Merge branch 'develop' into seqpool

cmake/generic.cmake

@@ -389,13 +389,60 @@ function(go_test TARGET_NAME)
     WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
 endfunction(go_test)
 
+# Modification of standard 'protobuf_generate_cpp()' with protobuf-lite support
+# Usage:
+#   paddle_protobuf_generate_cpp(<proto_srcs> <proto_hdrs> <proto_files>)
+
+function(paddle_protobuf_generate_cpp SRCS HDRS)
+  if(NOT ARGN)
+    message(SEND_ERROR "Error: paddle_protobuf_generate_cpp() called without any proto files")
+    return()
+  endif()
+
+  set(${SRCS})
+  set(${HDRS})
+
+  if (MOBILE_INFERENCE)
+      set(EXTRA_FLAG "lite:")  
+  else()
+      set(EXTRA_FLAG "") 
+  endif()
+
+  foreach(FIL ${ARGN})
+    get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
+    get_filename_component(FIL_WE ${FIL} NAME_WE)
+    
+    set(_protobuf_protoc_src "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}.pb.cc")
+    set(_protobuf_protoc_hdr "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}.pb.h")
+    list(APPEND ${SRCS} "${_protobuf_protoc_src}")
+    list(APPEND ${HDRS} "${_protobuf_protoc_hdr}")
+    
+    add_custom_command(
+      OUTPUT "${_protobuf_protoc_src}"
+             "${_protobuf_protoc_hdr}"
+
+      COMMAND ${CMAKE_COMMAND} -E make_directory "${CMAKE_CURRENT_BINARY_DIR}"
+      COMMAND ${PROTOBUF_PROTOC_EXECUTABLE} 
+      -I${CMAKE_CURRENT_SOURCE_DIR}
+      --cpp_out "${EXTRA_FLAG}${CMAKE_CURRENT_BINARY_DIR}" ${ABS_FIL}
+      DEPENDS ${ABS_FIL} protoc
+      COMMENT "Running C++ protocol buffer compiler on ${FIL}"
+      VERBATIM )
+  endforeach()
+
+  set_source_files_properties(${${SRCS}} ${${HDRS}} PROPERTIES GENERATED TRUE)
+  set(${SRCS} ${${SRCS}} PARENT_SCOPE)
+  set(${HDRS} ${${HDRS}} PARENT_SCOPE)
+endfunction()
+
+
 function(proto_library TARGET_NAME)
   set(oneValueArgs "")
   set(multiValueArgs SRCS DEPS)
   cmake_parse_arguments(proto_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
   set(proto_srcs)
   set(proto_hdrs)
-  protobuf_generate_cpp(proto_srcs proto_hdrs ${proto_library_SRCS})
+  paddle_protobuf_generate_cpp(proto_srcs proto_hdrs ${proto_library_SRCS})
   cc_library(${TARGET_NAME} SRCS ${proto_srcs} DEPS ${proto_library_DEPS} protobuf)
 endfunction()
 

paddle/api/CMakeLists.txt

@@ -26,7 +26,7 @@ FILE(GLOB PY_PADDLE_PYTHON_FILES ${PADDLE_SOURCE_DIR}/paddle/py_paddle/*.py)
 SET_SOURCE_FILES_PROPERTIES(Paddle.i PROPERTIES CPLUSPLUS ON)
 
 SET(CMAKE_SWIG_OUTDIR ${CMAKE_CURRENT_BINARY_DIR})
-SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-parentheses-equality -Wno-missing-field-initializers -Wno-self-assign")
+SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-parentheses-equality -Wno-missing-field-initializers -Wno-self-assign -ftls-model=global-dynamic")
 
 SET(SWIG_MODULE_swig_paddle_EXTRA_DEPS
     paddle_parameter

paddle/framework/CMakeLists.txt

@@ -42,11 +42,13 @@ add_custom_command(TARGET framework_py_proto POST_BUILD
 cc_library(backward SRCS backward.cc DEPS net_op)
 cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context)
 
-cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto backward ${GLOB_OP_LIB})
+cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto backward)
+set(EXECUTOR_TEST_OP elementwise_add_op gaussian_random_op feed_op fetch_op
+    mul_op sum_op squared_l2_distance_op fill_constant_op sgd_op)
 if(WITH_GPU)
-    nv_test(executor_test SRCS executor_test.cc DEPS executor)
+    nv_test(executor_test SRCS executor_test.cc DEPS executor ${EXECUTOR_TEST_OP})
 else()
-    cc_test(executor_test SRCS executor_test.cc DEPS executor)
+    cc_test(executor_test SRCS executor_test.cc DEPS executor ${EXECUTOR_TEST_OP})
 endif()
 
 cc_library(tensor_array SRCS tensor_array.cc DEPS lod_tensor)

paddle/framework/backward.h

@@ -27,6 +27,8 @@ extern std::unique_ptr<OperatorBase> Backward(
     const OperatorBase& forwardOp,
     const std::unordered_set<std::string>& no_grad_vars);
 
+// TODO(jiayi): Add target as parameter and generate backward op
+// according to target.
 void AppendBackward(ProgramDescBind& program_desc,
                     const std::unordered_set<std::string>& no_grad_vars);
 

paddle/framework/op_desc.cc

@@ -211,6 +211,15 @@ static InferShapeFuncMap &InferShapeFuncs() {
   return *g_map;
 }
 
+void OpDescBind::CheckAttrs() {
+  PADDLE_ENFORCE(!Type().empty(),
+                 "CheckAttr() can not be called before type is setted.");
+  const auto *checker = OpInfoMap::Instance().Get(Type()).Checker();
+  PADDLE_ENFORCE_NOT_NULL(checker, "Operator \"%s\" has no registered checker.",
+                          Type());
+  checker->Check(attrs_);
+}
+
 void OpDescBind::InferShape(const BlockDescBind &block) const {
   auto &funcs = InferShapeFuncs();
   auto it = funcs.find(this->Type());

paddle/framework/op_desc.h

@@ -100,6 +100,8 @@ class OpDescBind {
     return &this->attrs_;
   }
 
+  void CheckAttrs();
+
   void InferShape(const BlockDescBind &block) const;
 
  private:

paddle/framework/operator.h

@@ -289,6 +289,15 @@ class ExecutionContext {
     return device_context_;
   }
 
+#ifdef PADDLE_WITH_CUDA
+  const platform::CUDADeviceContext& cuda_device_context() const {
+    PADDLE_ENFORCE(platform::is_gpu_place(device_context_.GetPlace()));
+    auto cuda_ctx =
+        reinterpret_cast<const platform::CUDADeviceContext*>(&device_context_);
+    return *cuda_ctx;
+  }
+#endif
+
  private:
   const OperatorBase& op_;
   const Scope& scope_;

paddle/framework/tensor.h

@@ -87,26 +87,31 @@ class Tensor {
   /**
    * @brief   Copy the content of external tensor to a new place.
    *
-   * @param[in] src   The external tensor.
-   * @param[in] ctx   The device context contains place where to store.
+   * @param[in] src        The external tensor.
+   * @param[in] dst_place  The dst place.
+   * @param[in] ctx        The device context contains device resources.
    *
    * @note    CopyFrom supports CPU <-> GPU, GPU <-> GPU.
    */
+  // TODO(qijun): https://github.com/PaddlePaddle/Paddle/issues/4647
+  // Remove `CopyFrom` and `CopyFromVector` from Tensor interface
+  // and make them global functions
   template <typename T>
-  inline void CopyFrom(const Tensor& src, const platform::Place& dst_place);
+  inline void CopyFrom(const Tensor& src, const platform::Place& dst_place,
+                       const platform::DeviceContext& ctx);
 
   /**
    * @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.
+   * @param[in] src        The external tensor.
+   * @param[in] ctx        The device context contains device resources.
    *
    * * @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);
+                             const platform::DeviceContext& ctx);
 
   /**
    * @brief   Return the slice of the tensor.

paddle/framework/tensor_array.cc

@@ -95,7 +95,8 @@ void TensorArray::Write(size_t index, const LoDTensor& value) {
 
   values_[index].Resize(value.dims());
   values_[index].mutable_data<value_type>(platform::CPUPlace());
-  values_[index].CopyFrom<value_type>(value, platform::CPUPlace());
+  values_[index].CopyFrom<value_type>(value, platform::CPUPlace(),
+                                      platform::CPUDeviceContext());
 }
 
 void TensorArray::WriteShared(size_t index, const LoDTensor& value) {
@@ -151,7 +152,8 @@ LoDTensor TensorArray::Stack() const {
 
   for (size_t idx = 0; idx < size(); idx++) {
     result.Slice<value_type>(idx, idx + 1)
-        .CopyFrom<value_type>(Read(idx), platform::CPUPlace());
+        .CopyFrom<value_type>(Read(idx), platform::CPUPlace(),
+                              platform::CPUDeviceContext());
   }
   return result;
 }
@@ -182,7 +184,8 @@ void TensorArray::Unstack(const LoDTensor& source, bool data_shared) const {
       // copy
       value.Resize(value_dims);
       value.CopyFrom<value_type>(source.Slice<value_type>(elem, elem + 1),
-                                 platform::CPUPlace());
+                                 platform::CPUPlace(),
+                                 platform::CPUDeviceContext());
     }
   }
 }
@@ -236,7 +239,8 @@ LoDTensor DynamicBatchUnpacker::GetBatch(size_t index) {
     auto target = result.Slice<value_type>(i, i + 1);
     auto source_ = source->Slice<value_type>(index, index + 1);
 
-    target.CopyFrom<value_type>(source_, platform::CPUPlace());
+    target.CopyFrom<value_type>(source_, platform::CPUPlace(),
+                                platform::CPUDeviceContext());
   }
 
   return result;
@@ -269,7 +273,8 @@ LoDTensor PackDynamicBatch(const std::vector<LoDTensor>& source,
       if (index >= seq_meta.end) break;
       auto source_ = source[batch_id].Slice<float>(seq_id, seq_id + 1);
       auto target = result.Slice<float>(index, index + 1);
-      target.CopyFrom<float>(source_, platform::CPUPlace());
+      target.CopyFrom<float>(source_, platform::CPUPlace(),
+                             platform::CPUDeviceContext());
     }
   }
 

paddle/framework/tensor_impl.h

@@ -88,7 +88,8 @@ inline Tensor& Tensor::ShareDataWith(const Tensor& src) {
 
 template <typename T>
 inline void Tensor::CopyFrom(const Tensor& src,
-                             const platform::Place& dst_place) {
+                             const platform::Place& dst_place,
+                             const platform::DeviceContext& ctx) {
   src.check_memory_size<T>();
   Resize(src.dims());
 
@@ -106,26 +107,45 @@ inline void Tensor::CopyFrom(const Tensor& src,
 #ifdef PADDLE_WITH_CUDA
   else if (platform::is_gpu_place(src_place) &&
            platform::is_cpu_place(dst_place)) {
-    memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
-                 boost::get<platform::GPUPlace>(src_place), src_ptr, size, 0);
+    auto src_gpu_place = boost::get<platform::GPUPlace>(src_place);
+    auto dst_cpu_place = boost::get<platform::CPUPlace>(dst_place);
+    auto ctx_place = ctx.GetPlace();
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
+    auto ctx_gpu_place = boost::get<platform::GPUPlace>(ctx_place);
+    PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place);
+    memory::Copy(
+        dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size,
+        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
   } else if (platform::is_cpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
-    memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
-                 boost::get<platform::CPUPlace>(src_place), src_ptr, size, 0);
+    auto src_cpu_place = boost::get<platform::CPUPlace>(src_place);
+    auto dst_gpu_place = boost::get<platform::GPUPlace>(dst_place);
+    auto ctx_place = ctx.GetPlace();
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
+    auto ctx_gpu_place = boost::get<platform::GPUPlace>(ctx_place);
+    PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place);
+    memory::Copy(
+        dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size,
+        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
   } else if (platform::is_gpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
-    memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
-                 boost::get<platform::GPUPlace>(src_place), src_ptr, size, 0);
+    auto src_gpu_place = boost::get<platform::GPUPlace>(src_place);
+    auto dst_gpu_place = boost::get<platform::GPUPlace>(dst_place);
+    auto ctx_place = ctx.GetPlace();
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
+    auto ctx_gpu_place = boost::get<platform::GPUPlace>(ctx_place);
+    PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place);
+    memory::Copy(
+        dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
+        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
   }
-  PADDLE_ENFORCE(cudaStreamSynchronize(0),
-                 "cudaStreamSynchronize failed in Tensor CopyFrom");
-
 #endif
 }
 
 template <typename T>
 inline void Tensor::CopyFromVector(const std::vector<T>& src,
-                                   const platform::Place& dst_place) {
+                                   const platform::DeviceContext& ctx) {
+  auto dst_place = ctx.GetPlace();
   auto src_ptr = static_cast<const void*>(src.data());
   platform::CPUPlace src_place;
   auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
@@ -137,12 +157,11 @@ inline void Tensor::CopyFromVector(const std::vector<T>& src,
   }
 #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);
+    memory::Copy(
+        boost::get<platform::GPUPlace>(dst_place), dst_ptr, src_place, src_ptr,
+        size,
+        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
   }
-  PADDLE_ENFORCE(cudaStreamSynchronize(0),
-                 "cudaStreamSynchronize failed in Tensor CopyFromVector");
-
 #endif
 }
 

paddle/framework/tensor_test.cc

@@ -194,6 +194,7 @@ TEST(Tensor, CopyFrom) {
   {
     Tensor src_tensor;
     Tensor dst_tensor;
+    CPUDeviceContext cpu_ctx((CPUPlace()));
 
     int* src_ptr = src_tensor.mutable_data<int>(make_ddim({3, 3}), CPUPlace());
 
@@ -201,7 +202,7 @@ TEST(Tensor, CopyFrom) {
     memcpy(src_ptr, arr, 9 * sizeof(int));
 
     auto cpu_place = new paddle::platform::CPUPlace();
-    dst_tensor.CopyFrom<int>(src_tensor, *cpu_place);
+    dst_tensor.CopyFrom<int>(src_tensor, *cpu_place, cpu_ctx);
 
     const int* dst_ptr = dst_tensor.data<int>();
     ASSERT_NE(src_ptr, dst_ptr);
@@ -210,7 +211,7 @@ TEST(Tensor, CopyFrom) {
     }
 
     Tensor slice_tensor = src_tensor.Slice<int>(1, 2);
-    dst_tensor.CopyFrom<int>(slice_tensor, *cpu_place);
+    dst_tensor.CopyFrom<int>(slice_tensor, *cpu_place, cpu_ctx);
     const int* slice_ptr = slice_tensor.data<int>();
     dst_ptr = dst_tensor.data<int>();
     ASSERT_NE(dst_ptr, slice_ptr);
@@ -231,13 +232,15 @@ TEST(Tensor, CopyFrom) {
 
     // CPU Tensor to GPU Tensor
     auto gpu_place = new paddle::platform::GPUPlace(0);
-    gpu_tensor.CopyFrom<int>(src_tensor, *gpu_place);
+    CUDADeviceContext gpu_ctx(*gpu_place);
+    gpu_tensor.CopyFrom<int>(src_tensor, *gpu_place, gpu_ctx);
 
     // GPU Tensor to CPU Tensor
     auto cpu_place = new paddle::platform::CPUPlace();
-    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
+    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place, gpu_ctx);
 
-    // Compare Tensors
+    // Sync before Compare Tensors
+    gpu_ctx.Wait();
     const int* dst_ptr = dst_tensor.data<int>();
     ASSERT_NE(src_ptr, dst_ptr);
     for (size_t i = 0; i < 9; ++i) {
@@ -247,12 +250,13 @@ TEST(Tensor, CopyFrom) {
     Tensor slice_tensor = src_tensor.Slice<int>(1, 2);
 
     // CPU Slice Tensor to GPU Tensor
-    gpu_tensor.CopyFrom<int>(slice_tensor, *gpu_place);
+    gpu_tensor.CopyFrom<int>(slice_tensor, *gpu_place, gpu_ctx);
 
     // GPU Tensor to CPU Tensor
-    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
+    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place, gpu_ctx);
 
-    // Compare Slice Tensors
+    // Sync before Compare Slice Tensors
+    gpu_ctx.Wait();
     const int* slice_ptr = slice_tensor.data<int>();
     dst_ptr = dst_tensor.data<int>();
     ASSERT_NE(dst_ptr, slice_ptr);
@@ -273,7 +277,8 @@ TEST(Tensor, CopyFromVector) {
     // 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);
+    CPUDeviceContext cpu_ctx(*cpu_place);
+    cpu_tensor.CopyFromVector<int>(src_vec, cpu_ctx);
 
     // Compare Tensors
     const int* cpu_ptr = cpu_tensor.data<int>();
@@ -285,7 +290,7 @@ TEST(Tensor, CopyFromVector) {
 
     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_tensor.CopyFromVector<int>(src_vec, cpu_ctx);
     cpu_ptr = cpu_tensor.data<int>();
     src_ptr = src_vec.data();
     ASSERT_NE(src_ptr, cpu_ptr);
@@ -306,16 +311,19 @@ TEST(Tensor, CopyFromVector) {
     // 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);
+    CPUDeviceContext cpu_ctx(*cpu_place);
+    cpu_tensor.CopyFromVector<int>(src_vec, cpu_ctx);
 
     // 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);
+    CUDADeviceContext gpu_ctx(*gpu_place);
+    gpu_tensor.CopyFromVector<int>(src_vec, gpu_ctx);
     // Copy from GPU to CPU tensor for comparison
-    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
+    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place, gpu_ctx);
 
-    // Compare Tensors
+    // Sync before Compare Tensors
+    gpu_ctx.Wait();
     const int* src_ptr = src_vec.data();
     const int* cpu_ptr = cpu_tensor.data<int>();
     const int* dst_ptr = dst_tensor.data<int>();
@@ -329,11 +337,13 @@ TEST(Tensor, CopyFromVector) {
     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_tensor.CopyFromVector<int>(src_vec, cpu_ctx);
     gpu_tensor.Resize(make_ddim({2, 2}));
-    gpu_tensor.CopyFromVector<int>(src_vec, *gpu_place);
-    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
+    gpu_tensor.CopyFromVector<int>(src_vec, gpu_ctx);
+    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place, gpu_ctx);
 
+    // Sync before Compare Tensors
+    gpu_ctx.Wait();
     src_ptr = src_vec.data();
     cpu_ptr = cpu_tensor.data<int>();
     dst_ptr = dst_tensor.data<int>();

paddle/operators/activation_op.cc

@@ -321,6 +321,23 @@ class STanhOpMaker : public framework::OpProtoAndCheckerMaker {
   }
 };
 
+template <typename AttrType>
+class ThresholdedReluOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ThresholdedReluOpMaker(framework::OpProto *proto,
+                         framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "Input of ThresholdedRelu operator");
+    AddOutput("Y", "Output of ThresholdedRelu operator");
+    AddComment(
+        "ThresholdedRelu activation operator, "
+        "thresholded_relu = x for x > threshold, "
+        "thresholded_relu = 0 otherwise.");
+    AddAttr<AttrType>("threshold", "The threshold location of activation")
+        .SetDefault(static_cast<AttrType>(1.0));
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -392,6 +409,10 @@ REGISTER_OP(stanh, ops::ActivationOp, ops::STanhOpMaker<float>, stanh_grad,
 REGISTER_OP(hard_shrink, ops::ActivationOp, ops::HardShrinkOpMaker<float>,
             hard_shrink_grad, ops::ActivationOpGrad);
 
+REGISTER_OP(thresholded_relu, ops::ActivationOp,
+            ops::ThresholdedReluOpMaker<float>, thresholded_relu_grad,
+            ops::ActivationOpGrad);
+
 #define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor)        \
   REGISTER_OP_CPU_KERNEL(                                                      \
       act_type,                                                                \

paddle/operators/activation_op.h

@@ -590,6 +590,32 @@ struct STanhGradFunctor : public BaseActivationFunctor<T> {
   }
 };
 
+template <typename T>
+struct ThresholdedReluFunctor : public BaseActivationFunctor<T> {
+  float threshold;
+  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 > static_cast<T>(threshold)).template cast<T>() * x;
+  }
+};
+
+template <typename T>
+struct ThresholdedReluGradFunctor : 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>(threshold)).template cast<T>();
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -615,4 +641,5 @@ struct STanhGradFunctor : public BaseActivationFunctor<T> {
   __macro(leaky_relu, LeakyReluFunctor, LeakyReluGradFunctor);    \
   __macro(tanh_shrink, TanhShrinkFunctor, TanhShrinkGradFunctor); \
   __macro(elu, ELUFunctor, ELUGradFunctor);                       \
-  __macro(hard_shrink, HardShrinkFunctor, HardShrinkGradFunctor)
+  __macro(hard_shrink, HardShrinkFunctor, HardShrinkGradFunctor); \
+  __macro(thresholded_relu, ThresholdedReluFunctor, ThresholdedReluGradFunctor);

paddle/operators/conv2d_op.cc

@@ -12,111 +12,91 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#include "paddle/operators/gemm_conv2d_op.h"
+#include "paddle/operators/conv2d_op.h"
 
 namespace paddle {
 namespace operators {
 
-int outputSize(int input_size, int filter_size, int padding, int stride) {
-  int output_size = (input_size - filter_size + 2 * padding) / stride + 1;
-  return output_size;
+void Conv2DOp::InferShape(framework::InferShapeContext* ctx) const {
+  PADDLE_ENFORCE(ctx->HasInput("Input"),
+                 "Input(Input) of Conv2DOp should not be null.");
+  PADDLE_ENFORCE(ctx->HasInput("Filter"),
+                 "Input(Filter) of Conv2DOp should not be null.");
+  PADDLE_ENFORCE(ctx->HasOutput("Output"),
+                 "Output(Output) of Conv2DOp should not be null.");
+
+  auto in_dims = ctx->GetInputDim("Input");
+  auto filter_dims = ctx->GetInputDim("Filter");
+  std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
+  std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
+  int groups = ctx->Attrs().Get<int>("groups");
+  int input_channels = in_dims[1];
+  int output_channels = filter_dims[0];
+
+  PADDLE_ENFORCE_EQ(in_dims.size(), 4, "Conv2DOp input should be 4-D.");
+  PADDLE_ENFORCE_EQ(filter_dims.size(), 4, "Conv2DOp filter should be 4-D.");
+  PADDLE_ENFORCE_EQ(input_channels, filter_dims[1] * groups,
+                    "The number of input channels should be equal to filter "
+                    "channels * groups.");
+  PADDLE_ENFORCE_EQ(
+      output_channels % groups, 0,
+      "The number of output channels should be divided by groups.");
+
+  auto output_height =
+      OutputSize(in_dims[2], filter_dims[2], paddings[0], strides[0]);
+  auto output_width =
+      OutputSize(in_dims[3], filter_dims[3], paddings[1], strides[1]);
+  ctx->SetOutputDim("Output",
+                    {in_dims[0], filter_dims[0], output_height, output_width});
 }
 
-class Conv2DOp : public framework::OperatorWithKernel {
- public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
- protected:
-  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"),
-                   "Input(Filter) of Conv2DOp should not be null.");
-    PADDLE_ENFORCE(ctx->HasOutput("Output"),
-                   "Output(Output) of Conv2DOp should not be null.");
-
-    auto in_dims = ctx->GetInputDim("Input");
-    auto filter_dims = ctx->GetInputDim("Filter");
-    std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
-    std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
-    int groups = ctx->Attrs().Get<int>("groups");
-    int input_channels = in_dims[1];
-    int output_channels = filter_dims[0];
-
-    PADDLE_ENFORCE_EQ(in_dims.size(), 4, "Conv2DOp input should be 4-D.");
-    PADDLE_ENFORCE_EQ(filter_dims.size(), 4, "Conv2DOp filter should be 4-D.");
-    PADDLE_ENFORCE_EQ(input_channels, filter_dims[1] * groups,
-                      "The number of input channels should be equal to filter "
-                      "channels * groups.");
-    PADDLE_ENFORCE_EQ(
-        output_channels % groups, 0,
-        "The number of output channels should be divided by groups.");
-
-    auto output_height =
-        outputSize(in_dims[2], filter_dims[2], paddings[0], strides[0]);
-    auto output_width =
-        outputSize(in_dims[3], filter_dims[3], paddings[1], strides[1]);
-    ctx->SetOutputDim(
-        "Output", {in_dims[0], filter_dims[0], output_height, output_width});
-  }
-};
-
-class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker {
- public:
-  Conv2DOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput(
-        "Input",
-        "The input tensor of convolution operator. "
-        "The format of input tensor is NCHW. Where N is batch size, C is the "
-        "number of channels, H and W is the height and width of image.");
-    AddInput(
-        "Filter",
-        "The filter tensor of convolution operator."
-        "The format of the filter tensor is MCHW, where M is the number of "
-        "output image channels, C is the number of input image channels, "
-        "H and W is height and width of filter. "
-        "If the groups attribute is greater than 1, C equal the number of "
-        "input image channels divided by the groups.");
-    AddOutput("Output",
-              "The output tensor of convolution operator."
-              "The format of output tensor is also NCHW.");
-    AddAttr<std::vector<int>>("strides", "strides of convolution operator.")
-        .SetDefault({1, 1});
-    AddAttr<std::vector<int>>("paddings", "paddings of convolution operator.")
-        .SetDefault({0, 0});
-    AddAttr<int>(
-        "groups",
-        "group size of convolution operator. "
-        "Refer to grouped convolution in Alex Krizhevsky's paper: "
-        "when group=2, the first half of the filters are only connected to the "
-        "first half of the input channels, and the second half only connected "
-        "to the second half.")
-        .SetDefault(1);
-    AddComment(R"DOC(
+Conv2DOpMaker::Conv2DOpMaker(framework::OpProto* proto,
+                             framework::OpAttrChecker* op_checker)
+    : OpProtoAndCheckerMaker(proto, op_checker) {
+  AddInput(
+      "Input",
+      "The input tensor of convolution operator. "
+      "The format of input tensor is NCHW. Where N is batch size, C is the "
+      "number of channels, H and W is the height and width of image.");
+  AddInput("Filter",
+           "The filter tensor of convolution operator."
+           "The format of the filter tensor is MCHW, where M is the number of "
+           "output image channels, C is the number of input image channels, "
+           "H and W is height and width of filter. "
+           "If the groups attribute is greater than 1, C equal the number of "
+           "input image channels divided by the groups.");
+  AddOutput("Output",
+            "The output tensor of convolution operator."
+            "The format of output tensor is also NCHW.");
+  AddAttr<std::vector<int>>("strides", "strides of convolution operator.")
+      .SetDefault({1, 1});
+  AddAttr<std::vector<int>>("paddings", "paddings of convolution operator.")
+      .SetDefault({0, 0});
+  AddAttr<int>(
+      "groups",
+      "group size of convolution operator. "
+      "Refer to grouped convolution in Alex Krizhevsky's paper: "
+      "when group=2, the first half of the filters are only connected to the "
+      "first half of the input channels, and the second half only connected "
+      "to the second half.")
+      .SetDefault(1);
+  AddComment(R"DOC(
 The convolution operation calculates the output based on the input, filter
 and strides, paddings, groups parameters. The size of each dimension of the
 parameters is checked in the infer-shape.
 )DOC");
-  }
-};
-
-class Conv2DOpGrad : public framework::OperatorWithKernel {
- public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
+}
 
- protected:
-  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"))) {
-      ctx->SetOutputDim(framework::GradVarName("Input"), in_dims);
-    }
-    if (ctx->HasOutput(framework::GradVarName("Filter"))) {
-      ctx->SetOutputDim(framework::GradVarName("Filter"), filter_dims);
-    }
+void Conv2DOpGrad::InferShape(framework::InferShapeContext* ctx) const {
+  auto in_dims = ctx->GetInputDim("Input");
+  auto filter_dims = ctx->GetInputDim("Filter");
+  if (ctx->HasOutput(framework::GradVarName("Input"))) {
+    ctx->SetOutputDim(framework::GradVarName("Input"), in_dims);
   }
-};
+  if (ctx->HasOutput(framework::GradVarName("Filter"))) {
+    ctx->SetOutputDim(framework::GradVarName("Filter"), filter_dims);
+  }
+}
 
 }  // namespace operators
 }  // namespace paddle

paddle/operators/conv2d_op.cu

@@ -12,7 +12,7 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#include "paddle/operators/gemm_conv2d_op.h"
+#include "paddle/operators/conv2d_op.h"
 
 namespace ops = paddle::operators;
 

paddle/operators/gemm_conv2d_op.h → paddle/operators/conv2d_op.h

@@ -24,6 +24,38 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
+// Base convolution operator definations for other conv
+// like operators to reuse the implementation.
+inline int OutputSize(int input_size, int filter_size, int padding,
+                      int stride) {
+  int output_size = (input_size - filter_size + 2 * padding) / stride + 1;
+  return output_size;
+}
+
+// Define Op classes in .h file so that other conv
+// operator implementations can reuse the code.
+class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  Conv2DOpMaker(framework::OpProto* proto,
+                framework::OpAttrChecker* op_checker);
+};
+
+class Conv2DOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override;
+};
+
+class Conv2DOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override;
+};
+
 template <typename Place, typename T>
 class GemmConv2DKernel : public framework::OpKernel<T> {
  public:
@@ -74,7 +106,6 @@ class GemmConv2DKernel : public framework::OpKernel<T> {
 
     framework::DDim output_matrix_shape = {output_channels,
                                            output_height * output_width};
-
     // convolution operator: im2col + gemm
     int in_step = input_channels / groups;
     int out_step = output_channels / groups;

paddle/operators/conv_cudnn_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/conv2d_op.h"
+
+namespace paddle {
+namespace operators {
+
+class CudnnConvOpMaker : public Conv2DOpMaker {
+ public:
+  CudnnConvOpMaker(framework::OpProto* proto,
+                   framework::OpAttrChecker* op_checker)
+      : Conv2DOpMaker(proto, op_checker) {
+    AddAttr<std::vector<int>>("dilations", "dilations of convolution operator.")
+        .SetDefault(std::vector<int>{1, 1});
+    AddAttr<int>("workspace_size_MB",
+                 "workspace size for cudnn, in MB, "
+                 "workspace is a section of GPU memory which will be "
+                 "allocated/freed each time the operator runs, larger "
+                 "workspace size can increase performance but also requires "
+                 "better hardward. This size should be carefully setted.")
+        .SetDefault(4096);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(conv_cudnn, ops::Conv2DOp, ops::CudnnConvOpMaker, conv_cudnn_grad,
+            ops::Conv2DOpGrad);
+REGISTER_OP_CPU_KERNEL(
+    conv_cudnn, ops::GemmConv2DKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    conv_cudnn_grad,
+    ops::GemmConvGrad2DKernel<paddle::platform::CPUPlace, float>);

paddle/operators/conv_cudnn_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/memory/memory.h"
+#include "paddle/operators/conv2d_op.h"
+#include "paddle/platform/assert.h"
+#include "paddle/platform/cudnn_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using ScopedTensorDescriptor = platform::ScopedTensorDescriptor;
+using ScopedFilterDescriptor = platform::ScopedFilterDescriptor;
+using ScopedConvolutionDescriptor = platform::ScopedConvolutionDescriptor;
+using DataLayout = platform::DataLayout;
+using CUDADeviceContext = platform::CUDADeviceContext;
+
+static constexpr size_t kCONV_CUDNN_WORKSPACE_LIMIT_BYTES = 1024 * 1024 * 1024;
+
+// NOTE: framework::vectorize converts to type int64_t
+//       which does not fit cudnn inputs.
+std::vector<int> Dims2Vector(const framework::DDim& dims) {
+  std::vector<int> ret;
+  for (int i = 0; i < dims.size(); i++) {
+    ret.push_back(dims[i]);
+  }
+  return ret;
+}
+
+template <typename T>
+class CudnnConvOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use GPUPlace.");
+    auto* input = ctx.Input<Tensor>("Input");
+    auto* filter = ctx.Input<Tensor>("Filter");
+    auto* output = ctx.Output<Tensor>("Output");
+
+    std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
+    std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
+    std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
+    int groups = ctx.Attr<int>("groups");
+    int user_workspace_size = ctx.Attr<int>("workspace_size_MB");
+
+    const T* input_data = input->data<T>();
+    const T* filter_data = filter->data<T>();
+    T* output_data = output->mutable_data<T>(ctx.GetPlace());
+
+    // ------------------- cudnn descriptors ---------------------
+    ScopedTensorDescriptor input_desc;
+    ScopedTensorDescriptor output_desc;
+    ScopedFilterDescriptor filter_desc;
+    ScopedConvolutionDescriptor conv_desc;
+    DataLayout layout = DataLayout::kNCHW;
+
+    cudnnTensorDescriptor_t cudnn_input_desc =
+        input_desc.descriptor<T>(layout, Dims2Vector(input->dims()), groups);
+    cudnnTensorDescriptor_t cudnn_output_desc =
+        output_desc.descriptor<T>(layout, Dims2Vector(output->dims()), groups);
+    cudnnFilterDescriptor_t cudnn_filter_desc =
+        filter_desc.descriptor<T>(layout, Dims2Vector(filter->dims()), groups);
+    cudnnConvolutionDescriptor_t cudnn_conv_desc =
+        conv_desc.descriptor<T>(paddings, strides, dilations);
+
+    int input_channels = input->dims()[1];
+    int input_height = input->dims()[2];
+    int input_width = input->dims()[3];
+    int output_channels = output->dims()[1];
+    int output_height = output->dims()[2];
+    int output_width = output->dims()[3];
+
+    int group_offset_in = input_channels / groups * input_height * input_width;
+    int group_offset_out =
+        output_channels / groups * output_height * output_width;
+    int group_offset_filter = filter->numel() / groups;
+    // ------------------- cudnn conv workspace ---------------------
+    void* cudnn_workspace = nullptr;
+    size_t workspace_size_in_bytes;  // final workspace to allocate.
+    size_t workspace_size_limit = kCONV_CUDNN_WORKSPACE_LIMIT_BYTES;
+    if (user_workspace_size > 0) {
+      workspace_size_limit = user_workspace_size * 1024 * 1024;
+    }
+    // ------------------- cudnn conv algorithm ---------------------
+    cudnnConvolutionFwdAlgo_t algo;
+    auto handle = ctx.cuda_device_context().cudnn_handle();
+
+    PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm(
+        handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc,
+        cudnn_output_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
+        workspace_size_limit, &algo));
+    // get workspace size able to allocate
+    PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize(
+        handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc,
+        cudnn_output_desc, algo, &workspace_size_in_bytes));
+    // Allocate on GPU memory
+    platform::GPUPlace gpu = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
+    // ------------------- cudnn conv forward ---------------------
+    T alpha = 1.0f, beta = 0.0f;
+    for (int i = 0; i < groups; i++) {
+      PADDLE_ENFORCE(platform::dynload::cudnnConvolutionForward(
+          handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in,
+          cudnn_filter_desc, filter_data + i * group_offset_filter,
+          cudnn_conv_desc, algo, cudnn_workspace, workspace_size_in_bytes,
+          &beta, cudnn_output_desc, output_data + i * group_offset_out));
+    }
+    // Release the cudnn workspace
+    paddle::memory::Free(gpu, cudnn_workspace);
+  }
+};
+
+template <typename T>
+class CudnnConvGradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use GPUPlace.");
+    auto input = ctx.Input<Tensor>("Input");
+    auto filter = ctx.Input<Tensor>("Filter");
+    auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
+    auto input_grad = ctx.Output<Tensor>(framework::GradVarName("Input"));
+    auto filter_grad = ctx.Output<Tensor>(framework::GradVarName("Filter"));
+
+    const T* input_data = input->data<T>();
+    const T* output_grad_data = output_grad->data<T>();
+    const T* filter_data = filter->data<T>();
+
+    std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
+    std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
+    std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
+    int groups = ctx.Attr<int>("groups");
+    int user_workspace_size = ctx.Attr<int>("workspace_size_MB");
+
+    // ------------------- cudnn descriptors ---------------------
+    ScopedTensorDescriptor input_desc;
+    ScopedTensorDescriptor output_grad_desc;
+    ScopedTensorDescriptor input_grad_desc;
+
+    ScopedFilterDescriptor filter_desc;
+    ScopedFilterDescriptor filter_grad_desc;
+    ScopedConvolutionDescriptor conv_desc;
+    DataLayout layout = DataLayout::kNCHW;
+
+    cudnnTensorDescriptor_t cudnn_input_desc =
+        input_desc.descriptor<T>(layout, Dims2Vector(input->dims()), groups);
+    cudnnTensorDescriptor_t cudnn_output_grad_desc =
+        output_grad_desc.descriptor<T>(layout, Dims2Vector(output_grad->dims()),
+                                       groups);
+    cudnnFilterDescriptor_t cudnn_filter_desc =
+        filter_desc.descriptor<T>(layout, Dims2Vector(filter->dims()), groups);
+    cudnnTensorDescriptor_t cudnn_input_grad_desc = nullptr;
+    cudnnFilterDescriptor_t cudnn_filter_grad_desc = nullptr;
+
+    cudnnConvolutionDescriptor_t cudnn_conv_desc =
+        conv_desc.descriptor<T>(paddings, strides, dilations);
+
+    int input_channels = input->dims()[1];
+    int input_height = input->dims()[2];
+    int input_width = input->dims()[3];
+    int output_grad_channels = filter->dims()[0];
+    int output_grad_height = output_grad->dims()[2];
+    int output_grad_width = output_grad->dims()[3];
+
+    int group_offset_in = input_channels / groups * input_height * input_width;
+    int group_offset_out =
+        output_grad_channels / groups * output_grad_height * output_grad_width;
+    int group_offset_filter = filter->numel() / groups;
+    // ------------------- cudnn backward algorithm ---------------------
+    cudnnConvolutionBwdDataAlgo_t data_algo;
+    cudnnConvolutionBwdFilterAlgo_t filter_algo;
+    size_t workspace_size_in_bytes = 0, tmp_size = 0;
+    size_t workspace_size_limit = kCONV_CUDNN_WORKSPACE_LIMIT_BYTES;
+    if (user_workspace_size > 0) {
+      workspace_size_limit = user_workspace_size * 1024 * 1024;
+    }
+
+    auto handle = ctx.cuda_device_context().cudnn_handle();
+    if (input_grad) {
+      cudnn_input_grad_desc = input_grad_desc.descriptor<T>(
+          layout, Dims2Vector(input_grad->dims()), groups);
+      PADDLE_ENFORCE(
+          platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
+              handle, cudnn_filter_desc,
+              // dyDesc: Handle to the previously initialized input differential
+              // tensor descriptor.
+              cudnn_output_grad_desc, cudnn_conv_desc,
+              // dxDesc: Handle to the previously initialized output tensor
+              // descriptor.
+              cudnn_input_grad_desc,
+              CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
+              workspace_size_limit, &data_algo));
+      PADDLE_ENFORCE(
+          platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
+              handle, cudnn_filter_desc, cudnn_output_grad_desc,
+              cudnn_conv_desc, cudnn_input_grad_desc, data_algo, &tmp_size));
+      workspace_size_in_bytes = std::max(workspace_size_in_bytes, tmp_size);
+    }
+
+    if (filter_grad) {
+      cudnn_filter_grad_desc = filter_grad_desc.descriptor<T>(
+          layout, Dims2Vector(filter_grad->dims()), groups);
+      PADDLE_ENFORCE(
+          platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
+              handle, cudnn_input_desc, cudnn_output_grad_desc, cudnn_conv_desc,
+              cudnn_filter_desc,
+              CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
+              workspace_size_limit, &filter_algo));
+
+      PADDLE_ENFORCE(
+          platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(
+              handle, cudnn_input_desc, cudnn_output_grad_desc, cudnn_conv_desc,
+              cudnn_filter_desc, filter_algo, &tmp_size));
+      workspace_size_in_bytes = std::max(workspace_size_in_bytes, tmp_size);
+    }
+    // ------------------- cudnn conv workspace ---------------------
+    // Already on GPU
+    void* cudnn_workspace = nullptr;
+    platform::GPUPlace gpu = boost::get<platform::GPUPlace>(ctx.GetPlace());
+    cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
+    // ------------------- cudnn conv backward data ---------------------
+    // FIXME(typhoonzero): template type T may not be the same as cudnn call.
+    T alpha = 1.0f, beta = 0.0f;
+    if (input_grad) {
+      T* input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
+      auto t = framework::EigenVector<T>::Flatten(*input_grad);
+      t.device(ctx.GetEigenDevice<platform::GPUPlace>()) =
+          t.constant(static_cast<T>(0));
+      for (int i = 0; i < groups; i++) {
+        PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardData(
+            handle, &alpha, cudnn_filter_desc,
+            filter_data + i * group_offset_filter, cudnn_output_grad_desc,
+            output_grad_data + i * group_offset_out, cudnn_conv_desc, data_algo,
+            cudnn_workspace, workspace_size_in_bytes, &beta,
+            cudnn_input_grad_desc, input_grad_data + i * group_offset_in));
+      }
+    }
+    // ------------------- cudnn conv backward filter ---------------------
+    if (filter_grad) {
+      T* filter_grad_data = filter_grad->mutable_data<T>(ctx.GetPlace());
+      auto t = framework::EigenVector<T>::Flatten(*filter_grad);
+      t.device(ctx.GetEigenDevice<platform::GPUPlace>()) =
+          t.constant(static_cast<T>(0));
+      for (int i = 0; i < groups; i++) {
+        PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter(
+            handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in,
+            cudnn_output_grad_desc, output_grad_data + i * group_offset_out,
+            cudnn_conv_desc, filter_algo, cudnn_workspace,
+            workspace_size_in_bytes, &beta, cudnn_filter_grad_desc,
+            filter_grad_data + i * group_offset_filter));
+      }
+    }
+    // Release the cudnn workspace
+    paddle::memory::Free(gpu, cudnn_workspace);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_GPU_KERNEL(conv_cudnn, paddle::operators::CudnnConvOpKernel<float>);
+REGISTER_OP_GPU_KERNEL(conv_cudnn_grad,
+                       paddle::operators::CudnnConvGradOpKernel<float>);

paddle/operators/feed_op.h

@@ -34,7 +34,7 @@ class FeedKernel : public framework::OpKernel<T> {
     // TODO(qijun):
     //   check tensors[col].dims() with attribute,
     //   except the first dimenson.
-    out->CopyFrom<T>(tensors[col], ctx.GetPlace());
+    out->CopyFrom<T>(tensors[col], ctx.GetPlace(), ctx.device_context());
   }
 };
 

paddle/operators/fetch_op.h

@@ -35,7 +35,8 @@ class FetchKernel : public framework::OpKernel<T> {
     PADDLE_ENFORCE_GT(tensors->size(), static_cast<size_t>(col));
     (*tensors)[col].Resize(input->dims());
     (*tensors)[col].mutable_data<T>(platform::CPUPlace());
-    (*tensors)[col].CopyFrom<T>(*input, platform::CPUPlace());
+    (*tensors)[col].CopyFrom<T>(*input, platform::CPUPlace(),
+                                ctx.device_context());
     // TODO(qijun): need to handle LodTensor later
   }
 };

paddle/operators/margin_rank_loss_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/margin_rank_loss_op.h"
+
+namespace paddle {
+namespace operators {
+
+class MarginRankLossOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    // input check
+    PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) shouldn't be null.");
+    PADDLE_ENFORCE(ctx->HasInput("X1"), "Input(X1) shouldn't be null.");
+    PADDLE_ENFORCE(ctx->HasInput("X2"), "Input(X2) shouldn't be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) shouldn't be null.");
+    auto label_dims = ctx->GetInputDim("Label");
+    auto x1_dims = ctx->GetInputDim("X1");
+    auto x2_dims = ctx->GetInputDim("X2");
+    PADDLE_ENFORCE(
+        (label_dims == x1_dims) && (x1_dims == x2_dims) &&
+            (label_dims.size() == 2) && (label_dims[1] == 1),
+        "All inputs must be 2-D tensor with shape [batch_size x 1].");
+    ctx->SetOutputDim("Activated", label_dims);
+    ctx->SetOutputDim("Out", label_dims);
+  }
+};
+
+template <typename T>
+class MarginRankLossOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  MarginRankLossOpMaker(framework::OpProto *proto,
+                        framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X1",
+             "(2-D tensor with shape [batch_size x 1]) The score for "
+             "one item X1 to be ranked, from pairwise ranking model.");
+    AddInput("X2",
+             "(2-D tensor with shape [batch_size x 1]) The score for "
+             "another item X2 to be ranked, from pairwise ranking model.");
+    AddInput("Label",
+             "(2-D tensor with shape [batch_size x 1]) "
+             "The label indicating X1 ranked higher than X2 or not, "
+             "can only be +1 or -1.");
+    AddAttr<T>("margin", "(scalar, default 0) Margin for MarginRankLossOp.")
+        .SetDefault(static_cast<T>(0));
+    AddOutput("Activated",
+              "(2-D tensor with shape [batch_size x 1]) Intermediate tensor "
+              "to indicate whether each element of Output(Out) is activated.")
+        .AsIntermediate();
+    AddOutput("Out",
+              "(2-D tensor with shape [batch_size x 1]) "
+              "The output loss of MarginRankLoss operator.");
+    AddComment(R"DOC(
+
+MarginRankLoss operator measures the loss given a pair of training sample
+{`X1`, `X2`} and the `Label` with attribute `margin`, where `Label = +1` 
+indicating X1 is ranked higher than `X2`, otherwise `Label = -1`. The loss 
+turns out
+
+loss(X1, X2, Label) = max(0, -Label * (X1 - X2) + margin).
+
+The attribute `margin` involved here helps make the predictions more robust.
+Denote the item ranked higher as the positive sample, otherwise the negative 
+sample. If the score of the two samples satisfies 
+
+positive sample - negative sample < margin,
+
+the pair of samples will contribute to the final loss, which will backpropogate 
+and train the ranking model to enlarge the difference of the two score.
+
+For batch input with size `batch_size`, `X1`, `X2` and `Label`
+all have the same shape [batch_size x 1].
+
+)DOC");
+  }
+};
+
+class MarginRankLossGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) shouldn't be null.");
+    PADDLE_ENFORCE(ctx->HasInput("X1"), "Input(X1) shouldn't be null.");
+    PADDLE_ENFORCE(ctx->HasInput("X2"), "Input(X2) shouldn't be null.");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Input(Out@GRAD) shouldn't be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Activated"),
+                   "Intermediate(Activated) shouldn't be null.");
+    auto dims = ctx->GetInputDim("Label");
+    ctx->SetOutputDim(framework::GradVarName("X1"), dims);
+    ctx->SetOutputDim(framework::GradVarName("X2"), dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+
+REGISTER_OP(margin_rank_loss, ops::MarginRankLossOp,
+            ops::MarginRankLossOpMaker<float>, margin_rank_loss_grad,
+            ops::MarginRankLossGradOp);
+REGISTER_OP_CPU_KERNEL(
+    margin_rank_loss,
+    ops::MarginRankLossKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    margin_rank_loss_grad,
+    ops::MarginRankLossGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/margin_rank_loss_op.cu

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

paddle/operators/margin_rank_loss_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 T>
+struct ReLU {
+  HOSTDEVICE T operator()(const T& val) const {
+    return val > 0 ? val : static_cast<T>(0);
+  }
+};
+
+template <typename T>
+struct Heaviside {
+  HOSTDEVICE T operator()(const T& val) const {
+    return static_cast<T>(val > 0 ? 1 : 0);
+  }
+};
+
+template <typename Place, typename T>
+class MarginRankLossKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const {
+    auto* out_t = ctx.Output<framework::Tensor>("Out");
+    auto* act_t = ctx.Output<framework::Tensor>("Activated");
+
+    auto* label_t = ctx.Input<framework::Tensor>("Label");
+    auto* x1_t = ctx.Input<framework::Tensor>("X1");
+    auto* x2_t = ctx.Input<framework::Tensor>("X2");
+
+    out_t->mutable_data<T>(ctx.GetPlace());
+    act_t->mutable_data<T>(ctx.GetPlace());
+
+    auto margin = static_cast<T>(ctx.Attr<T>("margin"));
+    auto out = framework::EigenVector<T>::Flatten(*out_t);
+    auto act = framework::EigenVector<T>::Flatten(*act_t);
+
+    auto label = framework::EigenVector<T>::Flatten(*label_t);
+    auto x1 = framework::EigenVector<T>::Flatten(*x1_t);
+    auto x2 = framework::EigenVector<T>::Flatten(*x2_t);
+
+    auto& dev = ctx.GetEigenDevice<Place>();
+    out.device(dev) = (-label * (x1 - x2) + margin).unaryExpr(ReLU<T>());
+    act.device(dev) = out.unaryExpr(Heaviside<T>());
+  }
+};
+
+template <typename Place, typename T>
+class MarginRankLossGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const {
+    auto* d_x1_t =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X1"));
+    auto* d_x2_t =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X2"));
+
+    auto* act_t = ctx.Input<framework::Tensor>("Activated");
+    auto* d_out_t = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* label_t = ctx.Input<framework::Tensor>("Label");
+
+    auto d_out = framework::EigenVector<T>::Flatten(*d_out_t);
+    auto act = framework::EigenVector<T>::Flatten(*act_t);
+    auto label = framework::EigenVector<T>::Flatten(*label_t);
+    auto& dev = ctx.GetEigenDevice<Place>();
+
+    // compute d_x1
+    if (d_x1_t) {
+      d_x1_t->mutable_data<T>(ctx.GetPlace());
+      auto d_x1 = framework::EigenVector<T>::Flatten(*d_x1_t);
+      d_x1.device(dev) = -d_out * act * label;
+    }
+    // compute d_x2
+    if (d_x2_t) {
+      d_x2_t->mutable_data<T>(ctx.GetPlace());
+      auto d_x2 = framework::EigenVector<T>::Flatten(*d_x2_t);
+      d_x2.device(dev) = d_out * act * label;
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/im2col_test.cc

@@ -49,10 +49,22 @@ void testIm2col() {
   memcpy(input_ptr, arr, 6 * sizeof(float));
 
   auto* place = new Place();
+  paddle::platform::DeviceContext* context;
+  if (paddle::platform::is_cpu_place(*place)) {
+    context =
+        new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace());
+  } else {
+#ifdef PADDLE_WITH_CUDA
+    context =
+        new paddle::platform::CUDADeviceContext(paddle::platform::GPUPlace());
+#else
+    PADDLE_THROW("no GPU support");
+#endif  // PADDLE_ONLY_CPU
+  }
   if (paddle::platform::is_cpu_place(*place)) {
     input = input_tmp;
   } else {
-    input.CopyFrom<float>(input_tmp, *place);
+    input.CopyFrom<float>(input_tmp, *place, *context);
   }
   output_cfo.mutable_data<float>(
       {1, filter_size, filter_size, output_height, output_width}, *place);
@@ -66,18 +78,6 @@ void testIm2col() {
       paddle::operators::math::ColFormat::kOCF, Place, float>
       im2col_ocf;
 
-  paddle::platform::DeviceContext* context;
-  if (paddle::platform::is_cpu_place(*place)) {
-    context =
-        new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace());
-  } else {
-#ifdef PADDLE_WITH_CUDA
-    context =
-        new paddle::platform::CUDADeviceContext(paddle::platform::GPUPlace());
-#else
-    PADDLE_THROW("no GPU support");
-#endif  // PADDLE_ONLY_CPU
-  }
   im2col(*context, input, output_cfo, stride, stride, padding, padding);
   im2col_ocf(*context, input, output_ocf, stride, stride, padding, padding);
 
@@ -85,7 +85,8 @@ void testIm2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     out_cfo_ptr = output_cfo.data<float>();
   } else {
-    output_tmp.CopyFrom<float>(output_cfo, paddle::platform::CPUPlace());
+    output_tmp.CopyFrom<float>(output_cfo, paddle::platform::CPUPlace(),
+                               *context);
     out_cfo_ptr = output_tmp.data<float>();
   }
   EXPECT_EQ(out_cfo_ptr[0], 0);
@@ -101,7 +102,8 @@ void testIm2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     out_ocf_ptr = output_ocf.data<float>();
   } else {
-    output_tmp.CopyFrom<float>(output_ocf, paddle::platform::CPUPlace());
+    output_tmp.CopyFrom<float>(output_ocf, paddle::platform::CPUPlace(),
+                               *context);
     out_ocf_ptr = output_tmp.data<float>();
   }
   EXPECT_EQ(out_ocf_ptr[0], 0);

paddle/operators/math/math_function_test.cc

@@ -17,17 +17,18 @@ TEST(math_function, notrans_mul_trans) {
   auto* gpu_place = new paddle::platform::GPUPlace(0);
   paddle::platform::CUDADeviceContext context(*gpu_place);
 
-  input1_gpu.CopyFrom<float>(input1, *gpu_place);
-  input2_gpu.CopyFrom<float>(input1, *gpu_place);
+  input1_gpu.CopyFrom<float>(input1, *gpu_place, context);
+  input2_gpu.CopyFrom<float>(input1, *gpu_place, context);
 
   out_gpu.mutable_data<float>({2, 2}, *gpu_place);
 
   paddle::operators::math::matmul<paddle::platform::GPUPlace, float>(
       context, input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0);
 
-  out.CopyFrom<float>(out_gpu, *cpu_place);
+  out.CopyFrom<float>(out_gpu, *cpu_place, context);
 
   float* out_ptr = out.data<float>();
+  context.Wait();
   EXPECT_EQ(out_ptr[0], 5);
   EXPECT_EQ(out_ptr[1], 14);
   EXPECT_EQ(out_ptr[2], 14);
@@ -50,17 +51,18 @@ TEST(math_function, trans_mul_notrans) {
   auto* gpu_place = new paddle::platform::GPUPlace(0);
   paddle::platform::CUDADeviceContext context(*gpu_place);
 
-  input1_gpu.CopyFrom<float>(input1, *gpu_place);
-  input2_gpu.CopyFrom<float>(input1, *gpu_place);
+  input1_gpu.CopyFrom<float>(input1, *gpu_place, context);
+  input2_gpu.CopyFrom<float>(input1, *gpu_place, context);
 
   out_gpu.mutable_data<float>({3, 3}, *gpu_place);
 
   paddle::operators::math::matmul<paddle::platform::GPUPlace, float>(
       context, input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0);
 
-  out.CopyFrom<float>(out_gpu, *cpu_place);
+  out.CopyFrom<float>(out_gpu, *cpu_place, context);
 
   float* out_ptr = out.data<float>();
+  context.Wait();
   EXPECT_EQ(out_ptr[0], 9);
   EXPECT_EQ(out_ptr[1], 12);
   EXPECT_EQ(out_ptr[2], 15);
@@ -98,9 +100,9 @@ TEST(math_function, gemm_notrans_cublas) {
   auto* gpu_place = new paddle::platform::GPUPlace(0);
   paddle::platform::CUDADeviceContext context(*gpu_place);
 
-  input1_gpu.CopyFrom<float>(input1, *gpu_place);
-  input2_gpu.CopyFrom<float>(input2, *gpu_place);
-  input3_gpu.CopyFrom<float>(input3, *gpu_place);
+  input1_gpu.CopyFrom<float>(input1, *gpu_place, context);
+  input2_gpu.CopyFrom<float>(input2, *gpu_place, context);
+  input3_gpu.CopyFrom<float>(input3, *gpu_place, context);
   float* a = input1_gpu.data<float>();
   float* b = input2_gpu.data<float>();
   float* c = input3_gpu.mutable_data<float>(*gpu_place);
@@ -108,7 +110,7 @@ TEST(math_function, gemm_notrans_cublas) {
   paddle::operators::math::gemm<paddle::platform::GPUPlace, float>(
       context, false, false, m, n, k, 1, a, 3, b + 1, 4, 1, c + 1, 4);
 
-  input3.CopyFrom<float>(input3_gpu, *cpu_place);
+  input3.CopyFrom<float>(input3_gpu, *cpu_place, context);
 
   // numpy code:
   // a = np.arange(6).reshape(2, 3)
@@ -116,6 +118,7 @@ TEST(math_function, gemm_notrans_cublas) {
   // c = np.arange(8).reshape(2, 4)[:, 1:]
   // out = np.arange(8).reshape(2, 4)
   // out[:, 1:] = np.dot(a, b) + c
+  context.Wait();
   EXPECT_EQ(input3_ptr[0], 0);
   EXPECT_EQ(input3_ptr[1], 24);
   EXPECT_EQ(input3_ptr[2], 28);
@@ -152,9 +155,9 @@ TEST(math_function, gemm_trans_cublas) {
   auto* gpu_place = new paddle::platform::GPUPlace(0);
   paddle::platform::CUDADeviceContext context(*gpu_place);
 
-  input1_gpu.CopyFrom<float>(input1, *gpu_place);
-  input2_gpu.CopyFrom<float>(input2, *gpu_place);
-  input3_gpu.CopyFrom<float>(input3, *gpu_place);
+  input1_gpu.CopyFrom<float>(input1, *gpu_place, context);
+  input2_gpu.CopyFrom<float>(input2, *gpu_place, context);
+  input3_gpu.CopyFrom<float>(input3, *gpu_place, context);
   float* a = input1_gpu.data<float>();
   float* b = input2_gpu.data<float>();
   float* c = input3_gpu.mutable_data<float>(*gpu_place);
@@ -162,7 +165,8 @@ TEST(math_function, gemm_trans_cublas) {
   paddle::operators::math::gemm<paddle::platform::GPUPlace, float>(
       context, false, true, m, n, k, 1, a, 3, b + 3, 3, 1, c + 1, 4);
 
-  input3.CopyFrom<float>(input3_gpu, *cpu_place);
+  input3.CopyFrom<float>(input3_gpu, *cpu_place, context);
+  context.Wait();
 
   EXPECT_EQ(input3_ptr[0], 0);
   EXPECT_EQ(input3_ptr[1], 24);

paddle/operators/math/vol2col_test.cc

@@ -78,7 +78,7 @@ void testVol2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     input = input_tmp;
   } else {
-    input.CopyFrom<float>(input_tmp, *place);
+    input.CopyFrom<float>(input_tmp, *place, *context);
   }
   output.mutable_data<float>({1, filter_size, filter_size, filter_size,
                               output_depth, output_height, output_width},
@@ -93,7 +93,7 @@ void testVol2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     out_cfo_ptr = output.data<float>();
   } else {
-    output_tmp.CopyFrom<float>(output, paddle::platform::CPUPlace());
+    output_tmp.CopyFrom<float>(output, paddle::platform::CPUPlace(), *context);
     out_cfo_ptr = output_tmp.data<float>();
   }
 
@@ -107,7 +107,7 @@ void testVol2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     input = input_tmp;
   } else {
-    input.CopyFrom<float>(input_tmp, *place);
+    input.CopyFrom<float>(input_tmp, *place, *context);
   }
 
   paddle::operators::math::Col2VolFunctor<Place, float> col2vol;
@@ -118,7 +118,7 @@ void testVol2col() {
   if (paddle::platform::is_cpu_place(*place)) {
     in_ptr = input.data<float>();
   } else {
-    input_tmp.CopyFrom<float>(input, paddle::platform::CPUPlace());
+    input_tmp.CopyFrom<float>(input, paddle::platform::CPUPlace(), *context);
     in_ptr = input_tmp.data<float>();
   }
 

paddle/operators/multiplex_op.cu

@@ -33,7 +33,8 @@ class MultiplexGPUKernel : public framework::OpKernel<T> {
     auto cols = ins[0]->numel() / rows;
     // copy index to cpu
     Tensor index_t_cpu;
-    index_t_cpu.CopyFrom<int32_t>(*ids, platform::CPUPlace());
+    index_t_cpu.CopyFrom<int32_t>(*ids, platform::CPUPlace(),
+                                  ctx.device_context());
     auto* index = index_t_cpu.data<int32_t>();
     auto stream = reinterpret_cast<const platform::CUDADeviceContext&>(
                       ctx.device_context())
@@ -70,7 +71,8 @@ class MultiplexGradGPUKernel : public framework::OpKernel<T> {
     auto cols = ins[0]->numel() / rows;
     // copy index to cpu
     Tensor index_t_cpu;
-    index_t_cpu.CopyFrom<int32_t>(*ids, platform::CPUPlace());
+    index_t_cpu.CopyFrom<int32_t>(*ids, platform::CPUPlace(),
+                                  ctx.device_context());
     auto* index = index_t_cpu.data<int32_t>();
 
     auto stream = reinterpret_cast<const platform::CUDADeviceContext&>(

paddle/operators/recurrent_op.cc

@@ -46,7 +46,7 @@ void RecurrentAlgorithm::Run(const Scope& scope,
     }
     (*stepnet_)->Run(*step_scopes[step_id], dev_ctx);
   }
-  rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len);
+  rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len, dev_ctx);
 }
 
 void RecurrentAlgorithm::CreateScopes(const Scope& scope,
@@ -151,12 +151,12 @@ void RecurrentGradientAlgorithm::Run(
   auto& step_scopes = GetStepScopes(scope);
   rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len);
   for (int step_id = seq_len - 1; step_id >= 0; --step_id) {
-    if (step_id != seq_len - 1) {
+    if (static_cast<size_t>(step_id) != seq_len - 1) {
       rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1);
     }
     (*stepnet_)->Run(*step_scopes[step_id], dev_ctx);
   }
-  rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len);
+  rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len, dev_ctx);
   LinkBootMemoryGradients(step_scopes[0]);
 }
 

paddle/operators/reshape_op.h

@@ -33,7 +33,7 @@ class ReshapeKernel : public framework::OpKernel<T> {
     std::transform(shape.begin(), shape.end(), shape_int64.begin(),
                    [](int a) { return static_cast<int64_t>(a); });
     auto out_dims = framework::make_ddim(shape_int64);
-    out->CopyFrom<T>(*in, ctx.GetPlace());
+    out->CopyFrom<T>(*in, ctx.GetPlace(), ctx.device_context());
     out->Resize(out_dims);
   }
 };
@@ -47,7 +47,7 @@ class ReshapeGradKernel : public framework::OpKernel<T> {
     d_x->mutable_data<T>(ctx.GetPlace());
 
     auto in_dims = d_x->dims();
-    d_x->CopyFrom<T>(*d_out, ctx.GetPlace());
+    d_x->CopyFrom<T>(*d_out, ctx.GetPlace(), ctx.device_context());
     d_x->Resize(in_dims);
   }
 };

paddle/operators/rnn/recurrent_op_utils.cc

@@ -51,7 +51,7 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
 
 void ConcatOutputs(const std::vector<Scope*>& step_scopes,
                    const std::vector<std::string>& outlinks,
-                   const size_t seq_len) {
+                   const size_t seq_len, const platform::DeviceContext& ctx) {
   for (size_t i = 0; i < outlinks.size(); i++) {
     auto* output_var = step_scopes[0]->parent().FindVar(outlinks[i]);
     PADDLE_ENFORCE_NOT_NULL(output_var, "output link [%s] is not in scope.",
@@ -72,7 +72,7 @@ void ConcatOutputs(const std::vector<Scope*>& step_scopes,
       // TODO(luotao02) data type and platform::DeviceContext() should set
       // correctly
       (output->Slice<float>(j, j + 1))
-          .CopyFrom<float>(*step_output, platform::CPUPlace());
+          .CopyFrom<float>(*step_output, platform::CPUPlace(), ctx);
     }
   }
 }

paddle/operators/rnn/recurrent_op_utils.h

@@ -71,7 +71,7 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
  */
 void ConcatOutputs(const std::vector<Scope*>& step_scopes,
                    const std::vector<std::string>& outlinks,
-                   const size_t seq_len);
+                   const size_t seq_len, const platform::DeviceContext& ctx);
 
 void LinkMemories(const std::vector<Scope*>& step_scopes,
                   const std::vector<MemoryAttr>& memories, const size_t step_id,

paddle/operators/sequence_concat_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/sequence_concat_op.h"
+
+namespace paddle {
+namespace operators {
+
+class SequenceConcatOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInputs("X"),
+                   "Inputs(X) of SequenceConcatOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of SequenceConcatOp should not be null.");
+    const size_t level = static_cast<size_t>(ctx->Attrs().Get<int>("level"));
+    const size_t axis = static_cast<size_t>(ctx->Attrs().Get<int>("axis"));
+    PADDLE_ENFORCE(level == 0UL || level == 1UL,
+                   "The sequence_concat operator only accepts sequence "
+                   "or a nested sequence as its input.");
+    auto ins_dims = ctx->GetInputsDim("X");
+    framework::DDim out_dims = ins_dims[0];
+    const size_t n = ins_dims.size();
+    for (size_t i = 1; i < n; ++i) {
+      out_dims[axis] += ins_dims[i][axis];
+    }
+    ctx->SetOutputDim("Out", out_dims);
+  }
+};
+
+class SequenceConcatOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SequenceConcatOpMaker(framework::OpProto* proto,
+                        framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(A vector of LoDTensor), the input is a vector of LoDTensor, "
+             "each of which is a variable-length sequence or nested sequence.")
+        .AsDuplicable();
+    AddOutput("Out",
+              "(A LoDTensor), the variable-length output of "
+              "sequence_concat Op.");
+    AddAttr<int>("axis",
+                 "(int, default 0)"
+                 "The axis which the inputs will be joined with. "
+                 "If axis is 0, the inputs will be joined with LoD index.")
+        .SetDefault(0);
+    AddAttr<int>("level",
+                 "(int, default 0)"
+                 "The level at which the inputs will be joined. "
+                 "If the level is 0, the inputs will be joined at the nested "
+                 "sequence level. "
+                 "If the level is 1, the inputs will be joined at the "
+                 "sequence level. "
+                 "The level should be less than the level number of inputs.")
+        .SetDefault(0);
+    AddComment(R"DOC(
+    The sequence_concat operator concatenates multiple LoDTensors. 
+    It only supports sequence (LoD Tensor with level number is 1) 
+    or a nested sequence (LoD tensor with level number is 2) as its input.
+    - Case1:
+      If the axis is other than 0(here, axis is 1 and level is 1),
+      each input should have the same LoD information and the LoD 
+      information of the output keeps the same as the input.
+
+      LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)
+      LoD(x1) = {{0,2,4}, {0,1,2,3,4}}; Dims(x1) = (4,4,4)
+      LoD(Out) = {{0,2,4}, {0,1,2,3,4}}; Dims(Out) = (4,7,4)
+
+    - Case2:
+      If the axis is 0(here, leve is 0), the inputs are concatenated along 
+      time steps, the LoD information of the output need to re-compute.
+
+      LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)
+      LoD(x1) = {{0,3,5}, {0,1,2,3,5}}; Dims(x1) = (5,3,4)
+      LoD(Out) = {{0,5,9}, {0,1,2,3,4,5,6,7,9}}; Dims(Out) = (9,3,4)
+
+    - Case3:
+      If the axis is 0(here, level is 1).
+
+      LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)
+      LoD(x1) = {{0,3,5}, {0,1,3,4,5}}; Dims(x1) = (5,3,4)
+      LoD(Out) = {{0,5,9}, {0,2,5,7,9}}; Dims(Out) = (9,3,4)
+      
+    NOTE: The levels of all the inputs should be the same.
+    )DOC");
+  }
+};
+
+class SequenceConcatGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "The gradient of Out should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName("X")),
+                   "The gradient of X should not be null.");
+    ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X"));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(sequence_concat, ops::SequenceConcatOp, ops::SequenceConcatOpMaker,
+            sequence_concat_grad, ops::SequenceConcatGradOp);
+REGISTER_OP_CPU_KERNEL(
+    sequence_concat,
+    ops::SequenceConcatOpKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    sequence_concat_grad,
+    ops::SequenceConcatGradOpKernel<paddle::platform::CPUPlace, float>);

paddle/operators/sequence_concat_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/sequence_concat_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(
+    sequence_concat,
+    ops::SequenceConcatOpKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    sequence_concat_grad,
+    ops::SequenceConcatGradOpKernel<paddle::platform::GPUPlace, float>);

paddle/operators/sequence_concat_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/op_registry.h"
+#include "paddle/operators/strided_memcpy.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+using LoD = framework::LoD;
+
+template <typename T>
+LoD concatLoD(const std::vector<const T*> ins, const size_t axis,
+              const size_t level) {
+  auto out_lod = ins[0]->lod();
+  const size_t n = ins.size();
+  if (axis == 0UL) {
+    for (size_t i = 1; i < n; ++i) {
+      for (size_t j = 0; j < ins[i]->lod()[0].size(); ++j) {
+        out_lod[0][j] += ins[i]->lod()[0][j];
+      }
+
+      if (ins[0]->NumLevels() == 2) {
+        for (size_t j = 1; j < ins[i]->lod()[1].size(); ++j) {
+          if (level == 0UL) {
+            out_lod[1].push_back(out_lod[1].back() + ins[i]->lod()[1][j] -
+                                 ins[i]->lod()[1][j - 1]);
+          } else if (level == 1UL) {
+            out_lod[1][j] += ins[1]->lod()[1][j];
+          }
+        }
+      }
+    }
+  }
+  return out_lod;
+}
+
+template <typename Place, typename T>
+class SequenceConcatOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto ins = ctx.MultiInput<LoDTensor>("X");
+    auto* out = ctx.Output<LoDTensor>("Out");
+    const size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
+    const size_t level = static_cast<size_t>(ctx.Attr<int>("level"));
+    const size_t n = ins.size();
+
+    for (size_t i = 1; i < n; ++i) {
+      PADDLE_ENFORCE_EQ(ins[0]->NumLevels(), ins[i]->NumLevels(),
+                        "The levels of all the input LoDTensors "
+                        "should be the same.");
+      PADDLE_ENFORCE_EQ(ins[0]->dims().size(), ins[i]->dims().size(),
+                        "The dimension size of all the input LoDTensors "
+                        "should be the same.");
+
+      const size_t dims_size = ins[i]->dims().size();
+      for (size_t j = 0; j < dims_size; ++j) {
+        if (j == axis) continue;
+        PADDLE_ENFORCE_EQ(ins[0]->dims()[j], ins[i]->dims()[j],
+                          "Except for the dimension of the specified "
+                          "axis along which all the inputs are concatenated, "
+                          "dimensions of all the other axises of the input "
+                          "LoDTensors should be the same.");
+      }
+    }
+    PADDLE_ENFORCE_GT(ins[0]->NumLevels(), level,
+                      "The levels of all the input LoDTensors "
+                      "should be greater than the specify level");
+
+    out->mutable_data<T>(ctx.GetPlace());
+    auto out_lod = concatLoD<LoDTensor>(ins, axis, level);
+    out->set_lod(out_lod);
+
+    auto out_lod_level = out_lod[level];
+    for (size_t i = 0; i < out_lod_level.size() - 1; ++i) {
+      Tensor out_t = out->Slice<T>(static_cast<int>(out_lod_level[i]),
+                                   static_cast<int>(out_lod_level[i + 1]));
+      auto out_stride = framework::stride(out_t.dims());
+      size_t offset = 0;
+
+      for (size_t j = 0; j < n; ++j) {
+        auto in_lod_level = ins[j]->lod()[level];
+        auto in_stride = framework::stride(ins[j]->dims());
+        Tensor in_t = ins[j]->Slice<T>(static_cast<int>(in_lod_level[i]),
+                                       static_cast<int>(in_lod_level[i + 1]));
+        size_t axis_dim = in_t.dims()[axis];
+        StridedMemcpy<T>(ctx.device_context(), in_t.data<T>(), in_stride,
+                         in_t.dims(), out_stride, out_t.data<T>() + offset);
+        offset += axis_dim * in_stride[axis];
+      }
+    }
+  }
+};
+
+template <typename Place, typename T>
+class SequenceConcatGradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto ins = ctx.MultiInput<framework::LoDTensor>("X");
+    auto* out_grad =
+        ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
+    auto x_grads =
+        ctx.MultiOutput<framework::LoDTensor>(framework::GradVarName("X"));
+    size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
+    size_t level = static_cast<size_t>(ctx.Attr<int>("level"));
+    const size_t n = x_grads.size();
+
+    // Set Grad(X) LoD as X
+    for (size_t i = 0; i < n; i++) {
+      x_grads[i]->set_lod(ins[i]->lod());
+      x_grads[i]->mutable_data<T>(ctx.GetPlace());
+    }
+
+    auto out_lod = concatLoD<LoDTensor>(ins, axis, level);
+    auto out_lod_level = out_lod[level];
+
+    for (size_t i = 0; i < out_lod_level.size() - 1; ++i) {
+      Tensor out_grad_t =
+          out_grad->Slice<T>(static_cast<int>(out_lod_level[i]),
+                             static_cast<int>(out_lod_level[i + 1]));
+      auto out_grad_stride = framework::stride(out_grad_t.dims());
+      size_t offset = 0;
+
+      for (size_t j = 0; j < n; ++j) {
+        auto x_grad_lod_level = x_grads[j]->lod()[level];
+        auto x_grad_stride = framework::stride(x_grads[j]->dims());
+        Tensor x_grad_t =
+            x_grads[j]->Slice<T>(static_cast<int>(x_grad_lod_level[i]),
+                                 static_cast<int>(x_grad_lod_level[i + 1]));
+        size_t axis_dim = x_grad_t.dims()[axis];
+        StridedMemcpy<T>(ctx.device_context(), out_grad_t.data<T>() + offset,
+                         out_grad_stride, out_grad_t.dims(), x_grad_stride,
+                         x_grad_t.data<T>());
+        offset += axis_dim * out_grad_stride[axis];
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/platform/cudnn_helper.h

@@ -71,23 +71,32 @@ class ScopedTensorDescriptor {
 
   inline cudnnTensorDescriptor_t descriptor(const cudnnTensorFormat_t format,
                                             const cudnnDataType_t type,
-                                            const std::vector<int>& dims) {
-    // the format is not used now, but it maybe useful feature
+                                            const std::vector<int>& dims,
+                                            const int groups = 1) {
+    // the format is not used now, will add later
     std::vector<int> strides(dims.size());
     strides[dims.size() - 1] = 1;
     for (int i = dims.size() - 2; i >= 0; i--) {
       strides[i] = dims[i + 1] * strides[i + 1];
     }
+    // Update tensor descriptor dims setting if groups > 1
+    // FIXME(typhoonzero): Assume using NCHW order
+    std::vector<int> dims_with_group(dims.begin(), dims.end());  // copy
+    if (groups > 1) {
+      dims_with_group[1] = dims_with_group[1] / groups;
+    }
     PADDLE_ENFORCE(dynload::cudnnSetTensorNdDescriptor(
-        desc_, type, dims.size(), dims.data(), strides.data()));
+        desc_, type, dims_with_group.size(), dims_with_group.data(),
+        strides.data()));
     return desc_;
   }
 
   template <typename T>
   inline cudnnTensorDescriptor_t descriptor(const DataLayout& order,
-                                            const std::vector<int>& dims) {
-    return descriptor(GetCudnnTensorFormat(order), CudnnDataType<T>::type,
-                      dims);
+                                            const std::vector<int>& dims,
+                                            const int groups = 1) {
+    return descriptor(GetCudnnTensorFormat(order), CudnnDataType<T>::type, dims,
+                      groups);
   }
 
  private:
@@ -106,18 +115,29 @@ class ScopedFilterDescriptor {
 
   inline cudnnFilterDescriptor_t descriptor(const cudnnTensorFormat_t format,
                                             const cudnnDataType_t type,
-                                            const std::vector<int>& kernel) {
-    // filter layout: output input spatial_dim_y spatial_dim_x
+                                            const std::vector<int>& kernel,
+                                            const int groups = 1) {
+    // filter layout: MCHW, where M is the number of
+    // output image channels, C is the number of input image channels,
+    // H and W is height and width of filter.
+    std::vector<int> kernel_with_group(kernel.begin(), kernel.end());
+    if (groups > 1) {
+      // M /= groups
+      kernel_with_group[0] /= groups;
+      // NOTE: input filter(C) of the filter is already asserted to be C/groups.
+    }
     PADDLE_ENFORCE(dynload::cudnnSetFilterNdDescriptor(
-        desc_, type, format, kernel.size(), kernel.data()));
+        desc_, type, format, kernel_with_group.size(),
+        kernel_with_group.data()));
     return desc_;
   }
 
   template <typename T>
   inline cudnnFilterDescriptor_t descriptor(const DataLayout& order,
-                                            const std::vector<int>& kernel) {
+                                            const std::vector<int>& kernel,
+                                            const int groups = 1) {
     return descriptor(GetCudnnTensorFormat(order), CudnnDataType<T>::type,
-                      kernel);
+                      kernel, groups);
   }
 
  private:

paddle/pybind/CMakeLists.txt

 if(WITH_PYTHON)
   cc_library(paddle_pybind SHARED
     SRCS pybind.cc exception.cc protobuf.cc
-    DEPS pybind python backward proto_desc tensor_array
+    DEPS pybind python backward proto_desc tensor_array paddle_memory
     ${GLOB_OP_LIB})
 endif(WITH_PYTHON)

paddle/pybind/protobuf.cc

@@ -15,6 +15,7 @@ limitations under the License. */
 #include "paddle/pybind/protobuf.h"
 #include <deque>
 #include <iostream>
+#include "paddle/framework/backward.h"
 #include "paddle/framework/block_desc.h"
 #include "paddle/framework/op_desc.h"
 #include "paddle/framework/program_desc.h"
@@ -116,6 +117,11 @@ void BindProgramDesc(py::module &m) {
                   py::return_value_policy::reference)
       .def("append_block", &ProgramDescBind::AppendBlock,
            py::return_value_policy::reference)
+      .def("append_backward",
+           [](ProgramDescBind &program_desc,
+              const std::unordered_set<std::string> &no_grad_vars) {
+             AppendBackward(program_desc, no_grad_vars);
+           })
       .def("block", &ProgramDescBind::Block, py::return_value_policy::reference)
       .def("num_blocks", &ProgramDescBind::Size);
 }
@@ -199,6 +205,7 @@ void BindOpDesc(py::module &m) {
       .def("attr", &OpDescBind::GetAttr)
       .def("set_block_attr", &OpDescBind::SetBlockAttr)
       .def("get_block_attr", &OpDescBind::GetBlockAttr)
+      .def("check_attrs", &OpDescBind::CheckAttrs)
       .def("infer_shape", &OpDescBind::InferShape);
 }
 

paddle/pybind/tensor_py.h

@@ -57,7 +57,18 @@ struct CastToPyBufferImpl<true, I, ARGS...> {
       }
       framework::Tensor dst_tensor;
       if (paddle::platform::is_gpu_place(tensor.place())) {
-        dst_tensor.CopyFrom<CUR_TYPE>(tensor, platform::CPUPlace());
+#ifdef PADDLE_WITH_CUDA
+        auto *src_ptr = static_cast<const void *>(tensor.data<CUR_TYPE>());
+        auto *dst_ptr = static_cast<void *>(dst_tensor.mutable_data<CUR_TYPE>(
+            tensor.dims(), platform::CPUPlace()));
+        // TODO(qijun): Here we use default CUDA stream to set GPU Tensor to
+        // a Python numpy array. It's better to manage CDUA stream unifiedly.
+        paddle::platform::GpuMemcpySync(dst_ptr, src_ptr,
+                                        sizeof(CUR_TYPE) * tensor.numel(),
+                                        cudaMemcpyDeviceToHost);
+#else
+        PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
+#endif
       } else if (paddle::platform::is_cpu_place(tensor.place())) {
         dst_tensor = tensor;
       }
@@ -120,6 +131,8 @@ void PyCUDATensorSetFromArray(
 
   self.Resize(framework::make_ddim(dims));
   auto *dst = self.mutable_data<T>(place);
+  // TODO(qijun): Here we use default CUDA stream to set a Python numpy
+  // array to a GPU Tensor. It's better to manage CDUA stream unifiedly.
   paddle::platform::GpuMemcpySync(dst, array.data(), sizeof(T) * array.size(),
                                   cudaMemcpyHostToDevice);
 }

proto/CMakeLists.txt

-file(GLOB proto_filenames . *.proto)
+if (MOBILE_INFERENCE)
+    file(GLOB proto_filenames . ModelConfig.proto ParameterConfig.proto
+         TrainerConfig.proto DataConfig.proto)
+else()
+    file(GLOB proto_filenames . *.proto)
+endif()
+
 include_directories(${CMAKE_CURRENT_BINARY_DIR})
 proto_library(paddle_proto SRCS ${proto_filenames})
 

python/paddle/trainer_config_helpers/layers.py

@@ -318,7 +318,7 @@ class LayerOutput(object):
     :param activation: Layer Activation.
     :type activation: BaseActivation.
     :param parents: Layer's parents.
-    :type parents: list|tuple|collections.Sequence
+    :type parents: list | tuple | collections.Sequence
     """
 
     def __init__(self,
@@ -435,7 +435,7 @@ def full_matrix_projection(input, size=0, param_attr=None):
                                      size=100,
                                      param_attr=ParamAttr(name='_proj'))
 
-    :param input: input layer
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param size: The parameter size. Means the width of parameter.
     :type size: int
@@ -471,7 +471,7 @@ def trans_full_matrix_projection(input, size=0, param_attr=None):
                                                 initial_mean=0.0,
                                                 initial_std=0.01))
 
-    :param input: input layer
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param size: The parameter size. Means the width of parameter.
     :type size: int
@@ -516,7 +516,7 @@ def table_projection(input, size=0, param_attr=None):
                                param_attr=ParamAttr(name='_proj'))
 
 
-    :param input: Input layer, which must contains id fields.
+    :param input: The input of this layer, which must contains id fields.
     :type input: LayerOutput
     :param size: The parameter size. Means the width of parameter.
     :type size: int
@@ -561,7 +561,7 @@ def identity_projection(input, offset=None, size=None):
 
     Note that both of two projections should not have any parameter.
 
-    :param input: Input Layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param offset: Offset, None if use default.
     :type offset: int
@@ -596,7 +596,7 @@ def slice_projection(input, slices):
 
     Note that slice_projection should not have any parameter.
 
-    :param input: Input Layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param slices: An array of slice parameters.
                    Each slice contains the start and end offsets based
@@ -634,7 +634,7 @@ def scaling_projection(input, param_attr=None):
 
        proj = scaling_projection(input=layer)
 
-    :param input: Input Layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param param_attr: Parameter config, None if use default.
     :type param_attr: ParameterAttribute
@@ -663,7 +663,7 @@ def dotmul_projection(input, param_attr=None):
 
        proj = dotmul_projection(input=layer)
 
-    :param input: Input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param param_attr: Parameter config, None if use default.
     :type param_attr: ParameterAttribute
@@ -734,7 +734,7 @@ def context_projection(input,
     after context projection and not set padding_attr, sequence will
     be [ 0AB ABC BCD CDE DEF EFG FG0 ].
 
-    :param input: Input Sequence.
+    :param input: The input of this layer, which should be a sequence.
     :type input: LayerOutput
     :param context_len: context length.
     :type context_len: int
@@ -744,7 +744,7 @@ def context_projection(input,
     :param padding_attr: Padding Parameter Attribute. If false, it means padding
                          always be zero. Otherwise Padding is learnable, and
                          parameter attribute is set by this parameter.
-    :type padding_attr: bool|ParameterAttribute
+    :type padding_attr: bool | ParameterAttribute
     :return: Projection
     :rtype: Projection
     """
@@ -782,13 +782,13 @@ class MixedLayerType(LayerOutput):
         :type name: basestring
         :param size: layer size.
         :type size: int
-        :param act: activation type.
+        :param act: Activation type.
         :type act: BaseActivation
         :param bias_attr: The Bias Attribute. If the parameter is set to
                           False or something not type of ParameterAttribute,
                           no bias is defined. If the parameter is set to
                           True, the bias is initialized to zero.
-        :type bias_attr: ParameterAttribute|None|Bool|Any
+        :type bias_attr: ParameterAttribute | None | bool | Any
         :param layer_attr: Extra Layer Attribute.
         :type layer_attr: ExtraLayerAttribute or None
         """
@@ -880,15 +880,15 @@ def mixed_layer(size=0,
     :type name: basestring
     :param size: layer size.
     :type size: int
-    :param input: inputs layer. It is an optional parameter. If set,
+    :param input: The input of this layer. It is an optional parameter. If set,
                   then this function will just return layer's name.
-    :param act: Activation Type.
+    :param act: Activation Type. LinearActivation is the default.
     :type act: BaseActivation
     :param bias_attr: The Bias Attribute. If the parameter is set to
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: The extra layer config. Default is None.
     :type layer_attr: ExtraLayerAttribute
     :return: MixedLayerType object can add inputs or layer name.
@@ -929,9 +929,9 @@ def data_layer(name, size, depth=None, height=None, width=None,
     :param size: Size of this data layer.
     :type size: int
     :param height: Height of this data layer, used for image
-    :type height: int|None
+    :type height: int | None
     :param width: Width of this data layer, used for image
-    :type width: int|None
+    :type width: int | None
     :param layer_attr: Extra Layer Attribute.
     :type layer_attr: ExtraLayerAttribute.
     :return: LayerOutput object.
@@ -966,15 +966,15 @@ def embedding_layer(input, size, name=None, param_attr=None, layer_attr=None):
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input layer for this embedding. NOTE: must be Index Data.
+    :param input: The input of this layer, which must be Index Data.
     :type input: LayerOutput
     :param size: The embedding dimension.
     :type size: int
     :param param_attr: The embedding parameter attribute. See ParameterAttribute
                       for details.
-    :type param_attr: ParameterAttribute|None
+    :type param_attr: ParameterAttribute | None
     :param layer_attr: Extra layer Config. Default is None.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -1021,11 +1021,11 @@ def fc_layer(input,
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input layer. Could be a list/tuple of input layer.
-    :type input: LayerOutput|list|tuple
+    :param input: The input of this layer.
+    :type input: LayerOutput | list | tuple
     :param size: The layer dimension.
     :type size: int
-    :param act: Activation Type. Default is tanh.
+    :param act: Activation Type. TanhActivation is the default.
     :type act: BaseActivation
     :param param_attr: The Parameter Attribute|list.
     :type param_attr: ParameterAttribute
@@ -1033,9 +1033,9 @@ def fc_layer(input,
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: Extra Layer config.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -1072,8 +1072,8 @@ def printer_layer(input, format=None, name=None):
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input layer. Could be a list/tuple of input layer.
-    :type input: LayerOutput|list|tuple
+    :param input: The input of this layer.
+    :type input: LayerOutput | list | tuple
     :return: LayerOutput
     """
     if isinstance(input, LayerOutput):
@@ -1110,7 +1110,7 @@ def priorbox_layer(input,
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param image: The network input image.
     :type image: LayerOutput
@@ -1306,7 +1306,7 @@ def cross_channel_norm_layer(input, name=None, param_attr=None):
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param param_attr: The Parameter Attribute|list.
     :type param_attr: ParameterAttribute
@@ -1371,20 +1371,20 @@ def pooling_layer(input,
     :type agg_level: AggregateLevel
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: input layer name.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param pooling_type: Type of pooling, MaxPooling(default), AvgPooling,
                          SumPooling, SquareRootNPooling.
-    :type pooling_type: BasePoolingType|None
+    :type pooling_type: BasePoolingType | None
     :param stride: The step size between successive pooling regions.
     :type stride: Int
     :param bias_attr: The Bias Attribute. If the parameter is set to
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: The Extra Attributes for layer, such as dropout.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -1469,11 +1469,11 @@ def lstmemory(input,
     :type name: basestring
     :param size: DEPRECATED. size of the lstm cell
     :type size: int
-    :param input: input layer name.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param reverse: is sequence process reversed or not.
     :type reverse: bool
-    :param act: activation type, TanhActivation by default. :math:`h_t`
+    :param act: Activation type. TanhActivation is the default. :math:`h_t`
     :type act: BaseActivation
     :param gate_act: gate activation type, SigmoidActivation by default.
     :type gate_act: BaseActivation
@@ -1483,11 +1483,11 @@ def lstmemory(input,
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr: Parameter Attribute.
-    :type param_attr: ParameterAttribute|None|False
+    :type param_attr: ParameterAttribute | None | False
     :param layer_attr: Extra Layer attribute
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -1591,14 +1591,14 @@ def grumemory(input,
        gru = grumemory(input)
 
     :param name: The gru layer name.
-    :type name: None|basestring
-    :param input: input layer.
+    :type name: None | basestring
+    :param input: The input of this layer.
     :type input: LayerOutput.
     :param size: DEPRECATED. size of the gru cell
     :type size: int
     :param reverse: Whether sequence process is reversed or not.
     :type reverse: bool
-    :param act: activation type, TanhActivation by default. This activation
+    :param act: Activation type, TanhActivation is the default. This activation
                 affects the :math:`{\\tilde{h_t}}`.
     :type act: BaseActivation
     :param gate_act: gate activation type, SigmoidActivation by default.
@@ -1609,11 +1609,11 @@ def grumemory(input,
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr: Parameter Attribute.
-    :type param_attr: ParameterAttribute|None|False
+    :type param_attr: ParameterAttribute | None | False
     :param layer_attr: Extra Layer attribute
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -1670,7 +1670,7 @@ def last_seq(input,
     :param agg_level: Aggregated level
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: Input layer name.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param stride: The step size between successive pooling regions.
     :type stride: Int
@@ -1726,7 +1726,7 @@ def first_seq(input,
     :param agg_level: aggregation level
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: Input layer name.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param stride: The step size between successive pooling regions.
     :type stride: Int
@@ -1799,7 +1799,7 @@ def expand_layer(input,
                              expand_as=layer2,
                              expand_level=ExpandLevel.FROM_NO_SEQUENCE)
 
-    :param input: Input layer
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param expand_as: Expand as this layer's sequence info.
     :type expand_as: LayerOutput
@@ -1809,7 +1809,7 @@ def expand_layer(input,
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param expand_level: whether input layer is timestep(default) or sequence.
     :type expand_level: ExpandLevel
     :param layer_attr: extra layer attributes.
@@ -1858,7 +1858,7 @@ def repeat_layer(input,
 
        expand = repeat_layer(input=layer, num_repeats=4)
 
-    :param input: Input layer
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param num_repeats: Repeat the input so many times
     :type num_repeats: int
@@ -1869,7 +1869,7 @@ def repeat_layer(input,
                           False for treating input as column vector and repeating
                           in the row direction.
     :type as_row_vector: bool
-    :param act: Activation type.
+    :param act: Activation type. IdentityActivation is the default.
     :type act: BaseActivation
     :type name: basestring
     :param layer_attr: extra layer attributes.
@@ -1917,13 +1917,13 @@ def seq_reshape_layer(input,
 
        reshape = seq_reshape_layer(input=layer, reshape_size=4)
 
-    :param input: Input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param reshape_size: the size of reshaped sequence.
     :type reshape_size: int
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param act: Activation type.
+    :param act: Activation type. IdentityActivation is the default.
     :type act: BaseActivation
     :param layer_attr: extra layer attributes.
     :type layer_attr: ExtraLayerAttribute.
@@ -1931,7 +1931,7 @@ def seq_reshape_layer(input,
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -1970,8 +1970,8 @@ def interpolation_layer(input, weight, name=None, layer_attr=None):
 
        interpolation = interpolation_layer(input=[layer1, layer2], weight=layer3)
 
-    :param input: Input layer.
-    :type input: list|tuple
+    :param input: The input of this layer.
+    :type input: list | tuple
     :param weight: Weight layer.
     :type weight: LayerOutput
     :param name: The name of this layer. It is optional.
@@ -2023,11 +2023,11 @@ def bilinear_interp_layer(input,
     :param   input:        A input layer.
     :type    input:        LayerOutput.
     :param   out_size_x:   bilinear interpolation output width.
-    :type    out_size_x:   int|None
+    :type    out_size_x:   int | None
     :param   out_size_y:   bilinear interpolation output height.
-    :type    out_size_y:   int|None
+    :type    out_size_y:   int | None
     :param   name:         The layer's name, which cna not be specified.
-    :type    name:         None|basestring
+    :type    name:         None | basestring
     :param   layer_attr:   Extra Layer attribute.
     :type    layer_attr:   ExtraLayerAttribute
     :return: LayerOutput object.
@@ -2075,7 +2075,7 @@ def power_layer(input, weight, name=None, layer_attr=None):
 
        power = power_layer(input=layer1, weight=layer2)
 
-    :param input: Input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param weight: Weight layer.
     :type weight: LayerOutput
@@ -2119,7 +2119,7 @@ def scaling_layer(input, weight, name=None, layer_attr=None):
 
        scale = scaling_layer(input=layer1, weight=layer2)
 
-    :param input: Input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param weight: Weight layer.
     :type weight: LayerOutput
@@ -2159,7 +2159,7 @@ def trans_layer(input, name=None, layer_attr=None):
 
        trans = trans_layer(input=layer)
 
-    :param input: Input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param name: The name of this layer. It is optional.
     :type name: basestring
@@ -2197,7 +2197,7 @@ def rotate_layer(input, height, width, name=None, layer_attr=None):
                           height=100,
                           width=100)
 
-    :param input: Input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param height: The height of the sample matrix
     :type height: int
@@ -2306,22 +2306,21 @@ def hsigmoid(input,
         cost = hsigmoid(input=[layer1, layer2],
                         label=data_layer)
 
-    :param input: Input layers. It could be a LayerOutput or list/tuple of
-                 LayerOutput.
-    :type input: LayerOutput|list|tuple
+    :param input: The input of this layer.
+    :type input: LayerOutput | list | tuple
     :param label: Label layer.
     :type label: LayerOutput
     :param num_classes: number of classes.
-    :type num_classes: int|None
+    :type num_classes: int | None
     :param name: The name of this layer. It is optional.
     :type name: basestring
     :param bias_attr: The Bias Attribute. If the parameter is set to
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr: Parameter Attribute. None means default parameter.
-    :type param_attr: ParameterAttribute|None
+    :type param_attr: ParameterAttribute | None
     :param layer_attr: Extra Layer Attribute.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
@@ -2429,40 +2428,40 @@ def img_conv_layer(input,
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: Layer Input.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param filter_size: The x dimension of a filter kernel. Or input a tuple for
                         two image dimension.
-    :type filter_size: int|tuple|list
+    :type filter_size: int | tuple | list
     :param filter_size_y: The y dimension of a filter kernel. Since PaddlePaddle
                         currently supports rectangular filters, the filter's
                         shape will be (filter_size, filter_size_y).
-    :type filter_size_y: int|None
+    :type filter_size_y: int | None
     :param num_filters: Each filter group's number of filter
-    :param act: Activation type. Default is tanh
+    :param act: Activation type. ReluActivation is the default.
     :type act: BaseActivation
     :param groups: Group size of filters.
     :type groups: int
     :param stride: The x dimension of the stride. Or input a tuple for two image
                    dimension.
-    :type stride: int|tuple|list
+    :type stride: int | tuple | list
     :param stride_y: The y dimension of the stride.
     :type stride_y: int
     :param padding: The x dimension of the padding. Or input a tuple for two
                     image dimension
-    :type padding: int|tuple|list
+    :type padding: int | tuple | list
     :param padding_y: The y dimension of the padding.
     :type padding_y: int
     :param dilation: The x dimension of the dilation. Or input a tuple for two
                     image dimension
-    :type dilation: int|tuple|list
+    :type dilation: int | tuple | list
     :param dilation_y: The y dimension of the dilation.
     :type dilation_y: int
     :param bias_attr: The Bias Attribute. If the parameter is set to
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param num_channels: number of input channels. If None will be set
                         automatically from previous output.
     :type num_channels: int
@@ -2616,15 +2615,15 @@ def img_pool_layer(input,
     :param padding: pooling padding width.
     :type padding: int
     :param padding_y: pooling padding height. It's equal to padding by default.
-    :type padding_y: int|None
+    :type padding_y: int | None
     :param name: name of pooling layer
     :type name: basestring.
-    :param input: layer's input
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param pool_size: pooling window width
     :type pool_size: int
     :param pool_size_y: pooling window height. It's eaqual to pool_size by default.
-    :type pool_size_y: int|None
+    :type pool_size_y: int | None
     :param num_channels: number of input channel.
     :type num_channels: int
     :param pool_type: pooling type. MaxPooling or AvgPooling. Default is
@@ -2633,7 +2632,7 @@ def img_pool_layer(input,
     :param stride: stride width of pooling.
     :type stride: int
     :param stride_y: stride height of pooling. It is equal to stride by default.
-    :type stride_y: int|None
+    :type stride_y: int | None
     :param layer_attr: Extra Layer attribute.
     :type layer_attr: ExtraLayerAttribute
     :param ceil_mode: Wether to use ceil mode to calculate output height and with.
@@ -2743,20 +2742,20 @@ def img_pool3d_layer(input,
                                  pool_type=MaxPooling())
 
     :param padding: pooling padding width.
-    :type padding: int|tuple|list
+    :type padding: int | tuple | list
     :param name: name of pooling layer
     :type name: basestring.
-    :param input: layer's input
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param pool_size: pooling window width
-    :type pool_size: int|tuple|list
+    :type pool_size: int | tuple | list
     :param num_channels: number of input channel.
     :type num_channels: int
     :param pool_type: pooling type. MaxPooling or AvgPooling. Default is
                       MaxPooling.
     :type pool_type: BasePoolingType
     :param stride: stride width of pooling.
-    :type stride: int|tuple|list
+    :type stride: int | tuple | list
     :param layer_attr: Extra Layer attribute.
     :type layer_attr: ExtraLayerAttribute
     :param ceil_mode: Wether to use ceil mode to calculate output height and with.
@@ -2855,7 +2854,7 @@ def spp_layer(input,
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: layer's input.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param num_channels: number of input channel.
     :type num_channels: int
@@ -2948,8 +2947,8 @@ def img_cmrnorm_layer(input,
         norm = img_cmrnorm_layer(input=net, size=5)
 
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring
-    :param input: layer's input.
+    :type name: None | basestring
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param size: Normalize in number of :math:`size` feature maps.
     :type size: int
@@ -3024,7 +3023,7 @@ def batch_norm_layer(input,
                             batch_norm for CPU. Otherwise, select batch norm
                             type based on the specified type. If you use cudnn_batch_norm,
                             we suggested you use latest version, such as v5.1.
-    :type batch_norm_type: None|string, None or "batch_norm" or "cudnn_batch_norm"
+    :type batch_norm_type: None | string, None or "batch_norm" or "cudnn_batch_norm"
     :param act: Activation Type. Better be relu. Because batch
                      normalization will normalize input near zero.
     :type act: BaseActivation
@@ -3034,7 +3033,7 @@ def batch_norm_layer(input,
     :type num_channels: int
     :param bias_attr: :math:`\\beta`, better be zero when initialize. So the
                       initial_std=0, initial_mean=1 is best practice.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr: :math:`\\gamma`, better be one when initialize. So the
                        initial_std=0, initial_mean=1 is best practice.
     :type param_attr: ParameterAttribute
@@ -3046,7 +3045,7 @@ def batch_norm_layer(input,
                              testing. If False, it will use the mean
                              and variance of current batch of test data for
                              testing.
-    :type use_global_stats: bool|None.
+    :type use_global_stats: bool | None.
     :param moving_average_fraction: Factor used in the moving average
                                    computation, referred to as facotr,
                                    :math:`runningMean = newMean*(1-factor)
@@ -3107,7 +3106,7 @@ def sum_to_one_norm_layer(input, name=None, layer_attr=None):
 
        sum_to_one_norm = sum_to_one_norm_layer(input=layer)
 
-    :param input: Input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param name: The name of this layer. It is optional.
     :type name: basestring
@@ -3143,7 +3142,7 @@ def row_l2_norm_layer(input, name=None, layer_attr=None):
 
        row_l2_norm_layer = row_l2_norm_layer(input=layer)
 
-    :param input: Input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param name: The name of this layer. It is optional.
     :type name: basestring
@@ -3201,14 +3200,14 @@ def addto_layer(input, act=None, name=None, bias_attr=None, layer_attr=None):
     :type name: basestring
     :param input: Input layers. It could be a LayerOutput or list/tuple of
                  LayerOutput.
-    :type input: LayerOutput|list|tuple
-    :param act: Activation Type, default is tanh.
+    :type input: LayerOutput | list | tuple
+    :param act: Activation Type. LinearActivation is the default.
     :type act: BaseActivation
     :param bias_attr: The Bias Attribute. If the parameter is set to
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: Extra Layer attribute.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
@@ -3260,8 +3259,8 @@ def concat_layer(input, act=None, name=None, layer_attr=None, bias_attr=None):
     :param name: The name of this layer. It is optional.
     :type name: basestring
     :param input: input layers or projections
-    :type input: list|tuple|collections.Sequence
-    :param act: Activation type.
+    :type input: list | tuple | collections.Sequence
+    :param act: Activation type. IdentityActivation is the default.
     :type act: BaseActivation
     :param layer_attr: Extra Layer Attribute.
     :type layer_attr: ExtraLayerAttribute
@@ -3356,7 +3355,7 @@ def seq_concat_layer(a, b, act=None, name=None, layer_attr=None,
     :type a: LayerOutput
     :param b: input sequence layer
     :type b: LayerOutput
-    :param act: Activation type.
+    :param act: Activation type. IdentityActivation is the default.
     :type act: BaseActivation
     :param layer_attr: Extra Layer Attribute.
     :type layer_attr: ExtraLayerAttribute
@@ -3364,7 +3363,7 @@ def seq_concat_layer(a, b, act=None, name=None, layer_attr=None,
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -3440,9 +3439,9 @@ def memory(name,
     :param is_seq: DEPRECATED. is sequence for boot_layer
     :type is_seq: bool
     :param boot_layer: boot layer of memory.
-    :type boot_layer: LayerOutput|None
+    :type boot_layer: LayerOutput | None
     :param boot_bias: boot layer's bias
-    :type boot_bias: ParameterAttribute|None
+    :type boot_bias: ParameterAttribute | None
     :param boot_bias_active_type: boot layer's active type.
     :type boot_bias_active_type: BaseActivation
     :param boot_with_const_id: boot layer's id.
@@ -3537,19 +3536,17 @@ def lstm_step_layer(input,
     :type input: LayerOutput
     :param state: State Layer. :math:`c_{t-1}`
     :type state: LayerOutput
-    :param act: Activation type. Default is tanh
+    :param act: Activation type. TanhActivation is the default.
     :type act: BaseActivation
-    :param gate_act: Gate Activation Type. Default is sigmoid, and should
-                          be sigmoid only.
+    :param gate_act: Gate Activation Type. SigmoidActivation is the default.
     :type gate_act: BaseActivation
-    :param state_act: State Activation Type. Default is sigmoid, and should
-                           be sigmoid only.
+    :param state_act: State Activation Type. TanhActivation is the default.
     :type state_act: BaseActivation
     :param bias_attr: The Bias Attribute. If the parameter is set to
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: layer's extra attribute.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
@@ -3600,13 +3597,15 @@ def gru_step_layer(input,
     :param output_mem:
     :param size:
     :param act:
+    :type act: BaseActivation
     :param name: The name of this layer. It is optional.
-    :param gate_act:
+    :param gate_act: Activation type of this layer's two gates. Default is Sigmoid.
+    :type gate_act: BaseActivation
     :param bias_attr: The Bias Attribute. If the parameter is set to
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr: the parameter_attribute for transforming the output_mem
                        from previous step.
     :param layer_attr:
@@ -3662,12 +3661,14 @@ def gru_step_naive_layer(input,
     :param size:
     :param name: The name of this layer. It is optional.
     :param act:
-    :param gate_act:
+    :type act: BaseActivation
+    :param gate_act: Activation type of this layer's two gates. Default is Sigmoid.
+    :type gate_act: BaseActivation
     :param bias_attr: The Bias Attribute. If the parameter is set to
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr:
     :param layer_attr:
     :return:
@@ -3786,15 +3787,15 @@ def recurrent_layer(input,
         out_{i} = act(in_{i} + out_{i+1} * W) \\ \\ \\text{for} \\ start <= i < end
 
 
-    :param input: Input Layer
+    :param input: The input of this layer.
     :type input: LayerOutput
-    :param act: activation.
+    :param act: Activation type. TanhActivation is the default.
     :type act: BaseActivation
     :param bias_attr: The Bias Attribute. If the parameter is set to
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param param_attr: parameter attribute.
     :type param_attr: ParameterAttribute
     :param name: The name of this layer. It is optional.
@@ -3901,7 +3902,7 @@ def recurrent_group(step, input, reverse=False, name=None, targetInlink=None):
                   StaticInput will be imported to each time step, and doesn't change
                   through time. It's a mechanism to access layer outside step function.
 
-    :type input: LayerOutput|StaticInput|SubsequenceInput|list|tuple
+    :type input: LayerOutput | StaticInput | SubsequenceInput | list | tuple
 
     :param reverse: If reverse is set true, the recurrent unit will process the
                     input sequence in a reverse order.
@@ -3916,7 +3917,7 @@ def recurrent_group(step, input, reverse=False, name=None, targetInlink=None):
                          of words in each sentence) with all layer group's outputs.
                          targetInlink should be one of the layer group's input.
 
-    :type targetInlink: LayerOutput|SubsequenceInput
+    :type targetInlink: LayerOutput | SubsequenceInput
 
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -4034,7 +4035,7 @@ def maxid_layer(input, name=None, layer_attr=None):
 
        maxid = maxid_layer(input=layer)
 
-    :param input: Input layer name.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param name: The name of this layer. It is optional.
     :type name: basestring
@@ -4112,7 +4113,7 @@ def eos_layer(input, eos_id, name=None, layer_attr=None):
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: Input layer name.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param eos_id: end id of sequence
     :type eos_id: int
@@ -4504,7 +4505,7 @@ def conv_projection(input,
                               num_filters=64,
                               num_channels=64)
 
-    :param input: input layer
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param filter_size: The x dimension of a filter kernel.
     :type filter_size: int
@@ -4529,7 +4530,7 @@ def conv_projection(input,
     :param param_attr: Convolution param attribute. None means default attribute
     :type param_attr: ParameterAttribute
     :param trans: whether it is convTrans or conv
-    :type trans: boolean
+    :type trans: bool
     :return: A DotMulProjection Object.
     :rtype: DotMulProjection
     """
@@ -4637,14 +4638,14 @@ def pad_layer(input,
                        pad_h=[0,0],
                        pad_w=[2,2])
 
-    :param input: layer's input.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param pad_c: padding size in channel dimension.
-    :type pad_c: list|None
+    :type pad_c: list | None
     :param pad_h: padding size in height dimension.
-    :type pad_h: list|None
+    :type pad_h: list | None
     :param pad_w: padding size in width dimension.
-    :type pad_w: list|None
+    :type pad_w: list | None
     :param layer_attr: Extra Layer Attribute.
     :type layer_attr: ExtraLayerAttribute
     :param name: The name of this layer. It is optional.
@@ -4779,7 +4780,7 @@ def tensor_layer(a,
     :type b: LayerOutput
     :param size: the layer dimension.
     :type size: int.
-    :param act: Activation Type. Default is tanh.
+    :param act: Activation type. LinearActivation is the default.
     :type act: BaseActivation
     :param param_attr: The Parameter Attribute.
     :type param_attr: ParameterAttribute
@@ -4787,9 +4788,9 @@ def tensor_layer(a,
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: Extra Layer config.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -4836,15 +4837,15 @@ def selective_fc_layer(input,
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input layer.
-    :type input: LayerOutput|list|tuple
+    :param input: The input of this layer.
+    :type input: LayerOutput | list | tuple
     :param select: The select layer. The output of select layer should be a
                    sparse binary matrix, and treat as the mask of selective fc.
                    If is None, acts exactly like fc_layer.
     :type select: LayerOutput
     :param size: The layer dimension.
     :type size: int
-    :param act: Activation Type. Default is tanh.
+    :param act: Activation type. TanhActivation is the default.
     :type act: BaseActivation
     :param param_attr: The Parameter Attribute.
     :type param_attr: ParameterAttribute
@@ -4852,9 +4853,9 @@ def selective_fc_layer(input,
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: Extra Layer config.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -4906,12 +4907,12 @@ def sampling_id_layer(input, name=None, layer_attr=None):
 
        samping_id = sampling_id_layer(input=input)
 
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param name: The name of this layer. It is optional.
     :type name: basestring
     :param layer_attr: Extra Layer config.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -4944,7 +4945,7 @@ def slope_intercept_layer(input,
 
        scale = slope_intercept_layer(input=input, slope=-1.0, intercept=1.0)
 
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param name: The name of this layer. It is optional.
     :type name: basestring
@@ -4953,7 +4954,7 @@ def slope_intercept_layer(input,
     :param intercept: the offset.
     :type intercept: float.
     :param layer_attr: Extra Layer config.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -5013,7 +5014,7 @@ def linear_comb_layer(weights, vectors, size=None, name=None, layer_attr=None):
     :param name: The name of this layer. It is optional.
     :type name: basestring
     :param layer_attr: Extra Layer config.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -5077,10 +5078,10 @@ def block_expand_layer(input,
                                          block_x=1,
                                          block_x=3)
 
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param num_channels: The channel number of input layer.
-    :type num_channels: int|None
+    :type num_channels: int | None
     :param block_x: The width of sub block.
     :type block_x: int
     :param block_y: The width of sub block.
@@ -5094,9 +5095,9 @@ def block_expand_layer(input,
     :param padding_y: The padding size in vertical direction.
     :type padding_y: int
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring.
+    :type name: None | basestring.
     :param layer_attr: Extra Layer config.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -5155,15 +5156,15 @@ def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None):
                              num_channels=128,
                              groups=4)
 
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param num_channels: The channel number of input layer. If None will be set
                      automatically from previous output.
-    :type num_channels: int|None
+    :type num_channels: int | None
     :param groups: The group number of input layer.
     :type groups: int
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring.
+    :type name: None | basestring.
     :param layer_attr: Extra Layer attribute.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
@@ -5220,18 +5221,18 @@ def ctc_layer(input,
                       size=9055,
                       norm_by_times=True)
 
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param label: The data layer of label with variable length.
     :type label: LayerOutput
     :param size: category numbers + 1.
     :type size: int
     :param name: The name of this layer. It is optional.
-    :type name: basestring|None
+    :type name: basestring | None
     :param norm_by_times: Whether to normalization by times. False by default.
     :type norm_by_times: bool
     :param layer_attr: Extra Layer config.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -5297,20 +5298,20 @@ def warp_ctc_layer(input,
                            blank=1000,
                            norm_by_times=False)
 
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param label: The data layer of label with variable length.
     :type label: LayerOutput
     :param size: category numbers + 1.
     :type size: int
     :param name: The name of this layer. It is optional.
-    :type name: basestring|None
+    :type name: basestring | None
     :param blank: the 'blank' label used in ctc
     :type blank: int
     :param norm_by_times: Whether to normalization by times. False by default.
     :type norm_by_times: bool
     :param layer_attr: Extra Layer config.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -5368,11 +5369,11 @@ def crf_layer(input,
     :param param_attr: Parameter attribute. None means default attribute
     :type param_attr: ParameterAttribute
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring
+    :type name: None | basestring
     :param coeff: The coefficient affects the gradient in the backward.
     :type coeff: float
     :param layer_attr: Extra Layer config.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -5438,9 +5439,9 @@ def crf_decoding_layer(input,
     :param param_attr: Parameter attribute. None means default attribute
     :type param_attr: ParameterAttribute
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring
+    :type name: None | basestring
     :param layer_attr: Extra Layer config.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -5499,14 +5500,14 @@ def nce_layer(input,
     :param name: The name of this layer. It is optional.
     :type name: basestring
     :param input: The input layers. It could be a LayerOutput of list/tuple of LayerOutput.
-    :type input: LayerOutput|list|tuple|collections.Sequence
+    :type input: LayerOutput | list | tuple | collections.Sequence
     :param label: label layer
     :type label: LayerOutput
     :param weight: weight layer, can be None(default)
     :type weight: LayerOutput
     :param num_classes: number of classes.
     :type num_classes: int
-    :param act: Activation, default is Sigmoid.
+    :param act: Activation type. SigmoidActivation is the default.
     :type act: BaseActivation
     :param param_attr: The Parameter Attribute|list.
     :type param_attr: ParameterAttribute
@@ -5515,12 +5516,12 @@ def nce_layer(input,
     :param neg_distribution: The distribution for generating the random negative labels.
                              A uniform distribution will be used if not provided.
                              If not None, its length must be equal to num_classes.
-    :type neg_distribution: list|tuple|collections.Sequence|None
+    :type neg_distribution: list | tuple | collections.Sequence | None
     :param bias_attr: The Bias Attribute. If the parameter is set to
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param layer_attr: Extra Layer Attribute.
     :type layer_attr: ExtraLayerAttribute
     :return: layer name.
@@ -5636,7 +5637,7 @@ def rank_cost(left,
                    It is an optional argument.
     :type weight: LayerOutput
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring
+    :type name: None | basestring
     :param coeff: The coefficient affects the gradient in the backward.
     :type coeff: float
     :param layer_attr: Extra Layer Attribute.
@@ -5701,7 +5702,7 @@ def lambda_cost(input,
                           entire list of get gradient.
     :type max_sort_size: int
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring
+    :type name: None | basestring
     :param layer_attr: Extra Layer Attribute.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
@@ -5745,7 +5746,7 @@ def cross_entropy(input,
     :param label: The input label.
     :type input: LayerOutput.
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring.
+    :type name: None | basestring.
     :param coeff: The cost is multiplied with coeff.
                   The coefficient affects the gradient in the backward.
     :type coeff: float.
@@ -5793,7 +5794,7 @@ def cross_entropy_with_selfnorm(input,
     :param label: The input label.
     :type input: LayerOutput.
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring.
+    :type name: None | basestring.
     :param coeff: The coefficient affects the gradient in the backward.
     :type coeff: float.
     :param softmax_selfnorm_alpha: The scale factor affects the cost.
@@ -5830,10 +5831,10 @@ def sum_cost(input, name=None, layer_attr=None):
 
        cost = sum_cost(input=input_layer)
 
-    :param input: The first input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput.
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring.
+    :type name: None | basestring.
     :param layer_attr: Extra Layer Attribute.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
@@ -5878,7 +5879,7 @@ def huber_regression_cost(input,
     :param label: The input label.
     :type input: LayerOutput.
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring.
+    :type name: None | basestring.
     :param delta: The difference between the observed and predicted values.
     :type delta: float.
     :param coeff: The coefficient affects the gradient in the backward.
@@ -5928,7 +5929,7 @@ def huber_classification_cost(input,
     :param label: The input label.
     :type input: LayerOutput.
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring.
+    :type name: None | basestring.
     :param coeff: The coefficient affects the gradient in the backward.
     :type coeff: float.
     :param layer_attr: Extra Layer Attribute.
@@ -5971,7 +5972,7 @@ def multi_binary_label_cross_entropy(input,
     :param label: The input label.
     :type input: LayerOutput
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring
+    :type name: None | basestring
     :param coeff: The coefficient affects the gradient in the backward.
     :type coeff: float
     :param layer_attr: Extra Layer Attribute.
@@ -6139,7 +6140,7 @@ def smooth_l1_cost(input, label, name=None, coeff=1.0, layer_attr=None):
     :param label: The input label.
     :type input: LayerOutput
     :param name: The name of this layer. It is optional.
-    :type name: None|basestring
+    :type name: None | basestring
     :param coeff: The coefficient affects the gradient in the backward.
     :type coeff: float
     :param layer_attr: Extra Layer Attribute.
@@ -6226,7 +6227,7 @@ def dropout_layer(input, dropout_rate, name=None):
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param dropout_rate: The probability of dropout.
     :type dropout_rate: float
@@ -6285,18 +6286,18 @@ def row_conv_layer(input,
        row_conv = row_conv_layer(input=input_layer, context_len=3)
 
 
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param context_len: The context length equals the lookahead step number
                         plus one.
     :type context_len: int
-    :param act: Activation Type. Default is linear activation.
+    :param act: Activation Type. LinearActivation is the default.
     :type act: BaseActivation
     :param param_attr: The Parameter Attribute. If None, the parameter will be
                        initialized smartly. It's better to set it by yourself.
     :type param_attr: ParameterAttribute
     :param layer_attr: Extra Layer config.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
 
@@ -6342,7 +6343,7 @@ def prelu_layer(input,
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param partial_sum: this parameter makes a group of inputs share a same weight.
 
@@ -6352,9 +6353,9 @@ def prelu_layer(input,
 
     :type partial_sum: int
     :param param_attr: The parameter attribute. See ParameterAttribute for details.
-    :type param_attr: ParameterAttribute|None
+    :type param_attr: ParameterAttribute | None
     :param layer_attr: Extra layer configurations. Default is None.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -6407,37 +6408,37 @@ def gated_unit_layer(input,
     .. code-block:: python
         gated_unit = gated_unit_layer(size=128, input=input_layer))
 
-    :param input: input for this layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param size: output size of the gated unit.
     :type size: int
-    :param act: activation type of the projected input.
+    :param act: Activation type of the projected input. LinearActivation is the default.
     :type act: BaseActivation
     :param name: The name of this layer. It is optional.
     :type name: basestring
     :param gate_attr: Attributes to tune the gate output, for example, error
         clipping threshold, dropout and so on. See ExtraLayerAttribute for
         more details.
-    :type gate_attr: ExtraLayerAttribute|None
+    :type gate_attr: ExtraLayerAttribute | None
     :param gate_param_attr: Attributes to tune the learnable projected matrix
         parameter of the gate.
-    :type gate_param_attr: ParameterAttribute|None
+    :type gate_param_attr: ParameterAttribute | None
     :param gate_bias_attr: Attributes to tune the learnable bias of the gate.
-    :type gate_bias_attr: ParameterAttribute|None
+    :type gate_bias_attr: ParameterAttribute | None
     :param inproj_attr: Attributes to the tune the projected input, for
         example, error clipping threshold, dropout and so on. See
         ExtraLayerAttribute for more details.
-    :type inproj_attr: ExtraLayerAttribute|None
+    :type inproj_attr: ExtraLayerAttribute | None
     :param inproj_param_attr: Attributes to tune the learnable parameter of
         the projection of input.
-    :type inproj_param_attr: ParameterAttribute|None
+    :type inproj_param_attr: ParameterAttribute | None
     :param inproj_bias_attr: Attributes to tune the learnable bias of
         projection of the input.
-    :type inproj_bias_attr: ParameterAttribute|None
+    :type inproj_bias_attr: ParameterAttribute | None
     :param layer_attr: Attributes to tune the final output of the gated unit,
         for example, error clipping threshold, dropout and so on. See
         ExtraLayerAttribute for more details.
-    :type layer_attr: ExtraLayerAttribute|None
+    :type layer_attr: ExtraLayerAttribute | None
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -6487,7 +6488,7 @@ def switch_order_layer(input,
        switch = switch_order(input=layer, name='switch', reshape_axis=reshape_axis)
        reshape = {'height':[ 0, 1, 2], 'width':[3]}
 
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param name: The name of this layer. It is optional.
     :type name: basestring
@@ -6521,7 +6522,7 @@ def switch_order_layer(input,
 @layer_support()
 def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None):
     """
-    The crop layer crops images by offset and shape. User can set crop shape by
+    This layer crops images by offset and shape. User can set crop shape by
     args 'shape' explicitly or by reference input layer.
 
     The example usage is:
@@ -6529,10 +6530,10 @@ def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None):
     .. code-block:: python
     crop = crop_layer(input=[image_input, reference_input], axis=2, offset=[2, 3])
 
-    :param input: The input layer.If two inputs were setted,
-                    the second input will be regarded as reference input
-    :type input: LayerOutput or Sequence
-    :param offset: The crop offset
+    :param input: The input of this layer. If two inputs are given, the second input
+                  will be regarded as reference input.
+    :type input: LayerOutput | Sequence
+    :param offset: The crop offset.
     :type offset: Sequence
     :param axis: start axis to be cropped. To image input layer:
         - 0: batch size
@@ -6581,12 +6582,12 @@ def sub_nested_seq_layer(input, selected_indices, name=None):
 
     .. code-block:: python
 
-        sub_nest_seq = sub_nested_seq_layer(input=[data, selected_indices])
+        sub_nest_seq = sub_nested_seq_layer(input=data, selected_indices=selected_ids)
 
 
-    :param input: A nested sequence.
+    :param input: The input of this layer. It is a nested sequence.
     :type input: LayerOutput
-    :param selected_indices: a set of sequence indices in the nested sequence.
+    :param selected_indices: A set of sequence indices in the nested sequence.
     :type input: LayerOutput
     :param name: The name of this layer. It is optional.
     :type name: basestring
@@ -6628,7 +6629,7 @@ def clip_layer(input, min, max, name=None):
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input layer.
+    :param input: The input of this layer.
     :type input: LayerOutput.
     :param min: The lower threshold for clipping.
     :type min: double
@@ -6673,12 +6674,12 @@ def seq_slice_layer(input, starts, ends, name=None):
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: input for this layer, it should be a sequence.
+    :param input: The input of this layer, which should be a sequence.
     :type input: LayerOutput
     :param starts: start indices to slice the input sequence.
-    :type starts: LayerOutput|None
+    :type starts: LayerOutput | None
     :param ends: end indices to slice the input sequence.
-    :type ends: LayerOutput|None
+    :type ends: LayerOutput | None
     :return: LayerOutput object.
     :rtype: LayerOutput
 
@@ -6727,9 +6728,9 @@ def kmax_seq_score_layer(input, name=None, beam_size=1):
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input layer. It stores scores over a sequence or a nested
+    :param input: The input of this layer. It stores scores over a sequence or a nested
         sequence and its size must be 1.
-    :type input: LayerOutput.
+    :type input: LayerOutput
     :param beam_size: sequence indices with top beam_size scores are returned.
     :type beam_size: double
     :return: LayerOutput object.
@@ -6785,24 +6786,24 @@ def img_conv3d_layer(input,
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: Layer Input.
+    :param input: The input of this layer.
     :type input: LayerOutput
     :param filter_size: The x dimension of a filter kernel. Or input a list.
-    :type filter_size: int|tuple|list
+    :type filter_size: int | tuple | list
     :param num_filters: Each filter group's number of filter
-    :param act: Activation type. Default is tanh
+    :param act: Activation type. ReluActivation is the default.
     :type act: BaseActivation
     :param groups: Group size of filters.
     :type groups: int
     :param stride: The x dimension of the stride. Or input a tuple for two image
                    dimension.
-    :type stride: int|tuple|list
+    :type stride: int | tuple | list
     :param padding: The x dimension of the padding. Or input a tuple for two
                     image dimension
-    :type padding: int|tuple|list
+    :type padding: int | tuple | list
     :param bias_attr: Convolution bias attribute. None means default bias.
                       False means no bias.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :param num_channels: number of input channels. If None will be set
                         automatically from previous output.
     :type num_channels: int
@@ -6916,15 +6917,15 @@ def scale_shift_layer(input, name=None, param_attr=None, bias_attr=None):
 
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param input: The input layer.
-    :type input: LayerOutput.
+    :param input: The input of this layer.
+    :type input: LayerOutput
     :param param_attr: The parameter attribute of scaling.
     :type param_attr: ParameterAttribute
     :param bias_attr: The Bias Attribute. If the parameter is set to
                       False or something not type of ParameterAttribute,
                       no bias is defined. If the parameter is set to
                       True, the bias is initialized to zero.
-    :type bias_attr: ParameterAttribute|None|Bool|Any
+    :type bias_attr: ParameterAttribute | None | bool | Any
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -6944,11 +6945,11 @@ def resize_layer(input, size, name=None):
     into the output matrix with a shape of [Height x Width / size, size],
     where size is the parameter of this layer indicating the output dimension.
 
-    :param input: The input to this layer.
+    :param input: The input of this layer.
     :type input: LayerOutput.
     :param name: The name of this layer. It is optional.
     :type name: basestring
-    :param size: The resized output dimesion of this layer.
+    :param size: The resized output dimension of this layer.
     :type size: int
     :return: A LayerOutput object.
     :rtype: LayerOutput

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

@@ -363,5 +363,26 @@ class TestSoftsign(OpTest):
         self.check_grad(['X'], 'Y', max_relative_error=0.007)
 
 
+class TestThresholdedRelu(OpTest):
+    def setUp(self):
+        self.op_type = "thresholded_relu"
+        threshold = 0.25
+        self.relative_error = 0.005
+        X = np.random.uniform(-1, 1, [11, 17]).astype("float32")
+
+        # Same reason as TestAbs
+        X[np.abs(X - threshold) < self.relative_error] = threshold + 0.2
+
+        self.inputs = {'X': X}
+        self.attrs = {'threshold': threshold}
+        self.outputs = {'Y': (X > threshold) * X}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Y', max_relative_error=self.relative_error)
+
+
 if __name__ == "__main__":
     unittest.main()

python/paddle/v2/framework/tests/test_conv2d_op.py

@@ -6,7 +6,7 @@ from op_test import OpTest
 class TestConv2dOp(OpTest):
     def setUp(self):
         self.init_groups()
-        self.op_type = "conv2d"
+        self.init_optype()
         batch_size = 2
         input_channels = 3
         input_height = 5
@@ -32,6 +32,7 @@ class TestConv2dOp(OpTest):
         self.attrs = {
             'strides': [1, 1],
             'paddings': [0, 0],
+            'dilations': [1, 1],
             'groups': self.groups
         }
 
@@ -93,11 +94,27 @@ class TestConv2dOp(OpTest):
     def init_groups(self):
         self.groups = 1
 
+    def init_optype(self):
+        self.op_type = "conv2d"
+
 
 class TestWithGroup(TestConv2dOp):
     def init_groups(self):
         self.groups = 3
 
 
+class TestCudnn2d(TestConv2dOp):
+    def init_optype(self):
+        self.op_type = "conv_cudnn"
+
+
+class TestCudnn2dWithGroup(TestConv2dOp):
+    def init_optype(self):
+        self.op_type = "conv_cudnn"
+
+    def init_groups(self):
+        self.groups = 3
+
+
 if __name__ == '__main__':
     unittest.main()

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestMarginRankLossOp(OpTest):
+    def setUp(self):
+        self.op_type = "margin_rank_loss"
+        batch_size = 5
+        margin = 0.5
+        # labels_{i} = {-1, 1}
+        label = 2 * np.random.randint(
+            0, 2, size=(batch_size, 1)).astype("float32") - 1
+        x1 = np.random.random((batch_size, 1)).astype("float32")
+        x2 = np.random.random((batch_size, 1)).astype("float32")
+        # loss = max(0, -label * (x1 - x2) + margin)
+        loss = -label * (x1 - x2) + margin
+        loss = np.where(loss > 0, loss, 0)
+        act = np.where(loss > 0, 1., 0.)
+
+        self.attrs = {'margin': margin}
+        self.inputs = {'Label': label, 'X1': x1, 'X2': x2}
+        self.outputs = {'Activated': act, 'Out': loss}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X1", "X2"], "Out")
+
+    def test_check_grad_ignore_x1(self):
+        self.check_grad(["X2"], "Out", no_grad_set=set('X1'))
+
+    def test_check_grad_ignore_x2(self):
+        self.check_grad(["X1"], "Out", no_grad_set=set('X2'))
+
+
+if __name__ == '__main__':
+    unittest.main()

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

 import unittest
+
+import paddle.v2.framework.core as core
 from paddle.v2.framework.graph import g_program
 
 
@@ -31,6 +33,34 @@ class TestProgram(unittest.TestCase):
         self.assertEqual(1, b.idx)
         self.assertEqual(0, b.parent_idx)
 
+    def test_append_backward(self):
+        prog = core.ProgramDesc.__create_program_desc__()
+        self.assertIsNotNone(prog)
+        block = prog.block(0)
+        self.assertIsNotNone(block)
+
+        mul_op_desc = block.append_op()
+        mul_op_desc.set_type("mul")
+        mul_op_desc.set_input("X", ["x1"])
+        mul_op_desc.set_input("Y", ["y1"])
+        mul_op_desc.set_output("Out", ["out1"])
+
+        sum_op_desc = block.append_op()
+        sum_op_desc.set_type("elementwise_add")
+        sum_op_desc.set_input("X", ["out1"])
+        sum_op_desc.set_input("Y", ["b1"])
+        sum_op_desc.set_output("Out", ["out2"])
+
+        expect_ops = [
+            "mul", "elementwise_add", "elementwise_add_grad", "mul_grad"
+        ]
+        actual_ops = []
+        prog.append_backward(set())
+        for op in block.all_ops():
+            actual_ops.append(op.type())
+        print(actual_ops)
+        self.assertEqual(actual_ops, expect_ops)
+
 
 if __name__ == '__main__':
     unittest.main()

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

@@ -55,6 +55,12 @@ class TestOpDesc(unittest.TestCase):
         op.set_block_attr("block_attr", prog.block(0))
         self.assertEqual(0, op.get_block_attr("block_attr"))
 
+        mul_op = block.append_op()
+        mul_op.set_type("mul")
+        mul_op.check_attrs()
+        self.assertEqual(mul_op.attr("x_num_col_dims"), 1)
+        self.assertEqual(mul_op.attr("y_num_col_dims"), 1)
+
 
 class TestProgramDesc(unittest.TestCase):
     def test_instance(self):

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

+import unittest
+import numpy as np
+import sys
+from op_test import OpTest
+
+
+class TestConcatOp(OpTest):
+    def set_data(self):
+        # two level, batch size is 3
+        x0 = np.random.random((4, 6, 3)).astype('float32')
+        lod0 = [[0, 2, 4], [0, 1, 2, 3, 4]]
+        x1 = np.random.random((4, 8, 3)).astype('float32')
+        lod1 = [[0, 2, 4], [0, 1, 2, 3, 4]]
+        axis = 1
+        level = 1
+        self.inputs = {'X': [('x0', (x0, lod0)), ('x1', (x1, lod1))]}
+        self.attrs = {'axis': axis, 'level': level}
+        outs = []
+        for i in range(4):
+            sub_x0 = x0[lod0[level][i]:lod0[level][i + 1], :]
+            sub_x1 = x1[lod1[level][i]:lod1[level][i + 1], :]
+            outs.append(np.concatenate((sub_x0, sub_x1), axis=axis))
+
+        self.outputs = {'Out': np.concatenate(outs, axis=0)}
+
+    def setUp(self):
+        self.op_type = "sequence_concat"
+        self.set_data()
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['x0'], 'Out')
+
+
+class TestConcatOpDiffLod(TestConcatOp):
+    def set_data(self):
+        # two level, batch size is 3
+        x0 = np.random.random((4, 6, 3)).astype('float32')
+        lod0 = [[0, 2, 4], [0, 1, 2, 3, 4]]
+        x1 = np.random.random((5, 6, 3)).astype('float32')
+        lod1 = [[0, 3, 5], [0, 1, 2, 3, 5]]
+        axis = 0
+        level = 1
+        self.inputs = {'X': [('x0', (x0, lod0)), ('x1', (x1, lod1))]}
+        self.attrs = {'axis': axis, 'level': level}
+        outs = []
+        for i in range(4):
+            sub_x0 = x0[lod0[level][i]:lod0[level][i + 1], :]
+            sub_x1 = x1[lod1[level][i]:lod1[level][i + 1], :]
+            outs.append(np.concatenate((sub_x0, sub_x1), axis=axis))
+
+        self.outputs = {'Out': np.concatenate(outs, axis=0)}
+
+
+class TestConcatOpLevelZero(TestConcatOp):
+    def set_data(self):
+        # two level, batch size is 3
+        x0 = np.random.random((4, 3, 4)).astype('float32')
+        lod0 = [[0, 2, 4], [0, 1, 2, 3, 4]]
+        x1 = np.random.random((5, 3, 4)).astype('float32')
+        lod1 = [[0, 3, 5], [0, 1, 3, 4, 5]]
+        axis = 0
+        level = 0
+        self.inputs = {'X': [('x0', (x0, lod0)), ('x1', (x1, lod1))]}
+        self.attrs = {'axis': axis, 'level': level}
+        outs = []
+        for i in range(2):
+            sub_x0 = x0[lod0[level][i]:lod0[level][i + 1], :]
+            sub_x1 = x1[lod1[level][i]:lod1[level][i + 1], :]
+            outs.append(np.concatenate((sub_x0, sub_x1), axis=axis))
+
+        self.outputs = {'Out': np.concatenate(outs, axis=0)}
+
+
+if __name__ == '__main__':
+    sys.exit(0)
+    unittest.main()