean', 'reduce_max', 'reduce_min',
-    'sequence_first_step', 'sequence_last_step', 'dropout', 'split',
-    'l2_normalize', 'matmul', 'warpctc', 'sequence_reshape', 'multiplex'
+    'fc',
+    'embedding',
+    'dynamic_lstm',
+    'dynamic_gru',
+    'gru_unit',
+    'linear_chain_crf',
+    'crf_decoding',
+    'cos_sim',
+    'cross_entropy',
+    'square_error_cost',
+    'accuracy',
+    'chunk_eval',
+    'sequence_conv',
+    'conv2d',
+    'sequence_pool',
+    'pool2d',
+    'batch_norm',
+    'beam_search_decode',
+    'conv2d_transpose',
+    'sequence_expand',
+    'lstm_unit',
+    'reduce_sum',
+    'reduce_mean',
+    'reduce_max',
+    'reduce_min',
+    'sequence_first_step',
+    'sequence_last_step',
+    'dropout',
+    'split',
+    'ctc_greedy_decoder',
+    'edit_distance',
+    'l2_normalize',
+    'matmul',
+    'warpctc',
+    'sequence_reshape',
+    'transpose',
+    'im2sequence',
+    'nce',
+    'beam_search',
+    'row_conv',
+    'multiplex',
 ]
 
 
@@ -43,14 +78,14 @@ def fc(input,
     **Fully Connected Layer**
 
     The fully connected layer can take multiple tensors as its inputs. It
-    creates a variable (one for each input tensor) called weights for each input
-    tensor, which represents a fully connected weight matrix from each input
-    unit to each output unit. The fully connected layer multiplies each input
-    tensor with its coresponding weight to produce an output Tensor. If
-    multiple input tensors are given, the results of multiple multiplications
-    will be sumed up. If bias_attr is not None, a biases variable will be
-    created and added to the output. Finally, if activation is not None,
-    it will be applied to the output as well.
+    creates a variable (one for each input tensor) called weights for each
+    input tensor, which represents a fully connected weight matrix from
+    each input unit to each output unit. The fully connected layer
+    multiplies each input tensor with its coresponding weight to produce
+    an output Tensor. If multiple input tensors are given, the results of
+    multiple multiplications will be sumed up. If bias_attr is not None,
+    a biases variable will be created and added to the output. Finally,
+    if activation is not None, it will be applied to the output as well.
 
     This process can be formulated as follows:
 
@@ -131,10 +166,8 @@ def fc(input,
         tmp = helper.create_tmp_variable(dtype)
         helper.append_op(
             type="mul",
-            inputs={
-                "X": input_var,
-                "Y": w,
-            },
+            inputs={"X": input_var,
+                    "Y": w},
             outputs={"Out": tmp},
             attrs={"x_num_col_dims": num_flatten_dims,
                    "y_num_col_dims": 1})
@@ -153,22 +186,35 @@ def fc(input,
     return helper.append_activation(pre_activation)
 
 
-def embedding(input, size, is_sparse=False, param_attr=None, dtype='float32'):
+def embedding(input,
+              size,
+              is_sparse=False,
+              padding_idx=None,
+              param_attr=None,
+              dtype='float32'):
     """
     **Embedding Layer**
 
-    This layer is used to lookup a vector of IDs, provided by *input*, in a lookup table.
-    The result of this lookup is the embedding of each ID in the *input*.
+    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
+    a lookup table. The result of this lookup is the embedding of each ID in the
+    :attr:`input`.
 
     All the input variables are passed in as local variables to the LayerHelper
     constructor.
 
     Args:
-       input(Variable): Input to the function
-       size(tuple|list|None): Shape of the look up table parameter
-       is_sparse(bool): Boolean flag that specifying whether the input is sparse
-       param_attr(ParamAttr): Parameters for this layer
-       dtype(np.dtype|core.DataType|str): The type of data : float32, float_16, int etc
+        input(Variable): The tensor variable containing the IDs.
+        size(tuple|list): The shape of the look up table parameter. It should
+            have two elements which indicate the size of the dictionary of
+            embeddings and the size of each embedding vector respectively.
+        is_sparse(bool): The flag indicating whether to use sparse update.
+        padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup.
+            Otherwise the given :attr:`padding_idx` indicates padding the output
+            with zeros whenever lookup encounters it in :attr:`input`. If
+            :math:`padding_idx < 0`, the padding_idx to use in lookup is
+            :math:`size[0] + dim`.
+        param_attr(ParamAttr): Parameters for this layer
+        dtype(np.dtype|core.DataType|str): The type of data : float32, float_16, int etc
 
     Returns:
         Variable: The tensor variable storing the embeddings of the \
@@ -186,12 +232,15 @@ def embedding(input, size, is_sparse=False, param_attr=None, dtype='float32'):
     w = helper.create_parameter(
         attr=helper.param_attr, shape=size, dtype=dtype, is_bias=False)
     tmp = helper.create_tmp_variable(dtype)
+    padding_idx = -1 if padding_idx is None else padding_idx if padding_idx >= 0 else (
+        size[0] + padding_idx)
     helper.append_op(
         type='lookup_table',
         inputs={'Ids': input,
                 'W': w},
         outputs={'Out': tmp},
-        attrs={'is_sparse': is_sparse})
+        attrs={'is_sparse': is_sparse,
+               'padding_idx': padding_idx})
     return tmp
 
 
@@ -338,6 +387,113 @@ def dynamic_lstm(input,
     return hidden, cell
 
 
+def dynamic_gru(input,
+                size,
+                param_attr=None,
+                bias_attr=None,
+                is_reverse=False,
+                gate_activation='sigmoid',
+                candidate_activation='tanh',
+                h_0=None):
+    """
+    **Dynamic GRU Layer**
+
+    Refer to `Empirical Evaluation of Gated Recurrent Neural Networks on
+    Sequence Modeling <https://arxiv.org/abs/1412.3555>`_
+
+    The formula is as follows:
+
+    .. math::
+
+        u_t & = act_g(W_{ux}x_{t} + W_{uh}h_{t-1} + b_u)
+
+        r_t & = act_g(W_{rx}x_{t} + W_{rh}h_{t-1} + b_r)
+
+        \\tilde{h_t} & = act_c(W_{cx}x_{t} + W_{ch}(r_t \odot h_{t-1}) + b_c)
+
+        h_t & = (1-u_t) \odot h_{t-1} + u_t \odot \\tilde{h_t}
+
+    The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
+    is the update gate and reset gate activation function and :math:`sigmoid`
+    is usually used for it. :math:`act_c` is the activation function for
+    candidate hidden state and :math:`tanh` is usually used for it.
+
+    Note that these :math:`W_{ux}x_{t}, W_{rx}x_{t}, W_{cx}x_{t}` operations on
+    the input :math:`x_{t}` are NOT included in this operator. Users can choose
+    to use fully-connect layer before GRU layer.
+
+    Args:
+        input(Variable): The input of dynamic_gru layer, which supports
+            variable-time length input sequence. The underlying tensor in this
+            Variable is a matrix with shape :math:`(T \\times 3D)`, where
+            :math:`T` is the total time steps in this mini-batch, :math:`D`
+            is the hidden size.
+        size(int): The dimension of the gru cell.
+        param_attr(ParamAttr|None): The parameter attribute for the learnable
+            hidden-hidden weight matrix. Note:
+
+            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
+              :math:`D` is the hidden size.
+            - All elements in the weight matrix can be divided into two parts.
+              The first part are weights of the update gate and reset gate with
+              shape :math:`(D \\times 2D)`, and the second part are weights for
+              candidate hidden state with shape :math:`(D \\times D)`.
+        bias_attr(ParamAttr): The parameter attribute for learnable the
+            hidden-hidden bias.
+        is_reverse(bool): Whether to compute reversed GRU, default
+            :attr:`False`.
+        gate_activation(str): The activation for update gate and reset gate.
+            Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid".
+        activation(str): The activation for candidate hidden state.
+            Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh".
+
+    Returns:
+        Variable: The hidden state of GRU. The shape is (T \\times D), and lod \
+            is the same with the input.
+
+    Examples:
+        .. code-block:: python
+
+            hidden_dim = 512
+            x = fluid.layers.fc(input=data, size=hidden_dim * 3)
+            hidden = fluid.layers.dynamic_gru(input=x, dim=hidden_dim)
+    """
+
+    helper = LayerHelper('gru', **locals())
+    dtype = helper.input_dtype()
+
+    weight = helper.create_parameter(
+        attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
+    bias = helper.create_parameter(
+        attr=helper.bias_attr, shape=[1, 3 * size], dtype=dtype, is_bias=True)
+    inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
+    if h_0 != None:
+        assert h_0.shape == (
+            size, size), 'The shape of h0 should be(%d, %d)' % (size, size)
+        inputs['h0'] = h_0
+
+    hidden = helper.create_tmp_variable(dtype)
+    batch_gate = helper.create_tmp_variable(dtype)
+    batch_reset_hidden_prev = helper.create_tmp_variable(dtype)
+    batch_hidden = helper.create_tmp_variable(dtype)
+
+    helper.append_op(
+        type='gru',
+        inputs=inputs,
+        outputs={
+            'Hidden': hidden,
+            'BatchGate': batch_gate,
+            'BatchResetHiddenPrev': batch_reset_hidden_prev,
+            'BatchHidden': batch_hidden
+        },
+        attrs={
+            'is_reverse': is_reverse,
+            'gate_activation': gate_activation,
+            'activation': candidate_activation
+        })
+    return hidden
+
+
 def gru_unit(input,
              hidden,
              size,
@@ -762,8 +918,8 @@ def conv2d(input,
     <http://ufldl.stanford.edu/tutorial/supervised/FeatureExtractionUsingConvolution/>`_ .
     If bias attribution and activation type are provided, bias is added to the output of the convolution,
     and the corresponding activation function is applied to the final result.
-    For each input :math:`X`, the equation is:
 
+    For each input :math:`X`, the equation is:
 
     .. math::
 
@@ -771,51 +927,54 @@ def conv2d(input,
 
     In the above equation:
 
-        * :math:`X`: Input value, a tensor with NCHW format.
-        * :math:`W`: Filter value, a tensor with MCHW format.
-        * :math:`\\ast`: Convolution operation.
-        * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
-        * :math:`\\sigma`: Activation function.
-        * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
+    * :math:`X`: Input value, a tensor with NCHW format.
+    * :math:`W`: Filter value, a tensor with MCHW format.
+    * :math:`\\ast`: Convolution operation.
+    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
+    * :math:`\\sigma`: Activation function.
+    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
 
     Example:
 
-        Input:
-            Input shape: $(N, C_{in}, H_{in}, W_{in})$
+        - Input:
+
+          Input shape: $(N, C_{in}, H_{in}, W_{in})$
+
+          Filter shape: $(C_{out}, C_{in}, H_f, W_f)$
 
-            Filter shape: $(C_{out}, C_{in}, H_f, W_f)$
+        - Output:
+          Output shape: $(N, C_{out}, H_{out}, W_{out})$
 
-        Output:
-            Output shape: $(N, C_{out}, H_{out}, W_{out})$
         Where
-    .. math::
+
+        .. math::
 
         H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\
         W_{out}&= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]} + 1
 
     Args:
-        input(Variable): The input image with [N, C, H, W] format.
-        num_filters(int): The number of filter. It is as same as the output
-            image channel.
-        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
-            it must contain two integers, (filter_size_H, filter_size_W).
-            Otherwise, the filter will be a square.
-        stride(int|tuple): The stride size. If stride is a tuple, it must
-            contain two integers, (stride_H, stride_W). Otherwise, the
-            stride_H = stride_W = stride. Default: stride = 1.
-        padding(int|tuple): The padding size. If padding is a tuple, it must
-            contain two integers, (padding_H, padding_W). Otherwise, the
-            padding_H = padding_W = padding. Default: padding = 0.
-        groups(int): The groups number of the Conv2d Layer. According to grouped
-            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
-            the first half of the filters is only connected to the first half
-            of the input channels, while the second half of the filters is only
-            connected to the second half of the input channels. Default: groups=1
-        param_attr(ParamAttr): The parameters to the Conv2d Layer. Default: None
-        bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None
-        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
-            library is installed. Default: True
-        act(str): Activation type. Default: None
+       input(Variable): The input image with [N, C, H, W] format.
+       num_filters(int): The number of filter. It is as same as the output
+           image channel.
+       filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
+           it must contain two integers, (filter_size_H, filter_size_W).
+           Otherwise, the filter will be a square.
+       stride(int|tuple): The stride size. If stride is a tuple, it must
+           contain two integers, (stride_H, stride_W). Otherwise, the
+           stride_H = stride_W = stride. Default: stride = 1.
+       padding(int|tuple): The padding size. If padding is a tuple, it must
+           contain two integers, (padding_H, padding_W). Otherwise, the
+           padding_H = padding_W = padding. Default: padding = 0.
+       groups(int): The groups number of the Conv2d Layer. According to grouped
+           convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
+           the first half of the filters is only connected to the first half
+           of the input channels, while the second half of the filters is only
+           connected to the second half of the input channels. Default: groups=1
+       param_attr(ParamAttr): The parameters to the Conv2d Layer. Default: None
+       bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None
+       use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
+           library is installed. Default: True
+       act(str): Activation type. Default: None
 
     Returns:
         Variable: The tensor variable storing the convolution and \
@@ -830,7 +989,6 @@ def conv2d(input,
           data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
           conv2d = fluid.layers.conv2d(input=data, num_filters=2, filter_size=3, act="relu")
     """
-
     if stride is None:
         stride = [1, 1]
     helper = LayerHelper('conv2d', **locals())
@@ -1184,38 +1342,85 @@ def conv2d_transpose(input,
                      use_cudnn=True,
                      name=None):
     """
-    The transpose of conv2d layer.
+    **Convlution2D transpose layer**
+
+    The convolution2D transpose layer calculates the output based on the input,
+    filter, and dilations, strides, paddings. Input(Input) and output(Output)
+    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(dilations, strides, paddings) are two elements. These two elements
+    represent height and width, respectively. The details of convolution transpose
+    layer, please refer to the following explanation and references `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
+
+    For each input :math:`X`, the equation is:
+
+    .. math::
+
+        Out = W \\ast X
+
+    In the above equation:
+
+    * :math:`X`: Input value, a tensor with NCHW format.
+    * :math:`W`: Filter value, a tensor with MCHW format.
+    * :math:`\\ast` : Convolution transpose operation.
+    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
+
+    Example:
+
+        - Input:
+
+          Input shape: $(N, C_{in}, H_{in}, W_{in})$
 
-    This layer is also known as deconvolution layer.
+          Filter shape: $(C_{in}, C_{out}, H_f, W_f)$
+
+        - Output:
+
+          Output shape: $(N, C_{out}, H_{out}, W_{out})$
+
+        Where
+
+        .. math::
+
+           H_{out} &= (H_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (H_f - 1) + 1 \\\\
+           W_{out} &= (W_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (W_f - 1) + 1
 
     Args:
-        input(Variable): The input image with [N, C, H, W] format.
-        num_filters(int): The number of filter. It is as same as the output
-            image channel.
-        output_size(int|tuple|None): The output image size. If output size is a
-            tuple, it must contain two integers, (image_H, image_W). This
-            parameter only works when filter_size is None.
-        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
-            it must contain two integers, (filter_size_H, filter_size_W).
-            Otherwise, the filter will be a square.  None if use output size to
-            calculate filter_size
-        padding(int|tuple): The padding size. If padding is a tuple, it must
-            contain two integers, (padding_H, padding_W). Otherwise, the
-            padding_H = padding_W = padding.
-        stride(int|tuple): The stride size. If stride is a tuple, it must
-            contain two integers, (stride_H, stride_W). Otherwise, the
-            stride_H = stride_W = stride.
-        dilation(int|tuple): The dilation size. If dilation is a tuple, it must
-            contain two integers, (dilation_H, dilation_W). Otherwise, the
-            dilation_H = dilation_W = dilation.
-        param_attr: Parameter Attribute.
-        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
-            library is installed. Default: True
-        name(str|None): A name for this layer(optional). If set None, the layer
-                       will be named automatically.
+       input(Variable): The input image with [N, C, H, W] format.
+       num_filters(int): The number of the filter. It is as same as the output
+           image channel.
+       output_size(int|tuple|None): The output image size. If output size is a
+           tuple, it must contain two integers, (image_H, image_W). This
+           parameter only works when filter_size is None.
+       filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
+           it must contain two integers, (filter_size_H, filter_size_W).
+           Otherwise, the filter will be a square. None if use output size to
+           calculate filter_size.
+       padding(int|tuple): The padding size. If padding is a tuple, it must
+           contain two integers, (padding_H, padding_W). Otherwise, the
+           padding_H = padding_W = padding. Default: padding = 0.
+       stride(int|tuple): The stride size. If stride is a tuple, it must
+           contain two integers, (stride_H, stride_W). Otherwise, the
+           stride_H = stride_W = stride. Default: stride = 1.
+       dilation(int|tuple): The dilation size. If dilation is a tuple, it must
+           contain two integers, (dilation_H, dilation_W). Otherwise, the
+           dilation_H = dilation_W = dilation. Default: dilation = 1.
+       param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer. Default: None
+       use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
+           library is installed. Default: True
+       name(str|None): A name for this layer(optional). If set None, the layer
+           will be named automatically.
 
     Returns:
-        Variable: Output image.
+       Variable: The tensor variable storing the convolution transpose result.
+
+    Raises:
+       ValueError: If the shapes of input, filter_size, stride, padding and groups mismatch.
+
+    Examples:
+       .. code-block:: python
+
+          data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
+          conv2d_transpose = fluid.layers.conv2d_transpose(input=data, num_filters=2, filter_size=3)
     """
     helper = LayerHelper("conv2d_transpose", **locals())
     if not isinstance(input, Variable):
@@ -1347,6 +1552,38 @@ def sequence_expand(x, y, name=None):
     return tmp
 
 
+def beam_search(pre_ids, ids, scores, beam_size, end_id, level=0):
+    '''
+    This function implements the beam search algorithm.
+    '''
+    helper = LayerHelper('beam_search', **locals())
+    score_type = scores.dtype
+    id_type = ids.dtype
+
+    selected_scores = helper.create_tmp_variable(dtype=score_type)
+    selected_ids = helper.create_tmp_variable(dtype=id_type)
+
+    helper.append_op(
+        type='beam_search',
+        inputs={
+            'pre_ids': pre_ids,
+            'ids': ids,
+            'scores': scores,
+        },
+        outputs={
+            'selected_ids': selected_ids,
+            'selected_scores': selected_scores,
+        },
+        attrs={
+            # TODO(ChunweiYan) to assure other value support
+            'level': level,
+            'beam_size': beam_size,
+            'end_id': end_id,
+        })
+
+    return selected_ids, selected_scores
+
+
 def lstm_unit(x_t,
               hidden_t_prev,
               cell_t_prev,
@@ -1866,6 +2103,146 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
     return out
 
 
+def edit_distance(input,
+                  label,
+                  normalized=False,
+                  ignored_tokens=None,
+                  name=None):
+    """
+    EditDistance operator computes the edit distances between a batch of hypothesis strings and their references. Edit distance, also called Levenshtein distance, measures how dissimilar two strings are by counting the minimum number of operations to transform one string into anthor. Here the operations include insertion, deletion, and substitution. For example, given hypothesis string A = "kitten" and reference B = "sitting", the edit distance is 3 for A will be transformed into B at least after two substitutions and one insertion:
+
+       "kitten" -> "sitten" -> "sittin" -> "sitting"
+
+    Input(Hyps) is a LoDTensor consisting of all the hypothesis strings with the total number denoted by `batch_size`, and the separation is specified by the LoD information. And the `batch_size` reference strings are arranged in order in the same way in the LoDTensor Input(Refs).
+
+    Output(Out) contains the `batch_size` results and each stands for the edit stance for a pair of strings respectively. If Attr(normalized) is true, the edit distance will be divided by the length of reference string.
+
+    Args:
+
+        input(Variable): The indices for hypothesis strings.
+
+        label(Variable): The indices for reference strings.
+
+        normalized(bool): Indicated whether to normalize the edit distance by the length of reference string.
+
+        ignored_tokens(list of int): Tokens that should be removed before calculating edit distance.
+
+    Returns:
+        Variable: sequence-to-sequence edit distance in shape [batch_size, 1].
+
+    Examples:
+        .. code-block:: python
+
+            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
+            y = fluid.layers.data(name='y', shape=[7], dtype='float32')
+
+            cost = fluid.layers.edit_distance(input=x,label=y)
+    """
+    helper = LayerHelper("edit_distance", **locals())
+
+    # remove some tokens from input and labels
+    if ignored_tokens is not None and len(ignored_tokens) > 0:
+        erased_input = helper.create_tmp_variable(dtype="int64")
+        erased_label = helper.create_tmp_variable(dtype="int64")
+
+        helper.append_op(
+            type="sequence_erase",
+            inputs={"X": [input]},
+            outputs={"Out": [erased_input]},
+            attrs={"tokens": ignored_tokens})
+        input = erased_input
+
+        helper.append_op(
+            type="sequence_erase",
+            inputs={"X": [label]},
+            outputs={"Out": [erase_label]},
+            attrs={"tokens": ignored_tokens})
+        label = erased_label
+
+    # edit distance op
+    edit_distance_out = helper.create_tmp_variable(dtype="int64")
+    sequence_num = helper.create_tmp_variable(dtype="int64")
+    helper.append_op(
+        type="edit_distance",
+        inputs={"Hyps": [input],
+                "Refs": [label]},
+        outputs={"Out": [edit_distance_out],
+                 "SequenceNum": [sequence_num]},
+        attrs={"normalized": normalized})
+
+    return edit_distance_out, sequence_num
+
+
+def ctc_greedy_decoder(input, blank, name=None):
+    """
+    This op is used to decode sequences by greedy policy by below steps:
+    1. Get the indexes of max value for each row in input. a.k.a. numpy.argmax(input, axis=0).
+    2. For each sequence in result of step1, merge repeated tokens between two blanks and delete all blanks.
+
+    A simple example as below:
+
+    .. code-block:: text
+
+        Given:
+
+        input.data = [[0.6, 0.1, 0.3, 0.1],
+                      [0.3, 0.2, 0.4, 0.1],
+                      [0.1, 0.5, 0.1, 0.3],
+                      [0.5, 0.1, 0.3, 0.1],
+
+                      [0.5, 0.1, 0.3, 0.1],
+                      [0.2, 0.2, 0.2, 0.4],
+                      [0.2, 0.2, 0.1, 0.5],
+                      [0.5, 0.1, 0.3, 0.1]]
+
+        input.lod = [[0, 4, 8]]
+
+        Then:
+
+        output.data = [[2],
+                       [1],
+                       [3]]
+
+        output.lod = [[0, 2, 3]]
+
+    Args:
+
+        input(Variable): (LoDTensor<float>), the probabilities of variable-length sequences, which is a 2-D Tensor with LoD information. It's shape is [Lp, num_classes + 1], where Lp is the sum of all input sequences' length and num_classes is the true number of classes. (not including the blank label).
+
+        blank(int): the blank label index of Connectionist Temporal Classification (CTC) loss, which is in thehalf-opened interval [0, num_classes + 1).
+
+    Returns:
+        Variable: CTC greedy decode result.
+
+    Examples:
+        .. code-block:: python
+
+            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
+
+            cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
+    """
+    helper = LayerHelper("ctc_greedy_decoder", **locals())
+    # top 1 op
+    topk_out = helper.create_tmp_variable(dtype=input.dtype)
+    topk_indices = helper.create_tmp_variable(dtype="int64")
+    helper.append_op(
+        type="top_k",
+        inputs={"X": [input]},
+        outputs={"Out": [topk_out],
+                 "Indices": [topk_indices]},
+        attrs={"k": 1})
+
+    # ctc align op
+    ctc_out = helper.create_tmp_variable(dtype="int64")
+    helper.append_op(
+        type="ctc_align",
+        inputs={"Input": [topk_indices]},
+        outputs={"Output": [ctc_out]},
+        attrs={"merge_repeated": True,
+               "blank": blank})
+    return ctc_out
+
+
 def warpctc(input, label, blank=0, norm_by_times=False, **kwargs):
     """
     An operator integrating the open source Warp-CTC library
@@ -1890,7 +2267,7 @@ def warpctc(input, label, blank=0, norm_by_times=False, **kwargs):
          Temporal Classification (CTC) loss, which is in the
          half-opened interval [0, num_classes + 1).
        norm_by_times: (bool, default: false), whether to normalize
-       the gradients by the number of time-step,which is also the
+       the gradients by the number of time-step, which is also the
        sequence's length. There is no need to normalize the gradients
        if warpctc layer was follewed by a mean_op.
 
@@ -1973,6 +2350,277 @@ def sequence_reshape(input, new_dim):
     return out
 
 
+@autodoc()
+def nce(input,
+        label,
+        num_total_classes,
+        sample_weight=None,
+        param_attr=None,
+        bias_attr=None,
+        num_neg_samples=None):
+    helper = LayerHelper('nce', **locals())
+    assert isinstance(input, Variable)
+    dim = input.shape[1]
+    assert isinstance(label, Variable)
+    num_true_class = label.shape[1]
+    w = helper.create_parameter(
+        attr=helper.param_attr,
+        shape=[num_total_classes, dim],
+        is_bias=False,
+        dtype=input.dtype)
+    b = helper.create_parameter(
+        attr=helper.bias_attr,
+        shape=[num_total_classes, 1],
+        is_bias=True,
+        dtype=input.dtype)
+    cost = helper.create_tmp_variable(dtype=input.dtype)
+    sample_logits = helper.create_tmp_variable(dtype=input.dtype)
+    sample_labels = helper.create_tmp_variable(dtype=label.dtype)
+
+    if num_neg_samples is None:
+        num_neg_samples = 10
+    else:
+        num_neg_samples = int(num_neg_samples)
+
+    attrs = {
+        'num_total_classes': int(num_total_classes),
+        'num_neg_samples': num_neg_samples
+    }
+
+    helper.append_op(
+        type='nce',
+        inputs={
+            'Input': input,
+            'Label': label,
+            'Weight': w,
+            'Bias': b,
+            'SampleWeight': sample_weight if sample_weight is not None else []
+        },
+        outputs={
+            'Cost': cost,
+            'SampleLogits': sample_logits,
+            'SampleLabels': sample_labels
+        },
+        attrs=attrs)
+    return cost / (num_neg_samples + 1)
+
+
+def transpose(x, perm, name=None):
+    """
+    **transpose Layer**
+
+    Permute the dimensions of `input` according to `perm`.
+
+    The `i`-th dimension  of the returned tensor will correspond to the
+    perm[i]-th dimension of `input`.
+
+    Args:
+       input (Variable): (Tensor), A Tensor.
+       perm (list): A permutation of the dimensions of `input`.
+
+    Returns:
+        Variable: A transposed Tensor.
+
+    Examples:
+        .. code-block:: python
+
+            x = fluid.layers.data(name='x', shape=[5, 10, 15], dtype='float32')
+            x_transposed = layers.transpose(x, perm=[1, 0, 2])
+    """
+
+    if len(perm) != len(x.shape):
+        raise ValueError(
+            "Input(perm) is the permutation of dimensions of Input(input). "
+            "It's length shoud be equal to Input(input)'s rank.")
+
+    helper = LayerHelper('transpose', **locals())
+    out = helper.create_tmp_variable(x.dtype)
+    helper.append_op(
+        type='transpose',
+        inputs={'X': [x]},
+        outputs={'Out': [out]},
+        attrs={'axis': perm})
+    return out
+
+
+def im2sequence(input, filter_size=1, stride=1, padding=0, name=None):
+    """
+    Extracts image patches from the input tensor to form a tensor of shape
+    {input.batch_size * output_height * output_width, filter_size_H *
+    filter_size_W * input.channels} which is similar with im2col.
+    This op use filter / kernel to scan images and convert these images to
+    sequences. After expanding, the number of time step are
+    output_height * output_width for an image, in which output_height and
+    output_width are calculated by below equation:
+
+    .. math::
+
+        output\_size = 1 + \
+            (2 * padding + img\_size - block\_size + stride - 1) / stride
+
+    And the dimension of each time step is block_y * block_x * input.channels.
+
+    Args:
+        input (Variable): The input should be a tensor in NCHW format.
+
+        filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
+            it must contain two integers, (filter_size_H, filter_size_W).
+            Otherwise, the filter will be a square.
+
+        stride(int|tuple): The stride size. If stride is a tuple, it must
+            contain two integers, (stride_H, stride_W). Otherwise, the
+            stride_H = stride_W = stride. Default: stride = 1.
+
+        padding(int|tuple): The padding size. If padding is a tuple, it can
+            contain two integers like (padding_H, padding_W) which means
+            padding_up = padding_down = padding_H and
+            padding_left = padding_right = padding_W. Or it can use
+            (padding_up, padding_left, padding_down, padding_right) to indicate
+            paddings of four direction. Otherwise, a scalar padding means
+            padding_up = padding_down = padding_left = padding_right = padding
+            Default: padding = 0.
+
+        name (int): The name of this layer. It is optional.
+
+    Returns:
+        output: The output is a LoDTensor with shape
+        {input.batch_size * output_height * output_width,
+        filter_size_H * filter_size_W * input.channels}.
+        If we regard output as a matrix, each row of this matrix is
+        a step of a sequence.
+
+    Examples:
+
+    As an example:
+
+        .. code-block:: text
+
+            Given:
+
+            x = [[[[ 6.  2.  1.]
+                   [ 8.  3.  5.]
+                   [ 0.  2.  6.]]
+
+                  [[ 2.  4.  4.]
+                   [ 6.  3.  0.]
+                   [ 6.  4.  7.]]]
+
+                 [[[ 6.  7.  1.]
+                   [ 5.  7.  9.]
+                   [ 2.  4.  8.]]
+
+                  [[ 1.  2.  1.]
+                   [ 1.  3.  5.]
+                   [ 9.  0.  8.]]]]
+
+            x.dims = {2, 2, 3, 3}
+
+            And:
+
+            filter = [2, 2]
+            stride = [1, 1]
+            padding = [0, 0]
+
+            Then:
+
+            output.data = [[ 6.  2.  8.  3.  2.  4.  6.  3.]
+                           [ 2.  1.  3.  5.  4.  4.  3.  0.]
+                           [ 8.  3.  0.  2.  6.  3.  6.  4.]
+                           [ 3.  5.  2.  6.  3.  0.  4.  7.]
+                           [ 6.  7.  5.  7.  1.  2.  1.  3.]
+                           [ 7.  1.  7.  9.  2.  1.  3.  5.]
+                           [ 5.  7.  2.  4.  1.  3.  9.  0.]
+                           [ 7.  9.  4.  8.  3.  5.  0.  8.]]
+
+            output.dims = {8, 9}
+
+            output.lod = [[0, 4, 8]]
+
+        The simple usage is:
+
+        .. code-block:: python
+
+            output = fluid.layers.im2sequence(input=layer, stride=[1, 1], filter_size=[2, 2])
+
+    """
+
+    if isinstance(filter_size, int):
+        filter_size = [filter_size, filter_size]
+    if isinstance(stride, int):
+        stride = [stride, stride]
+    if isinstance(padding, int):
+        padding = [padding, padding]
+    if len(padding) == 2:
+        padding.append(padding[0])
+        padding.append(padding[1])
+
+    helper = LayerHelper('im2sequence', **locals())
+    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    helper.append_op(
+        type='im2sequence',
+        inputs={'X': input},
+        outputs={'Out': out},
+        attrs={
+            'kernels': filter_size,
+            'strides': stride,
+            'paddings': padding,
+        })
+    return out
+
+
+def row_conv(input, future_context_size, param_attr=None, act=None):
+    """Row Conv Operator. This layer will apply lookahead convolution to
+    **input**. The input variable should be a 2D LoDTensor with shape [T, D].
+    Parameters with shape [future_context_size + 1, D] will be created. The math
+    equation of row convolution is as follows:
+
+    .. math::
+        Out_{i} = \sum_{j = i} ^ {i + \\tau} X_{j} \odot W_{i - j}
+
+    In the above equation:
+
+    * :math:`Out_{i}`: The i-th row of output variable with shape [1, D].
+    * :math:`\\tau`: Future context size.
+    * :math:`X_{j}`: The j-th row of input variable with shape [1, D].
+    * :math:`W_{i-j}`: The (i-j)-th row of parameters with shape [1, D].
+
+    More details about row_conv please refer to the paper \
+    (http://www.cs.cmu.edu/~dyogatam/papers/wang+etal.iclrworkshop2016.pdf) and
+    the design document \
+    (https://github.com/PaddlePaddle/Paddle/issues/2228#issuecomment-303903645).
+
+    Args:
+        input (Variable): Input variable, a 2D LoDTensor with shape [T, D].
+        future_context_size (int): Future context size. Please note, the shape
+            of convolution kernel is [future_context_size + 1, D].
+        param_attr (ParamAttr): Attributes of parameters, including
+            name, initializer etc.
+        act (str): Non-linear activation to be applied to output variable.
+
+    Returns:
+        Variable: The output tensor with same shape as input tensor.
+
+    Examples:
+        .. code-block:: python
+
+            x = fluid.layers.data(name='x', shape=[16],
+                            dtype='float32', lod_level=1)
+            out = fluid.layers.row_conv(input=x, future_context_size=2)
+    """
+    helper = LayerHelper('row_conv', **locals())
+    dtype = helper.input_dtype()
+    filter_shape = [future_context_size + 1, input.shape[1]]
+    filter_param = helper.create_parameter(
+        attr=helper.param_attr, shape=filter_shape, dtype=dtype)
+    out = helper.create_tmp_variable(dtype)
+    helper.append_op(
+        type='row_conv',
+        inputs={'X': [input],
+                'Filter': [filter_param]},
+        outputs={'Out': [out]})
+    return helper.append_activation(out)
+
+
 def multiplex(inputs, index):
     """
     **Multiplex Layer**
@@ -1996,8 +2644,8 @@ def multiplex(inputs, index):
     Args:
        inputs (list): Input variables which are tensors with same shape and the
                 rank is at least 2.
-       index (Variable): Tensor<int>, index variable which is a 2-D tensor with
-                shape [M, 1] where M for batch size.
+       index (Variable): Tensor<int32>, index variable which is a 2-D tensor
+                with shape [M, 1] where M for batch size.
 
     Returns:
         Variable: Multiplex variable gathered from input variables.
diff --git a/python/paddle/v2/fluid/layers/ops.py b/python/paddle/v2/fluid/layers/ops.py
index b517f8be6a3e5558dd01afe094fb3989cfb3af44..022a94cad440f13383a927233195bb008a688843 100644
--- a/python/paddle/v2/fluid/layers/ops.py
+++ b/python/paddle/v2/fluid/layers/ops.py
@@ -45,10 +45,20 @@ __activations__ = [
 ]
 
 __all__ = [
-    'mean', 'mul', 'reshape', 'scale', 'transpose',
-    'sigmoid_cross_entropy_with_logits', 'elementwise_add', 'elementwise_div',
-    'elementwise_sub', 'elementwise_mul', 'elementwise_max', 'elementwise_min',
-    'clip', 'clip_by_norm', 'sequence_softmax'
+    'mean',
+    'mul',
+    'reshape',
+    'scale',
+    'sigmoid_cross_entropy_with_logits',
+    'elementwise_add',
+    'elementwise_div',
+    'elementwise_sub',
+    'elementwise_mul',
+    'elementwise_max',
+    'elementwise_min',
+    'clip',
+    'clip_by_norm',
+    'sequence_softmax',
 ] + __activations__
 
 for _OP in set(__all__):
diff --git a/python/paddle/v2/fluid/memory_optimization_transpiler.py b/python/paddle/v2/fluid/memory_optimization_transpiler.py
index 1b4b64755963b5edc3d07d861c2a9b6cc3f23587..956c5b66da28fd8e74d4fd12f249688daa72d8ac 100644
--- a/python/paddle/v2/fluid/memory_optimization_transpiler.py
+++ b/python/paddle/v2/fluid/memory_optimization_transpiler.py
@@ -31,10 +31,12 @@ dtype_to_size = {
 
 
 class ControlFlowGraph(object):
-    def __init__(self, Program):
+    def __init__(self, Program, ops, forward_num):
         self._program = Program
-        self._succesors = defaultdict(set)
-        self._presucessors = defaultdict(set)
+        self._ops = ops
+        self._forward_num = forward_num
+        self._successors = defaultdict(set)
+        self._presuccessors = defaultdict(set)
         self._uses = defaultdict(set)
         self._defs = defaultdict(set)
         self._live_in = defaultdict(set)
@@ -45,25 +47,16 @@ class ControlFlowGraph(object):
             self._add(node1, node2)
 
     def _add(self, node1, node2):
-        self._succesors[node1].add(node2)
-        self._presucessors[node2].add(node1)
+        self._successors[node1].add(node2)
+        self._presuccessors[node2].add(node1)
 
     def _build_graph(self):
-        program_desc = self._program.get_desc()
-        block_size = program_desc.num_blocks()
-
-        # TODO(qijun) handle Program with if/while operators
-        self.global_block_desc = program_desc.block(0)
-        self.op_size = self.global_block_desc.op_size()
-
+        self.op_size = len(self._ops)
         op_node_connections = [(i, i + 1) for i in range(self.op_size - 1)]
         self._add_connections(op_node_connections)
-
-        self.ops = [self.global_block_desc.op(i) for i in range(self.op_size)]
-
         for i in range(self.op_size):
-            self._uses[i].update(self.ops[i].input_arg_names())
-            self._defs[i].update(self.ops[i].output_arg_names())
+            self._uses[i].update(self._ops[i].input_arg_names())
+            self._defs[i].update(self._ops[i].output_arg_names())
 
     def _update_graph(self, old_name, new_name, begin_idx=0):
         for i in range(begin_idx, self.op_size):
@@ -103,7 +96,7 @@ class ControlFlowGraph(object):
                 live_out[i] = set(self._live_out[i])
                 self._live_in[i] = self._uses[i] | (
                     self._live_out[i] - self._defs[i])
-                for s in self._succesors[i]:
+                for s in self._successors[i]:
                     self._live_out[i] |= self._live_in[s]
 
             if self._reach_fixed_point(live_in, live_out):
@@ -113,39 +106,76 @@ class ControlFlowGraph(object):
         u = a & b
         return a - u, b - u
 
+    def _has_var(self, block_desc, var_name, is_forward):
+        if is_forward:
+            return block_desc.has_var(str(var_name))
+        else:
+            return block_desc.has_var_recursive(str(var_name))
+
+    def _find_var(self, block_desc, var_name, is_forward):
+        if is_forward:
+            return block_desc.find_var(str(var_name))
+        else:
+            return block_desc.find_var_recursive(str(var_name))
+
     def memory_optimize(self):
+        def check_var_validity(block_desc, x, is_forward):
+            if str(x) == "@EMPTY@":
+                return False
+            if not self._has_var(block_desc, x, is_forward):
+                return False
+            if self._find_var(block_desc, x, is_forward).persistable():
+                return False
+            if self._find_var(
+                    block_desc, x,
+                    is_forward).type() != core.VarDesc.VarType.LOD_TENSOR:
+                return False
+            return True
+
         self._build_graph()
         self._dataflow_analyze()
         self.pool = []
         for i in range(self.op_size):
+            op = self._ops[i]
+            if op.type() == "while" or op.type() == "while_grad":
+                continue
+            block_desc = op.block()
+            is_forward = i < self._forward_num
             if self.pool:
-                out_pair = [(x, self.global_block_desc.var(str(x)).shape())
-                            for x in self._defs[i]]
+                defs_can_optimize = filter(
+                    lambda x: check_var_validity(block_desc, x, is_forward),
+                    self._defs[i])
+                out_pair = [
+                    (x, self._find_var(block_desc, x, is_forward).shape())
+                    for x in defs_can_optimize
+                ]
                 for x, x_shape in out_pair:
-                    if not self.global_block_desc.var(str(x)).persistable():
-                        for index, cache_pair in enumerate(self.pool):
-                            cache_var = cache_pair[0]
-                            cache_shape = cache_pair[1]
-                            if x_shape == cache_shape:
-                                x_dtype = self.global_block_desc.var(str(
-                                    x)).dtype()
-                                cache_dtype = self.global_block_desc.var(
-                                    str(cache_var)).dtype()
+                    for index, cache_pair in enumerate(self.pool):
+                        cache_var = cache_pair[0]
+                        cache_shape = cache_pair[1]
+                        if x_shape == cache_shape:
+                            if self._has_var(block_desc, cache_var, is_forward):
+                                x_dtype = self._find_var(block_desc, x,
+                                                         is_forward).dtype()
+                                cache_dtype = self._find_var(
+                                    block_desc, cache_var, is_forward).dtype()
                                 # TODO(qijun): actually, we should compare dtype_to_size[x_dtype]
                                 # and dtype_to_size[cache_dtype]
                                 if x_dtype == cache_dtype:
-                                    print(
-                                        ("Hit Cache !!!! cache pool index "
-                                         "is %d, var name is %s, "
-                                         "cached var name is %s, "
-                                         "var shape is %s ") %
-                                        (index, x, cache_var, str(cache_shape)))
+                                    print(("Hit Cache !!!! cache pool index "
+                                           "is %d, var name is %s, "
+                                           "cached var name is %s, "
+                                           "var shape is %s ") %
+                                          (index, x, cache_var,
+                                           str(cache_shape)))
                                     self.pool.pop(index)
+                                    if x == cache_var:
+                                        break
                                     _rename_arg_(
-                                        self.ops, x, cache_var, begin_idx=i)
-                                    self._program.current_block().var(str(
-                                        x)).desc = self.global_block_desc.var(
-                                            str(cache_var))
+                                        self._ops, x, cache_var, begin_idx=i)
+                                    self._program.block(block_desc.id).var(
+                                        str(x)).desc = self._find_var(
+                                            block_desc, cache_var, is_forward)
                                     self._update_graph(
                                         x, cache_var, begin_idx=i)
                                     break
@@ -153,20 +183,70 @@ class ControlFlowGraph(object):
             in_diff, out_diff = self._get_diff(self._live_in[i],
                                                self._live_out[i])
             can_optimize = filter(
-                lambda x: not self.global_block_desc.var(str(x)).persistable(),
+                lambda x: check_var_validity(block_desc, x, is_forward),
                 in_diff)
             if can_optimize:
                 for var_name in can_optimize:
-                    self.pool.append(
-                        (var_name,
-                         self.global_block_desc.var(str(var_name)).shape()))
-
-    def get_program(self):
-        return self._program
+                    self.pool.append((var_name, self._find_var(
+                        block_desc, var_name, is_forward).shape()))
+
+
+def get_cfgs(input_program):
+    ops_list = []
+    pdesc = input_program.get_desc()
+    block_desc = pdesc.block(0)
+    op_size = block_desc.op_size()
+    # Get global block ops
+    ops_list.append(([block_desc.op(i) for i in range(op_size)], op_size))
+
+    while_sub_block_ids = []
+    while_grad_sub_block_ids = []
+    while_pair = []
+
+    for i in range(op_size):
+        op = block_desc.op(i)
+        if op.type() == "while":
+            while_sub_block_ids.append(op.attr("sub_block").id)
+        elif op.type() == "while_grad":
+            while_grad_sub_block_ids.append(op.attr("sub_block").id)
+
+    # Find while/while_grad block pair
+    for grad_id in while_grad_sub_block_ids:
+        parent_id = pdesc.block(grad_id).parent
+        if parent_id in while_sub_block_ids:
+            while_pair.append((parent_id, grad_id))
+            while_sub_block_ids.remove(parent_id)
+
+    # Get while/while_grad block ops
+    for parent_id, grad_id in while_pair:
+        while_block_ops = []
+        while_block = pdesc.block(parent_id)
+        while_block_op_size = while_block.op_size()
+        for i in range(while_block_op_size):
+            while_block_ops.append(while_block.op(i))
+
+        while_grad_block = pdesc.block(grad_id)
+        while_grad_block_op_size = while_grad_block.op_size()
+        for i in range(while_grad_block_op_size):
+            while_block_ops.append(while_grad_block.op(i))
+
+        ops_list.append((while_block_ops, while_block_op_size))
+
+    # Process rest while block ops
+    for parent_id in while_sub_block_ids:
+        while_block_ops = []
+        while_block = pdesc.block(parent_id)
+        while_block_op_size = while_block.op_size()
+        for i in range(while_block_op_size):
+            while_block_ops.append(while_block.op(i))
+
+        ops_list.append((while_block_ops, while_block_op_size))
+
+    cfgs = [ControlFlowGraph(input_program, i, j) for i, j in ops_list]
+    return cfgs
 
 
 def memory_optimize(input_program):
-    graph = ControlFlowGraph(input_program)
-    graph.memory_optimize()
-    result_program = graph.get_program()
-    return result_program
+    cfgs = get_cfgs(input_program)
+    for cfg in cfgs:
+        cfg.memory_optimize()
diff --git a/python/paddle/v2/fluid/nets.py b/python/paddle/v2/fluid/nets.py
index a30e646d8cbccb397d11c1f6164946e748f40c5e..6146e3711d3c62d22591b2855d73b5791e4b47d0 100644
--- a/python/paddle/v2/fluid/nets.py
+++ b/python/paddle/v2/fluid/nets.py
@@ -56,7 +56,7 @@ def img_conv_group(input,
                    conv_act=None,
                    param_attr=None,
                    conv_with_batchnorm=False,
-                   conv_batchnorm_drop_rate=None,
+                   conv_batchnorm_drop_rate=0.0,
                    pool_stride=1,
                    pool_type=None,
                    use_cudnn=True):
@@ -127,21 +127,21 @@ def sequence_conv_pool(input,
 
 def glu(input, dim=-1):
     """
-    The gated linear unit composed by split, sigmoid activation and elementwise 
-    multiplication. Specifically, Split the input into two equal sized parts 
-    :math:`a` and :math:`b` along the given dimension and then compute as 
+    The gated linear unit composed by split, sigmoid activation and elementwise
+    multiplication. Specifically, Split the input into two equal sized parts
+    :math:`a` and :math:`b` along the given dimension and then compute as
     following:
 
         .. math::
 
             {GLU}(a, b)= a \otimes \sigma(b)
 
-    Refer to `Language Modeling with Gated Convolutional Networks 
+    Refer to `Language Modeling with Gated Convolutional Networks
     <https://arxiv.org/pdf/1612.08083.pdf>`_.
-    
+
     Args:
         input (Variable): The input variable which is a Tensor or LoDTensor.
-        dim (int): The dimension along which to split. If :math:`dim < 0`, the 
+        dim (int): The dimension along which to split. If :math:`dim < 0`, the
             dimension to split along is :math:`rank(input) + dim`.
 
     Returns:
@@ -164,24 +164,24 @@ def dot_product_attention(querys, keys, values):
     """
     The dot-product attention.
 
-    Attention mechanism can be seen as mapping a query and a set of key-value 
-    pairs to an output. The output is computed as a weighted sum of the values, 
-    where the weight assigned to each value is computed by a compatibility 
+    Attention mechanism can be seen as mapping a query and a set of key-value
+    pairs to an output. The output is computed as a weighted sum of the values,
+    where the weight assigned to each value is computed by a compatibility
     function (dot-product here) of the query with the corresponding key.
-    
-    The dot-product attention can be implemented through (batch) matrix 
+
+    The dot-product attention can be implemented through (batch) matrix
     multipication as follows:
 
         .. math::
 
             Attention(Q, K, V)= softmax(QK^\mathrm{T})V
 
-    Refer to `Attention Is All You Need 
+    Refer to `Attention Is All You Need
     <https://arxiv.org/pdf/1706.03762.pdf>`_.
 
-    Note that batch data containing sequences with different lengths is not 
+    Note that batch data containing sequences with different lengths is not
     supported by this because of the (batch) matrix multipication.
-    
+
     Args:
         query (Variable): The input variable which is a Tensor or LoDTensor.
         key (Variable): The input variable which is a Tensor or LoDTensor.
diff --git a/python/paddle/v2/fluid/tests/book/test_fit_a_line.py b/python/paddle/v2/fluid/tests/book/test_fit_a_line.py
index 462669c262f285a7c6d36cf60f2f3f952c83f6b3..0b954c60b6bc2d721c0373243e747056f8f572cf 100644
--- a/python/paddle/v2/fluid/tests/book/test_fit_a_line.py
+++ b/python/paddle/v2/fluid/tests/book/test_fit_a_line.py
@@ -49,7 +49,7 @@ for pass_id in range(PASS_NUM):
         avg_loss_value, = exe.run(fluid.default_main_program(),
                                   feed=feeder.feed(data),
                                   fetch_list=[avg_cost])
-
+        print(avg_loss_value)
         if avg_loss_value[0] < 10.0:
             exit(0)  # if avg cost less than 10.0, we think our code is good.
 exit(1)
diff --git a/python/paddle/v2/fluid/tests/book/test_machine_translation.py b/python/paddle/v2/fluid/tests/book/test_machine_translation.py
index 53ae200a2387712c63ab67f44d4e9da03ebbe4b2..82b760d693560dae1ab1fa39afdc186f60423e65 100644
--- a/python/paddle/v2/fluid/tests/book/test_machine_translation.py
+++ b/python/paddle/v2/fluid/tests/book/test_machine_translation.py
@@ -17,7 +17,7 @@ import paddle.v2 as paddle
 import paddle.v2.fluid as fluid
 import paddle.v2.fluid.core as core
 import paddle.v2.fluid.framework as framework
-import paddle.v2.fluid.layers as layers
+import paddle.v2.fluid.layers as pd
 from paddle.v2.fluid.executor import Executor
 
 dict_size = 30000
@@ -26,53 +26,136 @@ src_dict, trg_dict = paddle.dataset.wmt14.get_dict(dict_size)
 hidden_dim = 32
 word_dim = 16
 IS_SPARSE = True
-batch_size = 10
-max_length = 50
+batch_size = 2
+max_length = 8
 topk_size = 50
 trg_dic_size = 10000
+beam_size = 2
 
 decoder_size = hidden_dim
 
+place = core.CPUPlace()
 
-def encoder_decoder():
+
+def encoder():
     # encoder
-    src_word_id = layers.data(
+    src_word_id = pd.data(
         name="src_word_id", shape=[1], dtype='int64', lod_level=1)
-    src_embedding = layers.embedding(
+    src_embedding = pd.embedding(
         input=src_word_id,
         size=[dict_size, word_dim],
         dtype='float32',
         is_sparse=IS_SPARSE,
         param_attr=fluid.ParamAttr(name='vemb'))
 
-    fc1 = fluid.layers.fc(input=src_embedding, size=hidden_dim * 4, act='tanh')
-    lstm_hidden0, lstm_0 = layers.dynamic_lstm(input=fc1, size=hidden_dim * 4)
-    encoder_out = layers.sequence_last_step(input=lstm_hidden0)
+    fc1 = pd.fc(input=src_embedding, size=hidden_dim * 4, act='tanh')
+    lstm_hidden0, lstm_0 = pd.dynamic_lstm(input=fc1, size=hidden_dim * 4)
+    encoder_out = pd.sequence_last_step(input=lstm_hidden0)
+    return encoder_out
+
 
+def decoder_train(context):
     # decoder
-    trg_language_word = layers.data(
+    trg_language_word = pd.data(
         name="target_language_word", shape=[1], dtype='int64', lod_level=1)
-    trg_embedding = layers.embedding(
+    trg_embedding = pd.embedding(
         input=trg_language_word,
         size=[dict_size, word_dim],
         dtype='float32',
         is_sparse=IS_SPARSE,
         param_attr=fluid.ParamAttr(name='vemb'))
 
-    rnn = fluid.layers.DynamicRNN()
+    rnn = pd.DynamicRNN()
     with rnn.block():
         current_word = rnn.step_input(trg_embedding)
-        mem = rnn.memory(init=encoder_out)
-        fc1 = fluid.layers.fc(input=[current_word, mem],
+        pre_state = rnn.memory(init=context)
+        current_state = pd.fc(input=[current_word, pre_state],
                               size=decoder_size,
                               act='tanh')
-        out = fluid.layers.fc(input=fc1, size=target_dict_dim, act='softmax')
-        rnn.update_memory(mem, fc1)
-        rnn.output(out)
+
+        current_score = pd.fc(input=current_state,
+                              size=target_dict_dim,
+                              act='softmax')
+        rnn.update_memory(pre_state, current_state)
+        rnn.output(current_score)
 
     return rnn()
 
 
+def decoder_decode(context):
+    init_state = context
+    array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length)
+    counter = pd.zeros(shape=[1], dtype='int64')
+
+    # fill the first element with init_state
+    state_array = pd.create_array('float32')
+    pd.array_write(init_state, array=state_array, i=counter)
+
+    # ids, scores as memory
+    ids_array = pd.create_array('int64')
+    scores_array = pd.create_array('float32')
+
+    init_ids = pd.data(name="init_ids", shape=[1], dtype="int64", lod_level=2)
+    init_scores = pd.data(
+        name="init_scores", shape=[1], dtype="float32", lod_level=2)
+
+    pd.array_write(init_ids, array=ids_array, i=counter)
+    pd.array_write(init_scores, array=scores_array, i=counter)
+
+    cond = pd.less_than(x=counter, y=array_len)
+
+    while_op = pd.While(cond=cond)
+    with while_op.block():
+        pre_ids = pd.array_read(array=ids_array, i=counter)
+        pre_state = pd.array_read(array=state_array, i=counter)
+        pre_score = pd.array_read(array=scores_array, i=counter)
+
+        # expand the lod of pre_state to be the same with pre_score
+        pre_state_expanded = pd.sequence_expand(pre_state, pre_score)
+
+        pre_ids_emb = pd.embedding(
+            input=pre_ids,
+            size=[dict_size, word_dim],
+            dtype='float32',
+            is_sparse=IS_SPARSE)
+
+        # use rnn unit to update rnn
+        current_state = pd.fc(input=[pre_ids_emb, pre_state_expanded],
+                              size=decoder_size,
+                              act='tanh')
+
+        # use score to do beam search
+        current_score = pd.fc(input=current_state,
+                              size=target_dict_dim,
+                              act='softmax')
+        topk_scores, topk_indices = pd.topk(current_score, k=50)
+        selected_ids, selected_scores = pd.beam_search(
+            pre_ids, topk_indices, topk_scores, beam_size, end_id=10, level=0)
+
+        pd.increment(x=counter, value=1, in_place=True)
+
+        # update the memories
+        pd.array_write(current_state, array=state_array, i=counter)
+        pd.array_write(selected_ids, array=ids_array, i=counter)
+        pd.array_write(selected_scores, array=scores_array, i=counter)
+
+        pd.less_than(x=counter, y=array_len, cond=cond)
+
+    translation_ids, translation_scores = pd.beam_search_decode(
+        ids=ids_array, scores=scores_array)
+
+    # return init_ids, init_scores
+
+    return translation_ids, translation_scores
+
+
+def set_init_lod(data, lod, place):
+    res = core.LoDTensor()
+    res.set(data, place)
+    res.set_lod(lod)
+    return res
+
+
 def to_lodtensor(data, place):
     seq_lens = [len(seq) for seq in data]
     cur_len = 0
@@ -88,12 +171,13 @@ def to_lodtensor(data, place):
     return res
 
 
-def main():
-    rnn_out = encoder_decoder()
-    label = layers.data(
+def train_main():
+    context = encoder()
+    rnn_out = decoder_train(context)
+    label = pd.data(
         name="target_language_next_word", shape=[1], dtype='int64', lod_level=1)
-    cost = layers.cross_entropy(input=rnn_out, label=label)
-    avg_cost = fluid.layers.mean(x=cost)
+    cost = pd.cross_entropy(input=rnn_out, label=label)
+    avg_cost = pd.mean(x=cost)
 
     optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4)
     optimizer.minimize(avg_cost)
@@ -103,13 +187,12 @@ def main():
             paddle.dataset.wmt14.train(dict_size), buf_size=1000),
         batch_size=batch_size)
 
-    place = core.CPUPlace()
     exe = Executor(place)
 
     exe.run(framework.default_startup_program())
 
     batch_id = 0
-    for pass_id in xrange(2):
+    for pass_id in xrange(1):
         for data in train_data():
             word_data = to_lodtensor(map(lambda x: x[0], data), place)
             trg_word = to_lodtensor(map(lambda x: x[1], data), place)
@@ -125,9 +208,48 @@ def main():
             print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) +
                   " avg_cost=" + str(avg_cost_val))
             if batch_id > 3:
-                exit(0)
+                break
             batch_id += 1
 
 
+def decode_main():
+    context = encoder()
+    translation_ids, translation_scores = decoder_decode(context)
+
+    exe = Executor(place)
+    exe.run(framework.default_startup_program())
+
+    init_ids_data = np.array([1 for _ in range(batch_size)], dtype='int64')
+    init_scores_data = np.array(
+        [1. for _ in range(batch_size)], dtype='float32')
+    init_ids_data = init_ids_data.reshape((batch_size, 1))
+    init_scores_data = init_scores_data.reshape((batch_size, 1))
+    init_lod = [i for i in range(batch_size)] + [batch_size]
+    init_lod = [init_lod, init_lod]
+
+    train_data = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.wmt14.train(dict_size), buf_size=1000),
+        batch_size=batch_size)
+    for _, data in enumerate(train_data()):
+        init_ids = set_init_lod(init_ids_data, init_lod, place)
+        init_scores = set_init_lod(init_scores_data, init_lod, place)
+
+        src_word_data = to_lodtensor(map(lambda x: x[0], data), place)
+
+        result_ids, result_scores = exe.run(
+            framework.default_main_program(),
+            feed={
+                'src_word_id': src_word_data,
+                'init_ids': init_ids,
+                'init_scores': init_scores
+            },
+            fetch_list=[translation_ids, translation_scores],
+            return_numpy=False)
+        print result_ids.lod()
+        break
+
+
 if __name__ == '__main__':
-    main()
+    # train_main()
+    decode_main()
diff --git a/python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py b/python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py
index 618191424150eb7c5a24407fc2e106ee8825fedb..117f74c59ad5bf6bb67711801cd7b9a41f39f1f8 100644
--- a/python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py
+++ b/python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py
@@ -65,13 +65,13 @@ def lstm_net(dict_dim, class_dim=2, emb_dim=32, seq_len=80, batch_size=50):
 
     emb = fluid.layers.embedding(input=data, size=[dict_dim, emb_dim])
     emb = fluid.layers.reshape(x=emb, shape=[batch_size, seq_len, emb_dim])
-    emb = fluid.layers.transpose(x=emb, axis=[1, 0, 2])
+    emb = fluid.layers.transpose(x=emb, perm=[1, 0, 2])
 
     c_pre_init = fluid.layers.fill_constant(
         dtype=emb.dtype, shape=[batch_size, emb_dim], value=0.0)
     c_pre_init.stop_gradient = False
     layer_1_out = lstm(emb, c_pre_init=c_pre_init, hidden_dim=emb_dim)
-    layer_1_out = fluid.layers.transpose(x=layer_1_out, axis=[1, 0, 2])
+    layer_1_out = fluid.layers.transpose(x=layer_1_out, perm=[1, 0, 2])
 
     prediction = fluid.layers.fc(input=layer_1_out,
                                  size=class_dim,
diff --git a/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py b/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py
index 52c7ecdeb3646fdce36937b84ba8956947371d87..9774edebfb1de0ae73970d582c620f8a984a4ebf 100644
--- a/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py
+++ b/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py
@@ -68,10 +68,10 @@ else:
         fluid.io.save_persistables(exe, "./fit_a_line.model/")
         fluid.io.load_persistables(exe, "./fit_a_line.model/")
         for data in train_reader():
-            avg_loss_value, = exe.run(trainer_prog,
-                                      feed=feeder.feed(data),
-                                      fetch_list=[avg_cost])
-
+            avg_loss_value = exe.run(trainer_prog,
+                                     feed=feeder.feed(data),
+                                     fetch_list=[avg_cost])
+            print("loss:" + str(avg_loss_value))
             if avg_loss_value[0] < 10.0:
                 exit(0)
 exit(1)
diff --git a/python/paddle/v2/fluid/tests/book_distribute/notest_machine_translation.py b/python/paddle/v2/fluid/tests/book_distribute/notest_machine_translation.py
new file mode 100644
index 0000000000000000000000000000000000000000..adeacd4adf2150e0302965d80457e26d07c6b96d
--- /dev/null
+++ b/python/paddle/v2/fluid/tests/book_distribute/notest_machine_translation.py
@@ -0,0 +1,157 @@
+#  Copyright (c) 2018 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.
+
+import numpy as np
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+import paddle.v2.fluid.core as core
+import paddle.v2.fluid.framework as framework
+import paddle.v2.fluid.layers as layers
+from paddle.v2.fluid.executor import Executor
+import os
+
+dict_size = 30000
+source_dict_dim = target_dict_dim = dict_size
+src_dict, trg_dict = paddle.dataset.wmt14.get_dict(dict_size)
+hidden_dim = 32
+word_dim = 16
+IS_SPARSE = True
+batch_size = 10
+max_length = 50
+topk_size = 50
+trg_dic_size = 10000
+
+decoder_size = hidden_dim
+
+
+def encoder_decoder():
+    # encoder
+    src_word_id = layers.data(
+        name="src_word_id", shape=[1], dtype='int64', lod_level=1)
+    src_embedding = layers.embedding(
+        input=src_word_id,
+        size=[dict_size, word_dim],
+        dtype='float32',
+        is_sparse=IS_SPARSE,
+        param_attr=fluid.ParamAttr(name='vemb'))
+
+    fc1 = fluid.layers.fc(input=src_embedding, size=hidden_dim * 4, act='tanh')
+    lstm_hidden0, lstm_0 = layers.dynamic_lstm(input=fc1, size=hidden_dim * 4)
+    encoder_out = layers.sequence_last_step(input=lstm_hidden0)
+
+    # decoder
+    trg_language_word = layers.data(
+        name="target_language_word", shape=[1], dtype='int64', lod_level=1)
+    trg_embedding = layers.embedding(
+        input=trg_language_word,
+        size=[dict_size, word_dim],
+        dtype='float32',
+        is_sparse=IS_SPARSE,
+        param_attr=fluid.ParamAttr(name='vemb'))
+
+    rnn = fluid.layers.DynamicRNN()
+    with rnn.block():
+        current_word = rnn.step_input(trg_embedding)
+        mem = rnn.memory(init=encoder_out)
+        fc1 = fluid.layers.fc(input=[current_word, mem],
+                              size=decoder_size,
+                              act='tanh')
+        out = fluid.layers.fc(input=fc1, size=target_dict_dim, act='softmax')
+        rnn.update_memory(mem, fc1)
+        rnn.output(out)
+
+    return rnn()
+
+
+def to_lodtensor(data, place):
+    seq_lens = [len(seq) for seq in data]
+    cur_len = 0
+    lod = [cur_len]
+    for l in seq_lens:
+        cur_len += l
+        lod.append(cur_len)
+    flattened_data = np.concatenate(data, axis=0).astype("int64")
+    flattened_data = flattened_data.reshape([len(flattened_data), 1])
+    res = core.LoDTensor()
+    res.set(flattened_data, place)
+    res.set_lod([lod])
+    return res
+
+
+def main():
+    rnn_out = encoder_decoder()
+    label = layers.data(
+        name="target_language_next_word", shape=[1], dtype='int64', lod_level=1)
+    cost = layers.cross_entropy(input=rnn_out, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+
+    optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4)
+    optimize_ops, params_grads = optimizer.minimize(avg_cost)
+
+    train_data = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.wmt14.train(dict_size), buf_size=1000),
+        batch_size=batch_size)
+
+    place = core.CPUPlace()
+    exe = Executor(place)
+
+    t = fluid.DistributeTranspiler()
+    # all parameter server endpoints list for spliting parameters
+    pserver_endpoints = os.getenv("PSERVERS")
+    # server endpoint for current node
+    current_endpoint = os.getenv("SERVER_ENDPOINT")
+    # run as trainer or parameter server
+    training_role = os.getenv(
+        "TRAINING_ROLE", "TRAINER")  # get the training role: trainer/pserver
+    t.transpile(
+        optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
+
+    if training_role == "PSERVER":
+        if not current_endpoint:
+            print("need env SERVER_ENDPOINT")
+            exit(1)
+        pserver_prog = t.get_pserver_program(current_endpoint)
+        pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
+        exe.run(pserver_startup)
+        exe.run(pserver_prog)
+    elif training_role == "TRAINER":
+        trainer_prog = t.get_trainer_program()
+        exe.run(framework.default_startup_program())
+
+        batch_id = 0
+        for pass_id in xrange(2):
+            for data in train_data():
+                word_data = to_lodtensor(map(lambda x: x[0], data), place)
+                trg_word = to_lodtensor(map(lambda x: x[1], data), place)
+                trg_word_next = to_lodtensor(map(lambda x: x[2], data), place)
+                outs = exe.run(trainer_prog,
+                               feed={
+                                   'src_word_id': word_data,
+                                   'target_language_word': trg_word,
+                                   'target_language_next_word': trg_word_next
+                               },
+                               fetch_list=[avg_cost])
+                avg_cost_val = np.array(outs[0])
+                print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) +
+                      " avg_cost=" + str(avg_cost_val))
+                if batch_id > 3:
+                    exit(0)
+                batch_id += 1
+    else:
+        print("environment var TRAINER_ROLE should be TRAINER os PSERVER")
+
+
+if __name__ == '__main__':
+    main()
diff --git a/python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py b/python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py
index cf054bb0fe778d34add4ac456f672a8b47483e84..7ad5e2c594f24999e298533b6c05ba688a935f0b 100644
--- a/python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py
+++ b/python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py
@@ -16,6 +16,11 @@ import numpy as np
 import paddle.v2 as paddle
 import paddle.v2.fluid as fluid
 
+# need to fix random seed and training data to compare the loss
+# value accurately calculated by the default and the memory optimization
+# version.
+fluid.default_startup_program().random_seed = 111
+
 x = fluid.layers.data(name='x', shape=[13], dtype='float32')
 
 y_predict = fluid.layers.fc(input=x, size=1, act=None)
@@ -28,15 +33,18 @@ avg_cost = fluid.layers.mean(x=cost)
 sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1)
 sgd_optimizer.minimize(avg_cost)
 
-# memopt_program = fluid.default_main_program()
-memopt_program = fluid.memory_optimize(fluid.default_main_program())
+fluid.memory_optimize(fluid.default_main_program())
 
 BATCH_SIZE = 200
 
+# fix the order of training data
 train_reader = paddle.batch(
-    paddle.reader.shuffle(
-        paddle.dataset.uci_housing.train(), buf_size=500),
-    batch_size=BATCH_SIZE)
+    paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE)
+
+# train_reader = paddle.batch(
+#     paddle.reader.shuffle(
+#         paddle.dataset.uci_housing.train(), buf_size=500),
+#     batch_size=BATCH_SIZE)
 
 place = fluid.CPUPlace()
 feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
@@ -49,7 +57,7 @@ for pass_id in range(PASS_NUM):
     fluid.io.save_persistables(exe, "./fit_a_line.model/")
     fluid.io.load_persistables(exe, "./fit_a_line.model/")
     for data in train_reader():
-        avg_loss_value, = exe.run(memopt_program,
+        avg_loss_value, = exe.run(fluid.default_main_program(),
                                   feed=feeder.feed(data),
                                   fetch_list=[avg_cost])
 
diff --git a/python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py b/python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py
index 42b3cb81ce67d38494677f3ecbfb1e07f7c0c3ad..26673afd83c48328c3f354e82bfa3725aa4805b5 100644
--- a/python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py
+++ b/python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py
@@ -19,6 +19,11 @@ import sys
 import paddle.v2 as paddle
 import paddle.v2.fluid as fluid
 
+# need to fix random seed and training data to compare the loss
+# value accurately calculated by the default and the memory optimization
+# version.
+fluid.default_startup_program().random_seed = 111
+
 
 def resnet_cifar10(input, depth=32):
     def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'):
@@ -117,31 +122,37 @@ opts = optimizer.minimize(avg_cost)
 
 accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
 
-# memopt_program = fluid.default_main_program()
-memopt_program = fluid.memory_optimize(fluid.default_main_program())
+fluid.memory_optimize(fluid.default_main_program())
 
 BATCH_SIZE = 128
 PASS_NUM = 1
 
+# fix the order of training data
 train_reader = paddle.batch(
-    paddle.reader.shuffle(
-        paddle.dataset.cifar.train10(), buf_size=128 * 10),
-    batch_size=BATCH_SIZE)
+    paddle.dataset.cifar.train10(), batch_size=BATCH_SIZE)
+
+# train_reader = paddle.batch(
+#     paddle.reader.shuffle(
+#         paddle.dataset.cifar.train10(), buf_size=128 * 10),
+#     batch_size=BATCH_SIZE)
 
 place = fluid.CPUPlace()
 exe = fluid.Executor(place)
 feeder = fluid.DataFeeder(place=place, feed_list=[images, label])
 exe.run(fluid.default_startup_program())
 
+i = 0
 for pass_id in range(PASS_NUM):
     accuracy.reset(exe)
     for data in train_reader():
-        loss, acc = exe.run(memopt_program,
+        loss, acc = exe.run(fluid.default_main_program(),
                             feed=feeder.feed(data),
                             fetch_list=[avg_cost] + accuracy.metrics)
         pass_acc = accuracy.eval(exe)
         print("loss:" + str(loss) + " acc:" + str(acc) + " pass_acc:" + str(
             pass_acc))
         # this model is slow, so if we can train two mini batch, we think it works properly.
-        exit(0)
+        if i > 2:
+            exit(0)
+        i += 1
 exit(1)
diff --git a/python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py b/python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py
new file mode 100644
index 0000000000000000000000000000000000000000..ffd53e7a78142162317a677de49c1821635a65b5
--- /dev/null
+++ b/python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py
@@ -0,0 +1,144 @@
+#   Copyright (c) 2018 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.
+
+import numpy as np
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+import paddle.v2.fluid.core as core
+import paddle.v2.fluid.framework as framework
+import paddle.v2.fluid.layers as layers
+from paddle.v2.fluid.executor import Executor
+
+dict_size = 30000
+source_dict_dim = target_dict_dim = dict_size
+src_dict, trg_dict = paddle.dataset.wmt14.get_dict(dict_size)
+hidden_dim = 32
+word_dim = 16
+IS_SPARSE = True
+batch_size = 10
+max_length = 50
+topk_size = 50
+trg_dic_size = 10000
+
+decoder_size = hidden_dim
+
+# need to fix random seed and training data to compare the loss
+# value accurately calculated by the default and the memory optimization
+# version.
+fluid.default_startup_program().random_seed = 111
+
+
+def encoder_decoder():
+    # encoder
+    src_word_id = layers.data(
+        name="src_word_id", shape=[1], dtype='int64', lod_level=1)
+    src_embedding = layers.embedding(
+        input=src_word_id,
+        size=[dict_size, word_dim],
+        dtype='float32',
+        is_sparse=IS_SPARSE,
+        param_attr=fluid.ParamAttr(name='vemb'))
+
+    fc1 = fluid.layers.fc(input=src_embedding, size=hidden_dim * 4, act='tanh')
+    lstm_hidden0, lstm_0 = layers.dynamic_lstm(input=fc1, size=hidden_dim * 4)
+    encoder_out = layers.sequence_last_step(input=lstm_hidden0)
+
+    # decoder
+    trg_language_word = layers.data(
+        name="target_language_word", shape=[1], dtype='int64', lod_level=1)
+    trg_embedding = layers.embedding(
+        input=trg_language_word,
+        size=[dict_size, word_dim],
+        dtype='float32',
+        is_sparse=IS_SPARSE,
+        param_attr=fluid.ParamAttr(name='vemb'))
+
+    rnn = fluid.layers.DynamicRNN()
+    with rnn.block():
+        current_word = rnn.step_input(trg_embedding)
+        mem = rnn.memory(init=encoder_out)
+        fc1 = fluid.layers.fc(input=[current_word, mem],
+                              size=decoder_size,
+                              act='tanh')
+        out = fluid.layers.fc(input=fc1, size=target_dict_dim, act='softmax')
+        rnn.update_memory(mem, fc1)
+        rnn.output(out)
+
+    return rnn()
+
+
+def to_lodtensor(data, place):
+    seq_lens = [len(seq) for seq in data]
+    cur_len = 0
+    lod = [cur_len]
+    for l in seq_lens:
+        cur_len += l
+        lod.append(cur_len)
+    flattened_data = np.concatenate(data, axis=0).astype("int64")
+    flattened_data = flattened_data.reshape([len(flattened_data), 1])
+    res = core.LoDTensor()
+    res.set(flattened_data, place)
+    res.set_lod([lod])
+    return res
+
+
+def main():
+    rnn_out = encoder_decoder()
+    label = layers.data(
+        name="target_language_next_word", shape=[1], dtype='int64', lod_level=1)
+    cost = layers.cross_entropy(input=rnn_out, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+
+    optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4)
+    optimizer.minimize(avg_cost)
+
+    fluid.memory_optimize(fluid.default_main_program())
+
+    # fix the order of training data
+    train_data = paddle.batch(
+        paddle.dataset.wmt14.train(dict_size), batch_size=batch_size)
+
+    # train_data = paddle.batch(
+    #     paddle.reader.shuffle(
+    #         paddle.dataset.wmt14.train(dict_size), buf_size=1000),
+    #     batch_size=batch_size)
+
+    place = core.CPUPlace()
+    exe = Executor(place)
+
+    exe.run(framework.default_startup_program())
+
+    batch_id = 0
+    for pass_id in xrange(10):
+        for data in train_data():
+            word_data = to_lodtensor(map(lambda x: x[0], data), place)
+            trg_word = to_lodtensor(map(lambda x: x[1], data), place)
+            trg_word_next = to_lodtensor(map(lambda x: x[2], data), place)
+            outs = exe.run(fluid.default_main_program(),
+                           feed={
+                               'src_word_id': word_data,
+                               'target_language_word': trg_word,
+                               'target_language_next_word': trg_word_next
+                           },
+                           fetch_list=[avg_cost])
+            avg_cost_val = np.array(outs[0])
+            print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) +
+                  " avg_cost=" + str(avg_cost_val))
+            if batch_id > 2:
+                exit(0)
+            batch_id += 1
+
+
+if __name__ == '__main__':
+    main()
diff --git a/python/paddle/v2/fluid/tests/test_bipartite_match_op.py b/python/paddle/v2/fluid/tests/test_bipartite_match_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..74138298978c7c18936f53761b313887f07aea81
--- /dev/null
+++ b/python/paddle/v2/fluid/tests/test_bipartite_match_op.py
@@ -0,0 +1,100 @@
+#  Copyright (c) 2018 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.
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+def bipartite_match(distance, match_indices, match_dist):
+    """Bipartite Matching algorithm.
+    Arg:
+        distance (numpy.array) : The distance of two entries with shape [M, N].
+        match_indices (numpy.array): the matched indices from column to row
+            with shape [1, N], it must be initialized to -1.
+        match_dist (numpy.array): The matched distance from column to row
+            with shape [1, N], it must be initialized to 0.
+    """
+    match_pair = []
+    row, col = distance.shape
+    for i in range(row):
+        for j in range(col):
+            match_pair.append((i, j, distance[i][j]))
+
+    match_sorted = sorted(match_pair, key=lambda tup: tup[2], reverse=True)
+
+    row_indices = -1 * np.ones((row, ), dtype=np.int)
+
+    idx = 0
+    for i, j, dist in match_sorted:
+        if idx >= row:
+            break
+        if match_indices[j] == -1 and row_indices[i] == -1 and dist > 0:
+            match_indices[j] = i
+            row_indices[i] = j
+            match_dist[j] = dist
+            idx += 1
+
+
+def batch_bipartite_match(distance, lod):
+    """Bipartite Matching algorithm for batch input.
+    Arg:
+        distance (numpy.array) : The distance of two entries with shape [M, N].
+        lod (list of int): The offsets of each input in this batch.
+    """
+    n = len(lod) - 1
+    m = distance.shape[1]
+    match_indices = -1 * np.ones((n, m), dtype=np.int)
+    match_dist = np.zeros((n, m), dtype=np.float32)
+    for i in range(len(lod) - 1):
+        bipartite_match(distance[lod[i]:lod[i + 1], :], match_indices[i, :],
+                        match_dist[i, :])
+    return match_indices, match_dist
+
+
+class TestBipartiteMatchOpForWithLoD(OpTest):
+    def setUp(self):
+        self.op_type = 'bipartite_match'
+        lod = [[0, 5, 11, 23]]
+        dist = np.random.random((23, 217)).astype('float32')
+        match_indices, match_dist = batch_bipartite_match(dist, lod[0])
+
+        self.inputs = {'DistMat': (dist, lod)}
+        self.outputs = {
+            'ColToRowMatchIndices': (match_indices),
+            'ColToRowMatchDis': (match_dist),
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestBipartiteMatchOpWithoutLoD(OpTest):
+    def setUp(self):
+        self.op_type = 'bipartite_match'
+        lod = [[0, 8]]
+        dist = np.random.random((8, 17)).astype('float32')
+        match_indices, match_dist = batch_bipartite_match(dist, lod[0])
+
+        self.inputs = {'DistMat': dist}
+        self.outputs = {
+            'ColToRowMatchIndices': match_indices,
+            'ColToRowMatchDis': match_dist,
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/fluid/tests/test_detection_output_op.py b/python/paddle/v2/fluid/tests/test_detection_output_op.py
index 4a9cd474b81a419bfb42c202327df04c0d2e5bd9..8a5e06b38f5ed5336ef02bac7876610758b44258 100644
--- a/python/paddle/v2/fluid/tests/test_detection_output_op.py
+++ b/python/paddle/v2/fluid/tests/test_detection_output_op.py
@@ -68,4 +68,6 @@ class TestUnpoolOp(OpTest):
 
 
 if __name__ == '__main__':
-    unittest.main()
+    # FIXME: detection_output_op will be rewritten. This unittest should be
+    # enabled after rewriting.
+    exit(0)  # temporary disable this unittest
diff --git a/python/paddle/v2/fluid/tests/test_edit_distance_op.py b/python/paddle/v2/fluid/tests/test_edit_distance_op.py
index 11cb85a151d1a4e213bcc52592d2f860f69b457f..bebdc5cba36fc96d31162d0d7d43e52064ca8e2d 100644
--- a/python/paddle/v2/fluid/tests/test_edit_distance_op.py
+++ b/python/paddle/v2/fluid/tests/test_edit_distance_op.py
@@ -61,6 +61,7 @@ class TestEditDistanceOp(OpTest):
 
         num_strs = len(x1_lod) - 1
         distance = np.zeros((num_strs, 1)).astype("float32")
+        sequence_num = np.array(2).astype("int64")
         for i in range(0, num_strs):
             distance[i] = Levenshtein(
                 hyp=x1[x1_lod[i]:x1_lod[i + 1]],
@@ -70,7 +71,7 @@ class TestEditDistanceOp(OpTest):
                 distance[i] = distance[i] / len_ref
         self.attrs = {'normalized': normalized}
         self.inputs = {'Hyps': (x1, [x1_lod]), 'Refs': (x2, [x2_lod])}
-        self.outputs = {'Out': distance}
+        self.outputs = {'Out': distance, 'SequenceNum': sequence_num}
 
     def test_check_output(self):
         self.check_output()
@@ -89,6 +90,7 @@ class TestEditDistanceOpNormalized(OpTest):
 
         num_strs = len(x1_lod) - 1
         distance = np.zeros((num_strs, 1)).astype("float32")
+        sequence_num = np.array(3).astype("int64")
         for i in range(0, num_strs):
             distance[i] = Levenshtein(
                 hyp=x1[x1_lod[i]:x1_lod[i + 1]],
@@ -98,7 +100,7 @@ class TestEditDistanceOpNormalized(OpTest):
                 distance[i] = distance[i] / len_ref
         self.attrs = {'normalized': normalized}
         self.inputs = {'Hyps': (x1, [x1_lod]), 'Refs': (x2, [x2_lod])}
-        self.outputs = {'Out': distance}
+        self.outputs = {'Out': distance, 'SequenceNum': sequence_num}
 
     def test_check_output(self):
         self.check_output()
diff --git a/python/paddle/v2/fluid/tests/test_gradient_clip.py b/python/paddle/v2/fluid/tests/test_gradient_clip.py
index 4e6e6a1ef6961d8f087dfc1ac5a4c4a8ad90032e..9337791c21183fd7c2e5d6b9d47c99d762c93d46 100644
--- a/python/paddle/v2/fluid/tests/test_gradient_clip.py
+++ b/python/paddle/v2/fluid/tests/test_gradient_clip.py
@@ -40,7 +40,8 @@ p_g = fluid.backward.append_backward(loss=avg_cost)
 p_g_clip = fluid.backward.append_backward(loss=avg_cost_clip)
 
 with fluid.program_guard(main_program=prog_clip):
-    fluid.clip.gradient_clip_by_global_norm(clip_norm=CLIP)
+    fluid.clip.set_gradient_clip(
+        fluid.clip.GradientClipByGlobalNorm(clip_norm=CLIP))
     p_g_clip = fluid.clip.append_gradient_clip_ops(p_g_clip)
 
 grad_list = [elem[1] for elem in p_g]
diff --git a/python/paddle/v2/fluid/tests/test_im2sequence_op.py b/python/paddle/v2/fluid/tests/test_im2sequence_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..2cab3e31a50034e3b1b362b59690e425aef1c399
--- /dev/null
+++ b/python/paddle/v2/fluid/tests/test_im2sequence_op.py
@@ -0,0 +1,167 @@
+#  Copyright (c) 2018 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.
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+def get_output_shape(attrs, in_shape):
+    img_height = in_shape[2]
+    img_width = in_shape[3]
+
+    paddings = attrs['paddings']
+    kernels = attrs['kernels']
+    strides = attrs['strides']
+
+    output_height = \
+      1 +  \
+      (img_height + paddings[0] + paddings[2] - kernels[0] + strides[0] - 1) / \
+          strides[0]
+
+    output_width = \
+      1 + \
+      (img_width + paddings[1] + paddings[3] - kernels[1] + strides[1] - 1) / \
+          strides[1]
+
+    return output_height, output_width
+
+
+def im2col(attrs, im, col):
+    """
+    im: {CHW}
+    col:
+        {outputHeight, outputWidth, inputChannels, filterHeight, filterWidth}
+    """
+    input_channels, input_height, input_width = im.shape
+    output_height, output_width, _, filter_height, filter_width = col.shape
+
+    stride_height, stride_width = attrs['strides']
+    padding_height, padding_width = attrs['paddings'][0:2]
+
+    for col_row_idx in range(0, output_height):
+        for col_col_idx in range(0, output_width):
+            for channel in range(0, input_channels):
+                for filter_row_idx in range(0, filter_height):
+                    for filter_col_idx in range(0, filter_width):
+                        im_row_offset = col_row_idx * stride_height \
+                            + filter_row_idx - padding_height
+
+                        im_col_offset = col_col_idx * stride_width \
+                            + filter_col_idx - padding_width
+
+                        if (im_row_offset < 0 or
+                                im_row_offset >= input_height or
+                                im_col_offset < 0 or
+                                im_col_offset >= input_width):
+                            col[col_row_idx][col_col_idx][channel][\
+                                filter_row_idx][filter_col_idx] = 0.0
+                        else:
+                            im_offset = (channel * input_height + im_row_offset \
+                                         ) * input_width + im_col_offset
+
+                            col[col_row_idx][col_col_idx][channel][\
+                                filter_row_idx][filter_col_idx] = im[channel][ \
+                                    im_row_offset][im_col_offset]
+
+
+def Im2Sequence(inputs, attrs):
+    output_height, output_width = get_output_shape(attrs, inputs.shape)
+    img_channels = inputs.shape[1]
+    batch_size = inputs.shape[0]
+    out = np.zeros([
+        batch_size, output_height, output_width, img_channels,
+        attrs['kernels'][0], attrs['kernels'][1]
+    ]).astype("float32")
+
+    for i in range(len(inputs)):
+        im2col(attrs, inputs[i], out[i])
+
+    out = out.reshape([
+        batch_size * output_height * output_width,
+        img_channels * attrs['kernels'][0] * attrs['kernels'][1]
+    ])
+    return out
+
+
+class TestBlockExpandOp(OpTest):
+    def config(self):
+        self.batch_size = 1
+        self.img_channels = 3
+        self.img_height = 4
+        self.img_width = 4
+        self.attrs = {
+            'kernels': [2, 2],
+            'strides': [1, 1],
+            'paddings': [1, 1, 1, 1]
+        }
+
+    def setUp(self):
+        self.config()
+        self.op_type = "im2sequence"
+        x = np.random.uniform(0.1, 1, [
+            self.batch_size, self.img_channels, self.img_height, self.img_width
+        ]).astype("float32")
+
+        out = Im2Sequence(x, self.attrs)
+        self.inputs = {'X': x}
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X'], 'Out')
+
+
+class TestBlockExpandOpCase2(TestBlockExpandOp):
+    def config(self):
+        self.batch_size = 2
+        self.img_channels = 3
+        self.img_height = 4
+        self.img_width = 5
+        self.attrs = {
+            'kernels': [2, 1],
+            'strides': [2, 1],
+            'paddings': [2, 1, 2, 1]
+        }
+
+
+class TestBlockExpandOpCase3(TestBlockExpandOp):
+    def config(self):
+        self.batch_size = 3
+        self.img_channels = 1
+        self.img_height = 4
+        self.img_width = 5
+        self.attrs = {
+            'kernels': [2, 1],
+            'strides': [2, 1],
+            'paddings': [2, 0, 2, 0]
+        }
+
+
+class TestBlockExpandOpCase4(TestBlockExpandOp):
+    def config(self):
+        self.batch_size = 2
+        self.img_channels = 2
+        self.img_height = 3
+        self.img_width = 3
+        self.attrs = {
+            'kernels': [2, 2],
+            'strides': [1, 1],
+            'paddings': [0, 0, 0, 0]
+        }
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/fluid/tests/test_iou_similarity_op.py b/python/paddle/v2/fluid/tests/test_iou_similarity_op.py
new file mode 100755
index 0000000000000000000000000000000000000000..128f2e4977195a563efcd26364cc6261da2dd685
--- /dev/null
+++ b/python/paddle/v2/fluid/tests/test_iou_similarity_op.py
@@ -0,0 +1,55 @@
+#   Copyright (c) 2018 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.
+
+import unittest
+import numpy as np
+import sys
+import math
+from op_test import OpTest
+
+
+class TestIOUSimilarityOp(OpTest):
+    def test_check_output(self):
+        self.check_output()
+
+    def setUp(self):
+        self.op_type = "iou_similarity"
+        self.boxes1 = np.array(
+            [[4.0, 3.0, 7.0, 5.0], [5.0, 6.0, 10.0, 7.0]]).astype('float32')
+        self.boxes2 = np.array([[3.0, 4.0, 6.0, 8.0], [14.0, 14.0, 15.0, 15.0],
+                                [0.0, 0.0, 20.0, 20.0]]).astype('float32')
+        self.output = np.array(
+            [[2.0 / 16.0, 0, 6.0 / 400.0],
+             [1.0 / 16.0, 0.0, 5.0 / 400.0]]).astype('float32')
+
+        self.inputs = {'X': self.boxes1, 'Y': self.boxes2}
+
+        self.outputs = {'Out': self.output}
+
+
+class TestIOUSimilarityOpWithLoD(TestIOUSimilarityOp):
+    def test_check_output(self):
+        self.check_output()
+
+    def setUp(self):
+        super(TestIOUSimilarityOpWithLoD, self).setUp()
+        self.boxes1_lod = [[0, 1, 2]]
+        self.output_lod = [[0, 1, 2]]
+
+        self.inputs = {'X': (self.boxes1, self.boxes1_lod), 'Y': self.boxes2}
+        self.outputs = {'Out': (self.output, self.output_lod)}
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/fluid/tests/test_layers.py b/python/paddle/v2/fluid/tests/test_layers.py
index dc143f6b9f34c912a4c8782f5da7ba24e08041e7..4e863625422c93c77ad4fb65be35580943d1cf54 100644
--- a/python/paddle/v2/fluid/tests/test_layers.py
+++ b/python/paddle/v2/fluid/tests/test_layers.py
@@ -17,8 +17,9 @@ import unittest
 
 import paddle.v2.fluid.layers as layers
 import paddle.v2.fluid.nets as nets
-from paddle.v2.fluid.framework import Program, program_guard
+from paddle.v2.fluid.framework import Program, program_guard, default_main_program
 from paddle.v2.fluid.param_attr import ParamAttr
+import decorators
 
 
 class TestBook(unittest.TestCase):
@@ -225,6 +226,59 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(out)
         print(str(program))
 
+    def test_im2sequence(self):
+        print("test_im2sequence")
+        program = Program()
+        with program_guard(program):
+            x = layers.data(name='x', shape=[3, 128, 128], dtype='float32')
+            output = layers.im2sequence(
+                input=x, stride=[1, 1], filter_size=[2, 2])
+            self.assertIsNotNone(output)
+        print(str(program))
+
+    @decorators.prog_scope()
+    def test_nce(self):
+        window_size = 5
+        words = []
+        for i in xrange(window_size):
+            words.append(
+                layers.data(
+                    name='word_{0}'.format(i), shape=[1], dtype='int64'))
+
+        dict_size = 10000
+        label_word = int(window_size / 2) + 1
+
+        embs = []
+        for i in xrange(window_size):
+            if i == label_word:
+                continue
+
+            emb = layers.embedding(
+                input=words[i],
+                size=[dict_size, 32],
+                param_attr='emb.w',
+                is_sparse=True)
+
+            embs.append(emb)
+
+        embs = layers.concat(input=embs, axis=1)
+        loss = layers.nce(input=embs,
+                          label=words[label_word],
+                          num_total_classes=dict_size,
+                          param_attr='nce.w',
+                          bias_attr='nce.b')
+        avg_loss = layers.mean(x=loss)
+        self.assertIsNotNone(avg_loss)
+        print(str(default_main_program()))
+
+    def test_row_conv(self):
+        program = Program()
+        with program_guard(program):
+            x = layers.data(name='x', shape=[16], dtype='float32', lod_level=1)
+            out = layers.row_conv(input=x, future_context_size=2)
+            self.assertIsNotNone(out)
+        print(str(program))
+
     def test_multiplex(self):
         program = Program()
         with program_guard(program):
diff --git a/python/paddle/v2/fluid/tests/test_lookup_table_op.py b/python/paddle/v2/fluid/tests/test_lookup_table_op.py
index d5255ba31f7c9e45cf29f412546146234f822026..0c566c76c91dce8dcfc882eed998f492ae3cde76 100644
--- a/python/paddle/v2/fluid/tests/test_lookup_table_op.py
+++ b/python/paddle/v2/fluid/tests/test_lookup_table_op.py
@@ -33,5 +33,19 @@ class TestLookupTableOp(OpTest):
         self.check_grad(['W'], 'Out', no_grad_set=set('Ids'))
 
 
+class TestLookupTableOpWithPadding(TestLookupTableOp):
+    def test_check_output(self):
+        ids = np.squeeze(self.inputs['Ids'])
+        padding_idx = np.random.choice(ids, 1)[0]
+        self.outputs['Out'][ids == padding_idx] = np.zeros(31)
+        self.attrs = {'padding_idx': long(padding_idx)}
+        self.check_output()
+
+    def test_check_grad(self):
+        # Since paddings are not trainable and fixed in forward, the gradient of 
+        # paddings makes no sense and we don't test the gradient here.
+        pass
+
+
 if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/fluid/tests/test_nce.py b/python/paddle/v2/fluid/tests/test_nce.py
index 3ae727a573855b3cb618a8fab70404adf3d92f51..9a51c1f612a0d5363d36e6642ed3b409970025b1 100644
--- a/python/paddle/v2/fluid/tests/test_nce.py
+++ b/python/paddle/v2/fluid/tests/test_nce.py
@@ -109,4 +109,6 @@ class TestNCECase1(TestNCE):
 
 
 if __name__ == '__main__':
+    # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/7778
+    exit(0)
     unittest.main()
diff --git a/python/paddle/v2/fluid/tests/test_prior_box_op.py b/python/paddle/v2/fluid/tests/test_prior_box_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..ca8d2bca74ce2d4be8160c8851e393489691ae56
--- /dev/null
+++ b/python/paddle/v2/fluid/tests/test_prior_box_op.py
@@ -0,0 +1,148 @@
+#   Copyright (c) 2018 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.
+
+import unittest
+import numpy as np
+import sys
+import math
+from op_test import OpTest
+
+
+class TestPriorBoxOp(OpTest):
+    def set_data(self):
+        self.init_test_params()
+        self.init_test_input()
+        self.init_test_output()
+        self.inputs = {'Input': self.input, 'Image': self.image}
+
+        self.attrs = {
+            'min_sizes': self.min_sizes,
+            'max_sizes': self.max_sizes,
+            'aspect_ratios': self.aspect_ratios,
+            'variances': self.variances,
+            'flip': self.flip,
+            'clip': self.clip,
+            'step_w': self.step_w,
+            'step_h': self.step_h,
+            'offset': self.offset
+        }
+
+        self.outputs = {'Boxes': self.out_boxes, 'Variances': self.out_var}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        return
+
+    def setUp(self):
+        self.op_type = "prior_box"
+        self.set_data()
+
+    def init_test_params(self):
+        self.layer_w = 4
+        self.layer_h = 4
+
+        self.image_w = 20
+        self.image_h = 20
+
+        self.step_w = float(self.image_w) / float(self.layer_w)
+        self.step_h = float(self.image_h) / float(self.layer_h)
+
+        self.input_channels = 2
+        self.image_channels = 3
+        self.batch_size = 10
+
+        self.min_sizes = [2, 4]
+        self.min_sizes = np.array(self.min_sizes).astype('int64')
+        self.max_sizes = [5, 10]
+        self.max_sizes = np.array(self.max_sizes).astype('int64')
+        self.aspect_ratios = [2.0, 3.0]
+        self.flip = True
+        self.real_aspect_ratios = [1, 2.0, 1.0 / 2.0, 3.0, 1.0 / 3.0]
+        self.aspect_ratios = np.array(
+            self.aspect_ratios, dtype=np.float).flatten()
+        self.variances = [0.1, 0.1, 0.2, 0.2]
+        self.variances = np.array(self.variances, dtype=np.float).flatten()
+
+        self.clip = True
+
+        self.num_priors = len(self.real_aspect_ratios) * len(self.min_sizes)
+        if len(self.max_sizes) > 1:
+            self.num_priors += len(self.max_sizes)
+        self.offset = 0.5
+
+    def init_test_input(self):
+        self.image = np.random.random(
+            (self.batch_size, self.image_channels, self.image_w,
+             self.image_h)).astype('float32')
+
+        self.input = np.random.random(
+            (self.batch_size, self.input_channels, self.layer_w,
+             self.layer_h)).astype('float32')
+
+    def init_test_output(self):
+        out_dim = (self.layer_h, self.layer_w, self.num_priors, 4)
+        out_boxes = np.zeros(out_dim).astype('float32')
+        out_var = np.zeros(out_dim).astype('float32')
+
+        idx = 0
+        for h in range(self.layer_h):
+            for w in range(self.layer_w):
+                c_x = (w + self.offset) * self.step_w
+                c_y = (h + self.offset) * self.step_h
+                idx = 0
+                for s in range(len(self.min_sizes)):
+                    min_size = self.min_sizes[s]
+                    c_w = c_h = min_size / 2.
+                    out_boxes[h, w, idx, :] = [
+                        (c_x - c_w) / self.image_w, (c_y - c_h) / self.image_h,
+                        (c_x + c_w) / self.image_w, (c_y + c_h) / self.image_h
+                    ]
+                    idx += 1
+
+                    if len(self.max_sizes) > 0:
+                        max_size = self.max_sizes[s]
+                        # second prior: aspect_ratio = 1,
+                        c_w = c_h = math.sqrt(min_size * max_size) / 2
+                        out_boxes[h, w, idx, :] = [(c_x - c_w) / self.image_w,
+                                                   (c_y - c_h) / self.image_h,
+                                                   (c_x + c_w) / self.image_w,
+                                                   (c_y + c_h) / self.image_h]
+                        idx += 1
+
+                    # rest of priors
+                    for r in range(len(self.real_aspect_ratios)):
+                        ar = self.real_aspect_ratios[r]
+                        if math.fabs(ar - 1.) < 1e-6:
+                            continue
+                        c_w = min_size * math.sqrt(ar) / 2
+                        c_h = (min_size / math.sqrt(ar)) / 2
+                        out_boxes[h, w, idx, :] = [(c_x - c_w) / self.image_w,
+                                                   (c_y - c_h) / self.image_h,
+                                                   (c_x + c_w) / self.image_w,
+                                                   (c_y + c_h) / self.image_h]
+                        idx += 1
+        # clip the prior's coordidate such that it is within[0, 1]
+        if self.clip:
+            out_boxes = np.clip(out_boxes, 0.0, 1.0)
+        # set the variance.
+        out_var = np.tile(self.variances, (self.layer_h, self.layer_w,
+                                           self.num_priors, 1))
+        self.out_boxes = out_boxes.astype('float32')
+        self.out_var = out_var.astype('float32')
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/fluid/tests/test_reduce_op.py b/python/paddle/v2/fluid/tests/test_reduce_op.py
index 1a4af39fb9dbc3de7d6746ee92a8e0c232e76c9f..c669f73a7c6de0735b3c580ed4f0ed8ba359a040 100644
--- a/python/paddle/v2/fluid/tests/test_reduce_op.py
+++ b/python/paddle/v2/fluid/tests/test_reduce_op.py
@@ -20,7 +20,7 @@ from op_test import OpTest
 class TestSumOp(OpTest):
     def setUp(self):
         self.op_type = "reduce_sum"
-        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float32")}
+        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")}
         self.outputs = {'Out': self.inputs['X'].sum(axis=0)}
 
     def test_check_output(self):
@@ -33,7 +33,7 @@ class TestSumOp(OpTest):
 class TestMeanOp(OpTest):
     def setUp(self):
         self.op_type = "reduce_mean"
-        self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float32")}
+        self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float64")}
         self.attrs = {'dim': 1}
         self.outputs = {'Out': self.inputs['X'].mean(axis=self.attrs['dim'])}
 
@@ -49,7 +49,7 @@ class TestMaxOp(OpTest):
 
     def setUp(self):
         self.op_type = "reduce_max"
-        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float32")}
+        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")}
         self.attrs = {'dim': -1}
         self.outputs = {'Out': self.inputs['X'].max(axis=self.attrs['dim'])}
 
@@ -62,7 +62,7 @@ class TestMinOp(OpTest):
 
     def setUp(self):
         self.op_type = "reduce_min"
-        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float32")}
+        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")}
         self.attrs = {'dim': 2}
         self.outputs = {'Out': self.inputs['X'].min(axis=self.attrs['dim'])}
 
@@ -73,7 +73,7 @@ class TestMinOp(OpTest):
 class TestKeepDimReduce(OpTest):
     def setUp(self):
         self.op_type = "reduce_sum"
-        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float32")}
+        self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")}
         self.attrs = {'dim': -2, 'keep_dim': True}
         self.outputs = {
             'Out': self.inputs['X'].sum(axis=self.attrs['dim'], keepdims=True)
@@ -89,7 +89,7 @@ class TestKeepDimReduce(OpTest):
 class Test1DReduce(OpTest):
     def setUp(self):
         self.op_type = "reduce_sum"
-        self.inputs = {'X': np.random.random(20).astype("float32")}
+        self.inputs = {'X': np.random.random(20).astype("float64")}
         self.outputs = {'Out': self.inputs['X'].sum(axis=0)}
 
     def test_check_output(self):
@@ -102,7 +102,7 @@ class Test1DReduce(OpTest):
 class TestReduceAll(OpTest):
     def setUp(self):
         self.op_type = "reduce_sum"
-        self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float32")}
+        self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float64")}
         self.attrs = {'reduce_all': True}
         self.outputs = {'Out': self.inputs['X'].sum()}
 
diff --git a/python/paddle/v2/image.py b/python/paddle/v2/image.py
index 1429d6b1e08fe4ab2d1c5a0f19f1cedbcbc85abd..e5000e440cc8d822dbd38dce3978d2722d32ebe4 100644
--- a/python/paddle/v2/image.py
+++ b/python/paddle/v2/image.py
@@ -176,7 +176,6 @@ def resize_short(im, size):
     :param size: the shorter edge size of image after resizing.
     :type size: int
     """
-    assert im.shape[-1] == 1 or im.shape[-1] == 3
     h, w = im.shape[:2]
     h_new, w_new = size, size
     if h > w:
@@ -267,7 +266,7 @@ def random_crop(im, size, is_color=True):
     return im
 
 
-def left_right_flip(im):
+def left_right_flip(im, is_color=True):
     """
     Flip an image along the horizontal direction.
     Return the flipped image.
@@ -278,13 +277,15 @@ def left_right_flip(im):
 
         im = left_right_flip(im)
     
-    :paam im: input image with HWC layout
+    :param im: input image with HWC layout or HW layout for gray image
     :type im: ndarray
+    :param is_color: whether input image is color or not
+    :type is_color: bool
     """
-    if len(im.shape) == 3:
+    if len(im.shape) == 3 and is_color:
         return im[:, ::-1, :]
     else:
-        return im[:, ::-1, :]
+        return im[:, ::-1]
 
 
 def simple_transform(im,
@@ -319,11 +320,12 @@ def simple_transform(im,
     """
     im = resize_short(im, resize_size)
     if is_train:
-        im = random_crop(im, crop_size)
+        im = random_crop(im, crop_size, is_color=is_color)
         if np.random.randint(2) == 0:
-            im = left_right_flip(im)
+            im = left_right_flip(im, is_color)
     else:
-        im = center_crop(im, crop_size)
+        im = center_crop(im, crop_size, is_color)
+        im = center_crop(im, crop_size, is_color=is_color)
     if len(im.shape) == 3:
         im = to_chw(im)
 
@@ -331,8 +333,10 @@ def simple_transform(im,
     if mean is not None:
         mean = np.array(mean, dtype=np.float32)
         # mean value, may be one value per channel 
-        if mean.ndim == 1:
+        if mean.ndim == 1 and is_color:
             mean = mean[:, np.newaxis, np.newaxis]
+        elif mean.ndim == 1:
+            mean = mean
         else:
             # elementwise mean
             assert len(mean.shape) == len(im)
@@ -372,6 +376,6 @@ def load_and_transform(filename,
                  mean values per channel.
     :type mean: numpy array | list
     """
-    im = load_image(filename)
+    im = load_image(filename, is_color)
     im = simple_transform(im, resize_size, crop_size, is_train, is_color, mean)
     return im
diff --git a/tools/manylinux1/Dockerfile.x64 b/tools/manylinux1/Dockerfile.x64
index 2c6ba650a5d7996bef212e88a16f2a159ca377e7..0f1b8331309248aaaf0ed32cf14c583a4cdb7437 100644
--- a/tools/manylinux1/Dockerfile.x64
+++ b/tools/manylinux1/Dockerfile.x64
@@ -35,7 +35,7 @@ RUN cd /opt && wget -q --no-check-certificate https://github.com/google/protobuf
     cd protobuf-3.1.0 && ./configure && make -j4 && make install && cd .. && rm -f protobuf-cpp-3.1.0.tar.gz
 
 
-RUN yum install -y sqlite-devel zlib-devel openssl-devel boost boost-devel pcre-devel vim tk-devel tkinter libtool
+RUN yum install -y sqlite-devel zlib-devel openssl-devel pcre-devel vim tk-devel tkinter libtool
 
 RUN wget -O /root/requirements.txt https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/python/requirements.txt