Merge branch 'develop' into kmax_score_layer

Dockerfile

@@ -38,17 +38,16 @@ RUN apt-get update && \
 RUN pip --no-cache-dir install 'numpy>=1.12.0'
 
 # Install Go and glide
-RUN wget -O go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz && \
-    tar -C /usr/local -xzf go.tgz && \
+RUN wget -qO- https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \
+    tar -xz -C /usr/local && \
     mkdir /root/gopath && \
     mkdir /root/gopath/bin && \
-    mkdir /root/gopath/src && \
-    rm go.tgz
+    mkdir /root/gopath/src
 ENV GOROOT=/usr/local/go GOPATH=/root/gopath
 # should not be in the same line with GOROOT definition, otherwise docker build could not find GOROOT.
 ENV PATH=${PATH}:${GOROOT}/bin:${GOPATH}/bin
 # install glide
-RUN curl -q https://glide.sh/get | sh
+RUN curl -s -q https://glide.sh/get | sh
 
 # git credential to skip password typing
 RUN git config --global credential.helper store

cmake/external/any.cmake

@@ -8,7 +8,7 @@ ExternalProject_Add(
     extern_lib_any
     ${EXTERNAL_PROJECT_LOG_ARGS}
     GIT_REPOSITORY  "https://github.com/PaddlePaddle/any.git"
-    GIT_TAG         "8fef1e93710a0edf8d7658999e284a1142c4c020"
+    GIT_TAG         "15595d8324be9e8a9a80d9ae442fdd12bd66df5d"
     PREFIX          ${ANY_SOURCE_DIR}
     UPDATE_COMMAND  ""
     CONFIGURE_COMMAND ""

cmake/external/mklml.cmake

@@ -43,8 +43,8 @@ SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib")
 
 INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
 
-SET(mklml_cmakefile ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt)
-FILE(WRITE ${mklml_cmakefile} "PROJECT(MKLML)\n"
+FILE(WRITE ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt
+  "PROJECT(MKLML)\n"
   "cmake_minimum_required(VERSION 3.0)\n"
   "install(DIRECTORY ${MKLML_VER}\n"
   "        DESTINATION ${MKLML_DST_DIR})\n")
@@ -54,8 +54,7 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     PREFIX                ${MKLML_SOURCE_DIR}
     DOWNLOAD_DIR          ${MKLML_DOWNLOAD_DIR}
-    DOWNLOAD_COMMAND      wget --no-check-certificate -O ${MKLML_DOWNLOAD_DIR}/${MKLML_VER}.tgz ${MKLML_URL}
-                          && tar -xzf ${MKLML_DOWNLOAD_DIR}/${MKLML_VER}.tgz
+    DOWNLOAD_COMMAND      wget --no-check-certificate -qO- ${MKLML_URL} | tar xz -C ${MKLML_DOWNLOAD_DIR}
     DOWNLOAD_NO_PROGRESS  1
     UPDATE_COMMAND        ""
     CMAKE_ARGS            -DCMAKE_INSTALL_PREFIX=${MKLML_INSTALL_ROOT}

doc/api/v2/config/layer.rst

@@ -262,6 +262,11 @@ kmax_sequence_score
 ..  autoclass:: paddle.v2.layer.kmax_sequence_score
     :noindex:
 
+sub_nested_seq
+--------------
+..  autoclass:: paddle.v2.layer.sub_nested_seq
+    :noindex:
+
 Reshaping Layers
 ================
 

doc/design/releasing_process.md

@@ -11,6 +11,15 @@ Paddle每次发新的版本，遵循以下流程:
 	* 编译这个版本的Ubuntu Deb包。如果失败，修复Ubuntu Deb包编译问题，Patch号加一，返回第二步。
 	* 使用Regression Test List作为检查列表，测试Docker镜像/ubuntu安装包的功能正确性
 		* 如果失败，记录下所有失败的例子，在这个`release/版本号`分支中，修复所有bug后，Patch号加一，返回第二步
+	* 编译这个版本的python wheel包，并发布到pypi。
+		* 由于pypi.python.org目前遵循[严格的命名规范PEP 513](https://www.python.org/dev/peps/pep-0513)，在使用twine上传之前，需要重命名wheel包中platform相关的后缀，比如将`linux_x86_64`修改成`manylinux1_x86_64`。
+		* pypi上的package名称为paddlepaddle和paddlepaddle_gpu，如果要上传GPU版本的包，需要修改build/python/setup.py中，name: "paddlepaddle_gpu"并重新打包wheel包：`python setup.py bdist_wheel`。
+		* 上传方法：
+			```
+			cd build/python
+			pip install twine
+			twine upload dist/[package to upload]
+			```
 4. 第三步完成后，将`release/版本号`分支合入master分支，并删除`release/版本号`分支。将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。最后删除`release/版本号`分支。
 5. 编译master分支的Docker发行镜像，发布到dockerhub。编译ubuntu的deb包，发布到github release页面
 6. 协同完成Release Note的书写

doc/getstarted/build_and_install/docker_install_cn.rst

@@ -3,6 +3,43 @@ 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。
 
+Docker使用入门
+------------------------------
+
+几个基础的概念帮助理解和使用Docker：
+
+- *镜像*：一个Docker镜像是一个打包好的软件。它包含了这个软件本身和它所依赖的运行环境。PaddlePaddle的Docker镜像就包含了PaddlePaddle的Python库以及其依赖的多个Python库。这样我们可以直接在Docker中运行需要的程序而不需要安装后在执行。可以执行：
+
+  .. code-block:: bash
+
+     docker images
+
+  来列出当前系统中的所有镜像，同样可以执行：
+
+  .. code-block:: bash
+		  
+     docker pull paddlepaddle/paddle:0.10.0
+
+  来下载Docker镜像，paddlepaddle/paddle是从官方镜像源Dockerhub.com下载的，推荐国内用户使用ocker.paddlepaddle.org/paddle下载。
+
+- *容器*： 如果说一个Docker镜像就是一个程序，那容器就是这个程序运行时产生的“进程”。
+  实际上，一个容器就是一个操作系统的进程，但是是运行在独立的进程空间，文件系统以及网络之上。
+  可以执行：
+
+  .. code-block:: bash
+
+     docker run paddlepaddle/paddle:0.10.0
+
+  来使用一个镜像启动一个容器。
+
+- 默认情况下，Docker容器会运行在独立的文件系统空间之上，我们无法在Docker容器中
+  访问到主机上的文件。可以通过*挂载Volume*的方式，将主机上的文件或目录挂载到
+  Docker容器中。下面的命令把当前目录挂载到了容器中的 /data 目录下，容器使用
+  debian镜像，并且启动后执行 :code:`ls /data`。
+
+  .. code-block:: bash
+
+     docker run --rm -v $(pwd):/data debian ls /data
 
 PaddlePaddle发布的Docker镜像使用说明
 ------------------------------
@@ -12,11 +49,11 @@ PaddlePaddle需要的所有编译工具。把编译出来的PaddlePaddle也打
 像，称为生产镜像，里面涵盖了PaddlePaddle运行所需的所有环境。每次
 PaddlePaddle发布新版本的时候都会发布对应版本的生产镜像以及开发镜像。运
 行镜像包括纯CPU版本和GPU版本以及其对应的非AVX版本。我们会在
-`dockerhub.com <https://hub.docker.com/r/paddlepaddle/paddle/tags/>`_ 提供最新
-的Docker镜像，可以在"tags"标签下找到最新的Paddle镜像版本。为了方便在国
-内的开发者下载Docker镜像，我们提供了国内的镜像服务器供大家使用。如果您
-在国内，请把文档里命令中的paddlepaddle/paddle替换成
-docker.paddlepaddle.org/paddle。
+`dockerhub.com <https://hub.docker.com/r/paddlepaddle/paddle/tags/>`_ 
+和国内镜像`docker.paddlepaddle.org` 提供最新
+的Docker镜像，可以在"tags"标签下找到最新的Paddle镜像版本。
+
+**注意：为了方便在国内的开发者下载Docker镜像，我们提供了国内的镜像服务器供大家使用。如果您在国内，请把文档里命令中的paddlepaddle/paddle替换成docker.paddlepaddle.org/paddle。**
 
 1. 开发镜像：:code:`paddlepaddle/paddle:0.10.0-dev`
 
@@ -68,6 +105,8 @@ docker.paddlepaddle.org/paddle。
 
    如果输出是No，就需要选择使用no-AVX的镜像
 
+   **注：在0.10.0之后的版本，PaddlePaddle都可以自动判断硬件是否支持AVX，所以无需判断AVX即可使用**
+
    以上方法在GPU镜像里也能用，只是请不要忘记提前在物理机上安装GPU最新驱动。
    为了保证GPU驱动能够在镜像里面正常运行，我们推荐使用[nvidia-docker](https://github.com/NVIDIA/nvidia-docker)来运行镜像。
 

doc/getstarted/build_and_install/docker_install_en.rst

@@ -63,12 +63,35 @@ CPU-only version and a CUDA GPU version and their no-AVX versions.
 
 We put the docker images on `dockerhub.com
 <https://hub.docker.com/r/paddlepaddle/paddle/tags/>`_. You can find the
-latest versions under "tags" tab at dockerhub.com. If you are in
-China, you can use our Docker image registry mirror to speed up the
-download process. To use it, please replace all paddlepaddle/paddle in
-the commands to docker.paddlepaddle.org/paddle.
+latest versions under "tags" tab at dockerhub.com. 
 
-1. Production images, this image might have multiple variants:
+** NOTE: If you are in China, you can use our Docker image registry mirror to speed up the download process. To use it, please replace all paddlepaddle/paddle in the commands to docker.paddlepaddle.org/paddle.**
+
+
+1. development image :code:`paddlepaddle/paddle:<version>-dev`
+
+   This image has packed related develop tools and runtime
+   environment. Users and developers can use this image instead of
+   their own local computer to accomplish development, build,
+   releasing, document writing etc. While different version of paddle
+   may depends on different version of libraries and tools, if you
+   want to setup a local environment, you must pay attention to the
+   versions.  The development image contains:
+   
+   - gcc/clang
+   - nvcc
+   - Python
+   - sphinx
+   - woboq
+   - sshd
+     
+   Many developers use servers with GPUs, they can use ssh to login to
+   the server and run :code:`docker exec` to enter the docker
+   container and start their work.  Also they can start a development
+   docker image with SSHD service, so they can login to the container
+   and start work.
+
+2. Production images, this image might have multiple variants:
 
    - GPU/AVX：:code:`paddlepaddle/paddle:<version>-gpu`
    - GPU/no-AVX：:code:`paddlepaddle/paddle:<version>-gpu-noavx`
@@ -84,7 +107,7 @@ the commands to docker.paddlepaddle.org/paddle.
 
       if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi
 
-   
+   **NOTE：versions after 0.10.0 will automatically detect system AVX support, so manual detect is not needed in this case.**
    To run the CPU-only image as an interactive container:
 
    .. code-block:: bash
@@ -103,29 +126,6 @@ the commands to docker.paddlepaddle.org/paddle.
 
       nvidia-docker run -it --rm paddlepaddle/paddle:0.10.0-gpu /bin/bash
 
-2. development image :code:`paddlepaddle/paddle:<version>-dev`
-
-   This image has packed related develop tools and runtime
-   environment. Users and developers can use this image instead of
-   their own local computer to accomplish development, build,
-   releasing, document writing etc. While different version of paddle
-   may depends on different version of libraries and tools, if you
-   want to setup a local environment, you must pay attention to the
-   versions.  The development image contains:
-   
-   - gcc/clang
-   - nvcc
-   - Python
-   - sphinx
-   - woboq
-   - sshd
-     
-   Many developers use servers with GPUs, they can use ssh to login to
-   the server and run :code:`docker exec` to enter the docker
-   container and start their work.  Also they can start a development
-   docker image with SSHD service, so they can login to the container
-   and start work.
-
 
 Train Model Using Python API
 ----------------------------

go/cmd/pserver/pserver.go

@@ -32,7 +32,7 @@ import (
 
 func main() {
 	port := flag.Int("port", 0, "port of the pserver")
-	index := flag.Int("index", -1, "index of this pserver, should be larger or equal than 0")
+	index := flag.Int("index", -1, "index of the pserver, set to -1 if use etcd for auto pserver index registry")
 	etcdEndpoint := flag.String("etcd-endpoint", "http://127.0.0.1:2379",
 		"comma separated endpoint string for pserver to connect to etcd")
 	dialTimeout := flag.Duration("dial-timeout", 5*time.Second, "dial timeout")
@@ -60,12 +60,12 @@ func main() {
 		idx, err = e.Register(*port)
 		candy.Must(err)
 
-		cp, err = pserver.NewCheckpointFromFile(*checkpointPath, idx, e)
+		cp, err = pserver.LoadCheckpoint(e, idx)
 		if err != nil {
 			if err == pserver.ErrCheckpointNotFound {
 				log.Infof("Could not find the pserver checkpoint.")
 			} else {
-				log.Errorf("Fetch checkpoint failed, %s", err)
+				panic(err)
 			}
 		}
 	}

go/glide.lock

-hash: 2a1c0eca5c07a130e3d224f9821f96cfa37a39bf6bce141c855bbc57ef569f1c
-updated: 2017-07-29T07:34:48.722757905+08:00
+hash: 1b9b07408ca7fac27a374dc2ccd2433e4bff090484008a037df967284949a582
+updated: 2017-08-03T21:46:51.744995189Z
 imports:
 - name: github.com/beorn7/perks
   version: 4c0e84591b9aa9e6dcfdf3e020114cd81f89d5f9
@@ -145,6 +145,8 @@ imports:
   version: a1dba9ce8baed984a2495b658c82687f8157b98f
   subpackages:
   - xfs
+- name: github.com/satori/go.uuid
+  version: 879c5887cd475cd7864858769793b2ceb0d44feb
 - name: github.com/sirupsen/logrus
   version: a3f95b5c423586578a4e099b11a46c2479628cac
 - name: github.com/topicai/candy

go/glide.yaml

@@ -14,11 +14,13 @@ import:
   version: ^1.0.0
 - package: github.com/topicai/candy
 - package: golang.org/x/crypto
-  vcs: git
   repo: https://github.com/golang/crypto.git
-- package: golang.org/x/sys
   vcs: git
+- package: golang.org/x/sys
   repo: https://github.com/golang/sys.git
-- package: golang.org/x/text
   vcs: git
+- package: golang.org/x/text
   repo: https://github.com/golang/text.git
+  vcs: git
+- package: github.com/satori/go.uuid
+  version: v1.1.0

go/master/service.go

@@ -77,11 +77,12 @@ type taskEntry struct {
 	NumFailure int
 }
 
-type taskQueues struct {
+type masterState struct {
 	Todo    []taskEntry
 	Pending map[int]taskEntry // map from task ID to task entry
 	Done    []taskEntry
 	Failed  []taskEntry
+	CurPass int
 }
 
 // Service is the master server service.
@@ -95,10 +96,10 @@ type Service struct {
 	initDone bool
 
 	mu sync.Mutex
-	taskQueues taskQueues
-	currPass   int
-	jobTasks   []taskEntry
-
+	// State to be persisted to snapshot.
+	state masterState
+	// The trainer that is currently saving model. This state is
+	// transient, does not need to be persisted to snapshot.
 	savingTrainer string
 }
 
@@ -141,8 +142,8 @@ func NewService(store Store, chunksPerTask int, timeoutDur time.Duration, failur
 	s.chunksPerTask = chunksPerTask
 	s.timeoutDur = timeoutDur
 	s.failureMax = failureMax
-	s.taskQueues = taskQueues{}
-	s.taskQueues.Pending = make(map[int]taskEntry)
+	s.state = masterState{}
+	s.state.Pending = make(map[int]taskEntry)
 	s.ready = make(chan struct{})
 	s.store = store
 	recovered, err := s.recover()
@@ -180,7 +181,7 @@ func (s *Service) recover() (bool, error) {
 	}
 
 	dec := gob.NewDecoder(gr)
-	var tqs taskQueues
+	var tqs masterState
 	err = dec.Decode(&tqs)
 	if err != nil {
 		return false, err
@@ -193,7 +194,12 @@ func (s *Service) recover() (bool, error) {
 		log.Errorln(err)
 	}
 
-	s.taskQueues = tqs
+	s.state = tqs
+	log.WithFields(s.logFields()).Infof("Master recovered from snapshot, scheduling pending task timeout check.")
+	for _, t := range s.state.Pending {
+		time.AfterFunc(s.timeoutDur, s.checkTimeoutFunc(t.Task.Meta.ID, t.Task.Meta.Epoch))
+	}
+
 	return true, nil
 }
 
@@ -208,7 +214,7 @@ func (s *Service) snapshot() error {
 	var buf bytes.Buffer
 	gw := gzip.NewWriter(&buf)
 	enc := gob.NewEncoder(gw)
-	err := enc.Encode(s.taskQueues)
+	err := enc.Encode(s.state)
 	if err != nil {
 		return err
 	}
@@ -290,8 +296,7 @@ func (s *Service) SetDataset(globPaths []string, _ *int) error {
 		return err
 	}
 
-	s.jobTasks = partition(chunks, s.chunksPerTask)
-	s.taskQueues.Todo = s.jobTasks
+	s.state.Todo = partition(chunks, s.chunksPerTask)
 
 	err = s.snapshot()
 	if err != nil {
@@ -319,17 +324,17 @@ func (s *Service) processFailedTask(t taskEntry, epoch int) {
 		}
 	}()
 
-	delete(s.taskQueues.Pending, t.Task.Meta.ID)
+	delete(s.state.Pending, t.Task.Meta.ID)
 
 	t.NumFailure++
 	if t.NumFailure > s.failureMax {
 		log.Warningf("Task %v failed %d times, discard.", t.Task, t.NumFailure)
-		s.taskQueues.Failed = append(s.taskQueues.Failed, t)
+		s.state.Failed = append(s.state.Failed, t)
 		return
 	}
 
 	log.Warningf("Task %v failed %d times, re-dispatch.", t.Task, t.NumFailure)
-	s.taskQueues.Todo = append(s.taskQueues.Todo, t)
+	s.state.Todo = append(s.state.Todo, t)
 	return
 }
 
@@ -338,7 +343,7 @@ func (s *Service) checkTimeoutFunc(taskID int, epoch int) func() {
 		s.mu.Lock()
 		defer s.mu.Unlock()
 
-		t, ok := s.taskQueues.Pending[taskID]
+		t, ok := s.state.Pending[taskID]
 		if !ok {
 			return
 		}
@@ -350,10 +355,11 @@ func (s *Service) checkTimeoutFunc(taskID int, epoch int) func() {
 // must be called with lock held.
 func (s *Service) logFields() log.Fields {
 	return log.Fields{
-		"todoLen":    len(s.taskQueues.Todo),
-		"pendingLen": len(s.taskQueues.Pending),
-		"doneLen":    len(s.taskQueues.Done),
-		"failedLen":  len(s.taskQueues.Failed),
+		"todoLen":    len(s.state.Todo),
+		"pendingLen": len(s.state.Pending),
+		"doneLen":    len(s.state.Done),
+		"failedLen":  len(s.state.Failed),
+		"curPass":    s.state.CurPass,
 	}
 }
 
@@ -366,17 +372,17 @@ func (s *Service) GetTask(passID int, task *Task) error {
 
 	s.mu.Lock()
 	defer s.mu.Unlock()
-	if passID < s.currPass {
+	if passID < s.state.CurPass {
 		return ErrPassBefore
 	}
-	if passID > s.currPass {
+	if passID > s.state.CurPass {
 		// Client may get run to pass after master when one client faster than the
 		// other
 		return ErrPassAfter
 	}
 
-	if len(s.taskQueues.Todo) == 0 {
-		if len(s.taskQueues.Done) == 0 && len(s.taskQueues.Pending) == 0 {
+	if len(s.state.Todo) == 0 {
+		if len(s.state.Done) == 0 && len(s.state.Pending) == 0 {
 			log.WithFields(s.logFields()).Warningln("All tasks failed, may start next pass")
 			return ErrAllTaskFailed
 		}
@@ -384,10 +390,10 @@ func (s *Service) GetTask(passID int, task *Task) error {
 		return ErrNoMoreAvailable
 	}
 
-	t := s.taskQueues.Todo[0]
+	t := s.state.Todo[0]
 	t.Task.Meta.Epoch++
-	s.taskQueues.Todo = s.taskQueues.Todo[1:]
-	s.taskQueues.Pending[t.Task.Meta.ID] = t
+	s.state.Todo = s.state.Todo[1:]
+	s.state.Pending[t.Task.Meta.ID] = t
 	err := s.snapshot()
 	if err != nil {
 		return err
@@ -409,7 +415,7 @@ func (s *Service) TaskFinished(taskID int, dummy *int) error {
 	s.mu.Lock()
 	defer s.mu.Unlock()
 
-	t, ok := s.taskQueues.Pending[taskID]
+	t, ok := s.state.Pending[taskID]
 	if !ok {
 		log.WithFields(s.logFields()).Warningln("Pending task #%d not found.", taskID)
 		return nil
@@ -417,18 +423,18 @@ func (s *Service) TaskFinished(taskID int, dummy *int) error {
 
 	// task finished, reset timeout
 	t.NumFailure = 0
-	s.taskQueues.Done = append(s.taskQueues.Done, t)
-	delete(s.taskQueues.Pending, taskID)
+	s.state.Done = append(s.state.Done, t)
+	delete(s.state.Pending, taskID)
 
 	log.WithFields(s.logFields()).Infof("Task #%d finished.", taskID)
-	if len(s.taskQueues.Todo) == 0 && len(s.taskQueues.Pending) == 0 {
+	if len(s.state.Todo) == 0 && len(s.state.Pending) == 0 {
 		// increase master side pass count if all tasks finished
-		s.currPass++
-		s.taskQueues.Todo = s.jobTasks
-		s.taskQueues.Done = []taskEntry{}
+		s.state.CurPass++
+		s.state.Todo = append(s.state.Done, s.state.Failed...)
+		s.state.Done = []taskEntry{}
 		// TODO(typhoonzero): deal with failed tasks
-		s.taskQueues.Failed = []taskEntry{}
-		log.WithFields(s.logFields()).Warningf("all task finished, add new pass data, newpass: %d.", s.currPass)
+		s.state.Failed = []taskEntry{}
+		log.WithFields(s.logFields()).Warningf("all task finished, add new pass data, newpass: %d.", s.state.CurPass)
 	}
 
 	err := s.snapshot()
@@ -447,7 +453,7 @@ func (s *Service) TaskFailed(meta TaskMeta, dummy *int) error {
 	s.mu.Lock()
 	defer s.mu.Unlock()
 
-	t, ok := s.taskQueues.Pending[meta.ID]
+	t, ok := s.state.Pending[meta.ID]
 	if !ok {
 		log.WithFields(s.logFields()).Warningln("TaskFailed:Pending task #%v not found.", t.Task.Meta)
 		return nil

go/pserver/client/client_test.go

@@ -59,7 +59,7 @@ func initClient() [numPserver]int {
 
 		go func(l net.Listener) {
 			var cp pserver.Checkpoint
-			s, err := pserver.NewService(0, 1, "", nil, cp)
+			s, err := pserver.NewService(0, time.Hour, "", nil, cp)
 			if err != nil {
 				panic(err)
 			}

go/pserver/client/etcd_client.go

@@ -103,7 +103,7 @@ func (p *EtcdClient) List() []Server {
 				time.Sleep(p.timeout)
 				continue
 			}
-			log.Infof("got value (%s) for key: %s", psAddr, psKey)
+			log.Debugf("got value (%s) for key: %s", psAddr, psKey)
 			servers[i].Index = i
 			servers[i].Addr = psAddr
 		}

go/pserver/etcd_client.go

@@ -206,6 +206,7 @@ func (e *EtcdClient) GetKey(key string, timeout time.Duration) ([]byte, error) {
 	if err != nil {
 		return []byte{}, err
 	}
+
 	kvs := resp.Kvs
 	if len(kvs) == 0 {
 		return []byte{}, nil
@@ -215,9 +216,14 @@ func (e *EtcdClient) GetKey(key string, timeout time.Duration) ([]byte, error) {
 }
 
 // PutKey put into etcd with value by key specified
-func (e *EtcdClient) PutKey(key string, value []byte, timeout time.Duration) error {
+func (e *EtcdClient) PutKey(key string, value []byte, timeout time.Duration, withLease bool) error {
 	ctx, cancel := context.WithTimeout(context.Background(), timeout)
-	_, err := e.client.Put(ctx, key, string(value), clientv3.WithLease(e.sess.Lease()))
+	var err error
+	if withLease {
+		_, err = e.client.Put(ctx, key, string(value), clientv3.WithLease(e.sess.Lease()))
+	} else {
+		_, err = e.client.Put(ctx, key, string(value))
+	}
 	cancel()
 	return err
 }

go/pserver/optimizer.go

@@ -32,6 +32,7 @@ type optimizer struct {
 	opt         *C.struct_paddle_optimizer
 	elementType ElementType
 	contentLen  int
+	config      []byte
 }
 
 func cArrayToSlice(p unsafe.Pointer, len int) []byte {
@@ -70,6 +71,7 @@ func newOptimizer(paramWithConfigs ParameterWithConfig, State []byte) *optimizer
 		cstate = unsafe.Pointer(&s[0])
 	}
 
+	o.config = c
 	o.opt = C.paddle_create_optimizer((*C.uchar)(&c[0]), C.int(len(c)),
 		C.paddle_element_type(p.ElementType), cbuffer, C.int(paramBufferSize), (*C.char)(cstate), C.int(len(s)))
 	return o

go/pserver/service.go

@@ -25,11 +25,13 @@ import (
 	"fmt"
 	"io/ioutil"
 	"os"
-	"path/filepath"
+	"path"
 	"strconv"
 	"sync"
 	"time"
 
+	uuid "github.com/satori/go.uuid"
+
 	log "github.com/sirupsen/logrus"
 )
 
@@ -44,7 +46,7 @@ var ErrCheckpointNotFound = errors.New("checkpoint not found")
 const (
 	AlreadyInitialized = "pserver already initialized"
 	Uninitialized      = "pserver not fully initialized"
-	CheckpointMD5Failed = "checkpoint file MD5 validation failed"
+	WrongChecksum      = "checkpoint file checksum validation failed"
 )
 
 // Supported element types.
@@ -73,11 +75,12 @@ type ParameterWithConfig struct {
 // checkpointMeta saves checkpoint metadata
 type checkpointMeta struct {
 	UUID      string `json:"uuid"`
+	Path      string `json:"path"`
 	MD5       string `json:"md5"`
 	Timestamp int64  `json:"timestamp"`
 }
 
-// Checkpoint is the pserver shard persist in file
+// Checkpoint is the pserver shard persist in file.
 type Checkpoint []parameterCheckpoint
 
 // Gradient is the gradient of the parameter.
@@ -90,50 +93,58 @@ type Service struct {
 	checkpointInterval time.Duration
 	checkpointPath     string
 	client             *EtcdClient
+
 	mu     sync.Mutex
 	optMap map[string]*optimizer
 }
 
-// parameterCheckpoint saves parameter checkpoint
+// parameterCheckpoint saves parameter checkpoint.
 type parameterCheckpoint struct {
 	ParameterWithConfig
 	State []byte
 }
 
-// NewCheckpointFromFile loads parameters and state from checkpoint file
-func NewCheckpointFromFile(cpPath string, idx int, e *EtcdClient) (Checkpoint, error) {
-	v, err := e.GetKey(PsPath+string(idx), 3*time.Second)
+func loadMeta(e *EtcdClient, idx int) (meta checkpointMeta, err error) {
+	v, err := e.GetKey(PsCheckpoint+strconv.Itoa(idx), 3*time.Second)
 	if err != nil {
-		return nil, err
+		return
 	}
 
 	if len(v) == 0 {
-		return nil, ErrCheckpointNotFound
+		err = ErrCheckpointNotFound
+		return
 	}
 
-	var cpMeta checkpointMeta
-	if err = json.Unmarshal(v, &cpMeta); err != nil {
-		return nil, err
+	if err = json.Unmarshal(v, &meta); err != nil {
+		return
 	}
 
-	fn := filepath.Join(cpPath, cpMeta.UUID)
-	if _, err = os.Stat(fn); os.IsNotExist(err) {
+	return
+}
+
+// LoadCheckpoint loads checkpoint from file.
+func LoadCheckpoint(e *EtcdClient, idx int) (Checkpoint, error) {
+	cpMeta, err := loadMeta(e, idx)
+	if err != nil {
 		return nil, err
 	}
-	content, err := ioutil.ReadFile(fn)
+
+	content, err := ioutil.ReadFile(cpMeta.Path)
 	if err != nil {
 		return nil, err
 	}
 
+	// TODO(helin): change MD5 to CRC since CRC is better for file
+	// checksum in our use case (emphasize speed over security).
 	h := md5.New()
 	md5 := hex.EncodeToString(h.Sum(content))
 	if md5 != cpMeta.MD5 {
-		return nil, errors.New(CheckpointMD5Failed)
+		return nil, errors.New(WrongChecksum)
 	}
 
 	dec := gob.NewDecoder(bytes.NewReader(content))
-	cp := Checkpoint{}
-	if err = dec.Decode(cp); err != nil {
+	var cp Checkpoint
+	if err = dec.Decode(&cp); err != nil {
 		return nil, err
 	}
 	return cp, nil
@@ -193,6 +204,15 @@ func (s *Service) FinishInitParams(_ int, _ *int) error {
 	}
 
 	close(s.initialized)
+	go func() {
+		t := time.Tick(s.checkpointInterval)
+		for range t {
+			err := s.checkpoint()
+			if err != nil {
+				log.Errorln(err)
+			}
+		}
+	}()
 	return nil
 }
 
@@ -240,23 +260,36 @@ func (s *Service) GetParam(name string, parameter *Parameter) error {
 	return nil
 }
 
-// pserver save checkpoint
-func (s *Service) doCheckpoint() (err error) {
-	<-s.initialized
-	s.mu.Lock()
-	defer s.mu.Unlock()
+func traceTime(start time.Time, name string) {
+	elapsed := time.Since(start)
+	log.Infof("%s took %v", name, elapsed)
+}
+
+// checkpoint saves checkpoint to disk.
+//
+// checkpoint should be only called after the parameters are
+// initialized.
+func (s *Service) checkpoint() (err error) {
+	log.Infoln("Begin save checkpoint.")
+	defer traceTime(time.Now(), "save checkpoint")
 
+	s.mu.Lock()
 	cp := make([]parameterCheckpoint, len(s.optMap))
 	index := 0
+	// TODO(helin): write checkpoint incrementally to reduce memory
+	// footprint during checkpoint.
 	for name, opt := range s.optMap {
 		var pc parameterCheckpoint
 		pc.Param.Name = name
 		pc.Param.ElementType = opt.elementType
 		pc.Param.Content = opt.GetWeights()
+		pc.Config = opt.config
 		pc.State = opt.GetStates()
 		cp[index] = pc
 		index++
 	}
+	s.mu.Unlock()
+
 	var buf bytes.Buffer
 	encoder := gob.NewEncoder(&buf)
 	err = encoder.Encode(cp)
@@ -264,32 +297,9 @@ func (s *Service) doCheckpoint() (err error) {
 		return
 	}
 
-	cpMeta := checkpointMeta{}
-	cpMeta.UUID = s.checkpointPath + strconv.Itoa(s.idx)
-	cpMeta.Timestamp = time.Now().UnixNano()
-	h := md5.New()
-	cpMeta.MD5 = hex.EncodeToString(h.Sum(buf.Bytes()))
-
-	cpMetajson, err := json.Marshal(cpMeta)
-	if err != nil {
-		return
-	}
-
-	err = s.client.PutKey(filepath.Join(PsCheckpoint, strconv.Itoa(s.idx)), cpMetajson, 3*time.Second)
-	if err != nil {
-		return
-	}
-	if _, err = os.Stat(cpMeta.UUID); os.IsNotExist(err) {
-		log.Info("checkpoint does not exists.")
-	} else {
-		err = os.Remove(cpMeta.UUID)
-		if err != nil {
-			log.Infof("Removing checkpoint %s failed", cpMeta.UUID)
-		} else {
-			log.Infof("checkpoint %s already exsits, removing ", cpMeta.UUID)
-		}
-	}
-	f, err := os.Create(cpMeta.UUID)
+	id := uuid.NewV4().String()
+	p := path.Join(s.checkpointPath, id)
+	f, err := os.Create(p)
 	if err != nil {
 		return
 	}
@@ -317,5 +327,43 @@ func (s *Service) doCheckpoint() (err error) {
 		return
 	}
 
+	oldMeta, err := loadMeta(s.client, s.idx)
+	if err == ErrCheckpointNotFound {
+		log.Infoln("Do not have existing checkpoint.")
+		err = nil
+	}
+
+	if err != nil {
+		return
+	}
+
+	h := md5.New()
+	md5 := hex.EncodeToString(h.Sum(buf.Bytes()))
+	cpMeta := checkpointMeta{
+		UUID:      id,
+		Timestamp: time.Now().UnixNano(),
+		MD5:       md5,
+		Path:      p,
+	}
+
+	json, err := json.Marshal(cpMeta)
+	if err != nil {
+		return
+	}
+
+	err = s.client.PutKey(PsCheckpoint+strconv.Itoa(s.idx), json, 3*time.Second, false)
+	if err != nil {
+		return
+	}
+
+	if oldMeta.Path != "" {
+		rmErr := os.Remove(oldMeta.Path)
+		if rmErr != nil {
+			// log error, but still treat checkpoint as
+			// successful.
+			log.Errorln(rmErr)
+		}
+	}
+
 	return
 }

go/pserver/service_test.go

@@ -30,7 +30,7 @@ const (
 
 func TestServiceFull(t *testing.T) {
 	var cp pserver.Checkpoint
-	s, err := pserver.NewService(0, 1, "", nil, cp)
+	s, err := pserver.NewService(0, time.Hour, "", nil, cp)
 	if err != nil {
 		t.Error(err)
 	}
@@ -102,7 +102,7 @@ func TestServiceFull(t *testing.T) {
 
 func TestMultipleInit(t *testing.T) {
 	var cp pserver.Checkpoint
-	s, err := pserver.NewService(0, 1, "", nil, cp)
+	s, err := pserver.NewService(0, time.Hour, "", nil, cp)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -119,7 +119,7 @@ func TestMultipleInit(t *testing.T) {
 
 func TestUninitialized(t *testing.T) {
 	var cp pserver.Checkpoint
-	s, err := pserver.NewService(0, 1, "", nil, cp)
+	s, err := pserver.NewService(0, time.Hour, "", nil, cp)
 	err = s.SendGrad(pserver.Gradient{}, nil)
 	if err.Error() != pserver.Uninitialized {
 		t.Fatal(err)
@@ -128,7 +128,7 @@ func TestUninitialized(t *testing.T) {
 
 func TestBlockUntilInitialized(t *testing.T) {
 	var cp pserver.Checkpoint
-	s, err := pserver.NewService(0, 1, "", nil, cp)
+	s, err := pserver.NewService(0, time.Hour, "", nil, cp)
 	if err != nil {
 		t.Error(err)
 	}

paddle/CMakeLists.txt

@@ -22,7 +22,5 @@ if(WITH_C_API)
 endif()
 
 if(WITH_SWIG_PY)
-  configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in
-          ${CMAKE_CURRENT_SOURCE_DIR}/setup.py)
   add_subdirectory(api)
 endif()

paddle/api/CMakeLists.txt

@@ -82,9 +82,7 @@ SWIG_LINK_LIBRARIES(swig_paddle
 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
+    COMMAND ${CMAKE_COMMAND} -E touch .timestamp
     WORKING_DIRECTORY ${PROJ_ROOT}/paddle
     DEPENDS _swig_paddle
 )
@@ -92,10 +90,6 @@ add_custom_command(OUTPUT ${PROJ_ROOT}/paddle/py_paddle/_swig_paddle.so
 # 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 ${CMAKE_SOURCE_DIR}/paddle/dist/
-    DESTINATION opt/paddle/share/wheels
-)
-
 if(WITH_TESTING)
     IF(NOT PY_PIP_FOUND)
         SET(PIP_SOURCES_DIR ${PYTHON_SOURCES_DIR}/pip)
@@ -108,7 +102,7 @@ if(WITH_TESTING)
             BUILD_COMMAND       ""
             INSTALL_COMMAND     env ${py_env} ${PYTHON_EXECUTABLE} setup.py install
             BUILD_IN_SOURCE     1
-            DEPENDS python setuptools python_api_wheel
+            #DEPENDS python setuptools python_api_wheel
         )
     ENDIF()
     add_subdirectory(test)

paddle/cuda/CMakeLists.txt

@@ -39,6 +39,7 @@ set(CUDA_CU_SOURCES
     src/hl_cuda_lstm.cu
     src/hl_top_k.cu
     src/hl_batch_transpose.cu
+    src/hl_batch_norm.cu
     src/hl_cuda_sequence.cu
     src/hl_table_apply.cu)
 

paddle/cuda/include/hl_batch_norm.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. */
+
+#ifndef HL_BATCH_NORM_H_
+#define HL_BATCH_NORM_H_
+
+#include "hl_base.h"
+
+/**
+ * @brief   batch norm inferece.
+ *
+ * @param[in]   input         input data.
+ * @param[out]  output        output data.
+ * @param[in]   scale         batch normalization scale parameter (in original
+ *                            paper scale is referred to as gamma).
+ * @param[in]   bias          batch normalization bias parameter (in original
+ *                            paper scale is referred to as beta).
+ * @param[in]   estimatedMean
+ * @param[in]   estimatedVar  The moving mean and variance
+ *                            accumulated during the training phase are passed
+ *                            as inputs here.
+ * @param[in]   epsilon       Epsilon value used in the batch
+ *                            normalization formula.
+ */
+extern void hl_batch_norm_cuda_inference(const real* input,
+                                         real* output,
+                                         const real* scale,
+                                         const real* bias,
+                                         const real* estimatedMean,
+                                         const real* estimatedVar,
+                                         const double epsilon,
+                                         size_t batchSize,
+                                         size_t channel,
+                                         size_t height,
+                                         size_t width);
+
+#endif  // HL_BATCH_NORM_H_

paddle/cuda/src/hl_batch_norm.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 "hl_batch_norm.h"
+
+__global__ void batchNormInference(real* output,
+                                   const real* input,
+                                   const real* scale,
+                                   const real* bias,
+                                   const real* estimatedMean,
+                                   const real* estimatedVar,
+                                   const double epsilon,
+                                   size_t batchSize,
+                                   size_t channel,
+                                   size_t height,
+                                   size_t width) {
+  const int tid = threadIdx.x;
+  const int num = channel * height * width;
+  const int batch = blockIdx.x;
+  for (int i = tid; i < num; i += blockDim.x) {
+    const int c = i / (height * width);
+    const int id = batch * num + i;
+    real val = input[id] - estimatedMean[c];
+    val /= sqrt(estimatedVar[c] + epsilon);
+    val *= scale[c];
+    val += bias[c];
+    output[id] = val;
+  }
+}
+
+void hl_batch_norm_cuda_inference(const real* input,
+                                  real* output,
+                                  const real* scale,
+                                  const real* bias,
+                                  const real* estimatedMean,
+                                  const real* estimatedVar,
+                                  const double epsilon,
+                                  size_t batchSize,
+                                  size_t channel,
+                                  size_t height,
+                                  size_t width) {
+  batchNormInference<<<batchSize, 256, 0, STREAM_DEFAULT>>>(output,
+                                                            input,
+                                                            scale,
+                                                            bias,
+                                                            estimatedMean,
+                                                            estimatedVar,
+                                                            epsilon,
+                                                            batchSize,
+                                                            channel,
+                                                            height,
+                                                            width);
+
+  CHECK_SYNC("hl_batch_norm_cuda_inference failed!");
+}

paddle/cuda/src/hl_cuda_cudnn.cc

@@ -1023,14 +1023,6 @@ void hl_batch_norm_forward_inference(hl_tensor_descriptor inputDesc,
   real beta = 1.0f;
   cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL;
 
-  int batch_size = ((cudnn_tensor_descriptor)inputDesc)->batch_size;
-  if (batch_size > 1024 && g_cudnn_lib_version < 6000) {
-    LOG(INFO) << " To process current batch data with size " << batch_size
-              << " (>1024), cudnnBatchNorm requires cuDNN version >= 6000."
-              << " If there is an error complaining CUDNN_STATUS_NOT_SUPPORTED,"
-              << " just recompile PaddlePaddle with cuDNN >= 6000, replacing"
-              << " current version " << g_cudnn_lib_version;
-  }
   CHECK_CUDNN(
       dynload::cudnnBatchNormalizationForwardInference(t_resource.cudnn_handle,
                                                        mode,

paddle/framework/CMakeLists.txt

@@ -35,6 +35,8 @@ add_dependencies(framework_py_proto framework_py_proto_init)
 
 cc_library(backward SRCS backward.cc DEPS net_op)
 cc_test(backward_test SRCS backward_test.cc DEPS backward)
+
+if(WITH_PYTHON)
 cc_library(paddle_pybind SHARED
     SRCS pybind.cc
     DEPS pybind python backward
@@ -43,4 +45,6 @@ cc_library(paddle_pybind SHARED
 	add_op
 	mean_op
 	cross_entropy_op
+	fill_zeros_like_op
 	recurrent_op)
+endif(WITH_PYTHON)

paddle/framework/op_registry.h

@@ -260,6 +260,12 @@ class OpRegistry {
     return CreateOp(op_desc.type(), inputs, outputs, attrs);
   }
 
+  static bool SupportGPU(const std::string& op_type) {
+    OperatorWithKernel::OpKernelKey key;
+    key.place_ = platform::GPUPlace();
+    return OperatorWithKernel::AllOpKernels().at(op_type).count(key) != 0;
+  }
+
   static std::shared_ptr<OperatorBase> CreateGradOp(const OperatorBase& op) {
     PADDLE_ENFORCE(!op.IsNetOp(),
                    "Use framework::Backward to get backward ops");

paddle/framework/operator.cc

@@ -34,7 +34,7 @@ ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
 #endif
 
 const std::string& OperatorBase::Input(const std::string& name) const {
-  PADDLE_ENFORCE(in_out_idxs_ != nullptr,
+  PADDLE_ENFORCE_NOT_NULL(in_out_idxs_,
                           "Input Output Indices could not be nullptr");
   auto it = in_out_idxs_->find(name);
   PADDLE_ENFORCE(it != in_out_idxs_->end(), "no key [%s] in in_out_idxs_",
@@ -49,7 +49,7 @@ const std::string& OperatorBase::Input(const std::string& name) const {
 }
 
 std::vector<std::string> OperatorBase::Inputs(const std::string& name) const {
-  PADDLE_ENFORCE(in_out_idxs_ != nullptr, "IO Idx could not be nullptr");
+  PADDLE_ENFORCE_NOT_NULL(in_out_idxs_, "IO Idx could not be nullptr");
   auto input_format = GetAttr<std::vector<int>>("input_format");
   auto offset = in_out_idxs_->at(name);
   PADDLE_ENFORCE(input_format.at(static_cast<size_t>(offset) + 1) <=
@@ -62,7 +62,7 @@ std::vector<std::string> OperatorBase::Inputs(const std::string& name) const {
 }
 
 const std::string& OperatorBase::Output(const std::string& name) const {
-  PADDLE_ENFORCE(in_out_idxs_ != nullptr, "InOut Indice could not be nullptr");
+  PADDLE_ENFORCE_NOT_NULL(in_out_idxs_, "InOut Indice could not be nullptr");
   auto it = in_out_idxs_->find(name);
   PADDLE_ENFORCE(it != in_out_idxs_->end(), "no key [%s] in in_out_idxs_",
                  name);
@@ -76,7 +76,7 @@ const std::string& OperatorBase::Output(const std::string& name) const {
 }
 
 std::vector<std::string> OperatorBase::Outputs(const std::string& name) const {
-  PADDLE_ENFORCE(in_out_idxs_ != nullptr, "InOut Indice could not be nullptr");
+  PADDLE_ENFORCE_NOT_NULL(in_out_idxs_, "InOut Indice could not be nullptr");
   auto output_format = GetAttr<std::vector<int>>("output_format");
   auto offset = in_out_idxs_->at(name);
   PADDLE_ENFORCE(output_format.at(static_cast<size_t>(offset) + 1) <=

paddle/framework/operator.h

@@ -167,15 +167,15 @@ class OperatorContext {
   template <typename T>
   const T* Input(const size_t index) const {
     auto var = InputVar(index);
-    PADDLE_ENFORCE(var != nullptr, "Input(%d) should not be nullptr", index);
+    PADDLE_ENFORCE_NOT_NULL(var, "Input(%d) should not be nullptr", index);
     return &var->Get<T>();
   }
 
   template <typename T>
   T* Output(const size_t index) const {
     auto var = OutputVar(index);
-    PADDLE_ENFORCE(
-        var != nullptr,
+    PADDLE_ENFORCE_NOT_NULL(
+        var,
         "Output(%d) not be nullptr, which means variable [%s] does not "
         "exist in scope",
         index, op_.outputs_[index]);
@@ -185,14 +185,14 @@ class OperatorContext {
   template <typename T>
   const T* Input(const std::string& name) const {
     auto var = InputVar(name);
-    PADDLE_ENFORCE(var != nullptr, "Input(%s) should not be nullptr", name);
+    PADDLE_ENFORCE_NOT_NULL(var, "Input(%s) should not be nullptr", name);
     return &var->Get<T>();
   }
 
   template <typename T>
   T* Output(const std::string& name) const {
     auto var = OutputVar(name);
-    PADDLE_ENFORCE(var != nullptr, "Output(%s) should not be nullptr", name);
+    PADDLE_ENFORCE_NOT_NULL(var, "Output(%s) should not be nullptr", name);
     return var->GetMutable<T>();
   }
 
@@ -204,9 +204,9 @@ class OperatorContext {
     std::transform(names.begin(), names.end(), std::back_inserter(res),
                    [&](const std::string& sub_name) {
                      auto var = scope_.FindVar(sub_name);
-                     PADDLE_ENFORCE(var != nullptr,
-                                    "MultiInput(%s:%s) should not be nullptr",
-                                    name, sub_name);
+                     PADDLE_ENFORCE_NOT_NULL(
+                         var, "MultiInput(%s:%s) should not be nullptr", name,
+                         sub_name);
                      return &var->Get<T>();
                    });
     return res;
@@ -220,9 +220,9 @@ class OperatorContext {
     std::transform(names.begin(), names.end(), std::back_inserter(res),
                    [&](const std::string& sub_name) {
                      auto var = scope_.FindVar(sub_name);
-                     PADDLE_ENFORCE(var != nullptr,
-                                    "MultiOutput(%s:%s) should not be nullptr",
-                                    name, sub_name);
+                     PADDLE_ENFORCE_NOT_NULL(
+                         var, "MultiOutput(%s:%s) should not be nullptr", name,
+                         sub_name);
                      return var->GetMutable<T>();
                    });
     return res;

paddle/framework/pybind.cc

@@ -32,7 +32,7 @@ limitations under the License. */
 namespace py = pybind11;
 
 USE_OP(add_two);
-USE_OP(onehot_cross_entropy);
+USE_OP_CPU(onehot_cross_entropy);
 USE_OP_WITHOUT_KERNEL(fc);
 USE_OP(sgd);
 USE_OP(mul);
@@ -40,6 +40,7 @@ USE_OP(mean);
 USE_OP(sigmoid);
 USE_OP(softmax);
 USE_OP(rowwise_add);
+USE_OP(fill_zeros_like);
 USE_OP_WITHOUT_KERNEL(recurrent_op);
 namespace paddle {
 namespace framework {
@@ -200,6 +201,8 @@ All parameter, weight, gradient are variables in Paddle.
     return OpRegistry::CreateOp(desc);
   });
 
+  operator_base.def_static("support_gpu", &OpRegistry::SupportGPU);
+
   operator_base.def("backward",
                     [](const OperatorBase &forwardOp,
                        const std::unordered_set<std::string> &no_grad_vars) {

paddle/framework/tensor.h

@@ -127,8 +127,8 @@ class Tensor {
                memory::PODDeleter<T, Place>(place)),
           place_(place),
           size_(size) {
-      PADDLE_ENFORCE(ptr_ != nullptr, "Insufficient %s memory to allocation.",
-                     is_cpu_place(place_) ? "CPU" : "GPU");
+      PADDLE_ENFORCE_NOT_NULL(ptr_, "Insufficient %s memory to allocation.",
+                              (is_cpu_place(place_) ? "CPU" : "GPU"));
     }
 
     virtual size_t size() const { return size_; }

paddle/framework/tensor_impl.h

@@ -14,15 +14,16 @@ limitations under the License. */
 
 #pragma once
 #include "paddle/memory/memcpy.h"
+#include "paddle/platform/enforce.h"
 
 namespace paddle {
 namespace framework {
 
 template <typename T>
 inline void Tensor::check_memory_size() const {
-  PADDLE_ENFORCE(holder_ != nullptr,
-                 "Tenosr holds no memory. Call Tensor::mutable_data first.");
-  PADDLE_ENFORCE(holder_->size() >= product(dims_) * sizeof(T) + offset_,
+  PADDLE_ENFORCE_NOT_NULL(
+      holder_, "Tenosr holds no memory. Call Tensor::mutable_data first.");
+  PADDLE_ENFORCE_GE(holder_->size(), product(dims_) * sizeof(T) + offset_,
                     "Tensor's dims_ is out of bound. Call Tensor::mutable_data "
                     "first to re-allocate memory.");
 }
@@ -51,7 +52,7 @@ inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
 template <typename T>
 inline T* Tensor::mutable_data(platform::Place place) {
   static_assert(std::is_pod<T>::value, "T must be POD");
-  PADDLE_ENFORCE(product(dims_) > 0,
+  PADDLE_ENFORCE_GT(product(dims_), 0,
                     "Tensor's numel must be larger than zero to call "
                     "Tensor::mutable_data. Call Tensor::set_dim first.");
   /* some versions of boost::variant don't have operator!= */
@@ -120,11 +121,11 @@ inline void Tensor::CopyFrom(const Tensor& src,
 template <typename T>
 inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
   check_memory_size<T>();
-  PADDLE_ENFORCE(begin_idx >= 0, "Slice begin index is less than zero.");
-  PADDLE_ENFORCE(end_idx <= dims_[0], "Slice end index is out of bound.");
-  PADDLE_ENFORCE(begin_idx < end_idx,
+  PADDLE_ENFORCE_GE(begin_idx, 0, "Slice begin index is less than zero.");
+  PADDLE_ENFORCE_LE(end_idx, dims_[0], "Slice end index is out of bound.");
+  PADDLE_ENFORCE_LT(begin_idx, end_idx,
                     "Begin index must be less than end index.");
-  PADDLE_ENFORCE(dims_[0] != 1, "Can not slice a tensor with dims_[0] = 1.");
+  PADDLE_ENFORCE_NE(dims_[0], 1, "Can not slice a tensor with dims_[0] = 1.");
   int base = product(dims_) / dims_[0];
   Tensor dst;
   dst.holder_ = holder_;

paddle/framework/tensor_test.cc

@@ -36,7 +36,8 @@ TEST(Tensor, DataAssert) {
   } catch (paddle::platform::EnforceNotMet err) {
     caught = true;
     std::string msg =
-        "Tenosr holds no memory. Call Tensor::mutable_data first.";
+        "holder_ should not be null\nTenosr holds no memory. Call "
+        "Tensor::mutable_data first.";
     const char* what = err.what();
     for (size_t i = 0; i < msg.length(); ++i) {
       ASSERT_EQ(what[i], msg[i]);
@@ -111,7 +112,8 @@ TEST(Tensor, ShareDataWith) {
     } catch (paddle::platform::EnforceNotMet err) {
       caught = true;
       std::string msg =
-          "Tenosr holds no memory. Call Tensor::mutable_data first.";
+          "holder_ should not be null\nTenosr holds no memory. Call "
+          "Tensor::mutable_data first.";
       const char* what = err.what();
       for (size_t i = 0; i < msg.length(); ++i) {
         ASSERT_EQ(what[i], msg[i]);

paddle/gserver/layers/CudnnBatchNormLayer.cpp

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "CudnnBatchNormLayer.h"
 #include "Layer.h"
+#include "paddle/cuda/include/hl_batch_norm.h"
 #include "paddle/utils/Stat.h"
 
 namespace paddle {
@@ -79,6 +80,7 @@ void CudnnBatchNormLayer::forward(PassType passType) {
                                    savedInvVar);
   } else {
     // used movingMean and movingVar in testing
+    if (batchSize <= 1024) {
       hl_batch_norm_forward_inference(ioDesc_,
                                       input,
                                       ioDesc_,
@@ -89,6 +91,22 @@ void CudnnBatchNormLayer::forward(PassType passType) {
                                       movingMean,
                                       movingVar,
                                       EPS);
+    } else {
+      // There is a limitation in cudnn library.
+      // When the batch size is larger than 1024 in cuDNN v5.1,
+      // the cudnnBatchNormalizationForwardInference will fail.
+      hl_batch_norm_cuda_inference(input,
+                                   output,
+                                   gamma,
+                                   beta,
+                                   movingMean,
+                                   movingVar,
+                                   EPS,
+                                   batchSize,
+                                   channels_,
+                                   imageH_,
+                                   imageW_);
+    }
   }
 
   /* activation */ {

paddle/gserver/layers/SubNestedSequenceLayer.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 "Layer.h"
+#include "paddle/math/Matrix.h"
+#include "paddle/math/Vector.h"
+#include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+class SubNestedSequenceLayer : public Layer {
+public:
+  explicit SubNestedSequenceLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+
+private:
+  /*
+   * This functions generates the indices of rows in a batch according to the
+   * indices of selected sub-sequence in each sequence.
+   *
+   * Examples:
+   * selectedIndices:
+   *   [
+   *     [0, 1, -1],
+   *     [0, 1, 2],
+   *     [0, -1, -1],
+   *     [0, 2, 3],
+   *   ]
+   * inputSeqInfo:
+   *   [
+   *     [0,3,4],
+   *     [4,5,7,10,15],
+   *     [15,20],
+   *     [20,22,23,25,28]
+   *   ]
+   *
+   * ths output is saved to private member rowIndice_;
+   * [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,
+   *  16,17,18,19,20,21,22,23,24,25,26,27]
+   */
+
+  void calSelectedCols(const MatrixPtr selectedIndices,
+                       const std::vector<std::vector<int>>& inputSeqInfo);
+
+  // if the second input of this layer is on GPU memory, copy it to CPU memory.
+  MatrixPtr selIdsCpu_;
+
+  // reorganized sequenceStartPositions and subSequenceStartPositions
+  // into a 2d vector to facilitate the sequence selection process.
+  std::vector<std::vector<int>> inputSeqInfoVec_;
+
+  // the final selected row indices in a batch,
+  // rowIdx_ and selectedRows_ actually share a same memory.
+  IVectorPtr rowIndice_;
+  std::vector<int> selectedRows_;
+};
+
+REGISTER_LAYER(sub_nested_seq, SubNestedSequenceLayer);
+
+bool SubNestedSequenceLayer::init(const LayerMap& layerMap,
+                                  const ParameterMap& parameterMap) {
+  /* Initialize the basic parent class */
+  Layer::init(layerMap, parameterMap);
+  CHECK_EQ(2U, inputLayers_.size());
+  setNeedSequenceInfo(false);
+  return true;
+}
+
+void SubNestedSequenceLayer::calSelectedCols(
+    const MatrixPtr selectedIndices,
+    const std::vector<std::vector<int>>& inputSeqInfo) {
+  selectedRows_.clear();
+
+  std::vector<int> outSeqStartInfo(1, 0);
+  std::vector<int> outSubSeqStartInfo(1, 0);
+
+  size_t seqNum = selectedIndices->getHeight();
+  size_t beamSize = selectedIndices->getWidth();
+  for (size_t i = 0; i < seqNum; ++i) {
+    for (size_t j = 0; j < beamSize; ++j) {
+      if (selectedIndices->getElement(i, j) == -1.) break;
+      int selSubSeqIdx = selectedIndices->getElement(i, j);
+      CHECK_GT(inputSeqInfoVec_[i].size() - 1, selSubSeqIdx);
+
+      size_t subSeqLen = inputSeqInfoVec_[i][selSubSeqIdx + 1] -
+                         inputSeqInfoVec_[i][selSubSeqIdx];
+      for (size_t k = 0; k < subSeqLen; ++k)
+        selectedRows_.push_back(inputSeqInfoVec_[i][selSubSeqIdx] + k);
+      outSubSeqStartInfo.push_back(outSubSeqStartInfo.back() + subSeqLen);
+    }
+    outSeqStartInfo.push_back(outSubSeqStartInfo.back());
+  }
+
+  if (useGpu_) {
+    rowIndice_ = IVector::create(selectedRows_.size(), useGpu_);
+    rowIndice_->copyFrom(selectedRows_.data(), selectedRows_.size());
+  } else {
+    rowIndice_ =
+        IVector::create(selectedRows_.data(), selectedRows_.size(), useGpu_);
+  }
+
+  // create the sequence information for the output.
+  ICpuGpuVector::resizeOrCreate(
+      output_.sequenceStartPositions, outSeqStartInfo.size(), false);
+  output_.sequenceStartPositions->copyFrom(
+      outSeqStartInfo.data(), outSeqStartInfo.size(), false);
+
+  ICpuGpuVector::resizeOrCreate(
+      output_.subSequenceStartPositions, outSubSeqStartInfo.size(), false);
+  output_.subSequenceStartPositions->copyFrom(
+      outSubSeqStartInfo.data(), outSubSeqStartInfo.size(), false);
+}
+
+void SubNestedSequenceLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  const Argument& inputSeq = getInput(0);
+  CHECK(inputSeq.hasSubseq()) << "The first input of SubNestSequence layer "
+                              << "must be a nested sequence.";
+  const MatrixPtr selectedIndices = getInputValue(1);
+  CHECK_EQ(inputSeq.getNumSequences(), selectedIndices->getHeight());
+
+  if (dynamic_cast<GpuMatrix*>(selectedIndices.get())) {
+    /*
+     * Currently, the second input for this layer is generated by
+     * kmax_sequence_score_layer whose output is always stored on CPU,
+     * or a data_layer which canbe on GPU.
+     *
+     * If the second input is on GPU, copy it to CPU memory, because this
+     * input always uses very few memory, and operations related to it are
+     * all logic control, not computations.
+     */
+    Matrix::resizeOrCreate(selIdsCpu_,
+                           selectedIndices->getHeight(),
+                           selectedIndices->getWidth(),
+                           false /* trans */,
+                           false /* useGpu */);
+    selIdsCpu_->copyFrom(*selectedIndices);
+  } else {
+    selIdsCpu_ = selectedIndices;
+  }
+
+  Argument::reorganizeSeqInfo(inputSeq.sequenceStartPositions,
+                              inputSeq.subSequenceStartPositions,
+                              inputSeqInfoVec_);
+  calSelectedCols(selIdsCpu_, inputSeqInfoVec_);
+
+  resetOutput(selectedRows_.size(), getSize());
+  getOutputValue()->selectRows(*getInputValue(0), *rowIndice_);
+}
+
+void SubNestedSequenceLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inputSeqGrad = getInputGrad(0);
+  MatrixPtr outputGrad = getOutputGrad();
+
+  if (inputSeqGrad) outputGrad->addToRows(*inputSeqGrad, *rowIndice_);
+}
+
+}  // namespace paddle

paddle/gserver/tests/test_BatchNorm.cpp

@@ -21,6 +21,8 @@ limitations under the License. */
 #include "paddle/utils/GlobalConstants.h"
 
 #include "LayerGradUtil.h"
+#include "paddle/cuda/include/hl_batch_norm.h"
+#include "paddle/math/tests/TensorCheck.h"
 #include "paddle/testing/TestUtil.h"
 
 using namespace paddle;  // NOLINT
@@ -117,6 +119,74 @@ TEST(Layer, batchNorm) {
   CHECK_EQ(static_cast<int>(convLayer->getOutputValue()->getWidth()), 576);
 }
 
+#ifndef PADDLE_ONLY_CPU
+void batchNormInference(int n, int c, int h, int w) {
+  MatrixPtr input = std::make_shared<GpuMatrix>(n, c * h * w);
+  MatrixPtr cudnnOut = std::make_shared<GpuMatrix>(n, c * h * w);
+  MatrixPtr cudaOut = std::make_shared<GpuMatrix>(n, c * h * w);
+  MatrixPtr cudnnCheck = std::make_shared<CpuMatrix>(n, c * h * w);
+  MatrixPtr cudaCheck = std::make_shared<CpuMatrix>(n, c * h * w);
+  input->randomizeUniform();
+  cudnnOut->zeroMem();
+  cudaOut->zeroMem();
+
+  MatrixPtr scale = std::make_shared<GpuMatrix>(1, c);
+  scale->randomizeUniform();
+  MatrixPtr bias = std::make_shared<GpuMatrix>(1, c);
+  bias->randomizeUniform();
+
+  MatrixPtr movingMean = std::make_shared<GpuMatrix>(1, c);
+  movingMean->randomizeUniform();
+
+  MatrixPtr movingVar = std::make_shared<GpuMatrix>(1, c);
+  movingVar->randomizeUniform();
+  movingVar->clip(0.01, 50);
+
+  hl_tensor_descriptor ioDesc;
+  hl_tensor_descriptor bnDesc;
+  hl_create_tensor_descriptor(&ioDesc);
+  hl_create_tensor_descriptor(&bnDesc);
+  hl_tensor_reshape(ioDesc, n, c, h, w);
+  hl_tensor_reshape(bnDesc, 1, c, 1, 1);
+
+  double EPS = 1E-5;
+  hl_batch_norm_forward_inference(ioDesc,
+                                  input->getData(),
+                                  ioDesc,
+                                  cudnnOut->getData(),
+                                  bnDesc,
+                                  scale->getData(),
+                                  bias->getData(),
+                                  movingMean->getData(),
+                                  movingVar->getData(),
+                                  EPS);
+
+  hl_batch_norm_cuda_inference(input->getData(),
+                               cudaOut->getData(),
+                               scale->getData(),
+                               bias->getData(),
+                               movingMean->getData(),
+                               movingVar->getData(),
+                               EPS,
+                               n,
+                               c,
+                               h,
+                               w);
+
+  cudnnCheck->copyFrom(*cudnnOut);
+  cudaCheck->copyFrom(*cudaOut);
+  autotest::TensorCheckErr(*cudnnCheck, *cudaCheck);
+
+  hl_destroy_tensor_descriptor(ioDesc);
+  hl_destroy_tensor_descriptor(bnDesc);
+}
+
+TEST(BatchNorm, Inference) {
+  batchNormInference(33, 267, 1, 1);
+  batchNormInference(19, 105, 4, 4);
+}
+#endif
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1899,6 +1899,84 @@ TEST(Layer, CropLayer) {
   }
 }
 
+vector<real> randSampling(real range, int n) {
+  CHECK_GE(range, n);
+  vector<real> num(range);
+  iota(begin(num), end(num), 0.);
+  if (range == n) return num;
+
+  random_shuffle(begin(num), end(num));
+  num.resize(n);
+  sort(begin(num), end(num));
+  return num;
+}
+
+TEST(Layer, SubNestedSequenceLayer) {
+  // layer size is not crutial for this layer,
+  // so use a small layer size in unittest
+  const int layerSize = 4;
+
+  const int maxSeqNum = 50;
+  const int maxSeqLen = 50;
+  const int maxBeamSize = 32;
+
+  srand((size_t)(time(NULL)));
+  int beamSize = 1 + (rand() % maxBeamSize);
+
+  TestConfig config;
+  config.layerConfig.set_type("sub_nested_seq");
+  config.layerConfig.set_name("sub_nested_seq_layer");
+  config.layerConfig.set_size(layerSize);
+
+  int seqNum = 1 + (rand() % maxSeqNum);
+
+  // sequence information for the first input, it is a nested sequence
+  vector<int> seqStartPos(seqNum + 1, 0);
+  vector<int> subSeqStartPos(1, 0);
+
+  // selected indices
+  MatrixPtr selectedIndices = Matrix::create(seqNum, beamSize, false, false);
+  selectedIndices->one();
+  selectedIndices->mulScalar(-1.);
+  real* indicesData = selectedIndices->getData();
+
+  for (int i = 0; i < seqNum; ++i) {
+    int subSeqNum = 1 + (rand() % maxSeqNum);
+    for (int j = 0; j < subSeqNum; ++j) {
+      subSeqStartPos.push_back(subSeqStartPos.back() +
+                               (1 + (rand() % maxSeqLen)));
+    }
+    vector<real> selSeqs =
+        randSampling(static_cast<real>(subSeqNum), min(beamSize, subSeqNum));
+    memcpy(indicesData + (i * beamSize),
+           selSeqs.data(),
+           selSeqs.size() * sizeof(real));
+    seqStartPos[i + 1] = subSeqStartPos.back();
+  }
+
+  MatrixPtr seqInputPtr =
+      Matrix::create(seqStartPos.back(), layerSize, false, false);
+  seqInputPtr->randomizeUniform();
+  config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA,
+                              "nested_seq_input",
+                              seqInputPtr,
+                              seqStartPos,
+                              subSeqStartPos});
+  config.layerConfig.add_inputs();
+  config.inputDefs.push_back(
+      {INPUT_SELF_DEFINE_DATA, "selected_indices", selectedIndices});
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config,
+                  "sub_nested_seq",
+                  /* batchSize */ seqNum,
+                  /* trans */ false,
+                  /* useGpu*/ useGpu,
+                  /* useWeight */ false);
+  }
+}
+
 TEST(Layer, ClipLayer) {
   const size_t batchSize = 128;
   const size_t size = 512;

paddle/operators/add_op.cc

@@ -22,8 +22,7 @@ class AddOp : public OperatorWithKernel {
   void InferShape(const InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_EQ(ctx.InputSize(), 2);
     PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1);
-    PADDLE_ENFORCE(ctx.InputVar(0) != nullptr && ctx.InputVar(1) != nullptr,
-                   "Inputs of AddOp must all be set");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "Inputs of AddOp must all be set");
     PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr,
                    "Outputs of AddOp must all be set");
     PADDLE_ENFORCE(ctx.Input<Tensor>(0)->dims() == ctx.Input<Tensor>(1)->dims(),

paddle/operators/cross_entropy_op.cc

@@ -20,17 +20,18 @@ namespace operators {
 class OnehotCrossEntropyOp : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 2,
+    PADDLE_ENFORCE_EQ(ctx.InputSize(), 2,
                       "Input size of OnehotCrossEntropyOp must be two");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1,
+    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1,
                       "Output size of OnehotCrossEntropyOp must be one");
-    PADDLE_ENFORCE(ctx.InputVar(0) != nullptr && ctx.InputVar(1) != nullptr,
-                   "Inputs of OnehotCrossEntropyOp must all be set");
-    PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr,
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0),
+                            "0-th input of OnehotCrossEntropyOp should be set");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(1),
+                            "1-th input of OnehotCrossEntropyOp should be set");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0),
                             "Outputs of OnehotCrossEntropyOp must all be set");
-    PADDLE_ENFORCE(ctx.Input<Tensor>(0)->dims().size() == 2,
-                   "X's dimension must be 2.");
-    PADDLE_ENFORCE(ctx.Output<Tensor>(0)->dims().size() == 1,
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>(0)->dims().size(), 2);
+    PADDLE_ENFORCE_EQ(ctx.Output<Tensor>(0)->dims().size(), 1,
                       "label's dimension must be 1.");
     ctx.Output<Tensor>(0)->Resize({ctx.Input<Tensor>(0)->dims()[0]});
   }

paddle/operators/cross_entropy_op.cu

@@ -14,6 +14,3 @@
 
 #define EIGEN_USE_GPU
 #include "paddle/operators/cross_entropy_op.h"
-
-REGISTER_OP_GPU_KERNEL(onehot_cross_entropy,
-                       ops::OnehotCrossEntropyOpKernel<ops::GPUPlace, float>);

paddle/operators/cross_entropy_op.h

@@ -18,7 +18,24 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-static const float kCrossEntropyLogThreshold{1e-20};
+template <typename T>
+T tolerable_value(T x) {
+  static_assert(std::is_floating_point<T>::value,
+                "tolerable_value works only on float, "
+                "double and double double.");
+
+  const T kApproInf = 1e20;
+
+  if (x == INFINITY) {
+    return kApproInf;
+  }
+
+  if (x == -INFINITY) {
+    return -kApproInf;
+  }
+
+  return x;
+}
 
 template <typename Place, typename T>
 class OnehotCrossEntropyOpKernel : public OpKernel {
@@ -36,10 +53,9 @@ class OnehotCrossEntropyOpKernel : public OpKernel {
     int batch_size = X->dims()[0];
     int class_num = X->dims()[1];
 
-    // Y[i] = -log(X[i][j])
     for (int i = 0; i < batch_size; ++i) {
-      Ydata[i] = -std::log(std::max(Xdata[i * class_num + label_data[i]],
-                                    kCrossEntropyLogThreshold));
+      int index = i * class_num + label_data[i];
+      Ydata[i] = -tolerable_value(std::log(Xdata[index]));
     }
   }
 };
@@ -62,9 +78,8 @@ class OnehotCrossEntropyGradientOpKernel : public OpKernel {
     const int class_num = X->dims()[1];
 
     for (int i = 0; i < batch_size; ++i) {
-      dXdata[i * class_num + label_data[i]] =
-          -dYdata[i] / std::max(Xdata[i * class_num + label_data[i]],
-                                kCrossEntropyLogThreshold);
+      int index = i * class_num + label_data[i];
+      dXdata[index] = -tolerable_value(dYdata[i] / Xdata[index]);
     }
   }
 };

paddle/operators/fill_zeros_like_op.cc

@@ -13,8 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/operators/fill_zeros_like_op.h"
-#include "paddle/framework/op_registry.h"
-#include "paddle/framework/tensor.h"
 
 namespace paddle {
 namespace operators {
@@ -22,13 +20,13 @@ namespace operators {
 class FillZerosLikeOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 1UL,
+    PADDLE_ENFORCE_EQ(ctx.InputSize(), 1UL,
                       "Input size of FillZerosLikeOp must be one.");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1UL,
+    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL,
                       "Output size of AddOp must be one.");
-    PADDLE_ENFORCE(ctx.InputVar(0) != nullptr,
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0),
                             "Input of FillZerosLikeOp must be set.");
-    PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr,
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0),
                             "Output of FillZerosLikeOp must be set.");
     ctx.Output<framework::Tensor>(0)->Resize(
         ctx.Input<framework::Tensor>(0)->dims());

paddle/operators/fill_zeros_like_op.cu

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

paddle/operators/fill_zeros_like_op.h

@@ -13,9 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#include "glog/logging.h"
-#include "paddle/framework/eigen.h"
-#include "paddle/framework/operator.h"
+#include "paddle/operators/type_alias.h"
 
 namespace paddle {
 namespace operators {
@@ -26,7 +24,8 @@ class FillZerosLikeKernel : public framework::OpKernel {
   void Compute(const framework::ExecutionContext& context) const override {
     auto* output = context.Output<framework::Tensor>(0);
     output->mutable_data<T>(context.GetPlace());
-    framework::EigenVector<T>::Flatten(*output).setZero();
+    auto t = framework::EigenVector<T>::Flatten(*output);
+    t.device(context.GetEigenDevice<Place>()) = t.constant(T(0));
   }
 };
 

paddle/operators/mean_op.cc

@@ -20,10 +20,10 @@ namespace operators {
 class MeanOp : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 1, "Input size of AddOp must be one");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1, "Output size of AddOp must be one");
-    PADDLE_ENFORCE(ctx.InputVar(0) != nullptr && ctx.OutputVar(0) != nullptr,
-                   "Input/Output of MeanOp must be initialized.");
+    PADDLE_ENFORCE_EQ(ctx.InputSize(), 1, "Input size of AddOp must be one");
+    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "Output size of AddOp must be one");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "input should be set");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0), "output should be set");
     ctx.Output<Tensor>(0)->Resize(framework::make_ddim({1}));
   }
 };

paddle/operators/net_op.h

@@ -70,15 +70,15 @@ class NetOp : public framework::OperatorBase {
    */
   void AddOp(const std::shared_ptr<OperatorBase>& op) {
     PADDLE_ENFORCE(!add_op_done_, "Cannot AddOp when this network is sealed");
-    PADDLE_ENFORCE(op != nullptr, "Cannot Insert Null op");
+    PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op");
     ops_.push_back(op);
   }
 
   void InsertOp(size_t pos, const std::shared_ptr<OperatorBase>& op) {
     PADDLE_ENFORCE(!add_op_done_,
                    "Cannot InsertOp when this network is sealed");
-    PADDLE_ENFORCE(op != nullptr, "Cannot Insert Null op");
-    PADDLE_ENFORCE(pos <= ops_.size(), "Out of range");
+    PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op");
+    PADDLE_ENFORCE_LE(pos, ops_.size(), "Out of range");
     ops_.insert(ops_.begin() + pos, op);
   }
 

paddle/operators/sgd_op.cc

@@ -20,11 +20,11 @@ namespace operators {
 class SGDOp : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 2, "Input size of SGDOp must be two");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1, "Output size of SGDOp must be one");
-    PADDLE_ENFORCE(ctx.InputVar(0) != nullptr, "inputs[0] mast be set");
-    PADDLE_ENFORCE(ctx.InputVar(1) != nullptr, "inputs[1] mast be set");
-    PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr, "outputs[0] mast be set");
+    PADDLE_ENFORCE_EQ(ctx.InputSize(), 2, "Input size of SGDOp must be two");
+    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "Output size of SGDOp must be one");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "inputs[0] mast be set");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(1), "inputs[1] mast be set");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0), "outputs[0] mast be set");
     PADDLE_ENFORCE(ctx.Input<Tensor>(0)->dims() == ctx.Input<Tensor>(1)->dims(),
                    "Two input of SGD Op's dimension must be same.");
     ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());

paddle/operators/softmax_op.cc

@@ -20,11 +20,11 @@ namespace operators {
 class SoftmaxOp : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 1UL,
+    PADDLE_ENFORCE_EQ(ctx.InputSize(), 1UL,
                       "Only one input is need for softmax");
-    PADDLE_ENFORCE(ctx.Input<Tensor>("X")->dims().size() == 2UL,
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims().size(), 2UL,
                       "The input of softmax op must be matrix");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1UL,
+    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL,
                       "Only one output is need for softmax");
     ctx.Output<Tensor>("Y")->Resize(ctx.Input<Tensor>("X")->dims());
   }
@@ -43,12 +43,12 @@ class SoftmaxOpMaker : public OpProtoAndCheckerMaker {
 class SoftmaxOpGrad : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 3UL,
+    PADDLE_ENFORCE_EQ(ctx.InputSize(), 3UL,
                       "Input of SoftmaxOpGrad should be 3, X, Y, YG");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1UL,
+    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL,
                       "Output of SoftmaxOpGrad should be 1");
-    PADDLE_ENFORCE(ctx.InputVar("Y") != nullptr, "Input(Y) should not be null");
-    PADDLE_ENFORCE(ctx.InputVar(framework::GradVarName("Y")) != nullptr,
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Y")),
                             "Input(Y@GRAD) should not be null");
     PADDLE_ENFORCE(ctx.Input<Tensor>("Y")->dims() ==
                        ctx.Input<Tensor>(framework::GradVarName("Y"))->dims(),

paddle/parameter/Argument.cpp

@@ -666,4 +666,24 @@ void Argument::subArgFrom(const Argument& input,
   }
 }
 
+void Argument::reorganizeSeqInfo(
+    const ICpuGpuVectorPtr seqStartPos,
+    const ICpuGpuVectorPtr subSeqStartPos,
+    std::vector<std::vector<int>>& reorganizedSeqInfo) {
+  int* seqStarts = seqStartPos->getMutableData(false);
+  int* subSeqStarts = subSeqStartPos->getMutableData(false);
+
+  int seqNum = seqStartPos->getSize() - 1;
+  reorganizedSeqInfo.resize(seqNum, std::vector<int>());
+  int seqIdx = 0;
+  for (size_t i = 0; i < subSeqStartPos->getSize(); ++i) {
+    reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]);
+    if (subSeqStarts[i] == seqStarts[seqIdx + 1]) {
+      seqIdx++;
+      if (seqIdx == seqNum) return;
+      reorganizedSeqInfo[seqIdx].push_back(subSeqStarts[i]);
+    }
+  }
+}
+
 }  // namespace paddle

paddle/parameter/Argument.h

@@ -317,6 +317,30 @@ struct Argument {
    */
   void printValueString(std::ostream& stream,
                         const std::string& prefix = "") const;
+
+  /**
+   * @brief reorganizeSeqInfo will reorganize sequenceStartPositions and
+   * subSequenceStartPositions into a 2 dimensional arrary: reorganizedSeqInfo.
+   *
+   * @param seqStartPos: sequenceStartPositions of an Argument.
+   * @param subSeqStartPos: subSequenceStartPositions of an Argument.
+   * @param the reorganized sequence start position information.
+   *
+   * Examples:
+   * seqStartPos: [0, 4, 15, 20, 28]
+   * subSeqStartPos: [0, 3, 4, 5, 7, 10, 15, 20, 22, 23, 25, 28]
+   * reorganizedSeqInfo:
+   *   [
+   *     [0,3,4],
+   *     [4,5,7,10,15],
+   *     [15,20],
+   *     [20,22,23,25,28]
+   *   ]
+   */
+  static void reorganizeSeqInfo(
+      const ICpuGpuVectorPtr seqStartPos,
+      const ICpuGpuVectorPtr subSeqStartPos,
+      std::vector<std::vector<int>>& reorganizedSeqInfo);
 };
 
 }  // namespace paddle

paddle/platform/enforce.h

@@ -187,25 +187,16 @@ inline void throw_on_error(T e) {
   __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, <, >=, __VA_ARGS__)
 #define PADDLE_ENFORCE_LE(__VAL0, __VAL1, ...) \
   __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, <=, >, __VA_ARGS__)
-
-// if two values have different data types, choose a compatible type for them.
-template <typename T1, typename T2>
-struct CompatibleType {
-  static const bool t1_to_t2 = std::is_convertible<T1, T2>::value;
-  typedef typename std::conditional<t1_to_t2, T2, T1>::type type;
-};
+#define PADDLE_ENFORCE_NOT_NULL(__VAL, ...)                            \
+  PADDLE_ENFORCE(nullptr != (__VAL), #__VAL " should not be null\n%s", \
+                 paddle::string::Sprintf("" __VA_ARGS__));
 
 #define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...)        \
-  PADDLE_ENFORCE(__COMPATIBLE_TYPE(__VAL0, __VAL1, __VAL0)                    \
-                     __CMP __COMPATIBLE_TYPE(__VAL0, __VAL1, __VAL1),         \
+  PADDLE_ENFORCE(__VAL0 __CMP __VAL1,                                         \
                  "enforce %s " #__CMP " %s failed, %s " #__INV_CMP " %s\n%s", \
                  #__VAL0, #__VAL1, std::to_string(__VAL0),                    \
                  std::to_string(__VAL1),                                      \
                  paddle::string::Sprintf("" __VA_ARGS__));
 
-#define __COMPATIBLE_TYPE(__VAL0, __VAL1, __VAL)              \
-  typename paddle::platform::CompatibleType<decltype(__VAL0), \
-                                            decltype(__VAL1)>::type(__VAL)
-
 }  // namespace platform
 }  // namespace paddle

paddle/platform/enforce_test.cc

@@ -9,8 +9,10 @@ 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/platform/enforce.h"
+#include <memory>
+
 #include "gtest/gtest.h"
+#include "paddle/platform/enforce.h"
 
 TEST(ENFORCE, OK) {
   PADDLE_ENFORCE(true, "Enforce is ok %d now %f", 123, 0.345);
@@ -196,3 +198,27 @@ TEST(ENFORCE_LT, FAIL) {
 
   ASSERT_TRUE(in_catch);
 }
+
+TEST(ENFORCE_NOT_NULL, OK) {
+  int* a = new int;
+  PADDLE_ENFORCE_NOT_NULL(a);
+  delete a;
+}
+TEST(ENFORCE_NOT_NULL, FAIL) {
+  bool in_catch = false;
+  int* a{nullptr};
+
+  try {
+    PADDLE_ENFORCE_NOT_NULL(a);
+
+  } catch (paddle::platform::EnforceNotMet error) {
+    in_catch = true;
+    const std::string msg = "a should not be null";
+    const char* what = error.what();
+    for (size_t i = 0; i < msg.length(); ++i) {
+      ASSERT_EQ(what[i], msg[i]);
+    }
+  }
+
+  ASSERT_TRUE(in_catch);
+}

paddle/pybind/CMakeLists.txt

@@ -6,4 +6,5 @@ cc_library(paddle_pybind SHARED
 	add_op
 	mean_op
 	cross_entropy_op
-	recurrent_op)
+	recurrent_op
+	fill_zeros_like_op)

paddle/scripts/CMakeLists.txt

 configure_file(submit_local.sh.in
-    submit_local.sh
+    paddle
     @ONLY)
 
-install(FILES ${CMAKE_CURRENT_BINARY_DIR}/submit_local.sh DESTINATION bin
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/paddle DESTINATION bin
         PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ
-            GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ
-        RENAME paddle)
+            GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ)
 
 configure_file(tools/usage_stat/usage.sh
-    usage.sh
+    paddle_usage
     @ONLY)
 
-install(FILES ${CMAKE_CURRENT_BINARY_DIR}/usage.sh DESTINATION opt/paddle/bin
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/paddle_usage DESTINATION opt/paddle/bin
         PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ
-            GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ
-        RENAME paddle_usage)
+            GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ)

paddle/scripts/docker/build.sh

@@ -33,6 +33,9 @@ Configuring cmake in /paddle/build ...
       -DWITH_AVX=${WITH_AVX:-OFF}
       -DWITH_GOLANG=${WITH_GOLANG:-OFF}
       -DWITH_SWIG_PY=ON
+      -DWITH_C_API=${WITH_C_API:-OFF}
+      -DWITH_PYTHON=${WITH_PYTHON:-ON}
+      -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON}
       -DCUDNN_ROOT=/usr/
       -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF}
       -DWITH_TESTING=${WITH_TESTING:-OFF}
@@ -49,7 +52,9 @@ cmake .. \
       -DWITH_GPU=${WITH_GPU:-OFF} \
       -DWITH_AVX=${WITH_AVX:-OFF} \
       -DWITH_GOLANG=${WITH_GOLANG:-OFF} \
-      -DWITH_SWIG_PY=ON \
+      -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} \
+      -DWITH_C_API=${WITH_C_API:-OFF} \
+      -DWITH_PYTHON=${WITH_PYTHON:-ON} \
       -DCUDNN_ROOT=/usr/ \
       -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} \
       -DWITH_TESTING=${WITH_TESTING:-OFF} \

paddle/scripts/docker/build_android.sh

@@ -20,4 +20,4 @@ cmake -DCMAKE_SYSTEM_NAME=Android \
       -DWITH_SWIG_PY=OFF \
       ..
 make -j `nproc`
-make install
+make install -j `nproc`

paddle/setup.py.in

-# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-
-from setuptools import setup, Extension
-
-setup(name="py_paddle",
-      version="${PADDLE_VERSION}",
-      packages=['py_paddle'],
-      include_package_data=True,
-      package_data={'py_paddle':['*.py','_swig_paddle.so']},
-      install_requires = [
-        'nltk>=3.2.2',
-        # We use `numpy.flip` in `test_image.py`.
-        # `numpy.flip` is introduced in `1.12.0`
-        'numpy>=1.12.0',      # The numpy is required.
-        'protobuf==${PROTOBUF_VERSION}'    # The paddle protobuf version
-      ],
-      url='http://www.paddlepaddle.org/',
-      license='Apache 2.0',
-)

paddle/trainer/tests/simple_sparse_neural_network.py

 from paddle.trainer_config_helpers import *
 
-settings(batch_size=128, learning_method=AdaGradOptimizer(), learning_rate=1e-4)
+settings(batch_size=17, learning_method=AdaGradOptimizer(), learning_rate=1e-4)
 
 file_list = 'trainer/tests/fake_file_list.list'
 
@@ -12,7 +12,7 @@ define_py_data_sources2(
 
 embedding = embedding_layer(
     input=data_layer(
-        name="word_ids", size=65536),
+        name="word_ids", size=8191),
     size=128,
     param_attr=ParamAttr(sparse_update=True))
 prediction = fc_layer(input=embedding, size=10, act=SoftmaxActivation())

paddle/trainer/tests/simple_sparse_neural_network_dp.py

@@ -7,15 +7,15 @@ def init_hook(settings, is_train, **kwargs):
 
 
 @provider(
-    input_types={'word_ids': integer_value(65536),
+    input_types={'word_ids': integer_value(8191),
                  'label': integer_value(10)},
     min_pool_size=0,
     init_hook=init_hook)
 def process(settings, filename):
     if settings.is_train:
-        data_size = 2**20
-    else:
         data_size = 2**10
+    else:
+        data_size = 2**5
 
     for _ in xrange(data_size):
-        yield random.randint(0, 65535), random.randint(0, 9)
+        yield random.randint(0, 8190), random.randint(0, 9)

paddle/trainer/tests/test_TrainerOnePass.cpp

@@ -100,25 +100,25 @@ TEST(average_window, gpu) {
 }
 
 TEST(average_window, gpu2) {
-  FLAGS_num_passes = 100;
+  FLAGS_num_passes = 20;
   trainerOnePassTest(configFile1, true, false, 2, 0.01);
   FLAGS_num_passes = 1;
 }
 
 TEST(average_window, gpu4) {
-  FLAGS_num_passes = 100;
+  FLAGS_num_passes = 20;
   trainerOnePassTest(configFile1, true, false, 4, 0.01);
   FLAGS_num_passes = 1;
 }
 
 TEST(average_window_cpu, gpu2) {
-  FLAGS_num_passes = 100;
+  FLAGS_num_passes = 20;
   trainerOnePassTest(configFile1, true, false, 2, 0.01, true);
   FLAGS_num_passes = 1;
 }
 
 TEST(average_window_cpu, gpu4) {
-  FLAGS_num_passes = 100;
+  FLAGS_num_passes = 20;
   trainerOnePassTest(configFile1, true, false, 4, 0.01, true);
   FLAGS_num_passes = 1;
 }

python/CMakeLists.txt

@@ -39,7 +39,7 @@ add_custom_command(OUTPUT ${OUTPUT_DIR}/.timestamp
     DEPENDS gen_proto_py copy_paddle_pybind framework_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
 
 add_custom_target(paddle_python ALL DEPENDS
-    ${OUTPUT_DIR}/.timestamp)
+    ${OUTPUT_DIR}/.timestamp paddle_pserver_main paddle_trainer paddle_merge_model python_api_wheel)
 
 set(PADDLE_PYTHON_PACKAGE_DIR ${CMAKE_CURRENT_BINARY_DIR}/dist/)
 

python/paddle/trainer/config_parser.py

@@ -2657,6 +2657,31 @@ class SubSequenceLayer(LayerBase):
         self.create_bias_parameter(bias, size)
 
 
+@config_layer('sub_nested_seq')
+class SubNestedSequenceLayer(LayerBase):
+    def __init__(self, name, inputs, selected_indices, bias=False, **xargs):
+        if isinstance(inputs, list):
+            assert len(inputs) == 1, ('the first input of sub_nested_seq '
+                                      'layer is a single nested sequence.')
+            inputs = inputs[0]
+        if isinstance(selected_indices, list):
+            assert len(selected_indices) == 1, (
+                'the second input of '
+                'sub_nested_seq layer is a single layer which is a '
+                'set of selected indices.')
+            selected_indices = selected_indices[0]
+
+        super(SubNestedSequenceLayer, self).__init__(
+            name,
+            'sub_nested_seq',
+            0,
+            inputs=[inputs, selected_indices],
+            **xargs)
+        input_layer0 = self.get_input_layer(0)
+        size = input_layer0.size
+        self.set_layer_size(size)
+
+
 @config_layer('out_prod')
 class OuterProdLayer(LayerBase):
     def __init__(self, name, inputs, device=None):

python/paddle/trainer_config_helpers/layers.py

@@ -129,6 +129,7 @@ __all__ = [
     'prelu_layer',
     'gated_unit_layer',
     'crop_layer',
+    'sub_nested_seq_layer',
     'clip_layer',
     'slice_projection',
     'kmax_sequence_score_layer',
@@ -225,6 +226,7 @@ class LayerType(object):
 
     PRELU = 'prelu'
     CROP_LAYER = 'crop'
+    SUB_NESTED_SEQ = 'sub_nested_seq'
     CLIP_LAYER = 'clip'
 
     KMAX_SEQ_SCORE = 'kmax_seq_score'
@@ -6091,6 +6093,53 @@ def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None):
         size=l.config.size)
 
 
+@wrap_name_default()
+@layer_support()
+def sub_nested_seq_layer(input, selected_indices, name=None):
+    """
+    The sub_nested_seq_layer accepts two inputs: the first one is a nested
+    sequence; the second one is a set of selceted indices in the nested sequence.
+
+    Then sub_nest_seq_layer trims the first nested sequence input according
+    to the selected indices to form a new output. This layer is useful in
+    beam training.
+
+    The example usage is:
+
+    .. code-block:: python
+
+        sub_nest_seq = sub_nested_seq_layer(input=[data, selected_indices])
+
+
+    :param input: A nested sequence.
+    :type input: LayerOutput
+    :param selected_indices: a set of sequence indices in the nested sequence.
+    :type input: LayerOutput
+    :param name: name of this layer.
+    :type name: basestring
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+
+    assert isinstance(input, LayerOutput), (
+        'The first input of '
+        'sub_nested_seq_layer must be a Paddle layer.')
+    assert isinstance(selected_indices, LayerOutput), (
+        'The second input of '
+        'sub_nested_seq_layer must be a Paddle layer.')
+
+    l = Layer(
+        inputs=input.name,
+        selected_indices=selected_indices.name,
+        name=name,
+        type=LayerType.SUB_NESTED_SEQ)
+    return LayerOutput(
+        name=name,
+        layer_type=LayerType.SUB_NESTED_SEQ,
+        parents=input,
+        size=l.config.size)
+
+
 @wrap_name_default("clip")
 def clip_layer(input, min, max, name=None):
     """

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

@@ -8,6 +8,6 @@ test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops
 test_seq_concat_reshape test_pad test_smooth_l1 test_multiplex_layer
 test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_layer
 test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer
-test_kmax_seq_socre_layer)
+test_kmax_seq_socre_layer test_seq_select_layers)
 
 export whole_configs=(test_split_datasource)

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

+type: "nn"
+layers {
+  name: "input_seq"
+  type: "data"
+  size: 300
+  active_type: ""
+}
+layers {
+  name: "input"
+  type: "data"
+  size: 5
+  active_type: ""
+}
+layers {
+  name: "__sub_nested_seq_layer_0__"
+  type: "sub_nested_seq"
+  size: 300
+  active_type: ""
+  inputs {
+    input_layer_name: "input_seq"
+  }
+  inputs {
+    input_layer_name: "input"
+  }
+}
+input_layer_names: "input_seq"
+output_layer_names: "__sub_nested_seq_layer_0__"
+sub_models {
+  name: "root"
+  layer_names: "input_seq"
+  layer_names: "input"
+  layer_names: "__sub_nested_seq_layer_0__"
+  input_layer_names: "input_seq"
+  output_layer_names: "__sub_nested_seq_layer_0__"
+  is_recurrent_layer_group: false
+}
+

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

+#!/usr/bin/env python
+#coding=utf-8
+from paddle.trainer_config_helpers import *
+
+beam_size = 5
+
+data = data_layer(name='input_seq', size=300)
+selected_ids = data_layer(name='input', size=beam_size)
+sub_nest_seq = sub_nested_seq_layer(input=data, selected_indices=selected_ids)
+
+outputs(sub_nest_seq)

python/paddle/v2/framework/network.py

-import paddle.v2.framework.core as core
-from paddle.v2.framework.create_op_creation_methods import op_creations
-from default_scope_funcs import new_var, find_var, get_cur_scope
-
-__all__ = ['Network']  # Only expose Network
-
-
-class NetworkFunctor(object):
-    """
-    Network Op Creation Function. Used internally in this module.
-    It convert string input to Variable. If it is not created before, just 
-    create in scope.
-    
-    It is a functor object. means the instances are callable.
-    
-    :param func: The op creation function which generated in Python.
-    :param net: The Network instance.
-    """
-
-    def __init__(self, func, net):
-        self.func = func
-        self.net = net
-
-    def __call__(self, *args, **kwargs):
-        if len(args) != 0:
-            raise ValueError("Paddle must use keyword argument")
-        inputs = self.func.all_input_args
-        for ipt in inputs:
-            if ipt in kwargs:
-                var = kwargs[ipt]
-                if isinstance(var, basestring):
-                    tmp = new_var(var)
-                    self.net.var_names[tmp] = var
-                    var = tmp
-
-                if not isinstance(var, core.Variable):
-                    raise TypeError(
-                        "Input of op creation must be string or variable")
-
-                kwargs[ipt] = self.net.var_names[var]
-
-        notemp_outputs = self.func.all_not_temp_output_args
-
-        for name in notemp_outputs:
-            if name not in kwargs:
-                kwargs[
-                    name] = self.func.__name__ + "@OUT@%d" % core.unique_integer(
-                    )
-
-        outputs = self.func.all_output_args
-        for opt in outputs:
-            if opt in kwargs:
-                var = kwargs[opt]
-                if isinstance(var, basestring):
-                    tmp = new_var(var)
-                    self.net.var_names[tmp] = var
-                    var = tmp
-
-                if not isinstance(var, core.Variable):
-                    raise TypeError(
-                        "Output of op creation must be string or variable")
-                kwargs[opt] = self.net.var_names[var]
-
-        op = self.func(**kwargs)
-
-        self.net.net.add_op(op)
-
-        lst = [find_var(kwargs[opt]) for opt in notemp_outputs]
-        if len(lst) == 1:
-            return lst[0]
-        elif len(lst) == 0:
-            return None
-        else:
-            return lst
-
-
-class Network(object):
-    """
-    The network concept. It avoid user to manually create operator, create 
-    variable, and combine them into a Net. Just use Network.xxx can create the
-    operator, create variables in default scope, and add them into `self.net`.
-    
-    For example:
-    
-    ..  code-block: python
-    
-        net = Network()
-        out = net.add_two(X="a", Y="b")
-        fc_out = net.fc(X="out", W="fc.w")
-        
-        net.run(...)
-    """
-
-    def __init__(self):
-        self.net = core.Net.create()
-        funcs = (func_name for func_name in dir(op_creations)
-                 if not func_name.startswith("__"))
-        self.var_names = dict()
-
-        # TODO(yuyang18): This code can work, but do not generate a good
-        # docstring, try to give a better way generate function in runtime
-        # later.
-        for func_name in funcs:
-            func = getattr(op_creations, func_name)
-            impl = NetworkFunctor(func, self)
-            setattr(self, func_name, impl.__call__)
-        self.__complete_add_op__ = False
-
-    def infer_shape(self):
-        self.complete_add_op()
-        self.net.infer_shape(get_cur_scope())
-
-    def run(self, device_context):
-        self.complete_add_op()
-        self.net.run(get_cur_scope(), device_context)
-
-    def __str__(self):
-        return str(self.net)
-
-    def complete_add_op(self):
-        if not self.__complete_add_op__:
-            self.net.complete_add_op()
-            self.__complete_add_op__ = True
-
-
-if __name__ == '__main__':
-    net = Network()
-    out = net.add_two(X="a", Y="b")
-    fc_out = net.fc(X=out, W="fc.w", b="fc.b", activation="softmax")
-    net.complete_add_op()
-    print net

python/paddle/v2/framework/create_op_creation_methods.py → python/paddle/v2/framework/op.py

@@ -2,7 +2,6 @@ import paddle.v2.framework.core as core
 import paddle.v2.framework.proto.op_proto_pb2 as op_proto_pb2
 import paddle.v2.framework.proto.op_desc_pb2 as op_desc_pb2
 import paddle.v2.framework.proto.attribute_pb2 as attribute_pb2
-import cStringIO
 
 
 def get_all_op_protos():
@@ -146,64 +145,14 @@ class OpDescCreationMethod(object):
         return False
 
 
-def get_docstring_from_op_proto(op_proto):
-    """
-    Generate docstring from a OpProto
-    :param op_proto: a OpProto instance.
-    :type op_proto: op_proto_pb2.OpProto
-    :return: docstring
-    """
-    if not isinstance(op_proto, op_proto_pb2.OpProto):
-        raise TypeError("Input must be OpProto")
-    f = cStringIO.StringIO()
-    f.write(op_proto.comment)
-    f.write("\n")
-
-    def __append_param__(name, comment, type):
-        # Maybe replace the following line with template engine is better.
-        f.write(":param ")
-        f.write(name)
-        f.write(": ")
-        f.write(comment)
-        f.write("\n")
-        f.write(":type ")
-        f.write(name)
-        f.write(": ")
-        f.write(type)
-        f.write("\n")
-
-    for ipt in op_proto.inputs:
-        __append_param__(ipt.name, ipt.comment, "list | basestr"
-                         if ipt.multiple else "basestr")
-
-    temp_var_prefix = \
-        "This is a temporary variable. It does not have to set by user. "
-    for opt in op_proto.outputs:
-        __append_param__(opt.name, opt.comment if not opt.temporary else
-                         temp_var_prefix + opt.comment, "list | basestr"
-                         if opt.multiple else "basestr")
-
-    for attr in op_proto.attrs:
-        attr_type = None
-        if attr.type == attribute_pb2.INT:
-            attr_type = "int"
-        elif attr.type == attribute_pb2.FLOAT:
-            attr_type = "float"
-        elif attr.type == attribute_pb2.STRING:
-            attr_type = "basestr"
-        elif attr.type == attribute_pb2.INTS:
-            attr_type = "list of int"
-        elif attr.type == attribute_pb2.FLOATS:
-            attr_type = "list of float"
-        elif attr.type == attribute_pb2.STRINGS:
-            attr_type = "list of basestr"
-
-        if attr_type is None:
-            raise RuntimeError("Not supported attribute type " + attr.type)
-
-        __append_param__(attr.name, attr.comment, attr_type)
-
-    return f.getvalue()
+class OpInfo(object):
+    def __init__(self, name, method, inputs, outputs, attrs, no_temp_outputs):
+        self.name = name
+        self.method = method
+        self.inputs = inputs
+        self.outputs = outputs
+        self.attrs = attrs
+        self.no_temp_outputs = no_temp_outputs
 
 
 def create_op_creation_method(op_proto):
@@ -216,38 +165,57 @@ def create_op_creation_method(op_proto):
         opdesc = method(*args, **kwargs)
         return core.Operator.create(opdesc.SerializeToString())
 
-    __impl__.__doc__ = get_docstring_from_op_proto(op_proto)
-    __impl__.all_input_args = [var.name for var in op_proto.inputs]
-    __impl__.all_output_args = [var.name for var in op_proto.outputs]
-    __impl__.all_attr_args = [attr.name for attr in op_proto.attrs]
-    __impl__.all_not_temp_output_args = [
+    return OpInfo(
+        method=__impl__,
+        name=op_proto.type,
+        inputs=[var.name for var in op_proto.inputs],
+        outputs=[var.name for var in op_proto.outputs],
+        attrs=[attr.name for attr in op_proto.attrs],
+        no_temp_outputs=[
             var.name for var in op_proto.outputs if not var.temporary
-    ]
+        ])
 
-    return __impl__
 
+class OperatorFactory(object):
+    def __init__(self):
+        self.op_methods = dict()
+        for op_proto in get_all_op_protos():
+            method = create_op_creation_method(op_proto)
+            self.op_methods[method.name] = method
 
-class OpCreationsHolder(object):
-    """
-    A object will holds all op creation methods.
+    def __call__(self, *args, **kwargs):
+        if 'type' in kwargs:
+            if len(args) != 0:
+                raise ValueError("All Paddle argument should be key-word "
+                                 "argument except type")
+            t = kwargs.pop('type')
+        else:
+            if len(args) != 1:
+                raise ValueError("All Paddle argument should be key-word "
+                                 "argument except type")
+            t = args[0]
 
-    Use `op_creations.xxx_op` to access them.
-    """
-    pass
+        return self.get_op_info(t).method(**kwargs)
 
+    def types(self):
+        return self.op_methods.keys()
 
-op_creations = OpCreationsHolder()
+    def get_op_info(self, t):
+        if t not in self.op_methods:
+            raise ValueError("operator %s is not registered", t)
+        return self.op_methods.get(t)
 
+    def get_op_input_names(self, type):
+        return self.get_op_info(type).inputs
 
-def __bootstrap__():
-    """
-    Bootstrap function for this module. It will dynamic create all op creation
-    methods in runtime.
-    """
-    for op_proto in get_all_op_protos():
-        func = create_op_creation_method(op_proto)
-        func.__name__ = str(op_proto.type)
-        setattr(op_creations, func.__name__, func)
+    def get_op_output_names(self, type):
+        return self.get_op_info(type).outputs
+
+    def get_op_attr_names(self, type):
+        return self.get_op_info(type).attrs
+
+    def get_op_no_temp_output_names(self, type):
+        return self.get_op_info(type).no_temp_outputs
 
 
-__bootstrap__()
+Operator = OperatorFactory()  # Default global factory

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

@@ -6,7 +6,6 @@ py_test(test_scope SRCS test_scope.py)
 py_test(test_tensor SRCS test_tensor.py)
 py_test(test_mul_op SRCS test_mul_op.py)
 
-py_test(test_network SRCS test_network.py)
 py_test(test_mean_op SRCS test_mean_op.py)
 
 py_test(test_protobuf SRCS test_protobuf.py)
@@ -14,10 +13,11 @@ py_test(test_protobuf SRCS test_protobuf.py)
 py_test(test_add_two_op SRCS test_add_two_op.py)
 py_test(test_sigmoid_op SRCS test_sigmoid_op.py)
 py_test(test_softmax_op SRCS test_softmax_op.py)
+py_test(test_fill_zeros_like_op SRCS test_fill_zeros_like_op.py)
 
 py_test(gradient_checker SRCS gradient_checker.py)
 
 py_test(test_rowwise_add_op SRCS test_rowwise_add_op.py)
 
 py_test(test_default_scope_funcs SRCS test_default_scope_funcs.py)
-py_test(test_op_creation_methods SRCS test_op_creation_methods.py)
+py_test(test_operator SRCS test_operator.py)

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

 import paddle.v2.framework.core as core
-from paddle.v2.framework.create_op_creation_methods import op_creations
+from paddle.v2.framework.op import Operator
 import numpy
 import unittest
 
@@ -80,7 +80,7 @@ if __name__ == '__main__':
 
     class GetNumericGradientTest(unittest.TestCase):
         def test_add_op(self):
-            add_op = op_creations.add_two(X="X", Y="Y", Out="Z")
+            add_op = Operator('add_two', X="X", Y="Y", Out="Z")
             x = numpy.random.random((10, 1)).astype("float32")
             y = numpy.random.random((10, 1)).astype("float32")
 

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

 import paddle.v2.framework.core as core
 import unittest
 import numpy
-import paddle.v2.framework.create_op_creation_methods as creation
+from paddle.v2.framework.op import Operator
 
 
 class OpTestMeta(type):
@@ -21,18 +21,14 @@ class OpTestMeta(type):
         obj = super(OpTestMeta, cls).__new__(cls, name, bases, attrs)
 
         def test_all(self):
-            func = getattr(creation.op_creations, self.type, None)
-            self.assertIsNotNone(func)
-
             scope = core.Scope()
             kwargs = dict()
-            places = []
-            places.append(core.CPUPlace())
-            if core.is_compile_gpu():
+            places = [core.CPUPlace()]
+            if core.is_compile_gpu() and core.Operator.support_gpu(self.type):
                 places.append(core.GPUPlace(0))
 
             for place in places:
-                for in_name in func.all_input_args:
+                for in_name in Operator.get_op_input_names(self.type):
                     if hasattr(self, "inputs") and in_name in self.inputs:
                         kwargs[in_name] = in_name
                         var = scope.new_var(in_name).get_tensor()
@@ -42,7 +38,7 @@ class OpTestMeta(type):
                     else:
                         kwargs[in_name] = "@EMPTY@"
 
-                for out_name in func.all_output_args:
+                for out_name in Operator.get_op_output_names(self.type):
                     if not hasattr(self, "outputs"):
                         raise ValueError(
                             "The test op must set self.outputs dict.")
@@ -52,21 +48,23 @@ class OpTestMeta(type):
                     kwargs[out_name] = out_name
                     scope.new_var(out_name).get_tensor()
 
-                for attr_name in func.all_attr_args:
+                for attr_name in Operator.get_op_attr_names(self.type):
                     if hasattr(self, "attrs") and attr_name in self.attrs:
                         kwargs[attr_name] = self.attrs[attr_name]
 
-                op = func(**kwargs)
+                op = Operator(self.type, **kwargs)
 
                 op.infer_shape(scope)
 
                 ctx = core.DeviceContext.create(place)
                 op.run(scope, ctx)
 
-                for out_name in func.all_output_args:
+                for out_name in Operator.get_op_output_names(self.type):
                     actual = numpy.array(scope.find_var(out_name).get_tensor())
                     expect = self.outputs[out_name]
-                    numpy.isclose(actual, expect)
+                    self.assertTrue(
+                        numpy.allclose(actual, expect),
+                        "output name: " + out_name + "has diff")
 
         obj.test_all = test_all
         return obj

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

@@ -2,7 +2,7 @@ import unittest
 
 import numpy
 import paddle.v2.framework.core as core
-import paddle.v2.framework.create_op_creation_methods as creation
+from paddle.v2.framework.op import Operator
 
 from op_test_util import OpTestMeta
 
@@ -21,7 +21,7 @@ class TestAddOp(unittest.TestCase):
 
 class TestAddGradOp(unittest.TestCase):
     def test_add_grad(self):
-        op = creation.op_creations.add_two(X="X", Y="Y", Out="Out")
+        op = Operator('add_two', X="X", Y="Y", Out="Out")
         backward_op = core.Operator.backward(op, set())
         self.assertEqual(backward_op.type(), "add_two_grad")
         expected = '''Op(add_two_grad), inputs:(X, Y, Out, Out@GRAD), outputs:(X@GRAD, Y@GRAD).'''

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

 import paddle.v2.framework.core as core
 import unittest
 import numpy
-import paddle.v2.framework.create_op_creation_methods as creation
+from paddle.v2.framework.op import Operator
 
 
 class TestFc(unittest.TestCase):
@@ -24,7 +24,7 @@ class TestFc(unittest.TestCase):
         # Set a real numpy array here.
         # x_tensor.set(numpy.array([]))
 
-        op = creation.op_creations.fc(X="X", Y="Y", W="W")
+        op = Operator("fc", X="X", Y="Y", W="W")
 
         for out in op.outputs():
             if scope.find_var(out) is None:

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

+import unittest
+from op_test_util import OpTestMeta
+import numpy
+
+
+class TestFillZerosLikeOp(unittest.TestCase):
+    __metaclass__ = OpTestMeta
+
+    def setUp(self):
+        self.type = "fill_zeros_like"
+        self.inputs = {'Src': numpy.random.random((219, 232)).astype("float32")}
+        self.outputs = {'Dst': numpy.zeros_like(self.inputs['Src'])}
+
+
+if __name__ == '__main__':
+    unittest.main()

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

 import paddle.v2.framework.core as core
-from paddle.v2.framework.create_op_creation_methods import op_creations
+from paddle.v2.framework.op import Operator
 import unittest
 
 
 class TestNet(unittest.TestCase):
     def test_net_all(self):
         net = core.Net.create()
-        op1 = op_creations.add_two(X="X", Y="Y", Out="Out")
+        op1 = Operator("add_two", X="X", Y="Y", Out="Out")
         net.add_op(op1)
 
         net2 = core.Net.create()
-        net2.add_op(op_creations.fc(X="X", W="w", Y="fc.out"))
+        net2.add_op(Operator("fc", X="X", W="w", Y="fc.out"))
         net2.complete_add_op(True)
         net.add_op(net2)
         net.complete_add_op(True)

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

-from paddle.v2.framework.network import Network
-import paddle.v2.framework.core as core
-import unittest
-
-
-class TestNet(unittest.TestCase):
-    def test_net_all(self):
-        net = Network()
-        out = net.add_two(X="X", Y="Y")
-        fc_out = net.fc(X=out, W="w")
-        net.complete_add_op()
-        self.assertTrue(isinstance(fc_out, core.Variable))
-        self.assertEqual(
-            '''Op(plain_net), inputs:(@EMPTY@, X, Y, w), outputs:(@TEMP@fc@0, add_two@OUT@0, fc@OUT@1).
-    Op(add_two), inputs:(X, Y), outputs:(add_two@OUT@0).
-    Op(fc), inputs:(add_two@OUT@0, w, @EMPTY@), outputs:(fc@OUT@1, @TEMP@fc@0).
-        Op(mul), inputs:(add_two@OUT@0, w), outputs:(@TEMP@fc@0).
-        Op(sigmoid), inputs:(@TEMP@fc@0), outputs:(fc@OUT@1).
-''', str(net))
-
-        net2 = Network()
-        tmp = net2.add_two(X="X", Y="Y")
-        self.assertTrue(isinstance(tmp, core.Variable))
-        net2.complete_add_op()
-        self.assertEqual(
-            '''Op(plain_net), inputs:(X, Y), outputs:(add_two@OUT@2).
-    Op(add_two), inputs:(X, Y), outputs:(add_two@OUT@2).
-''', str(net2))
-
-
-if __name__ == '__main__':
-    unittest.main()

python/paddle/v2/framework/tests/test_op_creation_methods.py → python/paddle/v2/framework/tests/test_operator.py

 import unittest
-import paddle.v2.framework.create_op_creation_methods as creation
+import paddle.v2.framework.op as op
 import paddle.v2.framework.core as core
 import paddle.v2.framework.proto.op_proto_pb2 as op_proto_pb2
 import paddle.v2.framework.proto.op_desc_pb2 as op_desc_pb2
@@ -8,7 +8,7 @@ import paddle.v2.framework.proto.attribute_pb2 as attribute_pb2
 
 class TestGetAllProtos(unittest.TestCase):
     def test_all(self):
-        all_protos = creation.get_all_op_protos()
+        all_protos = op.get_all_op_protos()
         self.assertNotEqual(0, len(all_protos))
 
         for each in all_protos:
@@ -17,25 +17,25 @@ class TestGetAllProtos(unittest.TestCase):
 
 class TestOpDescCreationMethod(unittest.TestCase):
     def test_plain_input_output(self):
-        op = op_proto_pb2.OpProto()
-        op.type = "test"
-        ipt = op.inputs.add()
+        op_proto = op_proto_pb2.OpProto()
+        op_proto.type = "test"
+        ipt = op_proto.inputs.add()
         ipt.name = "X"
         ipt.comment = "not matter"
 
-        ipt = op.inputs.add()
+        ipt = op_proto.inputs.add()
         ipt.name = "Y"
         ipt.comment = "not matter"
 
-        opt = op.outputs.add()
+        opt = op_proto.outputs.add()
         opt.name = "Z"
         opt.comment = "not matter"
 
-        op.comment = "not matter"
+        op_proto.comment = "not matter"
 
-        self.assertTrue(op.IsInitialized())
+        self.assertTrue(op_proto.IsInitialized())
 
-        method = creation.OpDescCreationMethod(op)
+        method = op.OpDescCreationMethod(op_proto)
         output = method(X="a", Y="b", Z="c")
 
         expected = op_desc_pb2.OpDesc()
@@ -45,29 +45,29 @@ class TestOpDescCreationMethod(unittest.TestCase):
         self.assertEqual(expected, output)
 
     def test_multiple_input_plain_output(self):
-        op = op_proto_pb2.OpProto()
-        op.type = "fc"
-        ipt = op.inputs.add()
+        op_proto = op_proto_pb2.OpProto()
+        op_proto.type = "fc"
+        ipt = op_proto.inputs.add()
         ipt.name = "X"
         ipt.comment = ""
         ipt.multiple = True
 
-        ipt = op.inputs.add()
+        ipt = op_proto.inputs.add()
         ipt.name = "W"
         ipt.comment = ""
         ipt.multiple = True
 
-        ipt = op.inputs.add()
+        ipt = op_proto.inputs.add()
         ipt.name = "b"
         ipt.comment = ""
 
-        out = op.outputs.add()
+        out = op_proto.outputs.add()
         out.name = "Y"
         out.comment = ""
 
-        op.comment = ""
-        self.assertTrue(op.IsInitialized())
-        method = creation.OpDescCreationMethod(op)
+        op_proto.comment = ""
+        self.assertTrue(op_proto.IsInitialized())
+        method = op.OpDescCreationMethod(op_proto)
 
         generated1 = method(X="x", W="w", b="b", Y="y")
         expected1 = op_desc_pb2.OpDesc()
@@ -93,14 +93,14 @@ class TestOpDescCreationMethod(unittest.TestCase):
         self.assertEqual(expected2, generated2)
 
     def test_attrs(self):
-        op = op_proto_pb2.OpProto()
-        op.type = "test"
-        ipt = op.inputs.add()
+        op_proto = op_proto_pb2.OpProto()
+        op_proto.type = "test"
+        ipt = op_proto.inputs.add()
         ipt.name = 'X'
         ipt.comment = ""
 
         def __add_attr__(name, type):
-            attr = op.attrs.add()
+            attr = op_proto.attrs.add()
             attr.name = name
             attr.comment = ""
             attr.type = type
@@ -112,10 +112,10 @@ class TestOpDescCreationMethod(unittest.TestCase):
         __add_attr__("floats_attr", attribute_pb2.FLOATS)
         __add_attr__("strings_attr", attribute_pb2.STRINGS)
 
-        op.comment = ""
-        self.assertTrue(op.IsInitialized())
+        op_proto.comment = ""
+        self.assertTrue(op_proto.IsInitialized())
 
-        method = creation.OpDescCreationMethod(op)
+        method = op.OpDescCreationMethod(op_proto)
 
         generated = method(
             X="a",
@@ -162,23 +162,23 @@ class TestOpDescCreationMethod(unittest.TestCase):
         self.assertEqual(expected, generated)
 
     def test_input_temporary_output(self):
-        op = op_proto_pb2.OpProto()
-        op.type = "test"
-        out = op.outputs.add()
+        op_proto = op_proto_pb2.OpProto()
+        op_proto.type = "test"
+        out = op_proto.outputs.add()
         out.name = "OUT"
         out.comment = ""
 
-        out = op.outputs.add()
+        out = op_proto.outputs.add()
         out.name = "TMP"
         out.comment = ""
         out.temporary = True
 
-        out = op.outputs.add()
+        out = op_proto.outputs.add()
         out.name = "OUT2"
         out.comment = ""
-        op.comment = ""
+        op_proto.comment = ""
 
-        method = creation.OpDescCreationMethod(op)
+        method = op.OpDescCreationMethod(op_proto)
         generated = method(OUT="a", OUT2="b")
         desc = op_desc_pb2.OpDesc()
         desc.outputs.extend(["a", core.var_names.temp(), "b"])
@@ -190,60 +190,9 @@ class TestOpDescCreationMethod(unittest.TestCase):
         self.assertEqual(generated, desc)
 
 
-class TestOpCreationDocStr(unittest.TestCase):
-    def test_all(self):
-        op = op_proto_pb2.OpProto()
-        op.type = "test"
-        op.comment = """Test Op.
-
-This op is used for unit test, not a real op.
-"""
-        a = op.inputs.add()
-        a.name = "a"
-        a.comment = "Input a for test op"
-        a.multiple = True
-
-        b = op.inputs.add()
-        b.name = "b"
-        b.comment = "Input b for test op"
-        self.assertTrue(op.IsInitialized())
-
-        o1 = op.outputs.add()
-        o1.name = "output"
-        o1.comment = "The output of test op"
-
-        o2 = op.outputs.add()
-        o2.name = "temp output"
-        o2.comment = "The temporary output of test op"
-        o2.temporary = True
-
-        test_str = op.attrs.add()
-        test_str.name = "str_attr"
-        test_str.type = attribute_pb2.STRING
-        test_str.comment = "A string attribute for test op"
-
-        actual = creation.get_docstring_from_op_proto(op)
-        expected_docstring = '''Test Op.
-
-This op is used for unit test, not a real op.
-
-:param a: Input a for test op
-:type a: list | basestr
-:param b: Input b for test op
-:type b: basestr
-:param output: The output of test op
-:type output: basestr
-:param temp output: This is a temporary variable. It does not have to set by user. The temporary output of test op
-:type temp output: basestr
-:param str_attr: A string attribute for test op
-:type str_attr: basestr
-'''
-        self.assertEqual(expected_docstring, actual)
-
-
 class TestOpCreations(unittest.TestCase):
     def test_all(self):
-        add_op = creation.op_creations.add_two(X="a", Y="b", Out="z")
+        add_op = op.Operator("add_two", X="a", Y="b", Out="z")
         self.assertIsNotNone(add_op)
         # Invoke C++ DebugString()
         self.assertEqual('Op(add_two), inputs:(a, b), outputs:(z).',

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

@@ -2,7 +2,7 @@ import unittest
 
 import numpy as np
 import paddle.v2.framework.core as core
-import paddle.v2.framework.create_op_creation_methods as creation
+from paddle.v2.framework.op import Operator
 
 from op_test_util import OpTestMeta
 
@@ -27,7 +27,7 @@ class TestSoftmaxOp(unittest.TestCase):
 
 class TestSoftmaxGradOp(unittest.TestCase):
     def test_softmax_grad(self):
-        op = creation.op_creations.softmax(X="X", Y="Y")
+        op = Operator('softmax', X="X", Y="Y")
         backward_op = core.Operator.backward(op, set())
         self.assertEqual(backward_op.type(), "softmax_grad")
         expected = '''Op(softmax_grad), inputs:(X, Y, Y@GRAD), outputs:(X@GRAD).'''

python/setup.py.in

-from setuptools import setup
+from setuptools import setup, Distribution
+
+class BinaryDistribution(Distribution):
+    def has_ext_modules(foo):
+        return True
 
 packages=['paddle',
           'paddle.proto',
@@ -11,7 +15,8 @@ packages=['paddle',
           'paddle.v2.master',
           'paddle.v2.plot',
           'paddle.v2.framework',
-          'paddle.v2.framework.proto']
+          'paddle.v2.framework.proto',
+          'py_paddle']
 
 setup_requires=["requests",
                 "numpy>=1.12",
@@ -21,23 +26,33 @@ setup_requires=["requests",
                 "rarfile",
                 "scipy>=0.19.0",
                 "Pillow",
-                "nltk"]
+                "nltk>=3.2.2"]
 
 if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']:
     setup_requires+=["opencv-python"]
 
-setup(name='paddle',
+setup(name='paddlepaddle',
       version='${PADDLE_VERSION}',
       description='Parallel Distributed Deep Learning',
       install_requires=setup_requires,
       packages=packages,
-      package_data={'paddle.v2.master': ['libpaddle_master.so'],
-            'paddle.v2.framework': ['core.so']
+      package_data={
+        'paddle.v2.master': ['libpaddle_master.so'],
+        'paddle.v2.framework': ['core.so'],
+        'py_paddle':['*.py','_swig_paddle.so']
       },
       package_dir={
           '': '${CMAKE_CURRENT_SOURCE_DIR}',
           # The paddle.v2.framework.proto will be generated while compiling.
           # So that package points to other directory.
-          'paddle.v2.framework.proto': '${PROJ_BINARY_ROOT}/paddle/framework'
+          'paddle.v2.framework.proto': '${PROJ_BINARY_ROOT}/paddle/framework',
+          'py_paddle': '${PROJ_ROOT}/paddle/py_paddle'
       },
+      scripts=['${PROJ_BINARY_ROOT}/paddle/scripts/paddle'],
+      distclass=BinaryDistribution,
+      data_files=[('/usr/local/opt/paddle/bin',
+                       ['${PROJ_BINARY_ROOT}/paddle/scripts/paddle_usage',
+                        '${PROJ_BINARY_ROOT}/paddle/trainer/paddle_trainer',
+                        '${PROJ_BINARY_ROOT}/paddle/trainer/paddle_merge_model',
+                        '${PROJ_BINARY_ROOT}/paddle/pserver/paddle_pserver_main'])]
 )