Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into pixel_softmax_layer

.pre-commit-config.yaml

@@ -21,3 +21,10 @@
     sha: 28c0ea8a67a3e2dbbf4822ef44e85b63a0080a29
     hooks:
     -   id: clang-formater
+-   repo: https://github.com/dnephin/pre-commit-golang
+    sha: e4693a4c282b4fc878eda172a929f7a6508e7d16
+    hooks:
+      -   id: go-fmt
+          files: (.*\.go)
+      -   id: go-lint
+          files: (.*\.go)

.travis.yml

@@ -33,16 +33,17 @@ addons:
       - ccache
 before_install:
   - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
-  # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python 
+  # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python
   # protobuf version.
   - pip install numpy wheel 'protobuf==3.1' sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit requests==2.9.2 LinkChecker
   - pip install rarfile
+  - curl https://glide.sh/get | bash
   - eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
   - |
     function timeout() { perl -e 'alarm shift; exec @ARGV' "$@"; }
 script:
   - |
-    export WITH_GOLANG=ON && timeout 2580 paddle/scripts/travis/${JOB}.sh # 43min timeout
+    timeout 2580 paddle/scripts/travis/${JOB}.sh # 43min timeout
     RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true; else false; fi;
 notifications:
   email:

CMakeLists.txt

@@ -16,6 +16,7 @@ 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})
 
 include(system)
 

cmake/external/glog.cmake

@@ -38,12 +38,14 @@ ExternalProject_Add(
     CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
     CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
     CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${GLOG_INSTALL_DIR}
+    CMAKE_ARGS      -DCMAKE_INSTALL_LIBDIR=${GLOG_INSTALL_DIR}/lib
     CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
     CMAKE_ARGS      -DWITH_GFLAGS=ON
     CMAKE_ARGS      -Dgflags_DIR=${GFLAGS_INSTALL_DIR}/lib/cmake/gflags
     CMAKE_ARGS      -DBUILD_TESTING=OFF
     CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
     CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GLOG_INSTALL_DIR}
+                     -DCMAKE_INSTALL_LIBDIR:PATH=${GLOG_INSTALL_DIR}/lib
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                      -DCMAKE_BUILD_TYPE:STRING=Release
 )

cmake/external/protobuf.cmake

@@ -17,6 +17,65 @@ INCLUDE(ExternalProject)
 FIND_PACKAGE(Protobuf QUIET)
 SET(PROTOBUF_FOUND "OFF")
 
+if(NOT COMMAND protobuf_generate_python)  # before cmake 3.4, protobuf_genrerate_python is not defined.
+    function(protobuf_generate_python SRCS)
+        # shameless copy from https://github.com/Kitware/CMake/blob/master/Modules/FindProtobuf.cmake
+        if(NOT ARGN)
+            message(SEND_ERROR "Error: PROTOBUF_GENERATE_PYTHON() called without any proto files")
+            return()
+        endif()
+
+        if(PROTOBUF_GENERATE_CPP_APPEND_PATH)
+            # Create an include path for each file specified
+            foreach(FIL ${ARGN})
+                get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
+                get_filename_component(ABS_PATH ${ABS_FIL} PATH)
+                list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
+                if(${_contains_already} EQUAL -1)
+                    list(APPEND _protobuf_include_path -I ${ABS_PATH})
+                endif()
+            endforeach()
+        else()
+            set(_protobuf_include_path -I ${CMAKE_CURRENT_SOURCE_DIR})
+        endif()
+
+        if(DEFINED PROTOBUF_IMPORT_DIRS AND NOT DEFINED Protobuf_IMPORT_DIRS)
+            set(Protobuf_IMPORT_DIRS "${PROTOBUF_IMPORT_DIRS}")
+        endif()
+
+        if(DEFINED Protobuf_IMPORT_DIRS)
+            foreach(DIR ${Protobuf_IMPORT_DIRS})
+                get_filename_component(ABS_PATH ${DIR} ABSOLUTE)
+                list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
+                if(${_contains_already} EQUAL -1)
+                    list(APPEND _protobuf_include_path -I ${ABS_PATH})
+                endif()
+            endforeach()
+        endif()
+
+        set(${SRCS})
+        foreach(FIL ${ARGN})
+            get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
+            get_filename_component(FIL_WE ${FIL} NAME_WE)
+            if(NOT PROTOBUF_GENERATE_CPP_APPEND_PATH)
+                get_filename_component(FIL_DIR ${FIL} DIRECTORY)
+                if(FIL_DIR)
+                    set(FIL_WE "${FIL_DIR}/${FIL_WE}")
+                endif()
+            endif()
+
+            list(APPEND ${SRCS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py")
+            add_custom_command(
+                    OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py"
+                    COMMAND  ${Protobuf_PROTOC_EXECUTABLE} --python_out ${CMAKE_CURRENT_BINARY_DIR} ${_protobuf_include_path} ${ABS_FIL}
+                    DEPENDS ${ABS_FIL} ${Protobuf_PROTOC_EXECUTABLE}
+                    COMMENT "Running Python protocol buffer compiler on ${FIL}"
+                    VERBATIM )
+        endforeach()
+
+        set(${SRCS} ${${SRCS}} PARENT_SCOPE)
+    endfunction()
+endif()
 
 # Print and set the protobuf library information,
 # finish this cmake process and exit from this file.

cmake/generic.cmake

@@ -88,7 +88,7 @@
 #
 
 # including binary directory for generated headers.
-include_directories(${CMAKE_BINARY_DIR})
+include_directories(${CMAKE_CURRENT_BINARY_DIR})
 
 if(NOT APPLE)
     find_package(Threads REQUIRED)
@@ -99,25 +99,44 @@ function(merge_static_libs TARGET_NAME)
   set(libs ${ARGN})
   list(REMOVE_DUPLICATES libs)
 
-  # First get the file names of the libraries to be merged
+  # Get all propagation dependencies from the merged libraries
   foreach(lib ${libs})
-    set(libfiles ${libfiles} $<TARGET_FILE:${lib}>)
+    list(APPEND libs_deps ${${lib}_LIB_DEPENDS})
   endforeach()
 
   if(APPLE) # Use OSX's libtool to merge archives
+    # To produce a library we need at least one source file.
+    # It is created by add_custom_command below and will helps
+    # also help to track dependencies.
     set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}_dummy.c)
+
+    # Make the generated dummy source file depended on all static input
+    # libs. If input lib changes,the source file is touched
+    # which causes the desired effect (relink).
+    add_custom_command(OUTPUT ${dummyfile}
+      COMMAND ${CMAKE_COMMAND} -E touch ${dummyfile}
+      DEPENDS ${libs})
+
+    # Generate dummy staic lib
     file(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
     add_library(${TARGET_NAME} STATIC ${dummyfile})
+    target_link_libraries(${TARGET_NAME} ${libs_deps})
+
+    foreach(lib ${libs})
+      # Get the file names of the libraries to be merged
+      set(libfiles ${libfiles} $<TARGET_FILE:${lib}>)
+    endforeach()
 		add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
       COMMAND rm "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a"
       COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles})
-	else() # general UNIX: use "ar" to extract objects and re-add to a common lib
+  else() # general UNIX: use "ar" to extract objects and re-add to a common lib
     foreach(lib ${libs})
       set(objlistfile ${lib}.objlist) # list of objects in the input library
       set(objdir ${lib}.objdir)
 
       add_custom_command(OUTPUT ${objdir}
-        COMMAND ${CMAKE_COMMAND} -E make_directory ${objdir})
+        COMMAND ${CMAKE_COMMAND} -E make_directory ${objdir}
+        DEPENDS ${lib})
 
       add_custom_command(OUTPUT ${objlistfile}
         COMMAND ${CMAKE_AR} -x "$<TARGET_FILE:${lib}>"
@@ -134,18 +153,18 @@ function(merge_static_libs TARGET_NAME)
       list(APPEND mergebases "${mergebase}")
     endforeach()
 
-    # We need a target for the output merged library
     add_library(${TARGET_NAME} STATIC ${mergebases})
+    target_link_libraries(${TARGET_NAME} ${libs_deps})
+
+    # Get the file name of the generated library
     set(outlibfile "$<TARGET_FILE:${TARGET_NAME}>")
 
     foreach(lib ${libs})
       add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
-      COMMAND ${CMAKE_AR} ru ${outlibfile} @"../${lib}.objlist"
-      WORKING_DIRECTORY ${lib}.objdir)
+        COMMAND ${CMAKE_AR} cr ${outlibfile} *.o
+        COMMAND ${CMAKE_RANLIB} ${outlibfile}
+        WORKING_DIRECTORY ${lib}.objdir)
     endforeach()
-
-    add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
-      COMMAND ${CMAKE_RANLIB} ${outlibfile})
   endif()
 endfunction(merge_static_libs)
 
@@ -194,7 +213,7 @@ function(cc_test TARGET_NAME)
     add_executable(${TARGET_NAME} ${cc_test_SRCS})
     target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} gtest gtest_main)
     add_dependencies(${TARGET_NAME} ${cc_test_DEPS} gtest gtest_main)
-    add_test(${TARGET_NAME} ${TARGET_NAME})
+    add_test(NAME ${TARGET_NAME} COMMAND ${TARGET_NAME} WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
   endif()
 endfunction(cc_test)
 
@@ -281,10 +300,11 @@ function(go_library TARGET_NAME)
 
   file(GLOB GO_SOURCE RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.go")
   string(REPLACE "${PADDLE_GO_PATH}/" "" CMAKE_CURRENT_SOURCE_REL_DIR ${CMAKE_CURRENT_SOURCE_DIR})
+  # FIXME: link path
   add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
     COMMAND rm "${${TARGET_NAME}_LIB_PATH}"
     # Golang build source code
-    COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} build ${BUILD_MODE}
+    COMMAND GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} build ${BUILD_MODE}
     -o "${${TARGET_NAME}_LIB_PATH}"
     "./${CMAKE_CURRENT_SOURCE_REL_DIR}/${GO_SOURCE}"
     # must run under GOPATH
@@ -299,11 +319,13 @@ function(go_binary TARGET_NAME)
   cmake_parse_arguments(go_binary "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
   string(REPLACE "${PADDLE_GO_PATH}/" "" CMAKE_CURRENT_SOURCE_REL_DIR ${CMAKE_CURRENT_SOURCE_DIR})
 
+  # FIXME: link path
   add_custom_command(OUTPUT ${TARGET_NAME}_timestamp
-    COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} build
+      COMMAND env LIBRARY_PATH=${CMAKE_BINARY_DIR}/go/pserver/client/c/:$ENV{LIBRARY_PATH}
+      GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} build
     -o "${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}"
     "./${CMAKE_CURRENT_SOURCE_REL_DIR}/${go_binary_SRCS}"
-  WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go")
+    WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go")
   # TODO: don't know what ${TARGET_NAME}_link does
   add_custom_target(${TARGET_NAME} ALL DEPENDS go_vendor ${TARGET_NAME}_timestamp ${go_binary_DEPS})
   install(PROGRAMS ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME} DESTINATION bin)
@@ -332,3 +354,12 @@ function(proto_library TARGET_NAME)
   protobuf_generate_cpp(proto_srcs proto_hdrs ${proto_library_SRCS})
   cc_library(${TARGET_NAME} SRCS ${proto_srcs} DEPS ${proto_library_DEPS} protobuf)
 endfunction()
+
+function(py_proto_compile TARGET_NAME)
+  set(oneValueArgs "")
+  set(multiValueArgs SRCS)
+  cmake_parse_arguments(py_proto_compile "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+  set(py_srcs)
+  protobuf_generate_python(py_srcs ${py_proto_compile_SRCS})
+  add_custom_target(${TARGET_NAME} ALL DEPENDS ${py_srcs})
+endfunction()

doc_theme/templates/layout.html

@@ -101,7 +101,7 @@
     </div>
     <div class="site-nav-links">
       <div class="site-menu">
-        <a class="fork-on-github" href="https://github.com/PaddlePaddle/Paddle" target="_blank"><i class="fa fa-github"></i>Folk me on Github</a>
+        <a class="fork-on-github" href="https://github.com/PaddlePaddle/Paddle" target="_blank"><i class="fa fa-github"></i>Fork me on Github</a>
         <div class="language-switcher dropdown">
           <a type="button" data-toggle="dropdown">
             <span>English</span>

go/cmd/master/CMakeLists.txt

@@ -12,4 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-go_binary(master SRC master.go)
+go_binary(master SRC master.go DEPS paddle_go_optimizer)

go/cmd/pserver/CMakeLists.txt

@@ -12,4 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-go_binary(pserver SRCS pserver.go)
+go_binary(pserver SRCS pserver.go DEPS paddle_go_optimizer)

go/master/c/client.go

@@ -104,11 +104,22 @@ func paddle_set_dataset(client C.paddle_master_client, path **C.char, size C.int
 	return C.PADDLE_MASTER_OK
 }
 
+// return value:
+//     0:ok
+//    -1:error
 //export paddle_next_record
 func paddle_next_record(client C.paddle_master_client, record **C.uchar) C.int {
 	c := get(client)
-	r := c.NextRecord()
+	r, err := c.NextRecord()
+	if err != nil {
+		// Error
+		// TODO: return the type of error?
+		*record = (*C.uchar)(nullPtr)
+		return -1
+	}
+
 	if len(r) == 0 {
+		// Empty record
 		*record = (*C.uchar)(nullPtr)
 		return 0
 	}

go/master/client.go

@@ -11,7 +11,12 @@ import (
 // Client is the client of the master server.
 type Client struct {
 	conn *connection.Conn
-	ch   chan []byte
+	ch   chan record
+}
+
+type record struct {
+	r   []byte
+	err error
 }
 
 // NewClient creates a new Client.
@@ -21,7 +26,7 @@ type Client struct {
 func NewClient(addrCh <-chan string, bufSize int) *Client {
 	c := &Client{}
 	c.conn = connection.New()
-	c.ch = make(chan []byte, bufSize)
+	c.ch = make(chan record, bufSize)
 	go c.monitorMaster(addrCh)
 	go c.getRecords()
 	return c
@@ -46,10 +51,11 @@ func (c *Client) getRecords() {
 
 			s := recordio.NewRangeScanner(f, &chunk.Index, -1, -1)
 			for s.Scan() {
-				c.ch <- s.Record()
+				c.ch <- record{s.Record(), nil}
 			}
 
 			if s.Err() != nil {
+				c.ch <- record{nil, s.Err()}
 				log.Errorln(err, chunk.Path)
 			}
 
@@ -116,6 +122,7 @@ func (c *Client) taskFinished(taskID int) error {
 //
 // NextRecord will block until the next record is available. It is
 // thread-safe.
-func (c *Client) NextRecord() []byte {
-	return <-c.ch
+func (c *Client) NextRecord() ([]byte, error) {
+	r := <-c.ch
+	return r.r, r.err
 }

go/master/client_test.go

@@ -68,12 +68,17 @@ func TestNextRecord(t *testing.T) {
 	for pass := 0; pass < 50; pass++ {
 		received := make(map[byte]bool)
 		for i := 0; i < total; i++ {
-			r := c.NextRecord()
+			r, err := c.NextRecord()
+			if err != nil {
+				t.Fatal(pass, i, "Read error:", err)
+			}
+
 			if len(r) != 1 {
-				t.Fatal("Length should be 1.", r)
+				t.Fatal(pass, i, "Length should be 1.", r)
 			}
+
 			if received[r[0]] {
-				t.Fatal("Received duplicate.", received, r)
+				t.Fatal(pass, i, "Received duplicate.", received, r)
 			}
 			received[r[0]] = true
 		}

go/pserver/client/c/CMakeLists.txt

 cc_library(paddle_go_optimizer DEPS paddle_optimizer paddle_proto glog gflags protobuf)
+target_link_libraries(paddle_go_optimizer stdc++ m)
 go_library(paddle_pserver_cclient STATIC DEPS paddle_go_optimizer)
 if(WITH_TESTING)
-  add_subdirectory(test)
+  # FIXME: this test requires pserver which is not managed by the test
+  # we need some kind of e2e testing machanism.
+  # add_subdirectory(test)
 endif()

go/pserver/client/c/test/CMakeLists.txt

-cc_test(test_cclient SRCS test_cclient.c DEPS paddle_pserver_cclient)
+cc_test(test_cclient SRCS test_cclient.c DEPS paddle_pserver_cclient paddle_go_optimizer)
 add_style_check_target(test_cclient test_cclient.c)

go/pserver/optimizer.go

@@ -2,7 +2,7 @@ package pserver
 
 // #cgo CFLAGS: -I ../../
 // //FIXME: ldflags contain "build" path
-// #cgo LDFLAGS: ../../build/go/pserver/client/c/libpaddle_go_optimizer.a -lstdc++ -lm
+// #cgo LDFLAGS: ${SRCDIR}/../../build/go/pserver/client/c/libpaddle_go_optimizer.a -lstdc++ -lm
 // #include "paddle/optimizer/optimizer.h"
 // #include <stdlib.h>
 // #include <string.h>
@@ -56,8 +56,8 @@ func newOptimizer(paramWithConfigs ParameterWithConfig) *optimizer {
 
 func (o *optimizer) GetWeights() []byte {
 	var buffer unsafe.Pointer
-	buffer_len := C.paddle_optimizer_get_weights(o.opt, &buffer)
-	return cArrayToSlice(buffer, int(buffer_len)*C.sizeof_float)
+	bufferLen := C.paddle_optimizer_get_weights(o.opt, &buffer)
+	return cArrayToSlice(buffer, int(bufferLen)*C.sizeof_float)
 }
 
 func (o *optimizer) UpdateParameter(g Gradient) error {

go/pserver/service.go

@@ -10,8 +10,10 @@ import (
 type ElementType int
 
 const (
+	// AlreadyInitialized is true if pserver is initialized
 	AlreadyInitialized = "pserver already initialized"
-	Uninitialized      = "pserver not fully initialized"
+	// Uninitialized is true if pserver not fully initialized
+	Uninitialized = "pserver not fully initialized"
 )
 
 // Supported element types
@@ -55,7 +57,7 @@ func NewService(idx int) (*Service, error) {
 	s := &Service{
 		idx: idx,
 	}
-  s.optMap = make(map[string]*optimizer)
+	s.optMap = make(map[string]*optimizer)
 	s.initialized = make(chan struct{})
 	return s, nil
 }

paddle/api/CMakeLists.txt

@@ -66,6 +66,7 @@ SWIG_LINK_LIBRARIES(swig_paddle
     paddle_trainer_lib
     paddle_network
     paddle_parameter
+    paddle_optimizer
     paddle_math
     paddle_utils
     paddle_proto

paddle/framework/CMakeLists.txt

@@ -9,6 +9,10 @@ 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_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf)
+cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_proto op_desc)
+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)

paddle/framework/attr_checker.h

+#pragma once
+
+#include <boost/variant.hpp>
+#include <functional>
+#include <string>
+#include <unordered_map>
+#include <vector>
+#include "paddle/framework/enforce.h"
+
+namespace paddle {
+namespace framework {
+
+typedef boost::variant<boost::blank, int, float, std::string, std::vector<int>,
+                       std::vector<float>, std::vector<std::string>>
+    Attribute;
+typedef std::unordered_map<std::string, Attribute> AttributeMap;
+
+// check whether a value(attribute) fit a certain limit
+template <typename T>
+class LargerThanChecker {
+ public:
+  LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
+  void operator()(T& value) const {
+    PADDLE_ENFORCE(value > lower_bound_, "larger_than check fail");
+  }
+
+ private:
+  T lower_bound_;
+};
+
+// we can provide users more common Checker, like 'LessThanChecker',
+// 'BetweenChecker'...
+
+template <typename T>
+class DefaultValueSetter {
+ public:
+  DefaultValueSetter(T default_value) : default_value_(default_value) {}
+  void operator()(T& value) const { value = default_value_; }
+
+ private:
+  T default_value_;
+};
+
+// check whether a certain attribute fit its limits
+// an attribute can have more than one limits
+template <typename T>
+class TypedAttrChecker {
+  typedef std::function<void(T&)> ValueChecker;
+
+ public:
+  TypedAttrChecker(const std::string& attr_name) : attr_name_(attr_name) {}
+
+  TypedAttrChecker& LargerThan(const T& lower_bound) {
+    value_checkers_.push_back(LargerThanChecker<T>(lower_bound));
+    return *this;
+  }
+
+  // we can add more common limits, like LessThan(), Between()...
+
+  TypedAttrChecker& SetDefault(const T& default_value) {
+    PADDLE_ENFORCE(default_value_setter_.empty(),
+                   "%s can't have more than one default value!", attr_name_);
+    default_value_setter_.push_back(DefaultValueSetter<T>(default_value));
+    return *this;
+  }
+
+  // allow users provide their own checker
+  TypedAttrChecker& AddCustomChecker(const ValueChecker& checker) {
+    value_checkers_.push_back(checker);
+    return *this;
+  }
+
+  void operator()(AttributeMap& attr_map) const {
+    if (!attr_map.count(attr_name_)) {
+      // user do not set this attr
+      PADDLE_ENFORCE(!default_value_setter_.empty(),
+                     "Attribute '%s' is required!", attr_name_);
+      // default_value_setter_ has no more than one element
+      T val;
+      (default_value_setter_[0])(val);
+      attr_map[attr_name_] = val;
+    }
+    Attribute& attr = attr_map.at(attr_name_);
+    T& attr_value = boost::get<T>(attr);
+    for (const auto& checker : value_checkers_) {
+      checker(attr_value);
+    }
+  }
+
+ private:
+  std::string attr_name_;
+  std::vector<ValueChecker> value_checkers_;
+  std::vector<ValueChecker> default_value_setter_;
+};
+
+// check whether op's all attributes fit their own limits
+class OpAttrChecker {
+  typedef std::function<void(AttributeMap&)> AttrChecker;
+
+ public:
+  template <typename T>
+  TypedAttrChecker<T>& AddAttrChecker(const std::string& attr_name) {
+    attr_checkers_.push_back(TypedAttrChecker<T>(attr_name));
+    AttrChecker& checker = attr_checkers_.back();
+    return *(checker.target<TypedAttrChecker<T>>());
+  }
+
+  void Check(AttributeMap& attr_map) const {
+    for (const auto& checker : attr_checkers_) {
+      checker(attr_map);
+    }
+  }
+
+ private:
+  std::vector<AttrChecker> attr_checkers_;
+};
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/op_registry.h

+#pragma once
+
+#include "paddle/framework/attr_checker.h"
+
+//#include "paddle/framework/op_base.h"
+#include "paddle/framework/op_desc.pb.h"
+#include "paddle/framework/op_proto.pb.h"
+
+namespace paddle {
+namespace framework {
+
+//==================For test================//
+class OpBase {
+ public:
+  std::vector<std::string> inputs_;
+  std::vector<std::string> outputs_;
+  AttributeMap attr_map_;
+
+  virtual std::string Run() const = 0;
+  virtual ~OpBase() {}
+};
+//=========================================//
+
+// helper class to set attribute type
+struct AttrTypeHelper {
+  template <typename T>
+  static void SetAttrType(AttrProto* attr);
+
+  static Attribute GetAttrValue(const AttrDesc& attr_desc) {
+    switch (attr_desc.type()) {
+      case paddle::framework::AttrType::INT: {
+        return attr_desc.i();
+      }
+      case paddle::framework::AttrType::FLOAT: {
+        return attr_desc.f();
+      }
+      case paddle::framework::AttrType::STRING: {
+        return attr_desc.s();
+      }
+      case paddle::framework::AttrType::INTS: {
+        std::vector<int> val(attr_desc.ints_size());
+        for (int i = 0; i < attr_desc.ints_size(); ++i) {
+          val[i] = attr_desc.ints(i);
+        }
+        return val;
+      }
+      case paddle::framework::AttrType::FLOATS: {
+        std::vector<float> val(attr_desc.floats_size());
+        for (int i = 0; i < attr_desc.floats_size(); ++i) {
+          val[i] = attr_desc.floats(i);
+        }
+        return val;
+      }
+      case paddle::framework::AttrType::STRINGS: {
+        std::vector<std::string> val(attr_desc.strings_size());
+        for (int i = 0; i < attr_desc.strings_size(); ++i) {
+          val[i] = attr_desc.strings(i);
+        }
+        return val;
+      }
+    }
+    PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !");
+    return boost::blank();
+  }
+};
+
+template <>
+void AttrTypeHelper::SetAttrType<int>(AttrProto* attr) {
+  attr->set_type(paddle::framework::AttrType::INT);
+}
+
+template <>
+void AttrTypeHelper::SetAttrType<float>(AttrProto* attr) {
+  attr->set_type(paddle::framework::AttrType::FLOAT);
+}
+
+template <>
+void AttrTypeHelper::SetAttrType<std::string>(AttrProto* attr) {
+  attr->set_type(paddle::framework::AttrType::STRING);
+}
+
+template <>
+void AttrTypeHelper::SetAttrType<std::vector<int>>(AttrProto* attr) {
+  attr->set_type(paddle::framework::AttrType::INTS);
+}
+
+template <>
+void AttrTypeHelper::SetAttrType<std::vector<float>>(AttrProto* attr) {
+  attr->set_type(paddle::framework::AttrType::FLOATS);
+}
+
+template <>
+void AttrTypeHelper::SetAttrType<std::vector<std::string>>(AttrProto* attr) {
+  attr->set_type(paddle::framework::AttrType::STRINGS);
+}
+
+// this class not only make proto but also init attribute checkers.
+class OpProtoAndCheckerMaker {
+ public:
+  OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : proto_(proto), op_checker_(op_checker) {}
+
+ protected:
+  void AddInput(const std::string& name, const std::string& comment) {
+    auto input = proto_->mutable_inputs()->Add();
+    *(input->mutable_name()) = name;
+    *(input->mutable_comment()) = comment;
+  }
+
+  void AddOutput(const std::string& name, const std::string& comment) {
+    auto output = proto_->mutable_outputs()->Add();
+    *(output->mutable_name()) = name;
+    *(output->mutable_comment()) = comment;
+  }
+
+  template <typename T>
+  TypedAttrChecker<T>& AddAttr(const std::string& name,
+                               const std::string& comment) {
+    auto attr = proto_->mutable_attrs()->Add();
+    *(attr->mutable_name()) = name;
+    *(attr->mutable_comment()) = comment;
+    AttrTypeHelper::SetAttrType<T>(attr);
+    return op_checker_->AddAttrChecker<T>(name);
+  }
+
+  void AddType(const std::string& op_type) { proto_->set_type(op_type); }
+
+  void AddComment(const std::string& comment) {
+    *(proto_->mutable_comment()) = comment;
+  }
+
+  OpProto* proto_;
+  OpAttrChecker* op_checker_;
+};
+
+class OpRegistry {
+  typedef std::function<OpBase*()> OpCreator;
+
+ public:
+  template <typename OpType, typename ProtoMakerType>
+  static void RegisterOp(const std::string& op_type) {
+    creators_[op_type] = []() { return new OpType; };
+    OpProto& op_proto = protos_[op_type];
+    OpAttrChecker& op_checker = op_checkers_[op_type];
+    ProtoMakerType(&op_proto, &op_checker);
+    PADDLE_ENFORCE(op_proto.IsInitialized() == true,
+                   "Fail to initialize %s's OpProto !", op_type);
+  }
+
+  static OpBase* CreateOp(const OpDesc& op_desc) {
+    std::string op_type = op_desc.type();
+    OpBase* op = (creators_.at(op_type))();
+    (op->inputs_).resize(op_desc.inputs_size());
+    for (int i = 0; i < op_desc.inputs_size(); ++i) {
+      (op->inputs_)[i] = op_desc.inputs(i);
+    }
+    (op->outputs_).resize(op_desc.outputs_size());
+    for (int i = 0; i < op_desc.outputs_size(); ++i) {
+      (op->outputs_)[i] = op_desc.outputs(i);
+    }
+    for (int i = 0; i < op_desc.attrs_size(); ++i) {
+      const AttrDesc& ith_attr = op_desc.attrs(i);
+      std::string name = ith_attr.name();
+      (op->attr_map_)[name] = AttrTypeHelper::GetAttrValue(ith_attr);
+    }
+    const OpAttrChecker& op_checker = op_checkers_.at(op_type);
+    op_checker.Check(op->attr_map_);
+    return op;
+  }
+
+ private:
+  static std::unordered_map<std::string, OpCreator> creators_;
+  static std::unordered_map<std::string, OpProto> protos_;
+  static std::unordered_map<std::string, OpAttrChecker> op_checkers_;
+};
+
+std::unordered_map<std::string, std::function<OpBase*()>> OpRegistry::creators_;
+std::unordered_map<std::string, OpProto> OpRegistry::protos_;
+std::unordered_map<std::string, OpAttrChecker> OpRegistry::op_checkers_;
+
+template <typename OpType, typename ProtoMakerType>
+class OpRegisterHelper {
+ public:
+  OpRegisterHelper(std::string op_type) {
+    OpRegistry::RegisterOp<OpType, ProtoMakerType>(op_type);
+  }
+};
+
+#define REGISTER_OP(__op_class, __op_maker_class, __op_type)         \
+  class __op_class##Register {                                       \
+   private:                                                          \
+    const static OpRegisterHelper<__op_class, __op_maker_class> reg; \
+  };                                                                 \
+  const OpRegisterHelper<__op_class, __op_maker_class>               \
+      __op_class##Register::reg(#__op_type);
+
+// Demos
+
+class CosineOp : public OpBase {
+ public:
+  virtual std::string Run() const {
+    std::string msg = "CosineOp runs! scale = " +
+                      std::to_string(boost::get<float>(attr_map_.at("scale")));
+    return msg;
+  }
+};
+
+class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
+ public:
+  CosineOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("input", "input of cosine op");
+    AddOutput("output", "output of cosine op");
+    AddAttr<float>("scale", "scale of cosine op")
+        .SetDefault(1.0)
+        .LargerThan(0.0);
+    AddType("cos");
+    AddComment("This is cos op");
+  }
+};
+
+REGISTER_OP(CosineOp, CosineOpProtoAndCheckerMaker, cos_sim)
+
+class MyTestOp : public OpBase {
+ public:
+  virtual std::string Run() const {
+    std::string msg =
+        "MyTestOp runs! test_attr = " +
+        std::to_string(boost::get<int>(attr_map_.at("test_attr")));
+    return msg;
+  }
+};
+
+class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
+ public:
+  MyTestOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("input", "input of cosine op");
+    AddOutput("output", "output of cosine op");
+    auto my_checker = [](int i) {
+      PADDLE_ENFORCE(i % 2 == 0, "'test_attr' must be even!");
+    };
+    AddAttr<int>("test_attr", "a simple test attribute")
+        .AddCustomChecker(my_checker);
+    AddType("my_test_op");
+    AddComment("This is my_test op");
+  }
+};
+
+REGISTER_OP(MyTestOp, MyTestOpProtoAndCheckerMaker, my_test_op)
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/op_registry_test.cc

+#include "paddle/framework/op_registry.h"
+#include <gtest/gtest.h>
+
+TEST(OpRegistry, CreateOp) {
+  paddle::framework::OpDesc op_desc;
+  op_desc.set_type("cos_sim");
+  op_desc.add_inputs("aa");
+  op_desc.add_outputs("bb");
+
+  auto attr = op_desc.mutable_attrs()->Add();
+  attr->set_name("scale");
+  attr->set_type(paddle::framework::AttrType::FLOAT);
+  attr->set_f(3.3);
+
+  paddle::framework::OpBase* op =
+      paddle::framework::OpRegistry::CreateOp(op_desc);
+  std::string debug_str = op->Run();
+  std::string str = "CosineOp runs! scale = " + std::to_string(3.3);
+  ASSERT_EQ(str.size(), debug_str.size());
+  for (size_t i = 0; i < debug_str.length(); ++i) {
+    ASSERT_EQ(debug_str[i], str[i]);
+  }
+}
+
+TEST(OpRegistry, IllegalAttr) {
+  paddle::framework::OpDesc op_desc;
+  op_desc.set_type("cos_sim");
+  op_desc.add_inputs("aa");
+  op_desc.add_outputs("bb");
+
+  auto attr = op_desc.mutable_attrs()->Add();
+  attr->set_name("scale");
+  attr->set_type(paddle::framework::AttrType::FLOAT);
+  attr->set_f(-2.0);
+
+  bool caught = false;
+  try {
+    paddle::framework::OpBase* op __attribute__((unused)) =
+        paddle::framework::OpRegistry::CreateOp(op_desc);
+  } catch (paddle::framework::EnforceNotMet err) {
+    caught = true;
+    std::string msg = "larger_than check fail";
+    const char* err_msg = err.what();
+    for (size_t i = 0; i < msg.length(); ++i) {
+      ASSERT_EQ(err_msg[i], msg[i]);
+    }
+  }
+  ASSERT_TRUE(caught);
+}
+
+TEST(OpRegistry, DefaultValue) {
+  paddle::framework::OpDesc op_desc;
+  op_desc.set_type("cos_sim");
+  op_desc.add_inputs("aa");
+  op_desc.add_outputs("bb");
+
+  paddle::framework::OpBase* op =
+      paddle::framework::OpRegistry::CreateOp(op_desc);
+  std::string debug_str = op->Run();
+  float default_value = 1.0;
+  std::string str = "CosineOp runs! scale = " + std::to_string(default_value);
+  ASSERT_EQ(str.size(), debug_str.size());
+  for (size_t i = 0; i < debug_str.length(); ++i) {
+    ASSERT_EQ(debug_str[i], str[i]);
+  }
+}
+
+TEST(OpRegistry, CustomChecker) {
+  paddle::framework::OpDesc op_desc;
+  op_desc.set_type("my_test_op");
+  op_desc.add_inputs("ii");
+  op_desc.add_outputs("oo");
+
+  // attr 'test_attr' is not set
+  bool caught = false;
+  try {
+    paddle::framework::OpBase* op __attribute__((unused)) =
+        paddle::framework::OpRegistry::CreateOp(op_desc);
+  } catch (paddle::framework::EnforceNotMet err) {
+    caught = true;
+    std::string msg = "Attribute 'test_attr' is required!";
+    const char* err_msg = err.what();
+    for (size_t i = 0; i < msg.length(); ++i) {
+      ASSERT_EQ(err_msg[i], msg[i]);
+    }
+  }
+  ASSERT_TRUE(caught);
+
+  // set 'test_attr' set to an illegal value
+  auto attr = op_desc.mutable_attrs()->Add();
+  attr->set_name("test_attr");
+  attr->set_type(paddle::framework::AttrType::INT);
+  attr->set_i(3);
+  caught = false;
+  try {
+    paddle::framework::OpBase* op __attribute__((unused)) =
+        paddle::framework::OpRegistry::CreateOp(op_desc);
+  } catch (paddle::framework::EnforceNotMet err) {
+    caught = true;
+    std::string msg = "'test_attr' must be even!";
+    const char* err_msg = err.what();
+    for (size_t i = 0; i < msg.length(); ++i) {
+      ASSERT_EQ(err_msg[i], msg[i]);
+    }
+  }
+  ASSERT_TRUE(caught);
+
+  // set 'test_attr' set to a legal value
+  op_desc.mutable_attrs()->Clear();
+  attr = op_desc.mutable_attrs()->Add();
+  attr->set_name("test_attr");
+  attr->set_type(paddle::framework::AttrType::INT);
+  attr->set_i(4);
+  paddle::framework::OpBase* op =
+      paddle::framework::OpRegistry::CreateOp(op_desc);
+  std::string debug_str = op->Run();
+  std::string str = "MyTestOp runs! test_attr = " + std::to_string(4);
+  ASSERT_EQ(str.size(), debug_str.size());
+  for (size_t i = 0; i < debug_str.length(); ++i) {
+    ASSERT_EQ(debug_str[i], str[i]);
+  }
+}
\ No newline at end of file

paddle/gserver/layers/AverageLayer.h

@@ -25,6 +25,10 @@ namespace paddle {
  * If SequenceLevel = kNonSeq:
  *    Output: output size is the number of input sequences (NOT input instances)
  *    output[i] = average_{for each instance in this sequence}{input[i]}
+ *    If stride_ > 0:
+ *      Output: a shorten sequence. Stride is the step size by which we slide a
+ *              window upon the input sequence, and the average pooling
+ *              operation is then applied to each interval independently.
  * If SequenceLevel = kSeq:
  *    Check input sequence must has sub-sequence
  *    Output: output size is the number of input sub-sequences

paddle/gserver/layers/CrossChannelNormLayer.cpp

@@ -36,6 +36,16 @@ MatrixPtr CrossChannelNormLayer::createSpatialMatrix(MatrixPtr data,
       data->getData() + iter * spatialDim, 1, spatialDim, false, useGpu_);
 }
 
+bool CrossChannelNormLayer::init(const LayerMap& layerMap,
+                                 const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+  CHECK(parameters_[0]);
+  const NormConfig& conf = config_.inputs(0).norm_conf();
+  channels_ = conf.channels();
+  scale_.reset(new Weight(channels_, 1, parameters_[0]));
+  return true;
+}
+
 void CrossChannelNormLayer::forward(PassType passType) {
   Layer::forward(passType);
   MatrixPtr inV = getInputValue(0);
@@ -51,9 +61,7 @@ void CrossChannelNormLayer::forward(PassType passType) {
   Matrix::resizeOrCreate(dataBuffer_, batchSize, dataDim, false, useGpu_);
   Matrix::resizeOrCreate(spatialBuffer_, 1, spatialDim, false, useGpu_);
   Matrix::resizeOrCreate(normBuffer_, batchSize, spatialDim, false, useGpu_);
-  normBuffer_->zeroMem();
-  // add eps to avoid overflow
-  normBuffer_->addScalar(*normBuffer_, 1e-6);
+
   inV->square2(*dataBuffer_);
   for (size_t i = 0; i < batchSize; i++) {
     const MatrixPtr inVTmp = createSampleMatrix(inV, i, spatialDim);
@@ -63,6 +71,8 @@ void CrossChannelNormLayer::forward(PassType passType) {
 
     // compute norm.
     spatialBuffer_->sumCols(*dataTmp, 1, 0);
+    // add eps to avoid overflow
+    spatialBuffer_->add(1e-6);
     spatialBuffer_->sqrt2(*spatialBuffer_);
     normTmp->copyFrom(*spatialBuffer_);
     outVTmp->copyFrom(*inVTmp);
@@ -82,6 +92,9 @@ void CrossChannelNormLayer::backward(const UpdateCallback& callback) {
   size_t dataDim = inG->getWidth();
   size_t spatialDim = dataDim / channels_;
 
+  MatrixPtr inGBuffer;
+  Matrix::resizeOrCreate(inGBuffer, channels_, spatialDim, false, useGpu_);
+
   dataBuffer_->dotMul(*outG, *outV);
   Matrix::resizeOrCreate(scaleDiff_, channels_, 1, false, useGpu_);
   Matrix::resizeOrCreate(channelBuffer_, channels_, 1, false, useGpu_);
@@ -100,22 +113,24 @@ void CrossChannelNormLayer::backward(const UpdateCallback& callback) {
     scaleDiff_->add(*channelBuffer_, 1.);
 
     sampleBuffer_->dotMul(*inVTmp, *outGTmp);
-    spatialBuffer_->sumCols(*sampleBuffer_, 1., 1.);
+    spatialBuffer_->sumCols(*sampleBuffer_, 1., 0.);
     // scale the grad
-    inGTmp->copyFrom(*inVTmp);
-    inGTmp->mulRowVector(*spatialBuffer_);
+    inGBuffer->copyFrom(*inVTmp);
+    inGBuffer->mulRowVector(*spatialBuffer_);
     // divide by square of norm
     spatialBuffer_->dotMul(*normTmp, *normTmp);
-    inGTmp->divRowVector(*spatialBuffer_);
+    inGBuffer->divRowVector(*spatialBuffer_);
     // subtract
-    inGTmp->add(*outGTmp, -1, 1);
+    inGBuffer->add(*outGTmp, -1, 1);
     // divide by norm
-    inGTmp->divRowVector(*normTmp);
+    inGBuffer->divRowVector(*normTmp);
     // scale the diff
-    inGTmp->mulColVector(*scale_->getW());
+    inGBuffer->mulColVector(*scale_->getW());
+
+    inGTmp->add(*inGBuffer);
   }
   // updata scale
-  if (scale_->getWGrad()) scale_->getWGrad()->copyFrom(*scaleDiff_);
+  if (scale_->getWGrad()) scale_->getWGrad()->add(*scaleDiff_);
   scale_->getParameterPtr()->incUpdate(callback);
 }
 

paddle/gserver/layers/MaxLayer.h

@@ -26,6 +26,10 @@ namespace paddle {
  * If SequenceLevel = kNonSeq:
  *    Output: output size is the number of input sequences (NOT input instances)
  *    output[i] = max_{for each instance in this sequence}{input[i]}
+ *    If stride_ > 0:
+ *      Output: a shorten sequence. Stride is the step size by which we slide a
+ *              window upon the input sequence, and the max pooling operation is
+ *              then applied to each interval independently.
  * If SequenceLevel = kSeq:
  *    Check input sequence must has sub-sequence
  *    Output: output size is the number of input sub-sequences

paddle/gserver/layers/NormLayer.cpp

@@ -56,14 +56,4 @@ bool ResponseNormLayer::init(const LayerMap& layerMap,
   return true;
 }
 
-bool CrossChannelNormLayer::init(const LayerMap& layerMap,
-                                 const ParameterMap& parameterMap) {
-  Layer::init(layerMap, parameterMap);
-  CHECK(parameters_[0]);
-  const NormConfig& conf = config_.inputs(0).norm_conf();
-  channels_ = conf.channels();
-  scale_.reset(new Weight(channels_, 1, parameters_[0]));
-  return true;
-}
-
 }  // namespace paddle

paddle/gserver/layers/SequenceLastInstanceLayer.cpp

@@ -26,10 +26,9 @@ namespace paddle {
  * If SequenceLevel = kNonseq:
  *   Output: a sequence containing only the last instance of the input sequence
  *   If stride_ > 0:
- *      Output: a shorten sequence. The operation of getting last instance of a
- *              sequence is independently performed on every slice of the input
- *              sequence, which is obtained by sliding a window with the window
- *              size set to stride_.
+ *      Output: a shorten sequence. Stride is the step size by which we slide a
+ *              window upon the input sequence, and getting last instance
+ *              operation is then applied to each interval independently.
  * If SequenceLevel = kSeq:
  *   Check input sequence must has sub-sequence
  *   Output: a sequence containing only the last instance of each sub-sequence
@@ -73,8 +72,7 @@ bool SequenceLastInstanceLayer::init(const LayerMap& layerMap,
 void SequenceLastInstanceLayer::forward(PassType passType) {
   SequencePoolLayer::forward(passType);
 
-  auto starts = (stride_ > 0) ? stridePositions_->getData()
-                              : startPositions_->getData(false);
+  auto starts = startPositions_->getData(false);
   MatrixPtr inputValue = getInputValue(0);
   MatrixPtr outputValue = getOutputValue();
 

paddle/gserver/layers/SequencePoolLayer.cpp

@@ -72,9 +72,8 @@ void SequencePoolLayer::forward(PassType passType) {
   if (stride_ > 0) {
     CHECK_EQ(input.hasSubseq(), 0UL)
         << "sequence stride pooling is invalid for hasSubseq now";
-    output_.poolSequenceWithStride(
-        input, stride_, &stridePositions_, reversed_);
-    newBatchSize_ = stridePositions_->getSize() - 1;
+    output_.poolSequenceWithStride(input, stride_, &startPositions_, reversed_);
+    newBatchSize_ = startPositions_->getSize() - 1;
   }
 
   resetOutput(newBatchSize_, dim);

paddle/gserver/layers/SequencePoolLayer.h

@@ -28,8 +28,9 @@ namespace paddle {
  * sequence}{input[i]}
  *    If stride_ > 0:
  *        Check input sequence must not have sub-sequence
- *        Output: a shorten sequence, pooling is performed upon a small local
- *                area
+ *        Output: a shorten sequence. Stride is the step size by which we slide
+ *                a window upon the input sequence, and the pooling operation
+ *                is then applied to each interval independently.
  * If SequenceLevel = kSeq:
  *    Check input sequence must has sub-sequence
  *    Output: output size is the number of input sub-sequences
@@ -47,8 +48,6 @@ protected:
   size_t newBatchSize_;
   ICpuGpuVectorPtr startPositions_;
   int stride_;
-  // Store the start position of each window.
-  IVectorPtr stridePositions_;
   // Whether the input sequence is reversed or not.
   bool reversed_ = false;
 

paddle/gserver/tests/LayerGradUtil.cpp

@@ -241,7 +241,7 @@ void testBatchState(LayerPtr testLayer,
 
     std::vector<Argument> args;
     args.push_back(out);
-    EXPECT_EQ(0, Argument::sum(args)) << "testBatchState failed";
+    ASSERT_NEAR(0, Argument::sum(args), 1e-5) << "testBatchState failed";
     for (size_t seqId = 0; seqId < numSequences; ++seqId) {
       start[seqId] += seqLens[seqId];
     }
@@ -465,7 +465,6 @@ void initTestLayer(TestConfig testConf,
                            ParameterConfig paraConfig) {
     paraConfig.set_name(paraName);
     paraConfig.set_size(paraSize);
-    paraConfig.set_initial_std(1);
     paraConfig.set_is_static(isStatic);
     auto para =
         std::make_shared<Parameter>(paraConfig, FLAGS_use_gpu, initialize);
@@ -499,6 +498,9 @@ void initTestLayer(TestConfig testConf,
         paraConfig.add_dims((*layerMap)[input.input_layer_name()]->getSize());
         paraConfig.add_dims(testConf.layerConfig.size());
       }
+      CHECK_GE(testConf.paramInitialStd, 0);
+      paraConfig.set_initial_mean(testConf.paramInitialMean);
+      paraConfig.set_initial_std(testConf.paramInitialStd);
       initParameter(paraName, paraSize, inputDef.isStatic, false, paraConfig);
     }
   }

paddle/gserver/tests/LayerGradUtil.h

@@ -125,12 +125,16 @@ struct TestConfig {
   LayerConfig layerConfig;
   std::vector<InputDef> inputDefs;
   size_t biasSize;
+  real paramInitialMean;
+  real paramInitialStd;
   bool testAccumulate;
   bool testState;
   bool staticBias;
   bool testBatchState;
   TestConfig()
       : biasSize(0),
+        paramInitialMean(0.0),
+        paramInitialStd(1.0),
         testAccumulate(true),
         testState(false),
         staticBias(false),

paddle/gserver/tests/test_LayerGrad.cpp

@@ -845,8 +845,12 @@ void testDegradeLayer(bool hasSubseq,
 
 TEST(Layer, MaxLayer) {
   testDegradeLayer(false, "max", "non-seq", -1);  // seq max to non-seq
-  testDegradeLayer(true, "max", "non-seq", -1);   // hasSubseq max to non-seq
-  testDegradeLayer(true, "max", "seq", -1);       // hasSubseq max to seq
+  testDegradeLayer(false,
+                   "max",
+                   "non-seq",
+                   5);  // seq max to a shorten seq, stride window = 5
+  testDegradeLayer(true, "max", "non-seq", -1);  // hasSubseq max to non-seq
+  testDegradeLayer(true, "max", "seq", -1);      // hasSubseq max to seq
 }
 
 TEST(Layer, SequenceLastInstanceLayer) {
@@ -868,6 +872,10 @@ TEST(Layer, SequenceLastInstanceLayer) {
 
 TEST(Layer, AverageLayer) {
   testDegradeLayer(false, "average", "non-seq", -1);  // seq average to non-seq
+  testDegradeLayer(false,
+                   "average",
+                   "non-seq",
+                   5);  // seq average to a shorten seq, stride window = 5
   testDegradeLayer(
       true, "average", "non-seq", -1);           // hasSubseq average to non-seq
   testDegradeLayer(true, "average", "seq", -1);  // hasSubseq average to seq
@@ -1661,6 +1669,8 @@ TEST(Layer, PadLayer) {
 
 TEST(Layer, CrossChannelNormLayer) {
   TestConfig config;
+  config.paramInitialMean = 1.;
+  config.paramInitialStd = 0.;
   config.layerConfig.set_type("norm");
   config.layerConfig.set_size(100);
   LayerInputConfig* input = config.layerConfig.add_inputs();
@@ -1674,7 +1684,7 @@ TEST(Layer, CrossChannelNormLayer) {
   config.inputDefs.push_back({INPUT_DATA, "layer_0", 100, 10});
 
   for (auto useGpu : {false, true}) {
-    testLayerGrad(config, "cross-channel-norm", 10, false, useGpu, false, 5);
+    testLayerGrad(config, "cross-channel-norm", 10, false, useGpu, false);
   }
 }
 

paddle/optimizer/adadelta_optimizer.cc

@@ -27,22 +27,24 @@ void AdadeltaOptimizer::Update(const Tensor* gradient) {
 
 const char* AdadeltaOptimizer::SerializeState(int* state_len) {
   AdadeltaOptimizerState state;
-  // TODO(zhihong) : add lr_policy serialization
   state.set_num_sample_passed(num_sample_passed_);
+  std::string lr_str = this->lr_policy_->SerializeState(state_len);
+  state.mutable_lr_state()->ParseFromString(lr_str);
 
   TensorToProto(*parameter_, state.mutable_parameter());
   TensorToProto(*accum_gradient_, state.mutable_accum_gradient());
   TensorToProto(*accum_delta_, state.mutable_accum_delta());
   TensorToProto(*update_delta_, state.mutable_update_delta());
   auto str = state.SerializeAsString();
-  *state_len = str.size();
+  *state_len += str.size();
   return str.c_str();
 }
 
 void AdadeltaOptimizer::DeserializeState(const std::string& str) {
   AdadeltaOptimizerState state;
   state.ParseFromString(str);
-  // TODO(zhihong) : add lr_policy DeserializeState
+  auto lr_state = state.lr_state();
+  this->lr_policy_->DeserializeState(lr_state.SerializeAsString());
   num_sample_passed_ = state.num_sample_passed();
 
   ProtoToTensor(state.parameter(), parameter_);

paddle/optimizer/adagrad_optimizer.cc

@@ -19,20 +19,23 @@ void AdagradOptimizer::Update(const Tensor* gradient) {
 }
 const char* AdagradOptimizer::SerializeState(int* state_len) {
   AdagradOptimizerState state;
-  // TODO(zhihong) : add lr_policy serialization
   state.set_num_sample_passed(num_sample_passed_);
+  std::string lr_str = this->lr_policy_->SerializeState(state_len);
+  state.mutable_lr_state()->ParseFromString(lr_str);
 
   TensorToProto(*parameter_, state.mutable_parameter());
   TensorToProto(*accum_gradient_, state.mutable_accum_gradient());
   auto str = state.SerializeAsString();
-  *state_len = str.size();
+  *state_len += str.size();
   return str.c_str();
 }
 
 void AdagradOptimizer::DeserializeState(const std::string& str) {
   AdagradOptimizerState state;
   state.ParseFromString(str);
-  // TODO(zhihong) : add lr_policy DeserializeState
+  auto lr_state = state.lr_state();
+  this->lr_policy_->DeserializeState(lr_state.SerializeAsString());
+
   num_sample_passed_ = state.num_sample_passed();
   ProtoToTensor(state.parameter(), parameter_);
   ProtoToTensor(state.accum_gradient(), accum_gradient_);

paddle/optimizer/adam_optimizer.cc

@@ -24,20 +24,23 @@ void AdamOptimizer::Update(const Tensor *gradient) {
 
 const char *AdamOptimizer::SerializeState(int *state_len) {
   AdamOptimizerState state;
-  // TODO(zhihong) : add lr_policy serialization
+  std::string lr_str = this->lr_policy_->SerializeState(state_len);
+  state.mutable_lr_state()->ParseFromString(lr_str);
   state.set_num_sample_passed(num_sample_passed_);
+
   TensorToProto(*parameter_, state.mutable_parameter());
   TensorToProto(*momentums_, state.mutable_momentums());
   TensorToProto(*velocitys_, state.mutable_velocitys());
   auto str = state.SerializeAsString();
-  *state_len = str.size();
+  *state_len += str.size();
   return str.c_str();
 }
 
 void AdamOptimizer::DeserializeState(const std::string &str) {
   AdamOptimizerState state;
   state.ParseFromString(str);
-  // TODO(zhihong) : add lr_policy DeserializeState
+  auto lr_state = state.lr_state();
+  this->lr_policy_->DeserializeState(lr_state.SerializeAsString());
   num_sample_passed_ = state.num_sample_passed();
 
   ProtoToTensor(state.parameter(), parameter_);

paddle/optimizer/lr_policy.h

@@ -17,36 +17,56 @@ public:
 // constant learning rate policy
 class ConstLr final : public LrPolicy {
 public:
-  ConstLr(double lr) : learning_rate(lr){};
+  ConstLr(double lr) : learning_rate_(lr){};
   double LearningRate(const uint64_t num_sample_passed) {
-    return learning_rate;
+    return learning_rate_;
+  }
+  const char *SerializeState(int *state_len) {
+    LrPolicyState state;
+    state.set_learning_rate(learning_rate_);
+    auto str = state.SerializeAsString();
+    *state_len = str.size();
+    return str.c_str();
+  }
+  void DeserializeState(const std::string &str) {
+    LrPolicyState state;
+    state.ParseFromString(str);
+    learning_rate_ = state.learning_rate();
   }
-  const char *SerializeState(int *state_len) { return nullptr; }
-  void DeserializeState(const std::string &state) {}
 
 private:
-  double learning_rate;
+  double learning_rate_;
 };
 
 class LinearLr final : public LrPolicy {
 public:
   LinearLr(double lr, double lr_decay_a, double lr_decay_b)
-      : learning_rate(lr), lr_decay_a(lr_decay_a), lr_decay_b(lr_decay_b) {}
+      : learning_rate_(lr), lr_decay_a_(lr_decay_a), lr_decay_b_(lr_decay_b) {}
   double LearningRate(const uint64_t num_sample_passed) {
-    return std::max(learning_rate - lr_decay_a * num_sample_passed, lr_decay_b);
+    return std::max(learning_rate_ - lr_decay_a_ * num_sample_passed,
+                    lr_decay_b_);
   }
   const char *SerializeState(int *state_len) {
-    // TODO(zhihong) : add lr_policy serialization
-    return nullptr;
+    LrPolicyState state;
+    state.set_learning_rate(learning_rate_);
+    state.set_lr_decay_a(lr_decay_a_);
+    state.set_lr_decay_b(lr_decay_b_);
+    auto str = state.SerializeAsString();
+    *state_len = str.size();
+    return str.c_str();
   }
-  void DeserializeState(const std::string &state) {
-    // TODO(zhihong) : add lr_policy serialization
+  void DeserializeState(const std::string &str) {
+    LrPolicyState state;
+    state.ParseFromString(str);
+    learning_rate_ = state.learning_rate();
+    lr_decay_a_ = state.lr_decay_a();
+    lr_decay_b_ = state.lr_decay_b();
   }
 
 private:
-  double learning_rate;
-  double lr_decay_a;
-  double lr_decay_b;
+  double learning_rate_;
+  double lr_decay_a_;
+  double lr_decay_b_;
 };
 
 }  // namespace optimizer

paddle/optimizer/sgd_optimizer.cc

@@ -30,16 +30,20 @@ void SGDOptimizer::Update(const Tensor *gradient) {
 const char *SGDOptimizer::SerializeState(int *state_len) {
   SGDOptimizerState state;
   state.set_num_sample_passed(num_sample_passed_);
+  std::string lr_str = this->lr_policy_->SerializeState(state_len);
+  state.mutable_lr_state()->ParseFromString(lr_str);
   TensorToProto(*parameter_, state.mutable_parameter());
   if (momentum_ != 0.0) TensorToProto(*momentums_, state.mutable_momentums());
   auto str = state.SerializeAsString();
-  *state_len = str.size();
+  *state_len += str.size();
   return str.c_str();
 }
 
 void SGDOptimizer::DeserializeState(const std::string &str) {
   SGDOptimizerState state;
   state.ParseFromString(str);
+  auto lr_state = state.lr_state();
+  this->lr_policy_->DeserializeState(lr_state.SerializeAsString());
   num_sample_passed_ = state.num_sample_passed();
   ProtoToTensor(state.parameter(), parameter_);
   if (momentum_ != 0.0) ProtoToTensor(state.parameter(), momentums_);

paddle/parameter/Argument.cpp

@@ -561,7 +561,7 @@ void Argument::degradeSequence(const Argument& input) {
 
 void Argument::poolSequenceWithStride(const Argument& input,
                                       size_t stride,
-                                      IVectorPtr* stridePostions,
+                                      ICpuGpuVectorPtr* stridePostions,
                                       bool reversed) {
   // If input.sequenceStartPositions = [0, 9, 14, 17, 30] and stride = 5,
   // then sequenceStartPositions = [0, 2, 3, 4, 7].
@@ -598,8 +598,8 @@ void Argument::poolSequenceWithStride(const Argument& input,
   stridePos.emplace_back(starts[numSequences]);
   int size = stridePos.size();
   CHECK_EQ(size - 1, tgtBuf[numSequences]);
-  IVector::resizeOrCreate(*stridePostions, size, false);
-  (*stridePostions)->copyFrom(stridePos.data(), size);
+  ICpuGpuVector::resizeOrCreate(*stridePostions, size, false);
+  (*stridePostions)->getMutableVector(false)->copyFrom(stridePos.data(), size);
 }
 
 void Argument::getValueString(

paddle/parameter/Argument.h

@@ -299,7 +299,7 @@ struct Argument {
    */
   void poolSequenceWithStride(const Argument& input,
                               size_t stride,
-                              IVectorPtr* stridePositions,
+                              ICpuGpuVectorPtr* stridePositions,
                               bool reversed = false);
   /**
    * @brief getValueString will return the argument's output in string. There

paddle/parameter/tests/test_argument.cpp

@@ -31,7 +31,7 @@ TEST(Argument, poolSequenceWithStride) {
   int strideResultReversed[] = {0, 4, 9, 14, 17, 20, 25, 30};
 
   for (auto reversed : {false, true}) {
-    IVectorPtr stridePositions;
+    ICpuGpuVectorPtr stridePositions;
     output.poolSequenceWithStride(
         input, 5 /* stride */, &stridePositions, reversed);
 
@@ -45,7 +45,7 @@ TEST(Argument, poolSequenceWithStride) {
     CHECK_EQ(stridePositions->getSize(), 8UL);
     auto result = reversed ? strideResultReversed : strideResult;
     for (int i = 0; i < 8; i++) {
-      CHECK_EQ(stridePositions->getData()[i], result[i]);
+      CHECK_EQ(stridePositions->getData(false)[i], result[i]);
     }
   }
 }

paddle/pserver/LightNetwork.cpp

@@ -142,7 +142,7 @@ SocketServer::SocketServer(const std::string &addr, int port, int rdmaCpu)
   }
 
   /// trigger to initialize RDMA lib
-  PCHECK(RdmaClientDaemons::get()) << "initilizate RDMA failed\n";
+  CHECK(RdmaClientDaemons::get()) << "initilizate RDMA failed\n";
 }
 
 SocketServer::~SocketServer() {
@@ -168,7 +168,7 @@ void SocketServer::tcpServer() {
 
   /// First call to socket() function
   socket_ = socket(AF_INET, SOCK_STREAM, 0);
-  PCHECK(socket_ >= 0) << "ERROR opening socket";
+  CHECK(socket_ >= 0) << "ERROR opening socket";
 
   /// Initialize socket structure
   bzero((char *)&serv_addr, sizeof(serv_addr));
@@ -176,7 +176,7 @@ void SocketServer::tcpServer() {
   serv_addr.sin_port = htons(port_);
   if (!addr_.empty()) {
     server = gethostbyname(addr_.c_str());
-    PCHECK(server) << "ERROR, no such host: " << addr_;
+    CHECK(server) << "ERROR, no such host: " << addr_;
     bcopy((char *)server->h_addr,
           (char *)&serv_addr.sin_addr.s_addr,
           server->h_length);
@@ -187,7 +187,7 @@ void SocketServer::tcpServer() {
   setOption(socket_);
 
   /// Now bind the host address using bind() call.
-  PCHECK(bind(socket_, (struct sockaddr *)&serv_addr, sizeof(serv_addr)) >= 0)
+  CHECK(bind(socket_, (struct sockaddr *)&serv_addr, sizeof(serv_addr)) >= 0)
       << "ERROR on binding " << addr_;
 
   /// Now start listening for the clients, here process will
@@ -201,7 +201,7 @@ void SocketServer::tcpServer() {
     if (stopping_) {
       break;
     }
-    PCHECK(newsockfd >= 0) << "ERROR on accept";
+    CHECK(newsockfd >= 0) << "ERROR on accept";
     constexpr int kPeerNameLen = 128;
     char peerName[kPeerNameLen];
     CHECK(inet_ntop(AF_INET, &cli_addr.sin_addr, peerName, kPeerNameLen));
@@ -227,14 +227,14 @@ void SocketServer::rdmaServer() {
 
   /// First call to socket() function
   rdmaSocket_ = rdma::ssocket(rdmaCpu_);
-  PCHECK(rdmaSocket_) << "ERROR opening RDMA socket";
+  CHECK(rdmaSocket_) << "ERROR opening RDMA socket";
 
-  PCHECK(rdma::bind(rdmaSocket_, rdmaUri_.c_str()) == 0)
+  CHECK(rdma::bind(rdmaSocket_, rdmaUri_.c_str()) == 0)
       << "ERROR bind RDMA socket";
 
   /// Now start listening for the clients, here process will
   /// go in sleep mode and will wait for the incoming connection
-  PCHECK(rdma::listen(rdmaSocket_) == 0) << "ERROR listen RDMA socket";
+  CHECK(rdma::listen(rdmaSocket_) == 0) << "ERROR listen RDMA socket";
 
   while (true) {
     /// Accept actual connection from the client
@@ -242,7 +242,7 @@ void SocketServer::rdmaServer() {
     if (stopping_) {
       break;
     }
-    PCHECK(newsock) << "ERROR on accept";
+    CHECK(newsock) << "ERROR on accept";
 
     constexpr int kPeerNameLen = 128;
     char peerName[kPeerNameLen];
@@ -290,7 +290,7 @@ RdmaClientDaemons::RdmaClientDaemons() {
     onlineCpus_ = rdma::numCpus();
     for (auto i = 0; i < onlineCpus_; i++) {
       socket = rdma::csocket(i);
-      PCHECK(socket) << "ERROR open client socket daemon";
+      CHECK(socket) << "ERROR open client socket daemon";
 
       rdmaClientSocket_.push_back(socket);
     }
@@ -355,7 +355,7 @@ void SocketClient::TcpClient(const std::string &serverAddr, int serverPort) {
 
   /// Create a socket point
   int sockfd = socket(AF_INET, SOCK_STREAM, 0);
-  PCHECK(sockfd >= 0) << "ERROR opening socket";
+  CHECK(sockfd >= 0) << "ERROR opening socket";
 
 #if defined(__OSX__) || defined(__APPLE__)
   server = getipnodebyname(serverAddr.c_str(), AF_INET, AI_DEFAULT, &errRet);
@@ -396,8 +396,8 @@ void SocketClient::TcpClient(const std::string &serverAddr, int serverPort) {
       }
       std::this_thread::sleep_for(std::chrono::seconds(1));
     } else {
-      PCHECK(errno != 0) << "ERROR connecting to " << serverAddr << ":"
-                         << serverPort << "errorno: " << errno;
+      CHECK(errno != 0) << "ERROR connecting to " << serverAddr << ":"
+                        << serverPort << "errorno: " << errno;
     }
   } while (errno == ECONNREFUSED);
 
@@ -426,7 +426,7 @@ void SocketClient::RdmaClient(const std::string &serverAddr, int serverPort) {
 
   /// connect to server with socket daemon
   sock = rdma::connect(socketDaemon_, rdmaUri.c_str());
-  PCHECK(sock) << "ERROR connect to server" << rdmaUri;
+  CHECK(sock) << "ERROR connect to server" << rdmaUri;
 
   std::vector<std::string> seg;
   str::split(rdmaUri, '/', &seg);

paddle/pserver/SocketChannel.cpp

@@ -51,7 +51,7 @@ size_t SocketChannel::read(void* buf, size_t size) {
     else
       len = rdma::read(rdmaSocket_, (char*)buf + total, size - total);
 
-    PCHECK(len >= 0) << " peer=" << peerName_;
+    CHECK(len >= 0) << " peer=" << peerName_;
     if (len <= 0) {
       return total;
     }
@@ -69,7 +69,7 @@ size_t SocketChannel::write(const void* buf, size_t size) {
     else
       len = rdma::write(rdmaSocket_, (char*)buf + total, size - total);
 
-    PCHECK(len >= 0) << " peer=" << peerName_;
+    CHECK(len >= 0) << " peer=" << peerName_;
     if (len <= 0) {
       return total;
     }
@@ -98,10 +98,10 @@ static size_t readwritev(IOFunc iofunc,
   while (size < total) {
     ssize_t len =
         iofunc(socket, &iovs[curIov], std::min(iovcnt - curIov, maxiovs));
-    PCHECK(len > 0) << " peer=" << peerName << " curIov=" << curIov
-                    << " iovCnt=" << iovcnt
-                    << " iovs[curIov].base=" << iovs[curIov].iov_base
-                    << " iovs[curIov].iov_len=" << iovs[curIov].iov_len;
+    CHECK(len > 0) << " peer=" << peerName << " curIov=" << curIov
+                   << " iovCnt=" << iovcnt
+                   << " iovs[curIov].base=" << iovs[curIov].iov_base
+                   << " iovs[curIov].iov_len=" << iovs[curIov].iov_len;
     size += len;
 
     /// restore iovs[curIov] to the original value
@@ -183,7 +183,7 @@ void SocketChannel::writeMessage(const std::vector<struct iovec>& userIovs) {
     header.totalLength += iov.iov_len;
   }
 
-  PCHECK(writev(iovs) == (size_t)header.totalLength);
+  CHECK(writev(iovs) == (size_t)header.totalLength);
 }
 
 std::unique_ptr<MsgReader> SocketChannel::readMessage() {
@@ -194,7 +194,7 @@ std::unique_ptr<MsgReader> SocketChannel::readMessage() {
     return nullptr;
   }
 
-  PCHECK(len == sizeof(header));
+  CHECK(len == sizeof(header));
 
   std::unique_ptr<MsgReader> msgReader(new MsgReader(this, header.numIovs));
 
@@ -209,7 +209,7 @@ std::unique_ptr<MsgReader> SocketChannel::readMessage() {
 MsgReader::MsgReader(SocketChannel* channel, size_t numBlocks)
     : channel_(channel), blockLengths_(numBlocks), currentBlockIndex_(0) {
   size_t size = numBlocks * sizeof(blockLengths_[0]);
-  PCHECK(channel_->read(&blockLengths_[0], size) == size);
+  CHECK(channel_->read(&blockLengths_[0], size) == size);
 }
 
 void MsgReader::readBlocks(const std::vector<void*>& bufs) {
@@ -223,12 +223,12 @@ void MsgReader::readBlocks(const std::vector<void*>& bufs) {
     ++currentBlockIndex_;
   }
 
-  PCHECK(channel_->readv(&iovs) == totalLength);
+  CHECK(channel_->readv(&iovs) == totalLength);
 }
 
 void MsgReader::readNextBlock(void* buf) {
   CHECK_LT(currentBlockIndex_, blockLengths_.size());
-  PCHECK(channel_->read(buf, getNextBlockLength()) == getNextBlockLength());
+  CHECK(channel_->read(buf, getNextBlockLength()) == getNextBlockLength());
   ++currentBlockIndex_;
 }
 

paddle/pserver/test/SocketTest.cpp

@@ -113,7 +113,7 @@ void SocketServer::run() {
 
   /* First call to socket() function */
   socket_ = socket(AF_INET, SOCK_STREAM, 0);
-  PCHECK(socket_ >= 0) << "ERROR opening socket";
+  CHECK(socket_ >= 0) << "ERROR opening socket";
 
   /* Initialize socket structure */
   bzero((char*)&serv_addr, sizeof(serv_addr));
@@ -122,7 +122,7 @@ void SocketServer::run() {
   serv_addr.sin_port = htons(port_);
 
   /* Now bind the host address using bind() call.*/
-  PCHECK(bind(socket_, (struct sockaddr*)&serv_addr, sizeof(serv_addr)) >= 0)
+  CHECK(bind(socket_, (struct sockaddr*)&serv_addr, sizeof(serv_addr)) >= 0)
       << "ERROR on binding";
 
   /* Now start listening for the clients, here process will
@@ -134,7 +134,7 @@ void SocketServer::run() {
   while (true) {
     /* Accept actual connection from the client */
     newsockfd = accept(socket_, (struct sockaddr*)&cli_addr, &clilen);
-    PCHECK(newsockfd >= 0) << "ERROR on accept";
+    CHECK(newsockfd >= 0) << "ERROR on accept";
 
     SocketWorker* worker = new SocketWorker(newsockfd);
     worker->start();
@@ -146,17 +146,17 @@ void SocketWorker::run() {
 
   while (true) {
     int64_t n = channel_.readAll(&header, sizeof(header));
-    PCHECK(n == sizeof(header)) << "ERROR reading from socket";
+    CHECK(n == sizeof(header)) << "ERROR reading from socket";
 
     buffer_.resize(header.dataLength);
     n = channel_.readAll(&buffer_[0], header.dataLength);
-    PCHECK(n == header.dataLength) << "ERROR reading from socket";
+    CHECK(n == header.dataLength) << "ERROR reading from socket";
 
     /* Write a response to the client */
     n = channel_.writeAll(&header, sizeof(header));
-    PCHECK(n == sizeof(header)) << "ERROR reading from socket";
+    CHECK(n == sizeof(header)) << "ERROR reading from socket";
     n = channel_.writeAll(buffer_.data(), buffer_.size());
-    PCHECK(n == header.dataLength) << "ERROR writing to socket";
+    CHECK(n == header.dataLength) << "ERROR writing to socket";
   }
 }
 
@@ -177,9 +177,9 @@ SocketClient::SocketClient(const std::string& serverAddr, int serverPort) {
 
   /* Create a socket point */
   int sockfd = socket(AF_INET, SOCK_STREAM, 0);
-  PCHECK(sockfd >= 0) << "ERROR opening socket";
+  CHECK(sockfd >= 0) << "ERROR opening socket";
   server = gethostbyname(serverAddr.c_str());
-  PCHECK(server) << "ERROR, no such host: " << serverAddr;
+  CHECK(server) << "ERROR, no such host: " << serverAddr;
 
   bzero((char*)&serv_addr, sizeof(serv_addr));
   serv_addr.sin_family = AF_INET;
@@ -189,7 +189,7 @@ SocketClient::SocketClient(const std::string& serverAddr, int serverPort) {
   serv_addr.sin_port = htons(serverPort);
 
   /* Now connect to the server */
-  PCHECK(connect(sockfd, (sockaddr*)&serv_addr, sizeof(serv_addr)) >= 0)
+  CHECK(connect(sockfd, (sockaddr*)&serv_addr, sizeof(serv_addr)) >= 0)
       << "ERROR connecting";
 
   channel_.reset(new SocketChannel(sockfd));
@@ -234,18 +234,18 @@ int main(int argc, char** argv) {
     cpuGrad.copyFrom(gpuGrad);
 
     header.dataLength = dataSize;
-    PCHECK(channel->writeAll(&header, sizeof(header)) == sizeof(header))
+    CHECK(channel->writeAll(&header, sizeof(header)) == sizeof(header))
         << "Client write header error";
 
-    PCHECK(channel->writeAll(cpuGrad.getData(), dataSize) == dataSize)
+    CHECK(channel->writeAll(cpuGrad.getData(), dataSize) == dataSize)
         << "Client write data error";
 
     /* Now read server response */
-    PCHECK(channel->readAll(&header, sizeof(header)) == sizeof(header))
+    CHECK(channel->readAll(&header, sizeof(header)) == sizeof(header))
         << "Client read header error";
 
     CHECK_EQ((uint64_t)header.dataLength, dataSize);
-    PCHECK(channel->readAll(cpuParam.getData(), dataSize) == dataSize)
+    CHECK(channel->readAll(cpuParam.getData(), dataSize) == dataSize)
         << "Client read data error";
 
     gpuParam.copyFrom(cpuParam);

paddle/scripts/docker/build.sh

@@ -78,7 +78,7 @@ paddle version
 # PaddlePaddle.  This awkwardness is due to
 # https://github.com/PaddlePaddle/Paddle/issues/1854.  It also
 # describes a solution.
-if [ ${WITH_DOC} == "ON" ]; then
+if [[ ${WITH_DOC} == "ON" ]]; then
     cat <<EOF
 ========================================
 Building documentation ...

paddle/scripts/travis/build_doc.sh

@@ -5,13 +5,14 @@ set -e
 mkdir -p $TRAVIS_BUILD_DIR/build
 cd $TRAVIS_BUILD_DIR/build
 
-# Compile Documentation only.
-cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_DOC=OFF -DWITH_STYLE_CHECK=OFF
+# Compile paddle binaries first
+cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_DOC=OFF -DWITH_GOLANG=ON -DWITH_STYLE_CHECK=OFF
 
 mkdir output
 make -j `nproc`
 find .. -name '*whl' | xargs pip install  # install all wheels.
 rm -rf *
+# Compile Documentation only.
 cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_DOC=ON
 make -j `nproc` paddle_docs paddle_docs_cn
 
@@ -25,7 +26,7 @@ SSH_REPO=${REPO/https:\/\/github.com\//git@github.com:}
 SHA=`git rev-parse --verify HEAD`
 
 # Documentation branch name
-# gh-pages branch is used for PaddlePaddle.org. The English version of 
+# gh-pages branch is used for PaddlePaddle.org. The English version of
 # documentation in `doc` directory, and the chinese version in `doc_cn`
 # directory.
 TARGET_BRANCH="gh-pages"
@@ -51,7 +52,7 @@ function deploy_docs() {
 
   # checkout github page branch
   git checkout $TARGET_BRANCH || git checkout --orphan $TARGET_BRANCH
-  
+
   mkdir -p ${DIR}
   # remove old docs. mv new docs.
   set +e
@@ -62,7 +63,7 @@ function deploy_docs() {
   git add .
 }
 
-deploy_docs "master" "." 
+deploy_docs "master" "."
 deploy_docs "develop" "./develop/"
 
 # Check is there anything changed.

paddle/trainer/Tester.cpp

@@ -175,7 +175,7 @@ real Tester::forwardOneBatch(const DataBatch& dataBatch,
     }
     hl_stream_synchronize(HPPL_STREAM_DEFAULT);
     FILE* fp = fopen(featFile.c_str(), "ab+");
-    PCHECK(!ferror(fp)) << "Fail to open " << featFile;
+    CHECK(!ferror(fp)) << "Fail to open " << featFile;
 
     size_t sampleNum = featMatrices[0]->getHeight();
     for (size_t i = 0; i < sampleNum; ++i) {

paddle/utils/ThreadLocal.h

@@ -51,7 +51,7 @@ template <class T>
 class ThreadLocal {
 public:
   ThreadLocal() {
-    PCHECK(pthread_key_create(&threadSpecificKey_, dataDestructor) == 0);
+    CHECK(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);
-      PCHECK(ret == 0);
+      CHECK(ret == 0);
     }
     return p;
   }
@@ -79,7 +79,7 @@ public:
     if (T* q = get(false)) {
       dataDestructor(q);
     }
-    PCHECK(pthread_setspecific(threadSpecificKey_, p) == 0);
+    CHECK(pthread_setspecific(threadSpecificKey_, p) == 0);
   }
 
   /**
@@ -112,7 +112,7 @@ private:
 template <class T>
 class ThreadLocalD {
 public:
-  ThreadLocalD() { PCHECK(pthread_key_create(&threadSpecificKey_, NULL) == 0); }
+  ThreadLocalD() { CHECK(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();
-      PCHECK(pthread_setspecific(threadSpecificKey_, p) == 0);
+      CHECK(pthread_setspecific(threadSpecificKey_, p) == 0);
       updateMap(p);
     }
     return p;
@@ -141,7 +141,7 @@ public:
     if (T* q = (T*)pthread_getspecific(threadSpecificKey_)) {
       dataDestructor(q);
     }
-    PCHECK(pthread_setspecific(threadSpecificKey_, p) == 0);
+    CHECK(pthread_setspecific(threadSpecificKey_, p) == 0);
     updateMap(p);
   }
 

proto/OptimizerConfig.proto

@@ -78,11 +78,15 @@ enum DataType {
   repeated bytes content = 2;
 }
 
+message LrPolicyState {
+  // learninRate Policy
+  optional double learning_rate = 1 [default = 1.0];
+  optional double lr_decay_a = 2;
+  optional double lr_decay_b = 3;
+}
+
 message SGDOptimizerState {
-  // learning rate policy
-  optional double learning_rate = 101;
-  optional double lr_decay_a = 102;
-  optional double lr_decay_b = 103;
+  optional LrPolicyState lr_state = 101;
   optional double num_sample_passed = 104;
   // state
   optional TensorProto parameter = 1;
@@ -91,9 +95,7 @@ message SGDOptimizerState {
 
 message AdadeltaOptimizerState {
   // learning rate policy
-  optional double learning_rate = 101;
-  optional double lr_decay_a = 102;
-  optional double lr_decay_b = 103;
+  optional LrPolicyState lr_state = 101;
   optional double num_sample_passed = 104;
   // state
   optional TensorProto parameter = 1;
@@ -102,11 +104,9 @@ message AdadeltaOptimizerState {
   optional TensorProto update_delta = 4;
 }
 
+
 message AdagradOptimizerState {
-  // learning rate policy
-  optional double learning_rate = 101;
-  optional double lr_decay_a = 102;
-  optional double lr_decay_b = 103;
+  optional LrPolicyState lr_state = 101;
   optional double num_sample_passed = 104;
   // state
   optional TensorProto parameter = 1;
@@ -114,10 +114,7 @@ message AdagradOptimizerState {
 }
 
 message AdamOptimizerState {
-  // learning rate policy
-  optional double learning_rate = 101;
-  optional double lr_decay_a = 102;
-  optional double lr_decay_b = 103;
+  optional LrPolicyState lr_state = 101;
   optional double num_sample_passed = 104;
   // state
   optional TensorProto parameter = 1;

python/CMakeLists.txt

@@ -29,7 +29,7 @@ configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in
 add_custom_command(OUTPUT ${OUTPUT_DIR}/.timestamp
     COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
     COMMAND ${CMAKE_COMMAND} -E touch ${OUTPUT_DIR}/.timestamp
-    DEPENDS gen_proto_py ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
+    DEPENDS gen_proto_py framework_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
 
 add_custom_target(paddle_python ALL DEPENDS
     ${OUTPUT_DIR}/.timestamp)
@@ -43,6 +43,7 @@ if (WITH_TESTING)
     add_subdirectory(paddle/v2/tests)
     add_subdirectory(paddle/v2/reader/tests)
     add_subdirectory(paddle/v2/plot/tests)
+    add_subdirectory(paddle/v2/framework/tests)
   endif()
 endif()
 install(DIRECTORY ${PADDLE_PYTHON_PACKAGE_DIR}

python/paddle/trainer/config_parser.py

@@ -1353,7 +1353,8 @@ class LayerBase(object):
             device=None,
             active_type="",
             drop_rate=0.,
-            coeff=None):
+            coeff=None,
+            error_clipping_threshold=None):
         config_assert('@' not in name,
                       "layer name: %s contain special character @" % name)
         global g_current_submodel
@@ -1387,6 +1388,9 @@ class LayerBase(object):
         elif g_default_device is not None:
             self.config.device = g_default_device
 
+        if error_clipping_threshold is not None:
+            self.config.error_clipping_threshold = error_clipping_threshold
+
         for input_index in xrange(len(self.inputs)):
             input = self.inputs[input_index]
             input_config = None
@@ -2466,10 +2470,14 @@ class MaxLayer(LayerBase):
                  trans_type='non-seq',
                  bias=False,
                  output_max_index=None,
+                 stride=-1,
                  **xargs):
         super(MaxLayer, self).__init__(name, 'max', 0, inputs=inputs, **xargs)
         config_assert(len(self.inputs) == 1, 'MaxLayer must have 1 input')
+        if trans_type == 'seq':
+            config_assert(stride == -1, 'subseq does not support stride window')
         self.config.trans_type = trans_type
+        self.config.seq_pool_stride = stride
         for input_index in xrange(len(self.inputs)):
             input_layer = self.get_input_layer(input_index)
             self.set_layer_size(input_layer.size)
@@ -2731,11 +2739,15 @@ class AverageLayer(LayerBase):
                  average_strategy='average',
                  trans_type='non-seq',
                  bias=False,
+                 stride=-1,
                  **xargs):
         super(AverageLayer, self).__init__(
             name, 'average', 0, inputs=inputs, **xargs)
         self.config.average_strategy = average_strategy
+        if trans_type == 'seq':
+            config_assert(stride == -1, 'subseq does not support stride window')
         self.config.trans_type = trans_type
+        self.config.seq_pool_stride = stride
         config_assert(len(inputs) == 1, 'AverageLayer must have 1 input')
         for input_index in xrange(len(self.inputs)):
             input_layer = self.get_input_layer(input_index)
@@ -2774,13 +2786,7 @@ class TensorLayer(LayerBase):
 
 @config_layer('mixed')
 class MixedLayer(LayerBase):
-    def __init__(self,
-                 name,
-                 inputs,
-                 size=0,
-                 bias=True,
-                 error_clipping_threshold=None,
-                 **xargs):
+    def __init__(self, name, inputs, size=0, bias=True, **xargs):
         config_assert(inputs, 'inputs cannot be empty')
         super(MixedLayer, self).__init__(
             name, 'mixed', size, inputs=inputs, **xargs)
@@ -2862,9 +2868,6 @@ class MixedLayer(LayerBase):
             self.config.bias_size = psize
             self.create_bias_parameter(bias, psize)
 
-        if error_clipping_threshold is not None:
-            self.config.error_clipping_threshold = error_clipping_threshold
-
 
 # like MixedLayer, but no bias parameter
 @config_func

python/paddle/trainer_config_helpers/layers.py

@@ -1247,10 +1247,19 @@ def pooling_layer(input,
                   name=None,
                   bias_attr=None,
                   agg_level=AggregateLevel.TO_NO_SEQUENCE,
+                  stride=-1,
                   layer_attr=None):
     """
     Pooling layer for sequence inputs, not used for Image.
 
+    If stride > 0, this layer slides a window whose size is determined by stride,
+    and return the pooling value of the window as the output. Thus, a long sequence
+    will be shorten.
+
+    The parameter stride specifies the intervals at which to apply the pooling
+    operation. Note that for sequence with sub-sequence, the default value
+    of stride is -1.
+
     The example usage is:
 
     .. code-block:: python
@@ -1269,6 +1278,8 @@ def pooling_layer(input,
     :param pooling_type: Type of pooling, MaxPooling(default), AvgPooling,
                          SumPooling, SquareRootNPooling.
     :type pooling_type: BasePoolingType|None
+    :param stride: The step size between successive pooling regions.
+    :type stride: Int
     :param bias_attr: Bias parameter attribute. False if no bias.
     :type bias_attr: ParameterAttribute|None|False
     :param layer_attr: The Extra Attributes for layer, such as dropout.
@@ -1286,12 +1297,16 @@ def pooling_layer(input,
         extra_dict['output_max_index'] = pooling_type.output_max_index
     extra_dict.update(ExtraLayerAttribute.to_kwargs(layer_attr))
 
+    if agg_level == AggregateLevel.TO_SEQUENCE:
+        assert stride == -1
+
     Layer(
         name=name,
         type=pooling_type.name,
         inputs=[Input(input.name)],
         bias=ParamAttr.to_bias(bias_attr),
         trans_type=agg_level,
+        stride=stride,
         **extra_dict)
 
     return LayerOutput(
@@ -1553,7 +1568,7 @@ def last_seq(input,
     :type name: basestring
     :param input: Input layer name.
     :type input: LayerOutput
-    :param stride: window size.
+    :param stride: The step size between successive pooling regions.
     :type stride: Int
     :param layer_attr: extra layer attributes.
     :type layer_attr: ExtraLayerAttribute.
@@ -1609,7 +1624,7 @@ def first_seq(input,
     :type name: basestring
     :param input: Input layer name.
     :type input: LayerOutput
-    :param stride: window size.
+    :param stride: The step size between successive pooling regions.
     :type stride: Int
     :param layer_attr: extra layer attributes.
     :type layer_attr: ExtraLayerAttribute.
@@ -4791,6 +4806,14 @@ def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None):
     So groups should be larger than 1, and the num of channels should be able
     to devided by groups.
 
+    .. math::
+       y_{si+j} = \max_k x_{gsi + sk + j}
+       g = groups
+       s = input.size / num_channels
+       0 \le i < num_channels / groups
+       0 \le j < s
+       0 \le k < groups
+
     Please refer to Paper:
       - Maxout Networks: http://www.jmlr.org/proceedings/papers/v28/goodfellow13.pdf
       - Multi-digit Number Recognition from Street View \

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

@@ -14,6 +14,7 @@ layers {
     input_layer_name: "dat_in"
   }
   trans_type: "seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__seq_pooling_1__"
@@ -24,6 +25,7 @@ layers {
     input_layer_name: "dat_in"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__seq_pooling_2__"
@@ -35,6 +37,7 @@ layers {
   }
   average_strategy: "average"
   trans_type: "seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__seq_pooling_3__"
@@ -46,6 +49,7 @@ layers {
   }
   average_strategy: "average"
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__seq_pooling_4__"
@@ -57,6 +61,7 @@ layers {
   }
   average_strategy: "sum"
   trans_type: "seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__seq_pooling_5__"
@@ -68,6 +73,7 @@ layers {
   }
   average_strategy: "sum"
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__seq_pooling_6__"
@@ -77,8 +83,44 @@ layers {
   inputs {
     input_layer_name: "dat_in"
   }
+  trans_type: "non-seq"
+  seq_pool_stride: 5
+}
+layers {
+  name: "__seq_pooling_7__"
+  type: "average"
+  size: 100
+  active_type: ""
+  inputs {
+    input_layer_name: "dat_in"
+  }
+  average_strategy: "average"
+  trans_type: "non-seq"
+  seq_pool_stride: 5
+}
+layers {
+  name: "__seq_pooling_8__"
+  type: "average"
+  size: 100
+  active_type: ""
+  inputs {
+    input_layer_name: "dat_in"
+  }
+  average_strategy: "sum"
+  trans_type: "non-seq"
+  seq_pool_stride: 5
+}
+layers {
+  name: "__seq_pooling_9__"
+  type: "max"
+  size: 100
+  active_type: ""
+  inputs {
+    input_layer_name: "dat_in"
+  }
   output_max_index: true
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 input_layer_names: "dat_in"
 output_layer_names: "__seq_pooling_0__"
@@ -88,6 +130,9 @@ output_layer_names: "__seq_pooling_3__"
 output_layer_names: "__seq_pooling_4__"
 output_layer_names: "__seq_pooling_5__"
 output_layer_names: "__seq_pooling_6__"
+output_layer_names: "__seq_pooling_7__"
+output_layer_names: "__seq_pooling_8__"
+output_layer_names: "__seq_pooling_9__"
 sub_models {
   name: "root"
   layer_names: "dat_in"
@@ -98,6 +143,9 @@ sub_models {
   layer_names: "__seq_pooling_4__"
   layer_names: "__seq_pooling_5__"
   layer_names: "__seq_pooling_6__"
+  layer_names: "__seq_pooling_7__"
+  layer_names: "__seq_pooling_8__"
+  layer_names: "__seq_pooling_9__"
   input_layer_names: "dat_in"
   output_layer_names: "__seq_pooling_0__"
   output_layer_names: "__seq_pooling_1__"
@@ -106,6 +154,9 @@ sub_models {
   output_layer_names: "__seq_pooling_4__"
   output_layer_names: "__seq_pooling_5__"
   output_layer_names: "__seq_pooling_6__"
+  output_layer_names: "__seq_pooling_7__"
+  output_layer_names: "__seq_pooling_8__"
+  output_layer_names: "__seq_pooling_9__"
   is_recurrent_layer_group: false
 }
 

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

@@ -14,6 +14,14 @@ for pt in POOL_TYPE:
     for al in AGG_LEVEL:
         opts.append(pooling_layer(input=din, agg_level=al, pooling_type=pt()))
 
+for pt in POOL_TYPE:
+    opts.append(
+        pooling_layer(
+            input=din,
+            agg_level=AggregateLevel.TO_NO_SEQUENCE,
+            pooling_type=pt(),
+            stride=5))
+
 opts.append(
     pooling_layer(
         input=din, pooling_type=MaxPooling(output_max_index=True)))

python/paddle/v2/framework/__init__.py

+__all__ = ['proto']

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

+add_python_test(test_framework test_protobuf.py)

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

+import paddle.v2.framework.proto.op_proto_pb2
+import paddle.v2.framework.proto.attr_type_pb2
+import unittest
+
+
+class TestFrameworkProto(unittest.TestCase):
+    def test_all(self):
+        op_proto_lib = paddle.v2.framework.proto.op_proto_pb2
+        attr_type_lib = paddle.v2.framework.proto.attr_type_pb2
+        op_proto = op_proto_lib.OpProto()
+        ipt0 = op_proto.inputs.add()
+        ipt0.name = "a"
+        ipt0.comment = "the input of cosine op"
+        ipt1 = op_proto.inputs.add()
+        ipt1.name = "b"
+        ipt1.comment = "the other input of cosine op"
+        opt = op_proto.outputs.add()
+        opt.name = "output"
+        opt.comment = "the output of cosine op"
+        op_proto.comment = "cosine op, output = scale*cos(a, b)"
+        attr = op_proto.attrs.add()
+        attr.name = "scale"
+        attr.comment = "scale of cosine op"
+        attr.type = attr_type_lib.FLOAT
+        op_proto.type = "cos"
+        self.assertTrue(op_proto.IsInitialized())

python/paddle/v2/master/client.py

@@ -26,14 +26,22 @@ class client(object):
             holder[idx] = c_ptr
         lib.paddle_set_dataset(self.c, holder, len(paths))
 
+    # return format: (record, errno)
+    # errno =  0: ok
+    #       <  0: error
     def next_record(self):
         p = ctypes.c_char_p()
         ret = ctypes.pointer(p)
         size = lib.paddle_next_record(self.c, ret)
+        if size < 0:
+            # Error
+            return None, size
+
         if size == 0:
             # Empty record
-            return ""
+            return "", 0
+
         record = ret.contents.value[:size]
         # Memory created from C should be freed.
         lib.mem_free(ret.contents)
-        return record
+        return record, 0

python/paddle/v2/reader/creator.py

@@ -57,17 +57,20 @@ def text_file(path):
     return reader
 
 
-def recordio(path):
+def recordio_local(paths, buf_size=100):
     """
-    Creates a data reader that outputs record one one by one from given recordio file
-    :path: path of recordio file
-    :returns: data reader of recordio file
+    Creates a data reader from given RecordIO file paths separated by ",", 
+        glob pattern is supported.
+    :path: path of recordio files.
+    :returns: data reader of recordio files.
     """
 
     import recordio as rec
+    import paddle.v2.reader.decorator as dec
 
     def reader():
-        f = rec.reader(path)
+        a = ','.join(paths)
+        f = rec.reader(a)
         while True:
             r = f.read()
             if r is None:
@@ -75,4 +78,38 @@ def recordio(path):
             yield r
         f.close()
 
+    return dec.buffered(reader, buf_size)
+
+
+def recordio(paths, buf_size=100):
+    """
+    Creates a data reader that outputs record one one by one 
+        from given local or cloud recordio path.
+    :path: path of recordio files.
+    :returns: data reader of recordio files.
+    """
+    import os
+    import paddle.v2.master.client as cloud
+
+    if "KUBERNETES_SERVICE_HOST" not in os.environ.keys():
+        return recordio_local(paths)
+
+    host_name = "MASTER_SERVICE_HOST"
+    if host_name not in os.environ.keys():
+        raise Exception('not find ' + host_name + ' in environ.')
+
+    addr = os.environ(host)
+
+    def reader():
+        c = cloud(addr, buf_size)
+        c.set_dataset(paths)
+
+        while True:
+            r, err = client.next_record()
+            if err < 0:
+                break
+            yield r
+
+        c.close()
+
     return reader

python/paddle/v2/reader/tests/creator_test.py

@@ -38,7 +38,7 @@ class TestRecordIO(unittest.TestCase):
     def test_recordio(self):
         path = os.path.join(
             os.path.dirname(__file__), "test_recordio_creator.dat")
-        reader = paddle.v2.reader.creator.recordio(path)
+        reader = paddle.v2.reader.creator.recordio([path])
         for idx, r in enumerate(reader()):
             self.assertSequenceEqual(r, str(idx))
 

python/setup.py.in

@@ -9,7 +9,9 @@ packages=['paddle',
           'paddle.v2.dataset',
           'paddle.v2.reader',
           'paddle.v2.master',
-          'paddle.v2.plot']
+          'paddle.v2.plot',
+          'paddle.v2.framework',
+          'paddle.v2.framework.proto']
 
 setup_requires=["requests",
                 "numpy",
@@ -27,8 +29,11 @@ setup(name='paddle',
       description='Parallel Distributed Deep Learning',
       install_requires=setup_requires,
       packages=packages,
-      package_data={'paddle.v2.master': ['${paddle_master_LIB_NAME}'], },
+      package_data={'paddle.v2.master': ['libpaddle_master.so'], },
       package_dir={
-          '': '${CMAKE_CURRENT_SOURCE_DIR}'
+          '': '${CMAKE_CURRENT_SOURCE_DIR}',
+          # The paddle.v2.framework.proto will be generated while compiling.
+          # So that package points to other directory.
+          'paddle.v2.framework.proto': '${PROJ_BINARY_ROOT}/paddle/framework'
       },
 )