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

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

README.md

@@ -62,9 +62,9 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl
 ## Installation
 
 It is recommended to check out the
-[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/docker_install_en.html)
+[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/docker_install_en.html)
 before looking into the
-[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/build_from_source_en.html).
+[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/build_from_source_en.html).
 
 ## Documentation
 

cmake/external/mkldnn.cmake

@@ -45,9 +45,9 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
 ELSE()
     MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
 ENDIF()
-
-SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow")
-SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} -Wno-error=strict-overflow")
+SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-unused-result")
+SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
+SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
 ExternalProject_Add(
     ${MKLDNN_PROJECT}
     ${EXTERNAL_PROJECT_LOG_ARGS}
@@ -61,6 +61,7 @@ ExternalProject_Add(
     CMAKE_ARGS          -DMKLROOT=${MKLML_ROOT}
     CMAKE_ARGS          -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
     CMAKE_ARGS          -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
+    CMAKE_ARGS          -DWITH_TEST=OFF -DWITH_EXAMPLE=OFF
     CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
                         -DMKLROOT:PATH=${MKLML_ROOT}
 )

cmake/external/mklml.cmake

@@ -27,7 +27,7 @@ ENDIF()
 INCLUDE(ExternalProject)
 
 SET(MKLML_PROJECT       "extern_mklml")
-SET(MKLML_VER           "mklml_lnx_2018.0.1.20171007")
+SET(MKLML_VER           "mklml_lnx_2018.0.3.20180406")
 SET(MKLML_URL           "http://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.tgz")
 SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
 SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")

doc/fluid/design/concepts/lod_tensor.md

@@ -155,7 +155,7 @@ into offsets
    3  2+3 4+5 1+9 2+10 3+12
 ```
 
-so we know that the first sentence is from word 0 to word 3, and the second sentence from work 3 to word 5.
+so we know that the first sentence is from word 0 to word 3, and the second sentence from word 3 to word 5.
 
 Similarly, the lengths in the top level LoD
 

paddle/fluid/framework/CMakeLists.txt

@@ -57,7 +57,7 @@ cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor
 cc_library(attribute SRCS attribute.cc DEPS framework_proto boost)
 cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc
 device_context)
-cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute)
+cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute glog)
 cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
 cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)

paddle/fluid/framework/data_device_transform_test.cu

@@ -32,8 +32,7 @@ struct AddFunctor {
 
 class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
-  OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("input", "input1 of test op");
     AddOutput("output", "output of test op");
     AddAttr<bool>("use_gpu", "force to use gpu kernel").SetDefault(false);

paddle/fluid/framework/details/computation_op_handle.h

@@ -36,7 +36,7 @@ struct ComputationOpHandle : public OpHandleBase {
  protected:
   void RunImpl() override;
 
-  virtual bool NeedWait(VarHandleBase *in_var);
+  bool NeedWait(VarHandleBase *in_var) override;
 
  private:
   std::unique_ptr<OperatorBase> op_;

paddle/fluid/framework/details/fetch_op_handle.h

@@ -42,7 +42,7 @@ struct FetchOpHandle : public OpHandleBase {
  protected:
   void RunImpl() override;
 
-  virtual void WaitInputVarGenerated(const platform::Place &place);
+  void WaitInputVarGenerated(const platform::Place &place) override;
 
  private:
   FeedFetchList *data_;

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -37,20 +37,26 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
     const std::vector<Scope *> &local_scopes,
-    platform::NCCLContextMap *nccl_ctxs, bool use_default_grad_scale)
+    platform::NCCLContextMap *nccl_ctxs, bool use_default_grad_scale,
+    bool balance_parameter_opt_between_cards)
     : loss_var_name_(loss_var_name),
       places_(places),
       local_scopes_(local_scopes),
-      nccl_ctxs_(nccl_ctxs) {
+      nccl_ctxs_(nccl_ctxs),
+      balance_parameter_opt_between_cards_(
+          balance_parameter_opt_between_cards) {
 #else
 MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::vector<platform::Place> &places,
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
-    const std::vector<Scope *> &local_scopes, bool use_default_grad_scale)
+    const std::vector<Scope *> &local_scopes, bool use_default_grad_scale,
+    bool balance_parameter_opt_between_cards)
     : loss_var_name_(loss_var_name),
       places_(places),
-      local_scopes_(local_scopes) {
+      local_scopes_(local_scopes),
+      balance_parameter_opt_between_cards_(
+          balance_parameter_opt_between_cards) {
 #endif
   for (auto &p : params) {
     grad_names_.insert(GradVarName(p));
@@ -124,6 +130,12 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
   // Find "send" op first for split is in front of send.
   OpDesc *send_op = GetSendOpDesc(program);
 
+  size_t cur_device_id = 0;
+  std::vector<std::unordered_set<std::string>> var_name_on_devices;
+  std::vector<std::unordered_set<std::string>> bcast_var_name_set;
+  var_name_on_devices.resize(places_.size());
+  bcast_var_name_set.resize(places_.size());
+
   bool is_forwarding = true;
   for (auto *op : program.Block(0).AllOps()) {
     if (op->Type() == "send") {
@@ -139,17 +151,33 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
       }
       is_forwarding = false;
     } else {
-      CreateComputationalOps(&result, *op, places_.size());
+      int op_dev_id = GetOpDeviceID(var_name_on_devices, *op);
+      if (op_dev_id == -1) {  // var on all device
+        CreateComputationalOps(&result, *op, places_.size());
+      } else {
+        CreateComputationalOp(&result, *op, op_dev_id);
+        for (auto &var_name : op->OutputArgumentNames()) {
+          var_name_on_devices[op_dev_id].emplace(var_name);
+        }
+      }
       if (!is_forwarding && places_.size() > 1) {
         // Currently, we assume that once gradient is generated, it can be
         // broadcast, and each gradient is only broadcast once.
         for (auto &og : op->OutputArgumentNames()) {
           if (IsParameterGradientOnce(og, &og_has_been_broadcast)) {
-            if (IsSparseGradient(var_types, og)) {
-              CreateReduceOp(&result, og, 0);
-              CreateBroadcastOp(&result, og, 0);
+            if (balance_parameter_opt_between_cards_) {
+              CreateReduceOp(&result, og, cur_device_id);
+              var_name_on_devices[cur_device_id].emplace(og);
+              bcast_var_name_set[cur_device_id].emplace(
+                  og.substr(0, og.size() - strlen(kGradVarSuffix)));
+              cur_device_id = (cur_device_id + 1) % places_.size();
             } else {
-              InsertNCCLAllReduceOp(&result, og);
+              if (IsSparseGradient(var_types, og)) {
+                CreateReduceOp(&result, og, 0);
+                CreateBroadcastOp(&result, og, 0);
+              } else {
+                InsertNCCLAllReduceOp(&result, og);
+              }
             }
           }
         }
@@ -157,6 +185,13 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
     }
   }
 
+  // Insert BCast Ops
+  for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
+    auto &to_bcast_set = bcast_var_name_set[dev_id];
+    for (auto &bcast_name : to_bcast_set) {
+      CreateBroadcastOp(&result, bcast_name, dev_id);
+    }
+  }
   /*
     Dependency graph has been constructed. However, there are still data
     harzaeds need to be handled.
@@ -265,6 +300,26 @@ bool MultiDevSSAGraphBuilder::IsParameterGradientOnce(
   return is_pg_once;
 }
 
+int MultiDevSSAGraphBuilder::GetOpDeviceID(
+    const std::vector<std::unordered_set<std::string>> &var_name_on_devices,
+    const OpDesc &op) const {
+  if (!balance_parameter_opt_between_cards_) {
+    return -1;
+  }
+
+  int var_dev_id = -1;
+  for (auto &var_name : op.InputArgumentNames()) {
+    if (var_dev_id != -1) break;
+    for (size_t i = 0; i < var_name_on_devices.size(); ++i) {
+      if (var_name_on_devices[i].count(var_name)) {
+        var_dev_id = static_cast<int>(i);
+        break;
+      }
+    }
+  }
+  return var_dev_id;
+}
+
 void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(SSAGraph *result) const {
   for (size_t i = 0; i < places_.size(); ++i) {
 // Insert ScaleCost OpHandle

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -36,13 +36,15 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
                           platform::NCCLContextMap *nccl_ctxs,
-                          bool use_default_grad_scale);
+                          bool use_default_grad_scale,
+                          bool balance_parameter_opt_between_cards);
 #else
   MultiDevSSAGraphBuilder(const std::vector<platform::Place> &places,
                           const std::string &loss_var_name,
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
-                          bool use_default_grad_scale);
+                          bool use_default_grad_scale,
+                          bool balance_parameter_opt_between_cards);
 #endif
 
   std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const override;
@@ -60,6 +62,7 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 #ifdef PADDLE_WITH_CUDA
   platform::NCCLContextMap *nccl_ctxs_;
 #endif
+  bool balance_parameter_opt_between_cards_;
   bool use_default_grad_scale_;
 
   bool IsScaleLossOp(const OpDesc &op) const;
@@ -84,6 +87,10 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
       const std::string &og,
       std::unordered_set<std::string> *og_has_been_broadcast) const;
 
+  int GetOpDeviceID(
+      const std::vector<std::unordered_set<std::string>> &var_name_on_devices,
+      const OpDesc &op) const;
+
   void InsertNCCLAllReduceOp(SSAGraph *result, const std::string &og) const;
 
   void CreateBroadcastOp(SSAGraph *result, const std::string &p_name,

paddle/fluid/framework/details/op_registry.h

@@ -95,7 +95,10 @@ struct OpInfoFiller<T, kOpProtoAndCheckerMaker> {
   void operator()(const char* op_type, OpInfo* info) const {
     info->proto_ = new proto::OpProto;
     info->checker_ = new OpAttrChecker();
-    auto maker = T(info->proto_, info->checker_);
+    T maker;
+    maker.SetProto(info->proto_);
+    maker.SetChecker(info->checker_);
+    maker.Make();
     maker.Validate();
     info->proto_->set_type(op_type);
     PADDLE_ENFORCE(

paddle/fluid/framework/op_proto_maker.h

@@ -14,56 +14,57 @@ limitations under the License. */
 #pragma once
 
 #include <string>
+#include "glog/logging.h"
 #include "paddle/fluid/framework/attribute.h"
 #include "paddle/fluid/framework/framework.pb.h"
-
 namespace paddle {
 namespace framework {
 
 // this class not only make proto but also init attribute checkers.
 class OpProtoAndCheckerMaker {
  public:
-  using OpProto = proto::OpProto;
-  using OpAttrChecker = framework::OpAttrChecker;
-  OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : proto_(proto), op_checker_(op_checker) {}
+  virtual void Make() = 0;
 
   virtual ~OpProtoAndCheckerMaker() {
-    PADDLE_ENFORCE(validated_, "should call Validate after build");
+    CHECK(validated_) << "should call Validate after build";
   }
 
+  void SetProto(proto::OpProto *proto) { proto_ = proto; }
+
+  void SetChecker(OpAttrChecker *attr_checker) { op_checker_ = attr_checker; }
+
   void Validate();
 
  protected:
   struct VariableBuilder {
-    OpProto::Var* var_;
+    proto::OpProto::Var *var_;
 
-    VariableBuilder& AsDuplicable() {
+    VariableBuilder &AsDuplicable() {
       var_->set_duplicable(true);
       return *this;
     }
 
-    VariableBuilder& AsIntermediate() {
+    VariableBuilder &AsIntermediate() {
       var_->set_intermediate(true);
       return *this;
     }
 
-    VariableBuilder& AsDispensable() {
+    VariableBuilder &AsDispensable() {
       var_->set_dispensable(true);
       return *this;
     }
   };
 
-  VariableBuilder AddInput(const std::string& name, const std::string& comment);
+  VariableBuilder AddInput(const std::string &name, const std::string &comment);
 
-  VariableBuilder AddOutput(const std::string& name,
-                            const std::string& comment);
+  VariableBuilder AddOutput(const std::string &name,
+                            const std::string &comment);
 
   template <typename T>
-  TypedAttrChecker<T>& AddAttr(const std::string& name,
-                               const std::string& comment,
+  TypedAttrChecker<T> &AddAttr(const std::string &name,
+                               const std::string &comment,
                                bool generated = false) {
-    auto* attr = proto_->add_attrs();
+    auto *attr = proto_->add_attrs();
     attr->set_name(name);
     attr->set_comment(comment);
     attr->set_generated(generated);
@@ -71,21 +72,14 @@ class OpProtoAndCheckerMaker {
     return op_checker_->AddAttrChecker<T>(name);
   }
 
-  void AddComment(const std::string& comment) { proto_->set_comment(comment); }
+  void AddComment(const std::string &comment) { proto_->set_comment(comment); }
 
  private:
   void CheckNoDuplicatedInOutAttrs();
 
-  OpProto* proto_;
-  OpAttrChecker* op_checker_;
+  proto::OpProto *proto_;
+  OpAttrChecker *op_checker_;
   bool validated_{false};
 };
-
-class NOPMaker : public OpProtoAndCheckerMaker {
- public:
-  NOPMaker(OpProto* proto, framework::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {}
-};
-
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/op_proto_maker_test.cc

@@ -18,9 +18,7 @@ limitations under the License. */
 
 class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
  public:
-  TestAttrProtoMaker(paddle::framework::proto::OpProto* proto,
-                     paddle::framework::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddAttr<float>("scale", "scale of test op");
     AddAttr<float>("scale", "scale of test op");
   }
@@ -29,15 +27,16 @@ class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
 TEST(ProtoMaker, DuplicatedAttr) {
   paddle::framework::proto::OpProto op_proto;
   paddle::framework::OpAttrChecker op_checker;
-  auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker);
+  TestAttrProtoMaker proto_maker;
+  proto_maker.SetProto(&op_proto);
+  proto_maker.SetChecker(&op_checker);
+  proto_maker.Make();
   ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
 }
 
 class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
  public:
-  TestInOutProtoMaker(paddle::framework::proto::OpProto* proto,
-                      paddle::framework::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("input", "input of test op");
     AddInput("input", "input of test op");
   }
@@ -46,6 +45,9 @@ class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
 TEST(ProtoMaker, DuplicatedInOut) {
   paddle::framework::proto::OpProto op_proto;
   paddle::framework::OpAttrChecker op_checker;
-  auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker);
+  TestAttrProtoMaker proto_maker;
+  proto_maker.SetProto(&op_proto);
+  proto_maker.SetChecker(&op_checker);
+  proto_maker.Make();
   ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
 }

paddle/fluid/framework/op_registry_test.cc

@@ -33,8 +33,7 @@ class CosineOp : public OperatorBase {
 
 class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
-  CosineOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("input", "input of cosine op");
     AddOutput("output", "output of cosine op");
     AddAttr<float>("scale", "scale of cosine op")
@@ -55,8 +54,7 @@ class MyTestOp : public OperatorBase {
 
 class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
-  MyTestOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("input", "input of cosine op").AsDuplicable();
     AddOutput("output", "output of cosine op").AsIntermediate();
     auto my_checker = [](int i) {
@@ -212,10 +210,7 @@ namespace framework {
 
 class OpKernelTestMaker : public OpProtoAndCheckerMaker {
  public:
-  OpKernelTestMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddComment("NoGradOp, same input output. no Grad");
-  }
+  void Make() { AddComment("NoGradOp, same input output. no Grad"); }
 };
 
 class OpWithKernelTest : public OperatorWithKernel {
@@ -275,9 +270,9 @@ TEST(OperatorRegistrar, CUDA) {
 
 static int op_test_value = 0;
 
-using paddle::platform::DeviceContext;
 using paddle::platform::CPUDeviceContext;
 using paddle::platform::CUDADeviceContext;
+using paddle::platform::DeviceContext;
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/operator_test.cc

@@ -46,8 +46,7 @@ class OpWithoutKernelTest : public OperatorBase {
 
 class OpWithoutKernelCheckerMaker : public OpProtoAndCheckerMaker {
  public:
-  OpWithoutKernelCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("input", "input of test op");
     AddOutput("output", "output of test op");
     AddAttr<float>("scale", "scale of cosine op");
@@ -98,8 +97,7 @@ namespace framework {
 
 class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
-  OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("x", "input of test op");
     AddOutput("y", "output of test op");
     AddAttr<float>("scale", "scale of cosine op")
@@ -137,9 +135,7 @@ class CPUKernelTest : public OpKernel<float> {
 class OpKernelTestMultiInputsProtoAndCheckerMaker
     : public OpProtoAndCheckerMaker {
  public:
-  OpKernelTestMultiInputsProtoAndCheckerMaker(OpProto* proto,
-                                              OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("xs", "inputs of test op").AsDuplicable();
     AddInput("k", "input of test op");
     AddOutput("ys", "outputs of test op").AsDuplicable();

paddle/fluid/framework/parallel_executor.cc

@@ -58,7 +58,7 @@ ParallelExecutor::ParallelExecutor(
     const std::unordered_set<std::string> &bcast_vars,
     const ProgramDesc &main_program, const std::string &loss_var_name,
     Scope *scope, const std::vector<Scope *> &local_scopes, bool allow_op_delay,
-    bool use_default_grad_scale)
+    bool use_default_grad_scale, bool balance_parameter_opt_between_cards)
     : member_(new ParallelExecutorPrivate(places)) {
   member_->global_scope_ = scope;
 
@@ -93,11 +93,12 @@ ParallelExecutor::ParallelExecutor(
 #ifdef PADDLE_WITH_CUDA
   details::MultiDevSSAGraphBuilder builder(
       member_->places_, loss_var_name, params, member_->local_scopes_,
-      member_->nccl_ctxs_.get(), use_default_grad_scale);
+      member_->nccl_ctxs_.get(), use_default_grad_scale,
+      balance_parameter_opt_between_cards);
 #else
-  details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name,
-                                           params, member_->local_scopes_,
-                                           use_default_grad_scale);
+  details::MultiDevSSAGraphBuilder builder(
+      member_->places_, loss_var_name, params, member_->local_scopes_,
+      use_default_grad_scale, balance_parameter_opt_between_cards);
 #endif
   auto graph = builder.Build(main_program);
 

paddle/fluid/framework/parallel_executor.h

@@ -40,7 +40,8 @@ class ParallelExecutor {
                             const ProgramDesc& main_program,
                             const std::string& loss_var_name, Scope* scope,
                             const std::vector<Scope*>& local_scopes,
-                            bool allow_op_delay, bool use_default_grad_scale);
+                            bool allow_op_delay, bool use_default_grad_scale,
+                            bool balance_parameter_opt_between_cards);
 
   ~ParallelExecutor();
 

paddle/fluid/framework/var_type_inference_test.cc

@@ -24,8 +24,7 @@ namespace framework {
 
 class SumOpMaker : public OpProtoAndCheckerMaker {
  public:
-  SumOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("X", "").AsDuplicable();
     AddOutput("Out", "");
     AddComment("");

paddle/fluid/operators/CMakeLists.txt

@@ -166,6 +166,8 @@ function(op_library TARGET)
       # NOTE(*): activation use macro to regist the kernels, set use_op manually.
       if(${TARGET} STREQUAL "activation")
         file(APPEND ${pybind_file} "USE_OP(relu);\n")
+      elseif(${TARGET} STREQUAL "reduce")
+        file(APPEND ${pybind_file} "USE_OP(reduce_sum);\n")
       else()
         file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
       endif()
@@ -268,6 +270,11 @@ foreach(src ${READER_LIBRARY})
     set(OP_LIBRARY ${src} ${OP_LIBRARY})
 endforeach()
 
+add_subdirectory(detection)
+foreach(src ${DETECTION_LIBRARY})
+    set(OP_LIBRARY ${src} ${OP_LIBRARY})
+endforeach()
+
 set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
 
 cc_test(gather_test SRCS gather_test.cc DEPS tensor)

paddle/fluid/operators/accuracy_op.cc

@@ -63,8 +63,7 @@ class AccuracyOp : public framework::OperatorWithKernel {
 
 class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AccuracyOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     // TODO(typhoonzero): support both inference value and indices.
     AddInput("Out", "The network output of topk (inferences)");
     AddInput("Indices", "The the network output of topk (indices)");

paddle/fluid/operators/activation_op.cc

@@ -19,19 +19,18 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-#define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT)                  \
-  class OP_NAME##OpMaker                                                   \
-      : public ::paddle::framework::OpProtoAndCheckerMaker {               \
-   public:                                                                 \
-    OP_NAME##OpMaker(OpProto *proto, OpAttrChecker *op_checker)            \
-        : ::paddle::framework::OpProtoAndCheckerMaker(proto, op_checker) { \
-      AddInput("X", "Input of " #OP_NAME "operator");                      \
-      AddOutput("Out", "Output of" #OP_NAME "operator");                   \
-      AddAttr<bool>("use_mkldnn",                                          \
-                    "(bool, default false) Only used in mkldnn kernel")    \
-          .SetDefault(false);                                              \
-      AddComment(#OP_COMMENT);                                             \
-    }                                                                      \
+#define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT)               \
+  class OP_NAME##OpMaker                                                \
+      : public ::paddle::framework::OpProtoAndCheckerMaker {            \
+   public:                                                              \
+    void Make() override {                                              \
+      AddInput("X", "Input of " #OP_NAME "operator");                   \
+      AddOutput("Out", "Output of" #OP_NAME "operator");                \
+      AddAttr<bool>("use_mkldnn",                                       \
+                    "(bool, default false) Only used in mkldnn kernel") \
+          .SetDefault(false);                                           \
+      AddComment(#OP_COMMENT);                                          \
+    }                                                                   \
   }
 
 #define REGISTER_ACTIVATION_OP_GRAD_MAKER(OP_NAME, KERNEL_TYPE)              \
@@ -204,8 +203,7 @@ $$out = \frac{x}{1 + |x|}$$
 
 class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LeakyReluOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of LeakyRelu operator");
     AddOutput("Out", "Output of LeakyRelu operator");
     AddAttr<float>("alpha", "The small negative slope").SetDefault(0.02f);
@@ -220,8 +218,7 @@ $out = \max(x, \alpha * x)$
 
 class SoftShrinkOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SoftShrinkOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of Softshrink operator");
     AddOutput("Out", "Output of Softshrink operator");
     AddAttr<float>("lambda", "non-negative offset").SetDefault(0.5f);
@@ -242,8 +239,7 @@ $$
 
 class HardShrinkOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  HardShrinkOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of HardShrink operator");
     AddOutput("Out", "Output of HardShrink operator");
     AddAttr<float>("threshold", "The value of threshold for HardShrink")
@@ -265,8 +261,7 @@ $$
 
 class BReluOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BReluOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of BRelu operator");
     AddOutput("Out", "Output of BRelu operator");
     AddAttr<float>("t_min", "The min marginal value of BRelu")
@@ -284,8 +279,7 @@ $out = \max(\min(x, t_{min}), t_{max})$
 
 class SoftReluOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SoftReluOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of SoftRelu operator");
     AddOutput("Out", "Output of SoftRelu operator");
     AddAttr<float>("threshold", "The threshold value of SoftRelu")
@@ -301,8 +295,7 @@ $out = \ln(1 + \exp(\max(\min(x, threshold), threshold))$
 
 class ELUOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ELUOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of ELU operator");
     AddOutput("Out", "Output of ELU operator");
     AddAttr<float>("alpha", "The alpha value of ELU").SetDefault(1.0f);
@@ -320,8 +313,7 @@ $out = \max(0, x) + \min(0, \alpha * (e^x - 1))$
 
 class Relu6OpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  Relu6OpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of Relu6 operator");
     AddOutput("Out", "Output of Relu6 operator");
     AddAttr<float>("threshold", "The threshold value of Relu6")
@@ -337,8 +329,7 @@ $out = \min(\max(0, x), 6)$
 
 class PowOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  PowOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of Pow operator");
     AddOutput("Out", "Output of Pow operator");
     AddAttr<float>("factor", "The exponential factor of Pow").SetDefault(1.0f);
@@ -353,8 +344,7 @@ $out = x^{factor}$
 
 class STanhOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  STanhOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of STanh operator");
     AddOutput("Out", "Output of STanh operator");
     AddAttr<float>("scale_a", "The scale parameter of a for the input")
@@ -372,8 +362,7 @@ $$out = b * \frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$
 
 class ThresholdedReluOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ThresholdedReluOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of ThresholdedRelu operator");
     AddOutput("Out", "Output of ThresholdedRelu operator");
     AddAttr<float>("threshold", "The threshold location of activation")
@@ -394,8 +383,7 @@ $$
 
 class HardSigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  HardSigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of HardSigmoid operator");
     AddOutput("Out", "Output of HardSigmoid operator");
     AddAttr<float>("slope", "Slope for linear approximation of sigmoid")
@@ -420,8 +408,7 @@ It is recommended to use the defaults for this activation.
 
 class SwishOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SwishOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of Swish operator");
     AddOutput("Out", "Output of Swish operator");
     AddAttr<float>("beta", "Constant beta of swish operator").SetDefault(1.0f);

paddle/fluid/operators/adadelta_op.cc

@@ -66,8 +66,7 @@ class AdadeltaOp : public framework::OperatorWithKernel {
 
 class AdadeltaOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AdadeltaOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Param", "(Tensor) Input parameter");
     AddInput("Grad", "(Tensor) Input gradient");
     AddInput("AvgSquaredGrad", "(Tensor) Input average of squared gradient");

paddle/fluid/operators/adagrad_op.cc

@@ -67,8 +67,7 @@ class AdagradOp : public framework::OperatorWithKernel {
 
 class AdagradOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AdagradOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Param", "(Tensor) Input parameter");
     AddInput("Grad", "(Tensor) Input gradient");
     AddInput("Moment", "(Tensor) Second moment");

paddle/fluid/operators/adam_op.cc

@@ -80,8 +80,7 @@ class AdamOp : public framework::OperatorWithKernel {
 
 class AdamOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AdamOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Param", "(Tensor) Input parameter");
     AddInput("Grad", "(Tensor) Input gradient");
     AddInput("LearningRate", "(Tensor) Learning rate");

paddle/fluid/operators/adamax_op.cc

@@ -74,8 +74,7 @@ class AdamaxOp : public framework::OperatorWithKernel {
 
 class AdamaxOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AdamaxOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Param", "(Tensor) Input parameter");
     AddInput("Grad", "(Tensor) Input gradient");
     AddInput("LearningRate", "(Tensor) Learning rate");

paddle/fluid/operators/array_to_lod_tensor_op.cc

@@ -123,8 +123,7 @@ class ArrayToLoDTensorOp : public framework::OperatorBase {
 
 class ArrayToLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ArrayToLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(std::vector<LodTensor>) A vector of tensors that is going to "
              "be casted to a big LoDTensor.");

paddle/fluid/operators/assign_op.cc

@@ -94,8 +94,7 @@ class AssignOp : public framework::OperatorBase {
 
 class AssignOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AssignOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LoDTensor, SelectedRows or LoDTensorArray) The input variable "
              "could be LoDTensor, SelectedRows or LoDTensorArray.")

paddle/fluid/operators/assign_value_op.cc

@@ -45,8 +45,7 @@ class AssignValueOp : public framework::OperatorWithKernel {
 
 class AssignValueOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AssignValueOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddOutput("Out", "(Tensor) Output tensor of assign_value operator.");
     AddAttr<std::vector<int>>("shape",
                               "(vector<int>) "

paddle/fluid/operators/auc_op.cc

@@ -50,8 +50,7 @@ class AucOp : public framework::OperatorWithKernel {
 
 class AucOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AucOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Out",
              "A floating point 2D tensor, values are in the range [0, 1]."
              "Each row is sorted in descending order. This input should be the"

paddle/fluid/operators/average_accumulates_op.cc

@@ -111,8 +111,7 @@ class AverageAccumulatesOp : public framework::OperatorWithKernel {
 
 class AverageAccumulatesOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AverageAccumulatesOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("param", "(Tensor), The parameter to be accumulated.");
     AddInput("in_sum_1",
              "(Tensor), A tensor used to store the parameter "

paddle/fluid/operators/batch_norm_op.cc

@@ -126,8 +126,7 @@ class BatchNormOp : public framework::OperatorWithKernel {
 
 class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BatchNormOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddAttr<bool>("is_test", "").SetDefault(false);
     AddAttr<float>("momentum", "").SetDefault(0.9);
     AddAttr<float>("epsilon", "")

paddle/fluid/operators/batch_size_like.h

@@ -53,8 +53,7 @@ class BatchSizeLikeOp : public framework::OperatorWithKernel {
 
 class BatchSizeLikeOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() final {
     AddInput("Input",
              "(Tensor) Tensor "
              "whose input_dim_idx'th dimension specifies the batch_size");
@@ -68,7 +67,11 @@ class BatchSizeLikeOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<int>("output_dim_idx",
                  "(int, default 0) The index of output's batch size dimension")
         .SetDefault(0);
+    Apply();
   }
+
+ protected:
+  virtual void Apply() = 0;
 };
 
 }  // namespace operators

paddle/fluid/operators/beam_search_decode_op.cc

@@ -134,8 +134,7 @@ class BeamSearchDecodeOp : public framework::OperatorBase {
 
 class BeamSearchDecodeOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BeamSearchDecodeOpProtoMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Ids",
              "(LodTensorArray)"
              "score of the candidate words in each step");

paddle/fluid/operators/beam_search_op.cc

@@ -197,8 +197,7 @@ std::string ItemToString(const BeamSearch::Item &item) {
 
 class BeamSearchOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BeamSearchOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     // inputs and outputs stored in proto
     AddInput("pre_ids", "ids in previous step");
     AddInput("ids", "a LoDTensor of shape of [None,k]");

paddle/fluid/operators/bilinear_interp_op.cc

@@ -41,8 +41,7 @@ class BilinearInterpOp : public framework::OperatorWithKernel {
 
 class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BilinearInterpOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor) The input tensor of bilinear interpolation, "
              "This is a 4-D tensor with shape of (N x C x h x w)");

paddle/fluid/operators/bilinear_tensor_product_op.cc

@@ -65,8 +65,7 @@ class BilinearTensorProductOp : public framework::OperatorWithKernel {
 
 class BilinearTensorProductOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BilinearTensorProductOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The first input of bilinear_tensor_product operator.");
     AddInput("Y", "The second input of bilinear_tensor_product operator.");
     AddInput("Weight",

paddle/fluid/operators/cast_op.cc

@@ -21,8 +21,7 @@ namespace operators {
 
 class CastOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CastOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input tensor of cast op");
     AddOutput("Out", "The output tensor of cast op");
     AddAttr<int>("out_dtype", "output data type");

paddle/fluid/operators/channel_close_op.cc

@@ -50,8 +50,7 @@ class ChannelCloseOpOpInferShape : public framework::InferShapeBase {
 
 class ChannelCloseOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ChannelCloseOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(kChannel,
              "The Channel Variable that should be closed by"
              " the ChannelClose Op.");

paddle/fluid/operators/channel_create_op.cc

@@ -91,8 +91,7 @@ class ChannelCreateOpOpInferShape : public framework::InferShapeBase {
 
 class ChannelCreateOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ChannelCreateOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddOutput(kOutput,
               "The object of a Channel type created by ChannelCreate Op.");
     AddAttr<int>("capacity", "The size of the buffer of Channel.")

paddle/fluid/operators/channel_recv_op.cc

@@ -72,8 +72,7 @@ class ChannelRecvOp : public framework::OperatorBase {
 
 class ChannelRecvOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ChannelRecvOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(Channel,
              "(Channel) A variable which \"receives\" the a value sent"
              "to it by a channel_send op.")

paddle/fluid/operators/channel_send_op.cc

@@ -57,8 +57,7 @@ class ChannelSendOp : public framework::OperatorBase {
 
 class ChannelSendOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ChannelSendOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(Channel,
              "(Channel) A variable which \"sends\" the passed in value to "
              "a listening receiver.")

paddle/fluid/operators/chunk_eval_op.cc

@@ -66,8 +66,7 @@ class ChunkEvalOp : public framework::OperatorWithKernel {
 
 class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ChunkEvalOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Inference",
              "(Tensor, default: Tensor<int64_t>). "
              "Predictions from the network.");

paddle/fluid/operators/clip_by_norm_op.cc

@@ -37,8 +37,7 @@ class ClipByNormOp : public framework::OperatorWithKernel {
 
 class ClipByNormOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ClipByNormOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor) The input of clip_by_norm op."
              "The number of dimensions must be between [1, 9].");

paddle/fluid/operators/clip_op.cc

@@ -38,8 +38,7 @@ class ClipOp : public framework::OperatorWithKernel {
 template <typename AttrType>
 class ClipOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ClipOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor)The input of clip op."
              "The number of dimensions must be between [1, 9].");

paddle/fluid/operators/compare_op.cc

@@ -21,8 +21,7 @@ namespace operators {
 template <typename OpComment>
 class CompareOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CompareOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     OpComment comment;
     AddInput("X",
              string::Sprintf("(LoDTensor) the left hand operand of %s operator",

paddle/fluid/operators/concat_op.cc

@@ -63,8 +63,7 @@ class ConcatOp : public framework::OperatorWithKernel {
 
 class ConcatOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ConcatOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input tensors of concat operator.").AsDuplicable();
     AddOutput("Out", "Output tensor of concat operator.");
     AddAttr<int>("axis",

paddle/fluid/operators/conditional_block_op.cc

@@ -108,8 +108,7 @@ class ConditionalBlockOp : public ConditionalOp {
 
 class ConditionalBlockOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ConditionalBlockOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "The conditional variable of this operator. If X is empty, the "
              "whole sub-block will not be executed.")

paddle/fluid/operators/conv_op.cc

@@ -106,8 +106,7 @@ framework::OpKernelType ConvOp::GetExpectedKernelType(
                                  library);
 }
 
-Conv2DOpMaker::Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-    : OpProtoAndCheckerMaker(proto, op_checker) {
+void Conv2DOpMaker::Make() {
   AddInput(
       "Input",
       "(Tensor) The input tensor of convolution operator. "
@@ -200,8 +199,7 @@ $$
 )DOC");
 }
 
-Conv3DOpMaker::Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-    : OpProtoAndCheckerMaker(proto, op_checker) {
+void Conv3DOpMaker::Make() {
   AddInput(
       "Input",
       "(Tensor) The input tensor of convolution operator. "

paddle/fluid/operators/conv_op.h

@@ -60,12 +60,12 @@ inline bool IsExpand(const std::vector<int64_t>& filter_dim,
 // operator implementations can reuse the code.
 class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker);
+  void Make() override;
 };
 
 class Conv3DOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker);
+  void Make() override;
 };
 
 class ConvOp : public framework::OperatorWithKernel {

paddle/fluid/operators/conv_shift_op.cc

@@ -75,8 +75,7 @@ class ConvShiftGradOp : public framework::OperatorWithKernel {
 
 class ConvShiftOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ConvShiftOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor, default Tensor<float>), a 2-D tensor with shape B x M, "
              "where B is the batch size and M is the data dimension.");

paddle/fluid/operators/conv_transpose_op.cc

@@ -84,9 +84,7 @@ framework::OpKernelType ConvTransposeOp::GetExpectedKernelType(
       layout_, library_);
 }
 
-Conv2DTransposeOpMaker::Conv2DTransposeOpMaker(OpProto* proto,
-                                               OpAttrChecker* op_checker)
-    : OpProtoAndCheckerMaker(proto, op_checker) {
+void Conv2DTransposeOpMaker::Make() {
   AddInput(
       "Input",
       "(Tensor) The input tensor of convolution transpose operator. "
@@ -168,9 +166,7 @@ Example:
 )DOC");
 }
 
-Conv3DTransposeOpMaker::Conv3DTransposeOpMaker(OpProto* proto,
-                                               OpAttrChecker* op_checker)
-    : OpProtoAndCheckerMaker(proto, op_checker) {
+void Conv3DTransposeOpMaker::Make() {
   AddInput("Input",
            "(Tensor) The input tensor of convolution transpose operator."
            "The format of input tensor is NCDHW. Where N is batch size, C is "

paddle/fluid/operators/conv_transpose_op.h

@@ -30,12 +30,12 @@ using DDim = framework::DDim;
 // operator implementations can reuse the code.
 class Conv2DTransposeOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  Conv2DTransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker);
+  void Make() override;
 };
 
 class Conv3DTransposeOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  Conv3DTransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker);
+  void Make() override;
 };
 
 class ConvTransposeOp : public framework::OperatorWithKernel {

paddle/fluid/operators/cos_sim_op.cc

@@ -62,8 +62,7 @@ class CosSimOp : public framework::OperatorWithKernel {
 
 class CosSimOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CosSimOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The 1st input of cos_sim op.");
     AddInput("Y", "The 2nd input of cos_sim op.");
     AddOutput("Out", "The output of cos_sim op.");

paddle/fluid/operators/crf_decoding_op.cc

@@ -18,8 +18,7 @@ namespace paddle {
 namespace operators {
 class CRFDecodingOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CRFDecodingOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Emission",
              "(LoDTensor, default: LoDTensor<float>). A LoDTensor with shape "
              "[N x D] where N is the size of the mini-batch and D is the total "

paddle/fluid/operators/crop_op.cc

@@ -52,8 +52,7 @@ class CropOp : public framework::OperatorWithKernel {
 
 class CropOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CropOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "The input of pad op. "
              "The input should be a k-D tensor(k > 0 and k < 7).");

paddle/fluid/operators/cross_entropy_op.cc

@@ -111,8 +111,7 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
 
 class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CrossEntropyOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor, default Tensor<float>), a 2-D tensor with shape [N x D],"
              " where N is the batch size and D is the number of classes. "

paddle/fluid/operators/ctc_align_op.cc

@@ -44,8 +44,7 @@ class CTCAlignOp : public framework::OperatorWithKernel {
 
 class CTCAlignOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CTCAlignOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Input",
              "(LodTensor, default: LoDTensor<int>), Its shape is "
              "[Lp, 1], where Lp is the sum of all input sequences' length.");

paddle/fluid/operators/cumsum_op.cc

@@ -29,8 +29,7 @@ class CumOp : public framework::OperatorWithKernel {
 
 class CumsumOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CumsumOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "Input of Cumsum operator");
     AddOutput("Out", "Output of Cumsum operator");
     AddAttr<int>("axis",

paddle/fluid/operators/decayed_adagrad_op.cc

@@ -62,8 +62,7 @@ class DecayedAdagradOp : public framework::OperatorWithKernel {
 
 class DecayedAdagradOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  DecayedAdagradOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Param", "(Tensor) Input parameter");
     AddInput("Grad", "(Tensor) Input gradient");
     AddInput("Moment", "(Tensor) Second moment");

paddle/fluid/operators/delete_var_op.cc

@@ -34,8 +34,7 @@ class DeleteVarOp : public framework::OperatorBase {
 
 class DeleteVarOpInfoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  DeleteVarOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input of delete op").AsDuplicable();
     AddComment(R"DOC(
 Delete Operator.

paddle/fluid/operators/detection/CMakeLists.txt

+set(LOCAL_DETECTION_LIBS)
+
+function(detection_library TARGET_NAME)
+    set(oneValueArgs "")
+    set(multiValueArgs SRCS DEPS)
+    set(options "")
+    set(common_deps op_registry)
+    set(pybind_flag 0)
+    cmake_parse_arguments(detection_library "${options}" "${oneValueArgs}"
+            "${multiValueArgs}" ${ARGN})
+    op_library(${TARGET_NAME} SRCS ${detection_library_SRCS} DEPS ${common_deps} ${detection_library_DEPS})
+    set(LOCAL_DETECTION_LIBS
+            ${TARGET_NAME}
+            ${LOCAL_DETECTION_LIBS}
+        PARENT_SCOPE)
+endfunction()
+
+detection_library(bipartite_match_op SRCS bipartite_match_op.cc)
+detection_library(box_coder_op SRCS box_coder_op.cc box_coder_op.cu)
+detection_library(iou_similarity_op SRCS iou_similarity_op.cc
+iou_similarity_op.cu)
+detection_library(mine_hard_examples_op SRCS mine_hard_examples_op.cc)
+detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc)
+detection_library(prior_box_op SRCS prior_box_op.cc prior_box_op.cu)
+detection_library(target_assign_op SRCS target_assign_op.cc
+target_assign_op.cu)
+
+# Export local libraries to parent
+set(DETECTION_LIBRARY ${LOCAL_DETECTION_LIBS} PARENT_SCOPE)

paddle/fluid/operators/bipartite_match_op.cc → paddle/fluid/operators/detection/bipartite_match_op.cc

@@ -182,8 +182,7 @@ class BipartiteMatchKernel : public framework::OpKernel<T> {
 
 class BipartiteMatchOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BipartiteMatchOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(
         "DistMat",
         "(LoDTensor or Tensor) this input is a 2-D LoDTensor with shape "

paddle/fluid/operators/box_coder_op.cc → paddle/fluid/operators/detection/box_coder_op.cc

@@ -9,7 +9,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/operators/box_coder_op.h"
+#include "paddle/fluid/operators/detection/box_coder_op.h"
 
 namespace paddle {
 namespace operators {
@@ -60,8 +60,7 @@ class BoxCoderOp : public framework::OperatorWithKernel {
 
 class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BoxCoderOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(
         "PriorBox",
         "(Tensor, default Tensor<float>) "

paddle/fluid/operators/box_coder_op.cu → paddle/fluid/operators/detection/box_coder_op.cu

@@ -9,7 +9,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/operators/box_coder_op.h"
+#include "paddle/fluid/operators/detection/box_coder_op.h"
 #include "paddle/fluid/platform/cuda_primitives.h"
 
 namespace paddle {

paddle/fluid/operators/iou_similarity_op.cc → paddle/fluid/operators/detection/iou_similarity_op.cc

@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/operators/iou_similarity_op.h"
+#include "paddle/fluid/operators/detection/iou_similarity_op.h"
 
 namespace paddle {
 namespace operators {
@@ -42,8 +42,7 @@ class IOUSimilarityOp : public framework::OperatorWithKernel {
 
 class IOUSimilarityOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  IOUSimilarityOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LoDTensor, default LoDTensor<float>) "
              "Box list X is a 2-D LoDTensor with shape [N, 4] holds N boxes, "

paddle/fluid/operators/iou_similarity_op.cu → paddle/fluid/operators/detection/iou_similarity_op.cu

@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/operators/iou_similarity_op.h"
+#include "paddle/fluid/operators/detection/iou_similarity_op.h"
 
 namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(

paddle/fluid/operators/mine_hard_examples_op.cc → paddle/fluid/operators/detection/mine_hard_examples_op.cc

@@ -253,8 +253,7 @@ class MineHardExamplesOp : public framework::OperatorWithKernel {
 
 class MineHardExamplesOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MineHardExamplesOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(
         "ClsLoss",
         "(Tensor, default Tensor<float>), The classification loss with shape "

paddle/fluid/operators/multiclass_nms_op.cc → paddle/fluid/operators/detection/multiclass_nms_op.cc

@@ -309,8 +309,7 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
 
 class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MultiClassNMSOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("BBoxes",
              "(Tensor) A 3-D Tensor with shape [N, M, 4] represents the "
              "predicted locations of M bounding bboxes, N is the batch size. "

paddle/fluid/operators/prior_box_op.cc → paddle/fluid/operators/detection/prior_box_op.cc

@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/operators/prior_box_op.h"
+#include "paddle/fluid/operators/detection/prior_box_op.h"
 
 namespace paddle {
 namespace operators {
@@ -79,8 +79,7 @@ class PriorBoxOp : public framework::OperatorWithKernel {
 
 class PriorBoxOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  PriorBoxOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Input",
              "(Tensor, default Tensor<float>), "
              "the input feature data of PriorBoxOp, The layout is NCHW.");

paddle/fluid/operators/prior_box_op.cu → paddle/fluid/operators/detection/prior_box_op.cu

@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/operators/prior_box_op.h"
+#include "paddle/fluid/operators/detection/prior_box_op.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/target_assign_op.cc → paddle/fluid/operators/detection/target_assign_op.cc

@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/operators/target_assign_op.h"
+#include "paddle/fluid/operators/detection/target_assign_op.h"
 
 namespace paddle {
 namespace operators {
@@ -65,8 +65,7 @@ class TargetAssignOp : public framework::OperatorWithKernel {
 
 class TargetAssignOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  TargetAssignOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LoDTensor), This input is a 3D LoDTensor with shape [M, P, K]. "
              "Some elements in X will be assigned to Out based on the "

paddle/fluid/operators/target_assign_op.cu → paddle/fluid/operators/detection/target_assign_op.cu

@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/operators/target_assign_op.h"
+#include "paddle/fluid/operators/detection/target_assign_op.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/detection_map_op.cc

@@ -78,8 +78,7 @@ class DetectionMAPOp : public framework::OperatorWithKernel {
 
 class DetectionMAPOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  DetectionMAPOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("DetectRes",
              "(LoDTensor) A 2-D LoDTensor with shape [M, 6] represents the "
              "detections. Each row has 6 values: "

paddle/fluid/operators/dropout_op.cc

@@ -37,8 +37,7 @@ class DropoutOp : public framework::OperatorWithKernel {
 
 class DropoutOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  DropoutOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input of dropout op.");
     AddOutput("Out", "The output of dropout op.");
     AddOutput("Mask", "The random sampled dropout mask.").AsIntermediate();

paddle/fluid/operators/edit_distance_op.cc

@@ -49,8 +49,7 @@ class EditDistanceOp : public framework::OperatorWithKernel {
 
 class EditDistanceOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  EditDistanceOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Hyps",
              "(2-D LoDTensor<int64_t>, 2nd dim. equal to 1) "
              "The indices for hypothesis strings.");

paddle/fluid/operators/elementwise_add_op.cc

@@ -14,26 +14,8 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/elementwise_add_op.h"
 #include "paddle/fluid/operators/elementwise_op.h"
-
-namespace paddle {
-namespace operators {
-class ElementwiseAddOpMaker : public ElementwiseOpMaker {
- public:
-  ElementwiseAddOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : ElementwiseOpMaker(proto, op_checker) {
-    SetComment("Add", "Out = X + Y");
-    AddComment(comment_);
-  }
-};
-}  // namespace operators
-}  // namespace paddle
-
 namespace ops = paddle::operators;
-REGISTER_OPERATOR(elementwise_add, ops::ElementwiseOp,
-                  ops::ElementwiseAddOpMaker, ops::ElementwiseOpInferVarType,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(elementwise_add_grad, ops::ElementwiseOpGrad);
-
+REGISTER_ELEMWISE_OP(elementwise_add, "Add", "Out = X + Y");
 REGISTER_OP_CPU_KERNEL(
     elementwise_add,
     ops::ElementwiseAddKernel<paddle::platform::CPUDeviceContext, float>,

paddle/fluid/operators/elementwise_div_op.cc

@@ -14,26 +14,8 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/elementwise_div_op.h"
 #include "paddle/fluid/operators/elementwise_op.h"
-
-namespace paddle {
-namespace operators {
-class ElementwiseDivOpMaker : public ElementwiseOpMaker {
- public:
-  ElementwiseDivOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : ElementwiseOpMaker(proto, op_checker) {
-    SetComment("Div", "Out = X / Y");
-    AddComment(comment_);
-  }
-};
-
-}  // namespace operators
-}  // namespace paddle
-
 namespace ops = paddle::operators;
-REGISTER_OPERATOR(elementwise_div, ops::ElementwiseOp,
-                  ops::ElementwiseDivOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(elementwise_div_grad, ops::ElementwiseOpGrad);
+REGISTER_ELEMWISE_OP(elementwise_div, "Div", "Out = X / Y");
 REGISTER_OP_CPU_KERNEL(
     elementwise_div,
     ops::ElementwiseDivKernel<paddle::platform::CPUDeviceContext, float>,

paddle/fluid/operators/elementwise_max_op.cc

@@ -14,25 +14,8 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/elementwise_max_op.h"
 #include "paddle/fluid/operators/elementwise_op.h"
-
-namespace paddle {
-namespace operators {
-class ElementwiseMaxOpMaker : public ElementwiseOpMaker {
- public:
-  ElementwiseMaxOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : ElementwiseOpMaker(proto, op_checker) {
-    SetComment("Max", "Out = max(X, Y)");
-    AddComment(comment_);
-  }
-};
-}  // namespace operators
-}  // namespace paddle
-
 namespace ops = paddle::operators;
-REGISTER_OPERATOR(elementwise_max, ops::ElementwiseOp,
-                  ops::ElementwiseMaxOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(elementwise_max_grad, ops::ElementwiseOpGrad);
+REGISTER_ELEMWISE_OP(elementwise_max, "Max", "Out = max(X, Y)");
 REGISTER_OP_CPU_KERNEL(
     elementwise_max,
     ops::ElementwiseMaxKernel<paddle::platform::CPUDeviceContext, float>,

paddle/fluid/operators/elementwise_min_op.cc

@@ -14,25 +14,8 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/elementwise_min_op.h"
 #include "paddle/fluid/operators/elementwise_op.h"
-
-namespace paddle {
-namespace operators {
-class ElementwiseMinOpMaker : public ElementwiseOpMaker {
- public:
-  ElementwiseMinOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : ElementwiseOpMaker(proto, op_checker) {
-    SetComment("Max", "Out = min(X, Y)");
-    AddComment(comment_);
-  }
-};
-}  // namespace operators
-}  // namespace paddle
-
 namespace ops = paddle::operators;
-REGISTER_OPERATOR(elementwise_min, ops::ElementwiseOp,
-                  ops::ElementwiseMinOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(elementwise_min_grad, ops::ElementwiseOpGrad);
+REGISTER_ELEMWISE_OP(elementwise_min, "Min", "Out = min(X, Y)");
 REGISTER_OP_CPU_KERNEL(
     elementwise_min,
     ops::ElementwiseMinKernel<paddle::platform::CPUDeviceContext, float>,

paddle/fluid/operators/elementwise_mul_op.cc

@@ -14,27 +14,8 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/elementwise_mul_op.h"
 #include "paddle/fluid/operators/elementwise_op.h"
-
-namespace paddle {
-namespace operators {
-
-class ElementwiseMulOpMaker : public ElementwiseOpMaker {
- public:
-  ElementwiseMulOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : ElementwiseOpMaker(proto, op_checker) {
-    SetComment("Mul", "Out = X \\odot\\ Y");
-    AddComment(comment_);
-  }
-};
-
-}  // namespace operators
-}  // namespace paddle
-
 namespace ops = paddle::operators;
-REGISTER_OPERATOR(elementwise_mul, ops::ElementwiseOp,
-                  ops::ElementwiseMulOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(elementwise_mul_grad, ops::ElementwiseOpGrad);
+REGISTER_ELEMWISE_OP(elementwise_mul, "Mul", "Out = X \\odot\\ Y");
 REGISTER_OP_CPU_KERNEL(
     elementwise_mul,
     ops::ElementwiseMulKernel<paddle::platform::CPUDeviceContext, float>,

paddle/fluid/operators/elementwise_op.h

@@ -54,8 +54,7 @@ class ElementwiseOpInferVarType : public framework::VarTypeInference {
 
 class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ElementwiseOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() final {
     AddInput("X", "(Tensor), The first input tensor of elementwise op.");
     AddInput("Y", "(Tensor), The second input tensor of elementwise op.");
     AddOutput("Out", "The output of elementwise op.");
@@ -64,12 +63,12 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker {
                  "for broadcasting Y onto X.")
         .SetDefault(-1)
         .EqualGreaterThan(-1);
-    comment_ = R"DOC(
-Limited Elementwise {name} Operator.
+    AddComment(string::Sprintf(R"DOC(
+Limited Elementwise %s Operator.
 
 The equation is:
 
-$${equation}$$
+$$%s$$
 
 $X$ is a tensor of any dimension and the dimensions of tensor $Y$ must be
 smaller than or equal to the dimensions of $X$.
@@ -100,26 +99,13 @@ For example
 Either of the inputs $X$ and $Y$ or none can carry the LoD (Level of Details)
 information. However, the output only shares the LoD information with input $X$.
 
-)DOC";
-    AddComment(comment_);
+)DOC",
+                               GetName(), GetEquation()));
   }
 
  protected:
-  std::string comment_;
-
-  void Replace(std::string* src, std::string from, std::string to) {
-    std::size_t len_from = std::strlen(from.c_str());
-    std::size_t len_to = std::strlen(to.c_str());
-    for (std::size_t pos = src->find(from); pos != std::string::npos;
-         pos = src->find(from, pos + len_to)) {
-      src->replace(pos, len_from, to);
-    }
-  }
-
-  void SetComment(std::string name, std::string equation) {
-    Replace(&comment_, "{name}", name);
-    Replace(&comment_, "{equation}", equation);
-  }
+  virtual std::string GetName() const = 0;
+  virtual std::string GetEquation() const = 0;
 };
 
 class ElementwiseOpGrad : public framework::OperatorWithKernel {
@@ -152,3 +138,16 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel {
 };
 }  // namespace operators
 }  // namespace paddle
+
+#define REGISTER_ELEMWISE_OP(op_type, op_name, equation)                \
+  class __ElemwiseOp##op_type##Maker__                                  \
+      : public ::paddle::operators::ElementwiseOpMaker {                \
+   protected:                                                           \
+    virtual std::string GetName() const { return op_name; }             \
+    virtual std::string GetEquation() const { return equation; }        \
+  };                                                                    \
+  REGISTER_OPERATOR(op_type, ::paddle::operators::ElementwiseOp,        \
+                    __ElemwiseOp##op_type##Maker__,                     \
+                    ::paddle::operators::ElementwiseOpInferVarType,     \
+                    ::paddle::framework::DefaultGradOpDescMaker<true>); \
+  REGISTER_OPERATOR(op_type##_grad, ::paddle::operators::ElementwiseOpGrad)

paddle/fluid/operators/elementwise_pow_op.cc

@@ -13,17 +13,15 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/elementwise_pow_op.h"
+#include <string>
 #include "paddle/fluid/operators/elementwise_op.h"
 
 namespace paddle {
 namespace operators {
 class ElementwisePowOpMaker : public ElementwiseOpMaker {
- public:
-  ElementwisePowOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : ElementwiseOpMaker(proto, op_checker) {
-    SetComment("Pow", "Out = X ^ Y");
-    AddComment(comment_);
-  }
+ protected:
+  std::string GetName() const override { return "Pow"; }
+  std::string GetEquation() const override { return "Out = X ^ Y"; }
 };
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/elementwise_sub_op.cc

@@ -14,25 +14,8 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/elementwise_sub_op.h"
 #include "paddle/fluid/operators/elementwise_op.h"
-
-namespace paddle {
-namespace operators {
-class ElementwiseSubOpMaker : public ElementwiseOpMaker {
- public:
-  ElementwiseSubOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : ElementwiseOpMaker(proto, op_checker) {
-    SetComment("Sub", "Out = X - Y");
-    AddComment(comment_);
-  }
-};
-}  // namespace operators
-}  // namespace paddle
-
 namespace ops = paddle::operators;
-REGISTER_OPERATOR(elementwise_sub, ops::ElementwiseOp,
-                  ops::ElementwiseSubOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(elementwise_sub_grad, ops::ElementwiseOpGrad);
+REGISTER_ELEMWISE_OP(elementwise_sub, "Sub", "Out = X - Y");
 REGISTER_OP_CPU_KERNEL(
     elementwise_sub,
     ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext, float>,

paddle/fluid/operators/expand_op.cc

@@ -56,8 +56,7 @@ class ExpandOp : public framework::OperatorWithKernel {
 
 class ExpandOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor, default Tensor<float>). A tensor with rank in [1, 6]."
              "X is the input to be expanded.");

paddle/fluid/operators/fc_op.cc

@@ -72,8 +72,7 @@ framework::OpKernelType FCOpGrad::GetExpectedKernelType(
       layout, library);
 }
 
-FCOpMaker::FCOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-    : OpProtoAndCheckerMaker(proto, op_checker) {
+void FCOpMaker::Make() {
   AddInput("Input", "(Tensor) The input tensor of fully connected operator. ");
   AddInput("W", "(Tensor), The second input tensor of fc op.");
   AddOutput("Out", "(Tensor) The output tensor of fully connected operator. ");

paddle/fluid/operators/fc_op.h

@@ -45,7 +45,7 @@ class FCOpGrad : public framework::OperatorWithKernel {
 
 class FCOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  FCOpMaker(OpProto* proto, OpAttrChecker* op_checker);
+  void Make() override;
 };
 
 }  // namespace operators

paddle/fluid/operators/feed_op.cc

@@ -66,8 +66,7 @@ class FeedOp : public framework::OperatorBase {
 
 class FeedOpInfoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  FeedOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input of feed op");
     AddOutput("Out", "The output of feed op");
     AddAttr<int>("col", "(int) The column of feed");

paddle/fluid/operators/fetch_op.cc

@@ -66,8 +66,7 @@ class FetchOp : public framework::OperatorBase {
 
 class FetchOpInfoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  FetchOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input of fetch op");
     AddOutput("Out", "The output of fetch op");
     AddAttr<int>("col", "(int) The column of fetch");

paddle/fluid/operators/fill_constant_batch_size_like_op.cc

@@ -30,9 +30,8 @@ class FillConstantBatchSizeLikeOp : public BatchSizeLikeOp {
 };
 
 class FillConstantBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker {
- public:
-  FillConstantBatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : BatchSizeLikeOpMaker(proto, op_checker) {
+ protected:
+  void Apply() override {
     AddAttr<int>("dtype",
                  "(int, default 5 (FP32)) "
                  "Output data type")

paddle/fluid/operators/fill_constant_op.cc

@@ -59,8 +59,7 @@ class FillConstantOp : public framework::OperatorBase {
 
 class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  FillConstantOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddAttr<int>("dtype",
                  "(int, default 5 (FP32)) "
                  "Output data type")

paddle/fluid/operators/fill_op.cc

@@ -82,8 +82,7 @@ class FillOp : public framework::OperatorBase {
 
 class FillOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  FillOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddComment(R"DOC(Fill operator
 
 Fill an tensor with `value` and `shape`. The type of the tensor is specify by

paddle/fluid/operators/fill_zeros_like_op.cc

@@ -33,8 +33,7 @@ class FillZerosLikeOp : public framework::OperatorWithKernel {
 
 class FillZerosLikeOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  FillZerosLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input of fill-zeros-like op.");
     AddOutput("Out", "The variable will be filled up with zeros.");
     AddComment(R"DOC(

paddle/fluid/operators/ftrl_op.cc

@@ -64,8 +64,7 @@ class FTRLOp : public framework::OperatorWithKernel {
 
 class FTRLOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  FTRLOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Param",
              "(Tensor, default Tensor<float>) "
              "Input parameter value that has to be updated.");

paddle/fluid/operators/gather_op.cc

@@ -67,8 +67,7 @@ class GatherGradOp : public framework::OperatorWithKernel {
 
 class GatherOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  GatherOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The source input of gather op");
     AddInput("Index", "The index input of gather op");
     AddOutput("Out", "The output of gather op");

paddle/fluid/operators/gaussian_random_batch_size_like_op.cc

@@ -32,9 +32,8 @@ class GaussianRandomBatchSizeLikeOp : public BatchSizeLikeOp {
 };
 
 class GaussianRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker {
- public:
-  GaussianRandomBatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : BatchSizeLikeOpMaker(proto, op_checker) {
+ protected:
+  void Apply() override {
     AddAttr<float>("mean",
                    "(float, default 0.0) "
                    "mean of random tensor.")

paddle/fluid/operators/gaussian_random_op.cc

@@ -70,8 +70,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
 
 class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  GaussianRandomOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddOutput("Out", "Output matrix of gaussian random op");
 
     AddAttr<std::vector<int>>("shape",

paddle/fluid/operators/get_places_op.cc

@@ -78,8 +78,7 @@ class GetPlacesOp : public framework::OperatorBase {
 
 class GetPlacesOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  GetPlacesOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddOutput("Out", "vector of Place");
     AddAttr<int>("device_count", "device count").SetDefault(0);
     AddAttr<std::string>("device_type", "device type")

paddle/fluid/operators/go_op.cc

@@ -89,8 +89,7 @@ class GoOp : public framework::OperatorBase {
 
 class GoOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  GoOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(kX,
              "A set of variables, which are required by operators inside the "
              "block of Go Op.")

paddle/fluid/operators/gru_op.cc

@@ -71,8 +71,7 @@ class GRUOp : public framework::OperatorWithKernel {
 
 class GRUOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  GRUOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Input",
              "(LoDTensor) The first input is a LodTensor, which supports "
              "variable-time length input sequence. The underlying tensor in "

paddle/fluid/operators/gru_unit_op.cc

@@ -71,8 +71,7 @@ class GRUUnitOp : public framework::OperatorWithKernel {
 
 class GRUUnitOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  GRUUnitOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Input",
              "(Tensor) Matrix with shape [batch_size, frame_size * 3] for the "
              "input.");

paddle/fluid/operators/hinge_loss_op.cc

@@ -46,8 +46,7 @@ class HingeLossOp : public framework::OperatorWithKernel {
 template <typename AttrType>
 class HingeLossOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  HingeLossOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Logits",
              "The input value (Logits) of Hinge loss op."
              "Logits is a 2-D tensor with shape [batch_size, 1].");

paddle/fluid/operators/huber_loss_op.cc

@@ -45,8 +45,7 @@ class HuberLossOp : public framework::OperatorWithKernel {
 template <typename AttrType>
 class HuberLossOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  HuberLossOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "The input value of huber loss op."
              "X is a 2-D tensor with shape [batch_size, 1].");

paddle/fluid/operators/im2sequence_op.cc

@@ -54,8 +54,7 @@ class Im2SequenceOp : public framework::OperatorWithKernel {
 
 class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  Im2SequenceOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor) The input tensor has NCHW format."
              "N: batch size"

paddle/fluid/operators/increment_op.cc

@@ -47,8 +47,7 @@ class IncrementOp : public framework::OperatorWithKernel {
 
 class IncrementOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  IncrementOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(Tensor) The input tensor of increment operator");
     AddOutput("Out", "(Tensor) The output tensor of increment operator.");
     AddAttr<float>("step",

paddle/fluid/operators/is_empty_op.cc

@@ -48,8 +48,7 @@ class IsEmptyOp : public framework::OperatorBase {
 
 class IsEmptyOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  IsEmptyOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(kInput, "(Tensor) Tensor which is to be checked.");
     AddOutput(kOutput, "(Tensor) a boolean Tensor that indicate empty or not.");
     AddComment(R"DOC(

paddle/fluid/operators/l1_norm_op.cc

@@ -48,8 +48,7 @@ class L1NormGradOp : public framework::OperatorWithKernel {
 
 class L1NormOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  L1NormOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(Tensor) The input of l1_norm op.");
     AddOutput("Out", "(Scalar) The output of l1_norm op.");
     AddComment(R"DOC(

paddle/fluid/operators/label_smooth_op.cc

@@ -47,8 +47,7 @@ class LabelSmoothOp : public framework::OperatorWithKernel {
 
 class LabelSmoothOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LabelSmoothOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LoDTensor) The input labels of LabelSmooth operator. This "
              "input can be batched labels in one-hot encoding or output from "

paddle/fluid/operators/layer_norm_op.cc

@@ -61,8 +61,7 @@ class LayerNormOp : public framework::OperatorWithKernel {
 
 class LayerNormOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LayerNormOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(LoDTensor) The input tensor.");
     AddInput("Scale",
              "(Tensor, optional) Scale is a 1-dimensional tensor of size "

paddle/fluid/operators/linear_chain_crf_op.cc

@@ -19,8 +19,7 @@ namespace operators {
 
 class LinearChainCRFOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LinearChainCRFOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Emission",
              "(LoDTensor, default LoDTensor<float>) "
              "A 2-D LoDTensor with shape [N x D], where N is the size of the "

paddle/fluid/operators/listen_and_serv_op.cc

@@ -343,8 +343,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
 
 class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ListenAndServOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("X", "(Tensor) Variables that server recv.").AsDuplicable();
     AddComment(R"DOC(
 ListenAndServ operator

paddle/fluid/operators/load_combine_op.cc

@@ -77,8 +77,7 @@ class LoadCombineOp : public framework::OperatorBase {
 
 class LoadCombineOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LoadCombineOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddOutput(
         "Out",
         "(vector) The output LoDTensors that will be read from the input file.")

paddle/fluid/operators/load_op.cc

@@ -73,8 +73,7 @@ class LoadOp : public framework::OperatorBase {
 
 class LoadOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LoadOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddOutput("Out", "(Tensor) The tensor need to be loaded");
     AddAttr<bool>(
         "load_as_fp16",

paddle/fluid/operators/lod_array_length_op.cc

@@ -40,8 +40,7 @@ class LoDArrayLengthOp : public framework::OperatorBase {
 
 class LoDArrayLengthProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LoDArrayLengthProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(LoDTensorArray) The input tensor array.");
     AddOutput("Out", "(Tensor) 1x1 CPU Tensor of length, int64_t");
     AddComment(R"DOC(

paddle/fluid/operators/lod_rank_table_op.cc

@@ -38,8 +38,7 @@ class LoDRankTableOp : public framework::OperatorBase {
 
 class LoDRankTableOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LoDRankTableOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LoDTensor) input lod tensor, must contain lod information.");
     AddOutput("Out", "(LoDRankTable) The rank table of specific level.");

paddle/fluid/operators/lod_reset_op.cc

@@ -47,8 +47,7 @@ class LoDResetOp : public framework::OperatorWithKernel {
 
 class LoDResetOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LoDResetOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor, LoDTensor) Input variable of LoDResetOp which "
              "could be a Tensor or LoDTensor, where the data of output "

paddle/fluid/operators/lod_tensor_to_array_op.cc

@@ -105,8 +105,7 @@ class LoDTensorToArrayOp : public framework::OperatorBase {
 
 class LoDTensorToArrayOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LoDTensorToArrayOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "");
     AddInput("RankTable", "");
     AddOutput("Out", "");

paddle/fluid/operators/log_loss_op.cc

@@ -46,8 +46,7 @@ class LogLossOp : public framework::OperatorWithKernel {
 template <typename AttrType>
 class LogLossOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LogLossOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Predicted",
              "The input value (Predicted) of Log loss op."
              "Predicted is a 2-D tensor with shape [batch_size, 1].");

paddle/fluid/operators/logical_op.cc

@@ -21,8 +21,7 @@ namespace operators {
 template <typename OpComment>
 class BinaryLogicalOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BinaryLogicalOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     OpComment comment;
     AddInput("X",
              string::Sprintf("(LoDTensor) Left hand operand of %s operator",
@@ -45,8 +44,7 @@ Each element of Out is calculated by %s
 template <typename OpComment>
 class UnaryLogicalOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  UnaryLogicalOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     OpComment comment;
     AddInput("X", string::Sprintf("(LoDTensor) Operand of %s operator",
                                   comment.type));

paddle/fluid/operators/lookup_sparse_table_op.cc

@@ -105,8 +105,7 @@ class LookupSparseTableOp : public framework::OperatorBase {
 
 class LookupSparseTableOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LookupSparseTableOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("W",
              "(SelectedRows) The input represents embedding table, "
              "which is a learnable parameter.");

paddle/fluid/operators/lookup_table_op.cc

@@ -58,8 +58,7 @@ class LookupTableOp : public framework::OperatorWithKernel {
 
 class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LookupTableOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("W",
              "(Tensor) The input represents embedding tensors, "
              "which is a learnable parameter.");

paddle/fluid/operators/lrn_op.cc

@@ -169,8 +169,7 @@ class LRNOp : public framework::OperatorWithKernel {
 template <typename T>
 class LRNOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LRNOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor) The input of LRN operator. "
              "It must be a 4D tenor with NCHW format.");

paddle/fluid/operators/lstm_op.cc

@@ -103,8 +103,7 @@ class LSTMOp : public framework::OperatorWithKernel {
 
 class LSTMOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LSTMOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Input",
              "(LoDTensor) the first input is a LodTensor, which support "
              "variable-time length input sequence. The underlying tensor in "

paddle/fluid/operators/lstm_unit_op.cc

@@ -48,8 +48,7 @@ class LstmUnitOp : public framework::OperatorWithKernel {
 
 class LstmUnitOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LstmUnitOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "Lstm unit only applies non-linear activations, please make sure"
              "that linear tranformation has already been applied to `X`. "

paddle/fluid/operators/lstmp_op.cc

@@ -120,8 +120,7 @@ class LSTMPOp : public framework::OperatorWithKernel {
 
 class LSTMPOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LSTMPOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Input",
              "(LoDTensor) the input for sequence data, which supports "
              "variable-time length input sequence. The underlying tensor in "

paddle/fluid/operators/margin_rank_loss_op.cc

@@ -42,8 +42,7 @@ class MarginRankLossOp : public framework::OperatorWithKernel {
 template <typename T>
 class MarginRankLossOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MarginRankLossOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X1",
              "(2-D tensor with shape [batch_size x 1]) The score for "
              "one item X1 to be ranked, from pairwise ranking model.");

paddle/fluid/operators/math/blas_impl.cu.h

@@ -96,10 +96,22 @@ struct CUBlas<platform::float16> {
                                        reinterpret_cast<__half *>(C), ldc));
   }
 
-  template <typename... ARGS>
-  static void GEMM_BATCH(ARGS... args) {
+  static void GEMM_BATCH(cublasHandle_t handle, cublasOperation_t transa,
+                         cublasOperation_t transb, int m, int n, int k,
+                         const float16 *alpha, const float16 *A, int lda,
+                         long long int strideA, const float16 *B,  // NOLINT
+                         int ldb, long long int strideB,           // NOLINT
+                         const float16 *beta, float16 *C, int ldc,
+                         long long int strideC,  // NOLINT
+                         int batchCount) {
 #if CUDA_VERSION >= 8000
-    PADDLE_ENFORCE(platform::dynload::cublasHgemmStridedBatched(args...));
+    PADDLE_ENFORCE(platform::dynload::cublasHgemmStridedBatched(
+        handle, transa, transb, m, n, k,
+        reinterpret_cast<const __half *>(alpha),
+        reinterpret_cast<const __half *>(A), lda, strideA,
+        reinterpret_cast<const __half *>(B), ldb, strideB,
+        reinterpret_cast<const __half *>(beta), reinterpret_cast<__half *>(C),
+        ldc, strideC, batchCount));
 #else
     PADDLE_THROW("HgemmStridedBatched is not supported on cuda <= 7.5");
 #endif

paddle/fluid/operators/math/blas_impl.h

@@ -172,9 +172,9 @@ void Blas<platform::CPUDeviceContext>::BatchedGEMM(
                        c_array.data(), &ldc, 1 /* group_count */, &batchCount);
 #else
   for (int k = 0; k < batchCount; ++k) {
-    const float *Ak = &A[k * strideA];
-    const float *Bk = &B[k * strideB];
-    float *Ck = &C[k * M * N];
+    auto *Ak = &A[k * strideA];
+    auto *Bk = &B[k * strideB];
+    auto *Ck = &C[k * M * N];
     this->template GEMM<T>(transA, transB, M, N, K, alpha, Ak, Bk, beta, Ck);
   }
 #endif

paddle/fluid/operators/math/math_function.cu

@@ -33,9 +33,10 @@ template struct SetConstant<platform::CUDADeviceContext, int>;
 template struct SetConstant<platform::CUDADeviceContext, int64_t>;
 template struct SetConstant<platform::CUDADeviceContext, bool>;
 
-#define DEFINE_GPU_TRANS(RANK)                                         \
-  template struct Transpose<platform::CUDADeviceContext, float, RANK>; \
-  template struct Transpose<platform::CUDADeviceContext, double, RANK>;
+#define DEFINE_GPU_TRANS(RANK)                                          \
+  template struct Transpose<platform::CUDADeviceContext, float, RANK>;  \
+  template struct Transpose<platform::CUDADeviceContext, double, RANK>; \
+  template struct Transpose<platform::CUDADeviceContext, float16, RANK>;
 
 DEFINE_GPU_TRANS(1);
 DEFINE_GPU_TRANS(2);

paddle/fluid/operators/matmul_op.cc

@@ -25,7 +25,7 @@ namespace operators {
  * Get row matrix shape from a vector shape. If the rank of x_dim > 1, the
  * original x_dim is returned.
  */
-static framework::DDim RowMatrixFromVector(const framework::DDim& x_dim) {
+static framework::DDim RowMatrixFromVector(const framework::DDim &x_dim) {
   if (x_dim.size() > 1) {
     return x_dim;
   }
@@ -36,7 +36,7 @@ static framework::DDim RowMatrixFromVector(const framework::DDim& x_dim) {
  * Get column matrix shape from a vector shape. If the ran of y_dim > 1, the
  * original y_dim is returned.
  */
-static framework::DDim ColumnMatrixFromVector(const framework::DDim& y_dim) {
+static framework::DDim ColumnMatrixFromVector(const framework::DDim &y_dim) {
   if (y_dim.size() > 1) {
     return y_dim;
   }
@@ -46,12 +46,12 @@ static framework::DDim ColumnMatrixFromVector(const framework::DDim& y_dim) {
 template <typename DeviceContext, typename T>
 class MatMulKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext& context) const override {
-    auto& x =
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto &x =
         detail::Ref(context.Input<framework::Tensor>("X"), "Cannot find X");
-    auto& y =
+    auto &y =
         detail::Ref(context.Input<framework::Tensor>("Y"), "Cannot find Y");
-    auto* out = context.Output<framework::Tensor>("Out");
+    auto *out = context.Output<framework::Tensor>("Out");
     out->mutable_data<T>(context.GetPlace());
 
     auto blas = math::GetBlas<DeviceContext, T>(context);
@@ -65,7 +65,7 @@ class MatMulKernel : public framework::OpKernel<T> {
 
 // Reshape a rank-3 tensor from P x M x N to (P * M) x N.
 // Identity op if the tensor is not of rank 3.
-static framework::Tensor FoldInitDims(const framework::Tensor& input) {
+static framework::Tensor FoldInitDims(const framework::Tensor &input) {
   auto output = input;
   auto in_dims = input.dims();
   if (in_dims.size() == 3) {
@@ -78,8 +78,8 @@ static framework::Tensor FoldInitDims(const framework::Tensor& input) {
 // (Warning: This requires transposing data and writes into new memory.)
 // Identity op if the tensor is not of rank 3.
 template <typename DeviceContext, typename T>
-static framework::Tensor FoldHeadAndLastDims(const DeviceContext& context,
-                                             const framework::Tensor& input) {
+static framework::Tensor FoldHeadAndLastDims(const DeviceContext &context,
+                                             const framework::Tensor &input) {
   auto in_dims = input.dims();
   if (in_dims.size() != 3) {
     return input;
@@ -102,7 +102,7 @@ static framework::Tensor FoldHeadAndLastDims(const DeviceContext& context,
  * If transposed, `H,W` will be swapped.
  */
 static void ReshapeTensorIntoMatrixSequence(
-    framework::Tensor* x, const math::MatDescriptor& descriptor) {
+    framework::Tensor *x, const math::MatDescriptor &descriptor) {
   int64_t h, w;
   h = descriptor.height_;
   w = descriptor.width_;
@@ -130,9 +130,9 @@ static void ReshapeTensorIntoMatrixSequence(
  * If any of `X` and `Y` has batch size BatchSize, the out will have the
  * BatchSize.
  */
-static void ReshapeXYOutIntoMatrixSequence(framework::Tensor* x,
-                                           framework::Tensor* y,
-                                           framework::Tensor* out, bool trans_x,
+static void ReshapeXYOutIntoMatrixSequence(framework::Tensor *x,
+                                           framework::Tensor *y,
+                                           framework::Tensor *out, bool trans_x,
                                            bool trans_y) {
   auto x_dim = RowMatrixFromVector(x->dims());
   auto y_dim = ColumnMatrixFromVector(y->dims());
@@ -177,10 +177,10 @@ static void ReshapeXYOutIntoMatrixSequence(framework::Tensor* x,
 template <typename DeviceContext, typename T>
 class MatMulGradKernel : public framework::OpKernel<T> {
  public:
-  void MatMul(const framework::ExecutionContext& context,
-              const framework::Tensor& a, bool trans_a,
-              const framework::Tensor& b, bool trans_b,
-              framework::Tensor* out) const {
+  void MatMul(const framework::ExecutionContext &context,
+              const framework::Tensor &a, bool trans_a,
+              const framework::Tensor &b, bool trans_b,
+              framework::Tensor *out) const {
     out->mutable_data<T>(context.GetPlace());
     auto blas = math::GetBlas<DeviceContext, T>(context);
     auto mat_dim_a = math::CreateMatrixDescriptor(a.dims(), 0, trans_a);
@@ -188,18 +188,18 @@ class MatMulGradKernel : public framework::OpKernel<T> {
     blas.MatMul(a, mat_dim_a, b, mat_dim_b, T(1), out, T(0));
   }
 
-  void CalcInputGrad(const framework::ExecutionContext& context,
-                     const framework::Tensor& a, bool trans_a,
-                     bool is_fold_init_dims_a, const framework::Tensor& b,
+  void CalcInputGrad(const framework::ExecutionContext &context,
+                     const framework::Tensor &a, bool trans_a,
+                     bool is_fold_init_dims_a, const framework::Tensor &b,
                      bool trans_b, bool is_fold_init_dims_b,
-                     framework::Tensor* out) const {
+                     framework::Tensor *out) const {
     if (out == nullptr) return;
     bool need_combine = (a.dims().size() == 3 || b.dims().size() == 3) &&
                         out->dims().size() == 2;
     if (!need_combine) {
       MatMul(context, a, trans_a, b, trans_b, out);
     } else {
-      auto& ctx = context.template device_context<DeviceContext>();
+      auto &ctx = context.template device_context<DeviceContext>();
       MatMul(context, is_fold_init_dims_a
                           ? FoldInitDims(a)
                           : FoldHeadAndLastDims<DeviceContext, T>(ctx, a),
@@ -210,13 +210,13 @@ class MatMulGradKernel : public framework::OpKernel<T> {
     }
   }
 
-  void Compute(const framework::ExecutionContext& context) const override {
+  void Compute(const framework::ExecutionContext &context) const override {
     auto x = *context.Input<framework::Tensor>("X");
     auto y = *context.Input<framework::Tensor>("Y");
     auto dout =
         *context.Input<framework::Tensor>(framework::GradVarName("Out"));
-    auto* dx = context.Output<framework::Tensor>(framework::GradVarName("X"));
-    auto* dy = context.Output<framework::Tensor>(framework::GradVarName("Y"));
+    auto *dx = context.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto *dy = context.Output<framework::Tensor>(framework::GradVarName("Y"));
     bool transpose_x = context.Attr<bool>("transpose_X");
     bool transpose_y = context.Attr<bool>("transpose_Y");
 
@@ -269,7 +269,7 @@ class MatMulOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContext* context) const override {
+  void InferShape(framework::InferShapeContext *context) const override {
     PADDLE_ENFORCE(context->HasInput("X"),
                    "Input(X) of MatMulOp should not be null.");
     PADDLE_ENFORCE(context->HasInput("Y"),
@@ -322,8 +322,7 @@ class MatMulOp : public framework::OperatorWithKernel {
 
 class MatMulOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MatMulOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The first input of MatMul op");
     AddInput("Y", "The second input of MatMul op");
     AddOutput("Out", "The output of MatMul op");
@@ -376,7 +375,7 @@ class MatMulOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContext* context) const override {
+  void InferShape(framework::InferShapeContext *context) const override {
     PADDLE_ENFORCE(context->HasInput("X"), "Input(X) should not be null");
     PADDLE_ENFORCE(context->HasInput("Y"), "Input(Y) should not be null");
     PADDLE_ENFORCE(context->HasInput(framework::GradVarName("Out")),
@@ -402,7 +401,7 @@ class MatMulOpGradMaker : public framework::SingleGradOpDescMaker {
 
  protected:
   std::unique_ptr<framework::OpDesc> Apply() const override {
-    auto* retv = new framework::OpDesc();
+    auto *retv = new framework::OpDesc();
     retv->SetType("matmul_grad");
     retv->SetInput("X", Input("X"));
     retv->SetInput("Y", Input("Y"));
@@ -421,15 +420,27 @@ REGISTER_OPERATOR(matmul, ops::MatMulOp, ops::MatMulOpMaker,
                   ops::MatMulOpGradMaker);
 REGISTER_OPERATOR(matmul_grad, ops::MatMulOpGrad);
 REGISTER_OP_CPU_KERNEL(
-    matmul, ops::MatMulKernel<paddle::platform::CPUDeviceContext, float>);
+    matmul, ops::MatMulKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::MatMulKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::MatMulKernel<paddle::platform::CPUDeviceContext,
+                      paddle::platform::float16>);
 REGISTER_OP_CPU_KERNEL(
     matmul_grad,
-    ops::MatMulGradKernel<paddle::platform::CPUDeviceContext, float>);
+    ops::MatMulGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::MatMulGradKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::MatMulGradKernel<paddle::platform::CPUDeviceContext,
+                          paddle::platform::float16>);
 
 #ifdef PADDLE_WITH_CUDA
 REGISTER_OP_CUDA_KERNEL(
-    matmul, ops::MatMulKernel<paddle::platform::CUDADeviceContext, float>);
+    matmul, ops::MatMulKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::MatMulKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::MatMulKernel<paddle::platform::CUDADeviceContext,
+                      paddle::platform::float16>);
 REGISTER_OP_CUDA_KERNEL(
     matmul_grad,
-    ops::MatMulGradKernel<paddle::platform::CUDADeviceContext, float>);
+    ops::MatMulGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::MatMulGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::MatMulGradKernel<paddle::platform::CUDADeviceContext,
+                          paddle::platform::float16>);
 #endif

paddle/fluid/operators/max_sequence_len_op.cc

@@ -41,8 +41,7 @@ class MaxSeqenceLenOp : public framework::OperatorBase {
 
 class MaxSeqenceLenOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MaxSeqenceLenOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("RankTable", "The lod_rank_table.");
     AddOutput("Out", "The max sequence length.");
     AddComment(

paddle/fluid/operators/maxout_op.cc

@@ -22,8 +22,7 @@ using framework::Tensor;
 
 class MaxOutOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MaxOutOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(
         "X",
         "(Tensor) The input tensor of maxout operator. "

paddle/fluid/operators/mean_op.cc

@@ -32,8 +32,7 @@ class MeanOp : public framework::OperatorWithKernel {
 
 class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MeanOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input of mean op");
     AddOutput("Out", "The output of mean op");
     AddComment(R"DOC(

paddle/fluid/operators/merge_lod_tensor_op.cc

@@ -121,8 +121,7 @@ class MergeLoDTensorOp : public framework::OperatorBase {
 
 class MergeLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MergeLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "The input LoDTensor, contains complete lod information to "
              "construct the output");

paddle/fluid/operators/minus_op.cc

@@ -48,8 +48,7 @@ class MinusOp : public framework::OperatorWithKernel {
 
 class MinusOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MinusOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The left tensor of minus operator.");
     AddInput("Y", "The right tensor of minus operator.");
     AddOutput("Out", "The output tensor of minus operator.");

paddle/fluid/operators/modified_huber_loss_op.cc

@@ -39,8 +39,7 @@ class ModifiedHuberLossOp : public framework::OperatorWithKernel {
 
 class ModifiedHuberLossOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ModifiedHuberLossOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "The input tensor of modified huber loss op. "
              "X is 2-D tensor with shape [batch_size, 1].");

paddle/fluid/operators/momentum_op.cc

@@ -62,8 +62,7 @@ class MomentumOp : public framework::OperatorWithKernel {
 
 class MomentumOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MomentumOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Param",
              "(Tensor, default Tensor<float>) "
              "Input parameter that has to be updated");

paddle/fluid/operators/mul_op.cc

@@ -96,8 +96,7 @@ class MulOp : public framework::OperatorWithKernel {
 
 class MulOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MulOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(Tensor), The first input tensor of mul op.");
     AddInput("Y", "(Tensor), The second input tensor of mul op.");
     AddOutput("Out", "(Tensor), The output tensor of mul op.");

paddle/fluid/operators/multiplex_op.cc

@@ -61,8 +61,7 @@ class MultiplexOp : public framework::OperatorWithKernel {
 
 class MultiplexOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MultiplexOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Ids", "The index tensor of multiplex operator.");
     AddInput("X", "The candidate tensors of multiplex operator.")
         .AsDuplicable();

paddle/fluid/operators/nccl_op.cc

@@ -76,8 +76,7 @@ class NCCLInitOpShapeInference : public framework::InferShapeBase {
 
 class NCCLInitOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  NCCLInitOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(kParallelScopes, "The working place of parallel do.");
     AddOutput("Communicator",
               "Create Communicator for communicating between gpus");
@@ -118,8 +117,7 @@ class NCCLAllReduceOp : public framework::OperatorWithKernel {
 // AllReduceOp
 class NCCLAllReduceOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  NCCLAllReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input of AllReduce op");
     AddInput("Communicator", "Communicator for communicating between gpus");
     AddOutput("Out", "The output of AllReduce op");
@@ -165,8 +163,7 @@ class NCCLReduceOp : public framework::OperatorWithKernel {
 // ReduceOp
 class NCCLReduceOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  NCCLReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input of Reduce op");
     AddInput("Communicator", "Communicator for communicating between gpus");
     AddOutput("Out", "The output of Reduce op");
@@ -214,8 +211,7 @@ class NCCLBcastOp : public framework::OperatorWithKernel {
 // BcastOp
 class NCCLBcastOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  NCCLBcastOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input of BcastSend op");
     AddInput("Communicator", "Communicator for communicating between gpus");
     AddOutput("Out", "The output of Bcast");

paddle/fluid/operators/nce_op.cc

@@ -75,8 +75,7 @@ class NCEOp : public framework::OperatorWithKernel {
 
 class NCEOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  NCEOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Input", "(Tensor) A tensor of shape [batch_size, dim].");
     AddInput(
         "Label",

paddle/fluid/operators/norm_op.cc

@@ -19,8 +19,7 @@ namespace operators {
 template <typename AttrType>
 class NormOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  NormOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(
         "X",
         "(Tensor) The input tensor of norm operator. "

paddle/fluid/operators/one_hot_op.cc

@@ -46,8 +46,7 @@ class OneHotOp : public framework::OperatorWithKernel {
 
 class OneHotOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  OneHotOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LoDTensor, LoDTensor<int>) Input variable with rank at least 2. "
              "The last dimension of X should be 1. Each value of X is an index "

paddle/fluid/operators/pad_op.cc

@@ -48,8 +48,7 @@ class PadOp : public framework::OperatorWithKernel {
 
 class PadOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  PadOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "The input of pad op. "
              "The input should be a k-D tensor(k > 0 and k < 7)");

paddle/fluid/operators/parallel_do_op.cc

@@ -196,8 +196,7 @@ class ParallelDoOp : public framework::OperatorBase {
 
 class ParallelDoOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ParallelDoOpProtoMaker(OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(kInputs, "").AsDuplicable();
     AddInput(kParameters, "").AsDuplicable();
     AddInput(kPlaces, "");

paddle/fluid/operators/pool_op.cc

@@ -135,8 +135,7 @@ framework::OpKernelType PoolOpGrad::GetExpectedKernelType(
                                  library_);
 }
 
-Pool2dOpMaker::Pool2dOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-    : OpProtoAndCheckerMaker(proto, op_checker) {
+void Pool2dOpMaker::Make() {
   AddInput(
       "X",
       "(Tensor) The input tensor of pooling operator. "
@@ -229,8 +228,7 @@ Example:
 )DOC");
 }
 
-Pool3dOpMaker::Pool3dOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-    : OpProtoAndCheckerMaker(proto, op_checker) {
+void Pool3dOpMaker::Make() {
   AddInput("X",
            "(Tensor) The input tensor of pooling operator. "
            "The format of input tensor is NCDHW, where N is batch size, C is "

paddle/fluid/operators/pool_op.h

@@ -50,12 +50,12 @@ class PoolOpGrad : public framework::OperatorWithKernel {
 
 class Pool2dOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  Pool2dOpMaker(OpProto* proto, OpAttrChecker* op_checker);
+  void Make() override;
 };
 
 class Pool3dOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  Pool3dOpMaker(OpProto* proto, OpAttrChecker* op_checker);
+  void Make() override;
 };
 
 template <typename DeviceContext, typename T>

paddle/fluid/operators/pool_with_index_op.cc

@@ -100,8 +100,7 @@ class MaxPoolWithIndexOpGrad : public framework::OperatorWithKernel {
 
 class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MaxPool2dWithIndexOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(
         "X",
         "(Tensor) The input tensor of pooling operator. "
@@ -177,8 +176,7 @@ Example:
 
 class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MaxPool3dWithIndexOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor) The input tensor of pooling operator. "
              "The format of input tensor is NCDHW, where N is batch size, C is "

paddle/fluid/operators/positive_negative_pair_op.cc

@@ -95,8 +95,7 @@ class PositiveNegativePairOp : public framework::OperatorWithKernel {
 
 class PositiveNegativePairOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  PositiveNegativePairOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Score",
              "(Tensor, float) Model Score on an item (with "
              "respect to QueryID). It's a 2-D tensor with shape [batch_size, "

paddle/fluid/operators/precision_recall_op.cc

@@ -90,8 +90,7 @@ class PrecisionRecallOp : public framework::OperatorWithKernel {
 
 class PrecisionRecallOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  PrecisionRecallOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("MaxProbs",
              "(Tensor, default Tensor<float>) A 2-D tensor with shape N x 1, "
              "where N is the batch size. Each row contains the max probability "

paddle/fluid/operators/prefetch_op.cc

@@ -64,8 +64,7 @@ class PrefetchOp : public framework::OperatorBase {
 
 class PrefetchOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  PrefetchOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("X", "(LoDTensor) Input Id variables to be sent").AsDuplicable();
     AddOutput("RPCClient",
               "(RPCClient) The RPC client object which will be"

paddle/fluid/operators/prelu_op.cc

@@ -38,8 +38,7 @@ class PReluOp : public framework::OperatorWithKernel {
 
 class PReluOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  PReluOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input tensor of prelu operator.");
     AddInput("Alpha", "The alpha weight of prelu operator.");
     AddOutput("Out", "The output tensor of prelu operator.");

paddle/fluid/operators/print_op.cc

@@ -209,8 +209,7 @@ class TensorPrintOp : public framework::OperatorBase {
 
 class PrintOpProtoAndCheckMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  PrintOpProtoAndCheckMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("In", "Input tensor to be displayed.");
     AddAttr<int>("first_n", "Only log `first_n` number of times.");
     AddAttr<std::string>("message", "A string message to print as a prefix.");

paddle/fluid/operators/proximal_adagrad_op.cc

@@ -66,8 +66,7 @@ class ProximalAdagradOp : public framework::OperatorWithKernel {
 
 class ProximalAdagradOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ProximalAdagradOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Param",
              "(Tensor, default Tensor<float>) "
              "Input parameter that has to be updated.");

paddle/fluid/operators/proximal_gd_op.cc

@@ -54,8 +54,7 @@ class ProximalGDOp : public framework::OperatorWithKernel {
 
 class ProximalGDOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ProximalGDOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Param",
              "(Tensor, default Tensor<float>) "
              "Input parameter value that has to be updated.");

paddle/fluid/operators/rank_loss_op.cc

@@ -46,8 +46,7 @@ class RankLossOp : public framework::OperatorWithKernel {
 
 class RankLossOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  RankLossOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Label",
              "(2-D Tensor with shape [batch_size x 1]) "
              "The label indicating A ranked higher than B or not.");

paddle/fluid/operators/read_op.cc

@@ -79,8 +79,7 @@ class ReadOp : public framework::OperatorBase {
 
 class ReadOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ReadOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(op_proto, op_checker) {
+  void Make() override {
     AddInput("Reader", "(ReaderHolder) The executed reader.");
     AddOutput("Out", "(LoDTensor) The output data.").AsDuplicable();
     AddComment(R"DOC(

paddle/fluid/operators/reader/create_batch_reader_op.cc

@@ -52,9 +52,8 @@ class CreateBatchReaderOp : public framework::OperatorBase {
 };
 
 class CreateBatchReaderOpMaker : public DecoratedReaderMakerBase {
- public:
-  CreateBatchReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
-      : DecoratedReaderMakerBase(op_proto, op_checker) {
+ protected:
+  void Apply() override {
     AddAttr<int>("batch_size",
                  "How many instances the batch reader yields each time.")
         .GreaterThan(0);

paddle/fluid/operators/reader/create_double_buffer_reader_op.cc

@@ -113,14 +113,13 @@ class CreateDoubleBufferReaderOp : public framework::OperatorBase {
 };
 
 class CreateDoubleBufferReaderOpMaker : public DecoratedReaderMakerBase {
- public:
-  CreateDoubleBufferReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
-      : DecoratedReaderMakerBase(op_proto, op_checker) {
+ protected:
+  void Apply() override {
     AddComment(R"DOC(
       CreateDoubleBufferReader Operator
 
       A double buffer reader takes another reader as its 'underlying reader'.
-      It launches another thread to execute the 'underlying reader' asynchronously, 
+      It launches another thread to execute the 'underlying reader' asynchronously,
       which prevents reading process from blocking subsequent training.
     )DOC");
     std::unordered_set<std::string> enum_range;

paddle/fluid/operators/reader/create_multi_pass_reader_op.cc

@@ -65,20 +65,19 @@ class CreateMultiPassReaderOp : public framework::OperatorBase {
 };
 
 class CreateMultiPassReaderOpMaker : public DecoratedReaderMakerBase {
- public:
-  CreateMultiPassReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
-      : DecoratedReaderMakerBase(op_proto, op_checker) {
+ protected:
+  void Apply() override {
     AddAttr<int>("pass_num", "The number of pass to run.").GreaterThan(0);
     AddComment(R"DOC(
       CreateMultiPassReader Operator
 
-      This operator creates a multi-pass reader. A multi-pass reader 
-      is used to yield data for several pass training continuously. 
+      This operator creates a multi-pass reader. A multi-pass reader
+      is used to yield data for several pass training continuously.
       It takes the number of passes to run as one of its attributes
-      ('pass_num'), and maintains a pass counter to record how many 
-      passes it has completed. When the underlying reader reaches the 
-      EOF, the multi-pass reader checks whether it has completed training 
-      of the given number of pass. If not, the underlying reader will 
+      ('pass_num'), and maintains a pass counter to record how many
+      passes it has completed. When the underlying reader reaches the
+      EOF, the multi-pass reader checks whether it has completed training
+      of the given number of pass. If not, the underlying reader will
       be re-initialized and starts a new pass automatically.
     )DOC");
   }

paddle/fluid/operators/reader/create_random_data_generator_op.cc

@@ -84,9 +84,8 @@ class CreateRandomDataGeneratorOp : public framework::OperatorBase {
 };
 
 class CreateRandomDataGeneratorOpMaker : public FileReaderMakerBase {
- public:
-  CreateRandomDataGeneratorOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
-      : FileReaderMakerBase(op_proto, op_checker) {
+ protected:
+  void Apply() override {
     AddAttr<float>("min", "The lower bound of reader's uniform distribution.");
     AddAttr<float>("max", "The upper bound of reader's uniform distribution.");
     AddComment(R"DOC(

paddle/fluid/operators/reader/create_recordio_file_reader_op.cc

@@ -76,9 +76,8 @@ class CreateRecordIOReaderOp : public framework::OperatorBase {
 };
 
 class CreateRecordIOReaderOpMaker : public FileReaderMakerBase {
- public:
-  CreateRecordIOReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
-      : FileReaderMakerBase(op_proto, op_checker) {
+ protected:
+  void Apply() override {
     AddAttr<std::string>("filename", "The filename of record io reader");
     AddComment(R"DOC(
       CreateRecordIOReader Operator

paddle/fluid/operators/reader/create_shuffle_reader_op.cc

@@ -92,9 +92,8 @@ class CreateShuffleReaderOp : public framework::OperatorBase {
 };
 
 class CreateShuffleReaderOpMaker : public DecoratedReaderMakerBase {
- public:
-  CreateShuffleReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
-      : DecoratedReaderMakerBase(op_proto, op_checker) {
+ protected:
+  void Apply() override {
     AddAttr<int>("buffer_size", "The shuffle buffer size.").GreaterThan(0);
     AddComment(R"DOC(
       CreateShuffleReader Operator

paddle/fluid/operators/reader/create_threaded_reader_op.cc

@@ -53,17 +53,16 @@ class CreateThreadedReaderOp : public framework::OperatorBase {
 };
 
 class CreateThreadedReaderOpMaker : public DecoratedReaderMakerBase {
- public:
-  CreateThreadedReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
-      : DecoratedReaderMakerBase(op_proto, op_checker) {
+ protected:
+  void Apply() override {
     AddComment(R"DOC(
       CreateThreadedReader Operator
 
-      This operator creates a threaded reader. A threaded reader's 
-      'ReadNext()' can be invoked by several threads at the same 
-      time. 
-      When the attribute 'safe_mode' is true, the threaded reader's 
-      'ReInit()' is disabled to avoid unexpected bugs in multi-thread 
+      This operator creates a threaded reader. A threaded reader's
+      'ReadNext()' can be invoked by several threads at the same
+      time.
+      When the attribute 'safe_mode' is true, the threaded reader's
+      'ReInit()' is disabled to avoid unexpected bugs in multi-thread
       environment.
     )DOC");
   }

paddle/fluid/operators/reader/open_files_op.cc

@@ -185,9 +185,8 @@ class OpenFilesOp : public framework::OperatorBase {
 };
 
 class OpenFilesOpMaker : public FileReaderMakerBase {
- public:
-  OpenFilesOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
-      : FileReaderMakerBase(op_proto, op_checker) {
+ protected:
+  void Apply() override {
     AddAttr<std::vector<std::string>>("file_names", "Files to be read.");
     AddAttr<int>("thread_num", "The maximal concurrent prefetch thread number.")
         .GreaterThan(0);
@@ -196,7 +195,7 @@ class OpenFilesOpMaker : public FileReaderMakerBase {
     AddComment(R"DOC(
       OpenFiles Operator
 
-      An OpenFilesOp creates a MultiFileReader, which is able to 
+      An OpenFilesOp creates a MultiFileReader, which is able to
       read data multi-threaded from multiple files.
     )DOC");
   }

paddle/fluid/operators/reader/reader_op_registry.cc

@@ -53,10 +53,7 @@ std::unique_ptr<framework::ReaderBase> CreateReaderByFileName(
   return std::unique_ptr<framework::ReaderBase>(reader);
 }
 
-FileReaderMakerBase::FileReaderMakerBase(
-    framework::OpProtoAndCheckerMaker::OpProto* op_proto,
-    framework::OpAttrChecker* op_checker)
-    : OpProtoAndCheckerMaker(op_proto, op_checker) {
+void FileReaderMakerBase::Make() {
   AddOutput("Out", "(ReaderHolder) The created random reader.").AsDuplicable();
   AddAttr<std::vector<int>>("shape_concat", "The concat of all data's shapes.");
   AddAttr<std::vector<int>>(
@@ -68,6 +65,7 @@ FileReaderMakerBase::FileReaderMakerBase(
       "It means the reader will generate two data each time,"
       "whose shapes are [2,3,4] and [5,6] respectively.");
   AddAttr<std::vector<int>>("lod_levels", "The LoD levels of each data.");
+  Apply();
 }
 
 void FileReaderInferShape::operator()(framework::InferShapeContext* ctx) const {
@@ -127,13 +125,11 @@ void DecoratedReaderInferVarType::operator()(
   out_reader->SetDataTypes(in_reader->GetDataTypes());
 }
 
-DecoratedReaderMakerBase::DecoratedReaderMakerBase(
-    framework::OpProtoAndCheckerMaker::OpProto* op_proto,
-    framework::OpAttrChecker* op_checker)
-    : OpProtoAndCheckerMaker(op_proto, op_checker) {
+void DecoratedReaderMakerBase::Make() {
   AddInput("UnderlyingReader",
            "(ReaderHolder) The underlying reader for creating a batch reader.");
   AddOutput("Out", "(ReaderHolder) The created batch reader.");
+  Apply();
 }
 
 }  // namespace reader

paddle/fluid/operators/reader/reader_op_registry.h

@@ -47,7 +47,10 @@ extern std::vector<framework::DDim> RestoreShapes(
 
 class FileReaderMakerBase : public framework::OpProtoAndCheckerMaker {
  public:
-  FileReaderMakerBase(OpProto* op_proto, OpAttrChecker* op_checker);
+  void Make() final;
+
+ protected:
+  virtual void Apply() = 0;
 };
 
 class FileReaderInferShape : public framework::InferShapeBase {
@@ -76,7 +79,10 @@ class DecoratedReaderInferVarType : public framework::VarTypeInference {
 
 class DecoratedReaderMakerBase : public framework::OpProtoAndCheckerMaker {
  public:
-  DecoratedReaderMakerBase(OpProto* op_proto, OpAttrChecker* op_checker);
+  void Make() final;
+
+ protected:
+  virtual void Apply() = 0;
 };
 
 }  // namespace reader

paddle/fluid/operators/recurrent_op.cc

@@ -508,8 +508,7 @@ class RecurrentGradOp : public RecurrentBase {
 
 class RecurrentOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  RecurrentOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(kInputs, "rnn inputs").AsDuplicable();
     AddInput(kInitialStates, "rnn initial states").AsDuplicable();
     AddInput(kParameters,

paddle/fluid/operators/recv_op.cc

@@ -53,8 +53,7 @@ class RecvOp : public framework::OperatorBase {
 
 class RecvOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  RecvOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddOutput("Out", "(Tensor) Variables to get from server.").AsDuplicable();
     AddComment(R"DOC(
 Recv operator

paddle/fluid/operators/reduce_op.cc

@@ -90,8 +90,7 @@ class ReduceGradOp : public framework::OperatorWithKernel {
 
 class ReduceOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() final {
     AddInput("X",
              "(Tensor) The input tensor. Tensors with rank at most 6 are "
              "supported.");
@@ -111,78 +110,20 @@ class ReduceOpMaker : public framework::OpProtoAndCheckerMaker {
                   "(bool, default false) "
                   "If true, output a scalar reduced along all dimensions.")
         .SetDefault(false);
-    comment_ = R"DOC(
-{ReduceOp} Operator.
+    AddComment(string::Sprintf(R"DOC(
+%s Operator.
 
-This operator computes the {reduce} of input tensor along the given dimension. 
+This operator computes the %s of input tensor along the given dimension.
 The result tensor has 1 fewer dimension than the input unless keep_dim is true.
 If reduce_all is true, just reduce along all dimensions and output a scalar.
 
-)DOC";
-    AddComment(comment_);
+)DOC",
+                               GetOpType(), GetName()));
   }
 
  protected:
-  std::string comment_;
-
-  void Replace(std::string *src, std::string from, std::string to) {
-    std::size_t len_from = std::strlen(from.c_str());
-    std::size_t len_to = std::strlen(to.c_str());
-    for (std::size_t pos = src->find(from); pos != std::string::npos;
-         pos = src->find(from, pos + len_to)) {
-      src->replace(pos, len_from, to);
-    }
-  }
-
-  void SetComment(std::string name, std::string op) {
-    Replace(&comment_, "{ReduceOp}", name);
-    Replace(&comment_, "{reduce}", op);
-  }
-};
-
-class ReduceSumOpMaker : public ReduceOpMaker {
- public:
-  ReduceSumOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : ReduceOpMaker(proto, op_checker) {
-    SetComment("ReduceSum", "sum");
-    AddComment(comment_);
-  }
-};
-
-class ReduceMeanOpMaker : public ReduceOpMaker {
- public:
-  ReduceMeanOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : ReduceOpMaker(proto, op_checker) {
-    SetComment("ReduceMean", "mean");
-    AddComment(comment_);
-  }
-};
-
-class ReduceMaxOpMaker : public ReduceOpMaker {
- public:
-  ReduceMaxOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : ReduceOpMaker(proto, op_checker) {
-    SetComment("ReduceMax", "max");
-    AddComment(comment_);
-  }
-};
-
-class ReduceMinOpMaker : public ReduceOpMaker {
- public:
-  ReduceMinOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : ReduceOpMaker(proto, op_checker) {
-    SetComment("ReduceMin", "min");
-    AddComment(comment_);
-  }
-};
-
-class ReduceProdOpMaker : public ReduceOpMaker {
- public:
-  ReduceProdOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : ReduceOpMaker(proto, op_checker) {
-    SetComment("ReduceProd", "production");
-    AddComment(comment_);
-  }
+  virtual std::string GetName() const = 0;
+  virtual std::string GetOpType() const = 0;
 };
 
 }  // namespace operators
@@ -190,25 +131,21 @@ class ReduceProdOpMaker : public ReduceOpMaker {
 
 namespace ops = paddle::operators;
 
-REGISTER_OPERATOR(reduce_sum, ops::ReduceOp, ops::ReduceSumOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(reduce_sum_grad, ops::ReduceGradOp);
-
-REGISTER_OPERATOR(reduce_mean, ops::ReduceOp, ops::ReduceMeanOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(reduce_mean_grad, ops::ReduceGradOp);
-
-REGISTER_OPERATOR(reduce_max, ops::ReduceOp, ops::ReduceMaxOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(reduce_max_grad, ops::ReduceGradOp);
-
-REGISTER_OPERATOR(reduce_min, ops::ReduceOp, ops::ReduceMinOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(reduce_min_grad, ops::ReduceGradOp);
-
-REGISTER_OPERATOR(reduce_prod, ops::ReduceOp, ops::ReduceProdOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
-REGISTER_OPERATOR(reduce_prod_grad, ops::ReduceGradOp);
+#define REGISTER_REDUCE_OP(op_name)                                        \
+  class __##op_name##Maker__ : public ops::ReduceOpMaker {                 \
+   protected:                                                              \
+    virtual std::string GetName() const { return #op_name; }               \
+    virtual std::string GetOpType() const { return "Reduce " #op_name; }   \
+  };                                                                       \
+  REGISTER_OPERATOR(reduce_##op_name, ops::ReduceOp, __##op_name##Maker__, \
+                    paddle::framework::DefaultGradOpDescMaker<true>);      \
+  REGISTER_OPERATOR(reduce_##op_name##_grad, ops::ReduceGradOp)
+
+REGISTER_REDUCE_OP(sum);
+REGISTER_REDUCE_OP(mean);
+REGISTER_REDUCE_OP(max);
+REGISTER_REDUCE_OP(min);
+REGISTER_REDUCE_OP(prod);
 
 #define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor)         \
   REGISTER_OP_CPU_KERNEL(reduce_type,                                          \

paddle/fluid/operators/reorder_lod_tensor_by_rank_op.cc

@@ -23,9 +23,7 @@ namespace operators {
 class ReorderLoDTensorByRankTableOpProtoMaker
     : public framework::OpProtoAndCheckerMaker {
  public:
-  ReorderLoDTensorByRankTableOpProtoMaker(OpProto *proto,
-                                          OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LoDTensor), the input lod tensor to be reordered according to "
              "Input(RankTable).");

paddle/fluid/operators/reshape_op.cc

@@ -22,8 +22,7 @@ namespace operators {
 
 class ReshapeOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ReshapeOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(Tensor). The input tensor of reshape operator.");
     AddInput("Shape",
              "(Tensor<int32>, optional). If provided, reshape according to "

paddle/fluid/operators/rmsprop_op.cc

@@ -63,8 +63,7 @@ class RmspropOp : public framework::OperatorWithKernel {
 
 class RmspropOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  RmspropOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Param",
              "(Tensor, default Tensor<float>) "
              "Input parameter value that has to be updated.");

paddle/fluid/operators/rnn_memory_helper_op.cc

@@ -59,8 +59,7 @@ class RNNMemoryHelperOpShapeInference : public framework::InferShapeBase {
 
 class RNNMemoryHelperOpInfoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  RNNMemoryHelperOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "");
     AddOutput("Out", "");
     AddAttr<int>("dtype",
@@ -117,8 +116,7 @@ class RNNMemoryHelperGradOp : public framework::OperatorBase {
 class RNNMemoryHelperGradOpInfoMaker
     : public framework::OpProtoAndCheckerMaker {
  public:
-  RNNMemoryHelperGradOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(framework::GradVarName("Out"), "");
     AddInput("X", "");
     AddInput("Out", "");

paddle/fluid/operators/roi_pool_op.cc

@@ -98,8 +98,7 @@ class ROIPoolGradOp : public framework::OperatorWithKernel {
 
 class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ROIPoolOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor), "
              "the input of ROIPoolOp. "

paddle/fluid/operators/row_conv_op.cc

@@ -76,8 +76,7 @@ class RowConvGradOp : public framework::OperatorWithKernel {
 
 class RowConvOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  RowConvOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LoDTensor), the input(X) is a LodTensor, which supports "
              "variable time-length input sequences. The underlying tensor "

paddle/fluid/operators/save_combine_op.cc

@@ -127,8 +127,7 @@ class SaveCombineOp : public framework::OperatorBase {
 
 class SaveCombineOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SaveCombineOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(
         "X",
         "(vector) Input LoDTensors that need to be saved together in a file.")

paddle/fluid/operators/save_op.cc

@@ -117,8 +117,7 @@ class SaveOp : public framework::OperatorBase {
 
 class SaveOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SaveOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(Tensor ) Input tensor to be saved");
     AddComment(R"DOC(
 Save operator

paddle/fluid/operators/scale_op.cc

@@ -37,8 +37,7 @@ class ScaleOp : public framework::OperatorWithKernel {
 
 class ScaleOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ScaleOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(Tensor) Input tensor of scale operator.");
     AddOutput("Out", "(Tensor) Output tensor of scale operator.");
     AddComment(R"DOC(

paddle/fluid/operators/scatter_op.cc

@@ -78,8 +78,7 @@ class ScatterGradOp : public framework::OperatorWithKernel {
 
 class ScatterOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ScatterOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The source input of scatter op");
     AddInput("Ids", "The index input of scatter op where X will be updated");
     AddInput("Updates", "The updated value of updates op");

paddle/fluid/operators/select_op.cc

@@ -380,8 +380,7 @@ class SelectOp : public framework::OperatorBase {
 
 class SelectOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SelectOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(kX,
              "A set of variables, which are required by operators inside the "
              "cases of Select Op")

paddle/fluid/operators/send_barrier_op.cc

@@ -57,8 +57,7 @@ class SendBarrierOp : public framework::OperatorBase {
 
 class SendBarrierOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SendBarrierOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddOutput("RPCClient",
               "(RPCClient) The RPC client object which is"
               "initialized at most once.");

paddle/fluid/operators/send_op.cc

@@ -92,8 +92,7 @@ class SendOp : public framework::OperatorBase {
 
 class SendOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SendOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("X", "(Tensor) Input tensor to be sent").AsDuplicable();
     AddOutput("Out", "(Tensor) Output tensor to be received from server")
         .AsDuplicable();

paddle/fluid/operators/send_vars_op.cc

@@ -66,8 +66,7 @@ class SendVarsOp : public framework::OperatorBase {
 
 class SendVarsOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SendVarsOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("X", "(Tensor, SelectedRows) Input variables to be sent")
         .AsDuplicable();
     AddOutput("RPCClient",

paddle/fluid/operators/sequence_concat_op.cc

@@ -43,8 +43,7 @@ class SequenceConcatOp : public framework::OperatorWithKernel {
 
 class SequenceConcatOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SequenceConcatOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LodTensorArray) Input is a vector of LoDTensor, "
              "each of which is a variable-length sequence or nested sequence.")

paddle/fluid/operators/sequence_conv_op.cc

@@ -102,8 +102,7 @@ class SequenceConvGradOp : public framework::OperatorWithKernel {
 
 class SequenceConvOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SequenceConvOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(
         "X",
         "(LoDTensor) the input(X) is a LodTensor, which supports "

paddle/fluid/operators/sequence_erase_op.cc

@@ -37,8 +37,7 @@ class SequenceEraseOp : public framework::OperatorWithKernel {
 
 class SequenceEraseOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SequenceEraseOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(2-D LoDTensor with the 2nd dim. equal to 1) "
              "Input LoDTensor of SequenceEraseOp.");

paddle/fluid/operators/sequence_expand_op.cc

@@ -94,8 +94,7 @@ class SequenceExpandOp : public framework::OperatorWithKernel {
 
 class SequenceExpandOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SequenceExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LoDTensor, default LoDTensor<float>) A 2-D LoDTensor whose lod "
              "level is at most 1.");

paddle/fluid/operators/sequence_pool_op.cc

@@ -38,8 +38,7 @@ class SequencePoolOp : public framework::OperatorWithKernel {
 
 class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SequencePoolOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(LoDTensor) The variable-length input of SequencePoolOp");
     AddOutput("Out",
               "(Tensor) The output of SequencePoolOp does not contain LoD "

paddle/fluid/operators/sequence_reshape_op.cc

@@ -42,8 +42,7 @@ class SequenceReshapeOp : public framework::OperatorWithKernel {
 
 class SequenceReshapeOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SequenceReshapeOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LoDTensor, default LoDTensor<float>) A 2-D LoDTensor with shape "
              "being [N, M].");

paddle/fluid/operators/sequence_slice_op.cc

@@ -79,8 +79,7 @@ class SequenceSliceGradOp : public framework::OperatorWithKernel {
 
 class SequenceSliceOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SequenceSliceOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LoDTensor), "
              "the input of SequenceSliceOp.");

paddle/fluid/operators/sequence_softmax_op.cc

@@ -57,8 +57,7 @@ class SequenceSoftmaxOp : public framework::OperatorWithKernel {
 
 class SequenceSoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SequenceSoftmaxOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(LoDTensor) 1-D or 2-D input LoDTensor with the 2-nd dimension "
              "of length 1.");

paddle/fluid/operators/sgd_op.cc

@@ -68,8 +68,7 @@ class SGDOpInferVarType : public framework::VarTypeInference {
 
 class SGDOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SGDOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Param", "(Tensor or SelectedRows) Input parameter");
     AddInput("LearningRate", "(Tensor) Learning rate of SGD");
     AddInput("Grad", "(Tensor or SelectedRows) Input gradient");

paddle/fluid/operators/shrink_rnn_memory_op.cc

@@ -69,8 +69,7 @@ class ShrinkRNNMemoryOp : public ArrayOp {
 
 class ShrinkRNNMemoryOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ShrinkRNNMemoryOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(LoDTensor) The RNN step memory to be shrinked.");
     AddInput("RankTable", "(LoDRankTable) The lod_rank_table of dynamic RNN.");
     AddInput("I",

paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc

@@ -86,9 +86,7 @@ class SigmoidCrossEntropyWithLogitsGradOp
 class SigmoidCrossEntropyWithLogitsOpMaker
     : public framework::OpProtoAndCheckerMaker {
  public:
-  SigmoidCrossEntropyWithLogitsOpMaker(OpProto* proto,
-                                       OpAttrChecker* op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor, default Tensor<float>), a 2-D tensor with shape N x D, "
              "where N is the batch size and D is the number of classes. "

paddle/fluid/operators/sign_op.cc

@@ -34,8 +34,7 @@ class SignOp : public framework::OperatorWithKernel {
 template <typename AttrType>
 class SignOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SignOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(Tensor) Input tensor of sign operator.");
     AddOutput("Out", "(Tensor) Output tensor of sign operator.");
     AddComment(R"DOC(

paddle/fluid/operators/smooth_l1_loss_op.cc

@@ -46,8 +46,7 @@ class SmoothL1LossOp : public framework::OperatorWithKernel {
 
 class SmoothL1LossOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SmoothL1LossOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "(Tensor, default Tensor<float>) A tensor with rank at least 2. "
              "The input value of smooth l1 loss op with shape "

paddle/fluid/operators/softmax_mkldnn_op.cc

@@ -53,25 +53,60 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
                    "Softmax input and output dimensions should match");
     // Same memory descriptor to be used for input and output
     memory::dims softmax_tz = {src_tz[0], src_tz[1]};
-    // Currently only supports NC data format
-    // TODO(jczaja-intel): support more formats
-    auto softmax_md =
-        MKLDNNMemDesc({softmax_tz}, memory::f32, memory::format::nc);
-    // Normalization is made after innermost dimension eg. C out of NC
-    auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring,
-                                              softmax_md, 1 /*dim: C*/);
-    // create memory primitives
-    auto softmax_src_memory =
-        memory({softmax_md, mkldnn_engine},
-               static_cast<void*>(const_cast<T*>(input_data)));
-    auto softmax_dst_memory =
-        memory({softmax_md, mkldnn_engine},
-               static_cast<void*>(const_cast<T*>(output_data)));
-    auto softmax_prim_desc =
-        softmax_forward::primitive_desc(softmax_desc, mkldnn_engine);
-    auto softmax = softmax_forward(softmax_prim_desc, softmax_src_memory,
-                                   softmax_dst_memory);
-    std::vector<primitive> pipeline{softmax};
+    // Generate keys for storing/retriving primitives for this operator
+    // TODO(jczaja): Each MKLDNN operator may have diffrent hashing function
+    auto gethash = [](memory::dims& operand_dims) {
+      return std::string(std::to_string(operand_dims[0]) + "-" +
+                         std::to_string(operand_dims[1]));
+    };
+    const std::string key = gethash(softmax_tz);
+    const std::string key_softmax_p = key + "@softmax_p";
+    const std::string key_softmax_src_mem_p = key + "@softmax_src_mem_p";
+    const std::string key_softmax_dst_mem_p = key + "@softmax_dst_mem_p";
+
+    std::shared_ptr<void> softmax_p = dev_ctx.GetBlob(key_softmax_p);
+    if (softmax_p == nullptr) {
+      // Currently only NC data format is supported
+      auto softmax_md =
+          MKLDNNMemDesc({softmax_tz}, memory::f32, memory::format::nc);
+      // Normalization is made after innermost dimension eg. C out of NC
+      auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring,
+                                                softmax_md, 1 /*dim: C*/);
+      // create memory primitives
+      auto softmax_src_memory_p = std::make_shared<memory>(
+          memory::primitive_desc{softmax_md, mkldnn_engine},
+          static_cast<void*>(const_cast<T*>(input_data)));
+      dev_ctx.SetBlob(key_softmax_src_mem_p, softmax_src_memory_p);
+      auto softmax_dst_memory_p = std::make_shared<memory>(
+          memory::primitive_desc{softmax_md, mkldnn_engine},
+          static_cast<void*>(output_data));
+      dev_ctx.SetBlob(key_softmax_dst_mem_p, softmax_dst_memory_p);
+
+      auto softmax_forward_pd =
+          std::make_shared<softmax_forward::primitive_desc>(softmax_desc,
+                                                            mkldnn_engine);
+      softmax_p = std::make_shared<softmax_forward>(
+          *(softmax_forward_pd.get()),
+          *(static_cast<memory*>(softmax_src_memory_p.get())),
+          *(static_cast<memory*>(softmax_dst_memory_p.get())));
+      dev_ctx.SetBlob(key_softmax_p, softmax_p);
+    } else {
+      // Primitives already exist
+      auto src_memory_p = std::static_pointer_cast<memory>(
+          dev_ctx.GetBlob(key_softmax_src_mem_p));
+      PADDLE_ENFORCE(src_memory_p != nullptr,
+                     "Fail to find softmax src mem_p in device context");
+      auto dst_memory_p = std::static_pointer_cast<memory>(
+          dev_ctx.GetBlob(key_softmax_dst_mem_p));
+      PADDLE_ENFORCE(dst_memory_p != nullptr,
+                     "Fail to find softmax dst mem_p in device context");
+      src_memory_p->set_data_handle(
+          reinterpret_cast<void*>(const_cast<T*>(input_data)));
+      dst_memory_p->set_data_handle(output_data);
+    }
+
+    std::vector<primitive> pipeline{
+        *(static_cast<softmax_forward::primitive*>(softmax_p.get()))};
     stream(stream::kind::eager).submit(pipeline).wait();
 
     const bool is_test = ctx.Attr<bool>("is_test");

paddle/fluid/operators/softmax_op.cc

@@ -77,8 +77,7 @@ class SoftmaxOp : public framework::OperatorWithKernel {
 
 class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SoftmaxOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X",
              "The input tensor of softmax. "
              "2-D with shape [batch_size, input_feature_dimensions].");

paddle/fluid/operators/softmax_with_cross_entropy_op.cc

@@ -20,8 +20,7 @@ namespace operators {
 class SoftmaxWithCrossEntropyOpMaker
     : public framework::OpProtoAndCheckerMaker {
  public:
-  SoftmaxWithCrossEntropyOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Logits",
              "(Tensor, default: Tensor<float>), The unscaled log probabilities "
              "which is a 2-D tensor with shape [N x K]. N is the batch_size, "

paddle/fluid/operators/split_byref_op.cc

@@ -64,8 +64,7 @@ class SplitByrefOp : public framework::OperatorWithKernel {
 
 class SplitByrefOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SplitByrefOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(Tensor) Input tensor of the split operator.");
     AddOutput("Out", "(Tensor) Output tensors of the split operator.")
         .AsDuplicable();

paddle/fluid/operators/split_ids_op.cc

@@ -19,8 +19,7 @@ namespace operators {
 
 class SplitIdsOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SplitIdsOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}");
     AddOutput("Out", "(LoDTensor) The outputs of the input Ids.")
         .AsDuplicable();

paddle/fluid/operators/split_lod_tensor_op.cc

@@ -125,8 +125,7 @@ class SplitLoDTensorOp : public framework::OperatorBase {
 
 class SplitLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SplitLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input LoDTensor");
     AddInput("Mask", "A bool column vector which mask the input");
     AddOutput("OutTrue", "True branch of input LoDTensor");

paddle/fluid/operators/split_op.cc

@@ -70,8 +70,7 @@ class SplitOp : public framework::OperatorWithKernel {
 
 class SplitOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SplitOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(Tensor) Input tensor of the split operator.");
     AddOutput("Out", "(Tensor) Output tensors of the split operator.")
         .AsDuplicable();

paddle/fluid/operators/split_selected_rows_op.cc

@@ -19,8 +19,7 @@ namespace operators {
 
 class SplitSelectedRowsOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SplitSelectedRowsOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "The input SelectedRows.");
     AddOutput("Out", "The outputs of the input SelectedRows.").AsDuplicable();
     AddAttr<std::vector<int>>("height_sections",

paddle/fluid/operators/spp_op.cc

@@ -20,8 +20,7 @@ namespace operators {
 
 class SppOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SppOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(
         "X",
         "(Tensor) The input tensor of spp operator. "

paddle/fluid/operators/squared_l2_distance_op.cc

@@ -56,8 +56,7 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
 
 class SquaredL2DistanceOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SquaredL2DistanceOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(Tensor) Input of SquaredL2DistanceOp.");
     AddInput("Y", "(Tensor) Target of SquaredL2DistanceOp.");
     AddOutput("sub_result",

paddle/fluid/operators/squared_l2_norm_op.cc

@@ -48,8 +48,7 @@ class SquaredL2NormGradOp : public framework::OperatorWithKernel {
 
 class SquaredL2NormOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SquaredL2NormOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(Tensor) The input of squared_l2_norm op.");
     AddOutput("Out", "(Scalar) The output of squared_l2_norm op.");
     AddComment(R"DOC(

paddle/fluid/operators/sum_op.cc

@@ -112,8 +112,7 @@ class SumOp : public framework::OperatorWithKernel {
 
 class SumOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SumOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(vector<Tensor>) The input tensors of sum operator.")
         .AsDuplicable();
     AddOutput("Out", "(Tensor) The output tensor of sum operator.");

paddle/fluid/operators/tensor_array_read_write_op.cc

@@ -57,8 +57,7 @@ class WriteToArrayOp : public ArrayOp {
 
 class WriteToArrayOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  WriteToArrayOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(LoDTensor) the tensor will be written to tensor array");
     AddInput(
         "I",
@@ -148,8 +147,7 @@ class ReadFromArrayOp : public ArrayOp {
 
 class ReadFromArrayProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ReadFromArrayProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(TensorArray) the array will be read from.");
     AddInput("I",
              "(Tensor) the subscript index in tensor array. The number of "

paddle/fluid/operators/top_k_op.cc

@@ -48,8 +48,7 @@ class TopkOp : public framework::OperatorWithKernel {
 
 class TopkOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  TopkOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("X", "(Tensor) The input of Topk op");
     AddOutput("Out", "(Tensor) The output tensor of Topk op");
     AddOutput("Indices", "(Tensor) The indices of Topk elements of input");

paddle/fluid/operators/transpose_op.cc

@@ -56,8 +56,7 @@ class TransposeOp : public framework::OperatorWithKernel {
 
 class TransposeOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  TransposeOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(
         "X",
         "(Tensor) The input tensor, tensors with rank up to 6 are supported.");

paddle/fluid/operators/uniform_random_batch_size_like_op.cc

@@ -32,9 +32,8 @@ class UniformRandomBatchSizeLikeOp : public BatchSizeLikeOp {
 };
 
 class UniformRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker {
- public:
-  UniformRandomBatchSizeLikeOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : BatchSizeLikeOpMaker(proto, op_checker) {
+ protected:
+  void Apply() override {
     AddComment(R"DOC(
 Uniform random operator
 

paddle/fluid/operators/uniform_random_op.cc

@@ -85,8 +85,7 @@ class UniformRandomOp : public framework::OperatorWithKernel {
 
 class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  UniformRandomOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddOutput("Out", "(Tensor) The output tensor of uniform random op");
     AddComment(R"DOC(
 Uniform random operator.

paddle/fluid/operators/unpool_op.cc

@@ -20,8 +20,7 @@ namespace operators {
 
 class Unpool2dOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  Unpool2dOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(
         "X",
         "(Tensor) The input tensor of unpool operator. "

paddle/fluid/operators/warpctc_op.cc

@@ -53,8 +53,7 @@ class WarpCTCOp : public framework::OperatorWithKernel {
 
 class WarpCTCOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  WarpCTCOpMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput("Logits",
              "(LodTensor, default: LoDTensor<float>), the unscaled "
              "probabilities of variable-length sequences, which is a 2-D "

paddle/fluid/operators/while_op.cc

@@ -68,8 +68,7 @@ class WhileOp : public framework::OperatorBase {
 
 class WhileOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  WhileOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() override {
     AddInput(kX,
              "A set of variables, which are required by operators inside the "
              "block of While Op.")

paddle/fluid/pybind/pybind.cc

@@ -502,11 +502,13 @@ All parameter, weight, gradient are variables in Paddle.
               const std::unordered_set<std::string> &bcast_vars,
               const ProgramDesc &main_program, const std::string &loss_var_name,
               Scope *scope, std::vector<Scope *> &local_scopes,
-              bool allow_op_delay, bool use_default_grad_scale) {
+              bool allow_op_delay, bool use_default_grad_scale,
+              bool balance_parameter_opt_between_cards) {
              new (&self) ParallelExecutor(
                  num_threads, use_event, places, params, bcast_vars,
                  main_program, loss_var_name, scope, local_scopes,
-                 allow_op_delay, use_default_grad_scale);
+                 allow_op_delay, use_default_grad_scale,
+                 balance_parameter_opt_between_cards);
            })
       .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs)
       // NOTE: even we return a vec<Scope*>* to Python use reference policy.

paddle/gserver/layers/PriorBox.cpp

@@ -28,7 +28,7 @@ namespace paddle {
  */
 
 class PriorBoxLayer : public Layer {
-public:
+public:  // NOLINT
   explicit PriorBoxLayer(const LayerConfig& config) : Layer(config) {}
   bool init(const LayerMap& layerMap,
             const ParameterMap& parameterMap) override;
@@ -36,7 +36,7 @@ public:
   void forward(PassType passType) override;
   void backward(const UpdateCallback& callback) override {}
 
-protected:
+protected:  // NOLINT
   int numPriors_;
   std::vector<int> minSize_;
   std::vector<int> maxSize_;
@@ -109,11 +109,18 @@ void PriorBoxLayer::forward(PassType passType) {
         real boxWidth = minSize;
         real boxHeight = minSize;
 
-        // priors with different aspect ratios
-        for (size_t r = 0; r < aspectRatio_.size(); r++) {
-          real ar = aspectRatio_[r];
-          boxWidth = minSize * sqrt(ar);
-          boxHeight = minSize / sqrt(ar);
+        // first prior: aspect_ratio == 1.0, compatible to old logic
+        tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth;
+        tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight;
+        tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth;
+        tmpPtr[idx++] = (centerY + boxHeight / 2.) / imageHeight;
+        // set the variance.
+        for (int t = 0; t < 4; t++) tmpPtr[idx++] = variance_[t];
+
+        if (maxSize_.size() > 0) {
+          // square prior with size sqrt(minSize * maxSize)
+          real maxSize = maxSize_[s];
+          boxWidth = boxHeight = sqrt(minSize * maxSize);
           tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth;
           tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight;
           tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth;
@@ -122,10 +129,14 @@ void PriorBoxLayer::forward(PassType passType) {
           for (int t = 0; t < 4; t++) tmpPtr[idx++] = variance_[t];
         }
 
-        if (maxSize_.size() > 0) {
-          // square prior with size sqrt(minSize * maxSize)
-          real maxSize = maxSize_[s];
-          boxWidth = boxHeight = sqrt(minSize * maxSize);
+        // priors with different aspect ratios
+        for (size_t r = 0; r < aspectRatio_.size(); r++) {
+          real ar = aspectRatio_[r];
+          if (fabs(ar - 1.0) < 1e-6) {
+            continue;
+          }
+          boxWidth = minSize * sqrt(ar);
+          boxHeight = minSize / sqrt(ar);
           tmpPtr[idx++] = (centerX - boxWidth / 2.) / imageWidth;
           tmpPtr[idx++] = (centerY - boxHeight / 2.) / imageHeight;
           tmpPtr[idx++] = (centerX + boxWidth / 2.) / imageWidth;

python/paddle/fluid/data_feeder.py

@@ -16,6 +16,7 @@ from __future__ import print_function
 import core
 import numpy
 import six.moves as six
+import multiprocessing
 
 from framework import Variable, default_main_program
 
@@ -116,3 +117,60 @@ class DataFeeder(object):
         for each_name, each_converter in six.zip(self.feed_names, converter):
             ret_dict[each_name] = each_converter.done()
         return ret_dict
+
+    def feed_parallel(self, iterable, num_places=None):
+        if isinstance(self.place, core.CUDAPlace):
+            places = [
+                core.CUDAPlace(i)
+                for i in six.xrange(self._get_number_of_places_(num_places))
+            ]
+        else:
+            places = [
+                core.CPUPlace()
+                for _ in six.xrange(self._get_number_of_places_(num_places))
+            ]
+
+        if len(iterable) != len(places):
+            raise ValueError("feed_parallel takes multiple mini-batches. Each "
+                             "mini-batch will be feed on each device. The "
+                             "number of devices and number of mini-batches "
+                             "must be same.")
+
+        place = self.place
+        for p, batch in six.zip(places, iterable):
+            self.place = p
+            yield self.feed(batch)
+        self.place = place
+
+    def _get_number_of_places_(self, num_places):
+        if num_places is not None:
+            return int(num_places)
+        elif isinstance(self.place, core.CUDAPlace):
+            return core.get_cuda_device_count()
+        else:
+            return multiprocessing.cpu_count()
+
+    def decorate_reader(self,
+                        reader,
+                        multi_devices,
+                        num_places=None,
+                        drop_last=True):
+        def __reader_creator__():
+            if not multi_devices:
+                for item in reader():
+                    yield self.feed(item)
+            else:
+                num = self._get_number_of_places_(num_places)
+                item = []
+                for batch in reader():
+                    item.append(batch)
+                    if len(item) == num:
+                        yield list(self.feed_parallel(item, num))
+                        item = []
+                if not drop_last and len(item) != 0:
+                    raise ValueError(
+                        "The data batch which cannot fit for devices will be "
+                        "dropped is not implementation. Other strategies are "
+                        "not implemented")
+
+        return __reader_creator__

python/paddle/fluid/inferencer.py

@@ -16,31 +16,42 @@ import core
 import framework
 import executor
 import io
+from trainer import check_and_get_place
+
 __all__ = ['Inferencer', ]
 
 
 class Inferencer(object):
-    def __init__(self, network_func, param_path=None, place=None):
-        # 1. we need to generate a framework.Program by calling
-        # network_func. Reference: fluid.program_guard in test_word2vec.py
-
-        # 2. move the default_main_program to self.program.
-
-        # 3. run the default_startup program.
-
-        # 4. load params from param_path into scope
+    def __init__(self, param_path, place=None):
+        """
+        :param param_path: the path where the inference model is saved by fluid.io.save_inference_model
+        :param place: place to do the inference
+        """
+        self.param_path = param_path
         self.scope = core.Scope()
-        self.place = place
-        self.startup_program = framework.Program()
-        # TODO: generate the startup_program with network_func
-
-        exe = executor.Executor(place)
-        exe.run(self.startup_program, scope=self.scope)
 
-        if param_path:
+        self.exe = executor.Executor(check_and_get_place(place))
+        with executor.scope_guard(self.scope):
             # load params from param_path into scope
-            io.load_persistables(exe, dirname=param_path)
-
-    def infer(self, inputs):
-        # run self.program
-        pass
+            [self.inference_program, _,
+             self.fetch_targets] = io.load_inference_model(
+                 executor=self.exe, dirname=param_path)
+
+    def infer(self, inputs, return_numpy=True):
+        """
+        :param inputs: a map of {"input_name": input_var} that will be feed into the inference program
+        to get the predict value
+        :param return_numpy: if return numpy value for row tensor
+        :return: the predict value of the inference model
+        """
+        if not isinstance(inputs, dict):
+            raise ValueError(
+                "inputs should be a map of {'input_name': input_var}")
+
+        with executor.scope_guard(self.scope):
+            results = self.exe.run(self.inference_program,
+                                   feed=inputs,
+                                   fetch_list=self.fetch_targets,
+                                   return_numpy=return_numpy)
+
+        return results

python/paddle/fluid/io.py

@@ -263,6 +263,9 @@ def get_inference_program(target_vars, main_program=None):
 def prepend_feed_ops(inference_program,
                      feed_target_names,
                      feed_holder_name='feed'):
+    if len(feed_target_names) == 0:
+        return
+
     global_block = inference_program.global_block()
     feed_var = global_block.create_var(
         name=feed_holder_name,
@@ -323,9 +326,10 @@ def save_inference_model(dirname,
     if isinstance(feeded_var_names, basestring):
         feeded_var_names = [feeded_var_names]
     else:
-        if not (bool(feeded_var_names) and all(
-                isinstance(name, basestring) for name in feeded_var_names)):
-            raise ValueError("'feed_var_names' should be a list of str.")
+        if len(feeded_var_names) > 0:
+            if not (bool(feeded_var_names) and all(
+                    isinstance(name, basestring) for name in feeded_var_names)):
+                raise ValueError("'feed_var_names' should be a list of str.")
 
     if isinstance(target_vars, Variable):
         target_vars = [target_vars]

python/paddle/fluid/layers/layer_function_generator.py

@@ -113,7 +113,7 @@ def generate_layer_fn(op_type):
 
     if len(not_intermediate_outputs) != 1:
         raise ValueError("Only one non intermediate output operator can be",
-                         "automatically generated.")
+                         "automatically generated. {0}".format(op_type))
 
     if not_intermediate_outputs[0].duplicable:
         raise ValueError(

python/paddle/fluid/parallel_executor.py

@@ -30,7 +30,8 @@ class ParallelExecutor(object):
                  num_threads=None,
                  allow_op_delay=False,
                  share_vars_from=None,
-                 use_default_grad_scale=True):
+                 use_default_grad_scale=True,
+                 balance_parameter_opt_between_cards=False):
         """
         ParallelExecutor can run program in parallel.
 
@@ -51,6 +52,9 @@ class ParallelExecutor(object):
                 gradients of each device and scaled gradients would be
                 aggregated. Otherwise, a customized scale value should be fed
                 to the network.
+            balance_parameter_opt_between_cards(bool, default True): Whether
+                updating different gradients on different cards. Currently, it
+                is not recommended.
 
         Returns:
             A ParallelExecutor object.
@@ -129,7 +133,8 @@ class ParallelExecutor(object):
             scope,
             local_scopes,
             allow_op_delay,
-            use_default_grad_scale)
+            use_default_grad_scale,
+            balance_parameter_opt_between_cards)
 
         self.scope = scope
 

python/paddle/fluid/tests/book/CMakeLists.txt

@@ -5,3 +5,5 @@ string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
 foreach(src ${TEST_OPS})
     py_test(${src} SRCS ${src}.py)
 endforeach()
+
+add_subdirectory(high-level-api)

python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt

+file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
+string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
+
+# default test
+foreach(src ${TEST_OPS})
+    py_test(${src} SRCS ${src}.py)
+endforeach()
+
+add_subdirectory(recognize_digits)

python/paddle/fluid/tests/book/high-level-api/recognize_digits/CMakeLists.txt

+file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
+string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
+
+# default test
+foreach(src ${TEST_OPS})
+    py_test(${src} SRCS ${src}.py)
+endforeach()

python/paddle/fluid/tests/book/high-level-api/recognize_digits/notest_recognize_digits_conv.py → python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py

@@ -21,7 +21,6 @@ import unittest
 import math
 import sys
 import os
-import paddle.v2.dataset as dataset
 
 BATCH_SIZE = 64
 
@@ -54,47 +53,65 @@ def train_program():
     predict = inference_program()
     cost = fluid.layers.cross_entropy(input=predict, label=label)
     avg_cost = fluid.layers.mean(cost)
-    acc = fluid.layers.accuracy(input=predict, label=label)
-    return avg_cost, acc
+    # acc = fluid.layers.accuracy(input=predict, label=label)
+    # return avg_cost, acc
+    return avg_cost
 
 
 def train(use_cuda, save_dirname):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-
     optimizer = fluid.optimizer.Adam(learning_rate=0.001)
-    trainer = fluid.Trainer(train_program, place=place, optimizer=optimizer)
+
+    trainer = fluid.Trainer(
+        train_func=train_program,
+        infer_func=inference_program,
+        place=place,
+        optimizer=optimizer)
 
     def event_handler(event):
-        if isinstance(event, fluid.EndIteration):
-            avg_cost, acc = event.values
-            print("avg_cost: %s" % avg_cost)
-            print("acc     : %s" % acc)
-
-            if (event.batch_id + 1) % 10 == 0:
-                test_metrics = trainer.test(reader=dataset.mnist.test())
-                avg_cost_set = test_metrics[0]
-                acc_set = test_metrics[1]
-
-                # get test acc and loss
-                acc = numpy.array(acc_set).mean()
-                avg_cost = numpy.array(avg_cost_set).mean()
-                if float(acc) > 0.2:  # Smaller value to increase CI speed
-                    trainer.save_params(save_dirname)
-                else:
-                    print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format(
-                        event.batch_id + 1, float(avg_cost), float(acc)))
-                    if math.isnan(float(avg_cost)):
-                        sys.exit("got NaN loss, training failed.")
+        if isinstance(event, fluid.EndEpochEvent):
+            # if (event.epoch + 1) % 10 == 0:
+            # trainer.save_params(save_dirname)
+            trainer.save_inference_model(save_dirname)
+
+            # TODO: Uncomment this part once we are sure that .train is working
+            # test_reader = paddle.batch(
+            #     paddle.dataset.mnist.test(), batch_size=BATCH_SIZE)
+            # test_metrics = trainer.test(reader=test_reader)
+            # avg_cost_set = test_metrics[0]
+            # acc_set = test_metrics[1]
+            #
+            # # get test acc and loss
+            # acc = numpy.array(acc_set).mean()
+            # avg_cost = numpy.array(avg_cost_set).mean()
+            #
+            # print("avg_cost: %s" % avg_cost)
+            # print("acc     : %s" % acc)
+            #
+            # if float(acc) > 0.2:  # Smaller value to increase CI speed
+            #     trainer.save_params(save_dirname)
+            # else:
+            #     print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format(
+            #         event.epoch + 1, float(avg_cost), float(acc)))
+            #     if math.isnan(float(avg_cost)):
+            #         sys.exit("got NaN loss, training failed.")
+
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.mnist.train(), buf_size=500),
+        batch_size=BATCH_SIZE)
 
     trainer.train(
-        reader=dataset.mnist.train(), num_pass=100, event_handler=event_handler)
+        num_epochs=1,
+        event_handler=event_handler,
+        reader=train_reader,
+        feed_order=['img', 'label'])
 
 
 def infer(use_cuda, save_dirname=None):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 
-    inferencer = fluid.Inferencer(
-        inference_program, param_path=save_dirname, place=place)
+    inferencer = fluid.Inferencer(param_path=save_dirname, place=place)
 
     batch_size = 1
     tensor_img = numpy.random.uniform(-1.0, 1.0,
@@ -114,5 +131,5 @@ def main(use_cuda):
 
 
 if __name__ == '__main__':
-    for use_cuda in (False, True):
-        main(use_cuda=use_cuda)
+    # for use_cuda in (False, True):
+    main(use_cuda=False)

python/paddle/fluid/tests/book/high-level-api/recognize_digits/notest_recognize_digits_mlp.py → python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py

@@ -21,7 +21,6 @@ import unittest
 import math
 import sys
 import os
-import paddle.v2.dataset as dataset
 
 BATCH_SIZE = 64
 
@@ -41,47 +40,64 @@ def train_program():
     predict = inference_program()
     cost = fluid.layers.cross_entropy(input=predict, label=label)
     avg_cost = fluid.layers.mean(cost)
-    acc = fluid.layers.accuracy(input=predict, label=label)
-    return avg_cost, acc
+    # acc = fluid.layers.accuracy(input=predict, label=label)
+    # return avg_cost, acc
+    return avg_cost
 
 
 def train(use_cuda, save_dirname):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-
     optimizer = fluid.optimizer.Adam(learning_rate=0.001)
-    trainer = fluid.Trainer(train_program, place=place, optimizer=optimizer)
+
+    trainer = fluid.Trainer(
+        train_func=train_program,
+        infer_func=inference_program,
+        place=place,
+        optimizer=optimizer)
 
     def event_handler(event):
-        if isinstance(event, fluid.EndIteration):
-            avg_cost, acc = event.values
-            print("avg_cost: %s" % avg_cost)
-            print("acc     : %s" % acc)
-
-            if (event.batch_id + 1) % 10 == 0:
-                test_metrics = trainer.test(reader=dataset.mnist.test())
-                avg_cost_set = test_metrics[0]
-                acc_set = test_metrics[1]
-
-                # get test acc and loss
-                acc = numpy.array(acc_set).mean()
-                avg_cost = numpy.array(avg_cost_set).mean()
-                if float(acc) > 0.2:  # Smaller value to increase CI speed
-                    trainer.save_params(save_dirname)
-                else:
-                    print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format(
-                        event.batch_id + 1, float(avg_cost), float(acc)))
-                    if math.isnan(float(avg_cost)):
-                        sys.exit("got NaN loss, training failed.")
+        if isinstance(event, fluid.EndEpochEvent):
+            # if (event.epoch + 1) % 10 == 0:
+            trainer.save_inference_model(save_dirname)
+
+            # TODO: Uncomment this part once we are sure that .train is working
+            # test_reader = paddle.batch(
+            #     paddle.dataset.mnist.test(), batch_size=BATCH_SIZE)
+            # test_metrics = trainer.test(reader=test_reader)
+            # avg_cost_set = test_metrics[0]
+            # acc_set = test_metrics[1]
+            #
+            # # get test acc and loss
+            # acc = numpy.array(acc_set).mean()
+            # avg_cost = numpy.array(avg_cost_set).mean()
+            #
+            # print("avg_cost: %s" % avg_cost)
+            # print("acc     : %s" % acc)
+            #
+            # if float(acc) > 0.2:  # Smaller value to increase CI speed
+            #     trainer.save_params(save_dirname)
+            # else:
+            #     print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format(
+            #         event.epoch + 1, float(avg_cost), float(acc)))
+            #     if math.isnan(float(avg_cost)):
+            #         sys.exit("got NaN loss, training failed.")
+
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.mnist.train(), buf_size=500),
+        batch_size=BATCH_SIZE)
 
     trainer.train(
-        reader=dataset.mnist.train(), num_pass=100, event_handler=event_handler)
+        num_epochs=1,
+        event_handler=event_handler,
+        reader=train_reader,
+        feed_order=['img', 'label'])
 
 
 def infer(use_cuda, save_dirname=None):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 
-    inferencer = fluid.Inferencer(
-        inference_program, param_path=save_dirname, place=place)
+    inferencer = fluid.Inferencer(param_path=save_dirname, place=place)
 
     batch_size = 1
     tensor_img = numpy.random.uniform(-1.0, 1.0,
@@ -101,5 +117,5 @@ def main(use_cuda):
 
 
 if __name__ == '__main__':
-    for use_cuda in (False, True):
-        main(use_cuda=use_cuda)
+    # for use_cuda in (False, True):
+    main(use_cuda=False)

python/paddle/fluid/tests/book/high-level-api/word2vec/no_test_word2vec_new_api.py

@@ -99,45 +99,45 @@ def train(use_cuda, is_sparse, save_path):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 
     def event_handler(event):
-        # print type(event)
         if isinstance(event, fluid.EndEpochEvent):
             outs = trainer.test(reader=test_reader)
             avg_cost = outs[0]
             print("loss= ", avg_cost)
 
             if avg_cost < 5.0:
-                trainer.save_params(save_path)
+                trainer.save_inference_model(save_path)
                 return
             if math.isnan(avg_cost):
                 sys.exit("got NaN loss, training failed.")
 
     trainer = fluid.Trainer(
         partial(train_program, is_sparse),
+        partial(inference_program, is_sparse),
         fluid.optimizer.SGD(learning_rate=0.001),
         place=place)
     trainer.train(
-        reader=train_reader, num_epochs=100, event_handler=event_handler)
+        reader=train_reader, num_epochs=1, event_handler=event_handler)
 
 
 def infer(use_cuda, is_sparse, save_path):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-    inferencer = fluid.Inferencer(
-        partial(inference_program, is_sparse),
-        param_path=save_path,
-        place=place)
+    inferencer = fluid.Inferencer(param_path=save_path, place=place)
 
     lod = [0, 1]
     first_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1)
     second_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1)
     third_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1)
     fourth_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1)
-    result = inferencer.infer({
-        'firstw': first_word,
-        'secondw': second_word,
-        'thirdw': third_word,
-        'forthw': fourth_word
-    })
-    print(result)
+
+    result = inferencer.infer(
+        {
+            'firstw': first_word,
+            'secondw': second_word,
+            'thirdw': third_word,
+            'forthw': fourth_word
+        },
+        return_numpy=False)
+    print(np.array(result[0]))
 
 
 def main(use_cuda, is_sparse):

python/paddle/fluid/tests/unittests/test_network_with_dtype.py

@@ -24,7 +24,7 @@ BATCH_SIZE = 20
 
 
 class TestNetWithDtype(unittest.TestCase):
-    def setUp(self):
+    def set_network(self):
         self.dtype = "float64"
         self.init_dtype()
         self.x = fluid.layers.data(name='x', shape=[13], dtype=self.dtype)
@@ -55,12 +55,14 @@ class TestNetWithDtype(unittest.TestCase):
         pass
 
     def test_cpu(self):
+        self.set_network()
         place = fluid.CPUPlace()
         self.run_net_on_place(place)
 
     def test_gpu(self):
         if not core.is_compiled_with_cuda():
             return
+        self.set_network()
         place = fluid.CUDAPlace(0)
         self.run_net_on_place(place)
 

python/paddle/fluid/tests/unittests/test_parallel_executor.py

@@ -205,7 +205,8 @@ class TestParallelExecutorBase(unittest.TestCase):
                                   allow_op_delay=False,
                                   feed_dict=None,
                                   seed=None,
-                                  use_parallel_executor=True):
+                                  use_parallel_executor=True,
+                                  balance_parameter_opt_between_cards=False):
         def run_executor(exe, feed, fetch_list, program=None):
             if isinstance(exe, fluid.ParallelExecutor):
                 res = exe.run(fetch_list=fetch_list, feed=feed)
@@ -234,7 +235,11 @@ class TestParallelExecutorBase(unittest.TestCase):
 
             if use_parallel_executor:
                 exe = fluid.ParallelExecutor(
-                    True, loss_name=loss.name, allow_op_delay=allow_op_delay)
+                    True,
+                    loss_name=loss.name,
+                    allow_op_delay=allow_op_delay,
+                    balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+                )
             else:
                 exe = fluid.Executor(place=place)
 
@@ -280,20 +285,27 @@ class TestMNIST(TestParallelExecutorBase):
             fluid.recordio_writer.convert_reader_to_recordio_file(
                 './mnist.recordio', reader, feeder)
 
-    def check_simple_fc_convergence(self):
+    def check_simple_fc_convergence(self, balance_parameter_opt_between_cards):
         self.check_network_convergence(simple_fc_net)
         self.check_network_convergence(simple_fc_net, allow_op_delay=True)
 
         img = np.zeros(shape=[32, 784], dtype='float32')
         label = np.ones(shape=[32, 1], dtype='int64')
         self.check_network_convergence(
-            simple_fc_net, feed_dict={"image": img,
-                                      "label": label})
+            simple_fc_net,
+            feed_dict={"image": img,
+                       "label": label},
+            balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+        )
 
     def test_simple_fc(self):
-        self.check_simple_fc_convergence()
+        self.check_simple_fc_convergence(False)
+
+    def test_simple_fc_with_new_strategy(self):
+        self.check_simple_fc_convergence(True)
 
-    def check_simple_fc_parallel_accuracy(self):
+    def check_simple_fc_parallel_accuracy(self,
+                                          balance_parameter_opt_between_cards):
         img = np.zeros(shape=[32, 784], dtype='float32')
         label = np.ones(shape=[32, 1], dtype='int64')
         single_first_loss, single_last_loss = self.check_network_convergence(
@@ -307,7 +319,9 @@ class TestMNIST(TestParallelExecutorBase):
             seed=1000,
             feed_dict={"image": img,
                        "label": label},
-            use_parallel_executor=True)
+            use_parallel_executor=True,
+            balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+        )
 
         for p_f in parallel_first_loss:
             self.assertAlmostEquals(p_f, single_first_loss[0], delta=1e-6)
@@ -315,18 +329,28 @@ class TestMNIST(TestParallelExecutorBase):
             self.assertAlmostEquals(p_l, single_last_loss[0], delta=1e-6)
 
     def test_simple_fc_parallel_accuracy(self):
-        self.check_simple_fc_parallel_accuracy()
+        self.check_simple_fc_parallel_accuracy(False)
 
-    def check_batchnorm_fc_convergence(self):
+    def test_simple_fc_parallel_accuracy_with_new_strategy(self):
+        self.check_simple_fc_parallel_accuracy(True)
+
+    def check_batchnorm_fc_convergence(self,
+                                       balance_parameter_opt_between_cards):
         self.check_network_convergence(fc_with_batchnorm)
         img = np.zeros(shape=[32, 784], dtype='float32')
         label = np.ones(shape=[32, 1], dtype='int64')
         self.check_network_convergence(
-            fc_with_batchnorm, feed_dict={"image": img,
-                                          "label": label})
+            fc_with_batchnorm,
+            feed_dict={"image": img,
+                       "label": label},
+            balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+        )
 
     def test_batchnorm_fc(self):
-        self.check_batchnorm_fc_convergence()
+        self.check_batchnorm_fc_convergence(False)
+
+    def test_batchnorm_fc_with_new_strategy(self):
+        self.check_batchnorm_fc_convergence(True)
 
 
 class TestResnet(TestParallelExecutorBase):
@@ -348,17 +372,22 @@ class TestResnet(TestParallelExecutorBase):
     #         fluid.recordio_writer.convert_reader_to_recordio_file(
     #             "./flowers.recordio", reader, feeder, compressor=fluid.core.RecordIOWriter.Compressor.NoCompress)
 
-    def check_resnet_convergence(self):
+    def check_resnet_convergence(self, balance_parameter_opt_between_cards):
         import functools
         batch_size = 2
         self.check_network_convergence(
             functools.partial(
                 SE_ResNeXt50Small, batch_size=batch_size),
             iter=20,
-            batch_size=batch_size)
+            batch_size=batch_size,
+            balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+        )
 
     def test_resnet(self):
-        self.check_resnet_convergence()
+        self.check_resnet_convergence(False)
+
+    def test_resnet_with_new_strategy(self):
+        self.check_resnet_convergence(True)
 
 
 class ModelHyperParams(object):
@@ -519,7 +548,7 @@ class TestTransformer(TestParallelExecutorBase):
 
 
 class ParallelExecutorTestingDuringTraining(unittest.TestCase):
-    def check_network_convergence(self):
+    def check_network_convergence(self, balance_parameter_opt_between_cards):
         main = fluid.Program()
         startup = fluid.Program()
         with fluid.program_guard(main, startup):
@@ -539,12 +568,18 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase):
             feed_dict = {'image': image, 'label': label}
 
             train_exe = fluid.ParallelExecutor(
-                use_cuda=True, loss_name=loss.name, main_program=main)
+                use_cuda=True,
+                loss_name=loss.name,
+                main_program=main,
+                balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+            )
 
             test_exe = fluid.ParallelExecutor(
                 use_cuda=True,
                 main_program=test_program,
-                share_vars_from=train_exe)
+                share_vars_from=train_exe,
+                balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+            )
 
             for i in xrange(5):
                 test_loss, = test_exe.run([loss.name], feed=feed_dict)
@@ -558,8 +593,11 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase):
                     "Train loss: " + str(train_loss) + "\n Test loss:" +
                     str(test_loss))
 
-    def test_parallel(self):
-        self.check_network_convergence()
+    def test_parallel_testing(self):
+        self.check_network_convergence(False)
+
+    def test_parallel_testing_with_new_strategy(self):
+        self.check_network_convergence(True)
 
 
 import paddle.dataset.conll05 as conll05
@@ -579,7 +617,7 @@ embedding_name = 'emb'
 
 
 def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark,
-            is_sparse, **ignored):
+            is_sparse, balance_parameter_opt_between_cards, **ignored):
     # 8 features
     predicate_embedding = fluid.layers.embedding(
         input=predicate,
@@ -648,7 +686,9 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark,
 
 
 class TestCRFModel(unittest.TestCase):
-    def check_network_convergence(self, is_sparse):
+    def check_network_convergence(self,
+                                  is_sparse,
+                                  balance_parameter_opt_between_cards=False):
         main = fluid.Program()
         startup = fluid.Program()
         with fluid.program_guard(main, startup):
@@ -696,7 +736,11 @@ class TestCRFModel(unittest.TestCase):
             exe = fluid.Executor(place)
             exe.run(startup)
 
-            pe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name)
+            pe = fluid.ParallelExecutor(
+                use_cuda=True,
+                loss_name=avg_cost.name,
+                balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+            )
 
             feeder = fluid.DataFeeder(
                 feed_list=[
@@ -718,6 +762,14 @@ class TestCRFModel(unittest.TestCase):
     def test_update_dense_parameter(self):
         self.check_network_convergence(is_sparse=False)
 
+    def test_update_sparse_parameter_with_new_strategy(self):
+        self.check_network_convergence(
+            is_sparse=False, balance_parameter_opt_between_cards=True)
+
+    def test_update_dense_parameter_with_new_strategy(self):
+        self.check_network_convergence(
+            is_sparse=False, balance_parameter_opt_between_cards=True)
+
 
 # test fetch all the variables of global_block
 
@@ -796,5 +848,42 @@ class TestFetchOp(unittest.TestCase):
         self.parallel_exe(train_inputs, seed=1)
 
 
+class TestFeedParallel(unittest.TestCase):
+    def test_main(self):
+        main = fluid.Program()
+        startup = fluid.Program()
+        startup.random_seed = 1
+        with fluid.scope_guard(fluid.core.Scope()):
+            with fluid.program_guard(main, startup):
+                data = fluid.layers.data(
+                    name='image', shape=[3, 224, 224], dtype='float32')
+                label = fluid.layers.data(
+                    name='label', shape=[1], dtype='int64')
+                out = Lenet(data, class_dim=102)
+                loss = fluid.layers.cross_entropy(input=out, label=label)
+                loss = fluid.layers.mean(loss)
+                opt = fluid.optimizer.Momentum(
+                    learning_rate=0.1,
+                    momentum=0.9,
+                    regularization=fluid.regularizer.L2Decay(1e-4))
+
+                opt.minimize(loss)
+        place = fluid.CUDAPlace(0)
+        feeder = fluid.DataFeeder(place=place, feed_list=[data, label])
+        reader = feeder.decorate_reader(
+            paddle.batch(
+                flowers.train(), batch_size=16), multi_devices=True)
+        exe = fluid.Executor(place)
+        exe.run(startup)
+        pe = fluid.ParallelExecutor(
+            use_cuda=True, loss_name=loss.name, main_program=main)
+
+        for batch_id, data in enumerate(reader()):
+            loss_np = np.array(pe.run(feed=data, fetch_list=[loss.name])[0])
+            print batch_id, loss_np
+            if batch_id == 2:
+                break
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/trainer.py

@@ -19,7 +19,7 @@ import executor
 import data_feeder
 import contextlib
 import io
-import transpiler
+import unique_name
 
 # optimizer is same as the parameter of Trainer.__init__. Rename it to opt_module
 import optimizer as opt_module
@@ -56,26 +56,62 @@ class EndStepEvent(object):
         self.step = step_id
 
 
+def check_and_get_place(place):
+    """
+    Check the type of place or get the default place
+    Args:
+        place(None|core.CUDAPlace|core.CPUPlace): the place that trainer will be executed on.
+
+    Raises:
+        TypeError if the type mismatched.
+
+    Returns:
+        the original place if it is not None.
+        if fluid is compiled with CUDA, returns CUDAPlace(0) by default.
+        Otherwise returns CPUPlace by default.
+    """
+    if place is None:
+        if core.is_compiled_with_cuda():
+            return core.CUDAPlace(0)
+        else:
+            return core.CPUPlace()
+    else:
+        if not isinstance(place, core.CUDAPlace) and not isinstance(
+                place, core.CPUPlace):
+            raise TypeError("Place should be either CUDAPlace or CPUPlace")
+        return place
+
+
 class Trainer(object):
     """
 
     Args:
-        program_func(callable): A function which will return loss. The loss must be a scaler.
+        train_func(callable): A function which will return loss. The loss must be a scalar.
+        infer_func(callable): A function which will return predict, used to save inference model
         optimizer(optimizer.Optimizer): The optimizer should be an instance of Optimizer
         place: The device place of this trainer.
     """
 
-    def __init__(self, program_func, optimizer, param_path=None, place=None):
+    def __init__(self,
+                 train_func,
+                 infer_func,
+                 optimizer,
+                 param_path=None,
+                 place=None):
         # 1. we need to generate a framework.Program by calling
         # program_func. Reference: fluid.program_guard in
         # test_word2vec.py
+        if not isinstance(optimizer, opt_module.Optimizer):
+            raise TypeError("The optimizer should be an instance of Optimizer")
+
+        self.infer_func = infer_func
         self.scope = core.Scope()
 
         self.startup_program = framework.Program()
         self.train_program = framework.Program()
 
         with framework.program_guard(self.train_program, self.startup_program):
-            program_func_outs = program_func()
+            program_func_outs = train_func()
             self.test_outputs = program_func_outs if isinstance(
                 program_func_outs, list) else [program_func_outs]
             self.test_program = self.train_program.clone()
@@ -86,9 +122,9 @@ class Trainer(object):
             loss = self.test_outputs[0]
             optimize_ops, params_grads = optimizer.minimize(loss)
 
-        self.place = Trainer._check_and_get_place(place)
+        self.place = check_and_get_place(place)
 
-        self.dist_transpile_if_necessary(optimize_ops, params_grads)
+        self._dist_transpile_if_necessary(optimize_ops, params_grads)
 
         # 2. move the default_main_program to self.program and run the
         # default_startup program on an empty core.Scope()
@@ -101,7 +137,7 @@ class Trainer(object):
             # load params from param_path into scope
             io.load_persistables(exe, dirname=param_path)
 
-    def dist_transpile_if_necessary(self, optimize_ops, params_grads):
+    def _dist_transpile_if_necessary(self, optimize_ops, params_grads):
         if "PADDLE_TRAINING_ROLE" not in os.environ:
             return
 
@@ -190,31 +226,14 @@ class Trainer(object):
             exe = executor.Executor(self.place)
             io.save_persistables(exe, dirname=param_path)
 
-    @staticmethod
-    def _check_and_get_place(place):
-        """
-        Check the type of place or get the default place
-        Args:
-            place(None|core.CUDAPlace|core.CPUPlace): the place that trainer will be executed on.
-
-        Raises:
-            TypeError if the type mismatched.
-
-        Returns:
-            the original place if it is not None.
-            if fluid is compiled with CUDA, returns CUDAPlace(0) by default.
-            Otherwise returns CPUPlace by default.
-        """
-        if place is None:
-            if core.is_compiled_with_cuda():
-                return core.CUDAPlace(0)
-            else:
-                return core.CPUPlace()
-        else:
-            if not isinstance(place, core.CUDAPlace) and not isinstance(
-                    place, core.CPUPlace):
-                raise TypeError("Place should be either CUDAPlace or CPUPlace")
-            return place
+    def save_inference_model(self, model_path):
+        inference_program = framework.Program()
+        with framework.program_guard(inference_program):
+            with unique_name.guard():
+                predict_var = self.infer_func()
+        predict_var = self.train_program.block(0).var(predict_var.name)
+        exe = executor.Executor(self.place)
+        io.save_inference_model(model_path, [], [predict_var], exe)
 
     @contextlib.contextmanager
     def _prog_and_scope_guard(self):