merge develop

CMakeLists.txt

@@ -86,6 +86,14 @@ if(ANDROID OR IOS)
         "Disable MKLDNN when cross-compiling for Android and iOS" FORCE)
     set(WITH_MKLML OFF CACHE STRING
         "Disable MKLML package when cross-compiling for Android and iOS" FORCE)
+
+    # Compile PaddlePaddle mobile inference library
+    if (NOT WITH_C_API)
+        set(WITH_C_API ON CACHE STRING
+            "Always compile the C_API when cross-compiling for Android and iOS" FORCE)
+    endif()
+    set(MOBILE_INFERENCE ON)
+    add_definitions(-DPADDLE_MOBILE_INFERENCE)
 endif()
 
 set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
@@ -160,9 +168,11 @@ endif(USE_NNPACK)
 
 add_subdirectory(proto)
 
-# "add_subdirectory(go)" should be placed after the following loine,
-# because it depends on paddle/optimizer.
-add_subdirectory(paddle/optimizer)
+if(NOT MOBILE_INFERENCE)
+    # "add_subdirectory(go)" should be placed after the following loine,
+    # because it depends on paddle/optimizer.
+    add_subdirectory(paddle/optimizer)
+endif()
 
 # "add_subdirectory(paddle)" and "add_subdirectory(python)" should be
 # placed after this block, because they depends on it.

cmake/util.cmake

@@ -73,25 +73,43 @@ function(link_paddle_exe TARGET_NAME)
         generate_rdma_links()
     endif()
 
-    target_circle_link_libraries(${TARGET_NAME}
-        ARCHIVE_START
-        paddle_gserver
-        paddle_function
-        ARCHIVE_END
-        paddle_pserver
-        paddle_trainer_lib
-        paddle_network
-        paddle_math
-        paddle_utils
-        paddle_parameter
-        paddle_proto
-        paddle_cuda
-        paddle_optimizer
-        ${EXTERNAL_LIBS}
-        ${CMAKE_THREAD_LIBS_INIT}
-        ${CMAKE_DL_LIBS}
-        ${RDMA_LD_FLAGS}
-        ${RDMA_LIBS})
+    if(MOBILE_INFERENCE)
+        target_circle_link_libraries(${TARGET_NAME}
+            ARCHIVE_START
+            paddle_gserver
+            paddle_function
+            ARCHIVE_END
+            paddle_math
+            paddle_utils
+            paddle_parameter
+            paddle_proto
+            paddle_cuda
+            ${EXTERNAL_LIBS}
+            ${CMAKE_THREAD_LIBS_INIT}
+            ${CMAKE_DL_LIBS}
+            ${RDMA_LD_FLAGS}
+            ${RDMA_LIBS})
+    else()
+        target_circle_link_libraries(${TARGET_NAME}
+            ARCHIVE_START
+            paddle_gserver
+            paddle_function
+            ARCHIVE_END
+            paddle_pserver
+            paddle_trainer_lib
+            paddle_network
+            paddle_math
+            paddle_utils
+            paddle_parameter
+            paddle_proto
+            paddle_cuda
+            paddle_optimizer
+            ${EXTERNAL_LIBS}
+            ${CMAKE_THREAD_LIBS_INIT}
+            ${CMAKE_DL_LIBS}
+            ${RDMA_LD_FLAGS}
+            ${RDMA_LIBS})
+    endif()
 
     if(ANDROID)
         target_link_libraries(${TARGET_NAME} log)

paddle/CMakeLists.txt

 add_subdirectory(cuda)
 add_subdirectory(function)
 add_subdirectory(utils)
-add_subdirectory(testing)
 add_subdirectory(math)
-add_subdirectory(parameter)
 add_subdirectory(gserver)
-add_subdirectory(pserver)
-add_subdirectory(trainer)
-add_subdirectory(scripts)
-add_subdirectory(string)
-
-if(Boost_FOUND)
-  add_subdirectory(memory)
-  add_subdirectory(platform)
-  add_subdirectory(framework)
-  add_subdirectory(operators)
-  add_subdirectory(pybind)
-endif()
+add_subdirectory(parameter)
+add_subdirectory(testing)
 
-if(WITH_C_API)
+if(MOBILE_INFERENCE)
   add_subdirectory(capi)
-endif()
+else()
+  add_subdirectory(pserver)
+  add_subdirectory(trainer)
+  add_subdirectory(string)
+  add_subdirectory(scripts)
+
+  if(WITH_C_API)
+    add_subdirectory(capi)
+  endif()
+
+  if(Boost_FOUND)
+    add_subdirectory(memory)
+    add_subdirectory(platform)
+    add_subdirectory(framework)
+    add_subdirectory(operators)
+    add_subdirectory(pybind)
+  endif()
 
-if(WITH_SWIG_PY)
-  add_subdirectory(api)
+  if(WITH_SWIG_PY)
+    add_subdirectory(api)
+  endif()
 endif()

paddle/capi/CMakeLists.txt

@@ -37,9 +37,7 @@ set(PADDLE_CAPI_INFER_LIBS
     paddle_cuda
     paddle_function
     paddle_gserver
-    paddle_proto
-    paddle_pserver
-    paddle_network)
+    paddle_proto)
 
 cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_CAPI_INFER_LIBS})
 

paddle/capi/tests/CMakeLists.txt

@@ -4,11 +4,12 @@ add_unittest(capi_test_mats test_Vector.cpp
 target_include_directories(capi_test_mats PUBLIC ${PADDLE_CAPI_INC_PATH})
 target_link_libraries(capi_test_mats paddle_capi)
 
-
-add_unittest_without_exec(capi_test_gradientMachine test_GradientMachine.cpp)
-target_include_directories(capi_test_gradientMachine PUBLIC
-  ${PADDLE_CAPI_INC_PATH})
-target_link_libraries(capi_test_gradientMachine paddle_capi)
-add_test(NAME capi_test_gradientMachine
-  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine
-  WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/capi/tests)
+if(NOT MOBILE_INFERENCE)
+    add_unittest_without_exec(capi_test_gradientMachine test_GradientMachine.cpp)
+    target_include_directories(capi_test_gradientMachine PUBLIC
+      ${PADDLE_CAPI_INC_PATH})
+    target_link_libraries(capi_test_gradientMachine paddle_capi)
+    add_test(NAME capi_test_gradientMachine
+      COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine
+      WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/capi/tests)
+endif()

paddle/framework/backward.cc

@@ -302,7 +302,7 @@ std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
     return grad_op_descs;  // empty vector
   }
 
-  grad_op_descs = OpRegistry::CreateGradOpDescs(*op_desc);
+  grad_op_descs = OpRegistry::CreateGradOpDescs(op_desc.get());
 
   std::list<std::unique_ptr<OpDescBind>> pending_fill_zeros_ops;
   for (auto& desc : grad_op_descs) {

paddle/framework/backward_test.cc

@@ -58,6 +58,8 @@ class MulOpMaker : public OpProtoAndCheckerMaker {
     AddInput("X", "A");
     AddInput("Y", "B");
     AddOutput("Out", "Out");
+    AddAttr<int>("x_num_col_dims", "").SetDefault(1).EqualGreaterThan(1);
+    AddAttr<int>("y_num_col_dims", "").SetDefault(1).EqualGreaterThan(1);
     AddComment("Mul");
   }
 };
@@ -440,6 +442,28 @@ TEST(Backward, simple_single_op) {
             std::vector<std::string>({f::GradVarName("b")}));
 }
 
+TEST(Backward, default_attribute) {
+  f::ProgramDesc *program_desc = GetNewProgramDesc();
+  f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
+  f::BlockDescBind *block = program.Block(0);
+  f::OpDescBind *op = block->AppendOp();
+  op->SetType("mul");
+  op->SetInput("X", {"x"});
+  op->SetInput("Y", {"y"});
+  op->SetOutput("Out", {"out"});
+
+  AppendBackward(program, {});
+
+  ASSERT_EQ(block->AllOps().size(), 2UL);
+  EXPECT_EQ(boost::get<int>(op->GetAttr("x_num_col_dims")), 1);
+  EXPECT_EQ(boost::get<int>(op->GetAttr("y_num_col_dims")), 1);
+
+  f::OpDescBind *grad_op = block->AllOps()[1];
+  ASSERT_EQ(grad_op->Type(), "mul_grad");
+  EXPECT_EQ(boost::get<int>(grad_op->GetAttr("x_num_col_dims")), 1);
+  EXPECT_EQ(boost::get<int>(grad_op->GetAttr("y_num_col_dims")), 1);
+}
+
 TEST(Backward, simple_mult_op) {
   f::ProgramDesc *program_desc = GetNewProgramDesc();
   f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);

paddle/framework/block_desc.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <deque>
+#include <memory>
 #include <unordered_map>
 #include <vector>
 #include "paddle/framework/op_desc.h"

paddle/framework/framework.proto

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 syntax = "proto2";
+option optimize_for = LITE_RUNTIME;
 package paddle.framework;
 
 enum AttrType {

paddle/framework/op_desc.h

@@ -52,8 +52,6 @@ class OpDescBind {
   void SetOutput(const std::string &param_name,
                  const std::vector<std::string> &args);
 
-  std::string DebugString() { return this->Proto()->DebugString(); }
-
   bool HasAttr(const std::string &name) const {
     return attrs_.find(name) != attrs_.end();
   }
@@ -97,6 +95,11 @@ class OpDescBind {
 
   const VariableNameMap &Outputs() const { return outputs_; }
 
+  AttributeMap *MutableAttrMap() {
+    this->need_update_ = true;
+    return &this->attrs_;
+  }
+
  private:
   template <typename MapType>
   static std::vector<typename MapType::key_type> MapKeys(const MapType &map) {

paddle/framework/op_registry.cc

@@ -60,9 +60,14 @@ std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDescBind& op_desc) {
 }
 
 std::vector<std::unique_ptr<OpDescBind>> OpRegistry::CreateGradOpDescs(
-    const OpDescBind& op_desc) {
-  auto& info = OpInfoMap::Instance().Get(op_desc.Type());
-  return info.grad_op_maker_(op_desc);
+    OpDescBind* op_desc) {
+  auto& info = OpInfoMap::Instance().Get(op_desc->Type());
+
+  if (info.Checker() != nullptr) {
+    info.Checker()->Check(*op_desc->MutableAttrMap());
+  }
+
+  return info.grad_op_maker_(*op_desc);
 }
 
 }  // namespace framework

paddle/framework/op_registry.h

@@ -80,7 +80,7 @@ class OpRegistry {
   static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
 
   static std::vector<std::unique_ptr<OpDescBind>> CreateGradOpDescs(
-      const OpDescBind& op_desc);
+      OpDescBind* op_desc);
 
   static std::unique_ptr<OperatorBase> CreateOp(const OpDescBind& op_desc);
 };

paddle/framework/operator.cc

@@ -205,13 +205,13 @@ void OperatorBase::GenerateTemporaryNames() {
 }
 
 template <>
-const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const {
+const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
   auto* var = InputVar(name);
   return var == nullptr ? nullptr : GetTensorFromVar(var);
 }
 
 template <>
-const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
+const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
     const std::string& name) const {
   auto names = op().Inputs(name);
   std::vector<const Tensor*> res;
@@ -225,13 +225,13 @@ const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
 }
 
 template <>
-Tensor* InferShapeContext::Output<Tensor>(const std::string& name) const {
+Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
   auto var = OutputVar(name);
   return var == nullptr ? nullptr : var->GetMutable<LoDTensor>();
 }
 
 template <>
-std::vector<Tensor*> InferShapeContext::MultiOutput<Tensor>(
+std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
     const std::string& name) const {
   auto names = op().Outputs(name);
   std::vector<Tensor*> res;

paddle/framework/operator.h

@@ -57,7 +57,6 @@ inline std::string GradVarName(const std::string& var_name) {
 }
 
 class OperatorBase;
-class InferShapeContext;
 class ExecutionContext;
 
 extern const Tensor* GetTensorFromVar(const Variable* var);
@@ -169,10 +168,11 @@ class NOP : public OperatorBase {
   }
 };
 
-class InferShapeContext {
+class ExecutionContext {
  public:
-  InferShapeContext(const OperatorBase& op, const Scope& scope)
-      : op_(op), scope_(scope) {}
+  ExecutionContext(const OperatorBase& op, const Scope& scope,
+                   const platform::DeviceContext& device_context)
+      : op_(op), scope_(scope), device_context_(device_context) {}
 
   const OperatorBase& op() const { return op_; }
 
@@ -278,31 +278,6 @@ class InferShapeContext {
     out_tensor->set_lod(in_tensor.lod());
   }
 
- private:
-  const OperatorBase& op_;
-  const Scope& scope_;
-};
-
-template <>
-const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const;
-
-template <>
-const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
-    const std::string& name) const;
-
-template <>
-Tensor* InferShapeContext::Output<Tensor>(const std::string& name) const;
-
-template <>
-std::vector<Tensor*> InferShapeContext::MultiOutput<Tensor>(
-    const std::string& name) const;
-
-class ExecutionContext : public InferShapeContext {
- public:
-  ExecutionContext(const OperatorBase& op, const Scope& scope,
-                   const platform::DeviceContext& device_context)
-      : InferShapeContext(op, scope), device_context_(device_context) {}
-
   template <typename PlaceType,
             typename DeviceType = typename platform::EigenDeviceConverter<
                 PlaceType>::EigenDeviceType>
@@ -315,10 +290,26 @@ class ExecutionContext : public InferShapeContext {
   }
 
  private:
+  const OperatorBase& op_;
+  const Scope& scope_;
   const platform::DeviceContext& device_context_;
 };
 
-class CompileTimeInferShapeContext : public InferShapeContextBase {
+template <>
+const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const;
+
+template <>
+const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
+    const std::string& name) const;
+
+template <>
+Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
+
+template <>
+std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
+    const std::string& name) const;
+
+class CompileTimeInferShapeContext : public InferShapeContext {
  public:
   CompileTimeInferShapeContext(const OpDescBind& op, const BlockDescBind& block)
       : op_(op), block_(block) {}
@@ -414,7 +405,7 @@ class CompileTimeInferShapeContext : public InferShapeContextBase {
   const BlockDescBind& block_;
 };
 
-class RuntimeInferShapeContext : public InferShapeContextBase {
+class RuntimeInferShapeContext : public InferShapeContext {
  public:
   RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope)
       : op_(op), scope_(scope) {}
@@ -612,7 +603,7 @@ class OperatorWithKernel : public OperatorBase {
                        });
   }
 
-  virtual void InferShape(InferShapeContextBase* ctx) const = 0;
+  virtual void InferShape(InferShapeContext* ctx) const = 0;
 
  protected:
   // indicate kernel DataType by input data. Defaultly all input data must be

paddle/framework/operator_test.cc

@@ -113,7 +113,7 @@ class OpWithKernelTest : public OperatorWithKernel {
   using OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {}
+  void InferShape(framework::InferShapeContext* ctx) const override {}
   DataType IndicateDataType(const ExecutionContext& ctx) const override {
     return DataType::FP32;
   }

paddle/framework/program_desc.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <memory>
 #include <vector>
 #include "paddle/framework/framework.pb.h"
 #include "paddle/platform/macros.h"
@@ -31,8 +32,6 @@ class ProgramDescBind {
 
   BlockDescBind *Block(size_t idx) { return blocks_[idx].get(); }
 
-  std::string DebugString() { return Proto()->DebugString(); }
-
   size_t Size() const { return blocks_.size(); }
 
   ProgramDesc *Proto();

paddle/framework/shape_inference.h

@@ -20,11 +20,11 @@ namespace paddle {
 namespace framework {
 
 // TODO(longfei): Once after both CompileTimeInferShapeContext and
-// RuntimeInferShapeContext get merged, we can rename InferShapeContextBase into
+// RuntimeInferShapeContext get merged, we can rename InferShapeContext into
 // InferShapeContext so to replace the current InferShapeContext.
-class InferShapeContextBase {
+class InferShapeContext {
  public:
-  virtual ~InferShapeContextBase() {}
+  virtual ~InferShapeContext() {}
   virtual bool HasInput(const std::string &name) const = 0;
   virtual bool HasOutput(const std::string &name) const = 0;
 

paddle/framework/tensor.h

@@ -95,6 +95,19 @@ class Tensor {
   template <typename T>
   inline void CopyFrom(const Tensor& src, const platform::Place& dst_place);
 
+  /**
+   * @brief   Copy the content of an external vector to a tensor.
+   *
+   * @param[in] src   The external vector.
+   * @param[in] ctx   The device context contains place where to store.
+   *
+   * * @note    CopyFromVector assumes that the tensor has been resized
+   *            before invoking.
+   */
+  template <typename T>
+  inline void CopyFromVector(const std::vector<T>& src,
+                             const platform::Place& dst_place);
+
   /**
    * @brief   Return the slice of the tensor.
    *

paddle/framework/tensor_impl.h

@@ -123,6 +123,29 @@ inline void Tensor::CopyFrom(const Tensor& src,
 #endif
 }
 
+template <typename T>
+inline void Tensor::CopyFromVector(const std::vector<T>& src,
+                                   const platform::Place& dst_place) {
+  auto src_ptr = static_cast<const void*>(src.data());
+  platform::CPUPlace src_place;
+  auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
+  auto size = src.size() * sizeof(T);
+
+  if (platform::is_cpu_place(dst_place)) {
+    memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr, src_place,
+                 src_ptr, size);
+  }
+#ifdef PADDLE_WITH_CUDA
+  else if (platform::is_gpu_place(dst_place)) {
+    memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr, src_place,
+                 src_ptr, size, 0);
+  }
+  PADDLE_ENFORCE(cudaStreamSynchronize(0),
+                 "cudaStreamSynchronize failed in Tensor CopyFromVector");
+
+#endif
+}
+
 template <typename T>
 inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
   check_memory_size<T>();

paddle/framework/tensor_test.cc

@@ -263,6 +263,93 @@ TEST(Tensor, CopyFrom) {
 #endif
 }
 
+TEST(Tensor, CopyFromVector) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  {
+    std::vector<int> src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9};
+    Tensor cpu_tensor;
+
+    // Copy to CPU Tensor
+    cpu_tensor.Resize(make_ddim({3, 3}));
+    auto cpu_place = new paddle::platform::CPUPlace();
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+
+    // Compare Tensors
+    const int* cpu_ptr = cpu_tensor.data<int>();
+    const int* src_ptr = src_vec.data();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    for (size_t i = 0; i < 9; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+    }
+
+    src_vec.erase(src_vec.begin(), src_vec.begin() + 5);
+    cpu_tensor.Resize(make_ddim({2, 2}));
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+    cpu_ptr = cpu_tensor.data<int>();
+    src_ptr = src_vec.data();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    for (size_t i = 0; i < 5; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+    }
+
+    delete cpu_place;
+  }
+
+#ifdef PADDLE_WITH_CUDA
+  {
+    std::vector<int> src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9};
+    Tensor cpu_tensor;
+    Tensor gpu_tensor;
+    Tensor dst_tensor;
+
+    // Copy to CPU Tensor
+    cpu_tensor.Resize(make_ddim({3, 3}));
+    auto cpu_place = new paddle::platform::CPUPlace();
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+
+    // Copy to GPUTensor
+    gpu_tensor.Resize(make_ddim({3, 3}));
+    auto gpu_place = new paddle::platform::GPUPlace();
+    gpu_tensor.CopyFromVector<int>(src_vec, *gpu_place);
+    // Copy from GPU to CPU tensor for comparison
+    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
+
+    // Compare Tensors
+    const int* src_ptr = src_vec.data();
+    const int* cpu_ptr = cpu_tensor.data<int>();
+    const int* dst_ptr = dst_tensor.data<int>();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    ASSERT_NE(src_ptr, dst_ptr);
+    for (size_t i = 0; i < 9; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+      EXPECT_EQ(src_ptr[i], dst_ptr[i]);
+    }
+
+    src_vec.erase(src_vec.begin(), src_vec.begin() + 5);
+
+    cpu_tensor.Resize(make_ddim({2, 2}));
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+    gpu_tensor.Resize(make_ddim({2, 2}));
+    gpu_tensor.CopyFromVector<int>(src_vec, *gpu_place);
+    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
+
+    src_ptr = src_vec.data();
+    cpu_ptr = cpu_tensor.data<int>();
+    dst_ptr = dst_tensor.data<int>();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    ASSERT_NE(src_ptr, dst_ptr);
+    for (size_t i = 0; i < 5; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+      EXPECT_EQ(src_ptr[i], dst_ptr[i]);
+    }
+
+    delete cpu_place;
+    delete gpu_place;
+  }
+#endif
+}
+
 TEST(Tensor, ReshapeToMatrix) {
   using namespace paddle::framework;
   using namespace paddle::platform;

paddle/gserver/CMakeLists.txt

@@ -60,6 +60,36 @@ if(NOT WITH_PYTHON)
             dataproviders/PyDataProvider.h)
 endif()
 
+if(MOBILE_INFERENCE)
+    # Remove evaluators
+    list(REMOVE_ITEM GSERVER_SOURCES
+         layers/ValidationLayer.cpp
+         evaluators/Evaluator.cpp
+         evaluators/DetectionMAPEvaluator.cpp
+         evaluators/CTCErrorEvaluator.cpp
+         evaluators/ChunkEvaluator.cpp)
+
+    # Remove dataproviders
+    list(REMOVE_ITEM GSERVER_SOURCES
+         dataproviders/DataProvider.cpp
+         dataproviders/MultiDataProvider.cpp
+         dataproviders/ProtoDataProvider.cpp
+         dataproviders/PyDataProvider2.cpp
+         dataproviders/PyDataProvider.cpp)
+
+    # Remove useless gradientmachines
+    list(REMOVE_ITEM GSERVER_SOURCES
+         gradientmachines/MultiNetwork.cpp
+         gradientmachines/RecurrentGradientMachine.cpp
+         gradientmachines/ParallelNeuralNetwork.cpp
+         gradientmachines/GradientMachineMode.cpp
+         gradientmachines/MultiGradientMachine.cpp)
+
+    # Remove useless layers
+    list(REMOVE_ITEM GSERVER_SOURCES
+    	 layers/RecurrentLayerGroup.cpp)
+endif()
+
 if(WITH_GPU)
     cuda_add_library(paddle_gserver ${GSERVER_SOURCES})
 else()

paddle/gserver/gradientmachines/GradientMachine.cpp

@@ -17,12 +17,15 @@ limitations under the License. */
 #include <fstream>
 #include "paddle/utils/Logging.h"
 
+#include "NeuralNetwork.h"
+#include "hl_gpu.h"
+
+#ifndef PADDLE_MOBILE_INFERENCE
 #include "GradientMachineMode.h"
 #include "MultiGradientMachine.h"
 #include "MultiNetwork.h"
-#include "NeuralNetwork.h"
 #include "ParallelNeuralNetwork.h"
-#include "hl_gpu.h"
+#endif
 
 namespace paddle {
 
@@ -30,13 +33,16 @@ GradientMachine* GradientMachine::create(
     const ModelConfig& config,
     int mode,
     const std::vector<ParameterType>& parameterTypes) {
+#ifndef PADDLE_MOBILE_INFERENCE
   if (auto gm = IGradientMachineMode::tryCreateGradientMachine(mode, config)) {
     return gm;
   }
   if (FLAGS_trainer_count > 1) {
     return new MultiGradientMachine(config, FLAGS_use_gpu);
   }
+#endif
   if (FLAGS_trainer_count == 1) {  // single
+#ifndef PADDLE_MOBILE_INFERENCE
     NeuralNetwork* nn;
     if (config.type() == "multi_nn") {
       /* multi submodel calculate, thread(s) will be initialized inside */
@@ -48,6 +54,9 @@ GradientMachine* GradientMachine::create(
       /* single thread calculate */
       nn = NeuralNetwork::create(config);
     }
+#else
+    NeuralNetwork* nn = NeuralNetwork::create(config);
+#endif
     ParamInitCallback testParamInitCb = [](int paramId, Parameter* para) {
       para->enableType(PARAMETER_VALUE);
     };

paddle/gserver/gradientmachines/GradientMachine.h

@@ -20,13 +20,16 @@ limitations under the License. */
 #include "ModelConfig.pb.h"
 #include "TrainerConfig.pb.h"
 #include "paddle/gserver/dataproviders/DataProvider.h"
-#include "paddle/gserver/evaluators/Evaluator.h"
 #include "paddle/gserver/layers/Layer.h"
 #include "paddle/math/Matrix.h"
 #include "paddle/parameter/Parameter.h"
 #include "paddle/parameter/ParameterUpdaterBase.h"
 #include "paddle/utils/Thread.h"
 
+#ifndef PADDLE_MOBILE_INFERENCE
+#include "paddle/gserver/evaluators/Evaluator.h"
+#endif
+
 namespace paddle {
 /**
  * @brief A gradient machine is capable of calculating some outputs given
@@ -147,6 +150,7 @@ public:
 
   virtual void onPassEnd() = 0;
 
+#ifndef PADDLE_MOBILE_INFERENCE
   /**
    * Create an evaluator which can be used for eval()
    */
@@ -156,6 +160,7 @@ public:
    * evaluate using the given evaluator
    */
   virtual void eval(Evaluator* evaluator) const = 0;
+#endif
 
   std::vector<ParameterPtr>& getParameters() { return parameters_; }
 

paddle/gserver/gradientmachines/NeuralNetwork.cpp

@@ -14,15 +14,17 @@ limitations under the License. */
 
 #include "paddle/utils/Util.h"
 
+#include "NeuralNetwork.h"
+#include "hl_gpu.h"
+#include "paddle/gserver/layers/AgentLayer.h"
 #include "paddle/utils/CustomStackTrace.h"
 #include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
 
+#ifndef PADDLE_MOBILE_INFERENCE
 #include "MultiNetwork.h"
-#include "NeuralNetwork.h"
 #include "RecurrentGradientMachine.h"
-#include "hl_gpu.h"
-#include "paddle/gserver/layers/AgentLayer.h"
-#include "paddle/utils/Stat.h"
+#endif
 
 namespace paddle {
 void parameterInitNN(int paramId,
@@ -54,6 +56,7 @@ void parameterInitNN(int paramId,
 }
 
 NeuralNetwork* NeuralNetwork::create(const ModelConfig& config) {
+#ifndef PADDLE_MOBILE_INFERENCE
   if (config.type() == "recurrent_nn") {
     return newNeuralNetwork("root");
   } else if (config.type() == "multi_nn") {
@@ -61,6 +64,9 @@ NeuralNetwork* NeuralNetwork::create(const ModelConfig& config) {
   } else {
     return newNeuralNetwork();
   }
+#else
+  return new NeuralNetwork();
+#endif
 }
 
 std::map<std::string, bool> NeuralNetwork::dllInitMap;
@@ -304,6 +310,8 @@ void NeuralNetwork::onPassEnd() {
   }
 }
 
+#ifndef PADDLE_MOBILE_INFERENCE
+
 class CombinedEvaluator : public Evaluator {
 public:
   void addEvaluator(std::unique_ptr<Evaluator>&& evaluator) {
@@ -466,6 +474,8 @@ Evaluator* NeuralNetwork::makeEvaluator() const {
 
 void NeuralNetwork::eval(Evaluator* evaluator) const { evaluator->eval(*this); }
 
+#endif
+
 void NeuralNetwork::setOutputGrad(const std::vector<Argument>& args) {
   CHECK_GE(outputLayers_.size(), args.size());
   for (size_t i = 0; i < args.size(); ++i) {

paddle/gserver/gradientmachines/NeuralNetwork.h

@@ -97,9 +97,12 @@ public:
 
   virtual void onPassEnd();
 
+#ifndef PADDLE_MOBILE_INFERENCE
   virtual Evaluator* makeEvaluator() const;
 
   virtual void eval(Evaluator* evaluator) const;
+#endif
+
   virtual void resetState();
   virtual void setOutputGrad(const std::vector<Argument>& args);
 

paddle/gserver/layers/Layer.cpp

@@ -15,11 +15,14 @@ limitations under the License. */
 #include "paddle/utils/Util.h"
 
 #include "CostLayer.h"
-#include "ValidationLayer.h"
 #include "paddle/math/SparseMatrix.h"
 #include "paddle/utils/Error.h"
 #include "paddle/utils/Logging.h"
 
+#ifndef PADDLE_MOBILE_INFERENCE
+#include "ValidationLayer.h"
+#endif
+
 DEFINE_bool(log_error_clipping, false, "enable log error clipping or not");
 
 namespace paddle {
@@ -103,10 +106,12 @@ LayerPtr Layer::create(const LayerConfig& config) {
     return LayerPtr(new MultiClassCrossEntropy(config));
   else if (type == "rank-cost")
     return LayerPtr(new RankingCost(config));
+#ifndef PADDLE_MOBILE_INFERENCE
   else if (type == "auc-validation")
     return LayerPtr(new AucValidation(config));
   else if (type == "pnpair-validation")
     return LayerPtr(new PnpairValidation(config));
+#endif
 
   return LayerPtr(registrar_.createByType(config.type(), config));
 }

paddle/gserver/tests/CMakeLists.txt

 # gserver pacakge unittests
 
+if(NOT MOBILE_INFERENCE)
 ################### test_ProtoDataProvider ############
-add_unittest_without_exec(test_ProtoDataProvider
-    test_ProtoDataProvider.cpp)
-
-# test_ProtoDataProvider will mkdir as same name,
-# so if WORKING_DIRECTORY is default directory, then
-# mkdir will get error.
-add_test(NAME test_ProtoDataProvider
-    COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoDataProvider
-    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
+    add_unittest_without_exec(test_ProtoDataProvider
+        test_ProtoDataProvider.cpp)
+
+    # test_ProtoDataProvider will mkdir as same name,
+    # so if WORKING_DIRECTORY is default directory, then
+    # mkdir will get error.
+    add_test(NAME test_ProtoDataProvider
+        COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoDataProvider
+        WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
+endif()
 
 ################# test_LayerGrad #######################
 add_unittest_without_exec(test_LayerGrad
@@ -98,9 +100,11 @@ add_unittest_without_exec(test_KmaxSeqScore
 add_test(NAME test_KmaxSeqScore
     COMMAND test_KmaxSeqScore)
 
+if(NOT MOBILE_INFERENCE)
 ################## test_Evaluator #######################
-add_unittest(test_Evaluator
-    test_Evaluator.cpp)
+    add_unittest(test_Evaluator
+        test_Evaluator.cpp)
+endif()
 
 ################ test_LinearChainCRF ####################
 add_simple_unittest(test_LinearChainCRF)
@@ -131,27 +135,31 @@ if(NOT WITH_DOUBLE)
         WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
 endif()
 
+if(NOT MOBILE_INFERENCE)
 ############### test_RecurrentGradientMachine ###############
-# TODO(yuyang18): There is some bug in test_RecurrentGradientMachine
-# I will fix it.
-add_unittest_without_exec(test_RecurrentGradientMachine
-    test_RecurrentGradientMachine.cpp)
-add_test(NAME test_RecurrentGradientMachine
-    COMMAND .set_python_path.sh -d
-            ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
-            ${CMAKE_CURRENT_BINARY_DIR}/test_RecurrentGradientMachine
-    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
-
-add_unittest_without_exec(test_NetworkCompare
-    test_NetworkCompare.cpp)
-if(WITH_GPU)
-    add_test(NAME test_NetworkCompare
-        COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=true
-        WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
-else()
-    add_test(NAME test_NetworkCompare
-        COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=false
-        WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
+  # TODO(yuyang18): There is some bug in test_RecurrentGradientMachine
+  # I will fix it.
+  add_unittest_without_exec(test_RecurrentGradientMachine
+      test_RecurrentGradientMachine.cpp)
+  add_test(NAME test_RecurrentGradientMachine
+      COMMAND .set_python_path.sh -d
+              ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
+              ${CMAKE_CURRENT_BINARY_DIR}/test_RecurrentGradientMachine
+      WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
+endif()
+
+if(NOT MOBILE_INFERENCE)
+    add_unittest_without_exec(test_NetworkCompare
+        test_NetworkCompare.cpp)
+    if(WITH_GPU)
+        add_test(NAME test_NetworkCompare
+            COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=true
+            WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
+    else()
+        add_test(NAME test_NetworkCompare
+            COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=false
+            WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
+    endif()
 endif()
 
 

paddle/gserver/tests/LayerGradUtil.h

@@ -15,7 +15,6 @@ limitations under the License. */
 #pragma once
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/trainer/Trainer.h"
 
 #include "paddle/testing/TestUtil.h"
 using namespace std;  // NOLINT

paddle/gserver/tests/test_ActivationGrad.cpp

@@ -17,7 +17,6 @@ limitations under the License. */
 #include <vector>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/trainer/Trainer.h"
 
 #include "LayerGradUtil.h"
 #include "paddle/testing/TestUtil.h"

paddle/gserver/tests/test_BatchNorm.cpp

@@ -17,7 +17,6 @@ limitations under the License. */
 #include <vector>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/trainer/Trainer.h"
 #include "paddle/utils/GlobalConstants.h"
 
 #include "LayerGradUtil.h"

paddle/gserver/tests/test_CRFLayerGrad.cpp

@@ -16,7 +16,6 @@ limitations under the License. */
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
 #include "paddle/gserver/layers/LinearChainCRF.h"
-#include "paddle/trainer/Trainer.h"
 
 #include "LayerGradUtil.h"
 #include "paddle/testing/TestUtil.h"

paddle/gserver/tests/test_ConvTrans.cpp

@@ -18,7 +18,6 @@ limitations under the License. */
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
 #include "paddle/math/MathUtils.h"
-#include "paddle/trainer/Trainer.h"
 #include "paddle/utils/GlobalConstants.h"
 
 #include "LayerGradUtil.h"

paddle/gserver/tests/test_ConvUnify.cpp

@@ -18,7 +18,6 @@ limitations under the License. */
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
 #include "paddle/math/MathUtils.h"
-#include "paddle/trainer/Trainer.h"
 #include "paddle/utils/GlobalConstants.h"
 
 #include "LayerGradUtil.h"

paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp

@@ -18,7 +18,6 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/trainer/Trainer.h"
 
 #include "LayerGradUtil.h"
 #include "paddle/testing/TestUtil.h"

paddle/gserver/tests/test_KmaxSeqScore.cpp

@@ -18,7 +18,6 @@ limitations under the License. */
 #include <vector>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/trainer/Trainer.h"
 #include "paddle/utils/GlobalConstants.h"
 
 #include "LayerGradUtil.h"

paddle/gserver/tests/test_LayerGrad.cpp

@@ -21,7 +21,6 @@ limitations under the License. */
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
 #include "paddle/math/MathUtils.h"
-#include "paddle/trainer/Trainer.h"
 
 #include "LayerGradUtil.h"
 #include "paddle/testing/TestUtil.h"

paddle/gserver/tests/test_SelectiveFCLayer.cpp

@@ -24,7 +24,6 @@ limitations under the License. */
 #include "paddle/gserver/layers/Layer.h"
 #include "paddle/gserver/layers/SelectiveFullyConnectedLayer.h"
 #include "paddle/math/CpuSparseMatrix.h"
-#include "paddle/trainer/Trainer.h"
 
 using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT

paddle/gserver/tests/test_SeqSliceLayerGrad.cpp

@@ -15,7 +15,6 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/trainer/Trainer.h"
 
 #include "LayerGradUtil.h"
 #include "paddle/testing/TestUtil.h"

paddle/operators/accuracy_op.cc

@@ -22,7 +22,7 @@ class AccuracyOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Inference"),
                    "Input(Inference) of AccuracyOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Label"),

paddle/operators/activation_op.cc

@@ -22,7 +22,7 @@ class ActivationOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     ctx->SetOutputDim("Y", ctx->GetInputDim("X"));
     ctx->ShareLoD("X", /*->*/ "Y");
   }
@@ -33,7 +33,7 @@ class ActivationOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Y"));
   }
 };
@@ -201,6 +201,19 @@ class SoftReluOpMaker : public framework::OpProtoAndCheckerMaker {
   }
 };
 
+template <typename AttrType>
+class Relu6OpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  Relu6OpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "Input of Relu6 operator");
+    AddOutput("Y", "Output of Relu6 operator");
+    AddComment("Relu6 activation operator, relu6 = min(max(0, x), 6)");
+    AddAttr<AttrType>("threshold", "The threshold value of Relu6")
+        .SetDefault(static_cast<AttrType>(6));
+  }
+};
+
 template <typename AttrType>
 class PowOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
@@ -276,6 +289,9 @@ REGISTER_OP(leaky_relu, ops::ActivationOp, ops::LeakyReluOpMaker<float>,
 REGISTER_OP(soft_relu, ops::ActivationOp, ops::SoftReluOpMaker<float>,
             soft_relu_grad, ops::ActivationOpGrad);
 
+REGISTER_OP(relu6, ops::ActivationOp, ops::Relu6OpMaker<float>, relu6_grad,
+            ops::ActivationOpGrad);
+
 REGISTER_OP(pow, ops::ActivationOp, ops::PowOpMaker<float>, pow_grad,
             ops::ActivationOpGrad);
 
@@ -285,11 +301,9 @@ REGISTER_OP(stanh, ops::ActivationOp, ops::STanhOpMaker<float>, stanh_grad,
 #define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor)        \
   REGISTER_OP_CPU_KERNEL(                                                      \
       act_type,                                                                \
-      paddle::operators::ActivationKernel<paddle::platform::CPUPlace,          \
-                                          paddle::operators::functor<float>>); \
+      ops::ActivationKernel<paddle::platform::CPUPlace, ops::functor<float>>); \
   REGISTER_OP_CPU_KERNEL(act_type##_grad,                                      \
-                         paddle::operators::ActivationGradKernel<              \
-                             paddle::platform::CPUPlace,                       \
-                             paddle::operators::grad_functor<float>>);
+                         ops::ActivationGradKernel<paddle::platform::CPUPlace, \
+                                                   ops::grad_functor<float>>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL);

paddle/operators/activation_op.cu

@@ -15,14 +15,14 @@
 #define EIGEN_USE_GPU
 #include "paddle/operators/activation_op.h"
 
+namespace ops = paddle::operators;
+
 #define REGISTER_ACTIVATION_GPU_KERNEL(act_type, functor, grad_functor)        \
   REGISTER_OP_GPU_KERNEL(                                                      \
       act_type,                                                                \
-      paddle::operators::ActivationKernel<paddle::platform::GPUPlace,          \
-                                          paddle::operators::functor<float>>); \
+      ops::ActivationKernel<paddle::platform::GPUPlace, ops::functor<float>>); \
   REGISTER_OP_GPU_KERNEL(act_type##_grad,                                      \
-                         paddle::operators::ActivationGradKernel<              \
-                             paddle::platform::GPUPlace,                       \
-                             paddle::operators::grad_functor<float>>);
+                         ops::ActivationGradKernel<paddle::platform::GPUPlace, \
+                                                   ops::grad_functor<float>>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_GPU_KERNEL);

paddle/operators/activation_op.h

@@ -280,6 +280,36 @@ struct BReluGradFunctor : public BaseActivationFunctor<T> {
   }
 };
 
+// relu6(x) = min(max(0, x), 6)
+template <typename T>
+struct Relu6Functor : public BaseActivationFunctor<T> {
+  float threshold;
+
+  // NOTE: Explicit hides the `BaseActivationFunctor<T>::GetAttrs`
+  // not polymorphism for speed.
+  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
+    return {{"threshold", &threshold}};
+  }
+
+  template <typename Device, typename X, typename Y>
+  void operator()(Device d, X x, Y y) const {
+    y.device(d) = x.cwiseMax(static_cast<T>(0)).cwiseMin(threshold);
+  }
+};
+
+template <typename T>
+struct Relu6GradFunctor : public BaseActivationFunctor<T> {
+  float threshold;
+  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
+    return {{"threshold", &threshold}};
+  }
+  template <typename Device, typename X, typename Y, typename dY, typename dX>
+  void operator()(Device d, X x, Y y, dY dy, dX dx) const {
+    dx.device(d) =
+        dy * ((x > static_cast<T>(0)) * (x < threshold)).template cast<T>();
+  }
+};
+
 // softsign(x) = x / (1 + |x|)
 template <typename T>
 struct SoftsignFunctor : public BaseActivationFunctor<T> {
@@ -425,5 +455,6 @@ struct STanhGradFunctor : public BaseActivationFunctor<T> {
   __macro(pow, PowFunctor, PowGradFunctor);                      \
   __macro(stanh, STanhFunctor, STanhGradFunctor);                \
   __macro(softsign, SoftsignFunctor, SoftsignGradFunctor);       \
+  __macro(relu6, Relu6Functor, Relu6GradFunctor);                \
   __macro(leaky_relu, LeakyReluFunctor, LeakyReluGradFunctor);   \
   __macro(tanh_shrink, TanhShrinkFunctor, TanhShrinkGradFunctor)

paddle/operators/adadelta_op.cc

@@ -22,7 +22,7 @@ class AdadeltaOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Param"),
                    "Input(Param) of AdadeltaOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Grad"),

paddle/operators/adagrad_op.cc

@@ -22,7 +22,7 @@ class AdagradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Param"),
                    "Input(Param) of AdagradOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Grad"),

paddle/operators/adamax_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/adamax_op.h"
+
+namespace paddle {
+namespace operators {
+
+class AdamaxOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Param"),
+                   "Input(Param) of AdamaxOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Grad"),
+                   "Input(Grad) of AdamaxOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Moment"),
+                   "Input(Moment) of AdamaxOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("InfNorm"),
+                   "Input(InfNorm) of AdamaxOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("LearningRate"),
+                   "Input(LearningRate) of AdamaxOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Beta1Pow"),
+                   "Input(Beta1Pow) of AdamaxOp should not be null.");
+
+    PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
+                   "Output(ParamOut) of AdamaxOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("MomentOut"),
+                   "Output(MomentOut) of AdamaxOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("InfNormOut"),
+                   "Output(InfNormOut) of AdamaxOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Beta1PowOut"),
+                   "Output(Beta1PowOut) of AdamaxOp should not be null.");
+
+    auto lr_dims = ctx->GetInputDim("LearningRate");
+    PADDLE_ENFORCE_EQ(framework::product(lr_dims), 1,
+                      "Learning rate should have 1 dimension");
+    auto beta1_pow_dims = ctx->GetInputDim("Beta1Pow");
+    PADDLE_ENFORCE_EQ(framework::product(beta1_pow_dims), 1,
+                      "Beta1 power accumulator should have 1 dimension");
+    auto param_dims = ctx->GetInputDim("Param");
+    PADDLE_ENFORCE_EQ(
+        param_dims, ctx->GetInputDim("Grad"),
+        "Param and Grad input of AdamaxOp should have same dimension");
+    PADDLE_ENFORCE_EQ(
+        param_dims, ctx->GetInputDim("Moment"),
+        "Param and Moment input of AdamaxOp should have same dimension");
+    PADDLE_ENFORCE_EQ(
+        param_dims, ctx->GetInputDim("InfNorm"),
+        "Param and InfNorm input of AdamaxOp should have same dimension");
+
+    ctx->SetOutputDim("ParamOut", param_dims);
+    ctx->SetOutputDim("MomentOut", param_dims);
+    ctx->SetOutputDim("InfNormOut", param_dims);
+    ctx->SetOutputDim("Beta1PowOut", beta1_pow_dims);
+  }
+};
+
+class AdamaxOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  AdamaxOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Param", "(Tensor) Input parameter");
+    AddInput("Grad", "(Tensor) Input gradient");
+    AddInput("LearningRate", "(Tensor) Learning rate");
+    AddInput("Moment", "(Tensor) First moment");
+    AddInput("InfNorm",
+             "(Tensor) "
+             "Input exponentially weighted infinity norm");
+    AddInput("Beta1Pow", "(Tensor) Input beta1 power accumulator");
+
+    AddOutput("ParamOut", "(Tensor) Output parameter");
+    AddOutput("MomentOut", "(Tensor) Output first moment");
+    AddOutput("InfNormOut",
+              "(Tensor) "
+              "Output exponentially weighted infinity norm");
+    AddOutput("Beta1PowOut", "(Tensor) Output beta1 power accumulator");
+
+    AddAttr<float>("beta1",
+                   "(float, default 0.9) "
+                   "Exponential decay rate for the "
+                   "1st moment estimates.")
+        .SetDefault(0.9f);
+    AddAttr<float>("beta2",
+                   "(float, default 0.999) "
+                   "exponential decay rate for the weighted "
+                   "infinity norm estimates.")
+        .SetDefault(0.999f);
+    AddAttr<float>("epsilon",
+                   "(float, default 1.0e-8) "
+                   "Constant for numerical stability")
+        .SetDefault(1.0e-8f);
+    AddComment(R"DOC(
+Adamax Updates Operator.
+
+This implements the Adamax optimizer from Section 7 of the Adam
+paper[1]. Adamax is a variant of the
+Adam algorithm based on the infinity norm.
+
+Adamax updates:
+
+moment_out = beta1 * moment + (1 - beta1) * grad
+inf_norm_out = max(beta2 * inf_norm + epsilon, abs(grad))
+beta1_pow_out = beta1_pow * beta1
+learning_rate_t = learning_rate/(1 - beta1_pow_out)
+param_out = param - learning_rate_t * moment_out/inf_norm_out
+
+The original paper does not have an epsilon attribute.
+However, it is added here for numerical stability
+by preventing divide by 0.
+
+References:
+  [1] Adam: A Method for Stochastic Optimization
+      (https://arxiv.org/abs/1412.6980)
+
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(adamax, ops::AdamaxOp, ops::AdamaxOpMaker);
+REGISTER_OP_CPU_KERNEL(adamax,
+                       ops::AdamaxOpKernel<paddle::platform::CPUPlace, float>);

paddle/operators/adamax_op.cu

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

paddle/operators/adamax_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class AdamaxOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
+    auto moment_out_tensor = ctx.Output<framework::Tensor>("MomentOut");
+    auto inf_norm_out_tensor = ctx.Output<framework::Tensor>("InfNormOut");
+    auto beta1_pow_out_tensor = ctx.Output<framework::Tensor>("Beta1PowOut");
+
+    param_out_tensor->mutable_data<T>(ctx.GetPlace());
+    moment_out_tensor->mutable_data<T>(ctx.GetPlace());
+    inf_norm_out_tensor->mutable_data<T>(ctx.GetPlace());
+    beta1_pow_out_tensor->mutable_data<T>(ctx.GetPlace());
+
+    float beta1 = ctx.Attr<float>("beta1");
+    float beta2 = ctx.Attr<float>("beta2");
+    float epsilon = ctx.Attr<float>("epsilon");
+
+    auto param = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("Param"));
+    auto grad = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("Grad"));
+    auto moment = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("Moment"));
+    auto inf_norm = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("InfNorm"));
+    auto lr = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("LearningRate"));
+    auto beta1_pow = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("Beta1Pow"));
+    auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
+    auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
+    auto inf_norm_out =
+        framework::EigenVector<T>::Flatten(*inf_norm_out_tensor);
+    auto beta1_pow_out =
+        framework::EigenVector<T>::Flatten(*beta1_pow_out_tensor);
+    auto place = ctx.GetEigenDevice<Place>();
+
+    moment_out.device(place) = beta1 * moment + (1 - beta1) * grad;
+    inf_norm_out.device(place) =
+        grad.abs().cwiseMax((beta2 * inf_norm) + epsilon);
+    beta1_pow_out.device(place) = beta1_pow * beta1;
+    auto lr_t = lr / (1 - beta1_pow_out);
+    Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
+    param_out.device(place) =
+        param - lr_t.broadcast(m_dsize) * (moment_out / inf_norm_out);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/clip_op.cc

@@ -22,7 +22,7 @@ class ClipOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of ClipOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
@@ -61,7 +61,7 @@ class ClipOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                    "Input(Out@GRAD) should not be null");

paddle/operators/concat_op.cc

@@ -24,7 +24,7 @@ class ConcatOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE_GE(ctx->Inputs("X").size(), 1UL,
                       "Inputs(X) of ConcatOp should be empty.")
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
@@ -83,7 +83,7 @@ class ConcatOpGrad : public framework::OperatorWithKernel {
       : OperatorWithKernel(type, inputs, outputs, attrs) {}
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X"));
   }
 };

paddle/operators/conv2d_op.cc

@@ -27,7 +27,7 @@ class Conv2DOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Input"),
                    "Input(Input) of Conv2DOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Filter"),
@@ -106,7 +106,7 @@ class Conv2DOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     auto in_dims = ctx->GetInputDim("Input");
     auto filter_dims = ctx->GetInputDim("Filter");
     if (ctx->HasOutput(framework::GradVarName("Input"))) {

paddle/operators/cos_sim_op.cc

@@ -24,7 +24,7 @@ class CosSimOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     // notnull check
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of CosSimOp should not be null.");
@@ -98,7 +98,7 @@ class CosSimOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     // notnull check
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) must not be null.");

paddle/operators/crop_op.cc

@@ -25,7 +25,7 @@ class CropOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of CropOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
@@ -115,7 +115,7 @@ class CropOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                    "Input(Out@GRAD) should not be null");

paddle/operators/cross_entropy_op.cc

@@ -22,7 +22,7 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
     PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
     PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null.");
@@ -60,7 +60,7 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
     PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")),

paddle/operators/dropout_op.cc

@@ -24,7 +24,7 @@ class DropoutOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null.");
     PADDLE_ENFORCE_GE(ctx->Attrs().Get<float>("dropout_prob"), 0);
     PADDLE_ENFORCE_LE(ctx->Attrs().Get<float>("dropout_prob"), 1);
@@ -70,7 +70,7 @@ class DropoutOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE_EQ(ctx->Attrs().Get<bool>("is_training"), 1,
                       "GradOp is only callable when is_training is true");
 

paddle/operators/elementwise_op.h

@@ -25,7 +25,7 @@ class ElementwiseOp : public framework::OperatorWithKernel {
 
  protected:
   using Tensor = framework::Tensor;
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of elementwise op should not be null");
     PADDLE_ENFORCE(ctx->HasInput("Y"),
@@ -106,7 +106,7 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel {
   using Tensor = framework::Tensor;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
     PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),

paddle/operators/fill_zeros_like_op.cc

@@ -22,7 +22,7 @@ class FillZerosLikeOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of FillZerosLikeOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Y"),

paddle/operators/gather_op.cc

@@ -23,7 +23,7 @@ class GatherOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of GatherOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Index"),
@@ -51,7 +51,7 @@ class GatherGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
   }
 

paddle/operators/gaussian_random_op.cc

@@ -43,7 +43,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of GaussianRandomOp should not be null.");
     auto dims = ctx->Attrs().Get<std::vector<int>>("dims");

paddle/operators/lookup_table_op.cc

@@ -22,7 +22,7 @@ class LookupTableOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("W"),
                    "Input(W) of LookupTableOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Ids"),
@@ -70,7 +70,7 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     auto table_dims = ctx->GetInputDim("W");
     ctx->SetOutputDim(framework::GradVarName("W"), table_dims);
   }

paddle/operators/lstm_unit_op.cc

@@ -22,7 +22,7 @@ class LstmUnitOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of LSTM should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("C_prev"),
                    "Input(C_prev) of LSTM should not be null.");
@@ -77,7 +77,7 @@ class LstmUnitGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("C")),
                    "Input(C@GRAD) should not be null");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("H")),

paddle/operators/mean_op.cc

@@ -22,7 +22,7 @@ class MeanOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of MeanOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
@@ -47,7 +47,7 @@ class MeanGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
   }
 };

paddle/operators/minus_op.cc

@@ -26,7 +26,7 @@ class MinusOp : public framework::OperatorWithKernel {
       : OperatorWithKernel(type, inputs, outputs, attrs) {}
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of MinusOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Y"),

paddle/operators/modified_huber_loss_op.cc

@@ -22,7 +22,7 @@ class ModifiedHuberLossOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "X must be initialized.");
     PADDLE_ENFORCE(ctx->HasInput("Y"), "Y must be initialized.");
 
@@ -74,7 +74,7 @@ class ModifiedHuberLossGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "X must be initialized.");
     PADDLE_ENFORCE(ctx->HasInput("Y"), "Y must be initialized.");
     PADDLE_ENFORCE(ctx->HasInput("IntermediateVal"),

paddle/operators/mul_op.cc

@@ -24,7 +24,7 @@ class MulOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of MulOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) of MulOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
@@ -97,7 +97,7 @@ class MulOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
     PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),

paddle/operators/multiplex_op.cc

@@ -24,7 +24,7 @@ class MultiplexOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Ids"), "Input(Ids) shouldn't be null.");
     PADDLE_ENFORCE(!ctx->Inputs("X").empty(),
                    "MultiInput(X) shouldn't be empty.");
@@ -90,7 +90,7 @@ class MultiplexGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(!ctx->Inputs("X").empty(), "Input(X) should not be null.");
     PADDLE_ENFORCE(!ctx->Outputs(framework::GradVarName("X")).empty(),
                    "Output(X@Grad) should not be null.");

paddle/operators/net_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <set>
 #include "paddle/framework/framework.pb.h"
 #include "paddle/framework/op_registry.h"
 

paddle/operators/pad_op.cc

@@ -24,7 +24,7 @@ class PadOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of PadOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of PadOp should not be null.");
@@ -98,7 +98,7 @@ class PadOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                    "Input(Out@GRAD) should not be null");

paddle/operators/pool_op.cc

@@ -27,7 +27,7 @@ class PoolOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "X(Input) of Pooling should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
@@ -74,7 +74,7 @@ class PoolOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "X(Input) of Pooling should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),

paddle/operators/prelu_op.cc

@@ -26,7 +26,7 @@ class PReluOp : public framework::OperatorWithKernel {
       : OperatorWithKernel(type, inputs, outputs, attrs) {}
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
     PADDLE_ENFORCE(ctx->HasInput("Alpha"), "Input(Alpha) should not be null");
     PADDLE_ENFORCE(product(ctx->GetInputDim("Alpha")) == 1,
@@ -63,7 +63,7 @@ class PReluGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null.");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                    "Input(Out@GRAD) should not be null");

paddle/operators/rank_loss_op.cc

@@ -25,7 +25,7 @@ class RankLossOp : public framework::OperatorWithKernel {
       : OperatorWithKernel(type, inputs, outputs, attrs) {}
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     // input check
     PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) shouldn't be null");
     PADDLE_ENFORCE(ctx->HasInput("Left"), "Input(Left) shouldn't be null");
@@ -90,7 +90,7 @@ class RankLossGradOp : public framework::OperatorWithKernel {
       : OperatorWithKernel(type, inputs, outputs, attrs) {}
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) shouldn't be null.");
     PADDLE_ENFORCE(ctx->HasInput("Left"), "Input(Left) shouldn't be null.");
     PADDLE_ENFORCE(ctx->HasInput("Right"), "Input(Right) shouldn't be null.");

paddle/operators/reduce_op.cc

@@ -24,7 +24,7 @@ class ReduceOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of ReduceOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
@@ -58,7 +58,7 @@ class ReduceGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null.");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                    "Input(Out@GRAD) should not be null.");
@@ -168,36 +168,22 @@ namespace ops = paddle::operators;
 
 REGISTER_OP(reduce_sum, ops::ReduceOp, ops::ReduceSumOpMaker, reduce_sum_grad,
             ops::ReduceGradOp);
-REGISTER_OP_CPU_KERNEL(
-    reduce_sum,
-    ops::ReduceKernel<paddle::platform::CPUPlace, float, ops::SumFunctor>);
-REGISTER_OP_CPU_KERNEL(reduce_sum_grad,
-                       ops::ReduceGradKernel<paddle::platform::CPUPlace, float,
-                                             ops::SumGradFunctor>);
 
 REGISTER_OP(reduce_mean, ops::ReduceOp, ops::ReduceMeanOpMaker,
             reduce_mean_grad, ops::ReduceGradOp);
-REGISTER_OP_CPU_KERNEL(
-    reduce_mean,
-    ops::ReduceKernel<paddle::platform::CPUPlace, float, ops::MeanFunctor>);
-REGISTER_OP_CPU_KERNEL(reduce_mean_grad,
-                       ops::ReduceGradKernel<paddle::platform::CPUPlace, float,
-                                             ops::MeanGradFunctor>);
 
 REGISTER_OP(reduce_max, ops::ReduceOp, ops::ReduceMaxOpMaker, reduce_max_grad,
             ops::ReduceGradOp);
-REGISTER_OP_CPU_KERNEL(
-    reduce_max,
-    ops::ReduceKernel<paddle::platform::CPUPlace, float, ops::MaxFunctor>);
-REGISTER_OP_CPU_KERNEL(reduce_max_grad,
-                       ops::ReduceGradKernel<paddle::platform::CPUPlace, float,
-                                             ops::MaxOrMinGradFunctor>);
-
-REGISTER_OP(reduce_min, ops::ReduceOp, ops::ReduceMaxOpMaker, reduce_min_grad,
+
+REGISTER_OP(reduce_min, ops::ReduceOp, ops::ReduceMinOpMaker, reduce_min_grad,
             ops::ReduceGradOp);
-REGISTER_OP_CPU_KERNEL(
-    reduce_min,
-    ops::ReduceKernel<paddle::platform::CPUPlace, float, ops::MinFunctor>);
-REGISTER_OP_CPU_KERNEL(reduce_min_grad,
-                       ops::ReduceGradKernel<paddle::platform::CPUPlace, float,
-                                             ops::MaxOrMinGradFunctor>);
+
+#define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor)     \
+  REGISTER_OP_CPU_KERNEL(                                                  \
+      reduce_type,                                                         \
+      ops::ReduceKernel<paddle::platform::CPUPlace, float, ops::functor>); \
+  REGISTER_OP_CPU_KERNEL(reduce_type##_grad,                               \
+                         ops::ReduceGradKernel<paddle::platform::CPUPlace, \
+                                               float, ops::grad_functor>);
+
+FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_CPU_KERNEL);

paddle/operators/reduce_op.cu

@@ -17,30 +17,12 @@
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(
-    reduce_sum,
-    ops::ReduceKernel<paddle::platform::GPUPlace, float, ops::SumFunctor>);
-REGISTER_OP_GPU_KERNEL(reduce_sum_grad,
-                       ops::ReduceGradKernel<paddle::platform::GPUPlace, float,
-                                             ops::SumGradFunctor>);
-
-REGISTER_OP_GPU_KERNEL(
-    reduce_mean,
-    ops::ReduceKernel<paddle::platform::GPUPlace, float, ops::MeanFunctor>);
-REGISTER_OP_GPU_KERNEL(reduce_mean_grad,
-                       ops::ReduceGradKernel<paddle::platform::GPUPlace, float,
-                                             ops::MeanGradFunctor>);
-
-REGISTER_OP_GPU_KERNEL(
-    reduce_max,
-    ops::ReduceKernel<paddle::platform::GPUPlace, float, ops::MaxFunctor>);
-REGISTER_OP_GPU_KERNEL(reduce_max_grad,
-                       ops::ReduceGradKernel<paddle::platform::GPUPlace, float,
-                                             ops::MaxOrMinGradFunctor>);
-
-REGISTER_OP_GPU_KERNEL(
-    reduce_min,
-    ops::ReduceKernel<paddle::platform::GPUPlace, float, ops::MinFunctor>);
-REGISTER_OP_GPU_KERNEL(reduce_min_grad,
-                       ops::ReduceGradKernel<paddle::platform::GPUPlace, float,
-                                             ops::MaxOrMinGradFunctor>);
+#define REGISTER_REDUCE_GPU_KERNEL(reduce_type, functor, grad_functor)     \
+  REGISTER_OP_GPU_KERNEL(                                                  \
+      reduce_type,                                                         \
+      ops::ReduceKernel<paddle::platform::GPUPlace, float, ops::functor>); \
+  REGISTER_OP_GPU_KERNEL(reduce_type##_grad,                               \
+                         ops::ReduceGradKernel<paddle::platform::GPUPlace, \
+                                               float, ops::grad_functor>);
+
+FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_GPU_KERNEL);

paddle/operators/reduce_op.h

@@ -198,3 +198,9 @@ class ReduceGradKernel : public framework::OpKernel<T> {
 
 }  // namespace operators
 }  // namespace paddle
+
+#define FOR_EACH_KERNEL_FUNCTOR(__macro)                \
+  __macro(reduce_sum, SumFunctor, SumGradFunctor);      \
+  __macro(reduce_mean, MeanFunctor, MeanGradFunctor);   \
+  __macro(reduce_max, MaxFunctor, MaxOrMinGradFunctor); \
+  __macro(reduce_min, MinFunctor, MaxOrMinGradFunctor);

paddle/operators/reshape_op.cc

@@ -26,7 +26,7 @@ class ReshapeOp : public framework::OperatorWithKernel {
       : OperatorWithKernel(type, inputs, outputs, attrs) {}
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     // input check
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of ReshapeOp should not be null.");
@@ -94,7 +94,7 @@ class ReshapeGradOp : public framework::OperatorWithKernel {
       : OperatorWithKernel(type, inputs, outputs, attrs) {}
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) shouldn't be null.");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                    "Input(Out@GRAD) shouldn't be null.");

paddle/operators/rmsprop_op.cc

@@ -22,7 +22,7 @@ class RmspropOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Param"),
                    "Input(Param) of RmspropOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("MeanSquare"),

paddle/operators/scale_op.cc

@@ -26,7 +26,7 @@ class ScaleOp : public framework::OperatorWithKernel {
       : OperatorWithKernel(type, inputs, outputs, attrs) {}
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of ScaleOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),

paddle/operators/scatter_op.cc

@@ -23,7 +23,7 @@ class ScatterOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Ref"),
                    "Input(Ref) of ScatterOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Index"),
@@ -60,7 +60,7 @@ class ScatterGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     ctx->SetOutputDim(framework::GradVarName("Updates"),
                       ctx->GetInputDim("Updates"));
     ctx->SetOutputDim(framework::GradVarName("Ref"), ctx->GetInputDim("Ref"));

paddle/operators/sequence_pool_op.cc

@@ -22,7 +22,7 @@ class SequencePoolOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of SequencePoolOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
@@ -74,7 +74,7 @@ class SequencePoolGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                    "Gradient of Out should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("X"), "The input X should not be null.");

paddle/operators/sequence_softmax_op.cc

@@ -22,7 +22,7 @@ class SequenceSoftmaxOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of SequenceSoftmaxOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
@@ -67,7 +67,7 @@ class SequenceSoftmaxGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Out"),
                    "Input(Out) of SequenceSoftmaxGradOp should not be null.");
     PADDLE_ENFORCE(

paddle/operators/sgd_op.cc

@@ -22,7 +22,7 @@ class SGDOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Param"),
                    "Input(Param) of SGDOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Grad"),

paddle/operators/sigmoid_cross_entropy_with_logits_op.cc

@@ -24,7 +24,7 @@ class SigmoidCrossEntropyWithLogitsOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
     PADDLE_ENFORCE(ctx->HasInput("Labels"),
                    "Input(Labels) should be not null.");
@@ -53,7 +53,7 @@ class SigmoidCrossEntropyWithLogitsGradOp
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
     PADDLE_ENFORCE(ctx->HasInput("Labels"),
                    "Input(Labels) should be not null.");

paddle/operators/smooth_l1_loss_op.cc

@@ -22,7 +22,7 @@ class SmoothL1LossOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "X must be initialized.");
     PADDLE_ENFORCE(ctx->HasInput("Y"), "Y must be initialized.");
 
@@ -94,7 +94,7 @@ class SmoothL1LossGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     auto in_dims = ctx->GetInputDim("X");
     auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
 

paddle/operators/softmax_op.cc

@@ -22,7 +22,7 @@ class SoftmaxOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of SoftmaxOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Y"),
@@ -69,7 +69,7 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should be not null.");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")),
                    "Input(Y@GRAD) should be not null.");

paddle/operators/softmax_with_cross_entropy_op.cc

@@ -83,7 +83,7 @@ class SoftmaxWithCrossEntropyOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Logits"),
                    "Input(Logits) should be not null.");
     PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
@@ -128,7 +128,7 @@ class SoftmaxWithCrossEntropyOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Loss")),
                    "Input(Loss@Grad) should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Softmax"),

paddle/operators/split_op.cc

@@ -24,7 +24,7 @@ class SplitOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of SplitOp should not be null.");
     PADDLE_ENFORCE_GE(ctx->Outputs("Out").size(), 1UL,

paddle/operators/squared_l2_distance_op.cc

@@ -22,7 +22,7 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of SquaredL2DistanceOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Y"),
@@ -86,7 +86,7 @@ class SquaredL2DistanceGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                    "Gradient of Out should not be null");
     auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));

paddle/operators/sum_op.cc

@@ -22,7 +22,7 @@ class SumOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInputs("X"), "Inputs(X) should not be null");
     auto x_dims = ctx->GetInputsDim("X");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),

paddle/operators/top_k_op.cc

@@ -22,7 +22,7 @@ class TopkOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase *ctx) const override {
+  void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of TopkOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),

paddle/operators/transpose_op.cc

@@ -24,7 +24,7 @@ class TransposeOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
     PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null");
     auto x_dims = ctx->GetInputDim("X");
@@ -93,7 +93,7 @@ class TransposeOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                    "Input(Out@GRAD) should not be null");

paddle/operators/uniform_random_op.cc

@@ -47,7 +47,7 @@ class UniformRandomOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of UniformRandomOp should not be null.");
 

paddle/pybind/protobuf.cc

@@ -117,7 +117,6 @@ void BindProgramDesc(py::module &m) {
       .def("append_block", &ProgramDescBind::AppendBlock,
            py::return_value_policy::reference)
       .def("block", &ProgramDescBind::Block, py::return_value_policy::reference)
-      .def("__str__", &ProgramDescBind::DebugString)
       .def("num_blocks", &ProgramDescBind::Size);
 }
 
@@ -191,8 +190,6 @@ void BindOpDesc(py::module &m) {
       .def("output", &OpDescBind::Output)
       .def("output_names", &OpDescBind::OutputNames)
       .def("set_output", &OpDescBind::SetOutput)
-      .def("__str__", &OpDescBind::DebugString)
-      .def("__repr__", &OpDescBind::DebugString)
       .def("has_attr", &OpDescBind::HasAttr)
       .def("attr_type", &OpDescBind::GetAttrType)
       .def("attr_names", &OpDescBind::AttrNames)

python/paddle/v2/event.py

@@ -10,7 +10,8 @@ There are:
 * EndPass
 """
 __all__ = [
-    'EndIteration', 'BeginIteration', 'BeginPass', 'EndPass', 'TestResult'
+    'EndIteration', 'BeginIteration', 'BeginPass', 'EndPass', 'TestResult',
+    'EndForwardBackward'
 ]
 
 
@@ -73,6 +74,17 @@ class BeginIteration(object):
         self.batch_id = batch_id
 
 
+class EndForwardBackward(object):
+    """
+    Event On One Batch ForwardBackward Complete.
+    """
+
+    def __init__(self, pass_id, batch_id, gm):
+        self.pass_id = pass_id
+        self.batch_id = batch_id
+        self.gm = gm
+
+
 class EndIteration(WithMetric):
     """
     Event On One Batch Training Complete.

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

@@ -137,21 +137,26 @@ class TestBRelu(OpTest):
         self.check_grad(['X'], 'Y', max_relative_error=0.02)
 
 
-class TestLeakyRelu(OpTest):
+class TestRelu6(OpTest):
     def setUp(self):
-        self.op_type = "leaky_relu"
-        alpha = 0.02
-        self.attrs = {'alpha': alpha}
-        self.inputs = {'X': np.random.uniform(-3, 3, [4, 4]).astype("float32")}
+        self.op_type = "relu6"
+        x = np.random.uniform(-1, 1, [4, 10]).astype("float32")
+        threshold = 6.0
+        # The same with TestAbs
+        x[np.abs(x) < 0.005] = 0.02
+        x[np.abs(x - threshold) < 0.005] = threshold + 0.02
+
+        self.inputs = {'X': x}
+        self.attrs = {'threshold': threshold}
         self.outputs = {
-            'Y': np.maximum(self.inputs['X'], alpha * self.inputs['X'])
+            'Y': np.minimum(np.maximum(self.inputs['X'], 0), threshold)
         }
 
     def test_check_output(self):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Y', max_relative_error=0.007)
+        self.check_grad(['X'], 'Y', max_relative_error=0.02)
 
 
 class TestSoftRelu(OpTest):

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestAdamaxOp1(OpTest):
+    def setUp(self):
+        '''Test Adamax Operator with supplied attributes
+        '''
+        self.op_type = "adamax"
+        param = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        grad = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        moment = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        # The infinity norm is positive
+        inf_norm = np.random.random((102, 105)).astype("float32")
+
+        learning_rate = 0.002
+        beta1 = 0.78
+        beta2 = 0.899
+        epsilon = 1e-5
+        beta1_pow = beta1**10
+
+        self.inputs = {
+            'Param': param,
+            'Grad': grad,
+            'Moment': moment,
+            'InfNorm': inf_norm,
+            'LearningRate': np.array([learning_rate]).astype("float32"),
+            'Beta1Pow': np.array([beta1_pow]).astype("float32")
+        }
+
+        self.attrs = {'beta1': beta1, 'beta2': beta2, 'epsilon': epsilon}
+
+        param_out, moment_out, inf_norm_out, beta1_pow_out = adamax_step(
+            self.inputs, self.attrs)
+
+        self.outputs = {
+            'ParamOut': param_out,
+            'MomentOut': moment_out,
+            'InfNormOut': inf_norm_out,
+            'Beta1PowOut': beta1_pow_out
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestAdamaxOp2(OpTest):
+    '''Test Adamax Operator with default attributes
+    '''
+
+    def setUp(self):
+        self.op_type = "adamax"
+        param = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        grad = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        moment = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        # The infinity norm is positive
+        inf_norm = np.random.random((102, 105)).astype("float32")
+
+        learning_rate = 0.002
+        beta1 = 0.9
+        beta2 = 0.999
+        epsilon = 1e-8
+        beta1_pow = beta1**8
+
+        self.inputs = {
+            'Param': param,
+            'Grad': grad,
+            'Moment': moment,
+            'InfNorm': inf_norm,
+            'LearningRate': np.array([learning_rate]).astype("float32"),
+            'Beta1Pow': np.array([beta1_pow]).astype("float32")
+        }
+
+        attrs = {'beta1': beta1, 'beta2': beta2, 'epsilon': epsilon}
+        param_out, moment_out, inf_norm_out, beta1_pow_out = adamax_step(
+            self.inputs, attrs)
+
+        self.outputs = {
+            'ParamOut': param_out,
+            'MomentOut': moment_out,
+            'InfNormOut': inf_norm_out,
+            'Beta1PowOut': beta1_pow_out
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestAdamaxOpMultipleSteps(OpTest):
+    def setUp(self):
+        '''Test Adamax Operator with supplied attributes
+        '''
+        self.op_type = "adamax"
+        self.num_steps = 10
+
+        param = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        grad = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        moment = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        # The infinity norm is positive
+        inf_norm = np.random.random((102, 105)).astype("float32")
+
+        learning_rate = 0.002
+        beta1 = 0.8
+        beta2 = 0.99
+        epsilon = 1e-5
+        beta1_pow = 1
+
+        self.inputs = {
+            'Param': param,
+            'Grad': grad,
+            'Moment': moment,
+            'InfNorm': inf_norm,
+            'LearningRate': np.array([learning_rate]).astype("float32"),
+            'Beta1Pow': np.array([beta1_pow]).astype("float32")
+        }
+
+        self.attrs = {'beta1': beta1, 'beta2': beta2, 'epsilon': epsilon}
+
+        param_out, moment_out, inf_norm_out, beta1_pow_out = adamax_step(
+            self.inputs, self.attrs)
+
+    def test_check_output(self):
+        for _ in range(self.num_steps):
+            param_out, moment_out, inf_norm_out, beta1_pow_out = adamax_step(
+                self.inputs, self.attrs)
+
+            self.outputs = {
+                'ParamOut': param_out,
+                'MomentOut': moment_out,
+                'InfNormOut': inf_norm_out,
+                'Beta1PowOut': beta1_pow_out
+            }
+
+            # Verify output for this step
+            self.check_output()
+
+            # Output of this step becomes input for next step
+            self.inputs['Param'] = param_out
+            self.inputs['Moment'] = moment_out
+            self.inputs['InfNorm'] = inf_norm_out
+            self.inputs['Beta1Pow'] = beta1_pow_out
+
+            # Randomize gradient for next step
+            self.inputs['Grad'] = np.random.uniform(
+                -1, 1, (102, 105)).astype("float32")
+
+
+def adamax_step(inputs, attributes):
+    '''
+    Simulate one step of the adamax optimizer
+    :param inputs: dict of inputs
+    :param attributes: dict of attributes
+    :return tuple: tuple of output param, moment, inf_norm and
+    beta1 power accumulator
+    '''
+    param = inputs['Param']
+    grad = inputs['Grad']
+    moment = inputs['Moment']
+    inf_norm = inputs['InfNorm']
+    lr = inputs['LearningRate']
+    beta1_pow = inputs['Beta1Pow']
+
+    beta1 = attributes['beta1']
+    beta2 = attributes['beta2']
+    epsilon = attributes['epsilon']
+
+    moment_out = beta1 * moment + (1 - beta1) * grad
+    inf_norm_out = np.maximum(beta2 * inf_norm + epsilon, np.abs(grad))
+    beta1_pow_out = beta1_pow * beta1
+    lr_t = (lr / (1 - beta1_pow_out))
+    param_out = param - lr_t * np.divide(moment_out, inf_norm_out)
+
+    return param_out, moment_out, inf_norm_out, beta1_pow_out
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/v2/trainer.py

@@ -164,11 +164,18 @@ class SGD(object):
                                                           pass_type)
                 self.__gradient_machine__.eval(pass_evaluator)
                 self.__gradient_machine__.eval(batch_evaluator)
+                event_handler(
+                    v2_event.EndForwardBackward(
+                        pass_id=pass_id,
+                        batch_id=batch_id,
+                        gm=self.__gradient_machine__))
                 for each_param in self.__gradient_machine__.getNonStaticParameters(
                 ):
                     self.__parameter_updater__.update(each_param)
                 cost_sum = out_args.sum()
                 cost = cost_sum / len(data_batch)
+                self.__parameter_updater__.finishBatch(cost)
+                batch_evaluator.finish()
                 event_handler(
                     v2_event.EndIteration(
                         pass_id=pass_id,
@@ -176,8 +183,6 @@ class SGD(object):
                         cost=cost,
                         evaluator=batch_evaluator,
                         gm=self.__gradient_machine__))
-                self.__parameter_updater__.finishBatch(cost)
-                batch_evaluator.finish()
 
             self.__parameter_updater__.finishPass()
             pass_evaluator.finish()