Merge branch 'develop' into build_arm

.dockerignore

-.gitignore
\ No newline at end of file

.dockerignore

+*.DS_Store
+build/
+*.user
+.vscode
+.idea
+.project
+.cproject
+.pydevproject
+Makefile
+.test_env/
+third_party/
+*~
+bazel-*
+
+!build/*.deb

.gitignore

@@ -7,6 +7,7 @@ build/
 .project
 .cproject
 .pydevproject
+.settings/
 Makefile
 .test_env/
 third_party/

.gitmodules

-[submodule "book"]
-	path = book
-	url = https://github.com/PaddlePaddle/book.git
-	branch = develop
\ No newline at end of file

CMakeLists.txt

@@ -39,7 +39,7 @@ option(WITH_RDMA        "Compile PaddlePaddle with RDMA support"        OFF)
 option(WITH_TIMER       "Compile PaddlePaddle with stats timer"         OFF)
 option(WITH_PROFILER    "Compile PaddlePaddle with GPU profiler"        OFF)
 option(WITH_DOC         "Compile PaddlePaddle with documentation"       OFF)
-option(ON_COVERALLS     "Compile PaddlePaddle with code coverage"       OFF)
+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)
 
@@ -81,14 +81,21 @@ include_directories("${PROJ_ROOT}/paddle/cuda/include")
 include_directories("${CMAKE_CURRENT_BINARY_DIR}/proto")
 
 set(EXTERNAL_LIBS
-    # have not include gtest here.
     ${GFLAGS_LIBRARIES}
     ${GLOG_LIBRARIES}
     ${CBLAS_LIBRARIES}
     ${PROTOBUF_LIBRARY}
     ${ZLIB_LIBRARIES}
+    ${PYTHON_LIBRARIES}
 )
 
+if(WITH_GPU)
+    list(APPEND EXTERNAL_LIB ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
+    if(NOT WITH_DSO)
+        list(APPEND EXTERNAL_LIB ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY})
+    endif(NOT WITH_DSO)
+endif(WITH_GPU)
+
 add_subdirectory(proto)
 add_subdirectory(paddle)
 add_subdirectory(python)

paddle/scripts/docker/Dockerfile.gpu → Dockerfile

+# A image for building paddle binaries
+# Use cuda devel base image for both cpu and gpu environment
 FROM nvidia/cuda:7.5-cudnn5-devel-ubuntu14.04
 MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
 
-ARG DEBIAN_FRONTEND=noninteractive
 ARG UBUNTU_MIRROR
-RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi'
+RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi'
 
 # ENV variables
 ARG BUILD_WOBOQ
-ARG BUILD_AND_INSTALL
+ARG WITH_GPU
 ARG WITH_AVX
 ARG WITH_DOC
 ARG WITH_STYLE_CHECK
 
 ENV BUILD_WOBOQ=${BUILD_WOBOQ:-OFF}
-ENV BUILD_AND_INSTALL=${BUILD_AND_INSTALL:-OFF}
-ENV WITH_GPU=ON
+ENV WITH_GPU=${WITH_AVX:-OFF}
 ENV WITH_AVX=${WITH_AVX:-ON}
 ENV WITH_DOC=${WITH_DOC:-OFF}
 ENV WITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF}
-ENV DOCKER_BUILD=TRUE
 
 ENV HOME /root
-
 # Add bash enhancements
 COPY ./paddle/scripts/docker/root/ /root/
 
@@ -30,7 +28,7 @@ RUN apt-get update && \
     apt-get install -y wget unzip tar xz-utils bzip2 gzip coreutils && \
     apt-get install -y curl sed grep graphviz libjpeg-dev zlib1g-dev && \
     apt-get install -y python-numpy python-matplotlib gcc g++ gfortran && \
-    apt-get install -y automake locales clang-format-3.8 && \
+    apt-get install -y automake locales clang-format-3.8 swig && \
     apt-get clean -y
 
 # git credential to skip password typing
@@ -50,10 +48,6 @@ RUN curl -sSL https://cmake.org/files/v3.4/cmake-3.4.1.tar.gz | tar -xz && \
     cd cmake-3.4.1 && ./bootstrap && make -j `nproc` && make install && \
     cd .. && rm -rf cmake-3.4.1
 
-COPY . /paddle/
-RUN cd /paddle/ && git submodule update --init --recursive
-RUN /paddle/paddle/scripts/docker/build.sh
-
 VOLUME ["/usr/share/nginx/html/data", "/usr/share/nginx/html/paddle"]
 
 # Configure OpenSSH server. c.f. https://docs.docker.com/engine/examples/running_ssh_service
@@ -63,9 +57,5 @@ RUN sed -ri 's/^PermitRootLogin\s+.*/PermitRootLogin yes/' /etc/ssh/sshd_config
 RUN sed -ri 's/UsePAM yes/#UsePAM yes/g' /etc/ssh/sshd_config
 EXPOSE 22
 
-# Jupyter Notebook: Paddle book
-EXPOSE 8888
-
-COPY ./paddle/scripts/docker/entrypoint /opt/bin/
-
-CMD ["/opt/bin/entrypoint"]
+# development image default do build work
+CMD ["bash", "/paddle/paddle/scripts/docker/build.sh"]

README.md

@@ -2,8 +2,8 @@
 
 
 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
-[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/)
-[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/cn/index.html)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/develop/doc/)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/doc_cn/)
 [![Coverage Status](https://coveralls.io/repos/github/PaddlePaddle/Paddle/badge.svg?branch=develop)](https://coveralls.io/github/PaddlePaddle/Paddle?branch=develop)
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
@@ -59,36 +59,36 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl
     the capability of PaddlePaddle to make a huge impact for your product.
 
 ## Installation
-Check out the [Install Guide](http://paddlepaddle.org/doc/build/) to install from
-pre-built packages (**docker image**, **deb package**) or
-directly build on **Linux** and **Mac OS X** from the source code.
+
+It is recommended to check out the
+[Docker installation guide](http://www.paddlepaddle.org/develop/doc/getstarted/build_and_install/docker_install_en.html)
+before looking into the
+[build from source guide](http://www.paddlepaddle.org/develop/doc/getstarted/build_and_install/build_from_source_en.html)
 
 ## Documentation
-Both [English Docs](http://paddlepaddle.org/doc/) and [Chinese Docs](http://paddlepaddle.org/doc_cn/) are provided for our users and developers.
 
-- [Quick Start](http://paddlepaddle.org/doc/demo/quick_start/index_en) <br>
-   You can follow the quick start tutorial to learn how use PaddlePaddle
-   step-by-step.
+We provide [English](http://www.paddlepaddle.org/develop/doc/) and
+[Chinese](http://www.paddlepaddle.org/doc_cn/) documentation.
+
+- [Deep Learning 101](http://book.paddlepaddle.org/index.en.html)
+
+  You might want to start from the this online interactive book that can run in Jupyter Notebook.
+
+- [Distributed Training](http://www.paddlepaddle.org/develop/doc/howto/usage/cluster/cluster_train_en.html)
+
+  You can run distributed training jobs on MPI clusters.
+
+- [Distributed Training on Kubernetes](http://www.paddlepaddle.org/develop/doc/howto/usage/k8s/k8s_en.html)
 
-- [Example and Demo](http://paddlepaddle.org/doc/demo/) <br>
-   We provide five demos, including: image classification, sentiment analysis,
-   sequence to sequence model, recommendation, semantic role labeling.
+   You can also run distributed training jobs on Kubernetes clusters.
 
-- [Distributed Training](http://paddlepaddle.org/doc/cluster) <br>
-  This system supports training deep learning models on multiple machines
-  with data parallelism.
+- [Python API](http://www.paddlepaddle.org/develop/doc/api/index_en.html)
 
-- [Python API](http://paddlepaddle.org/doc/ui/) <br>
-   PaddlePaddle supports using either Python interface or C++ to build your
-   system. We also use SWIG to wrap C++ source code to create a user friendly
-   interface for Python. You can also use SWIG to create interface for your
-   favorite programming language.
+   Our new API enables much shorter programs.
 
-- [How to Contribute](http://paddlepaddle.org/doc/build/contribute_to_paddle.html) <br>
-   We sincerely appreciate your interest and contributions. If you would like to
-   contribute, please read the contribution guide.
+- [How to Contribute](http://www.paddlepaddle.org/develop/doc/howto/dev/contribute_to_paddle_en.html)
 
-- [Source Code Documents](http://paddlepaddle.org/doc/source/) <br>
+   We appreciate your contributions!
 
 ## Ask Questions
 

book @ 6e3875eb

-Subproject commit 6e3875eb62533de1f2c1088a477719eb57b9732c

cmake/FindSphinx.cmake

@@ -72,7 +72,7 @@ function( Sphinx_add_target target_name builder conf cache source destination )
     ${source}
     ${destination}
     COMMENT "Generating sphinx documentation: ${builder}"
-    COMMAND cd ${destination} && ln -s ./index_*.html index.html
+    COMMAND cd ${destination} && ln -sf ./index_*.html index.html
     )
 
   set_property(

cmake/coveralls.cmake

@@ -61,7 +61,7 @@ function(code_coverage _COVERAGE_SRCS _COVERALLS_UPLOAD _CMAKE_SCRIPT_PATH)
     endif()
 endfunction()
 
-if(ON_COVERALLS)
+if(WITH_COVERAGE)
     set(CMAKE_BUILD_TYPE "Debug")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -O0 -fprofile-arcs -ftest-coverage")
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -g -O0 -fprofile-arcs -ftest-coverage")

cmake/coverallsGcovJsons.cmake

@@ -134,7 +134,7 @@ foreach(GCDA ${GCDA_FILES})
 	# If -p is not specified then the file is named only "the_file.c.gcov"
 	#
 	execute_process(
-		COMMAND "${GCOV_EXECUTABLE} -p -o ${GCDA_DIR} ${GCDA} >/dev/null"
+		COMMAND ${GCOV_EXECUTABLE} -p -o ${GCDA_DIR} ${GCDA} >/dev/null
 		WORKING_DIRECTORY ${GCDA_DIR}
 	)
 endforeach()

cmake/external/openblas.cmake

@@ -47,7 +47,7 @@ IF(NOT ${CBLAS_FOUND})
         PREFIX              ${CBLAS_SOURCES_DIR}
         INSTALL_DIR         ${CBLAS_INSTALL_DIR}
         BUILD_IN_SOURCE     1
-        BUILD_COMMAND       ${CMAKE_MAKE_PROGRAM} FC=${CMAKE_Fortran_COMPILER} CC=${CMAKE_C_COMPILER} HOSTCC=${CMAKE_C_COMPILER} NO_SHARED=1 libs netlib
+        BUILD_COMMAND       ${CMAKE_MAKE_PROGRAM} FC=${CMAKE_Fortran_COMPILER} CC=${CMAKE_C_COMPILER} HOSTCC=${CMAKE_C_COMPILER} DYNAMIC_ARCH=1 NO_SHARED=1 libs netlib
         INSTALL_COMMAND     ${CMAKE_MAKE_PROGRAM} install NO_SHARED=1 PREFIX=<INSTALL_DIR>
         UPDATE_COMMAND      ""
         CONFIGURE_COMMAND   ""

cmake/external/protobuf.cmake

@@ -14,7 +14,8 @@
 
 INCLUDE(ExternalProject)
 
-FIND_PACKAGE(Protobuf 3.1)
+set(PROTOBUF_VERSION 3.1)
+FIND_PACKAGE(Protobuf ${PROTOBUF_VERSION})
 
 IF(PROTOBUF_FOUND)
     EXEC_PROGRAM(${PROTOBUF_PROTOC_EXECUTABLE} ARGS --version OUTPUT_VARIABLE PROTOBUF_VERSION)

cmake/util.cmake

@@ -71,21 +71,10 @@ function(link_paddle_exe TARGET_NAME)
         generate_rdma_links()
     endif()
 
-    if(WITH_METRIC)
-        if(WITH_GPU)
-            set(METRIC_LIBS paddle_metric_learning paddle_dserver_lib metric metric_cpu)
-        else()
-            set(METRIC_LIBS paddle_metric_learning paddle_dserver_lib metric_cpu)
-        endif()
-    else()
-        set(METRIC_LIBS "")
-    endif()
-
     target_circle_link_libraries(${TARGET_NAME}
         ARCHIVE_START
         paddle_gserver
         paddle_function
-        ${METRIC_LIBS}
         ARCHIVE_END
         paddle_pserver
         paddle_trainer_lib
@@ -95,33 +84,12 @@ function(link_paddle_exe TARGET_NAME)
         paddle_parameter
         paddle_proto
         paddle_cuda
-        ${METRIC_LIBS}
         ${EXTERNAL_LIBS}
         ${CMAKE_THREAD_LIBS_INIT}
         ${CMAKE_DL_LIBS}
         ${RDMA_LD_FLAGS}
         ${RDMA_LIBS})
 
-    if(WITH_PYTHON)
-        target_link_libraries(${TARGET_NAME}
-            ${PYTHON_LIBRARIES} util)
-    endif()
-
-    if(WITH_GPU)
-        target_link_libraries(${TARGET_NAME} ${CUDA_CUDART_LIBRARY})
-        if(NOT WITH_DSO OR WITH_METRIC)
-            target_link_libraries(${TARGET_NAME}
-                ${CUDNN_LIBRARY}
-                ${CUDA_curand_LIBRARY})
-            CUDA_ADD_CUBLAS_TO_TARGET(${TARGET_NAME})
-        endif()
-
-        check_library_exists(rt clock_gettime "time.h" HAVE_CLOCK_GETTIME )
-        if(HAVE_CLOCK_GETTIME)
-            target_link_libraries(${TARGET_NAME} rt)
-        endif()
-    endif()
-
     add_dependencies(${TARGET_NAME} ${external_project_dependencies})
 endfunction()
 

demo/introduction/api_train_v2.py

@@ -14,7 +14,7 @@ def main():
                                 act=paddle.activation.Linear(),
                                 bias_attr=paddle.attr.Param(name='b'))
     y = paddle.layer.data(name='y', type=paddle.data_type.dense_vector(1))
-    cost = paddle.layer.regression_cost(input=y_predict, label=y)
+    cost = paddle.layer.mse_cost(input=y_predict, label=y)
 
     # create parameters
     parameters = paddle.parameters.create(cost)

demo/introduction/trainer_config.py

@@ -34,5 +34,5 @@ y_predict = fc_layer(
     size=1,
     act=LinearActivation(),
     bias_attr=ParamAttr(name='b'))
-cost = regression_cost(input=y_predict, label=y)
+cost = mse_cost(input=y_predict, label=y)
 outputs(cost)

demo/recommendation/api_train_v2.py

@@ -61,7 +61,7 @@ def main():
 
     inference = paddle.layer.cos_sim(
         a=usr_combined_features, b=mov_combined_features, size=1, scale=5)
-    cost = paddle.layer.regression_cost(
+    cost = paddle.layer.mse_cost(
         input=inference,
         label=paddle.layer.data(
             name='score', type=paddle.data_type.dense_vector(1)))

demo/recommendation/trainer_config.py

@@ -86,10 +86,7 @@ movie_feature = construct_feature("movie")
 user_feature = construct_feature("user")
 similarity = cos_sim(a=movie_feature, b=user_feature)
 if not is_predict:
-    outputs(
-        regression_cost(
-            input=similarity, label=data_layer(
-                'rating', size=1)))
+    outputs(mse_cost(input=similarity, label=data_layer('rating', size=1)))
 
     define_py_data_sources2(
         'data/train.list',

doc/api/v1/trainer_config_helpers/layers.rst

@@ -432,6 +432,12 @@ multi_binary_label_cross_entropy
     :members: multi_binary_label_cross_entropy
     :noindex:
 
+mse_cost
+---------
+..  automodule:: paddle.trainer_config_helpers.layers
+    :members: mse_cost
+    :noindex:
+
 huber_cost
 ----------
 ..  automodule:: paddle.trainer_config_helpers.layers
@@ -450,6 +456,12 @@ rank_cost
     :members: rank_cost
     :noindex:
 
+sum_cost
+---------
+..  automodule:: paddle.trainer_config_helpers.layers
+    :members: sum_cost
+    :noindex:
+
 crf_layer
 -----------------
 ..  automodule:: paddle.trainer_config_helpers.layers
@@ -486,12 +498,6 @@ hsigmoid
     :members: hsigmoid
     :noindex:
 
-sum_cost
----------
-..  automodule:: paddle.trainer_config_helpers.layers
-    :members: sum_cost
-    :noindex:
-
 Check Layer 
 ============
 

doc/design/multi_language_interface/why_plain_c.md

+# Paddle多语言接口实现
+## 背景
+
+Paddle需要一个多语言接口，这个接口需要做到:
+
+* 有标准的，良好的文档
+    * 例如Python可以使用[Sphinx](http://www.sphinx-doc.org/en/stable/)生成API文档，golang可以使用[GoDoc](https://godoc.org/golang.org/x/tools/cmd/godoc)生成文档。这都需要这个接口按照约定俗成的规则来注释完备。
+* 不同语言的接口适应不同语言的特性
+    * 例如Java与Python的错误处理是直接扔出来Exception，而对于golang错误处理应该使用返回值。
+
+## 基本要求
+
+Paddle的多语言接口实现包括一下几个方面:
+
+* 我们使用动态库来分发Paddle。在这个动态库中不嵌入任何其他语言的解释器，也不使用其他动态库。
+* 这个动态库使用C99标准的头文件导出一些函数，不使用/导出C++符号。
+* 不导出Paddle内部的结构体、类，仅仅使用`void*`指针作为类型的句柄(handler)。
+* 不使用SWIG这种代码生成器，而是手写多语言绑定。
+
+
+## 原因
+
+### 使用动态库来分发Paddle
+
+* Paddle的链接方式比较复杂
+    * 如果用户要把Paddle的静态库（libpaddle.a）链接到自己的程序里，得使用 `--whole-archive` (for GCC) 或者 `--force_load` (for Clang) 参数，来确保把 libpaddle.a 里所有的符号都写入自己的程序的二进制文件里。这是因为 Paddle 的源码里使用了[object factory design pattern](http://stackoverflow.com/a/1310326/724872)。
+* 编译型语言，例如C/C++使用静态库和动态库难度差不多。但是解释性语言，例如[Python](http://stackoverflow.com/questions/19560594/how-to-import-static-library-in-python)或者[Java](http://stackoverflow.com/questions/24493337/linking-static-library-with-jni)，只能调用Paddle的动态库，否则得把Paddle静态库链接到解释器里。
+    * 解释性语言实际运行的二进制是解释器本身，如果调用静态库只能将静态库与解释器链接。例如对于Java来说，便是将静态库加入JVM中。这对于通常的Java的开发者来说，是不常见的做法。
+
+### 动态库中不嵌入任何其他语言的解释器
+
+* 目前Paddle的进程模型是C++内部驱动Python解释器进行模型配置解析和数据读取
+* 我们最终的动态库中不嵌入Python或者其他任何语言的解释器。模型配置解析，数据读取均交由其他语言完成
+
+现阶段Paddle有一个问题是，Paddle内嵌的Python解释器和外部使用的Python如果版本不同，会直接报错退出。
+
+### Paddle动态库中，不引用其他动态库
+
+* 即这个动态库是不依赖于其他任何文件的，可以在任何机器上执行的。
+
+###  这个动态库使用C99标准的头文件导出一些函数，不使用/导出C++符号
+
+* 由于C++编译器没有[名字修饰](https://en.wikipedia.org/wiki/Name_mangling#C.2B.2B)的规范，不同版本的编译器之间，对于同一段C++代码生成的符号可能不一致。而多语言接口需要直接读取生成的二进制(动态库)，需要有稳定的导出符号。
+* C语言是有导出符号的标准的，并且在常见的平台上，都是ABI调用标准的。
+* 大多数语言都支持使用C语言API
+* 使用C99而不使用C89，是因为C99支持[Fixed-width integer types](https://en.wikipedia.org/wiki/C_data_types#Fixed-width_integer_types)和[Boolean type](https://en.wikipedia.org/wiki/C_data_types#Boolean_type)。
+* 使用C99而不使用C11的原因是，[C11](https://en.wikipedia.org/wiki/C11_(C_standard_revision))并没有Paddle特别需要的特性，且C99相对于C11使用更加广泛。
+
+### 不导出Paddle内部的结构体、类，仅仅使用`void*`指针作为类型的句柄(handler)
+
+* Paddle内部的类为C++书写，直接导出到C的接口比较困难。
+* 在C-API中使用`void*`来表示Paddle内部类。再在每一个API中自己检查类型。
+
+在C的头文件 `paddle_matrix.h` 中:
+
+```C
+typedef void* paddle_matrix;
+typedef int paddle_error;
+
+extern "C"
+paddle_error paddle_matrix_shape(paddle_matrix matrix,
+                                 uint64_t* width,
+                                 uint64_t* height);
+```
+而在CPP里面实现这个C的接口，文件 `paddle_matrix.cpp`
+
+```cpp
+#include "paddle/math/matrix.hpp"
+extern "C"
+paddle_error paddle_matrix_shape(paddle_matrix matrix,
+                                 uint64_t *width,
+                                 uint64_t *height) {
+  auto m = (paddle::math::matrix*)(matrix);
+  *width = m->width();
+  *height = m->height();
+}
+```
+
+其中`paddle/math/matrix.hpp`文件内容为:
+
+```cpp
+namespace paddle {
+namespace math {  
+
+class Matrix {
+  //...
+};
+
+}  // namespace math
+}  // namespace paddle
+```
+
+### 不使用SWIG这种代码生成器，而是手写多语言绑定
+
+* [SWIG](http://www.swig.org/)是一个多语言接口的代码生成器。他的目标是使用C/C++写代码，SWIG直接读取C/C++的头文件，生成各种语言的绑定代码。
+    * 对于多语言接口，SWIG需要写一个interface文件。这个文件具有独特的语法，学习成本高。且增加一个第三方语言，就需要对这个第三方语言增加一些定义。有的时候，interface文件的写法非常[tricky](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/api/Paddle.swig#L36)。社区贡献代码学习成本高。
+    * SWIG暴露的接口保留了C++的接口样式，很难保证多语言代码风格的一致性。(函数命名，错误处理)
+        * 因为SWIG在第三方语言中暴露的函数名，类名和C++中完全一致。C++的命名风格并不能适应其他第三方语言。如果使用SWIG我们需要将在interface文件里，将大量的`SomeCppClass`重命名成`some_python_class`，或者`SomeGoTypes`。
+        * 对于不同语言，错误处理的方式也不尽相同。例如对于Java或者Python，最常见的错误处理方式是Exception，而对于Golang，错误处理方式是返回值。而SWIG只能简单的暴露C++接口，无法做到对于各种语言错误处理方式的适配。
+    * 对于大多数语言，直接使用C语言的.h并不困难。例如Python的[cffi](https://cffi.readthedocs.io/en/latest/overview.html#simple-example-abi-level-in-line)或者[Cython](http://cython.org/), golang的[cgo](https://golang.org/cmd/cgo/)。
+    * SWIG支持的语言或者解释器有局限。例如对于Python，使用SWIG只支持CPython解释器，而不支持PyPy解释器。
+
+
+## 原因列表
+
+| 结论 | 对比 | 原因 |
+|---| --- | --- |
+| 使用动态库 | 不使用静态库 | 解释型语言只能调用动态库，Paddle静态库链接复杂 |
+| 不嵌入其他语言解释器 | 不嵌入Python解释器 | Paddle C++目前嵌入Python解释器，会导致不同版本Python在一个进程里的bug |
+| 不引用其他动态库 | | Paddle一个动态库可以在任何Linux系统上运行 |
+| 使用C99做接口 | 不使用C++做接口 | C有标准的ABI，C99是目前C最广泛的使用标准，且C99支持bool类型和定长整数(uint64_t等)类型 |
+| 使用void*作为类句柄 | 不显示的写每个类具体包含什么| 实现简单，并且让接口脱离实现细节 |
+| 手写多语言绑定 | 不使用SWIG | 使用SWIG需要多语言绑定的开发人员熟练掌握SWIG配置，社区参与困难。SWIG生成的代码不能保证多语言代码风格的一致性 |
+
+
+## 简单实现
+
+TBD

doc/faq/index_cn.rst

@@ -286,3 +286,16 @@ PaddlePaddle的参数使用名字 :code:`name` 作为参数的ID，相同名字
 ..      code-block:: bash
 
         paddle train --use_gpu=true --trainer_count=2 --gpu_id=2
+
+
+12. 训练过程中出现 :code:`Floating point exception`, 训练因此退出怎么办?
+------------------------------------------------------------------------
+
+Paddle二进制在运行时捕获了浮点数异常，只要出现浮点数异常(即训练过程中出现NaN或者Inf)，立刻退出。浮点异常通常的原因是浮点数溢出、除零等问题。
+主要原因包括两个方面:
+
+* 训练过程中参数或者训练过程中的梯度尺度过大，导致参数累加，乘除等时候，导致了浮点数溢出。
+* 模型一直不收敛，发散到了一个数值特别大的地方。
+* 训练数据有问题，导致参数收敛到了一些奇异的情况。或者输入数据尺度过大，有些特征的取值达到数百万，这时进行矩阵乘法运算就可能导致浮点数溢出。
+
+主要的解决办法是减小学习律或者对数据进行归一化处理。

doc/getstarted/basic_usage/index_cn.rst

@@ -55,7 +55,7 @@ PaddlePaddle是源于百度的一个深度学习平台。这份简短的介绍
     # 线性计算网络层: ȳ = wx + b
     ȳ = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b'))
     # 计算误差函数，即  ȳ 和真实 y 之间的距离
-    cost = regression_cost(input= ȳ, label=y)
+    cost = mse_cost(input= ȳ, label=y)
     outputs(cost)
 
 
@@ -69,7 +69,7 @@ PaddlePaddle是源于百度的一个深度学习平台。这份简短的介绍
     
     - **数据层**：数据层 `data_layer` 是神经网络的入口，它读入数据并将它们传输到接下来的网络层。这里数据层有两个，分别对应于变量 `x` 和 `y`。
     - **全连接层**：全连接层 `fc_layer` 是基础的计算单元，这里利用它建模变量之间的线性关系。计算单元是神经网络的核心，PaddlePaddle支持大量的计算单元和任意深度的网络连接，从而可以拟合任意的函数来学习复杂的数据关系。
-    - **回归误差代价层**：回归误差代价层 `regression_cost` 是众多误差代价函数层的一种，它们在训练过程作为网络的出口，用来计算模型的误差，是模型参数优化的目标函数。
+    - **回归误差代价层**：回归误差代价层 `mse_cost` 是众多误差代价函数层的一种，它们在训练过程作为网络的出口，用来计算模型的误差，是模型参数优化的目标函数。
 
 定义了网络结构并保存为 `trainer_config.py` 之后，运行以下训练命令：
 

doc/getstarted/basic_usage/index_en.rst

@@ -49,7 +49,7 @@ To recover this relationship between ``X`` and ``Y``, we use a neural network wi
         x = data_layer(name='x', size=1)
         y = data_layer(name='y', size=1)
         y_predict = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b'))
-        cost = regression_cost(input=y_predict, label=y)
+        cost = mse_cost(input=y_predict, label=y)
         outputs(cost)
 
 Some of the most fundamental usages of PaddlePaddle are demonstrated:

doc/getstarted/build_and_install/build_from_source_en.md

@@ -51,7 +51,7 @@ PaddlePaddle supports some build options.
 <tr><td class="left">WITH_TIMER</td><td class="left">Compile PaddlePaddle with stats timer</td></tr>
 <tr><td class="left">WITH_PROFILER</td><td class="left">Compile PaddlePaddle with GPU profiler</td></tr>
 <tr><td class="left">WITH_DOC</td><td class="left">Compile PaddlePaddle with documentation</td></tr>
-<tr><td class="left">ON_COVERALLS</td><td class="left">Compile PaddlePaddle with code coverage</td></tr>
+<tr><td class="left">WITH_COVERAGE</td><td class="left">Compile PaddlePaddle with code coverage</td></tr>
 <tr><td class="left">COVERALLS_UPLOAD</td><td class="left">Package code coverage data to coveralls</td></tr>
 <tr><td class="left">ON_TRAVIS</td><td class="left">Exclude special unit test on Travis CI</td></tr>
 </tbody>

doc/getstarted/build_and_install/docker_install_cn.rst

@@ -4,138 +4,137 @@ PaddlePaddle的Docker容器使用方式
 PaddlePaddle目前唯一官方支持的运行的方式是Docker容器。因为Docker能在所有主要操作系统（包括Linux，Mac OS X和Windows）上运行。 请注意，您需要更改 `Dockers设置 <https://github.com/PaddlePaddle/Paddle/issues/627>`_ 才能充分利用Mac OS X和Windows上的硬件资源。
 
 
-通过Docker容器开发PaddlePaddle
+纯CPU和GPU的docker镜像使用说明
 ------------------------------
 
-开发人员可以在Docker中开发PaddlePaddle。这样开发人员可以以一致的方式在不同的平台上工作 - Linux，Mac OS X和Windows。
+对于每一个PaddlePaddle版本，我们都会发布两个Docker镜像：纯CPU的和GPU的。
+我们通过设置 `dockerhub.com <https://hub.docker.com/r/paddledev/paddle/>`_ 自动生成最新的docker镜像：
+`paddledev/paddle:0.10.0rc1-cpu` 和 `paddledev/paddle:0.10.0rc1-gpu`。
 
-1. 将开发环境构建为Docker镜像
+以交互容器方式运行纯CPU的镜像：
 
-   .. code-block:: bash
+.. code-block:: bash
 
-      git clone --recursive https://github.com/PaddlePaddle/Paddle
-      cd Paddle
-      docker build -t paddle:dev -f paddle/scripts/docker/Dockerfile .
+    docker run -it --rm paddledev/paddle:0.10.0rc1-cpu /bin/bash
 
+或者，可以以后台进程方式运行容器：
 
-   请注意，默认情况下，:code:`docker build` 不会将源码导入到镜像中并编译它。如果我们想这样做，需要设置一个参数：
+.. code-block:: bash
 
-   .. code-block:: bash
-
-      docker build -t paddle:dev -f paddle/scripts/docker/Dockerfile --build-arg BUILD_AND_INSTALL=ON .
+    docker run -d -p 2202:22 -p 8888:8888 paddledev/paddle:0.10.0rc1-cpu
 
+然后用密码 :code:`root` SSH进入容器：
 
-2. 运行开发环境
+.. code-block:: bash
 
-   当我们编译好了 :code:`paddle:dev`， 我们可以在docker容器里做开发，源代码可以通过挂载本地文件来被载入Docker的开发环境里面：
+    ssh -p 2202 root@localhost
 
-   .. code-block:: bash
+SSH方式的一个优点是我们可以从多个终端进入容器。比如，一个终端运行vi，另一个终端运行Python。另一个好处是我们可以把PaddlePaddle容器运行在远程服务器上，并在笔记本上通过SSH与其连接。
 
-      docker run -d -p 2202:22 -v $PWD:/paddle paddle:dev
 
-   以上代码会启动一个带有PaddlePaddle开发环境的docker容器，源代码会被挂载到 :code:`/paddle` 。
+以上方法在GPU镜像里也能用－只是请不要忘记按装CUDA驱动，以及告诉Docker：
 
-   请注意， :code:`paddle:dev` 的默认入口是 :code:`sshd` 。以上的 :code:`docker run` 命令其实会启动一个在2202端口监听的SSHD服务器。这样，我们就能SSH进入我们的开发容器了：
+.. code-block:: bash
 
-   .. code-block:: bash
+    export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
+    export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
+    docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:0.10.0rc1-gpu
 
-      ssh root@localhost -p 2202
 
-3. 在Docker开发环境中编译与安装PaddlPaddle代码
+运行PaddlePaddle书籍
+---------------------
 
-   当在容器里面的时候，可以用脚本 :code:`paddle/scripts/docker/build.sh` 来编译、安装与测试PaddlePaddle：
+Jupyter Notebook是一个开源的web程序，大家可以通过它制作和分享带有代码、公式、图表、文字的交互式文档。用户可以通过网页浏览文档。
 
-   .. code-block:: bash
+PaddlePaddle书籍是为用户和开发者制作的一个交互式的Jupyter Nodebook。
+如果您想要更深入了解deep learning，PaddlePaddle书籍一定是您最好的选择。
 
-      /paddle/paddle/scripts/docker/build.sh
+当您进入容器内之后，只用运行以下命令：
 
-   以上指令会在 :code:`/paddle/build` 中编译PaddlePaddle。通过以下指令可以运行单元测试：
+.. code-block:: bash
         
-   .. code-block:: bash
+    jupyter notebook
 
-      cd /paddle/build
-      ctest
+然后在浏览器中输入以下网址：
     
-4. 在Docker容器中运行PaddlePaddle书籍
+.. code-block:: text
 
-   Jupyter Notebook是一个开源的web程序，大家可以通过它制作和分享带有代码、公式、图表、文字的交互式文档。用户可以通过网页浏览文档。
+    http://localhost:8888/
 
-   PaddlePaddle书籍是为用户和开发者制作的一个交互式的Jupyter Nodebook。
-   如果您想要更深入了解deep learning，PaddlePaddle书籍一定是您最好的选择。
+就这么简单，享受您的旅程！
 
-   当您进入容器内之后，只用运行以下命令：
 
-   .. code-block:: bash
+非AVX镜像
+---------
 
-      jupyter notebook
+纯CPU镜像以及GPU镜像都会用到AVX指令集，但是2008年之前生产的旧电脑不支持AVX。以下指令能检查Linux电脑是否支持AVX：
 
-   然后在浏览器中输入以下网址：
+.. code-block:: bash
 
-   .. code-block:: text
+   if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi
 
-      http://localhost:8888/
+如果输出是No，我们就需要手动编译一个非AVX版本的镜像：
 
-   就这么简单，享受您的旅程！
+.. code-block:: bash
 
-纯CPU和GPU的docker镜像
-----------------------
+   cd ~
+   git clone https://github.com/PaddlePaddle/Paddle.git
+   cd Paddle
+   docker build --build-arg WITH_AVX=OFF -t paddle:cpu-noavx -f paddle/scripts/docker/Dockerfile .
+   docker build --build-arg WITH_AVX=OFF -t paddle:gpu-noavx -f paddle/scripts/docker/Dockerfile.gpu .
 
-对于每一个PaddlePaddle版本，我们都会发布两个Docker镜像：纯CPU的和GPU的。我们通过设置 `dockerhub.com <https://hub.docker.com/r/paddledev/paddle/>`_ 自动运行以下两个命令：
 
-.. code-block:: bash
+通过Docker容器开发PaddlePaddle
+------------------------------
 
-   docker build -t paddle:cpu -f paddle/scripts/docker/Dockerfile --build-arg BUILD_AND_INSTALL=ON .
-   docker build -t paddle:gpu -f paddle/scripts/docker/Dockerfile.gpu --build-arg BUILD_AND_INSTALL=ON .
+开发人员可以在Docker中开发PaddlePaddle。这样开发人员可以以一致的方式在不同的平台上工作 - Linux，Mac OS X和Windows。
 
-以交互容器方式运行纯CPU的镜像：
+1. 将开发环境构建为Docker镜像
    
-.. code-block:: bash
+   .. code-block:: bash
 
-    docker run -it --rm paddledev/paddle:0.10.0rc1-cpu /bin/bash
+      git clone --recursive https://github.com/PaddlePaddle/Paddle
+      cd Paddle
+      docker build -t paddle:dev -f paddle/scripts/docker/Dockerfile .
 
-或者，可以以后台进程方式运行容器：
 
-.. code-block:: bash
+   请注意，默认情况下，:code:`docker build` 不会将源码导入到镜像中并编译它。如果我们想这样做，需要设置一个参数：
 
-    docker run -d -p 2202:22 paddledev/paddle:0.10.0rc1-cpu
+   .. code-block:: bash
 
-然后用密码 :code:`root` SSH进入容器：
+      docker build -t paddle:dev -f paddle/scripts/docker/Dockerfile --build-arg BUILD_AND_INSTALL=ON .
 
-.. code-block:: bash
 
-    ssh -p 2202 root@localhost
+2. 运行开发环境
 
-SSH方式的一个优点是我们可以从多个终端进入容器。比如，一个终端运行vi，另一个终端运行Python。另一个好处是我们可以把PaddlePaddle容器运行在远程服务器上，并在笔记本上通过SSH与其连接。
+   当我们编译好了 :code:`paddle:dev`， 我们可以在docker容器里做开发，源代码可以通过挂载本地文件来被载入Docker的开发环境里面：
    
+   .. code-block:: bash
 
-以上方法在GPU镜像里也能用－只是请不要忘记按装CUDA驱动，以及告诉Docker：
+      docker run -d -p 2202:22 -v $PWD:/paddle paddle:dev
 
-.. code-block:: bash
+   以上代码会启动一个带有PaddlePaddle开发环境的docker容器，源代码会被挂载到 :code:`/paddle` 。
 
-    export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
-    export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
-    docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:0.10.0rc1-gpu
+   请注意， :code:`paddle:dev` 的默认入口是 :code:`sshd` 。以上的 :code:`docker run` 命令其实会启动一个在2202端口监听的SSHD服务器。这样，我们就能SSH进入我们的开发容器了：
    
+   .. code-block:: bash
 
-非AVX镜像
----------
+      ssh root@localhost -p 2202
 
-纯CPU镜像以及GPU镜像都会用到AVX指令集，但是2008年之前生产的旧电脑不支持AVX。以下指令能检查Linux电脑是否支持AVX：
+3. 在Docker开发环境中编译与安装PaddlPaddle代码
 
+   当在容器里面的时候，可以用脚本 :code:`paddle/scripts/docker/build.sh` 来编译、安装与测试PaddlePaddle：
    
-.. code-block:: bash
+   .. code-block:: bash
 		      
-   if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi
+      /paddle/paddle/scripts/docker/build.sh
 
-如果输出是No，我们就需要手动编译一个非AVX版本的镜像：
+   以上指令会在 :code:`/paddle/build` 中编译PaddlePaddle。通过以下指令可以运行单元测试：
    
-.. code-block:: bash
+   .. code-block:: bash
 
-   cd ~
-   git clone https://github.com/PaddlePaddle/Paddle.git
-   cd Paddle
-   docker build --build-arg WITH_AVX=OFF -t paddle:cpu-noavx -f paddle/scripts/docker/Dockerfile .
-   docker build --build-arg WITH_AVX=OFF -t paddle:gpu-noavx -f paddle/scripts/docker/Dockerfile.gpu .
+      cd /paddle/build
+      ctest
 
 
 文档

doc/getstarted/build_and_install/docker_install_en.rst

@@ -9,6 +9,100 @@ Please be aware that you will need to change `Dockers settings
 of your hardware resource on Mac OS X and Windows.
 
 
+Usage of CPU-only and GPU Images
+----------------------------------
+
+For each version of PaddlePaddle, we release 2 Docker images, a
+CPU-only one and a CUDA GPU one.  We do so by configuring
+`dockerhub.com <https://hub.docker.com/r/paddledev/paddle/>`_
+automatically generate the latest docker images `paddledev/paddle:0.10.0rc1-cpu`
+and `paddledev/paddle:0.10.0rc1-gpu`.
+
+To run the CPU-only image as an interactive container:
+
+.. code-block:: bash
+
+    docker run -it --rm paddledev/paddle:0.10.0rc1-cpu /bin/bash
+
+or, we can run it as a daemon container
+
+.. code-block:: bash
+
+    docker run -d -p 2202:22 -p 8888:8888 paddledev/paddle:0.10.0rc1-cpu
+
+and SSH to this container using password :code:`root`:
+
+.. code-block:: bash
+
+    ssh -p 2202 root@localhost
+
+An advantage of using SSH is that we can connect to PaddlePaddle from
+more than one terminals.  For example, one terminal running vi and
+another one running Python interpreter.  Another advantage is that we
+can run the PaddlePaddle container on a remote server and SSH to it
+from a laptop.
+
+Above methods work with the GPU image too -- just please don't forget
+to install CUDA driver and let Docker knows about it:
+
+.. code-block:: bash
+
+    export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
+    export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
+    docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:0.10.0rc1-gpu
+
+
+PaddlePaddle Book
+------------------
+
+The Jupyter Notebook is an open-source web application that allows
+you to create and share documents that contain live code, equations,
+visualizations and explanatory text in a single browser.
+
+PaddlePaddle Book is an interactive Jupyter Notebook for users and developers. 
+We already exposed port 8888 for this book. If you want to
+dig deeper into deep learning, PaddlePaddle Book definitely is your best choice.
+
+Once you are inside the container, simply issue the command:
+
+.. code-block:: bash
+        
+    jupyter notebook
+
+Then, you would back and paste the address into the local browser:
+    
+.. code-block:: text
+
+    http://localhost:8888/
+
+That's all. Enjoy your journey!
+
+
+Non-AVX Images
+--------------
+
+Please be aware that the CPU-only and the GPU images both use the AVX
+instruction set, but old computers produced before 2008 do not support
+AVX.  The following command checks if your Linux computer supports
+AVX:
+
+.. code-block:: bash
+
+   if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi
+
+
+If it doesn't, we will need to build non-AVX images manually from
+source code:
+
+.. code-block:: bash
+
+   cd ~
+   git clone https://github.com/PaddlePaddle/Paddle.git
+   cd Paddle
+   docker build --build-arg WITH_AVX=OFF -t paddle:cpu-noavx -f paddle/scripts/docker/Dockerfile .
+   docker build --build-arg WITH_AVX=OFF -t paddle:gpu-noavx -f paddle/scripts/docker/Dockerfile.gpu .
+
+
 Development Using Docker
 ------------------------
 
@@ -82,103 +176,6 @@ Windows -- in a consistent way.
       cd /paddle/build
       ctest
 
-4. Run PaddlePaddle Book under Docker Container
-
-   The Jupyter Notebook is an open-source web application that allows
-   you to create and share documents that contain live code, equations,
-   visualizations and explanatory text in a single browser.
-
-   PaddlePaddle Book is an interactive Jupyter Notebook for users and developers. 
-   We already exposed port 8888 for this book. If you want to
-   dig deeper into deep learning, PaddlePaddle Book definitely is your best choice.
-
-   Once you are inside the container, simply issue the command:
-
-   .. code-block:: bash
-		   
-      jupyter notebook
-
-   Then, you would back and paste the address into the local browser:
-      
-   .. code-block:: text
-
-      http://localhost:8888/
-
-   That's all. Enjoy your journey!
-
-CPU-only and GPU Images
------------------------
-
-For each version of PaddlePaddle, we release 2 Docker images, a
-CPU-only one and a CUDA GPU one.  We do so by configuring
-`dockerhub.com <https://hub.docker.com/r/paddledev/paddle/>`_
-automatically runs the following commands:
-
-.. code-block:: bash
-
-   docker build -t paddle:cpu -f paddle/scripts/docker/Dockerfile --build-arg BUILD_AND_INSTALL=ON .
-   docker build -t paddle:gpu -f paddle/scripts/docker/Dockerfile.gpu --build-arg BUILD_AND_INSTALL=ON .
-
-
-To run the CPU-only image as an interactive container:
-
-.. code-block:: bash
-
-    docker run -it --rm paddledev/paddle:0.10.0rc1-cpu /bin/bash
-
-or, we can run it as a daemon container
-
-.. code-block:: bash
-
-    docker run -d -p 2202:22 paddledev/paddle:0.10.0rc1-cpu
-
-and SSH to this container using password :code:`root`:
-
-.. code-block:: bash
-
-    ssh -p 2202 root@localhost
-
-An advantage of using SSH is that we can connect to PaddlePaddle from
-more than one terminals.  For example, one terminal running vi and
-another one running Python interpreter.  Another advantage is that we
-can run the PaddlePaddle container on a remote server and SSH to it
-from a laptop.
-
-
-Above methods work with the GPU image too -- just please don't forget
-to install CUDA driver and let Docker knows about it:
-
-.. code-block:: bash
-
-    export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
-    export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
-    docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:0.10.0rc1-gpu
-
-
-Non-AVX Images
---------------
-
-Please be aware that the CPU-only and the GPU images both use the AVX
-instruction set, but old computers produced before 2008 do not support
-AVX.  The following command checks if your Linux computer supports
-AVX:
-
-.. code-block:: bash
-
-   if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi
-
-
-If it doesn't, we will need to build non-AVX images manually from
-source code:
-
-.. code-block:: bash
-
-   cd ~
-   git clone https://github.com/PaddlePaddle/Paddle.git
-   cd Paddle
-   docker build --build-arg WITH_AVX=OFF -t paddle:cpu-noavx -f paddle/scripts/docker/Dockerfile .
-   docker build --build-arg WITH_AVX=OFF -t paddle:gpu-noavx -f paddle/scripts/docker/Dockerfile.gpu .
-
 
 Documentation
 -------------

doc/howto/usage/cmd_parameter/arguments_cn.md

@@ -228,16 +228,6 @@
 <td class="left"></td><td class="left"></td><td class="left">√</td><td class="left">√</td>
 </tr>
 
-<tr>
-<td class="left" rowspan = "2">度量学习(metric learning)</td><td class="left">external</td>
-<td class="left">√</td><td class="left">√</td><td class="left">√</td><td class="left">√</td>
-</tr>
-
-<tr>
-<td class="left">data_server_port</td>
-<td class="left"></td><td class="left">√</td><td class="left"></td><td class="left">√</td>
-</tr>
-
 <tr>
 <td class="left" rowspan = "16">参数服务器(PServer)</td><td class="left">start_pserver</td>
 <td class="left"></td><td class="left">√</td><td class="left"></td><td class="left">√</td>

doc/howto/usage/cmd_parameter/arguments_en.md

@@ -228,16 +228,6 @@ It looks like there are a lot of arguments. However, most of them are for develo
 <td class="left"></td><td class="left"></td><td class="left">√</td><td class="left">√</td>
 </tr>
 
-<tr>
-<td class="left" rowspan = "2">metric learning</td><td class="left">external</td>
-<td class="left">√</td><td class="left">√</td><td class="left">√</td><td class="left">√</td>
-</tr>
-
-<tr>
-<td class="left">data_server_port</td>
-<td class="left"></td><td class="left">√</td><td class="left"></td><td class="left">√</td>
-</tr>
-
 <tr>
 <td class="left" rowspan = "16">PServer</td><td class="left">start_pserver</td>
 <td class="left"></td><td class="left">√</td><td class="left"></td><td class="left">√</td>

doc/howto/usage/cmd_parameter/detail_introduction_cn.md

@@ -180,15 +180,6 @@
   - 用户可以自定义beam search的方法，编译成动态库，供PaddlePaddle加载。 该参数用于指定动态库路径.
   - 类型: string (默认: "", null).
 
-## 度量学习(Metric Learning)
-* `--external`
-   - 指示是否使用外部机器进行度量学习.
-   - 类型: bool (默认: 0).
-
-* `--data_server_port`
-  - 数据服务器(data server)的监听端口，主要用在度量学习中.
-  - 类型: int32 (默认: 21134).
-
 ## 数据支持(DataProvider)
 
 * `--memory_threshold_on_load_data`

doc/howto/usage/cmd_parameter/detail_introduction_en.md

@@ -184,15 +184,6 @@
   - Specify shared dynamic library. It can be defined out of paddle by user.
   - type: string (default: "", null).
 
-## Metric Learning
-* `--external`
-   - Whether to use external machine for metric learning.
-   - type: bool (default: 0).
-
-* `--data_server_port`
-  - Listening port for dserver (data server), dserver is mainly used in metric learning.
-  - type: int32 (default: 21134).
-
 ## DataProvider
 
 * `--memory_threshold_on_load_data`

doc/howto/usage/k8s/k8s_distributed_cn.md

@@ -213,7 +213,7 @@ I1116 09:10:17.123440    50 Util.cpp:130] Calling runInitFunctions
 I1116 09:10:17.123764    50 Util.cpp:143] Call runInitFunctions done.
 [WARNING 2016-11-16 09:10:17,227 default_decorators.py:40] please use keyword arguments in paddle config.
 [INFO 2016-11-16 09:10:17,239 networks.py:1282] The input order is [movie_id, title, genres, user_id, gender, age, occupation, rating]
-[INFO 2016-11-16 09:10:17,239 networks.py:1289] The output order is [__regression_cost_0__]
+[INFO 2016-11-16 09:10:17,239 networks.py:1289] The output order is [__mse_cost_0__]
 I1116 09:10:17.392917    50 Trainer.cpp:170] trainer mode: Normal
 I1116 09:10:17.613910    50 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process
 I1116 09:10:17.680917    50 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process

paddle/api/CMakeLists.txt

-FUNCTION(generate_python_api target_name)
-    ADD_CUSTOM_COMMAND(OUTPUT ${PROJ_ROOT}/paddle/py_paddle/swig_paddle.py
-                              ${PROJ_ROOT}/paddle/Paddle_wrap.cxx
-                              ${PROJ_ROOT}/paddle/Paddle_wrap.h
-        COMMAND ${SWIG_EXECUTABLE} -python -c++ -outcurrentdir -I../ api/Paddle.swig
-                && mv ${PROJ_ROOT}/paddle/swig_paddle.py ${PROJ_ROOT}/paddle/py_paddle/swig_paddle.py
-        DEPENDS ${PROJ_ROOT}/paddle/api/Paddle.swig
-                ${PROJ_ROOT}/paddle/api/PaddleAPI.h
-                ${external_project_dependencies}
-        WORKING_DIRECTORY ${PROJ_ROOT}/paddle
-        COMMENT "Generate Python API from swig")
-    ADD_CUSTOM_TARGET(${target_name} ALL DEPENDS
-                ${PROJ_ROOT}/paddle/Paddle_wrap.cxx
-                ${PROJ_ROOT}/paddle/Paddle_wrap.h
-                ${PROJ_ROOT}/paddle/py_paddle/swig_paddle.py
-                ${external_project_dependencies})
-ENDFUNCTION(generate_python_api)
-
 set(API_SOURCES
     Arguments.cpp
     ConfigParser.cpp
@@ -33,65 +15,86 @@ set(API_HEADER
     PaddleAPI.h
     Internal.h)
 
-add_library(paddle_api STATIC
-        ${API_SOURCES})
+add_library(paddle_api STATIC ${API_SOURCES})
 add_dependencies(paddle_api gen_proto_cpp)
 
-list(LENGTH "${GFLAGS_LIBRARIES}" GFLAGS_LIBRARIES_LENGTH)
-
-if(${GFLAGS_LIBRARIES_LENGTH} EQUAL 0 AND TARGET "${GFLAGS_LIBRARIES}")
-# Because gflags compiled by cmake, so it is imported by cmake target,
-# not a real library path. Get the real library path here.
-message(STATUS "GFLAGS Libraries is ${GFLAGS_LIBRARIES}")
-get_target_property(GFLAGS_LOCATION ${GFLAGS_LIBRARIES} LOCATION)
-message(STATUS "GFLAGS Target location is ${GFLAGS_LOCATION}")
-else()
-set(GFLAGS_LOCATION ${GFLAGS_LIBRARIES})
-endif()
+INCLUDE(${SWIG_USE_FILE})
+INCLUDE_DIRECTORIES(${PROJ_ROOT}/paddle)
 
-configure_file(
-    paddle_api_config.py.in
-    ${PROJ_ROOT}/paddle/api/paddle_api_config.py
-)
+FILE(GLOB PY_PADDLE_PYTHON_FILES ${PROJ_ROOT}/paddle/py_paddle/*.py)
 
-generate_python_api(python_swig_sources)
+SET_SOURCE_FILES_PROPERTIES(Paddle.i PROPERTIES CPLUSPLUS ON)
 
-file(GLOB PY_PADDLE_PYTHON_FILES ${PROJ_ROOT}/paddle/py_paddle/*.py)
+SET(CMAKE_SWIG_OUTDIR ${CMAKE_CURRENT_BINARY_DIR})
+SET(CMAKE_CXX_FLAGS "-std=c++11 -fPIC -Wall")
+IF(WITH_COVERAGE)
+    SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -O0 -fprofile-arcs -ftest-coverage")
+ENDIF(WITH_COVERAGE)
 
-# TODO(yuyang18) : make wheel name calculated by cmake
-add_custom_command(OUTPUT ${PROJ_ROOT}/paddle/dist/.timestamp
-    COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
-    COMMAND ${CMAKE_COMMAND} -E touch dist/.timestamp
-    COMMAND rm -rf py_paddle.egg-info build
-    WORKING_DIRECTORY ${PROJ_ROOT}/paddle
-    DEPENDS python_swig_sources
+SET(SWIG_MODULE_swig_paddle_EXTRA_DEPS
     paddle_parameter
     paddle_function
     paddle_math
     paddle_utils
     paddle_gserver
     paddle_pserver
-            paddle_trainer
     paddle_api
     paddle_cuda
-        ${PY_PADDLE_PYTHON_FILES}
+    paddle_trainer_lib
+    paddle_network
+    paddle_proto
+    ${external_project_dependencies}
 )
 
-install(DIRECTORY ${PROJ_ROOT}/paddle/dist/
-    DESTINATION opt/paddle/share/wheels
-)
+IF(APPLE)
+    SET(MACOS_LD_FLAGS "-undefined dynamic_lookup -Wl,-all_load")
+ELSE(APPLE)
+    SET(START_GROUP "-Xlinker -start-group")
+    SET(END_GROUP "-Xlinker -end-group")
+    SET(ARCHIVE_START "-Wl,--whole-archive")
+    SET(ARCHIVE_END "-Wl,--no-whole-archive")
+ENDIF(APPLE)
 
-add_custom_target(python_api_wheel ALL DEPENDS
-  ${PROJ_ROOT}/paddle/dist/.timestamp)
-add_dependencies(python_api_wheel python_swig_sources
+SWIG_ADD_MODULE(swig_paddle python Paddle.i)
+SWIG_LINK_LIBRARIES(swig_paddle
+    ${MACOS_LD_FLAGS}
+    ${START_GROUP}
+    ${ARCHIVE_START}
+    paddle_gserver
+    paddle_function
+    ${METRIC_LIBS}
+    ${ARCHIVE_END}
+    paddle_pserver
+    paddle_trainer_lib
+    paddle_network
     paddle_parameter
     paddle_math
     paddle_utils
-  paddle_gserver
-  paddle_pserver
-  paddle_trainer
+    paddle_proto
+    paddle_cuda
     paddle_api
-  paddle_cuda)
+    ${CMAKE_DL_LIBS}
+    ${EXTERNAL_LIBS}
+    ${CMAKE_THREAD_LIBS_INIT}
+    ${RDMA_LD_FLAGS}
+    ${RDMA_LIBS}
+    ${START_END}
+)
+
+add_custom_command(OUTPUT ${PROJ_ROOT}/paddle/py_paddle/_swig_paddle.so
+    COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/swig_paddle.py ${PROJ_ROOT}/paddle/py_paddle
+    COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/_swig_paddle.so ${PROJ_ROOT}/paddle/py_paddle
+    COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
+    COMMAND ${CMAKE_COMMAND} -E touch dist/.timestamp
+    COMMAND rm -rf py_paddle.egg-info build
+    WORKING_DIRECTORY ${PROJ_ROOT}/paddle
+    DEPENDS _swig_paddle
+)
+
+# TODO(yuyang18) : make wheel name calculated by cmake
+add_custom_target(python_api_wheel ALL DEPENDS ${PROJ_ROOT}/paddle/py_paddle/_swig_paddle.so)
+
+install(DIRECTORY ${PROJ_ROOT}/paddle/dist/ DESTINATION opt/paddle/share/wheels)
 
 if(WITH_TESTING)
     IF(NOT PY_PIP_FOUND)

paddle/api/paddle_api_config.py.in

-PADDLE_BUILD_DIR="@CMAKE_CURRENT_BINARY_DIR@/../"
-WITH_GPU="@WITH_GPU@"
-PROTOBUF_LIBRARY="@PROTOBUF_LIBRARY@"
-ZLIB_LIBRARIES="@ZLIB_LIBRARIES@"
-CMAKE_THREAD_LIB="@CMAKE_THREAD_LIBS_INIT@"
-CMAKE_DL_LIBS="@CMAKE_DL_LIBS@"
-
-
-WITH_PYTHON="@WITH_PYTHON@"
-PYTHON_LIBRARIES="@PYTHON_LIBRARIES@"
-GLOG_LIBRARIES="@GLOG_LIBRARIES@"
-GFLAGS_LIBRARIES="@GFLAGS_LIBRARIES@"
-GFLAGS_LOCATION="@GFLAGS_LOCATION@"
-CBLAS_LIBRARIES="@CBLAS_LIBRARIES@"
-
-CUDA_LIBRARIES="@CUDA_CUDART_LIBRARY@"
-WITH_COVERALLS="@ON_COVERALLS@"

paddle/api/paddle_ld_flags.py

-# 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.
-
-try:
-    from paddle_api_config import *
-    import os.path
-    import platform
-
-    system = platform.system().lower()
-    is_osx = (system == 'darwin')
-    is_win = (system == 'windows')
-    is_lin = (system == 'linux')
-
-    if is_lin:
-        whole_start = "-Wl,--whole-archive"
-        whole_end = "-Wl,--no-whole-archive"
-    elif is_osx:
-        whole_start = ""
-        whole_end = ""
-
-    LIB_DIRS = [
-        "math", 'function', 'utils', 'parameter', "gserver", "api", "cuda",
-        "pserver", "trainer"
-    ]
-    PARENT_LIB_DIRS = ['proto']
-
-    class PaddleLDFlag(object):
-        def __init__(self):
-            self.paddle_build_dir = PADDLE_BUILD_DIR
-            self.paddle_build_dir = os.path.abspath(self.paddle_build_dir)
-            self.with_gpu = PaddleLDFlag.cmake_bool(WITH_GPU)
-            self.protolib = PROTOBUF_LIBRARY
-            self.zlib = ZLIB_LIBRARIES
-            self.thread = CMAKE_THREAD_LIB
-            self.dl_libs = CMAKE_DL_LIBS
-            self.with_python = PaddleLDFlag.cmake_bool(WITH_PYTHON)
-            self.python_libs = PYTHON_LIBRARIES
-
-            self.glog_libs = GLOG_LIBRARIES
-
-            self.with_coverage = PaddleLDFlag.cmake_bool(WITH_COVERALLS)
-            self.gflags_libs = GFLAGS_LIBRARIES
-            self.gflags_location = GFLAGS_LOCATION
-            self.cblas_libs = CBLAS_LIBRARIES
-            self.curt = CUDA_LIBRARIES
-
-        def ldflag_str(self):
-            return " ".join(
-                [self.libs_dir_str(), self.parent_dir_str(), self.libs_str()])
-
-        def libs_dir_str(self):
-            libdirs = LIB_DIRS
-            return " ".join(
-                map(lambda x: "-L" + os.path.join(self.paddle_build_dir, x),
-                    libdirs))
-
-        def parent_dir_str(self):
-            libdirs = PARENT_LIB_DIRS
-            return " ".join(
-                map(lambda x: "-L" + os.path.join(self.paddle_build_dir, '..', x),
-                    libdirs))
-
-        def libs_str(self):
-            libs = [
-                whole_start,
-                "-lpaddle_gserver",
-                "-lpaddle_function",
-                whole_end,
-                "-lpaddle_pserver",
-                "-lpaddle_trainer_lib",
-                "-lpaddle_network",
-                '-lpaddle_parameter',
-                "-lpaddle_math",
-                '-lpaddle_utils',
-                "-lpaddle_proto",
-                "-lpaddle_cuda",
-                "-lpaddle_api",
-                self.normalize_flag(self.protolib),
-                self.normalize_flag(self.glog_libs),
-                self.normalize_flag(self.gflags_libs),
-                self.normalize_flag(self.zlib),
-                self.normalize_flag(self.thread),
-                self.normalize_flag(self.dl_libs),
-                self.normalize_flag(self.cblas_libs),
-            ]
-
-            if self.with_python:
-                libs.append(self.normalize_flag(self.python_libs))
-            if self.with_gpu:
-                libs.append(self.normalize_flag(self.curt))
-            if self.with_coverage:
-                libs.append("-fprofile-arcs")
-            return " ".join(filter(lambda l: len(l) != 0, libs))
-
-        def normalize_flag(self, cmake_flag):
-            """
-            CMake flag string to ld flag
-            :type cmake_flag: str
-            """
-            if ";" in cmake_flag:
-                return " ".join(map(self.normalize_flag, cmake_flag.split(";")))
-            if cmake_flag.startswith("/"):  # is a path
-                return cmake_flag
-            elif cmake_flag.startswith("-l"):  # normal link command
-                return cmake_flag
-            elif cmake_flag in [
-                    "gflags-shared", "gflags-static", "gflags_nothreads-shared",
-                    "gflags_nothreads-static"
-            ]:  # special for gflags
-                assert PaddleLDFlag.cmake_bool(self.gflags_location)
-                return self.gflags_location
-            elif len(cmake_flag) != 0:
-                return "".join(["-l", cmake_flag])
-            else:
-                return ""
-
-        @staticmethod
-        def cmake_bool(cmake_str):
-            """
-            CMake bool string to bool
-            :param cmake_str: cmake boolean string
-            :type cmake_str: str
-            :rtype: bool
-            """
-            if cmake_str in ["FALSE", "OFF", "NO"] or cmake_str.endswith(
-                    "-NOTFOUND"):
-                return False
-            else:
-                return True
-
-        def c_flag(self):
-            if self.with_coverage:
-                return [
-                    "-fprofile-arcs", "-ftest-coverage", "-O0", "-g",
-                    "-std=c++11"
-                ]
-            else:
-                return ["-std=c++11"]
-except ImportError:
-
-    class PaddleLDFlag(object):
-        def ldflag_str(self):
-            pass
-
-        def c_flag(self):
-            pass

paddle/gserver/CMakeLists.txt

@@ -25,12 +25,16 @@ filter_test(GSERVER_HEADER)
 filter_test(GSERVER_SOURCES)
 if(NOT WITH_GPU)
     list(REMOVE_ITEM GSERVER_HEADER
+        layers/CudnnConvBaseLayer.h
         layers/CudnnConvLayer.h
+        layers/CudnnConvTransLayer.h
         layers/CudnnPoolLayer.h
         layers/CudnnBatchNormLayer.h)
 
     list(REMOVE_ITEM GSERVER_SOURCES
+        layers/CudnnConvBaseLayer.cpp
         layers/CudnnConvLayer.cpp
+        layers/CudnnConvTransLayer.cpp
         layers/CudnnPoolLayer.cpp
         layers/CudnnBatchNormLayer.cpp)
     compile_cu_as_cpp(layers/LstmCompute.cu)

paddle/gserver/dataproviders/DataProvider.h

@@ -164,15 +164,6 @@ public:
     argu.value = value;
     data_.push_back(argu);
   }
-  /**
-   * @brief Append user defined data
-   * @param[in]  ptr     user defined data
-   */
-  void appendUserDefinedPtr(UserDefinedVectorPtr ptr) {
-    Argument argu;
-    argu.udp = ptr;
-    data_.push_back(argu);
-  }
 
   /*
    * @brief Append argument

paddle/gserver/gradientmachines/MultiGradientMachine.cpp

@@ -24,9 +24,6 @@ limitations under the License. */
 DEFINE_bool(allow_only_one_model_on_one_gpu,
             true,
             "If true, do not allow multiple models on one GPU device");
-#ifdef PADDLE_METRIC_LEARNING
-DECLARE_bool(external);
-#endif
 
 namespace paddle {
 
@@ -45,11 +42,7 @@ MultiGradientMachine::MultiGradientMachine(const ModelConfig& config,
       trainerBarrier_(FLAGS_trainer_count),
       allBarrier_(FLAGS_trainer_count + 1),
       inArgsCopied_(false) {
-#ifdef PADDLE_METRIC_LEARNING
-  isPassGrad_ = FLAGS_external;
-#else
   isPassGrad_ = false;
-#endif
   numThreads_ = FLAGS_trainer_count;
   if (useGpu) {
     //! TODO(yuyang18): When useGpu=false && paddle is not compiled with gpu,

paddle/gserver/layers/ConvBaseOperator.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "ConvBaseOperator.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * @brief ConvBaseOperator takes two inputs to perform the convolution.
+ * The first input is the image, and the second input is the convolution kernel.
+ * The height of data for two inputs are the same. Each data of the first input
+ * is convolved with each data of the second input indepedently.
+ *
+ * The config file api is conv_operator.
+ */
+
+ConvBaseOperator::ConvBaseOperator(const OperatorConfig &config, bool useGpu)
+    : Operator(config, useGpu) {
+  CHECK(useGpu);
+  CHECK_EQ(config_.input_indices_size(), 2L);
+
+  caffeMode_ = true;
+  getConvParams();
+  computeConvSizes();
+
+  // initialize all to default algorithms
+  fwdAlgo_ = 0;
+  bwdFilterAlgo_ = 0;
+  bwdDataAlgo_ = 0;
+  fwdLimitBytes_ = 0;
+  bwdDataLimitBytes_ = 0;
+  bwdFilterLimitBytes_ = 0;
+  workSpaceInBytes_ = 0;
+  workSpace_ = nullptr;
+
+  isSelectAlgo_ = false;
+}
+
+void ConvBaseOperator::allocConvWorkSpace() {
+  hl_conv_workspace(imageDesc_,
+                    outputDesc_,
+                    filterDesc_,
+                    convDesc_,
+                    &fwdAlgo_,
+                    &fwdLimitBytes_,
+                    &bwdDataAlgo_,
+                    &bwdDataLimitBytes_,
+                    &bwdFilterAlgo_,
+                    &bwdFilterLimitBytes_);
+
+  size_t maxWorkSpace = 0;
+  maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_);
+  maxWorkSpace = std::max(maxWorkSpace, bwdFilterLimitBytes_);
+
+  if (maxWorkSpace > workSpaceInBytes_) {
+    if (workSpaceInBytes_ != 0) {
+      hl_free_mem_device(workSpace_);
+    }
+    // total amount of storage needed
+    workSpace_ = hl_malloc_device(maxWorkSpace);
+    workSpaceInBytes_ = maxWorkSpace;
+  }
+}
+
+void ConvBaseOperator::computeConvSizes() {
+  hl_create_filter_descriptor(
+      &filterDesc_, channels_, numFilters_, filterSizeY_, filterSize_);
+  hl_create_tensor_descriptor(&imageDesc_);
+  hl_create_tensor_descriptor(&outputDesc_);
+  hl_create_convolution_descriptor(&convDesc_,
+                                   imageDesc_,
+                                   filterDesc_,
+                                   paddingY_,
+                                   padding_,
+                                   strideY_,
+                                   stride_);
+}
+
+void ConvBaseOperator::reshapeImageDescriptors() {
+  hl_tensor_reshape(imageDesc_,
+                    1,
+                    channels_,
+                    imageH_,
+                    imageW_,
+                    channels_ * imageH_ * imageW_,
+                    imageH_ * imageW_,
+                    imageW_,
+                    1);
+  hl_tensor_reshape(outputDesc_,
+                    1,
+                    numFilters_,
+                    outputH_,
+                    outputW_,
+                    numFilters_ * outputH_ * outputW_,
+                    outputH_ * outputW_,
+                    outputW_,
+                    1);
+  hl_reset_convolution_descriptor(convDesc_,
+                                  imageDesc_,
+                                  filterDesc_,
+                                  paddingY_,
+                                  padding_,
+                                  strideY_,
+                                  stride_);
+}
+
+void ConvBaseOperator::getConvParams() {
+  configNumFilters_ = config_.num_filters();
+  const ConvConfig &conf = config_.conv_conf();
+  padding_ = conf.padding();
+  stride_ = conf.stride();
+  filterSize_ = conf.filter_size();
+  paddingY_ = conf.padding_y();
+  strideY_ = conf.stride_y();
+  filterSizeY_ = conf.filter_size_y();
+  filterPixels_ = filterSize_ * filterSizeY_;
+  configChannels_ = conf.channels();
+  imgSize_ = conf.img_size();
+  imgSizeY_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
+  imgPixels_ = imgSize_ * imgSizeY_;
+  CHECK_EQ(conf.groups(), 1U);
+  filterChannels_ = conf.filter_channels();
+  outputX_ = conf.output_x();
+  outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x();
+  outputs_ = outputX_ * outputX_;
+
+  isDeconv_ = (config_.type() == "conv") ? false : true;
+  if (isDeconv_) {
+    channels_ = configNumFilters_;
+    numFilters_ = configChannels_;
+  } else {
+    channels_ = configChannels_;
+    numFilters_ = configNumFilters_;
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/ConvBaseOperator.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+#pragma once
+
+#include "Operator.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * @brief ConvOperator takes two inputs to perform the convolution.
+ * The first input is the image, and the second input is the convolution kernel.
+ * The height of data for two inputs are the same. Each data of the first input
+ * is convolved with each data of the second input indepedently.
+ *
+ * The config file api is conv_operator.
+ */
+
+class ConvBaseOperator : public Operator {
+public:
+  ConvBaseOperator(const OperatorConfig &config, bool useGpu);
+  /**
+   * Free workspace in device and destroy cudnn tensor descriptor.
+   */
+  virtual ~ConvBaseOperator() {
+    if (workSpaceInBytes_ != 0) {
+      hl_free_mem_device(workSpace_);
+      workSpaceInBytes_ = 0;
+    }
+
+    hl_destroy_tensor_descriptor(imageDesc_);
+    hl_destroy_tensor_descriptor(outputDesc_);
+    hl_destroy_filter_descriptor(filterDesc_);
+    hl_destroy_convolution_descriptor(convDesc_);
+  }
+
+protected:
+  /**
+   * Get convolution parameters from layer config and
+   * initialize member variables.
+   */
+  void getConvParams();
+
+  /**
+   * Allocate Gpu Memory for cudnn convolution algorithms.
+   */
+  void allocConvWorkSpace();
+
+  /**
+   * Create cudnn tensor descriptor for convolution operation.
+   */
+  void computeConvSizes();
+
+  /**
+   * Reshape cudnn tensor descriptor.
+   */
+  void reshapeImageDescriptors();
+
+  /**
+   * Reshape cudnn tensor descriptor.
+   */
+  virtual void reshape(int batchSize) = 0;
+
+  /**
+   * Check filter size is equal to the size calculated by parameters from
+   * layer config.
+   */
+  void checkFilterSize(const MatrixPtr &filter) {
+    CHECK_EQ(static_cast<int>(filter->getWidth()),
+             filterSize_ * filterSizeY_ * channels_ * numFilters_);
+  }
+
+  /// Most of member variables are same with CudnnConvLayer.
+  /// There is no explanation here.
+  bool isDeconv_;
+  int imageH_, imageW_, outputH_, outputW_;
+  hl_tensor_descriptor imageDesc_;
+  hl_tensor_descriptor outputDesc_;
+  hl_filter_descriptor filterDesc_;
+  hl_convolution_descriptor convDesc_;
+  bool caffeMode_;
+  int inputOffset_, outputOffset_, weightOffset_;
+  int numFilters_, channels_;
+
+  /// from parsing config
+  int configNumFilters_, configChannels_;
+  int padding_, stride_, filterSize_, imgSize_, imgSizeY_;
+  int paddingY_, strideY_, filterSizeY_;
+  int imgPixels_, filterPixels_, filterChannels_, outputX_, outputY_, outputs_;
+
+  /// Following member variables are same with CudnnConvLayer.
+  /// There is no explanation here.
+  int fwdAlgo_, bwdFilterAlgo_, bwdDataAlgo_;
+  size_t fwdLimitBytes_, bwdDataLimitBytes_, bwdFilterLimitBytes_;
+  size_t workSpaceInBytes_;
+  void *workSpace_;
+  bool isSelectAlgo_;
+};
+
+}  // namespace paddle

paddle/gserver/layers/ConvBaseProjection.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "ConvBaseProjection.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+ThreadLocalD<std::vector<MemoryHandle *>> ConvBaseProjection::convMem_;
+
+ConvBaseProjection::ConvBaseProjection(const ProjectionConfig &config,
+                                       ParameterPtr parameter,
+                                       bool useGpu)
+    : Projection(config, parameter, useGpu) {
+  CHECK(useGpu);  // only support GPU
+  getConvParams();
+  initCudnn();
+
+  size_t height = filterH_ * filterW_ * channels_ / groups_;
+  size_t width = numFilters_;
+  weight_.reset(new Weight(height, width, parameter));
+  weightOffset_ = height * width / groups_;
+}
+
+void ConvBaseProjection::getConvParams() {
+  const ConvConfig &conf = config_.conv_conf();
+  paddingH_ = conf.padding_y();
+  paddingW_ = conf.padding();
+
+  strideH_ = conf.stride_y();
+  strideW_ = conf.stride();
+
+  filterH_ = conf.filter_size_y();
+  filterW_ = conf.filter_size();
+
+  configImgH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
+  configImgW_ = conf.img_size();
+
+  configOutH_ = conf.has_output_y() ? conf.output_y() : conf.output_x();
+  configOutW_ = conf.output_x();
+
+  configChannels_ = conf.channels();
+  configNumFilters_ = config_.num_filters();
+
+  isDeconv_ = (config_.type() == "conv") ? false : true;
+
+  channels_ = (isDeconv_) ? configNumFilters_ : configChannels_;
+  numFilters_ = (isDeconv_) ? configChannels_ : configNumFilters_;
+
+  groups_ = conf.groups();
+  CHECK_EQ(channels_ % groups_, 0);
+  CHECK_EQ(numFilters_ % groups_, 0);
+}
+
+void ConvBaseProjection::initCudnn() {
+  hl_create_filter_descriptor(&filterDesc_,
+                              channels_ / groups_,
+                              numFilters_ / groups_,
+                              filterH_,
+                              filterW_);
+  hl_create_tensor_descriptor(&imageDesc_);
+  hl_create_tensor_descriptor(&outputDesc_);
+  hl_create_convolution_descriptor(&convDesc_,
+                                   imageDesc_,
+                                   filterDesc_,
+                                   paddingH_,
+                                   paddingW_,
+                                   strideH_,
+                                   strideW_);
+
+  // initialize all to default algorithms
+  fwdAlgo_ = 0;
+  bwdFilterAlgo_ = 0;
+  bwdDataAlgo_ = 0;
+  fwdLimitBytes_ = 0;
+  bwdDataLimitBytes_ = 0;
+  bwdFilterLimitBytes_ = 0;
+  workSpaceInBytes_ = 0;
+
+  batchNum_ = 0;
+  isSelectAlgo_ = false;
+}
+
+void ConvBaseProjection::reshapeTensorDesc(int batchSize) {
+  // The stride between two consecutive samples in the output of ConvProjection
+  // may not be numFilters_ * outputH_ * outputW_ (conv) or
+  // channels_ * imageH_ * imageW_ (deconv)
+  // for example, in the case of layer ConcatenateLayer2 with two
+  // ConvProjection, the stride is the output_size of layer ConcatenateLayer2.
+  // So the calculation of nStride is different from CudnnConvLayer.
+  size_t nStrideImage, nStrideOutput;
+  if (isDeconv_) {
+    nStrideImage = out_->value->getStride();
+    nStrideOutput = numFilters_ * outputH_ * outputW_;
+  } else {
+    nStrideImage = channels_ * imageH_ * imageW_;
+    nStrideOutput = out_->value->getStride();
+  }
+
+  hl_tensor_reshape(imageDesc_,
+                    batchSize,
+                    channels_ / groups_,
+                    imageH_,
+                    imageW_,
+                    nStrideImage,
+                    imageH_ * imageW_,
+                    imageW_,
+                    1);
+
+  hl_tensor_reshape(outputDesc_,
+                    batchSize,
+                    numFilters_ / groups_,
+                    outputH_,
+                    outputW_,
+                    nStrideOutput,
+                    outputH_ * outputW_,
+                    outputW_,
+                    1);
+
+  hl_reset_convolution_descriptor(convDesc_,
+                                  imageDesc_,
+                                  filterDesc_,
+                                  paddingH_,
+                                  paddingW_,
+                                  strideH_,
+                                  strideW_);
+}
+
+void ConvBaseProjection::reshape(int batchSize) {
+  size_t width = calOutputSize();
+  CHECK_EQ(width, out_->value->getWidth());
+  CHECK_EQ(calInputSize(), in_->value->getWidth());
+
+  isSelectAlgo_ = (batchSize == batchNum_);
+  batchNum_ = batchSize;
+
+  if (!isSelectAlgo_) {
+    reshapeTensorDesc(batchSize);
+    hl_conv_workspace(imageDesc_,
+                      outputDesc_,
+                      filterDesc_,
+                      convDesc_,
+                      &fwdAlgo_,
+                      &fwdLimitBytes_,
+                      &bwdDataAlgo_,
+                      &bwdDataLimitBytes_,
+                      &bwdFilterAlgo_,
+                      &bwdFilterLimitBytes_);
+
+    size_t maxWorkSpace = 0;
+    maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_);
+    maxWorkSpace = std::max(maxWorkSpace, bwdFilterLimitBytes_);
+    workSpaceInBytes_ = maxWorkSpace;
+
+    VLOG(3) << getName() << " Fwd / BwdData / BwdFilter algo: " << fwdAlgo_
+            << " / " << bwdDataAlgo_ << " / " << bwdFilterAlgo_;
+  }
+
+  isSelectAlgo_ = true;
+}
+
+void *ConvBaseProjection::getSpaceBytes(size_t size) {
+  std::vector<MemoryHandle *> &convMem = *convMem_;
+  if (convMem.empty()) {
+    int numDevices = hl_get_device_count();
+    convMem.resize(numDevices);
+  }
+
+  int devId = hl_get_device();
+  MemoryHandle **localMem = &(convMem[devId]);
+  if (NULL == *localMem || size > (*localMem)->getAllocSize()) {
+    *localMem = new GpuMemoryHandle(size);
+  }
+  return (*localMem)->getBuf();
+}
+
+ConvBaseProjection::~ConvBaseProjection() {
+  hl_destroy_tensor_descriptor(imageDesc_);
+  hl_destroy_tensor_descriptor(outputDesc_);
+  hl_destroy_filter_descriptor(filterDesc_);
+  hl_destroy_convolution_descriptor(convDesc_);
+}
+
+}  // namespace paddle

paddle/gserver/layers/ConvBaseProjection.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "Projection.h"
+#include "paddle/math/MathUtils.h"
+
+namespace paddle {
+
+/**
+ * @brief Base class for ConvProjection and ConvTransProjection.
+ */
+class ConvBaseProjection : public Projection {
+public:
+  /**
+   * Constructor.
+   */
+  ConvBaseProjection(const ProjectionConfig& config,
+                     ParameterPtr parameter,
+                     bool useGpu);
+
+  ~ConvBaseProjection();
+
+protected:
+  void getConvParams();
+  void initCudnn();
+
+  void reshapeTensorDesc(int batchSize);
+  void reshape(int batchSize);
+
+  virtual size_t calOutputSize() = 0;
+  virtual size_t calInputSize() = 0;
+
+  static void* getSpaceBytes(size_t size);
+
+  /// True if it's deconv projection layer, false if it's ConvProjection layer
+  bool isDeconv_;
+  /// imageH_ and imageW_ / outputH_ and outputW_
+  /// is calculated from the input layer.
+  int imageH_, imageW_;
+  int outputH_, outputW_;
+  /// configImgH_ and configImgW_ / configOutH_ and configOutW_
+  /// is obtained from config.
+  int configImgH_, configImgW_;
+  int configOutH_, configOutW_;
+  /// channels_ and numFilters_ are defined in terms of convolution semantics
+  int channels_, numFilters_;
+  /// configChannels and configNumFilters_ are obtained from config
+  /// For Conv they are the same as channels_ and numFilters
+  /// For ConvTrans they are opposite to channels_ and numFilters
+  int configChannels_, configNumFilters_;
+  int paddingH_, paddingW_;
+  int strideH_, strideW_;
+  int filterH_, filterW_;
+  /// One group offset of input data.
+  int inputOffset_;
+  /// One group offset of output data.
+  int outputOffset_;
+  /// One group offset of weight.
+  int weightOffset_;
+  int groups_;
+
+  /// Cudnn tensor descriptor for input.
+  hl_tensor_descriptor imageDesc_;
+  /// Cudnn tensor descriptor for output.
+  hl_tensor_descriptor outputDesc_;
+  /// Cudnn tensor descriptor for filter.
+  hl_filter_descriptor filterDesc_;
+  /// Cudnn tensor descriptor for a convolution operation.
+  hl_convolution_descriptor convDesc_;
+
+  /// Record the algorithm for forward convolution, which is obtained by cudnn
+  /// api to search the best suited algorithm.
+  int fwdAlgo_;
+  /// Record the algorithm for computing convolution gradient with respect to
+  /// filter coefficients.
+  int bwdFilterAlgo_;
+  /// Record the algorithm for computing convolution gradient with respect to
+  /// the output.
+  int bwdDataAlgo_;
+  /// Amount of GPU memory needed as workspace to be able to execute a
+  /// forward convolution with the specified algo.
+  size_t fwdLimitBytes_;
+  /// Amount of GPU memory needed as workspace to be able to execute a
+  /// backwardFilter with the specified algo.
+  size_t bwdDataLimitBytes_;
+  /// Amount of GPU memory needed as workspace to be able to execute a
+  /// backwardData with the specified algo.
+  size_t bwdFilterLimitBytes_;
+  /// Size of total work space.
+  size_t workSpaceInBytes_;
+
+  /// Whether to call cuDNN api to choose conv algorithm.
+  bool isSelectAlgo_;
+  /// batchNum is used to record batch size. If the batch size is changed,
+  /// the selection algorithm will be called.
+  int batchNum_;
+  bool bias_;
+
+  std::unique_ptr<Weight> weight_;
+  static ThreadLocalD<std::vector<MemoryHandle*>> convMem_;
+};
+
+}  // namespace paddle

paddle/gserver/layers/ConvOperator.cpp

@@ -12,7 +12,7 @@ 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 "Operator.h"
+#include "ConvOperator.h"
 #include "paddle/math/MathUtils.h"
 #include "paddle/math/Matrix.h"
 
@@ -27,120 +27,8 @@ namespace paddle {
  * The config file api is conv_operator.
  */
 
-class ConvOperator : public Operator {
-public:
-  ConvOperator(const OperatorConfig &config, bool useGpu);
-  /**
-   * Free workspace in device and destroy cudnn tensor descriptor.
-   */
-  virtual ~ConvOperator() {
-    if (workSpaceInBytes_ != 0) {
-      hl_free_mem_device(workSpace_);
-      workSpaceInBytes_ = 0;
-    }
-
-    hl_destroy_tensor_descriptor(inputDesc_);
-    hl_destroy_tensor_descriptor(outputDesc_);
-    hl_destroy_filter_descriptor(filterDesc_);
-    hl_destroy_convolution_descriptor(convDesc_);
-  }
-  virtual void forward();
-  virtual void backward();
-
-private:
-  /**
-   * Get convolution parameters from layer config and
-   * initialize member variables.
-   */
-  void getConvParams();
-
-  /**
-   * Allocate Gpu Memory for cudnn convolution algorithms.
-   */
-  void allocConvWorkSpace(size_t maxWorkSpace);
-
-  /**
-   * Create cudnn tensor descriptor for convolution operation.
-   */
-  void computeConvSizes();
-
-  /**
-   * Reshape cudnn tensor descriptor.
-   */
-  void reshapeImageDescriptors();
-
-  /**
-   * Reshape cudnn tensor descriptor.
-   */
-  void reshape(int batchSize);
-
-  /**
-   * Check filter size is equal to the size calculated by parameters from
-   * layer config.
-   */
-  void checkFilterSize(const MatrixPtr &filter) {
-    CHECK_EQ(static_cast<int>(filter->getWidth()),
-             filterSize_ * filterSizeY_ * channels_ * numFilters_);
-  }
-
-  /// Most of member variables are same with CudnnConvLayer.
-  /// There is no explanation here.
-  int imageH_, imageW_, outputH_, outputW_;
-  hl_tensor_descriptor inputDesc_;
-  hl_tensor_descriptor outputDesc_;
-  hl_filter_descriptor filterDesc_;
-  hl_convolution_descriptor convDesc_;
-  bool caffeMode_;
-  int inputOffset_, outputOffset_, weightOffset_;
-  int numFilters_;
-  int padding_, stride_, filterSize_, channels_, imgSize_, imgSizeY_;
-  int paddingY_, strideY_, filterSizeY_;
-  int imgPixels_, filterPixels_, filterChannels_, outputX_, outputY_, outputs_;
-
-  /// Following member variables are same with CudnnConvLayer.
-  /// There is no explanation here.
-  int fwdAlgo_, bwdFilterAlgo_, bwdDataAlgo_;
-  size_t fwdLimitBytes_, bwdDataLimitBytes_, bwdFilterLimitBytes_;
-  size_t workSpaceInBytes_;
-  void *workSpace_;
-  bool isSelectAlgo_;
-};
-
 REGISTER_OPERATOR(conv, ConvOperator);
 
-ConvOperator::ConvOperator(const OperatorConfig &config, bool useGpu)
-    : Operator(config, useGpu) {
-  CHECK(useGpu);
-  CHECK_EQ(config_.input_indices_size(), 2L);
-
-  caffeMode_ = true;
-  getConvParams();
-  computeConvSizes();
-
-  // initialize all to default algorithms
-  fwdAlgo_ = 0;
-  bwdFilterAlgo_ = 0;
-  bwdDataAlgo_ = 0;
-  fwdLimitBytes_ = 0;
-  bwdDataLimitBytes_ = 0;
-  bwdFilterLimitBytes_ = 0;
-  workSpaceInBytes_ = 0;
-  workSpace_ = nullptr;
-
-  isSelectAlgo_ = false;
-}
-
-void ConvOperator::allocConvWorkSpace(size_t maxWorkSpace) {
-  if (maxWorkSpace > workSpaceInBytes_) {
-    if (workSpaceInBytes_ != 0) {
-      hl_free_mem_device(workSpace_);
-    }
-    // total amount of storage needed
-    workSpace_ = hl_malloc_device(maxWorkSpace);
-    workSpaceInBytes_ = maxWorkSpace;
-  }
-}
-
 void ConvOperator::reshape(int batchSize) {
   imageH_ = ins_[0]->getFrameHeight();
   imageW_ = ins_[0]->getFrameWidth();
@@ -148,106 +36,25 @@ void ConvOperator::reshape(int batchSize) {
   if (imageW_ == 0) imageW_ = imgSize_;
   outputH_ = outputSize(imageH_, filterSizeY_, paddingY_, strideY_, caffeMode_);
   outputW_ = outputSize(imageW_, filterSize_, padding_, stride_, caffeMode_);
-
+  /// Check that the outputSizes are consistent with config
+  CHECK_EQ(outputH_, outputY_);
+  CHECK_EQ(outputW_, outputX_);
   out_->setFrameHeight(outputH_);
   out_->setFrameWidth(outputW_);
 
   reshapeImageDescriptors();
 
-  if (!isSelectAlgo_) {
-    hl_conv_workspace(inputDesc_,
-                      outputDesc_,
-                      filterDesc_,
-                      convDesc_,
-                      &fwdAlgo_,
-                      &fwdLimitBytes_,
-                      &bwdDataAlgo_,
-                      &bwdDataLimitBytes_,
-                      &bwdFilterAlgo_,
-                      &bwdFilterLimitBytes_);
-
-    size_t maxWorkSpace = 0;
-    maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_);
-    maxWorkSpace = std::max(maxWorkSpace, bwdFilterLimitBytes_);
+  inputOffset_ = channels_ * imageH_ * imageW_;
+  outputOffset_ = numFilters_ * outputH_ * outputW_;
+  weightOffset_ = numFilters_ * channels_ * filterSize_ * filterSizeY_;
 
-    allocConvWorkSpace(maxWorkSpace);
+  if (!isSelectAlgo_) {
+    allocConvWorkSpace();
   }
 
   isSelectAlgo_ = true;
 }
 
-void ConvOperator::computeConvSizes() {
-  hl_create_filter_descriptor(
-      &filterDesc_, channels_, numFilters_, filterSizeY_, filterSize_);
-  hl_create_tensor_descriptor(&inputDesc_);
-  int outputX =
-      outputSize(imgSize_, filterSize_, padding_, stride_, caffeMode_);
-  int outputY =
-      outputSize(imgSizeY_, filterSizeY_, paddingY_, strideY_, caffeMode_);
-  CHECK_EQ(outputX, outputX_);
-  CHECK_EQ(outputY, outputY_);
-  hl_create_tensor_descriptor(&outputDesc_);
-  hl_create_convolution_descriptor(&convDesc_,
-                                   inputDesc_,
-                                   filterDesc_,
-                                   paddingY_,
-                                   padding_,
-                                   strideY_,
-                                   stride_);
-}
-
-void ConvOperator::reshapeImageDescriptors() {
-  hl_tensor_reshape(inputDesc_,
-                    1,
-                    channels_,
-                    imageH_,
-                    imageW_,
-                    channels_ * imageH_ * imageW_,
-                    imageH_ * imageW_,
-                    imageW_,
-                    1);
-  hl_tensor_reshape(outputDesc_,
-                    1,
-                    numFilters_,
-                    outputH_,
-                    outputW_,
-                    numFilters_ * outputH_ * outputW_,
-                    outputH_ * outputW_,
-                    outputW_,
-                    1);
-  hl_reset_convolution_descriptor(convDesc_,
-                                  inputDesc_,
-                                  filterDesc_,
-                                  paddingY_,
-                                  padding_,
-                                  strideY_,
-                                  stride_);
-  inputOffset_ = channels_ * imageH_ * imageW_;
-  outputOffset_ = numFilters_ * outputH_ * outputW_;
-  weightOffset_ = numFilters_ * channels_ * filterSize_ * filterSize_;
-}
-
-void ConvOperator::getConvParams() {
-  numFilters_ = config_.num_filters();
-  const ConvConfig &conf = config_.conv_conf();
-  padding_ = conf.padding();
-  stride_ = conf.stride();
-  filterSize_ = conf.filter_size();
-  paddingY_ = conf.padding_y();
-  strideY_ = conf.stride_y();
-  filterSizeY_ = conf.filter_size_y();
-  filterPixels_ = filterSize_ * filterSizeY_;
-  channels_ = conf.channels();
-  imgSize_ = conf.img_size();
-  imgSizeY_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
-  imgPixels_ = imgSize_ * imgSizeY_;
-  CHECK_EQ(conf.groups(), 1U);
-  filterChannels_ = conf.filter_channels();
-  outputX_ = conf.output_x();
-  outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x();
-  outputs_ = outputX_ * outputX_;
-}
-
 void ConvOperator::forward() {
   size_t batchSize = ins_[0]->value->getHeight();
   reshape(batchSize);
@@ -264,7 +71,7 @@ void ConvOperator::forward() {
       real *inputData = ins_[0]->value->getData() + inputOffset_ * batchId;
       real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId;
       real *outData = out_->value->getData() + outputOffset_ * batchId;
-      hl_convolution_forward(inputDesc_,
+      hl_convolution_forward(imageDesc_,
                              inputData,
                              outputDesc_,
                              outData,
@@ -287,7 +94,7 @@ void ConvOperator::backward() {
       if (ins_[1]->grad) {
         real *inputData = ins_[0]->value->getData() + inputOffset_ * batchId;
         real *weightGrad = ins_[1]->grad->getData() + weightOffset_ * batchId;
-        hl_convolution_backward_filter(inputDesc_,
+        hl_convolution_backward_filter(imageDesc_,
                                        inputData,
                                        outputDesc_,
                                        outGrad,
@@ -303,7 +110,7 @@ void ConvOperator::backward() {
       if (NULL != preGrad) {
         real *inputGrad = preGrad->getData() + inputOffset_ * batchId;
         real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId;
-        hl_convolution_backward_data(inputDesc_,
+        hl_convolution_backward_data(imageDesc_,
                                      inputGrad,
                                      outputDesc_,
                                      outGrad,

paddle/gserver/layers/ConvOperator.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+#pragma once
+
+#include "ConvBaseOperator.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * @brief ConvOperator takes two inputs to perform the convolution.
+ * The first input is the image, and the second input is the convolution kernel.
+ * The height of data for two inputs are the same. Each data of the first input
+ * is convolved with each data of the second input indepedently.
+ *
+ * The config file api is conv_operator.
+ */
+
+class ConvOperator : public ConvBaseOperator {
+public:
+  ConvOperator(const OperatorConfig &config, bool useGpu)
+      : ConvBaseOperator(config, useGpu) {}
+  /**
+   * Free workspace in device and destroy cudnn tensor descriptor.
+   */
+  virtual ~ConvOperator() {}
+  void forward() override;
+  void backward() override;
+  void reshape(int batchSize) override;
+};
+
+}  // namespace paddle

paddle/gserver/layers/ConvProjection.cpp

@@ -19,149 +19,32 @@ namespace paddle {
 
 REGISTER_PROJECTION(conv, ConvProjection);
 
-ThreadLocalD<std::vector<MemoryHandle *>> ConvProjection::convMem_;
-
-ConvProjection::ConvProjection(const ProjectionConfig &config,
-                               ParameterPtr parameter,
-                               bool useGpu)
-    : Projection(config, parameter, useGpu) {
-  CHECK(useGpu);  // only support GPU
-  getConvParams();
-  initCudnn();
-
-  size_t height = filterH_ * filterW_ * channels_ / groups_;
-  size_t width = numFilters_;
-  weight_.reset(new Weight(height, width, parameter));
-  weightOffset_ = height * width / groups_;
-}
-
-void ConvProjection::getConvParams() {
-  const ConvConfig &conf = config_.conv_conf();
-  paddingH_ = conf.padding_y();
-  paddingW_ = conf.padding();
-
-  strideH_ = conf.stride_y();
-  strideW_ = conf.stride();
-
-  filterH_ = conf.filter_size_y();
-  filterW_ = conf.filter_size();
-
-  configImgH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
-  configImgW_ = conf.img_size();
-
-  channels_ = conf.channels();
-  numFilters_ = config_.num_filters();
-
-  groups_ = conf.groups();
-  CHECK_EQ(channels_ % groups_, 0);
-  CHECK_EQ(numFilters_ % groups_, 0);
-}
-
-void ConvProjection::initCudnn() {
-  hl_create_filter_descriptor(&filterDesc_,
-                              channels_ / groups_,
-                              numFilters_ / groups_,
+size_t ConvProjection::calOutputSize() {
+  imageH_ = in_->getFrameHeight();
+  imageW_ = in_->getFrameWidth();
+  if (imageH_ == 0) imageH_ = configImgH_;
+  if (imageW_ == 0) imageW_ = configImgW_;
+  outputH_ = outputSize(imageH_,
                         filterH_,
-                              filterW_);
-  hl_create_tensor_descriptor(&inputDesc_);
-  hl_create_tensor_descriptor(&outputDesc_);
-  hl_create_convolution_descriptor(&convDesc_,
-                                   inputDesc_,
-                                   filterDesc_,
                         paddingH_,
-                                   paddingW_,
                         strideH_,
-                                   strideW_);
-
-  // initialize all to default algorithms
-  fwdAlgo_ = 0;
-  bwdFilterAlgo_ = 0;
-  bwdDataAlgo_ = 0;
-  fwdLimitBytes_ = 0;
-  bwdDataLimitBytes_ = 0;
-  bwdFilterLimitBytes_ = 0;
-  workSpaceInBytes_ = 0;
-
-  batchNum_ = 0;
-  isSelectAlgo_ = false;
-}
-
-void ConvProjection::reshapeTensorDesc(int batchSize) {
-  hl_tensor_reshape(inputDesc_,
-                    batchSize,
-                    channels_ / groups_,
-                    imageH_,
-                    imageW_,
-                    channels_ * imageH_ * imageW_,
-                    imageH_ * imageW_,
-                    imageW_,
-                    1);
-  hl_reset_convolution_descriptor(convDesc_,
-                                  inputDesc_,
-                                  filterDesc_,
-                                  paddingH_,
+                        /* caffeMode */ true);
+  outputW_ = outputSize(imageW_,
+                        filterW_,
                         paddingW_,
-                                  strideH_,
-                                  strideW_);
+                        strideW_,
+                        /* caffeMode */ true);
 
-  // The stride between two consecutive images in ConvProjection may not be 1,
-  // for example, in the case of layer ConcatenateLayer2 with two
-  // ConvProjection, the stride is the output_size of layer ConcatenateLayer2.
-  // So the calculation of nStride is different from CudnnConvLayer.
-  // In fact, only "nStride = out_->value->getStride()" is ok.
-  size_t nStride = numFilters_ * outputH_ * outputW_;
-  if (out_->value->isContiguous()) {
-    CHECK_EQ(nStride, out_->value->getWidth());
-  } else {
-    nStride = out_->value->getStride();
-  }
+  const_cast<Argument *>(out_)->setFrameHeight(outputH_);
+  const_cast<Argument *>(out_)->setFrameWidth(outputW_);
 
-  hl_tensor_reshape(outputDesc_,
-                    batchSize,
-                    numFilters_ / groups_,
-                    outputH_,
-                    outputW_,
-                    nStride,
-                    outputH_ * outputW_,
-                    outputW_,
-                    1);
+  inputOffset_ = (configChannels_ / groups_) * imageH_ * imageW_;
+  outputOffset_ = (configNumFilters_ / groups_) * outputH_ * outputW_;
+  return outputH_ * outputW_ * configNumFilters_;
 }
 
-void ConvProjection::reshape(int batchSize) {
-  size_t width = calOutputSize();
-  CHECK_EQ(width, out_->value->getWidth());
-  CHECK_EQ(static_cast<size_t>(channels_ * imageH_ * imageW_),
-           in_->value->getWidth())
-      << "Wrong input size for convolution"
-      << " channels=" << channels_ << " imageH=" << imageH_
-      << " imageW=" << imageW_ << " inputSize=" << in_->value->getWidth();
-
-  isSelectAlgo_ = (batchSize == batchNum_);
-  batchNum_ = batchSize;
-
-  if (!isSelectAlgo_) {
-    reshapeTensorDesc(batchSize);
-    hl_conv_workspace(inputDesc_,
-                      outputDesc_,
-                      filterDesc_,
-                      convDesc_,
-                      &fwdAlgo_,
-                      &fwdLimitBytes_,
-                      &bwdDataAlgo_,
-                      &bwdDataLimitBytes_,
-                      &bwdFilterAlgo_,
-                      &bwdFilterLimitBytes_);
-
-    size_t maxWorkSpace = 0;
-    maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_);
-    maxWorkSpace = std::max(maxWorkSpace, bwdFilterLimitBytes_);
-    workSpaceInBytes_ = maxWorkSpace;
-
-    VLOG(3) << getName() << " Fwd / BwdData / BwdFilter algo: " << fwdAlgo_
-            << " / " << bwdDataAlgo_ << " / " << bwdFilterAlgo_;
-  }
-
-  isSelectAlgo_ = true;
+size_t ConvProjection::calInputSize() {
+  return static_cast<size_t>(configChannels_ * imageH_ * imageW_);
 }
 
 void ConvProjection::forward() {
@@ -179,7 +62,7 @@ void ConvProjection::forward() {
     real *inputData = in_->value->getData() + g * inputOffset_;
     real *wgtData = weight_->getW()->getData() + g * weightOffset_;
     real *outData = out_->value->getData() + g * outputOffset_;
-    hl_convolution_forward(inputDesc_,
+    hl_convolution_forward(imageDesc_,
                            inputData,
                            outputDesc_,
                            outData,
@@ -205,7 +88,7 @@ void ConvProjection::backward(const UpdateCallback &callback) {
     if (weight_->getWGrad()) {
       real *inputData = in_->value->getData() + g * inputOffset_;
       real *weightGrad = weight_->getWGrad()->getData() + g * weightOffset_;
-      hl_convolution_backward_filter(inputDesc_,
+      hl_convolution_backward_filter(imageDesc_,
                                      inputData,
                                      outputDesc_,
                                      outGrad,
@@ -221,7 +104,7 @@ void ConvProjection::backward(const UpdateCallback &callback) {
     if (NULL != preGrad) {
       real *inputGrad = preGrad->getData() + g * inputOffset_;
       real *wgtData = weight_->getW()->getData() + g * weightOffset_;
-      hl_convolution_backward_data(inputDesc_,
+      hl_convolution_backward_data(imageDesc_,
                                    inputGrad,
                                    outputDesc_,
                                    outGrad,
@@ -237,26 +120,4 @@ void ConvProjection::backward(const UpdateCallback &callback) {
   weight_->getParameterPtr()->incUpdate(callback);
 }
 
-void *ConvProjection::getSpaceBytes(size_t size) {
-  std::vector<MemoryHandle *> &convMem = *convMem_;
-  if (convMem.empty()) {
-    int numDevices = hl_get_device_count();
-    convMem.resize(numDevices);
-  }
-
-  int devId = hl_get_device();
-  MemoryHandle **localMem = &(convMem[devId]);
-  if (NULL == *localMem || size > (*localMem)->getAllocSize()) {
-    *localMem = new GpuMemoryHandle(size);
-  }
-  return (*localMem)->getBuf();
-}
-
-ConvProjection::~ConvProjection() {
-  hl_destroy_tensor_descriptor(inputDesc_);
-  hl_destroy_tensor_descriptor(outputDesc_);
-  hl_destroy_filter_descriptor(filterDesc_);
-  hl_destroy_convolution_descriptor(convDesc_);
-}
-
 }  // namespace paddle

paddle/gserver/layers/ConvProjection.h

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #pragma once
 
-#include "Projection.h"
+#include "ConvBaseProjection.h"
 #include "paddle/math/MathUtils.h"
 
 namespace paddle {
@@ -22,109 +22,22 @@ namespace paddle {
 /**
  * @brief Convolution projection do the same calculation with CudnnConvLayer.
  */
-class ConvProjection : public Projection {
+class ConvProjection : public ConvBaseProjection {
 public:
   /**
    * Constructor.
    */
   ConvProjection(const ProjectionConfig& config,
                  ParameterPtr parameter,
-                 bool useGpu);
+                 bool useGpu)
+      : ConvBaseProjection(config, parameter, useGpu) {}
 
-  ~ConvProjection();
+  ~ConvProjection() {}
 
   virtual void forward();
   virtual void backward(const UpdateCallback& callback);
-
-protected:
-  void getConvParams();
-  void initCudnn();
-
-  void reshapeTensorDesc(int batchSize);
-  void reshape(int batchSize);
-
-  size_t calOutputSize() {
-    imageH_ = in_->getFrameHeight();
-    imageW_ = in_->getFrameWidth();
-    if (imageH_ == 0) imageH_ = configImgH_;
-    if (imageW_ == 0) imageW_ = configImgW_;
-    outputH_ = outputSize(imageH_,
-                          filterH_,
-                          paddingH_,
-                          strideH_,
-                          /* caffeMode */ true);
-    outputW_ = outputSize(imageW_,
-                          filterW_,
-                          paddingW_,
-                          strideW_,
-                          /* caffeMode */ true);
-
-    const_cast<Argument*>(out_)->setFrameHeight(outputH_);
-    const_cast<Argument*>(out_)->setFrameWidth(outputW_);
-
-    inputOffset_ = (channels_ / groups_) * imageH_ * imageW_;
-    outputOffset_ = (numFilters_ / groups_) * outputH_ * outputW_;
-    return outputH_ * outputW_ * numFilters_;
-  }
-
-  static void* getSpaceBytes(size_t size);
-
-  /// imageH_ and imageW_ is calculated from the input layer.
-  int imageH_, imageW_;
-  /// configImgH_ and configImgW_ is obtained from config.
-  int configImgH_, configImgW_;
-  int outputH_, outputW_;
-  int channels_, numFilters_;
-  int paddingH_, paddingW_;
-  int strideH_, strideW_;
-  int filterH_, filterW_;
-  /// One group offset of input data.
-  int inputOffset_;
-  /// One group offset of output data.
-  int outputOffset_;
-  /// One group offset of weight.
-  int weightOffset_;
-  int groups_;
-
-  /// Cudnn tensor descriptor for input.
-  hl_tensor_descriptor inputDesc_;
-  /// Cudnn tensor descriptor for output.
-  hl_tensor_descriptor outputDesc_;
-  /// Cudnn tensor descriptor for filter.
-  hl_filter_descriptor filterDesc_;
-  /// Cudnn tensor descriptor for a convolution operation.
-  hl_convolution_descriptor convDesc_;
-
-  /// Record the algorithm for forward convolution, which is obtained by cudnn
-  /// api to search the best suited algorithm.
-  int fwdAlgo_;
-  /// Record the algorithm for computing convolution gradient with respect to
-  /// filter coefficients.
-  int bwdFilterAlgo_;
-  /// Record the algorithm for computing convolution gradient with respect to
-  /// the output.
-  int bwdDataAlgo_;
-  /// Amount of GPU memory needed as workspace to be able to execute a
-  /// forward convolution with the specified algo.
-  size_t fwdLimitBytes_;
-  /// Amount of GPU memory needed as workspace to be able to execute a
-  /// backwardFilter with the specified algo.
-  size_t bwdDataLimitBytes_;
-  /// Amount of GPU memory needed as workspace to be able to execute a
-  /// backwardData with the specified algo.
-  size_t bwdFilterLimitBytes_;
-  /// Size of total work space.
-  size_t workSpaceInBytes_;
-
-  /// Whether to call cuDNN api to choose conv algorithm.
-  bool isSelectAlgo_;
-  /// batchNum is used to record batch size. If the batch size is changed,
-  /// the selection algorithm will be called.
-  int batchNum_;
-  bool bias_;
-
-  std::unique_ptr<Weight> weight_;
-  static ThreadLocalD<std::vector<MemoryHandle*>> convMem_;
+  virtual size_t calOutputSize();
+  virtual size_t calInputSize();
 };
 
 }  // namespace paddle

paddle/gserver/layers/ConvTransOperator.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "ConvTransOperator.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * @brief ConvTransOperator takes two inputs to perform the convolution.
+ * The first input is the image, and the second input is the convolution kernel.
+ * The height of data for two inputs are the same. Each data of the first input
+ * is convolved with each data of the second input indepedently.
+ *
+ * The config file api is conv_operator.
+ */
+
+REGISTER_OPERATOR(convt, ConvTransOperator);
+
+void ConvTransOperator::reshape(int batchSize) {
+  outputH_ = ins_[0]->getFrameHeight();
+  outputW_ = ins_[0]->getFrameWidth();
+  if (outputH_ == 0) outputH_ = outputY_;
+  if (outputW_ == 0) outputW_ = outputX_;
+  imageH_ = imageSize(outputH_, filterSizeY_, paddingY_, strideY_, caffeMode_);
+  imageW_ = imageSize(outputW_, filterSize_, padding_, stride_, caffeMode_);
+  /// Check that the imageSizes are consistent with config
+  CHECK_EQ(imageH_, imgSizeY_);
+  CHECK_EQ(imageW_, imgSize_);
+  out_->setFrameHeight(imageH_);
+  out_->setFrameWidth(imageW_);
+
+  reshapeImageDescriptors();
+
+  inputOffset_ = numFilters_ * outputH_ * outputW_;
+  outputOffset_ = channels_ * imageH_ * imageW_;
+  weightOffset_ = numFilters_ * channels_ * filterSize_ * filterSizeY_;
+
+  if (!isSelectAlgo_) {
+    allocConvWorkSpace();
+  }
+
+  isSelectAlgo_ = true;
+}
+
+void ConvTransOperator::forward() {
+  size_t batchSize = ins_[0]->value->getHeight();
+  reshape(batchSize);
+  CHECK_EQ(ins_[1]->value->getHeight(), batchSize);
+  checkFilterSize(ins_[1]->value);
+  Matrix::resizeOrCreate(
+      out_->value, batchSize, imageH_ * imageW_ * channels_, false, useGpu_);
+  {
+    AsyncGpuBlock block;
+    for (size_t batchId = 0; batchId < batchSize; ++batchId) {
+      real *inputData = ins_[0]->value->getData() + inputOffset_ * batchId;
+      real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId;
+      real *outData = out_->value->getData() + outputOffset_ * batchId;
+      hl_convolution_backward_data(imageDesc_,
+                                   outData,
+                                   outputDesc_,
+                                   inputData,
+                                   filterDesc_,
+                                   wgtData,
+                                   convDesc_,
+                                   workSpace_,
+                                   workSpaceInBytes_,
+                                   bwdDataAlgo_);
+    }
+  }
+}
+
+void ConvTransOperator::backward() {
+  size_t batchSize = ins_[0]->value->getHeight();
+  {
+    AsyncGpuBlock block;
+    for (size_t batchId = 0; batchId < batchSize; ++batchId) {
+      real *outGrad = out_->grad->getData() + outputOffset_ * batchId;
+      if (ins_[1]->grad) {
+        real *inputData = ins_[0]->value->getData() + inputOffset_ * batchId;
+        real *weightGrad = ins_[1]->grad->getData() + weightOffset_ * batchId;
+        hl_convolution_backward_filter(imageDesc_,
+                                       outGrad,
+                                       outputDesc_,
+                                       inputData,
+                                       filterDesc_,
+                                       weightGrad,
+                                       convDesc_,
+                                       workSpace_,
+                                       workSpaceInBytes_,
+                                       bwdFilterAlgo_);
+      }
+
+      MatrixPtr preGrad = ins_[0]->grad;
+      if (NULL != preGrad) {
+        real *inputGrad = preGrad->getData() + inputOffset_ * batchId;
+        real *wgtData = ins_[1]->value->getData() + weightOffset_ * batchId;
+        hl_convolution_forward(imageDesc_,
+                               outGrad,
+                               outputDesc_,
+                               inputGrad,
+                               filterDesc_,
+                               wgtData,
+                               convDesc_,
+                               workSpace_,
+                               workSpaceInBytes_,
+                               fwdAlgo_);
+      }
+    }
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/ConvTransOperator.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+#pragma once
+
+#include "ConvBaseOperator.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * @brief ConvTransOperator takes two inputs to perform the convolution.
+ * The first input is the image, and the second input is the convolution kernel.
+ * The height of data for two inputs are the same. Each data of the first input
+ * is convolved with each data of the second input indepedently.
+ *
+ * The config file api is conv_operator.
+ */
+
+class ConvTransOperator : public ConvBaseOperator {
+public:
+  ConvTransOperator(const OperatorConfig &config, bool useGpu)
+      : ConvBaseOperator(config, useGpu) {}
+  /**
+   * Free workspace in device and destroy cudnn tensor descriptor.
+   */
+  virtual ~ConvTransOperator() {}
+  void forward() override;
+  void backward() override;
+  void reshape(int batchSize) override;
+};
+
+}  // namespace paddle

paddle/gserver/layers/ConvTransProjection.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "ConvTransProjection.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+REGISTER_PROJECTION(convt, ConvTransProjection);
+size_t ConvTransProjection::calOutputSize() {
+  outputH_ = in_->getFrameHeight();
+  outputW_ = in_->getFrameWidth();
+  if (outputH_ == 0) outputH_ = configOutH_;
+  if (outputW_ == 0) outputW_ = configOutW_;
+  imageH_ = imageSize(outputH_,
+                      filterH_,
+                      paddingH_,
+                      strideH_,
+                      /* caffeMode */ true);
+
+  imageW_ = imageSize(outputW_,
+                      filterW_,
+                      paddingW_,
+                      strideW_,
+                      /* caffeMode */ true);
+
+  const_cast<Argument *>(out_)->setFrameHeight(imageH_);
+  const_cast<Argument *>(out_)->setFrameWidth(imageW_);
+
+  inputOffset_ = (configChannels_ / groups_) * outputH_ * outputW_;
+  outputOffset_ = (configNumFilters_ / groups_) * imageH_ * imageW_;
+  return imageH_ * imageW_ * configNumFilters_;
+}
+
+size_t ConvTransProjection::calInputSize() {
+  return static_cast<size_t>(configChannels_ * outputH_ * outputW_);
+}
+
+void ConvTransProjection::forward() {
+  int batchSize = in_->value->getHeight();
+  reshape(batchSize);
+
+  void *workSpace = NULL;
+  if (workSpaceInBytes_ > 0) {
+    workSpace = getSpaceBytes(workSpaceInBytes_);
+  }
+
+  for (int g = 0; g < groups_; ++g) {
+    REGISTER_TIMER_INFO("CudnnConvTransFwTimer", getName().c_str());
+
+    real *inData = in_->value->getData() + g * inputOffset_;
+    real *wgtData = weight_->getW()->getData() + g * weightOffset_;
+    real *outData = out_->value->getData() + g * outputOffset_;
+    hl_convolution_backward_data(imageDesc_,
+                                 outData,
+                                 outputDesc_,
+                                 inData,
+                                 filterDesc_,
+                                 wgtData,
+                                 convDesc_,
+                                 workSpace,
+                                 bwdDataLimitBytes_,
+                                 bwdDataAlgo_);
+  }
+}
+
+void ConvTransProjection::backward(const UpdateCallback &callback) {
+  REGISTER_TIMER_INFO("CudnnConvTransBpTimer", getName().c_str());
+
+  void *workSpace = NULL;
+  if (workSpaceInBytes_ > 0) {
+    workSpace = getSpaceBytes(workSpaceInBytes_);
+  }
+
+  for (int g = 0; g < groups_; ++g) {
+    real *outGrad = out_->grad->getData() + g * outputOffset_;
+    if (weight_->getWGrad()) {
+      real *inData = in_->value->getData() + g * inputOffset_;
+      real *weightGrad = weight_->getWGrad()->getData() + g * weightOffset_;
+      hl_convolution_backward_filter(imageDesc_,
+                                     outGrad,
+                                     outputDesc_,
+                                     inData,
+                                     filterDesc_,
+                                     weightGrad,
+                                     convDesc_,
+                                     workSpace,
+                                     bwdFilterLimitBytes_,
+                                     bwdFilterAlgo_);
+    }
+
+    MatrixPtr preGrad = in_->grad;
+    if (NULL != preGrad) {
+      real *inGrad = preGrad->getData() + g * inputOffset_;
+      real *wgtData = weight_->getW()->getData() + g * weightOffset_;
+      hl_convolution_forward(imageDesc_,
+                             outGrad,
+                             outputDesc_,
+                             inGrad,
+                             filterDesc_,
+                             wgtData,
+                             convDesc_,
+                             workSpace,
+                             fwdLimitBytes_,
+                             fwdAlgo_);
+    }
+  }
+
+  weight_->getParameterPtr()->incUpdate(callback);
+}
+
+}  // namespace paddle

paddle/gserver/layers/ConvTransProjection.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "ConvBaseProjection.h"
+#include "paddle/math/MathUtils.h"
+
+namespace paddle {
+
+/**
+ * @brief Convolution projection do the same calculation with CudnnConvLayer.
+ */
+class ConvTransProjection : public ConvBaseProjection {
+public:
+  /**
+   * Constructor.
+   */
+  ConvTransProjection(const ProjectionConfig& config,
+                      ParameterPtr parameter,
+                      bool useGpu)
+      : ConvBaseProjection(config, parameter, useGpu) {}
+
+  ~ConvTransProjection() {}
+
+  virtual void forward();
+  virtual void backward(const UpdateCallback& callback);
+  virtual size_t calOutputSize();
+  virtual size_t calInputSize();
+};
+
+}  // namespace paddle

paddle/gserver/layers/CostLayer.cpp

@@ -192,6 +192,59 @@ void SumOfSquaresCostLayer::backwardImp(Matrix& output,
   outputG.sumOfSquaresBp(output, *label.value);
 }
 
+//
+// class SmoothL1CostLayer
+//
+
+REGISTER_LAYER(smooth_l1, SmoothL1CostLayer);
+
+bool SmoothL1CostLayer::init(const LayerMap& layerMap,
+                             const ParameterMap& parameterMap) {
+  return CostLayer::init(layerMap, parameterMap);
+}
+
+void SmoothL1CostLayer::forwardImp(Matrix& output,
+                                   Argument& label,
+                                   Matrix& target) {
+  MatrixPtr targetCpu, outputCpu, labelCpu;
+  if (useGpu_) {
+    targetCpu =
+        Matrix::create(target.getHeight(), target.getWidth(), false, false);
+    outputCpu =
+        Matrix::create(output.getHeight(), output.getWidth(), false, false);
+    labelCpu = Matrix::create(
+        label.value->getHeight(), label.value->getWidth(), false, false);
+    targetCpu->copyFrom(target);
+    outputCpu->copyFrom(output);
+    labelCpu->copyFrom(*label.value);
+    targetCpu->smoothL1(*outputCpu, *(labelCpu));
+    target.copyFrom(*targetCpu);
+  } else {
+    target.smoothL1(output, *label.value);
+  }
+}
+
+void SmoothL1CostLayer::backwardImp(Matrix& output,
+                                    Argument& label,
+                                    Matrix& outputG) {
+  MatrixPtr outputGCpu, outputCpu, labelCpu;
+  if (useGpu_) {
+    outputGCpu =
+        Matrix::create(outputG.getHeight(), outputG.getWidth(), false, false);
+    outputCpu =
+        Matrix::create(output.getHeight(), output.getWidth(), false, false);
+    labelCpu = Matrix::create(
+        label.value->getHeight(), label.value->getWidth(), false, false);
+    outputGCpu->copyFrom(outputG);
+    outputCpu->copyFrom(output);
+    labelCpu->copyFrom(*label.value);
+    outputGCpu->smoothL1Bp(*outputCpu, *labelCpu);
+    outputG.copyFrom(*outputGCpu);
+  } else {
+    outputG.smoothL1Bp(output, *label.value);
+  }
+}
+
 //
 // class RankingCost
 //

paddle/gserver/layers/CostLayer.h

@@ -159,6 +159,29 @@ public:
                    Matrix& outputGrad) override;
 };
 
+/**
+ * This cost layer compute smooth L1 loss for real-valued regression
+ * tasks.
+ * \f[
+ * L =
+ *   (output - label)^2 * 0.5  / -1 < (output - label) < 1 /
+ *   (output - label) - 0.5    / otherwise  /
+ * \f]
+ */
+class SmoothL1CostLayer : public CostLayer {
+public:
+  explicit SmoothL1CostLayer(const LayerConfig& config) : CostLayer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
+
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
+};
+
 /**
  * A cost layer for learning to rank (LTR) task. This layer contains at leat
  * three inputs.

paddle/gserver/layers/CudnnConvLayer.cpp → paddle/gserver/layers/CudnnConvBaseLayer.cpp

@@ -12,15 +12,15 @@ 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 "CudnnConvLayer.h"
+#include "CudnnConvBaseLayer.h"
 #include "paddle/utils/Logging.h"
 #include "paddle/utils/Stat.h"
 
 namespace paddle {
+REGISTER_LAYER(cudnn_conv, CudnnConvBaseLayer);
+REGISTER_LAYER(cudnn_convt, CudnnConvBaseLayer);
 
-REGISTER_LAYER(cudnn_conv, CudnnConvLayer);
-
-bool CudnnConvLayer::init(const LayerMap &layerMap,
+bool CudnnConvBaseLayer::init(const LayerMap &layerMap,
                               const ParameterMap &parameterMap) {
   if (!ConvBaseLayer::init(layerMap, parameterMap)) return false;
   CHECK(useGpu_) << "CudnnConvLayer only support gpu";
@@ -33,7 +33,11 @@ bool CudnnConvLayer::init(const LayerMap &layerMap,
   CHECK(config_.shared_biases());
   for (size_t i = 0; i < inputLayers_.size(); i++) {
     ProjectionConfig *conf = new ProjectionConfig();
+    if (isDeconv_) {
+      conf->set_type("convt");
+    } else {
       conf->set_type("conv");
+    }
     conf->set_num_filters(numFilters_);
     ConvConfig *convConf = conf->mutable_conv_conf();
     *convConf = *(config_.mutable_inputs(i)->mutable_conv_conf());
@@ -47,14 +51,13 @@ bool CudnnConvLayer::init(const LayerMap &layerMap,
   if (biases_.get() && sharedBiases_) {
     hl_create_tensor_descriptor(&biasDesc_);
     hl_create_tensor_descriptor(&outputDesc_);
-    hl_tensor_reshape(biasDesc_, 1, numFilters_ / groups_[0], 1, 1);
-    biasOffset_ = numFilters_ / groups_[0];
+    hl_tensor_reshape(biasDesc_, 1, numFilters_, 1, 1);
   }
 
   return true;
 }
 
-void CudnnConvLayer::forward(PassType passType) {
+void CudnnConvBaseLayer::forward(PassType passType) {
   Layer::forward(passType);
 
   int batchSize = getInput(0).getBatchSize();
@@ -67,37 +70,41 @@ void CudnnConvLayer::forward(PassType passType) {
   if (biases_) {
     REGISTER_TIMER_INFO("CudnnConvBiasTimer", getName().c_str());
     int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
+    int outH, outW;
+    if (isDeconv_) {
+      outH = imgSizeH_[0];
+      outW = imgSizeW_[0];
+    } else {
+      outH = outputH_[0];
+      outW = outputW_[0];
+    }
+
     hl_tensor_reshape(outputDesc_,
                       batchSize,
-                      numFilters_ / groups_[0],
-                      outputH_[0],
-                      outputW_[0],
-                      numFilters_ * outputH_[0] * outputW_[0],
-                      outputH_[0] * outputW_[0],
-                      outputW_[0],
+                      numFilters_,
+                      outH,
+                      outW,
+                      numFilters_ * outH * outW,
+                      outH * outW,
+                      outW,
                       1);
-    outputOffset_ = getOutputValue()->getWidth() / groups_[0];
-    for (int g = 0; g < groups_[0]; ++g) {
-      real *biasData = biases_->getW()->getData() + biasOffset_ * g;
-      real *outData = getOutputValue()->getData() + outputOffset_ * g;
-      hl_convolution_forward_add_bias(
-          biasDesc_, biasData, outputDesc_, outData);
-    }
+    real *outData = getOutputValue()->getData();
+    real *biasData = biases_->getW()->getData();
+    hl_convolution_forward_add_bias(biasDesc_, biasData, outputDesc_, outData);
   }
 
   forwardActivation();
 }
 
-void CudnnConvLayer::backward(const UpdateCallback &callback) {
+void CudnnConvBaseLayer::backward(const UpdateCallback &callback) {
   backwardActivation();
 
   if (biases_ && biases_->getWGrad()) {
     REGISTER_TIMER_INFO("CudnnConvBpBiasTimer", getName().c_str());
-    for (int g = 0; g < groups_[0]; ++g) {
-      real *biasGrad = biases_->getWGrad()->getData() + biasOffset_ * g;
-      real *outGrad = getOutputGrad()->getData() + outputOffset_ * g;
+    real *biasGrad = biases_->getWGrad()->getData();
+    real *outGrad = getOutputGrad()->getData();
     hl_convolution_backward_bias(biasDesc_, biasGrad, outputDesc_, outGrad);
-    }
+
     biases_->getParameterPtr()->incUpdate(callback);
   }
 
@@ -106,7 +113,7 @@ void CudnnConvLayer::backward(const UpdateCallback &callback) {
   }
 }
 
-CudnnConvLayer::~CudnnConvLayer() {
+CudnnConvBaseLayer::~CudnnConvBaseLayer() {
   if (biases_) {
     hl_destroy_tensor_descriptor(biasDesc_);
     hl_destroy_tensor_descriptor(outputDesc_);

paddle/gserver/layers/CudnnConvLayer.h → paddle/gserver/layers/CudnnConvBaseLayer.h

@@ -30,27 +30,24 @@ namespace paddle {
  *
  * The config file api is img_conv_layer.
  */
-class CudnnConvLayer : public ConvBaseLayer {
+class CudnnConvBaseLayer : public ConvBaseLayer {
 protected:
   std::vector<std::unique_ptr<ProjectionConfig>> projConf_;
   std::vector<std::unique_ptr<Projection>> projections_;
 
   hl_tensor_descriptor biasDesc_;
   hl_tensor_descriptor outputDesc_;
-  int biasOffset_;
-  int outputOffset_;
 
 public:
-  explicit CudnnConvLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
+  explicit CudnnConvBaseLayer(const LayerConfig& config)
+      : ConvBaseLayer(config) {}
 
-  ~CudnnConvLayer();
+  ~CudnnConvBaseLayer();
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 
   bool init(const LayerMap& layerMap,
             const ParameterMap& parameterMap) override;
-  void forward(PassType passType) override;
-  void backward(const UpdateCallback& callback) override;
-  void addBiases();
-  void bpropBiases();
 };
 
 }  // namespace paddle

paddle/gserver/layers/Layer.cpp

@@ -381,8 +381,7 @@ void Layer::backwardActivation() {
 void Layer::forwardDropOut() {
   auto& outV = getOutputValue();
 
-  if (passType_ == PASS_TRAIN || passType_ == PASS_METRIC_TRAIN ||
-      passType_ == PASS_METRIC_TRAIN_WITH_NOERROR) {
+  if (passType_ == PASS_TRAIN) {
     // new dropOutMask_ if dropOutMask_ is null ptr
     Matrix::resizeOrCreate(dropOutMask_,
                            outV->getHeight(),

paddle/gserver/layers/SequencePoolLayer.cpp

@@ -56,17 +56,16 @@ void SequencePoolLayer::forward(PassType passType) {
   CHECK_EQ(newBatchSize_, starts->getSize() - 1);
 
   resetOutput(newBatchSize_, dim);
-  if (type_) {
-    CHECK(input.subSequenceStartPositions)
-        << "when trans_type = seq, input must hasSubseq";
-  }
+
   /* If type_ = kNonSeq, both seq has or not has sub-seq degrade to a non-seq,
    * thus, in this case, output_ has no sequenceStartPositions.
    * If type_ = kSeq, seq has sub-seq degrades to a seq, thus, only in this
    * case, we should compute the new sequenceStartPositions.
   */
   if (type_) {
-    output_.degradeSequence(input, useGpu_);
+    CHECK(input.subSequenceStartPositions)
+        << "when trans_type = seq, input must hasSubseq";
+    output_.degradeSequence(input);
   }
 }
 

paddle/gserver/tests/test_ConvUnify.cpp

@@ -34,8 +34,7 @@ DECLARE_double(checkgrad_eps);
 DECLARE_bool(thread_local_rand_use_global_seed);
 DECLARE_bool(prev_batch_state);
 
-// Do one forward pass of convTrans layer and check to see if its output
-// matches the given result
+// Do one forward pass of ConvLayer using either exconv or cudnn_conv
 MatrixPtr doOneConvTest(size_t imgSize,
                         size_t output_x,
                         size_t stride,
@@ -46,22 +45,35 @@ MatrixPtr doOneConvTest(size_t imgSize,
                         size_t groups,
                         MatrixPtr& inputData,
                         real* param,
-                        bool useGpu) {
+                        bool useGpu,
+                        bool isDeconv = false) {
   TestConfig config;
   config.biasSize = numfilters;
+  string layerType;
   if (useGpu) {
-    config.layerConfig.set_type("cudnn_conv");
+    layerType = (isDeconv) ? "cudnn_convt" : "cudnn_conv";
   } else {
-    config.layerConfig.set_type("exconv");
+    layerType = (isDeconv) ? "exconvt" : "exconv";
   }
+  config.layerConfig.set_type(layerType);
   config.layerConfig.set_num_filters(numfilters);
   config.layerConfig.set_partial_sum(1);
   config.layerConfig.set_shared_biases(true);
 
   size_t weightSize = channel * filter_size * filter_size *
                       config.layerConfig.num_filters() / groups;
+  if (isDeconv) {
+    config.inputDefs.push_back(
+        {INPUT_DATA, "layer_0", output_x * output_x * channel, weightSize});
+    config.layerConfig.set_size(imgSize * imgSize *
+                                config.layerConfig.num_filters());
+  } else {
     config.inputDefs.push_back(
         {INPUT_DATA, "layer_0", imgSize * imgSize * channel, weightSize});
+    config.layerConfig.set_size(output_x * output_x *
+                                config.layerConfig.num_filters());
+  }
+
   LayerInputConfig* input = config.layerConfig.add_inputs();
   ConvConfig* conv = input->mutable_conv_conf();
   conv->set_filter_size(filter_size);
@@ -72,12 +84,15 @@ MatrixPtr doOneConvTest(size_t imgSize,
   conv->set_stride(stride);
   conv->set_stride_y(stride);
   conv->set_groups(groups);
-  conv->set_filter_channels(channel / groups);
   conv->set_img_size(imgSize);
   conv->set_output_x(output_x);
 
-  config.layerConfig.set_size(conv->output_x() * conv->output_x() *
-                              config.layerConfig.num_filters());
+  if (isDeconv) {
+    conv->set_filter_channels(numfilters / groups);
+  } else {
+    conv->set_filter_channels(channel / groups);
+  }
+
   config.layerConfig.set_name("conv");
 
   std::vector<DataLayerPtr> dataLayers;
@@ -105,6 +120,8 @@ MatrixPtr doOneConvTest(size_t imgSize,
 TEST(Layer, convParaUnified) {
 #ifndef PADDLE_ONLY_CPU
   MatrixPtr input, resultCpu, resultGpu;
+
+  /// TEST1 for conv ///
   input = Matrix::create(1, 4 * 4, false, false);
   real inputData[] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16};
   real param[] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 9, 8, 7, 6, 5, 4, 3, 2, 1};
@@ -121,7 +138,38 @@ TEST(Layer, convParaUnified) {
                             /*groups*/ 1,
                             input,
                             param,
-                            false);
+                            /*useGpu*/ false);
+
+  resultGpu = doOneConvTest(/* imgSize */ 4,
+                            /* output_x */ 2,
+                            /* stride */ 1,
+                            /* padding */ 0,
+                            /* filter_size */ 3,
+                            /*channel*/ 1,
+                            /*numfilters*/ 2,
+                            /*groups*/ 1,
+                            input,
+                            param,
+                            /*useGpu*/ true);
+  checkMatrixEqual(resultCpu, resultGpu);
+
+  /// TEST1 for deconv ///
+  input = Matrix::create(1, 2 * 2, false, false);
+  real inputDataT[] = {1, 2, 3, 4};
+  input->setData(inputDataT);
+
+  resultCpu = doOneConvTest(/* imgSize */ 4,
+                            /* output_x */ 2,
+                            /* stride */ 1,
+                            /* padding */ 0,
+                            /* filter_size */ 3,
+                            /*channel*/ 1,
+                            /*numfilters*/ 2,
+                            /*groups*/ 1,
+                            input,
+                            param,
+                            /*useGpu*/ false,
+                            /*isDeconv*/ true);
 
   resultGpu = doOneConvTest(/* imgSize */ 4,
                             /* output_x */ 2,
@@ -133,9 +181,11 @@ TEST(Layer, convParaUnified) {
                             /*groups*/ 1,
                             input,
                             param,
-                            true);
+                            /*useGpu*/ true,
+                            /*isDeconv*/ true);
   checkMatrixEqual(resultCpu, resultGpu);
 
+  /// TEST2 for conv ///
   input = Matrix::create(1, 3 * 3 * 2, false, false);
   real inputData2[] = {
       1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18};
@@ -153,7 +203,7 @@ TEST(Layer, convParaUnified) {
                             /*groups*/ 1,
                             input,
                             param2,
-                            false);
+                            /*useGpu*/ false);
 
   resultGpu = doOneConvTest(/* imgSize */ 3,
                             /* output_x */ 2,
@@ -165,9 +215,10 @@ TEST(Layer, convParaUnified) {
                             /*groups*/ 1,
                             input,
                             param2,
-                            true);
+                            /*useGpu*/ true);
   checkMatrixEqual(resultCpu, resultGpu);
 
+  /// TEST3 for conv ///
   real param3[] = {1, 2, 3, 4, 4, 3, 2, 1};
 
   resultCpu = doOneConvTest(/* imgSize */ 3,
@@ -180,7 +231,66 @@ TEST(Layer, convParaUnified) {
                             /*groups*/ 2,
                             input,
                             param3,
-                            false);
+                            /*useGpu*/ false);
+
+  resultGpu = doOneConvTest(/* imgSize */ 3,
+                            /* output_x */ 2,
+                            /* stride */ 1,
+                            /* padding */ 0,
+                            /* filter_size */ 2,
+                            /*channel*/ 2,
+                            /*numfilters*/ 2,
+                            /*groups*/ 2,
+                            input,
+                            param3,
+                            /*useGpu*/ true);
+  checkMatrixEqual(resultCpu, resultGpu);
+
+  /// TEST2 for deconv ///
+  input = Matrix::create(1, 2 * 2 * 2, false, false);
+  real inputData2T[] = {1, 2, 3, 4, 5, 6, 7, 8};
+  input->setData(inputData2T);
+
+  resultCpu = doOneConvTest(/* imgSize */ 3,
+                            /* output_x */ 2,
+                            /* stride */ 1,
+                            /* padding */ 0,
+                            /* filter_size */ 2,
+                            /*channel*/ 2,
+                            /*numfilters*/ 2,
+                            /*groups*/ 1,
+                            input,
+                            param2,
+                            /*useGpu*/ false,
+                            /*isDeconv*/ true);
+
+  resultGpu = doOneConvTest(/* imgSize */ 3,
+                            /* output_x */ 2,
+                            /* stride */ 1,
+                            /* padding */ 0,
+                            /* filter_size */ 2,
+                            /*channel*/ 2,
+                            /*numfilters*/ 2,
+                            /*groups*/ 1,
+                            input,
+                            param2,
+                            /*useGpu*/ true,
+                            /*isDeconv*/ true);
+  checkMatrixEqual(resultCpu, resultGpu);
+
+  /// TEST3 for deconv ///
+  resultCpu = doOneConvTest(/* imgSize */ 3,
+                            /* output_x */ 2,
+                            /* stride */ 1,
+                            /* padding */ 0,
+                            /* filter_size */ 2,
+                            /*channel*/ 2,
+                            /*numfilters*/ 2,
+                            /*groups*/ 2,
+                            input,
+                            param3,
+                            /*useGpu*/ false,
+                            /*isDeconv*/ true);
 
   resultGpu = doOneConvTest(/* imgSize */ 3,
                             /* output_x */ 2,
@@ -192,7 +302,8 @@ TEST(Layer, convParaUnified) {
                             /*groups*/ 2,
                             input,
                             param3,
-                            true);
+                            /*useGpu*/ true,
+                            /*isDeconv*/ true);
   checkMatrixEqual(resultCpu, resultGpu);
 #endif
 }

paddle/gserver/tests/test_LayerGrad.cpp

@@ -166,15 +166,19 @@ TEST(Projection, scaling) {
   }
 }
 
-void testProjectionConv(size_t groups) {
+void testProjectionConv(size_t groups, bool isDeconv) {
   const int NUM_FILTERS = 18;
   const int FILTER_SIZE = 2;
-  const int FILTER_SIZE_Y = 3;
+  const int FILTER_SIZE_Y = 4;
   const int CHANNELS = 3;
   const int IMAGE_SIZE = 16;
 
   ProjectionConfig conf;
+  if (isDeconv) {
+    conf.set_type("convt");
+  } else {
     conf.set_type("conv");
+  }
   conf.set_num_filters(NUM_FILTERS);
 
   ConvConfig* conv = conf.mutable_conv_conf();
@@ -186,7 +190,11 @@ void testProjectionConv(size_t groups) {
   conv->set_stride(2);
   conv->set_stride_y(2);
   conv->set_groups(groups);
+  if (isDeconv) {
+    conv->set_filter_channels(NUM_FILTERS / conv->groups());
+  } else {
     conv->set_filter_channels(conv->channels() / conv->groups());
+  }
   conv->set_img_size(IMAGE_SIZE);
   int output_x = outputSize(conv->img_size(),
                             conv->filter_size(),
@@ -199,8 +207,14 @@ void testProjectionConv(size_t groups) {
                             conv->stride_y(),
                             /* caffeMode */ true);
   conv->set_output_x(output_x);
+  conv->set_output_y(output_y);
+  if (isDeconv) {
+    conf.set_input_size(output_x * output_y * CHANNELS);
+    conf.set_output_size(IMAGE_SIZE * IMAGE_SIZE * NUM_FILTERS);
+  } else {
     conf.set_input_size(IMAGE_SIZE * IMAGE_SIZE * CHANNELS);
     conf.set_output_size(output_x * output_y * NUM_FILTERS);
+  }
 
   testProjectionGrad(conf,
                      INPUT_DATA,
@@ -215,8 +229,12 @@ void testProjectionConv(size_t groups) {
 
 #ifndef PADDLE_ONLY_CPU
 TEST(Projection, conv) {
-  testProjectionConv(1);
-  testProjectionConv(3);
+  /// test ConvProjection
+  testProjectionConv(1, false);
+  testProjectionConv(3, false);
+  /// test ConvTransProjection
+  testProjectionConv(1, true);
+  testProjectionConv(3, true);
 }
 #endif
 
@@ -385,11 +403,11 @@ void testConvTransLayer(const string& type, bool trans, bool useGpu) {
   config.layerConfig.set_partial_sum(1);
   config.layerConfig.set_shared_biases(true);
 
-  config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 288});
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 384});
   LayerInputConfig* input = config.layerConfig.add_inputs();
   ConvConfig* conv = input->mutable_conv_conf();
   conv->set_filter_size(2);
-  conv->set_filter_size_y(3);
+  conv->set_filter_size_y(4);
   conv->set_channels(16);
   conv->set_padding(0);
   conv->set_padding_y(1);
@@ -416,6 +434,9 @@ TEST(Layer, convTransLayer) {
   for (auto useGpu : {false, true}) {
     testConvTransLayer("exconvt", /* trans= */ false, /* useGpu= */ useGpu);
   }
+#ifndef PADDLE_ONLY_CPU
+  testConvTransLayer("cudnn_convt", /* trans= */ false, /* useGpu= */ true);
+#endif
 }
 
 TEST(Layer, blockExpandLayer) {
@@ -1482,16 +1503,20 @@ TEST(Layer, BatchNormalizationLayer) {
 #endif
 }
 
-TEST(Operator, conv) {
+void testConvOperator(bool isDeconv) {
   TestConfig config;
   const int NUM_FILTERS = 16;
   const int FILTER_SIZE = 2;
   const int FILTER_SIZE_Y = 3;
   const int CHANNELS = 3;
   const int IMAGE_SIZE = 16;
-  const int IMAGE_SIZE_Y = 8;
+  const int IMAGE_SIZE_Y = 9;
   OperatorConfig& operatorConf = *config.layerConfig.add_operator_confs();
+  if (isDeconv) {
+    operatorConf.set_type("convt");
+  } else {
     operatorConf.set_type("conv");
+  }
   ConvConfig* conv = operatorConf.mutable_conv_conf();
   operatorConf.set_num_filters(NUM_FILTERS);
   conv->set_filter_size(FILTER_SIZE);
@@ -1502,7 +1527,6 @@ TEST(Operator, conv) {
   conv->set_stride(2);
   conv->set_stride_y(2);
   conv->set_groups(1);
-  conv->set_filter_channels(conv->channels() / conv->groups());
   conv->set_img_size(IMAGE_SIZE);
   conv->set_img_size_y(IMAGE_SIZE_Y);
   conv->set_output_x(outputSize(conv->img_size(),
@@ -1515,11 +1539,22 @@ TEST(Operator, conv) {
                                 conv->padding_y(),
                                 conv->stride_y(),
                                 /*  caffeMode */ true));
-  config.layerConfig.set_size(conv->output_x() * conv->output_y() *
-                              NUM_FILTERS);
 
+  if (isDeconv) {
+    conv->set_filter_channels(NUM_FILTERS / conv->groups());
+    config.inputDefs.push_back({INPUT_DATA,
+                                "layer_0",
+                                conv->output_x() * conv->output_y() * CHANNELS,
+                                0});
+    config.layerConfig.set_size(IMAGE_SIZE * IMAGE_SIZE_Y * NUM_FILTERS);
+  } else {
+    conv->set_filter_channels(conv->channels() / conv->groups());
     config.inputDefs.push_back(
         {INPUT_DATA, "layer_0", IMAGE_SIZE * IMAGE_SIZE_Y * CHANNELS, 0});
+    config.layerConfig.set_size(conv->output_x() * conv->output_y() *
+                                NUM_FILTERS);
+  }
+
   config.inputDefs.push_back(
       {INPUT_DATA,
        "layer_1",
@@ -1531,6 +1566,11 @@ TEST(Operator, conv) {
   testOperatorGrad(config, operatorConf, 100, /*useGpu*/ true, false);
 }
 
+TEST(Operator, conv) {
+  testConvOperator(/*isDeconv*/ true);
+  testConvOperator(/*isDeconv*/ false);
+}
+
 TEST(Layer, FeatureMapExpandLayer) {
   TestConfig config;
   config.layerConfig.set_type("featmap_expand");
@@ -1602,6 +1642,20 @@ TEST(Layer, PadLayer) {
   }
 }
 
+TEST(Layer, smooth_l1) {
+  TestConfig config;
+  config.layerConfig.set_type("smooth_l1");
+
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0});
+  config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 1, 0});
+  config.layerConfig.add_inputs();
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "smooth_l1", 100, false, useGpu, false, 2.0);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/math/Matrix.cpp

@@ -3557,6 +3557,55 @@ void CpuMatrix::sumOfSquaresBp(Matrix& output, Matrix& label) {
   }
 }
 
+void CpuMatrix::smoothL1(Matrix& output, Matrix& label) {
+  CHECK(output.useGpu_ == false && label.useGpu_ == false)
+      << "Matrix type are not equal";
+
+  size_t numSamples = getHeight();
+  size_t dim = output.getWidth();
+  CHECK_EQ(label.getHeight(), numSamples);
+  CHECK_EQ(output.getHeight(), numSamples);
+  CHECK_EQ(label.getWidth(), dim);
+  CHECK_EQ(getWidth(), (size_t)1);
+  real* out = output.getData();
+  real* cost = getData();
+  real* lbl = label.getData();
+
+  for (size_t i = 0; i < numSamples; ++i, out += dim, cost += dim, lbl += dim) {
+    for (size_t j = 0; j < dim; ++j) {
+      cost[j] = std::fabs(out[j] - lbl[j]);
+      if (cost[j] < 1.0)
+        cost[j] = 0.5 * cost[j] * cost[j];
+      else
+        cost[j] = cost[j] - 0.5;
+    }
+  }
+}
+
+void CpuMatrix::smoothL1Bp(Matrix& output, Matrix& label) {
+  CHECK(output.useGpu_ == false && label.useGpu_ == false)
+      << "Matrix type are not equal";
+
+  size_t numSamples = getHeight();
+  size_t dim = output.getWidth();
+  CHECK_EQ(label.getHeight(), numSamples);
+  CHECK_EQ(output.getHeight(), numSamples);
+  CHECK_EQ(label.getWidth(), dim);
+  CHECK_EQ(getWidth(), (size_t)1);
+  real* out = output.getData();
+  real* cost = getData();
+  real* lbl = label.getData();
+
+  // f'(x) = x         if |x| < 1
+  //       = sign(x)   otherwise
+  for (size_t i = 0; i < numSamples; ++i, out += dim, cost += dim, lbl += dim) {
+    for (size_t j = 0; j < dim; ++j) {
+      cost[j] = out[j] - lbl[j];
+      if (std::fabs(cost[j]) >= 1) cost[j] = (0 < cost[j]) - (cost[j] < 0);
+    }
+  }
+}
+
 void CpuMatrix::tanh(Matrix& output) {
   CHECK(isContiguous());
   CHECK(output.isContiguous());

paddle/math/Matrix.h

@@ -783,6 +783,14 @@ public:
     LOG(FATAL) << "Not implemented";
   }
 
+  virtual void smoothL1(Matrix& output, Matrix& label) {
+    LOG(FATAL) << "Not implemented";
+  }
+
+  virtual void smoothL1Bp(Matrix& outputV, Matrix& label) {
+    LOG(FATAL) << "Not implemented";
+  }
+
   virtual void tanh(Matrix& output) { LOG(FATAL) << "Not implemented"; }
 
   virtual void tanhDerivative(Matrix& output) {
@@ -1720,6 +1728,9 @@ public:
   /// gradient of sumOfSquares.
   void sumOfSquaresBp(Matrix& outputV, Matrix& label);
 
+  void smoothL1(Matrix& output, Matrix& label);
+  void smoothL1Bp(Matrix& output, Matrix& label);
+
   void tanh(Matrix& output);
   void tanhDerivative(Matrix& output);
 

paddle/parameter/Argument.cpp

@@ -123,46 +123,6 @@ static void resizeAndCopy(ICpuGpuVectorPtr& dest,
   }
 }
 
-static void resizeAndCopy(UserDefinedVectorPtr& dest,
-                          const UserDefinedVectorPtr& src,
-                          bool useGpu,
-                          hl_stream_t stream) {
-  if (src) {
-    CHECK(!useGpu) << "not implemented";
-    size_t height = src->size();
-    if (!dest) {
-      dest = std::make_shared<std::vector<void*>>(height);
-    } else {
-      dest->resize(height);
-    }
-    std::copy_n(src->begin(), height, dest->begin());
-  } else {
-    dest.reset();
-  }
-}
-
-static void resizeAndCopy(UserDefinedVectorPtr& dest,
-                          const UserDefinedVectorPtr& src,
-                          int32_t startPos,
-                          int32_t copySize,
-                          bool useGpu,
-                          hl_stream_t stream = HPPL_STREAM_DEFAULT) {
-  if (src) {
-    CHECK(!useGpu) << "not implemented";
-    CHECK_LE((size_t)startPos + copySize, src->size());
-
-    size_t height = copySize;
-    if (!dest) {
-      dest = std::make_shared<std::vector<void*>>(height);
-    } else {
-      dest->resize(height);
-    }
-    std::copy_n(src->begin() + startPos, height, dest->begin());
-  } else {
-    dest.reset();
-  }
-}
-
 static void resizeAndCopy(SVectorPtr& dest,
                           const SVectorPtr& src,
                           bool useGpu,
@@ -223,7 +183,6 @@ void Argument::resizeAndCopyFrom(const Argument& src,
                   false /* useGpu */,
                   stream);
   }
-  resizeAndCopy(udp, src.udp, useGpu, stream);
   resizeAndCopy(strs, src.strs, useGpu, stream);
   frameWidth = src.frameWidth;
   frameHeight = src.frameHeight;
@@ -255,7 +214,6 @@ int32_t Argument::resizeAndCopyFrom(const Argument& src,
     resizeAndCopy(value, src.value, startRow, copySize, useGpu, stream);
     resizeAndCopy(grad, src.grad, startRow, copySize, useGpu, stream);
     resizeAndCopy(ids, src.ids, startRow, copySize, useGpu, stream);
-    resizeAndCopy(udp, src.udp, startRow, copySize, useGpu, stream);
     resizeAndCopy(strs, src.strs, startRow, copySize, useGpu, stream);
     return copySize;
   } else {
@@ -268,7 +226,6 @@ int32_t Argument::resizeAndCopyFrom(const Argument& src,
     resizeAndCopy(value, src.value, startRow, copyFeatureSize, useGpu, stream);
     resizeAndCopy(grad, src.grad, startRow, copyFeatureSize, useGpu, stream);
     resizeAndCopy(ids, src.ids, startRow, copyFeatureSize, useGpu, stream);
-    resizeAndCopy(udp, src.udp, startRow, copySize, useGpu, stream);
     resizeAndCopy(sequenceStartPositions,
                   src.sequenceStartPositions,
                   startSeq,
@@ -583,7 +540,7 @@ void Argument::checkSubset() const {
   }
 }
 
-void Argument::degradeSequence(const Argument& input, bool useGpu) {
+void Argument::degradeSequence(const Argument& input) {
   CHECK_EQ(input.hasSubseq(), 1UL);
   size_t numSequences = input.getNumSequences();
   size_t numSubSequences = input.getNumSubSequences();

paddle/parameter/Argument.h

@@ -24,8 +24,6 @@ limitations under the License. */
 
 namespace paddle {
 
-// vector of user defined pointers
-typedef std::shared_ptr<std::vector<void*>> UserDefinedVectorPtr;
 typedef std::shared_ptr<std::vector<std::string>> SVectorPtr;
 
 struct Argument {
@@ -40,7 +38,6 @@ struct Argument {
         sequenceStartPositions(nullptr),
         subSequenceStartPositions(nullptr),
         cpuSequenceDims(nullptr),
-        udp(nullptr),
         deviceId(-1),
         allCount(0),
         valueCount(0),
@@ -63,7 +60,6 @@ struct Argument {
     sequenceStartPositions = argument.sequenceStartPositions;
     subSequenceStartPositions = argument.subSequenceStartPositions;
     cpuSequenceDims = argument.cpuSequenceDims;
-    udp = argument.udp;
     deviceId = argument.deviceId;
     allCount = argument.allCount;
     frameHeight = argument.frameHeight;
@@ -96,8 +92,6 @@ struct Argument {
   // dimension of sequence, stored only in CPU
   IVectorPtr cpuSequenceDims;
 
-  UserDefinedVectorPtr udp;  // user defined pointer
-
   int deviceId;            // the GPU device id which the argument in
   int allCount;            // the number of output layers using this argument
   mutable int valueCount;  // waiting this member when layer do forward
@@ -137,7 +131,6 @@ struct Argument {
     if (ids) return ids->getSize();
     if (grad) return grad->getHeight();
     if (in) return in->getHeight();
-    if (udp) return udp->size();
     if (strs) return strs->size();
     return 0;
   }
@@ -296,7 +289,7 @@ struct Argument {
   /*
    sequence has sub-sequence degrades to a sequence.
    */
-  void degradeSequence(const Argument& input, bool useGpu);
+  void degradeSequence(const Argument& input);
 
   /**
    * @brief getValueString will return the argument's output in string. There

paddle/pserver/BaseClient.h

@@ -30,9 +30,6 @@ namespace paddle {
  * the first solution arms with sendThreads_/recvThreads_ and sendJobQueue_/
  * recvJobQueue_. the second solution use some shared thread pool to manage
  * connections.
- * In addition to pserver, metric learning also uses network to exchange
- * features within multi-machines, so this class just abstracts some basic
- * threads and queue buffer creation for them
  */
 class BaseClient {
 protected:

paddle/pserver/ParameterServer2.cpp

@@ -369,11 +369,8 @@ void ParameterServer2::addGradient(const SendParameterRequest& request,
                                    std::vector<Buffer>* outputBuffers) {
   VLOG(1) << "pserver: addGradient";
 
-/// forwardbackward delta from all trainers
-/// indicate the fluctuation caused by forwardbackward.
-#ifndef PADDLE_METRIC_LEARNING
-  // @TODO(yanfei):
-  // add support tuning forwardbackward balance for metric learning
+  // forwardbackward delta from all trainers
+  // indicate the fluctuation caused by forwardbackward.
   if (!numPassFinishClients_) {
     REGISTER_BARRIER_DELTA_SERVER_SET(
         *statSet_,
@@ -383,7 +380,6 @@ void ParameterServer2::addGradient(const SendParameterRequest& request,
         request.forwardbackward_time(),
         isSparseServer_ ? "_sparseUpdater" : "_denseUpdater");
   }
-#endif
 
   {
     /// approximately pure network overhead

paddle/py_paddle/.gitignore

 swig_paddle.py
+_swig_paddle.so

paddle/py_paddle/dataprovider_converter.py

@@ -16,11 +16,25 @@ import paddle.trainer.PyDataProvider2 as dp2
 import collections
 import swig_paddle
 import numpy
+import itertools
 
 __all__ = ['DataProviderConverter']
 
 
 class IScanner(object):
+    """
+    The scanner will scan Python object two passes, then convert it to Paddle's
+    argument.
+
+    In the first pass, `pre_scan` will be invoked by every data instance, and
+    then invoke `finish_pre_scan` to arguments. And the second pass do the same
+    thing except the functions changed to `scan`, `finish_scan`.
+
+    During the first pass, a scanner may count the shape of input matrix and
+    allocate memory for this argument. Then fill the data into this  argument
+    in second pass.
+    """
+
     def __init__(self, input_type, pos):
         self.input_type = input_type
         if not isinstance(self.input_type, dp2.InputType):
@@ -36,10 +50,40 @@ class IScanner(object):
         self.data_in_gpu = swig_paddle.isUsingGpu(
         ) and swig_paddle.getTrainerCount() == 1
 
+    def pre_scan(self, dat):
+        """
+        First pass scan method. During this method, the scanner could count the
+        data number, and get the total memory size this batch would use.
+
+        :param dat: The python object.
+        """
+        pass
+
+    def finish_pre_scan(self, argument):
+        """
+        Finish first scan pass. Allocate the memory.
+
+        :param argument: Output arguments object.
+        :type argument: swig_paddle.Arguments
+        :return:
+        """
+        pass
+
     def scan(self, dat):
+        """
+        Second pass scan method. Copy the data to arguments.
+
+        :param dat: The python object.
+        """
         pass
 
     def finish_scan(self, argument):
+        """
+        Finish second pass. Finalize the resources, etc.
+
+        :param argument: Output arguments object.
+        :type argument: swig_paddle.Arguments
+        """
         pass
 
 
@@ -51,12 +95,19 @@ class DenseScanner(IScanner):
     def __init__(self, input_type, pos):
         IScanner.__init__(self, input_type, pos)
         self.__mat__ = None
+        self.__height__ = 0
+
+    def pre_scan(self, dat):
+        self.__height__ += 1
+
+    def finish_pre_scan(self, argument):
+        self.__mat__ = numpy.ndarray(
+            shape=(self.__height__, self.input_type.dim), dtype=numpy.float32)
+        self.__height__ = 0
 
     def scan(self, dat):
-        if self.__mat__ is None:
-            self.__mat__ = numpy.array([dat], dtype='float32')
-        else:
-            self.__mat__ = numpy.append(self.__mat__, [dat], axis=0)
+        self.__mat__[self.__height__] = dat
+        self.__height__ += 1
 
     def finish_scan(self, argument):
         assert isinstance(argument, swig_paddle.Arguments)
@@ -163,7 +214,14 @@ class DataProviderConverter(object):
         ]
 
         for each_sample in dat:
-            for each_step, scanner in zip(each_sample, scanners):
+            for each_step, scanner in itertools.izip(each_sample, scanners):
+                scanner.pre_scan(each_step)
+
+        for scanner in scanners:
+            scanner.finish_pre_scan(argument)
+
+        for each_sample in dat:
+            for each_step, scanner in itertools.izip(each_sample, scanners):
                 scanner.scan(each_step)
 
         for scanner in scanners:

paddle/scripts/deb/build_scripts/.gitignore

-dist

paddle/scripts/deb/build_scripts/Dockerfile

-FROM paddledev/paddle:gpu-latest
-MAINTAINER PaddlePaddle Dev Team <paddle-dev@baidu.com>
-COPY build.sh /root/
-CMD cd /root/ && bash build.sh
-

paddle/scripts/deb/build_scripts/build.sh

-#!/bin/bash
-set -e
-apt-get update
-apt-get install -y dh-make
-cd ~
-mkdir -p ~/dist/gpu
-mkdir -p ~/dist/cpu
-mkdir -p ~/dist/cpu-noavx
-mkdir -p ~/dist/gpu-noavx
-cd paddle
-mkdir build
-cd build
-cmake .. -DWITH_GPU=OFF -DWITH_SWIG_PY=ON -DWITH_AVX=ON
-make -j `nproc`
-cpack -D CPACK_GENERATOR='DEB' ..
-mv *.deb ~/dist/cpu
-
-rm -rf *
-cmake .. -DWITH_GPU=ON -DWITH_SWIG_PY=ON -DWITH_AVX=ON -DCUDNN_ROOT=/usr/
-make -j `nproc`
-cpack -D CPACK_GENERATOR='DEB' ..
-mv *.deb ~/dist/gpu
-
-
-rm -rf *
-cmake .. -DWITH_GPU=OFF -DWITH_SWIG_PY=ON -DWITH_AVX=OFF
-make -j `nproc`
-cpack -D CPACK_GENERATOR='DEB' ..
-mv *.deb ~/dist/cpu-noavx
-
-rm -rf *
-cmake .. -DWITH_GPU=ON -DWITH_SWIG_PY=ON -DWITH_AVX=OFF -DCUDNN_ROOT=/usr/
-make -j `nproc`
-cpack -D CPACK_GENERATOR='DEB' ..
-mv *.deb ~/dist/gpu-noavx

paddle/scripts/deb/build_scripts/build_deb.sh

-#!/bin/bash
-set -e
-docker build -t build_paddle_deb .
-rm -rf dist
-mkdir -p dist
-docker run -v$PWD/dist:/root/dist -v $PWD/../../../..:/root/paddle --name tmp_build_deb_container build_paddle_deb
-docker rm tmp_build_deb_container
-docker rmi build_paddle_deb

paddle/scripts/deb/postinst

-#!/bin/bash
-set -e
-echo "Post install paddle debian package."
-echo "Install some python package used for paddle. You can run "
-echo "  pip install /usr/opt/paddle/share/wheels/*.whl to install them."
-pip install /usr/opt/paddle/share/wheels/*.whl
-

paddle/scripts/docker/Dockerfile

-FROM ubuntu:14.04
-MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
-
-ARG DEBIAN_FRONTEND=noninteractive
-ARG UBUNTU_MIRROR
-RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi'
-
-# ENV variables
-ARG BUILD_WOBOQ
-ARG BUILD_AND_INSTALL
-ARG WITH_AVX
-ARG WITH_DOC
-ARG WITH_STYLE_CHECK
-
-ENV BUILD_WOBOQ=${BUILD_WOBOQ:-OFF}
-ENV BUILD_AND_INSTALL=${BUILD_AND_INSTALL:-OFF}
-ENV WITH_GPU=OFF
-ENV WITH_AVX=${WITH_AVX:-ON}
-ENV WITH_DOC=${WITH_DOC:-OFF}
-ENV WITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF}
-ENV DOCKER_BUILD=TRUE
-
-ENV HOME /root
-
-# Add bash enhancements
-COPY ./paddle/scripts/docker/root/ /root/
-
-RUN apt-get update && \
-    apt-get install -y git python-pip python-dev openssh-server bison && \
-    apt-get install -y wget unzip tar xz-utils bzip2 gzip coreutils && \
-    apt-get install -y curl sed grep graphviz libjpeg-dev zlib1g-dev && \
-    apt-get install -y python-numpy python-matplotlib gcc g++ gfortran && \
-    apt-get install -y automake locales clang-format-3.8 && \
-    apt-get clean -y
-
-# git credential to skip password typing
-RUN git config --global credential.helper store
-
-# Fix locales to en_US.UTF-8
-RUN localedef -i en_US -f UTF-8 en_US.UTF-8
-
-RUN pip install --upgrade pip && \
-    pip install -U 'protobuf==3.1.0' && \
-    pip install -U wheel pillow BeautifulSoup && \
-    pip install -U docopt PyYAML sphinx && \
-    pip install -U sphinx-rtd-theme==0.1.9 recommonmark && \
-    pip install -U pre-commit 'requests==2.9.2' jupyter
-
-RUN curl -sSL https://cmake.org/files/v3.4/cmake-3.4.1.tar.gz | tar -xz && \
-    cd cmake-3.4.1 && ./bootstrap && make -j `nproc` && make install && \
-    cd .. && rm -rf cmake-3.4.1
-
-COPY . /paddle/
-RUN cd /paddle/ && git submodule update --init --recursive
-RUN /paddle/paddle/scripts/docker/build.sh
-
-VOLUME ["/usr/share/nginx/html/data", "/usr/share/nginx/html/paddle"]
-
-# Configure OpenSSH server. c.f. https://docs.docker.com/engine/examples/running_ssh_service
-RUN mkdir /var/run/sshd
-RUN echo 'root:root' | chpasswd
-RUN sed -ri 's/^PermitRootLogin\s+.*/PermitRootLogin yes/' /etc/ssh/sshd_config
-RUN sed -ri 's/UsePAM yes/#UsePAM yes/g' /etc/ssh/sshd_config
-EXPOSE 22
-
-# Jupyter Notebook: Paddle book
-EXPOSE 8888
-
-COPY ./paddle/scripts/docker/entrypoint /opt/bin/
-
-CMD ["/opt/bin/entrypoint"]

paddle/scripts/docker/README.md

+# Building PaddlePaddle
+
+## Goals
+
+We want the building procedure generates Docker images so that we can run PaddlePaddle applications on Kubernetes clusters.
+
+We want to build .deb packages so that enterprise users can run PaddlePaddle applications without Docker.
+
+We want to minimize the size of generated Docker images and .deb packages so to reduce the download time.
+
+We want to encapsulate building tools and dependencies in a *development* Docker image so to ease the tools installation for developers.
+
+Developers use various editors (emacs, vim, Eclipse, Jupyter Notebook), so the development Docker image contains only building tools, not editing tools, and developers are supposed to git clone source code into their development computers and map the code into the development container.
+
+We want the procedure and tools also work with testing, continuous integration, and releasing.
+
+
+## Docker Images
+
+So we need two Docker images for each version of PaddlePaddle:
+
+1. `paddle:<version>-dev`
+
+   This a development image contains only the development tools and standardizes the building procedure.  Users include:
+
+   - developers -- no longer need to install development tools on the host, and can build their current work on the host (development computer).
+   - release engineers -- use this to build the official release from certain branch/tag on Github.com.
+   - document writers / Website developers -- Our documents are in the source repo in the form of .md/.rst files and comments in source code.  We need tools to extract the information, typeset, and generate Web pages.
+
+   Of course, developers can install building tools on their development computers.  But different versions of PaddlePaddle might require different set or version of building tools.  Also, it makes collaborative debugging easier if all developers use a unified development environment.
+
+  The development image should include the following tools:
+
+   - gcc/clang
+   - nvcc
+   - Python
+   - sphinx
+   - woboq
+   - sshd
+
+   Many developers work on a remote computer with GPU; they could ssh into the computer and  `docker exec` into the development container. However, running `sshd` in the container allows developers to ssh into the container directly.
+
+1. `paddle:<version>`
+
+   This is the production image, generated using the development image. This image might have multiple variants:
+
+   - GPU/AVX   `paddle:<version>-gpu`
+   - GPU/no-AVX  `paddle:<version>-gpu-noavx`
+   - no-GPU/AVX  `paddle:<version>`
+   - no-GPU/no-AVX  `paddle:<version>-noavx`
+
+   We allow users to choose between GPU and no-GPU because the GPU version image is much larger than then the no-GPU version.
+
+   We allow users the choice between AVX and no-AVX, because some cloud providers don't provide AVX-enabled VMs.
+
+
+## Development Environment
+
+Here we describe how to use above two images.  We start from considering our daily development environment.
+
+Developers work on a computer, which is usually a laptop or desktop:
+
+<img src="doc/paddle-development-environment.png" width=500 />
+
+or, they might rely on a more sophisticated box (like with GPUs):
+
+<img src="doc/paddle-development-environment-gpu.png" width=500 />
+
+A principle here is that source code lies on the development computer (host) so that editors like Eclipse can parse the source code to support auto-completion.
+
+
+## Usages
+
+### Build the Development Docker Image
+
+The following commands check out the source code to the host and build the development image `paddle:dev`:
+
+```bash
+git clone https://github.com/PaddlePaddle/Paddle paddle
+cd paddle
+docker build -t paddle:dev .
+```
+
+The `docker build` command assumes that `Dockerfile` is in the root source tree.  Note that in this design, this `Dockerfile` is this only one in our repo.
+
+Users can specify a Ubuntu mirror server for faster downloading:
+
+```bash
+docker build -t paddle:dev --build-arg UBUNTU_MIRROR=mirror://mirrors.ubuntu.com/mirrors.txt .
+```
+
+### Build PaddlePaddle from Source Code
+
+Given the development image `paddle:dev`, the following command builds PaddlePaddle from the source tree on the development computer (host):
+
+```bash
+docker run -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "TEST=OFF" paddle:dev
+```
+
+This command mounts the source directory on the host into `/paddle` in the container, so the default entry point of `paddle:dev`, `build.sh`, could build the source code with possible local changes.  When it writes to `/paddle/build` in the container, it writes to `$PWD/build` on the host indeed.
+
+`build.sh` builds the following:
+
+- PaddlePaddle binaries,
+- `$PWD/build/paddle-<version>.deb` for production installation, and
+- `$PWD/build/Dockerfile`, which builds the production Docker image.
+
+Users can specify the following Docker build arguments with either "ON" or "OFF" value:
+- `WITH_GPU`: ***Required***. Generates NVIDIA CUDA GPU code and relies on CUDA libraries.
+- `WITH_AVX`: ***Required***. Set to "OFF" prevents from generating AVX instructions. If you don't know what is AVX, you might want to set "ON".
+- `TEST`: ***Optional, default OFF***. Build unit tests and run them after building.
+
+### Build the Production Docker Image
+
+The following command builds the production image:
+
+```bash
+docker build -t paddle -f build/Dockerfile .
+```
+
+This production image is minimal -- it includes binary `paddle`, the shared library `libpaddle.so`, and Python runtime.
+
+### Run PaddlePaddle Applications
+
+Again the development happens on the host.  Suppose that we have a simple application program in `a.py`, we can test and run it using the production image:
+
+```bash
+docker run -it -v $PWD:/work paddle /work/a.py
+```
+
+But this works only if all dependencies of `a.py` are in the production image. If this is not the case, we need to build a new Docker image from the production image and with more dependencies installs.
+
+### Build and Run PaddlePaddle Applications
+
+We need a Dockerfile in https://github.com/paddlepaddle/book that builds Docker image `paddlepaddle/book:<version>`, basing on the PaddlePaddle production image:
+
+```
+FROM paddlepaddle/paddle:<version>
+RUN pip install -U matplotlib jupyter ...
+COPY . /book
+EXPOSE 8080
+CMD ["jupyter"]
+```
+
+The book image is an example of PaddlePaddle application image.  We can build it
+
+```bash
+git clone https://github.com/paddlepaddle/book
+cd book
+docker build -t book .
+```
+
+### Build and Run Distributed Applications
+
+In our [API design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/api.md#distributed-training), we proposed an API that starts a distributed training job on a cluster.  This API need to build a PaddlePaddle application into a Docker image as above and calls kubectl to run it on the cluster.  This API might need to generate a Dockerfile look like above and call `docker build`.
+
+Of course, we can manually build an application image and launch the job using the kubectl tool:
+
+```bash
+docker build -f some/Dockerfile -t myapp .
+docker tag myapp me/myapp
+docker push
+kubectl ...
+```

paddle/scripts/docker/build.sh

 #!/bin/bash
 
-function abort(){
-    echo "An error occurred. Exiting..." 1>&2
-    exit 1
-}
-
-trap 'abort' 0
 set -e
 
-# If Dockerfile.* sets BUILD_AND_INSTALL to 'ON', it would have copied
-# source tree to /paddle, and this scripts should build it into
-# /paddle/build.
-if [[ ${BUILD_AND_INSTALL:-OFF} == 'ON' ]]; then
-    if [[ ${WITH_GPU:-OFF} == 'ON' ]]; then
+# Set BASE_IMAGE according to env variables
+if [ ${WITH_GPU} == "ON" ]; then
+  BASE_IMAGE="nvidia/cuda:7.5-cudnn5-runtime-ubuntu14.04"
+  # additional packages to install when building gpu images
+  GPU_DOCKER_PKG="python-pip python-dev"
+else
+  BASE_IMAGE="python:2.7.13-slim"
+fi
+
+DOCKERFILE_GPU_ENV=""
+if [[ ${WITH_GPU:-OFF} == 'ON' ]]; then
+    DOCKERFILE_GPU_ENV="ENV LD_LIBRARY_PATH /usr/lib/x86_64-linux-gnu:${LD_LIBRARY_PATH}"
+
+    # for cmake to find cudnn
     ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so /usr/lib/libcudnn.so
-    fi
+fi
+
+mkdir -p /paddle/build
+cd /paddle/build
 
-    mkdir -p /paddle/build # -p means no error if exists
-    # clean local cmake and third_party cache
-    cd /paddle/build && rm -rf * && rm -rf ../third_party
-    cmake .. \
+# build script will not fail if *.deb does not exist
+rm *.deb 2>/dev/null || true
+
+cmake .. \
+      -DCMAKE_BUILD_TYPE=Release \
       -DWITH_DOC=${WITH_DOC:-OFF} \
       -DWITH_GPU=${WITH_GPU:-OFF} \
       -DWITH_AVX=${WITH_AVX:-OFF} \
       -DWITH_SWIG_PY=ON \
       -DCUDNN_ROOT=/usr/ \
-	  -DWITH_STYLE_CHECK=OFF \
+      -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} \
+      -DWITH_COVERAGE=${TEST:-OFF} \
       -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
-    make -j `nproc`
-    make install
+make -j `nproc`
+if [[ ${TEST:-OFF} == "ON" ]]; then
+    make coveralls
+fi
+make install
 
-    if [[ ${BUILD_WOBOQ:-OFF} == 'ON' ]]; then
+# generate deb package for current build
+# FIXME(typhoonzero): should we remove paddle/scripts/deb ?
+# FIXME: CPACK_DEBIAN_PACKAGE_DEPENDS removes all dev dependencies, must
+# install them in docker
+cpack -D CPACK_GENERATOR='DEB' -D CPACK_DEBIAN_PACKAGE_DEPENDS="" ..
+
+if [[ ${BUILD_WOBOQ:-OFF} == 'ON' ]]; then
     apt-get install -y clang-3.8 llvm-3.8 libclang-3.8-dev
     # Install woboq_codebrowser.
     git clone https://github.com/woboq/woboq_codebrowser /woboq
@@ -52,17 +69,34 @@ if [[ ${BUILD_AND_INSTALL:-OFF} == 'ON' ]]; then
     /woboq/indexgenerator/codebrowser_indexgenerator $WOBOQ_OUT
     cd /woboq
     make clean
-    fi
+fi
 
-    pip install /usr/local/opt/paddle/share/wheels/py_paddle*linux*.whl
-    pip install /usr/local/opt/paddle/share/wheels/paddle*.whl
-    paddle version
+paddle version
 
-    if [[ ${DOCKER_BUILD:-FALSE} == 'TRUE' ]]; then
-	# reduce docker image size
-	rm -rf /paddle/build
-	rm -rf /usr/local/opt/paddle/share/wheels/
-    fi
+# generate production docker image Dockerfile
+if [ ${USE_MIRROR} ]; then
+  MIRROR_UPDATE="sed 's@http:\/\/archive.ubuntu.com\/ubuntu\/@mirror:\/\/mirrors.ubuntu.com\/mirrors.txt@' -i /etc/apt/sources.list && \\"
+else
+  MIRROR_UPDATE="\\"
 fi
 
-trap : 0
+cat > /paddle/build/Dockerfile <<EOF
+FROM ${BASE_IMAGE}
+MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
+ENV HOME /root
+ENV LANG en_US.UTF-8
+# Use Fix locales to en_US.UTF-8
+RUN ${MIRROR_UPDATE}
+    apt-get update && \
+    apt-get install -y libgfortran3 libpython2.7 ${GPU_DOCKER_PKG} && \
+    apt-get clean -y && \
+    pip install --upgrade pip && \
+    pip install -U 'protobuf==3.1.0' requests numpy
+# Use different deb file when building different type of images
+ADD build/*.deb /usr/local/opt/paddle/deb/
+# run paddle version to install python packages first
+RUN dpkg -i /usr/local/opt/paddle/deb/*.deb && rm -f /usr/local/opt/paddle/deb/*.deb && paddle version
+${DOCKERFILE_GPU_ENV}
+# default command shows the paddle version and exit
+CMD ["paddle", "version"]
+EOF

paddle/scripts/submit_local.sh.in

@@ -95,8 +95,14 @@ EOF
 
 if [ $? -eq 1 ]; then  # Older version installed, or not installed at all
     echo "First time run paddle, need to install some python dependencies."
+    # setuptools normalizes package version, so we need to use normalized
+    # package version for paddle python package
+    PYTHON_PADDLE_VERSION=$(python -c 'import packaging
+import setuptools
+print str(packaging.version.Version("@PADDLE_VERSION@"))
+' 2>/dev/null)
     BASEDIR=$(dirname "$0")
-   pip install ${BASEDIR}/../opt/paddle/share/wheels/*-@PADDLE_VERSION@-*.whl
+    pip install ${BASEDIR}/../opt/paddle/share/wheels/*-${PYTHON_PADDLE_VERSION}-*.whl
     if [ $? -ne 0 ]; then
 	echo "pip install wheels failed. "
 	echo "Please use 'sudo paddle' at the first time you use PaddlePaddle"

paddle/scripts/travis/build_and_test.sh

@@ -5,7 +5,7 @@ NPROC=1
 export PYTHONPATH=/opt/python/2.7.12/lib/python2.7/site-packages
 export PYTHONHOME=/opt/python/2.7.12
 export PATH=/opt/python/2.7.12/bin:${PATH}
-cmake .. -DCMAKE_Fortran_COMPILER=/usr/bin/gfortran-4.8 -DON_TRAVIS=ON -DON_COVERALLS=ON -DCOVERALLS_UPLOAD=ON ${EXTRA_CMAKE_OPTS}
+cmake .. -DCMAKE_Fortran_COMPILER=/usr/bin/gfortran-4.8 -DON_TRAVIS=ON -DWITH_COVERAGE=ON -DCOVERALLS_UPLOAD=ON ${EXTRA_CMAKE_OPTS}
 NRPOC=`nproc`
 make -j $NPROC
 make coveralls

paddle/setup.py.in

@@ -12,68 +12,19 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# This file is used to build paddle python binding package.
-# It will be invoked by Makefile that generated by COMAKE
 
 from setuptools import setup, Extension
 
-import numpy as np
-import api.paddle_ld_flags
-import platform
-import os
-
-system = platform.system().lower()
-
-is_osx = (system == 'darwin')
-is_win = (system == 'windows')
-is_lin = (system == 'linux')
-
-
-# The extra links will passed from COMAKE
-#   because generate paddle LDFLAGS is too complicated to do in setup.py
-#   it just read COMAKE generated LDFLAGS.
-extra_comps = []
-extra_links = []
-obj = api.paddle_ld_flags.PaddleLDFlag()
-extra_comps = obj.c_flag()
-ldflags = obj.ldflag_str()
-if ldflags is not None:
-  extra_links.extend(ldflags.split(" "))
-
-try:
-  with open('.py_paddle_extra_link_flags', 'r') as f:
-    for line in f:
-      extra_links += line.split()
-except:
-  pass
-
-if is_lin == True:
-    extra_links = ["-Xlinker", '-start-group'] + extra_links + ["-Xlinker", "-end-group"]
-elif is_osx == True:
-    os.environ["ARCHFLAGS"] = "-arch x86_64"
-    extra_links = ["-Wl,-all_load"] + extra_links
-
-include_dirs = [np.get_include(), "../"]    # include numpy and paddle.
-
-os.environ["CC"] = "@CMAKE_C_COMPILER@"
-os.environ["CXX"] = "@CMAKE_CXX_COMPILER@"
-
 setup(name="py_paddle",
-  version="@PADDLE_VERSION@",
-  ext_modules=[
-    Extension('py_paddle._swig_paddle',      # Build SWIG Extension.
-       ['Paddle_wrap.cxx'],
-       language = "c++",
-       include_dirs = include_dirs,
-       extra_link_args = extra_links,
-       extra_compile_args = extra_comps
-    )
-  ],
+      version="${PADDLE_VERSION}",
       packages=['py_paddle'],
-  include_dirs = include_dirs,
+      include_package_data=True,
+      package_data={'py_paddle':['*.py','_swig_paddle.so']},
       install_requires = [
         'nltk>=3.2.2',
         'numpy>=1.8.0',      # The numpy is required.
-    'protobuf>=3.0.0'    # The paddle protobuf version
+        'protobuf>=${PROTOBUF_VERSION}'    # The paddle protobuf version
       ],
+      url='http://www.paddlepaddle.org/',
+      license='Apache 2.0',
 )

paddle/trainer/Trainer.h

@@ -30,10 +30,6 @@ limitations under the License. */
 #include "TrainerConfigHelper.h"
 #include "TrainerInternal.h"
 
-#ifdef PADDLE_METRIC_LEARNING
-#include "paddle/internals/metric_learning/MetricTrainer.h"
-#endif
-
 DECLARE_int32(num_passes);
 
 namespace paddle {
@@ -201,12 +197,8 @@ protected:
   // parameter util
   std::unique_ptr<ParameterUtil> paramUtil_;
 
-#ifdef PADDLE_METRIC_LEARNING
-  MetricTrainer trainerInternal_;
-#else
   // trainer Internal
   TrainerInternal trainerInternal_;
-#endif
 };
 
 }  // namespace paddle

paddle/utils/CpuId.h

@@ -12,6 +12,7 @@ limitations under the License. */
 #pragma once
 
 #include "Common.h"
+#include "Error.h"
 
 namespace paddle {
 
@@ -99,4 +100,37 @@ private:
 #define HAS_NEON    HAS_SIMD(SIMD_NEON)
 // clang-format on
 
+/**
+ * Invoke checkCPUFeature() before Paddle initialization to
+ * check target machine whether support compiled instructions.
+ * If not, simply throw out an error.
+ */
+inline Error __must_check checkCPUFeature() {
+  Error err;
+#ifndef __AVX__
+  if (HAS_AVX) {
+    LOG(WARNING) << "PaddlePaddle wasn't compiled to use avx instructions, "
+                 << "but these are available on your machine and could "
+                 << "speed up CPU computations via CMAKE .. -DWITH_AVX=ON";
+  }
+#else
+  if (!HAS_AVX) {
+    err = Error(
+        "PaddlePaddle was compiled to use avx instructions, "
+        "but these aren't available on your machine, please "
+        "disable it via CMAKE .. -DWITH_AVX=OFF");
+  }
+#endif  // __AVX__
+#ifdef __SSE3__
+  if (!HAS_SSE3) {
+    err = Error(
+        "PaddlePaddle was compiled to use sse3 instructions, "
+        "which is the minimum requirement of PaddlePaddle. "
+        "But these aren't available on your current machine.");
+  }
+#endif  // __SSE3__
+
+  return err;
+}
+
 }  // namespace paddle

paddle/utils/Flags.cpp

@@ -30,7 +30,6 @@ DEFINE_bool(parallel_nn,
 DEFINE_int32(trainer_count, 1, "Defined how many trainers to train");
 DEFINE_int32(gpu_id, 0, "Which gpu core to use");
 DEFINE_int32(port, 20134, "Listening port for pserver");
-DEFINE_int32(data_server_port, 21134, "Listening port for dserver");
 DEFINE_int32(ports_num,
              1,
              "Number of ports for sending dense parameter,"

paddle/utils/Flags.h

@@ -19,7 +19,6 @@ limitations under the License. */
 DECLARE_bool(parallel_nn);
 DECLARE_int32(async_count);
 DECLARE_int32(port);
-DECLARE_int32(data_server_port);
 DECLARE_bool(use_gpu);
 DECLARE_int32(gpu_id);
 DECLARE_int32(trainer_count);

paddle/utils/GlobalConstants.h

@@ -23,11 +23,6 @@ enum PassType {
   PASS_TEST,    // Test pass
   PASS_GC,      // Gradient Check pass
   PASS_METRIC,  // pass for generate template output with no drop rate.
-  // pass for metric learning training with metric learning error, only used
-  // when we are doing KNN evaluation.
-  PASS_METRIC_TRAIN,
-  PASS_METRIC_TRAIN_WITH_NOERROR,  // Pass for metric learning training
-                                   // with no evaluation.
 };
 
 enum ParameterType {

paddle/utils/PythonUtil.cpp.in

@@ -195,9 +195,14 @@ extern const char enable_virtualenv_py[];
 }
 void initPython(int argc, char** argv) {
 #ifndef PADDLE_NO_PYTHON
-  char pyHome[] = "@PYTHON_INSTALL_DIR@"; // NOLINT
-  if (strlen(pyHome)) {
-    Py_SetPythonHome(pyHome);
+  std::string pyHome;
+#if defined(__APPLE__) || defined(__OSX__)
+  pyHome = "/usr/local/Frameworks/Python.framework/Versions/2.7";
+  Py_SetPythonHome(const_cast<char*>(pyHome.c_str()));
+#endif
+  pyHome = "@PYTHON_INSTALL_DIR@"; // NOLINT
+  if (!pyHome.empty()) {
+    Py_SetPythonHome(const_cast<char*>(pyHome.c_str()));
   }
   Py_SetProgramName(argv[0]);
   Py_Initialize();

paddle/utils/Util.cpp

@@ -31,6 +31,7 @@ limitations under the License. */
 
 #include <gflags/gflags.h>
 
+#include "CpuId.h"
 #include "CustomStackTrace.h"
 #include "Logging.h"
 #include "StringUtil.h"
@@ -194,6 +195,7 @@ void initMain(int argc, char** argv) {
   }
 
   version::printVersion();
+  checkCPUFeature().check();
   runInitFunctions();
 }
 

python/paddle/trainer/config_parser.py

@@ -686,25 +686,17 @@ class ContextProjection(Projection):
 
 
 @config_class
-class ConvProjection(Projection):
-    type = 'conv'
-
+class ConvBaseProjection(Projection):
     def __init__(self,
                  input_layer_name,
                  num_filters=None,
                  conv_conf=None,
                  **xargs):
-        super(ConvProjection, self).__init__(input_layer_name, **xargs)
+        super(ConvBaseProjection, self).__init__(input_layer_name, **xargs)
 
         if num_filters is not None:
             self.proj_conf.num_filters = num_filters
 
-        parse_conv(conv_conf, input_layer_name, self.proj_conf.conv_conf,
-                   num_filters)
-        self.proj_conf.output_size = self.proj_conf.conv_conf.output_x * \
-                                     self.proj_conf.conv_conf.output_y * \
-                                     num_filters
-
     def calc_output_size(self, input_layer_config):
         return self.proj_conf.output_size
 
@@ -723,6 +715,48 @@ class ConvProjection(Projection):
         return None
 
 
+@config_class
+class ConvProjection(ConvBaseProjection):
+    type = 'conv'
+
+    def __init__(self,
+                 input_layer_name,
+                 num_filters=None,
+                 conv_conf=None,
+                 **xargs):
+        super(ConvProjection, self).__init__(input_layer_name, num_filters,
+                                             conv_conf, **xargs)
+
+        parse_conv(conv_conf, self.input_layer_name, self.proj_conf.conv_conf,
+                   num_filters)
+        self.proj_conf.output_size = self.proj_conf.conv_conf.output_x * \
+                                     self.proj_conf.conv_conf.output_y * \
+                                     num_filters
+
+
+@config_class
+class ConvTransProjection(ConvBaseProjection):
+    type = 'convt'
+
+    def __init__(self,
+                 input_layer_name,
+                 num_filters=None,
+                 conv_conf=None,
+                 **xargs):
+        super(ConvTransProjection, self).__init__(input_layer_name, num_filters,
+                                                  conv_conf, **xargs)
+
+        parse_conv(
+            conv_conf,
+            self.input_layer_name,
+            self.proj_conf.conv_conf,
+            num_filters,
+            trans=True)
+        self.proj_conf.output_size = self.proj_conf.conv_conf.img_size_y * \
+                                     self.proj_conf.conv_conf.img_size * \
+                                     num_filters
+
+
 # Define a operator for mixed layer
 @config_class
 class Operator(Cfg):
@@ -789,6 +823,36 @@ class ConvOperator(Operator):
         return self.operator_conf.output_size
 
 
+@config_class
+class ConvTransOperator(Operator):
+    type = 'convt'
+
+    def __init__(self,
+                 input_layer_names,
+                 num_filters=None,
+                 conv_conf=None,
+                 **xargs):
+        super(ConvTransOperator, self).__init__(input_layer_names, **xargs)
+        if num_filters is not None:
+            self.operator_conf.num_filters = num_filters
+
+        parse_conv(
+            conv_conf,
+            MakeLayerNameInSubmodel(input_layer_names[0]),
+            self.operator_conf.conv_conf,
+            num_filters,
+            trans=True)
+        self.operator_conf.output_size = \
+            self.operator_conf.conv_conf.img_size * \
+            self.operator_conf.conv_conf.img_size_y * \
+            num_filters
+
+        config_assert(len(input_layer_names) == 2, "Conv is binary operator")
+
+    def calc_output_size(self, input_sizes):
+        return self.operator_conf.output_size
+
+
 # please refer to the comments in proto/ModelConfig.proto
 @config_class
 class Conv(Cfg):
@@ -1772,7 +1836,16 @@ class ConvTransLayerBase(LayerBase):
         use_gpu = int(g_command_config_args.get("use_gpu", 0))
         parallel_nn = int(g_command_config_args.get("parallel_nn", 0))
 
-        # cudnn_convt has not been implemented so use exconvt only
+        # Automatically select cudnn_type for GPU and exconvt for CPU
+        # if set type=exconvt, but still reserve the way user specify
+        # exconvt or cudnn_convt manually.
+        if self.layer_type == "cudnn_convt":
+            config_assert(use_gpu, "cudnn_convt only support GPU")
+
+        if (use_gpu == 1 and self.layer_type != "exconvt" and
+            (parallel_nn == 0 or self.config.device > -1)):
+            self.layer_type = "cudnn_convt"
+        else:
             self.layer_type = "exconvt"
         # need to specify layer in config
         self.config.type = self.layer_type
@@ -1790,10 +1863,9 @@ class ConvTransLayerBase(LayerBase):
                 trans=True)
             conv_conf = self.config.inputs[input_index].conv_conf
             psize = self.calc_parameter_size(conv_conf)
-            print("output size for %s is %d " % (name, conv_conf.output_x))
             self.create_input_parameter(input_index, psize)
-            self.set_layer_size(
-                (conv_conf.img_size**2) * self.config.num_filters)
+            self.set_cnn_layer(name, conv_conf.img_size_y, conv_conf.img_size,
+                               self.config.num_filters)
 
         psize = self.config.size
         if shared_biases:
@@ -1810,6 +1882,11 @@ class ConvTransLayer(ConvTransLayerBase):
     layer_type = 'exconvt'
 
 
+@config_layer('cudnn_convt')
+class ConvTransLayer(ConvTransLayerBase):
+    layer_type = 'cudnn_convt'
+
+
 @config_layer('norm')
 class NormLayer(LayerBase):
     def __init__(self, name, inputs, **xargs):
@@ -2222,7 +2299,10 @@ def Link(
 
 # memory for recurrent layer group.
 # *name* and *size* are actual layer's name and size.
-# will return name of the memory,
+# If *name* is None, need to provide *memory_name* and need to use
+# SetMemoryInput() later to specify the layer which this memory remembers.
+#
+# return the name of the memory,
 # use this name if you assign the memory as other layer's input
 #
 # boot frame of memory is zeroed by default,
@@ -2234,15 +2314,18 @@ def Link(
 # can only be initailized by a *boot_layer* which is a sequence.
 #
 @config_func
-def Memory(
-        name,
+def Memory(name,
            size,
            is_sequence=False,
            boot_layer=None,
            boot_bias=False,
            boot_bias_active_type="",
-        boot_with_const_id=None, ):
-    agent_name = name + "+delay1"
+           boot_with_const_id=None,
+           memory_name=None):
+    if not memory_name:
+        config_assert(name is not None, "name needs cannot be None")
+        memory_name = name + "+delay1"
+    agent_name = memory_name
     if is_sequence:
         agent_layer = SequenceAgentLayer(agent_name, size)
     else:
@@ -2250,6 +2333,7 @@ def Memory(
     config_assert(g_current_submodel.is_recurrent_layer_group,
                   'Memory should be used in recurrent layer group only')
     memory = g_current_submodel.memories.add()
+    if name is not None:
         memory.layer_name = MakeLayerNameInSubmodel(name)
     memory.link_name = MakeLayerNameInSubmodel(agent_name)
     memory.is_sequence = is_sequence
@@ -2274,6 +2358,17 @@ def Memory(
     return agent_name
 
 
+@config_func
+def SetMemoryInput(memory_name, layer_name):
+    memory_name = MakeLayerNameInSubmodel(memory_name)
+    layer_name = MakeLayerNameInSubmodel(layer_name)
+    for mem in g_current_submodel.memories:
+        if mem.link_name == memory_name:
+            mem.layer_name = layer_name
+            return
+    logger.fatal("Nonexistent memory name: " + memory_name)
+
+
 # Generator for recurrent layer group, to use it:
 #  1. define a id layer as output of layer group
 #  2. define a memory of this id layer, and assign a boot id(begin of sequence)

python/paddle/trainer_config_helpers/default_decorators.py

@@ -97,13 +97,13 @@ def reset_hook():
 register_parse_config_hook(reset_hook)
 
 
-def wrap_name_default(name_prefix=None):
+def wrap_name_default(name_prefix=None, name_param="name"):
     """
     Decorator to set "name" arguments default to "{name_prefix}_{invoke_count}".
 
     ..  code:: python
 
-        @default_name("some_name")
+        @wrap_name_default("some_name")
         def func(name=None):
             print name      # name will never be None. If name is not set,
                             # name will be "some_name_%d"
@@ -115,7 +115,7 @@ def wrap_name_default(name_prefix=None):
     """
     factory = DefaultNameFactory(name_prefix)
     _name_factories.append(factory)
-    return wrap_param_default(["name"], factory)
+    return wrap_param_default([name_param], factory)
 
 
 def wrap_param_attr_default(param_names=None, default_factory=None):

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

@@ -34,11 +34,31 @@ flt = data_layer(name='filter', size=3 * 3 * 1 * 64)
 with mixed_layer() as m7:
     m7 += conv_operator(
         img=img, filter=flt, num_filters=64, num_channels=1, filter_size=3)
+    m7 += conv_projection(img, filter_size=3, num_filters=64, num_channels=1)
 
+with mixed_layer() as m8:
+    m8 += conv_operator(
+        img=img,
+        filter=flt,
+        num_filters=64,
+        num_channels=1,
+        filter_size=3,
+        stride=2,
+        padding=1,
+        trans=True)
+    m8 += conv_projection(
+        img,
+        filter_size=3,
+        num_filters=64,
+        num_channels=1,
+        stride=2,
+        padding=1,
+        trans=True)
 end = mixed_layer(
     input=[
         full_matrix_projection(input=m5),
-        trans_full_matrix_projection(input=m6), full_matrix_projection(input=m7)
+        trans_full_matrix_projection(input=m6),
+        full_matrix_projection(input=m7), full_matrix_projection(input=m8)
     ],
     size=100,
     layer_attr=ExtraAttr(

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

@@ -33,6 +33,8 @@ layers {
   bias_parameter_name: "___conv_0__.wbias"
   num_filters: 64
   shared_biases: true
+  height: 256
+  width: 256
 }
 layers {
   name: "__batch_norm_0__"
@@ -58,6 +60,8 @@ layers {
   }
   bias_parameter_name: "___batch_norm_0__.wbias"
   moving_average_fraction: 0.9
+  height: 256
+  width: 256
 }
 layers {
   name: "__crmnorm_0__"

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

@@ -45,7 +45,7 @@ layers {
   coeff: 1.0
 }
 layers {
-  name: "__regression_cost_0__"
+  name: "__mse_cost_0__"
   type: "square_error"
   size: 1
   active_type: ""
@@ -84,7 +84,7 @@ input_layer_names: "input"
 input_layer_names: "label"
 input_layer_names: "weight"
 output_layer_names: "__cost_0__"
-output_layer_names: "__regression_cost_0__"
+output_layer_names: "__mse_cost_0__"
 evaluators {
   name: "classification_error_evaluator"
   type: "classification_error"
@@ -99,12 +99,12 @@ sub_models {
   layer_names: "weight"
   layer_names: "__fc_layer_0__"
   layer_names: "__cost_0__"
-  layer_names: "__regression_cost_0__"
+  layer_names: "__mse_cost_0__"
   input_layer_names: "input"
   input_layer_names: "label"
   input_layer_names: "weight"
   output_layer_names: "__cost_0__"
-  output_layer_names: "__regression_cost_0__"
+  output_layer_names: "__mse_cost_0__"
   evaluator_names: "classification_error_evaluator"
   is_recurrent_layer_group: false
 }

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

@@ -331,6 +331,54 @@ layers {
   }
   trans_type: "non-seq"
 }
+layers {
+  name: "__recurrent_group_3__"
+  type: "recurrent_layer_group"
+  active_type: ""
+}
+layers {
+  name: "seq_input@__recurrent_group_3__"
+  type: "scatter_agent"
+  size: 100
+  active_type: ""
+}
+layers {
+  name: "__memory_6__@__recurrent_group_3__"
+  type: "agent"
+  size: 200
+  active_type: ""
+}
+layers {
+  name: "__fc_layer_0__@__recurrent_group_3__"
+  type: "fc"
+  size: 200
+  active_type: "tanh"
+  inputs {
+    input_layer_name: "seq_input@__recurrent_group_3__"
+    input_parameter_name: "___fc_layer_0__@__recurrent_group_3__.w0"
+  }
+  inputs {
+    input_layer_name: "__memory_6__@__recurrent_group_3__"
+    input_parameter_name: "___fc_layer_0__@__recurrent_group_3__.w1"
+  }
+  bias_parameter_name: "___fc_layer_0__@__recurrent_group_3__.wbias"
+}
+layers {
+  name: "__fc_layer_0__"
+  type: "gather_agent"
+  size: 200
+  active_type: ""
+}
+layers {
+  name: "__last_seq_4__"
+  type: "seqlastins"
+  size: 200
+  active_type: "linear"
+  inputs {
+    input_layer_name: "__fc_layer_0__"
+  }
+  trans_type: "non-seq"
+}
 parameters {
   name: "___mixed_0__.w0"
   size: 40000
@@ -481,6 +529,36 @@ parameters {
   initial_strategy: 0
   initial_smart: false
 }
+parameters {
+  name: "___fc_layer_0__@__recurrent_group_3__.w0"
+  size: 20000
+  initial_mean: 0.0
+  initial_std: 0.1
+  dims: 100
+  dims: 200
+  initial_strategy: 0
+  initial_smart: true
+}
+parameters {
+  name: "___fc_layer_0__@__recurrent_group_3__.w1"
+  size: 40000
+  initial_mean: 0.0
+  initial_std: 0.0707106781187
+  dims: 200
+  dims: 200
+  initial_strategy: 0
+  initial_smart: true
+}
+parameters {
+  name: "___fc_layer_0__@__recurrent_group_3__.wbias"
+  size: 200
+  initial_mean: 0.0
+  initial_std: 0.0
+  dims: 1
+  dims: 200
+  initial_strategy: 0
+  initial_smart: false
+}
 input_layer_names: "seq_input"
 input_layer_names: "sub_seq_input"
 output_layer_names: "__last_seq_0__"
@@ -488,6 +566,7 @@ output_layer_names: "__first_seq_0__"
 output_layer_names: "__last_seq_1__"
 output_layer_names: "__last_seq_2__"
 output_layer_names: "__last_seq_3__"
+output_layer_names: "__last_seq_4__"
 sub_models {
   name: "root"
   layer_names: "seq_input"
@@ -510,6 +589,9 @@ sub_models {
   layer_names: "__gru_group_0___recurrent_group"
   layer_names: "__gru_group_0__"
   layer_names: "__last_seq_3__"
+  layer_names: "__recurrent_group_3__"
+  layer_names: "__fc_layer_0__"
+  layer_names: "__last_seq_4__"
   input_layer_names: "seq_input"
   input_layer_names: "sub_seq_input"
   output_layer_names: "__last_seq_0__"
@@ -517,6 +599,7 @@ sub_models {
   output_layer_names: "__last_seq_1__"
   output_layer_names: "__last_seq_2__"
   output_layer_names: "__last_seq_3__"
+  output_layer_names: "__last_seq_4__"
   is_recurrent_layer_group: false
 }
 sub_models {
@@ -647,4 +730,28 @@ sub_models {
   }
   target_inlinkid: -1
 }
+sub_models {
+  name: "__recurrent_group_3__"
+  layer_names: "seq_input@__recurrent_group_3__"
+  layer_names: "__memory_6__@__recurrent_group_3__"
+  layer_names: "__fc_layer_0__@__recurrent_group_3__"
+  is_recurrent_layer_group: true
+  reversed: false
+  memories {
+    layer_name: "__fc_layer_0__@__recurrent_group_3__"
+    link_name: "__memory_6__@__recurrent_group_3__"
+    is_sequence: false
+  }
+  in_links {
+    layer_name: "seq_input"
+    link_name: "seq_input@__recurrent_group_3__"
+    has_subseq: false
+  }
+  out_links {
+    layer_name: "__fc_layer_0__@__recurrent_group_3__"
+    link_name: "__fc_layer_0__"
+    has_subseq: false
+  }
+  target_inlinkid: -1
+}
 

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

@@ -10,5 +10,5 @@ fc = fc_layer(input=data, size=10, act=SoftmaxActivation())
 outputs(
     classification_cost(
         input=fc, label=lbl, weight=wt),
-    regression_cost(
+    mse_cost(
         input=fc, label=lbl, weight=wt))