Merge branch 'develop' into huber_loss

.pre-commit-config.yaml

@@ -22,7 +22,7 @@
     -   id: clang-format-with-version-check
         name: clang-format
         description: Format files with ClangFormat.
-        entry: ./.clang_format.hook -i
+        entry: bash ./.clang_format.hook -i
         language: system
         files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$
 -   repo: https://github.com/PaddlePaddle/pre-commit-golang

Dockerfile

@@ -10,13 +10,11 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub
 ARG WITH_GPU
 ARG WITH_AVX
 ARG WITH_DOC
-ARG WITH_STYLE_CHECK
 
 ENV WOBOQ OFF
-ENV WITH_GPU=${WITH_GPU:-OFF}
+ENV WITH_GPU=${WITH_GPU:-ON}
 ENV WITH_AVX=${WITH_AVX:-ON}
 ENV WITH_DOC=${WITH_DOC:-OFF}
-ENV WITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF}
 
 ENV HOME /root
 # Add bash enhancements

doc/design/cluster_train/README.md

@@ -54,17 +54,18 @@ The life cycle of a single task is illustrated below:
 <img src="src/paddle-task-states.png"/>
 
 1. When a new pass of training starts, all tasks will be placed in the todo queue.
-1. The master server will dispatch few tasks to each trainer at a time, puts them in the pending queue and waits for completion.
-1. The trainer will work on its tasks and tell the master server once a task is completed. The master server will dispatch a new task to that trainer.
-1. If a task timeout. the master server will move it back to the todo queue. The timeout count will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, so it will be discarded.
+1. Upon trainer requests for new task, the master server will dispatch a task from todo queue to it, put the task in the pending queue and wait for completion.
+1. The trainer will work on its task and tell the master server once the task is completed and ask for new task. The master server will dispatch a new task to that trainer.
+1. If a task fails for any reason in trainer, or takes longer than a specific period of time,  the master server will move the task back to the todo queue. The timeout count for that task will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, then it will be discarded.
 1. The master server will move completed task to the done queue. When the todo queue is empty, the master server will start a new pass by moving all tasks in the done queue to todo queue and reset the timeout counter of all tasks to zero.
 
 ### Trainer Process
 
 The trainer process will:
 
-- Receive tasks from the master.
-- Work on the tasks: calculate and upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers.
+- Request tasks from the master.
+- Work on the tasks
+- Upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers.
 
 ### Parameter Server Process
 
@@ -119,8 +120,8 @@ When the master is started by the Kubernetes, it executes the following steps at
 
 1. Grabs a unique *master* lock in etcd, which prevents concurrent master instantiations.
 1. Recovers the task queues from etcd if they already exist, otherwise, the master will create them.
-1. Watches the trainer prefix keys `/trainer/` on etcd to find the live trainers.
-1. Starts dispatching the tasks to the trainers, and updates task queue using an etcd transaction to ensure lock is held during the update.
+1. Write its ip address to */master/addr* so that trainers can discover it.
+1. Listens to trainers' request of task, dispatch one upon request, and updates task queue using an etcd transaction to ensure lock is held during the update.
 
 When the master server process is dead for any reason, Kubernetes will restart it. It will be online again with all states recovered from etcd in few minutes.
 
@@ -128,13 +129,11 @@ When the master server process is dead for any reason, Kubernetes will restart i
 
 When the trainer is started by the Kubernetes, it executes the following steps at startup:
 
-1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count.
-1. Generates a unique ID, and sets key `/trainer/<unique ID>` with its contact address as value. The key will be deleted when the lease expires, so the master will be aware of the trainer being online and offline.
-1. Waits for tasks from the master to start training.
+1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count */ps_desired*.
+1. Finds and watches */master/addr* to get master's address.
+1. Requests for tasks from the master to start training.
 
-If trainer's etcd lease expires, it will try set key `/trainer/<unique ID>` again so that the master server can discover the trainer again.
-
-When a trainer fails, Kuberentes would try to restart it. The recovered trainer would fetch tasks from the TODO queue and go on training.
+When a trainer fails, Kuberentes would try to restart it. The recovered trainer would fetch tasks from master and go on training.
 
 ### Parameter Server Process
 

doc/howto/dev/build_cn.md

+# 编译PaddlePaddle和运行单元测试
+
+## 需要的软硬件
+
+为了开发PaddlePaddle，我们需要
+
+1. 一台电脑，可以装的是 Linux, BSD, Windows 或者 MacOS 操作系统，以及
+1. Docker。
+
+不需要依赖其他任何软件了。即便是 Python 和 GCC 都不需要，因为我们会把所有编译工具都安装进一个 Docker image 里。
+
+## 总体流程
+
+1. 获取源码
+
+   ```bash
+   git clone https://github.com/paddlepaddle/paddle
+   ```
+
+2. 安装开发工具到 Docker image 里
+
+   ```bash
+   cd paddle; docker build -t paddle:dev .
+   ```
+
+   请注意这个命令结尾处的 `.`；它表示 `docker build` 应该读取当前目录下的 [`Dockerfile`文件](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile)，按照其内容创建一个名为 `paddle:dev` 的 Docker image，并且把各种开发工具安装进去。
+
+3. 编译
+
+   以下命令启动一个 Docker container 来执行 `paddle:dev` 这个 Docker image，同时把当前目录（源码树根目录）映射为 container 里的 `/paddle` 目录，并且运行 `Dockerfile` 描述的默认入口程序 [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `cmake` 和 `make` 来编译 `/paddle` 里的源码，结果输出到 `/paddle/build`，也就是本地的源码树根目录里的 `build` 子目录。
+
+   ```bash
+   docker run --rm -v $PWD:/paddle paddle:dev
+   ```
+
+   上述命令编译出一个 CUDA-enabled 版本。如果我们只需要编译一个只支持 CPU 的版本，可以用
+
+   ```bash
+   docker run --rm -e WITH_GPU=OFF -v $PWD:/paddle paddle:dev
+   ```
+
+4. 运行单元测试
+
+   用本机的第一个 GPU 来运行包括 GPU 单元测试在内的所有单元测试：
+
+   ```bash
+   NV_GPU=0 nvidia-docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest"
+   ```
+
+   如果编译的时候我们用了 `WITH_GPU=OFF` 选项，那么编译过程只会产生 CPU-based 单元测试，那么我们也就不需要 nvidia-docker 来运行单元测试了。我们只需要：
+
+   ```bash
+   docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest"
+   ```
+
+   有时候我们只想运行一个特定的单元测试，比如 `memory_test`，我们可以
+
+   ```bash
+   nvidia-docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest -V -R memory_test"
+   ```
+
+5. 清理
+
+   有时候我们会希望清理掉已经下载的第三方依赖以及已经编译的二进制文件。此时只需要：
+
+   ```bash
+   rm -rf build
+   ```
+
+## 为什么要 Docker 呀？
+
+- 什么是 Docker?
+
+  如果您没有听说 Docker，可以把它想象为一个类似 virtualenv 的系统，但是虚拟的不仅仅是 Python 的运行环境。
+
+- Docker 还是虚拟机？
+
+  有人用虚拟机来类比 Docker。需要强调的是：Docker 不会虚拟任何硬件，Docker container 里运行的编译工具实际上都是在本机的 CPU 和操作系统上直接运行的，性能和把编译工具安装在本机运行一样。
+
+- 为什么用 Docker?
+
+  把工具和配置都安装在一个 Docker image 里可以标准化编译环境。这样如果遇到问题，其他人可以复现问题以便帮助。
+
+  另外，对于习惯使用Windows和MacOS的开发者来说，使用Docker就不用配置交叉编译环境了。
+
+- 我可以选择不用Docker吗？
+
+  当然可以。大家可以用把开发工具安装进入 Docker image 一样的方式，把这些工具安装到本机。这篇文档介绍基于 Docker 的开发流程，是因为这个流程比其他方法都更简便。
+
+- 学习 Docker 有多难？
+
+  理解 Docker 并不难，大概花十分钟看一下[这篇文章](https://zhuanlan.zhihu.com/p/19902938)。这可以帮您省掉花一小时安装和配置各种开发工具，以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。
+
+- 我可以用 IDE 吗？
+
+  当然可以，因为源码就在本机上。IDE 默认调用 make 之类的程序来编译源码，我们只需要配置 IDE 来调用 Docker 命令编译源码即可。
+
+  很多 PaddlePaddle 开发者使用 Emacs。他们在自己的 `~/.emacs` 配置文件里加两行
+
+  ```emacs
+  (global-set-key "\C-cc" 'compile)
+  (setq compile-command
+   "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev")
+  ```
+
+  就可以按 `Ctrl-C` 和 `c` 键来启动编译了。
+
+- 可以并行编译吗？
+
+  是的。我们的 Docker image 运行一个 [Bash 脚本](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。
+
+## 可能碰到的问题
+
+- Docker 需要 sudo
+
+  如果用自己的电脑开发，自然也就有管理员权限（sudo）了。如果用公用的电脑开发，需要请管理员安装和配置好 Docker。此外，PaddlePaddle 项目在努力开始支持其他不需要 sudo 的集装箱技术，比如 rkt。
+
+- 在 Windows/MacOS 上编译很慢
+
+  Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存，以保证编译高效。具体做法请参考[这个issue](https://github.com/PaddlePaddle/Paddle/issues/627)。
+
+- 磁盘不够
+
+  本文中的例子里，`docker run` 命令里都用了 `--rm` 参数，这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 `docker ps -a` 命令看到停止后但是没有删除的 containers。`docker build` 命令有时候会产生一些中间结果，是没有名字的 images，也会占用磁盘。可以参考[这篇文章](https://zaiste.net/posts/removing_docker_containers/)来清理这些内容。

doc/howto/dev/build_en.md

+# Build PaddlePaddle from Source Code and Run Unit Test
+
+## What Developers Need
+
+To contribute to PaddlePaddle, you need
+
+1. A computer -- Linux, BSD, Windows, MacOS, and
+1. Docker.
+
+Nothing else.  Not even Python and GCC, because you can install all build tools into a Docker image.  We run all the tools by running this image.
+
+## General Process
+
+1. Retrieve source code.
+
+   ```bash
+   git clone https://github.com/paddlepaddle/paddle
+   ```
+
+2. Install build tools into a Docker image.
+
+   ```bash
+   cd paddle; docker build -t paddle:dev .
+   ```
+
+   Please be aware of the `.` at the end of the command, which refers to the [`./Dockerfile` file](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile).  `docker build` follows instructions in this file to create a Docker image named `paddle:dev`, and installs building tools into it.
+
+3. Build from source.
+
+   This following command starts a Docker container that executes the Docker image `paddle:dev`, mapping the current directory to `/paddle/` in the container, and runs the default entry-point [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh) as specified in the Dockefile.  `build.sh` invokes `cmake` and `make` to build PaddlePaddle source code, which had been mapped to `/paddle`, and writes outputs to `/paddle/build`, which maps to `build` in the current source directory on the computer.
+
+   ```bash
+   docker run -v $PWD:/paddle paddle:dev
+   ```
+
+   Above command builds a CUDA-enabled version.  If we want to build a CPU-only version, we can type
+
+   ```bash
+   docker run -e WITH_GPU=OFF -v $PWD:/paddle paddle:dev
+   ```
+
+4. Run unit tests.
+
+   To run all unit tests using the first GPU of a node:
+
+   ```bash
+   NV_GPU=0 nvidia-docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest"
+   ```
+
+   If we used `WITH_GPU=OFF` at build time, it generates only CPU-based unit tests, and we don't need nvidia-docker to run them.  We can just run
+
+   ```bash
+   docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest"
+   ```
+
+   Sometimes we want to run a specific unit test, say `memory_test`, we can run
+
+   ```bash
+   nvidia-docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest -V -R memory_test"
+   ```
+
+5. Clean Build.
+
+   Sometimes, we might want to clean all thirt-party dependents and built binaries.  To do so, just
+
+   ```bash
+   rm -rf build
+   ```
+
+## Docker, Or Not?
+
+- What is Docker?
+
+  If you haven't heard of it, consider it something like Python's virtualenv.
+
+- Docker or virtual machine?
+
+  Some people compare Docker with VMs, but Docker doesn't virtualize any hardware nor running a guest OS, which means there is no compromise on the performance.
+
+- Why Docker?
+
+  Using a Docker image of build tools standardizes the building environment, which makes it easier for others to reproduce your problems and to help.
+
+  Also, some build tools don't run on Windows or Mac or BSD, but Docker runs almost everywhere, so developers can use whatever computer they want.
+
+- Can I choose not to use Docker?
+
+  Sure, you don't have to install build tools into a Docker image; instead, you can install them in your local computer.  This document exists because Docker would make the development way easier.
+
+- How difficult is it to learn Docker?
+
+    It takes you ten minutes to read [an introductory article](https://docs.docker.com/get-started) and saves you more than one hour to install all required build tools, configure them, especially when new versions of PaddlePaddle require some new tools.  Not even to mention the time saved when other people trying to reproduce the issue you have.
+
+- Can I use my favorite IDE?
+
+  Yes, of course.  The source code resides on your local computer, and you can edit it using whatever editor you like.
+
+  Many PaddlePaddle developers are using Emacs.  They add the following few lines into their `~/.emacs` configure file:
+
+  ```emacs
+  (global-set-key "\C-cc" 'compile)
+  (setq compile-command
+   "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev")
+  ```
+
+  so they could type `Ctrl-C` and `c` to build PaddlePaddle from source.
+
+- Does Docker do parallel building?
+
+  Our building Docker image runs a [Bash script](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh), which calls `make -j$(nproc)` to starts as many processes as the number of your CPU cores.
+
+## Some Gotchas
+
+- Docker requires sudo
+
+  An owner of a computer has the administrative privilege, a.k.a., sudo, and Docker requires this privilege to work properly.  If you use a shared computer for development, please ask the administrator to install and configure Docker.  We will do our best to support rkt, another container technology that doesn't require sudo.
+
+- Docker on Windows/MacOS builds slowly
+
+  On Windows and MacOS, Docker containers run in a Linux VM.  You might want to give this VM some more memory and CPUs so to make the building efficient.  Please refer to [this issue](https://github.com/PaddlePaddle/Paddle/issues/627) for details.
+
+- Not enough disk space
+
+  Examples in this article uses option `--rm` with the `docker run` command.  This option ensures that stopped containers do not exist on hard disks.  We can use `docker ps -a` to list all containers, including stopped.  Sometimes `docker build` generates some intermediate dangling images, which also take disk space.  To clean them, please refer to [this article](https://zaiste.net/posts/removing_docker_containers/).

doc/howto/index_cn.rst

@@ -19,6 +19,7 @@
 ..  toctree::
   :maxdepth: 1
 
+  dev/build_cn.rst
   dev/write_docs_cn.rst
   dev/contribute_to_paddle_cn.md
 

doc/howto/index_en.rst

@@ -18,6 +18,7 @@ Development
 ..  toctree::
   :maxdepth: 1
 
+  dev/build_en.rst
   dev/new_layer_en.rst
   dev/contribute_to_paddle_en.md
 

paddle/capi/CMakeLists.txt

@@ -53,7 +53,10 @@ add_custom_target(paddle_capi_whole ALL
 set_target_properties(paddle_capi_whole
   PROPERTIES IMPORTED_LOCATION ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library})
 
+set(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map")
+# TODO: merge mkl into paddle_capi_shared
 add_library(paddle_capi_shared SHARED ${CAPI_SOURCES})
+set_target_properties(paddle_capi_shared	PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
 target_include_directories(paddle_capi_shared PUBLIC ${CMAKE_CURRENT_BINARY_DIR})
 link_paddle_exe(paddle_capi_shared)
 

paddle/capi/export.map

+{
+	global:
+		paddle_*;
+	local:
+		*;
+};

paddle/cuda/include/hl_cuda_cudnn.h

@@ -214,7 +214,8 @@ extern void hl_conv_workspace(hl_tensor_descriptor input,
                               int* convBwdDataAlgo,
                               size_t* bwdDataLimitBytes,
                               int* convBwdFilterAlgo,
-                              size_t* bwdFilterLimitBytes);
+                              size_t* bwdFilterLimitBytes,
+                              bool useDilation);
 
 /**
  * @brief   destroy filter descriptor.
@@ -242,7 +243,9 @@ extern void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
                                              int padding_height,
                                              int padding_width,
                                              int stride_height,
-                                             int stride_width);
+                                             int stride_width,
+                                             int dilation_h = 1,
+                                             int dilation_w = 1);
 
 /**
  * @brief   reset convolution descriptor.
@@ -262,7 +265,9 @@ extern void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                             int padding_height,
                                             int padding_width,
                                             int stride_height,
-                                            int stride_width);
+                                            int stride_width,
+                                            int dilation_h = 1,
+                                            int dilation_w = 1);
 
 /**
  * @brief   destroy convolution descriptor.

paddle/cuda/include/stub/hl_cuda_cudnn_stub.h

@@ -78,7 +78,9 @@ inline void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
                                              int padding_height,
                                              int padding_width,
                                              int stride_height,
-                                             int stride_width) {}
+                                             int stride_width,
+                                             int dilation_h,
+                                             int dilation_w) {}
 
 inline void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                             hl_tensor_descriptor image,
@@ -86,7 +88,9 @@ inline void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                             int padding_height,
                                             int padding_width,
                                             int stride_height,
-                                            int stride_width) {}
+                                            int stride_width,
+                                            int dilation_h,
+                                            int dilation_w) {}
 
 inline void hl_destroy_convolution_descriptor(hl_convolution_descriptor conv) {}
 
@@ -99,7 +103,8 @@ inline void hl_conv_workspace(hl_tensor_descriptor input,
                               int* convBwdDataAlgo,
                               size_t* bwdDataLimitBytes,
                               int* convBwdFilterAlgo,
-                              size_t* bwdFilterLimitBytes) {}
+                              size_t* bwdFilterLimitBytes,
+                              bool useDilation) {}
 
 inline void hl_convolution_forward(hl_tensor_descriptor input,
                                    real* input_data,

paddle/cuda/src/hl_cuda_cudnn.cc

@@ -201,7 +201,8 @@ void hl_conv_workspace(hl_tensor_descriptor input,
                        int* convBwdDataAlgo,
                        size_t* bwdDataLimitBytes,
                        int* convBwdFilterAlgo,
-                       size_t* bwdFilterLimitBytes) {
+                       size_t* bwdFilterLimitBytes,
+                       bool useDilation) {
 #if CUDNN_VERSION >= 4000
 
   CHECK_NOTNULL(input);
@@ -213,12 +214,32 @@ void hl_conv_workspace(hl_tensor_descriptor input,
   size_t memoryLimitBytes =
       (1LL << 20) * FLAGS_cudnn_conv_workspace_limit_in_mb;
 
+  // For dilation
+  int algo = 0;
+
   // cudnn convolution forward configuration
   cudnnTensorDescriptor_t fwd_src_desc = GET_TENSOR_DESCRIPTOR(input);
   cudnnTensorDescriptor_t fwd_dest_desc = GET_TENSOR_DESCRIPTOR(output);
   cudnnFilterDescriptor_t fwd_filter_desc = GET_FILTER_DESCRIPTOR(filter);
   cudnnConvolutionDescriptor_t fwd_conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv);
+  // cudnn convolution backward data configuration
+  cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter);
+  cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output);
+  cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input);
+  cudnnConvolutionDescriptor_t bwd_data_conv_desc =
+      GET_CONVOLUTION_DESCRIPTOR(conv);
+  // cudnn convolution backward filter configuration
+  cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input);
+  cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output);
+  cudnnConvolutionDescriptor_t bwd_filter_conv_desc =
+      GET_CONVOLUTION_DESCRIPTOR(conv);
+  cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter);
 
+  if (useDilation) {
+    convFwdAlgo = &algo;
+    convBwdDataAlgo = &algo;
+    convBwdFilterAlgo = &algo;
+  } else {
     CHECK_CUDNN(dynload::cudnnGetConvolutionForwardAlgorithm(
         t_resource.cudnn_handle,
         fwd_src_desc,
@@ -228,23 +249,6 @@ void hl_conv_workspace(hl_tensor_descriptor input,
         CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
         memoryLimitBytes,
         reinterpret_cast<cudnnConvolutionFwdAlgo_t*>(convFwdAlgo)));
-
-  CHECK_CUDNN(dynload::cudnnGetConvolutionForwardWorkspaceSize(
-      t_resource.cudnn_handle,
-      fwd_src_desc,
-      fwd_filter_desc,
-      fwd_conv_desc,
-      fwd_dest_desc,
-      static_cast<cudnnConvolutionFwdAlgo_t>(*convFwdAlgo),
-      fwdLimitBytes));
-
-  // cudnn convolution backward data configuration
-  cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter);
-  cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output);
-  cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input);
-  cudnnConvolutionDescriptor_t bwd_data_conv_desc =
-      GET_CONVOLUTION_DESCRIPTOR(conv);
-
     CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataAlgorithm(
         t_resource.cudnn_handle,
         bwd_data_filter_desc,
@@ -254,23 +258,6 @@ void hl_conv_workspace(hl_tensor_descriptor input,
         CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
         memoryLimitBytes,
         reinterpret_cast<cudnnConvolutionBwdDataAlgo_t*>(convBwdDataAlgo)));
-
-  CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
-      t_resource.cudnn_handle,
-      bwd_data_filter_desc,
-      bwd_data_diff_desc,
-      bwd_data_conv_desc,
-      bwd_data_grad_desc,
-      static_cast<cudnnConvolutionBwdDataAlgo_t>(*convBwdDataAlgo),
-      bwdDataLimitBytes));
-
-  // cudnn convolution backward filter configuration
-  cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input);
-  cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output);
-  cudnnConvolutionDescriptor_t bwd_filter_conv_desc =
-      GET_CONVOLUTION_DESCRIPTOR(conv);
-  cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter);
-
     CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
         t_resource.cudnn_handle,
         bwd_filter_src_desc,
@@ -280,6 +267,25 @@ void hl_conv_workspace(hl_tensor_descriptor input,
         CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
         memoryLimitBytes,
         reinterpret_cast<cudnnConvolutionBwdFilterAlgo_t*>(convBwdFilterAlgo)));
+  }
+
+  CHECK_CUDNN(dynload::cudnnGetConvolutionForwardWorkspaceSize(
+      t_resource.cudnn_handle,
+      fwd_src_desc,
+      fwd_filter_desc,
+      fwd_conv_desc,
+      fwd_dest_desc,
+      static_cast<cudnnConvolutionFwdAlgo_t>(*convFwdAlgo),
+      fwdLimitBytes));
+
+  CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
+      t_resource.cudnn_handle,
+      bwd_data_filter_desc,
+      bwd_data_diff_desc,
+      bwd_data_conv_desc,
+      bwd_data_grad_desc,
+      static_cast<cudnnConvolutionBwdDataAlgo_t>(*convBwdDataAlgo),
+      bwdDataLimitBytes));
 
   CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(
       t_resource.cudnn_handle,
@@ -603,7 +609,9 @@ void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
                                       int padding_height,
                                       int padding_width,
                                       int stride_height,
-                                      int stride_width) {
+                                      int stride_width,
+                                      int dilation_h,
+                                      int dilation_w) {
   CHECK_NOTNULL(conv);
 
   cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)malloc(
@@ -625,18 +633,24 @@ void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
                                                        padding_width,
                                                        stride_height,
                                                        stride_width,
-                                                       1,
-                                                       1,
+                                                       dilation_h,
+                                                       dilation_w,
                                                        mode,
                                                        data_type));
 #else
+  if (dilation_h > 1 || dilation_w > 1) {
+    LOG(FATAL)
+        << "Current cuDNN version does't support for dilation convolution. "
+        << "The dilation convolution requires cuDNN >= v6.0.";
+  }
+
   CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(hl_conv->desc,
                                                        padding_height,
                                                        padding_width,
                                                        stride_height,
                                                        stride_width,
-                                                       1,
-                                                       1,
+                                                       dilation_h,
+                                                       dilation_w,
                                                        mode));
 #endif
 
@@ -659,7 +673,9 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                      int padding_height,
                                      int padding_width,
                                      int stride_height,
-                                     int stride_width) {
+                                     int stride_width,
+                                     int dilation_h,
+                                     int dilation_w) {
   CHECK_NOTNULL(conv);
   CHECK_NOTNULL(image);
   CHECK_NOTNULL(filter);
@@ -678,8 +694,8 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                                        padding_width,
                                                        stride_height,
                                                        stride_width,
-                                                       1,
-                                                       1,
+                                                       dilation_h,
+                                                       dilation_w,
                                                        mode,
                                                        data_type));
 #else
@@ -688,8 +704,8 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                                        padding_width,
                                                        stride_height,
                                                        stride_width,
-                                                       1,
-                                                       1,
+                                                       dilation_h,
+                                                       dilation_w,
                                                        mode));
 #endif
 

paddle/framework/backward.md

@@ -21,18 +21,32 @@ grad_op_builder(fengjiayi)
 
 given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`,`InputGradients`.
 
-1. bla bla bla (yuyang)
+1. Op 
+
+   when the input forward network is a Op, return its gradient Operator Immediately.
 
 2. NetOp 
 
-   when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively and ensure them done. During the process, we need to collect the `OutputGradients` name.
+   when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to forward NetOp.
+
+   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwirte their shared input variable.  
+
+   <p align="center">
+   <img src="./images/duplicate_op.png" width="70%" ><br/>
+
+   1. shared variable in two operators. 
+
+   </p>
+
+   Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator replace the overwirte links. 
+
+   <p align="center">
+   <img src="images/duplicate_op2.png" width="90%" ><br/>
 
-   We share variable in the same scope, as a result, duplicate operator `OutputGradients` will overwirte then duplicate variable.  
+   2. replace shared variable gradient with `Add` Operator
 
-   ![./images/duplicate_op]()
+   </p>
 
-    Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator instead. 
 
-![./images/duplicate_op2]()
 
-​	Then collect the sub graph OutputGradients/InputGradients as the NetOp's and return it.
+​	Then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it.

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -1344,7 +1344,7 @@ void RecurrentGradientMachine::fillGenOutputs() {
       CHECK(!finalPaths_[i].empty());
       Path& path = finalPaths_[i][0];
       generator_.ids.insert(
-          generator_.ids.begin(), path.ids.begin(), path.ids.end());
+          generator_.ids.end(), path.ids.begin(), path.ids.end());
       starts[i + 1] = starts[i] + path.ids.size();
     }
   }

paddle/gserver/layers/ConvBaseLayer.cpp

@@ -32,9 +32,11 @@ bool ConvBaseLayer::init(const LayerMap& layerMap,
     const ConvConfig& conf = inputConfig.conv_conf();
     padding_.push_back(conf.padding());
     stride_.push_back(conf.stride());
+    dilation_.push_back(conf.dilation());
     filterSize_.push_back(conf.filter_size());
     paddingY_.push_back(conf.padding_y());
     strideY_.push_back(conf.stride_y());
+    dilationY_.push_back(conf.dilation_y());
     filterSizeY_.push_back(conf.filter_size_y());
     filterPixels_.push_back(filterSize_.back() * filterSizeY_.back());
     channels_.push_back(conf.channels());
@@ -89,7 +91,11 @@ size_t ConvBaseLayer::calOutputSize() {
   size_t layerSize = 0;
 
   auto setLayerSize = [&](IntV& inH, IntV& inW, IntV& outH, IntV& outW) {
+    size_t filterSizeY;
+    size_t filterSize;
     for (size_t i = 0; i < inputLayers_.size(); i++) {
+      filterSizeY = (filterSizeY_[i] - 1) * dilationY_[i] + 1;
+      filterSize = (filterSize_[i] - 1) * dilation_[i] + 1;
       inH.push_back(inputLayers_[i]->getOutput().getFrameHeight());
       inW.push_back(inputLayers_[i]->getOutput().getFrameWidth());
       const ConvConfig& conf = config_.inputs(i).conv_conf();
@@ -98,17 +104,17 @@ size_t ConvBaseLayer::calOutputSize() {
           inH[i] = conf.has_output_y() ? conf.output_y() : conf.output_x();
         if (inW[i] == 0) inW[i] = conf.output_x();
         outH.push_back(imageSize(
-            inH[i], filterSizeY_[i], paddingY_[i], strideY_[i], caffeMode_));
-        outW.push_back(imageSize(
-            inW[i], filterSize_[i], padding_[i], stride_[i], caffeMode_));
+            inH[i], filterSizeY, paddingY_[i], strideY_[i], caffeMode_));
+        outW.push_back(
+            imageSize(inW[i], filterSize, padding_[i], stride_[i], caffeMode_));
       } else {
         if (inH[i] == 0)
           inH[i] = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
         if (inW[i] == 0) inW[i] = conf.img_size();
         outH.push_back(outputSize(
-            inH[i], filterSizeY_[i], paddingY_[i], strideY_[i], caffeMode_));
+            inH[i], filterSizeY, paddingY_[i], strideY_[i], caffeMode_));
         outW.push_back(outputSize(
-            inW[i], filterSize_[i], padding_[i], stride_[i], caffeMode_));
+            inW[i], filterSize, padding_[i], stride_[i], caffeMode_));
       }
       CHECK_EQ(outH[i], outH[0]);
       CHECK_EQ(outW[i], outW[0]);

paddle/gserver/layers/ConvBaseLayer.h

@@ -40,6 +40,10 @@ protected:
   IntV stride_;
   /// The y dimension of the stride.
   IntV strideY_;
+  /// The x dimension of the dilation.
+  IntV dilation_;
+  /// The y dimension of the dilation.
+  IntV dilationY_;
   /// The x dimension of a filter kernel.
   IntV filterSize_;
   /// The y dimension of a filter kernel.

paddle/gserver/layers/ConvBaseOperator.cpp

@@ -59,7 +59,8 @@ void ConvBaseOperator::allocConvWorkSpace() {
                     &bwdDataAlgo_,
                     &bwdDataLimitBytes_,
                     &bwdFilterAlgo_,
-                    &bwdFilterLimitBytes_);
+                    &bwdFilterLimitBytes_,
+                    /*useDilation*/ false);
 
   size_t maxWorkSpace = 0;
   maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_);

paddle/gserver/layers/ConvBaseProjection.cpp

@@ -41,6 +41,11 @@ void ConvBaseProjection::getConvParams() {
   strideH_ = conf.stride_y();
   strideW_ = conf.stride();
 
+  dilationH_ = conf.dilation_y();
+  dilationW_ = conf.dilation();
+  CHECK_GT(dilationH_, 0);
+  CHECK_GT(dilationW_, 0);
+
   filterH_ = conf.filter_size_y();
   filterW_ = conf.filter_size();
 
@@ -77,7 +82,9 @@ void ConvBaseProjection::initCudnn() {
                                    paddingH_,
                                    paddingW_,
                                    strideH_,
-                                   strideW_);
+                                   strideW_,
+                                   dilationH_,
+                                   dilationW_);
 
   // initialize all to default algorithms
   fwdAlgo_ = 0;
@@ -131,7 +138,9 @@ void ConvBaseProjection::reshapeTensorDesc(int batchSize) {
                                   paddingH_,
                                   paddingW_,
                                   strideH_,
-                                  strideW_);
+                                  strideW_,
+                                  dilationH_,
+                                  dilationW_);
 }
 
 void ConvBaseProjection::reshape(int batchSize) {
@@ -140,6 +149,10 @@ void ConvBaseProjection::reshape(int batchSize) {
   CHECK_EQ(calInputSize(), in_->value->getWidth());
 
   reshapeTensorDesc(batchSize);
+  bool useDilation = false;
+  if (dilationH_ > 1 || dilationW_ > 1) {
+    useDilation = true;
+  }
   hl_conv_workspace(imageDesc_,
                     outputDesc_,
                     filterDesc_,
@@ -149,7 +162,8 @@ void ConvBaseProjection::reshape(int batchSize) {
                     &bwdDataAlgo_,
                     &bwdDataLimitBytes_,
                     &bwdFilterAlgo_,
-                    &bwdFilterLimitBytes_);
+                    &bwdFilterLimitBytes_,
+                    useDilation);
 
   size_t maxWorkSpace = 0;
   maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_);

paddle/gserver/layers/ConvBaseProjection.h

@@ -63,6 +63,7 @@ protected:
   int configChannels_, configNumFilters_;
   int paddingH_, paddingW_;
   int strideH_, strideW_;
+  int dilationH_, dilationW_;
   int filterH_, filterW_;
   /// One group offset of input data.
   int inputOffset_;

paddle/gserver/layers/ConvProjection.cpp

@@ -25,12 +25,12 @@ size_t ConvProjection::calOutputSize() {
   if (imageH_ == 0) imageH_ = configImgH_;
   if (imageW_ == 0) imageW_ = configImgW_;
   outputH_ = outputSize(imageH_,
-                        filterH_,
+                        (filterH_ - 1) * dilationH_ + 1,
                         paddingH_,
                         strideH_,
                         /* caffeMode */ true);
   outputW_ = outputSize(imageW_,
-                        filterW_,
+                        (filterW_ - 1) * dilationW_ + 1,
                         paddingW_,
                         strideW_,
                         /* caffeMode */ true);

paddle/gserver/layers/SequenceSliceLayer.cpp

@@ -130,6 +130,8 @@ void SequenceSliceLayer::calSelectedRows(const MatrixPtr starts,
   CHECK(starts || ends) << "At least one of the start or end indices "
                         << "should be given.";
 
+  bool hasSubseq = getInput(0).hasSubseq();
+
   outSeqStartPos_.resize(1, 0);
   outSubSeqStartPos_.resize(1, 0);
   selectedRows_.clear();
@@ -151,14 +153,13 @@ void SequenceSliceLayer::calSelectedRows(const MatrixPtr starts,
         int seqLen = endPos - begPos + 1;
         CHECK_GT(seqLen, 0U);
         for (int m = begPos; m <= endPos; ++m) selectedRows_.push_back(m);
-        inputSeqInfoVec_.size() > 1
+        hasSubseq
             ? outSubSeqStartPos_.push_back(outSubSeqStartPos_.back() + seqLen)
             : outSeqStartPos_.push_back(outSeqStartPos_.back() + seqLen);
       }
       rowIdx++;
     }
-    if (inputSeqInfoVec_.size() > 1)
-      outSeqStartPos_.push_back(outSubSeqStartPos_.back());
+    if (hasSubseq) outSeqStartPos_.push_back(outSubSeqStartPos_.back());
   }
 
   if (useGpu_) {
@@ -175,7 +176,7 @@ void SequenceSliceLayer::calSelectedRows(const MatrixPtr starts,
   output_.sequenceStartPositions->copyFrom(
       outSeqStartPos_.data(), outSeqStartPos_.size(), false);
 
-  if (inputSeqInfoVec_.size() > 1) {
+  if (hasSubseq) {
     ICpuGpuVector::resizeOrCreate(
         output_.subSequenceStartPositions, outSubSeqStartPos_.size(), false);
     output_.subSequenceStartPositions->copyFrom(
@@ -204,10 +205,11 @@ void SequenceSliceLayer::forward(PassType passType) {
     copySliceIdsToCpu();
   }
 
-  // calculate the selected row indices in a batch,
-  // and build the output sequence information.
-  calSelectedRows(startIdsOnCpu_ ? startIdsOnCpu_ : nullptr,
-                  endIdsOnCpu_ ? endIdsOnCpu_ : nullptr);
+  /*
+   * calculate the selected row indices in a batch, and build the output
+   * sequence information.
+   */
+  calSelectedRows(startIdsOnCpu_, endIdsOnCpu_);
 
   resetOutput(selectedRows_.size(), getSize());
 

paddle/gserver/tests/test_LayerGrad.cpp

@@ -12,6 +12,9 @@ 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. */
 
+#ifndef PADDLE_ONLY_CPU
+#include <cudnn.h>
+#endif
 #include <gtest/gtest.h>
 #include <string>
 #include <vector>
@@ -189,10 +192,16 @@ TEST(Projection, scaling) {
 void testProjectionConv(size_t groups, bool isDeconv) {
   const int NUM_FILTERS = 18;
   const int FILTER_SIZE = 2;
-  const int FILTER_SIZE_Y = 4;
+  const int FILTER_SIZE_Y = 2;
   const int CHANNELS = 3;
   const int IMAGE_SIZE = 16;
 
+#if CUDNN_VERSION >= 6000
+  const int DILATION = 2;
+#else
+  const int DILATION = 1;
+#endif
+
   ProjectionConfig conf;
   if (isDeconv) {
     conf.set_type("convt");
@@ -209,6 +218,8 @@ void testProjectionConv(size_t groups, bool isDeconv) {
   conv->set_padding_y(1);
   conv->set_stride(2);
   conv->set_stride_y(2);
+  conv->set_dilation(DILATION);
+  conv->set_dilation_y(DILATION);
   conv->set_groups(groups);
   if (isDeconv) {
     conv->set_filter_channels(NUM_FILTERS / conv->groups());
@@ -217,12 +228,12 @@ void testProjectionConv(size_t groups, bool isDeconv) {
   }
   conv->set_img_size(IMAGE_SIZE);
   int output_x = outputSize(conv->img_size(),
-                            conv->filter_size(),
+                            (conv->filter_size() - 1) * DILATION + 1,
                             conv->padding(),
                             conv->stride(),
                             /* caffeMode */ true);
   int output_y = outputSize(conv->img_size(),
-                            conv->filter_size_y(),
+                            (conv->filter_size_y() - 1) * DILATION + 1,
                             conv->padding_y(),
                             conv->stride_y(),
                             /* caffeMode */ true);
@@ -424,27 +435,38 @@ void testConvLayer(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", 384, 288});
+  int dilation = 1;
+  if (type == "cudnn_conv") {
+#if CUDNN_VERSION >= 6000
+    dilation = 2;
+#else
+    dilation = 1;
+#endif
+  }
+
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 768, 192});
   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(2);
   conv->set_channels(3);
   conv->set_padding(0);
   conv->set_padding_y(1);
   conv->set_stride(2);
   conv->set_stride_y(2);
+  conv->set_dilation(dilation);
+  conv->set_dilation_y(dilation);
   conv->set_groups(1);
   conv->set_filter_channels(conv->channels() / conv->groups());
   conv->set_img_size(16);
-  conv->set_img_size_y(8);
+  conv->set_img_size_y(16);
   conv->set_output_x(outputSize(conv->img_size(),
-                                conv->filter_size(),
+                                (conv->filter_size() - 1) * dilation + 1,
                                 conv->padding(),
                                 conv->stride(),
                                 /* caffeMode */ true));
   conv->set_output_y(outputSize(conv->img_size_y(),
-                                conv->filter_size_y(),
+                                (conv->filter_size_y() - 1) * dilation + 1,
                                 conv->padding_y(),
                                 conv->stride_y(),
                                 /* caffeMode */ true));

paddle/gserver/tests/test_SeqSliceLayerGrad.cpp

@@ -30,6 +30,8 @@ const int MAX_SEQ_NUM = 17;
 const int MAX_SEQ_LEN = 23;
 const int MAX_BEAM_SIZE = 13;
 
+const size_t SEED = (size_t)(time(NULL));
+
 vector<real> randSampling(real range, int n) {
   CHECK_GE(range, n);
   vector<real> num(range);
@@ -46,7 +48,7 @@ void genSeqInfo(vector<int>& seqStartPos, vector<int>& subSeqStartPos) {
   seqStartPos.resize(1, 0);
   subSeqStartPos.resize(1, 0);
 
-  srand((size_t)(time(NULL)));
+  srand(SEED);
   int seqNum = 1 + (rand() % MAX_SEQ_NUM);
   for (int i = 0; i < seqNum; ++i) {
     int subSeqNum = 1 + (rand() % MAX_SEQ_NUM);

paddle/operators/CMakeLists.txt

@@ -42,10 +42,12 @@ function(op_library TARGET)
 endfunction()
 
 add_subdirectory(math)
+
 cc_test(gather_test SRCS gather_test.cc DEPS tensor)
 op_library(gather_op SRCS gather_op.cc gather_op.cu)
 
 cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
+op_library(scatter_op SRCS scatter_op.cc scatter_op.cu)
 
 cc_library(net_op SRCS net_op.cc DEPS op_registry)
 cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
@@ -67,7 +69,7 @@ op_library(sgd_op SRCS sgd_op.cc sgd_op.cu)
 
 op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
     DEPS framework_proto tensor op_registry operator net_op)
-op_library(uniform_random_op
-        SRCS uniform_random_op.cc uniform_random_op.cu)
+op_library(uniform_random_op SRCS uniform_random_op.cc uniform_random_op.cu)
+op_library(lookup_table_op SRCS lookup_table_op.cc lookup_table_op.cu)
 op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op)
 op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op)

paddle/operators/fill_zeros_like_op.h

@@ -26,7 +26,7 @@ class FillZerosLikeKernel : public framework::OpKernel {
     auto* output = context.Output<framework::Tensor>("Dst");
     output->mutable_data<T>(context.GetPlace());
     auto t = framework::EigenVector<T>::Flatten(*output);
-    t.device(context.GetEigenDevice<Place>()) = t.constant(T(0));
+    t.device(context.GetEigenDevice<Place>()) = t.constant(static_cast<T>(0));
   }
 };
 

paddle/operators/lookup_table_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/lookup_table_op.h"
+
+namespace paddle {
+namespace operators {
+
+class LookupTableOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &context) const override {
+    auto table_t = context.Input<Tensor>("W");
+    auto ids_t = context.Input<Tensor>("Ids");
+    auto output_t = context.Output<Tensor>("Out");
+
+    output_t->Resize({ids_t->dims()[0], table_t->dims()[1]});
+  }
+};
+
+class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  LookupTableOpMaker(framework::OpProto *proto,
+                     framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("W",
+             "An input represents embedding tensors,"
+             " which is a learnable parameter.");
+    AddInput("Ids",
+             "An input with type int32 or int64"
+             "contains the ids to be looked up in W.");
+    AddOutput("Out", "The lookup results, which have the same type with W.");
+    AddComment(
+        "This operator is used to perform lookups on the parameter W,"
+        "then concatenated into a dense tensor.");
+  }
+};
+
+class LookupTableOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &context) const override {
+    auto table = context.Input<Tensor>("W");
+    auto d_table = context.Output<Tensor>(framework::GradVarName("W"));
+    d_table->Resize(table->dims());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(lookup_table, ops::LookupTableOp, ops::LookupTableOpMaker,
+            lookup_table_grad, ops::LookupTableOpGrad);
+
+REGISTER_OP_CPU_KERNEL(lookup_table, ops::LookupTableKernel<float>);
+REGISTER_OP_CPU_KERNEL(lookup_table_grad, ops::LookupTableGradKernel<float>);

paddle/operators/lookup_table_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/platform/assert.h"
+#include "paddle/platform/cuda_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T, int BlockDimX, int BlockDimY, int GridDimX>
+__global__ void LookupTable(T* output, const T* table, const int32_t* ids,
+                            const int N, const int K, const int D) {
+  int idx = threadIdx.x;
+  int idy = blockIdx.x + threadIdx.y * GridDimX;
+
+  while (idy < K) {
+    int id = ids[idy];
+    PADDLE_ASSERT(id >= 0);
+    PADDLE_ASSERT(id < N);
+    T* out = output + idy * D;
+    const T* tab = table + id * D;
+    for (int i = idx; i < D; i += BlockDimX) {
+      out[i] = tab[i];
+    }
+    idy += BlockDimY * GridDimX;
+  }
+}
+
+template <typename T, int BlockDimX, int BlockDimY, int GridDimX>
+__global__ void LookupTableGrad(T* table, const T* output, const int32_t* ids,
+                                const int N, const int K, const int D) {
+  int idx = threadIdx.x;
+  int idy = blockIdx.x + threadIdx.y * GridDimX;
+
+  while (idy < K) {
+    int id = ids[idy];
+    PADDLE_ASSERT(id >= 0);
+    PADDLE_ASSERT(id < N);
+    const T* out = output + idy * D;
+    T* tab = table + id * D;
+    for (int i = idx; i < D; i += BlockDimX) {
+      paddle::platform::CudaAtomicAdd(&tab[i], out[i]);
+    }
+    idy += BlockDimY * GridDimX;
+  }
+}
+
+template <typename T>
+class LookupTableCUDAKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto table_t = context.Input<Tensor>("W");
+    auto ids_t = context.Input<Tensor>("Ids");
+    auto output_t = context.Output<Tensor>("Out");
+
+    size_t N = table_t->dims()[0];
+    size_t D = table_t->dims()[1];
+    size_t K = product(ids_t->dims());
+    auto ids = ids_t->data<int32_t>();
+    auto table = table_t->data<T>();
+    auto output = output_t->mutable_data<T>(context.GetPlace());
+
+    dim3 threads(128, 8);
+    dim3 grids(8, 1);
+    LookupTable<T, 128, 8, 8><<<grids, threads>>>(output, table, ids, N, K, D);
+  }
+};
+
+template <typename T>
+class LookupTableGradCUDAKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto ids_t = context.Input<Tensor>("Ids");
+    auto d_output_t = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto d_table_t = context.Output<Tensor>(framework::GradVarName("W"));
+
+    int N = d_table_t->dims()[0];
+    int D = d_table_t->dims()[1];
+    int K = product(ids_t->dims());
+    const int32_t* ids = ids_t->data<int32_t>();
+    const T* d_output = d_output_t->data<T>();
+    T* d_table = d_table_t->mutable_data<T>(context.GetPlace());
+
+    auto t = framework::EigenVector<T>::Flatten(*d_table_t);
+    t.device(context.GetEigenDevice<platform::GPUPlace>()) =
+        t.constant(static_cast<T>(0));
+
+    dim3 threads(128, 8);
+    dim3 grids(8, 1);
+    LookupTableGrad<T, 128, 8, 8><<<grids, threads>>>(d_table, d_output, ids, N,
+                                                      K, D);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(lookup_table, ops::LookupTableCUDAKernel<float>);
+REGISTER_OP_GPU_KERNEL(lookup_table_grad,
+                       ops::LookupTableGradCUDAKernel<float>);

paddle/operators/lookup_table_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T>
+class LookupTableKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto table_t = context.Input<Tensor>("W");      // float tensor
+    auto ids_t = context.Input<Tensor>("Ids");      // int tensor
+    auto output_t = context.Output<Tensor>("Out");  // float tensor
+
+    size_t N = table_t->dims()[0];
+    size_t D = table_t->dims()[1];
+    auto ids = ids_t->data<int32_t>();
+    auto table = table_t->data<T>();
+    auto output = output_t->mutable_data<T>(context.GetPlace());
+    for (size_t i = 0; i < product(ids_t->dims()); ++i) {
+      PADDLE_ENFORCE_LT(ids[i], N);
+      PADDLE_ENFORCE_GE(ids[i], 0);
+      memcpy(output + i * D, table + ids[i] * D, D * sizeof(T));
+    }
+  }
+};
+
+template <typename T>
+class LookupTableGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto ids_t = context.Input<Tensor>("Ids");
+    auto d_output_t = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto d_table_t = context.Output<Tensor>(framework::GradVarName("W"));
+
+    size_t N = d_table_t->dims()[0];
+    size_t D = d_table_t->dims()[1];
+    auto ids = ids_t->data<int32_t>();
+    const T* d_output = d_output_t->data<T>();
+    T* d_table = d_table_t->mutable_data<T>(context.GetPlace());
+
+    auto t = framework::EigenVector<T>::Flatten(*d_table_t);
+    t.device(context.GetEigenDevice<platform::CPUPlace>()) =
+        t.constant(static_cast<T>(0));
+
+    for (size_t i = 0; i < product(ids_t->dims()); ++i) {
+      PADDLE_ENFORCE_LT(ids[i], N);
+      PADDLE_ENFORCE_GE(ids[i], 0);
+      for (size_t j = 0; j < D; ++j) {
+        d_table[ids[i] * D + j] += d_output[i * D + j];
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/scatter_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/scatter_op.h"
+#include "paddle/framework/ddim.h"
+
+namespace paddle {
+namespace operators {
+
+class ScatterOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Index")->dims().size(), 1,
+                      "Update Index should be 1-D.");
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Ref")->dims().size(),
+                      ctx.Input<Tensor>("Updates")->dims().size(),
+                      "Reference and Updates should have the same shape size");
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Updates")->dims()[0],
+                      ctx.Input<Tensor>("Index")->dims()[0],
+                      "Updates and Index should have same batch-size.");
+    framework::DDim data_dim(ctx.Input<Tensor>("Updates")->dims());
+    for (int i = 1; i < data_dim.size(); ++i)
+      PADDLE_ENFORCE_EQ(data_dim[i], ctx.Input<Tensor>("Updates")->dims()[i]);
+    ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("Ref")->dims());
+  }
+};
+
+class ScatterGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto *dUpdates = ctx.Output<Tensor>(framework::GradVarName("Updates"));
+    auto *Updates = ctx.Input<Tensor>("Updates");
+    auto *dRef = ctx.Output<Tensor>(framework::GradVarName("Ref"));
+    auto *Ref = ctx.Input<Tensor>("Ref");
+
+    dRef->Resize(Ref->dims());
+    dUpdates->Resize(Updates->dims());
+  }
+};
+
+class ScatterOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ScatterOpMaker(framework::OpProto *proto,
+                 framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Ref", "The source input of scatter op");
+    AddInput("Index",
+             "The index input of scatter op where Ref will be updated");
+    AddInput("Updates", "The updated value of updates op");
+    AddOutput("Out", "The output of add op");
+    AddComment(R"DOC(
+Scatter Operator by selecting from the first axis, 
+
+Out = Ref
+Out[Index] = Ref[Index] + Updates
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(scatter, ops::ScatterOp, ops::ScatterOpMaker, scatter_grad,
+            ops::ScatterGradOp);
+REGISTER_OP_CPU_KERNEL(scatter,
+                       ops::ScatterOpKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    scatter_grad,
+    ops::ScatterGradientOpKernel<paddle::platform::CPUPlace, float>);

paddle/operators/scatter_op.cu

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

paddle/operators/scatter_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "gather.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "scatter.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename Place, typename T>
+class ScatterOpKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto *Ref = ctx.Input<Tensor>("Ref");
+    auto *Index = ctx.Input<Tensor>("Index");
+    auto *Updates = ctx.Input<Tensor>("Updates");
+    auto *Out = ctx.Output<Tensor>("Out");
+
+    // In place output: Out = Ref, Out[Index] += Updates
+    Out->ShareDataWith<T>(*Ref);
+    // Apply ScatterUpdate: Out[index] += Updates[:]
+    ScatterUpdate<T>(ctx.GetPlace(), Updates, Index, Out);
+  }
+};
+
+template <typename Place, typename T>
+class ScatterGradientOpKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto *dRef = ctx.Output<Tensor>(framework::GradVarName("Ref"));
+    auto *dUpdates = ctx.Output<Tensor>(framework::GradVarName("Updates"));
+    auto *Index = ctx.Input<Tensor>("Index");
+    auto *dOut = ctx.Input<Tensor>(framework::GradVarName("Out"));
+
+    // In place gradient: dRef = dO
+    dRef->ShareDataWith<T>(*dOut);
+    dUpdates->mutable_data<T>(ctx.GetPlace());
+    // Gradient by Gather: dUpdates += dO[Index]
+    Gather<T>(ctx.GetPlace(), dOut, Index, dUpdates);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/platform/cuda_helper.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 <cuda.h>
+
+namespace paddle {
+namespace platform {
+
+#define CUDA_ATOMIC_WRAPPER(op, T) \
+  __device__ __forceinline__ T CudaAtomic##op(T* address, const T val)
+
+#define USE_CUDA_ATOMIC(op, T) \
+  CUDA_ATOMIC_WRAPPER(op, T) { return atomic##op(address, val); }
+
+// For atomicAdd.
+USE_CUDA_ATOMIC(Add, float);
+
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 600
+USE_CUDA_ATOMIC(Add, double);
+#else
+CUDA_ATOMIC_WRAPPER(Add, double) {
+  unsigned long long int* address_as_ull =
+      reinterpret_cast<unsigned long long int*>(address);
+  unsigned long long int old = *address_as_ull, assumed;
+
+  do {
+    assumed = old;
+    old = atomicCAS(address_as_ull, assumed,
+                    __double_as_longlong(val + __longlong_as_double(assumed)));
+
+    // Note: uses integer comparison to avoid hang in case of NaN
+  } while (assumed != old);
+
+  return __longlong_as_double(old);
+}
+#endif
+
+}  // namespace platform
+}  // namespace paddle

paddle/pybind/CMakeLists.txt

@@ -4,6 +4,7 @@ cc_library(paddle_pybind SHARED
     DEPS pybind python backward
     sgd_op
     gather_op
+    scatter_op
     add_op
     mul_op
     rowwise_add_op
@@ -15,6 +16,7 @@ cc_library(paddle_pybind SHARED
     uniform_random_op
     gaussian_random_op
     fill_zeros_like_op
+    lookup_table_op
     scale_op
     minus_op)
 endif(WITH_PYTHON)

paddle/pybind/pybind.cc

@@ -42,10 +42,12 @@ USE_OP(fill_zeros_like);
 USE_OP_ITSELF(recurrent_op);
 USE_OP(gaussian_random);
 USE_OP(uniform_random);
+USE_OP(lookup_table);
 USE_OP(scale);
 USE_OP_ITSELF(identity);
 USE_OP(minus);
 USE_CPU_ONLY_OP(gather);
+USE_CPU_ONLY_OP(scatter);
 
 namespace paddle {
 namespace framework {

paddle/scripts/docker/build.sh

@@ -38,7 +38,7 @@ Configuring cmake in /paddle/build ...
       -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON}
       -DCUDNN_ROOT=/usr/
       -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF}
-      -DWITH_TESTING=${WITH_TESTING:-OFF}
+      -DWITH_TESTING=${WITH_TESTING:-ON}
       -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
 ========================================
 EOF
@@ -56,19 +56,18 @@ cmake .. \
       -DWITH_C_API=${WITH_C_API:-OFF} \
       -DWITH_PYTHON=${WITH_PYTHON:-ON} \
       -DCUDNN_ROOT=/usr/ \
-      -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} \
-      -DWITH_TESTING=${WITH_TESTING:-OFF} \
+      -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON} \
+      -DWITH_TESTING=${WITH_TESTING:-ON} \
       -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
 
 cat <<EOF
 ============================================
 Building in /paddle/build ...
-   Build unit tests: ${WITH_TESTING:-OFF}
 ============================================
 EOF
 make -j `nproc`
 
-if [ ${WITH_TESTING:-OFF} == "ON" ] && [ ${RUN_TEST:-OFF} == "ON" ] ; then
+if [ ${WITH_TESTING:-ON} == "ON" ] && [ ${RUN_TEST:-OFF} == "ON" ] ; then
 cat <<EOF
 ========================================
 Running unit tests ...

proto/ModelConfig.proto

@@ -82,6 +82,9 @@ message ConvConfig {
 
   // if not set, use img_size
   optional uint32 img_size_y = 14;
+
+  optional uint32 dilation = 15 [ default = 1 ];
+  optional uint32 dilation_y = 16 [ default = 1 ];
 }
 
 message PoolConfig {

python/paddle/trainer/config_parser.py

@@ -870,12 +870,16 @@ class Conv(Cfg):
                  caffe_mode=True,
                  filter_size_y=None,
                  padding_y=None,
-                 stride_y=None):
+                 stride_y=None,
+                 dilation=None,
+                 dilation_y=None):
         self.add_keys(locals())
         if filter_size_y is None:
             self.filter_size_y = filter_size
         if padding_y is None:
             self.padding_y = padding
+        if dilation_y is None:
+            self.dilation_y = dilation
         if stride_y is None:
             self.stride_y = stride
         if output_x is not None:

python/paddle/trainer_config_helpers/layers.py

@@ -2344,6 +2344,7 @@ def img_conv_layer(input,
                    groups=1,
                    stride=1,
                    padding=0,
+                   dilation=1,
                    bias_attr=None,
                    param_attr=None,
                    shared_biases=True,
@@ -2351,6 +2352,7 @@ def img_conv_layer(input,
                    filter_size_y=None,
                    stride_y=None,
                    padding_y=None,
+                   dilation_y=None,
                    trans=False,
                    layer_type=None):
     """
@@ -2415,6 +2417,11 @@ def img_conv_layer(input,
     :type padding: int|tuple|list
     :param padding_y: The y dimension of the padding.
     :type padding_y: int
+    :param dilation: The x dimension of the dilation. Or input a tuple for two
+                    image dimension
+    :type dilation: int|tuple|list
+    :param dilation_y: The y dimension of the dilation.
+    :type dilation_y: int
     :param bias_attr: Convolution bias attribute. None means default bias.
                       False means no bias.
     :type bias_attr: ParameterAttribute|False
@@ -2462,6 +2469,13 @@ def img_conv_layer(input,
         else:
             padding_y = padding
 
+    if dilation_y is None:
+        if isinstance(dilation, collections.Sequence):
+            assert len(dilation) == 2
+            dilation, dilation_y = dilation
+        else:
+            dilation_y = dilation
+
     if param_attr.attr.get('initial_smart'):
         # special initial for conv layers.
         init_w = (2.0 / (filter_size**2 * num_channels))**0.5
@@ -2471,6 +2485,8 @@ def img_conv_layer(input,
         param_attr.attr["initial_smart"] = False
 
     if layer_type:
+        if dilation > 1 or dilation_y > 1:
+            assert layer_type in ["cudnn_conv", "cudnn_convt"]
         if trans:
             assert layer_type in ["exconvt", "cudnn_convt"]
         else:
@@ -2486,11 +2502,13 @@ def img_conv_layer(input,
             conv=Conv(
                 filter_size=filter_size,
                 padding=padding,
+                dilation=dilation,
                 stride=stride,
                 channels=num_channels,
                 groups=groups,
                 filter_size_y=filter_size_y,
                 padding_y=padding_y,
+                dilation_y=dilation_y,
                 stride_y=stride_y),
             **param_attr.attr),
         active_type=act.name,
@@ -2591,15 +2609,15 @@ def img_pool_layer(input,
         assert input.num_filters is not None
         num_channels = input.num_filters
 
-    assert type(pool_type) in [AvgPooling, MaxPooling, CudnnAvgPooling,
-                               CudnnMaxPooling], \
-        "only (Cudnn)AvgPooling, (Cudnn)MaxPooling are supported"
-
     if pool_type is None:
         pool_type = MaxPooling()
     elif isinstance(pool_type, AvgPooling):
         pool_type.name = 'avg'
 
+    assert type(pool_type) in [AvgPooling, MaxPooling, CudnnAvgPooling,
+                               CudnnMaxPooling], \
+        "only (Cudnn)AvgPooling, (Cudnn)MaxPooling are supported"
+
     type_name = pool_type.name + '-projection' \
         if (
         isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \

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

@@ -12,6 +12,7 @@ img_conv = img_conv_layer(
     num_filters=64,
     filter_size=(32, 32),
     padding=(1, 1),
+    dilation=(1, 1),
     stride=(1, 1),
     act=LinearActivation())
 img_bn = batch_norm_layer(input=img_conv, act=ReluActivation())

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

@@ -14,6 +14,7 @@ py_test(test_sigmoid_op SRCS test_sigmoid_op.py)
 py_test(test_softmax_op SRCS test_softmax_op.py)
 py_test(test_cross_entropy_op SRCS test_cross_entropy_op.py)
 py_test(test_gather_op SRCS test_gather_op.py)
+py_test(test_scatter_op SRCS test_scatter_op.py)
 py_test(test_fill_zeros_like_op SRCS test_fill_zeros_like_op.py)
 
 py_test(gradient_checker SRCS gradient_checker.py)
@@ -28,5 +29,6 @@ py_test(test_uniform_random_op SRCS test_uniform_random_op.py)
 py_test(test_recurrent_op SRCS test_recurrent_op.py)
 py_test(test_sgd_op SRCS test_sgd_op.py)
 py_test(test_gradient_checker SRCS test_gradient_checker.py)
+py_test(test_lookup_table SRCS test_lookup_table.py)
 py_test(test_scale_and_identity_op SRCS test_scale_and_identity_op.py)
 py_test(mnist SRCS mnist.py)

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

@@ -23,12 +23,17 @@ def grad_var_name(var_name):
     return var_name + "@GRAD"
 
 
+def empty_var_name():
+    return "@EMPTY@"
+
+
 def get_numeric_gradient(op,
                          input_values,
                          output_name,
                          input_to_check,
                          delta=0.005,
-                         local_scope=None):
+                         local_scope=None,
+                         in_place=False):
     """
     Get Numeric Gradient for an operator's input.
     
@@ -77,6 +82,11 @@ def get_numeric_gradient(op,
     def product(dim):
         return reduce(lambda a, b: a * b, dim, 1)
 
+    def restore_inputs():
+        for var_name in input_values:
+            tensor_ = local_scope.find_var(var_name).get_tensor()
+            tensor_.set(numpy.copy(input_values[var_name]), core.CPUPlace())
+
     # get the input tensor that we want to get it's numeric gradient.
     tensor_to_check = local_scope.find_var(input_to_check).get_tensor()
     tensor_size = product(tensor_to_check.get_dims())
@@ -86,6 +96,8 @@ def get_numeric_gradient(op,
     # we only compute gradient of one element each time.
     # we use a for loop to compute the gradient of every element.
     for i in xrange(tensor_size):
+        if in_place:
+            restore_inputs()
         # get one input element throw it's index i.
         origin = tensor_to_check.get_float_element(i)
 
@@ -95,6 +107,8 @@ def get_numeric_gradient(op,
         y_pos = get_output()
 
         # plus delta to this element, run op and get the sum of the result tensor.
+        if in_place:
+            restore_inputs()
         x_neg = origin - delta
         tensor_to_check.set_float_element(i, x_neg)
         y_neg = get_output()
@@ -176,7 +190,7 @@ class GradientChecker(unittest.TestCase):
         ]
         return outs
 
-    def compare_grad(self, forward_op, input_value):
+    def compare_grad(self, forward_op, input_value, no_grad_set=None):
         """ Compare the input gradients between CPU and GPU for the given forward
         operator.
 
@@ -184,15 +198,20 @@ class GradientChecker(unittest.TestCase):
         :type forward_op: Operator
         :param input_value: input values.
         :type input_value: dict{string:numpy.array}
+        :param no_grad_set: the set of variables names without gradients.
+        :type no_grad_set: a set of string
         :raises: AssertionError, there is different gradient value.
         """
-        backward_op = core.Operator.backward(forward_op, set())
+        if no_grad_set is None:
+            no_grad_set = set()
+        backward_op = core.Operator.backward(forward_op, no_grad_set)
         # return if not compile with GPU or not implementing GPU kernel
         if not (core.is_compile_gpu() and backward_op.support_gpu()):
             return
 
         outputs = backward_op.outputs()
         out_names = [item for k in outputs for item in outputs[k]]
+        out_names = filter(lambda x: x != empty_var_name(), out_names)
         cpu_grads = self.__get_gradient(forward_op, backward_op, input_value,
                                         out_names, core.CPUPlace())
         gpu_grads = self.__get_gradient(forward_op, backward_op, input_value,
@@ -242,6 +261,7 @@ class GradientChecker(unittest.TestCase):
                    output_name,
                    no_grad_set=None,
                    only_cpu=False,
+                   in_place=False,
                    max_relative_error=0.005):
         """
         :param forward_op: used to create backward_op
@@ -274,7 +294,8 @@ class GradientChecker(unittest.TestCase):
 
         # get numerical gradients
         numeric_grads = [
-            get_numeric_gradient(forward_op, input_vars, output_name, name)
+            get_numeric_gradient(
+                forward_op, input_vars, output_name, name, in_place=in_place)
             for name in inputs_to_check
         ]
 

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

@@ -181,7 +181,7 @@ images = data_layer(name='pixel', dims=[BATCH_SIZE, 784])
 labels = data_layer(name='label', dims=[BATCH_SIZE])
 fc1 = fc_layer(net=forward_net, input=images, size=100, act="sigmoid")
 fc2 = fc_layer(net=forward_net, input=fc1, size=100, act="sigmoid")
-predict = fc_layer(net=forward_net, input=fc2, size=100, act="softmax")
+predict = fc_layer(net=forward_net, input=fc2, size=10, act="softmax")
 cost = cross_entropy_layer(net=forward_net, input=predict, label=labels)
 
 init_net.complete_add_op(True)

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

@@ -21,12 +21,9 @@ class TestGatherOp(unittest.TestCase):
 
 class TestGatherGradOp(GradientChecker):
     def test_gather_grad(self):
-        print 'creating op'
         op = create_op("gather")
-        print 'creating op done'
         xnp = numpy.random.random((10, 20)).astype("float32")
         inputs = {'X': xnp, 'Index': numpy.array([1, 3, 5]).astype("int32")}
-        print 'correct before check gradient'
         self.check_grad(op, inputs, set("X"), "Out")
 
 

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

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

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

+import unittest
+from op_test_util import OpTestMeta
+from gradient_checker import GradientChecker, create_op
+import numpy
+import paddle.v2.framework.core as core
+from paddle.v2.framework.op import Operator
+
+
+class TestScatterOp(unittest.TestCase):
+    __metaclass__ = OpTestMeta
+
+    def setUp(self):
+        self.type = "scatter"
+        ref_np = numpy.ones((3, 3)).astype("float32")
+        index_np = numpy.array([1, 2]).astype("int32")
+        updates_np = numpy.random.random((2, 3)).astype("float32")
+        output_np = numpy.copy(ref_np)
+        output_np[index_np] += updates_np
+        self.inputs = {'Ref': ref_np, 'Index': index_np, 'Updates': updates_np}
+        self.outputs = {'Out': output_np}
+
+
+class TestScatterGradOp(GradientChecker):
+    def test_scatter_grad(self):
+        op = create_op("scatter")
+        # test data setup
+        ref_np = numpy.ones((3, 10)).astype("float32")
+        index_np = numpy.array([1, 2]).astype("int32")
+        updates_np = numpy.random.random((2, 10)).astype("float32")
+        output_np = numpy.copy(ref_np)
+        output_np[index_np] += updates_np
+        inputs = {'Ref': ref_np, 'Index': index_np, 'Updates': updates_np}
+        self.check_grad(
+            op, inputs, set(["Updates", "Ref"]), "Out", in_place=True)
+
+
+if __name__ == "__main__":
+    unittest.main()