Fix paddle enforce special cases

.pre-commit-config.yaml

@@ -21,10 +21,10 @@
     sha: 28c0ea8a67a3e2dbbf4822ef44e85b63a0080a29
     hooks:
     -   id: clang-formater
--   repo: https://github.com/dnephin/pre-commit-golang
-    sha: e4693a4c282b4fc878eda172a929f7a6508e7d16
+-   repo: https://github.com/PaddlePaddle/pre-commit-golang
+    sha: 16398aeccf263adaf53b2495eed0406347d76281
     hooks:
       -   id: go-fmt
-          files: (.*\.go)
-      -   id: go-lint
-          files: (.*\.go)
+          types: [go]
+      -   id: gometalinter
+          types: [go]

.travis.yml

@@ -41,6 +41,8 @@ before_install:
   - pip install rarfile
   - curl https://glide.sh/get | bash
   - eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
+  - go get -u github.com/alecthomas/gometalinter
+  - gometalinter --install
   - |
     function timeout() { perl -e 'alarm shift; exec @ARGV' "$@"; }
 script:

CMakeLists.txt

@@ -13,7 +13,6 @@
 # limitations under the License
 
 cmake_minimum_required(VERSION 3.0)
-
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 set(PROJ_ROOT ${CMAKE_CURRENT_SOURCE_DIR})
 set(PROJ_BINARY_ROOT ${CMAKE_CURRENT_BINARY_DIR})
@@ -137,7 +136,8 @@ if(WITH_GPU)
 endif(WITH_GPU)
 
 if(USE_NNPACK)
-  list(APPEND EXTERNAL_LIBS ${NNPACK_LIB} ${PTHREADPOOL_LIB} "rt")
+    include(external/nnpack)
+    list(APPEND EXTERNAL_LIBS ${NNPACK_LIBS})
 endif(USE_NNPACK)
 
 add_subdirectory(proto)

Dockerfile

@@ -25,7 +25,7 @@ COPY ./paddle/scripts/docker/root/ /root/
 RUN apt-get update && \
     apt-get install -y \
     git python-pip python-dev openssh-server bison  \
-    wget unzip tar xz-utils bzip2 gzip coreutils ntp \
+    wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
     curl sed grep graphviz libjpeg-dev zlib1g-dev  \
     python-numpy python-matplotlib gcc g++ \
     automake locales clang-format-3.8 swig doxygen cmake  \

Dockerfile.android

@@ -14,6 +14,17 @@ RUN apt-get update && \
     wget curl tar unzip gcc g++ locales clang-format-3.8 swig cmake && \
     apt-get clean -y
 
+# 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 && \
+    mkdir /root/gopath && \
+    mkdir /root/gopath/bin && \
+    mkdir /root/gopath/src && \
+    rm go.tgz
+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
+
 # git credential to skip password typing
 RUN git config --global credential.helper store
 

cmake/configure.cmake

@@ -102,12 +102,19 @@ if(WITH_GOLANG)
       message(FATAL_ERROR "no glide executeble found: $ENV{GOPATH}/bin/glide")
     endif()
 
-    add_custom_target(go_vendor)
-    add_custom_command(TARGET go_vendor
+    # this command will only run when the file it depends is missing
+    # or has changed, or the output is missing.
+    add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/glide
       COMMAND env GOPATH=${GOPATH} ${GLIDE} install
+      COMMAND touch ${CMAKE_BINARY_DIR}/glide
+      DEPENDS ${PROJ_ROOT}/go/glide.lock
       WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go"
-    )
-    add_dependencies(go_vendor go_path)
+      )
+
+    # depends on the custom command which outputs
+    # ${CMAKE_BINARY_DIR}/glide, the custom command does not need to
+    # run every time this target is built.
+    add_custom_target(go_vendor DEPENDS ${CMAKE_BINARY_DIR}/glide go_path)
   endif()
 
 endif(WITH_GOLANG)

cmake/cpplint.cmake

@@ -27,7 +27,8 @@ set(IGNORE_PATTERN
     .*cblas\\.h.*
     .*\\.pb\\.txt
     .*LtrDataProvider.*
-    .*MultiDataProvider.*)
+    .*MultiDataProvider.*
+    .*pb.*)
 
 # add_style_check_target
 #
@@ -52,14 +53,13 @@ macro(add_style_check_target TARGET_NAME)
                 endif()
             endforeach()
             if(LINT MATCHES ON)
+                # cpplint code style
                 get_filename_component(base_filename ${filename} NAME)
                 set(CUR_GEN ${CMAKE_CURRENT_BINARY_DIR}/${base_filename}.cpplint)
-                add_custom_command(OUTPUT ${CUR_GEN}
-                    PRE_BUILD
-                    COMMAND env ${py_env} "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py"
-                                "--filter=${STYLE_FILTER}"
-                                "--write-success=${CUR_GEN}" ${filename}
-                    DEPENDS ${filename}
+                add_custom_command(TARGET ${TARGET_NAME} PRE_BUILD
+                    COMMAND "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py"
+                            "--filter=${STYLE_FILTER}"
+                            "--write-success=${CUR_GEN}" ${filename}
                     WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
             endif()
         endforeach()

cmake/cross_compiling/android.cmake

@@ -108,6 +108,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
         ENDIF()
         IF(ANDROID_ABI STREQUAL "arm64-v8a")
             SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android)
+            SET(CMAKE_SYSTEM_PROCESSOR aarch64)
         ENDIF()
         SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-")
     ENDIF()
@@ -166,7 +167,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
     ENDIF()
 
     IF(ANDROID_ABI STREQUAL "arm64-v8a")
-      LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a)
+        LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a)
     ENDIF()
 
     STRING(REPLACE ";" " " ANDROID_COMPILER_FLAGS "${ANDROID_COMPILER_FLAGS}")
@@ -193,6 +194,10 @@ ELSE()
         SET(CMAKE_ANDROID_STANDALONE_TOOLCHAIN ${ANDROID_STANDALONE_TOOLCHAIN})
     ENDIF()
     SET(CMAKE_ANDROID_ARCH_ABI ${ANDROID_ABI})
-    SET(CMAKE_ANDROID_ARM_MODE ${ANDROID_ARM_MODE})
-    SET(CMAKE_ANDROID_ARM_NEON ${ANDROID_ARM_NEON})
+    IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
+        SET(CMAKE_ANDROID_ARM_MODE ${ANDROID_ARM_MODE})
+        IF(ANDROID_ABI STREQUAL "armeabi-v7a")
+            SET(CMAKE_ANDROID_ARM_NEON ${ANDROID_ARM_NEON})
+        ENDIF()
+    ENDIF()
 ENDIF()

paddle/function/nnpack/nnpack.cmake → cmake/external/nnpack.cmake

@@ -7,10 +7,24 @@ set(NNPACK_ROOT $ENV{NNPACK_ROOT} CACHE PATH "Folder contains NNPACK")
 find_path(NNPACK_INC_DIR nnpack.h PATHS ${NNPACK_ROOT}/include)
 find_library(NNPACK_LIB NAMES nnpack PATHS ${NNPACK_ROOT}/lib)
 find_library(PTHREADPOOL_LIB NAMES pthreadpool PATHS ${NNPACK_ROOT}/lib)
+find_library(NNPACK_UKERNELS_LIB NAMES nnpack_ukernels PATHS ${NNPACK_ROOT}/lib)
+find_library(NNPACK_CPUFEATURES_LIB NAMES cpufeatures PATHS ${NNPACK_ROOT}/lib)
 
 if(NNPACK_INC_DIR AND NNPACK_LIB AND PTHREADPOOL_LIB)
   set(NNPACK_FOUND ON)
   INCLUDE_DIRECTORIES(${NNPACK_INC_DIR})
+
+  set(NNPACK_LIBS)
+  list(APPEND NNPACK_LIBS ${NNPACK_LIB} ${PTHREADPOOL_LIB})
+  if (NNPACK_UKERNELS_LIB)
+    list(APPEND NNPACK_LIBS ${NNPACK_UKERNELS_LIB})
+  endif()
+  if (NNPACK_CPUFEATURES_LIB)
+    list(APPEND NNPACK_LIBS ${NNPACK_CPUFEATURES_LIB})
+  endif()
+  if(NOT ANDROID)
+    list(APPEND NNPACK_LIBS "rt")
+  endif()
 else()
   message(FATAL_ERROR "Cannot find NNPACK in (${NNPACK_ROOT})")
 endif()

cmake/generic.cmake

@@ -185,6 +185,10 @@ function(cc_library TARGET_NAME)
       add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
       target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
     endif()
+    
+    # cpplint code style
+    add_style_check_target(${TARGET_NAME} ${cc_library_SRCS})
+
   else(cc_library_SRCS)
     if (cc_library_DEPS)
       merge_static_libs(${TARGET_NAME} ${cc_library_DEPS})
@@ -286,8 +290,22 @@ function(go_library TARGET_NAME)
     set(${TARGET_NAME}_LIB_NAME "${CMAKE_STATIC_LIBRARY_PREFIX}${TARGET_NAME}${CMAKE_STATIC_LIBRARY_SUFFIX}" CACHE STRING "output library name for target ${TARGET_NAME}")
   endif()
 
-  # Add dummy code to support `make target_name` under Terminal Command
   set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}_dummy.c)
+
+  # This custom command will always run since it depends on a not
+  # existing file.
+  add_custom_command(
+    OUTPUT dummy_rebulid_${TARGET_NAME}
+    COMMAND cmake -E touch ${dummyfile}
+    )
+  # Create a custom target that depends on the custom command output
+  # file, so the custom command can be referenced as a dependency by
+  # `add_dependencies`.
+  add_custom_target(rebuild_${TARGET_NAME}
+    DEPENDS dummy_rebulid_${TARGET_NAME}
+    )
+
+  # Add dummy code to support `make target_name` under Terminal Command
   file(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
   if (go_library_SHARED OR go_library_shared)
     add_library(${TARGET_NAME} SHARED ${dummyfile})
@@ -298,6 +316,12 @@ function(go_library TARGET_NAME)
     add_dependencies(${TARGET_NAME} ${go_library_DEPS})
   endif(go_library_DEPS)
 
+  # The "source file" of the library is `${dummyfile}` which never
+  # change, so the target will never rebuild. Make the target depends
+  # on the custom command that touches the library "source file", so
+  # rebuild will always happen.
+  add_dependencies(${TARGET_NAME} rebuild_${TARGET_NAME})
+
   set(${TARGET_NAME}_LIB_PATH "${CMAKE_CURRENT_BINARY_DIR}/${${TARGET_NAME}_LIB_NAME}" CACHE STRING "output library path for target ${TARGET_NAME}")
 
   file(GLOB GO_SOURCE RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.go")
@@ -338,7 +362,7 @@ function(go_test TARGET_NAME)
   string(REPLACE "${PADDLE_GO_PATH}" "" CMAKE_CURRENT_SOURCE_REL_DIR ${CMAKE_CURRENT_SOURCE_DIR})
   add_custom_target(${TARGET_NAME} ALL DEPENDS go_vendor ${go_test_DEPS})
   add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
-    COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} test
+    COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} test -race
     -c -o "${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}"
     ".${CMAKE_CURRENT_SOURCE_REL_DIR}"
     WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go")

go/cmd/master/master.go

@@ -11,6 +11,7 @@ import (
 
 	"github.com/namsral/flag"
 	log "github.com/sirupsen/logrus"
+	"github.com/topicai/candy"
 
 	"github.com/PaddlePaddle/Paddle/go/master"
 	"github.com/PaddlePaddle/Paddle/go/utils/networkhelper"
@@ -20,11 +21,18 @@ func main() {
 	port := flag.Int("port", 8080, "port of the master server.")
 	ttlSec := flag.Int("ttl", 60, "etcd lease TTL in seconds.")
 	endpoints := flag.String("endpoints", "http://127.0.0.1:2379", "comma separated etcd endpoints. If empty, fault tolerance will not be enabled.")
-	taskTimeoutDur := flag.Duration("task_timout_dur", 20*time.Minute, "task timout duration.")
-	taskTimeoutMax := flag.Int("task_timeout_max", 3, "max timtout count for each task before it being declared failed task.")
-	chunkPerTask := flag.Int("chunk_per_task", 10, "chunk per task.")
+	taskTimeoutDur := flag.Duration("task-timout-dur", 20*time.Minute, "task timout duration.")
+	taskTimeoutMax := flag.Int("task-timeout-max", 3, "max timtout count for each task before it being declared failed task.")
+	chunkPerTask := flag.Int("chunk-per-task", 10, "chunk per task.")
+	logLevel := flag.String("log-level", "info",
+		"log level, possible values: debug, info, warning, error, fatal, panic")
 	flag.Parse()
 
+	level, e := log.ParseLevel(*logLevel)
+	candy.Must(e)
+
+	log.SetLevel(level)
+
 	if *endpoints == "" {
 		log.Warningln("-endpoints not set, fault tolerance not be enabled.")
 	}

go/cmd/pserver/pserver.go

@@ -40,7 +40,7 @@ func main() {
 		idx = *index
 	} else {
 		e = pserver.NewEtcdClient(*etcdEndpoint, *numPservers, *etcdTimeout)
-		idx, err = e.Register()
+		idx, err = e.Register(*port)
 		candy.Must(err)
 
 		cp, err = pserver.NewCheckpointFromFile(*checkpointPath, idx, e)

go/master/c/client.go

@@ -23,7 +23,6 @@ import (
 	log "github.com/sirupsen/logrus"
 )
 
-var nullPtr = unsafe.Pointer(uintptr(0))
 var mu sync.Mutex
 var handleMap = make(map[C.paddle_master_client]*master.Client)
 var curHandle C.paddle_master_client
@@ -114,13 +113,13 @@ func paddle_next_record(client C.paddle_master_client, record **C.uchar) C.int {
 	if err != nil {
 		// Error
 		// TODO: return the type of error?
-		*record = (*C.uchar)(nullPtr)
+		*record = (*C.uchar)(nil)
 		return -1
 	}
 
 	if len(r) == 0 {
 		// Empty record
-		*record = (*C.uchar)(nullPtr)
+		*record = (*C.uchar)(nil)
 		return 0
 	}
 

go/master/client.go

@@ -2,6 +2,7 @@ package master
 
 import (
 	"os"
+	"time"
 
 	"github.com/PaddlePaddle/Paddle/go/connection"
 	"github.com/PaddlePaddle/recordio"
@@ -36,9 +37,9 @@ func (c *Client) getRecords() {
 	for {
 		t, err := c.getTask()
 		if err != nil {
-			// TODO(helin): wait before move on with next
 			// getTask call.
-			log.Errorln(err)
+			log.Errorf("Get task failed, sleep 3 seconds and continue, %s", err)
+			time.Sleep(3 * time.Second)
 			continue
 		}
 
@@ -68,7 +69,10 @@ func (c *Client) getRecords() {
 		// We treat a task as finished whenever the last data
 		// instance of the task is read. This is not exactly
 		// correct, but a reasonable approximation.
-		c.taskFinished(t.Meta.ID)
+		err = c.taskFinished(t.Meta.ID)
+		if err != nil {
+			log.Errorln(err)
+		}
 	}
 }
 

go/master/client_internal_test.go

@@ -66,11 +66,21 @@ func TestGetFinishTask(t *testing.T) {
 
 	for i := 0; i < totalTask*chunkPerTask; i++ {
 		w := recordio.NewWriter(f, -1, -1)
-		w.Write(nil)
+		_, err = w.Write(nil)
+		if err != nil {
+			panic(err)
+		}
+
 		// call Close to force RecordIO writing a chunk.
-		w.Close()
+		err = w.Close()
+		if err != nil {
+			panic(err)
+		}
+	}
+	err = f.Close()
+	if err != nil {
+		panic(err)
 	}
-	f.Close()
 
 	// Manually intialize client to avoid calling c.getRecords()
 	c := &Client{}
@@ -79,7 +89,11 @@ func TestGetFinishTask(t *testing.T) {
 	ch := make(chan string, 1)
 	ch <- addr
 	go c.monitorMaster(ch)
-	c.SetDataset([]string{path})
+	err = c.SetDataset([]string{path})
+	if err != nil {
+		panic(err)
+	}
+
 	checkOnePass := func(i int) {
 		var tasks []Task
 		for idx := 0; idx < totalTask; idx++ {

go/master/client_test.go

@@ -57,14 +57,30 @@ func TestNextRecord(t *testing.T) {
 
 	w := recordio.NewWriter(f, -1, -1)
 	for i := 0; i < total; i++ {
-		w.Write([]byte{byte(i)})
+		_, err = w.Write([]byte{byte(i)})
+		if err != nil {
+			panic(err)
+		}
+	}
+
+	err = w.Close()
+	if err != nil {
+		panic(err)
+	}
+
+	err = f.Close()
+	if err != nil {
+		panic(err)
 	}
-	w.Close()
-	f.Close()
+
 	curAddr := make(chan string, 1)
 	curAddr <- fmt.Sprintf(":%d", p)
 	c := master.NewClient(curAddr, 10)
-	c.SetDataset([]string{path})
+	err = c.SetDataset([]string{path})
+	if err != nil {
+		panic(err)
+	}
+
 	for pass := 0; pass < 50; pass++ {
 		received := make(map[byte]bool)
 		for i := 0; i < total; i++ {

go/master/etcd_client.go

@@ -30,7 +30,7 @@ type EtcdClient struct {
 // NewEtcdClient creates a new EtcdClient.
 func NewEtcdClient(endpoints []string, addr string, lockPath, addrPath, statePath string, ttlSec int) (*EtcdClient, error) {
 	log.Debugf("Connecting to etcd at %v", endpoints)
-	// TODO(helin): gracefully shutdown etcd store. Becuase etcd
+	// TODO(helin): gracefully shutdown etcd store. Because etcd
 	// store holds a etcd lock, even though the lock will expire
 	// when the lease timeout, we need to implement graceful
 	// shutdown to release the lock.
@@ -60,7 +60,7 @@ func NewEtcdClient(endpoints []string, addr string, lockPath, addrPath, statePat
 	}
 	log.Debugf("Successfully acquired lock at %s.", lockPath)
 
-	put := clientv3.OpPut(addrPath, string(addr))
+	put := clientv3.OpPut(addrPath, addr)
 	resp, err := cli.Txn(context.Background()).If(lock.IsOwner()).Then(put).Commit()
 	if err != nil {
 		return nil, err

go/master/inmem_store.go

@@ -4,7 +4,7 @@ import "sync"
 
 // InMemStore is an in memory implementation of Store interface.
 //
-// It does not tolerate the fault that casues the program to crash.
+// It does not tolerate the fault that causes the program to crash.
 type InMemStore struct {
 	mu  sync.Mutex
 	buf []byte

go/master/service.go

@@ -160,7 +160,7 @@ func (s *Service) recover() (bool, error) {
 
 // snapshot *must* be called with s.mu being held.
 func (s *Service) snapshot() error {
-	// TOOD(helin): etcd request has a size limit, so the snapshot
+	// TODO(helin): etcd request has a size limit, so the snapshot
 	// size is limited by the max request size. We should either
 	// divide the snapshot into smaller chunks and save under
 	// different keys, or configure the request size to be big
@@ -215,6 +215,7 @@ func readChunks(globPaths []string) ([]Chunk, error) {
 		}
 
 		count := index.NumChunks()
+		log.Infof("readChunks: file %s has %d chunks", path, count)
 		for i := 0; i < count; i++ {
 			chunk := Chunk{
 				Path:  path,
@@ -288,7 +289,6 @@ func (s *Service) processFailedTask(t taskEntry, epoch int) {
 
 	log.Warningf("Task %v failed %d times, discard.", t.Task, t.NumFailure)
 	s.taskQueues.Todo = append(s.taskQueues.Todo, t)
-	return
 }
 
 func (s *Service) checkTimeoutFunc(taskID int, epoch int) func() {

go/pserver/client/c/cclient.go

@@ -34,7 +34,6 @@ import (
 	log "github.com/sirupsen/logrus"
 )
 
-var nullPtr = unsafe.Pointer(uintptr(0))
 var mu sync.Mutex
 var handleMap = make(map[C.paddle_pserver_client]*client.Client)
 var curHandle C.paddle_pserver_client
@@ -63,7 +62,7 @@ func remove(client C.paddle_pserver_client) *client.Client {
 }
 
 func cArrayToSlice(p unsafe.Pointer, len int) []byte {
-	if p == nullPtr {
+	if p == nil {
 		return nil
 	}
 
@@ -101,11 +100,11 @@ func paddle_new_pserver_client(addrs *C.char, selected int) C.paddle_pserver_cli
 }
 
 //export paddle_new_etcd_pserver_client
-func paddle_new_etcd_pserver_client(etcd_endpoints *C.char, selected int) C.paddle_pserver_client {
+func paddle_new_etcd_pserver_client(etcdEndpoints *C.char, selected int) C.paddle_pserver_client {
 	// TODO(Longfei: use etcd lock to decide which trainer to initialize the parameters)
-	addr := C.GoString(etcd_endpoints)
-	etcd_client := client.NewEtcd(addr)
-	c := client.NewClient(etcd_client, etcd_client.Desired(), selector(selected != 0))
+	addr := C.GoString(etcdEndpoints)
+	etcdClient := client.NewEtcd(addr)
+	c := client.NewClient(etcdClient, etcdClient.Desired(), selector(selected != 0))
 	return add(c)
 }
 
@@ -124,20 +123,20 @@ func paddle_begin_init_params(client C.paddle_pserver_client) C.int {
 }
 
 //export paddle_init_param
-func paddle_init_param(client C.paddle_pserver_client, param C.paddle_parameter, param_config unsafe.Pointer, config_len C.int) C.int {
+func paddle_init_param(client C.paddle_pserver_client, param C.paddle_parameter, paramConfig unsafe.Pointer, configLen C.int) C.int {
 	et := pserver.ElementType(param.element_type)
 	name := C.GoString(param.name)
 	content := cArrayToSlice(unsafe.Pointer(param.content), int(param.content_len))
 	pc := pserver.ParameterWithConfig{
 		Param:  pserver.Parameter{Name: name, ElementType: et, Content: content},
-		Config: cArrayToSlice(param_config, int(config_len)),
+		Config: cArrayToSlice(paramConfig, int(configLen)),
 	}
 	c := get(client)
 	err := c.InitParam(pc)
 
 	if err != nil {
 		if err.Error() == pserver.AlreadyInitialized {
-			log.Warningf("parameter %s already initialized, treat paddle_init_param as sucessful.", name)
+			log.Warningf("parameter %s already initialized, treat paddle_init_param as successful.", name)
 			return C.PSERVER_OK
 		}
 		log.Errorln(err)
@@ -153,7 +152,7 @@ func paddle_finish_init_params(client C.paddle_pserver_client) C.int {
 	err := c.FinishInitParams()
 	if err != nil {
 		if err.Error() == pserver.AlreadyInitialized {
-			log.Warningln("parameters already initialized, treat paddle_finish_init_params as sucessful.")
+			log.Warningln("parameters already initialized, treat paddle_finish_init_params as successful.")
 			return C.PSERVER_OK
 		}
 
@@ -223,12 +222,12 @@ func paddle_get_params(client C.paddle_pserver_client, dst **C.paddle_parameter,
 		p := ps[i]
 		param := *(**C.paddle_parameter)(unsafe.Pointer((uintptr(unsafe.Pointer(dst)) + uintptr(i)*unsafe.Sizeof(*dst))))
 
-		if unsafe.Pointer(param) == nullPtr {
+		if unsafe.Pointer(param) == nil {
 			log.Errorln("must pre-allocate parameter.")
 			return C.PSERVER_ERROR
 		}
 
-		if unsafe.Pointer(param.content) != nullPtr {
+		if unsafe.Pointer(param.content) != nil {
 			if int(param.content_len) != len(p.Content) {
 				log.Errorf("the pre-allocated content len does not match parameter content len. Pre-allocated len: %d, returned len: %d", param.content_len, len(p.Content))
 				return C.PSERVER_ERROR

go/pserver/client/c/test/test_train.py

 import paddle.v2 as paddle
 import paddle.v2.dataset.uci_housing as uci_housing
+import paddle.v2.master as master
+import os
+import cPickle as pickle
+
+etcd_ip = os.getenv("MASTER_IP", "127.0.0.1")
+etcd_endpoint = "http://" + etcd_ip + ":2379"
+
+
+def cloud_reader():
+    print "connecting to master, etcd endpoints: ", etcd_endpoint
+    master_client = master.client(etcd_endpoint, 5, 64)
+    master_client.set_dataset(
+        ["/pfs/dlnel/public/dataset/uci_housing/uci_housing-*-of-*"])
+    while 1:
+        r, e = master_client.next_record()
+        if not r:
+            break
+        yield pickle.loads(r)
 
 
 def main():
@@ -22,13 +40,13 @@ def main():
     # create optimizer of new remote updater to pserver
     optimizer = paddle.optimizer.Momentum(momentum=0)
 
-    #TODO(zhihong) : replace optimizer with new OptimizerConfig
-
+    print "etcd endoint: ", etcd_endpoint
     trainer = paddle.trainer.SGD(cost=cost,
                                  parameters=parameters,
                                  update_equation=optimizer,
                                  is_local=False,
-                                 pserver_spec="localhost:3000")
+                                 pserver_spec=etcd_endpoint,
+                                 use_etcd=True)
 
     # event_handler to print training and testing info
     def event_handler(event):
@@ -47,11 +65,11 @@ def main():
                 print "Test %d, %.2f" % (event.pass_id, result.cost)
 
     # training
+    # NOTE: use uci_housing.train() as reader for non-paddlecloud training
     trainer.train(
         reader=paddle.batch(
             paddle.reader.shuffle(
-                uci_housing.train(), buf_size=500),
-            batch_size=2),
+                cloud_reader, buf_size=500), batch_size=2),
         feeding={'x': 0,
                  'y': 1},
         event_handler=event_handler,

go/pserver/client/client.go

@@ -233,7 +233,7 @@ func (c *Client) Save(path string) error {
 
 func strHash(s string) uint32 {
 	h := fnv.New32a()
-	h.Write([]byte(s))
+	_, _ = h.Write([]byte(s))
 	return h.Sum32()
 }
 

go/pserver/client/client_test.go

@@ -79,15 +79,33 @@ func initEtcdClient() {
 		log.Errorf("err %v", err)
 	}
 	ctx, cancel := context.WithTimeout(context.Background(), timeout)
-	client.Delete(ctx, pserver.PsDesired)
-	client.Delete(ctx, pserver.PsPath)
-	client.Put(ctx, pserver.PsDesired, strconv.Itoa(numPserver))
+	_, err = client.Delete(ctx, pserver.PsDesired)
+	if err != nil {
+		panic(err)
+	}
+
+	_, err = client.Delete(ctx, pserver.PsPath)
+	if err != nil {
+		panic(err)
+	}
+
+	_, err = client.Put(ctx, pserver.PsDesired, strconv.Itoa(numPserver))
+	if err != nil {
+		panic(err)
+	}
+
 	ports := initClient()
 	for i := 0; i < numPserver; i++ {
-		client.Put(ctx, pserver.PsPath+strconv.Itoa(i), ":"+strconv.Itoa(ports[i]))
+		_, err = client.Put(ctx, pserver.PsPath+strconv.Itoa(i), ":"+strconv.Itoa(ports[i]))
+		if err != nil {
+			panic(err)
+		}
 	}
 	cancel()
-	client.Close()
+	err = client.Close()
+	if err != nil {
+		panic(err)
+	}
 }
 
 type selector bool
@@ -164,7 +182,7 @@ func testClient(t *testing.T, c *client.Client) {
 
 		wg.Add(1)
 		go func(gs []pserver.Gradient) {
-			err = c.SendGrads(gs)
+			err := c.SendGrads(gs)
 			if err != nil {
 				t.Fatal(err)
 			}

go/pserver/client/etcd_client.go

@@ -12,7 +12,7 @@ import (
 )
 
 const (
-	DefaultEtcdTimeout time.Duration = 5 * time.Second
+	defaultEtcdTimeout time.Duration = 5 * time.Second
 )
 
 // EtcdClient is used by pserver client that is a part of trainer process.
@@ -47,7 +47,7 @@ func (p *EtcdClient) Desired() int {
 
 		psDesired, err = strconv.Atoi(string(resp.Kvs[0].Value))
 		if err != nil {
-			log.Errorf("psDesired %s invalid %v", psDesired, err)
+			log.Errorf("psDesired %d invalid %v", psDesired, err)
 			time.Sleep(p.timeout)
 			continue
 		}
@@ -106,11 +106,11 @@ func NewEtcd(endpoints string) *EtcdClient {
 	for {
 		cli, err = clientv3.New(clientv3.Config{
 			Endpoints:   ep,
-			DialTimeout: DefaultEtcdTimeout,
+			DialTimeout: defaultEtcdTimeout,
 		})
 		if err != nil {
 			log.Errorf("Init etcd connection failed: %v", err)
-			time.Sleep(DefaultEtcdTimeout)
+			time.Sleep(defaultEtcdTimeout)
 			continue
 		}
 		break
@@ -118,7 +118,7 @@ func NewEtcd(endpoints string) *EtcdClient {
 	log.Infof("Connected to etcd: %s\n", endpoints)
 	client := &EtcdClient{
 		client:    cli,
-		timeout:   DefaultEtcdTimeout,
+		timeout:   defaultEtcdTimeout,
 		endpoints: ep,
 	}
 	return client

go/pserver/etcd_client.go

@@ -49,7 +49,7 @@ func NewEtcdClient(endpoints string, numPservers int, timeout time.Duration) *Et
 // Register registers the pserver on etcd
 //
 // Register returns the index of the current pserver.
-func (e *EtcdClient) Register() (int, error) {
+func (e *EtcdClient) Register(port int) (int, error) {
 
 	var err error
 	e.externalIP, err = networkhelper.GetExternalIP()
@@ -116,7 +116,7 @@ func (e *EtcdClient) Register() (int, error) {
 	for {
 		ctx, cancel := context.WithTimeout(context.Background(), time.Second)
 		var err error
-		pserverIdx, err = e.registerPserverEtcd(ctx)
+		pserverIdx, err = e.registerPserverEtcd(ctx, port)
 		cancel()
 		if err != nil {
 			log.Warn(err)
@@ -140,7 +140,7 @@ func (e *EtcdClient) initDesiredPservers(ctx context.Context, numPservers int) (
 }
 
 // registerPserverEtcd registers pserver node on etcd using transaction.
-func (e *EtcdClient) registerPserverEtcd(ctx context.Context) (int, error) {
+func (e *EtcdClient) registerPserverEtcd(ctx context.Context, port int) (int, error) {
 	var idx int
 	_, err := concurrency.NewSTM(e.etcdClient, func(c concurrency.STM) error {
 		registered := false
@@ -156,8 +156,9 @@ func (e *EtcdClient) registerPserverEtcd(ctx context.Context) (int, error) {
 					log.Fatal(err)
 				}
 				// find the first id and write info
-				c.Put(psKey, e.externalIP, clientv3.WithLease(resp.ID))
-				log.Debugf("set pserver node %s with value %s", psKey, e.externalIP)
+				pserverAddr := e.externalIP + ":" + strconv.Itoa(port)
+				c.Put(psKey, pserverAddr, clientv3.WithLease(resp.ID))
+				log.Debugf("set pserver node %s with value %s", psKey, pserverAddr)
 				ch, kaerr := e.etcdClient.KeepAlive(context.TODO(), resp.ID)
 				if kaerr != nil {
 					log.Errorf("keepalive etcd node error: %v", kaerr)
@@ -176,10 +177,10 @@ func (e *EtcdClient) registerPserverEtcd(ctx context.Context) (int, error) {
 				break
 			}
 		}
-		if registered == true {
+		if registered {
 			return nil
 		}
-		return errors.New("not registerd, may due to already have enough pservers")
+		return errors.New("not registered, may due to already have enough pservers")
 	}, concurrency.WithAbortContext(ctx), concurrency.WithIsolation(concurrency.RepeatableReads))
 
 	if err != nil {
@@ -210,8 +211,5 @@ func (e *EtcdClient) PutKey(key string, value []byte, timeout time.Duration) err
 	ctx, cancel := context.WithTimeout(context.Background(), timeout)
 	_, err := e.etcdClient.Put(ctx, key, string(value))
 	cancel()
-	if err != nil {
-		return err
-	}
-	return nil
+	return err
 }

go/pserver/optimizer.go

@@ -14,8 +14,6 @@ import (
 	log "github.com/sirupsen/logrus"
 )
 
-var nullPtr = unsafe.Pointer(uintptr(0))
-
 type optimizer struct {
 	opt         *C.struct_paddle_optimizer
 	elementType ElementType
@@ -23,7 +21,7 @@ type optimizer struct {
 }
 
 func cArrayToSlice(p unsafe.Pointer, len int) []byte {
-	if p == nullPtr {
+	if p == nil {
 		return nil
 	}
 
@@ -92,8 +90,8 @@ func (o *optimizer) UpdateParameter(g Gradient) error {
 }
 
 func (o *optimizer) Cleanup() {
-	if unsafe.Pointer(o.opt) != nullPtr {
+	if unsafe.Pointer(o.opt) != nil {
 		C.paddle_release_optimizer(o.opt)
-		o.opt = (*C.struct_paddle_optimizer)(nullPtr)
+		o.opt = (*C.struct_paddle_optimizer)(nil)
 	}
 }

go/pserver/service.go

@@ -211,7 +211,7 @@ func (s *Service) GetParam(name string, parameter *Parameter) error {
 	// learning optimization methods are stochastic in
 	// nature. This race condition is allowed deliberately
 	// to save the program from making a copy of the
-	// paramter content.
+	// parameter content.
 	parameter.Name = name
 	parameter.ElementType = opt.elementType
 	parameter.Content = opt.GetWeights()
@@ -219,7 +219,7 @@ func (s *Service) GetParam(name string, parameter *Parameter) error {
 }
 
 // pserver save checkpoint
-func (s *Service) doCheckpoint() error {
+func (s *Service) doCheckpoint() (err error) {
 	<-s.initialized
 	s.mu.Lock()
 	defer s.mu.Unlock()
@@ -237,9 +237,9 @@ func (s *Service) doCheckpoint() error {
 	}
 	var buf bytes.Buffer
 	encoder := gob.NewEncoder(&buf)
-	err := encoder.Encode(cp)
+	err = encoder.Encode(cp)
 	if err != nil {
-		return err
+		return
 	}
 
 	cpMeta := checkpointMeta{}
@@ -248,10 +248,14 @@ func (s *Service) doCheckpoint() error {
 	h := md5.New()
 	cpMeta.MD5 = hex.EncodeToString(h.Sum(buf.Bytes()))
 
-	cpMetajson, _ := json.Marshal(cpMeta)
+	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 err
+		return
 	}
 	if _, err = os.Stat(cpMeta.UUID); os.IsNotExist(err) {
 		log.Info("checkpoint does not exists.")
@@ -264,15 +268,32 @@ func (s *Service) doCheckpoint() error {
 		}
 	}
 	f, err := os.Create(cpMeta.UUID)
-	defer f.Close()
 	if err != nil {
-		return err
+		return
 	}
+
+	defer func() {
+		closeErr := f.Close()
+		if closeErr != nil {
+			if err != nil {
+				log.Errorln(closeErr)
+			} else {
+				// Set closeErr as return value.
+				err = closeErr
+			}
+		}
+	}()
+
 	writer := bufio.NewWriter(f)
 	_, err = writer.Write(buf.Bytes())
-	writer.Flush()
 	if err != nil {
-		return err
+		return
 	}
-	return nil
+
+	err = writer.Flush()
+	if err != nil {
+		return
+	}
+
+	return
 }

paddle/api/PaddleAPI.h

@@ -843,7 +843,8 @@ public:
                                                bool useSparseUpdater);
   static ParameterUpdater* createNewRemoteUpdater(
       OptimizationConfig* config,
-      const std::string pserverSpec) throw(UnsupportError);
+      const std::string pserverSpec,
+      const bool useEtcd) throw(UnsupportError);
   ~ParameterUpdater();
 
   /**

paddle/api/ParameterUpdater.cpp

@@ -33,11 +33,12 @@ ParameterUpdater *ParameterUpdater::createLocalUpdater(
 
 ParameterUpdater *ParameterUpdater::createNewRemoteUpdater(
     OptimizationConfig *config,
-    const std::string pserverSpec) throw(UnsupportError) {
+    const std::string pserverSpec,
+    const bool useEtcd) throw(UnsupportError) {
 #ifndef PADDLE_WITHOUT_GOLANG
   auto updater = new ParameterUpdater();
   updater->m->updater.reset(new paddle::NewRemoteParameterUpdater(
-      config->m->getConfig(), pserverSpec));
+      config->m->getConfig(), pserverSpec, useEtcd));
   return updater;
 #else
   throw UnsupportError();

paddle/framework/CMakeLists.txt

 # ddim lib
-cc_library(ddim SRCS ddim.cc)
+cc_library(ddim SRCS ddim.cc DEPS eigen3)
 cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
 nv_test(dim_test SRCS dim_test.cu DEPS ddim)
-cc_test(tensor_test SRCS tensor_test.cc DEPS ddim)
+
+cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory)
+cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
+cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
+
 cc_test(variable_test SRCS variable_test.cc)
 cc_test(scope_test SRCS scope_test.cc)
-cc_test(enforce_test SRCS enforce_test.cc)
+
 proto_library(attr_type SRCS attr_type.proto)
 proto_library(op_proto SRCS op_proto.proto DEPS attr_type)
-cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf)
 proto_library(op_desc SRCS op_desc.proto DEPS attr_type)
+cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf)
 cc_test(op_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf)
-cc_library(operator SRCS operator.cc DEPS op_desc device_context)
+
+cc_library(operator SRCS operator.cc DEPS op_desc device_context tensor)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
+
 cc_library(op_registry SRCS op_registry.cc DEPS op_proto op_desc)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry operator)
+
 py_proto_compile(framework_py_proto SRCS attr_type.proto op_proto.proto op_desc.proto)
 # Generate an empty __init__.py to make framework_py_proto as a valid python module.
 add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
 add_dependencies(framework_py_proto framework_py_proto_init)
 
 proto_library(net_proto SRCS net_proto.proto DEPS op_proto)
-cc_library(net SRCS net.cc DEPS net_proto)
+cc_library(net SRCS net.cc DEPS operator net_proto op_registry)
+cc_test(net_op_test SRCS net_op_test.cc DEPS net)

paddle/framework/attr_checker.h

@@ -4,8 +4,9 @@
 #include <functional>
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 #include <vector>
-#include "paddle/framework/enforce.h"
+#include "paddle/platform/enforce.h"
 
 namespace paddle {
 namespace framework {
@@ -41,6 +42,35 @@ class DefaultValueSetter {
   T default_value_;
 };
 
+template <typename T>
+class EnumInContainer {
+ public:
+  explicit EnumInContainer(const std::unordered_set<T>& c) : container_(c) {}
+  void operator()(T& val) const {
+    PADDLE_ENFORCE(container_.find(val) != container_.end(),
+                   "Value %s is not in enum container %s", val,
+                   ContainerDebugString());
+  }
+
+ private:
+  std::string ContainerDebugString() const {
+    std::ostringstream sout;
+    sout << "[";
+    size_t cnt = 0;
+    for (auto& v : container_) {
+      sout << v;
+      ++cnt;
+      if (cnt != container_.size()) {
+        sout << " ,";
+      }
+    }
+    sout << "]";
+    return sout.str();
+  }
+
+  std::unordered_set<T> container_;
+};
+
 // check whether a certain attribute fit its limits
 // an attribute can have more than one limits
 template <typename T>
@@ -50,6 +80,11 @@ class TypedAttrChecker {
  public:
   TypedAttrChecker(const std::string& attr_name) : attr_name_(attr_name) {}
 
+  TypedAttrChecker& InEnum(const std::unordered_set<T>& range) {
+    value_checkers_.push_back(EnumInContainer<T>(range));
+    return *this;
+  }
+
   TypedAttrChecker& LargerThan(const T& lower_bound) {
     value_checkers_.push_back(LargerThanChecker<T>(lower_bound));
     return *this;

paddle/framework/ddim.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
 #include "paddle/framework/ddim.h"
+#include "paddle/platform/enforce.h"
 
 namespace paddle {
 namespace framework {
 
-///@cond HIDDEN
+/// @cond HIDDEN
 
 template <int i>
 Dim<i> make_dim(const int* d) {
@@ -50,7 +65,7 @@ void make_ddim(DDim& ddim, const int* dims, int n) {
   }
 }
 
-///@endcond
+/// @endcond
 
 DDim make_ddim(std::initializer_list<int> dims) {
   DDim result(make_dim(0));
@@ -64,11 +79,11 @@ DDim make_ddim(const std::vector<int>& dims) {
   return result;
 }
 
-///@cond HIDDEN
+/// @cond HIDDEN
 // XXX For some reason, putting this in an anonymous namespace causes errors
 class DynamicMutableIndexer : public boost::static_visitor<int&> {
  public:
-  DynamicMutableIndexer(int idx) : idx_(idx) {}
+  explicit DynamicMutableIndexer(int idx) : idx_(idx) {}
 
   template <int D>
   int& operator()(Dim<D>& dim) const {
@@ -81,7 +96,7 @@ class DynamicMutableIndexer : public boost::static_visitor<int&> {
 
 class DynamicConstIndexer : public boost::static_visitor<int> {
  public:
-  DynamicConstIndexer(int idx) : idx_(idx) {}
+  explicit DynamicConstIndexer(int idx) : idx_(idx) {}
 
   template <int D>
   int operator()(const Dim<D>& dim) const {
@@ -92,7 +107,7 @@ class DynamicConstIndexer : public boost::static_visitor<int> {
   int idx_;
 };
 
-///@endcond
+/// @endcond
 
 int& DDim::operator[](int idx) {
   return boost::apply_visitor(DynamicMutableIndexer(idx), var);
@@ -102,6 +117,8 @@ int DDim::operator[](int idx) const {
   return boost::apply_visitor(DynamicConstIndexer(idx), var);
 }
 
+ssize_t DDim::size() const { return arity(*this); }
+
 bool DDim::operator==(DDim d) const {
   if (var.which() != d.getVar().which()) {
     return false;
@@ -155,11 +172,11 @@ int get(const DDim& ddim, int idx) { return ddim[idx]; }
 
 void set(DDim& ddim, int idx, int value) { ddim[idx] = value; }
 
-///@cond HIDDEN
+/// @cond HIDDEN
 struct VectorizeVisitor : public boost::static_visitor<> {
   std::vector<int>& vector;
 
-  VectorizeVisitor(std::vector<int>& v) : vector(v) {}
+  explicit VectorizeVisitor(std::vector<int>& v) : vector(v) {}
 
   template <typename T>
   void operator()(const T& t) {
@@ -169,7 +186,7 @@ struct VectorizeVisitor : public boost::static_visitor<> {
 
   void operator()(const Dim<1>& t) { vector.push_back(t.head); }
 };
-///@endcond
+/// @endcond
 
 std::vector<int> vectorize(const DDim& ddim) {
   std::vector<int> result;
@@ -178,16 +195,59 @@ std::vector<int> vectorize(const DDim& ddim) {
   return result;
 }
 
+struct ProductVisitor : public boost::static_visitor<ssize_t> {
+  template <int D>
+  ssize_t operator()(const Dim<D>& dim) {
+    return product(dim);
+  }
+};
+
 ssize_t product(const DDim& ddim) {
-  ssize_t result = 1;
-  std::vector<int> v = vectorize(ddim);
-  for (auto i : v) {
-    result *= i;
+  ProductVisitor visitor;
+  return boost::apply_visitor(visitor, ddim);
+}
+
+struct SliceVectorizeVisitor : public boost::static_visitor<> {
+  std::vector<int>& vector;
+  int begin;
+  int end;
+
+  SliceVectorizeVisitor(std::vector<int>& v, int b, int e)
+      : vector(v), begin(b), end(e) {
+    PADDLE_ENFORCE(begin < end,
+                   "Begin index must be less than end index in ddim slice.");
+    PADDLE_ENFORCE(begin >= 0,
+                   "Begin index can't be less than zero in ddim slice.");
   }
-  return result;
+
+  template <int S>
+  void operator()(const Dim<S>& dim) {
+    if (begin == 0) {
+      vector.push_back(dim.head);
+    } else {
+      --begin;
+    }
+    --end;
+    if (end > 0) {
+      this->operator()(dim.tail);
+    }
+  }
+
+  void operator()(const Dim<1>& dim) {
+    PADDLE_ENFORCE(end == 1, "End index in ddim slice is out of bound.");
+    vector.push_back(dim.head);
+  }
+};
+
+DDim slice_ddim(const DDim& dim, int begin, int end) {
+  std::vector<int> vec;
+  vec.reserve(end - begin);
+  SliceVectorizeVisitor visitor(vec, begin, end);
+  boost::apply_visitor(visitor, dim);
+  return make_ddim(vec);
 }
 
-///\cond HIDDEN
+/// \cond HIDDEN
 
 struct ArityVisitor : boost::static_visitor<int> {
   template <int D>
@@ -196,15 +256,15 @@ struct ArityVisitor : boost::static_visitor<int> {
   }
 };
 
-///\endcond
+/// \endcond
 
 int arity(const DDim& d) { return boost::apply_visitor(ArityVisitor(), d); }
 
-///\cond HIDDEN
+/// \cond HIDDEN
 
 struct DDimPrinter : boost::static_visitor<void> {
   std::ostream& os;
-  DDimPrinter(std::ostream& os_) : os(os_) {}
+  explicit DDimPrinter(std::ostream& os_) : os(os_) {}
 
   template <typename T>
   void operator()(const T& t) {
@@ -212,7 +272,7 @@ struct DDimPrinter : boost::static_visitor<void> {
   }
 };
 
-///\endcond
+/// \endcond
 
 std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
   DDimPrinter printer(os);
@@ -220,5 +280,9 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
   return os;
 }
 
+DDim::DDim(std::initializer_list<int> init_list) {
+  *this = make_ddim(init_list);
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/ddim.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
 #pragma once
 
 #include <boost/variant.hpp>
 #include <initializer_list>
 #include <stdexcept>
 #include <vector>
-
 #include "paddle/framework/dim.h"
+#include "paddle/platform/enforce.h"
+#include "unsupported/Eigen/CXX11/Tensor"
 
 namespace paddle {
 namespace framework {
@@ -27,7 +42,9 @@ struct DDim {
   DDim() : var(Dim<1>()) {}
 
   template <int D>
-  DDim(const Dim<D>& in) : var(in) {}
+  explicit DDim(const Dim<D>& in) : var(in) {}
+
+  /*implicit*/ DDim(std::initializer_list<int> init_list);
 
   template <int D>
   DDim& operator=(const Dim<D>& in) {
@@ -57,6 +74,8 @@ struct DDim {
   DDim operator+(DDim d) const;
 
   DDim operator*(DDim d) const;
+
+  ssize_t size() const;
 };
 
 /**
@@ -81,6 +100,15 @@ std::vector<int> vectorize(const DDim& ddim);
 
 ssize_t product(const DDim& ddim);
 
+/**
+ * \brief Slice a ddim
+ *
+ * Slice dim with [begin, end).
+ * e.g.  DDim d = make_ddim({1,2,3,4,5});
+ *       slice_ddim(d, 1, 3); ====> {2,3}
+ */
+DDim slice_ddim(const DDim& dim, int begin, int end);
+
 /**
  * \brief What is the length of this dimension?
  *

paddle/framework/ddim_test.cc

@@ -49,9 +49,30 @@ TEST(DDim, Equality) {
 
   // arity of a DDim
   EXPECT_EQ(paddle::framework::arity(ddim), 3);
+  EXPECT_EQ(ddim.size(), 3);
 
   // product of a DDim
   EXPECT_EQ(paddle::framework::product(vddim), 45);
+  EXPECT_EQ(
+      paddle::framework::product(paddle::framework::make_ddim({3, 2, 5, 3})),
+      90);
+
+  // slice a DDim
+  paddle::framework::DDim ddim2 =
+      paddle::framework::make_ddim({1, 2, 3, 4, 5, 6});
+  paddle::framework::DDim ss = paddle::framework::slice_ddim(ddim2, 2, 5);
+  EXPECT_EQ(arity(ss), 3);
+  EXPECT_EQ(ss[0], 3);
+  EXPECT_EQ(ss[1], 4);
+  EXPECT_EQ(ss[2], 5);
+  paddle::framework::DDim ss2 = paddle::framework::slice_ddim(ddim2, 0, 6);
+  EXPECT_EQ(arity(ss2), 6);
+  EXPECT_EQ(ss2[0], 1);
+  EXPECT_EQ(ss2[1], 2);
+  EXPECT_EQ(ss2[2], 3);
+  EXPECT_EQ(ss2[3], 4);
+  EXPECT_EQ(ss2[4], 5);
+  EXPECT_EQ(ss2[5], 6);
 }
 
 TEST(DDim, Print) {

paddle/framework/dim_test.cu

 #include <thrust/device_vector.h>
 #include <sstream>
 
-#include "paddle/framework/dim.h"
 #include "gtest/gtest.h"
+#include "paddle/framework/dim.h"
 
 __global__ void test(paddle::framework::Dim<2>* o) {
-    o[0] = paddle::framework::make_dim(5, 6);
+  o[0] = paddle::framework::make_dim(5, 6);
 }
 
 __global__ void dyn_idx_gpu(int* o) {
-    auto d = paddle::framework::make_dim(5, 6);
-    o[0] = d[1];
+  auto d = paddle::framework::make_dim(5, 6);
+  o[0] = d[1];
 }
 
 TEST(Dim, Equality) {
-    // construct a Dim on the CPU
-    auto a = paddle::framework::make_dim(3, 4);
-    EXPECT_EQ(paddle::framework::get<0>(a), 3);
-    EXPECT_EQ(paddle::framework::get<1>(a), 4);
-
-    // construct a Dim on the GPU
-    thrust::device_vector<paddle::framework::Dim<2>> t(2);
-    test<<<1,1>>>(thrust::raw_pointer_cast(t.data()));
-    a = t[0];
-    EXPECT_EQ(paddle::framework::get<0>(a), 5);
-    EXPECT_EQ(paddle::framework::get<1>(a), 6);
-
-    // linearization
-    auto b = paddle::framework::make_dim(7, 8);
-    EXPECT_EQ(paddle::framework::linearize(a, b), 83);
-
-    // product
-    EXPECT_EQ(paddle::framework::product(a), 30);
-
-    // mutate a Dim
-    paddle::framework::get<1>(b) = 10;
-    EXPECT_EQ(paddle::framework::get<0>(b), 7);
-    EXPECT_EQ(paddle::framework::get<1>(b), 10);
-
-    // dynamic access
-    paddle::framework::get(b, 0) = 8;
-    b[1] = 11;
-    EXPECT_EQ(paddle::framework::get<0>(b), 8);
-    EXPECT_EQ(paddle::framework::get<1>(b), 11);
-    EXPECT_EQ(paddle::framework::get(b, 0), 8);
-    EXPECT_EQ(b[1], 11);
-
-    // dynamic access on GPU
-    thrust::device_vector<int> r(1);
-    dyn_idx_gpu<<<1,1>>>(thrust::raw_pointer_cast(r.data()));
-    int res = r[0];
-    EXPECT_EQ(res, 6);
-
-    // ex_prefix_mul
-    paddle::framework::Dim<3> c = paddle::framework::ex_prefix_mul(paddle::framework::Dim<3>(3, 4, 5));
-    EXPECT_EQ(paddle::framework::get<0>(c), 1);
-    EXPECT_EQ(paddle::framework::get<1>(c), 3);
-    EXPECT_EQ(paddle::framework::get<2>(c), 12);
-
-    // generate from an index
-    auto size = paddle::framework::make_dim(4, 5, 2);
-    c = paddle::framework::Dim<3>(14, size);
-    EXPECT_EQ(paddle::framework::get<0>(c), 2);
-    EXPECT_EQ(paddle::framework::get<1>(c), 3);
-    EXPECT_EQ(paddle::framework::get<2>(c), 0);
-    c = paddle::framework::Dim<3>(25, size);
-    EXPECT_EQ(paddle::framework::get<0>(c), 1);
-    EXPECT_EQ(paddle::framework::get<1>(c), 1);
-    EXPECT_EQ(paddle::framework::get<2>(c), 1);
+  // construct a Dim on the CPU
+  auto a = paddle::framework::make_dim(3, 4);
+  EXPECT_EQ(paddle::framework::get<0>(a), 3);
+  EXPECT_EQ(paddle::framework::get<1>(a), 4);
+
+  // construct a Dim on the GPU
+  thrust::device_vector<paddle::framework::Dim<2>> t(2);
+  test<<<1, 1>>>(thrust::raw_pointer_cast(t.data()));
+  a = t[0];
+  EXPECT_EQ(paddle::framework::get<0>(a), 5);
+  EXPECT_EQ(paddle::framework::get<1>(a), 6);
+
+  // linearization
+  auto b = paddle::framework::make_dim(7, 8);
+  EXPECT_EQ(paddle::framework::linearize(a, b), 83);
+
+  // product
+  EXPECT_EQ(paddle::framework::product(a), 30);
+
+  // mutate a Dim
+  paddle::framework::get<1>(b) = 10;
+  EXPECT_EQ(paddle::framework::get<0>(b), 7);
+  EXPECT_EQ(paddle::framework::get<1>(b), 10);
+
+  // dynamic access
+  paddle::framework::get(b, 0) = 8;
+  b[1] = 11;
+  EXPECT_EQ(paddle::framework::get<0>(b), 8);
+  EXPECT_EQ(paddle::framework::get<1>(b), 11);
+  EXPECT_EQ(paddle::framework::get(b, 0), 8);
+  EXPECT_EQ(b[1], 11);
+
+  // dynamic access on GPU
+  thrust::device_vector<int> r(1);
+  dyn_idx_gpu<<<1, 1>>>(thrust::raw_pointer_cast(r.data()));
+  int res = r[0];
+  EXPECT_EQ(res, 6);
+
+  // ex_prefix_mul
+  paddle::framework::Dim<3> c =
+      paddle::framework::ex_prefix_mul(paddle::framework::Dim<3>(3, 4, 5));
+  EXPECT_EQ(paddle::framework::get<0>(c), 1);
+  EXPECT_EQ(paddle::framework::get<1>(c), 3);
+  EXPECT_EQ(paddle::framework::get<2>(c), 12);
+
+  // generate from an index
+  auto size = paddle::framework::make_dim(4, 5, 2);
+  c = paddle::framework::Dim<3>(14, size);
+  EXPECT_EQ(paddle::framework::get<0>(c), 2);
+  EXPECT_EQ(paddle::framework::get<1>(c), 3);
+  EXPECT_EQ(paddle::framework::get<2>(c), 0);
+  c = paddle::framework::Dim<3>(25, size);
+  EXPECT_EQ(paddle::framework::get<0>(c), 1);
+  EXPECT_EQ(paddle::framework::get<1>(c), 1);
+  EXPECT_EQ(paddle::framework::get<2>(c), 1);
 }
 
 TEST(Dim, Bool) {
-    auto a = paddle::framework::make_dim(3, 4);
-    auto b = paddle::framework::make_dim(5, 6);
-    auto c = paddle::framework::make_dim(3, 4);
-
-    // in_bounds check
-    EXPECT_TRUE(paddle::framework::contained(a, b));
-    EXPECT_FALSE(paddle::framework::contained(b, a));
-
-    // comparison
-    EXPECT_TRUE(a == a);
-    EXPECT_FALSE(a == b);
-    EXPECT_TRUE(a == c);
+  auto a = paddle::framework::make_dim(3, 4);
+  auto b = paddle::framework::make_dim(5, 6);
+  auto c = paddle::framework::make_dim(3, 4);
+
+  // in_bounds check
+  EXPECT_TRUE(paddle::framework::contained(a, b));
+  EXPECT_FALSE(paddle::framework::contained(b, a));
+
+  // comparison
+  EXPECT_TRUE(a == a);
+  EXPECT_FALSE(a == b);
+  EXPECT_TRUE(a == c);
 }
 
 TEST(Dim, Print) {
-    {
-        std::stringstream ss;
-        auto a = paddle::framework::make_dim(2, 3);
-        ss << a;
-        EXPECT_EQ(ss.str(), "2, 3");
-    }
-    {
-        std::stringstream ss;
-        ss << paddle::framework::make_dim(8);
-        EXPECT_EQ(ss.str(), "8");
-    }
+  {
+    std::stringstream ss;
+    auto a = paddle::framework::make_dim(2, 3);
+    ss << a;
+    EXPECT_EQ(ss.str(), "2, 3");
+  }
+  {
+    std::stringstream ss;
+    ss << paddle::framework::make_dim(8);
+    EXPECT_EQ(ss.str(), "8");
+  }
 }

paddle/framework/eigen.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "paddle/framework/tensor.h"
+#include "unsupported/Eigen/CXX11/Tensor"
+
+namespace paddle {
+namespace framework {
+
+// EigenDim converts paddle::platform::DDim into Eigen::DSizes.
+template <int D>
+struct EigenDim {
+  using Type = Eigen::DSizes<Eigen::DenseIndex, D>;
+
+  static Type From(const DDim& dims) {
+    PADDLE_ENFORCE(arity(dims) == D, "D must match arity(DDim)");
+    Type ret;
+    for (int d = 0; d < arity(dims); d++) {
+      ret[d] = dims[d];
+    }
+    return ret;
+  }
+};
+
+// Interpret paddle::platform::Tensor as EigenTensor and EigenConstTensor.
+template <typename T, size_t D, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+struct EigenTensor {
+  // TODO(qijun) Now, default type in unaligned, and we will make a benchmark on
+  // the speed of aligned and unaligned version in future.
+  using Type = Eigen::TensorMap<Eigen::Tensor<T, D, MajorType, IndexType>>;
+
+  using ConstType =
+      Eigen::TensorMap<Eigen::Tensor<const T, D, MajorType, IndexType>>;
+
+  static Type From(Tensor& tensor, DDim dims) {
+    return Type(tensor.data<T>(), EigenDim<D>::From(dims));
+  }
+
+  static Type From(Tensor& tensor) { return From(tensor, tensor.dims_); }
+
+  static ConstType From(const Tensor& tensor, DDim dims) {
+    return ConstType(tensor.data<T>(), EigenDim<D>::From(dims));
+  }
+
+  static ConstType From(const Tensor& tensor) {
+    return From(tensor, tensor.dims_);
+  }
+};
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+struct EigenVector : public EigenTensor<T, 1, MajorType, IndexType> {
+  // Flatten is to reshape a Tensor into a one dimension EigenVector
+  static typename EigenTensor<T, 1>::Type Flatten(Tensor& tensor) {
+    return EigenTensor<T, 1>::From(
+        tensor, make_ddim({static_cast<int>(product(tensor.dims_))}));
+  }
+
+  static typename EigenTensor<T, 1>::ConstType Flatten(const Tensor& tensor) {
+    return EigenTensor<T, 1>::From(
+        tensor, make_ddim({static_cast<int>(product(tensor.dims_))}));
+  }
+};
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = EigenTensor<T, 2, MajorType, IndexType>;
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/eigen_test.cc

+/*
+  Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+  Licensed under the Apache License, Version 2.0 (the "License");
+  you may not use this file except in compliance with the License.
+  You may obtain a copy of the License at
+  http://www.apache.org/licenses/LICENSE-2.0
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+*/
+
+#include "paddle/framework/eigen.h"
+#include <gtest/gtest.h>
+
+namespace paddle {
+namespace framework {
+
+TEST(EigenDim, From) {
+  EigenDim<3>::Type ed = EigenDim<3>::From(make_ddim({1, 2, 3}));
+  ASSERT_EQ(1, ed[0]);
+  ASSERT_EQ(2, ed[1]);
+  ASSERT_EQ(3, ed[2]);
+}
+
+TEST(Eigen, Tensor) {
+  Tensor t;
+  float* p = t.mutable_data<float>(make_ddim({1, 2, 3}), platform::CPUPlace());
+  for (int i = 0; i < 1 * 2 * 3; i++) {
+    p[i] = static_cast<float>(i);
+  }
+
+  EigenTensor<float, 3>::Type et = EigenTensor<float, 3>::From(t);
+
+  ASSERT_EQ(1, et.dimension(0));
+  ASSERT_EQ(2, et.dimension(1));
+  ASSERT_EQ(3, et.dimension(2));
+
+  for (int i = 0; i < 1; i++) {
+    for (int j = 0; j < 2; j++) {
+      for (int k = 0; k < 3; k++) {
+        ASSERT_NEAR((i * 2 + j) * 3 + k, et(i, j, k), 1e-6f);
+      }
+    }
+  }
+}
+
+TEST(Eigen, VectorFrom) {
+  Tensor t;
+  float* p = t.mutable_data<float>(make_ddim({6}), platform::CPUPlace());
+  for (int i = 0; i < 6; i++) {
+    p[i] = static_cast<float>(i);
+  }
+
+  EigenVector<float>::Type ev = EigenVector<float>::From(t);
+
+  ASSERT_EQ(6, ev.dimension(0));
+
+  for (int i = 0; i < 6; i++) {
+    ASSERT_NEAR(i, ev(i), 1e-6f);
+  }
+}
+
+TEST(Eigen, VectorFlatten) {
+  Tensor t;
+  float* p = t.mutable_data<float>(make_ddim({1, 2, 3}), platform::CPUPlace());
+  for (int i = 0; i < 1 * 2 * 3; i++) {
+    p[i] = static_cast<float>(i);
+  }
+
+  EigenVector<float>::Type ev = EigenVector<float>::Flatten(t);
+
+  ASSERT_EQ(1 * 2 * 3, ev.dimension(0));
+
+  for (int i = 0; i < 1 * 2 * 3; i++) {
+    ASSERT_NEAR(i, ev(i), 1e-6f);
+  }
+}
+
+TEST(Eigen, Matrix) {
+  Tensor t;
+  float* p = t.mutable_data<float>(make_ddim({2, 3}), platform::CPUPlace());
+  for (int i = 0; i < 2 * 3; i++) {
+    p[i] = static_cast<float>(i);
+  }
+
+  EigenMatrix<float>::Type em = EigenMatrix<float>::From(t);
+
+  ASSERT_EQ(2, em.dimension(0));
+  ASSERT_EQ(3, em.dimension(1));
+
+  for (int i = 0; i < 2; i++) {
+    for (int j = 0; j < 3; j++) {
+      ASSERT_NEAR(i * 3 + j, em(i, j), 1e-6f);
+    }
+  }
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/enforce.h

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-    http://www.apache.org/licenses/LICENSE-2.0
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#pragma once
-#include <paddle/string/printf.h>
-#include <exception>
-#include <sstream>
-
-namespace paddle {
-namespace framework {
-
-/**
- * @brief Enforce exception. Inherits std::exception
- *
- * All enforce condition not met, will throw an EnforceNotMet exception.
- */
-class EnforceNotMet : public std::exception {
- public:
-  EnforceNotMet(const std::string& msg, const char* file, int fileline) {
-    std::ostringstream sout;
-    sout << msg << " at [" << file << ":" << fileline << "];";
-    all_msg_ = sout.str();
-  }
-
-  const char* what() const noexcept override { return all_msg_.c_str(); }
-
- private:
-  std::string all_msg_;
-};
-
-// From https://stackoverflow.com/questions/30130930/
-// __buildin_expect is in C++ 11 standard. Since the condition which enforced
-// should be true in most situation, it will make the compiler generate faster
-// code by adding `UNLIKELY` macro.
-#define UNLIKELY(condition) __builtin_expect(static_cast<bool>(condition), 0)
-
-/**
- * @brief Throw a EnforceNotMet exception, automatically filled __FILE__ &
- * __LINE__
- *
- * This macro take __VA_ARGS__, user can pass any type if that type can
- * serialize to std::ostream
- */
-#define PADDLE_THROW(...)                                            \
-  do {                                                               \
-    throw ::paddle::framework::EnforceNotMet(                        \
-        ::paddle::string::Sprintf(__VA_ARGS__), __FILE__, __LINE__); \
-  } while (0)
-
-/**
- * @brief Enforce a condition, otherwise throw an EnforceNotMet
- */
-#define PADDLE_ENFORCE(condition, ...) \
-  do {                                 \
-    if (UNLIKELY(!(condition))) {      \
-      PADDLE_THROW(__VA_ARGS__);       \
-    }                                  \
-  } while (0)
-
-}  // namespace framework
-}  // namespace paddle

paddle/framework/net.cc

+/*
+  Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+  Licensed under the Apache License, Version 2.0 (the "License");
+  you may not use this file except in compliance with the License.
+  You may obtain a copy of the License at
+
+  http://www.apache.org/licenses/LICENSE-2.0
+
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+*/
+
 #include "paddle/framework/net.h"
 
 namespace paddle {
 namespace framework {
 
-PlainNet::PlainNet(const NetDesc& def) {}
+void PlainNet::CompleteAddOp(bool calc) {
+  add_op_done_ = true;
+  if (!calc) return;
 
-void PlainNet::InferShape(const ScopePtr& scope) const {
+  std::unordered_set<std::string> input_set;
+  std::unordered_set<std::string> output_set;
+  std::unordered_set<std::string> temp_output;
   for (auto& op : ops_) {
-    op.InferShape();
+    for (auto& ipt : op->inputs_) {
+      if (!Contains(output_set, ipt)) {  // Not other op's output
+        input_set.insert(ipt);
+      } else {
+        temp_output.insert(ipt);
+      }
+    }
+
+    for (auto& opt : op->outputs_) {
+      output_set.insert(opt);
+    }
   }
+  inputs_.reserve(input_set.size());
+  std::copy(input_set.begin(), input_set.end(), std::back_inserter(inputs_));
+
+  outputs_.reserve(output_set.size());
+  std::vector<int> tmp_index;
+  tmp_index.reserve(temp_output.size());
+  int idx = 0;
+  for (auto& opt : output_set) {
+    if (Contains(temp_output, opt)) {
+      tmp_index.push_back(idx);
+    }
+    outputs_.push_back(opt);
+    ++idx;
+  }
+
+  attrs_["temporary_index"] = tmp_index;
 }
 
-void PlainNet::Run(const ScopePtr& scope, const DeviceContext& ctx) const {
+std::string PlainNet::DebugString() const {
+  std::ostringstream os;
+  os << this->type_ << ":" << std::endl;
   for (auto& op : ops_) {
-    op.Run(ctx);
+    os << "\t" << op->DebugString() << std::endl;
   }
+  return os.str();
 }
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/net.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
+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
+    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. */
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
 
 #pragma once
 
-#include "paddle/framework/net_proto.pb.h"
+#include <paddle/framework/op_desc.pb.h>
+#include <paddle/framework/operator.h>
 #include "paddle/framework/op_proto.pb.h"
+#include "paddle/framework/op_registry.h"
 #include "paddle/framework/scope.h"
 #include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace framework {
-using namespace paddle::platform;
-
-// operator's index stored in a network.
-typedef int OpIndex;
-/**
- * NOTE following codes are some definitions of unimplemented concepts.
- * We write some basic implementation to make Net compilable. These APIs will
- * keep updating if the concepts related are implemented.
- */
-
-struct OpDesc;
-struct OpAttrs {};
-
-class Operator {
- public:
-  Operator(const OpDesc &def) {}
-  void InferShape() const {}
-  void Run(const DeviceContext &ctx) const {}
-};
-
 /**
- * @brief Network that manage the operators it has.
+ * @brief Network is also a type of Operator
+ *
+ * It will manage the operators it has.
  *
- * Network is the container and controller of a set of operators, user can build
- * a real network from a NetDesc which is a protobuf message and use
- * Network.Run() * to run all the operators in the network.
+ * Network is the container and controller of a set of operators.
 
  * A network object knows all Operators belonging to this network. Variables,
  * which are inputs and outputs of these operators, are created and managed by a
  * hierarchy of Scope objects.
  *
- * This is the base class of network, all the networks should implement the apis
+ * This is the base class of network, all the networks should implement the APIs
  * it defines.
  */
-class Net {
+class Net : public OperatorBase {
  public:
-  /**
-   * @brief Infer shapes of all inputs and outputs of operators.
-   */
-  virtual void InferShape(const ScopePtr &scope) const = 0;
-  /**
-   * @brief Run the network.
-   *
-   * Run all the operators and return success(true) or not, with all the
-   * variables are located in `scope`. `context` describes the detail execution
-   * environment for ops. `begin` and `end` specify the scope of `ops_` to run,
-   * If no positive indexes are provided, all operators in `ops_` will run.
-   */
-  virtual void Run(const ScopePtr &scope, const DeviceContext &ctx) const = 0;
-
-  /**
-   * @brief Add an Operator according to `def`.
-   */
-  virtual OpIndex AddOp(const OpProto &def) = 0;
-
-  /**
-   * @brief Add optimizer operators acctording to `attrs`.
-   */
-  virtual void AddOptimizerOps(const OpAttrs &attrs) = 0;
-
-  /**
-   * @brief Add backward operators.
-   */
-  virtual void AddBackwardOps() = 0;
-
-  /**
-   * @brief Create a network.
-   */
-  static std::unique_ptr<Net> Create(const NetDesc &def = NetDesc());
-
-  virtual ~Net() {}
+  virtual void AddOp(const OperatorPtr& op) = 0;
+  virtual void CompleteAddOp(bool calc) = 0;
 };
 
+using NetPtr = std::shared_ptr<Net>;
+
 /**
  * @brief a basic implementation of Net.
  *
@@ -103,18 +54,14 @@ class Net {
 class PlainNet : public Net {
  public:
   /**
-   * @brief Initialize a PlainNet.
-   *
-   * Initialize from  a network describe by `def`. NetDesc is the definition of
-   * a network.
-   */
-  PlainNet(const NetDesc &def);
-
-  /**
-   * Infer all the operators' input and output varialbes' shapes, will be called
+   * Infer all the operators' input and output variables' shapes, will be called
    * before every mini-batch
    */
-  virtual void InferShape(const ScopePtr &scope) const override;
+  void InferShape(const ScopePtr& scope) const override {
+    for (auto& op : ops_) {
+      op->InferShape(scope);
+    }
+  }
 
   /**
    * @brief Run the network.
@@ -123,49 +70,34 @@ class PlainNet : public Net {
    * scope will be used instead. If no OpContext is provicded, default context
    * will be used.
    */
-  virtual void Run(const ScopePtr &scope,
-                   const DeviceContext &ctx) const override;
+  void Run(const ScopePtr& scope,
+           const platform::DeviceContext& dev_ctx) const override {
+    for (auto& op : ops_) {
+      op->Run(scope, dev_ctx);
+    }
+  }
 
   /**
-   * @brief Add an operator to this network.
+   * @brief Add an operator by ptr
    */
-  virtual OpIndex AddOp(const OpProto &def) override;
+  void AddOp(const OperatorPtr& op) override {
+    PADDLE_ENFORCE(!add_op_done_, "Cannot AddOp when this network is sealed");
+    ops_.push_back(op);
+  }
 
-  /**
-   * @brief Add all optimizer operators related into the network.
-   */
-  virtual void AddOptimizerOps(const OpAttrs &attrs) override;
+  void CompleteAddOp(bool calculate = true) override;
 
-  /**
-   * @brief Add all backward operators related into the network.
-   */
-  virtual void AddBackwardOps() override;
-
-  virtual ~PlainNet() override {}
+  std::string DebugString() const override;
 
- protected:
-  /**
-   * @brief Build the network.
-   *
-   * Create operators accordding to `def`, will be called by the constructor.
-   */
-  void BuildNet(const NetDesc &def);
-
-  /**
-   * @brief Add an operator into this network.
-   *
-   * Add a operator which is identified as `type` and has attributes described
-   * in `attrs`, the `inputs` are the keys of readonly input variables,
-   * `outputs` are keys of mutable output variables. An `OpIndex` will be
-   * returned to indicate the offset of the new operator in `ops_`.
-   */
-  OpIndex AddOp(const std::string &type, const std::vector<std::string> &inputs,
-                const std::vector<std::string> &outputs,
-                const OpAttrs &attrs = OpAttrs());
+  std::vector<OperatorPtr> ops_;
 
  private:
-  // the operators owned by `Network`.
-  std::vector<Operator> ops_;
+  bool add_op_done_{false};
+
+  template <typename T, typename KeyType>
+  static bool Contains(T container, KeyType key) {
+    return container.find(key) != container.end();
+  }
 };
 
 }  // namespace framework

paddle/framework/net_op_test.cc

+#include <gtest/gtest.h>
+#include <paddle/framework/net.h>
+#include <paddle/framework/op_registry.h>
+#include <paddle/framework/operator.h>
+
+namespace pd = paddle::framework;
+
+static int infer_shape_cnt = 0;
+static int run_cnt = 0;
+
+class TestOp : public pd::OperatorBase {
+ public:
+  void InferShape(const paddle::framework::ScopePtr& scope) const override {
+    ++infer_shape_cnt;
+  }
+  void Run(const paddle::framework::ScopePtr& scope,
+           const paddle::platform::DeviceContext& dev_ctx) const override {
+    ++run_cnt;
+  }
+};
+
+template <typename T>
+void AssertSameVectorWithoutOrder(const std::vector<T>& expected,
+                                  const std::vector<T>& actual) {
+  ASSERT_EQ(expected.size(), actual.size());
+  std::unordered_set<T> expected_set;
+  for (auto& tmp : expected) {
+    expected_set.insert(tmp);
+  }
+  for (auto& act : actual) {
+    ASSERT_NE(expected_set.end(), expected_set.find(act));
+  }
+}
+
+TEST(OpKernel, all) {
+  auto net = std::make_shared<paddle::framework::PlainNet>();
+  ASSERT_NE(net, nullptr);
+
+  auto op1 = std::make_shared<TestOp>();
+  op1->inputs_ = {"x", "w1", "b1"};
+  op1->outputs_ = {"y"};
+  net->AddOp(op1);
+
+  auto op2 = std::make_shared<TestOp>();
+  op2->inputs_ = {"y", "w2", "b2"};
+  op2->outputs_ = {"z"};
+  net->AddOp(op2);
+
+  net->CompleteAddOp();
+  AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"}, net->inputs_);
+  AssertSameVectorWithoutOrder({"y", "z"}, net->outputs_);
+  auto tmp_idx_iter = net->attrs_.find("temporary_index");
+  ASSERT_NE(net->attrs_.end(), tmp_idx_iter);
+  auto& tmp_idx = boost::get<std::vector<int>>(tmp_idx_iter->second);
+  ASSERT_EQ(1UL, tmp_idx.size());
+  ASSERT_EQ("y", net->outputs_[tmp_idx[0]]);
+
+  auto scope = std::make_shared<pd::Scope>();
+  paddle::platform::CPUDeviceContext dev_ctx;
+
+  net->InferShape(scope);
+  net->Run(scope, dev_ctx);
+  ASSERT_EQ(2, infer_shape_cnt);
+  ASSERT_EQ(2, run_cnt);
+
+  ASSERT_THROW(net->AddOp(op2), std::runtime_error);
+}

paddle/framework/op_registry.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
 #include <paddle/framework/op_registry.h>
 
 namespace paddle {
@@ -33,4 +47,4 @@ void AttrTypeHelper::SetAttrType<std::vector<std::string>>(AttrProto* attr) {
   attr->set_type(paddle::framework::AttrType::STRINGS);
 }
 }  // namespace framework
-}  // namespace paddle
\ No newline at end of file
+}  // namespace paddle

paddle/framework/op_registry.h

 #pragma once
 
 #include <algorithm>
+#include <atomic>
 #include <type_traits>
 #include <unordered_map>
 #include <unordered_set>
@@ -61,7 +62,14 @@ class OpProtoAndCheckerMaker {
   OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
       : proto_(proto), op_checker_(op_checker) {}
 
-  ~OpProtoAndCheckerMaker() { CheckNoDuplicatedAttrs(); }
+  ~OpProtoAndCheckerMaker() {
+    PADDLE_ENFORCE(validated_, "should call Validate after build");
+  }
+
+  void Validate() {
+    validated_ = true;
+    CheckNoDuplicatedInOutAttrs();
+  }
 
  protected:
   void AddInput(const std::string& name, const std::string& comment,
@@ -163,19 +171,26 @@ Add a mark to which output is temporary is helpful for future optimization.
     }
   }
 
-  void CheckNoDuplicatedAttrs() {
+  void CheckNoDuplicatedInOutAttrs() {
     std::unordered_set<std::string> names;
-    size_t cnt = 0;
+    auto checker = [&](const std::string& name) {
+      PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name);
+      names.insert(name);
+    };
     for (auto& attr : proto_->attrs()) {
-      names.insert(attr.name());
-      ++cnt;
+      checker(attr.name());
+    }
+    for (auto& input : proto_->inputs()) {
+      checker(input.name());
+    }
+    for (auto& output : proto_->outputs()) {
+      checker(output.name());
     }
-    PADDLE_ENFORCE(names.size() == cnt,
-                   "Cannot register two attribute in same name!");
   }
 
   OpProto* proto_;
   OpAttrChecker* op_checker_;
+  bool validated_{false};
   bool has_multiple_input_{false};
   bool has_multiple_output_{false};
   bool has_temporary_output_{false};
@@ -183,6 +198,8 @@ Add a mark to which output is temporary is helpful for future optimization.
 
 class OpRegistry {
   using OpCreator = std::function<OperatorBase*()>;
+  using VarIndexMap = std::unordered_map<std::string, int>;
+  using VarNameList = std::vector<std::string>;
 
  public:
   template <typename OpType, typename ProtoMakerType>
@@ -190,36 +207,71 @@ class OpRegistry {
     creators()[op_type] = [] { return new OpType; };
     OpProto& op_proto = protos()[op_type];
     OpAttrChecker& op_checker = op_checkers()[op_type];
-    ProtoMakerType(&op_proto, &op_checker);
+    auto maker = ProtoMakerType(&op_proto, &op_checker);
+    maker.Validate();
     *op_proto.mutable_type() = op_type;
     PADDLE_ENFORCE(
         op_proto.IsInitialized(),
         "Fail to initialize %s's OpProto, because %s is not initialized",
         op_type, op_proto.InitializationErrorString());
+
+    VarIndexMaps()[op_type].reset(new VarIndexMap());
+    auto& varmap = *VarIndexMaps()[op_type];
+    int idx = 0;
+    for (auto& var : op_proto.inputs()) {
+      varmap[var.name()] = idx++;
+    }
+    idx = 0;
+    for (auto& var : op_proto.outputs()) {
+      varmap[var.name()] = idx++;
+    }
+  }
+
+  static OperatorPtr CreateOp(const std::string& type,
+                              const VarNameList& inputs,
+                              const VarNameList& outputs,
+                              const AttributeMap& attrs) {
+    auto op_create_it = creators().find(type);
+    PADDLE_ENFORCE(op_create_it != creators().end(),
+                   "Operator %s cannot be found", type);
+
+    auto op = op_create_it->second();
+    op->type_ = type;
+    op->inputs_ = inputs;
+    op->outputs_ = outputs;
+    op->attrs_ = attrs;
+    op_checkers().at(type).Check(op->attrs_);
+
+    GenerateTempVariableName(op);
+
+    {
+      auto var_index_it = VarIndexMaps().find(type);
+      if (var_index_it != VarIndexMaps().end()) {
+        op->in_out_idxs_ = var_index_it->second;
+      }
+    }
+
+    op->Init();
+    return OperatorPtr(op);
   }
 
   static OperatorPtr CreateOp(const OpDesc& op_desc) {
-    std::string op_type = op_desc.type();
-    OperatorPtr op(creators().at(op_type)());
-    op->desc_ = op_desc;
-    op->inputs_.reserve((size_t)op_desc.inputs_size());
+    std::vector<std::string> inputs;
+    inputs.reserve((size_t)op_desc.inputs_size());
     std::copy(op_desc.inputs().begin(), op_desc.inputs().end(),
-              std::back_inserter(op->inputs_));
-    op->outputs_.reserve((size_t)op_desc.outputs_size());
+              std::back_inserter(inputs));
+
+    std::vector<std::string> outputs;
+    outputs.reserve((size_t)op_desc.outputs_size());
     std::copy(op_desc.outputs().begin(), op_desc.outputs().end(),
-              std::back_inserter(op->outputs_));
+              std::back_inserter(outputs));
+
+    AttributeMap attrs;
     for (auto& attr : op_desc.attrs()) {
-      op->attrs_[attr.name()] = AttrTypeHelper::GetAttrValue(attr);
+      attrs[attr.name()] = AttrTypeHelper::GetAttrValue(attr);
     }
-    op_checkers().at(op_type).Check(op->attrs_);
-    op->Init();
-    return op;
-  }
 
- private:
-  static std::unordered_map<std::string, OpCreator>& creators() {
-    static std::unordered_map<std::string, OpCreator> creators_;
-    return creators_;
+    return CreateOp(op_desc.type(), inputs, outputs, attrs);
   }
 
   static std::unordered_map<std::string, OpProto>& protos() {
@@ -227,6 +279,29 @@ class OpRegistry {
     return protos_;
   };
 
+ private:
+  static std::unordered_map<std::string, std::shared_ptr<VarIndexMap>>&
+  VarIndexMaps() {
+    static std::unordered_map<std::string, std::shared_ptr<VarIndexMap>> maps_;
+    return maps_;
+  }
+
+  static void GenerateTempVariableName(OperatorBase* op) {
+    static std::atomic<size_t> gUniqId(0UL);
+    for (auto& outname : op->outputs_) {
+      if (outname == OperatorBase::TMP_VAR_NAME()) {
+        outname += op->type_;
+        outname += "@";
+        outname += std::to_string(gUniqId.fetch_add(1));
+      }
+    }
+  }
+
+  static std::unordered_map<std::string, OpCreator>& creators() {
+    static std::unordered_map<std::string, OpCreator> creators_;
+    return creators_;
+  }
+
   static std::unordered_map<std::string, OpAttrChecker>& op_checkers() {
     static std::unordered_map<std::string, OpAttrChecker> op_checkers_;
     return op_checkers_;
@@ -241,12 +316,18 @@ class OpRegisterHelper {
   }
 };
 
+/**
+ * check if MACRO is used in GLOBAL NAMESPACE.
+ */
 #define STATIC_ASSERT_GLOBAL_NAMESPACE(uniq_name, msg)                        \
   struct __test_global_namespace_##uniq_name##__ {};                          \
   static_assert(std::is_same<::__test_global_namespace_##uniq_name##__,       \
                              __test_global_namespace_##uniq_name##__>::value, \
                 msg)
 
+/**
+ * Macro to Register Operator.
+ */
 #define REGISTER_OP(__op_type, __op_class, __op_maker_class)                 \
   STATIC_ASSERT_GLOBAL_NAMESPACE(__reg_op__##__op_type,                      \
                                  "REGISTER_OP must be in global namespace"); \
@@ -254,27 +335,36 @@ class OpRegisterHelper {
       __op_register_##__op_type##__(#__op_type);                             \
   int __op_register_##__op_type##_handle__() { return 0; }
 
-#define REGISTER_OP_KERNEL(type, GPU_OR_CPU, PlaceType, KernelType)       \
+/**
+ * Macro to Register OperatorKernel.
+ */
+#define REGISTER_OP_KERNEL(type, DEVICE_TYPE, PlaceType, ...)             \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
-      __reg_op_kernel_##type##_##GPU_OR_CPU##__,                          \
+      __reg_op_kernel_##type##_##DEVICE_TYPE##__,                         \
       "REGISTER_OP_KERNEL must be in global namespace");                  \
   struct __op_kernel_register__##type##__ {                               \
     __op_kernel_register__##type##__() {                                  \
       ::paddle::framework::OperatorWithKernel::OpKernelKey key;           \
       key.place_ = PlaceType();                                           \
       ::paddle::framework::OperatorWithKernel::AllOpKernels()[#type][key] \
-          .reset(new KernelType());                                       \
+          .reset(new __VA_ARGS__());                                      \
     }                                                                     \
   };                                                                      \
   static __op_kernel_register__##type##__ __reg_kernel_##type##__;        \
-  int __op_kernel_register_##type##_handle_##GPU_OR_CPU##__() { return 0; }
+  int __op_kernel_register_##type##_handle_##DEVICE_TYPE##__() { return 0; }
 
-#define REGISTER_OP_GPU_KERNEL(type, KernelType) \
-  REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, KernelType)
+// (type, KernelType)
+#define REGISTER_OP_GPU_KERNEL(type, ...) \
+  REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__)
 
-#define REGISTER_OP_CPU_KERNEL(type, KernelType) \
-  REGISTER_OP_KERNEL(type, CPU, ::paddle::platform::CPUPlace, KernelType)
+// (type, KernelType)
+#define REGISTER_OP_CPU_KERNEL(type, ...) \
+  REGISTER_OP_KERNEL(type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
 
+/**
+ * Macro to mark what Operator and Kernel we will use and tell the compiler to
+ * link them into target.
+ */
 #define USE_OP_WITHOUT_KERNEL(op_type)                      \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                           \
       __use_op_without_kernel_##op_type,                    \
@@ -292,15 +382,16 @@ class OpRegisterHelper {
       __attribute__((unused)) =                                           \
           __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__()
 
-#ifdef PADDLE_ONLY_CPU
-#define USE_OP(op_type)           \
+// use Operator with only cpu kernel.
+#define USE_OP_CPU(op_type)       \
   USE_OP_WITHOUT_KERNEL(op_type); \
-  USE_OP_KERNEL(op_type, CPU);
+  USE_OP_KERNEL(op_type, CPU)
 
+#ifdef PADDLE_ONLY_CPU
+#define USE_OP(op_type) USE_OP_CPU(op_type)
 #else
-#define USE_OP(op_type)           \
-  USE_OP_WITHOUT_KERNEL(op_type); \
-  USE_OP_KERNEL(op_type, CPU);    \
+#define USE_OP(op_type) \
+  USE_OP_CPU(op_type);  \
   USE_OP_KERNEL(op_type, GPU)
 #endif
 

paddle/framework/op_registry_test.cc

 #include "paddle/framework/op_registry.h"
 #include <gtest/gtest.h>
 
+namespace pd = paddle::framework;
+
 namespace paddle {
 namespace framework {
 class CosineOp : public OperatorBase {
@@ -28,8 +30,6 @@ class MyTestOp : public OperatorBase {
   void InferShape(const ScopePtr& scope) const override {}
   void Run(const ScopePtr& scope,
            const platform::DeviceContext& dev_ctx) const override {}
-
- public:
 };
 
 class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
@@ -91,7 +91,7 @@ TEST(OpRegistry, IllegalAttr) {
   try {
     paddle::framework::OperatorPtr op __attribute__((unused)) =
         paddle::framework::OpRegistry::CreateOp(op_desc);
-  } catch (paddle::framework::EnforceNotMet err) {
+  } catch (std::runtime_error& err) {
     caught = true;
     std::string msg = "larger_than check fail";
     const char* err_msg = err.what();
@@ -138,7 +138,7 @@ TEST(OpRegistry, CustomChecker) {
   try {
     paddle::framework::OperatorPtr op __attribute__((unused)) =
         paddle::framework::OpRegistry::CreateOp(op_desc);
-  } catch (paddle::framework::EnforceNotMet err) {
+  } catch (std::runtime_error& err) {
     caught = true;
     std::string msg = "Attribute 'test_attr' is required!";
     const char* err_msg = err.what();
@@ -157,7 +157,7 @@ TEST(OpRegistry, CustomChecker) {
   try {
     paddle::framework::OperatorPtr op __attribute__((unused)) =
         paddle::framework::OpRegistry::CreateOp(op_desc);
-  } catch (paddle::framework::EnforceNotMet err) {
+  } catch (std::runtime_error& err) {
     caught = true;
     std::string msg = "'test_attr' must be even!";
     const char* err_msg = err.what();
@@ -182,3 +182,35 @@ TEST(OpRegistry, CustomChecker) {
   int test_attr = op->GetAttr<int>("test_attr");
   ASSERT_EQ(test_attr, 4);
 }
+
+class TestAttrProtoMaker : public pd::OpProtoAndCheckerMaker {
+ public:
+  TestAttrProtoMaker(pd::OpProto* proto, pd::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddAttr<float>("scale", "scale of test op");
+    AddAttr<float>("scale", "scale of test op");
+  }
+};
+
+TEST(ProtoMaker, DuplicatedAttr) {
+  pd::OpProto op_proto;
+  pd::OpAttrChecker op_checker;
+  auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker);
+  ASSERT_THROW(proto_maker.Validate(), std::runtime_error);
+}
+
+class TestInOutProtoMaker : public pd::OpProtoAndCheckerMaker {
+ public:
+  TestInOutProtoMaker(pd::OpProto* proto, pd::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("input", "input of test op");
+    AddInput("input", "input of test op");
+  }
+};
+
+TEST(ProtoMaker, DuplicatedInOut) {
+  pd::OpProto op_proto;
+  pd::OpAttrChecker op_checker;
+  auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker);
+  ASSERT_THROW(proto_maker.Validate(), std::runtime_error);
+}

paddle/framework/operator.cc

@@ -12,32 +12,92 @@ 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 <algorithm>
+
 #include "paddle/framework/operator.h"
 
 namespace paddle {
 namespace framework {
 
+template <>
+Eigen::DefaultDevice* KernelContext::GetEigenDevice<
+    platform::CPUPlace, Eigen::DefaultDevice>() const {
+  return device_context_.get_eigen_device<Eigen::DefaultDevice>();
+}
+
+#ifndef PADDLE_ONLY_CPU
+template <>
+Eigen::GpuDevice*
+KernelContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
+  return device_context_.get_eigen_device<Eigen::GpuDevice>();
+}
+#endif
+
+const std::string& OperatorBase::Input(const std::string& name) const {
+  auto it = in_out_idxs_->find(name);
+  PADDLE_ENFORCE(it != in_out_idxs_->end(), "no key [%s] in in_out_idxs_",
+                 name);
+
+  if (attrs_.count("input_format") == 0) {
+    return inputs_[it->second];
+  } else {
+    const auto& input_format = GetAttr<std::vector<int>>("input_format");
+    int idx = input_format[it->second];
+    return inputs_.at(idx);
+  }
+}
+
+std::vector<std::string> OperatorBase::Inputs(const std::string& name) const {
+  auto input_format = GetAttr<std::vector<int>>("input_format");
+  auto offset = in_out_idxs_->at(name);
+
+  return std::vector<std::string>{
+      inputs_.begin() + input_format.at(offset),
+      inputs_.begin() + input_format.at(offset + 1)};
+}
+
+const std::string& OperatorBase::Output(const std::string& name) const {
+  auto it = in_out_idxs_->find(name);
+  PADDLE_ENFORCE(it != in_out_idxs_->end(), "no key [%s] in in_out_idxs_",
+                 name);
+
+  if (attrs_.count("output_format") == 0) {
+    return outputs_[it->second];
+  } else {
+    const auto& output_format = GetAttr<std::vector<int>>("output_format");
+    int idx = output_format[it->second];
+    return outputs_.at(idx);
+  }
+}
+
+std::vector<std::string> OperatorBase::Outputs(const std::string& name) const {
+  auto output_format = GetAttr<std::vector<int>>("output_format");
+  auto offset = in_out_idxs_->at(name);
+
+  return std::vector<std::string>{
+      outputs_.begin() + output_format.at(offset),
+      outputs_.begin() + output_format.at(offset + 1)};
+}
+
 std::string OperatorBase::DebugString() const {
   std::stringstream ss;
-  ss << "=================\n";
-  ss << "type = " << desc_.type() << "\n";
-  ss << "inputs = [";
-  for (auto& ipt : inputs_) {
-    ss << ipt << ", ";
-  }
-  ss << "]\n";
-  ss << "outputs = [";
-  for (auto& opt : outputs_) {
-    ss << opt << ", ";
+  ss << "Op(" << type_ << "), inputs:(";
+  for (size_t i = 0; i < inputs_.size(); ++i) {
+    ss << inputs_[i];
+    if (i != inputs_.size() - 1) {
+      ss << ", ";
+    }
   }
-  ss << "]\n";
-  ss << "attr_keys = [";
-  for (auto& attr : attrs_) {
-    ss << attr.first << ", ";
+  ss << "), outputs:(";
+  for (size_t i = 0; i < outputs_.size(); ++i) {
+    ss << outputs_[i];
+    if (i != outputs_.size() - 1) {
+      ss << ", ";
+    }
   }
-  ss << "]\n";
+  ss << ").";
   return ss.str();
 }
 
 }  // namespace framework
-}  // namespace paddle
\ No newline at end of file
+}  // namespace paddle

paddle/framework/operator.h

@@ -14,21 +14,38 @@ limitations under the License. */
 
 #pragma once
 
-#include <paddle/framework/attr_checker.h>
-#include <paddle/framework/op_desc.pb.h>
-#include <paddle/framework/scope.h>
-#include <paddle/framework/tensor.h>
-#include <paddle/platform/device_context.h>
-#include <paddle/platform/place.h>
-#include <paddle/utils/Error.h>
 #include <boost/variant.hpp>
 #include <string>
 #include <unordered_map>
 #include <vector>
 
+#include "paddle/framework/attr_checker.h"
+#include "paddle/framework/op_desc.pb.h"
+#include "paddle/framework/op_proto.pb.h"
+#include "paddle/framework/scope.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/device_context.h"
+#include "paddle/platform/place.h"
+#include "paddle/utils/Error.h"
+
 namespace paddle {
 namespace framework {
 
+template <typename T>
+struct EigenDeviceConverter;
+
+template <>
+struct EigenDeviceConverter<platform::CPUPlace> {
+  using EigenDeviceType = Eigen::DefaultDevice;
+};
+
+#ifndef PADDLE_ONLY_CPU
+template <>
+struct EigenDeviceConverter<platform::GPUPlace> {
+  using EigenDeviceType = Eigen::GpuDevice;
+};
+#endif
+
 class OperatorBase;
 using OperatorPtr = std::shared_ptr<OperatorBase>;
 /**
@@ -39,6 +56,13 @@ using OperatorPtr = std::shared_ptr<OperatorBase>;
  */
 class OperatorBase {
  public:
+  /// If a variable is a empty variable, that name will be used.
+  static std::string EMPTY_VAR_NAME() { return "@EMPTY@"; }
+
+  /// If a variable is a temporary variable, that name will be set in Python,
+  /// but it will be convert to a unique name in scope after OpCreator.
+  static std::string TMP_VAR_NAME() { return "@TEMP@"; }
+
   virtual ~OperatorBase() {}
 
   template <typename T>
@@ -48,7 +72,7 @@ class OperatorBase {
     return boost::get<T>(attrs_.at(name));
   }
 
-  std::string DebugString() const;
+  virtual std::string DebugString() const;
 
   /// Init will be called after CreateOperator, you can put some initialization
   /// logic here.
@@ -62,14 +86,76 @@ class OperatorBase {
   virtual void Run(const ScopePtr& scope,
                    const platform::DeviceContext& dev_ctx) const = 0;
 
- protected:
-  std::string Type() const { return desc_.type(); }
+  // Get a input with argument's name described in `op_proto`
+  const std::string& Input(const std::string& name) const;
+  // Get a input which has multiple variables.
+  // TODO add a vector_view to prevent memory copy.
+  std::vector<std::string> Inputs(const std::string& name) const;
+  // Get a output with argument's name described in `op_proto`
+  const std::string& Output(const std::string& name) const;
+  // Get an output which has multiple variables.
+  // TODO add a vector_view to prevent memory copy.
+  std::vector<std::string> Outputs(const std::string& name) const;
 
  public:
-  OpDesc desc_;
+  std::string type_;
   std::vector<std::string> inputs_;
   std::vector<std::string> outputs_;
   AttributeMap attrs_;
+  // store the arguments' offset described in op_desc.
+  std::shared_ptr<std::unordered_map<std::string, int>> in_out_idxs_;
+};
+
+class KernelContext {
+ public:
+  KernelContext(const OperatorBase* op, const std::shared_ptr<Scope>& scope,
+                const platform::DeviceContext& device_context)
+      : op_(*op), scope_(scope), device_context_(device_context) {}
+
+  const Variable* Input(int index) const {
+    return scope_->GetVariable(op_.inputs_[index]);
+  }
+
+  Variable* Output(int index) const {
+    return scope_->GetVariable(op_.outputs_[index]);
+  }
+
+  const Variable* Input(const std::string& name) const {
+    return scope_->GetVariable(op_.Input(name));
+  }
+
+  const Variable* Output(const std::string& name) const {
+    return scope_->GetVariable(op_.Output(name));
+  }
+
+  const std::vector<const Variable*> Inputs(const std::string& name) const {
+    auto names = op_.Inputs(name);
+    std::vector<const Variable*> res;
+    std::transform(
+        names.begin(), names.end(), res.begin(),
+        [this](const std::string& name) { return scope_->GetVariable(name); });
+    return res;
+  }
+
+  const std::vector<const Variable*> Outputs(const std::string& name) const {
+    auto names = op_.Outputs(name);
+    std::vector<const Variable*> res;
+    std::transform(
+        names.begin(), names.end(), res.begin(),
+        [this](const std::string& name) { return scope_->GetVariable(name); });
+    return res;
+  }
+
+  template <typename PlaceType,
+            typename DeviceType =
+                typename EigenDeviceConverter<PlaceType>::EigenDeviceType>
+  DeviceType* GetEigenDevice() const;
+
+  platform::Place GetPlace() const { return device_context_.GetPlace(); }
+
+  const OperatorBase& op_;
+  const std::shared_ptr<Scope>& scope_;
+  const platform::DeviceContext& device_context_;
 };
 
 class OpKernel {
@@ -80,24 +166,6 @@ class OpKernel {
    * device resource such as CUDA stream, cublas handle, etc. from
    * KernelContext. User should construct it before run the Operator.
    */
-  class KernelContext {
-   public:
-    KernelContext(const OperatorBase* op, const ScopePtr& scope,
-                  const platform::DeviceContext& device_context)
-        : op_(*op), scope_(scope), device_context_(device_context) {}
-
-    const Variable* Input(int index) const {
-      return scope_->GetVariable(op_.inputs_[index]);
-    }
-
-    Variable* Output(int index) const {
-      return scope_->GetVariable(op_.outputs_[index]);
-    }
-
-    const OperatorBase& op_;
-    const ScopePtr& scope_;
-    const platform::DeviceContext& device_context_;
-  };
 
   virtual void Compute(const KernelContext& context) const = 0;
 
@@ -142,8 +210,8 @@ class OperatorWithKernel : public OperatorBase {
 
   void Run(const ScopePtr& scope,
            const platform::DeviceContext& dev_ctx) const final {
-    auto& opKernel = AllOpKernels().at(Type()).at(OpKernelKey(dev_ctx));
-    opKernel->Compute(OpKernel::KernelContext(this, scope, dev_ctx));
+    auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
+    opKernel->Compute(KernelContext(this, scope, dev_ctx));
   }
 
   static std::unordered_map<std::string /* op_type */, OpKernelMap>&
@@ -151,6 +219,7 @@ class OperatorWithKernel : public OperatorBase {
     static std::unordered_map<std::string, OpKernelMap> g_all_op_kernels;
     return g_all_op_kernels;
   }
+
   void InferShape(const std::shared_ptr<Scope>& scope) const final {
     std::vector<const Tensor*> ins;
     VarNamesToTensors(scope, inputs_, &ins);

paddle/framework/operator_test.cc

@@ -19,14 +19,17 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-class OperatorTest : public OperatorBase {
+static int op_run_num = 0;
+
+class OpWithoutKernelTest : public OperatorBase {
  public:
   void Init() override { x = 1; }
   void InferShape(const ScopePtr& scope) const override {}
   void Run(const ScopePtr& scope,
            const platform::DeviceContext& dev_ctx) const override {
-    float scale = GetAttr<float>("scale");
-    ASSERT_NEAR(scale, 3.14, 1e-5);
+    op_run_num++;
+    ASSERT_EQ((int)inputs_.size(), 1);
+    ASSERT_EQ((int)outputs_.size(), 1);
     ASSERT_EQ(scope->GetVariable(inputs_[0]), nullptr);
     ASSERT_EQ(x, 1);
     ASSERT_NE(scope->GetVariable(outputs_[0]), nullptr);
@@ -36,15 +39,14 @@ class OperatorTest : public OperatorBase {
   float x = 0;
 };
 
-class OperatorTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
+class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
-  OperatorTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
+  OpeWithoutKernelTestProtoAndCheckerMaker(OpProto* proto,
+                                           OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("input", "input of test op");
     AddOutput("output", "output of test op");
-    AddAttr<float>("scale", "scale of cosine op")
-        .SetDefault(1.0)
-        .LargerThan(0.0);
+    AddAttr<float>("scale", "scale of cosine op");
     AddComment("This is test op");
   }
 };
@@ -52,8 +54,8 @@ class OperatorTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 }  // namespace framework
 }  // namespace paddle
 
-REGISTER_OP(test_operator, paddle::framework::OperatorTest,
-            paddle::framework::OperatorTestProtoAndCheckerMaker);
+REGISTER_OP(test_operator, paddle::framework::OpWithoutKernelTest,
+            paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
 
 TEST(OperatorBase, all) {
   paddle::framework::OpDesc op_desc;
@@ -63,18 +65,17 @@ TEST(OperatorBase, all) {
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
   attr->set_type(paddle::framework::AttrType::FLOAT);
-  float scale = 3.14;
-  attr->set_f(scale);
+  attr->set_f(3.14);
 
   paddle::platform::CPUDeviceContext device_context;
   auto scope = std::make_shared<paddle::framework::Scope>();
 
   paddle::framework::OperatorPtr op =
       paddle::framework::OpRegistry::CreateOp(op_desc);
-  ASSERT_EQ(op->GetAttr<float>("scale"), scale);
   scope->CreateVariable("OUT1");
+  ASSERT_EQ(paddle::framework::op_run_num, 0);
   op->Run(scope, device_context);
-  std::cout << op->DebugString() << std::endl;
+  ASSERT_EQ(paddle::framework::op_run_num, 1);
 }
 
 namespace paddle {
@@ -84,8 +85,8 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
   OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("input", "input of test op");
-    AddOutput("output", "output of test op");
+    AddInput("x", "input of test op");
+    AddOutput("y", "output of test op");
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
         .LargerThan(0.0);
@@ -93,19 +94,76 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
   }
 };
 
+static int cpu_kernel_run_num = 0;
+
 class OpWithKernelTest : public OperatorWithKernel {
  protected:
   void InferShape(const std::vector<const Tensor*>& inputs,
                   const std::vector<Tensor*>& outputs) const override {}
 };
 
+template <typename T1, typename T2>
 class CPUKernelTest : public OpKernel {
  public:
-  void Compute(const KernelContext& context) const {
-    float scale = context.op_.GetAttr<float>("scale");
-    ASSERT_NEAR(scale, 3.14, 1e-5);
+  void Compute(const KernelContext& ctx) const {
     std::cout << "this is cpu kernel" << std::endl;
-    std::cout << context.op_.DebugString() << std::endl;
+    std::cout << ctx.op_.DebugString() << std::endl;
+    cpu_kernel_run_num++;
+    ASSERT_EQ(ctx.op_.Input("x"), "IN1");
+    ASSERT_EQ(ctx.op_.Output("y"), "OUT1");
+  }
+};
+
+// multiple inputs test
+class OperatorMultiInputsTest : public OperatorBase {
+ public:
+  void Init() override { x = 1; }
+  void InferShape(const std::shared_ptr<Scope>& scope) const override {}
+  void Run(const std::shared_ptr<Scope>& scope,
+           const platform::DeviceContext& dev_ctx) const override {
+    ASSERT_EQ(scope->GetVariable(inputs_[0]), nullptr);
+    ASSERT_EQ(x, 1);
+    ASSERT_NE(scope->GetVariable(outputs_[0]), nullptr);
+    ASSERT_EQ(Input("x"), "IN1");
+    ASSERT_EQ(Input("y"), "OUT1");
+  }
+
+ public:
+  float x = 0;
+};
+
+class OpKernelTestMultiInputsProtoAndCheckerMaker
+    : public OpProtoAndCheckerMaker {
+ public:
+  OpKernelTestMultiInputsProtoAndCheckerMaker(OpProto* proto,
+                                              OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInputs("xs", "inputs of test op");
+    AddInput("k", "input of test op");
+    AddOutputs("ys", "outputs of test op");
+    AddAttr<float>("scale", "scale of cosine op")
+        .SetDefault(1.0)
+        .LargerThan(0.0);
+    AddComment("This is test op");
+  }
+};
+
+class CPUKernalMultiInputsTest : public OpKernel {
+ public:
+  void Compute(const KernelContext& ctx) const {
+    auto xs = ctx.op_.Inputs("xs");
+    ASSERT_EQ(xs.size(), 3UL);
+    ASSERT_EQ(xs[0], "x0");
+    ASSERT_EQ(xs[1], "x1");
+    ASSERT_EQ(xs[2], "x2");
+
+    auto k = ctx.op_.Input("k");
+    ASSERT_EQ(k, "k0");
+
+    auto ys = ctx.op_.Outputs("ys");
+    ASSERT_EQ(ys.size(), 2UL);
+    ASSERT_EQ(ys[0], "y0");
+    ASSERT_EQ(ys[1], "y1");
   }
 };
 
@@ -114,8 +172,10 @@ class CPUKernelTest : public OpKernel {
 
 REGISTER_OP(op_with_kernel, paddle::framework::OpWithKernelTest,
             paddle::framework::OpKernelTestProtoAndCheckerMaker);
-REGISTER_OP_CPU_KERNEL(op_with_kernel, paddle::framework::CPUKernelTest);
+REGISTER_OP_CPU_KERNEL(op_with_kernel,
+                       paddle::framework::CPUKernelTest<float, float>);
 
+// test with single input
 TEST(OpKernel, all) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("op_with_kernel");
@@ -131,5 +191,51 @@ TEST(OpKernel, all) {
 
   paddle::framework::OperatorPtr op =
       paddle::framework::OpRegistry::CreateOp(op_desc);
+  ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 0);
+  op->Run(scope, cpu_device_context);
+  ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
+}
+
+REGISTER_OP(op_multi_inputs_with_kernel, paddle::framework::OpWithKernelTest,
+            paddle::framework::OpKernelTestMultiInputsProtoAndCheckerMaker);
+REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel,
+                       paddle::framework::CPUKernalMultiInputsTest);
+
+// test with multi inputs
+TEST(OpKernel, multi_inputs) {
+  using namespace paddle::framework;
+
+  OpDesc op_desc;
+  op_desc.set_type("op_multi_inputs_with_kernel");
+  *op_desc.mutable_inputs()->Add() = "x0";
+  *op_desc.mutable_inputs()->Add() = "x1";
+  *op_desc.mutable_inputs()->Add() = "x2";
+  *op_desc.mutable_inputs()->Add() = "k0";
+  *op_desc.mutable_outputs()->Add() = "y0";
+  *op_desc.mutable_outputs()->Add() = "y1";
+  auto attr = op_desc.mutable_attrs()->Add();
+  attr->set_name("scale");
+  attr->set_type(paddle::framework::AttrType::FLOAT);
+  attr->set_f(3.14);
+
+  auto attr0 = op_desc.mutable_attrs()->Add();
+  attr0->set_name("input_format");
+  attr0->set_type(paddle::framework::AttrType::INTS);
+  auto input_format = attr0->mutable_ints();
+  input_format->Add(0);  // x0
+  input_format->Add(3);  // k
+  input_format->Add(4);  // end
+
+  auto attr1 = op_desc.mutable_attrs()->Add();
+  attr1->set_name("output_format");
+  attr1->set_type(paddle::framework::AttrType::INTS);
+  auto output_format = attr1->mutable_ints();
+  output_format->Add(0);  // y0
+  output_format->Add(2);  // y1
+
+  paddle::platform::CPUDeviceContext cpu_device_context;
+  auto scope = std::make_shared<Scope>();
+
+  OperatorPtr op(paddle::framework::OpRegistry::CreateOp(op_desc));
   op->Run(scope, cpu_device_context);
 }

paddle/framework/tensor.cc

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

paddle/framework/tensor.h

@@ -17,19 +17,34 @@ limitations under the License. */
 #include <cstdint>
 #include <cstring>
 #include <memory>
+#include <typeindex>
 #include "paddle/framework/ddim.h"
-#include "paddle/framework/enforce.h"
 #include "paddle/memory/memory.h"
+#include "paddle/platform/enforce.h"
 #include "paddle/platform/place.h"
+#include "unsupported/Eigen/CXX11/Tensor"
 
 namespace paddle {
+namespace pybind {
+namespace details {  // forward declare
+template <bool less, size_t i, typename... args>
+struct CastToPyBufferImpl;
+}  // namespace details
+}  // namespace pybind
 namespace framework {
 
 class Tensor {
- public:
-  Tensor() : numel_(0), offset_(0) {}
+  template <bool less, size_t i, typename... args>
+  friend struct paddle::pybind::details::CastToPyBufferImpl;
+
+  template <typename T, size_t D, int MajorType, typename IndexType>
+  friend struct EigenTensor;
 
-  Tensor& operator=(const Tensor& src) = delete;
+  template <typename T, int MajorType, typename IndexType>
+  friend struct EigenVector;
+
+ public:
+  Tensor() : offset_(0) {}
 
   template <typename T>
   const T* data() const {
@@ -39,21 +54,40 @@ class Tensor {
   }
 
   template <typename T>
-  T* mutable_data(DDim dims, paddle::platform::Place place) {
+  T* data() {
+    CheckDims<T>();
+    return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
+                                offset_);
+  }
+
+  template <typename T>
+  T* mutable_data(DDim dims, platform::Place place) {
     set_dims(dims);
     return mutable_data<T>(place);
   }
 
   template <typename T>
-  T* mutable_data(paddle::platform::Place place) {
-    PADDLE_ENFORCE(numel_ > 0,
-                   "Tensor::numel_ must be larger than zero to call "
+  T* mutable_data(platform::Place place) {
+    PADDLE_ENFORCE(product(dims_) > 0,
+                   "Tensor's numel must be larger than zero to call "
                    "Tensor::mutable_data. Call Tensor::set_dim first.");
     if (holder_ == nullptr ||
         !(holder_->place() ==
           place) /* some versions of boost::variant don't have operator!= */
-        || holder_->size() < numel_ * sizeof(T) + offset_) {
-      holder_.reset(new PlaceholderImpl<T>(place, numel_ * sizeof(T)));
+        || holder_->size() < product(dims_) * sizeof(T) + offset_) {
+      if (platform::is_cpu_place(place)) {
+        holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
+            boost::get<platform::CPUPlace>(place), product(dims_) * sizeof(T)));
+      } else if (platform::is_gpu_place(place)) {
+#ifdef PADDLE_ONLY_CPU
+        PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
+#else
+        holder_.reset(new PlaceholderImpl<T, platform::GPUPlace>(
+            boost::get<platform::GPUPlace>(place), product(dims_) * sizeof(T)));
+#endif
+      } else {
+        PADDLE_THROW("Unknown 'place'.");
+      }
       offset_ = 0;
     }
     return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
@@ -69,12 +103,12 @@ class Tensor {
   }
 
   template <typename T>
-  void CopyFrom(const Tensor& src, paddle::platform::Place dst_place) {
+  void CopyFrom(const Tensor& src, platform::Place dst_place) {
     PADDLE_ENFORCE(platform::is_cpu_place(src.holder_->place()) &&
                        platform::is_cpu_place(dst_place),
                    "Tensor::CopyFrom only support CPU now.");
     src.CheckDims<T>();
-    size_t size = src.numel_ * sizeof(T);
+    size_t size = product(src.dims_) * sizeof(T);
     set_dims(src.dims());
     const void* src_ptr = static_cast<const void*>(src.data<T>());
     void* dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
@@ -108,7 +142,6 @@ class Tensor {
       return;
     }
     dims_ = dims;
-    numel_ = product(dims_);
   }
 
   DDim dims() const { return dims_; }
@@ -119,52 +152,52 @@ class Tensor {
   struct Placeholder {
     virtual ~Placeholder() {}
     virtual void* ptr() const = 0;
-    virtual paddle::platform::Place place() const = 0;
+    virtual platform::Place place() const = 0;
     virtual size_t size() const = 0;
+    virtual std::type_index type() const = 0;
   };
 
-  template <typename T>
+  template <typename T, typename PlaceType>
   struct PlaceholderImpl : public Placeholder {
    private:
+    template <typename PType>
     class Deleter {
      public:
-      Deleter(platform::Place place) : place_(place) {}
-      void operator()(T* ptr) {
-        paddle::memory::Free(place_, static_cast<void*>(ptr));
-      }
+      Deleter(PType place) : place_(place) {}
+      void operator()(T* ptr) { memory::Free(place_, static_cast<void*>(ptr)); }
 
      private:
-      paddle::platform::Place place_;
+      PType place_;
     };
 
    public:
-    PlaceholderImpl(paddle::platform::Place place, size_t size)
-        : ptr_(static_cast<T*>(paddle::memory::Alloc(place, size)),
-               Deleter(place)),
+    PlaceholderImpl(PlaceType place, size_t size)
+        : ptr_(static_cast<T*>(memory::Alloc(place, size)),
+               Deleter<PlaceType>(place)),
           place_(place),
           size_(size) {}
 
     virtual void* ptr() const { return static_cast<void*>(ptr_.get()); }
     virtual size_t size() const { return size_; }
     virtual paddle::platform::Place place() const { return place_; }
+    virtual std::type_index type() const { return std::type_index(typeid(T)); }
 
-    std::unique_ptr<T, Deleter> ptr_;
-    paddle::platform::Place place_;  // record the place of ptr_.
-    size_t size_;                    // size of the memory block.
+    std::unique_ptr<T, Deleter<PlaceType>> ptr_;
+    platform::Place place_;  // record the place of ptr_.
+    size_t size_;            // size of the memory block.
   };
 
   template <typename T>
   inline void CheckDims() const {
     PADDLE_ENFORCE(holder_ != nullptr,
                    "Tenosr holds no memory. Call Tensor::mutable_data first.");
-    PADDLE_ENFORCE(holder_->size() >= numel_ * sizeof(T) + offset_,
+    PADDLE_ENFORCE(holder_->size() >= product(dims_) * sizeof(T) + offset_,
                    "Tensor's dims_ is out of bound. Call Tensor::mutable_data "
                    "first to re-allocate memory.");
   }
 
   std::shared_ptr<Placeholder> holder_;  // holds the memory block if allocated.
   DDim dims_;
-  size_t numel_;   // cache of `product(dims_)`
   size_t offset_;  // marks the begin of tensor data area.
 };
 

paddle/framework/tensor_test.cc

@@ -33,7 +33,7 @@ TEST(Tensor, DataAssert) {
   bool caught = false;
   try {
     src_tensor.data<double>();
-  } catch (paddle::framework::EnforceNotMet err) {
+  } catch (std::runtime_error& err) {
     caught = true;
     std::string msg =
         "Tenosr holds no memory. Call Tensor::mutable_data first.";
@@ -47,7 +47,7 @@ TEST(Tensor, DataAssert) {
 
 /* following tests are not available at present
    because Memory::Alloc() and Memory::Free() have not been ready.
-
+*/
 TEST(Tensor, MutableData) {
   using namespace paddle::framework;
   using namespace paddle::platform;
@@ -72,7 +72,7 @@ TEST(Tensor, MutableData) {
     p2 = src_tensor.mutable_data<float>(make_ddim({2, 2}), CPUPlace());
     EXPECT_EQ(p1, p2);
   }
-
+#ifdef __CUDACC__
   {
     Tensor src_tensor;
     float* p1 = nullptr;
@@ -94,6 +94,7 @@ TEST(Tensor, MutableData) {
     p2 = src_tensor.mutable_data<float>(make_ddim({2, 2}), GPUPlace());
     EXPECT_EQ(p1, p2);
   }
+#endif
 }
 
 TEST(Tensor, ShareDataFrom) {
@@ -106,11 +107,13 @@ TEST(Tensor, ShareDataFrom) {
     bool caught = false;
     try {
       dst_tensor.ShareDataFrom<float>(src_tensor);
-    } catch (EnforceNotMet err) {
+    } catch (std::runtime_error& err) {
       caught = true;
-      std::string msg = "Tenosr 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]);
+      std::string msg =
+          "Tenosr 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]);
       }
     }
     ASSERT_TRUE(caught);
@@ -120,6 +123,7 @@ first."; const char* what = err.what(); for (size_t i = 0; i < msg.length();
     ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
   }
 
+#ifdef __CUDACC__
   {
     Tensor src_tensor;
     Tensor dst_tensor;
@@ -127,6 +131,7 @@ first."; const char* what = err.what(); for (size_t i = 0; i < msg.length();
     dst_tensor.ShareDataFrom<int>(src_tensor);
     ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
   }
+#endif
 }
 
 TEST(Tensor, Slice) {
@@ -155,6 +160,7 @@ TEST(Tensor, Slice) {
     EXPECT_EQ(src_data_address + 3 * 4 * 1 * sizeof(int), slice_data_address);
   }
 
+#ifdef __CUDACC__
   {
     Tensor src_tensor;
     src_tensor.mutable_data<double>(make_ddim({6, 9}), GPUPlace());
@@ -176,6 +182,7 @@ TEST(Tensor, Slice) {
     EXPECT_EQ(slice_data_address, slice_mutable_data_address);
     EXPECT_EQ(src_data_address + 9 * 2 * sizeof(double), slice_data_address);
   }
+#endif
 }
 
 TEST(Tensor, CopyFrom) {
@@ -203,4 +210,3 @@ TEST(Tensor, CopyFrom) {
     EXPECT_EQ(dst_ptr[i], slice_ptr[i]);
   }
 }
-*/
\ No newline at end of file

paddle/function/CMakeLists.txt

@@ -11,7 +11,6 @@ if(WITH_GPU)
 endif()
 
 if(USE_NNPACK)
-  include(nnpack/nnpack.cmake)
   list(APPEND cpp_files nnpack/NNPACKConvOp.cpp)
   if(WITH_TESTING)
     add_unittest(NNPACKConvOpTest nnpack/NNPACKConvOpTest.cpp)
@@ -37,6 +36,7 @@ if(WITH_GPU)
     add_simple_unittest(MulOpTest)
     add_simple_unittest(CosSimOpTest)
     add_simple_unittest(RowConvOpTest)
+    add_simple_unittest(CropOpTest)
 endif()
 
 add_simple_unittest(ConvOpTest)

paddle/function/CropOp.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 "CropOp.h"
+#include "paddle/function/TensorShape.h"
+#include "paddle/math/Vector.h"
+
+namespace paddle {
+
+template <>
+void Crop<DEVICE_TYPE_CPU>(real* outputs,
+                           const real* inputs,
+                           const TensorShape inShape,
+                           const TensorShape outShape,
+                           const FuncConfig& conf) {
+  std::vector<uint32_t> crop_corner =
+      conf.get<std::vector<uint32_t>>("crop_corner");
+  int cCrop = crop_corner[1];
+  int hCrop = crop_corner[2];
+  int wCrop = crop_corner[3];
+
+  int num = inShape[0];
+  int inC = inShape[1];
+  int inH = inShape[2];
+  int inW = inShape[3];
+
+  int outC = outShape[1];
+  int outH = outShape[2];
+  int outW = outShape[3];
+
+  for (int n = 0; n < num; n++) {
+    for (int c = 0; c < outC; c++) {
+      for (int h = 0; h < outH; h++) {
+        int outoff = ((n * outC + c) * outH + h) * outW;
+        int inoff = ((n * inC + c + cCrop) * inH + h + hCrop) * inW + wCrop;
+        memcpy(outputs + outoff, inputs + inoff, outW * sizeof(real));
+      }
+    }
+  }
+}
+
+template <>
+void CropGrad<DEVICE_TYPE_CPU>(const real* inGrad,
+                               real* outGrad,
+                               const TensorShape inShape,
+                               const TensorShape outShape,
+                               const FuncConfig& conf) {
+  std::vector<uint32_t> crop_corner =
+      conf.get<std::vector<uint32_t>>("crop_corner");
+  int cCrop = crop_corner[1];
+  int hCrop = crop_corner[2];
+  int wCrop = crop_corner[3];
+
+  int num = outShape[0];
+  int outC = outShape[1];
+  int outH = outShape[2];
+  int outW = outShape[3];
+
+  int inC = inShape[1];
+  int inH = inShape[2];
+  int inW = inShape[3];
+
+  for (int n = 0; n < num; n++) {
+    for (int c = 0; c < inC; c++) {
+      for (int h = 0; h < inH; h++) {
+        int outoff = ((n * outC + c + cCrop) * outH + h + hCrop) * outW + wCrop;
+        int inoff = ((n * inC + c) * inH + h) * inW;
+        CpuVector inG = CpuVector(inW, const_cast<real*>(inGrad + inoff));
+        CpuVector outG = CpuVector(inW, outGrad + outoff);
+        outG += inG;
+      }
+    }
+  }
+}
+
+/**
+ * \brief Crop input according to the specify corner and shape.
+ *        The input and output is a 4D tensor. In CropFunc, we only
+ *        crop the 2nd to 4th dimension.
+ *
+ * Argument in this Function:
+ * \param pad_    A struct object contains the cropping corner and shape.
+ * \param inputs  A 4D tensor, only one input.
+ * \param outputs A 4D tensor, the output value after cropping.
+ *
+ * For example,
+ * Input(2,2,2,3) = [
+ *                    [ [[1,2,3], [3,4,5]],
+ *                      [[2,3,5], [1,6,7]] ],
+ *                    [ [[4,3,1], [1,8,7]],
+ *                      [[3,8,9], [2,3,5]] ]
+ *                  ] # the input shape is (2,2,2,3)
+ *
+ * pad_: if corner = (0,1,1) and crop_shape = (2,1,2)
+ * Output(2,2,1,2) = [
+ *                    [ [[4,5]],
+ *                      [[6,7]] ],
+ *                    [ [[8,7]],
+ *                      [[3,5]] ]
+ *                  ] # the input shape is (2,2,2,3)
+ */
+template <DeviceType Device>
+class CropFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override { conf_ = config; }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(1UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+    CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO);
+
+    TensorShape inShape = inputs[0].shape();
+    TensorShape outShape = outputs[0].shape();
+
+    Crop<Device>(outputs[0].data<real>(),
+                 inputs[0].data<real>(),
+                 inShape,
+                 outShape,
+                 conf_);
+  }
+
+private:
+  FuncConfig conf_;
+};
+
+/**
+ * \brief The backward propagation of cropping Function.
+ *
+ * Argument in this Function:
+ * \param crop_    The same meaning as it in CropFunc.
+ * \param inputs  The gradient with respect to the output value of CropFunc.
+ * \param outputs The gradient with respect to the input value of CropFunc.
+ */
+
+template <DeviceType Device>
+class CropGradFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override { conf_ = config; }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(1UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+    CHECK_EQ(outputs[0].getArgType(), ADD_TO);
+
+    TensorShape outShape = outputs[0].shape();
+    TensorShape inShape = inputs[0].shape();
+
+    CropGrad<Device>(inputs[0].data<real>(),
+                     outputs[0].data<real>(),
+                     inShape,
+                     outShape,
+                     conf_);
+  }
+
+private:
+  FuncConfig conf_;
+};
+
+REGISTER_TYPED_FUNC(Crop, CPU, CropFunc);
+REGISTER_TYPED_FUNC(CropGrad, CPU, CropGradFunc);
+#ifndef PADDLE_ONLY_CPU
+REGISTER_TYPED_FUNC(Crop, GPU, CropFunc);
+REGISTER_TYPED_FUNC(CropGrad, GPU, CropGradFunc);
+#endif
+
+}  // namespace paddle

paddle/function/CropOp.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "Function.h"
+
+namespace paddle {
+
+/**
+ * \brief  This funtion crops inputs according to the specify start point and
+ *shape.
+ *
+ * \param[out] outputs	save results.
+ * \param[in]  inputs	input data.
+ * \param[in]  inShape  the shape of input tensor.
+ * \param[in]  conf     the cropping config
+ */
+template <DeviceType Device>
+void Crop(real* outputs,
+          const real* inputs,
+          const TensorShape inShape,
+          const TensorShape outShape,
+          const FuncConfig& conf);
+
+/**
+ * \brief   Cropping operation backward.
+ *
+ * \param[out] inGrad	gradients of previous layer
+ * \param[in]  outGrad  output gradient
+ * \param[in]  inShape  the shape of input tensor.
+ * \param[in]  conf     the cropping config
+ */
+template <DeviceType Device>
+void CropGrad(const real* inGrad,
+              real* outGrad,
+              const TensorShape inShape,
+              const TensorShape outShape,
+              const FuncConfig& conf);
+}  // namespace paddle

paddle/function/CropOpGpu.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_base.h"
+#include "CropOp.h"
+
+namespace paddle {
+
+__global__ void KeCrop(real* outputs, const real* inputs,
+                      int inC, int inH, int inW,
+                      int cropC, int cropH, int cropW,
+                      int outC, int outH, int outW, int nthreads) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int w = idx % outW;
+    const int h = (idx / outW) % outH;
+    const int c = (idx / outW / outH) % outC;
+    const int n = idx / outW / outH / outC;
+
+    const int off = ((n * inC + c + cropC) * inH + h + cropH) * inW + cropW + w;
+    outputs[idx] = inputs[off];
+  }
+}
+
+template <>
+void Crop<DEVICE_TYPE_GPU>(real* outputs,
+                          const real* inputs,
+						  const TensorShape inShape,
+						  const TensorShape outShape,
+                          const FuncConfig& conf) {
+  std::vector<uint32_t> crop_corner = conf.get<std::vector<uint32_t>>("crop_corner");
+  int cropC = crop_corner[1];
+  int cropH = crop_corner[2];
+  int cropW = crop_corner[3];
+
+  int num = inShape[0];
+  int inC = inShape[1];
+  int inH = inShape[2];
+  int inW = inShape[3];
+
+  int outC = outShape[1];
+  int outH = outShape[2];
+  int outW = outShape[3];
+
+  size_t nth = num * outC * outH * outW;
+  int blockSize = 1024;
+  int gridSize = (nth + blockSize - 1) / blockSize;
+
+  KeCrop<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>
+    (outputs, inputs, inC, inH, inW, cropC, cropH, cropW,
+     outC, outH, outW, nth);
+  CHECK_SYNC("Crop");
+}
+
+__global__ void KeCropDiff(const real* inGrad, real* outGrad,
+                          int inC, int inH, int inW,
+                          int cropC, int cropH, int cropW,
+                          int outC, int outH, int outW, int nthreads) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int w = idx % inW;
+    const int h = (idx / inW) % inH;
+    const int c = (idx / inW / inH) % inC;
+    const int n = idx / inW / inH / inC;
+
+    const int off = ((n * outC + c + cropC) * outH + h + cropH) * outW + cropW + w;
+
+    outGrad[off] += inGrad[idx];
+  }
+}
+
+template <>
+void CropGrad<DEVICE_TYPE_GPU>(const real* inGrad,
+                              real* outGrad,
+                              const TensorShape inShape,
+                              const TensorShape outShape,
+                              const FuncConfig& conf) {
+  std::vector<uint32_t> crop_corner = conf.get<std::vector<uint32_t>>("crop_corner");
+  int cropC = crop_corner[1];
+  int cropH = crop_corner[2];
+  int cropW = crop_corner[3];
+
+  int num = outShape[0];
+  int outC = outShape[1];
+  int outH = outShape[2];
+  int outW = outShape[3];
+
+  int inC = inShape[1];
+  int inH = inShape[2];
+  int inW = inShape[3];
+
+  size_t nth = num * inC * inH * inW;
+  int blockSize = 1024;
+  int gridSize = (nth + blockSize - 1) / blockSize;
+
+  KeCropDiff <<<gridSize, blockSize, 0, STREAM_DEFAULT>>>
+    (inGrad, outGrad, inC, inH, inW, cropC, cropH, cropW,
+     outC, outH, outW, nth);
+  CHECK_SYNC("CropGrad");
+}
+
+}  // namespace paddle

paddle/function/CropOpTest.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "FunctionTest.h"
+
+namespace paddle {
+
+TEST(Crop, real) {
+  for (size_t numSamples : {5, 32}) {
+    for (size_t channels : {5, 5, 32}) {
+      for (size_t imgSizeH : {5, 33, 100}) {
+        for (size_t imgSizeW : {5, 32, 96}) {
+          VLOG(3) << " numSamples=" << numSamples << " channels=" << channels
+                  << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW;
+          for (bool test_grad : {false, true}) {
+            CpuGpuFuncCompare compare(
+                test_grad ? "CropGrad" : "Crop",
+                FuncConfig()
+                    .set<std::vector<uint32_t>>("crop_corner", {0, 1, 1, 1})
+                    .set<std::vector<uint32_t>>("crop_shape", {0, 2, 3, 3}));
+            TensorShape inDims{numSamples, channels, imgSizeH, imgSizeW};
+            TensorShape outDims{numSamples, 2, 3, 3};
+            compare.addInputs(
+                BufferArg(VALUE_TYPE_FLOAT, test_grad ? outDims : inDims));
+            compare.addOutputs(BufferArg(VALUE_TYPE_FLOAT,
+                                         test_grad ? inDims : outDims,
+                                         test_grad ? ADD_TO : ASSIGN_TO),
+                               test_grad ? ADD_TO : ASSIGN_TO);
+            compare.run();
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace paddle

paddle/function/GemmConvOp.cpp

@@ -117,8 +117,7 @@ public:
     ConvFunctionBase::init(config);
   }
 
-  virtual void check(const BufferArgs& inputs,
-                     const BufferArgs& outputs) override {
+  void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
     const TensorShape& input = inputs[0].shape();
     const TensorShape& filter = inputs[1].shape();
     const TensorShape& output = outputs[0].shape();
@@ -217,8 +216,7 @@ public:
     ConvFunctionBase::init(config);
   }
 
-  virtual void check(const BufferArgs& inputs,
-                     const BufferArgs& outputs) override {
+  void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
     const TensorShape& output = inputs[0].shape();
     const TensorShape& filter = inputs[1].shape();
     const TensorShape& input = outputs[0].shape();
@@ -311,8 +309,7 @@ public:
     ConvFunctionBase::init(config);
   }
 
-  virtual void check(const BufferArgs& inputs,
-                     const BufferArgs& outputs) override {
+  void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
     const TensorShape& output = inputs[0].shape();
     const TensorShape& input = inputs[1].shape();
     const TensorShape& filter = outputs[0].shape();

paddle/function/NaiveConvOp.cpp

@@ -90,8 +90,7 @@ public:
     ConvFunctionBase::init(config);
   }
 
-  virtual void check(const BufferArgs& inputs,
-                     const BufferArgs& outputs) override {
+  void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
     const TensorShape& input = inputs[0].shape();
     const TensorShape& filter = inputs[1].shape();
     const TensorShape& output = outputs[0].shape();

paddle/function/RowConvOpGpu.cu

@@ -32,7 +32,7 @@ __global__ void KeRowConv(real* y, const real* x,  const real* w,
   for (int i = tidy; i < context; i += blky) {
     sw[i][tidx] = gidx + tidx < width ? w[i*width + gidx + tidx] : 0.0;
   }
-  
+
   __syncthreads();
 
   for (int i = 0; i < numSeq; ++i) {
@@ -144,12 +144,15 @@ __global__ void KeRowConvBwWeight(real* dw, const real* x, const real* dy,
       int yoff = start + j;
 
       // transpose
-      sh_x[tidx][tidy] = (xoff < width && yoff < end) ? x[yoff * width + xoff] : 0.0;
-      sh_dy[tidx][tidy + context - 1] = (xoff < width && yoff < end) ? dy[yoff * width + xoff] : 0.0;
+      sh_x[tidx][tidy] = (xoff < width && yoff < end) ?
+      x[yoff * width + xoff] : 0.0;
+      sh_dy[tidx][tidy + context - 1] = (xoff < width && yoff < end) ?
+      dy[yoff * width + xoff] : 0.0;
       __syncthreads();
       if (tidy < (context - 1)) {
         yoff = yoff - context + 1;
-        sh_dy[tidx][tidy] = (xoff < width && yoff >= start) ? dy[yoff * width + xoff] : 0.0;
+        sh_dy[tidx][tidy] = (xoff < width && yoff >= start) ?
+        dy[yoff * width + xoff] : 0.0;
       }
       __syncthreads();
 
@@ -199,11 +202,13 @@ __global__ void KeRowConvBwWeight2(real* dw, const real* x, const real* dy,
       int yoff = start + j;
 
       // transpose
-      sh_x[tidx][tidy] = (xoff < width && yoff < end) ? x[yoff * width + xoff] : 0.0;
+      sh_x[tidx][tidy] = (xoff < width && yoff < end) ?
+      x[yoff * width + xoff] : 0.0;
       __syncthreads();
 
       for (int t = 0; t < context; t++) {
-        sh_dy[tidx][tidy] = (xoff < width && (yoff - t) >= start && yoff - t < end) ? dy[(yoff - t) * width + xoff] : 0.0;
+        sh_dy[tidx][tidy] = (xoff < width && (yoff - t) >= start &&
+        yoff - t < end) ? dy[(yoff - t) * width + xoff] : 0.0;
         __syncthreads();
 
         real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx];
@@ -239,7 +244,7 @@ __global__ void KeRowConvBwData(real* dx, const real* w, const real* dy,
   for (int i = tidy; i < context; i += blky) {
     sw[i][tidx] = gidx + tidx < width ? w[i*width + gidx + tidx] : 0.0;
   }
-  
+
   __syncthreads();
 
   for (int i = 0; i < numSeq; ++i) {
@@ -312,7 +317,7 @@ void RowConvGrad<DEVICE_TYPE_GPU>(const GpuMatrix& outG,
     dim3 dimBlock(32, 32);
     dim3 dimGrid(DIVUP(width, dimBlock.x), 1);
     real* dw = filterG.getData();
-    if (contextLength <= 32) { 
+    if (contextLength <= 32) {
       KeRowConvBwWeight<32, 32, 32>
         <<<dimGrid, dimBlock, 0, STREAM_DEFAULT>>>
         (dw, x, dy, starts, height, width, numSeq, contextLength);

paddle/function/nnpack/NNPACKConvOp.cpp

@@ -16,7 +16,7 @@ limitations under the License. */
 #include "paddle/function/ConvOp.h"
 
 DEFINE_bool(nnpack_allocate_outside,
-            false,
+            true,
             "Allocate and free workspace memory outside the NNPACK interface.");
 DEFINE_int32(nnpack_num_threads,
              0,
@@ -58,18 +58,10 @@ public:
     workspaceBuffer_ = nullptr;
     workspaceSize_ = 0;
 
-    threadpool_ = nullptr;
-    if (FLAGS_nnpack_num_threads) {
-      threadpool_ = pthreadpool_create(FLAGS_nnpack_num_threads);
-      VLOG(3) << "Number of threads "
-              << pthreadpool_get_threads_count(threadpool_);
-    }
+    create_nnpack_threadpool();
   }
 
   ~NNPACKConvFunction() {
-    if (threadpool_) {
-      pthreadpool_destroy(threadpool_);
-    }
     if (workspaceBuffer_) {
       free(workspaceBuffer_);
     }
@@ -225,14 +217,25 @@ public:
     }
   }
 
+  static void create_nnpack_threadpool() {
+    if (FLAGS_nnpack_num_threads && threadpool_ == nullptr) {
+      threadpool_ = pthreadpool_create(FLAGS_nnpack_num_threads);
+      VLOG(3) << "Number of threads "
+              << pthreadpool_get_threads_count(threadpool_);
+    }
+  }
+
 private:
   nnp_convolution_algorithm algorithm_;
   nnp_convolution_transform_strategy transform_strategy_;
   void* workspaceBuffer_;
   size_t workspaceSize_;
-  pthreadpool_t threadpool_;
+  static pthreadpool_t threadpool_;
 };
 
+template <DeviceType Device>
+pthreadpool_t NNPACKConvFunction<Device>::threadpool_ = nullptr;
+
 REGISTER_TYPED_FUNC(NNPACKConv, CPU, NNPACKConvFunction);
 
 }  // namespace paddle

paddle/gserver/gradientmachines/NeuralNetwork.cpp

@@ -403,7 +403,7 @@ public:
       : layerName_(layerName) {
     addEvaluator(std::move(evaluator));
   }
-  virtual void eval(const NeuralNetwork& nn) override {
+  void eval(const NeuralNetwork& nn) override {
     const LayerPtr& layer = nn.getLayer(layerName_);
     CHECK(layer) << "Nonexisted layer: " << layerName_ << " in submodel "
                  << nn.getName();

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -636,7 +636,7 @@ void lenToStarts(std::vector<int>& starts) {
   }
   starts.back() = pos;
 }
-}
+}  // namespace
 
 void RecurrentGradientMachine::calcSequenceStartPositions() {
   std::vector<int> starts(commonSeqInfo_.size() + 1);

paddle/gserver/layers/AgentLayer.cpp

@@ -124,7 +124,7 @@ void copyElements(const IVector& srcVec,
     dest[index[i]] = src[i];
   }
 }
-}
+}  // namespace
 
 void GatherAgentLayer::forwardIds(PassType passType) {
   IVectorPtr realId = realLayers_[0]->getOutputLabel();

paddle/gserver/layers/CropLayer.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 "CropLayer.h"
+#include "paddle/utils/Stat.h"
+namespace paddle {
+
+REGISTER_LAYER(crop, CropLayer);
+
+bool CropLayer::init(const LayerMap& layerMap,
+                     const ParameterMap& parameterMap) {
+  /* Initialize the basic parent class */
+  Layer::init(layerMap, parameterMap);
+  CHECK_LE(static_cast<int>(inputLayers_.size()), 2);
+  CHECK_GE(static_cast<int>(inputLayers_.size()), 1);
+  crop_axis_ = config_.axis();
+  for (int i = 0; i < config_.offset_size(); i++) {
+    crop_offsets_.push_back(config_.offset(i));
+  }
+
+  // 1. get input_0 shape
+  auto& input0_img_conf = config_.inputs(0).image_conf();
+  inDims_ = TensorShape({0,
+                         input0_img_conf.channels(),
+                         input0_img_conf.has_img_size_y()
+                             ? input0_img_conf.img_size_y()
+                             : input0_img_conf.img_size(),
+                         input0_img_conf.img_size()});
+  // 2. get target dims from config
+  if (config_.inputs_size() == 1) {
+    targetDims_ = TensorShape({config_.shape(0),
+                               config_.shape(1),
+                               config_.shape(2),
+                               config_.shape(3)});
+  } else {
+    // 2. get input_1 shape
+    auto& input1_img_conf = config_.inputs(1).image_conf();
+    targetDims_ = TensorShape({0,
+                               input1_img_conf.channels(),
+                               input1_img_conf.has_img_size_y()
+                                   ? input1_img_conf.img_size_y()
+                                   : input1_img_conf.img_size(),
+                               input1_img_conf.img_size()});
+  }
+
+  // 3. get final crop corner
+  int dimSize = 4;
+  crop_corner_ = {0, 0, 0, 0};
+  for (int i = 0; i < dimSize; i++) {
+    if (i >= crop_axis_) {
+      if (crop_offsets_.size() > 1) {
+        crop_corner_[i] = crop_offsets_[i - crop_axis_];
+      } else {
+        crop_corner_[i] = crop_offsets_[0];
+      }
+    }
+  }
+
+  outDims_ = TensorShape(4);
+
+  createFunction(
+      forward_, "Crop", FuncConfig().set("crop_corner", crop_corner_));
+  createFunction(
+      backward_, "CropGrad", FuncConfig().set("crop_corner", crop_corner_));
+
+  return true;
+}
+
+void CropLayer::setOutDims() {
+  MatrixPtr input = inputLayers_[1]->getOutputValue();
+  size_t batchSize = input->getHeight();
+  // get target dims from input_1
+  if (config_.inputs_size() == 2) {
+    targetDims_.setDim(0, batchSize);
+    int ch = config_.inputs(0).image_conf().channels();
+    if (ch != 0) targetDims_.setDim(1, ch);
+    int h = inputLayers_[1]->getOutput().getFrameHeight();
+    if (h != 0) targetDims_.setDim(2, h);
+    int w = inputLayers_[1]->getOutput().getFrameWidth();
+    if (w != 0) targetDims_.setDim(3, w);
+  }
+  // get final crop shape from target dims and crop axis
+  std::vector<uint32_t> crop_shape;
+  int dimSize = 4;
+  for (int i = 0; i < dimSize; i++) {
+    if (i >= crop_axis_) {
+      crop_shape.push_back(targetDims_[i]);
+    } else {
+      crop_shape.push_back(inDims_[i]);
+    }
+  }
+
+  outDims_.reshape(
+      {crop_shape[0], crop_shape[1], crop_shape[2], crop_shape[3]});
+  output_.setFrameHeight(crop_shape[2]);
+  output_.setFrameWidth(crop_shape[3]);
+}
+
+void CropLayer::setInDims() {
+  MatrixPtr input = inputLayers_[0]->getOutputValue();
+  size_t batchSize = input->getHeight();
+  inDims_.setDim(0, batchSize);
+  int h = inputLayers_[0]->getOutput().getFrameHeight();
+  if (h != 0) inDims_.setDim(2, h);
+  int w = inputLayers_[0]->getOutput().getFrameWidth();
+  if (w != 0) inDims_.setDim(3, w);
+}
+
+void CropLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  setInDims();
+  setOutDims();
+  int size = outDims_[1] * outDims_[2] * outDims_[3];
+  resetOutput(outDims_[0], size);
+  MatrixPtr outV = getOutputValue();
+  REGISTER_TIMER_INFO("CropForward", getName().c_str());
+
+  BufferArgs inputs;
+  BufferArgs outputs;
+  inputs.addArg(*getInputValue(0), inDims_);
+  outputs.addArg(*getOutputValue(), outDims_, ASSIGN_TO);
+  forward_[0]->calc(inputs, outputs);
+}
+
+void CropLayer::backward(const UpdateCallback& callback) {
+  (void)callback;
+  REGISTER_TIMER_INFO("CropBackward", getName().c_str());
+
+  BufferArgs inputs;
+  BufferArgs outputs;
+  inputs.addArg(*getOutputGrad(), outDims_);
+  outputs.addArg(*getInputGrad(0), inDims_, ADD_TO);
+  backward_[0]->calc(inputs, outputs);
+}
+}  // namespace paddle

paddle/gserver/layers/CropLayer.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "Layer.h"
+
+namespace paddle {
+
+/**
+ * \brief  This layer crop input according to the specify conf.
+ *         input_0: input to be cropped
+ *         input_1: optional reference input
+ *         axis: start dimension to be croped
+ *         offset: offset of cropping  in each dimension
+ *         shape: if reference input layer was not setted,
+ *                  crop input as this shape conf
+ */
+class CropLayer : public Layer {
+public:
+  explicit CropLayer(const LayerConfig& config) : Layer(config) {}
+
+  ~CropLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+
+protected:
+  void setOutDims();
+  void setInDims();
+
+  int32_t crop_axis_;
+  std::vector<uint32_t> crop_offsets_;
+  std::vector<uint32_t> crop_corner_;
+  TensorShape inDims_;
+  TensorShape targetDims_;
+  TensorShape outDims_;
+};
+}  // namespace paddle

paddle/gserver/layers/Layer.cpp

@@ -359,12 +359,11 @@ void Layer::backwardActivation() {
   /* Do error clipping */
   if (config_.error_clipping_threshold() > 0.0f) {
     if (FLAGS_log_error_clipping) {
-      CpuVector outGradVec(0, nullptr);
-      outGradVec.subVecFrom(
-          output_.grad->getData(), 0, output_.grad->getElementCnt());
-      real maxAbsGrad = outGradVec.getAbsMax();
+      VectorPtr outGradVec = Vector::create(
+          output_.grad->getData(), output_.grad->getElementCnt(), useGpu_);
+      real maxAbsGrad = outGradVec->getAbsMax();
       if (maxAbsGrad > config_.error_clipping_threshold()) {
-        real avgAbsGrad = outGradVec.getAbsSum() / outGradVec.getSize();
+        real avgAbsGrad = outGradVec->getAbsSum() / outGradVec->getSize();
         LOG(INFO) << " layer=" << config_.name() << " need clipping,"
                   << " max error=" << maxAbsGrad << " avg error=" << avgAbsGrad;
       }

paddle/gserver/tests/CMakeLists.txt

@@ -56,7 +56,7 @@ add_test(NAME test_DetectionOutput
 add_unittest_without_exec(test_ConvUnify
     test_ConvUnify.cpp
     LayerGradUtil.cpp)
-    
+
 add_test(NAME test_ConvUnify
     COMMAND test_ConvUnify)
 ################# test_BatchNorm #######################

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1802,6 +1802,34 @@ TEST(Layer, RowConvLayer) {
   }
 }
 
+TEST(Layer, CropLayer) {
+  TestConfig config;
+  // config input_0
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 0});
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ImageConfig* img = input->mutable_image_conf();
+  img->set_channels(4);
+  img->set_img_size(16);
+  config.layerConfig.set_axis(2);
+  config.layerConfig.add_offset(0);
+  config.layerConfig.add_offset(0);
+
+  // config input_1
+  config.inputDefs.push_back({INPUT_DATA, "layer_1", 128, 0});
+  input = config.layerConfig.add_inputs();
+  img = input->mutable_image_conf();
+  img->set_channels(2);
+  img->set_img_size(8);
+
+  // config crop layer
+  config.layerConfig.set_type("crop");
+  config.layerConfig.set_name("cropLayer");
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "crop", 100, false, useGpu, false);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/memory/detail/memory_block.cc

@@ -152,6 +152,6 @@ MemoryBlock* MemoryBlock::metadata() const {
       reinterpret_cast<const Metadata*>(this) - 1));
 }
 
-}  // detail
-}  // memory
-}  // paddle
+}  // namespace detail
+}  // namespace memory
+}  // namespace paddle

paddle/memory/detail/system_allocator.cc

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "paddle/memory/detail/system_allocator.h"
 #include "paddle/platform/assert.h"
-#include "paddle/platform/error.h"
+#include "paddle/platform/enforce.h"
 #include "paddle/platform/gpu_info.h"
 
 #include <stdlib.h>    // for malloc and free
@@ -128,8 +128,7 @@ void GPUAllocator::Free(void* p, size_t size, size_t index) {
   // process is terminating, in which case we don't care if
   // cudaFree succeeds.
   if (err != cudaErrorCudartUnloading) {
-    platform::throw_on_error(err,
-                             "cudaFree{Host} failed in GPUAllocator::Free.");
+    PADDLE_ENFORCE(err, "cudaFree{Host} failed in GPUAllocator::Free.");
   }
 }
 

paddle/operators/CMakeLists.txt

-if(WITH_GPU)
-    nv_library(add_op SRCS add_op.cc add_op.cu DEPS operator op_registry glog ddim)
-else()
-    cc_library(add_op SRCS add_op.cc DEPS operator op_registry glog ddim)
-endif()
+function(op_library TARGET)
+    # op_library is a function to create op library. The interface is same as
+    # cc_library. But it handle split GPU/CPU code and link some common library
+    # for ops.
+    set(cc_srcs)
+    set(cu_srcs)
+    set(op_common_deps operator op_registry)
+    set(options "")
+    set(oneValueArgs "")
+    set(multiValueArgs SRCS DEPS)
+    cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
+            "${multiValueArgs}" ${ARGN})
+
+    foreach(src ${op_library_SRCS})
+        if (${src} MATCHES ".*\\.cu$")
+            list(APPEND cu_srcs ${src})
+        elseif(${src} MATCHES ".*\\.cc$")
+            list(APPEND cc_srcs ${src})
+        else()
+            message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
+        endif()
+    endforeach()
+
+    list(LENGTH cc_srcs cc_srcs_len)
+    if (${cc_srcs_len} EQUAL 0)
+        message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
+    endif()
+
+    list(LENGTH cu_srcs cu_srcs_len)
+    list(LENGTH op_library_DEPS dep_len)
+    if (${cu_srcs_len} EQUAL 0 AND ${dep_len} EQUAL 0)
+        message(WARNING "The op library ${TARGET} not support GPU!")
+    endif()
+
+    if (WITH_GPU)
+        nv_library(${TARGET} SRCS ${cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
+                ${op_common_deps})
+    else()
+        cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${op_library_DEPS}
+                ${op_common_deps})
+    endif()
+endfunction()
+
+op_library(add_op SRCS add_op.cc add_op.cu)
 cc_test(add_op_test SRCS add_op_test.cc DEPS add_op)
+
+op_library(mul_op SRCS mul_op.cc mul_op.cu)
+op_library(rowwise_add_op SRCS rowwise_add_op.cu rowwise_add_op.cc)
+op_library(sigmoid_op SRCS sigmoid_op.cu sigmoid_op.cc)
+op_library(softmax_op SRCS softmax_op.cc softmax_op.cu)
+
+op_library(fc_op SRCS fc_op.cc DEPS mul_op rowwise_add_op sigmoid_op
+        softmax_op net)
+
+op_library(sgd_op SRCS sgd_op.cc sgd_op.cu)

paddle/operators/add_op.cc

-#include <paddle/framework/op_registry.h>
-#include <paddle/framework/tensor.h>
-#include <paddle/operators/add_op.h>
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/add_op.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/framework/tensor.h"
 
 namespace paddle {
 namespace operators {
@@ -17,8 +31,7 @@ protected:
         "Inputs/Outputs of AddOp must all be set");
     PADDLE_ENFORCE(inputs[0]->dims() == inputs[1]->dims(),
                    "Two input of Add Op's dimension must be same.");
-    // Need set dims in Tensor
-    // outputs[0]->set_dims(inputs[0]->dims())
+    outputs[0]->set_dims(inputs[0]->dims());
   }
 };
 
@@ -36,9 +49,10 @@ The equation is: Out = X + Y
 )DOC");
   }
 };
-}  // namespace op
+}  // namespace operators
 }  // namespace paddle
 
 REGISTER_OP(add_two, paddle::operators::AddOp, paddle::operators::AddOpMaker);
-REGISTER_OP_CPU_KERNEL(
-    add_two, ::paddle::operators::AddKernel<::paddle::platform::CPUPlace>);
\ No newline at end of file
+typedef paddle::operators::AddKernel<::paddle::platform::CPUPlace, float>
+    AddKernel_CPU_float;
+REGISTER_OP_CPU_KERNEL(add_two, AddKernel_CPU_float);

paddle/operators/add_op.cu

-#include <paddle/operators/add_op.h>
-#include <paddle/framework/op_registry.h>
+#include "paddle/operators/add_op.h"
+#include "paddle/framework/op_registry.h"
 
+typedef paddle::operators::AddKernel<::paddle::platform::GPUPlace, float> AddKernel_GPU_float;
 REGISTER_OP_GPU_KERNEL(add_two,
-                       paddle::operators::AddKernel<paddle::platform::GPUPlace>);
\ No newline at end of file
+                       AddKernel_GPU_float);
\ No newline at end of file

paddle/operators/add_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
 #pragma once
-#include <glog/logging.h>
-#include <paddle/framework/operator.h>
+#include "glog/logging.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/operator.h"
 
 namespace paddle {
 namespace operators {
 
-template <typename Place>
+template <typename Place, typename T>
 class AddKernel : public framework::OpKernel {
 public:
-  void Compute(const KernelContext &context) const override {
-    LOG(INFO) << "Add kernel in " << typeid(Place).name();
+  void Compute(const framework::KernelContext& context) const override {
+    auto input0 = context.Input(0)->Get<framework::Tensor>();
+    auto input1 = context.Input(1)->Get<framework::Tensor>();
+    auto* output = context.Output(0)->GetMutable<framework::Tensor>();
+
+    output->mutable_data<T>(context.GetPlace());
+
+    framework::EigenVector<T>::Flatten(*output).device(
+        *(context.GetEigenDevice<Place>())) =
+        framework::EigenVector<T>::Flatten(input0) +
+        framework::EigenVector<T>::Flatten(input1);
   }
 };
 
-}  // namespace op
+}  // namespace operators
 }  // namespace paddle

paddle/operators/add_op_test.cc

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

paddle/operators/fc_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/framework/net.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/framework/operator.h"
+
+namespace paddle {
+namespace operators {
+
+class FullyConnectedOp : public framework::PlainNet {
+public:
+  void Init() override {
+    AddOp(framework::OpRegistry::CreateOp("mul",
+                                          {
+                                              Input("X"), Input("W"),
+                                          },
+                                          {Output("before_act")},
+                                          {}));
+    auto b = Input("b");
+    if (b != framework::OperatorBase::EMPTY_VAR_NAME()) {
+      AddOp(framework::OpRegistry::CreateOp("rowwise_add",
+                                            {Output("before_act"), Input("b")},
+                                            {Output("before_act")},
+                                            {}));
+    }
+
+    auto activation = GetAttr<std::string>("activation");
+    AddOp(framework::OpRegistry::CreateOp(
+        activation, {Output("before_act")}, {Output("Y")}, {}));
+    CompleteAddOp(false);
+  }
+};
+
+class FullyConnectedOpMaker : public framework::OpProtoAndCheckerMaker {
+public:
+  FullyConnectedOpMaker(framework::OpProto *proto,
+                        framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "the input of fc operator");
+    AddInput("W", "the weight of fc operator");
+    AddInput("b", "the bias of fc operator");
+
+    AddOutput("Y", "the output of fc operator");
+    AddOutput(
+        "before_act", "the before activation output of fc operator", true);
+    AddAttr<std::string>("activation", "The activation key for fc layer")
+        .SetDefault("sigmoid")
+        .InEnum({"sigmoid", "softmax"});
+
+    //! TODO(yuyang18): Complete comment;
+    AddComment("FullyConnected Operator");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+USE_OP(mul);
+USE_OP(rowwise_add);
+USE_OP(sigmoid);
+USE_OP(softmax);
+
+REGISTER_OP(fc,
+            paddle::operators::FullyConnectedOp,
+            paddle::operators::FullyConnectedOpMaker);

paddle/operators/mul_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include <paddle/framework/op_registry.h>
+#include <paddle/framework/tensor.h>
+#include <paddle/operators/mul_op.h>
+
+namespace paddle {
+namespace operators {
+
+class MulOp : public framework::OperatorWithKernel {
+protected:
+  void InferShape(
+      const std::vector<const framework::Tensor *> &inputs,
+      const std::vector<framework::Tensor *> &outputs) const override {
+    PADDLE_ENFORCE(inputs.size() == 2, "The mul op must take two inputs");
+    auto dim0 = inputs[0]->dims();
+    auto dim1 = inputs[1]->dims();
+    PADDLE_ENFORCE(dim0.size() == 2 && dim1.size() == 2,
+                   "The input of mul op must be matrix");
+    PADDLE_ENFORCE(
+        dim0[1] == dim1[0],
+        "First matrix's width must be equal with second matrix's height.");
+    PADDLE_ENFORCE(outputs.size() == 1, "The mul op must take one output");
+    outputs[0]->set_dims({dim0[0], dim1[1]});
+  }
+};
+
+class MulOpMaker : public framework::OpProtoAndCheckerMaker {
+public:
+  MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The first input of mul op");
+    AddInput("Y", "The second input of mul op");
+    AddOutput("Out", "The output of mul op");
+    AddComment(R"DOC(
+Two Element Mul Operator.
+
+The equation is: Out = X * Y
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP(mul, paddle::operators::MulOp, paddle::operators::MulOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    mul, paddle::operators::MulKernel<paddle::platform::CPUPlace>);

paddle/operators/mul_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include <paddle/operators/mul_op.h>
+#include <paddle/framework/op_registry.h>
+
+REGISTER_OP_GPU_KERNEL(mul,
+                       paddle::operators::MulKernel<paddle::platform
+                       ::GPUPlace>);
\ No newline at end of file

paddle/operators/mul_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+
+#include <glog/logging.h>
+#include <paddle/framework/operator.h>
+
+namespace paddle {
+namespace operators {
+
+template <typename Place>
+class MulKernel : public framework::OpKernel {
+public:
+  void Compute(const framework::KernelContext &context) const override {
+    LOG(INFO) << "Mul kernel in " << typeid(Place).name();
+  }
+};
+}  // namespace operators
+}  // namespace paddle

paddle/operators/rowwise_add_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include <paddle/framework/op_registry.h>
+#include <paddle/operators/rowwise_add_op.h>
+namespace paddle {
+namespace operators {
+
+class RowWiseAddOp : public framework::OperatorWithKernel {
+protected:
+  void InferShape(
+      const std::vector<const framework::Tensor *> &inputs,
+      const std::vector<framework::Tensor *> &outputs) const override {
+    PADDLE_ENFORCE(inputs.size() == 2UL, "Two inputs is needed by rowwise add");
+    auto dim0 = inputs[0]->dims();
+    auto dim1 = inputs[1]->dims();
+
+    PADDLE_ENFORCE(dim0.size() == 2, "Input 0 must be matrix");
+    PADDLE_ENFORCE(dim1.size() == 1, "The second input must be vector");
+    PADDLE_ENFORCE(dim0[1] == dim1[0], "The width of two input must be same");
+    PADDLE_ENFORCE(outputs.size() == 1, "The output size must be 1");
+    outputs[0]->set_dims(inputs[0]->dims());
+  }
+};
+
+class RowWiseAddOpMaker : public framework::OpProtoAndCheckerMaker {
+public:
+  RowWiseAddOpMaker(framework::OpProto *proto,
+                    framework::OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The left input of row-wise add op, must be matrix");
+    AddInput("b", "The right input of row-wise add op, must be vector");
+    AddOutput("Out", "The output of row-wise add op");
+    AddComment(R"DOC(Row-wise Add operator
+
+for i in xrange(X.shape[0]):
+  Out = X[i] + b
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP(rowwise_add,
+            paddle::operators::RowWiseAddOp,
+            paddle::operators::RowWiseAddOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    rowwise_add,
+    paddle::operators::RowWiseAddKernel<paddle::platform::CPUPlace>);

paddle/operators/rowwise_add_op.cu

+#include <paddle/framework/op_registry.h>
+#include <paddle/operators/rowwise_add_op.h>
+
+REGISTER_OP_GPU_KERNEL(
+    rowwise_add,
+    paddle::operators::RowWiseAddKernel<paddle::platform ::GPUPlace>);

paddle/operators/rowwise_add_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+#include <glog/logging.h>
+#include <paddle/framework/operator.h>
+
+namespace paddle {
+namespace operators {
+
+template <typename Place>
+class RowWiseAddKernel : public framework::OpKernel {
+public:
+  void Compute(const framework::KernelContext &context) const override {
+    LOG(INFO) << "RowWiseAdd kernel in " << typeid(Place).name();
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/sgd_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/sgd_op.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/framework/tensor.h"
+
+namespace paddle {
+namespace operators {
+
+class SGDOp : public framework::OperatorWithKernel {
+protected:
+  void InferShape(
+      const std::vector<const framework::Tensor *> &inputs,
+      const std::vector<framework::Tensor *> &outputs) const override {
+    PADDLE_ENFORCE(inputs.size() == 2, "Input size of SGDOp must be two");
+    PADDLE_ENFORCE(outputs.size() == 1, "Output size of SGDOp must be one");
+    PADDLE_ENFORCE(inputs[0] != nullptr, "inputs[0] mast be set");
+    PADDLE_ENFORCE(inputs[1] != nullptr, "inputs[1] mast be set");
+    PADDLE_ENFORCE(outputs[0] != nullptr, "outputs[0] mast be set");
+    PADDLE_ENFORCE(inputs[0]->dims() == inputs[1]->dims(),
+                   "Two input of SGD Op's dimension must be same.");
+    outputs[0]->set_dims(inputs[0]->dims());
+  }
+};
+
+class SGDOpMaker : public framework::OpProtoAndCheckerMaker {
+public:
+  SGDOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("param", "input parameter");
+    AddInput("grad", "input gradient");
+    AddOutput("param_out", "output parameter");
+    AddAttr<float>("learning_rate", "learning rate of sgd");
+    AddComment(R"DOC(
+
+Simplest sgd algorithm.
+
+param_out = param - learning_rate * grad;
+
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP(sgd, paddle::operators::SGDOp, paddle::operators::SGDOpMaker);
+typedef paddle::operators::SGDOpKernel<::paddle::platform::CPUPlace, float>
+    SGDOpKernel_CPU_float;
+REGISTER_OP_CPU_KERNEL(sgd, SGDOpKernel_CPU_float);

paddle/operators/sgd_op.cu

+#include "paddle/operators/sgd_op.h"
+#include "paddle/framework/op_registry.h"
+
+typedef paddle::operators::SGDOpKernel<::paddle::platform::GPUPlace, float> SGDOpKernel_GPU_float;
+REGISTER_OP_GPU_KERNEL(sgd, SGDOpKernel_GPU_float);
\ No newline at end of file

paddle/operators/sgd_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "glog/logging.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/operator.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class SGDOpKernel : public framework::OpKernel {
+public:
+  void Compute(const framework::KernelContext& ctx) const override {
+    auto param = ctx.Input("param")->Get<framework::Tensor>();
+    auto grad = ctx.Input("grad")->Get<framework::Tensor>();
+    auto* param_out = ctx.Output(0)->GetMutable<framework::Tensor>();
+    float lr = ctx.op_.GetAttr<float>("learning_rate");
+
+    param_out->mutable_data<T>(ctx.GetPlace());
+
+    framework::EigenVector<T>::Flatten(*param_out)
+        .device(*(ctx.GetEigenDevice<Place>())) =
+        framework::EigenVector<T>::Flatten(param) -
+        lr * framework::EigenVector<T>::Flatten(grad);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/sgd_op_test.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include <paddle/framework/op_registry.h>
+USE_OP(sgd);
+TEST(SGDOp, GetOpProto) {
+  auto& protos = paddle::framework::OpRegistry::protos();
+  auto it = protos.find("sgd");
+  ASSERT_NE(it, protos.end());
+}

paddle/operators/sigmoid_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include <paddle/framework/op_registry.h>
+#include <paddle/operators/sigmoid_op.h>
+namespace paddle {
+namespace operators {
+
+class SigmoidOp : public framework::OperatorWithKernel {
+protected:
+  void InferShape(
+      const std::vector<const framework::Tensor *> &inputs,
+      const std::vector<framework::Tensor *> &outputs) const override {
+    PADDLE_ENFORCE(inputs.size() == 1, "Sigmoid Op only have one input");
+    PADDLE_ENFORCE(outputs.size() == 1, "Sigmoid Op only have one output");
+    outputs[0]->set_dims(inputs[0]->dims());
+  }
+};
+
+class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
+public:
+  SigmoidOpMaker(framework::OpProto *proto,
+                 framework::OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "sigmoid input");
+    AddInput("Y", "sigmoid output");
+    AddComment("Sigmoid function");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP(sigmoid,
+            paddle::operators::SigmoidOp,
+            paddle::operators::SigmoidOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    sigmoid, paddle::operators::SigmoidKernel<paddle::platform::CPUPlace>);

paddle/operators/sigmoid_op.cu

+#include <paddle/operators/sigmoid_op.h>
+#include <paddle/framework/op_registry.h>
+
+REGISTER_OP_GPU_KERNEL(
+    sigmoid, paddle::operators::SigmoidKernel<paddle::platform::GPUPlace>);

paddle/operators/sigmoid_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+
+#include <glog/logging.h>
+#include <paddle/framework/operator.h>
+
+namespace paddle {
+namespace operators {
+
+template <typename Place>
+class SigmoidKernel : public framework::OpKernel {
+public:
+  void Compute(const framework::KernelContext &context) const override {
+    LOG(INFO) << "Sigmoid kernel in " << typeid(Place).name();
+  }
+};
+}  // namespace operators
+}  // namespace paddle

paddle/operators/softmax_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+#include <paddle/framework/op_registry.h>
+#include <paddle/operators/softmax_op.h>
+
+namespace paddle {
+namespace operators {
+
+class SoftmaxOp : public framework::OperatorWithKernel {
+protected:
+  void InferShape(
+      const std::vector<const framework::Tensor *> &inputs,
+      const std::vector<framework::Tensor *> &outputs) const override {
+    PADDLE_ENFORCE(inputs.size() == 1, "Only one input is need for softmax");
+    PADDLE_ENFORCE(outputs.size() == 1, "Only one output is need for softmax");
+
+    outputs[0]->set_dims(inputs[0]->dims());
+  }
+};
+
+class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
+public:
+  SoftmaxOpMaker(framework::OpProto *proto,
+                 framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "input of softmax");
+    AddOutput("Y", "output of softmax");
+    AddComment("Softmax Op");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker);
+REGISTER_OP_CPU_KERNEL(softmax, ops::SoftmaxKernel<paddle::platform::CPUPlace>);

paddle/operators/softmax_op.cu

+#include <paddle/framework/op_registry.h>
+#include <paddle/operators/softmax_op.h>
+
+REGISTER_OP_GPU_KERNEL(
+    softmax, paddle::operators::SoftmaxKernel<paddle::platform::GPUPlace>);

paddle/operators/softmax_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+
+#include <glog/logging.h>
+#include <paddle/framework/operator.h>
+
+namespace paddle {
+namespace operators {
+
+template <typename Place>
+class SoftmaxKernel : public framework::OpKernel {
+public:
+  void Compute(const framework::KernelContext &context) const override {
+    LOG(INFO) << "Softmax kernel in " << typeid(Place).name();
+  }
+};
+}  // namespace operators
+}  // namespace paddle

paddle/optimizer/parameter_optimizer_test.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 "parameter_optimizer.h"
 #include <cmath>
 #include <map>
@@ -5,21 +21,18 @@
 #include "gtest/gtest.h"
 #include "lr_policy.h"
 
-using namespace paddle;
-using namespace paddle::optimizer;
-
-Tensor* FillTensor(size_t size) {
-  Tensor* param = new Tensor(size);
-  Tensor& p = *param;
+paddle::optimizer::Tensor* FillTensor(size_t size) {
+  paddle::optimizer::Tensor* param = new paddle::optimizer::Tensor(size);
+  paddle::optimizer::Tensor& p = *param;
   for (size_t i = 0; i < p.size(); ++i) {
     p[i] = (float)rand() / (float)RAND_MAX;
   }
   return param;
 }
 
-Tensor* FixedTensor(size_t size) {
-  Tensor* param = new Tensor(size);
-  Tensor& p = *param;
+paddle::optimizer::Tensor* FixedTensor(size_t size) {
+  paddle::optimizer::Tensor* param = new paddle::optimizer::Tensor(size);
+  paddle::optimizer::Tensor& p = *param;
   for (size_t i = 0; i < p.size(); ++i) {
     p[i] = i;
   }
@@ -28,7 +41,8 @@ Tensor* FixedTensor(size_t size) {
 
 class OptimizerTest : public testing::Test {
 public:
-  // init tensor shape
+  virtual ~OptimizerTest() {}
+  // init paddle::optimizer::Tensor shape
   const size_t kSize = 5;
 
   virtual void SetUp() {
@@ -38,34 +52,36 @@ public:
   virtual void TearDown() {}
 
   void CreateSGD() {
-    Tensor* parameter = FixedTensor(kSize);
-    config_.set_optimizer(OptimizerConfig::SGD);
+    paddle::optimizer::Tensor* parameter = FixedTensor(kSize);
+    config_.set_optimizer(paddle::OptimizerConfig::SGD);
     config_.mutable_sgd()->set_momentum(0.0);
     config_.mutable_sgd()->set_decay(0.0);
     config_.mutable_sgd()->set_nesterov(false);
-    config_.set_lr_policy(OptimizerConfig::Const);
+    config_.set_lr_policy(paddle::OptimizerConfig::Const);
     config_.mutable_const_lr()->set_learning_rate(0.1);
     std::string str = config_.SerializeAsString();
-    ParameterOptimizer* opt = ParameterOptimizer::Create(str, parameter);
+    paddle::optimizer::ParameterOptimizer* opt =
+        paddle::optimizer::ParameterOptimizer::Create(str, parameter);
     opts_.push_back(opt);
   }
 
   void CreateAdam() {
-    Tensor* parameter = FixedTensor(kSize);
-    config_.set_optimizer(OptimizerConfig::Adam);
+    paddle::optimizer::Tensor* parameter = FixedTensor(kSize);
+    config_.set_optimizer(paddle::OptimizerConfig::Adam);
     config_.mutable_adam()->set_beta_1(0.9);
     config_.mutable_adam()->set_beta_2(0.1);
     config_.mutable_adam()->set_epsilon(1e-3);
     config_.mutable_adam()->set_decay(0.0);
-    config_.set_lr_policy(OptimizerConfig::Const);
+    config_.set_lr_policy(paddle::OptimizerConfig::Const);
     config_.mutable_const_lr()->set_learning_rate(0.1);
     std::string str = config_.SerializeAsString();
-    ParameterOptimizer* opt = ParameterOptimizer::Create(str, parameter);
+    paddle::optimizer::ParameterOptimizer* opt =
+        paddle::optimizer::ParameterOptimizer::Create(str, parameter);
     opts_.push_back(opt);
   }
 
   void TestGetWeight() {
-    Tensor* p = FixedTensor(kSize);
+    paddle::optimizer::Tensor* p = FixedTensor(kSize);
     for (size_t i = 0; i < opts_.size(); ++i) {
       int s = 0;
       float* newp = (float*)opts_[i]->get_weight(&s);
@@ -76,7 +92,7 @@ public:
   }
 
   void TestUpdate() {
-    Tensor* g = FixedTensor(kSize);
+    paddle::optimizer::Tensor* g = FixedTensor(kSize);
     for (size_t i = 0; i < opts_.size(); ++i) {
       opts_[i]->Update(g);
     }
@@ -91,8 +107,8 @@ public:
   }
 
 private:
-  std::vector<ParameterOptimizer*> opts_;
-  OptimizerConfig config_;
+  std::vector<paddle::optimizer::ParameterOptimizer*> opts_;
+  paddle::OptimizerConfig config_;
 };
 
 TEST_F(OptimizerTest, TestGetWeight) { TestGetWeight(); }

paddle/optimizer/serialization_test.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 "serialization.h"
 #include "gtest/gtest.h"
 
-using namespace paddle;
-using namespace paddle::optimizer;
-
 TEST(TensorToProto, Case1) {
-  Tensor t(3), t1(3);
+  paddle::optimizer::Tensor t(3), t1(3);
   for (size_t i = 0; i < t.size(); ++i) {
     t[i] = i;
     t1[i] = 0;
   }
 
-  TensorProto proto;
-  TensorToProto(t, &proto);
-  ProtoToTensor(proto, &t1);
+  paddle::TensorProto proto;
+  paddle::optimizer::TensorToProto(t, &proto);
+  paddle::optimizer::ProtoToTensor(proto, &t1);
   for (size_t i = 0; i < t1.size(); ++i) {
     EXPECT_EQ(t1[i], t[i]);
   }

paddle/platform/CMakeLists.txt

@@ -8,6 +8,8 @@ cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
 
 add_subdirectory(dynload)
 
+cc_test(enforce_test SRCS enforce_test.cc)
+
 IF(WITH_GPU)
     set(GPU_CTX_DEPS dynload_cuda dynamic_loader)
 ELSE()

paddle/platform/cpu_info.cc

@@ -22,7 +22,6 @@ limitations under the License. */
 #endif
 
 #include "gflags/gflags.h"
-#include "paddle/platform/error.h"
 
 DEFINE_double(fraction_of_cpu_memory_to_use, 1,
               "Default use 100% of CPU memory for PaddlePaddle,"
@@ -41,8 +40,8 @@ inline size_t CpuTotalPhysicalMemory() {
   if (sysctl(mib, 2, &size, &len, NULL, 0) == 0) return (size_t)size;
   return 0L;
 #else
-  long pages = sysconf(_SC_PHYS_PAGES);
-  long page_size = sysconf(_SC_PAGE_SIZE);
+  int64_t pages = sysconf(_SC_PHYS_PAGES);
+  int64_t page_size = sysconf(_SC_PAGE_SIZE);
   return pages * page_size;
 #endif
 }

paddle/platform/device_context.cc

@@ -15,14 +15,15 @@ namespace paddle {
 namespace platform {
 
 template <>
-Eigen::DefaultDevice* DeviceContext::get_eigen_device<Eigen::DefaultDevice>() {
-  return reinterpret_cast<CPUDeviceContext*>(this)->eigen_device();
+Eigen::DefaultDevice* DeviceContext::get_eigen_device<Eigen::DefaultDevice>()
+    const {
+  return reinterpret_cast<const CPUDeviceContext*>(this)->eigen_device();
 }
 
 #ifndef PADDLE_ONLY_CPU
 template <>
-Eigen::GpuDevice* DeviceContext::get_eigen_device<Eigen::GpuDevice>() {
-  return reinterpret_cast<CUDADeviceContext*>(this)->eigen_device();
+Eigen::GpuDevice* DeviceContext::get_eigen_device<Eigen::GpuDevice>() const {
+  return reinterpret_cast<const CUDADeviceContext*>(this)->eigen_device();
 }
 #endif
 

paddle/platform/device_context.h

@@ -11,18 +11,19 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/framework/enforce.h"
+#include "paddle/platform/enforce.h"
+#include "paddle/platform/place.h"
+
 #ifndef PADDLE_ONLY_CPU
 #include "paddle/platform/dynload/cublas.h"
 #include "paddle/platform/dynload/cudnn.h"
 #include "paddle/platform/dynload/curand.h"
-#include "paddle/platform/error.h"
 #include "paddle/platform/gpu_info.h"
 #define EIGEN_USE_GPU
 #endif
-#include <paddle/platform/place.h>
 #include <memory>
-#include <unsupported/Eigen/CXX11/Tensor>
+#include "paddle/platform/place.h"
+#include "unsupported/Eigen/CXX11/Tensor"
 
 namespace paddle {
 namespace platform {
@@ -33,17 +34,14 @@ class DeviceContext {
   virtual Place GetPlace() const = 0;
 
   template <typename DeviceType>
-  DeviceType* get_eigen_device();
+  DeviceType* get_eigen_device() const;
 };
 
 class CPUDeviceContext : public DeviceContext {
  public:
-  Eigen::DefaultDevice* eigen_device() {
-    if (!eigen_device_) {
-      eigen_device_.reset(new Eigen::DefaultDevice());
-    }
-    return eigen_device_.get();
-  }
+  CPUDeviceContext() { eigen_device_.reset(new Eigen::DefaultDevice()); }
+
+  Eigen::DefaultDevice* eigen_device() const { return eigen_device_.get(); }
 
   Place GetPlace() const override {
     Place retv = CPUPlace();
@@ -74,8 +72,7 @@ class CUDADeviceContext : public DeviceContext {
  public:
   explicit CUDADeviceContext(const GPUPlace gpu_place) : gpu_place_(gpu_place) {
     GPUPlaceGuard guard(gpu_place_);
-    paddle::platform::throw_on_error(cudaStreamCreate(&stream_),
-                                     "cudaStreamCreate failed");
+    PADDLE_ENFORCE(cudaStreamCreate(&stream_), "cudaStreamCreate failed");
     eigen_stream_.reset(new Eigen::CudaStreamDevice(&stream_));
     eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
   }
@@ -86,23 +83,22 @@ class CUDADeviceContext : public DeviceContext {
   }
 
   void Wait() {
-    paddle::platform::throw_on_error(cudaStreamSynchronize(stream_),
-                                     "cudaStreamSynchronize failed");
+    PADDLE_ENFORCE(cudaStreamSynchronize(stream_),
+                   "cudaStreamSynchronize failed");
   }
 
   cudaStream_t stream() { return stream_; }
 
-  Eigen::GpuDevice* eigen_device() { return eigen_device_.get(); }
+  Eigen::GpuDevice* eigen_device() const { return eigen_device_.get(); }
 
   cublasHandle_t cublas_handle() {
     if (!blas_handle_) {
       GPUPlaceGuard guard(gpu_place_);
-      PADDLE_ENFORCE(paddle::platform::dynload::cublasCreate(&blas_handle_) ==
-                         CUBLAS_STATUS_SUCCESS,
+      PADDLE_ENFORCE(paddle::platform::dynload::cublasCreate(&blas_handle_),
                      "cublasCreate failed");
-      PADDLE_ENFORCE(paddle::platform::dynload::cublasSetStream(
-                         blas_handle_, stream_) == CUBLAS_STATUS_SUCCESS,
-                     "cublasSetStream failed");
+      PADDLE_ENFORCE(
+          paddle::platform::dynload::cublasSetStream(blas_handle_, stream_),
+          "cublasSetStream failed");
     }
     return blas_handle_;
   }
@@ -110,12 +106,11 @@ class CUDADeviceContext : public DeviceContext {
   cudnnHandle_t cudnn_handle() {
     if (!dnn_handle_) {
       GPUPlaceGuard guard(gpu_place_);
-      PADDLE_ENFORCE(paddle::platform::dynload::cudnnCreate(&dnn_handle_) ==
-                         CUDNN_STATUS_SUCCESS,
+      PADDLE_ENFORCE(paddle::platform::dynload::cudnnCreate(&dnn_handle_),
                      "cudnnCreate failed");
-      PADDLE_ENFORCE(paddle::platform::dynload::cudnnSetStream(
-                         dnn_handle_, stream_) == CUDNN_STATUS_SUCCESS,
-                     "cudnnSetStream failed");
+      PADDLE_ENFORCE(
+          paddle::platform::dynload::cudnnSetStream(dnn_handle_, stream_),
+          "cudnnSetStream failed");
     }
     return dnn_handle_;
   }
@@ -124,16 +119,15 @@ class CUDADeviceContext : public DeviceContext {
     if (!rand_generator_) {
       GPUPlaceGuard guard(gpu_place_);
       PADDLE_ENFORCE(paddle::platform::dynload::curandCreateGenerator(
-                         &rand_generator_, CURAND_RNG_PSEUDO_DEFAULT) ==
-                         CURAND_STATUS_SUCCESS,
+                         &rand_generator_, CURAND_RNG_PSEUDO_DEFAULT),
                      "curandCreateGenerator failed");
       PADDLE_ENFORCE(
           paddle::platform::dynload::curandSetPseudoRandomGeneratorSeed(
-              rand_generator_, random_seed_) == CURAND_STATUS_SUCCESS,
+              rand_generator_, random_seed_),
           "curandSetPseudoRandomGeneratorSeed failed");
-      PADDLE_ENFORCE(paddle::platform::dynload::curandSetStream(
-                         rand_generator_, stream_) == CURAND_STATUS_SUCCESS,
-                     "curandSetStream failed");
+      PADDLE_ENFORCE(
+          paddle::platform::dynload::curandSetStream(rand_generator_, stream_),
+          "curandSetStream failed");
     }
     return rand_generator_;
   }
@@ -141,26 +135,23 @@ class CUDADeviceContext : public DeviceContext {
   ~CUDADeviceContext() {
     Wait();
     if (blas_handle_) {
-      PADDLE_ENFORCE(paddle::platform::dynload::cublasDestroy(blas_handle_) ==
-                         CUBLAS_STATUS_SUCCESS,
+      PADDLE_ENFORCE(paddle::platform::dynload::cublasDestroy(blas_handle_),
                      "cublasDestroy failed");
     }
 
     if (dnn_handle_) {
-      PADDLE_ENFORCE(paddle::platform::dynload::cudnnDestroy(dnn_handle_) ==
-                         CUDNN_STATUS_SUCCESS,
+      PADDLE_ENFORCE(paddle::platform::dynload::cudnnDestroy(dnn_handle_),
                      "cudnnDestroy failed");
     }
 
     if (rand_generator_) {
-      PADDLE_ENFORCE(paddle::platform::dynload::curandDestroyGenerator(
-                         rand_generator_) == CURAND_STATUS_SUCCESS,
-                     "curandDestroyGenerator failed");
+      PADDLE_ENFORCE(
+          paddle::platform::dynload::curandDestroyGenerator(rand_generator_),
+          "curandDestroyGenerator failed");
     }
     eigen_stream_.reset();
     eigen_device_.reset();
-    paddle::platform::throw_on_error(cudaStreamDestroy(stream_),
-                                     "cudaStreamDestroy failed");
+    PADDLE_ENFORCE(cudaStreamDestroy(stream_), "cudaStreamDestroy failed");
   }
 
  private:

paddle/platform/dynload/dynamic_loader.cc

@@ -19,7 +19,7 @@ limitations under the License. */
 #include <string>
 #include "gflags/gflags.h"
 #include "glog/logging.h"
-#include "paddle/framework/enforce.h"
+#include "paddle/platform/enforce.h"
 
 DEFINE_string(cudnn_dir, "",
               "Specify path for loading libcudnn.so. For instance, "

paddle/platform/enforce.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <paddle/string/printf.h>
+#include <sstream>
+#include <stdexcept>
+#include <string>
+
+#ifndef PADDLE_ONLY_CPU
+
+#include "paddle/platform/dynload/cublas.h"
+#include "paddle/platform/dynload/cudnn.h"
+#include "paddle/platform/dynload/curand.h"
+
+#include <cublas_v2.h>
+#include <cudnn.h>
+#include <curand.h>
+#include <thrust/system/cuda/error.h>
+#include <thrust/system_error.h>
+
+#endif  // PADDLE_ONLY_CPU
+
+namespace paddle {
+namespace platform {
+
+// Because most enforce conditions would evaluate to true, we can use
+// __builtin_expect to instruct the C++ compiler to generate code that
+// always forces branch prediction of true.
+// This generates faster binary code. __builtin_expect is since C++11.
+// For more details, please check https://stackoverflow.com/a/43870188/724872.
+#define UNLIKELY(condition) __builtin_expect(static_cast<bool>(condition), 0)
+
+template <typename T>
+inline void throw_on_error(T e) {
+  throw_on_error(e, "");
+}
+
+template <typename... Args>
+inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
+    int stat, const Args&... args) {
+  if (UNLIKELY(!(stat))) {
+    throw std::runtime_error(
+        string::Sprintf(args...) +
+        string::Sprintf(" at [%s:%s];", __FILE__, __LINE__));
+  }
+}
+
+#ifndef PADDLE_ONLY_CPU
+
+template <typename... Args>
+inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
+    cudaError_t e, const Args&... args) {
+  if (UNLIKELY(e)) {
+    // clang-format off
+    throw thrust::system_error(
+        e, thrust::cuda_category(),
+        string::Sprintf(args...) +
+        string::Sprintf(" at [%s:%s];", __FILE__, __LINE__));
+    // clang-format on
+  }
+}
+
+template <typename... Args>
+inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
+    curandStatus_t stat, const Args&... args) {
+  if (stat != CURAND_STATUS_SUCCESS) {
+    // clang-format off
+    throw thrust::system_error(
+        cudaErrorLaunchFailure, thrust::cuda_category(),
+        string::Sprintf(args...) +
+        string::Sprintf(" at [%s:%s];", __FILE__, __LINE__));
+    // clang-format on
+  }
+}
+
+template <typename... Args>
+inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
+    cudnnStatus_t stat, const Args&... args) {
+  if (stat == CUDNN_STATUS_SUCCESS) {
+    return;
+  } else {
+    // clang-format off
+    throw std::runtime_error(
+        platform::dynload::cudnnGetErrorString(stat) +
+        string::Sprintf(args...) +
+        string::Sprintf(" at [%s:%s];", __FILE__, __LINE__));
+    // clang-format on
+  }
+}
+
+template <typename... Args>
+inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
+    cublasStatus_t stat, const Args&... args) {
+  std::string err;
+  if (stat == CUBLAS_STATUS_SUCCESS) {
+    return;
+  } else if (stat == CUBLAS_STATUS_NOT_INITIALIZED) {
+    err = "CUBLAS: not initialized, ";
+  } else if (stat == CUBLAS_STATUS_ALLOC_FAILED) {
+    err = "CUBLAS: alloc failed, ";
+  } else if (stat == CUBLAS_STATUS_INVALID_VALUE) {
+    err = "CUBLAS: invalid value, ";
+  } else if (stat == CUBLAS_STATUS_ARCH_MISMATCH) {
+    err = "CUBLAS: arch mismatch, ";
+  } else if (stat == CUBLAS_STATUS_MAPPING_ERROR) {
+    err = "CUBLAS: mapping error, ";
+  } else if (stat == CUBLAS_STATUS_EXECUTION_FAILED) {
+    err = "CUBLAS: execution failed, ";
+  } else if (stat == CUBLAS_STATUS_INTERNAL_ERROR) {
+    err = "CUBLAS: internal error, ";
+  } else if (stat == CUBLAS_STATUS_NOT_SUPPORTED) {
+    err = "CUBLAS: not supported, ";
+  } else if (stat == CUBLAS_STATUS_LICENSE_ERROR) {
+    err = "CUBLAS: license error, ";
+  }
+  throw std::runtime_error(err + string::Sprintf(args...) +
+                           string::Sprintf(" at [%s:%s];", __FILE__, __LINE__));
+}
+
+#endif  // PADDLE_ONLY_CPU
+
+#define PADDLE_THROW(...)                                     \
+  do {                                                        \
+    throw std::runtime_error(                                 \
+        string::Sprintf(__VA_ARGS__) +                        \
+        string::Sprintf(" at [%s:%s];", __FILE__, __LINE__)); \
+  } while (0)
+
+#define PADDLE_ENFORCE(...)                          \
+  do {                                               \
+    ::paddle::platform::throw_on_error(__VA_ARGS__); \
+  } while (0)
+
+}  // namespace platform
+}  // namespace paddle

paddle/framework/enforce_test.cc → paddle/platform/enforce_test.cc

@@ -9,8 +9,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include <gtest/gtest.h>
-#include <paddle/framework/enforce.h>
+#include "paddle/platform/enforce.h"
+#include "gtest/gtest.h"
 
 TEST(ENFORCE, OK) {
   PADDLE_ENFORCE(true, "Enforce is ok %d now %f", 123, 0.345);
@@ -23,13 +23,14 @@ TEST(ENFORCE, FAILED) {
   bool in_catch = false;
   try {
     PADDLE_ENFORCE(false, "Enforce is not ok %d at all", 123);
-  } catch (paddle::framework::EnforceNotMet err) {
+  } catch (const std::runtime_error& error) {
+    // your error handling code here
     in_catch = true;
     std::string msg = "Enforce is not ok 123 at all";
-    const char* what = err.what();
+    const char* what = error.what();
     for (size_t i = 0; i < msg.length(); ++i) {
       ASSERT_EQ(what[i], msg[i]);
     }
   }
   ASSERT_TRUE(in_catch);
-}
\ No newline at end of file
+}

paddle/platform/error.h

-#pragma once
-
-#include <sstream>
-#include <stdexcept>
-#include <string>
-
-#ifndef PADDLE_ONLY_CPU
-
-#include <cublas_v2.h>
-#include <cudnn.h>
-#include <curand.h>
-#include <thrust/system/cuda/error.h>
-#include <thrust/system_error.h>
-
-#endif  // PADDLE_ONLY_CPU
-
-namespace paddle {
-namespace platform {
-
-#ifndef PADDLE_ONLY_CPU
-
-inline void throw_on_error(cudaError_t e, const char* message) {
-  if (e) {
-    throw thrust::system_error(e, thrust::cuda_category(), message);
-  }
-}
-
-inline void throw_on_error(curandStatus_t stat, const char* message) {
-  if (stat != CURAND_STATUS_SUCCESS) {
-    throw thrust::system_error(cudaErrorLaunchFailure, thrust::cuda_category(),
-                               message);
-  }
-}
-
-inline void throw_on_error(cudnnStatus_t stat, const char* message) {
-  std::stringstream ss;
-  if (stat == CUDNN_STATUS_SUCCESS) {
-    return;
-  } else {
-    ss << cudnnGetErrorString(stat);
-    ss << ", " << message;
-    throw std::runtime_error(ss.str());
-  }
-}
-
-inline void throw_on_error(cublasStatus_t stat, const char* message) {
-  std::stringstream ss;
-  if (stat == CUBLAS_STATUS_SUCCESS) {
-    return;
-  } else if (stat == CUBLAS_STATUS_NOT_INITIALIZED) {
-    ss << "CUBLAS: not initialized";
-  } else if (stat == CUBLAS_STATUS_ALLOC_FAILED) {
-    ss << "CUBLAS: alloc failed";
-  } else if (stat == CUBLAS_STATUS_INVALID_VALUE) {
-    ss << "CUBLAS: invalid value";
-  } else if (stat == CUBLAS_STATUS_ARCH_MISMATCH) {
-    ss << "CUBLAS: arch mismatch";
-  } else if (stat == CUBLAS_STATUS_MAPPING_ERROR) {
-    ss << "CUBLAS: mapping error";
-  } else if (stat == CUBLAS_STATUS_EXECUTION_FAILED) {
-    ss << "CUBLAS: execution failed";
-  } else if (stat == CUBLAS_STATUS_INTERNAL_ERROR) {
-    ss << "CUBLAS: internal error";
-  } else if (stat == CUBLAS_STATUS_NOT_SUPPORTED) {
-    ss << "CUBLAS: not supported";
-  } else if (stat == CUBLAS_STATUS_LICENSE_ERROR) {
-    ss << "CUBLAS: license error";
-  }
-  ss << ", " << message;
-  throw std::runtime_error(ss.str());
-}
-
-inline void throw_on_error(cublasStatus_t stat) {
-  const char* message = "";
-  throw_on_error(stat, message);
-}
-
-#endif  // PADDLE_ONLY_CPU
-
-inline void throw_on_error(int stat, const char* message) {
-  if (stat) {
-    throw std::runtime_error(message + (", stat = " + std::to_string(stat)));
-  }
-}
-
-}  // namespace platform
-}  // namespace paddle

paddle/platform/gpu_info.cc

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "paddle/platform/gpu_info.h"
 #include "gflags/gflags.h"
-#include "paddle/platform/error.h"
+#include "paddle/platform/enforce.h"
 
 DEFINE_double(fraction_of_gpu_memory_to_use, 0.95,
               "Default use 95% of GPU memory for PaddlePaddle,"
@@ -25,7 +25,7 @@ namespace platform {
 
 int GetDeviceCount() {
   int count;
-  throw_on_error(
+  PADDLE_ENFORCE(
       cudaGetDeviceCount(&count),
       "cudaGetDeviceCount failed in paddle::platform::GetDeviceCount");
   return count;
@@ -33,19 +33,19 @@ int GetDeviceCount() {
 
 int GetCurrentDeviceId() {
   int device_id;
-  throw_on_error(
+  PADDLE_ENFORCE(
       cudaGetDevice(&device_id),
       "cudaGetDevice failed in paddle::platform::GetCurrentDeviceId");
   return device_id;
 }
 
 void SetDeviceId(int id) {
-  throw_on_error(cudaSetDevice(id),
+  PADDLE_ENFORCE(cudaSetDevice(id),
                  "cudaSetDevice failed in paddle::platform::SetDeviceId");
 }
 
 void GpuMemoryUsage(size_t &available, size_t &total) {
-  throw_on_error(cudaMemGetInfo(&available, &total),
+  PADDLE_ENFORCE(cudaMemGetInfo(&available, &total),
                  "cudaMemGetInfo failed in paddle::platform::GetMemoryUsage");
 }
 
@@ -84,21 +84,26 @@ size_t GpuMaxChunkSize() {
 
 void GpuMemcpyAsync(void *dst, const void *src, size_t count,
                     enum cudaMemcpyKind kind, cudaStream_t stream) {
-  PADDLE_ENFORCE(cudaMemcpyAsync(dst, src, count, kind, stream));
+  PADDLE_ENFORCE(cudaMemcpyAsync(dst, src, count, kind, stream),
+                 "cudaMemcpyAsync failed in paddle::platform::GpuMemcpyAsync");
 }
 
 void GpuMemcpySync(void *dst, const void *src, size_t count,
                    enum cudaMemcpyKind kind) {
-  PADDLE_ENFORCE(cudaMemcpy(dst, src, count, kind));
+  PADDLE_ENFORCE(cudaMemcpy(dst, src, count, kind),
+                 "cudaMemcpy failed in paddle::platform::GpuMemcpySync");
   // note: cudaMemcpy may actually be asynchronous with respect to the caller,
   //       block on stream 0 to make sure the copy has completed
-  PADDLE_ENFORCE(cudaStreamSynchronize(0));
+  PADDLE_ENFORCE(
+      cudaStreamSynchronize(0),
+      "cudaStreamSynchronize failed in paddle::platform::GpuMemcpySync");
 }
 
 void GpuMemcpyPeer(void *dst, int dst_device, const void *src, int src_device,
                    size_t count, cudaStream_t stream) {
   PADDLE_ENFORCE(
-      cudaMemcpyPeerAsync(dst, dst_device, src, src_device, count, stream));
+      cudaMemcpyPeerAsync(dst, dst_device, src, src_device, count, stream),
+      "cudaMemcpyPeerAsync failed in paddle::platform::GpuMemcpyPeer");
 }
 }  // namespace platform
 }  // namespace paddle

paddle/platform/place.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
 #include "paddle/platform/place.h"
 
 namespace paddle {
@@ -7,7 +21,7 @@ namespace detail {
 
 class PlacePrinter : public boost::static_visitor<> {
  public:
-  PlacePrinter(std::ostream &os) : os_(os) {}
+  explicit PlacePrinter(std::ostream &os) : os_(os) {}
   void operator()(const CPUPlace &) { os_ << "CPUPlace"; }
   void operator()(const GPUPlace &p) { os_ << "GPUPlace(" << p.device << ")"; }
 

paddle/pybind/CMakeLists.txt

-cc_library(paddle_pybind SHARED SRCS pybind.cc DEPS pybind python)
+cc_library(paddle_pybind SHARED SRCS pybind.cc DEPS pybind python
+        add_op fc_op sgd_op)

paddle/pybind/pybind.cc

@@ -13,15 +13,49 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <Python.h>
+#include <paddle/framework/op_registry.h>
+#include <paddle/framework/operator.h>
 #include <paddle/framework/scope.h>
+#include <paddle/pybind/tensor_bind.h>
+#include <pybind11/numpy.h>
 #include <pybind11/pybind11.h>
+#include <pybind11/stl.h>
+#include <fstream>
+#include <vector>
 
 namespace py = pybind11;
 namespace pd = paddle::framework;
 
+USE_OP(add_two);
+USE_OP_WITHOUT_KERNEL(fc);
+USE_OP(sgd);
+
 PYBIND11_PLUGIN(core) {
   py::module m("core", "C++ core of Paddle Paddle");
 
+  py::class_<pd::Tensor>(m, "Tensor", py::buffer_protocol())
+      .def_buffer([](pd::Tensor& self) -> py::buffer_info {
+        return paddle::pybind::CastToPyBuffer(self);
+      })
+      .def("get_dims",
+           [](const pd::Tensor& self) { return pd::vectorize(self.dims()); })
+      .def("set_dims",
+           [](pd::Tensor& self, const std::vector<int>& dim) {
+             self.set_dims(pd::make_ddim(dim));
+           })
+      .def("alloc_float",
+           [](pd::Tensor& self) {
+             self.mutable_data<float>(paddle::platform::CPUPlace());
+           })
+      .def("alloc_int",
+           [](pd::Tensor& self) {
+             self.mutable_data<int>(paddle::platform::CPUPlace());
+           })
+      .def("set", paddle::pybind::PyTensorSetFromArray<float>)
+      .def("set", paddle::pybind::PyTensorSetFromArray<int>)
+      .def("shape",
+           [](pd::Tensor& self) { return pd::vectorize(self.dims()); });
+
   py::class_<pd::Variable>(m, "Variable", R"DOC(Variable Class.
 
 All parameter, weight, gradient are variables in Paddle.
@@ -32,7 +66,12 @@ All parameter, weight, gradient are variables in Paddle.
              *var.GetMutable<int>() = val;
            })
       .def("get_int",
-           [](const pd::Variable& var) -> int { return var.Get<int>(); });
+           [](const pd::Variable& var) -> int { return var.Get<int>(); })
+      .def("get_tensor",
+           [](pd::Variable& self) -> pd::Tensor* {
+             return self.GetMutable<pd::Tensor>();
+           },
+           py::return_value_policy::reference);
 
   py::class_<pd::Scope, std::shared_ptr<pd::Scope>>(m, "Scope")
       .def(py::init<const std::shared_ptr<pd::Scope>&>())
@@ -43,5 +82,47 @@ All parameter, weight, gradient are variables in Paddle.
            &pd::Scope::CreateVariable,
            py::return_value_policy::reference);
 
+  //! @note: Be careful! PyBind will return std::string as an unicode, not
+  //! Python str. If you want a str object, you should cast them in Python.
+  m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
+    auto& protos = pd::OpRegistry::protos();
+    std::vector<py::bytes> ret_values;
+    for (auto it = protos.begin(); it != protos.end(); ++it) {
+      PADDLE_ENFORCE(it->second.IsInitialized(),
+                     "OpProto must all be initialized");
+      std::string str;
+      PADDLE_ENFORCE(it->second.SerializeToString(&str),
+                     "Serialize OpProto Error. This could be a bug of Paddle.");
+      ret_values.push_back(py::bytes(str));
+    }
+    return ret_values;
+  });
+  m.def_submodule(
+       "var_names",
+       "The module will return special predefined variable name in Paddle")
+      .def("empty", pd::OperatorBase::EMPTY_VAR_NAME)
+      .def("temp", pd::OperatorBase::TMP_VAR_NAME);
+
+  py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
+      .def_static("cpu_context", []() -> paddle::platform::DeviceContext* {
+        return new paddle::platform::CPUDeviceContext();
+      });
+
+  py::class_<pd::OperatorBase, pd::OperatorPtr>(m, "Operator")
+      .def("__str__", &pd::OperatorBase::DebugString)
+      .def_static("create",
+                  [](py::bytes protobin) {
+                    pd::OpDesc desc;
+                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                                   "Cannot parse user input to OpDesc");
+                    PADDLE_ENFORCE(desc.IsInitialized(),
+                                   "User OpDesc is not initialized, reason %s",
+                                   desc.InitializationErrorString());
+                    return pd::OpRegistry::CreateOp(desc);
+                  })
+      .def("infer_shape", &pd::OperatorBase::InferShape)
+      .def("run", &pd::OperatorBase::Run)
+      .def("outputs", [](const pd::OperatorPtr& op) { return op->outputs_; });
+
   return m.ptr();
 }

paddle/pybind/tensor_bind.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+#include <paddle/framework/tensor.h>
+#include <pybind11/numpy.h>
+#include <pybind11/pybind11.h>
+
+namespace py = pybind11;
+
+namespace paddle {
+
+namespace pybind {
+
+namespace details {
+
+template <bool less, size_t I, typename... ARGS>
+struct CastToPyBufferImpl;
+
+template <size_t I, typename... ARGS>
+struct CastToPyBufferImpl<false, I, ARGS...> {
+  py::buffer_info operator()(framework::Tensor &tensor) {
+    PADDLE_THROW("This type of tensor cannot be expose to Python");
+    return py::buffer_info();
+  }
+};
+
+template <size_t I, typename... ARGS>
+struct CastToPyBufferImpl<true, I, ARGS...> {
+  using CUR_TYPE = typename std::tuple_element<I, std::tuple<ARGS...>>::type;
+  py::buffer_info operator()(framework::Tensor &tensor) {
+    PADDLE_ENFORCE(paddle::platform::is_cpu_place(tensor.holder_->place()),
+                   "Only CPU tensor can cast to numpy array");
+
+    if (std::type_index(typeid(CUR_TYPE)) == tensor.holder_->type()) {
+      auto dim_vec = framework::vectorize(tensor.dims());
+      std::vector<size_t> dims_outside;
+      std::vector<size_t> strides;
+      dims_outside.resize(dim_vec.size());
+      strides.resize(dim_vec.size());
+
+      size_t prod = 1;
+      for (size_t i = dim_vec.size(); i != 0; --i) {
+        dims_outside[i - 1] = (size_t)dim_vec[i - 1];
+        strides[i - 1] = sizeof(CUR_TYPE) * prod;
+        prod *= dims_outside[i - 1];
+      }
+
+      return py::buffer_info(
+          tensor.mutable_data<CUR_TYPE>(tensor.holder_->place()),
+          sizeof(CUR_TYPE),
+          py::format_descriptor<CUR_TYPE>::format(),
+          (size_t)framework::arity(tensor.dims()),
+          dims_outside,
+          strides);
+    } else {
+      constexpr bool less = I + 1 < std::tuple_size<std::tuple<ARGS...>>::value;
+      return CastToPyBufferImpl<less, I + 1, ARGS...>()(tensor);
+    }
+  }
+};
+}  // namespace details
+inline py::buffer_info CastToPyBuffer(framework::Tensor &tensor) {
+  auto buffer_info = details::CastToPyBufferImpl<true, 0, float, int>()(tensor);
+  return buffer_info;
+}
+
+template <typename T>
+void PyTensorSetFromArray(
+    framework::Tensor &self,
+    py::array_t<T, py::array::c_style | py::array::forcecast> array) {
+  std::vector<int> dims;
+  dims.reserve(array.ndim());
+  for (size_t i = 0; i < array.ndim(); ++i) {
+    dims.push_back((int)array.shape()[i]);
+  }
+
+  self.set_dims(framework::make_ddim(dims));
+  auto *dst = self.mutable_data<T>(paddle::platform::CPUPlace());
+  std::memcpy(dst, array.data(), sizeof(T) * array.size());
+}
+
+}  // namespace pybind
+}  // namespace paddle

paddle/scripts/docker/build.sh

@@ -155,7 +155,8 @@ RUN apt-get update &&\
     paddle version
 ${DOCKERFILE_CUDNN_DSO}
 ${DOCKERFILE_GPU_ENV}
-
+ADD go/cmd/pserver/pserver /usr/bin/
+ADD go/cmd/master/master /usr/bin/
 # default command shows the paddle version and exit
 CMD ["paddle", "version"]
 EOF

paddle/scripts/docker/build_android.sh

@@ -2,9 +2,9 @@
 
 set -xe
 
-mkdir -p /paddle/build
-cd /paddle/build
-rm -f /paddle/install 2>/dev/null || true
+mkdir -p /paddle/build_android
+cd /paddle/build_android
+rm -rf /paddle/install 2>/dev/null || true
 cmake -DCMAKE_SYSTEM_NAME=Android \
       -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_STANDALONE_TOOLCHAIN \
       -DANDROID_ABI=armeabi-v7a \
@@ -21,6 +21,3 @@ cmake -DCMAKE_SYSTEM_NAME=Android \
       ..
 make -j `nproc`
 make install
-
-export PATH=/paddle/install/bin:/paddle/install/opt/paddle/bin:$PATH
-paddle version

paddle/scripts/travis/check_style.sh

 #!/bin/bash
 function abort(){
     echo "Your change doesn't follow PaddlePaddle's code style." 1>&2
-    echo "Please use pre-commit to reformat your code and git push again." 1>&2
+    echo "Please use pre-commit to check what is wrong." 1>&2
     exit 1
 }
 
@@ -13,8 +13,14 @@ export PATH=/usr/bin:$PATH
 pre-commit install
 clang-format --version
 
+# set up go environment for running gometalinter
+mkdir -p $GOPATH/src/github.com/PaddlePaddle/
+ln -sf $TRAVIS_BUILD_DIR $GOPATH/src/github.com/PaddlePaddle/Paddle
+cd  $GOPATH/src/github.com/PaddlePaddle/Paddle/go; glide install; cd -
+
 if ! pre-commit run -a ; then
-  git diff  --exit-code
+    git diff
+    exit 1
 fi
 
 trap : 0

paddle/trainer/NewRemoteParameterUpdater.cpp

@@ -28,6 +28,17 @@ NewRemoteParameterUpdater::NewRemoteParameterUpdater(
       newGradients_(nullptr),
       pserverSpec_(pserverSpec) {}
 
+NewRemoteParameterUpdater::NewRemoteParameterUpdater(
+    const OptimizationConfig &config,
+    const std::string pserverSpec,
+    const bool useEtcd)
+    : trainerConfig_(config),
+      parameterClient_(-1),
+      newParameters_(nullptr),
+      newGradients_(nullptr),
+      pserverSpec_(pserverSpec),
+      useEtcd_(useEtcd) {}
+
 void NewRemoteParameterUpdater::init(
     const std::vector<ParameterPtr> &parameters) {
   ParameterUpdater::init(parameters);
@@ -38,8 +49,13 @@ void NewRemoteParameterUpdater::init(
   }
 
   // create parameter server client.
-  parameterClient_ = paddle_new_pserver_client((char *)pserverSpec_.c_str(),
-                                               FLAGS_trainer_id == 0);
+  if (useEtcd_) {
+    parameterClient_ = paddle_new_etcd_pserver_client(
+        (char *)pserverSpec_.c_str(), FLAGS_trainer_id == 0);
+  } else {
+    parameterClient_ = paddle_new_pserver_client((char *)pserverSpec_.c_str(),
+                                                 FLAGS_trainer_id == 0);
+  }
 
   // init new parameter and gradient.
   newParameters_ = initNewParameter(PARAMETER_VALUE);

paddle/trainer/NewRemoteParameterUpdater.h

@@ -32,6 +32,9 @@ class NewRemoteParameterUpdater : public ParameterUpdater {
 public:
   NewRemoteParameterUpdater(const OptimizationConfig& config,
                             const std::string pserverSpec);
+  NewRemoteParameterUpdater(const OptimizationConfig& config,
+                            const std::string pserverSpec,
+                            const bool useEtcd);
   ~NewRemoteParameterUpdater() {
     releaseNewParameter(newParameters_);
     releaseNewParameter(newGradients_);
@@ -111,6 +114,8 @@ protected:
   paddle_parameter** newGradients_;
   /// the specification of parameter server "host1:port,host1:port"
   std::string pserverSpec_;
+  /// true if pserverSpec_ is etcd endpoint, else pserverSpec_ is pserver addr
+  bool useEtcd_;
 };
 
 }  // namespace paddle

paddle/utils/DynamicLoader.h

@@ -12,8 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#ifndef DYNAMIC_LOAD_H_
-#define DYNAMIC_LOAD_H_
+#pragma once
 
 #include <dlfcn.h>
 #include <memory>
@@ -59,5 +58,3 @@ void GetWarpCTCDsoHandle(void** dso_handle);
  *
  */
 void GetLapackDsoHandle(void** dso_handle);
-
-#endif  // DYNAMIC_LOAD_H_

paddle/utils/ThreadLocal.h

@@ -51,7 +51,7 @@ template <class T>
 class ThreadLocal {
 public:
   ThreadLocal() {
-    CHECK(pthread_key_create(&threadSpecificKey_, dataDestructor) == 0);
+    CHECK_EQ(pthread_key_create(&threadSpecificKey_, dataDestructor), 0);
   }
   ~ThreadLocal() { pthread_key_delete(threadSpecificKey_); }
 
@@ -65,7 +65,7 @@ public:
     if (!p && createLocal) {
       p = new T();
       int ret = pthread_setspecific(threadSpecificKey_, p);
-      CHECK(ret == 0);
+      CHECK_EQ(ret, 0);
     }
     return p;
   }
@@ -79,7 +79,7 @@ public:
     if (T* q = get(false)) {
       dataDestructor(q);
     }
-    CHECK(pthread_setspecific(threadSpecificKey_, p) == 0);
+    CHECK_EQ(pthread_setspecific(threadSpecificKey_, p), 0);
   }
 
   /**
@@ -112,7 +112,7 @@ private:
 template <class T>
 class ThreadLocalD {
 public:
-  ThreadLocalD() { CHECK(pthread_key_create(&threadSpecificKey_, NULL) == 0); }
+  ThreadLocalD() { CHECK_EQ(pthread_key_create(&threadSpecificKey_, NULL), 0); }
   ~ThreadLocalD() {
     pthread_key_delete(threadSpecificKey_);
     for (auto t : threadMap_) {
@@ -127,7 +127,7 @@ public:
     T* p = (T*)pthread_getspecific(threadSpecificKey_);
     if (!p) {
       p = new T();
-      CHECK(pthread_setspecific(threadSpecificKey_, p) == 0);
+      CHECK_EQ(pthread_setspecific(threadSpecificKey_, p), 0);
       updateMap(p);
     }
     return p;
@@ -141,7 +141,7 @@ public:
     if (T* q = (T*)pthread_getspecific(threadSpecificKey_)) {
       dataDestructor(q);
     }
-    CHECK(pthread_setspecific(threadSpecificKey_, p) == 0);
+    CHECK_EQ(pthread_setspecific(threadSpecificKey_, p), 0);
     updateMap(p);
   }
 

proto/ModelConfig.proto

@@ -472,10 +472,16 @@ message LayerConfig {
   // blank label used in ctc loss
   optional uint32 blank = 52 [default = 0];
 
-  // stride parameter for seqlastins layer, AverageLayer, MaxLayer, which 
+  // stride parameter for seqlastins layer, AverageLayer, MaxLayer, which
   // controls the scope of pooling operation. can be set > 0.
   // leave empty or set to -1 to disable this stride pooling.
   optional int32 seq_pool_stride = 53 [default = -1];
+
+  // for crop layer
+  optional int32 axis = 54 [default = 2];
+  repeated uint32 offset = 55;
+  repeated uint32 shape = 56;
+
 }
 
 message EvaluatorConfig {

python/paddle/trainer/config_parser.py

@@ -1575,7 +1575,13 @@ class MultiClassCrossEntropySelfNormCostLayer(LayerBase):
 
 @config_layer('fc')
 class FCLayer(LayerBase):
-    def __init__(self, name, size, inputs, bias=True, **xargs):
+    def __init__(self,
+                 name,
+                 size,
+                 inputs,
+                 bias=True,
+                 error_clipping_threshold=None,
+                 **xargs):
         super(FCLayer, self).__init__(name, 'fc', size, inputs=inputs, **xargs)
         for input_index in xrange(len(self.inputs)):
             input_layer = self.get_input_layer(input_index)
@@ -1592,6 +1598,8 @@ class FCLayer(LayerBase):
             self.create_input_parameter(input_index, psize, dims, sparse,
                                         format)
         self.create_bias_parameter(bias, self.config.size)
+        if error_clipping_threshold is not None:
+            self.config.error_clipping_threshold = error_clipping_threshold
 
 
 @config_layer('selective_fc')
@@ -1990,6 +1998,23 @@ class PadLayer(LayerBase):
         self.config.size = out_ch * out_h * out_w
 
 
+@config_layer('crop')
+class CropLayer(LayerBase):
+    def __init__(self, name, inputs, axis, offset, shape, **xargs):
+        super(CropLayer, self).__init__(name, 'crop', 0, inputs=inputs, **xargs)
+        self.config.axis = axis
+        self.config.offset.extend(offset)
+        self.config.shape.extend(shape)
+
+        # get channel, width and height from input_0 layer
+        input_layer = self.get_input_layer(0)
+        image_conf = self.config.inputs[0].image_conf
+        image_conf.img_size = input_layer.width
+        image_conf.img_size_y = input_layer.height
+        image_conf.channels = input_layer.size / (input_layer.width *
+                                                  input_layer.height)
+
+
 @config_layer('batch_norm')
 class BatchNormLayer(LayerBase):
     layer_type = 'batch_norm'

python/paddle/trainer_config_helpers/layers.py

@@ -127,6 +127,7 @@ __all__ = [
     'dropout_layer',
     'prelu_layer',
     'gated_unit_layer',
+    'crop_layer',
 ]
 
 
@@ -218,6 +219,7 @@ class LayerType(object):
     SMOOTH_L1 = 'smooth_l1'
 
     PRELU = 'prelu'
+    CROP_LAYER = 'crop'
 
     @staticmethod
     def is_layer_type(type_name):
@@ -5970,3 +5972,52 @@ def gated_unit_layer(input,
         name="%s_gated_act" % name,
         input=dotmul_operator(input_proj, gate),
         layer_attr=layer_attr)
+
+
+@wrap_name_default()
+@layer_support()
+def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None):
+    """
+    The crop layer crops images by offset and shape. User can set crop shape by
+    args 'shape' explicitly or by reference input layer.
+    
+    The example usage is:
+
+    .. code-block:: python
+    crop = crop_layer(input=[image_input, reference_input], axis=2, offset=[2, 3])
+
+    :param input: The input layer.If two inputs were setted,
+                    the second input will be regarded as reference input
+    :type input: LayerOutput or Sequence
+    :param offset: The crop offset
+    :type offset: Sequence
+    :param axis: start axis to be cropped. To image input layer:
+        - 0: batch size
+        - 1: channels
+        - 2: height
+        - 3: width
+    :type partial_sum: int
+    :param shape: The shape to be cropped. Default is None.
+    :type shape: Sequence | None
+    :param name: Name of this layer.
+    :type name: basestring
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+    if isinstance(input, LayerOutput):
+        input = [input]
+    else:
+        assert isinstance(input, collections.Sequence)
+    l = Layer(
+        inputs=[x.name for x in input],
+        axis=axis,
+        offset=offset,
+        shape=shape,
+        name=name,
+        type=LayerType.CROP_LAYER,
+        **ExtraLayerAttribute.to_kwargs(layer_attr))
+    return LayerOutput(
+        name=name,
+        layer_type=LayerType.CROP_LAYER,
+        parents=input,
+        size=l.config.size)

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

+from paddle.trainer_config_helpers import *
+
+settings(batch_size=1000, learning_rate=1e-5)
+
+data = data_layer(name='data', size=2016, height=48, width=42)
+refernce_data = data_layer(name='data', size=768, height=16, width=16)
+
+conv = img_conv_layer(
+    input=data,
+    filter_size=3,
+    num_channels=1,
+    num_filters=16,
+    padding=1,
+    act=LinearActivation(),
+    bias_attr=True)
+
+pool = img_pool_layer(input=conv, pool_size=2, stride=2, pool_type=MaxPooling())
+
+crop = crop_layer(input=[pool, refernce_data], axis=2)
+
+outputs(pad)

python/paddle/v2/__init__.py

@@ -20,7 +20,6 @@ import trainer
 import event
 import data_type
 import topology
-import data_feeder
 import networks
 import evaluator
 from . import dataset
@@ -31,7 +30,6 @@ import op
 import pooling
 import inference
 import networks
-import py_paddle.swig_paddle as api
 import minibatch
 import plot
 import image
@@ -47,7 +45,6 @@ __all__ = [
     'data_type',
     'attr',
     'pooling',
-    'data_feeder',
     'dataset',
     'reader',
     'topology',
@@ -61,6 +58,7 @@ __all__ = [
 
 
 def init(**kwargs):
+    import py_paddle.swig_paddle as api
     args = []
     args_dict = {}
     # NOTE: append arguments if they are in ENV

python/paddle/v2/data_feeder.py

@@ -11,7 +11,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 from py_paddle import DataProviderConverter
 import collections
 import paddle.trainer.PyDataProvider2 as pydp2

python/paddle/v2/dataset/__init__.py

@@ -26,8 +26,9 @@ import sentiment
 import wmt14
 import mq2007
 import flowers
+import voc2012
 
 __all__ = [
     'mnist', 'imikolov', 'imdb', 'cifar', 'movielens', 'conll05', 'sentiment'
-    'uci_housing', 'wmt14', 'mq2007', 'flowers'
+    'uci_housing', 'wmt14', 'mq2007', 'flowers', 'voc2012'
 ]

python/paddle/v2/dataset/common.py

@@ -22,6 +22,8 @@ import importlib
 import paddle.v2.dataset
 import cPickle
 import glob
+import cPickle as pickle
+import random
 
 __all__ = [
     'DATA_HOME', 'download', 'md5file', 'split', 'cluster_files_reader',
@@ -170,8 +172,6 @@ def convert(output_path,
             name_prefix,
             max_lines_to_shuffle=1000):
     import recordio
-    import cPickle as pickle
-    import random
     """
     Convert data from reader to recordio format files.
 
@@ -201,8 +201,10 @@ def convert(output_path,
     def write_data(w, lines):
         random.shuffle(lines)
         for i, d in enumerate(lines):
-            d = pickle.dumps(d, pickle.HIGHEST_PROTOCOL)
-            w[i % num_shards].write(d)
+            # FIXME(Yancey1989):
+            # dumps with protocol: pickle.HIGHEST_PROTOCOL
+            o = pickle.dumps(d)
+            w[i % num_shards].write(o)
 
     w = open_writers()
     lines = []

python/paddle/v2/dataset/mq2007.py

@@ -212,19 +212,19 @@ def gen_pair(querylist, partial_order="full"):
         for j in range(i + 1, len(querylist)):
             query_right = querylist[j]
             if query_left.relevance_score > query_right.relevance_score:
-                labels.append(1)
+                labels.append([1])
                 docpairs.append([
                     np.array(query_left.feature_vector),
                     np.array(query_right.feature_vector)
                 ])
             elif query_left.relevance_score < query_right.relevance_score:
-                labels.append(1)
+                labels.append([1])
                 docpairs.append([
                     np.array(query_right.feature_vector),
                     np.array(query_left.feature_vector)
                 ])
     for label, pair in zip(labels, docpairs):
-        yield label, pair[0], pair[1]
+        yield np.array(label), pair[0], pair[1]
 
 
 def gen_list(querylist):

python/paddle/v2/dataset/tests/voc2012_test.py

+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle.v2.dataset.voc2012
+import unittest
+
+
+class TestVOC(unittest.TestCase):
+    def check_reader(self, reader):
+        sum = 0
+        label = 0
+        for l in reader():
+            self.assertEqual(l[0].size, 3 * l[1].size)
+            sum += 1
+        return sum
+
+    def test_train(self):
+        count = self.check_reader(paddle.v2.dataset.voc_seg.train())
+        self.assertEqual(count, 2913)
+
+    def test_test(self):
+        count = self.check_reader(paddle.v2.dataset.voc_seg.test())
+        self.assertEqual(count, 1464)
+
+    def test_val(self):
+        count = self.check_reader(paddle.v2.dataset.voc_seg.val())
+        self.assertEqual(count, 1449)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/dataset/voc2012.py

+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Image dataset for segmentation.
+The 2012 dataset contains images from 2008-2011 for which additional
+segmentations have been prepared. As in previous years the assignment
+to training/test sets has been maintained. The total number of images
+with segmentation has been increased from 7,062 to 9,993.
+"""
+
+import tarfile
+import io
+import numpy as np
+from paddle.v2.dataset.common import download
+from paddle.v2.image import *
+from PIL import Image
+
+__all__ = ['train', 'test', 'val']
+
+VOC_URL = 'http://host.robots.ox.ac.uk/pascal/VOC/voc2012/\
+VOCtrainval_11-May-2012.tar'
+
+VOC_MD5 = '6cd6e144f989b92b3379bac3b3de84fd'
+SET_FILE = 'VOCdevkit/VOC2012/ImageSets/Segmentation/{}.txt'
+DATA_FILE = 'VOCdevkit/VOC2012/JPEGImages/{}.jpg'
+LABEL_FILE = 'VOCdevkit/VOC2012/SegmentationClass/{}.png'
+
+CACHE_DIR = 'voc2012'
+
+
+def reader_creator(filename, sub_name):
+
+    tarobject = tarfile.open(filename)
+    name2mem = {}
+    for ele in tarobject.getmembers():
+        name2mem[ele.name] = ele
+
+    def reader():
+        set_file = SET_FILE.format(sub_name)
+        sets = tarobject.extractfile(name2mem[set_file])
+        for line in sets:
+            line = line.strip()
+            data_file = DATA_FILE.format(line)
+            label_file = LABEL_FILE.format(line)
+            data = tarobject.extractfile(name2mem[data_file]).read()
+            label = tarobject.extractfile(name2mem[label_file]).read()
+            data = Image.open(io.BytesIO(data))
+            label = Image.open(io.BytesIO(label))
+            data = np.array(data)
+            label = np.array(label)
+            yield data, label
+
+    return reader
+
+
+def train():
+    """
+    Create a train dataset reader containing 2913 images in HWC order.
+    """
+    return reader_creator(download(VOC_URL, CACHE_DIR, VOC_MD5), 'trainval')
+
+
+def test():
+    """
+    Create a test dataset reader containing 1464 images in HWC order.
+    """
+    return reader_creator(download(VOC_URL, CACHE_DIR, VOC_MD5), 'train')
+
+
+def val():
+    """
+    Create a val dataset reader containing 1449 images in HWC order.
+    """
+    return reader_creator(download(VOC_URL, CACHE_DIR, VOC_MD5), 'val')

python/paddle/v2/event.py

@@ -9,8 +9,6 @@ There are:
 * BeginPass
 * EndPass
 """
-import py_paddle.swig_paddle as api
-
 __all__ = [
     'EndIteration', 'BeginIteration', 'BeginPass', 'EndPass', 'TestResult'
 ]
@@ -18,6 +16,7 @@ __all__ = [
 
 class WithMetric(object):
     def __init__(self, evaluator):
+        import py_paddle.swig_paddle as api
         if not isinstance(evaluator, api.Evaluator):
             raise TypeError("Evaluator should be api.Evaluator type")
         self.__evaluator__ = evaluator

python/paddle/v2/framework/create_op_creation_methods.py

+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.attr_type_pb2 as attr_type_pb2
+import cStringIO
+
+
+def get_all_op_protos():
+    """
+    Get all registered op proto from Paddle C++
+    :return: list of OpProto
+    """
+    protostrs = core.get_all_op_protos()
+    ret_values = []
+    for pbstr in protostrs:
+        op_proto = op_proto_pb2.OpProto.FromString(str(pbstr))
+        ret_values.append(op_proto)
+    return ret_values
+
+
+class OpDescCreationMethod(object):
+    """
+    A Functor object to convert user input(use key word args) to OpDesc based on
+    OpProto.
+    
+    :param op_proto: The OpProto object.
+    :type op_proto: op_proto_pb2.OpProto
+    """
+
+    def __init__(self, op_proto):
+        if not isinstance(op_proto, op_proto_pb2.OpProto):
+            raise TypeError("Argument should be OpProto")
+        self.__op_proto__ = op_proto
+
+    def __call__(self, *args, **kwargs):
+        """
+        Convert user input to OpDesc. Only key-word args are supported. 
+        :return: OpDesc based on user input
+        :rtype: op_desc_pb2.OpDesc
+        """
+        if len(args) != 0:
+            raise ValueError("Only keyword arguments is supported by Paddle")
+        op_desc = op_desc_pb2.OpDesc()
+
+        # Inputs
+        ipts, ipt_format, _ = OpDescCreationMethod.extract_input_or_output(
+            "input", kwargs, self.__op_proto__.inputs)
+        op_desc.inputs.extend(ipts)
+        if ipt_format is not None:
+            op_desc.attrs.extend([ipt_format])
+
+        # Outputs
+        outs, out_format, tmp_index = OpDescCreationMethod.extract_input_or_output(
+            "output", kwargs, self.__op_proto__.outputs)
+        op_desc.outputs.extend(outs)
+        if out_format is not None:
+            op_desc.attrs.extend([out_format])
+        if len(tmp_index) != 0:
+            tmp_index_attr = op_desc.attrs.add()
+            tmp_index_attr.type = attr_type_pb2.INTS
+            tmp_index_attr.name = "temporary_index"
+            tmp_index_attr.ints.extend(tmp_index)
+
+        # Types
+        op_desc.type = self.__op_proto__.type
+
+        # Attrs
+        for attr in self.__op_proto__.attrs:
+            if attr.generated:
+                continue
+            user_defined_attr = kwargs.get(attr.name, None)
+            if user_defined_attr is not None:
+                new_attr = op_desc.attrs.add()
+                new_attr.name = attr.name
+                new_attr.type = attr.type
+                if attr.type == attr_type_pb2.INT:
+                    new_attr.i = user_defined_attr
+                elif attr.type == attr_type_pb2.FLOAT:
+                    new_attr.f = user_defined_attr
+                elif attr.type == attr_type_pb2.STRING:
+                    new_attr.s = user_defined_attr
+                elif attr.type == attr_type_pb2.INTS:
+                    new_attr.ints.extend(user_defined_attr)
+                elif attr.type == attr_type_pb2.FLOATS:
+                    new_attr.floats.extend(user_defined_attr)
+                elif attr.type == attr_type_pb2.STRINGS:
+                    new_attr.strings.extend(user_defined_attr)
+                else:
+                    raise NotImplementedError("Not support attribute type " +
+                                              attr.type)
+
+        return op_desc
+
+    @staticmethod
+    def extract_input_or_output(in_out, kwargs, meta):
+        """
+        Extract input variable names or output variable names from key-word 
+        arguments, which base on VarProtos.
+        
+        :param in_out: "input" or "output"
+        :param kwargs: key-word arguments that user inputted.
+        :param meta: a list of VarProto
+        :return: The three object will be return. The variable names. The 
+        input_format or output_format attribute(None if the input or output is 
+        not multiple). The temporary variable index list.
+        """
+        multiple = OpDescCreationMethod.any_is_true((m.multiple for m in meta))
+        tmp_index = []
+        retv = []
+        if multiple:
+            var_format = op_desc_pb2.AttrDesc()
+            var_format.type = attr_type_pb2.INTS
+            var_format.name = "%s_format" % in_out
+            var_format.ints.append(0)
+
+            for var in meta:
+                var_name = var.name
+
+                if var.temporary:
+                    var_name = [core.var_names.temp()]
+                    tmp_index.append(len(retv))
+                else:
+                    var_name = kwargs.get(var_name, [])
+                if not isinstance(var_name, list):
+                    var_name = [var_name]
+                retv.extend(var_name)
+                var_format.ints.append(len(var_name) + var_format.ints[-1])
+            return retv, var_format, tmp_index
+        else:
+            for var in meta:
+                if var.temporary:
+                    retv.append(kwargs.get(var.name, core.var_names.temp()))
+                    tmp_index.append(len(retv))
+                else:
+                    retv.append(kwargs.get(var.name, core.var_names.empty()))
+            return retv, None, tmp_index
+
+    @staticmethod
+    def any_is_true(generator):
+        """
+        Reduce a bool array to one. If any of them is True, then return True.
+        """
+        for flag in generator:
+            if flag:
+                return True
+        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 == attr_type_pb2.INT:
+            attr_type = "int"
+        elif attr.type == attr_type_pb2.FLOAT:
+            attr_type = "float"
+        elif attr.type == attr_type_pb2.STRING:
+            attr_type = "basestr"
+        elif attr.type == attr_type_pb2.INTS:
+            attr_type = "list of int"
+        elif attr.type == attr_type_pb2.FLOATS:
+            attr_type = "list of float"
+        elif attr.type == attr_type_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()
+
+
+def create_op_creation_method(op_proto):
+    """
+    Generate op creation method for an OpProto
+    """
+    method = OpDescCreationMethod(op_proto)
+
+    def __impl__(*args, **kwargs):
+        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]
+
+    return __impl__
+
+
+class OpCreationsHolder(object):
+    """
+    A object will holds all op creation methods.
+    
+    Use `op_creations.xxx_op` to access them.
+    """
+    pass
+
+
+op_creations = OpCreationsHolder()
+
+
+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)
+
+
+__bootstrap__()

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

 add_python_test(test_framework test_protobuf.py test_scope.py
-    test_default_scope_funcs.py)
+    test_default_scope_funcs.py test_op_creation_methods.py
+    test_tensor.py test_fc_op.py test_add_two_op.py test_sgd_op.py)

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
+
+
+class OpTestMeta(type):
+    def __new__(cls, name, bases, attrs):
+        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(None)
+            kwargs = dict()
+
+            for in_name in func.all_input_args:
+                if hasattr(self, in_name):
+                    kwargs[in_name] = in_name
+                    var = scope.create_var(in_name).get_tensor()
+                    arr = getattr(self, in_name)
+                    var.set_dims(arr.shape)
+                    var.set(arr)
+                else:
+                    kwargs[in_name] = "@EMPTY@"
+
+            for out_name in func.all_output_args:
+                if hasattr(self, out_name):
+                    kwargs[out_name] = out_name
+                    scope.create_var(out_name).get_tensor()
+
+            for attr_name in func.all_attr_args:
+                if hasattr(self, attr_name):
+                    kwargs[attr_name] = getattr(self, attr_name)
+
+            op = func(**kwargs)
+
+            op.infer_shape(scope)
+
+            ctx = core.DeviceContext.cpu_context()
+            op.run(scope, ctx)
+
+            for out_name in func.all_output_args:
+                actual = numpy.array(scope.get_var(out_name).get_tensor())
+                expect = getattr(self, out_name)
+                numpy.testing.assert_almost_equal(actual, expect)
+
+        obj.test_all = test_all
+        return obj

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

+import unittest
+from op_test_util import OpTestMeta
+import numpy
+
+
+class TestAddOp(unittest.TestCase):
+    __metaclass__ = OpTestMeta
+
+    def setUp(self):
+        self.type = "add_two"
+        self.X = numpy.random.random((342, 345)).astype("float32")
+        self.Y = numpy.random.random((342, 345)).astype("float32")
+        self.Out = self.X + self.Y
+
+
+if __name__ == '__main__':
+    unittest.main()

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
+
+
+class TestFc(unittest.TestCase):
+    def test_fc(self):
+        scope = core.Scope(None)
+        x = scope.create_var("X")
+        x_tensor = x.get_tensor()
+        x_tensor.set_dims([1000, 784])
+        x_tensor.alloc_float()
+
+        w = scope.create_var("W")
+        w_tensor = w.get_tensor()
+        w_tensor.set_dims([784, 100])
+        w_tensor.alloc_float()
+
+        w_tensor.set(numpy.random.random((784, 100)).astype("float32"))
+
+        # Set a real numpy array here.
+        # x_tensor.set(numpy.array([]))
+
+        op = creation.op_creations.fc(X="X", Y="Y", W="W")
+
+        for out in op.outputs():
+            if scope.get_var(out) is None:
+                scope.create_var(out).get_tensor()
+
+        tensor = scope.get_var("Y").get_tensor()
+        op.infer_shape(scope)
+        self.assertEqual([1000, 100], tensor.shape())
+
+        ctx = core.DeviceContext.cpu_context()
+
+        op.run(scope, ctx)
+
+        # After complete all ops, check Y is expect or not.
+
+
+if __name__ == '__main__':
+    unittest.main()

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

+import unittest
+import paddle.v2.framework.create_op_creation_methods as creation
+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.attr_type_pb2 as attr_type_pb2
+
+
+class TestGetAllProtos(unittest.TestCase):
+    def test_all(self):
+        all_protos = creation.get_all_op_protos()
+        self.assertNotEqual(0, len(all_protos))
+
+        for each in all_protos:
+            self.assertTrue(each.IsInitialized())
+
+
+class TestOpDescCreationMethod(unittest.TestCase):
+    def test_plain_input_output(self):
+        op = op_proto_pb2.OpProto()
+        op.type = "test"
+        ipt = op.inputs.add()
+        ipt.name = "X"
+        ipt.comment = "not matter"
+
+        ipt = op.inputs.add()
+        ipt.name = "Y"
+        ipt.comment = "not matter"
+
+        opt = op.outputs.add()
+        opt.name = "Z"
+        opt.comment = "not matter"
+
+        op.comment = "not matter"
+
+        self.assertTrue(op.IsInitialized())
+
+        method = creation.OpDescCreationMethod(op)
+        output = method(X="a", Y="b", Z="c")
+
+        expected = op_desc_pb2.OpDesc()
+        expected.type = "test"
+        expected.inputs.extend(["a", "b"])
+        expected.outputs.append("c")
+        self.assertEqual(expected, output)
+
+    def test_multiple_input_plain_output(self):
+        op = op_proto_pb2.OpProto()
+        op.type = "fc"
+        ipt = op.inputs.add()
+        ipt.name = "X"
+        ipt.comment = ""
+        ipt.multiple = True
+
+        ipt = op.inputs.add()
+        ipt.name = "W"
+        ipt.comment = ""
+        ipt.multiple = True
+
+        ipt = op.inputs.add()
+        ipt.name = "b"
+        ipt.comment = ""
+
+        out = op.outputs.add()
+        out.name = "Y"
+        out.comment = ""
+
+        op.comment = ""
+        self.assertTrue(op.IsInitialized())
+        method = creation.OpDescCreationMethod(op)
+
+        generated1 = method(X="x", W="w", b="b", Y="y")
+        expected1 = op_desc_pb2.OpDesc()
+        expected1.inputs.extend(['x', 'w', 'b'])
+        expected1.outputs.extend(['y'])
+        expected1.type = 'fc'
+        attr = expected1.attrs.add()
+        attr.name = 'input_format'
+        attr.type = attr_type_pb2.INTS
+        attr.ints.extend([0, 1, 2, 3])
+        self.assertEqual(expected1, generated1)
+
+        generated2 = method(
+            X=['x1', 'x2', 'x3'], b='b', W=['w1', 'w2', 'w3'], Y='y')
+        expected2 = op_desc_pb2.OpDesc()
+        expected2.inputs.extend(['x1', 'x2', 'x3', 'w1', 'w2', 'w3', 'b'])
+        expected2.outputs.extend(['y'])
+        expected2.type = 'fc'
+        attr = expected2.attrs.add()
+        attr.name = 'input_format'
+        attr.type = attr_type_pb2.INTS
+        attr.ints.extend([0, 3, 6, 7])
+        self.assertEqual(expected2, generated2)
+
+    def test_attrs(self):
+        op = op_proto_pb2.OpProto()
+        op.type = "test"
+        ipt = op.inputs.add()
+        ipt.name = 'X'
+        ipt.comment = ""
+
+        def __add_attr__(name, type):
+            attr = op.attrs.add()
+            attr.name = name
+            attr.comment = ""
+            attr.type = type
+
+        __add_attr__("int_attr", attr_type_pb2.INT)
+        __add_attr__("float_attr", attr_type_pb2.FLOAT)
+        __add_attr__("string_attr", attr_type_pb2.STRING)
+        __add_attr__("ints_attr", attr_type_pb2.INTS)
+        __add_attr__("floats_attr", attr_type_pb2.FLOATS)
+        __add_attr__("strings_attr", attr_type_pb2.STRINGS)
+
+        op.comment = ""
+        self.assertTrue(op.IsInitialized())
+
+        method = creation.OpDescCreationMethod(op)
+
+        generated = method(
+            X="a",
+            int_attr=10,
+            float_attr=3.2,
+            string_attr="test_str",
+            ints_attr=[0, 1, 2, 3, 4],
+            floats_attr=[0.2, 3.2, 4.5],
+            strings_attr=["a", "b", "c"])
+
+        expected = op_desc_pb2.OpDesc()
+        expected.type = "test"
+        expected.inputs.extend(['a'])
+        attr = expected.attrs.add()
+        attr.name = "int_attr"
+        attr.type = attr_type_pb2.INT
+        attr.i = 10
+
+        attr = expected.attrs.add()
+        attr.name = "float_attr"
+        attr.type = attr_type_pb2.FLOAT
+        attr.f = 3.2
+
+        attr = expected.attrs.add()
+        attr.name = "string_attr"
+        attr.type = attr_type_pb2.STRING
+        attr.s = "test_str"
+
+        attr = expected.attrs.add()
+        attr.name = "ints_attr"
+        attr.type = attr_type_pb2.INTS
+        attr.ints.extend([0, 1, 2, 3, 4])
+
+        attr = expected.attrs.add()
+        attr.name = 'floats_attr'
+        attr.type = attr_type_pb2.FLOATS
+        attr.floats.extend([0.2, 3.2, 4.5])
+
+        attr = expected.attrs.add()
+        attr.name = 'strings_attr'
+        attr.type = attr_type_pb2.STRINGS
+        attr.strings.extend(['a', 'b', 'c'])
+
+        self.assertEqual(expected, generated)
+
+    def test_input_temporary_output(self):
+        op = op_proto_pb2.OpProto()
+        op.type = "test"
+        out = op.outputs.add()
+        out.name = "OUT"
+        out.comment = ""
+
+        out = op.outputs.add()
+        out.name = "TMP"
+        out.comment = ""
+        out.temporary = True
+
+        out = op.outputs.add()
+        out.name = "OUT2"
+        out.comment = ""
+        op.comment = ""
+
+        method = creation.OpDescCreationMethod(op)
+        generated = method(OUT="a", OUT2="b")
+        desc = op_desc_pb2.OpDesc()
+        desc.outputs.extend(["a", core.var_names.temp(), "b"])
+        desc.type = "test"
+        attr = desc.attrs.add()
+        attr.name = "temporary_index"
+        attr.type = attr_type_pb2.INTS
+        attr.ints.append(2)
+        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 = attr_type_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")
+        self.assertIsNotNone(add_op)
+        # Invoke C++ DebugString()
+        self.assertEqual('Op(add_two), inputs:(a, b), outputs:(z).',
+                         str(add_op))
+
+
+if __name__ == "__main__":
+    unittest.main()

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

+import unittest
+import numpy
+from op_test_util import OpTestMeta
+
+
+class TestSGD(unittest.TestCase):
+    __metaclass__ = OpTestMeta
+
+    def setUp(self):
+        self.type = "sgd"
+        self.param = numpy.random.random((342, 345)).astype("float32")
+        self.grad = numpy.random.random((342, 345)).astype("float32")
+        self.learning_rate = 0.1
+        self.param_out = self.param - self.learning_rate * self.grad
+
+
+if __name__ == "__main__":
+    unittest.main()

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

+import paddle.v2.framework.core as core
+import unittest
+import numpy
+
+
+class TestScope(unittest.TestCase):
+    def test_int_tensor(self):
+        scope = core.Scope(None)
+        var = scope.create_var("test_tensor")
+        tensor = var.get_tensor()
+
+        tensor.set_dims([1000, 784])
+        tensor.alloc_int()
+
+        tensor_array = numpy.array(tensor)
+        self.assertEqual((1000, 784), tensor_array.shape)
+        tensor_array[3, 9] = 1
+        tensor_array[19, 11] = 2
+        tensor.set(tensor_array)
+
+        tensor_array_2 = numpy.array(tensor)
+        self.assertEqual(1.0, tensor_array_2[3, 9])
+        self.assertEqual(2.0, tensor_array_2[19, 11])
+
+    def test_float_tensor(self):
+        scope = core.Scope(None)
+        var = scope.create_var("test_tensor")
+        tensor = var.get_tensor()
+
+        tensor.set_dims([1000, 784])
+        tensor.alloc_float()
+
+        tensor_array = numpy.array(tensor)
+        self.assertEqual((1000, 784), tensor_array.shape)
+        tensor_array[3, 9] = 1.0
+        tensor_array[19, 11] = 2.0
+        tensor.set(tensor_array)
+
+        tensor_array_2 = numpy.array(tensor)
+        self.assertAlmostEqual(1.0, tensor_array_2[3, 9])
+        self.assertAlmostEqual(2.0, tensor_array_2[19, 11])
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/inference.py

 import numpy
-import py_paddle.swig_paddle as api
 import collections
 import topology
 import minibatch
-from data_feeder import DataFeeder
 
 __all__ = ['infer', 'Inference']
 
@@ -28,6 +26,7 @@ class Inference(object):
     """
 
     def __init__(self, output_layer, parameters):
+        import py_paddle.swig_paddle as api
         topo = topology.Topology(output_layer)
         gm = api.GradientMachine.createFromConfigProto(
             topo.proto(), api.CREATE_MODE_TESTING, [api.PARAMETER_VALUE])
@@ -40,6 +39,7 @@ class Inference(object):
         self.__data_types__ = topo.data_type()
 
     def iter_infer(self, input, feeding=None):
+        from data_feeder import DataFeeder
         feeder = DataFeeder(self.__data_types__, feeding)
         batch_size = len(input)
 

python/paddle/v2/master/client.py

@@ -10,8 +10,9 @@ class client(object):
     client is a client to the master server.
     """
 
-    def __init__(self, addr, buf_size):
-        self.c = lib.paddle_new_master_client(addr, buf_size)
+    def __init__(self, etcd_endpoints, timeout, buf_size):
+        self.c = lib.paddle_new_etcd_master_client(etcd_endpoints, timeout,
+                                                   buf_size)
 
     def close(self):
         lib.paddle_release_master_client(self.c)

python/paddle/v2/optimizer.py

-import py_paddle.swig_paddle as swig_api
-
 import paddle.trainer_config_helpers.config_parser_utils as config_parser_utils
 import paddle.trainer_config_helpers.optimizers as v1_optimizers
 """
@@ -18,6 +16,7 @@ __all__ = [
 
 class Optimizer(object):
     def __init__(self, **kwargs):
+        import py_paddle.swig_paddle as swig_api
         if 'batch_size' in kwargs:
             del kwargs['batch_size']  # not important for python library.
 
@@ -36,23 +35,27 @@ class Optimizer(object):
         For each optimizer(SGD, Adam), GradientMachine should enable different
         buffers.
         """
+        import py_paddle.swig_paddle as swig_api
         tmp = swig_api.ParameterOptimizer.create(self.__opt_conf__)
         assert isinstance(tmp, swig_api.ParameterOptimizer)
         return tmp.getParameterTypes()
 
     def __create_local_updater__(self):
+        import py_paddle.swig_paddle as swig_api
         return swig_api.ParameterUpdater.createLocalUpdater(self.__opt_conf__)
 
     def __create_remote_updater__(self, pass_num, use_sparse_updater):
+        import py_paddle.swig_paddle as swig_api
         return swig_api.ParameterUpdater.createRemoteUpdater(
             self.__opt_conf__, pass_num, use_sparse_updater)
 
-    def __create_new_remote_updater__(self, pserver_spec):
+    def __create_new_remote_updater__(self, pserver_spec, use_etcd):
+        import py_paddle.swig_paddle as swig_api
         return swig_api.ParameterUpdater.createNewRemoteUpdater(
-            self.__opt_conf__, pserver_spec)
+            self.__opt_conf__, pserver_spec, use_etcd)
 
     def create_updater(self, is_local, num_passes, use_sparse_updater,
-                       pserver_spec):
+                       pserver_spec, use_etcd):
         """
         create proper parameter_updater by configuration.
         :param is_local: create local or remote parameter updater
@@ -78,7 +81,7 @@ class Optimizer(object):
                     num_passes, use_sparse_updater)
             else:
                 parameter_updater = self.__create_new_remote_updater__(
-                    pserver_spec)
+                    pserver_spec, use_etcd)
         return parameter_updater
 
 
@@ -268,6 +271,7 @@ ModelAverage = v1_optimizers.ModelAverage
 L2Regularization = v1_optimizers.L2Regularization
 
 if __name__ == '__main__':
+    import py_paddle.swig_paddle as swig_api
     swig_api.initPaddle('--use_gpu=false')
     for opt in [
             Momentum(), Adam(), Adamax(), AdaGrad(), DecayedAdaGrad(),

python/paddle/v2/parameters.py

 import numpy as np
-import py_paddle.swig_paddle as api
 from paddle.proto.ParameterConfig_pb2 import ParameterConfig
 import paddle.trainer.config_parser as cp
 import struct
@@ -124,6 +123,7 @@ class Parameters(object):
         :return: parameter value
         :rtype: np.ndarray
         """
+        import py_paddle.swig_paddle as api
         shape = self.get_shape(key)
 
         if len(self.__gradient_machines__) == 0:
@@ -223,7 +223,7 @@ class Parameters(object):
         :type gradient_machine: api.GradientMachine
         :return:
         """
-
+        import py_paddle.swig_paddle as api
         if not isinstance(gradient_machine, api.GradientMachine):
             raise ValueError("gradient_machine should be api.GradientMachine")
 
@@ -359,6 +359,7 @@ def __copy_parameter_to_gradient_machine__(gradient_machine, name, arr):
     :return:
     :rtype: api.Parameter
     """
+    import py_paddle.swig_paddle as api
     param = __get_parameter_in_gradient_machine__(gradient_machine, name)
     vec = param.getBuf(api.PARAMETER_VALUE)
     assert isinstance(vec, api.Vector)

python/paddle/v2/trainer.py

@@ -2,12 +2,6 @@
 Module Trainer
 """
 import collections
-import gzip
-import os
-
-import py_paddle.swig_paddle as api
-
-from data_feeder import DataFeeder
 from topology import Topology
 from . import event as v2_event
 from . import optimizer as v2_optimizer
@@ -51,7 +45,8 @@ class SGD(object):
                  update_equation,
                  extra_layers=None,
                  is_local=True,
-                 pserver_spec=None):
+                 pserver_spec=None,
+                 use_etcd=True):
 
         if not isinstance(parameters, v2_parameters.Parameters):
             raise TypeError('parameters should be parameters')
@@ -59,6 +54,7 @@ class SGD(object):
         if not isinstance(update_equation, v2_optimizer.Optimizer):
             raise TypeError("update equation parameter must be "
                             "paddle.v2.optimizer.Optimizer")
+        import py_paddle.swig_paddle as api
         topology = Topology(cost, extra_layers=extra_layers)
         self.__optimizer__ = update_equation
         self.__topology__ = topology
@@ -66,6 +62,7 @@ class SGD(object):
         self.__topology_in_proto__ = topology.proto()
         self.__is_local__ = is_local
         self.__pserver_spec__ = pserver_spec
+        self.__use_etcd__ = use_etcd
 
         self.__use_sparse_updater__ = self.__topology__.use_sparse_updater()
         # # In local mode, disable sparse_remote_update.
@@ -124,13 +121,15 @@ class SGD(object):
         :type feeding: dict|list
         :return:
         """
+        import py_paddle.swig_paddle as api
+        from data_feeder import DataFeeder
         if event_handler is None:
             event_handler = default_event_handler
         __check_train_args__(**locals())
 
         self.__parameter_updater__ = self.__optimizer__.create_updater(
             self.__is_local__, num_passes, self.__use_sparse_updater__,
-            self.__pserver_spec__)
+            self.__pserver_spec__, self.__use_etcd__)
         self.__parameter_updater__.init(self.__gradient_machine__)
 
         self.__gradient_machine__.start()
@@ -187,6 +186,8 @@ class SGD(object):
         :type feeding: dict
         :return:
         """
+        import py_paddle.swig_paddle as api
+        from data_feeder import DataFeeder
         feeder = DataFeeder(self.__data_types__, feeding)
         evaluator = self.__gradient_machine__.makeEvaluator()
         out_args = api.Arguments.createArguments(0)

python/setup.py.in

@@ -19,7 +19,9 @@ setup_requires=["requests",
                 "recordio",
                 "matplotlib",
                 "rarfile",
-                "scipy>=0.19.0"]
+                "scipy>=0.19.0",
+                "Pillow",
+                "nltk"]
 
 if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']:
     setup_requires+=["opencv-python"]