Merge branch 'develop' of https://github.com/paddlepaddle/paddle into add_trt_plugin

test=develop

CMakeLists.txt

@@ -41,6 +41,7 @@ option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_F
 option(WITH_AMD_GPU     "Compile PaddlePaddle with AMD GPU"             OFF)
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
 option(WITH_MKL         "Compile PaddlePaddle with MKL support."        ${AVX_FOUND})
+option(WITH_NGRAPH      "Compile PaddlePaddle with nGraph support."     OFF)
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        OFF)
 option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)
@@ -103,6 +104,8 @@ if(ANDROID OR IOS)
         "Disable RDMA when cross-compiling for Android and iOS" FORCE)
     set(WITH_MKL OFF CACHE STRING
         "Disable MKL when cross-compiling for Android and iOS" FORCE)
+    set(WITH_NGRAPH OFF CACHE STRING
+        "Disable nGraph when cross-compiling for Android and iOS" FORCE)
     set(WITH_GOLANG OFF CACHE STRING
         "Disable golang when cross-compiling for Android and iOS" FORCE)
 
@@ -171,6 +174,7 @@ include(external/protobuf)  # download, build, install protobuf
 include(external/python)    # download, build, install python
 include(external/openblas)  # download, build, install openblas
 include(external/mkldnn)    # download, build, install mkldnn
+include(external/ngraph)    # download, build, install nGraph
 include(external/swig)      # download, build, install swig
 include(external/boost)     # download boost
 include(external/any)       # download libn::any

cmake/external/mkldnn.cmake

@@ -37,7 +37,6 @@ SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
 SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
 
 INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR}) # For MKLDNN code to include internal headers.
-INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include mkldnn.h
 
 IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
     SET(MKLDNN_DEPENDS   ${MKLML_PROJECT})

cmake/external/ngraph.cmake

+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+add_library(ngraph INTERFACE)
+
+IF(WIN32 OR APPLE)
+    MESSAGE(WARNING
+        "Windows or Mac is not supported with nGraph in Paddle yet."
+        "Force WITH_NGRAPH=OFF")
+    SET(WITH_NGRAPH OFF CACHE STRING "Disable nGraph in Windows and MacOS" FORCE)
+ENDIF()
+
+IF(${WITH_NGRAPH} AND NOT ${WITH_MKLDNN})
+    MESSAGE(WARNING
+        "nGraph needs mkl-dnn to be enabled."
+        "Force WITH_NGRAPH=OFF")
+    SET(WITH_NGRAPH OFF CACHE STRING "Disable nGraph if mkl-dnn is disabled" FORCE)
+ENDIF()
+
+IF(NOT ${WITH_NGRAPH})
+    return()
+ENDIF()
+
+INCLUDE(ExternalProject)
+
+SET(NGRAPH_PROJECT         "extern_ngraph")
+SET(NGRAPH_VERSION         "0.9")
+SET(NGRAPH_GIT_TAG         "f9fd9d4cc318dc59dd4b68448e7fbb5f67a28bd0")
+SET(NGRAPH_SOURCES_DIR     ${THIRD_PARTY_PATH}/ngraph)
+SET(NGRAPH_INSTALL_DIR     ${THIRD_PARTY_PATH}/install/ngraph)
+SET(NGRAPH_INC_DIR         ${NGRAPH_INSTALL_DIR}/include)
+SET(NGRAPH_SHARED_LIB_NAME libngraph.so.${NGRAPH_VERSION})
+SET(NGRAPH_CPU_LIB_NAME    libcpu_backend.so)
+SET(NGRAPH_TBB_LIB_NAME    libtbb.so.2)
+SET(NGRAPH_GIT_REPO        "https://github.com/NervanaSystems/ngraph.git")
+
+ExternalProject_Add(
+    ${NGRAPH_PROJECT}
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    DEPENDS             ${MKLDNN_PROJECT} ${MKLML_PROJECT}
+    GIT_REPOSITORY      ${NGRAPH_GIT_REPO}
+    GIT_TAG             ${NGRAPH_GIT_TAG}
+    PREFIX              ${NGRAPH_SOURCES_DIR}
+    UPDATE_COMMAND      ""
+    CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${NGRAPH_INSTALL_DIR}
+    CMAKE_ARGS          -DNGRAPH_UNIT_TEST_ENABLE=FALSE
+    CMAKE_ARGS          -DNGRAPH_TOOLS_ENABLE=FALSE
+    CMAKE_ARGS          -DNGRAPH_INTERPRETER_ENABLE=FALSE
+    CMAKE_ARGS          -DNGRAPH_DEX_ONLY=TRUE
+    CMAKE_ARGS          -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
+    CMAKE_ARGS          -DMKLDNN_INCLUDE_DIR=${MKLDNN_INC_DIR}
+    CMAKE_ARGS          -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/lib
+)
+
+if(UNIX AND NOT APPLE)
+    include(GNUInstallDirs)
+    SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR})
+else()
+    SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/lib)
+endif()
+MESSAGE(STATUS "nGraph lib will be installed at: ${NGRAPH_LIB_DIR}")
+
+SET(NGRAPH_SHARED_LIB      ${NGRAPH_LIB_DIR}/${NGRAPH_SHARED_LIB_NAME})
+SET(NGRAPH_CPU_LIB         ${NGRAPH_LIB_DIR}/${NGRAPH_CPU_LIB_NAME})
+SET(NGRAPH_TBB_LIB         ${NGRAPH_LIB_DIR}/${NGRAPH_TBB_LIB_NAME})
+
+# Workaround for nGraph expecting mklml to be in mkldnn install directory.
+ExternalProject_Add_Step(
+    ${NGRAPH_PROJECT}
+    PrepareMKL
+    COMMAND ${CMAKE_COMMAND} -E create_symlink ${MKLML_LIB} ${MKLDNN_INSTALL_DIR}/lib/libmklml_intel.so
+    COMMAND ${CMAKE_COMMAND} -E create_symlink ${MKLML_IOMP_LIB} ${MKLDNN_INSTALL_DIR}/lib/libiomp5.so
+    DEPENDEES download
+    DEPENDERS configure
+)
+
+add_dependencies(ngraph ${NGRAPH_PROJECT})
+target_compile_definitions(ngraph INTERFACE -DPADDLE_WITH_NGRAPH)
+target_include_directories(ngraph INTERFACE ${NGRAPH_INC_DIR})
+target_link_libraries(ngraph INTERFACE ${NGRAPH_SHARED_LIB})
+LIST(APPEND external_project_dependencies ngraph)

cmake/external/protobuf.cmake

@@ -30,66 +30,61 @@ UNSET_VAR(PROTOBUF_LITE_LIBRARY)
 UNSET_VAR(PROTOBUF_LIBRARY)
 UNSET_VAR(PROTOBUF_INCLUDE_DIR)
 UNSET_VAR(Protobuf_PROTOC_EXECUTABLE)
+function(protobuf_generate_python SRCS)
+    # shameless copy from https://github.com/Kitware/CMake/blob/master/Modules/FindProtobuf.cmake
+    if(NOT ARGN)
+        message(SEND_ERROR "Error: PROTOBUF_GENERATE_PYTHON() called without any proto files")
+        return()
+    endif()
 
-if(NOT COMMAND protobuf_generate_python)  # before cmake 3.4, protobuf_genrerate_python is not defined.
-    function(protobuf_generate_python SRCS)
-        # shameless copy from https://github.com/Kitware/CMake/blob/master/Modules/FindProtobuf.cmake
-        if(NOT ARGN)
-            message(SEND_ERROR "Error: PROTOBUF_GENERATE_PYTHON() called without any proto files")
-            return()
-        endif()
-
-        if(PROTOBUF_GENERATE_CPP_APPEND_PATH)
-            # Create an include path for each file specified
-            foreach(FIL ${ARGN})
-                get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
-                get_filename_component(ABS_PATH ${ABS_FIL} PATH)
-                list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
-                if(${_contains_already} EQUAL -1)
-                    list(APPEND _protobuf_include_path -I ${ABS_PATH})
-                endif()
-            endforeach()
-        else()
-            set(_protobuf_include_path -I ${CMAKE_CURRENT_SOURCE_DIR})
-        endif()
-
-        if(DEFINED PROTOBUF_IMPORT_DIRS AND NOT DEFINED Protobuf_IMPORT_DIRS)
-            set(Protobuf_IMPORT_DIRS "${PROTOBUF_IMPORT_DIRS}")
-        endif()
-
-        if(DEFINED Protobuf_IMPORT_DIRS)
-            foreach(DIR ${Protobuf_IMPORT_DIRS})
-                get_filename_component(ABS_PATH ${DIR} ABSOLUTE)
-                list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
-                if(${_contains_already} EQUAL -1)
-                    list(APPEND _protobuf_include_path -I ${ABS_PATH})
-                endif()
-            endforeach()
-        endif()
-
-        set(${SRCS})
+    if(PROTOBUF_GENERATE_CPP_APPEND_PATH)
+        # Create an include path for each file specified
         foreach(FIL ${ARGN})
             get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
-            get_filename_component(FIL_WE ${FIL} NAME_WE)
-            if(NOT PROTOBUF_GENERATE_CPP_APPEND_PATH)
-                get_filename_component(FIL_DIR ${FIL} DIRECTORY)
-                if(FIL_DIR)
-                    set(FIL_WE "${FIL_DIR}/${FIL_WE}")
-                endif()
+            get_filename_component(ABS_PATH ${ABS_FIL} PATH)
+            list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
+            if(${_contains_already} EQUAL -1)
+                list(APPEND _protobuf_include_path -I ${ABS_PATH})
             endif()
+        endforeach()
+    else()
+        set(_protobuf_include_path -I ${CMAKE_CURRENT_SOURCE_DIR})
+    endif()
+    if(DEFINED PROTOBUF_IMPORT_DIRS AND NOT DEFINED Protobuf_IMPORT_DIRS)
+        set(Protobuf_IMPORT_DIRS "${PROTOBUF_IMPORT_DIRS}")
+    endif()
 
-            list(APPEND ${SRCS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py")
-            add_custom_command(
-                    OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py"
-                    COMMAND  ${Protobuf_PROTOC_EXECUTABLE} --python_out ${CMAKE_CURRENT_BINARY_DIR} ${_protobuf_include_path} ${ABS_FIL}
-                    DEPENDS ${ABS_FIL} ${Protobuf_PROTOC_EXECUTABLE}
-                    COMMENT "Running Python protocol buffer compiler on ${FIL}"
-                    VERBATIM )
+    if(DEFINED Protobuf_IMPORT_DIRS)
+        foreach(DIR ${Protobuf_IMPORT_DIRS})
+            get_filename_component(ABS_PATH ${DIR} ABSOLUTE)
+            list(FIND _protobuf_include_path ${ABS_PATH} _contains_already)
+            if(${_contains_already} EQUAL -1)
+                list(APPEND _protobuf_include_path -I ${ABS_PATH})
+            endif()
         endforeach()
+    endif()
 
-        set(${SRCS} ${${SRCS}} PARENT_SCOPE)
-    endfunction()
-endif()
+    set(${SRCS})
+    foreach(FIL ${ARGN})
+        get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
+        get_filename_component(FIL_WE ${FIL} NAME_WE)
+        if(NOT PROTOBUF_GENERATE_CPP_APPEND_PATH)
+            get_filename_component(FIL_DIR ${FIL} DIRECTORY)
+            if(FIL_DIR)
+                set(FIL_WE "${FIL_DIR}/${FIL_WE}")
+            endif()
+        endif()
+        list(APPEND ${SRCS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py")
+        add_custom_command(
+                OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py"
+                COMMAND  ${PROTOBUF_PROTOC_EXECUTABLE} --python_out ${CMAKE_CURRENT_BINARY_DIR} ${_protobuf_include_path} ${ABS_FIL}
+                DEPENDS ${ABS_FIL} ${PROTOBUF_PROTOC_EXECUTABLE}
+                COMMENT "Running Python protocol buffer compiler on ${FIL}"
+                VERBATIM )
+    endforeach()
+
+    set(${SRCS} ${${SRCS}} PARENT_SCOPE)
+endfunction()
 
 # Print and set the protobuf library information,
 # finish this cmake process and exit from this file.
@@ -126,6 +121,7 @@ macro(PROMPT_PROTOBUF_LIB)
     # FIND_Protobuf.cmake uses `Protobuf_PROTOC_EXECUTABLE`.
     # make `protobuf_generate_cpp` happy.
     SET(Protobuf_PROTOC_EXECUTABLE ${PROTOBUF_PROTOC_EXECUTABLE})
+
     FOREACH(dep ${protobuf_DEPS})
         ADD_DEPENDENCIES(protobuf ${dep})
         ADD_DEPENDENCIES(protobuf_lite ${dep})

paddle/fluid/API.spec

@@ -103,7 +103,7 @@ paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 's
 paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None))
-paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode'], varargs=None, keywords=None, defaults=(False, -100, False))
+paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode', 'return_softmax'], varargs=None, keywords=None, defaults=(False, -100, False, False))
 paddle.fluid.layers.smooth_l1 ArgSpec(args=['x', 'y', 'inside_weight', 'outside_weight', 'sigma'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.one_hot ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.autoincreased_step_counter ArgSpec(args=['counter_name', 'begin', 'step'], varargs=None, keywords=None, defaults=(None, 1, 1))
@@ -179,10 +179,12 @@ paddle.fluid.layers.space_to_depth ArgSpec(args=['x', 'blocksize', 'name'], vara
 paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None))
+paddle.fluid.layers.similarity_focus ArgSpec(args=['input', 'axis', 'indexes', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.hash ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None))
 paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None))
 paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act', 'name', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
 paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)
@@ -201,6 +203,7 @@ paddle.fluid.layers.create_tensor ArgSpec(args=['dtype', 'name', 'persistable'],
 paddle.fluid.layers.create_parameter ArgSpec(args=['shape', 'dtype', 'name', 'attr', 'is_bias', 'default_initializer'], varargs=None, keywords=None, defaults=(None, None, False, None))
 paddle.fluid.layers.create_global_var ArgSpec(args=['shape', 'value', 'dtype', 'persistable', 'force_cpu', 'name'], varargs=None, keywords=None, defaults=(False, False, None))
 paddle.fluid.layers.cast ArgSpec(args=['x', 'dtype'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.layers.tensor_array_to_tensor ArgSpec(args=['input', 'axis', 'name'], varargs=None, keywords=None, defaults=(1, None))
 paddle.fluid.layers.concat ArgSpec(args=['input', 'axis', 'name'], varargs=None, keywords=None, defaults=(0, None))
 paddle.fluid.layers.sums ArgSpec(args=['input', 'out'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.assign ArgSpec(args=['input', 'output'], varargs=None, keywords=None, defaults=(None,))
@@ -271,6 +274,7 @@ paddle.fluid.layers.hard_shrink ArgSpec(args=['x', 'threshold'], varargs=None, k
 paddle.fluid.layers.cumsum ArgSpec(args=['x', 'axis', 'exclusive', 'reverse'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.thresholded_relu ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.prior_box ArgSpec(args=['input', 'image', 'min_sizes', 'max_sizes', 'aspect_ratios', 'variance', 'flip', 'clip', 'steps', 'offset', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, [1.0], [0.1, 0.1, 0.2, 0.2], False, False, [0.0, 0.0], 0.5, None, False))
+paddle.fluid.layers.density_prior_box ArgSpec(args=['input', 'image', 'densities', 'fixed_sizes', 'fixed_ratios', 'variance', 'clip', 'steps', 'offset', 'name'], varargs=None, keywords=None, defaults=(None, None, None, [0.1, 0.1, 0.2, 0.2], False, [0.0, 0.0], 0.5, None))
 paddle.fluid.layers.multi_box_head ArgSpec(args=['inputs', 'image', 'base_size', 'num_classes', 'aspect_ratios', 'min_ratio', 'max_ratio', 'min_sizes', 'max_sizes', 'steps', 'step_w', 'step_h', 'offset', 'variance', 'flip', 'clip', 'kernel_size', 'pad', 'stride', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, 0.5, [0.1, 0.1, 0.2, 0.2], True, False, 1, 0, 1, None, False))
 paddle.fluid.layers.bipartite_match ArgSpec(args=['dist_matrix', 'match_type', 'dist_threshold', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.target_assign ArgSpec(args=['input', 'matched_indices', 'negative_indices', 'mismatch_value', 'name'], varargs=None, keywords=None, defaults=(None, None, None))

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -167,10 +167,12 @@ struct HitGroup {
 
   bool Match(Node *node, PDNode *pat) {
     if (nodes_.count(node)) {
-      if (!roles.count(pat)) return false;
-      return roles[pat] == node;
+      if (roles.count(pat) && roles[pat] == node) return true;
+      return false;
+    } else {
+      if (roles.count(pat) && roles[pat] != node) return false;
+      return true;
     }
-    return !roles.count(pat) || roles.at(pat) == node;
   }
 
   void Register(Node *node, PDNode *pat) {
@@ -198,7 +200,6 @@ GraphPatternDetector::DetectPatterns() {
   std::vector<GraphPatternDetector::subgraph_t> result;
   std::vector<HitGroup> init_groups;
   std::array<std::vector<HitGroup>, 2> bi_records;
-  // PADDLE_ENFORCE(!pattern_.edges().empty(), "At least one edge is needed");
   auto *first_pnode = pattern_.edges().empty() ? pattern().nodes().front().get()
                                                : pattern_.edges().front().first;
   if (!pdnodes2nodes_.count(first_pnode)) return result;
@@ -228,11 +229,12 @@ GraphPatternDetector::DetectPatterns() {
         VLOG(80) << "check " << source->id() << " -- " << target->id();
         // TODO(Superjomn) add some prune strategies.
         for (const auto &group : pre_groups) {
-          HitGroup new_group = group;
-          if (IsNodesLink(source, target) &&
-              new_group.Match(source, edge.first)) {
-            new_group.Register(source, edge.first);
-            if (new_group.Match(target, edge.second)) {
+          if (IsNodesLink(source, target)) {
+            HitGroup new_group = group;
+            bool flag = new_group.Match(source, edge.first) &&
+                        new_group.Match(target, edge.second);
+            if (flag) {
+              new_group.Register(source, edge.first);
               new_group.Register(target, edge.second);
               cur_groups.push_back(new_group);
               // TODO(Superjomn) need to unique
@@ -261,14 +263,16 @@ GraphPatternDetector::DetectPatterns() {
   return result;
 }
 
-bool GraphItemCMP(const std::pair<PDNode *, Node *> &a,
+struct GraphItemLessThan {
+  bool operator()(const std::pair<PDNode *, Node *> &a,
                   const std::pair<PDNode *, Node *> &b) {
-  if (a.first != b.first) {
-    return a.first < b.first;
-  } else {
-    return a.second < b.second;
+    if (a.first != b.first) {
+      return a.first < b.first;
+    } else {
+      return a.second < b.second;
+    }
   }
-}
+};
 
 // TODO(Superjomn) enhance the function as it marks unique unique as duplicates
 // see https://github.com/PaddlePaddle/Paddle/issues/13550
@@ -282,7 +286,7 @@ void GraphPatternDetector::UniquePatterns(
   for (auto &g : *subgraphs) {
     // Sort the items in the sub-graph, and transform to a string key.
     std::vector<std::pair<PDNode *, Node *>> sorted_keys(g.begin(), g.end());
-    std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemCMP);
+    std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemLessThan());
     std::stringstream ss;
     for (auto &item : sorted_keys) {
       ss << item.first << ":" << item.second;

paddle/fluid/framework/operator.cc

@@ -259,6 +259,8 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
           if (row_size >= 0) {
             ss << "[row_size=" << row_size << "]";
           }
+          std::string dtype = GetDtype(*scope, output.second[i]);
+          ss << ":" << dtype;
           ss << "[" << GetDims(*scope, var_name, true) << "]";
           ss << "(" << GetLoD(*scope, var_name) << ")";
         }

paddle/fluid/framework/var_type_inference.h

@@ -13,6 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <string>
+#include "paddle/fluid/framework/block_desc.h"
+#include "paddle/fluid/framework/op_desc.h"
 #include "paddle/fluid/framework/type_defs.h"
 
 namespace paddle {
@@ -24,5 +27,27 @@ class VarTypeInference {
   virtual void operator()(const OpDesc& op_desc, BlockDesc* block) const = 0;
 };
 
+class PassInDtypeAndVarTypeToOutput : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc& op_desc,
+                  framework::BlockDesc* block) const final {
+    auto in_out_var_names = this->GetInputOutputWithSameType();
+
+    for (auto& i_o_n : in_out_var_names) {
+      auto& x_name = op_desc.Input(i_o_n.first).at(0);
+      auto& out_name = op_desc.Output(i_o_n.second).at(0);
+
+      auto& x = block->FindRecursiveOrCreateVar(x_name);
+      auto& out = block->FindRecursiveOrCreateVar(out_name);
+      out.SetType(x.GetType());
+      out.SetDataType(x.GetDataType());
+    }
+  }
+
+ protected:
+  virtual std::unordered_map<std::string, std::string>
+  GetInputOutputWithSameType() const = 0;
+};
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/inference/analysis/analyzer.cc

@@ -113,7 +113,9 @@ void Analyzer::Run(Argument* argument) {
   passes.push_back("infer_clean_graph_pass");
   passes.push_back("graph_viz_pass");  // add graphviz for debug.
   for (auto& pass : ir_passes_) {
-    if (!disabled_ir_passes_.count(pass)) {
+    // skip mkldnn pass when use_mkldnn_ = false;
+    bool skip_pass = (!use_mkldnn_) && pass.find("mkldnn") != std::string::npos;
+    if (!disabled_ir_passes_.count(pass) && !skip_pass) {
       passes.push_back(pass);
       passes.push_back("graph_viz_pass");  // add graphviz for debug.
     }

paddle/fluid/inference/analysis/data_flow_graph.cc

@@ -112,8 +112,8 @@ void DataFlowGraph::Build(const framework::proto::ProgramDesc &prog) {
           out_alias->SetPbMsg(out->pb_msg());
           var2id[out_alias->name()] =
               out_alias->id();  // update variable's alias Node
-          LOG(INFO) << "loop found in graph, create SSA alias node ["
-                    << out_alias->repr() << "] for [" << out->repr() << "]";
+          VLOG(40) << "loop found in graph, create SSA alias node ["
+                   << out_alias->repr() << "] for [" << out->repr() << "]";
           out = out_alias;
         }
         out->inlinks.push_back(o);

paddle/fluid/inference/tensorrt/convert/activation_op.cc

@@ -27,7 +27,7 @@ class ActivationOpConverter : public OpConverter {
     // Here the two nullptr looks strange, that's because the
     // framework::OpDesc's constructor is strange.
     framework::OpDesc op_desc(op, nullptr);
-    LOG(INFO)
+    VLOG(3)
         << "convert a fluid Activation op to tensorrt activation layer whose "
            "type is "
         << op_type_;

paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc

@@ -23,7 +23,7 @@ class BatchNormOpConverter : public OpConverter {
  public:
   void operator()(const framework::proto::OpDesc& op,
                   const framework::Scope& scope, bool test_mode) override {
-    LOG(INFO) << "convert a fluid batch norm op to tensorrt batch_norm";
+    VLOG(3) << "convert a fluid batch norm op to tensorrt batch_norm";
 
     framework::OpDesc op_desc(op, nullptr);
     PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);

paddle/fluid/inference/tensorrt/convert/concat_op.cc

@@ -25,7 +25,7 @@ class ConcatOpConverter : public OpConverter {
  public:
   void operator()(const framework::proto::OpDesc& op,
                   const framework::Scope& scope, bool test_mode) override {
-    VLOG(40) << "convert a fluid mul op to tensorrt mul layer without bias";
+    VLOG(3) << "convert a fluid mul op to tensorrt mul layer without bias";
 
     framework::OpDesc op_desc(op, nullptr);
     // Declare inputs

paddle/fluid/inference/tensorrt/convert/conv2d_op.cc

@@ -37,8 +37,7 @@ class Conv2dOpConverter : public OpConverter {
  public:
   void operator()(const framework::proto::OpDesc& op,
                   const framework::Scope& scope, bool test_mode) override {
-    LOG(INFO)
-        << "convert a fluid conv2d op to tensorrt conv layer without bias";
+    VLOG(3) << "convert a fluid conv2d op to tensorrt conv layer without bias";
 
     framework::OpDesc op_desc(op, nullptr);
     PADDLE_ENFORCE_EQ(op_desc.Input("Input").size(), 1);

paddle/fluid/inference/tensorrt/convert/dropout_op.cc

@@ -25,7 +25,7 @@ class DropoutOpConverter : public OpConverter {
  public:
   void operator()(const framework::proto::OpDesc& op,
                   const framework::Scope& scope, bool test_mode) override {
-    VLOG(40) << "convert a fluid dropout op to tensorrt dropout layer";
+    VLOG(3) << "convert a fluid dropout op to tensorrt dropout layer";
     framework::OpDesc op_desc(op, nullptr);
     // Declare inputs
     auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]);

paddle/fluid/inference/tensorrt/convert/elementwise_op.cc

@@ -26,7 +26,7 @@ class ElementwiseWeightOpConverter : public OpConverter {
     // Here the two nullptr looks strange, that's because the
     // framework::OpDesc's constructor is strange.
     framework::OpDesc op_desc(op, nullptr);
-    LOG(INFO) << "convert a fluid elementwise op to tensorrt IScaleLayer";
+    VLOG(3) << "convert a fluid elementwise op to tensorrt IScaleLayer";
 
     PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
     PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1);  // Y is a weight
@@ -108,7 +108,7 @@ class ElementwiseTensorOpConverter : public OpConverter {
     // Here the two nullptr looks strange, that's because the
     // framework::OpDesc's constructor is strange.
     framework::OpDesc op_desc(op, nullptr);
-    LOG(INFO) << "convert a fluid elementwise op to tensorrt IScaleLayer";
+    VLOG(3) << "convert a fluid elementwise op to tensorrt IScaleLayer";
 
     PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
     PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1);  // Y is a weight

paddle/fluid/inference/tensorrt/convert/fc_op.cc

@@ -52,7 +52,7 @@ class FcOpConverter : public OpConverter {
  public:
   void operator()(const framework::proto::OpDesc& op,
                   const framework::Scope& scope, bool test_mode) override {
-    VLOG(40) << "convert a fluid fc op to tensorrt fc layer without bias";
+    VLOG(3) << "convert a fluid fc op to tensorrt fc layer without bias";
 
     framework::OpDesc op_desc(op, nullptr);
     PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);

paddle/fluid/inference/tensorrt/convert/mul_op.cc

@@ -25,7 +25,7 @@ class MulOpConverter : public OpConverter {
  public:
   void operator()(const framework::proto::OpDesc& op,
                   const framework::Scope& scope, bool test_mode) override {
-    VLOG(40) << "convert a fluid mul op to tensorrt mul layer without bias";
+    VLOG(3) << "convert a fluid mul op to tensorrt mul layer without bias";
 
     framework::OpDesc op_desc(op, nullptr);
     // Declare inputs

paddle/fluid/inference/tensorrt/convert/pad_op.cc

@@ -25,7 +25,7 @@ class PadOpConverter : public OpConverter {
  public:
   void operator()(const framework::proto::OpDesc& op,
                   const framework::Scope& scope, bool test_mode) override {
-    VLOG(40) << "convert a fluid transpose op to tensorrt tranpose layer";
+    VLOG(3) << "convert a fluid transpose op to tensorrt tranpose layer";
 
     framework::OpDesc op_desc(op, nullptr);
     // Declare inputs

paddle/fluid/inference/tensorrt/convert/pool2d_op.cc

@@ -25,7 +25,7 @@ class Pool2dOpConverter : public OpConverter {
  public:
   void operator()(const framework::proto::OpDesc& op,
                   const framework::Scope& scope, bool test_mode) override {
-    VLOG(40)
+    VLOG(3)
         << "convert a fluid pool2d op to tensorrt pool2d layer without bias";
     framework::OpDesc op_desc(op, nullptr);
     // Declare inputs

paddle/fluid/inference/tensorrt/convert/softmax_op.cc

@@ -25,7 +25,7 @@ class SoftMaxOpConverter : public OpConverter {
  public:
   void operator()(const framework::proto::OpDesc& op,
                   const framework::Scope& scope, bool test_mode) override {
-    VLOG(40)
+    VLOG(3)
         << "convert a fluid softmax op to tensorrt softmax layer without bias";
     framework::OpDesc op_desc(op, nullptr);
     // Declare inputs

paddle/fluid/inference/tensorrt/helper.h

@@ -52,7 +52,7 @@ class NaiveLogger : public nvinfer1::ILogger {
   void log(nvinfer1::ILogger::Severity severity, const char* msg) override {
     switch (severity) {
       case Severity::kINFO:
-        LOG(INFO) << msg;
+        VLOG(3) << msg;
         break;
       case Severity::kWARNING:
         LOG(WARNING) << msg;

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -110,5 +110,5 @@ if(WITH_GPU AND TENSORRT_FOUND)
    endif()
    cc_test(test_trt_models SRCS trt_models_tester.cc  
      ARGS --dirname=${TRT_MODEL_INSTALL_DIR}/trt_test_models
-     DEPS paddle_inference_tensorrt_subgraph_engine)
+     DEPS paddle_inference_tensorrt_subgraph_engine SERIAL)
 endif()

paddle/fluid/operators/CMakeLists.txt

@@ -5,6 +5,8 @@ list(REMOVE_DUPLICATES GENERAL_OPS)
 set(DEPS_OPS "")
 set(pybind_file ${PADDLE_BINARY_DIR}/paddle/fluid/pybind/pybind.h)
 file(WRITE ${pybind_file} "// Generated by the paddle/fluid/operator/CMakeLists.txt.  DO NOT EDIT!\n\n")
+
+set(PART_CUDA_KERNEL_FILES)
 function(op_library TARGET)
     # op_library is a function to create op library. The interface is same as
     # cc_library. But it handle split GPU/CPU code and link some common library
@@ -37,6 +39,12 @@ function(op_library TARGET)
         if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
             list(APPEND cu_srcs ${TARGET}.cu)
         endif()
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu)
+            set(PART_CUDA_KERNEL_FILES ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu
+                    ${PART_CUDA_KERNEL_FILES} PARENT_SCOPE)
+            list(APPEND cu_srcs ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu)
+        endif()
+
         if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.hip.cu)
             list(APPEND hip_cu_srcs ${TARGET}.hip.cu)
         endif()
@@ -317,6 +325,7 @@ op_library(save_op DEPS lod_tensor)
 op_library(load_op DEPS lod_tensor)
 op_library(save_combine_op DEPS lod_tensor)
 op_library(load_combine_op DEPS lod_tensor)
+op_library(tensor_array_to_tensor_op DEPS concat_op)
 op_library(concat_op DEPS concat_and_split)
 
 list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
@@ -326,6 +335,8 @@ foreach(src ${GENERAL_OPS})
 endforeach()
 
 file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")
+
+
 if (NOT WIN32)
 add_subdirectory(reader)
 endif(NOT WIN32)
@@ -352,3 +363,14 @@ if(NOT WIN32)
     nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
 endif()
 nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor)
+
+if(WITH_GPU)
+    foreach(CUDA_KERNEL_FILE ${PART_CUDA_KERNEL_FILES})
+        file(READ ${CUDA_KERNEL_FILE} TARGET_CONTENT)
+        string(REGEX MATCH "REGISTER_OP_CUDA_KERNEL\\(\\n?([^,]+),.*" MATCHED ${TARGET_CONTENT})
+        if (MATCHED)
+            string(STRIP ${CMAKE_MATCH_1} MATCHED)
+            file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${MATCHED}, CUDA);\n")
+        endif()
+    endforeach()
+endif()

paddle/fluid/operators/activation_op.cc

@@ -91,16 +91,12 @@ class ActivationOp : public framework::OperatorWithKernel {
   }
 };
 
-class ActivationOpInferVarType : public framework::VarTypeInference {
- public:
-  void operator()(const framework::OpDesc& op_desc,
-                  framework::BlockDesc* block) const override {
-    auto x_name = op_desc.Input("X")[0];
-    auto out_name = op_desc.Output("Out")[0];
-    auto& x = block->FindRecursiveOrCreateVar(x_name);
-    auto& out = block->FindRecursiveOrCreateVar(out_name);
-    out.SetType(x.GetType());
-    out.SetDataType(x.GetDataType());
+class ActivationOpInferVarType
+    : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
+      const override {
+    return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Out"}};
   }
 };
 

paddle/fluid/operators/batch_norm_op.cc

@@ -170,6 +170,15 @@ The required data format for this layer is one of the following:
   }
 };
 
+class BatchNormOpInferVarType
+    : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
+      const override {
+    return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Y"}};
+  }
+};
+
 template <typename T>
 class BatchNormKernel<platform::CPUDeviceContext, T>
     : public framework::OpKernel<T> {
@@ -525,7 +534,7 @@ class BatchNormGradMaker : public framework::SingleGradOpDescMaker {
 
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker,
-                  ops::BatchNormGradMaker);
+                  ops::BatchNormOpInferVarType, ops::BatchNormGradMaker);
 REGISTER_OPERATOR(batch_norm_grad, ops::BatchNormGradOp);
 
 REGISTER_OP_CPU_KERNEL(

paddle/fluid/operators/conv_op.cc

@@ -224,6 +224,15 @@ $$
 )DOC");
 }
 
+class ConvOpInferVarType : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
+      const override {
+    return std::unordered_map<std::string, std::string>{
+        {"Input", /*->*/ "Output"}};
+  }
+};
+
 void Conv3DOpMaker::Make() {
   AddInput(
       "Input",
@@ -365,6 +374,7 @@ framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
 
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(conv2d, ops::ConvOp, ops::Conv2DOpMaker,
+                  ops::ConvOpInferVarType,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 REGISTER_OPERATOR(conv2d_grad, ops::ConvOpGrad);
 
@@ -372,7 +382,9 @@ REGISTER_OPERATOR(conv2d_grad, ops::ConvOpGrad);
 REGISTER_OPERATOR(depthwise_conv2d, ops::ConvOp, ops::Conv2DOpMaker,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 REGISTER_OPERATOR(depthwise_conv2d_grad, ops::ConvOpGrad);
+
 REGISTER_OPERATOR(conv3d, ops::ConvOp, ops::Conv3DOpMaker,
+                  ops::ConvOpInferVarType,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 REGISTER_OPERATOR(conv3d_grad, ops::ConvOpGrad);
 

paddle/fluid/operators/cross_entropy_op.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/cross_entropy_op.h"
+#include <string>
 
 namespace paddle {
 namespace operators {
@@ -179,6 +180,15 @@ or not. But the output only shares the LoD information with input X.
 )DOC");
   }
 };
+
+class CrossEntropyOpInferVarType
+    : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
+      const override {
+    return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Y"}};
+  }
+};
 }  // namespace operators
 }  // namespace paddle
 
@@ -186,6 +196,7 @@ namespace ops = paddle::operators;
 using CPUCtx = paddle::platform::CPUDeviceContext;
 
 REGISTER_OPERATOR(cross_entropy, ops::CrossEntropyOp, ops::CrossEntropyOpMaker,
+                  ops::CrossEntropyOpInferVarType,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 REGISTER_OPERATOR(cross_entropy_grad, ops::CrossEntropyGradientOp);
 REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel<CPUCtx, float>,

paddle/fluid/operators/detection/CMakeLists.txt

@@ -22,6 +22,7 @@ iou_similarity_op.cu)
 detection_library(mine_hard_examples_op SRCS mine_hard_examples_op.cc)
 detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc poly_util.cc gpc.cc)
 detection_library(prior_box_op SRCS prior_box_op.cc prior_box_op.cu)
+detection_library(density_prior_box_op SRCS density_prior_box_op.cc)
 detection_library(anchor_generator_op SRCS anchor_generator_op.cc
 anchor_generator_op.cu)
 detection_library(target_assign_op SRCS target_assign_op.cc

paddle/fluid/operators/detection/density_prior_box_op.cc

+/*Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+    http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/detection/density_prior_box_op.h"
+
+namespace paddle {
+namespace operators {
+
+class DensityPriorBoxOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"),
+                   "Input(Input) of DensityPriorBoxOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Image"),
+                   "Input(Image) of DensityPriorBoxOp should not be null.");
+
+    auto image_dims = ctx->GetInputDim("Image");
+    auto input_dims = ctx->GetInputDim("Input");
+    PADDLE_ENFORCE(image_dims.size() == 4, "The layout of image is NCHW.");
+    PADDLE_ENFORCE(input_dims.size() == 4, "The layout of input is NCHW.");
+
+    PADDLE_ENFORCE_LT(input_dims[2], image_dims[2],
+                      "The height of input must smaller than image.");
+
+    PADDLE_ENFORCE_LT(input_dims[3], image_dims[3],
+                      "The width of input must smaller than image.");
+    auto variances = ctx->Attrs().Get<std::vector<float>>("variances");
+
+    auto fixed_sizes = ctx->Attrs().Get<std::vector<float>>("fixed_sizes");
+    auto fixed_ratios = ctx->Attrs().Get<std::vector<float>>("fixed_ratios");
+    auto densities = ctx->Attrs().Get<std::vector<int>>("densities");
+
+    PADDLE_ENFORCE_EQ(fixed_sizes.size(), densities.size(),
+                      "The number of fixed_sizes and densities must be equal.");
+    size_t num_priors = 0;
+    if ((fixed_sizes.size() > 0) && (densities.size() > 0)) {
+      for (size_t i = 0; i < densities.size(); ++i) {
+        if (fixed_ratios.size() > 0) {
+          num_priors += (fixed_ratios.size()) * (pow(densities[i], 2));
+        }
+      }
+    }
+    std::vector<int64_t> dim_vec(4);
+    dim_vec[0] = input_dims[2];
+    dim_vec[1] = input_dims[3];
+    dim_vec[2] = num_priors;
+    dim_vec[3] = 4;
+    ctx->SetOutputDim("Boxes", framework::make_ddim(dim_vec));
+    ctx->SetOutputDim("Variances", framework::make_ddim(dim_vec));
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<framework::Tensor>("Input")->type()),
+        platform::CPUPlace());
+  }
+};
+
+class DensityPriorBoxOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput(
+        "Input",
+        "(Tensor, default Tensor<float>), "
+        "the input feature data of DensityPriorBoxOp, the layout is NCHW.");
+    AddInput("Image",
+             "(Tensor, default Tensor<float>), "
+             "the input image data of DensityPriorBoxOp, the layout is NCHW.");
+    AddOutput("Boxes",
+              "(Tensor, default Tensor<float>), the output prior boxes of "
+              "DensityPriorBoxOp. The layout is [H, W, num_priors, 4]. "
+              "H is the height of input, W is the width of input, num_priors "
+              "is the box count of each position.");
+    AddOutput("Variances",
+              "(Tensor, default Tensor<float>), the expanded variances of "
+              "DensityPriorBoxOp. The layout is [H, W, num_priors, 4]. "
+              "H is the height of input, W is the width of input, num_priors "
+              "is the box count of each position.");
+    AddAttr<std::vector<float>>("variances",
+                                "(vector<float>) List of variances to be "
+                                "encoded in density prior boxes.")
+        .AddCustomChecker([](const std::vector<float>& variances) {
+          PADDLE_ENFORCE_EQ(variances.size(), 4,
+                            "Must and only provide 4 variance.");
+          for (size_t i = 0; i < variances.size(); ++i) {
+            PADDLE_ENFORCE_GT(variances[i], 0.0,
+                              "variance[%d] must be greater than 0.", i);
+          }
+        });
+    AddAttr<bool>("clip", "(bool) Whether to clip out-of-boundary boxes.")
+        .SetDefault(true);
+
+    AddAttr<float>(
+        "step_w",
+        "Density prior boxes step across width, 0.0 for auto calculation.")
+        .SetDefault(0.0)
+        .AddCustomChecker([](const float& step_w) {
+          PADDLE_ENFORCE_GE(step_w, 0.0, "step_w should be larger than 0.");
+        });
+    AddAttr<float>(
+        "step_h",
+        "Density prior boxes step across height, 0.0 for auto calculation.")
+        .SetDefault(0.0)
+        .AddCustomChecker([](const float& step_h) {
+          PADDLE_ENFORCE_GE(step_h, 0.0, "step_h should be larger than 0.");
+        });
+
+    AddAttr<float>("offset",
+                   "(float) "
+                   "Density prior boxes center offset.")
+        .SetDefault(0.5);
+    AddAttr<std::vector<float>>("fixed_sizes",
+                                "(vector<float>) List of fixed sizes "
+                                "of generated density prior boxes.")
+        .SetDefault(std::vector<float>{})
+        .AddCustomChecker([](const std::vector<float>& fixed_sizes) {
+          for (size_t i = 0; i < fixed_sizes.size(); ++i) {
+            PADDLE_ENFORCE_GT(fixed_sizes[i], 0.0,
+                              "fixed_sizes[%d] should be larger than 0.", i);
+          }
+        });
+
+    AddAttr<std::vector<float>>("fixed_ratios",
+                                "(vector<float>) List of fixed ratios "
+                                "of generated density prior boxes.")
+        .SetDefault(std::vector<float>{})
+        .AddCustomChecker([](const std::vector<float>& fixed_ratios) {
+          for (size_t i = 0; i < fixed_ratios.size(); ++i) {
+            PADDLE_ENFORCE_GT(fixed_ratios[i], 0.0,
+                              "fixed_ratios[%d] should be larger than 0.", i);
+          }
+        });
+
+    AddAttr<std::vector<int>>("densities",
+                              "(vector<float>) List of densities "
+                              "of generated density prior boxes.")
+        .SetDefault(std::vector<int>{})
+        .AddCustomChecker([](const std::vector<int>& densities) {
+          for (size_t i = 0; i < densities.size(); ++i) {
+            PADDLE_ENFORCE_GT(densities[i], 0,
+                              "densities[%d] should be larger than 0.", i);
+          }
+        });
+    AddComment(R"DOC(
+        Density Prior box operator
+        Each position of the input produce N density prior boxes, N is determined by
+        the count of fixed_ratios, densities, the calculation of N is as follows:
+        for density in densities:
+        N += size(fixed_ratios)*density^2
+        )DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(density_prior_box, ops::DensityPriorBoxOp,
+                  ops::DensityPriorBoxOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+
+REGISTER_OP_CPU_KERNEL(density_prior_box, ops::DensityPriorBoxOpKernel<float>,
+                       ops::DensityPriorBoxOpKernel<double>);

paddle/fluid/operators/detection/density_prior_box_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+    http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <algorithm>
+#include <vector>
+#include "paddle/fluid/operators/detection/prior_box_op.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* input = ctx.Input<paddle::framework::Tensor>("Input");
+    auto* image = ctx.Input<paddle::framework::Tensor>("Image");
+    auto* boxes = ctx.Output<paddle::framework::Tensor>("Boxes");
+    auto* vars = ctx.Output<paddle::framework::Tensor>("Variances");
+
+    auto variances = ctx.Attr<std::vector<float>>("variances");
+    auto clip = ctx.Attr<bool>("clip");
+
+    auto fixed_sizes = ctx.Attr<std::vector<float>>("fixed_sizes");
+    auto fixed_ratios = ctx.Attr<std::vector<float>>("fixed_ratios");
+    auto densities = ctx.Attr<std::vector<int>>("densities");
+
+    T step_w = static_cast<T>(ctx.Attr<float>("step_w"));
+    T step_h = static_cast<T>(ctx.Attr<float>("step_h"));
+    T offset = static_cast<T>(ctx.Attr<float>("offset"));
+
+    auto img_width = image->dims()[3];
+    auto img_height = image->dims()[2];
+
+    auto feature_width = input->dims()[3];
+    auto feature_height = input->dims()[2];
+
+    T step_width, step_height;
+    if (step_w == 0 || step_h == 0) {
+      step_width = static_cast<T>(img_width) / feature_width;
+      step_height = static_cast<T>(img_height) / feature_height;
+    } else {
+      step_width = step_w;
+      step_height = step_h;
+    }
+    int num_priors = 0;
+    if (fixed_sizes.size() > 0 && densities.size() > 0) {
+      for (size_t i = 0; i < densities.size(); ++i) {
+        if (fixed_ratios.size() > 0) {
+          num_priors += (fixed_ratios.size()) * (pow(densities[i], 2));
+        }
+      }
+    }
+
+    boxes->mutable_data<T>(ctx.GetPlace());
+    vars->mutable_data<T>(ctx.GetPlace());
+    auto e_boxes = framework::EigenTensor<T, 4>::From(*boxes).setConstant(0.0);
+
+    int step_average = static_cast<int>((step_width + step_height) * 0.5);
+
+    for (int h = 0; h < feature_height; ++h) {
+      for (int w = 0; w < feature_width; ++w) {
+        T center_x = (w + offset) * step_width;
+        T center_y = (h + offset) * step_height;
+        int idx = 0;
+        // Generate density prior boxes with fixed sizes.
+        for (size_t s = 0; s < fixed_sizes.size(); ++s) {
+          auto fixed_size = fixed_sizes[s];
+          int density = densities[s];
+          // Generate density prior boxes with fixed ratios.
+          if (fixed_ratios.size() > 0) {
+            for (size_t r = 0; r < fixed_ratios.size(); ++r) {
+              float ar = fixed_ratios[r];
+              int shift = step_average / density;
+              float box_width_ratio = fixed_size * sqrt(ar);
+              float box_height_ratio = fixed_size / sqrt(ar);
+              for (int di = 0; di < density; ++di) {
+                for (int dj = 0; dj < density; ++dj) {
+                  float center_x_temp =
+                      center_x - step_average / 2. + shift / 2. + dj * shift;
+                  float center_y_temp =
+                      center_y - step_average / 2. + shift / 2. + di * shift;
+                  e_boxes(h, w, idx, 0) =
+                      (center_x_temp - box_width_ratio / 2.) / img_width >= 0
+                          ? (center_x_temp - box_width_ratio / 2.) / img_width
+                          : 0;
+                  e_boxes(h, w, idx, 1) =
+                      (center_y_temp - box_height_ratio / 2.) / img_height >= 0
+                          ? (center_y_temp - box_height_ratio / 2.) / img_height
+                          : 0;
+                  e_boxes(h, w, idx, 2) =
+                      (center_x_temp + box_width_ratio / 2.) / img_width <= 1
+                          ? (center_x_temp + box_width_ratio / 2.) / img_width
+                          : 1;
+                  e_boxes(h, w, idx, 3) =
+                      (center_y_temp + box_height_ratio / 2.) / img_height <= 1
+                          ? (center_y_temp + box_height_ratio / 2.) / img_height
+                          : 1;
+                  idx++;
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+    if (clip) {
+      platform::Transform<platform::CPUDeviceContext> trans;
+      ClipFunctor<T> clip_func;
+      trans(ctx.template device_context<platform::CPUDeviceContext>(),
+            boxes->data<T>(), boxes->data<T>() + boxes->numel(),
+            boxes->data<T>(), clip_func);
+    }
+    framework::Tensor var_t;
+    var_t.mutable_data<T>(
+        framework::make_ddim({1, static_cast<int>(variances.size())}),
+        ctx.GetPlace());
+
+    auto var_et = framework::EigenTensor<T, 2>::From(var_t);
+
+    for (size_t i = 0; i < variances.size(); ++i) {
+      var_et(0, i) = variances[i];
+    }
+
+    int box_num = feature_height * feature_width * num_priors;
+    auto var_dim = vars->dims();
+    vars->Resize({box_num, static_cast<int>(variances.size())});
+
+    auto e_vars = framework::EigenMatrix<T, Eigen::RowMajor>::From(*vars);
+
+    e_vars = var_et.broadcast(Eigen::DSizes<int, 2>(box_num, 1));
+
+    vars->Resize(var_dim);
+  }
+};  // namespace operators
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/elementwise_op.h

@@ -75,16 +75,12 @@ class ElementwiseOp : public framework::OperatorWithKernel {
   }
 };
 
-class ElementwiseOpInferVarType : public framework::VarTypeInference {
- public:
-  void operator()(const framework::OpDesc &op_desc,
-                  framework::BlockDesc *block) const override {
-    auto x_name = op_desc.Input("X")[0];
-    auto out_name = op_desc.Output("Out")[0];
-    auto &x = block->FindRecursiveOrCreateVar(x_name);
-    auto &out = block->FindRecursiveOrCreateVar(out_name);
-    out.SetType(x.GetType());
-    out.SetDataType(x.GetDataType());
+class ElementwiseOpInferVarType
+    : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
+      const override {
+    return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Out"}};
   }
 };
 

paddle/fluid/operators/gather.cu.h

@@ -50,7 +50,9 @@ void GPUGather(const platform::DeviceContext& ctx, const Tensor& src,
                const Tensor& index, Tensor* output) {
   // PADDLE_ENFORCE(platform::is_gpu_place(place));
   // check index of shape 1-D
-  PADDLE_ENFORCE(index.dims().size() == 1);
+  PADDLE_ENFORCE(index.dims().size() == 1 ||
+                 (index.dims().size() == 2 && index.dims()[1] == 1));
+
   int index_size = index.dims()[0];
 
   auto src_dims = src.dims();

paddle/fluid/operators/gather.h

@@ -38,7 +38,8 @@ void CPUGather(const platform::DeviceContext& ctx, const Tensor& src,
                const Tensor& index, Tensor* output) {
   PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()));
   // check index of shape 1-D
-  PADDLE_ENFORCE(index.dims().size() == 1);
+  PADDLE_ENFORCE(index.dims().size() == 1 ||
+                 (index.dims().size() == 2 && index.dims()[1] == 1));
   int64_t index_size = index.dims()[0];
 
   auto src_dims = src.dims();

paddle/fluid/operators/gather_op.cc

@@ -31,7 +31,8 @@ class GatherOp : public framework::OperatorWithKernel {
                    "Output(Out) of GatherOp should not be null.");
 
     auto index_dims = ctx->GetInputDim("Index");
-    PADDLE_ENFORCE(index_dims.size() == 1);
+    PADDLE_ENFORCE(index_dims.size() == 1 ||
+                   (index_dims.size() == 2 && index_dims[1] == 1));
     int batch_size = ctx->GetInputDim("Index")[0];
     framework::DDim output_dims(ctx->GetInputDim("X"));
     output_dims[0] = batch_size;
@@ -53,6 +54,7 @@ class GatherGradOp : public framework::OperatorWithKernel {
 
   void InferShape(framework::InferShapeContext* ctx) const override {
     ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
+    ctx->ShareLoD("X", /*-->*/ framework::GradVarName("X"));
   }
 
  protected:
@@ -75,7 +77,7 @@ Gather Operator.
 
 $Out = X[Index]$
 
-Out is obtained by gathering entries of the outer-most dimension 
+Out is obtained by gathering entries of the outer-most dimension
 of X indexed by Index and concatenate them together.
 
 Example:

paddle/fluid/operators/mean_op.cc

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/mean_op.h"
-
+#include <string>
 namespace paddle {
 namespace operators {
 
@@ -42,6 +42,14 @@ Mean Operator calculates the mean of all elements in X.
   }
 };
 
+class MeanOpInferVarType : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
+      const override {
+    return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Out"}};
+  }
+};
+
 class MeanGradOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -50,6 +58,14 @@ class MeanGradOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
     ctx->ShareLoD("X", framework::GradVarName("X"));
   }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    auto input_data_type =
+        framework::ToDataType(ctx.Input<Tensor>("X")->type());
+
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
 };
 
 class MeanGradMaker : public framework::SingleGradOpDescMaker {
@@ -71,7 +87,8 @@ class MeanGradMaker : public framework::SingleGradOpDescMaker {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OPERATOR(mean, ops::MeanOp, ops::MeanOpMaker, ops::MeanGradMaker);
+REGISTER_OPERATOR(mean, ops::MeanOp, ops::MeanOpMaker, ops::MeanOpInferVarType,
+                  ops::MeanGradMaker);
 REGISTER_OPERATOR(mean_grad, ops::MeanGradOp);
 REGISTER_OP_CPU_KERNEL(
     mean, ops::MeanKernel<paddle::platform::CPUDeviceContext, float>,

paddle/fluid/operators/mul_op.cc

@@ -126,6 +126,14 @@ or not. But the output only shares the LoD information with input $X$.
   }
 };
 
+class MulOpInferVarType : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
+      const override {
+    return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Out"}};
+  }
+};
+
 class MulGradOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -178,7 +186,8 @@ class MulOpGradMaker : public framework::SingleGradOpDescMaker {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OPERATOR(mul, ops::MulOp, ops::MulOpMaker, ops::MulOpGradMaker);
+REGISTER_OPERATOR(mul, ops::MulOp, ops::MulOpMaker, ops::MulOpInferVarType,
+                  ops::MulOpGradMaker);
 REGISTER_OPERATOR(mul_grad, ops::MulGradOp);
 REGISTER_OP_CPU_KERNEL(
     mul, ops::MulKernel<paddle::platform::CPUDeviceContext, float>,

paddle/fluid/operators/pool_op.cc

@@ -40,7 +40,7 @@ int PoolOutputSize(int input_size, int filter_size, int padding, int stride,
   return output_size;
 }
 
-void PoolOp::InferShape(framework::InferShapeContext *ctx) const {
+void PoolOp::InferShape(framework::InferShapeContext* ctx) const {
   PADDLE_ENFORCE(ctx->HasInput("X"), "X(Input) of Pooling should not be null.");
   PADDLE_ENFORCE(ctx->HasOutput("Out"),
                  "Out(Output) of Pooling should not be null.");
@@ -81,7 +81,7 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const {
 }
 
 framework::OpKernelType PoolOp::GetExpectedKernelType(
-    const framework::ExecutionContext &ctx) const {
+    const framework::ExecutionContext& ctx) const {
   framework::LibraryType library_{framework::LibraryType::kPlain};
   std::string data_format = ctx.Attr<std::string>("data_format");
   framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
@@ -104,7 +104,7 @@ framework::OpKernelType PoolOp::GetExpectedKernelType(
       layout_, library_);
 }
 
-void PoolOpGrad::InferShape(framework::InferShapeContext *ctx) const {
+void PoolOpGrad::InferShape(framework::InferShapeContext* ctx) const {
   PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null.");
   PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
                  "Input(X@GRAD) should not be null.");
@@ -112,7 +112,7 @@ void PoolOpGrad::InferShape(framework::InferShapeContext *ctx) const {
 }
 
 framework::OpKernelType PoolOpGrad::GetExpectedKernelType(
-    const framework::ExecutionContext &ctx) const {
+    const framework::ExecutionContext& ctx) const {
   framework::LibraryType library_{framework::LibraryType::kPlain};
   std::string data_format = ctx.Attr<std::string>("data_format");
   framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
@@ -262,6 +262,14 @@ Example:
 )DOC");
 }
 
+class PoolOpInferVarType : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
+      const override {
+    return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Out"}};
+  }
+};
+
 void Pool3dOpMaker::Make() {
   AddInput("X",
            "(Tensor) The input tensor of pooling operator. "
@@ -372,6 +380,7 @@ Example:
 namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(pool2d, ops::PoolOp, ops::Pool2dOpMaker,
+                  ops::PoolOpInferVarType,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 REGISTER_OPERATOR(pool2d_grad, ops::PoolOpGrad);
 
@@ -383,6 +392,7 @@ REGISTER_OP_CPU_KERNEL(
     ops::PoolGradKernel<paddle::platform::CPUDeviceContext, double>);
 
 REGISTER_OPERATOR(pool3d, ops::PoolOp, ops::Pool3dOpMaker,
+                  ops::PoolOpInferVarType,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 REGISTER_OPERATOR(pool3d_grad, ops::PoolOpGrad);
 

paddle/fluid/operators/reduce_max_op.cu

@@ -23,12 +23,3 @@ REGISTER_OP_CUDA_KERNEL(reduce_max,
                                           int, ops::MaxFunctor>,
                         ops::ReduceKernel<paddle::platform::CUDADeviceContext,
                                           int64_t, ops::MaxFunctor>);
-REGISTER_OP_CUDA_KERNEL(
-    reduce_max_grad, ops::ReduceGradKernel<paddle::platform::CUDADeviceContext,
-                                           float, ops::MaxOrMinGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, double,
-                          ops::MaxOrMinGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int,
-                          ops::MaxOrMinGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int64_t,
-                          ops::MaxOrMinGradFunctor>);

paddle/fluid/operators/reduce_max_op.part.cu

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/operators/reduce_min_max_op.h"
+
+REGISTER_OP_CUDA_KERNEL(
+    reduce_max_grad, ops::ReduceGradKernel<paddle::platform::CUDADeviceContext,
+                                           float, ops::MaxOrMinGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, double,
+                          ops::MaxOrMinGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int,
+                          ops::MaxOrMinGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int64_t,
+                          ops::MaxOrMinGradFunctor>);

paddle/fluid/operators/reduce_mean_op.cu

@@ -69,13 +69,3 @@ REGISTER_OP_CUDA_KERNEL(reduce_mean, ops::ReduceMeanKernel<float>,
                         ops::ReduceMeanKernel<double>,
                         ops::ReduceMeanKernel<int>,
                         ops::ReduceMeanKernel<int64_t>);
-
-REGISTER_OP_CUDA_KERNEL(
-    reduce_mean_grad, ops::ReduceGradKernel<paddle::platform::CUDADeviceContext,
-                                            float, ops::MeanGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, double,
-                          ops::MeanGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int,
-                          ops::MeanGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int64_t,
-                          ops::MeanGradFunctor>);

paddle/fluid/operators/reduce_mean_op.part.cu

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// .part used to speed up nvcc compile
+#include "paddle/fluid/operators/reduce_mean_op.h"
+
+REGISTER_OP_CUDA_KERNEL(
+    reduce_mean_grad, ops::ReduceGradKernel<paddle::platform::CUDADeviceContext,
+                                            float, ops::MeanGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, double,
+                          ops::MeanGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int,
+                          ops::MeanGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int64_t,
+                          ops::MeanGradFunctor>);

paddle/fluid/operators/reduce_min_op.cu

@@ -23,12 +23,3 @@ REGISTER_OP_CUDA_KERNEL(reduce_min,
                                           int, ops::MinFunctor>,
                         ops::ReduceKernel<paddle::platform::CUDADeviceContext,
                                           int64_t, ops::MinFunctor>);
-REGISTER_OP_CUDA_KERNEL(
-    reduce_min_grad, ops::ReduceGradKernel<paddle::platform::CUDADeviceContext,
-                                           float, ops::MaxOrMinGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, double,
-                          ops::MaxOrMinGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int,
-                          ops::MaxOrMinGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int64_t,
-                          ops::MaxOrMinGradFunctor>);

paddle/fluid/operators/reduce_min_op.part.cu

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/operators/reduce_min_max_op.h"
+
+REGISTER_OP_CUDA_KERNEL(
+    reduce_min_grad, ops::ReduceGradKernel<paddle::platform::CUDADeviceContext,
+                                           float, ops::MaxOrMinGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, double,
+                          ops::MaxOrMinGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int,
+                          ops::MaxOrMinGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int64_t,
+                          ops::MaxOrMinGradFunctor>);

paddle/fluid/operators/reduce_prod_op.cu

@@ -23,12 +23,3 @@ REGISTER_OP_CUDA_KERNEL(reduce_prod,
                                           int, ops::ProdFunctor>,
                         ops::ReduceKernel<paddle::platform::CUDADeviceContext,
                                           int64_t, ops::ProdFunctor>);
-REGISTER_OP_CUDA_KERNEL(
-    reduce_prod_grad, ops::ReduceGradKernel<paddle::platform::CUDADeviceContext,
-                                            float, ops::ProdGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, double,
-                          ops::ProdGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int,
-                          ops::ProdGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int64_t,
-                          ops::ProdGradFunctor>);

paddle/fluid/operators/reduce_prod_op.part.cu

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/operators/reduce_prod_op.h"
+
+REGISTER_OP_CUDA_KERNEL(
+    reduce_prod_grad, ops::ReduceGradKernel<paddle::platform::CUDADeviceContext,
+                                            float, ops::ProdGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, double,
+                          ops::ProdGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int,
+                          ops::ProdGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int64_t,
+                          ops::ProdGradFunctor>);

paddle/fluid/operators/reduce_sum_op.cu

@@ -64,13 +64,3 @@ class ReduceSumKernel : public framework::OpKernel<T> {
 REGISTER_OP_CUDA_KERNEL(reduce_sum, ops::ReduceSumKernel<float>,
                         ops::ReduceSumKernel<double>, ops::ReduceSumKernel<int>,
                         ops::ReduceSumKernel<int64_t>);
-
-REGISTER_OP_CUDA_KERNEL(
-    reduce_sum_grad, ops::ReduceGradKernel<paddle::platform::CUDADeviceContext,
-                                           float, ops::SumGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, double,
-                          ops::SumGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int,
-                          ops::SumGradFunctor>,
-    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int64_t,
-                          ops::SumGradFunctor>);

paddle/fluid/operators/reduce_sum_op.part.cu

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/operators/cub_reduce.h"
+#include "paddle/fluid/operators/reduce_sum_op.h"
+
+REGISTER_OP_CUDA_KERNEL(
+    reduce_sum_grad, ops::ReduceGradKernel<paddle::platform::CUDADeviceContext,
+                                           float, ops::SumGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, double,
+                          ops::SumGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int,
+                          ops::SumGradFunctor>,
+    ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int64_t,
+                          ops::SumGradFunctor>);

paddle/fluid/operators/scatter.cu.h

@@ -51,7 +51,8 @@ void GPUScatterAssign(const platform::DeviceContext& ctx, const Tensor& src,
                       const Tensor& index, Tensor* output) {
   // PADDLE_ENFORCE(platform::is_gpu_place(place));
   // check index of shape 1-D
-  PADDLE_ENFORCE(index.dims().size() == 1);
+  PADDLE_ENFORCE(index.dims().size() == 1 ||
+                 (index.dims().size() == 2 && index.dims()[1] == 1));
   int index_size = index.dims()[0];
 
   auto src_dims = src.dims();

paddle/fluid/operators/scatter.h

@@ -37,7 +37,8 @@ void ScatterAssign(const platform::DeviceContext& ctx, const Tensor& src,
                    const Tensor& index, Tensor* output) {
   PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()));
   // check index of shape 1-D
-  PADDLE_ENFORCE(index.dims().size() == 1);
+  PADDLE_ENFORCE(index.dims().size() == 1 ||
+                 (index.dims().size() == 2 && index.dims()[1] == 1));
   int index_size = index.dims()[0];
 
   auto src_dims = src.dims();

paddle/fluid/operators/similarity_focus_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/similarity_focus_op.h"
+
+namespace paddle {
+namespace operators {
+class SimilarityFocusOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor, default Tensor<float>), a 4-D tensor with shape,"
+             " [BatchSize, X, Y, Z]");
+    AddOutput("Out",
+              "(Tensor, default Tensor<float>), the similarity focus mask"
+              " with the same shape of input X.");
+    AddAttr<int>("axis",
+                 "(int32), indicating the dimension to be select. It can"
+                 " only be 1, 2, or 3.");
+    AddAttr<std::vector<int>>("indexes",
+                              "(std::vector<int32>), indicating the indexes"
+                              " of the selected dimension.");
+    AddComment(R"DOC(
+SimilarityFocus Operator.
+
+Generate a similarity focus mask with the same shape of input using the following method:
+1. Extract the 3-D tensor(here the first dimension is BatchSize) corresponding 
+   to the axis according to the indexes. For example, if axis=1 and indexes=[a], 
+   it will get the matrix T=X[:, a, :, :]. In this case, if the shape of input X 
+   is (BatchSize, A, B, C), the shape of tensor T is (BatchSize, B, C).
+2. For each index, find the largest numbers in the tensor T, so that the same 
+   row and same column has at most one number(what it means is that if the 
+   largest number has been found in the i-th row and the j-th column, then 
+   the numbers in the i-th row or j-th column will be skipped. And then the 
+   next largest number will be selected from the remaining numbers. Obviously 
+   there will be min(B, C) numbers), and mark the corresponding position of the 
+   3-D similarity focus mask as 1, otherwise as 0. Do elementwise-or for 
+   each index.
+3. Broadcast the 3-D similarity focus mask to the same shape of input X.
+
+Refer to `Similarity Focus Layer <http://www.aclweb.org/anthology/N16-1108>`_
+)DOC");
+  }
+};
+
+class SimilarityFocusOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should be not null.");
+    auto x_dims = ctx->GetInputDim("X");
+    PADDLE_ENFORCE_EQ(x_dims.size(), 4, "Input(X)'s rank should be 4.");
+    ctx->SetOutputDim("Out", x_dims);
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("X")->type()),
+        platform::CPUPlace());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(similarity_focus, ops::SimilarityFocusOp,
+                  ops::SimilarityFocusOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+REGISTER_OP_CPU_KERNEL(similarity_focus, ops::SimilarityFocusKernel<float>,
+                       ops::SimilarityFocusKernel<double>);

paddle/fluid/operators/similarity_focus_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <algorithm>
+#include <cstring>
+#include <utility>
+#include <vector>
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+using Tensor = framework::Tensor;
+
+template <typename T>
+class SimilarityFocusKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    Tensor* out = context.Output<Tensor>("Out");
+    const Tensor* x = context.Input<Tensor>("X");
+    T* out_data = out->mutable_data<T>(context.GetPlace());
+    const T* x_data = x->data<T>();
+
+    int axis = context.Attr<int>("axis");
+    std::vector<int> indexes = context.Attr<std::vector<int>>("indexes");
+
+    int64_t batch_size = x->dims()[0];
+    int64_t dim[4];
+    for (int i = 1; i <= 3; ++i) {
+      dim[i] = x->dims()[i];
+    }
+
+    if (indexes.size() < 1) {
+      PADDLE_THROW("Indexes' size can not be 0.");
+    }
+    for (auto index : indexes) {
+      if (dim[axis] < index) {
+        PADDLE_THROW("Index exceeds tensor shape limit.");
+      }
+    }
+
+    int64_t array_size = 1;
+    for (int i = 1; i <= 3; ++i) {
+      if (i != axis) {
+        array_size *= dim[i];
+      }
+    }
+
+    std::vector<std::pair<T, int64_t>> array(array_size);
+
+    bool (*cmp)(std::pair<T, int64_t>, std::pair<T, int64_t>) = [](
+        std::pair<T, int64_t> x, std::pair<T, int64_t> y) {
+      return x.first > y.first;
+    };
+
+    int64_t (*compute_index)(int64_t*, int, int, int, int) = [](
+        int64_t* dim, int d1, int d2, int d3, int d4) {
+      return d1 * dim[1] * dim[2] * dim[3] + d2 * dim[2] * dim[3] +
+             d3 * dim[3] + d4;
+    };
+
+    memset(out_data, 0, sizeof(T) * batch_size * dim[1] * dim[2] * dim[3]);
+    for (int i = 0; i < batch_size; ++i) {
+      for (auto index : indexes) {
+        if (axis == 1) {
+          for (int j = 0; j < dim[2]; ++j) {
+            for (int k = 0; k < dim[3]; ++k) {
+              array[j * dim[3] + k] = std::make_pair(
+                  x_data[compute_index(dim, i, index, j, k)], j * dim[3] + k);
+            }
+          }
+
+          std::sort(array.begin(), array.end(), cmp);
+          int tag_num = 0;
+          std::vector<bool> tag2(dim[2]), tag3(dim[3]);
+          for (auto x : array) {
+            int idx2 = x.second / dim[3];
+            int idx3 = x.second % dim[3];
+            if (tag2[idx2] || tag3[idx3]) {
+              continue;
+            }
+            tag_num++;
+            tag2[idx2] = true;
+            tag3[idx3] = true;
+            for (int j = 0; j < dim[1]; ++j) {
+              out_data[compute_index(dim, i, j, idx2, idx3)] = 1;
+            }
+            if (tag_num == std::min(dim[2], dim[3])) {
+              break;
+            }
+          }
+        } else if (axis == 2) {
+          for (int j = 0; j < dim[1]; ++j) {
+            for (int k = 0; k < dim[3]; ++k) {
+              array[j * dim[3] + k] = std::make_pair(
+                  x_data[compute_index(dim, i, j, index, k)], j * dim[3] + k);
+            }
+          }
+
+          std::sort(array.begin(), array.end(), cmp);
+          int tag_num = 0;
+          std::vector<bool> tag1(dim[1]), tag3(dim[3]);
+          for (auto x : array) {
+            int idx1 = x.second / dim[3];
+            int idx3 = x.second % dim[3];
+            if (tag1[idx1] || tag3[idx3]) {
+              continue;
+            }
+            tag_num++;
+            tag1[idx1] = true;
+            tag3[idx3] = true;
+            for (int j = 0; j < dim[2]; ++j) {
+              out_data[compute_index(dim, i, idx1, j, idx3)] = 1;
+            }
+            if (tag_num == std::min(dim[1], dim[3])) {
+              break;
+            }
+          }
+        } else if (axis == 3) {
+          for (int j = 0; j < dim[1]; ++j) {
+            for (int k = 0; k < dim[2]; ++k) {
+              array[j * dim[2] + k] = std::make_pair(
+                  x_data[compute_index(dim, i, j, k, index)], j * dim[2] + k);
+            }
+          }
+
+          std::sort(array.begin(), array.end(), cmp);
+          int tag_num = 0;
+          std::vector<bool> tag1(dim[1]), tag2(dim[2]);
+          for (auto x : array) {
+            int idx1 = x.second / dim[2];
+            int idx2 = x.second % dim[2];
+            if (tag1[idx1] || tag2[idx2]) {
+              continue;
+            }
+            tag_num++;
+            tag1[idx1] = true;
+            tag2[idx2] = true;
+            for (int j = 0; j < dim[3]; ++j) {
+              out_data[compute_index(dim, i, idx1, idx2, j)] = 1;
+            }
+            if (tag_num == std::min(dim[1], dim[2])) {
+              break;
+            }
+          }
+        } else {
+          PADDLE_THROW("Axis must be 1 or 2 or 3");
+        }
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/softmax_op.cc

@@ -124,6 +124,14 @@ For each row $i$ and each column $j$ in the matrix, we have:
   }
 };
 
+class SoftmaxOpInferVarType : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
+      const override {
+    return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Out"}};
+  }
+};
+
 class SoftmaxOpGrad : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -196,7 +204,7 @@ class SoftmaxOpGradMaker : public framework::SingleGradOpDescMaker {
 namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker,
-                  ops::SoftmaxOpGradMaker);
+                  ops::SoftmaxOpInferVarType, ops::SoftmaxOpGradMaker);
 REGISTER_OPERATOR(softmax_grad, ops::SoftmaxOpGrad);
 REGISTER_OP_CPU_KERNEL(
     softmax, ops::SoftmaxKernel<paddle::platform::CPUDeviceContext, float>,

paddle/fluid/operators/stack_op.cc

@@ -21,8 +21,12 @@ REGISTER_OPERATOR(stack, ops::StackOp, ops::StackOpMaker,
 REGISTER_OPERATOR(stack_grad, ops::StackOpGrad);
 
 REGISTER_OP_CPU_KERNEL(stack, ops::StackKernel<plat::CPUDeviceContext, float>,
-                       ops::StackKernel<plat::CPUDeviceContext, double>);
+                       ops::StackKernel<plat::CPUDeviceContext, double>,
+                       ops::StackKernel<plat::CPUDeviceContext, int>,
+                       ops::StackKernel<plat::CPUDeviceContext, int64_t>);
 
 REGISTER_OP_CPU_KERNEL(stack_grad,
                        ops::StackGradKernel<plat::CPUDeviceContext, float>,
-                       ops::StackGradKernel<plat::CPUDeviceContext, double>);
+                       ops::StackGradKernel<plat::CPUDeviceContext, double>,
+                       ops::StackGradKernel<plat::CPUDeviceContext, int>,
+                       ops::StackGradKernel<plat::CPUDeviceContext, int64_t>);

paddle/fluid/operators/stack_op.cu

@@ -18,8 +18,12 @@ namespace plat = paddle::platform;
 namespace ops = paddle::operators;
 
 REGISTER_OP_CUDA_KERNEL(stack, ops::StackKernel<plat::CUDADeviceContext, float>,
-                        ops::StackKernel<plat::CUDADeviceContext, double>);
+                        ops::StackKernel<plat::CUDADeviceContext, double>,
+                        ops::StackKernel<plat::CUDADeviceContext, int>,
+                        ops::StackKernel<plat::CUDADeviceContext, int64_t>);
 
 REGISTER_OP_CUDA_KERNEL(stack_grad,
                         ops::StackGradKernel<plat::CUDADeviceContext, float>,
-                        ops::StackGradKernel<plat::CUDADeviceContext, double>);
+                        ops::StackGradKernel<plat::CUDADeviceContext, double>,
+                        ops::StackGradKernel<plat::CUDADeviceContext, int>,
+                        ops::StackGradKernel<plat::CUDADeviceContext, int64_t>);

paddle/fluid/operators/tensor_array_to_tensor_op.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <string>
+#include <vector>
+
+#include "paddle/fluid/framework/lod_tensor_array.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/variable.h"
+
+namespace paddle {
+namespace operators {
+using framework::Tensor;
+
+void LodTensorArray2LodTensorVector(const framework::Scope &scope,
+                                    const std::string &base_name,
+                                    const std::string &lod_tensor_array_name,
+                                    std::vector<std::string> *res_names) {
+  auto &inx =
+      scope.FindVar(lod_tensor_array_name)->Get<framework::LoDTensorArray>();
+  for (size_t i = 0; i < inx.size(); i++) {
+    std::string var_name = base_name + std::to_string(i);
+    framework::Variable *g_feed_value =
+        const_cast<framework::Scope &>(scope).Var(var_name);
+    auto &feed_input =
+        *(g_feed_value->GetMutable<paddle::framework::LoDTensor>());
+    feed_input.ShareDataWith(inx[i]);
+    res_names->push_back(var_name);
+  }
+}
+
+void LodTensorVectorResizeFromLodTensorArray(
+    const framework::Scope &scope, const std::string &base_name,
+    const std::string &lod_tensor_array_name,
+    std::vector<std::string> *res_names) {
+  auto &inx =
+      scope.FindVar(lod_tensor_array_name)->Get<framework::LoDTensorArray>();
+  for (size_t i = 0; i < inx.size(); i++) {
+    std::string var_name = base_name + std::to_string(i);
+    framework::Variable *g_feed_value =
+        const_cast<framework::Scope &>(scope).Var(var_name);
+    auto &feed_input =
+        *(g_feed_value->GetMutable<paddle::framework::LoDTensor>());
+    auto dims = inx[i].dims();
+    feed_input.Resize(dims);
+    res_names->push_back(var_name);
+  }
+}
+
+void LodTensorArrayCreateFromLodTensorArray(
+    const framework::Scope &scope,
+    const std::string &input_lod_tensor_array_name,
+    const std::string &output_lod_tensor_array_name) {
+  auto &inx = scope.FindVar(input_lod_tensor_array_name)
+                  ->Get<framework::LoDTensorArray>();
+  auto &grad_inx = *scope.FindVar(output_lod_tensor_array_name)
+                        ->GetMutable<framework::LoDTensorArray>();
+
+  for (size_t i = 0; i < inx.size(); i++) {
+    std::string var_name = output_lod_tensor_array_name + std::to_string(i);
+    framework::Variable *g_feed_value =
+        const_cast<framework::Scope &>(scope).Var(var_name);
+    auto &feed_input =
+        *(g_feed_value->GetMutable<paddle::framework::LoDTensor>());
+    grad_inx.push_back(feed_input);
+  }
+}
+
+class LoDTensorArray2TensorOp : public framework::OperatorBase {
+ public:
+  using OperatorBase::OperatorBase;
+
+ private:
+  void RunImpl(const framework::Scope &scope,
+               const platform::Place &place) const override {
+    auto axis = Attr<int>("axis");
+
+    framework::AttributeMap attrs;
+    attrs["axis"] = axis;
+
+    auto &inx = scope.FindVar(Input("X"))->Get<framework::LoDTensorArray>();
+    auto &out =
+        *scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();
+    auto &out_inx =
+        *scope.FindVar(Output("OutIndex"))->GetMutable<framework::LoDTensor>();
+
+    const size_t n = inx.size();
+    PADDLE_ENFORCE_GT(n, 0, "Input tensorarray size should > 0.");
+
+    std::string base_name = Inputs("X")[0];
+    std::vector<std::string> names;
+
+    // get the input tensorarray items' dim in out_inx
+    auto out_inx_dim = out_inx.dims();
+    out_inx_dim[0] = inx.size();
+    out_inx.Resize(out_inx_dim);
+
+    std::string var_name = "out_index";
+    framework::Variable *tmp_index_var =
+        const_cast<framework::Scope &>(scope).Var(var_name);
+    auto &tmp_index_tensor =
+        *(tmp_index_var->GetMutable<paddle::framework::LoDTensor>());
+    tmp_index_tensor.Resize(out_inx_dim);
+    int *tmp_index_data =
+        tmp_index_tensor.mutable_data<int>(platform::CPUPlace());
+
+    auto out_dims = inx[0].dims();
+    size_t out_dim_sum = 0;
+    for (size_t index = 0; index < inx.size(); index++) {
+      auto inx_dims = inx[index].dims();
+      out_dim_sum += inx_dims[axis];
+      tmp_index_data[index] = inx_dims[axis];
+    }
+    out_inx.ShareDataWith(tmp_index_tensor);
+
+    // get input array items' dims
+    out_dims[axis] = out_dim_sum;
+    out.Resize(out_dims);
+
+    LodTensorArray2LodTensorVector(scope, base_name, Input("X"), &names);
+    // Invoke Reshape Op
+    auto concat_op = framework::OpRegistry::CreateOp(
+        "concat", {{"X", names}}, {{"Out", {Output("Out")}}}, attrs);
+
+    concat_op->Run(scope, place);
+  }
+};
+
+class LoDTensorArray2TensorOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X", "Input LoDTensorArray of tensor_array_to_tensor operator.");
+    AddOutput("Out", "Output tensor of tensor_array_to_tensor operator.");
+    AddOutput("OutIndex",
+              "Output input LoDTensorArray items' dims of "
+              "tensor_array_to_tensor operator.");
+    AddAttr<int>("axis",
+                 "The axis along which the input tensors will be concatenated.")
+        .SetDefault(0);
+    AddComment(R"DOC(
+tensor_array_to_tensor Operator.
+
+Concatenate the input LoDTensorArray along dimension axis to the output Tensor.
+Examples:
+  Input = {[1,2], [3,4], [5,6]}
+  axis = 0
+  Output = [[1,2],
+            [3,4],
+            [5,6]]
+  OutputIndex = [1,1,1]
+
+)DOC");
+  }
+};
+
+class LoDTensorArray2TensorOpInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *ctx) const override {}
+};
+
+class LoDTensorArray2TensorGradInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *context) const override {}
+};
+
+class LoDTensorArray2TensorGradInferVarType
+    : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc &op_desc,
+                  framework::BlockDesc *block) const override {
+    for (auto &out_var : op_desc.Output(framework::GradVarName("X"))) {
+      block->Var(out_var)->SetType(framework::proto::VarType::LOD_TENSOR_ARRAY);
+    }
+  }
+};
+
+class LoDTensorArray2TensorGradOp : public framework::OperatorBase {
+ public:
+  using OperatorBase::OperatorBase;
+
+ private:
+  void RunImpl(const framework::Scope &scope,
+               const platform::Place &place) const override {
+    auto axis = Attr<int>("axis");
+    framework::AttributeMap attrs;
+    attrs["axis"] = axis;
+
+    auto &inx = scope.FindVar(Input("X"))->Get<framework::LoDTensorArray>();
+    const size_t n = inx.size();
+    PADDLE_ENFORCE_GT(n, 0, "Input tensorarray size should > 0.");
+
+    std::string base_name = Inputs("X")[0];
+    std::vector<std::string> names;
+
+    LodTensorArray2LodTensorVector(scope, base_name, Input("X"), &names);
+
+    // grad
+    auto dx_name = Output(framework::GradVarName("X"));
+    auto dout_name = Input(framework::GradVarName("Out"));
+
+    std::vector<std::string> grad_names;
+
+    LodTensorVectorResizeFromLodTensorArray(scope, "grad_name", Input("X"),
+                                            &grad_names);
+
+    auto concat_grad_op = framework::OpRegistry::CreateOp(
+        "concat_grad", {{"X", names}, {"Out@GRAD", {dout_name}}},
+        {{"X@GRAD", grad_names}}, attrs);
+
+    concat_grad_op->Run(scope, place);
+
+    LodTensorArrayCreateFromLodTensorArray(scope, Input("X"), dx_name);
+    auto &grad_inx =
+        *scope.FindVar(dx_name)->GetMutable<framework::LoDTensorArray>();
+
+    for (size_t i = 0; i < grad_names.size(); i++) {
+      std::string var_name = grad_names[i];
+      auto &feed_input = scope.FindVar(var_name)->Get<framework::LoDTensor>();
+      grad_inx[i].ShareDataWith(feed_input);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+USE_OP(concat);
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(tensor_array_to_tensor, ops::LoDTensorArray2TensorOp,
+                  ops::LoDTensorArray2TensorOpMaker,
+                  ops::LoDTensorArray2TensorOpInferShape,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+REGISTER_OPERATOR(tensor_array_to_tensor_grad, ops::LoDTensorArray2TensorGradOp,
+                  ops::LoDTensorArray2TensorGradInferShape,
+                  ops::LoDTensorArray2TensorGradInferVarType);

paddle/scripts/paddle_build.sh

@@ -139,6 +139,7 @@ function cmake_gen() {
         -DWITH_AMD_GPU=${WITH_AMD_GPU:-OFF}
         -DWITH_DISTRIBUTE=${WITH_DISTRIBUTE:-OFF}
         -DWITH_MKL=${WITH_MKL:-ON}
+        -DWITH_NGRAPH=${WITH_NGRAPH:-OFF}
         -DWITH_AVX=${WITH_AVX:-OFF}
         -DWITH_GOLANG=${WITH_GOLANG:-OFF}
         -DCUDA_ARCH_NAME=${CUDA_ARCH_NAME:-All}
@@ -171,6 +172,7 @@ EOF
         -DWITH_AMD_GPU=${WITH_AMD_GPU:-OFF} \
         -DWITH_DISTRIBUTE=${WITH_DISTRIBUTE:-OFF} \
         -DWITH_MKL=${WITH_MKL:-ON} \
+        -DWITH_NGRAPH=${WITH_NGRAPH:-OFF} \
         -DWITH_AVX=${WITH_AVX:-OFF} \
         -DWITH_GOLANG=${WITH_GOLANG:-OFF} \
         -DCUDA_ARCH_NAME=${CUDA_ARCH_NAME:-All} \

python/paddle/fluid/__init__.py

@@ -34,6 +34,7 @@ from . import regularizer
 from . import average
 from . import metrics
 from . import transpiler
+from . import distribute_lookup_table
 from .param_attr import ParamAttr, WeightNormParamAttr
 from .data_feeder import DataFeeder
 from .core import LoDTensor, LoDTensorArray, CPUPlace, CUDAPlace, CUDAPinnedPlace, Scope

python/paddle/fluid/distribute_lookup_table.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+LOOKUP_TABLE_TYPE = "lookup_table"
+
+
+def find_distributed_lookup_table(program):
+    """
+    Find distribute lookup table in program.
+    We only support one distribute table now.
+    :param program:
+    :return: table_name or None
+    """
+    table_name = None
+
+    for op in program.global_block().ops:
+        if op.type == LOOKUP_TABLE_TYPE:
+            if op.attr('is_distributed') is True:
+                if table_name is None:
+                    table_name = op.input("W")[0]
+                if table_name != op.input("W")[0]:
+                    raise RuntimeError("all distributed lookup_table_ops"
+                                       " should have only one table")
+            else:
+                if table_name is not None:
+                    assert op.input("W")[0] != table_name
+
+    return table_name

python/paddle/fluid/layers/detection.py

@@ -31,6 +31,7 @@ from functools import reduce
 
 __all__ = [
     'prior_box',
+    'density_prior_box',
     'multi_box_head',
     'bipartite_match',
     'target_assign',
@@ -1023,6 +1024,135 @@ def prior_box(input,
     return box, var
 
 
+def density_prior_box(input,
+                      image,
+                      densities=None,
+                      fixed_sizes=None,
+                      fixed_ratios=None,
+                      variance=[0.1, 0.1, 0.2, 0.2],
+                      clip=False,
+                      steps=[0.0, 0.0],
+                      offset=0.5,
+                      name=None):
+    """
+    **Density Prior Box Operator**
+
+    Generate density prior boxes for SSD(Single Shot MultiBox Detector) 
+    algorithm. Each position of the input produce N prior boxes, N is 
+    determined by the count of densities, fixed_sizes and fixed_ratios. 
+    Boxes center at grid points around each input position is generated by 
+    this operator, and the grid points is determined by densities and 
+    the count of density prior box is determined by fixed_sizes and fixed_ratios. 
+    Obviously, the number of fixed_sizes is equal to the number of densities.
+    For densities_i in densities:
+    N_density_prior_box =sum(N_fixed_ratios * densities_i^2),
+
+    Args:
+       input(Variable): The Input Variables, the format is NCHW.
+       image(Variable): The input image data of PriorBoxOp,
+            the layout is NCHW.
+       densities(list|tuple|None): the densities of generated density prior 
+            boxes, this attribute should be a list or tuple of integers. 
+            Default: None.
+       fixed_sizes(list|tuple|None): the fixed sizes of generated density
+            prior boxes, this attribute should a list or tuple of same 
+            length with :attr:`densities`. Default: None.
+       fixed_ratios(list|tuple|None): the fixed ratios of generated density
+            prior boxes, if this attribute is not set and :attr:`densities`
+            and :attr:`fix_sizes` is set, :attr:`aspect_ratios` will be used
+            to generate density prior boxes.
+       variance(list|tuple): the variances to be encoded in density prior boxes.
+            Default:[0.1, 0.1, 0.2, 0.2].
+       clip(bool): Whether to clip out-of-boundary boxes. Default: False.
+       step(list|turple): Prior boxes step across width and height, If
+            step[0] == 0.0/step[1] == 0.0, the density prior boxes step across
+            height/weight of the input will be automatically calculated.
+            Default: [0., 0.]
+       offset(float): Prior boxes center offset. Default: 0.5
+       name(str): Name of the density prior box op. Default: None.
+
+    Returns:
+        tuple: A tuple with two Variable (boxes, variances)
+
+        boxes: the output density prior boxes of PriorBox.
+        The layout is [H, W, num_priors, 4].
+        H is the height of input, W is the width of input,
+        num_priors is the total
+        box count of each position of input.
+
+        variances: the expanded variances of PriorBox.
+        The layout is [H, W, num_priors, 4].
+        H is the height of input, W is the width of input
+        num_priors is the total
+        box count of each position of input
+
+
+    Examples:
+        .. code-block:: python
+
+            box, var = fluid.layers.density_prior_box(
+                input=conv1,
+                image=images,
+                min_sizes=[100.],
+                max_sizes=[200.],
+                aspect_ratios=[1.0, 1.0 / 2.0, 2.0],
+                densities=[3, 4],
+                fixed_sizes=[50., 60.],
+                fixed_ratios=[1.0, 3.0, 1.0 / 3.0],
+                flip=True,
+                clip=True)
+    """
+    helper = LayerHelper("density_prior_box", **locals())
+    dtype = helper.input_dtype()
+
+    def _is_list_or_tuple_(data):
+        return (isinstance(data, list) or isinstance(data, tuple))
+
+    if not _is_list_or_tuple_(densities):
+        raise TypeError('densities should be a list or a tuple or None.')
+    if not _is_list_or_tuple_(fixed_sizes):
+        raise TypeError('fixed_sizes should be a list or a tuple or None.')
+    if not _is_list_or_tuple_(fixed_ratios):
+        raise TypeError('fixed_ratios should be a list or a tuple or None.')
+    if len(densities) != len(fixed_sizes):
+        raise ValueError('densities and fixed_sizes length should be euqal.')
+    if not (_is_list_or_tuple_(steps) and len(steps) == 2):
+        raise ValueError('steps should be a list or tuple ',
+                         'with length 2, (step_width, step_height).')
+
+    densities = list(map(int, densities))
+    fixed_sizes = list(map(float, fixed_sizes))
+    fixed_ratios = list(map(float, fixed_ratios))
+    steps = list(map(float, steps))
+
+    attrs = {
+        'variances': variance,
+        'clip': clip,
+        'step_w': steps[0],
+        'step_h': steps[1],
+        'offset': offset,
+    }
+    if densities is not None and len(densities) > 0:
+        attrs['densities'] = densities
+    if fixed_sizes is not None and len(fixed_sizes) > 0:
+        attrs['fixed_sizes'] = fixed_sizes
+    if fixed_ratios is not None and len(fixed_ratios) > 0:
+        attrs['fixed_ratios'] = fixed_ratios
+
+    box = helper.create_variable_for_type_inference(dtype)
+    var = helper.create_variable_for_type_inference(dtype)
+    helper.append_op(
+        type="density_prior_box",
+        inputs={"Input": input,
+                "Image": image},
+        outputs={"Boxes": box,
+                 "Variances": var},
+        attrs=attrs, )
+    box.stop_gradient = True
+    var.stop_gradient = True
+    return box, var
+
+
 def multi_box_head(inputs,
                    image,
                    base_size,

python/paddle/fluid/layers/nn.py

@@ -160,10 +160,12 @@ __all__ = [
     'affine_grid',
     'sequence_reverse',
     'affine_channel',
+    'similarity_focus',
     'hash',
     'grid_sampler',
     'log_loss',
     'add_position_encoding',
+    'bilinear_tensor_product',
 ]
 
 
@@ -4065,8 +4067,8 @@ def edit_distance(input, label, normalized=True, ignored_tokens=None):
     Examples:
         .. code-block:: python
 
-            x = fluid.layers.data(name='x', shape=[8], dtype='float32')
-            y = fluid.layers.data(name='y', shape=[7], dtype='float32')
+            x = fluid.layers.data(name='x', shape=[1], dtype='float32')
+            y = fluid.layers.data(name='y', shape=[1], dtype='float32')
             cost = fluid.layers.edit_distance(input=x,label=y)
     """
     helper = LayerHelper("edit_distance", **locals())
@@ -4740,7 +4742,8 @@ def softmax_with_cross_entropy(logits,
                                label,
                                soft_label=False,
                                ignore_index=-100,
-                               numeric_stable_mode=False):
+                               numeric_stable_mode=False,
+                               return_softmax=False):
     """
     **Softmax With Cross Entropy Operator.**
 
@@ -4804,9 +4807,15 @@ def softmax_with_cross_entropy(logits,
                                     the algorithm is always numerically stable. 
                                     Note that the speed may be slower when use 
                                     stable algorithm. Default: False
+        return_softmax (bool): A flag indicating whether to return the softmax 
+                               along with the cross entropy loss. Default: False
 
     Returns:
-        Variable: The cross entropy loss is a 2-D tensor with shape [N x 1].
+        Variable or Tuple of two Variables: Return the cross entropy loss if 
+                              `return_softmax` is False, otherwise the tuple 
+                              (loss, softmax), where the cross entropy loss is 
+                              a 2-D tensor with shape [N x 1], and softmax is a 
+                              2-D tensor with shape [N x K].
 
     Examples:
         .. code-block:: python
@@ -4831,6 +4840,10 @@ def softmax_with_cross_entropy(logits,
             'ignore_index': ignore_index,
             'numeric_stable_mode': numeric_stable_mode
         })
+
+    if return_softmax:
+        return loss, softmax
+
     return loss
 
 
@@ -7933,6 +7946,118 @@ def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None):
     return out
 
 
+def similarity_focus(input, axis, indexes, name=None):
+    """  
+    SimilarityFocus Operator
+
+    Generate a similarity focus mask with the same shape of input using the following method:
+    1. Extract the 3-D tensor(here the first dimension is BatchSize) corresponding 
+       to the axis according to the indexes. For example, if axis=1 and indexes=[a], 
+       it will get the matrix T=X[:, a, :, :]. In this case, if the shape of input X 
+       is (BatchSize, A, B, C), the shape of tensor T is (BatchSize, B, C).
+    2. For each index, find the largest numbers in the tensor T, so that the same 
+       row and same column has at most one number(what it means is that if the 
+       largest number has been found in the i-th row and the j-th column, then 
+       the numbers in the i-th row or j-th column will be skipped. And then the 
+       next largest number will be selected from the remaining numbers. Obviously 
+       there will be min(B, C) numbers), and mark the corresponding position of the 
+       3-D similarity focus mask as 1, otherwise as 0. Do elementwise-or for 
+       each index.
+    3. Broadcast the 3-D similarity focus mask to the same shape of input X.
+
+    Refer to `Similarity Focus Layer <http://www.aclweb.org/anthology/N16-1108>`_
+
+    .. code-block:: text
+
+        * Example :
+
+            Given a 4-D tensor x with the shape (BatchSize, C, A, B), where C is
+            the number of channels and the shape of feature map is (A, B):
+                x.shape = (2, 3, 2, 2)
+                x.data = [[[[0.8, 0.1],
+                            [0.4, 0.5]],
+
+                           [[0.9, 0.7],
+                            [0.9, 0.9]],
+
+                           [[0.8, 0.9],
+                            [0.1, 0.2]]],
+
+
+                          [[[0.2, 0.5],
+                            [0.3, 0.4]],
+
+                           [[0.9, 0.7],
+                            [0.8, 0.4]],
+
+                           [[0.0, 0.2],
+                            [0.4, 0.7]]]]
+
+            Given axis: 1 (the axis of the channel)
+            Given indexes: [0]
+
+            then we get a 4-D tensor out with the same shape of input x:
+                out.shape = (2, 3, 2, 2)
+                out.data = [[[[1.0, 0.0],
+                              [0.0, 1.0]],
+
+                             [[1.0, 0.0],
+                              [0.0, 1.0]],
+
+                             [[1.0, 0.0],
+                              [0.0, 1.0]]],
+
+                            [[[0.0, 1.0],
+                              [1.0, 0.0]],
+
+                             [[0.0, 1.0],
+                              [1.0, 0.0]],
+
+                             [[0.0, 1.0],
+                              [1.0, 0.0]]]]
+
+    Args:
+        input(Variable): The input tensor variable(default float). It should 
+            be a 4-D tensor with shape [BatchSize, A, B, C].
+        axis(int): Indicating the dimension to be selected. It can only be
+            1, 2 or 3.
+        indexes(list): Indicating the indexes of the selected dimension.
+
+    Returns:
+        Variable: A tensor variable with the same shape and same type 
+            as the input.
+        
+    Examples:
+        .. code-block:: python
+            data = fluid.layers.data(
+              name='data', shape=[2, 3, 2, 2], dtype='float32')
+            x = fluid.layers.layer_norm(input=data, axis=1, indexes=[0])
+    """
+    helper = LayerHelper('similarity_focus', **locals())
+    # check attrs
+    if isinstance(axis, int) is False:
+        raise TypeError("axis must be int type.")
+    if isinstance(indexes, list) is False:
+        raise TypeError("indexes must be list type.")
+    if axis != 1 and axis != 2 and axis != 3:
+        raise ValueError("axis must be 1, 2 or 3.")
+    if len(indexes) == 0:
+        raise ValueError("indexes can not be empty.")
+
+    if name is None:
+        out = helper.create_variable_for_type_inference(dtype=input.dtype)
+    else:
+        out = helper.create_variable(
+            name=name, dtype=input.dtype, persistable=False)
+    helper.append_op(
+        type='similarity_focus',
+        inputs={'X': input},
+        outputs={'Out': out},
+        attrs={"axis": axis,
+               "indexes": indexes})
+    return out
+
+
 def hash(input, hash_size, num_hash=1, name=None):
     """
     Hash the input to an integer whose value is less than the given hash size.
@@ -8176,3 +8301,72 @@ def add_position_encoding(input, alpha, beta, name=None):
         attrs={"alpha": alpha,
                "beta": beta})
     return out
+
+
+def bilinear_tensor_product(x,
+                            y,
+                            size,
+                            act=None,
+                            name=None,
+                            param_attr=None,
+                            bias_attr=None):
+    """
+    **Add Bilinear Tensor Product Layer**
+
+    This layer performs bilinear tensor product on two inputs.
+    For example:
+
+    .. math::
+       out{i} = x * W_{i} * {y^\mathrm{T}}, i=0,1,...,size-1
+
+    In this formula:
+      - :math:`x`: the first input contains M elements, shape is [batch_size, M].
+      - :math:`y`: the second input contains N elements, shape is [batch_size, N].
+      - :math:`W_{i}`: the i-th learned weight, shape is [M, N]
+      - :math:`out{i}`: the i-th element of out, shape is [batch_size, size].
+      - :math:`y^\mathrm{T}`: the transpose of :math:`y_{2}`.
+
+    Args:
+        x (Variable): 2-D input tensor with shape [batch_size, M]
+        y (Variable): 2-D input tensor with shape [batch_size, N]
+        size (int): The dimension of this layer.
+        act (str, default None): Activation to be applied to the output of this layer.
+        name (str, default None): The name of this layer.
+        param_attr (ParamAttr, default None): The parameter attribute for the learnable w.
+            parameters/weights of this layer.
+        bias_attr (ParamAttr, default None): The parameter attribute for the bias
+            of this layer. If it is set to False, no bias will be added to the output units.
+            If it is set to None, the bias is initialized zero. Default: None.
+
+    Returns:
+        Variable: A 2-D Tensor of shape [batch_size, size].
+
+    Examples:
+        .. code-block:: python
+
+          tensor = bilinear_tensor_product(x=layer1, y=layer2, size=1000)
+    """
+    helper = LayerHelper('bilinear_tensor_product', **locals())
+    dtype = helper.input_dtype('x')
+
+    param_shape = [size, x.shape[1], y.shape[1]]
+
+    w = helper.create_parameter(
+        attr=helper.param_attr, shape=param_shape, dtype=dtype, is_bias=False)
+
+    if name is None:
+        out = helper.create_variable_for_type_inference(dtype=dtype)
+    else:
+        out = helper.create_variable(name=name, dtype=dtype, persistable=False)
+
+    inputs = {"X": x, "Y": y, "Weight": w}
+    if helper.bias_attr:
+        bias_size = [1, size]
+        bias = helper.create_parameter(
+            attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
+        inputs["Bias"] = bias
+    helper.append_op(
+        type="bilinear_tensor_product", inputs=inputs, outputs={"Out": out})
+
+    # add activation
+    return helper.append_activation(out)

python/paddle/fluid/layers/tensor.py

@@ -24,10 +24,10 @@ from .layer_function_generator import templatedoc
 import numpy
 
 __all__ = [
-    'create_tensor', 'create_parameter', 'create_global_var', 'cast', 'concat',
-    'sums', 'assign', 'fill_constant_batch_size_like', 'fill_constant',
-    'argmin', 'argmax', 'argsort', 'ones', 'zeros', 'reverse', 'has_inf',
-    'has_nan', 'isfinite'
+    'create_tensor', 'create_parameter', 'create_global_var', 'cast',
+    'tensor_array_to_tensor', 'concat', 'sums', 'assign',
+    'fill_constant_batch_size_like', 'fill_constant', 'argmin', 'argmax',
+    'argsort', 'ones', 'zeros', 'reverse', 'has_inf', 'has_nan', 'isfinite'
 ]
 
 
@@ -193,6 +193,60 @@ def concat(input, axis=0, name=None):
     return out
 
 
+def tensor_array_to_tensor(input, axis=1, name=None):
+    """
+    This function concatenates the input LodTensorArray along the axis mentioned
+    and returns that as the output.
+
+    A simple example as below:
+    
+    .. code-block:: text
+    
+        Given:
+
+        input.data = {[[0.6, 0.1, 0.3],
+                       [0.5, 0.3, 0.2]],
+                      [[1.3],
+                       [1.8]],
+                      [[2.3, 2.1],
+                       [2.5, 2.4]]}
+        
+        axis = 1
+    
+        Then:
+
+        output.data = [[0.6, 0.1, 0.3, 1.3, 2.3, 2.1],
+                       [0.5, 0.3, 0.2, 1.8, 2.5, 2.4]]
+
+        output_index.data = [3, 1, 2]
+
+    Args:
+        input(list): Input LodTensorArray
+        axis(int): Integer axis along which the tensors will be concatenated
+        name(str|None): A name for this layer(optional). If set None, the layer
+                       will be named automatically.
+
+    Returns:
+        Variable: Output variable of the concatenation
+        Variable: The input LodTensorArray items' dims along the axis
+
+    Examples:
+        .. code-block:: python
+
+           output, output_index = fluid.layers.tensor_array_to_tensor(input=tensor_array)
+    """
+    helper = LayerHelper('tensor_array_to_tensor', **locals())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
+    out_index = helper.create_variable_for_type_inference(dtype="int32")
+    helper.append_op(
+        type='tensor_array_to_tensor',
+        inputs={'X': input},
+        outputs={'Out': [out],
+                 'OutIndex': [out_index]},
+        attrs={'axis': axis})
+    return out, out_index
+
+
 def sums(input, out=None):
     """
     This function performs the sum operation on the input and returns the

python/paddle/fluid/optimizer.py

@@ -13,21 +13,23 @@
 # limitations under the License.
 
 from __future__ import print_function
-import re
-import sys
+
 from collections import defaultdict
+from contextlib import contextmanager
+
 from paddle.fluid.framework import Program, Variable, name_scope, default_main_program
+from paddle.fluid.distribute_lookup_table import find_distributed_lookup_table
+
 from . import framework
 from . import layers
+from . import unique_name
 from .backward import append_backward
+from .clip import append_gradient_clip_ops, error_clip_callback
 from .framework import program_guard
-from . import unique_name
 from .initializer import Constant
 from .layer_helper import LayerHelper
-from .regularizer import append_regularization_ops
-from .clip import append_gradient_clip_ops, error_clip_callback
-from contextlib import contextmanager
 from .layers import ops
+from .regularizer import append_regularization_ops
 
 __all__ = [
     'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'Ftrl',
@@ -85,7 +87,7 @@ class Optimizer(object):
             name=unique_name.generate("learning_rate"),
             shape=[1],
             value=float(self._learning_rate),
-            dtype='float32' if self._dtype == None else self._dtype,
+            dtype='float32' if self._dtype is None else self._dtype,
             persistable=True)
 
     def _global_learning_rate(self, program=None):
@@ -245,6 +247,50 @@ class Optimizer(object):
             end = len(global_block.ops)
             return global_block._slice_ops(start, end)
 
+    def _process_distribute_lookuptable(self, param_grads, loss,
+                                        startup_program):
+        """
+        Because distribute lookup table only support SGD optimizer for now, not support
+        other optimizer and regularization, so we should find the table parameter out,
+        and avoid to add regularization and other op for it, and add sgd optimize op
+        for it independently.
+        :param param_grads(list((Var, Var))): list of (param, grad) pair.
+        :param loss: the loss variable.
+        :param startup_program: the startup program
+        """
+        program = loss.block.program
+        table_name = find_distributed_lookup_table(program)
+        table_param = None
+        table_grad = None
+        new_param_grads = []
+        for p, g in param_grads:
+            if p.name == table_name:
+                if table_param is not None:
+                    raise RuntimeError(
+                        "multi dist table var found, only support one now!")
+                table_param = p
+                table_grad = g
+            else:
+                new_param_grads.append((p, g))
+        sgd_op = None
+        if table_param is not None:
+            with program_guard(program, startup_program):
+                param_and_grad = [table_param, table_grad]
+                with table_param.block.program._optimized_guard(param_and_grad), \
+                     framework.name_scope("optimizer"):
+                    self._create_global_learning_rate()
+                    # create the optimize op
+                    sgd_op = loss.block.append_op(
+                        type='sgd',
+                        inputs={
+                            "Param": table_param,
+                            "Grad": table_grad,
+                            "LearningRate":
+                            self._create_param_lr(param_and_grad)
+                        },
+                        outputs={"ParamOut": param_and_grad[0]})
+        return new_param_grads, (table_param, table_grad), sgd_op
+
     def minimize(self,
                  loss,
                  startup_program=None,
@@ -260,6 +306,9 @@ class Optimizer(object):
 
         params_grads = sorted(params_grads, key=lambda x: x[0].name)
 
+        params_grads, table_param_and_grad, table_optimize_op = \
+            self._process_distribute_lookuptable(params_grads, loss, startup_program)
+
         params_grads = append_gradient_clip_ops(params_grads)
 
         # Add regularization if any
@@ -268,6 +317,9 @@ class Optimizer(object):
 
         optimize_ops = self._create_optimization_pass(params_grads, loss,
                                                       startup_program)
+        if table_optimize_op is not None:
+            optimize_ops.append(table_optimize_op)
+            params_grads.append(table_param_and_grad)
         return optimize_ops, params_grads
 
 

python/paddle/fluid/tests/book/test_label_semantic_roles.py

@@ -38,7 +38,7 @@ depth = 8
 mix_hidden_lr = 1e-3
 
 IS_SPARSE = True
-PASS_NUM = 10
+PASS_NUM = 1
 BATCH_SIZE = 10
 
 embedding_name = 'emb'

python/paddle/fluid/tests/test_detection.py

@@ -128,6 +128,24 @@ class TestPriorBox(unittest.TestCase):
         assert box.shape[3] == 4
 
 
+class TestDensityPriorBox(unittest.TestCase):
+    def test_density_prior_box(self):
+        data_shape = [3, 224, 224]
+        images = fluid.layers.data(
+            name='pixel', shape=data_shape, dtype='float32')
+        conv1 = fluid.layers.conv2d(images, 3, 3, 2)
+        box, var = layers.density_prior_box(
+            input=conv1,
+            image=images,
+            densities=[3, 4],
+            fixed_sizes=[50., 60.],
+            fixed_ratios=[1.0],
+            clip=True)
+        assert len(box.shape) == 4
+        assert box.shape == var.shape
+        assert box.shape[3] == 4
+
+
 class TestAnchorGenerator(unittest.TestCase):
     def test_anchor_generator(self):
         data_shape = [3, 224, 224]

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

+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+import sys
+import math
+from op_test import OpTest
+
+
+class TestDensityPriorBoxOp(OpTest):
+    def set_data(self):
+        self.init_test_params()
+        self.init_test_input()
+        self.init_test_output()
+        self.inputs = {'Input': self.input, 'Image': self.image}
+
+        self.attrs = {
+            'variances': self.variances,
+            'clip': self.clip,
+            'step_w': self.step_w,
+            'step_h': self.step_h,
+            'offset': self.offset,
+            'densities': self.densities,
+            'fixed_sizes': self.fixed_sizes,
+            'fixed_ratios': self.fixed_ratios
+        }
+        self.outputs = {'Boxes': self.out_boxes, 'Variances': self.out_var}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def setUp(self):
+        self.op_type = "density_prior_box"
+        self.set_data()
+
+    def set_density(self):
+        self.densities = []
+        self.fixed_sizes = []
+        self.fixed_ratios = []
+
+    def init_test_params(self):
+        self.layer_w = 32
+        self.layer_h = 32
+
+        self.image_w = 40
+        self.image_h = 40
+
+        self.step_w = float(self.image_w) / float(self.layer_w)
+        self.step_h = float(self.image_h) / float(self.layer_h)
+
+        self.input_channels = 2
+        self.image_channels = 3
+        self.batch_size = 10
+
+        self.variances = [0.1, 0.1, 0.2, 0.2]
+        self.variances = np.array(self.variances, dtype=np.float).flatten()
+
+        self.set_density()
+
+        self.clip = True
+        self.num_priors = 0
+        if len(self.fixed_sizes) > 0 and len(self.densities) > 0:
+            for density in self.densities:
+                if len(self.fixed_ratios) > 0:
+                    self.num_priors += len(self.fixed_ratios) * (pow(density,
+                                                                     2))
+        self.offset = 0.5
+
+    def init_test_input(self):
+        self.image = np.random.random(
+            (self.batch_size, self.image_channels, self.image_w,
+             self.image_h)).astype('float32')
+
+        self.input = np.random.random(
+            (self.batch_size, self.input_channels, self.layer_w,
+             self.layer_h)).astype('float32')
+
+    def init_test_output(self):
+        out_dim = (self.layer_h, self.layer_w, self.num_priors, 4)
+        out_boxes = np.zeros(out_dim).astype('float32')
+        out_var = np.zeros(out_dim).astype('float32')
+
+        step_average = int((self.step_w + self.step_h) * 0.5)
+        for h in range(self.layer_h):
+            for w in range(self.layer_w):
+                idx = 0
+                c_x = (w + self.offset) * self.step_w
+                c_y = (h + self.offset) * self.step_h
+                # Generate density prior boxes with fixed size
+                for density, fixed_size in zip(self.densities,
+                                               self.fixed_sizes):
+                    if (len(self.fixed_ratios) > 0):
+                        for ar in self.fixed_ratios:
+                            shift = int(step_average / density)
+                            box_width_ratio = fixed_size * math.sqrt(ar)
+                            box_height_ratio = fixed_size / math.sqrt(ar)
+                            for di in range(density):
+                                for dj in range(density):
+                                    c_x_temp = c_x - step_average / 2.0 + shift / 2.0 + dj * shift
+                                    c_y_temp = c_y - step_average / 2.0 + shift / 2.0 + di * shift
+                                    out_boxes[h, w, idx, :] = [
+                                        max((c_x_temp - box_width_ratio / 2.0) /
+                                            self.image_w, 0),
+                                        max((c_y_temp - box_height_ratio / 2.0)
+                                            / self.image_h, 0),
+                                        min((c_x_temp + box_width_ratio / 2.0) /
+                                            self.image_w, 1),
+                                        min((c_y_temp + box_height_ratio / 2.0)
+                                            / self.image_h, 1)
+                                    ]
+                                    idx += 1
+        if self.clip:
+            out_boxes = np.clip(out_boxes, 0.0, 1.0)
+        out_var = np.tile(self.variances,
+                          (self.layer_h, self.layer_w, self.num_priors, 1))
+        self.out_boxes = out_boxes.astype('float32')
+        self.out_var = out_var.astype('float32')
+
+
+class TestDensityPriorBox(TestDensityPriorBoxOp):
+    def set_density(self):
+        self.densities = [3, 4]
+        self.fixed_sizes = [1.0, 2.0]
+        self.fixed_ratios = [1.0]
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -567,7 +567,6 @@ class TestDistLookupTable(TestDistLookupTableBase):
             'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
             'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
             'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
-            'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
             'fill_constant', 'fill_constant', 'uniform_random',
             'uniform_random', 'recv', 'recv', 'recv', 'fetch_barrier', 'concat',
             'fake_init'
@@ -639,7 +638,7 @@ class TestAsyncDistLookupTable(TestDistLookupTableBase):
         # 5 save table
         self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"])
 
-        trainer, _ = self.get_trainer(config)
+        trainer, trainer_startup = self.get_trainer(config)
         self.assertEqual(len(trainer.blocks), 1)
         ops = [
             'split_ids', 'prefetch', 'merge_ids', 'sequence_pool',
@@ -653,6 +652,16 @@ class TestAsyncDistLookupTable(TestDistLookupTableBase):
             'recv', 'concat'
         ]
         self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
+        startup_ops = [
+            'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
+            'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
+            'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
+            'fill_constant', 'fill_constant', 'uniform_random',
+            'uniform_random', 'recv', 'recv', 'recv', 'fetch_barrier', 'concat',
+            'fake_init'
+        ]
+        self.assertEqual([op.type for op in trainer_startup.blocks[0].ops],
+                         startup_ops)
 
 
 class TestDistLookupTableSliceSize(TestDistLookupTableBase):

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

@@ -369,6 +369,10 @@ class TestBook(unittest.TestCase):
         with program_guard(program):
             x = layers.data(name='x', shape=[16], dtype='float32')
             y = layers.data(name='label', shape=[1], dtype='int64')
+            loss, softmax = layers.softmax_with_cross_entropy(
+                x, y, return_softmax=True)
+            self.assertIsNotNone(loss)
+            self.assertIsNotNone(softmax)
             loss = layers.softmax_with_cross_entropy(x, y)
             self.assertIsNotNone(loss)
         print(str(program))
@@ -911,6 +915,16 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(data_1)
         print(str(program))
 
+    def test_bilinear_tensor_product_layer(self):
+        program = Program()
+        with program_guard(program):
+            data = layers.data(name='data', shape=[4], dtype="float32")
+
+            theta = layers.data(name="theta", shape=[5], dtype="float32")
+            out = layers.bilinear_tensor_product(data, theta, 6)
+
+        print(str(program))
+
 
 if __name__ == '__main__':
     unittest.main()

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

@@ -21,8 +21,8 @@ import six
 class TestBase(unittest.TestCase):
     def main(self,
              network_func,
-             iter=100,
-             iter_per_pe=100,
+             iter=10,
+             iter_per_pe=10,
              use_gpu=True,
              use_experimental_executor=False):
         if use_gpu and not fluid.core.is_compiled_with_cuda():
@@ -45,7 +45,7 @@ class TestBase(unittest.TestCase):
             exe_strategy._dry_run = True
             exe_strategy.use_experimental_executor = use_experimental_executor
             pe = fluid.ParallelExecutor(
-                use_cuda=True,
+                use_cuda=use_gpu,
                 loss_name=loss.name,
                 main_program=main_prog,
                 exec_strategy=exe_strategy)

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

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+import paddle.fluid.core as core
+from op_test import OpTest
+
+
+class TestSimilarityFocusOp(OpTest):
+    def setUp(self):
+        self.op_type = "similarity_focus"
+        batch_size = 2
+        x_dim, y_dim, z_dim = 3, 2, 2
+        self.inputs = {
+            'X': np.array([[[[0.8, 0.1], [0.4, 0.5]], [[0.9, 0.7], [0.9, 0.9]],
+                            [[0.8, 0.9], [0.1, 0.2]]],
+                           [[[0.2, 0.5], [0.3, 0.4]], [[0.9, 0.7], [0.8, 0.4]],
+                            [[0.0, 0.2], [0.4, 0.7]]]]),
+        }
+        self.attrs = {
+            'axis': 1,
+            'indexes': [0],
+        }
+
+        output = None
+        for batch in range(batch_size):
+            res = np.zeros((1, y_dim, z_dim)).astype("float32").reshape(-1)
+            for index in self.attrs['indexes']:
+                channel = self.inputs['X'][batch, index, :, :].reshape(-1).copy(
+                )
+                tag1 = [0 for i in range(y_dim)]
+                tag2 = [0 for i in range(z_dim)]
+                cnt = 0
+                for i in range(channel.size):
+                    index = channel.argmax()
+                    idx1 = index // z_dim
+                    idx2 = index % z_dim
+                    if tag1[idx1] + tag2[idx2] == 0:
+                        tag1[idx1] = 1
+                        tag2[idx2] = 1
+                        res[index] = 1
+                        cnt += 1
+                        if cnt == min(y_dim, z_dim):
+                            break
+                    channel[index] = -1
+            res = res.reshape(1, y_dim, z_dim).repeat([x_dim], axis=0)
+            res = res.reshape(1, x_dim, y_dim, z_dim)
+            if output is not None:
+                output = np.concatenate((output, res), axis=0)
+            else:
+                output = res
+        self.outputs = {'Out': output}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestSimilarityFocusOp_axis1(OpTest):
+    def setUp(self):
+        self.op_type = "similarity_focus"
+        batch_size = 3
+        x_dim, y_dim, z_dim = 4, 5, 6
+        self.inputs = {
+            'X': np.random.random(
+                (batch_size, x_dim, y_dim, z_dim)).astype("float32"),
+        }
+        self.attrs = {
+            'axis': 1,
+            'indexes': [0, 3],
+        }
+
+        output = None
+        for batch in range(batch_size):
+            res = np.zeros((1, y_dim, z_dim)).astype("float32").reshape(-1)
+            for index in self.attrs['indexes']:
+                channel = self.inputs['X'][batch, index, :, :].reshape(-1).copy(
+                )
+                tag1 = [0 for i in range(y_dim)]
+                tag2 = [0 for i in range(z_dim)]
+                cnt = 0
+                for i in range(channel.size):
+                    index = channel.argmax()
+                    idx1 = index // z_dim
+                    idx2 = index % z_dim
+                    if tag1[idx1] + tag2[idx2] == 0:
+                        tag1[idx1] = 1
+                        tag2[idx2] = 1
+                        res[index] = 1
+                        cnt += 1
+                        if cnt == min(y_dim, z_dim):
+                            break
+                    channel[index] = -1
+            res = res.reshape(1, y_dim, z_dim)
+            res = res.repeat([x_dim], axis=0)
+            res = res.reshape(1, x_dim, y_dim, z_dim)
+            if output is not None:
+                output = np.concatenate((output, res), axis=0)
+            else:
+                output = res
+        self.outputs = {'Out': output}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestSimilarityFocusOp_axis2(OpTest):
+    def setUp(self):
+        self.op_type = "similarity_focus"
+        batch_size = 6
+        x_dim, y_dim, z_dim = 7, 8, 9
+        self.inputs = {
+            'X': np.random.random(
+                (batch_size, x_dim, y_dim, z_dim)).astype("float32"),
+        }
+        self.attrs = {
+            'axis': 2,
+            'indexes': [0, 3, 5],
+        }
+
+        output = None
+        for batch in range(batch_size):
+            res = np.zeros((x_dim, 1, z_dim)).astype("float32").reshape(-1)
+            for index in self.attrs['indexes']:
+                channel = self.inputs['X'][batch, :, index, :].reshape(-1).copy(
+                )
+                tag1 = [0 for i in range(x_dim)]
+                tag2 = [0 for i in range(z_dim)]
+                cnt = 0
+                for i in range(channel.size):
+                    index = channel.argmax()
+                    idx1 = index // z_dim
+                    idx2 = index % z_dim
+                    if tag1[idx1] + tag2[idx2] == 0:
+                        tag1[idx1] = 1
+                        tag2[idx2] = 1
+                        res[index] = 1
+                        cnt += 1
+                        if cnt == min(x_dim, z_dim):
+                            break
+                    channel[index] = -1
+            res = res.reshape(x_dim, 1, z_dim)
+            res = res.repeat([y_dim], axis=1)
+            res = res.reshape(1, x_dim, y_dim, z_dim)
+            if output is not None:
+                output = np.concatenate((output, res), axis=0)
+            else:
+                output = res
+        self.outputs = {'Out': output}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestSimilarityFocusOp_axis3(OpTest):
+    def setUp(self):
+        self.op_type = "similarity_focus"
+        batch_size = 64
+        x_dim, y_dim, z_dim = 48, 48, 13
+        self.inputs = {
+            'X': np.random.random(
+                (batch_size, x_dim, y_dim, z_dim)).astype("float32"),
+        }
+        self.attrs = {
+            'axis': 3,
+            'indexes': [0, 2, 7, 9],
+        }
+
+        output = None
+        for batch in range(batch_size):
+            res = np.zeros((x_dim, y_dim, 1)).astype("float32").reshape(-1)
+            for index in self.attrs['indexes']:
+                channel = self.inputs['X'][batch, :, :, index].reshape(-1).copy(
+                )
+                tag1 = [0 for i in range(x_dim)]
+                tag2 = [0 for i in range(y_dim)]
+                cnt = 0
+                for i in range(channel.size):
+                    index = channel.argmax()
+                    idx1 = index // y_dim
+                    idx2 = index % y_dim
+                    if tag1[idx1] + tag2[idx2] == 0:
+                        tag1[idx1] = 1
+                        tag2[idx2] = 1
+                        res[index] = 1
+                        cnt += 1
+                        if cnt == min(x_dim, y_dim):
+                            break
+                    channel[index] = -1
+            res = res.reshape(x_dim, y_dim, 1)
+            res = res.repeat([z_dim], axis=2)
+            res = res.reshape(1, x_dim, y_dim, z_dim)
+            if output is not None:
+                output = np.concatenate((output, res), axis=0)
+            else:
+                output = res
+        self.outputs = {'Out': output}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == "__main__":
+    unittest.main()

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

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import numpy
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+from paddle.fluid.op import Operator
+from paddle.fluid.executor import Executor
+
+
+class TestLoDTensorArrayConcat(unittest.TestCase):
+    def setUp(self):
+        self.op_type = "tensor_array_to_tensor"
+        self.attrs = {"axis": 0}
+        self.outputs = ["Out"]
+
+    def test_get_set(self):
+        scope = core.Scope()
+        program = fluid.Program()
+        block = program.global_block()
+
+        input_arr = block.create_var(
+            name="tmp_lod_tensor_array",
+            type=core.VarDesc.VarType.LOD_TENSOR_ARRAY)
+        input_arr.persistable = True
+        input_arr_var = scope.var('tmp_lod_tensor_array')
+        input_tensor_array = input_arr_var.get_lod_tensor_array()
+        self.assertEqual(0, len(input_tensor_array))
+
+        cpu = core.CPUPlace()
+        for i in range(10):
+            t = core.LoDTensor()
+            if i == 0:
+                t.set(numpy.array([[i], [i]], dtype='float32'), cpu)
+            else:
+                t.set(numpy.array([[i]], dtype='float32'), cpu)
+            input_tensor_array.append(t)
+
+        self.assertEqual(10, len(input_tensor_array))
+
+        random_grad = numpy.random.random_sample([11]).astype(numpy.float32)
+
+        y_out = block.create_var(name="Out")
+        y_out.persistable = True
+        y_out_index = block.create_var(name="OutIndex")
+        y_out_index.persistable = True
+
+        y_grad_arr = block.create_var(
+            name='Out@GRAD', dtype='float32', shape=[11])
+        y_grad_arr.persistable = True
+        y_grad = scope.var('Out@GRAD')
+        y_grad_tensor = y_grad.get_tensor()
+        y_grad_tensor.set(random_grad, cpu)
+
+        op = block.append_op(
+            type=self.op_type,
+            inputs={"X": input_arr},
+            outputs={"Out": y_out,
+                     "OutIndex": y_out_index},
+            attrs=self.attrs)
+
+        out_grad = block.create_var(
+            name="tmp_lod_tensor_array@GRAD",
+            type=core.VarDesc.VarType.LOD_TENSOR_ARRAY)
+        out_grad.persistable = True
+
+        grad_op_desc_list, op_grad_to_var = core.get_grad_op_desc(op.desc,
+                                                                  set(), [])
+        grad_op_desc = grad_op_desc_list[0]
+        new_op_desc = block.desc.append_op()
+        new_op_desc.copy_from(grad_op_desc)
+        for var_name in grad_op_desc.output_arg_names():
+            block.desc.var(var_name.encode("ascii"))
+
+        grad_op_desc.infer_var_type(block.desc)
+        grad_op_desc.infer_shape(block.desc)
+        for arg in grad_op_desc.output_arg_names():
+            grad_var = block.desc.find_var(arg.encode("ascii"))
+            grad_var.set_dtype(core.VarDesc.VarType.FP32)
+
+        fetch_list = []
+        fetch_list.append(block.var('Out'))
+        fetch_list.append(block.var('OutIndex'))
+
+        exe = fluid.Executor(fluid.CPUPlace())
+        out = exe.run(program, fetch_list=fetch_list, scope=scope)
+        #print ("index: ", numpy.array(out[1]))  
+
+        # test forward
+        tensor_res = numpy.array(out[0])
+        tensor_res_out_idx = numpy.array(out[1])
+        tensor_gt = numpy.array(
+            [0] + [0, 1, 2, 3, 4, 5, 6, 7, 8, 9], dtype='float32')
+
+        self.assertEqual(len(tensor_res), len(tensor_gt))
+        self.assertEqual(len(tensor_res_out_idx), 10)
+
+        for i in range(len(tensor_res)):
+            self.assertEqual(tensor_res[i], tensor_gt[i])
+
+        for i in range(len(tensor_res_out_idx)):
+            if i == 0:
+                self.assertEqual(tensor_res_out_idx[i], 2)
+            else:
+                self.assertEqual(tensor_res_out_idx[i], 1)
+
+        # test backward
+        grad_tensor = scope.var('tmp_lod_tensor_array@GRAD')
+        grad_tensor_array = grad_tensor.get_lod_tensor_array()
+
+        self.assertEqual(10, len(grad_tensor_array))
+
+        for i in range(len(grad_tensor_array)):
+            if i == 0:
+                self.assertEqual(
+                    numpy.array(grad_tensor_array[i])[0],
+                    numpy.array(random_grad[i]))
+                self.assertEqual(
+                    numpy.array(grad_tensor_array[i])[1],
+                    numpy.array(random_grad[i + 1]))
+            if i == 1:
+                self.assertEqual(
+                    numpy.array(grad_tensor_array[i]),
+                    numpy.array(random_grad[i + 1]))
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -31,18 +31,17 @@ Steps to transpile pserver:
 """
 
 import math
-import sys
 import numpy as np
 import collections
-import six
 import logging
 
-from .ps_dispatcher import RoundRobin, HashName, PSDispatcher
+from .ps_dispatcher import RoundRobin, PSDispatcher
 from .. import core, framework, unique_name
 from ..framework import Program, default_main_program, \
     default_startup_program, Block, \
     Parameter, grad_var_name
 from .details import *
+from ..distribute_lookup_table import find_distributed_lookup_table
 from functools import reduce
 
 LOOKUP_TABLE_TYPE = "lookup_table"
@@ -292,7 +291,8 @@ class DistributeTranspiler(object):
         self.optimize_ops, self.params_grads = self._get_optimize_pass()
 
         ps_dispatcher = self.config.split_method(self.pserver_endpoints)
-        self.has_distributed_lookup_table = self._has_distributed_lookup_table()
+        self.table_name = find_distributed_lookup_table(self.origin_program)
+        self.has_distributed_lookup_table = self.table_name != None
         self.param_name_to_grad_name = dict()
         self.grad_name_to_param_name = dict()
         for param_var, grad_var in self.params_grads:
@@ -966,28 +966,6 @@ to transpile() call.")
 
     # ====================== private transpiler functions =====================
 
-    def _has_distributed_lookup_table(self):
-        # process lookup_table_op
-        # 1. check all lookup_table_op is distributed
-        # 2. check all lookup_table_op share the same table.
-        distributed_lookup_table_ops = []
-        # support only one distributed_lookup_table now
-        self.table_name = None
-        for op in self.origin_program.global_block().ops:
-            if op.type == LOOKUP_TABLE_TYPE:
-                if op.attr('is_distributed') is True:
-                    if self.table_name is None:
-                        self.table_name = op.input("W")[0]
-                    if self.table_name != op.input("W")[0]:
-                        raise RuntimeError("all distributed lookup_table_ops"
-                                           " should have only one table")
-                    distributed_lookup_table_ops.append(op)
-                else:
-                    if self.table_name is not None:
-                        assert op.input("W")[0] != self.table_name
-
-        return len(distributed_lookup_table_ops) > 0
-
     def _update_dist_lookup_table_vars(self, param_list, grad_list,
                                        params_grads):
         # TODO(wuyi): put find a way to put dist lookup table stuff all together.
@@ -1341,7 +1319,6 @@ to transpile() call.")
         """
         create a new block to handle save checkpoint.
         """
-        import os
 
         pserver_program.global_block().create_var(
             name="kLookupTablePath",

python/setup.py.in

@@ -174,6 +174,18 @@ if '${CMAKE_BUILD_TYPE}' == 'Release':
             raise Exception("patch libmkldnn.so failed, command: %s" % command)
         package_data['paddle.libs']+=['libmkldnn.so.0']
         shutil.copy('${MKLDNN_SHARED_LIB}', libs_path)
+if '${WITH_NGRAPH}' == 'ON':
+    if '${CMAKE_BUILD_TYPE}' == 'Release':
+        # only change rpath in Release mode.
+        command = "patchelf --set-rpath '$ORIGIN/' ${NGRAPH_SHARED_LIB}"
+        if os.system(command) != 0:
+            raise Exception("patch ${NGRAPH_SHARED_LIB_NAME} failed, command: %s" % command)
+    shutil.copy('${NGRAPH_SHARED_LIB}', libs_path)
+    shutil.copy('${NGRAPH_CPU_LIB}', libs_path)
+    shutil.copy('${NGRAPH_TBB_LIB}', libs_path)
+    package_data['paddle.libs']+=['${NGRAPH_SHARED_LIB_NAME}',
+                                  '${NGRAPH_CPU_LIB_NAME}',
+                                  '${NGRAPH_TBB_LIB_NAME}']
 # remove unused paddle/libs/__init__.py
 os.remove(libs_path+'/__init__.py')
 package_dir['paddle.libs']=libs_path