Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into optimize-cpp-reader

test=develop

cmake/external/gflags.cmake

@@ -63,6 +63,15 @@ ADD_DEPENDENCIES(gflags extern_gflags)
 
 LIST(APPEND external_project_dependencies gflags)
 
+# On Windows (including MinGW), the Shlwapi library is used by gflags if available.
+if (WIN32)
+  include(CheckIncludeFileCXX)
+  check_include_file_cxx("shlwapi.h" HAVE_SHLWAPI)
+  if (HAVE_SHLWAPI)
+    set_property(GLOBAL PROPERTY OS_DEPENDENCY_MODULES shlwapi.lib)
+  endif(HAVE_SHLWAPI)
+endif (WIN32)
+
 IF(WITH_C_API)
   INSTALL(DIRECTORY ${GFLAGS_INCLUDE_DIR} DESTINATION third_party/gflags)
   IF(ANDROID)

cmake/external/ngraph.cmake

@@ -37,14 +37,18 @@ INCLUDE(GNUInstallDirs)
 INCLUDE(ExternalProject)
 
 SET(NGRAPH_PROJECT         "extern_ngraph")
-SET(NGRAPH_GIT_TAG         "08851c2c45fcf9fa9c74871dd3dbc3fe38f37cc9")
+SET(NGRAPH_GIT_TAG         "20bd8bbc79ae3a81c57313846a2be7313e5d1dab")
 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_LIB_DIR         ${NGRAPH_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR})
 SET(NGRAPH_SHARED_LIB_NAME libngraph.so)
 SET(NGRAPH_CPU_LIB_NAME    libcpu_backend.so)
-SET(NGRAPH_TBB_LIB_NAME    libtbb.so.2)
+if(CMAKE_BUILD_TYPE STREQUAL "Debug")
+    SET(NGRAPH_TBB_LIB_NAME    libtbb_debug.so.2)
+else()
+    SET(NGRAPH_TBB_LIB_NAME    libtbb.so.2)
+endif()
 SET(NGRAPH_GIT_REPO        "https://github.com/NervanaSystems/ngraph.git")
 SET(NGRAPH_SHARED_LIB      ${NGRAPH_LIB_DIR}/${NGRAPH_SHARED_LIB_NAME})
 SET(NGRAPH_CPU_LIB         ${NGRAPH_LIB_DIR}/${NGRAPH_CPU_LIB_NAME})
@@ -66,16 +70,7 @@ ExternalProject_Add(
     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
-)
-
-# 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} ${MKLDNN_INSTALL_DIR}/lib/libiomp5.so
-    DEPENDEES download
-    DEPENDERS configure
+    CMAKE_ARGS          -DMKLML_LIB_DIR=${MKLML_INSTALL_DIR}/lib
 )
 
 add_dependencies(ngraph ${NGRAPH_PROJECT})

cmake/generic.cmake

@@ -359,6 +359,8 @@ function(cc_binary TARGET_NAME)
     add_dependencies(${TARGET_NAME} ${cc_binary_DEPS})
     common_link(${TARGET_NAME})
   endif()
+  get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
+  target_link_libraries(${TARGET_NAME} ${os_dependency_modules})
 endfunction(cc_binary)
 
 function(cc_test TARGET_NAME)
@@ -367,18 +369,15 @@ function(cc_test TARGET_NAME)
     set(oneValueArgs "")
     set(multiValueArgs SRCS DEPS ARGS)
     cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    add_executable(${TARGET_NAME} ${cc_test_SRCS})
     if(WIN32)
-      list(APPEND win32_deps shlwapi)
       if("${cc_test_DEPS};" MATCHES "python;")
         list(REMOVE_ITEM cc_test_DEPS python)
-        list(APPEND win32_deps ${PYTHON_LIBRARIES})
+        target_link_libraries(${TARGET_NAME} ${PYTHON_LIBRARIES})
       endif()
     endif(WIN32)
-    add_executable(${TARGET_NAME} ${cc_test_SRCS})
-    target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
-    if(WIN32)
-      target_link_libraries(${TARGET_NAME} ${win32_deps})
-    endif(WIN32)
+    get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
+    target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} ${os_dependency_modules} paddle_gtest_main lod_tensor memory gtest gflags glog)
     add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
     common_link(${TARGET_NAME})
     add_test(NAME ${TARGET_NAME}
@@ -451,7 +450,8 @@ function(nv_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS)
     cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
-    target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
+    get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
+    target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog ${os_dependency_modules})
     add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
     common_link(${TARGET_NAME})
     add_test(${TARGET_NAME} ${TARGET_NAME})
@@ -538,7 +538,8 @@ function(hip_test TARGET_NAME)
     endif()
     add_executable(${TARGET_NAME} ${_cmake_options} ${_generated_files} ${_sources})
     set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP)
-    target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
+    get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
+    target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags ${os_dependency_modules})
     add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
     common_link(${TARGET_NAME})
     add_test(${TARGET_NAME} ${TARGET_NAME})

paddle/fluid/API.spec

@@ -88,6 +88,7 @@ paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'poo
 paddle.fluid.layers.adaptive_pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None))
 paddle.fluid.layers.adaptive_pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None))
 paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False))
+paddle.fluid.layers.data_norm ArgSpec(args=['input', 'act', 'epsilon', 'param_attr', 'data_layout', 'in_place', 'use_mkldnn', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var'], varargs=None, keywords=None, defaults=(None, 1e-05, None, 'NCHW', False, False, None, None, None, False))
 paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
 paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
@@ -210,6 +211,7 @@ paddle.fluid.layers.get_tensor_from_selected_rows ArgSpec(args=['x', 'name'], va
 paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1))
 paddle.fluid.layers.py_func ArgSpec(args=['func', 'x', 'out', 'backward_func', 'skip_vars_in_backward_input'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.psroi_pool ArgSpec(args=['input', 'rois', 'output_channels', 'spatial_scale', 'pooled_height', 'pooled_width', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.teacher_student_sigmoid_loss ArgSpec(args=['input', 'label', 'soft_max_up_bound', 'soft_max_lower_bound'], varargs=None, keywords=None, defaults=(15.0, -15.0))
 paddle.fluid.layers.huber_loss ArgSpec(args=['input', 'label', 'delta'], varargs=None, keywords=None, defaults=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))
@@ -406,28 +408,50 @@ paddle.fluid.nets.glu ArgSpec(args=['input', 'dim'], varargs=None, keywords=None
 paddle.fluid.nets.scaled_dot_product_attention ArgSpec(args=['queries', 'keys', 'values', 'num_heads', 'dropout_rate'], varargs=None, keywords=None, defaults=(1, 0.0))
 paddle.fluid.nets.img_conv_group ArgSpec(args=['input', 'conv_num_filter', 'pool_size', 'conv_padding', 'conv_filter_size', 'conv_act', 'param_attr', 'conv_with_batchnorm', 'conv_batchnorm_drop_rate', 'pool_stride', 'pool_type', 'use_cudnn'], varargs=None, keywords=None, defaults=(1, 3, None, None, False, 0.0, 1, 'max', True))
 paddle.fluid.optimizer.SGDOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'regularization', 'name'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.optimizer.SGDOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.optimizer.SGDOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.SGDOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.MomentumOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'momentum', 'use_nesterov', 'regularization', 'name'], varargs=None, keywords=None, defaults=(False, None, None))
+paddle.fluid.optimizer.MomentumOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.optimizer.MomentumOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.MomentumOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.AdagradOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1e-06, None, None))
+paddle.fluid.optimizer.AdagradOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.optimizer.AdagradOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.AdagradOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.AdamOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name', 'lazy_mode'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None, False))
+paddle.fluid.optimizer.AdamOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.optimizer.AdamOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.AdamOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.AdamaxOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None))
+paddle.fluid.optimizer.AdamaxOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.optimizer.AdamaxOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.AdamaxOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.DecayedAdagradOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'decay', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.95, 1e-06, None, None))
+paddle.fluid.optimizer.DecayedAdagradOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.optimizer.DecayedAdagradOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.DecayedAdagradOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.FtrlOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'l1', 'l2', 'lr_power', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.0, 0.0, -0.5, None, None))
+paddle.fluid.optimizer.FtrlOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.optimizer.FtrlOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.FtrlOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.RMSPropOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'rho', 'epsilon', 'momentum', 'centered', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.95, 1e-06, 0.0, False, None, None))
+paddle.fluid.optimizer.RMSPropOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.optimizer.RMSPropOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.RMSPropOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.AdadeltaOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'rho', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1e-06, 0.95, None, None))
+paddle.fluid.optimizer.AdadeltaOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.optimizer.AdadeltaOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.AdadeltaOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.ModelAverage.__init__ ArgSpec(args=['self', 'average_window_rate', 'min_average_window', 'max_average_window', 'regularization', 'name'], varargs=None, keywords=None, defaults=(10000, 10000, None, None))
 paddle.fluid.optimizer.ModelAverage.apply ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
+paddle.fluid.optimizer.ModelAverage.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.optimizer.ModelAverage.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.ModelAverage.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.ModelAverage.restore ArgSpec(args=['self', 'executor'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.optimizer.LarsMomentumOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'momentum', 'lars_coeff', 'lars_weight_decay', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.0005, None, None))
+paddle.fluid.optimizer.LarsMomentumOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.optimizer.LarsMomentumOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.LarsMomentumOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.backward.append_backward ArgSpec(args=['loss', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.regularizer.L1DecayRegularizer.__init__ ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,))

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -226,7 +226,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
    * Only variables should be the leaves of graph.
    */
   AddOutputToLeafOps(&result);
-  result.Erase<GraphOps>(kGraphOps);
+  result.Erase(kGraphOps);
   return graph;
 }
 

paddle/fluid/framework/ir/CMakeLists.txt

@@ -42,11 +42,23 @@ pass_library(seq_concat_fc_fuse_pass inference)
 pass_library(multi_batch_merge_pass base)
 pass_library(conv_bn_fuse_pass inference)
 pass_library(seqconv_eltadd_relu_fuse_pass inference)
+pass_library(seqpool_concat_fuse_pass inference)
 pass_library(is_test_pass base)
 pass_library(conv_elementwise_add_act_fuse_pass inference)
 pass_library(conv_elementwise_add2_act_fuse_pass inference)
 pass_library(conv_elementwise_add_fuse_pass inference)
 pass_library(conv_affine_channel_fuse_pass inference)
+pass_library(transpose_flatten_concat_fuse_pass inference)
+
+# There may be many transpose-flatten structures in a model, and the output of
+# these structures will be used as inputs to the concat Op. This pattern will
+# be detected by our pass. The index here represents the number of structures in the
+# pattern. We use index 3 ~ 6, because these quantities of structures are
+# common in the models.
+foreach (index RANGE 3 6)
+   file(APPEND ${pass_file} "USE_PASS(transpose_flatten${index}_concat_fuse_pass);\n")
+endforeach()
+
 if(WITH_MKLDNN)
     pass_library(mkldnn_placement_pass base)
     pass_library(depthwise_conv_mkldnn_pass base)
@@ -68,6 +80,7 @@ cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph graph_helper op_r
 cc_test(graph_to_program_pass_test SRCS graph_to_program_pass_test.cc DEPS graph_to_program_pass)
 cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS graph_pattern_detector)
 cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
+cc_test(test_seqpool_concat_fuse_pass SRCS seqpool_concat_fuse_pass_tester.cc DEPS seqpool_concat_fuse_pass framework_proto)
 cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass)
 if (WITH_MKLDNN)
     cc_test(test_depthwise_conv_mkldnn_pass SRCS depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)

paddle/fluid/framework/ir/graph.h

@@ -109,7 +109,6 @@ class Graph {
     attr_dels_[attr_name] = []() {};
   }
 
-  template <typename AttrType>
   void Erase(const std::string &attr_name) {
     PADDLE_ENFORCE(attrs_.count(attr_name) != 0, "%s not set in the graph",
                    attr_name);

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -1306,6 +1306,69 @@ PDNode *patterns::ConvAffineChannel::operator()(
   return ac_out_var;
 }
 
+// a -> transpose_op(1) -> transpose_out_a -> flatten_op(1) -> flatten_out_a
+// b -> transpose_op(2) -> transpose_out_b -> flatten_op(2) -> flatten_out_b
+// ...
+// z -> transpose_op(n) -> transpose_out_z -> flatten_op(n) -> flatten_out_z
+// flatten_out_a -> concat_op  flatten_out_b -> concat_op ... flatten_out_z ->
+// concat_op
+PDNode *patterns::TransposeFlattenConcat::operator()(
+    std::vector<PDNode *> conv_in, int times) {
+  // The times represents the repeat times of the
+  // {trans, trans_out, flatten, flatten_out}
+  const int kNumFields = 4;
+  const int kTransOutOffset = 1;
+  const int kFlattenOffset = 2;
+  const int kFlattenOutOffset = 3;
+
+  std::vector<PDNode *> nodes;
+
+  for (int i = 0; i < times; i++) {
+    nodes.push_back(
+        pattern->NewNode(GetNodeName("transpose" + std::to_string(i)))
+            ->assert_is_op("transpose2"));
+    nodes.push_back(
+        pattern->NewNode(GetNodeName("transpose_out" + std::to_string(i)))
+            ->assert_is_op_output("transpose2")
+            ->assert_is_op_input("flatten2", "X")
+            ->AsIntermediate());
+    nodes.push_back(pattern->NewNode(GetNodeName("flatten" + std::to_string(i)))
+                        ->assert_is_op("flatten2"));
+
+    nodes.push_back(
+        pattern->NewNode(GetNodeName("flatten_out" + std::to_string(i)))
+            ->assert_is_op_output("flatten2")
+            ->assert_is_op_nth_input("concat", "X", i)
+            ->AsIntermediate());
+  }
+
+  auto concat_op = pattern->NewNode(GetNodeName("concat"))
+                       ->assert_is_op("concat")
+                       ->assert_op_has_n_inputs("concat", times);
+  auto concat_out = pattern->NewNode(GetNodeName("concat_out"))
+                        ->assert_is_op_output("concat")
+                        ->AsOutput();
+
+  std::vector<PDNode *> flatten_outs;
+  for (int i = 0; i < times; i++) {
+    conv_in[i]->AsInput();
+    // trans
+    nodes[i * kNumFields]->LinksFrom({conv_in[i]});
+    // trans_out
+    nodes[i * kNumFields + kTransOutOffset]->LinksFrom({nodes[i * kNumFields]});
+    // flatten
+    nodes[i * kNumFields + kFlattenOffset]->LinksFrom(
+        {nodes[i * kNumFields + kTransOutOffset]});
+    // flatten_out
+    nodes[i * kNumFields + kFlattenOutOffset]->LinksFrom(
+        {nodes[i * kNumFields + kFlattenOffset]});
+    flatten_outs.push_back(nodes[i * kNumFields + kFlattenOutOffset]);
+  }
+
+  concat_op->LinksFrom(flatten_outs).LinksTo({concat_out});
+  return concat_out;
+}
+
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -766,6 +766,21 @@ struct ConvAffineChannel : public PatternBase {
   PATTERN_DECL_NODE(ac_out);  // Out
 };
 
+struct TransposeFlattenConcat : public PatternBase {
+  TransposeFlattenConcat(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "transpose_flatten_concat") {}
+
+  PDNode* operator()(std::vector<PDNode*> conv_inputs, int times);
+
+  std::string GetNodeName(const std::string& op_type) {
+    return PDNodeName(name_scope_, repr_, id_, op_type);
+  }
+
+  PDNode* GetPDNode(const std::string& op_type) {
+    return pattern->RetrieveNode(GetNodeName(op_type));
+  }
+};
+
 }  // namespace patterns
 
 // Link two ir::Nodes from each other.

paddle/fluid/framework/ir/seqpool_concat_fuse_pass.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/framework/ir/seqpool_concat_fuse_pass.h"
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/lod_tensor.h"
+
+#define MAX_CONCAT_INPUTS 200
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+PDNode* BuildSeqPoolConcatPattern(PDPattern* pattern,
+                                  const std::string& name_scope,
+                                  int num_inputs) {
+  auto is_concat_op_with_inputs = [](Node* x, int num) -> bool {
+    return x && x->IsOp() && x->Op()->Type() == "concat" &&
+           x->Op()->Input("X").size() == static_cast<size_t>(num);
+  };
+
+  auto is_nth_input_var_of_concat = [=](Node* x, int idx) -> bool {
+    return x && x->IsVar() && VarLinksToOp(x, "concat") &&
+           x->outputs.size() == 1 && IsNthInput(x, x->outputs[0], "X", idx) &&
+           is_concat_op_with_inputs(x->outputs[0], num_inputs);
+  };
+
+  auto is_seqpool_op_with_pootype_of_nth_input_of_concat = [=](
+      Node* x, const std::string& type, int idx) -> bool {
+    bool this_is_seqpool_op =
+        x && x->IsOp() && x->Op()->Type() == "sequence_pool" &&
+        x->Op()->HasAttr("pooltype") &&
+        boost::get<std::string>(x->Op()->GetAttr("pooltype")) == type &&
+        x->outputs.size() == 2;  // seqpool should only have 2 outputs
+    bool satisfied_all = this_is_seqpool_op;
+    if (this_is_seqpool_op) {
+      // Only one output of seqpool_op is nth_input_var of concat,
+      // the other one should be unused empty var.
+      if (is_nth_input_var_of_concat(x->outputs[0], idx)) {
+        satisfied_all = satisfied_all && x->outputs[1]->IsVar() &&
+                        x->outputs[1]->outputs.empty();
+      } else {
+        satisfied_all =
+            satisfied_all && is_nth_input_var_of_concat(x->outputs[1], idx) &&
+            x->outputs[0]->IsVar() && x->outputs[0]->outputs.size() == 0;
+      }
+    }
+    return satisfied_all;
+  };
+
+  auto* concat_op = pattern->NewNode(
+      [=](Node* x) { return is_concat_op_with_inputs(x, num_inputs); },
+      name_scope + "/concat_op");
+  concat_op->assert_op_attr<int>("axis", 1);
+
+  auto* concat_out_var = pattern->NewNode(
+      [=](Node* x) {
+        return x && x->IsVar() && VarLinksFromOp(x, "concat") &&
+               x->inputs.size() == 1 &&
+               is_concat_op_with_inputs(x->inputs[0], num_inputs);
+      },
+      name_scope + "/concat_out_var");
+  concat_out_var->assert_is_only_output_of_op("concat");
+
+  std::vector<PDNode*> seqpool_ops_input_var(num_inputs);
+  std::vector<PDNode*> seqpool_ops_output_var(num_inputs);
+  std::vector<PDNode*> seqpool_ops_output_unused_var(num_inputs);
+  std::vector<PDNode*> seqpool_ops(num_inputs);
+
+  for (int i = 0; i < num_inputs; ++i) {
+    seqpool_ops_output_var[i] = pattern->NewNode(
+        [=](Node* x) {
+          return x && x->IsVar() && is_nth_input_var_of_concat(x, i) &&
+                 x->inputs.size() == 1 &&
+                 is_seqpool_op_with_pootype_of_nth_input_of_concat(x->inputs[0],
+                                                                   "SUM", i);
+        },
+        name_scope + "/sequence_pool_out_" + std::to_string(i));
+
+    seqpool_ops_output_unused_var[i] = pattern->NewNode(
+        [=](Node* x) {
+          return x && x->IsVar() && x->inputs.size() == 1 &&
+                 x->outputs.size() == 0 &&
+                 is_seqpool_op_with_pootype_of_nth_input_of_concat(x->inputs[0],
+                                                                   "SUM", i);
+        },
+        name_scope + "/sequence_pool_unused_out_" + std::to_string(i));
+
+    seqpool_ops[i] = pattern->NewNode(
+        [=](Node* x) {
+          return x && x->IsOp() &&
+                 is_seqpool_op_with_pootype_of_nth_input_of_concat(x, "SUM", i);
+        },
+        name_scope + "/sequence_pool_op_" + std::to_string(i));
+
+    seqpool_ops_input_var[i] = pattern->NewNode(
+        [=](Node* x) {
+          bool basic = x && x->IsVar() && x->outputs.size() >= 1;
+          bool next_is_fine = false;
+          for (auto* o : x->outputs) {
+            if (is_seqpool_op_with_pootype_of_nth_input_of_concat(o, "SUM",
+                                                                  i)) {
+              next_is_fine = true;
+              break;
+            }
+          }
+          return basic && next_is_fine;
+        },
+        name_scope + "/sequence_pool_in_" + std::to_string(i));
+
+    // Links
+    seqpool_ops[i]
+        ->LinksFrom({seqpool_ops_input_var[i]})
+        .LinksTo({seqpool_ops_output_var[i], seqpool_ops_output_unused_var[i]});
+  }
+  concat_op->LinksFrom(seqpool_ops_output_var).LinksTo({concat_out_var});
+  return concat_out_var;
+}
+
+int BuildFusion(Graph* graph, const std::string& name_scope, int num_inputs) {
+  GraphPatternDetector gpd;
+  auto* pattern = gpd.mutable_pattern();
+  BuildSeqPoolConcatPattern(pattern, name_scope, num_inputs);
+
+  auto retrieve_node = [](const std::string& name,
+                          const GraphPatternDetector::subgraph_t& subgraph,
+                          const PDPattern& pat) -> Node* {
+    PADDLE_ENFORCE(subgraph.count(pat.RetrieveNode(name)),
+                   "pattern has no Node called %s", name.c_str());
+    Node* p = subgraph.at(pat.RetrieveNode(name));
+    PADDLE_ENFORCE_NOT_NULL(p, "subgraph has no node %s", name.c_str());
+    return p;
+  };
+
+  int fusion_count{0};
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    VLOG(4) << "handle SeqPool Concat fuse";
+    std::vector<std::string> input_names(num_inputs);
+    std::vector<Node*> input_vars(num_inputs);
+    auto& fused_pattern = gpd.pattern();
+    for (int i = 0; i < num_inputs; ++i) {
+      input_vars[i] =
+          retrieve_node(name_scope + "/sequence_pool_in_" + std::to_string(i),
+                        subgraph, fused_pattern);
+      input_names[i] = input_vars[i]->Name();
+    }
+    auto* concat_op =
+        retrieve_node(name_scope + "/concat_op", subgraph, fused_pattern);
+    auto* concat_out_var =
+        retrieve_node(name_scope + "/concat_out_var", subgraph, fused_pattern);
+    auto* seqpool_op0 = retrieve_node(name_scope + "/sequence_pool_op_0",
+                                      subgraph, fused_pattern);
+
+    // Create New OpDesc
+    OpDesc op_desc;
+    op_desc.SetType("fusion_seqpool_concat");
+    op_desc.SetInput("X", input_names);
+    op_desc.SetAttr("pooltype", seqpool_op0->Op()->GetAttr("pooltype"));
+    op_desc.SetAttr("axis", concat_op->Op()->GetAttr("axis"));
+    op_desc.SetOutput("Out", {concat_out_var->Name()});
+    auto* op = graph->CreateOpNode(&op_desc);
+    for (size_t i = 0; i < input_vars.size(); ++i) {
+      IR_NODE_LINK_TO(input_vars[i], op);
+    }
+    IR_NODE_LINK_TO(op, concat_out_var);
+
+    std::unordered_set<const Node*> marked_nodes;
+    for (auto& item : subgraph) {
+      marked_nodes.insert(item.second);
+    }
+    for (size_t i = 0; i < input_vars.size(); ++i) {
+      marked_nodes.erase(input_vars[i]);
+    }
+    marked_nodes.erase(concat_out_var);
+    GraphSafeRemoveNodes(graph, marked_nodes);
+    ++fusion_count;
+  };
+
+  gpd(graph, handler);
+  return fusion_count;
+}
+
+std::unique_ptr<ir::Graph> SeqPoolConcatFusePass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  FusePassBase::Init(name_scope_, graph.get());
+  int fusion_count = 0;
+  for (int i = MAX_CONCAT_INPUTS; i > 0; --i) {
+    fusion_count +=
+        BuildFusion(graph.get(), name_scope_ + "/" + std::to_string(i), i);
+  }
+  AddStatis(fusion_count);
+
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(seqpool_concat_fuse_pass,
+              paddle::framework::ir::SeqPoolConcatFusePass);

paddle/fluid/framework/ir/seqpool_concat_fuse_pass.h

+/* 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. */
+
+#pragma once
+
+#include <string>
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+/**
+ * Fuse SequencePool(with sum pooltype yet) and Concat;
+ *
+ * Before fuse:
+ *    |         |             |
+ * seq_pool, seq_pool, ... seq_pool
+ *    \         |      ...   /
+ *            concat
+ *              |
+ * After fuse:
+ *    \      |       /
+ *   FusionSeqPoolConcat
+ *           |
+ */
+class SeqPoolConcatFusePass : public FusePassBase {
+ public:
+  virtual ~SeqPoolConcatFusePass() {}
+
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
+
+  const std::string name_scope_{"seqpool_concat_fuse"};
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/seqpool_concat_fuse_pass_tester.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/framework/ir/seqpool_concat_fuse_pass.h"
+#include <gtest/gtest.h>
+#include "paddle/fluid/framework/op_proto_maker.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+void SetOp(ProgramDesc* prog, const std::string& type,
+           const std::vector<std::string>& inputs,
+           const std::vector<std::string>& outputs) {
+  auto* op = prog->MutableBlock(0)->AppendOp();
+  op->SetType(type);
+  if (type == "sequence_pool") {
+    op->SetInput("X", {inputs[0]});
+    std::string pooltype = "SUM";
+    op->SetAttr("pooltype", pooltype);
+    op->SetOutput("MaxIndex", {outputs[0]});
+    op->SetOutput("Out", {outputs[1]});
+  } else if (type == "concat") {
+    op->SetInput("X", inputs);
+    op->SetAttr("axis", 1);
+    op->SetOutput("Out", {outputs[0]});
+  } else {
+    op->SetInput("X", inputs);
+    op->SetOutput("Out", outputs);
+  }
+  op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
+              static_cast<int>(OpRole::kForward));
+}
+
+int CountOpType(const ir::Graph* graph,
+                const std::string& op_type = "fusion_seqpool_concat") {
+  int count = 0;
+  for (auto* node : graph->Nodes()) {
+    if (node->IsOp() && node->Op()->Type() == op_type) {
+      ++count;
+    }
+  }
+  return count;
+}
+
+std::unique_ptr<ir::Graph> GetNumNodesOfBeforeAfter(
+    std::unique_ptr<ir::Graph> graph, int* before, int* after,
+    const std::string& pass_type = "seqpool_concat_fuse_pass") {
+  auto pass = PassRegistry::Instance().Get(pass_type);
+  *before = graph->Nodes().size();
+  graph = pass->Apply(std::move(graph));
+  *after = graph->Nodes().size();
+  return graph;
+}
+
+/*
+ * Before fuse:
+ *    a         b         c
+ *    |         |         |
+ *   op1       op2       op3
+ *   / \       / \       / \
+ *  d  e      f   g     h   i
+ *      \         |        /
+ *            concat
+ *              |
+ *              j
+ * Type of op1, op2 and op3 are sequence_pool, with "SUM" pooltype attr
+ *
+ * After fuse:
+ *    a         b         c
+ *    \         |        /
+ *    fusion_seqpool_concat
+ *              |
+ *              j
+ */
+TEST(SeqPoolConcatFusePass, basic) {
+  ProgramDesc prog;
+  for (auto& v : std::vector<std::string>(
+           {"a", "b", "c", "d", "e", "f", "g", "h", "i", "j"})) {
+    auto* var = prog.MutableBlock(0)->Var(v);
+    var->SetType(proto::VarType::LOD_TENSOR);
+  }
+
+  SetOp(&prog, "sequence_pool", std::vector<std::string>({"a"}),
+        std::vector<std::string>({"d", "e"}));
+  SetOp(&prog, "sequence_pool", std::vector<std::string>({"b"}),
+        std::vector<std::string>({"f", "g"}));
+  SetOp(&prog, "sequence_pool", std::vector<std::string>({"c"}),
+        std::vector<std::string>({"h", "i"}));
+  SetOp(&prog, "concat", std::vector<std::string>({"e", "g", "i"}),
+        std::vector<std::string>({"j"}));
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+  int before, after;
+  graph = GetNumNodesOfBeforeAfter(std::move(graph), &before, &after);
+  // Remove 10 Nodes: op1, op2, op3, d, e, f, g, h, i, concat_op
+  // Add 1 Node: fusion_seqpool_concat
+  EXPECT_EQ(after, before - 9);
+  EXPECT_EQ(CountOpType(graph.get()), 1);
+}
+
+/*
+ * Before fuse:
+ *    a            b
+ *    |           /  \
+ *   op1        op2  op3
+ *   / \        / \    \
+ *  c  d       e   f    g
+ *      \         /
+ *        concat
+ *          |
+ *          h
+ * Type of op1 and op2 are sequence_pool, with "SUM" pooltype attr
+ *
+ * After fuse:
+ *   a                         b
+ *    \                     /     \
+ *    fusion_seqpool_concat       op3
+ *              |                  |
+ *              h                  g
+ */
+TEST(SeqPoolConcatFusePass, advanced) {
+  ProgramDesc prog;
+  for (auto& v :
+       std::vector<std::string>({"a", "b", "c", "d", "e", "f", "g", "h"})) {
+    auto* var = prog.MutableBlock(0)->Var(v);
+    var->SetType(proto::VarType::LOD_TENSOR);
+  }
+
+  SetOp(&prog, "sequence_pool", std::vector<std::string>({"a"}),
+        std::vector<std::string>({"c", "d"}));
+  SetOp(&prog, "sequence_pool", std::vector<std::string>({"b"}),
+        std::vector<std::string>({"e", "f"}));
+  SetOp(&prog, "op3", std::vector<std::string>({"b"}),
+        std::vector<std::string>({"g"}));
+  SetOp(&prog, "concat", std::vector<std::string>({"d", "f"}),
+        std::vector<std::string>({"h"}));
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+  int before, after;
+  graph = GetNumNodesOfBeforeAfter(std::move(graph), &before, &after);
+  // Remove 7 Nodes: op1, op2, c, d, e, f concat_op
+  // Add 1 Node: fusion_seqpool_concat
+  EXPECT_EQ(after, before - 6);
+  EXPECT_EQ(CountOpType(graph.get()), 1);
+}
+
+ProgramDesc BuildProgramDesc(int num_inputs_of_concat) {
+  ProgramDesc prog;
+  auto new_var = [&](const std::string& name) {
+    auto* var = prog.MutableBlock(0)->Var(name);
+    var->SetType(proto::VarType::LOD_TENSOR);
+  };
+  std::vector<std::string> concat_inputs;
+  for (int i = 0; i < num_inputs_of_concat; ++i) {
+    std::string prefix = "seqpool_op_" + i;
+    new_var(prefix + "in");
+    new_var(prefix + "out");
+    new_var(prefix + "out_unused");
+    SetOp(&prog, "sequence_pool", std::vector<std::string>({prefix + "in"}),
+          std::vector<std::string>({prefix + "out", prefix + "out_unused"}));
+    concat_inputs.push_back(prefix + "out");
+  }
+  SetOp(&prog, "concat", concat_inputs,
+        std::vector<std::string>({"concat_out"}));
+  return prog;
+}
+
+// test more inputs of concat
+TEST(SeqPoolConcatFusePass, more_inputs) {
+  for (int num : {1, 2, 10}) {
+    ProgramDesc prog = BuildProgramDesc(num);
+    std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+    int before, after;
+    graph = GetNumNodesOfBeforeAfter(std::move(graph), &before, &after);
+    // Remove Nodes: n * (seqpool_op, out, out_unused), and concat_op
+    // Add Node: fusion_seqpool_concat op
+    EXPECT_EQ(after, before - num * 3);
+    EXPECT_EQ(CountOpType(graph.get()), 1);
+  }
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+USE_PASS(seqpool_concat_fuse_pass);

paddle/fluid/framework/ir/transpose_flatten_concat_fuse_pass.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/ir/graph_viz_pass.h"
+#include "paddle/fluid/framework/ir/node.h"
+#include "paddle/fluid/framework/ir/transpose_flatten_concat_fuse_pass.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+template <int times>
+std::unique_ptr<ir::Graph> TransposeFlattenConcatFusePass<times>::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  const std::string pattern_name =
+      "transpose_flatten" + std::to_string(times) + "_concat_fuse";
+  FusePassBase::Init(pattern_name, graph.get());
+
+  GraphPatternDetector gpd;
+  std::vector<PDNode *> input_nodes;
+  for (int i = 0; i < times; i++) {
+    input_nodes.push_back(gpd.mutable_pattern()
+                              ->NewNode("x" + std::to_string(i))
+                              ->assert_is_op_input("transpose2", "X")
+                              ->AsInput());
+  }
+
+  patterns::TransposeFlattenConcat pattern(gpd.mutable_pattern(), pattern_name);
+  pattern(input_nodes, times);
+
+  auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph,
+                     Graph *g) {
+    const int kNumFields = 5;
+    const int kTransOffset = 1;
+    const int kTransOutOffset = 2;
+    const int kFlattenOffset = 3;
+    const int kFlattenOutOffset = 4;
+    std::vector<Node *> nodes;
+
+    for (int i = 0; i < times; i++) {
+      PADDLE_ENFORCE(
+          subgraph.at(pattern.GetPDNode("transpose" + std::to_string(i))));
+      PADDLE_ENFORCE(
+          subgraph.at(pattern.GetPDNode("transpose_out" + std::to_string(i))));
+      PADDLE_ENFORCE(
+          subgraph.at(pattern.GetPDNode("flatten" + std::to_string(i))));
+      PADDLE_ENFORCE(
+          subgraph.at(pattern.GetPDNode("flatten_out" + std::to_string(i))));
+      PADDLE_ENFORCE(subgraph.at(input_nodes[i]));
+
+      nodes.push_back(subgraph.at(input_nodes[i]));
+      nodes.push_back(
+          subgraph.at(pattern.GetPDNode("transpose" + std::to_string(i))));
+      nodes.push_back(
+          subgraph.at(pattern.GetPDNode("transpose_out" + std::to_string(i))));
+      nodes.push_back(
+          subgraph.at(pattern.GetPDNode("flatten" + std::to_string(i))));
+      nodes.push_back(
+          subgraph.at(pattern.GetPDNode("flatten_out" + std::to_string(i))));
+    }
+
+    Node *concat_op = subgraph.at(pattern.GetPDNode("concat"));
+    Node *concat_out = subgraph.at(pattern.GetPDNode("concat_out"));
+    std::vector<std::string> input_names;
+    std::vector<int> trans_axis = boost::get<std::vector<int>>(
+        nodes[kTransOffset]->Op()->GetAttr("axis"));
+    int flatten_axis =
+        boost::get<int>(nodes[kFlattenOffset]->Op()->GetAttr("axis"));
+    int concat_axis = boost::get<int>(concat_op->Op()->GetAttr("axis"));
+    std::string output_name = concat_out->Name();
+
+    for (int i = 0; i < times; i++) {
+      input_names.push_back(nodes[i * kNumFields]->Name());
+    }
+
+    framework::OpDesc new_op_desc;
+    new_op_desc.SetType("fusion_transpose_flatten_concat");
+    new_op_desc.SetInput("X", input_names);
+    new_op_desc.SetAttr("trans_axis", trans_axis);
+    new_op_desc.SetAttr("flatten_axis", flatten_axis);
+    new_op_desc.SetAttr("concat_axis", concat_axis);
+    new_op_desc.SetOutput("Out", {output_name});
+    new_op_desc.Flush();
+
+    // Create a new node for the fused op.
+    auto *new_conv_op = graph->CreateOpNode(&new_op_desc);
+
+    std::unordered_set<const Node *> delete_nodes;
+
+    for (int i = 0; i < times; i++) {
+      nodes[i * kNumFields]->outputs.push_back(new_conv_op);
+      new_conv_op->inputs.push_back(nodes[i * kNumFields]);
+      delete_nodes.insert(nodes[i * kNumFields + kTransOffset]);
+      delete_nodes.insert(nodes[i * kNumFields + kTransOutOffset]);
+      delete_nodes.insert(nodes[i * kNumFields + kFlattenOffset]);
+      delete_nodes.insert(nodes[i * kNumFields + kFlattenOutOffset]);
+    }
+    delete_nodes.insert(concat_op);
+
+    new_conv_op->outputs.push_back(concat_out);
+    concat_out->inputs.push_back(new_conv_op);
+
+    // Delete the unneeded nodes.
+    GraphSafeRemoveNodes(graph.get(), delete_nodes);
+  };
+
+  gpd(graph.get(), handler);
+  return graph;
+}
+
+template class TransposeFlattenConcatFusePass<1>;
+template class TransposeFlattenConcatFusePass<3>;
+template class TransposeFlattenConcatFusePass<4>;
+template class TransposeFlattenConcatFusePass<5>;
+template class TransposeFlattenConcatFusePass<6>;
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(transpose_flatten_concat_fuse_pass,
+              paddle::framework::ir::TransposeFlattenConcatFusePass<1>);
+
+REGISTER_PASS(transpose_flatten3_concat_fuse_pass,
+              paddle::framework::ir::TransposeFlattenConcatFusePass<3>);
+
+REGISTER_PASS(transpose_flatten4_concat_fuse_pass,
+              paddle::framework::ir::TransposeFlattenConcatFusePass<4>);
+
+REGISTER_PASS(transpose_flatten5_concat_fuse_pass,
+              paddle::framework::ir::TransposeFlattenConcatFusePass<5>);
+
+REGISTER_PASS(transpose_flatten6_concat_fuse_pass,
+              paddle::framework::ir::TransposeFlattenConcatFusePass<6>);

paddle/fluid/framework/ir/transpose_flatten_concat_fuse_pass.h

+// 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.
+
+#pragma once
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+// There may be many transpose-flatten structures in a model, and the output of
+// these structures will be used as inputs to the concat Op. This pattern will
+// be detected by our pass. The times here represents the repeat times of this
+// structure.
+template <int times>
+class TransposeFlattenConcatFusePass : public FusePassBase {
+ public:
+  virtual ~TransposeFlattenConcatFusePass() {}
+
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/op_proto_maker.cc

@@ -82,6 +82,10 @@ void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto,
   AddAttr<std::string>(OpNamescopeAttrName(), "Operator name with namesope.")
       .SetDefault("");
 
+  AddAttr<std::vector<std::string>>(OpCreationCallstackAttrName(),
+                                    "Callstack for Op Creatation.")
+      .SetDefault({});
+
   Validate();
 }
 

paddle/fluid/framework/op_proto_maker.h

@@ -47,6 +47,7 @@ class OpProtoAndCheckerMaker {
   static const char *OpRoleAttrName() { return "op_role"; }
   static const char *OpRoleVarAttrName() { return "op_role_var"; }
   static const char *OpNamescopeAttrName() { return "op_namescope"; }
+  static const char *OpCreationCallstackAttrName() { return "op_callstack"; }
 
   void operator()(proto::OpProto *proto, OpAttrChecker *attr_checker);
 

paddle/fluid/framework/operator.cc

@@ -16,9 +16,15 @@ limitations under the License. */
 #include <glog/logging.h>
 
 #include <algorithm>
+#include <sstream>
+#include <string>
+#include <vector>
+#include "gflags/gflags.h"
+#include "glog/logging.h"
 #include "paddle/fluid/framework/data_transform.h"
 #include "paddle/fluid/framework/executor.h"
 #include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/op_proto_maker.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/shape_inference.h"
 #include "paddle/fluid/framework/transfer_scope_cache.h"
@@ -156,27 +162,55 @@ RuntimeContext::RuntimeContext(const VariableNameMap& innames,
 }
 
 void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
-  VLOG(4) << place << " " << DebugStringEx(&scope);
-  if (platform::is_gpu_place(place)) {
+  try {
+    VLOG(4) << place << " " << DebugStringEx(&scope);
+    if (platform::is_gpu_place(place)) {
 #ifndef PADDLE_WITH_CUDA
-    PADDLE_THROW("Cannot run operator on place %s", place);
+      PADDLE_THROW("Cannot run operator on place %s", place);
 #else
-    auto dev_id = boost::get<platform::CUDAPlace>(place).device;
-    platform::SetDeviceId(dev_id);
+      auto dev_id = boost::get<platform::CUDAPlace>(place).device;
+      platform::SetDeviceId(dev_id);
 #endif
-  }
+    }
 
-  // The profile has a process-wide mutex, results in serious performance issue
-  // in concurrency scenerio. Here use an `if` to fix this issue.
-  // Please not remove the `if`, ask @Superjomn if there are any concern.
-  if (platform::IsProfileEnabled()) {
-    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    platform::RecordEvent record_event(Type(), pool.Get(place));
-    RunImpl(scope, place);
-  } else {
-    RunImpl(scope, place);
+    // The profile has a process-wide mutex, results in serious performance
+    // issue
+    // in concurrency scenerio. Here use an `if` to fix this issue.
+    // Please not remove the `if`, ask @Superjomn if there are any concern.
+    if (platform::IsProfileEnabled()) {
+      platform::DeviceContextPool& pool =
+          platform::DeviceContextPool::Instance();
+      platform::RecordEvent record_event(Type(), pool.Get(place));
+      RunImpl(scope, place);
+    } else {
+      RunImpl(scope, place);
+    }
+
+    VLOG(3) << place << " " << DebugStringEx(&scope);
+  } catch (platform::EnforceNotMet exception) {
+    if (Attrs().count("sub_block") != 0) {
+      throw exception;
+    }
+
+    auto& callstack = Attr<std::vector<std::string>>(
+        OpProtoAndCheckerMaker::OpCreationCallstackAttrName());
+
+    if (callstack.empty()) {
+      throw exception;
+    }
+    std::ostringstream sout;
+    sout << "Invoke operator " << Type() << " error.\n";
+    sout << "Python Callstacks: \n";
+    for (auto& line : callstack) {
+      sout << line;
+    }
+    sout << "C++ Callstacks: \n";
+    sout << exception.err_str_;
+    exception.err_str_ = sout.str();
+    throw exception;
+  } catch (...) {
+    std::rethrow_exception(std::current_exception());
   }
-  VLOG(3) << place << " " << DebugStringEx(&scope);
 }
 
 bool OperatorBase::HasInputs(const std::string& name) const {

paddle/fluid/framework/operator.h

@@ -391,7 +391,7 @@ class ExecutionContext {
     PADDLE_ENFORCE(
         dynamic_cast<platform::TemporaryAllocation*>(allocation_ptr) != nullptr,
         "The AllocationPtr must be TemporaryAllocation.");
-    PADDLE_ENFORCE_EQ(allocation_ptr->size(),
+    PADDLE_ENFORCE_GE(allocation_ptr->size(),
                       framework::product(dim) * sizeof(T));
 
     paddle::framework::Tensor temp_tensor(

paddle/fluid/framework/python_headers.h

+/* 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. */
+
+#pragma once
+
+// workaround for Python 2 issue: https://bugs.python.org/issue17120
+#pragma push_macro("_XOPEN_SOURCE")
+#pragma push_macro("_POSIX_C_SOURCE")
+#undef _XOPEN_SOURCE
+#undef _POSIX_C_SOURCE
+
+#include "pybind11/pybind11.h"
+
+#pragma pop_macro("_XOPEN_SOURCE")
+#pragma pop_macro("_POSIX_C_SOURCE")

paddle/fluid/imperative/CMakeLists.txt

+if(WITH_PYTHON)
 cc_library(layer SRCS layer.cc DEPS proto_desc operator)
 cc_library(tracer SRCS tracer.cc DEPS proto_desc)
 cc_library(engine SRCS engine.cc)
+endif()

paddle/fluid/imperative/layer.cc

@@ -27,6 +27,8 @@
 namespace paddle {
 namespace imperative {
 
+std::map<int, py::object> py_funcs_;
+
 using framework::Variable;
 
 void AddTo(Variable* src, Variable* dst) {
@@ -42,7 +44,7 @@ void AddTo(Variable* src, Variable* dst) {
                  src_tensor->numel());
   float* dst_data = dst_tensor->mutable_data<float>(platform::CPUPlace());
   const float* src_data = src_tensor->data<float>();
-  for (size_t i = 0; i < src_tensor->numel(); ++i) {
+  for (int64_t i = 0; i < src_tensor->numel(); ++i) {
     dst_data[i] += src_data[i];
   }
 }
@@ -55,6 +57,7 @@ class Autograd {
     if (var->stop_gradient_) {
       return;
     }
+    VLOG(3) << "start autograd";
 
     std::deque<OpBase*> ready;
     ready.push_back(var->pre_op_);
@@ -114,57 +117,63 @@ class Autograd {
   }
 };
 
-framework::LoDTensor& VarBase::Grad() {
+framework::LoDTensor& VarBase::GradValue() {
   VLOG(3) << "get var grad " << var_desc_->Name();
-  return *grads_->GetMutable<framework::LoDTensor>();
+  return *(grads_->var_->GetMutable<framework::LoDTensor>());
 }
 
 std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad() {
-  if (!grad_op_desc_) {
+  if (!grad_op_desc_ && backward_id_ <= 0) {
     LOG(WARNING) << "op with no grad: " << op_desc_->Type();
     return {};
   }
-  VLOG(3) << "op grad " << grad_op_desc_->Type();
 
-  std::vector<std::unique_ptr<framework::Variable>> tmp_vars;
   std::map<std::string, std::vector<framework::Variable*>> grad_outputs;
-  for (auto it : grad_output_vars_) {
-    auto& outputs = grad_outputs[it.first];
-    for (size_t i = 0; i < it.second.size(); ++i) {
-      // Allocate a new variable
-      Variable* tmp_var = new framework::Variable();
-      tmp_var->GetMutable<framework::LoDTensor>();
-
-      tmp_vars.emplace_back(tmp_var);
-      outputs.push_back(tmp_var);
+  if (backward_id_ > 0) {
+    VLOG(3) << "py_layer_grad";
+    grad_outputs["Out@GRAD"] =
+        PyLayer::ApplyGrad(backward_id_, grad_input_vars_["X@GRAD"]);
+  } else {
+    VLOG(3) << "op grad " << grad_op_desc_->Type();
+    for (auto it : grad_output_vars_) {
+      auto& outputs = grad_outputs[it.first];
+      for (size_t i = 0; i < it.second.size(); ++i) {
+        // Allocate a new variable
+        Variable* tmp_var = new framework::Variable();
+        tmp_var->GetMutable<framework::LoDTensor>();
+        outputs.push_back(tmp_var);
+      }
     }
-  }
 
-  framework::RuntimeContext ctx(grad_input_vars_, grad_outputs);
+    framework::RuntimeContext ctx(grad_input_vars_, grad_outputs);
 
-  // No need to do compile time infer shape here.
-  // grad_op_desc_->InferShape(*block_);
-  grad_op_desc_->InferVarType(block_);
+    // No need to do compile time infer shape here.
+    // grad_op_desc_->InferShape(*block_);
+    grad_op_desc_->InferVarType(block_);
 
-  std::unique_ptr<framework::OperatorBase> opbase =
-      framework::OpRegistry::CreateOp(*grad_op_desc_);
-  framework::OperatorWithKernel* op_kernel =
-      dynamic_cast<framework::OperatorWithKernel*>(opbase.get());
-  PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel");
+    std::unique_ptr<framework::OperatorBase> opbase =
+        framework::OpRegistry::CreateOp(*grad_op_desc_);
+    framework::OperatorWithKernel* op_kernel =
+        dynamic_cast<framework::OperatorWithKernel*>(opbase.get());
+    PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel");
 
-  framework::Scope scope;
-  platform::CPUPlace place;
-  PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place);
-  p.op.RuntimeInferShape(scope, place, ctx);
-  p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx));
+    framework::Scope scope;
+    platform::CPUPlace place;
+    PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place);
+    p.op.RuntimeInferShape(scope, place, ctx);
+    p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx));
+  }
 
   for (auto it : grad_output_vars_) {
     auto& outputs = grad_outputs[it.first];
     auto& origin_outputs = it.second;
+    PADDLE_ENFORCE_EQ(outputs.size(), origin_outputs.size());
 
     for (size_t i = 0; i < outputs.size(); ++i) {
+      framework::Variable* grad = outputs[i];
       framework::Variable* orig_grad = origin_outputs[i];
-      AddTo(outputs[i], orig_grad);
+      AddTo(grad, orig_grad);
+      delete grad;
     }
   }
   return input_vars_;
@@ -173,7 +182,8 @@ std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad() {
 void VarBase::RunBackward() {
   if (!pre_op_) return;
 
-  auto grads_t = grads_->GetMutable<framework::LoDTensor>();
+  VLOG(3) << "start backward";
+  auto grads_t = grads_->var_->GetMutable<framework::LoDTensor>();
   float* data = grads_t->mutable_data<float>(platform::CPUPlace());
   std::fill(data, data + grads_t->numel(), 1.0);
 
@@ -183,5 +193,65 @@ void VarBase::RunBackward() {
   Autograd().RunBackward(this);
 }
 
+void PyLayer::RegisterFunc(int func_id, const py::object& py_func) {
+  py_funcs_[func_id] = py_func;
+}
+
+int PyLayer::NumFuncs() { return py_funcs_.size(); }
+
+std::vector<VarBase*> PyLayer::Apply(int func_id,
+                                     const std::vector<VarBase*>& inputs) {
+  std::vector<framework::Variable*> invars;
+  for (const VarBase* in : inputs) {
+    invars.push_back(in->var_);
+  }
+  PADDLE_ENFORCE(py_funcs_.find(func_id) != py_funcs_.end());
+  std::vector<Variable*> outvars = CallPythonFunc(py_funcs_[func_id], invars);
+  std::vector<VarBase*> ret;
+  for (Variable* v : outvars) {
+    ret.push_back(new VarBase(v, new VarBase(true)));
+  }
+  return ret;
+}
+
+std::vector<Variable*> PyLayer::ApplyGrad(
+    int func_id, const std::vector<framework::Variable*>& inputs) {
+  PADDLE_ENFORCE(py_funcs_.find(func_id) != py_funcs_.end());
+  return CallPythonFunc(py_funcs_[func_id], inputs);
+}
+
+std::vector<framework::Variable*> PyLayer::CallPythonFunc(
+    const py::object& callable, const std::vector<framework::Variable*>& ins) {
+  py::gil_scoped_acquire guard;
+  py::tuple in_args(ins.size());
+  for (size_t i = 0; i < ins.size(); ++i) {
+    const framework::LoDTensor& t = ins[i]->Get<framework::LoDTensor>();
+    in_args[i] = t.IsInitialized() ? py::cast(t) : py::cast(nullptr);
+  }
+  VLOG(3) << "pyfunc in " << py::len(in_args);
+
+  // TODO(panyx0718): Who owns the returned LoDTensor.
+  auto ret = callable(in_args);
+  auto ret_tuple = py::cast<py::tuple>(ret);
+  size_t ret_num = py::len(ret_tuple);
+  std::vector<framework::Variable*> outs;
+  VLOG(3) << "pyfunc out " << ret_num;
+  for (size_t i = 0; i < ret_num; ++i) {
+    try {
+      auto* py_out_tensor = py::cast<framework::LoDTensor*>(ret_tuple[i]);
+      PADDLE_ENFORCE_NOT_NULL(py_out_tensor,
+                              "Output tensor %d should not be nullptr", i);
+      auto* var = new framework::Variable();
+      auto* tensor = var->GetMutable<framework::LoDTensor>();
+      tensor->ShareDataWith(*py_out_tensor);
+      tensor->set_lod(py_out_tensor->lod());
+      outs.push_back(var);
+    } catch (py::cast_error&) {
+      PADDLE_THROW("The %d-th output must be LoDTensor", i);
+    }
+  }
+  return outs;
+}
+
 }  // namespace imperative
 }  // namespace paddle

paddle/fluid/imperative/layer.h

@@ -14,17 +14,26 @@
 
 #pragma once
 
-#include <map>
-#include <string>
-#include <vector>
+// clang-format off
+#include "paddle/fluid/framework/python_headers.h"
+// clang-format on
+
+#include <map>     // NOLINT
+#include <string>  // NOLINT
+#include <vector>  // NOLINT
+
 #include "paddle/fluid/framework/op_desc.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/var_desc.h"
 #include "paddle/fluid/platform/enforce.h"
 
+#include "paddle/fluid/imperative/type_defs.h"
+
 namespace paddle {
 namespace imperative {
 
+namespace py = ::pybind11;
+
 class PreparedOp {
  public:
   PreparedOp(const framework::OperatorBase& op,
@@ -80,29 +89,47 @@ class PreparedOp {
 
 class OpBase;
 
+/* The wrapper for Variable which holds a Variable and a VarBase of its
+ * gradient. This object should be managed totally by Python intepreter.
+ *
+ * Nearly all interface should be implemented in C++.
+ */
 class VarBase {
  public:
-  VarBase()
+  VarBase() : VarBase(new framework::Variable(), new VarBase(true)) {}
+
+  // Owns `var` and `grad`
+  VarBase(framework::Variable* var, VarBase* grad)
       : pre_op_(nullptr),
+        pre_op_out_name_(),
         pre_op_out_idx_(-1),
         var_desc_(nullptr),
-        var_(new framework::Variable()),
-        grads_(new framework::Variable()),
+        var_(var),
+        grads_(grad),
         stop_gradient_(false) {}
 
   explicit VarBase(bool stop_gradient)
       : pre_op_(nullptr),
+        pre_op_out_name_(),
         pre_op_out_idx_(-1),
         var_desc_(nullptr),
         var_(new framework::Variable()),
-        grads_(new framework::Variable()),
+        grads_(stop_gradient ? nullptr : new VarBase(true)),
         stop_gradient_(stop_gradient) {}
 
-  virtual ~VarBase() {}
+  virtual ~VarBase() {
+    if (var_) {
+      delete var_;
+    }
+
+    if (grads_) {
+      delete grads_;
+    }
+  }
 
   void RunBackward();
 
-  framework::LoDTensor& Grad();
+  framework::LoDTensor& GradValue();
 
   inline std::string GradName() const {
     PADDLE_ENFORCE(
@@ -116,15 +143,23 @@ class VarBase {
   int pre_op_out_idx_;
 
   framework::VarDesc* var_desc_;
+
   framework::Variable* var_;
-  framework::Variable* grads_;
+  VarBase* grads_;
 
   bool stop_gradient_;
 };
 
+/* The wrapper for OpDesc which holds a OpDesc and a OpDesc of its
+ * gradient. This object should be managed totally by Python intepreter.
+ */
 class OpBase {
  public:
-  OpBase() : op_desc_(nullptr), grad_op_desc_(nullptr) {}
+  OpBase()
+      : op_desc_(nullptr),
+        forward_id_(-1),
+        grad_op_desc_(nullptr),
+        backward_id_(-1) {}
 
   virtual ~OpBase() {
     if (grad_op_desc_) delete grad_op_desc_;
@@ -132,16 +167,22 @@ class OpBase {
 
   std::map<std::string, std::vector<VarBase*>> ApplyGrad();
 
+  // One of `op_desc_` or `forward_id_` is set, not both.
+  // For pure python PyLayer, use `forward_id_`, otherwise, use op_desc_.
   framework::OpDesc* op_desc_;
+  int forward_id_;
+  // When has backward, one of `grad_op_desc_` or `backward_id_` is set,
+  // not both.
   framework::OpDesc* grad_op_desc_;
+  int backward_id_;
 
-  std::map<std::string, std::vector<VarBase*>> input_vars_;
-  std::map<std::string, std::vector<VarBase*>> output_vars_;
-  std::map<std::string, std::vector<OpBase*>> pre_ops_;
+  VarBasePtrMap input_vars_;
+  VarBasePtrMap output_vars_;
+  OpBasePtrMap pre_ops_;
   std::map<std::string, std::vector<int>> pre_ops_out_idx_;
 
-  std::map<std::string, std::vector<framework::Variable*>> grad_input_vars_;
-  std::map<std::string, std::vector<framework::Variable*>> grad_output_vars_;
+  framework::VariableValueMap grad_input_vars_;
+  framework::VariableValueMap grad_output_vars_;
   framework::BlockDesc* block_;
 };
 
@@ -153,8 +194,25 @@ class Layer {
     std::vector<VarBase> vars;
     return vars;
   }
+};
+
+class PyLayer {
+ public:
+  virtual ~PyLayer() {}
+
+  static void RegisterFunc(int func_id, const py::object& py_func);
+
+  static int NumFuncs();
+
+  static std::vector<VarBase*> Apply(int func_id,
+                                     const std::vector<VarBase*>& inputs);
+
+  static std::vector<framework::Variable*> ApplyGrad(
+      int func_id, const std::vector<framework::Variable*>& inputs);
 
-  virtual void Backward() { LOG(ERROR) << "To support customize"; }
+ private:
+  static std::vector<framework::Variable*> CallPythonFunc(
+      const py::object& callable, const std::vector<framework::Variable*>& ins);
 };
 
 }  // namespace imperative

paddle/fluid/imperative/tracer.cc

@@ -15,5 +15,199 @@
 #include "paddle/fluid/imperative/tracer.h"
 
 namespace paddle {
-namespace imperative {}  // namespace imperative
+namespace imperative {
+
+void CreateGradOp(const framework::OpDesc& op_desc,
+                  const std::unordered_set<std::string>& no_grad_set,
+                  const std::vector<framework::BlockDesc*>& grad_sub_block,
+                  framework::OpDesc** grad_op_desc,
+                  std::unordered_map<std::string, std::string>* grad_to_var) {
+  std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs =
+      framework::OpInfoMap::Instance()
+          .Get(op_desc.Type())
+          .GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block);
+  PADDLE_ENFORCE(grad_op_descs.size() == 1, "Only support 1 grad op now.");
+  // TODO(panyx0718): Leak?
+  *grad_op_desc = grad_op_descs[0].release();
+}
+
+void InitVar(framework::Variable* var, framework::Variable* grad_var) {
+  auto& var_t = var->Get<framework::LoDTensor>();
+  float* data =
+      grad_var->GetMutable<framework::LoDTensor>()->mutable_data<float>(
+          var_t.dims(), platform::CPUPlace());
+  std::fill(data, data + var_t.numel(), 0.0);
+}
+
+void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
+                   const VarBasePtrMap& outputs, framework::BlockDesc* block,
+                   const bool stop_gradient) {
+  std::map<std::string, VarBase*> vars;
+
+  framework::OpDesc* op_desc = op->op_desc_;
+  VLOG(3) << "tracer tracing " << op_desc->Type();
+  op_desc->InferShape(*block);
+  op_desc->InferVarType(block);
+  std::unique_ptr<framework::OperatorBase> op_base =
+      framework::OpRegistry::CreateOp(*op_desc);
+
+  framework::VariableValueMap invars_map;
+  framework::VariableValueMap outvars_map;
+
+  op->input_vars_ = inputs;
+  for (auto it : op->input_vars_) {
+    auto& invars = invars_map[it.first];
+    for (VarBase* inp : it.second) {
+      PADDLE_ENFORCE_NOT_NULL(inp->var_, "op %s input %s nullptr",
+                              op->op_desc_->Type(), inp->var_desc_->Name());
+
+      invars.push_back(inp->var_);
+      vars[inp->var_desc_->Name()] = inp;
+      if (inp->pre_op_) {
+        op->pre_ops_[it.first].push_back(inp->pre_op_);
+        op->pre_ops_out_idx_[it.first].push_back(inp->pre_op_out_idx_);
+      } else {
+        op->pre_ops_[it.first].push_back(nullptr);
+      }
+      VLOG(3) << "input vname " << inp->var_desc_->Name() << " "
+              << inp->var_->IsInitialized();
+    }
+  }
+
+  op->output_vars_ = outputs;
+  for (auto it : op->output_vars_) {
+    auto& outvars = outvars_map[it.first];
+    const std::vector<VarBase*>& outputs = it.second;
+    for (size_t i = 0; i < outputs.size(); ++i) {
+      VarBase* out = outputs[i];
+      outvars.push_back(out->var_);
+      vars[out->var_desc_->Name()] = out;
+
+      framework::VarDesc* var_desc = block->FindVar(out->var_desc_->Name());
+      if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
+        out->var_->GetMutable<framework::LoDTensor>();
+      } else {
+        LOG(ERROR) << "tracer doesn't support yet";
+      }
+      out->stop_gradient_ = stop_gradient;
+      out->pre_op_ = op;
+      out->pre_op_out_name_ = it.first;
+      out->pre_op_out_idx_ = i;
+
+      VLOG(3) << "output vname " << out->var_desc_->Name() << " "
+              << out->var_->IsInitialized();
+    }
+  }
+
+  VLOG(3) << "tracer running " << op_desc->Type();
+  framework::RuntimeContext ctx(invars_map, outvars_map);
+
+  // TODO(panyx0718): Cache p.
+  framework::OperatorWithKernel* op_kernel =
+      dynamic_cast<framework::OperatorWithKernel*>(op_base.get());
+  PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel");
+
+  framework::Scope scope;
+  platform::CPUPlace place;
+  PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place);
+  p.op.RuntimeInferShape(scope, place, ctx);
+  p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx));
+
+  if (!stop_gradient) {
+    framework::OpDesc* grad_op_desc;
+    // TODO(panyx): Is this leaked?
+    std::unique_ptr<std::unordered_map<std::string, std::string>> grad_to_var(
+        new std::unordered_map<std::string, std::string>());
+    CreateGradOp(*op_desc, {}, {block}, &grad_op_desc, grad_to_var.get());
+    op->grad_op_desc_ = grad_op_desc;
+
+    for (auto it : grad_op_desc->Inputs()) {
+      auto& grad_in_vars = op->grad_input_vars_[it.first];
+      for (const std::string& grad_invar : it.second) {
+        block->FindRecursiveOrCreateVar(grad_invar);
+        auto var_it = grad_to_var->find(grad_invar);
+        if (var_it == grad_to_var->end()) {
+          auto fwd_var_it = vars.find(grad_invar);
+          PADDLE_ENFORCE(fwd_var_it != vars.end());
+          // Forward inputs or outputs.
+          grad_in_vars.push_back(fwd_var_it->second->var_);
+        } else {
+          VarBase* var = vars[var_it->second];
+          if (!var->grads_->var_->IsInitialized()) {
+            InitVar(var->var_, var->grads_->var_);
+          }
+          // Douts.
+          grad_in_vars.push_back(var->grads_->var_);
+        }
+      }
+    }
+
+    for (auto it : grad_op_desc->Outputs()) {
+      auto& grad_out_vars = op->grad_output_vars_[it.first];
+      for (const std::string& grad_outvar : it.second) {
+        block->FindRecursiveOrCreateVar(grad_outvar);
+        auto var_it = grad_to_var->find(grad_outvar);
+        PADDLE_ENFORCE(var_it != grad_to_var->end());
+        VarBase* var = vars[var_it->second];
+        if (!var->grads_->var_->IsInitialized()) {
+          InitVar(var->var_, var->grads_->var_);
+        }
+        grad_out_vars.push_back(var->grads_->var_);
+      }
+    }
+  }
+
+  op->block_ = block;
+}
+
+std::vector<VarBase*> Tracer::PyTrace(OpBase* op,
+                                      const std::vector<VarBase*>& inputs,
+                                      bool stop_gradient) {
+  VLOG(3) << "py_trace";
+  op->input_vars_["X"] = inputs;
+  op->output_vars_["Out"] = PyLayer::Apply(op->forward_id_, inputs);
+  for (VarBase* inp : inputs) {
+    if (inp->pre_op_) {
+      op->pre_ops_["X"].push_back(inp->pre_op_);
+      op->pre_ops_out_idx_["X"].push_back(inp->pre_op_out_idx_);
+    } else {
+      op->pre_ops_["X"].push_back(nullptr);
+    }
+  }
+
+  auto& outputs = op->output_vars_["Out"];
+  for (size_t i = 0; i < outputs.size(); ++i) {
+    VarBase* out = outputs[i];
+    out->stop_gradient_ = stop_gradient;
+    out->pre_op_ = op;
+    out->pre_op_out_name_ = "Out";
+    out->pre_op_out_idx_ = i;
+  }
+  if (!stop_gradient) {
+    auto& grad_input_vars = op->grad_input_vars_["X@GRAD"];
+    auto& grad_output_vars = op->grad_output_vars_["Out@GRAD"];
+
+    for (const VarBase* inp : inputs) {
+      grad_input_vars.push_back(inp->var_);
+    }
+    for (VarBase* out : outputs) {
+      grad_input_vars.push_back(out->var_);
+    }
+    for (VarBase* out : outputs) {
+      grad_input_vars.push_back(out->grads_->var_);
+      if (!grad_input_vars.back()->IsInitialized()) {
+        InitVar(out->var_, grad_input_vars.back());
+      }
+    }
+    for (const VarBase* inp : inputs) {
+      grad_output_vars.push_back(inp->grads_->var_);
+      if (!grad_output_vars.back()->IsInitialized()) {
+        InitVar(inp->var_, grad_output_vars.back());
+      }
+    }
+  }
+  return outputs;
+}
+
+}  // namespace imperative
 }  // namespace paddle

paddle/fluid/imperative/tracer.h

@@ -30,23 +30,9 @@ void CreateGradOp(const framework::OpDesc& op_desc,
                   const std::unordered_set<std::string>& no_grad_set,
                   const std::vector<framework::BlockDesc*>& grad_sub_block,
                   framework::OpDesc** grad_op_desc,
-                  std::unordered_map<std::string, std::string>* grad_to_var) {
-  std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs =
-      framework::OpInfoMap::Instance()
-          .Get(op_desc.Type())
-          .GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block);
-  PADDLE_ENFORCE(grad_op_descs.size() == 1, "Only support 1 grad op now.");
-  // TODO(panyx0718): Leak?
-  *grad_op_desc = grad_op_descs[0].release();
-}
+                  std::unordered_map<std::string, std::string>* grad_to_var);
 
-void InitVar(framework::Variable* var, framework::Variable* grad_var) {
-  auto& var_t = var->Get<framework::LoDTensor>();
-  float* data =
-      grad_var->GetMutable<framework::LoDTensor>()->mutable_data<float>(
-          var_t.dims(), platform::CPUPlace());
-  std::fill(data, data + var_t.numel(), 0.0);
-}
+void InitVar(framework::Variable* var, framework::Variable* grad_var);
 
 class Tracer {
  public:
@@ -57,120 +43,10 @@ class Tracer {
   void Trace(OpBase* op,
              const std::map<std::string, std::vector<VarBase*>>& inputs,
              const std::map<std::string, std::vector<VarBase*>>& outputs,
-             framework::BlockDesc* block, const bool stop_gradient = false) {
-    std::map<std::string, VarBase*> vars;
+             framework::BlockDesc* block, const bool stop_gradient = false);
 
-    framework::OpDesc* op_desc = op->op_desc_;
-    VLOG(3) << "tracer tracing " << op_desc->Type();
-    op_desc->InferShape(*block);
-    op_desc->InferVarType(block);
-    std::unique_ptr<framework::OperatorBase> op_base =
-        framework::OpRegistry::CreateOp(*op_desc);
-
-    framework::VariableValueMap invars_map;
-    framework::VariableValueMap outvars_map;
-
-    op->input_vars_ = inputs;
-    for (auto it : op->input_vars_) {
-      auto& invars = invars_map[it.first];
-      for (VarBase* inp : it.second) {
-        PADDLE_ENFORCE_NOT_NULL(inp->var_, "op %s input %s nullptr",
-                                op->op_desc_->Type(), inp->var_desc_->Name());
-
-        invars.push_back(inp->var_);
-        vars[inp->var_desc_->Name()] = inp;
-        if (inp->pre_op_) {
-          op->pre_ops_[it.first].push_back(inp->pre_op_);
-          op->pre_ops_out_idx_[it.first].push_back(inp->pre_op_out_idx_);
-        } else {
-          op->pre_ops_[it.first].push_back(nullptr);
-        }
-        VLOG(3) << "input vname " << inp->var_desc_->Name() << " "
-                << inp->var_->IsInitialized();
-      }
-    }
-
-    op->output_vars_ = outputs;
-    for (auto it : op->output_vars_) {
-      auto& outvars = outvars_map[it.first];
-      const std::vector<VarBase*>& outputs = it.second;
-      for (size_t i = 0; i < outputs.size(); ++i) {
-        VarBase* out = outputs[i];
-        outvars.push_back(out->var_);
-        vars[out->var_desc_->Name()] = out;
-
-        framework::VarDesc* var_desc = block->FindVar(out->var_desc_->Name());
-        if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
-          out->var_->GetMutable<framework::LoDTensor>();
-        } else {
-          LOG(ERROR) << "tracer doesn't support yet";
-        }
-        out->stop_gradient_ = stop_gradient;
-        out->pre_op_ = op;
-        out->pre_op_out_name_ = it.first;
-        out->pre_op_out_idx_ = i;
-
-        VLOG(3) << "output vname " << out->var_desc_->Name() << " "
-                << out->var_->IsInitialized();
-      }
-    }
-
-    VLOG(3) << "tracer running " << op_desc->Type();
-    framework::RuntimeContext ctx(invars_map, outvars_map);
-
-    // TODO(panyx0718): Cache p.
-    framework::OperatorWithKernel* op_kernel =
-        dynamic_cast<framework::OperatorWithKernel*>(op_base.get());
-    PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel");
-
-    framework::Scope scope;
-    platform::CPUPlace place;
-    PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place);
-    p.op.RuntimeInferShape(scope, place, ctx);
-    p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx));
-
-    if (!stop_gradient) {
-      framework::OpDesc* grad_op_desc;
-      auto grad_to_var = new std::unordered_map<std::string, std::string>();
-      CreateGradOp(*op_desc, {}, {block}, &grad_op_desc, grad_to_var);
-      op->grad_op_desc_ = grad_op_desc;
-
-      for (auto it : grad_op_desc->Inputs()) {
-        auto& grad_in_vars = op->grad_input_vars_[it.first];
-        for (const std::string& grad_invar : it.second) {
-          block->FindRecursiveOrCreateVar(grad_invar);
-          auto var_it = grad_to_var->find(grad_invar);
-          if (var_it == grad_to_var->end()) {
-            auto fwd_var_it = vars.find(grad_invar);
-            PADDLE_ENFORCE(fwd_var_it != vars.end());
-            grad_in_vars.push_back(fwd_var_it->second->var_);
-          } else {
-            VarBase* var = vars[var_it->second];
-            if (!var->grads_->IsInitialized()) {
-              InitVar(var->var_, var->grads_);
-            }
-            grad_in_vars.push_back(var->grads_);
-          }
-        }
-      }
-
-      for (auto it : grad_op_desc->Outputs()) {
-        auto& grad_out_vars = op->grad_output_vars_[it.first];
-        for (const std::string& grad_outvar : it.second) {
-          block->FindRecursiveOrCreateVar(grad_outvar);
-          auto var_it = grad_to_var->find(grad_outvar);
-          PADDLE_ENFORCE(var_it != grad_to_var->end());
-          VarBase* var = vars[var_it->second];
-          if (!var->grads_->IsInitialized()) {
-            InitVar(var->var_, var->grads_);
-          }
-          grad_out_vars.push_back(var->grads_);
-        }
-      }
-    }
-
-    op->block_ = block;
-  }
+  std::vector<VarBase*> PyTrace(OpBase* op, const std::vector<VarBase*>& inputs,
+                                bool stop_gradient = false);
 
  private:
   framework::BlockDesc* root_block_;

paddle/fluid/imperative/type_defs.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 <map>
+#include <string>
+#include <vector>
+
+namespace paddle {
+namespace imperative {
+
+class VarBase;
+class OpBase;
+
+typedef std::map<std::string, std::vector<VarBase*>> VarBasePtrMap;
+typedef std::map<std::string, std::vector<OpBase*>> OpBasePtrMap;
+
+}  // namespace imperative
+}  // namespace paddle

paddle/fluid/inference/api/analysis_config.cc

@@ -127,6 +127,7 @@ void contrib::AnalysisConfig::EnableTensorRtEngine(int workspace_size,
   use_tensorrt_ = true;
   tensorrt_workspace_size_ = workspace_size;
   tensorrt_max_batchsize_ = max_batch_size;
+  Update();
 }
 
 void contrib::AnalysisConfig::Update() {

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

@@ -92,10 +92,10 @@ if(WITH_MKL)
   if(NOT WIN32)
     set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
                  ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
-  else(WIN32)
+  else()
     set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml${CMAKE_SHARED_LIBRARY_SUFFIX}
             ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5md${CMAKE_SHARED_LIBRARY_SUFFIX})
-  endif(WIN32)
+  endif()
   set(MKLDNN_PATH "${PADDLE_LIB}/third_party/install/mkldnn")
   if(EXISTS ${MKLDNN_PATH})
     include_directories("${MKLDNN_PATH}/include")
@@ -128,8 +128,8 @@ else()
       ${CMAKE_STATIC_LIBRARY_PREFIX}glog  ${CMAKE_STATIC_LIBRARY_PREFIX}gflags  ${CMAKE_STATIC_LIBRARY_PREFIX}protobuf
       ${CMAKE_STATIC_LIBRARY_PREFIX}snappy ${CMAKE_STATIC_LIBRARY_PREFIX}z ${CMAKE_STATIC_LIBRARY_PREFIX}xxhash
       snappystream ${EXTERNAL_LIB})
-  # NOTE(dzhwinter) shlwapi is deprecated.
-  set(DEPS ${DEPS} libcmt shlwapi)
+  get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
+  set(DEPS ${DEPS} libcmt ${os_dependency_modules})
 endif(NOT WIN32)
 
 if(WITH_GPU)

paddle/fluid/inference/api/demo_ci/run.sh

@@ -116,6 +116,10 @@ D
       --modeldir=$DATA_DIR/mobilenet/model \
       --data=$DATA_DIR/mobilenet/data.txt \
       --refer=$DATA_DIR/mobilenet/result.txt 
+    if [ $? -ne 0 ]; then
+      echo "trt demo trt_mobilenet_demo runs fail."
+      exit 1
+    fi
   fi
 done
 set +x

paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc

@@ -38,8 +38,8 @@ void Main() {
   std::unique_ptr<PaddlePredictor> predictor;
   paddle::contrib::AnalysisConfig config;
   config.EnableUseGpu(100, 0);
-  config.SetModel(FLAGS_modeldir + "/__params__",
-                  FLAGS_modeldir + "/__model__");
+  config.SetModel(FLAGS_modeldir + "/__model__",
+                  FLAGS_modeldir + "/__params__");
   config.EnableTensorRtEngine();
   predictor = CreatePaddlePredictor(config);
 

paddle/fluid/inference/api/helper.h

@@ -204,11 +204,14 @@ static std::string DescribeTensor(const PaddleTensor &tensor) {
     os << to_string(l) << "; ";
   }
   os << "\n";
-  os << " - data: ";
+  os << " - memory length: " << tensor.data.length();
+  os << "\n";
 
+  os << " - data: ";
   int dim = VecReduceToInt(tensor.shape);
+  float *pdata = static_cast<float *>(tensor.data.data());
   for (int i = 0; i < dim; i++) {
-    os << static_cast<float *>(tensor.data.data())[i] << " ";
+    os << pdata[i] << " ";
   }
   os << '\n';
   return os.str();
@@ -224,10 +227,12 @@ static std::string DescribeZeroCopyTensor(const ZeroCopyTensor &tensor) {
     os << to_string(l) << "; ";
   }
   os << "\n";
-  os << " - data: ";
   PaddlePlace place;
   int size;
   const auto *data = tensor.data<float>(&place, &size);
+  os << " - numel: " << size;
+  os << "\n";
+  os << " - data: ";
   for (int i = 0; i < size; i++) {
     os << data[i] << " ";
   }

paddle/fluid/inference/api/paddle_api.h

@@ -123,7 +123,8 @@ class ZeroCopyTensor {
    */
   template <typename T>
   T* mutable_data(PaddlePlace place);
-  /** Get the memory directly, will return the place and memory size by pointer.
+  /** Get the memory directly, will return the place and element size by
+   * pointer.
    * This is for reading the output tensor.
    */
   template <typename T>

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -89,6 +89,7 @@ class CpuPassStrategy : public PassStrategy {
     passes_.assign({
         "infer_clean_graph_pass",         //
         "attention_lstm_fuse_pass",       //
+        "seqpool_concat_fuse_pass",       //
         "seqconv_eltadd_relu_fuse_pass",  //
         // "embedding_fc_lstm_fuse_pass", //
         "fc_lstm_fuse_pass",             //
@@ -140,6 +141,10 @@ class GpuPassStrategy : public PassStrategy {
         "conv_elementwise_add_fuse_pass",            //
     });
 
+    for (int i = 6; i >= 3; i--) {
+      passes_.push_back("transpose_flatten" + std::to_string(i) +
+                        "_concat_fuse_pass");
+    }
     use_gpu_ = true;
   }
 

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

@@ -39,6 +39,7 @@ class ElementwiseWeightOpConverter : public OpConverter {
                   const framework::Scope& scope, bool test_mode) override {
     // Here the two nullptr looks strange, that's because the
     // framework::OpDesc's constructor is strange.
+    nvinfer1::ILayer* layer = nullptr;
     framework::OpDesc op_desc(op, nullptr);
     VLOG(3) << "Convert a fluid elementwise op to TensorRT IScaleLayer";
 
@@ -98,13 +99,21 @@ class ElementwiseWeightOpConverter : public OpConverter {
                                          0};
     TensorRTEngine::Weight power_weights{nvinfer1::DataType::kFLOAT, nullptr,
                                          0};
+    if (op_type_ == "add") {
+      nvinfer1::IScaleLayer* scale_layer = TRT_ENGINE_ADD_LAYER(
+          engine_, Scale, *X, scale_mode, shift_weights.get(),
+          scale_weights.get(), power_weights.get());
+      layer = scale_layer;
+    } else if (op_type_ == "mul") {
+      nvinfer1::IScaleLayer* scale_layer = TRT_ENGINE_ADD_LAYER(
+          engine_, Scale, *X, scale_mode, scale_weights.get(),
+          shift_weights.get(), power_weights.get());
+      layer = scale_layer;
+    }
 
-    nvinfer1::IScaleLayer* layer = TRT_ENGINE_ADD_LAYER(
-        engine_, Scale, *const_cast<nvinfer1::ITensor*>(X), scale_mode,
-        shift_weights.get(), scale_weights.get(), power_weights.get());
     auto output_name = op_desc.Output("Out")[0];
-
-    layer->setName(("elementwise_add (Output: " + output_name + ")").c_str());
+    layer->setName(
+        ("elementwise_" + op_type_ + "(Output: " + output_name + ")").c_str());
     layer->getOutput(0)->setName(output_name.c_str());
     engine_->weight_map[op_desc.Input("Y").front()] = std::move(weight_tensor);
     engine_->SetITensor(output_name, layer->getOutput(0));
@@ -113,6 +122,9 @@ class ElementwiseWeightOpConverter : public OpConverter {
       engine_->DeclareOutput(output_name);
     }
   }
+
+ protected:
+  std::string op_type_;
 };
 
 class ElementwiseTensorOpConverter : public OpConverter {
@@ -188,6 +200,16 @@ const std::unordered_map<std::string, nvinfer1::ElementWiseOperation>
         {"max", nvinfer1::ElementWiseOperation::kMAX},
 };
 
+class ElementwiseWeightAddOpConverter : public ElementwiseWeightOpConverter {
+ public:
+  ElementwiseWeightAddOpConverter() { op_type_ = "add"; }
+};
+
+class ElementwiseWeightMulOpConverter : public ElementwiseWeightOpConverter {
+ public:
+  ElementwiseWeightMulOpConverter() { op_type_ = "mul"; }
+};
+
 class ElementwiseTensorAddOpConverter : public ElementwiseTensorOpConverter {
  public:
   ElementwiseTensorAddOpConverter() { op_type_ = "add"; }
@@ -227,7 +249,10 @@ class ElementwiseTensorPowOpConverter : public ElementwiseTensorOpConverter {
 }  // namespace inference
 }  // namespace paddle
 
-REGISTER_TRT_OP_CONVERTER(elementwise_add_weight, ElementwiseWeightOpConverter);
+REGISTER_TRT_OP_CONVERTER(elementwise_add_weight,
+                          ElementwiseWeightAddOpConverter);
+REGISTER_TRT_OP_CONVERTER(elementwise_mul_weight,
+                          ElementwiseWeightMulOpConverter);
 
 REGISTER_TRT_OP_CONVERTER(elementwise_add_tensor,
                           ElementwiseTensorAddOpConverter);

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

@@ -100,14 +100,14 @@ set(OCR_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/ocr")
 if (NOT EXISTS ${OCR_INSTALL_DIR})
     inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Focr.tar.gz")
 endif()
-inference_analysis_api_test_with_refer_result(test_analyzer_ocr ${OCR_INSTALL_DIR} analyzer_vis_tester.cc)
+inference_analysis_api_test_with_refer_result(test_analyzer_ocr ${OCR_INSTALL_DIR} analyzer_vis_tester.cc SERIAL)
 
 # mobilenet with transpose op
 set(MOBILENET_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/mobilenet")
 if (NOT EXISTS ${MOBILENET_INSTALL_DIR})
     inference_download_and_uncompress(${MOBILENET_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Fmobilenet.tar.gz")
 endif()
-inference_analysis_api_test_with_refer_result(test_analyzer_mobilenet_transpose ${MOBILENET_INSTALL_DIR} analyzer_vis_tester.cc)
+inference_analysis_api_test_with_refer_result(test_analyzer_mobilenet_transpose ${MOBILENET_INSTALL_DIR} analyzer_vis_tester.cc SERIAL)
 
 # resnet50
 inference_analysis_api_test_with_fake_data(test_analyzer_resnet50

paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc

@@ -283,7 +283,7 @@ TEST(Analyzer_rnn1, multi_thread) {
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
   TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
-                 input_slots_all, &outputs, 4 /* multi_thread */);
+                 input_slots_all, &outputs, 2 /* multi_thread */);
 }
 
 // Validate that the AnalysisPredictor + ZeroCopyTensor really works by testing
@@ -351,10 +351,10 @@ TEST(Analyzer_rnn1, ZeroCopy) {
   ASSERT_TRUE(native_predictor->Run(native_inputs.front(), &native_outputs));
   LOG(INFO) << "native output " << DescribeTensor(native_outputs.front());
 
-  int output_size{0};
+  int output_size{0};  // this is the number of elements not memory size
   auto *zero_copy_data = output_tensor->data<float>(&place, &output_size);
   auto *native_data = static_cast<float *>(native_outputs.front().data.data());
-  for (size_t i = 0; i < output_size / sizeof(float); i++) {
+  for (int i = 0; i < output_size; i++) {
     EXPECT_NEAR(zero_copy_data[i], native_data[i], 1e-3);
   }
 }
@@ -370,15 +370,12 @@ TEST(Analyzer_rnn1, ZeroCopyMultiThread) {
   auto base_predictor = CreatePaddlePredictor<AnalysisConfig>(config);
   double total_time_of_threads{0};
   std::vector<std::thread> threads;
-  std::vector<std::unique_ptr<PaddlePredictor>> predictors;
-  for (int tid = 0; tid < FLAGS_num_threads; tid++) {
-    predictors.emplace_back(CreatePaddlePredictor<AnalysisConfig>(config));
-  }
 
   for (int tid = 0; tid < FLAGS_num_threads; tid++) {
-    threads.emplace_back([config, &total_time_of_threads, &predictors, tid] {
-      // auto predictor = base_predictor->Clone();
-      auto &predictor = predictors[tid];
+    threads.emplace_back([&, tid] {
+      // To ensure the thread binding correctly,
+      // please clone inside the threadpool.
+      auto predictor = base_predictor->Clone();
       NEW_TENSOR(data_lod_attention);
       NEW_TENSOR(cell_init);
       NEW_TENSOR(data);

paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc

@@ -121,14 +121,6 @@ void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data) {
   }
 }
 
-void SetConfig(AnalysisConfig *cfg) {
-  cfg->SetModel(FLAGS_infer_model + "/model", FLAGS_infer_model + "/params");
-  cfg->DisableGpu();
-  cfg->SwitchSpecifyInputNames();
-  cfg->pass_builder()->TurnOnDebug();
-  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
-}
-
 void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
   DataRecord data(FLAGS_infer_data, FLAGS_batch_size);
   std::vector<PaddleTensor> input_slots;
@@ -141,15 +133,22 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
   }
 }
 
+void SetConfig(AnalysisConfig *cfg, bool use_mkldnn = false) {
+  cfg->SetModel(FLAGS_infer_model + "/model", FLAGS_infer_model + "/params");
+  cfg->DisableGpu();
+  cfg->SwitchSpecifyInputNames();
+  cfg->pass_builder()->TurnOnDebug();
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
+  if (use_mkldnn) {
+    cfg->EnableMKLDNN();
+  }
+}
+
 void profile(bool use_mkldnn = false) {
   AnalysisConfig cfg;
-  SetConfig(&cfg);
+  SetConfig(&cfg, use_mkldnn);
 
-  if (use_mkldnn) {
-    cfg.EnableMKLDNN();
-  }
   std::vector<PaddleTensor> outputs;
-
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
   TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
@@ -169,16 +168,162 @@ TEST(Analyzer_seq_pool1, compare) {
       reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
-// Check the fuse status
-TEST(Analyzer_seq_pool1, fuse_statis) {
+// Compare Deterministic result
+TEST(Analyzer_seq_pool1, compare_determine) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                       input_slots_all);
+}
+
+void analysis_fuse_statis(bool use_zerocopy) {
   AnalysisConfig cfg;
   SetConfig(&cfg);
+  cfg.SwitchUseFeedFetchOps(!use_zerocopy);
   int num_ops;
   auto predictor = CreatePaddlePredictor<AnalysisConfig>(cfg);
-  auto fuse_statis = GetFuseStatis(
-      static_cast<AnalysisPredictor *>(predictor.get()), &num_ops);
+  auto fuse_statis = GetFuseStatis(predictor.get(), &num_ops);
+  ASSERT_TRUE(fuse_statis.count("fc_fuse"));
+  ASSERT_EQ(fuse_statis.at("fc_fuse"), 10);
+  ASSERT_TRUE(fuse_statis.count("seqpool_concat_fuse"));
+  EXPECT_EQ(fuse_statis.at("seqpool_concat_fuse"), 2);
   LOG(INFO) << "num_ops: " << num_ops;
-  EXPECT_EQ(num_ops, 349);
+  EXPECT_EQ(num_ops, 195);
+}
+
+// Check the fuse status
+TEST(Analyzer_seq_pool1, fuse_statis) { analysis_fuse_statis(false); }
+
+void PrepareZeroCopyInputs(
+    const std::unique_ptr<PaddlePredictor> &predictor,
+    std::vector<std::unique_ptr<ZeroCopyTensor>> *inputs) {
+  DataRecord data(FLAGS_infer_data, FLAGS_batch_size);
+  // only feed one batch
+  const auto &one_batch = data.NextBatch();
+  inputs->clear();
+  for (size_t i = 0; i < one_batch.size(); ++i) {
+    auto &slot = one_batch[i];
+    auto tensor = predictor->GetInputTensor(slot.name + "_embed");
+    tensor->Reshape(slot.shape);
+    tensor->SetLoD({slot.lod});
+    ZeroCopyTensorAssignData<float>(tensor.get(), slot.data);
+    inputs->emplace_back(std::move(tensor));
+  }
+}
+
+// diff: similarity_norm.tmp_0, // speed: fc_4.tmp_1
+static const char out_var_name[] = "reduce_sum_0.tmp_0";
+
+// return the output values
+std::vector<float> zerocopy_profile(int repeat_times) {
+  AnalysisConfig config;
+  SetConfig(&config);
+  config.SwitchUseFeedFetchOps(false);
+  auto predictor = CreatePaddlePredictor<AnalysisConfig>(config);
+  std::vector<std::unique_ptr<ZeroCopyTensor>> inputs;
+  PrepareZeroCopyInputs(predictor, &inputs);
+  auto output_tensor = predictor->GetOutputTensor(out_var_name);
+  Timer timer;
+  LOG(INFO) << "Warm up run...";
+  timer.tic();
+  predictor->ZeroCopyRun();
+  PrintTime(FLAGS_batch_size, 1, 1, 0, timer.toc(), 1);
+  if (FLAGS_profile) {
+    paddle::platform::ResetProfiler();
+  }
+  LOG(INFO) << "Run " << repeat_times << " times...";
+  timer.tic();
+  for (int i = 0; i < repeat_times; i++) {
+    predictor->ZeroCopyRun();
+  }
+  PrintTime(FLAGS_batch_size, repeat_times, 1, 0, timer.toc() / repeat_times,
+            1);
+
+  LOG(INFO) << "ZeroCopy output: " << DescribeZeroCopyTensor(*output_tensor);
+  PaddlePlace place;
+  int output_size{0};
+  auto *pdata = output_tensor->data<float>(&place, &output_size);
+  std::vector<float> res(output_size);
+  for (int i = 0; i < output_size; ++i) {
+    res[i] = pdata[i];
+  }
+  return res;
+}
+
+TEST(Analyzer_seq_pool1, zerocopy_profile) { zerocopy_profile(FLAGS_repeat); }
+
+TEST(Analyzer_seq_pool1, zerocopy_profile_threads) {
+  AnalysisConfig config;
+  SetConfig(&config);
+  config.SwitchUseFeedFetchOps(false);
+
+  auto base_predictor = CreatePaddlePredictor<AnalysisConfig>(config);
+  double total_time_of_threads{0};
+  std::vector<std::thread> threads;
+
+  for (int tid = 0; tid < FLAGS_num_threads; tid++) {
+    threads.emplace_back([&, tid] {
+      // To ensure the thread binding correctly,
+      // please clone inside the threadpool.
+      auto predictor = base_predictor->Clone();
+      std::vector<std::unique_ptr<ZeroCopyTensor>> inputs;
+      PrepareZeroCopyInputs(predictor, &inputs);
+      auto output_tensor = predictor->GetOutputTensor(out_var_name);
+      Timer timer;
+      double total_time{0};
+
+      LOG(INFO) << "Warm up run...";
+      timer.tic();
+      predictor->ZeroCopyRun();
+      PrintTime(FLAGS_batch_size, 1, FLAGS_num_threads, tid, timer.toc(), 1);
+      if (FLAGS_profile) {
+        paddle::platform::ResetProfiler();
+      }
+      int repeat_times = FLAGS_repeat;
+      LOG(INFO) << "Run " << repeat_times << " times...";
+      timer.tic();
+
+      for (int i = 0; i < repeat_times; i++) {
+        predictor->ZeroCopyRun();
+      }
+      total_time += timer.toc();
+      total_time_of_threads += total_time;
+
+      LOG(INFO) << "thread time: " << total_time / repeat_times;
+    });
+  }
+
+  for (auto &t : threads) {
+    t.join();
+  }
+
+  LOG(INFO) << "average time: "
+            << total_time_of_threads / FLAGS_num_threads / FLAGS_repeat;
+}
+
+TEST(Analyzer_seq_pool1, zerocopy_fuse_statis) { analysis_fuse_statis(true); }
+
+TEST(Analyzer_seq_pool1, zerocopy_compare_native) {
+  AnalysisConfig config;
+  SetConfig(&config);
+  config.SwitchUseFeedFetchOps(true);
+  auto predictor = CreatePaddlePredictor<NativeConfig>(config.ToNativeConfig());
+  std::vector<PaddleTensor> native_outputs;
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  ASSERT_TRUE(predictor->Run(input_slots_all[0], &native_outputs));
+  EXPECT_EQ(native_outputs.size(), 1UL);
+
+  auto zerocopy_output = zerocopy_profile(1);
+  EXPECT_EQ(zerocopy_output.size() * sizeof(float),
+            native_outputs.front().data.length());
+  auto *native_data = static_cast<float *>(native_outputs.front().data.data());
+  for (size_t i = 0; i < zerocopy_output.size(); ++i) {
+    EXPECT_NEAR(zerocopy_output[i], native_data[i], 1e-3);
+  }
 }
 
 }  // namespace analysis

paddle/fluid/inference/tests/api/config_printer.h

@@ -62,7 +62,7 @@ std::ostream &operator<<(std::ostream &os,
                          const contrib::AnalysisConfig &config) {
   os << GenSpaces(num_spaces) << "contrib::AnalysisConfig {\n";
   num_spaces++;
-  os << *reinterpret_cast<const NativeConfig *>(&config);
+  os << config.ToNativeConfig();
   if (!config.model_from_memory()) {
     os << GenSpaces(num_spaces) << "prog_file: " << config.prog_file() << "\n";
     os << GenSpaces(num_spaces) << "param_file: " << config.params_file()

paddle/fluid/inference/tests/api/tester_helper.h

@@ -54,11 +54,13 @@ namespace paddle {
 namespace inference {
 
 void PrintConfig(const PaddlePredictor::Config *config, bool use_analysis) {
+  const auto *analysis_config =
+      reinterpret_cast<const contrib::AnalysisConfig *>(config);
   if (use_analysis) {
-    LOG(INFO) << *reinterpret_cast<const contrib::AnalysisConfig *>(config);
+    LOG(INFO) << *analysis_config;
     return;
   }
-  LOG(INFO) << *reinterpret_cast<const NativeConfig *>(config);
+  LOG(INFO) << analysis_config->ToNativeConfig();
 }
 
 void CompareResult(const std::vector<PaddleTensor> &outputs,
@@ -96,12 +98,13 @@ void CompareResult(const std::vector<PaddleTensor> &outputs,
 
 std::unique_ptr<PaddlePredictor> CreateTestPredictor(
     const PaddlePredictor::Config *config, bool use_analysis = true) {
+  const auto *analysis_config =
+      reinterpret_cast<const contrib::AnalysisConfig *>(config);
   if (use_analysis) {
-    return CreatePaddlePredictor<contrib::AnalysisConfig>(
-        *(reinterpret_cast<const contrib::AnalysisConfig *>(config)));
+    return CreatePaddlePredictor<contrib::AnalysisConfig>(*analysis_config);
   }
-  return CreatePaddlePredictor<NativeConfig>(
-      *(reinterpret_cast<const NativeConfig *>(config)));
+  auto native_config = analysis_config->ToNativeConfig();
+  return CreatePaddlePredictor<NativeConfig>(native_config);
 }
 
 size_t GetSize(const PaddleTensor &out) { return VecReduceToInt(out.shape); }
@@ -310,13 +313,12 @@ void CompareDeterministic(
   int num_times = FLAGS_repeat;
   auto predictor = CreateTestPredictor(config, FLAGS_use_analysis);
 
-  // warmup run
   std::vector<PaddleTensor> warmup_outputs, outputs;
-  predictor->Run(inputs[0], &warmup_outputs, batch_size);
-
   // run num_times to Compare Deterministic Result.
-  for (int i = 0; i < num_times; i++) {
-    for (size_t j = 0; j < inputs.size(); j++) {
+  for (size_t j = 0; j < inputs.size(); j++) {
+    // warmup run
+    predictor->Run(inputs[j], &warmup_outputs, batch_size);
+    for (int i = 0; i < num_times; i++) {
       predictor->Run(inputs[j], &outputs, batch_size);
       CompareResult(outputs, warmup_outputs);
     }
@@ -328,10 +330,7 @@ void CompareNativeAndAnalysis(
     const std::vector<std::vector<PaddleTensor>> &inputs) {
   PrintConfig(config, true);
   std::vector<PaddleTensor> native_outputs, analysis_outputs;
-  const auto *analysis_config =
-      reinterpret_cast<const contrib::AnalysisConfig *>(config);
-  auto native_config = analysis_config->ToNativeConfig();
-  TestOneThreadPrediction(&native_config, inputs, &native_outputs, false);
+  TestOneThreadPrediction(config, inputs, &native_outputs, false);
   TestOneThreadPrediction(config, inputs, &analysis_outputs, true);
   CompareResult(analysis_outputs, native_outputs);
 }

paddle/fluid/inference/tests/api/trt_models_tester.cc

@@ -99,24 +99,12 @@ void compare(std::string model_dir, bool use_tensorrt) {
     SetFakeImageInput(&inputs_all, model_dir, false, "__model__", "");
   }
 
-  std::vector<PaddleTensor> native_outputs;
-  NativeConfig native_config;
-  SetConfig<NativeConfig>(&native_config, model_dir, true, false,
-                          FLAGS_batch_size);
-  TestOneThreadPrediction(
-      reinterpret_cast<PaddlePredictor::Config*>(&native_config), inputs_all,
-      &native_outputs, false);
-
-  std::vector<PaddleTensor> analysis_outputs;
   contrib::AnalysisConfig analysis_config;
-  analysis_config.EnableUseGpu(50, 0);
   SetConfig<contrib::AnalysisConfig>(&analysis_config, model_dir, true,
                                      use_tensorrt, FLAGS_batch_size);
-  TestOneThreadPrediction(
-      reinterpret_cast<PaddlePredictor::Config*>(&analysis_config), inputs_all,
-      &analysis_outputs, true);
-
-  CompareResult(native_outputs, analysis_outputs);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config*>(&analysis_config),
+      inputs_all);
 }
 
 TEST(TensorRT_mobilenet, compare) {

paddle/fluid/inference/utils/CMakeLists.txt

@@ -2,6 +2,3 @@ cc_library(benchmark SRCS benchmark.cc DEPS enforce)
 cc_test(test_benchmark SRCS benchmark_tester.cc DEPS benchmark)
 cc_binary(visualizer SRCS visualizer.cc DEPS analysis
     paddle_pass_builder ir_pass_manager pass graph_viz_pass analysis_passes)
-if(WIN32)
-  target_link_libraries(visualizer shlwapi)
-endif(WIN32)

paddle/fluid/operators/conv_cudnn_op.cu.cc

@@ -137,7 +137,6 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
     // ------------------- cudnn conv algorithm ---------------------
     cudnnConvolutionFwdAlgo_t algo;
     auto handle = dev_ctx.cudnn_handle();
-    auto workspace_handle = dev_ctx.cudnn_workspace_handle();
 
     bool half_float = false;
 #if CUDA_VERSION >= 9000 && CUDNN_VERSION_MIN(7, 0, 1)
@@ -158,6 +157,8 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
       VLOG(5) << "NOT use cudnn_tensor_op_math";
     }
 #endif
+    Tensor cudnn_workspace;
+    void* cudnn_workspace_ptr = nullptr;
 
     auto x_dims = framework::vectorize(input->dims());
     auto f_dims = framework::vectorize(filter->dims());
@@ -180,21 +181,26 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
                 .Var(kCUDNNFwdAlgoCache)
                 ->GetMutable<AlgorithmsCache<cudnnConvolutionFwdAlgo_t>>();
       }
+      cudnn_workspace =
+          ctx.AllocateTmpTensor<int8_t, platform::CUDADeviceContext>(
+              framework::make_ddim(
+                  {static_cast<int64_t>(workspace_size_limit)}),
+              dev_ctx);
+      cudnn_workspace_ptr = static_cast<void*>(cudnn_workspace.data<int8_t>());
+
       algo = algo_cache->GetAlgorithm(
           x_dims, f_dims, strides, paddings, dilations, 0, [&]() {
             int returned_algo_count;
             std::array<cudnnConvolutionFwdAlgoPerf_t, kNUM_CUDNN_FWD_ALGS>
                 fwd_perf_stat;
-            auto cudnn_find_func = [&](void* cudnn_workspace) {
-              CUDNN_ENFORCE(
-                  platform::dynload::cudnnFindConvolutionForwardAlgorithmEx(
-                      handle, cudnn_input_desc, input_data, cudnn_filter_desc,
-                      filter_data, cudnn_conv_desc, cudnn_output_desc,
-                      output_data, kNUM_CUDNN_FWD_ALGS, &returned_algo_count,
-                      fwd_perf_stat.data(), cudnn_workspace,
-                      workspace_size_limit));
-            };
-            workspace_handle.RunFunc(cudnn_find_func, workspace_size_limit);
+
+            CUDNN_ENFORCE(
+                platform::dynload::cudnnFindConvolutionForwardAlgorithmEx(
+                    handle, cudnn_input_desc, input_data, cudnn_filter_desc,
+                    filter_data, cudnn_conv_desc, cudnn_output_desc,
+                    output_data, kNUM_CUDNN_FWD_ALGS, &returned_algo_count,
+                    fwd_perf_stat.data(), cudnn_workspace_ptr,
+                    workspace_size_limit));
 
             VLOG(3) << "Perf result: (algo: stat, time, memory)";
             for (int i = 0; i < returned_algo_count; ++i) {
@@ -219,17 +225,23 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
     PADDLE_ENFORCE_LE(workspace_size_in_bytes, workspace_size_limit,
                       "workspace_size to be allocated exceeds the limit");
 
+    // Allocate on GPU memory
+    if (!cudnn_workspace_ptr) {
+      cudnn_workspace =
+          ctx.AllocateTmpTensor<int8_t, platform::CUDADeviceContext>(
+              framework::make_ddim(
+                  {static_cast<int64_t>(workspace_size_in_bytes)}),
+              dev_ctx);
+      cudnn_workspace_ptr = static_cast<void*>(cudnn_workspace.data<int8_t>());
+    }
     // ------------------- cudnn conv forward ---------------------
     ScalingParamType<T> alpha = 1.0f, beta = 0.0f;
     for (int i = 0; i < groups; i++) {
-      auto cudnn_func = [&](void* cudnn_workspace) {
-        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward(
-            handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in,
-            cudnn_filter_desc, filter_data + i * group_offset_filter,
-            cudnn_conv_desc, algo, cudnn_workspace, workspace_size_in_bytes,
-            &beta, cudnn_output_desc, output_data + i * group_offset_out));
-      };
-      workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
+      CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward(
+          handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in,
+          cudnn_filter_desc, filter_data + i * group_offset_filter,
+          cudnn_conv_desc, algo, cudnn_workspace_ptr, workspace_size_in_bytes,
+          &beta, cudnn_output_desc, output_data + i * group_offset_out));
     }
   }
 };
@@ -297,6 +309,21 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
     cudnnFilterDescriptor_t cudnn_filter_desc = filter_desc.descriptor<T>(
         layout, framework::vectorize2int(filter->dims()), groups);
 
+#if CUDA_VERSION >= 9000 && CUDNN_VERSION_MIN(7, 0, 1)
+    // Enable Tensor Core for cudnn backward
+    if (dev_ctx.GetComputeCapability() >= 70 &&
+        std::type_index(typeid(T)) ==
+            std::type_index(typeid(platform::float16))) {
+      CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType(
+          cudnn_conv_desc, CUDNN_TENSOR_OP_MATH));
+      VLOG(5) << "use cudnn_tensor_op_math for backward";
+    } else {
+      CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType(
+          cudnn_conv_desc, CUDNN_DEFAULT_MATH));
+      VLOG(5) << "NOT use cudnn_tensor_op_math for backward";
+    }
+#endif
+
     int input_channels = input->dims()[1];
     int input_height, input_width, input_depth;
     if (input->dims().size() == 5) {
@@ -338,10 +365,20 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
       workspace_size_limit = max_user_size * 1024 * 1024;
     }
 
+    Tensor cudnn_workspace;
+    void* cudnn_workspace_ptr = nullptr;
+    if ((input_data || filter_data) && exhaustive_search) {
+      cudnn_workspace =
+          ctx.AllocateTmpTensor<int8_t, platform::CUDADeviceContext>(
+              framework::make_ddim(
+                  {static_cast<int64_t>(workspace_size_limit)}),
+              dev_ctx);
+      cudnn_workspace_ptr = static_cast<void*>(cudnn_workspace.data<int8_t>());
+    }
+
     auto x_dims = framework::vectorize(input->dims());
     auto f_dims = framework::vectorize(filter->dims());
     auto handle = dev_ctx.cudnn_handle();
-    auto workspace_handle = dev_ctx.cudnn_workspace_handle();
     if (input_grad) {
       T* input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
       if (exhaustive_search) {
@@ -359,25 +396,22 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
                   ->GetMutable<
                       AlgorithmsCache<cudnnConvolutionBwdDataAlgo_t>>();
         }
+
         data_algo = data_algo_cache->GetAlgorithm(
             x_dims, f_dims, strides, paddings, dilations, 0, [&]() {
               int returned_algo_count;
               std::array<cudnnConvolutionBwdDataAlgoPerf_t,
                          kNUM_CUDNN_BWD_DATA_ALGS>
                   data_perf_stat;
-              auto cudnn_find_bd_data_func = [&](void* cudnn_workspace) {
-                CUDNN_ENFORCE(
-                    platform::dynload::
-                        cudnnFindConvolutionBackwardDataAlgorithmEx(
-                            handle, cudnn_filter_desc, filter_data,
-                            cudnn_output_grad_desc, output_grad_data,
-                            cudnn_conv_desc, cudnn_input_desc, input_grad_data,
-                            kNUM_CUDNN_BWD_DATA_ALGS, &returned_algo_count,
-                            data_perf_stat.data(), cudnn_workspace,
-                            workspace_size_limit));
-              };
-              workspace_handle.RunFunc(cudnn_find_bd_data_func,
-                                       workspace_size_limit);
+
+              CUDNN_ENFORCE(platform::dynload::
+                                cudnnFindConvolutionBackwardDataAlgorithmEx(
+                                    handle, cudnn_filter_desc, filter_data,
+                                    cudnn_output_grad_desc, output_grad_data,
+                                    cudnn_conv_desc, cudnn_input_desc,
+                                    input_grad_data, kNUM_CUDNN_BWD_DATA_ALGS,
+                                    &returned_algo_count, data_perf_stat.data(),
+                                    cudnn_workspace_ptr, workspace_size_limit));
 
               VLOG(3) << "Perf result: (algo: stat, time, memory)";
               for (int i = 0; i < returned_algo_count; ++i) {
@@ -428,25 +462,23 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
                   ->GetMutable<
                       AlgorithmsCache<cudnnConvolutionBwdFilterAlgo_t>>();
         }
+
         filter_algo = f_algo_cache->GetAlgorithm(
             x_dims, f_dims, strides, paddings, dilations, 0, [&]() {
               int returned_algo_count;
               std::array<cudnnConvolutionBwdFilterAlgoPerf_t,
                          kNUM_CUDNN_BWD_FILTER_ALGS>
                   filter_perf_stat;
-              auto cudnn_find_bd_f_func = [&](void* cudnn_workspace) {
-                CUDNN_ENFORCE(
-                    platform::dynload::
-                        cudnnFindConvolutionBackwardFilterAlgorithmEx(
-                            handle, cudnn_input_desc, input_data,
-                            cudnn_output_grad_desc, output_grad_data,
-                            cudnn_conv_desc, cudnn_filter_desc,
-                            filter_grad_data, kNUM_CUDNN_BWD_FILTER_ALGS,
-                            &returned_algo_count, filter_perf_stat.data(),
-                            cudnn_workspace, workspace_size_limit));
-              };
-              workspace_handle.RunFunc(cudnn_find_bd_f_func,
-                                       workspace_size_limit);
+
+              CUDNN_ENFORCE(
+                  platform::dynload::
+                      cudnnFindConvolutionBackwardFilterAlgorithmEx(
+                          handle, cudnn_input_desc, input_data,
+                          cudnn_output_grad_desc, output_grad_data,
+                          cudnn_conv_desc, cudnn_filter_desc, filter_grad_data,
+                          kNUM_CUDNN_BWD_FILTER_ALGS, &returned_algo_count,
+                          filter_perf_stat.data(), cudnn_workspace_ptr,
+                          workspace_size_limit));
               return filter_perf_stat[0].algo;
             });
         VLOG(3) << "cuDNN backward filter algo " << filter_algo;
@@ -467,6 +499,16 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
       workspace_size_in_bytes = std::max(workspace_size_in_bytes, tmp_size);
     }
 
+    // ------------------- cudnn conv workspace ---------------------
+    if (!cudnn_workspace_ptr) {
+      cudnn_workspace =
+          ctx.AllocateTmpTensor<int8_t, platform::CUDADeviceContext>(
+              framework::make_ddim(
+                  {static_cast<int64_t>(workspace_size_in_bytes)}),
+              dev_ctx);
+      cudnn_workspace_ptr = static_cast<void*>(cudnn_workspace.data<int8_t>());
+    }
+
     // ------------------- cudnn conv backward data ---------------------
     ScalingParamType<T> alpha = 1.0f, beta = 0.0f;
     if (input_grad) {
@@ -474,15 +516,12 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
       // Because beta is zero, it is unnecessary to reset input_grad.
 
       for (int i = 0; i < groups; i++) {
-        auto cudnn_func = [&](void* cudnn_workspace) {
-          CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardData(
-              handle, &alpha, cudnn_filter_desc,
-              filter_data + i * group_offset_filter, cudnn_output_grad_desc,
-              output_grad_data + i * group_offset_out, cudnn_conv_desc,
-              data_algo, cudnn_workspace, workspace_size_in_bytes, &beta,
-              cudnn_input_desc, input_grad_data + i * group_offset_in));
-        };
-        workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
+        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardData(
+            handle, &alpha, cudnn_filter_desc,
+            filter_data + i * group_offset_filter, cudnn_output_grad_desc,
+            output_grad_data + i * group_offset_out, cudnn_conv_desc, data_algo,
+            cudnn_workspace_ptr, workspace_size_in_bytes, &beta,
+            cudnn_input_desc, input_grad_data + i * group_offset_in));
       }
     }
     // ------------------- cudnn conv backward filter ---------------------
@@ -490,15 +529,12 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
       T* filter_grad_data = filter_grad->mutable_data<T>(ctx.GetPlace());
       // Because beta is zero, it is unnecessary to reset filter_grad.
       for (int i = 0; i < groups; i++) {
-        auto cudnn_func = [&](void* cudnn_workspace) {
-          CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter(
-              handle, &alpha, cudnn_input_desc,
-              input_data + i * group_offset_in, cudnn_output_grad_desc,
-              output_grad_data + i * group_offset_out, cudnn_conv_desc,
-              filter_algo, cudnn_workspace, workspace_size_in_bytes, &beta,
-              cudnn_filter_desc, filter_grad_data + i * group_offset_filter));
-        };
-        workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
+        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter(
+            handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in,
+            cudnn_output_grad_desc, output_grad_data + i * group_offset_out,
+            cudnn_conv_desc, filter_algo, cudnn_workspace_ptr,
+            workspace_size_in_bytes, &beta, cudnn_filter_desc,
+            filter_grad_data + i * group_offset_filter));
       }
     }
   }

paddle/fluid/operators/conv_mkldnn_op.cc

@@ -318,10 +318,14 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
     std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
     int groups = ctx.Attr<int>("groups");
-
     bool fuse_relu = ctx.Attr<bool>("fuse_relu");
+    bool fuse_residual_conn = ctx.Attr<bool>("fuse_residual_connection");
 
     bool force_fp32_output = ctx.Attr<bool>("force_fp32_output");
+    if (fuse_residual_conn) {
+      PADDLE_ENFORCE(force_fp32_output != true,
+                     "residual fusion does not support force output with fp32");
+    }
 
     bool is_conv3d = strides.size() == 3U;
     // TODO(tpatejko): add support for dilation
@@ -355,14 +359,23 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
           framework::DataTypeTrait<float>::DataType);
     }
 
+    if (fuse_residual_conn) {
+      auto residual = ctx.Input<Tensor>("ResidualData");
+      auto residual_dt = paddle::framework::ToMKLDNNDataType(residual->type());
+      if (dst_dt != residual_dt) dst_dt = residual_dt;
+    }
+
     // Get unique name for storing MKLDNN primitives
     std::string key;
     key.reserve(MaxKeyLength);
     platform::ConvMKLDNNHandler::AppendKey(
         &key, src_tz, weights_tz, strides, paddings, dilations, groups, src_dt,
-        input->format(), dst_dt, ctx.op().Output("Output"));
+        input->format(), fuse_relu, fuse_residual_conn,
+        ctx.op().Output("Output"));
     const std::string key_conv_pd = key + "@conv_pd";
 
+    bool need_s8_to_u8 = false;
+
     std::shared_ptr<mkldnn::convolution_forward> conv_p = nullptr;
     std::shared_ptr<mkldnn::memory> src_memory_p = nullptr;
     std::shared_ptr<mkldnn::memory> user_src_memory_p = nullptr;
@@ -377,14 +390,20 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto src_key = key + "@src_mem_p";
     auto user_src_key = key + "@user_src_mem_p";
     auto src_reorder_key = key + "@src_mem_preorder_p";
+    auto residual_reorder_key = key + "@residual_data_mem_preorder_p";
+
     conv_p = std::static_pointer_cast<mkldnn::convolution_forward>(
         dev_ctx.GetBlob(prim_key));
+
     if (conv_p == nullptr || !is_test) {
       const K* filter_data = filter->data<K>();
       auto scale_in_data = ctx.Attr<float>("Scale_in");
+      auto scale_in_eltwise_data = ctx.Attr<float>("Scale_in_eltwise");
       auto scale_weights_data = ctx.Attr<std::vector<float>>("Scale_weights");
       auto scale_out_data =
           force_fp32_output ? 1.0f : ctx.Attr<float>("Scale_out");
+      float sum_scale =
+          fuse_residual_conn ? scale_out_data / scale_in_eltwise_data : 1.0f;
 
       bool is_multi_channel = scale_weights_data.size() > 1;
 
@@ -427,6 +446,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
           weights_tz, memory::data_type::s8, chosen_memory_format);
       auto dst_md =
           platform::MKLDNNMemDesc(dst_tz, dst_dt, chosen_memory_format);
+
       // create a conv primitive descriptor and save it for usage in backward
       if (bias) {
         bias_tz = paddle::framework::vectorize2int(bias->dims());
@@ -434,11 +454,13 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
                                                memory::format::x);
         conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, bias_md, dst_md,
                                        strides, paddings, mkldnn_engine,
-                                       fuse_relu, output_shift_scale, is_test);
+                                       fuse_relu, fuse_residual_conn,
+                                       output_shift_scale, sum_scale, is_test);
       } else {
-        conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides,
-                                       paddings, mkldnn_engine, fuse_relu,
-                                       output_shift_scale, is_test);
+        conv_pd =
+            ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides, paddings,
+                                 mkldnn_engine, fuse_relu, fuse_residual_conn,
+                                 output_shift_scale, sum_scale, is_test);
       }
       // Save conv_pd/src_memory/weights_memory for backward pass
       dev_ctx.SetBlob(key_conv_pd, conv_pd);
@@ -463,7 +485,41 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
           user_weights_memory_p, pipeline, is_test, true, scale_weights_data,
           mask_reorder);
 
-      if (!force_fp32_output) {
+      if (fuse_residual_conn) {
+        auto residual_param = ctx.Input<Tensor>("ResidualData");
+        PADDLE_ENFORCE_EQ(output->dims(), residual_param->dims(),
+                          "Output and elementwise parameter need to have the "
+                          "same dimension sizes");
+        auto residual_dt =
+            paddle::framework::ToMKLDNNDataType(residual_param->type());
+        if (residual_param->format() != handler->GetDstFormat()) {
+          auto residual_data_tz =
+              paddle::framework::vectorize2int(residual_param->dims());
+
+          auto user_residual_md = platform::MKLDNNMemDesc(
+              residual_data_tz, residual_dt, residual_param->format());
+
+          if (residual_dt == mkldnn::memory::data_type::u8) {
+            dst_memory_p = platform::SetDstMemory<uint8_t>(
+                ctx, output, residual_param, user_residual_md, handler,
+                &pipeline);
+          } else {
+            need_s8_to_u8 = fuse_relu;
+            dst_memory_p = platform::SetDstMemory<int8_t>(
+                ctx, output, residual_param, user_residual_md, handler,
+                &pipeline);
+          }
+        } else {
+          output->ShareDataWith(*residual_param);
+          if (residual_dt == mkldnn::memory::data_type::u8) {
+            dst_memory_p =
+                platform::SetDstMemory<uint8_t>(ctx, output, handler);
+          } else {
+            need_s8_to_u8 = fuse_relu;
+            dst_memory_p = platform::SetDstMemory<int8_t>(ctx, output, handler);
+          }
+        }
+      } else if (!force_fp32_output) {
         if (fuse_relu) {
           dst_memory_p = platform::SetDstMemory<uint8_t>(ctx, output, handler);
         } else {
@@ -476,11 +532,11 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
       // create convolution op primitive
       auto scale_bias_key = key + "@scale_bias";
       if (bias) {
-        const float* bias_data = bias->data<float>();
+        const K* bias_data = bias->data<K>();
         auto user_bias_md = platform::MKLDNNMemDesc(
-            {bias_tz}, platform::MKLDNNGetDataType<float>(), memory::format::x);
+            {bias_tz}, platform::MKLDNNGetDataType<K>(), memory::format::x);
         auto user_bias_memory_p = handler->AcquireBiasMemory(
-            user_bias_md, to_void_cast<float>(bias_data));
+            user_bias_md, to_void_cast<K>(bias_data));
         std::shared_ptr<mkldnn::memory> bias_memory_p;
         int mask_reorder = is_multi_channel ? 1 << 0 : 1;
         int count =
@@ -526,26 +582,51 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         handler.reset(new platform::ConvMKLDNNHandler(conv_pd, dev_ctx,
                                                       mkldnn_engine, key));
       }
-      if (!force_fp32_output) {
+
+      if (fuse_residual_conn) {
+        auto residual_param = ctx.Input<Tensor>("ResidualData");
+        auto residual_dt =
+            paddle::framework::ToMKLDNNDataType(residual_param->type());
+        output->ShareDataWith(*residual_param);
+        if (residual_dt == mkldnn::memory::data_type::u8) {
+          platform::SetDstMemoryHandler<uint8_t>(ctx, output, handler,
+                                                 &dst_memory_p);
+        } else {
+          platform::SetDstMemoryHandler<int8_t>(ctx, output, handler,
+                                                &dst_memory_p);
+        }
+      } else if (!force_fp32_output) {
         if (fuse_relu) {
-          dst_memory_p =
-              platform::SetDstMemoryHandler<uint8_t>(ctx, output, handler);
+          platform::SetDstMemoryHandler<uint8_t>(ctx, output, handler,
+                                                 &dst_memory_p);
         } else {
-          dst_memory_p =
-              platform::SetDstMemoryHandler<int8_t>(ctx, output, handler);
+          platform::SetDstMemoryHandler<int8_t>(ctx, output, handler,
+                                                &dst_memory_p);
         }
       } else {
-        dst_memory_p =
-            platform::SetDstMemoryHandler<float>(ctx, output, handler);
+        platform::SetDstMemoryHandler<float>(ctx, output, handler,
+                                             &dst_memory_p);
       }
+
       if (src_memory_reorder_p) {
         pipeline.push_back(*src_memory_reorder_p);
       }
+
+      auto residual_reorder_p = std::static_pointer_cast<mkldnn::memory>(
+          dev_ctx.GetBlob(residual_reorder_key));
+      if (residual_reorder_p) {
+        pipeline.push_back(*residual_reorder_p);
+      }
+
       pipeline.push_back(*conv_p);
     }
     // push primitive to stream and wait until it's executed
     stream(stream::kind::eager).submit(pipeline).wait();
 
+    if (need_s8_to_u8) {
+      output->mutable_data<uint8_t>(ctx.GetPlace());
+    }
+
     output->set_layout(DataLayout::kMKLDNN);
     output->set_format(GetMKLDNNFormat(*dst_memory_p));
   }
@@ -577,11 +658,15 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
   }
 
   mkldnn::primitive_attr CreatePostOps(
-      bool fuse_relu, const std::vector<float> output_shift_scale) const {
+      bool fuse_relu, bool fuse_residual_conn,
+      const std::vector<float> output_shift_scale, float sum_scale) const {
     mkldnn::primitive_attr conv_attr;
     mkldnn::post_ops post_operations;
     int mask = output_shift_scale.size() > 1 ? 1 << 1 : 0;
     conv_attr.set_output_scales(mask, output_shift_scale);
+    if (fuse_residual_conn) {
+      post_operations.append_sum(sum_scale);
+    }
     if (fuse_relu) {
       constexpr float scale = 1.0f;
       constexpr float negative_slope = 0.0f;
@@ -622,8 +707,9 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
                        const memory::desc& dst, const std::vector<int>& strides,
                        const std::vector<int>& paddings,
                        const mkldnn::engine& engine, const bool fuse_relu,
+                       const bool fuse_residual_conn,
                        const std::vector<float> output_shift_scale,
-                       bool is_test) const {
+                       const float sum_scale, bool is_test) const {
     memory::dims stride_dims = {strides[0], strides[1]};
     memory::dims padding_dims = {paddings[0], paddings[1]};
 
@@ -634,8 +720,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         propagation, mkldnn::convolution_direct, src, weights, dst, stride_dims,
         padding_dims, padding_dims, mkldnn::padding_kind::zero);
 
-    mkldnn::primitive_attr conv_attr =
-        CreatePostOps(fuse_relu, output_shift_scale);
+    mkldnn::primitive_attr conv_attr = CreatePostOps(
+        fuse_relu, fuse_residual_conn, output_shift_scale, sum_scale);
 
     auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
         conv_desc, conv_attr, engine);
@@ -675,8 +761,9 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
                        const std::vector<int>& strides,
                        const std::vector<int>& paddings,
                        const mkldnn::engine& engine, const bool fuse_relu,
+                       const bool fuse_residual_conn,
                        const std::vector<float> output_shift_scale,
-                       bool is_test) const {
+                       const float sum_scale, bool is_test) const {
     memory::dims stride_dims = {strides[0], strides[1]};
     memory::dims padding_dims = {paddings[0], paddings[1]};
 
@@ -687,8 +774,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         propagation, mkldnn::convolution_direct, src, weights, bias, dst,
         stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero);
 
-    mkldnn::primitive_attr conv_attr =
-        CreatePostOps(fuse_relu, output_shift_scale);
+    mkldnn::primitive_attr conv_attr = CreatePostOps(
+        fuse_relu, fuse_residual_conn, output_shift_scale, sum_scale);
 
     auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
         conv_desc, conv_attr, engine);
@@ -891,7 +978,7 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
       input_grad->set_format(GetMKLDNNFormat(*diff_src_memory_p));
     }
     stream(stream::kind::eager).submit(pipeline).wait();
-  }  // Compute()
+  }
 };
 
 }  // namespace operators

paddle/fluid/operators/data_norm_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/data_norm_op.h"
+#include <string>
+#include "paddle/fluid/framework/data_layout.h"
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+using DataLayout = framework::DataLayout;
+
+template <typename T>
+using EigenArrayMap =
+    Eigen::Map<Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
+template <typename T>
+using ConstEigenArrayMap =
+    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, Eigen::Dynamic>>;
+template <typename T>
+using EigenVectorArrayMap = Eigen::Map<Eigen::Array<T, Eigen::Dynamic, 1>>;
+template <typename T>
+using ConstEigenVectorArrayMap =
+    Eigen::Map<const Eigen::Array<T, Eigen::Dynamic, 1>>;
+
+class DataNormOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "");
+    PADDLE_ENFORCE(ctx->HasInput("BatchSize"), "");
+    PADDLE_ENFORCE(ctx->HasInput("BatchSum"), "");
+    PADDLE_ENFORCE(ctx->HasInput("BatchSquareSum"), "");
+    PADDLE_ENFORCE(ctx->HasOutput("Means"), "");
+    PADDLE_ENFORCE(ctx->HasOutput("Scales"), "");
+    PADDLE_ENFORCE(ctx->HasOutput("Y"), "");
+
+    const auto x_dims = ctx->GetInputDim("X");
+    const DataLayout data_layout = framework::StringToDataLayout(
+        ctx->Attrs().Get<std::string>("data_layout"));
+
+    PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
+                   "Input X must have 2 to 5 dimensions.");
+
+    const int64_t C =
+        (data_layout == DataLayout::kNCHW ? x_dims[1]
+                                          : x_dims[x_dims.size() - 1]);
+
+    PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSize").size(), 1UL);
+    PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSum").size(), 1UL);
+    PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSquareSum").size(), 1UL);
+    PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSize")[0], C);
+    PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSum")[0], C);
+    PADDLE_ENFORCE_EQ(ctx->GetInputDim("BatchSquareSum")[0], C);
+
+    ctx->SetOutputDim("Y", x_dims);
+    ctx->SetOutputDim("Means", {C});
+    ctx->SetOutputDim("Scales", {C});
+    ctx->ShareLoD("X", "Y");
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    auto input_data_type = ctx.Input<Tensor>("X")->type();
+    // By default, the type of the scale, bias, mean,
+    // and var tensors should both be float. (For float or float16 input tensor)
+    // or double (For double input tensor).
+    auto dn_param_type = framework::proto::VarType::FP32;
+    if (input_data_type == framework::proto::VarType::FP64) {
+      dn_param_type = framework::proto::VarType::FP64;
+    }
+    PADDLE_ENFORCE_EQ(dn_param_type, ctx.Input<Tensor>("BatchSize")->type(),
+                      "BatchSize input should be of float type");
+    PADDLE_ENFORCE_EQ(dn_param_type, ctx.Input<Tensor>("BatchSum")->type(),
+                      "BatchSum input should be of float type");
+    PADDLE_ENFORCE_EQ(dn_param_type,
+                      ctx.Input<Tensor>("BatchSquareSum")->type(),
+                      "BatchSquareSum input should be of float type");
+
+    // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
+    framework::LibraryType library = framework::LibraryType::kPlain;
+    framework::DataLayout layout = framework::DataLayout::kAnyLayout;
+#ifdef PADDLE_WITH_MKLDNN
+    if (library == framework::LibraryType::kPlain &&
+        platform::CanMKLDNNBeUsed(ctx)) {
+      library = framework::LibraryType::kMKLDNN;
+      layout = framework::DataLayout::kMKLDNN;
+    }
+#endif
+
+    return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout,
+                                   library);
+  }
+};
+
+class DataNormOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    // AddAttr<bool>("is_test", "").SetDefault(false);
+    AddAttr<float>("epsilon", "")
+        .SetDefault(1e-4)
+        .AddCustomChecker([](const float &epsilon) {
+          PADDLE_ENFORCE(epsilon >= 0.0f && epsilon <= 0.001f,
+                         "'epsilon' should be between 0.0 and 0.001.");
+        });
+    AddAttr<std::string>("data_layout", "").SetDefault("NCHW");
+    AddInput("X", "The input tensor");
+    AddInput("BatchSize",
+             "BatchSize is a 1-dimensional tensor of size C "
+             "that is applied to the output");
+    AddInput("BatchSum",
+             "BatchSum is a 1-dimensional tensor of size C "
+             "that is applied to the output");
+    AddInput("BatchSquareSum",
+             "The global BatchSquareSum (for training) or "
+             "estimated BatchSquareSum (for testing)");
+    AddOutput("Y", "result after normalization");
+    AddOutput("Means",
+              "Mean of the history data batch, "
+              "will apply to output when training")
+        .AsIntermediate();
+    AddOutput("Scales",
+              "Scales of the history data batch, "
+              "will apply to output when training")
+        .AsIntermediate();
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
+    AddComment(R"DOC(
+Data Normalization.
+
+Can be used as a normalizer function for data
+The required data format for this layer is one of the following:
+1. NHWC `[batch, in_height, in_width, in_channels]`
+2. NCHW `[batch, in_channels, in_height, in_width]`
+
+)DOC");
+  }
+};
+
+template <typename T>
+class DataNormKernel<platform::CPUDeviceContext, T>
+    : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    // const bool is_test = ctx.Attr<bool>("is_test");
+    const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
+    const DataLayout data_layout =
+        framework::StringToDataLayout(data_layout_str);
+
+    const auto *x = ctx.Input<Tensor>("X");
+    const auto &x_dims = x->dims();
+    PADDLE_ENFORCE(x_dims.size() == 2, "The Input dim size should be 2");
+    const int N = x_dims[0];
+    const int C =
+        (data_layout == DataLayout::kNCHW ? x_dims[1]
+                                          : x_dims[x_dims.size() - 1]);
+    auto *y = ctx.Output<Tensor>("Y");
+    auto *mean_out = ctx.Output<Tensor>("Means");
+    auto *scales = ctx.Output<Tensor>("Scales");
+
+    // alloc memory
+    y->mutable_data<T>(ctx.GetPlace());
+
+    Eigen::Array<T, Eigen::Dynamic, 1> inv_std(C);
+    ConstEigenVectorArrayMap<T> b_size_arr(
+        ctx.Input<Tensor>("BatchSize")->data<T>(), C);
+    ConstEigenVectorArrayMap<T> b_sum_arr(
+        ctx.Input<Tensor>("BatchSum")->data<T>(), C);
+    ConstEigenVectorArrayMap<T> b_square_sum_arr(
+        ctx.Input<Tensor>("BatchSquareSum")->data<T>(), C);
+    EigenVectorArrayMap<T> means_arr(mean_out->mutable_data<T>(ctx.GetPlace()),
+                                     C);
+    EigenVectorArrayMap<T> scales_arr(scales->mutable_data<T>(ctx.GetPlace()),
+                                      C);
+    means_arr = b_sum_arr / b_size_arr;
+    scales_arr = (b_size_arr / b_square_sum_arr).sqrt();
+
+    switch (data_layout) {
+      case DataLayout::kNCHW:  // because it's two dimensions, so make no
+                               // difference
+      case DataLayout::kNHWC: {
+        EigenArrayMap<T>(y->mutable_data<T>(ctx.GetPlace()), C, N) =
+            (ConstEigenArrayMap<T>(x->data<T>(), C, N).colwise() - means_arr)
+                .colwise() *
+            scales_arr;
+        break;
+      }
+      default:
+        PADDLE_THROW("Unknown storage order: %d", data_layout);
+    }
+  }
+};
+
+class DataNormGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    // check input
+    PADDLE_ENFORCE(ctx->HasInput("X"));
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")), "");
+    PADDLE_ENFORCE(ctx->HasInput("BatchSize"), "");
+    PADDLE_ENFORCE(ctx->HasInput("BatchSum"), "");
+    PADDLE_ENFORCE(ctx->HasInput("BatchSquareSum"), "");
+    PADDLE_ENFORCE(ctx->HasInput("Means"), "");
+    PADDLE_ENFORCE(ctx->HasInput("Scales"), "");
+
+    // check output
+    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), "");
+    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("BatchSize")), "");
+    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("BatchSum")), "");
+    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("BatchSquareSum")),
+                   "");
+
+    const auto x_dims = ctx->GetInputDim("X");
+    const DataLayout data_layout = framework::StringToDataLayout(
+        ctx->Attrs().Get<std::string>("data_layout"));
+    const int C =
+        (data_layout == DataLayout::kNCHW ? x_dims[1]
+                                          : x_dims[x_dims.size() - 1]);
+
+    ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
+    ctx->SetOutputDim(framework::GradVarName("BatchSize"), {C});
+    ctx->SetOutputDim(framework::GradVarName("BatchSum"), {C});
+    ctx->SetOutputDim(framework::GradVarName("BatchSquareSum"), {C});
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    const auto *var = ctx.InputVar(framework::GradVarName("Y"));
+    if (var == nullptr) {
+      PADDLE_THROW("can't find Y@GRAD");
+    }
+    const Tensor *t = nullptr;
+    if (var->IsType<Tensor>()) {
+      t = &var->Get<Tensor>();
+    } else if (var->IsType<LoDTensor>()) {
+      t = &var->Get<LoDTensor>();
+    }
+    if (t == nullptr) {
+      PADDLE_THROW("can't find Y@GRAD");
+    }
+
+    // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
+    framework::LibraryType library = framework::LibraryType::kPlain;
+    framework::DataLayout layout = framework::DataLayout::kAnyLayout;
+
+#ifdef PADDLE_WITH_MKLDNN
+    if (library == framework::LibraryType::kPlain &&
+        platform::CanMKLDNNBeUsed(ctx)) {
+      library = framework::LibraryType::kMKLDNN;
+      layout = framework::DataLayout::kMKLDNN;
+    }
+#endif
+
+    return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
+                                   ctx.GetPlace(), layout, library);
+  }
+};
+
+template <typename T>
+class DataNormGradKernel<platform::CPUDeviceContext, T>
+    : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    const auto *x = ctx.Input<Tensor>("X");
+    const auto *d_y = ctx.Input<Tensor>(framework::GradVarName("Y"));
+    const auto *batch_size = ctx.Input<Tensor>("BatchSize");
+    const auto *batch_sum = ctx.Input<Tensor>("BatchSum");
+    const auto *batch_square_sum = ctx.Input<Tensor>("BatchSquareSum");
+    const auto *scales = ctx.Input<Tensor>("Scales");
+    const auto *means = ctx.Input<Tensor>("Means");
+
+    const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
+    const DataLayout data_layout =
+        framework::StringToDataLayout(data_layout_str);
+
+    // Get the size for each dimension.
+    // NCHW [batch_size, in_channels, in_height, in_width]
+    const auto &x_dims = x->dims();
+    PADDLE_ENFORCE(x_dims.size() == 2, "The Input dim size should be 2");
+    const int N = x_dims[0];
+    const int C =
+        (data_layout == DataLayout::kNCHW ? x_dims[1]
+                                          : x_dims[x_dims.size() - 1]);
+
+    // init output
+    auto *d_x = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto *d_batch_size =
+        ctx.Output<Tensor>(framework::GradVarName("BatchSize"));
+    auto *d_batch_sum = ctx.Output<Tensor>(framework::GradVarName("BatchSum"));
+    auto *d_batch_square_sum =
+        ctx.Output<Tensor>(framework::GradVarName("BatchSquareSum"));
+
+    EigenVectorArrayMap<T> d_batch_size_arr(
+        d_batch_size->mutable_data<T>(ctx.GetPlace()), C);
+    EigenVectorArrayMap<T> d_batch_sum_arr(
+        d_batch_sum->mutable_data<T>(ctx.GetPlace()), C);
+    EigenVectorArrayMap<T> d_batch_square_sum_arr(
+        d_batch_square_sum->mutable_data<T>(ctx.GetPlace()), C);
+
+    d_batch_size_arr.setZero();
+    d_batch_sum_arr.setZero();
+    d_batch_square_sum_arr.setZero();
+
+    const float epsilon = ctx.Attr<float>("epsilon");
+    switch (
+        data_layout) {  // because it's two dimensions, so make no difference
+      case DataLayout::kNCHW:
+      case DataLayout::kNHWC: {
+        ConstEigenVectorArrayMap<T> scales_arr(scales->data<T>(), C);
+        ConstEigenVectorArrayMap<T> means_arr(means->data<T>(), C);
+        ConstEigenArrayMap<T> x_arr(x->data<T>(), C, N);
+        ConstEigenArrayMap<T> d_y_arr(d_y->data<T>(), C, N);
+        EigenArrayMap<T> d_x_arr(d_x->mutable_data<T>(ctx.GetPlace()), C, N);
+        d_x_arr.setZero();
+        for (int nc = 0; nc < N; ++nc) {
+          d_x_arr.col(nc) = d_y_arr.col(nc) * scales_arr;
+        }
+
+        // calculate data sum and squre sum
+        ConstEigenVectorArrayMap<T> batch_size_arr(batch_size->data<T>(), C);
+        ConstEigenVectorArrayMap<T> batch_sum_arr(batch_sum->data<T>(), C);
+        ConstEigenVectorArrayMap<T> batch_square_sum_arr(
+            batch_square_sum->data<T>(), C);
+        Eigen::Array<T, Eigen::Dynamic, 1> sample_sum(C);
+        Eigen::Array<T, Eigen::Dynamic, 1> sample_square_sum(C);
+        // calculate data sample sum and square sum
+        sample_sum.setZero();
+        sample_square_sum.setZero();
+        for (int nc = 0; nc < N; ++nc) {
+          sample_sum += x_arr.col(nc);
+          sample_square_sum += (x_arr.col(nc) - means_arr).square();
+        }
+        // calculate gradient
+        d_batch_size_arr.setConstant(N);
+        d_batch_sum_arr = sample_sum;
+        d_batch_square_sum_arr = sample_square_sum + d_batch_size_arr * epsilon;
+        break;
+      }
+      default:
+        PADDLE_THROW("Unknown storage order: %s", data_layout_str);
+    }
+  }
+};
+
+class DataNormGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto *op = new framework::OpDesc();
+    op->SetType("data_norm_grad");
+    op->SetInput("X", Input("X"));
+    op->SetInput(framework::GradVarName("Y"), OutputGrad("Y"));
+
+    op->SetInput("BatchSize", Input("BatchSize"));
+    op->SetInput("BatchSum", Input("BatchSum"));
+    op->SetInput("BatchSquareSum", Input("BatchSquareSum"));
+    op->SetInput("Scales", Output("Scales"));
+    op->SetInput("Means", Output("Means"));
+
+    op->SetAttrMap(Attrs());
+
+    op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
+    op->SetOutput(framework::GradVarName("BatchSize"), InputGrad("BatchSize"));
+    op->SetOutput(framework::GradVarName("BatchSum"), InputGrad("BatchSum"));
+    op->SetOutput(framework::GradVarName("BatchSquareSum"),
+                  InputGrad("BatchSquareSum"));
+
+    return std::unique_ptr<framework::OpDesc>(op);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(data_norm, ops::DataNormOp, ops::DataNormOpMaker,
+                  ops::DataNormGradMaker);
+REGISTER_OPERATOR(data_norm_grad, ops::DataNormGradOp);
+
+REGISTER_OP_CPU_KERNEL(
+    data_norm, ops::DataNormKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::DataNormKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    data_norm_grad,
+    ops::DataNormGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::DataNormGradKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/data_norm_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 "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class DataNormKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override;
+};
+
+template <typename DeviceContext, typename T>
+class DataNormGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override;
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/elementwise/elementwise_sub_op.cu

@@ -12,18 +12,23 @@ 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/elementwise/elementwise_sub_op.h"
+#include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
 
 REGISTER_OP_CUDA_KERNEL(
     elementwise_sub,
     ops::ElementwiseSubKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseSubKernel<paddle::platform::CUDADeviceContext,
+                              paddle::platform::float16>,
     ops::ElementwiseSubKernel<paddle::platform::CUDADeviceContext, double>,
     ops::ElementwiseSubKernel<paddle::platform::CUDADeviceContext, int>,
     ops::ElementwiseSubKernel<paddle::platform::CUDADeviceContext, int64_t>);
 REGISTER_OP_CUDA_KERNEL(
     elementwise_sub_grad,
     ops::ElementwiseSubGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseSubGradKernel<paddle::platform::CUDADeviceContext,
+                                  paddle::platform::float16>,
     ops::ElementwiseSubGradKernel<paddle::platform::CUDADeviceContext, double>,
     ops::ElementwiseSubGradKernel<paddle::platform::CUDADeviceContext, int>,
     ops::ElementwiseSubGradKernel<paddle::platform::CUDADeviceContext,

paddle/fluid/operators/expand_op.cc

@@ -48,7 +48,7 @@ class ExpandOp : public framework::OperatorWithKernel {
     }
 
     // set the first dim to -1 in compile time
-    if (!ctx->IsRuntime()) {
+    if (!ctx->IsRuntime() && x_dims[0] < 0) {
       out_shape[0] = x_dims[0];
     }
 
@@ -115,7 +115,7 @@ class ExpandGradOp : public framework::OperatorWithKernel {
     auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
 
     size_t start_pos = 0u;
-    if (!ctx->IsRuntime()) {
+    if (!ctx->IsRuntime() && x_dims[0] < 0) {
       PADDLE_ENFORCE_EQ(
           x_dims[0], out_dims[0],
           "The first dimension size of Input(Out@GRAD) should be "

paddle/fluid/operators/fused/fusion_seqpool_concat_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/fused/fusion_seqpool_concat_op.h"
+#include <string>
+#include <vector>
+#include "paddle/fluid/operators/jit/kernels.h"
+
+namespace paddle {
+namespace operators {
+
+void FusionSeqPoolConcatOp::InferShape(
+    framework::InferShapeContext* ctx) const {
+  PADDLE_ENFORCE_GE(ctx->Inputs("X").size(), 1UL,
+                    "Inputs(X) of FusionSeqPoolConcatOp should not be empty.");
+  PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                 "Output(Out) of FusionSeqPoolConcatOp should not be null.");
+  int axis = ctx->Attrs().Get<int>("axis");
+  PADDLE_ENFORCE_EQ(axis, 1,
+                    "FusionSeqPoolConcatOp only supports concat axis=1 yet.");
+
+  auto ins_dims = ctx->GetInputsDim("X");
+  const size_t n = ins_dims.size();
+  PADDLE_ENFORCE_GT(n, 0UL, "Input tensors count should > 0.");
+  if (n == 1) {
+    LOG(WARNING) << "Only have one input, may waste memory";
+  }
+
+  // The output height should be confirmed in Compute,
+  // since input lod is not accessible here.
+  PADDLE_ENFORCE_EQ(ins_dims[0].size(), 2UL,
+                    "The dims size of first input should be 2.");
+  ctx->SetOutputDim("Out", {-1, ins_dims[0][axis] * static_cast<int>(n)});
+}
+
+framework::OpKernelType FusionSeqPoolConcatOp::GetExpectedKernelType(
+    const framework::ExecutionContext& ctx) const {
+  return framework::OpKernelType(
+      framework::GetDataTypeOfVar(ctx.MultiInputVar("X")[0]), ctx.GetPlace());
+}
+
+void FusionSeqPoolConcatOpMaker::Make() {
+  AddInput("X", "(LoDTensor) Input tensors of this operator.").AsDuplicable();
+  AddOutput("Out", "(LoDTensor) Output tensor of concat operator.");
+  AddAttr<std::string>("pooltype",
+                       "(string, default 'SUM') some of the pooling "
+                       "pooltype of SequencePoolOp.")
+      .SetDefault("SUM")
+      .InEnum({"AVERAGE", "SUM", "SQRT"});
+  AddAttr<int>("axis",
+               "The axis along which the input tensors will be concatenated. "
+               "Only supports concat axis=1 yet.")
+      .SetDefault(1);
+  AddComment(R"DOC(
+Fusion Sequence Pool of pooltype(sum, average and sqrt) and Concat Operator.
+)DOC");
+}
+
+template <typename T>
+class FusionSeqPoolConcatKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto ins = ctx.MultiInput<LoDTensor>("X");
+    auto* out = ctx.Output<LoDTensor>("Out");
+    std::string pooltype = ctx.Attr<std::string>("pooltype");
+    auto x0_lod = ins[0]->lod();
+    auto x0_dims = ins[0]->dims();
+    auto y_dims = out->dims();
+    size_t bs = x0_lod[0].size() - 1;
+    out->Resize({static_cast<int64_t>(bs), y_dims[1]});
+    framework::LoD y_lod(1);
+    y_lod[0].resize(bs + 1);
+    for (size_t i = 0; i <= bs; ++i) {
+      y_lod[0][i] = i;
+    }
+    out->set_lod(y_lod);
+    auto place = ctx.GetPlace();
+    T* y_data = out->mutable_data<T>(place);
+
+    int w = ins[0]->numel() / x0_dims[0];
+    PADDLE_ENFORCE_EQ(y_dims[1] % w, 0,
+                      "The output of dims[1] should be dividable of w");
+    jit::seq_pool_attr_t attr(w, jit::SeqPoolType::kSum);
+    if (pooltype == "AVERAGE") {
+      attr.type = jit::SeqPoolType::kAvg;
+    } else if (pooltype == "SQRT") {
+      attr.type = jit::SeqPoolType::kSqrt;
+    }
+    auto seqpool =
+        jit::Get<jit::kSeqPool, jit::SeqPoolTuples<T>, platform::CPUPlace>(
+            attr);
+    size_t n = ins.size();
+    size_t dst_step_size = n * w;
+    for (size_t i = 0; i < n; ++i) {
+      auto x_dims = ins[i]->dims();
+      auto x_lod = ins[i]->lod()[0];
+      const T* src = ins[i]->data<T>();
+      T* dst = y_data + i * w;
+      PADDLE_ENFORCE_EQ(static_cast<int>(ins[i]->numel() / x_dims[0]), w,
+                        "Width of all inputs should be equal.");
+      PADDLE_ENFORCE_EQ(x_lod.size(), bs + 1,
+                        "Batchsize of all inputs should be equal.");
+      for (size_t j = 0; j < bs; ++j) {
+        attr.h = static_cast<int>(x_lod[j + 1] - x_lod[j]);
+        seqpool(src, dst, &attr);
+        dst += dst_step_size;
+        src += attr.h * attr.w;
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(fusion_seqpool_concat, ops::FusionSeqPoolConcatOp,
+                  ops::FusionSeqPoolConcatOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+
+REGISTER_OP_CPU_KERNEL(fusion_seqpool_concat,
+                       ops::FusionSeqPoolConcatKernel<float>,
+                       ops::FusionSeqPoolConcatKernel<double>);

paddle/fluid/operators/fused/fusion_seqpool_concat_op.h

+/* 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. */
+
+#pragma once
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using LoDTensor = framework::LoDTensor;
+using Tensor = framework::Tensor;
+
+class FusionSeqPoolConcatOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override;
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override;
+};
+
+class FusionSeqPoolConcatOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override;
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/benchmark.cc

@@ -52,11 +52,11 @@ struct BenchFunc {
     for (int i = 0; i < FLAGS_burning; ++i) {
       tgt(args...);
     }
-    auto start = paddle::platform::PosixInNsec() / 1e-3;
+    auto start = paddle::platform::PosixInNsec() * 1e-3;
     for (int i = 0; i < FLAGS_repeat; ++i) {
       tgt(args...);
     }
-    auto end = paddle::platform::PosixInNsec() / 1e-3;
+    auto end = paddle::platform::PosixInNsec() * 1e-3;
     return static_cast<double>(end - start) / FLAGS_repeat;
   }
 };

paddle/fluid/operators/math/selected_rows_functor.cc

@@ -195,6 +195,10 @@ struct SelectedRowsAddToTensor<platform::CPUDeviceContext, T> {
   void operator()(const platform::CPUDeviceContext& context,
                   const framework::SelectedRows& input1,
                   framework::Tensor* input2) {
+    if (UNLIKELY(input1.rows().size() == 0)) {
+      LOG(WARNING) << "input selected rows is empty!";
+      return;
+    }
     auto in1_height = input1.height();
     auto in2_dims = input2->dims();
     PADDLE_ENFORCE_EQ(in1_height, in2_dims[0]);

paddle/fluid/operators/math/softmax.h

@@ -49,6 +49,7 @@ class SoftmaxGradCUDNNFunctor {
                   const framework::Tensor* Y, const framework::Tensor* y_grad,
                   framework::Tensor* x_grad);
 };
+
 #endif
 
 }  // namespace math

paddle/fluid/operators/py_func_op.cc

@@ -13,10 +13,10 @@
 // limitations under the License.
 
 #include "paddle/fluid/operators/py_func_op.h"
+
 #include <set>
 #include <string>
 #include <vector>
-#include "Python.h"
 #include "paddle/fluid/framework/op_registry.h"
 
 namespace paddle {

paddle/fluid/operators/py_func_op.h

@@ -13,8 +13,7 @@
 // limitations under the License.
 
 #pragma once
-
-#include "pybind11/pybind11.h"
+#include "paddle/fluid/framework/python_headers.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/softmax_with_cross_entropy_op.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.
@@ -58,12 +55,24 @@ __global__ void SoftCrossEntropyGradientKernel(T* logit_grad,
 
 }  // namespace
 
-static __device__ __forceinline__ float real_exp(float x) { return expf(x); }
-static __device__ __forceinline__ double real_exp(double x) { return exp(x); }
-static __device__ __forceinline__ float real_log(float x) {
+static __device__ __forceinline__ platform::float16 exp_on_device(
+    platform::float16 x) {
+  return ::Eigen::numext::exp(x);
+}
+static __device__ __forceinline__ float exp_on_device(float x) {
+  return expf(x);
+}
+static __device__ __forceinline__ double exp_on_device(double x) {
+  return exp(x);
+}
+static __device__ __forceinline__ platform::float16 log_on_device(
+    platform::float16 x) {
+  return math::TolerableValue<platform::float16>()(::Eigen::numext::log(x));
+}
+static __device__ __forceinline__ float log_on_device(float x) {
   return math::TolerableValue<float>()(logf(x));
 }
-static __device__ __forceinline__ double real_log(double x) {
+static __device__ __forceinline__ double log_on_device(double x) {
   return math::TolerableValue<double>()(log(x));
 }
 
@@ -72,25 +81,20 @@ static __device__ __forceinline__ double real_log(double x) {
 /*
   Supposing the x is `logits` and y is `labels`, the equations are as
 followings:
-
   cross\_entropy_i = \sum_{j}[- y_i_j * log({e^{x_i_j}/\sum_{j}e^{x_i_j}})]
         = \sum_{j}[- y_i_j * log({e^{x_i_j - max_i}/\sum_{j}e^{x_i_j-max_i}})]
         = \sum_{j}[-y_i_j * (x_i_j - max_i - log\sum_{j}e^{x_i_j - max_i})]
         = \sum_{j}[-y_i_j * (x_i_j - max_i - logDiffMaxSum_i)]
         = \sum_{j}(-y_i_j * tmp_i_j)
-
   softmax_i_j = e^{tmp_i_j}
-
 where:
   max_i = \max_{j}{x_i_j}
   logDiffMaxSum_i = log\sum_{j}e^{x_i_j - max_i}
   tmp_i_j = x_i_j - max_i - logDiffMaxSum_i
-
 Therefore, the calculation can be separated into 3 steps:
 Step 1: row-wise operation to calculate max_i
 Step 2: row-wise operation to calculate logDiffMaxSum_i
 Step 3: caculate tmp_i_j, and finally get softmax_i_j and cross\_entropy_i
-
 To save memory, we can share memory among max_i, logDiffMaxSum_i and
 cross\_entropy_i.
 In this way, the 3 steps should be changed to:
@@ -134,7 +138,8 @@ static __global__ void RowReductionForMax(const T* logits_data, T* max_data,
   cur_max = BlockReduce<T, BlockDim>(temp_storage).Reduce(cur_max, cub::Max());
 
   if (threadIdx.x == 0) {
-    max_data[blockIdx.x] = cur_max < -64 ? -64 : cur_max;
+    max_data[blockIdx.x] =
+        cur_max < static_cast<T>(-64) ? static_cast<T>(-64) : cur_max;
   }
 }
 
@@ -151,17 +156,17 @@ static __global__ void RowReductionForDiffMaxSum(const T* logits_data,
   auto block_max = max_data[blockIdx.x];
 
   softmax[beg_idx] = logits_data[beg_idx] - block_max;
-  T diff_max_sum = real_exp(softmax[beg_idx]);
+  T diff_max_sum = exp_on_device(softmax[beg_idx]);
   auto idx = beg_idx + BlockDim;
   while (idx < end_idx) {
     softmax[idx] = logits_data[idx] - block_max;
-    diff_max_sum += real_exp(softmax[idx]);
+    diff_max_sum += exp_on_device(softmax[idx]);
     idx += BlockDim;
   }
 
   diff_max_sum =
       BlockReduce<T, BlockDim>(temp_storage).Reduce(diff_max_sum, cub::Sum());
-  if (threadIdx.x == 0) max_data[blockIdx.x] = real_log(diff_max_sum);
+  if (threadIdx.x == 0) max_data[blockIdx.x] = log_on_device(diff_max_sum);
 
   if (!CalculateLogSoftmax) return;
   __syncthreads();
@@ -188,12 +193,12 @@ static __global__ void RowReductionForSoftmaxAndCrossEntropy(
   // log_diff_max_sum shares memory with loss
   auto block_log_diff_max_sum = loss_data[blockIdx.x];
   auto tmp = softmax[beg_idx] - block_log_diff_max_sum;
-  softmax[beg_idx] = real_exp(tmp);
+  softmax[beg_idx] = exp_on_device(tmp);
   auto loss = -labels_data[beg_idx] * tmp;
   beg_idx += BlockDim;
   while (beg_idx < end_idx) {
     tmp = softmax[beg_idx] - block_log_diff_max_sum;
-    softmax[beg_idx] = real_exp(tmp);
+    softmax[beg_idx] = exp_on_device(tmp);
     loss -= (labels_data[beg_idx] * tmp);
     beg_idx += BlockDim;
   }
@@ -218,10 +223,10 @@ struct HardLabelSoftmaxWithCrossEntropyFunctor {
     auto row_idx = idx / feature_size_;
     auto col_idx = idx % feature_size_;
     if (col_idx != labels_[row_idx]) {
-      log_softmax_[idx] = real_exp(log_softmax_[idx]);
+      log_softmax_[idx] = exp_on_device(log_softmax_[idx]);
     } else {
       auto softmax = log_softmax_[idx];
-      log_softmax_[idx] = real_exp(softmax);
+      log_softmax_[idx] = exp_on_device(softmax);
       loss_[row_idx] = -softmax;
     }
   }
@@ -253,10 +258,10 @@ struct HardLabelSoftmaxWithCrossEntropyFunctorWithIgnoreIdx {
     auto row_idx = idx / feature_size_;
     auto col_idx = idx % feature_size_;
     if (col_idx != labels_[row_idx] || col_idx == ignore_idx_) {
-      log_softmax_[idx] = real_exp(log_softmax_[idx]);
+      log_softmax_[idx] = exp_on_device(log_softmax_[idx]);
     } else {
       auto softmax = log_softmax_[idx];
-      log_softmax_[idx] = real_exp(softmax);
+      log_softmax_[idx] = exp_on_device(softmax);
       loss_[row_idx] = -softmax;
     }
   }
@@ -464,9 +469,12 @@ class SoftmaxWithCrossEntropyGradCUDAKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_CUDA_KERNEL(softmax_with_cross_entropy,
-                        ops::SoftmaxWithCrossEntropyCUDAKernel<float>,
-                        ops::SoftmaxWithCrossEntropyCUDAKernel<double>);
-REGISTER_OP_CUDA_KERNEL(softmax_with_cross_entropy_grad,
-                        ops::SoftmaxWithCrossEntropyGradCUDAKernel<float>,
-                        ops::SoftmaxWithCrossEntropyGradCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(
+    softmax_with_cross_entropy, ops::SoftmaxWithCrossEntropyCUDAKernel<float>,
+    ops::SoftmaxWithCrossEntropyCUDAKernel<paddle::platform::float16>,
+    ops::SoftmaxWithCrossEntropyCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(
+    softmax_with_cross_entropy_grad,
+    ops::SoftmaxWithCrossEntropyGradCUDAKernel<float>,
+    ops::SoftmaxWithCrossEntropyGradCUDAKernel<paddle::platform::float16>,
+    ops::SoftmaxWithCrossEntropyGradCUDAKernel<double>);

paddle/fluid/operators/sum_op.cc

@@ -41,7 +41,9 @@ class SumOp : public framework::OperatorWithKernel {
       return;  // skip runtime infershape when is tensor array;
     }
 
+    auto x_var_types = ctx->GetInputsVarType("X");
     auto x_dims = ctx->GetInputsDim("X");
+
     size_t N = x_dims.size();
     PADDLE_ENFORCE_GT(N, 0, "Input tensors count should > 0.");
     if (N == 1) {
@@ -49,7 +51,13 @@ class SumOp : public framework::OperatorWithKernel {
     }
 
     framework::DDim in_dim({0});
-    for (auto& x_dim : x_dims) {
+    for (size_t i = 0; i < x_dims.size(); ++i) {
+      auto& x_dim = x_dims[i];
+      // x_dim.size() == 1 means the real dim of selected rows is [0]
+      if (x_var_types[i] == framework::proto::VarType::SELECTED_ROWS &&
+          x_dim.size() == 1) {
+        continue;
+      }
       if (framework::product(x_dim) == 0) {
         continue;
       }

paddle/fluid/operators/teacher_student_sigmoid_loss_op.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/teacher_student_sigmoid_loss_op.h"
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+class TeacherStudentSigmoidLossOp : 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->HasInput("Label"), "Input(Label) should be not null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto label_dims = ctx->GetInputDim("Label");
+    PADDLE_ENFORCE_EQ(x_dims.size(), 2UL, "Input(X)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(label_dims.size(), 2UL,
+                      "Input(Label)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0],
+                      "The 1st dimension of Input(X) and Input(Label) should "
+                      "be equal.");
+    PADDLE_ENFORCE_EQ(label_dims[1], 1UL,
+                      "The 2nd dimension of "
+                      "Input(Label) should be 1.");
+    ctx->SetOutputDim("Y", {x_dims[0], 1});
+    ctx->ShareLoD("X", /*->*/ "Y");
+  }
+
+ protected:
+  // Explicitly set that the data type of computation kernel of
+  // teacher_student_sigmoid_loss
+  // is determined by its input "X".
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
+                                   ctx.device_context());
+  }
+};
+
+class TeacherStudentSigmoidLossGradientOp
+    : 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->HasInput("Label"), "Input(Label) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")),
+                   "Input(Y@GRAD) should be not null.");
+    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
+                   "Output(X@GRAD) should be not null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto label_dims = ctx->GetInputDim("Label");
+    auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y"));
+    PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(dy_dims.size(), 2, "Input(Y@Grad)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(label_dims.size(), 2, "Input(Label)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0],
+                      "The 1st dimension of Input(X) and Input(Label) should "
+                      "be equal.");
+    PADDLE_ENFORCE_EQ(x_dims[0], dy_dims[0],
+                      "The 1st dimension of Input(X) and Input(Y@Grad) should "
+                      "be equal.");
+    PADDLE_ENFORCE_EQ(dy_dims[1], 1,
+                      "The 2nd dimension of Input(Y@Grad) should be 1.");
+    PADDLE_ENFORCE_EQ(label_dims[1], 1,
+                      "When Attr(soft_label) == false, the 2nd dimension of "
+                      "Input(Label) should be 1.");
+    ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
+    ctx->ShareLoD("X", framework::GradVarName("X"));
+  }
+
+ protected:
+  // Explicitly set that the data type of computation kernel of
+  // teacher_student_sigmoid_loss
+  // is determined by its input "X".
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
+                                   ctx.device_context());
+  }
+};
+
+class TeacherStudentSigmoidLossOpMaker
+    : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor, default Tensor<float>), a 2-D tensor with shape [N x 1],"
+             " where N is the batch size and D is the output. "
+             "This input is a probability computed by the previous operator, "
+             "which is almost always the result of a softmax operator.");
+    AddInput("Label",
+             "(Tensor), the ground truth which is a 2-D tensor. "
+             "Label is a Tensor<float> with shape [N x 1]. ");
+    AddOutput("Y",
+              "(Tensor, default Tensor<float>), a 2-D tensor with shape "
+              "[N x 1]. The teacher student sigmoid loss.");
+    AddAttr<float>(
+        "soft_max_up_bound",
+        "fp32, if input > soft_max_up_bound, will be bound, default 15.0")
+        .SetDefault(15.0);
+    AddAttr<float>(
+        "soft_max_lower_bound",
+        "fp32, if input < soft_max_lower_bound, will be bound, default -15.0")
+        .SetDefault(-15.0);
+    AddComment(R"DOC(
+TeacherStudentSigmoidLoss Operator.
+
+It's similarity to SigmoidCrossEntropyWithLogits Operator. The difference is that
+we add another label(z') to original.
+        loss = max(x, 0) - x * z + log(1 + exp(-abs(x))) + max(x, 0) - x * z' + log(1 + exp(-abs(x)))
+        z is click or not
+        z' is teacher value 
+        label = {-2, -1, [0, 2]}
+        when z' is not exist, clk = 0 : label = -2;
+        when z' is not exist, clk = 1 : label = -1;
+        when z' is exist    , clk = 0 : label = 0 + z';
+        when z' is exist    , clk = 1 : label = 1 + z';
+
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(teacher_student_sigmoid_loss,
+                  ops::TeacherStudentSigmoidLossOp,
+                  ops::TeacherStudentSigmoidLossOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+
+REGISTER_OPERATOR(teacher_student_sigmoid_loss_grad,
+                  ops::TeacherStudentSigmoidLossGradientOp);
+
+REGISTER_OP_CPU_KERNEL(teacher_student_sigmoid_loss,
+                       ops::TeacherStudentSigmoidLossOpKernel<float>,
+                       ops::TeacherStudentSigmoidLossOpKernel<double>);
+
+REGISTER_OP_CPU_KERNEL(teacher_student_sigmoid_loss_grad,
+                       ops::TeacherStudentSigmoidLossGradOpKernel<float>,
+                       ops::TeacherStudentSigmoidLossGradOpKernel<double>);

paddle/fluid/operators/teacher_student_sigmoid_loss_op.h

+/* 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. */
+
+#pragma once
+#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 TeacherStudentSigmoidLossOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    Tensor* y = context.Output<Tensor>("Y");
+    const Tensor* x = context.Input<Tensor>("X");
+    const Tensor* labels = context.Input<Tensor>("Label");
+    T* y_data = y->mutable_data<T>(context.GetPlace());
+    const T* x_data = x->data<T>();
+    const T* label_data = labels->data<T>();
+    int64_t batch_size = x->dims()[0];
+    // loss = max(x, 0) - x * z + log(1 + exp(-abs(x))) + max(x, 0) - x * z' +
+    // log(1 + exp(-abs(x)))
+    // z is click or not
+    // z' is value q of feed_fine
+    // label = {-2, -1, [0, 2]}
+    // when z' is not exist, clk = 0 : label = -2;
+    // when z' is not exist, clk = 1 : label = -1;
+    // when z' is exist    , clk = 0 : label = 0 + z';
+    // when z' is exist    , clk = 1 : label = 1 + z';
+    for (int i = 0; i < batch_size; ++i) {
+      if (label_data[i] < -1.0) {
+        y_data[i] = (x_data[i] > 0 ? x_data[i] : 0.0) +
+                    log(1.0 + exp(-fabs(x_data[i])));
+      } else if (label_data[i] < 0.0) {
+        y_data[i] = (x_data[i] > 0 ? x_data[i] : 0.0) - x_data[i] +
+                    log(1.0 + exp(-fabs(x_data[i])));
+      } else if (label_data[i] < 1.0) {
+        y_data[i] = (x_data[i] > 0 ? x_data[i] : 0.0) +
+                    log(1.0 + exp(-fabs(x_data[i]))) +
+                    (x_data[i] > 0 ? x_data[i] : 0.0) -
+                    x_data[i] * label_data[i] +
+                    log(1.0 + exp(-fabs(x_data[i])));
+      } else {
+        y_data[i] = (x_data[i] > 0 ? x_data[i] : 0.0) - x_data[i] +
+                    log(1.0 + exp(-fabs(x_data[i]))) +
+                    (x_data[i] > 0 ? x_data[i] : 0.0) -
+                    x_data[i] * (label_data[i] - 1.0) +
+                    log(1.0 + exp(-fabs(x_data[i])));
+      }
+    }
+  }
+};
+
+template <typename T>
+class TeacherStudentSigmoidLossGradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const Tensor* x = context.Input<Tensor>("X");
+    const T* x_data = x->data<T>();
+
+    Tensor* dx = context.Output<Tensor>(framework::GradVarName("X"));
+    T* dx_data = dx->mutable_data<T>(context.GetPlace());
+
+    const Tensor* labels = context.Input<Tensor>("Label");
+    const T* label_data = labels->data<T>();
+
+    T soft_max_up_bound =
+        static_cast<T>(context.Attr<float>("soft_max_up_bound"));
+    T soft_max_lower_bound =
+        static_cast<T>(context.Attr<float>("soft_max_lower_bound"));
+
+    int64_t batch_size = x->dims()[0];
+
+    const framework::Tensor* dOut =
+        context.Input<framework::Tensor>(framework::GradVarName("Y"));
+
+    const T* dout_data = dOut->data<T>();
+
+    for (int i = 0; i < batch_size; ++i) {
+      T sum_val = x_data[i];
+      if (sum_val > soft_max_up_bound) {
+        sum_val = soft_max_up_bound;
+      } else {
+        if (sum_val < soft_max_lower_bound) {
+          sum_val = soft_max_lower_bound;
+        }
+      }
+
+      T pred = 1.0 / (1.0 + exp(-sum_val));
+      if (label_data[i] < -1.0) {
+        dx_data[i] = 0.0 - pred;
+      } else if (label_data[i] < 0.0) {
+        dx_data[i] = 1.0 - pred;
+      } else {
+        dx_data[i] = label_data[i] - 2.0 * pred;
+      }
+      if (sum_val >= soft_max_up_bound || sum_val <= soft_max_lower_bound) {
+        dx_data[i] = 0;
+      }
+      dx_data[i] *= dout_data[i] * -1;
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/platform/cpu_info.cc

@@ -35,20 +35,8 @@ limitations under the License. */
 DEFINE_double(fraction_of_cpu_memory_to_use, 1,
               "Default use 100% of CPU memory for PaddlePaddle,"
               "reserve the rest for page tables, etc");
-#if !defined(_WIN32)
-DEFINE_uint64(initial_cpu_memory_in_mb,
-#ifdef PADDLE_WITH_MKLDNN
-              /* Aligned with mozga-intel, MKLDNN need at least 5000 MB
-               * to obtain the best performance*/
-              5000ul,
-#else
-              500ul,
-#endif
-              "Initial CPU memory for PaddlePaddle, in MD unit.");
-#else
 DEFINE_uint64(initial_cpu_memory_in_mb, 500ul,
               "Initial CPU memory for PaddlePaddle, in MD unit.");
-#endif  // !defined(_WIN32)
 
 DEFINE_double(
     fraction_of_cuda_pinned_memory_to_use, 0.5,

paddle/fluid/platform/cuda_helper_test.cu

@@ -15,6 +15,9 @@
 #include <gtest/gtest.h>
 #include <algorithm>
 #include <iostream>
+#ifdef _WIN32
+#include <numeric>
+#endif
 #include <random>
 
 #define PADDLE_CUDA_FP16

paddle/fluid/platform/device_context.cc

@@ -92,26 +92,24 @@ platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
     const platform::Place& place, const cudaStream_t& stream) {
   PADDLE_ENFORCE(platform::is_gpu_place(place));
   auto place_stream = std::make_pair(place, stream);
-  {
-    std::unique_lock<std::mutex> lock(mtx_);
-    if (!device_allocator_.count(place_stream)) {
-      device_allocator_[place_stream].reset(new TemporaryAllocator(place));
-      device_allocator_[place_stream]->SetCallback([stream]() {
-        PADDLE_ENFORCE(cudaStreamSynchronize(stream));
-        PADDLE_ENFORCE(cudaGetLastError());
-      });
-    }
+  std::unique_lock<std::mutex> lock(mtx_);
+  auto it = device_allocator_.find(place_stream);
+  if (it == device_allocator_.end()) {
+    auto tmp_allocator = new TemporaryAllocator(place);
+    tmp_allocator->SetCallback([stream]() {
+      PADDLE_ENFORCE(cudaStreamSynchronize(stream));
+      PADDLE_ENFORCE(cudaGetLastError());
+    });
+    device_allocator_[place_stream].reset(tmp_allocator);
+    return *tmp_allocator;
+  } else {
+    return *it->second;
   }
-  return *device_allocator_.at(place_stream);
 }
 
 template <>
 platform::TemporaryAllocator& DeviceTemporaryAllocator::Get(
     const platform::CUDADeviceContext& dev_ctx) {
-  auto place_stream = std::make_pair(dev_ctx.GetPlace(), dev_ctx.stream());
-  if (device_allocator_.count(place_stream)) {
-    return *device_allocator_.at(place_stream);
-  }
   return Get(dev_ctx.GetPlace(), dev_ctx.stream());
 }
 #endif
@@ -325,7 +323,7 @@ Place CUDADeviceContext::GetPlace() const { return place_; }
 void CUDADeviceContext::Wait() const {
   auto& allocator =
       DeviceTemporaryAllocator::Instance().Get<CUDADeviceContext>(*this);
-  allocator.Release([=]() {
+  allocator.Release([this]() {
     PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
     PADDLE_ENFORCE(cudaGetLastError());
   });

paddle/fluid/platform/device_context.h

@@ -61,7 +61,7 @@ namespace platform {
  * the allocations of temp_allocation_queue:
  *  - when the Stream calls cudaStreamSynchronize;
  *  - when the allocation size of opportunities exceeds a certain threshold
- *    (defined by FLAGS_limit_of_temporary_allocation).
+ *    (defined by FLAGS_limit_of_tmp_allocation).
  *
  * */
 class DeviceTemporaryAllocator {

paddle/fluid/platform/enforce.h

@@ -263,6 +263,10 @@ inline void throw_on_error(T e) {
 #define __THROW_ON_ERROR_ONE_ARG(COND, ARG) \
   ::paddle::platform::throw_on_error(COND, ::paddle::string::Sprintf(ARG));
 
+#ifdef _WIN32
+#define __PADDLE_THROW_ON_ERROR(COND, ...) \
+  __THROW_ON_ERROR_ONE_ARG(COND, __VA_ARGS__)
+#else  // _WIN32
 #define __PADDLE_THROW_ON_ERROR(COND, ...)                                \
   __PADDLE_THROW_ERROR_I(                                                 \
       __VA_ARGS__, ::paddle::platform::throw_on_error(COND, __VA_ARGS__), \
@@ -274,6 +278,7 @@ inline void throw_on_error(T e) {
       ::paddle::platform::throw_on_error(COND, __VA_ARGS__),              \
       ::paddle::platform::throw_on_error(COND, __VA_ARGS__),              \
       __THROW_ON_ERROR_ONE_ARG(COND, __VA_ARGS__))
+#endif  // _WIN32
 
 #define __PADDLE_UNARY_COMPARE(COND, ...)                 \
   do {                                                    \

paddle/fluid/platform/float16.h

@@ -59,7 +59,7 @@ limitations under the License. */
 #if !defined(_WIN32)
 #define PADDLE_ALIGN(x) __attribute__((aligned(x)))
 #else
-#define PADDLE_ALIGN(x) /*do nothing*/
+#define PADDLE_ALIGN(x) __declspec(align(x))
 #endif
 
 namespace paddle {

paddle/fluid/platform/float16_test.cu

@@ -271,11 +271,13 @@ TEST(float16, isinf) {
   float16 b = float16(INFINITY);
   // underflow to 0
   float16 native_a(5e-40f);
-  // overflow to inf
-  float16 native_b(5e40f);
   EXPECT_EQ(std::isinf(a), true);
   EXPECT_EQ(std::isinf(b), true);
+#ifndef _WIN32
+  // overflow to inf
+  float16 native_b(5e40f);
   EXPECT_EQ(std::isinf(native_b), true);
+#endif
   EXPECT_EQ(native_a, float16(0));
 }
 

paddle/fluid/platform/mkldnn_reuse.h

@@ -210,13 +210,15 @@ class MKLDNNHandler {
     dst_memory.reset(new mkldnn::memory(*dst_pd, to_void_cast<T>(output_data)));
   }
 
-  static void AppendKey(
-      std::string* key, const mkldnn::memory::dims& input_dims,
-      const mkldnn::memory::dims& weights_dims, const std::vector<int>& strides,
-      const std::vector<int>& paddings, const std::vector<int>& dilations,
-      const int& groups, const mkldnn::memory::data_type& srcdt,
-      const mkldnn::memory::format& format,
-      const mkldnn::memory::data_type& dstdt, const std::string& suffix) {
+  static void AppendKey(std::string* key,
+                        const mkldnn::memory::dims& input_dims,
+                        const mkldnn::memory::dims& weights_dims,
+                        const std::vector<int>& strides,
+                        const std::vector<int>& paddings,
+                        const std::vector<int>& dilations, const int& groups,
+                        const mkldnn::memory::data_type& srcdt,
+                        const mkldnn::memory::format& format, const bool& relu,
+                        const bool& residual, const std::string& suffix) {
     AppendKeyDims(key, input_dims);
     AppendKeyDims(key, weights_dims);
     AppendKeyVec(key, strides);
@@ -225,7 +227,8 @@ class MKLDNNHandler {
     AppendKey(key, std::to_string(groups));
     AppendKey(key, std::to_string(srcdt));
     AppendKey(key, std::to_string(format));
-    AppendKey(key, std::to_string(dstdt));
+    AppendKey(key, std::to_string(relu));
+    AppendKey(key, std::to_string(residual));
     AppendKey(key, suffix);
   }
 
@@ -664,15 +667,35 @@ static std::shared_ptr<mkldnn::memory> SetDstMemory(
 }
 
 template <typename T>
-static std::shared_ptr<mkldnn::memory> SetDstMemoryHandler(
+static std::shared_ptr<mkldnn::memory> SetDstMemory(
     const framework::ExecutionContext& ctx, framework::Tensor* output,
-    const std::shared_ptr<ConvMKLDNNHandler>& handler) {
+    const framework::Tensor* residual_param,
+    const mkldnn::memory::desc& user_residual_md,
+    const std::shared_ptr<ConvMKLDNNHandler>& handler,
+    std::vector<mkldnn::primitive>* pipeline) {
+  const T* residual_param_data = residual_param->data<T>();
+  PADDLE_ENFORCE(residual_param_data != nullptr,
+                 "Provide data if you want MKLDNN conv+elementwise_add fusion");
+  std::shared_ptr<mkldnn::memory> user_residual_memory_p =
+      handler->AcquireResidualDataMemory(user_residual_md,
+                                         to_void_cast<T>(residual_param_data));
+  T* output_data = output->mutable_data<T>(ctx.GetPlace());
+  std::shared_ptr<mkldnn::memory> dst_memory_p =
+      handler->AcquireDstMemoryFromResidualDataMemory(
+          user_residual_memory_p, to_void_cast<T>(output_data), *pipeline);
+  return dst_memory_p;
+}
+
+template <typename T>
+static void SetDstMemoryHandler(
+    const framework::ExecutionContext& ctx, framework::Tensor* output,
+    const std::shared_ptr<ConvMKLDNNHandler>& handler,
+    std::shared_ptr<mkldnn::memory>* dst_memory_p) {
   T* output_data = output->mutable_data<T>(
       ctx.GetPlace(), ::paddle::memory::Allocator::kDefault,
       handler->GetDstMemorySize());
-  std::shared_ptr<mkldnn::memory> dst_memory_p;
-  dst_memory_p->set_data_handle(to_void_cast<T>(output_data));
-  return dst_memory_p;
+  (*dst_memory_p)->set_data_handle(to_void_cast<T>(output_data));
 }
+
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/platform/temporary_allocator.cc

@@ -15,8 +15,15 @@
 #include "paddle/fluid/platform/temporary_allocator.h"
 #include "paddle/fluid/memory/allocation/allocator_facade.h"
 
-DEFINE_double(limit_of_temporary_allocation, -1,
-              "The up limit of temporary_allocation size.");
+DEFINE_int64(limit_of_tmp_allocation, -1,
+             "The up limit of temporary_allocation size.");
+DEFINE_double(times_excess_than_required_tmp_allocation, 2,
+              "times_excess_than_required_tmp_allocation indicates the "
+              "max size the TemporaryAllocator can return. For example, "
+              "if the required memory size is N, and "
+              "times_excess_than_required_tmp_allocation is 2.0, "
+              "the TemporaryAllocator will return the available allocation "
+              "that the range of size is N ~ 2*N.");
 
 namespace paddle {
 namespace platform {
@@ -29,24 +36,25 @@ TemporaryAllocation::TemporaryAllocation(
       underlying_allocation_(std::move(underlying_allocation)) {}
 
 TemporaryAllocator::TemporaryAllocator(platform::Place place) : place_(place) {
-  temp_mem_queue_.reset(new std::deque<TemporaryAllocation *>());
+  temp_mem_map_.reset(new std::multimap<size_t, TemporaryAllocation *>());
 }
 
 bool TemporaryAllocator::IsAllocThreadSafe() const { return true; }
 
 void TemporaryAllocator::Release(const std::function<void()> &callback) {
-  std::shared_ptr<std::deque<TemporaryAllocation *>> t_allocations;
+  std::unique_ptr<std::multimap<size_t, TemporaryAllocation *>> t_allocations;
   {
     std::unique_lock<std::mutex> lock(mtx_);
     callback();
-    t_allocations = temp_mem_queue_;
-    temp_mem_queue_.reset(new std::deque<TemporaryAllocation *>());
+    t_allocations.swap(temp_mem_map_);
+    temp_mem_map_.reset(new std::multimap<size_t, TemporaryAllocation *>());
     wait_delete_mem_ = 0;
   }
+
   for (auto tmp : *t_allocations) {
-    VLOG(10) << "Delete temporary allocation " << tmp->ptr()
-             << " size: " << tmp->size();
-    delete tmp;
+    VLOG(10) << "Delete temporary allocation " << tmp.second->ptr()
+             << " size: " << tmp.second->size();
+    delete tmp.second;
   }
 }
 
@@ -54,28 +62,34 @@ void TemporaryAllocator::Free(alloc::Allocation *allocation) {
   auto *temp_allocation = dynamic_cast<TemporaryAllocation *>(allocation);
   PADDLE_ENFORCE_NOT_NULL(temp_allocation);
   if (platform::is_gpu_place(temp_allocation->place())) {
+    PADDLE_ENFORCE(platform::is_same_place(temp_allocation->place(), place_),
+                   "The place should be the same.");
     size_t wait_delete_mem = 0;
     {
       std::unique_lock<std::mutex> lock(mtx_);
-      temp_mem_queue_->emplace_back(temp_allocation);
+      temp_mem_map_->emplace(temp_allocation->size(), temp_allocation);
       wait_delete_mem_ += temp_allocation->size();
       wait_delete_mem = wait_delete_mem_;
       VLOG(10) << "Move temporary allocation: " << temp_allocation->ptr()
                << " to delete queue: " << temp_allocation->size() << "; "
-               << "wait_delete_mem: " << wait_delete_mem_;
+               << "wait_delete_mem: " << wait_delete_mem;
     }
-    if (FLAGS_limit_of_temporary_allocation > 0 &&
-        wait_delete_mem > FLAGS_limit_of_temporary_allocation) {
+
+    if (FLAGS_limit_of_tmp_allocation > 0 &&
+        wait_delete_mem > static_cast<size_t>(FLAGS_limit_of_tmp_allocation)) {
+      PADDLE_ENFORCE(callback_ != nullptr, "The callback is non-initialized.");
       Release(callback_);
     }
     return;
   }
+  VLOG(10) << "Delete temporary allocation " << temp_allocation->ptr()
+           << " size: " << temp_allocation->size();
   delete temp_allocation;
 }
 
 size_t TemporaryAllocator::TemporaryAllocationQueueSize() {
   std::unique_lock<std::mutex> lock(mtx_);
-  return temp_mem_queue_ ? temp_mem_queue_->size() : 0;
+  return temp_mem_map_ ? temp_mem_map_->size() : 0;
 }
 
 void TemporaryAllocator::SetCallback(const std::function<void()> &callback) {
@@ -84,6 +98,27 @@ void TemporaryAllocator::SetCallback(const std::function<void()> &callback) {
 
 alloc::Allocation *TemporaryAllocator::AllocateImpl(
     size_t size, alloc::Allocator::Attr attr) {
+  {
+    // Find available allocation in temp_mem_map.
+    std::unique_lock<std::mutex> lock(mtx_);
+    if (temp_mem_map_->size()) {
+      auto it = temp_mem_map_->lower_bound(size);
+      // FIXME(zcd): Not sure the best value of excess fraction.
+      if (it != temp_mem_map_->end() &&
+          it->first <
+              static_cast<size_t>(
+                  size * FLAGS_times_excess_than_required_tmp_allocation)) {
+        auto tmp_ptr = it->second;
+        temp_mem_map_->erase(it);
+        wait_delete_mem_ -= tmp_ptr->size();
+        VLOG(10) << "Reuse temporary allocation: " << tmp_ptr->ptr() << ": "
+                 << tmp_ptr->size();
+        return tmp_ptr;
+      }
+    }
+  }
+  // If not find the the available allocation, get allocation from
+  // AllocatorFacadeInstance.
   auto raw_allocation =
       alloc::AllocatorFacade::Instance().Alloc(place_, size, attr);
   auto temp_mem = new TemporaryAllocation(std::move(raw_allocation));

paddle/fluid/platform/temporary_allocator.h

@@ -15,6 +15,7 @@
 #pragma once
 #include <condition_variable>  // NOLINT
 #include <deque>
+#include <map>
 #include <mutex>  // NOLINT
 #include "paddle/fluid/memory/allocation/allocator.h"
 #include "paddle/fluid/platform/lock_guard_ptr.h"
@@ -39,7 +40,7 @@ class TemporaryAllocation : public memory::allocation::Allocation {
  *
  * There is one opportunity to free the allocations of temp_allocation_queue:
  *   - when the allocation size of opportunities exceeds a certain threshold
- *     (defined by FLAGS_limit_of_temporary_allocation).
+ *     (defined by FLAGS_limit_of_tmp_allocation).
  *
  * */
 class TemporaryAllocator : public memory::allocation::Allocator {
@@ -62,11 +63,10 @@ class TemporaryAllocator : public memory::allocation::Allocator {
 
  private:
   platform::Place place_;
-
   // When the allocation is not held by any variable, it should be placed
-  // to temp_mem_queue immediately.
-  std::shared_ptr<std::deque<TemporaryAllocation *>> temp_mem_queue_{nullptr};
-
+  // to temp_mem_map immediately.
+  std::unique_ptr<std::multimap<size_t, TemporaryAllocation *>> temp_mem_map_{
+      nullptr};
   std::mutex mtx_;
   size_t wait_delete_mem_{0};
   std::function<void()> callback_;

paddle/fluid/platform/temporary_allocator_test.cc

@@ -18,7 +18,8 @@
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/tensor_util.h"
 
-DECLARE_double(limit_of_temporary_allocation);
+DECLARE_int64(limit_of_tmp_allocation);
+DECLARE_double(times_excess_than_required_tmp_allocation);
 
 namespace paddle {
 namespace platform {
@@ -35,7 +36,7 @@ class DummyOp : public framework::OperatorBase {
                const platform::Place& place) const override {}
 };
 
-TEST(temporary_allocator, temporary_allocator) {
+TEST(temporary_allocator, test_base_function) {
   platform::CPUPlace cpu_place;
   TemporaryAllocator alloc(cpu_place);
   alloc.Allocate(100);
@@ -59,10 +60,10 @@ TEST(temporary_allocator, temporary_allocator) {
 #endif
 }
 
-TEST(temporary_allocator, add_callback) {
+TEST(temporary_allocator, test_flags_function) {
 #ifdef PADDLE_WITH_CUDA
-  const double limit = FLAGS_limit_of_temporary_allocation;
-  FLAGS_limit_of_temporary_allocation = 10;
+  const int64_t limit = FLAGS_limit_of_tmp_allocation;
+  FLAGS_limit_of_tmp_allocation = 10;
   platform::CUDAPlace gpu_place(0);
   TemporaryAllocator gpu_alloc(gpu_place);
 
@@ -78,7 +79,52 @@ TEST(temporary_allocator, add_callback) {
   });
   { gpu_alloc.Allocate(100); }
   PADDLE_ENFORCE(deleted);
-  FLAGS_limit_of_temporary_allocation = limit;
+  FLAGS_limit_of_tmp_allocation = limit;
+#endif
+}
+
+TEST(temporary_allocator, test_reuse_tmp_allocation) {
+#ifdef PADDLE_WITH_CUDA
+  platform::CUDAPlace gpu_place(0);
+  TemporaryAllocator gpu_alloc(gpu_place);
+  gpu_alloc.SetCallback([]() {});
+
+  void* tmp_allocation_ptr1 = nullptr;
+  {
+    PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
+    auto tmp_allocation1 = gpu_alloc.Allocate(100);
+    tmp_allocation_ptr1 = tmp_allocation1->ptr();
+  }
+  PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 1);
+  auto tmp_allocation2 = gpu_alloc.Allocate(100);
+  void* tmp_allocation_ptr2 = tmp_allocation2->ptr();
+  PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
+  PADDLE_ENFORCE_EQ(tmp_allocation_ptr1, tmp_allocation_ptr2);
+
+  auto tmp_allocation3 = gpu_alloc.Allocate(100);
+  void* tmp_allocation_ptr3 = tmp_allocation2->ptr();
+  PADDLE_ENFORCE_EQ(tmp_allocation_ptr1, tmp_allocation_ptr3);
+#endif
+}
+
+TEST(temporary_allocator, test_times_excess_than_required_tmp_allocation) {
+#ifdef PADDLE_WITH_CUDA
+  platform::CUDAPlace gpu_place(0);
+  TemporaryAllocator gpu_alloc(gpu_place);
+  gpu_alloc.SetCallback([]() {});
+  double excess_fraction = FLAGS_times_excess_than_required_tmp_allocation;
+  void* tmp_allocation_ptr1 = nullptr;
+  {
+    PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
+    auto tmp_allocation1 =
+        gpu_alloc.Allocate(static_cast<size_t>(100 * excess_fraction - 1));
+    tmp_allocation_ptr1 = tmp_allocation1->ptr();
+  }
+  PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 1);
+  auto tmp_allocation2 = gpu_alloc.Allocate(100);
+  void* tmp_allocation_ptr2 = tmp_allocation2->ptr();
+  PADDLE_ENFORCE_EQ(gpu_alloc.TemporaryAllocationQueueSize(), 0);
+  PADDLE_ENFORCE_EQ(tmp_allocation_ptr1, tmp_allocation_ptr2);
 #endif
 }
 

paddle/fluid/pybind/CMakeLists.txt

-
-set(PYBIND_DEPS pybind python proto_desc memory executor async_executor prune feed_fetch_method pass_builder parallel_executor profiler layer scope_pool)
+set(PYBIND_DEPS pybind python proto_desc memory executor async_executor prune
+  feed_fetch_method pass_builder parallel_executor profiler layer scope_pool
+  tracer)
 if(WITH_PYTHON)
   list(APPEND PYBIND_DEPS py_func_op)
 endif()
-set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc async_executor_py.cc imperative.cc)
+set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc async_executor_py.cc imperative.cc ir.cc)
 
 if(WITH_PYTHON)
   if(WITH_AMD_GPU)
@@ -21,9 +22,8 @@ if(WITH_PYTHON)
     endif(NOT APPLE AND NOT ANDROID AND NOT WIN32)
   endif(WITH_AMD_GPU)
 
-  if(WIN32)
-    target_link_libraries(paddle_pybind shlwapi)
-  endif(WIN32)
+  get_property (os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
+  target_link_libraries(paddle_pybind ${os_dependency_modules})
 
   cc_test(tensor_py_test SRCS tensor_py_test.cc DEPS python)
 endif(WITH_PYTHON)

paddle/fluid/pybind/const_value.cc

@@ -49,6 +49,9 @@ void BindConstValue(pybind11::module* m) {
   op_proto_and_checker_maker.def(
       "kOpNameScopeAttrName",
       framework::OpProtoAndCheckerMaker::OpNamescopeAttrName);
+  op_proto_and_checker_maker.def(
+      "kOpCreationCallstackAttrName",
+      framework::OpProtoAndCheckerMaker::OpCreationCallstackAttrName);
 }
 
 }  // namespace pybind

paddle/fluid/pybind/imperative.cc

@@ -26,7 +26,9 @@ void BindTracer(pybind11::module *m) {
            [](imperative::Tracer &self, framework::BlockDesc *root_block) {
              new (&self) imperative::Tracer(root_block);
            })
-      .def("trace", &imperative::Tracer::Trace);
+      .def("trace", &imperative::Tracer::Trace)
+      .def("py_trace", &imperative::Tracer::PyTrace,
+           pybind11::return_value_policy::take_ownership);
 }
 
 }  // namespace pybind

paddle/fluid/pybind/imperative.h

@@ -22,7 +22,7 @@ limitations under the License. */
 namespace paddle {
 namespace pybind {
 
-class PyLayer : public imperative::Layer {
+class Layer : public imperative::Layer {
  public:
   using imperative::Layer::Layer;  // Inherit constructors
 
@@ -31,10 +31,6 @@ class PyLayer : public imperative::Layer {
     PYBIND11_OVERLOAD(std::vector<imperative::VarBase>, Layer, Forward,
                       inputs);  // NOLINT
   }
-
-  void Backward() override {
-    PYBIND11_OVERLOAD(void, Layer, Backward, );  // NOLINT
-  }
 };
 
 class PyOpBase : public imperative::OpBase {

paddle/fluid/pybind/ir.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/pybind/ir.h"
+#include <string>
+#include <unordered_map>
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/node.h"
+#include "paddle/fluid/framework/op_desc.h"
+#include "paddle/fluid/framework/var_desc.h"
+#include "pybind11/stl.h"
+
+namespace py = pybind11;
+using paddle::framework::ir::Graph;
+using paddle::framework::ir::Node;
+using paddle::framework::OpDesc;
+using paddle::framework::ProgramDesc;
+using paddle::framework::VarDesc;
+using pybind11::return_value_policy;
+
+namespace paddle {
+namespace pybind {
+void BindGraph(py::module *m) {
+  py::class_<Graph, std::shared_ptr<Graph>>(
+      *m, "Graph",
+      "The graph is a Directed Acyclic Single Static Assignment Graph, see "
+      "`paddle::ir::Graph` for details.")
+      .def(py::init<const ProgramDesc &>())
+      .def("has", &Graph::Has)
+      .def("get_int", &Graph::Get<int>)
+      .def("get_float", &Graph::Get<float>)
+      .def("get_double", &Graph::Get<double>)
+      .def("get_string", &Graph::Get<std::string>)
+      .def("set", [](Graph &self, const std::string &attr_name,
+                     int attr) { return self.Set(attr_name, new int(attr)); })
+      .def("set",
+           [](Graph &self, const std::string &attr_name,
+              const std::string &attr) {
+             return self.Set(attr_name, new std::string(attr));
+           })
+      .def("set",
+           [](Graph &self, const std::string &attr_name, float attr) {
+             return self.Set(attr_name, new float(attr));
+           })
+      .def("set",
+           [](Graph &self, const std::string &attr_name, double attr) {
+             return self.Set(attr_name, new double(attr));
+           })
+      .def("erase", &Graph::Erase)
+      .def("nodes", &Graph::Nodes, return_value_policy::reference)
+      .def("create_var_node",
+           [](Graph &self, VarDesc &var_desc) {
+             return self.CreateVarNode(&var_desc);
+           },
+           return_value_policy::reference)
+      .def("create_op_node",
+           [](Graph &self, OpDesc &op_desc) {
+             return self.CreateOpNode(&op_desc);
+           },
+           return_value_policy::reference)
+      .def("create_control_dep_var", &Graph::CreateControlDepVar,
+           return_value_policy::reference)
+      .def("create_empty_node", &Graph::CreateEmptyNode,
+           return_value_policy::reference)
+      .def("release_nodes", &Graph::ReleaseNodes)
+      .def("remove_node",
+           [](Graph &self, Node &node) { return self.RemoveNode(&node); })
+      .def("retrieve_node", &Graph::RetrieveNode,
+           return_value_policy::reference)
+      .def("resolve_hazard", &Graph::ResolveHazard);
+}
+
+void BindNode(py::module *m) {
+  py::class_<Node> node(*m, "Node");
+  node.def("name", &Node::Name)
+      .def("node_type", &Node::NodeType)
+      .def("var", &Node::Var)
+      .def("op", &Node::Op)
+      .def("id", &Node::id)
+      .def("is_op", &Node::IsOp)
+      .def("is_var", &Node::IsVar)
+      .def("is_ctrl_var", &Node::IsCtrlVar)
+      .def_readwrite("inputs", &Node::inputs)
+      .def_readwrite("outputs", &Node::outputs);
+
+  py::enum_<Node::Type>(node, "Type")
+      .value("Operation", Node::Type::kOperation)
+      .value("Variable", Node::Type::kVariable)
+      .export_values();
+}
+}  // namespace pybind
+}  // namespace paddle

paddle/fluid/pybind/ir.h

+// 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.
+
+#pragma once
+
+#include <pybind11/pybind11.h>
+#include "paddle/fluid/framework/ir/graph.h"
+
+namespace paddle {
+namespace pybind {
+void BindGraph(pybind11::module *m);
+void BindNode(pybind11::module *m);
+}  // namespace pybind
+}  // namespace paddle

paddle/fluid/pybind/pybind.cc

@@ -49,6 +49,7 @@ limitations under the License. */
 #include "paddle/fluid/pybind/const_value.h"
 #include "paddle/fluid/pybind/exception.h"
 #include "paddle/fluid/pybind/imperative.h"
+#include "paddle/fluid/pybind/ir.h"
 #include "paddle/fluid/pybind/protobuf.h"
 #include "paddle/fluid/pybind/pybind.h"  // NOLINT
 #include "paddle/fluid/pybind/recordio.h"
@@ -125,26 +126,18 @@ PYBIND11_MODULE(core, m) {
   m.add_object("_cleanup",
                py::capsule([]() { ScopePool::Instance().Clear(); }));
 
-  py::class_<imperative::VarBase, std::shared_ptr<imperative::VarBase>>(
-      m, "VarBase", R"DOC()DOC")
+  py::class_<imperative::VarBase>(m, "VarBase", R"DOC()DOC")
       // .def(py::init<>())
       .def(py::init<bool>(), py::arg("stop_gradient") = false)
       .def("_run_backward",
            [](imperative::VarBase &self) { self.RunBackward(); })
       .def("_grad_name", &imperative::VarBase::GradName)
-      .def("_grad", &imperative::VarBase::Grad)
-      .def_property("grad_value",
-                    [](const imperative::VarBase &self) { return self.grads_; },
-                    [](imperative::VarBase &self, framework::Variable *grad) {
-                      self.grads_ = grad;
-                    },
-                    py::return_value_policy::reference)
-      .def_property("value",
-                    [](const imperative::VarBase &self) { return self.var_; },
-                    [](imperative::VarBase &self, framework::Variable *var) {
-                      self.var_ = var;
-                    },
-                    py::return_value_policy::reference)
+      .def("_grad_value", &imperative::VarBase::GradValue)
+      .def("_grad_ivar",
+           [](const imperative::VarBase &self) { return self.grads_; },
+           py::return_value_policy::reference)
+      .def("value", [](const imperative::VarBase &self) { return self.var_; },
+           py::return_value_policy::reference)
       .def_property(
           "desc",
           [](const imperative::VarBase &self) { return self.var_desc_; },
@@ -168,16 +161,44 @@ PYBIND11_MODULE(core, m) {
               self.op_desc_ = op_desc;
             }
           },
+          py::return_value_policy::reference)
+      .def_property(
+          "forward_id",
+          [](const imperative::OpBase &self) { return self.forward_id_; },
+          [](imperative::OpBase &self, int forward_id) {
+            self.forward_id_ = forward_id;
+          },
+          py::return_value_policy::reference)
+      .def_property(
+          "backward_id",
+          [](const imperative::OpBase &self) { return self.backward_id_; },
+          [](imperative::OpBase &self, int backward_id) {
+            self.backward_id_ = backward_id;
+          },
           py::return_value_policy::reference);
 
-  py::class_<imperative::Layer, PyLayer /* <--- trampoline*/> layer(m, "Layer");
+  py::class_<imperative::Layer, Layer /* <--- trampoline*/> layer(m, "Layer");
   layer.def(py::init<>())
-      .def("forward",
-           [](imperative::Layer &self,
-              const std::vector<imperative::VarBase> &inputs) {
-             return self.Forward(inputs);
-           })
-      .def("backward", &imperative::Layer::Backward);
+      .def("forward", [](imperative::Layer &self,
+                         const std::vector<imperative::VarBase> &inputs) {
+        return self.Forward(inputs);
+      });
+
+  py::class_<imperative::PyLayer>(m, "PyLayer")
+      .def(py::init<>())
+      .def_static(
+          "apply",
+          [](int func_id, const std::vector<imperative::VarBase *> &inputs)
+              -> std::vector<imperative::VarBase *> {
+                return imperative::PyLayer::Apply(func_id, inputs);
+              },
+          py::return_value_policy::take_ownership)
+      .def_static("register_func",
+                  [](int func_id, const py::object &callable) {
+                    imperative::PyLayer::RegisterFunc(func_id, callable);
+                  })
+      .def_static("num_funcs", &imperative::PyLayer::NumFuncs);
+
   BindTracer(&m);
 
   py::class_<Tensor>(m, "Tensor", py::buffer_protocol())
@@ -769,7 +790,12 @@ All parameter, weight, gradient are variables in Paddle.
           })
       .def("set_int", [](ir::Pass &self, const std::string &name,
                          int val) { self.Set<const int>(name, new int(val)); })
-      .def("type", &ir::Pass::Type);
+      .def("type", &ir::Pass::Type)
+      .def("apply", [](ir::Pass &self, std::shared_ptr<ir::Graph> graph) {
+        std::unique_ptr<ir::Graph> origin_graph(graph.get());
+        auto optim_graph = self.Apply(std::move(origin_graph));
+        graph.reset(optim_graph.release());
+      });
 
   py::class_<ir::PassBuilder, std::shared_ptr<ir::PassBuilder>> pb(
       m, "PassBuilder");
@@ -1036,6 +1062,9 @@ All parameter, weight, gradient are variables in Paddle.
 
   BindRecordIOWriter(&m);
   BindAsyncExecutor(&m);
+
+  BindGraph(&m);
+  BindNode(&m);
 }
 }  // namespace pybind
 }  // namespace paddle

paddle/scripts/paddle_build.sh

@@ -490,7 +490,8 @@ function assert_api_spec_approvals() {
         BRANCH="develop"
     fi
 
-    API_FILES=("paddle/fluid/API.spec"
+    API_FILES=("cmake/external"
+               "paddle/fluid/API.spec"
                "paddle/fluid/framework/operator.h"
                "paddle/fluid/framework/tensor.h"
                "paddle/fluid/framework/lod_tensor.h"

python/paddle/dataset/mnist.py

@@ -21,10 +21,9 @@ parse training set and test set into paddle reader creators.
 from __future__ import print_function
 
 import paddle.dataset.common
-import subprocess
+import gzip
 import numpy
-import platform
-import tempfile
+import struct
 from six.moves import range
 __all__ = ['train', 'test', 'convert']
 
@@ -41,51 +40,47 @@ TRAIN_LABEL_MD5 = 'd53e105ee54ea40749a09fcbcd1e9432'
 
 def reader_creator(image_filename, label_filename, buffer_size):
     def reader():
-        if platform.system() == 'Darwin':
-            zcat_cmd = 'gzcat'
-        elif platform.system() == 'Linux':
-            zcat_cmd = 'zcat'
-        else:
-            raise NotImplementedError()
-
-        # According to http://stackoverflow.com/a/38061619/724872, we
-        # cannot use standard package gzip here.
-        tmp_image_file = tempfile.TemporaryFile(prefix='paddle_dataset')
-        m = subprocess.Popen(
-            [zcat_cmd, image_filename], stdout=tmp_image_file).communicate()
-        tmp_image_file.seek(16)  # skip some magic bytes
-
-        # Python3 will not take stdout as file
-        tmp_label_file = tempfile.TemporaryFile(prefix='paddle_dataset')
-        l = subprocess.Popen(
-            [zcat_cmd, label_filename], stdout=tmp_label_file).communicate()
-        tmp_label_file.seek(8)  # skip some magic bytes
-
-        try:  # reader could be break.
-            while True:
-                labels = numpy.fromfile(
-                    tmp_label_file, 'ubyte', count=buffer_size).astype("int")
-
-                if labels.size != buffer_size:
-                    break  # numpy.fromfile returns empty slice after EOF.
-
-                images = numpy.fromfile(
-                    tmp_image_file, 'ubyte', count=buffer_size * 28 *
-                    28).reshape((buffer_size, 28 * 28)).astype('float32')
-
-                images = images / 255.0 * 2.0 - 1.0
-
-                for i in range(buffer_size):
-                    yield images[i, :], int(labels[i])
-        finally:
-            try:
-                m.terminate()
-            except:
-                pass
-            try:
-                l.terminate()
-            except:
-                pass
+        with gzip.GzipFile(image_filename, 'rb') as image_file:
+            img_buf = image_file.read()
+            with gzip.GzipFile(label_filename, 'rb') as label_file:
+                lab_buf = label_file.read()
+
+                step_label = 0
+
+                offset_img = 0
+                # read from Big-endian
+                # get file info from magic byte
+                # image file : 16B
+                magic_byte_img = '>IIII'
+                magic_img, image_num, rows, cols = struct.unpack_from(
+                    magic_byte_img, img_buf, offset_img)
+                offset_img += struct.calcsize(magic_byte_img)
+
+                offset_lab = 0
+                # label file : 8B
+                magic_byte_lab = '>II'
+                magic_lab, label_num = struct.unpack_from(magic_byte_lab,
+                                                          lab_buf, offset_lab)
+                offset_lab += struct.calcsize(magic_byte_lab)
+
+                while True:
+                    if step_label >= label_num:
+                        break
+                    fmt_label = '>' + str(buffer_size) + 'B'
+                    labels = struct.unpack_from(fmt_label, lab_buf, offset_lab)
+                    offset_lab += struct.calcsize(fmt_label)
+                    step_label += buffer_size
+
+                    fmt_images = '>' + str(buffer_size * rows * cols) + 'B'
+                    images_temp = struct.unpack_from(fmt_images, img_buf,
+                                                     offset_img)
+                    images = numpy.reshape(images_temp, (
+                        buffer_size, rows * cols)).astype('float32')
+                    offset_img += struct.calcsize(fmt_images)
+
+                    images = images / 255.0 * 2.0 - 1.0
+                    for i in range(buffer_size):
+                        yield images[i, :], int(labels[i])
 
     return reader
 

python/paddle/fluid/__init__.py

@@ -155,7 +155,8 @@ def __bootstrap__():
             'fraction_of_gpu_memory_to_use', 'cudnn_deterministic',
             'enable_cublas_tensor_op_math', 'conv_workspace_size_limit',
             'cudnn_exhaustive_search', 'memory_optimize_debug', 'selected_gpus',
-            'sync_nccl_allreduce'
+            'sync_nccl_allreduce', 'limit_of_tmp_allocation',
+            'times_excess_than_required_tmp_allocation'
         ]
 
     core.init_gflags([sys.argv[0]] +

python/paddle/fluid/data_feeder.py

@@ -71,10 +71,25 @@ class DataToLoDTensorConverter(object):
             for each_data in data:
                 self._feed_impl_(each_data, lod[1:], lod_level - 1)
 
+    def _check_shape(self, shape):
+        for s1, s2 in zip(self.shape, shape):
+            if s1 != s2 and s1 >= 0 and s2 >= 0:
+                raise ValueError(
+                    "Shape not match. What is defined in data layer is {}, but receive {}".
+                    format(self.shape, shape))
+
     def done(self):
         arr = numpy.array(self.data, dtype=self.dtype)
-        if self.shape and len(arr.shape) != len(self.shape):
-            arr = arr.reshape(self.shape)
+        if self.shape:
+            if len(arr.shape) != len(self.shape):
+                try:
+                    arr = arr.reshape(self.shape)
+                except ValueError:
+                    raise ValueError(
+                        "Reshape error. What is defined in data layer is {}, but receive {}"
+                        .format(self.shape, arr.shape))
+            else:
+                self._check_shape(arr.shape)
         t = core.LoDTensor()
         t.set(arr, self.place)
         if self.lod_level > 0:
@@ -152,17 +167,8 @@ class DataFeeder(object):
                 raise TypeError("Feed list should contain a list of variable")
             self.feed_dtypes.append(each_var.dtype)
             self.feed_names.append(each_var.name)
-            shape = each_var.shape
-            batch_size_dim = -1
-            for i, s in enumerate(shape):
-                if s < 0:
-                    batch_size_dim = i
-                    break
-            if batch_size_dim == -1:
-                raise ValueError("Variable {0} must has a batch size dimension",
-                                 each_var.name)
             self.feed_lod_level.append(each_var.lod_level)
-            self.feed_shapes.append(shape)
+            self.feed_shapes.append(each_var.shape)
 
         self.place = place
 

python/paddle/fluid/executor.py

@@ -382,9 +382,11 @@ class Executor(object):
         """
         Close this executor.
 
-        You can no long use this executor after calling this method.
+        You can no longer use this executor after calling this method.
         For the distributed training, this method would free the resource on PServers related to
         the current Trainer.
+        TODO(typhoonzero): Define "no longer use" meaning? Can user create
+        a new Executor for the same program and run?
         TODO(panyx0718): Why ParallelExecutor doesn't have close?
 
         Example:
@@ -397,7 +399,7 @@ class Executor(object):
             self.executor.close()
             self._closed = True
 
-    def _run_parallel(self, scope, feed, fetch_list, fetch_var_name,
+    def _run_parallel(self, program, scope, feed, fetch_list, fetch_var_name,
                       return_numpy):
         if isinstance(feed, dict):
             feed_tensor_dict = dict()
@@ -413,7 +415,7 @@ class Executor(object):
             self.executor.feed_and_split_tensor_into_local_scopes(
                 feed_tensor_dict)
         elif isinstance(feed, list) or isinstance(feed, tuple):
-            if len(feed) != len(self._places):
+            if len(feed) != len(program._places):
                 raise ValueError(
                     "Feed a list of tensor, the list should be the same size as places"
                 )
@@ -428,7 +430,7 @@ class Executor(object):
                     tensor = each[feed_name]
                     if not isinstance(tensor, core.LoDTensor):
                         tmp = core.LoDTensor()
-                        tmp.set(tensor, self._places[i])
+                        tmp.set(tensor, program._places[i])
                         tensor = tmp
                     res_dict[feed_name] = tensor
                 res.append(res_dict)
@@ -462,7 +464,7 @@ class Executor(object):
 
         Args:
             program(Program|CompiledProgram): the program that need to run,
-                if not provided, then default_main_program will be used.
+                if not provided, then default_main_program (not compiled) will be used.
             feed(dict): feed variable map, e.g. {"image": ImageData, "label": LabelData}
             fetch_list(list): a list of variable or variable names that user want to get, run will return them according to this list.
             feed_var_name(str): the name for the input variable of feed Operator.
@@ -525,6 +527,7 @@ class Executor(object):
         self.executor = program._executor
         if program._is_data_parallel:
             return self._run_parallel(
+                program,
                 scope=scope,
                 feed=feed,
                 fetch_list=fetch_list,

python/paddle/fluid/framework.py

@@ -19,6 +19,7 @@ from collections import defaultdict
 import contextlib
 import os
 import re
+import traceback
 import six
 
 import numpy as np
@@ -372,27 +373,21 @@ class Variable(object):
         self.stop_gradient = stop_gradient
         self.is_data = is_data
         if _in_imperative_mode():
-            self._ivar = core.VarBase()
+            self._ivar = kwargs.get("ivar", None)
+            if not self._ivar:
+                self._ivar = core.VarBase()
             self._ivar.desc = self.desc
             self._ivar.stop_gradient = stop_gradient
 
     def _numpy(self):
-        tensor = self._ivar.value.get_tensor()
+        tensor = self._ivar.value().get_tensor()
         return np.array(tensor)
 
     def _backward(self):
         self._ivar._run_backward()
 
     def _gradient(self):
-        return np.array(self._ivar._grad())
-
-    @property
-    def _value(self):
-        return self._ivar.value
-
-    @_value.setter
-    def _value(self, v):
-        self._ivar.value = v
+        return np.array(self._ivar._grad_value())
 
     def __str__(self):
         return self.to_string(True)
@@ -631,6 +626,11 @@ class Operator(object):
         if type is None:
             raise ValueError(
                 "`type` to initilized an Operator can not be None.")
+        else:
+            callstack_var_name = op_maker.kOpCreationCallstackAttrName()
+            op_attrs[callstack_var_name] = list(
+                reversed(traceback.format_stack()))[1:]
+
         self.desc.set_type(type)
         proto = OpProtoHolder.instance().get_op_proto(type)
 

python/paddle/fluid/imperative/base.py

@@ -45,7 +45,7 @@ def to_variable(value, block=None):
             name=None,
             shape=value.shape,
             dtype=value.dtype)
-        var = py_var._ivar.value
+        var = py_var._ivar.value()
         tensor = var.get_tensor()
         tensor.set(value, core.CPUPlace())
         return py_var

python/paddle/fluid/imperative/layers.py

@@ -20,10 +20,12 @@ from paddle.fluid import core
 from paddle.fluid import framework
 from paddle.fluid.imperative import base
 
-__all__ = ['PyLayer']
+__all__ = ['Layer', 'PyLayer']
 
 
-class PyLayer(core.Layer):
+class Layer(core.Layer):
+    """Layers composed of operators."""
+
     def __init__(self, dtype=core.VarDesc.VarType.FP32, name=None):
         self._once_built = False
         self._dtype = dtype
@@ -37,8 +39,56 @@ class PyLayer(core.Layer):
             self._once_built = True
 
         outputs = self.forward(*inputs)
-
         return outputs
 
     def forward(self, *inputs):
         raise NotImplementedError
+
+    def backward(self, *inputs):
+        raise ValueError("Layer shouldn't implement backward")
+
+
+class PyLayer(core.PyLayer):
+    """Layers composed of user-defined python codes."""
+
+    def __init__(self):
+        super(PyLayer, self).__init__()
+
+    @staticmethod
+    def forward(*inputs):
+        raise NotImplementedError
+
+    @staticmethod
+    def backward(*douts):
+        raise NotImplementedError
+
+    @classmethod
+    def __call__(cls, *inputs):
+        tracer = framework._imperative_tracer()
+        block = framework.default_main_program().current_block()
+        ivar_inputs = [x._ivar for x in inputs]
+
+        if not hasattr(cls, 'forward_id'):
+            cls.forward_id = core.PyLayer.num_funcs() + 1
+            PyLayer.register_func(cls.forward_id, cls.forward)
+            cls.backward_id = core.PyLayer.num_funcs() + 1
+            PyLayer.register_func(cls.backward_id, cls.backward)
+
+        iop = core.OpBase()
+        iop.forward_id = cls.forward_id
+        iop.backward_id = cls.backward_id
+        block.ops.append(iop)
+        ivars = tracer.py_trace(iop, ivar_inputs, False)
+        # ivars = core.PyLayer.apply(cls.forward, inputs)
+        ret = []
+        for ivar in ivars:
+            tensor = ivar.value().get_tensor()
+            py_var = framework.Variable(
+                block,
+                type=core.VarDesc.VarType.LOD_TENSOR,
+                name=None,
+                shape=tensor.shape(),
+                dtype=tensor._dtype(),
+                ivar=ivar)
+            ret.append(py_var)
+        return ret

python/paddle/fluid/imperative/nn.py

@@ -30,7 +30,7 @@ __all__ = [
 ]
 
 
-class Conv2D(layers.PyLayer):
+class Conv2D(layers.Layer):
     def __init__(self,
                  num_channels,
                  num_filters,
@@ -143,7 +143,7 @@ class Conv2D(layers.PyLayer):
         return self._helper.append_activation(pre_act)
 
 
-class Pool2D(layers.PyLayer):
+class Pool2D(layers.Layer):
     def __init__(self,
                  pool_size=-1,
                  pool_type="max",
@@ -205,7 +205,7 @@ class Pool2D(layers.PyLayer):
         return pool_out
 
 
-class FC(layers.PyLayer):
+class FC(layers.Layer):
     def __init__(self,
                  size,
                  param_attr=None,

python/paddle/fluid/layers/nn.py

@@ -58,6 +58,7 @@ __all__ = [
     'adaptive_pool2d',
     'adaptive_pool3d',
     'batch_norm',
+    'data_norm',
     'beam_search_decode',
     'conv2d_transpose',
     'conv3d_transpose',
@@ -180,6 +181,7 @@ __all__ = [
     'lstm',
     'py_func',
     'psroi_pool',
+    'teacher_student_sigmoid_loss',
     'huber_loss',
 ]
 
@@ -2896,6 +2898,133 @@ def batch_norm(input,
     return helper.append_activation(batch_norm_out)
 
 
+def data_norm(input,
+              act=None,
+              epsilon=1e-05,
+              param_attr=None,
+              data_layout='NCHW',
+              in_place=False,
+              use_mkldnn=False,
+              name=None,
+              moving_mean_name=None,
+              moving_variance_name=None,
+              do_model_average_for_mean_and_var=False):
+    """
+    **Data Normalization Layer**
+
+    Can be used as a normalizer function for conv2d and fully_connected operations.
+    The required data format for this layer is one of the following:
+
+    1. NHWC `[batch, in_height, in_width, in_channels]`
+
+    2. NCHW `[batch, in_channels, in_height, in_width]`
+
+    :math:`input` is the input features over a mini-batch.
+
+    ..  math::
+
+        \\mu_{\\beta} &\\gets \\frac{1}{m} \\sum_{i=1}^{m} x_i \\qquad &//\\
+        \ mini-batch\ mean \\\\
+        \\sigma_{\\beta}^{2} &\\gets \\frac{1}{m} \\sum_{i=1}^{m}(x_i - \\
+        \\mu_{\\beta})^2 \\qquad &//\ mini-batch\ variance \\\\
+        \\hat{x_i} &\\gets \\frac{x_i - \\mu_\\beta} {\\sqrt{\\
+        \\sigma_{\\beta}^{2} + \\epsilon}} \\qquad &//\ normalize \\\\
+        y_i &\\gets \\gamma \\hat{x_i} + \\beta \\qquad &//\ scale\ and\ shift
+
+    Args:
+        input(variable): The input variable which is a LoDTensor.
+        act(string, Default None): Activation type, linear|relu|prelu|...
+        epsilon(float, Default 1e-05):
+        param_attr(ParamAttr): The parameter attribute for Parameter `scale`.
+        data_layout(string, default NCHW): NCHW|NHWC
+        in_place(bool, Default False): Make the input and output of batch norm reuse memory.
+        use_mkldnn(bool, Default false): ${use_mkldnn_comment}
+        name(string, Default None): A name for this layer(optional). If set None, the layer
+            will be named automatically.
+        moving_mean_name(string, Default None): The name of moving_mean which store the global Mean.
+        moving_variance_name(string, Default None): The name of the moving_variance which store the global Variance.
+        do_model_average_for_mean_and_var(bool, Default False): Do model average for mean and variance or not.
+
+    Returns:
+        Variable: A tensor variable which is the result after applying data normalization on the input.
+
+    Examples:
+
+        .. code-block:: python
+
+            data = fluid.layers.data(input=x, size=200, param_attr='fc1.w')
+            hidden2 = fluid.layers.data_norm(input=hidden1)
+    """
+    helper = LayerHelper('data_norm', **locals())
+    dtype = helper.input_dtype()
+
+    input_shape = input.shape
+    if data_layout == 'NCHW':
+        channel_num = input_shape[1]
+    else:
+        if data_layout == 'NHWC':
+            channel_num = input_shape[-1]
+        else:
+            raise ValueError("unsupported data layout:" + data_layout)
+
+    param_shape = [channel_num]
+
+    batch_size_default = 1e4
+    batch_sum_default = 0.0
+    batch_square_sum_default = 1e4
+
+    if param_attr and isinstance(param_attr, dict):
+        batch_size_default = param_attr.get("batch_size", 1e4)
+        batch_sum_default = param_attr.get("batch_sum", 0.0)
+        batch_square_sum_default = param_attr.get("batch_square", 1e4)
+
+    # create parameter
+    batch_size = helper.create_parameter(
+        attr=ParamAttr(
+            name=name + '.batch_size',
+            initializer=Constant(value=float(batch_size_default)),
+            trainable=True),
+        shape=param_shape,
+        dtype=input.dtype)
+
+    batch_sum = helper.create_parameter(
+        attr=ParamAttr(
+            name=name + '.batch_sum',
+            initializer=Constant(value=float(batch_sum_default)),
+            trainable=True),
+        shape=param_shape,
+        dtype=input.dtype)
+
+    batch_square_sum = helper.create_parameter(
+        attr=ParamAttr(
+            name=name + '.batch_square_sum',
+            initializer=Constant(value=float(batch_square_sum_default)),
+            trainable=True),
+        shape=param_shape,
+        dtype=input.dtype)
+
+    means = helper.create_variable(dtype=dtype, stop_gradient=True)
+    scales = helper.create_variable(dtype=dtype, stop_gradient=True)
+
+    data_norm_out = input if in_place else helper.create_variable(dtype=dtype)
+
+    helper.append_op(
+        type="data_norm",
+        inputs={
+            "X": input,
+            "BatchSize": batch_size,
+            "BatchSum": batch_sum,
+            "BatchSquareSum": batch_square_sum
+        },
+        outputs={"Y": data_norm_out,
+                 "Means": means,
+                 "Scales": scales},
+        attrs={"epsilon": epsilon,
+               "use_mkldnn": use_mkldnn})
+
+    return helper.append_activation(data_norm_out)
+
+
 @templatedoc()
 def layer_norm(input,
                scale=True,
@@ -3064,9 +3193,9 @@ def group_norm(input,
         inputs['Bias'] = bias
 
     # create output
-    mean_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
-    variance_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
-    group_norm_out = helper.create_tmp_variable(dtype)
+    mean_out = helper.create_variable(dtype=dtype, stop_gradient=True)
+    variance_out = helper.create_variable(dtype=dtype, stop_gradient=True)
+    group_norm_out = helper.create_variable(dtype)
 
     helper.append_op(
         type="group_norm",
@@ -8350,8 +8479,7 @@ def shape(input):
     """
 
     helper = LayerHelper('shape', **locals())
-    out = helper.create_variable_for_type_inference(
-        dtype=helper.input_dtype('input'))
+    out = helper.create_variable_for_type_inference(dtype='int32')
     helper.append_op(
         type='shape', inputs={'Input': input}, outputs={'Out': out})
 
@@ -9264,6 +9392,47 @@ def log_loss(input, label, epsilon=1e-4, name=None):
     return loss
 
 
+def teacher_student_sigmoid_loss(input,
+                                 label,
+                                 soft_max_up_bound=15.0,
+                                 soft_max_lower_bound=-15.0):
+    """
+    **Teacher Student Log Loss Layer**
+
+    This layer accepts input predictions and target label and returns the
+    teacher_student loss.
+
+    .. math::
+        loss = max(x, 0) - x * z + log(1 + exp(-abs(x))) + max(x, 0) - x * z' + log(1 + exp(-abs(x)))
+
+    Args:
+        input (Variable|list):  a 2-D tensor with shape [N x 1], where N is the
+                                batch size. This input is a probability computed
+                                by the previous operator.
+        label (Variable|list):  the ground truth which is a 2-D tensor with
+                                shape [N x 1], where N is the batch size.
+        soft_max_up_bound  (float):  if input > soft_max_up_bound, will be bound 
+        soft_max_lower_bound (float): if input < soft_max_lower_bound, will be bound
+
+    Returns:
+        Variable: A 2-D tensor with shape [N x 1], the teacher_student_sigmoid_loss.
+
+    Examples:
+        .. code-block:: python
+          cost = fluid.layers.teacher_student_sigmoid_loss(input=similarity, label=label)
+    """
+    helper = LayerHelper('teacher_student_sigmoid_loss', **locals())
+    out = helper.create_variable(dtype=input.dtype)
+    helper.append_op(
+        type='teacher_student_sigmoid_loss',
+        inputs={'X': [input],
+                'Label': [label]},
+        outputs={'Y': [out]},
+        attrs={"soft_max_lower_bound": float(soft_max_lower_bound), \
+                "soft_max_up_bound": float(soft_max_up_bound)})
+    return out
+
+
 def add_position_encoding(input, alpha, beta, name=None):
     """
     **Add Position Encoding Layer**

python/paddle/fluid/optimizer.py

-#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+# 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.
@@ -195,22 +195,18 @@ class Optimizer(object):
                             format(name, param.name))
         return self._accumulators[name][param.name]
 
-    def _create_optimization_pass(self,
-                                  parameters_and_grads,
-                                  loss,
-                                  startup_program=None):
+    def _create_optimization_pass(self, parameters_and_grads):
         """Add optimization operators to update gradients to variables.
 
         Args:
-          loss(Variable): the target that this optimization is for.
           parameters_and_grads(list(tuple(Variable, Variable))):
-          a list of (variable, gradient) pair to update.
+            a list of (variable, gradient) pair to update.
 
         Returns:
           return_op_list: a list of operators that will complete one step of
-          optimization. This will include parameter update ops, global step
-          update ops and any other custom ops required by subclasses to manage
-          their internal state.
+            optimization. This will include parameter update ops, global step
+            update ops and any other custom ops required by subclasses to manage
+            their internal state.
         """
         # This is a default implementation of create_optimization_pass that
         # can be shared by most optimizers. This implementation assumes that
@@ -219,37 +215,33 @@ class Optimizer(object):
         # _create_accumulators method if it needs to create accumulators
         # for parameters and extend _finish_update method to add custom ops.
 
-        # Create any accumulators
-        program = loss.block.program
-        self._dtype = loss.dtype
-        with program_guard(program, startup_program):
-            global_block = framework.default_main_program().global_block()
-            start = len(global_block.ops)
-            self.helper = LayerHelper(self.__class__.__name__)
-            self._create_accumulators(loss.block,
-                                      [p[0] for p in parameters_and_grads])
-            self._create_global_learning_rate()
-
-            optimize_ops = []
-            for param_and_grad in parameters_and_grads:
-                if param_and_grad[1] is None:
-                    continue
-                with param_and_grad[0].block.program._optimized_guard(
-                        param_and_grad), name_scope("optimizer"):
-                    if param_and_grad[0].trainable is True:
-                        optimize_op = self._append_optimize_op(loss.block,
-                                                               param_and_grad)
-                        optimize_ops.append(optimize_op)
-
-            # Get custom finish ops for subclasses
-            # FIXME: Need to fix this once we figure out how to handle dependencies
-            self._finish_update(loss.block, parameters_and_grads)
-
-            end = len(global_block.ops)
-            return global_block._slice_ops(start, end)
-
-    def _process_distribute_lookuptable(self, param_grads, loss,
-                                        startup_program):
+        # Allways called under program_guard use global block as loss block
+        global_block = framework.default_main_program().global_block()
+        start = len(global_block.ops)
+        self.helper = LayerHelper(self.__class__.__name__)
+        self._create_accumulators(global_block,
+                                  [p[0] for p in parameters_and_grads])
+        self._create_global_learning_rate()
+
+        optimize_ops = []
+        for param_and_grad in parameters_and_grads:
+            if param_and_grad[1] is None:
+                continue
+            with param_and_grad[0].block.program._optimized_guard(
+                    param_and_grad), name_scope("optimizer"):
+                if param_and_grad[0].trainable is True:
+                    optimize_op = self._append_optimize_op(global_block,
+                                                           param_and_grad)
+                    optimize_ops.append(optimize_op)
+
+        # Get custom finish ops for subclasses
+        # FIXME: Need to fix this once we figure out how to handle dependencies
+        self._finish_update(global_block, parameters_and_grads)
+
+        end = len(global_block.ops)
+        return global_block._slice_ops(start, end)
+
+    def _process_distribute_lookuptable(self, param_grads):
         """
         Because distribute lookup table only support SGD optimizer for now, not support
         other optimizer and regularization, so we should find the table parameter out,
@@ -259,7 +251,8 @@ class Optimizer(object):
         :param loss: the loss variable.
         :param startup_program: the startup program
         """
-        program = loss.block.program
+        program = framework.default_main_program()
+        global_block = framework.default_main_program().global_block()
         table_name = find_distributed_lookup_table(program)
         table_param = None
         table_grad = None
@@ -275,38 +268,121 @@ class Optimizer(object):
                 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]})
+            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 = global_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 backward(self,
+                 loss,
+                 startup_program=None,
+                 parameter_list=None,
+                 no_grad_set=None,
+                 callbacks=None):
+        """
+        First part of `minimize`, do auto-diff to append backward ops for
+        the current program.
+
+        Args:
+            loss (Variable): loss variable to run optimizations.
+            startup_program (Program): startup_program for initializing parameters
+                in `parameter_list`.
+            parameter_list (list): list of Variables to update.
+            no_grad_set (set|None): set of Variables should be ignored.
+            callbacks (list|None): list of callables to run when appending backward
+                operator for one parameter.
+        
+        Return:
+            list: list of (param, grad) pair, grad is the output of backward.
+        
+        Examples:
+            See examples in `apply_gradients`.
+        """
+        if callbacks is None:
+            callbacks = [error_clip_callback]
+        else:
+            assert (isinstance(callbacks, list))
+            callbacks.append(error_clip_callback)
+        return append_backward(loss, parameter_list, no_grad_set, callbacks)
+
+    def apply_gradients(self, params_grads):
+        """
+        Second part of `minimize`, appending optimization operators for
+        given `params_grads` pairs.
+
+        Args:
+            params_grads (list): list of (param, grad) pair to do optimization.
+        
+        Returns:
+            list: A list of operators appended to the current program.
+        
+        Examples:
+            .. code-block:: python
+
+                loss = network()
+                optimizer = fluid.optimizer.SGD(learning_rate=0.1)
+                params_grads = optimizer.backward(loss)
+                # you may append operations for params_grads here
+                # ...
+                optimizer.apply_gradients(params_grads)
+        """
+        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)
+
+        params_grads = append_gradient_clip_ops(params_grads)
+
+        # Add regularization if any
+        params_grads = append_regularization_ops(params_grads,
+                                                 self.regularization)
+
+        optimize_ops = self._create_optimization_pass(params_grads)
+        if table_optimize_op is not None:
+            optimize_ops.append(table_optimize_op)
+            params_grads.append(table_param_and_grad)
+
+        return optimize_ops
+
     def minimize(self,
                  loss,
                  startup_program=None,
                  parameter_list=None,
                  no_grad_set=None):
-        """Add operations to minimize `loss` by updating `parameter_list`.
+        """
+        Add operations to minimize `loss` by updating `parameter_list`.
 
-        This method combines interface `append_backward()` and
-        `create_optimization_pass()` into one.
+        This method combines interface `backward()` and
+        `apply_gradients()` into one.
+        
+        Args:
+            loss (Variable): loss variable to run optimizations.
+            startup_program (Program): startup_program for initializing parameters
+                in `parameter_list`.
+            parameter_list (list): list of Variables to update.
+            no_grad_set (set|None): set of Variables should be ignored.
+
+        Returns:
+            tuple: (optimize_ops, params_grads) which are, list of operators appended;
+            and list of (param, grad) Variables pair for optimization.
         """
+        self._dtype = loss.dtype
+        program = loss.block.program
+        optimize_ops = []
         if imperative_base.enabled():
             if parameter_list is not None:
                 params_grads = parameter_list
             else:
-                program = loss.block.program
                 parameters = program.global_block().all_parameters()
                 params_grads = []
                 for param in parameters:
@@ -314,32 +390,16 @@ class Optimizer(object):
                     grad_var = Variable(
                         block=loss.block,
                         name=param._ivar._grad_name(),
-                        stop_gradient=True)
-                    grad_var._value = param._ivar.grad_value
+                        stop_gradient=True,
+                        ivar=param._ivar._grad_ivar())
                     params_grads.append((param, grad_var))
-
-            optimize_ops = self._create_optimization_pass(params_grads, loss,
-                                                          startup_program)
+            with program_guard(program, startup_program):
+                optimize_ops = self._create_optimization_pass(params_grads)
         else:
-            params_grads = append_backward(loss, parameter_list, no_grad_set,
-                                           [error_clip_callback])
-
-            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
-            params_grads = append_regularization_ops(params_grads,
-                                                     self.regularization)
-
-            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)
+            with program_guard(program, startup_program):
+                params_grads = self.backward(loss, startup_program,
+                                             parameter_list, no_grad_set)
+                optimize_ops = self.apply_gradients(params_grads)
 
         return optimize_ops, params_grads
 

python/paddle/fluid/tests/test_data_feeder.py

@@ -30,6 +30,12 @@ class TestDataFeeder(unittest.TestCase):
         self.assertEqual(result['image'].recursive_sequence_lengths(), [])
         self.assertEqual(result['label'].recursive_sequence_lengths(), [])
 
+        try:
+            result = feeder.feed([([0] * 783, [9]), ([1] * 783, [1])])
+            self.assertTrue(False)
+        except ValueError:
+            self.assertTrue(True)
+
     def test_lod_level_1_converter(self):
         # lod_level = 1
         # each sentence has a different number of words

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

@@ -107,7 +107,7 @@ if(WITH_DISTRIBUTE)
 endif()
 py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SERIAL)
 py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL)
-set_tests_properties(test_parallel_executor_fetch_feed PROPERTIES TIMEOUT 150)
+set_tests_properties(test_parallel_executor_fetch_feed PROPERTIES TIMEOUT 450)
 py_test_modules(test_parallel_executor_transformer MODULES test_parallel_executor_transformer SERIAL)
 if(NOT APPLE)
     py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL)

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

@@ -31,6 +31,7 @@ fluid.default_main_program().random_seed = 1
 
 class TestDistCTR2x2(TestDistRunnerBase):
     def get_model(self, batch_size=2):
+
         dnn_input_dim, lr_input_dim = dist_ctr_reader.load_data_meta()
         """ network definition """
         dnn_data = fluid.layers.data(
@@ -97,7 +98,14 @@ class TestDistCTR2x2(TestDistRunnerBase):
 
         inference_program = paddle.fluid.default_main_program().clone()
 
-        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.0001)
+        regularization = None
+        use_l2_decay = bool(os.getenv('USE_L2_DECAY', 0))
+        if use_l2_decay:
+            regularization = fluid.regularizer.L2DecayRegularizer(
+                regularization_coeff=1e-1)
+
+        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.0001,
+                                            regularization=regularization)
         sgd_optimizer.minimize(avg_cost)
 
         dataset = dist_ctr_reader.Dataset()

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

@@ -235,7 +235,6 @@ class DistSeResneXt2x2(TestDistRunnerBase):
 
         bd = [step * e for e in epochs]
         base_lr = 0.1
-        lr = []
         lr = [base_lr * (0.1**i) for i in range(len(bd) + 1)]
 
         optimizer = fluid.optimizer.Momentum(

python/paddle/fluid/tests/unittests/test_conv2d_int8_mkldnn_op.py

@@ -25,6 +25,15 @@ from test_conv2d_op import conv2d_forward_naive, TestConv2dOp
 def conv2d_forward_refer(input, filter, group, conv_param):
     out, in_n, out_h, out_w, out_c = conv2d_forward_naive(input, filter, group,
                                                           conv_param)
+    size = [in_n, out_c, out_h, out_w]
+    return format_reorder(out, size)
+
+
+def format_reorder(out, size):
+    in_n = size[0]
+    out_h = size[2]
+    out_w = size[3]
+    out_c = size[1]
     out_tmp = np.zeros((in_n, out_h, out_w, out_c))
     for n in range(in_n):
         for i in range(out_h):
@@ -48,6 +57,7 @@ class TestConv2dInt8Op(TestConv2dOp):
         self.init_dilation()
         self.init_test_case()
         self.init_fuse_relu()
+        self.init_fuse_residual()
         self.init_data_type()
 
         conv2d_param = {
@@ -79,11 +89,24 @@ class TestConv2dInt8Op(TestConv2dOp):
                 np.round((input_shift) * self.scale_in).astype(np.int32),
                 filter_int, self.groups,
                 conv2d_param).astype(np.float32) * scale_output_shift
-            if self.fuse_relu:
-                output = np.maximum(np.round(output1 - output2),
-                                    0).astype(self.dsttype)
+            if self.fuse_residual:
+                input_residual = np.random.randint(
+                    -5, 5, self.input_residual_size).astype(self.srctype)
+                output_tmp = np.round(output1 - output2 + format_reorder(
+                    input_residual, self.input_residual_size).astype(
+                        self.srctype) * (self.scale_out / self.scale_in_eltwise
+                                         ))
+                if self.fuse_relu:
+                    output = np.maximum(output_tmp, 0).astype(self.dsttype)
+                else:
+                    output = output_tmp.astype(self.dsttype)
             else:
-                output = np.round(output1 - output2).astype(self.dsttype)
+                if self.fuse_relu:
+                    output = np.maximum(np.round(output1 - output2),
+                                        0).astype(self.dsttype)
+                else:
+                    output = np.round(output1 - output2).astype(self.dsttype)
+
         else:
             filter_int = np.round(filter *
                                   self.scale_weights[0]).astype(np.int32)
@@ -92,21 +115,35 @@ class TestConv2dInt8Op(TestConv2dOp):
             output1 = conv2d_forward_refer(
                 input.astype(np.int32), filter_int, self.groups,
                 conv2d_param).astype(np.float32)
-            if self.fuse_relu:
-                output = np.maximum(
-                    np.round(output1 * (self.scale_out / (
-                        self.scale_in * self.scale_weights[0]))),
-                    0).astype(self.dsttype)
+            if self.fuse_residual:
+                input_residual = np.random.randint(
+                    0, 10, self.input_residual_size).astype(self.srctype)
+                output_tmp = np.round(output1 * (self.scale_out / (
+                    self.scale_in * self.scale_weights[0])) + format_reorder(
+                        input_residual, self.input_residual_size).astype(
+                            np.int32) * (self.scale_out / self.scale_in_eltwise
+                                         ))
+                output_tmp2 = np.round(output1 * (
+                    self.scale_out / (self.scale_in * self.scale_weights[0])))
+                if self.fuse_relu:
+                    output = np.maximum(output_tmp, 0).astype(self.dsttype)
+                else:
+                    output = output_tmp.astype(self.dsttype)
             else:
-                output = np.round(output1 * (self.scale_out / (
-                    self.scale_in *
-                    self.scale_weights[0]))).astype(self.dsttype)
+                if self.fuse_relu:
+                    output = np.maximum(output_tmp2, 0).astype(self.dsttype)
+                else:
+                    output = output_tmp2.astype(self.dsttype)
 
         self.inputs = {
             'Input':
             OpTest.np_dtype_to_fluid_dtype(input.astype(self.srctype)),
             'Filter': OpTest.np_dtype_to_fluid_dtype(filter)
         }
+        if self.fuse_residual:
+            self.inputs['ResidualData'] = OpTest.np_dtype_to_fluid_dtype(
+                input_residual)
+
         self.attrs = {
             'strides': self.stride,
             'paddings': self.pad,
@@ -119,7 +156,9 @@ class TestConv2dInt8Op(TestConv2dOp):
             'Scale_in': self.scale_in,
             'Scale_out': self.scale_out,
             'Scale_weights': self.scale_weights,
-            'fuse_relu': self.fuse_relu
+            'Scale_in_eltwise': self.scale_in_eltwise,
+            'fuse_relu': self.fuse_relu,
+            'fuse_residual_connection': self.fuse_residual
         }
         self.outputs = {'Output': output}
 
@@ -137,11 +176,14 @@ class TestConv2dInt8Op(TestConv2dOp):
 
     def init_test_case(self):
         TestConv2dOp.init_test_case(self)
+        self.input_size = [1, 1, 5, 5]  # NCHW
         f_c = self.input_size[1] // self.groups
-        self.filter_size = [1, f_c, 3, 3]
+        self.input_residual_size = [1, 2, 3, 3]
+        self.filter_size = [2, f_c, 3, 3]
         self.scale_in = 1.0
         self.scale_out = 0.5
         self.scale_weights = [10.0]
+        self.scale_in_eltwise = 0.6
 
     def init_data_type(self):
         self.srctype = np.uint8
@@ -150,8 +192,11 @@ class TestConv2dInt8Op(TestConv2dOp):
     def init_fuse_relu(self):
         self.fuse_relu = True
 
+    def init_fuse_residual(self):
+        self.fuse_residual = True
+
 
-#--------------------test conv2d u8 in and u8 out--------------------
+#--------------------test conv2d u8 in and u8 out with residual fuse--------------------
 
 
 class TestConv2d(TestConv2dInt8Op):
@@ -159,18 +204,21 @@ class TestConv2d(TestConv2dInt8Op):
         self.pad = [0, 0]
         self.stride = [1, 1]
         self.input_size = [2, 3, 5, 5]  # NCHW
+        self.input_residual_size = [2, 6, 3, 3]
         assert np.mod(self.input_size[1], self.groups) == 0
         f_c = self.input_size[1] // self.groups
         self.filter_size = [6, f_c, 3, 3]
         self.scale_in = 1.0
         self.scale_out = 0.5
         self.scale_weights = [10.0]
+        self.scale_in_eltwise = 0.6
 
 
 class TestWithPad(TestConv2d):
     def init_test_case(self):
         TestConv2d.init_test_case(self)
         self.pad = [1, 1]
+        self.input_residual_size = [2, 6, 5, 5]
 
 
 class TestWithGroup(TestConv2d):
@@ -183,12 +231,14 @@ class TestWithStride(TestConv2dInt8Op):
         self.pad = [1, 1]
         self.stride = [2, 2]
         self.input_size = [2, 3, 6, 6]
+        self.input_residual_size = [2, 6, 3, 3]
         assert np.mod(self.input_size[1], self.groups) == 0
         f_c = self.input_size[1] // self.groups
         self.filter_size = [6, f_c, 3, 3]
         self.scale_in = 1.0
         self.scale_out = 0.8
         self.scale_weights = [10.0]
+        self.scale_in_eltwise = 0.5
 
 
 class TestWith1x1(TestConv2dInt8Op):
@@ -196,12 +246,14 @@ class TestWith1x1(TestConv2dInt8Op):
         self.pad = [0, 0]
         self.stride = [1, 1]
         self.input_size = [1, 3, 5, 5]
+        self.input_residual_size = [1, 6, 5, 5]
         assert np.mod(self.input_size[1], self.groups) == 0
         f_c = self.input_size[1] // self.groups
         self.filter_size = [6, f_c, 1, 1]
         self.scale_in = 1.0
         self.scale_out = 0.5
         self.scale_weights = [12.0]
+        self.scale_in_eltwise = 0.5
 
 
 class TestWithInput1x1Filter1x1(TestConv2dInt8Op):
@@ -209,24 +261,29 @@ class TestWithInput1x1Filter1x1(TestConv2dInt8Op):
         self.pad = [0, 0]
         self.stride = [1, 1]
         self.input_size = [2, 3, 1, 1]
+        self.input_residual_size = [2, 6, 1, 1]
         assert np.mod(self.input_size[1], self.groups) == 0
         f_c = self.input_size[1] // self.groups
         self.filter_size = [6, f_c, 1, 1]
         self.scale_in = 1.0
         self.scale_out = 0.5
         self.scale_weights = [10.0]
+        self.scale_in_eltwise = 0.8
 
     def init_group(self):
         self.groups = 3
 
 
-def init_data_type_with_fusion(self, input_dt, fuse_relu):
+def init_data_type_with_fusion(self, input_dt, fuse_relu, fuse_residual):
     self.srctype = input_dt
     self.dsttype = np.uint8 if fuse_relu else np.int8
 
     def init_fuse_relu(self):
         self.fuse_relu = fuse_relu
 
+    def init_fuse_residual(self):
+        self.fuse_residual = fuse_residual
+
 
 def create_test_int8_class(parent):
 
@@ -234,29 +291,68 @@ def create_test_int8_class(parent):
 
     class TestS8U8Case(parent):
         def init_data_type(self):
-            init_data_type_with_fusion(self, np.int8, True)
+            init_data_type_with_fusion(self, np.int8, True, False)
 
     #--------------------test conv2d s8 in and s8 out--------------------
 
     class TestS8S8Case(parent):
         def init_data_type(self):
-            init_data_type_with_fusion(self, np.int8, False)
+            init_data_type_with_fusion(self, np.int8, False, False)
 
     #--------------------test conv2d u8 in and s8 out--------------------
 
     class TestU8S8Case(parent):
         def init_data_type(self):
-            init_data_type_with_fusion(self, np.uint8, False)
+            init_data_type_with_fusion(self, np.uint8, False, False)
+
+    #--------------------test conv2d u8 in and u8 out without residual fuse--------------------
+
+    class TestU8U8Case(parent):
+        def init_data_type(self):
+            init_data_type_with_fusion(self, np.uint8, True, False)
 
-    cls_name_s8u8 = "{0}_relu_{1}".format(parent.__name__, "1")
-    cls_name_s8s8 = "{0}_relu_{1}".format(parent.__name__, "0")
-    cls_name_u8s8 = "{0}_relu_{1}".format(parent.__name__, "0")
+    #--------------------test conv2d s8 in and u8 out with residual fuse--------------------
+
+    class TestS8U8ResCase(parent):
+        def init_data_type(self):
+            init_data_type_with_fusion(self, np.int8, True, True)
+
+    #--------------------test conv2d s8 in and s8 out with residual fuse--------------------
+
+    class TestS8S8ResCase(parent):
+        def init_data_type(self):
+            init_data_type_with_fusion(self, np.int8, False, True)
+
+    #--------------------test conv2d u8 in and s8 out with residual fuse--------------------
+
+    class TestU8S8ResCase(parent):
+        def init_data_type(self):
+            init_data_type_with_fusion(self, np.uint8, False, True)
+
+    cls_name_s8u8 = "{0}_relu_{1}_residual_0".format(parent.__name__, "1")
+    cls_name_s8s8 = "{0}_relu_{1}_residual_0".format(parent.__name__, "0")
+    cls_name_u8s8 = "{0}_relu_{1}_residual_0".format(parent.__name__, "0")
+    cls_name_u8u8 = "{0}_relu_{1}_residual_0".format(parent.__name__, "1")
+    cls_name_s8u8_re_1 = "{0}_relu_{1}_residual_{2}".format(parent.__name__,
+                                                            "1", "1")
+    cls_name_s8s8_re_1 = "{0}_relu_{1}_residual_{2}".format(parent.__name__,
+                                                            "0", "1")
+    cls_name_u8s8_re_1 = "{0}_relu_{1}_residual_{2}".format(parent.__name__,
+                                                            "0", "1")
     TestS8U8Case.__name__ = cls_name_s8u8
     TestS8S8Case.__name__ = cls_name_s8s8
     TestU8S8Case.__name__ = cls_name_u8s8
+    TestU8U8Case.__name__ = cls_name_u8u8
+    TestS8U8ResCase.__name__ = cls_name_s8u8_re_1
+    TestS8S8ResCase.__name__ = cls_name_s8s8_re_1
+    TestU8S8ResCase.__name__ = cls_name_u8s8_re_1
     globals()[cls_name_s8u8] = TestS8U8Case
     globals()[cls_name_s8s8] = TestS8S8Case
     globals()[cls_name_u8s8] = TestU8S8Case
+    globals()[cls_name_u8u8] = TestU8U8Case
+    globals()[cls_name_s8u8_re_1] = TestS8U8ResCase
+    globals()[cls_name_s8s8_re_1] = TestS8S8ResCase
+    globals()[cls_name_u8s8_re_1] = TestU8S8ResCase
 
 
 create_test_int8_class(TestConv2dInt8Op)

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

@@ -18,7 +18,6 @@ import unittest
 from test_dist_base import TestDistBase
 
 
-# FIXME(tangwei): sum op can not handle when inputs is empty.
 class TestDistCTR2x2(TestDistBase):
     def _setup_config(self):
         self._sync_mode = True
@@ -28,5 +27,19 @@ class TestDistCTR2x2(TestDistBase):
         self.check_with_place("dist_ctr.py", delta=1e-7, check_error_log=False)
 
 
+class TestDistCTRWithL2Decay2x2(TestDistBase):
+    def _setup_config(self):
+        self._sync_mode = True
+        self._enforce_place = "CPU"
+
+    def test_dist_ctr(self):
+        need_envs = {"USE_L2_DECAY": "1"}
+        self.check_with_place(
+            "dist_ctr.py",
+            delta=1e-7,
+            check_error_log=False,
+            need_envs=need_envs)
+
+
 if __name__ == "__main__":
     unittest.main()

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

@@ -22,6 +22,7 @@ import unittest
 import paddle
 import paddle.fluid.core as core
 import paddle.fluid as fluid
+from paddle.fluid import compiler
 
 
 def train(network, use_cuda, use_parallel_executor, batch_size=32, pass_num=2):
@@ -57,19 +58,19 @@ def train(network, use_cuda, use_parallel_executor, batch_size=32, pass_num=2):
     exe = fluid.Executor(place)
     exe.run(fluid.default_startup_program())
 
+    train_cp = compiler.CompiledProgram(fluid.default_main_program())
     if use_parallel_executor:
-        train_exe = fluid.ParallelExecutor(
-            use_cuda=use_cuda, loss_name=cost.name)
+        train_cp = train_cp.with_data_parallel(loss_name=cost.name)
         fetch_list = [cost.name]
     else:
-        train_exe = exe
         fetch_list = [cost]
 
     for pass_id in six.moves.xrange(pass_num):
         batch_id = 0
         for data in reader():
-            train_exe.run(feed=data,
-                          fetch_list=fetch_list if batch_id % 4 == 0 else [])
+            exe.run(train_cp,
+                    feed=data,
+                    fetch_list=fetch_list if batch_id % 4 == 0 else [])
             batch_id += 1
             if batch_id > 16:
                 break

python/paddle/fluid/tests/unittests/test_fusion_seqpool_concat_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
+from op_test import OpTest
+from test_reorder_lod_tensor import convert_to_offset
+from test_seq_pool import compute_seqpool_sum, compute_seqpool_avg, compute_seqpool_sqrt
+
+
+class TestFusionSeqPoolConcatOp(OpTest):
+    def setUp(self):
+        self.w = 11
+        self.lods = [[[2, 3, 5]], [[1, 5, 2]]]
+        self.set_conf()
+        self.set_pooltype()
+        self.op_type = 'fusion_seqpool_concat'
+        self.axis = 1
+        bs = len(self.lods[0][0])
+        inputs = []
+        outs = []
+        i = 0
+        for lod in self.lods:
+            assert bs == len(lod[0]), 'All lod size should be equal'
+            x = np.random.uniform(0.1, 1,
+                                  [sum(lod[0]), self.w]).astype('float32')
+            offset = convert_to_offset(lod)
+            out = np.zeros((bs, self.w)).astype('float32')
+            if self.pooltype == "SUM":
+                compute_seqpool_sum(x, offset, out)
+            elif self.pooltype == "AVERAGE":
+                compute_seqpool_avg(x, offset, out)
+            elif self.pooltype == "SQRT":
+                compute_seqpool_sqrt(x, offset, out)
+            else:
+                raise Exception("Unsupported pool type!")
+            inputs.append(('x_{0}'.format(i), (x, lod)))
+            outs.append(out)
+            i = i + 1
+
+        self.inputs = {'X': inputs}
+        self.outputs = {'Out': np.concatenate(outs, axis=self.axis)}
+        self.attrs = {
+            'pooltype': self.pooltype,
+            'axis': self.axis,
+        }
+
+    def set_pooltype(self):
+        self.pooltype = "SUM"
+
+    def set_conf(self):
+        pass
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestFusionSeqPoolConcatOpCase1(TestFusionSeqPoolConcatOp):
+    def set_conf(self):
+        self.lods = [[[1]]]
+
+
+class TestFusionSeqPoolConcatOpCase2(TestFusionSeqPoolConcatOp):
+    def set_conf(self):
+        self.lods = [[[1]], [[1]], [[1]]]
+
+
+class TestFusionSeqPoolConcatOpCase3(TestFusionSeqPoolConcatOp):
+    def set_conf(self):
+        self.lods = [[[1, 3, 4, 6]]]
+        self.w = 10
+
+
+class TestFusionSeqPoolConcatOpCase4(TestFusionSeqPoolConcatOp):
+    def set_conf(self):
+        self.lods = [[[2, 13, 4]], [[1, 1, 1]], [[5, 3, 1]], [[9, 10, 3]]]
+        self.w = 3
+
+
+## test avg pool and sqrt
+def create_test_avg_sqrt_class(parent):
+    class TestSeqPoolAvgCase(parent):
+        def set_pooltype(self):
+            self.pooltype = "AVERAGE"
+
+    class TestSeqPoolSqrtCase(parent):
+        def set_pooltype(self):
+            self.pooltype = "SQRT"
+
+    cls_name_avg = "{0}_{1}".format(parent.__name__, "avg")
+    cls_name_sqrt = "{0}_{1}".format(parent.__name__, "sqrt")
+    TestSeqPoolAvgCase.__name__ = cls_name_avg
+    TestSeqPoolSqrtCase.__name__ = cls_name_sqrt
+    globals()[cls_name_avg] = TestSeqPoolAvgCase
+    globals()[cls_name_sqrt] = TestSeqPoolSqrtCase
+
+
+create_test_avg_sqrt_class(TestFusionSeqPoolConcatOp)
+create_test_avg_sqrt_class(TestFusionSeqPoolConcatOpCase1)
+create_test_avg_sqrt_class(TestFusionSeqPoolConcatOpCase2)
+create_test_avg_sqrt_class(TestFusionSeqPoolConcatOpCase3)
+create_test_avg_sqrt_class(TestFusionSeqPoolConcatOpCase4)
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -15,6 +15,7 @@
 import contextlib
 import unittest
 import numpy as np
+import sys
 
 import paddle.fluid as fluid
 from paddle.fluid import core
@@ -22,7 +23,7 @@ from paddle.fluid.imperative.nn import FC
 from test_imperative_base import new_program_scope
 
 
-class MyLayer(fluid.imperative.PyLayer):
+class MyLayer(fluid.imperative.Layer):
     def __init__(self):
         super(MyLayer, self).__init__()
 
@@ -34,7 +35,35 @@ class MyLayer(fluid.imperative.PyLayer):
         return [x]
 
 
-class MLP(fluid.imperative.PyLayer):
+class MyPyLayer(fluid.imperative.PyLayer):
+    def __init__(self):
+        super(MyPyLayer, self).__init__()
+
+    @staticmethod
+    def forward(inputs):
+        sys.stderr.write('before forward\n')
+        ret = np.tanh(inputs[0])
+        sys.stderr.write('after forward: %s\n' % ret)
+        tensor = core.LoDTensor()
+        tensor.set(ret, core.CPUPlace())
+        return tuple([tensor])
+
+    @staticmethod
+    def backward(inputs):
+        sys.stderr.write('calling into backward: %s\n' % str(inputs))
+        inp, out, dout = inputs
+        inp = np.array(inp)
+        out = np.array(out)
+        dout = np.array(dout)
+        sys.stderr.write('calling into backward: %s, %s, %s\n' %
+                         (inp, out, dout))
+        ret = np.array(dout) * (1 - np.square(np.array(out)))
+        tensor = core.LoDTensor()
+        tensor.set(ret, core.CPUPlace())
+        return tuple([tensor])
+
+
+class MLP(fluid.imperative.Layer):
     def __init__(self):
         super(MLP, self).__init__()
         self._fc1 = FC(3,
@@ -56,9 +85,77 @@ class TestImperative(unittest.TestCase):
         with fluid.imperative.guard():
             cl = core.Layer()
             cl.forward([])
-            l = fluid.imperative.PyLayer()
+            l = fluid.imperative.Layer()
             self.assertRaises(NotImplementedError, l.forward, [])
 
+    def test_pylayer_func_id(self):
+
+        with fluid.imperative.guard():
+
+            class PyLayer1(fluid.imperative.PyLayer):
+                def __init__(self):
+                    super(PyLayer1, self).__init__()
+
+                @staticmethod
+                def forward(input):
+                    return input
+
+                @staticmethod
+                def backward(input):
+                    return input
+
+            class PyLayer2(fluid.imperative.PyLayer):
+                def __init__(self):
+                    super(PyLayer2, self).__init__()
+
+                @staticmethod
+                def forward(input):
+                    return input
+
+                @staticmethod
+                def backward(input):
+                    return input
+
+            py_layer_1 = PyLayer1()
+            py_layer_2 = PyLayer2()
+            py_layer_1(fluid.imperative.base.to_variable(np.ones([2, 2])))
+            py_layer_2(fluid.imperative.base.to_variable(np.ones([2, 2])))
+            id = py_layer_1.forward_id
+            self.assertGreater(id, 0)
+            self.assertEqual(py_layer_1.backward_id, id + 1)
+            self.assertEqual(py_layer_2.forward_id, id + 2)
+            self.assertEqual(py_layer_2.backward_id, id + 3)
+            py_layer_1(fluid.imperative.base.to_variable(np.ones([2, 2])))
+            self.assertEqual(py_layer_1.forward_id, id)
+
+    def test_pylayer(self):
+        np_inp = np.ones([2, 2], np.float32)
+        with fluid.imperative.guard():
+            my_py_layer = MyPyLayer()
+            var_inp = fluid.imperative.base.to_variable(np_inp)
+            outs = my_py_layer(var_inp)
+            dy_out = np.sum(outs[0]._numpy())
+            outs[0]._backward()
+            dy_grad = var_inp._gradient()
+
+        with new_program_scope():
+            inp = fluid.layers.data(
+                name="inp", shape=[2, 2], append_batch_size=False)
+            # TODO(panyx0718): Paddle doesn't diff against data `inp`.
+            x1 = inp * 1
+            # TODO(panyx0718): If reduce_sum is skipped, the result is wrong.
+            x = fluid.layers.reduce_sum(fluid.layers.tanh(x1))
+            param_grads = fluid.backward.append_backward(
+                x, parameter_list=[x1.name])[0]
+            exe = fluid.Executor(fluid.CPUPlace())
+
+            static_out, static_grad = exe.run(
+                feed={inp.name: np_inp},
+                fetch_list=[x.name, param_grads[1].name])
+
+        self.assertTrue(np.allclose(dy_out, static_out))
+        self.assertTrue(np.allclose(dy_grad, static_grad))
+
     def test_layer_in_out(self):
         np_inp = np.array([1.0, 2.0, -1.0], dtype=np.float32)
         with fluid.imperative.guard():

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

@@ -26,7 +26,7 @@ from paddle.fluid.imperative.base import to_variable
 from test_imperative_base import new_program_scope
 
 
-class SimpleImgConvPool(fluid.imperative.PyLayer):
+class SimpleImgConvPool(fluid.imperative.Layer):
     def __init__(self,
                  num_channels,
                  num_filters,
@@ -72,7 +72,7 @@ class SimpleImgConvPool(fluid.imperative.PyLayer):
         return x
 
 
-class MNIST(fluid.imperative.PyLayer):
+class MNIST(fluid.imperative.Layer):
     def __init__(self, param_attr=None, bias_attr=None):
         super(MNIST, self).__init__()
 
@@ -105,7 +105,6 @@ class TestImperativeMnist(unittest.TestCase):
             fluid.default_startup_program().random_seed = seed
             fluid.default_main_program().random_seed = seed
 
-            #  mnist = Conv2D(1, 20, 5)
             mnist = MNIST()
             sgd = SGDOptimizer(learning_rate=1e-3)
             train_reader = paddle.batch(
@@ -126,16 +125,17 @@ class TestImperativeMnist(unittest.TestCase):
                 label._stop_gradient = True
 
                 cost = mnist(img)
-                loss = fluid.layers.reduce_mean(cost)
-                dy_out = loss._numpy()
+                loss = fluid.layers.cross_entropy(cost, label)
+                avg_loss = fluid.layers.mean(loss)
+                dy_out = avg_loss._numpy()
 
                 if batch_id == 0:
                     for param in fluid.default_main_program().global_block(
                     ).all_parameters():
                         dy_param_init_value[param.name] = param._numpy()
 
-                loss._backward()
-                sgd.minimize(loss)
+                avg_loss._backward()
+                sgd.minimize(avg_loss)
                 dy_param_value = {}
                 for param in fluid.default_main_program().global_block(
                 ).all_parameters():
@@ -147,7 +147,6 @@ class TestImperativeMnist(unittest.TestCase):
 
             exe = fluid.Executor(fluid.CPUPlace())
 
-            #  mnist = Conv2D(1, 20, 5)
             mnist = MNIST()
             sgd = SGDOptimizer(learning_rate=1e-3)
             train_reader = paddle.batch(
@@ -157,8 +156,9 @@ class TestImperativeMnist(unittest.TestCase):
                 name='pixel', shape=[1, 28, 28], dtype='float32')
             label = fluid.layers.data(name='label', shape=[1], dtype='int64')
             cost = mnist(img)
-            loss = fluid.layers.reduce_mean(cost)
-            sgd.minimize(loss)
+            loss = fluid.layers.cross_entropy(cost, label)
+            avg_loss = fluid.layers.mean(loss)
+            sgd.minimize(avg_loss)
 
             # initialize params and fetch them
             static_param_init_value = {}
@@ -182,7 +182,7 @@ class TestImperativeMnist(unittest.TestCase):
                 y_data = np.array([x[1] for x in data]).astype('int64').reshape(
                     [128, 1])
 
-                fetch_list = [loss.name]
+                fetch_list = [avg_loss.name]
                 fetch_list.extend(static_param_name_list)
                 out = exe.run(fluid.default_main_program(),
                               feed={"pixel": x_data,

python/paddle/fluid/tests/unittests/test_ir_graph.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.
+
+import os
+import unittest
+import six
+from paddle import fluid
+
+
+class TestIRGraph(unittest.TestCase):
+    """
+    TODO(fc500110): `resolve_hazard` api will be tested when it can be used.
+    """
+
+    def test_nodes(self):
+        graph = build_graph()
+        self.assertTrue(
+            {node.name()
+             for node in graph.nodes()} == {"x1", "x2", "out", "sum"})
+
+    def test_has_set_get(self):
+        graph = build_graph()
+        for attr_name in ["int", "float", "string"]:
+            self.assertFalse(graph.has(attr_name))
+        graph.set("int", 1)
+        graph.set("float", 0.5)
+        graph.set("string", "string")
+        for attr_name in ["int", "float", "string"]:
+            self.assertTrue(graph.has(attr_name))
+
+        self.assertTrue(graph.get_int("int") == 1)
+        self.assertTrue(graph.get_float("float") == 0.5)
+        self.assertTrue(graph.get_string("string") == "string")
+
+    def test_erase(self):
+        graph = build_graph()
+        graph.set("test", 0)
+        self.assertTrue(graph.has("test"))
+        graph.erase("test")
+        self.assertFalse(graph.has("test"))
+
+    def test_create_var_node(self):
+        prog = fluid.core.ProgramDesc()
+        block = prog.block(0)
+        shape = [10, 20]
+        x1 = block.var(six.b("x1"))
+        x1.set_type(fluid.core.VarDesc.VarType.LOD_TENSOR)
+        x1.set_shape(shape)
+        graph = fluid.core.Graph(prog)
+        node = graph.create_var_node(x1)
+        self.assertTrue(node.node_type() == fluid.core.Node.Type.Variable)
+
+    def test_create_op_node(self):
+        prog = fluid.core.ProgramDesc()
+        block = prog.block(0)
+        sum_op_desc = block.append_op()
+        graph = fluid.core.Graph(prog)
+        node = graph.create_op_node(sum_op_desc)
+        self.assertTrue(node.node_type() == fluid.core.Node.Type.Operation)
+
+    def test_create_control_dep_var(self):
+        graph = build_graph()
+        name = "__control_var@{}".format(len(graph.nodes()))
+        node = graph.create_control_dep_var()
+        self.assertTrue(node.name() == name)
+
+    def test_create_empty_node(self):
+        prog = fluid.core.ProgramDesc()
+        graph = fluid.core.Graph(prog)
+        n1 = graph.create_empty_node('x', fluid.core.Node.Type.Operation)
+        self.assertTrue(n1.name() == 'x')
+        n2 = graph.create_empty_node('y', fluid.core.Node.Type.Variable)
+        self.assertTrue(n2.name() == 'y')
+
+    def test_release_nodes(self):
+        graph = build_graph()
+        nodes = graph.release_nodes()
+        self.assertTrue(len(graph.nodes()) == 0)
+        self.assertTrue({node.name()
+                         for node in nodes} == {"x1", "x2", "out", "sum"})
+
+    def test_remove_node(self):
+        graph = build_graph()
+        nodes = graph.nodes()
+        for node in nodes:
+            if node.name() == "sum":
+                break
+        self.assertTrue({node.name()
+                         for node in nodes} == {"x1", "x2", "out", "sum"})
+        nodes.remove(node)
+        self.assertTrue({node.name() for node in nodes} == {"x1", "x2", "out"})
+
+    def test_retrieve_node(self):
+        graph = build_graph()
+        nodes = []
+        for i in range(len(graph.nodes())):
+            nodes.append(graph.retrieve_node(i))
+
+        for node in nodes:
+            self.assertTrue(node in graph.nodes())
+
+    def resolve_hazard(self):
+        pass
+
+
+def build_graph():
+    prog = fluid.core.ProgramDesc()
+    block = prog.block(0)
+
+    shape = [10, 20]
+
+    # prepare input/output
+    x1 = block.var(six.b("x1"))
+    x1.set_type(fluid.core.VarDesc.VarType.LOD_TENSOR)
+    x1.set_shape(shape)
+    x2 = block.var(six.b("x2"))
+    x2.set_type(fluid.core.VarDesc.VarType.LOD_TENSOR)
+    x2.set_shape(shape)
+
+    out = block.var(six.b("out"))
+    out.set_type(fluid.core.VarDesc.VarType.LOD_TENSOR)
+
+    sum_op_desc = block.append_op()
+    sum_op_desc.set_type("sum")
+    sum_op_desc.set_input("X", ["x1", "x2"])
+    sum_op_desc.set_output("Out", ["out"])
+
+    sum_op_desc.check_attrs()
+    sum_op_desc.infer_shape(block)
+    graph = fluid.core.Graph(prog)
+    return graph
+
+
+if __name__ == "__main__":
+    unittest.main()

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

@@ -69,7 +69,7 @@ class TestOperator(unittest.TestCase):
             set(mul_op.attr_names),
             set([
                 "x_num_col_dims", "y_num_col_dims", "op_role", "op_role_var",
-                "op_namescope"
+                "op_namescope", "op_callstack"
             ]))
         self.assertEqual(mul_op.has_attr("x_num_col_dims"), True)
         self.assertEqual(mul_op.attr_type("x_num_col_dims"), core.AttrType.INT)

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

@@ -61,6 +61,48 @@ class TestOptimizer(unittest.TestCase):
         self.assertEqual([op.type for op in opts], ["sgd"])
 
 
+class TestOptimizerBackwardApplygrad(unittest.TestCase):
+    def test_sgd_optimizer(self):
+        def check_sgd_optimizer(optimizer_attr):
+            init_program = framework.Program()
+            program = framework.Program()
+            block = program.global_block()
+            mul_x = block.create_parameter(
+                dtype="float32",
+                shape=[5, 10],
+                lod_level=0,
+                name="mul.x",
+                optimize_attr=optimizer_attr)
+            mul_y = block.create_var(
+                dtype="float32", shape=[10, 8], lod_level=0, name="mul.y")
+            mul_out = block.create_var(
+                dtype="float32", shape=[5, 8], lod_level=0, name="mul.out")
+            mean_out = block.create_var(
+                dtype="float32", shape=[1], lod_level=0, name="mean.out")
+            block.append_op(
+                type="mul",
+                inputs={"X": mul_x,
+                        "Y": mul_y},
+                outputs={"Out": mul_out},
+                attrs={"x_num_col_dims": 1})
+            block.append_op(
+                type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out})
+            sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.01)
+            with framework.program_guard(program, init_program):
+                p_g = sgd_optimizer.backward(mean_out)
+                opts = sgd_optimizer.apply_gradients(p_g)
+            return opts
+
+        opts = check_sgd_optimizer({'learning_rate': 1.1})
+        self.assertEqual(len(opts), 3)
+        self.assertEqual([op.type for op in opts],
+                         ["fill_constant", "elementwise_mul", "sgd"])
+
+        opts = check_sgd_optimizer({'learning_rate': 1.0})
+        self.assertEqual(len(opts), 1)
+        self.assertEqual([op.type for op in opts], ["sgd"])
+
+
 class TestMomentumOptimizer(unittest.TestCase):
     class MockMomentum(optimizer.MomentumOptimizer):
         def get_accumulators(self):
@@ -99,8 +141,8 @@ class TestMomentumOptimizer(unittest.TestCase):
         params_grads = append_backward(mean_out)
         self.assertEqual(len(params_grads), 1)
         self.assertEqual(len(momentum_optimizer.get_accumulators()), 0)
-        opts = momentum_optimizer._create_optimization_pass(
-            params_grads, mul_out, init_program)
+        with framework.program_guard(program, init_program):
+            opts = momentum_optimizer.apply_gradients(params_grads)
         self.assertEqual(len(opts), 3)
         sgd_op = opts[-1]
         self.assertEqual([op.type for op in opts],
@@ -153,8 +195,8 @@ class TestMomentumOptimizer(unittest.TestCase):
         params_grads = append_backward(mean_out)
         self.assertEqual(len(params_grads), 1)
         self.assertEqual(len(momentum_optimizer.get_accumulators()), 0)
-        opts = momentum_optimizer._create_optimization_pass(
-            params_grads, mul_out, init_program)
+        with framework.program_guard(program, init_program):
+            opts = momentum_optimizer.apply_gradients(params_grads)
         self.assertEqual(len(opts), 3)
         sgd_op = opts[-1]
         self.assertEqual([op.type for op in opts],
@@ -216,8 +258,8 @@ class TestAdagradOptimizer(unittest.TestCase):
         params_grads = append_backward(mean_out)
         self.assertEqual(len(params_grads), 1)
         self.assertEqual(len(adagrad_optimizer.get_accumulators()), 0)
-        opts = adagrad_optimizer._create_optimization_pass(
-            params_grads, mul_out, init_program)
+        with framework.program_guard(program, init_program):
+            opts = adagrad_optimizer.apply_gradients(params_grads)
         self.assertEqual(len(opts), 3)
         self.assertEqual([op.type for op in opts],
                          ["fill_constant", "elementwise_mul", "adagrad"])
@@ -280,8 +322,8 @@ class TestAdamOptimizer(unittest.TestCase):
         params_grads = append_backward(mean_out)
         self.assertEqual(len(params_grads), 1)
         self.assertEqual(len(adam_optimizer.get_accumulators()), 0)
-        opts = adam_optimizer._create_optimization_pass(params_grads, mul_out,
-                                                        init_program)
+        with framework.program_guard(program, init_program):
+            opts = adam_optimizer.apply_gradients(params_grads)
         self.assertEqual(len(opts), 5)
         self.assertEqual(
             [op.type for op in opts],
@@ -347,8 +389,8 @@ class TestAdamaxOptimizer(unittest.TestCase):
         params_grads = append_backward(mean_out)
         self.assertEqual(len(params_grads), 1)
         self.assertEqual(len(adamax_optimizer.get_accumulators()), 0)
-        opts = adamax_optimizer._create_optimization_pass(params_grads, mul_out,
-                                                          init_program)
+        with framework.program_guard(program, init_program):
+            opts = adamax_optimizer.apply_gradients(params_grads)
         self.assertEqual(len(opts), 4)
         self.assertEqual(
             [op.type for op in opts],
@@ -411,8 +453,8 @@ class TestDecayedAdagradOptimizer(unittest.TestCase):
         params_grads = append_backward(mean_out)
         self.assertEqual(len(params_grads), 1)
         self.assertEqual(len(decayed_adagrad_optimizer.get_accumulators()), 0)
-        opts = decayed_adagrad_optimizer._create_optimization_pass(
-            params_grads, mul_out, init_program)
+        with framework.program_guard(program, init_program):
+            opts = decayed_adagrad_optimizer.apply_gradients(params_grads)
         self.assertEqual(len(opts), 3)
         self.assertEqual(
             [op.type for op in opts],
@@ -477,8 +519,8 @@ class TestFtrlOptimizer(unittest.TestCase):
         params_grads = append_backward(mean_out)
         self.assertEqual(len(params_grads), 1)
         self.assertEqual(len(ftrl_optimizer.get_accumulators()), 0)
-        opts = ftrl_optimizer._create_optimization_pass(params_grads, mul_out,
-                                                        init_program)
+        with framework.program_guard(program, init_program):
+            opts = ftrl_optimizer.apply_gradients(params_grads)
         self.assertEqual(len(opts), 3)
         self.assertEqual([op.type for op in opts],
                          ["fill_constant", "elementwise_mul", "ftrl"])

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

@@ -16,6 +16,7 @@ from __future__ import print_function
 
 import paddle.dataset.conll05 as conll05
 import paddle.fluid as fluid
+from paddle.fluid import compiler
 import paddle.fluid.core as core
 import unittest
 import paddle
@@ -157,10 +158,8 @@ class TestCRFModel(unittest.TestCase):
             exe = fluid.Executor(place)
             exe.run(startup)
 
-            pe = fluid.ParallelExecutor(
-                use_cuda=use_cuda,
-                loss_name=avg_cost.name,
-                build_strategy=build_strategy)
+            train_cp = compiler.CompiledProgram(main).with_data_parallel(
+                loss_name=avg_cost.name, build_strategy=build_strategy)
 
             feeder = fluid.DataFeeder(
                 feed_list=[
@@ -172,8 +171,9 @@ class TestCRFModel(unittest.TestCase):
             data = train_data()
             for i in range(10):
                 cur_batch = next(data)
-                print(pe.run(feed=feeder.feed(cur_batch),
-                             fetch_list=[avg_cost.name])[0])
+                print(exe.run(train_cp,
+                              feed=feeder.feed(cur_batch),
+                              fetch_list=[avg_cost.name])[0])
 
     def _new_build_strategy(self, use_reduce=False):
         build_strategy = fluid.BuildStrategy()

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

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import paddle.fluid as fluid
+from paddle.fluid import compiler
 import unittest
 import logging
 import six
@@ -36,21 +37,18 @@ class TestBase(unittest.TestCase):
         with fluid.program_guard(main_prog, startup_prog):
             with fluid.scope_guard(scope):
                 loss = network_func()
-                fluid.Executor(
-                    fluid.CUDAPlace(0)
-                    if use_gpu else fluid.CPUPlace()).run(startup_prog)
+                exe = fluid.Executor(
+                    fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace())
+                exe.run(startup_prog)
 
         for _ in six.moves.xrange(iter):
             exe_strategy = fluid.ExecutionStrategy()
             exe_strategy._dry_run = True
             exe_strategy.use_experimental_executor = use_experimental_executor
-            pe = fluid.ParallelExecutor(
-                use_cuda=use_gpu,
-                loss_name=loss.name,
-                main_program=main_prog,
-                exec_strategy=exe_strategy)
+            train_cp = compiler.CompiledProgram(main_prog).with_data_parallel(
+                loss_name=loss.name, exec_strategy=exe_strategy)
             for _ in six.moves.xrange(iter_per_pe):
-                pe.run([])
+                exe.run(train_cp)
 
 
 class TestMNISTDryRun(TestBase):

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

@@ -14,13 +14,12 @@
 
 from __future__ import print_function
 
-import paddle.dataset.flowers as flowers
 import math
 import paddle.fluid as fluid
+from paddle.fluid import compiler
 import paddle.fluid.core as core
 import unittest
 import numpy as np
-import paddle
 import os
 
 
@@ -38,114 +37,112 @@ def Lenet(data, class_dim):
     return fc2
 
 
-class TestFetchOp(unittest.TestCase):
-    def parallel_exe(self, train_inputs, seed, use_cuda):
-        main = fluid.Program()
+class TestFetchAndFeed(unittest.TestCase):
+    def parallel_exe(self, use_cuda, run_parallel_exe, seed=1):
+        main_program = fluid.Program()
         startup = fluid.Program()
         startup.random_seed = seed
-        with fluid.program_guard(main, startup):
+        with fluid.program_guard(main_program, startup):
             data = fluid.layers.data(
                 name='image', shape=[3, 224, 224], dtype='float32')
             label = fluid.layers.data(name='label', shape=[1], dtype='int64')
             out = Lenet(data, class_dim=102)
             loss = fluid.layers.cross_entropy(input=out, label=label)
             loss = fluid.layers.mean(loss)
-
             opt = fluid.optimizer.Momentum(
                 learning_rate=0.1,
                 momentum=0.9,
                 regularization=fluid.regularizer.L2Decay(1e-4))
-
             opt.minimize(loss)
 
-            # TODO(zcd): I found that onece the memory optimizer is open,
-            # parallel_exe doesn't fetch some variable, such as conv2d_0.b_0@GRAD,
-            # conv2d_1.b_0@GRAD. Those variables should not be pruned.
-            # fluid.memory_optimize(main)
-
-            place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-            exe = fluid.Executor(place)
-            exe.run(startup)
-
-            feeder = fluid.DataFeeder(place=place, feed_list=[data, label])
-            pe = fluid.ParallelExecutor(
-                use_cuda=use_cuda, loss_name=loss.name, main_program=main)
-
-            fetch_list = []
-            all_vars = main.global_block().vars
-            for k, v in all_vars.items():
-                if 'tmp' not in k and k[0] is not '_' or v.persistable:
-                    fetch_list.append(k)
-
-            for data in train_inputs:
-                ret = pe.run(fetch_list,
-                             feed=feeder.feed(data),
-                             return_numpy=True)
-                for i in range(len(fetch_list)):
-                    assert not math.isnan(np.sum(ret[i])) and \
-                           not math.isinf(np.sum(ret[i]))
-
-    @unittest.skip(reason="CI timeout")
-    def test_fetch_op(self):
-        tst_reader = paddle.batch(flowers.test(use_xmap=False), batch_size=16)
-        tst_reader_iter = tst_reader()
-
-        iters = 3
-        train_inputs = []
-        for i in range(iters):
-            train_inputs.append(next(tst_reader_iter))
+        place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
+        exe = fluid.Executor(place)
+        exe.run(startup)
 
-        os.environ['CPU_NUM'] = str(4)
-        if core.is_compiled_with_cuda():
-            self.parallel_exe(train_inputs, seed=1, use_cuda=True)
-        self.parallel_exe(train_inputs, seed=1, use_cuda=False)
+        train_cp = compiler.CompiledProgram(main_program).with_data_parallel(
+            loss_name=loss.name)
+
+        run_parallel_exe(train_cp, exe, use_cuda, data, label, loss)
+
+    def run_parallel_exe_with_fetch(self, compiled_program, exe, use_cuda, data,
+                                    label, loss):
+        def get_data(batch_size=8):
+            np.random.seed(5)
+            while True:
+                img = np.random.random(
+                    size=[batch_size, 3, 224, 224]).astype(np.float32)
+                l = (np.random.random(size=[batch_size, 1]) *
+                     10).astype(np.int64)
+                yield img, l
+
+        # TODO(zcd): I found that onece the memory optimizer is open,
+        # parallel_exe doesn't fetch some variable, such as conv2d_0.b_0@GRAD,
+        # conv2d_1.b_0@GRAD. Those variables should not be pruned.
+        # fluid.memory_optimize(main)
+        fetch_list = []
+        all_vars = compiled_program._program.global_block().vars
+
+        for k, v in all_vars.items():
+            if ('tmp' not in k) and (
+                    k[0] is not '_' or v.persistable
+            ) and v.type == core.VarDesc.VarType.LOD_TENSOR:
+                fetch_list.append(k)
+
+        for batch_id, img_label in enumerate(get_data()):
+            img, l = img_label
+            train_inputs = {data.name: img, label.name: l}
+            ret = exe.run(compiled_program,
+                          fetch_list=fetch_list,
+                          feed=train_inputs,
+                          return_numpy=True)
+            for i in range(len(fetch_list)):
+                assert not math.isnan(np.sum(ret[i])) and \
+                       not math.isinf(np.sum(ret[i]))
+            if batch_id == 2:
+                break
 
-
-class TestFeedParallel(unittest.TestCase):
-    def parallel_exe(self, use_cuda, seed):
-        main = fluid.Program()
-        startup = fluid.Program()
-        startup.random_seed = seed
-        with fluid.scope_guard(fluid.core.Scope()):
-            with fluid.program_guard(main, startup):
-                data = fluid.layers.data(
-                    name='image', shape=[3, 224, 224], dtype='float32')
-                label = fluid.layers.data(
-                    name='label', shape=[1], dtype='int64')
-                out = Lenet(data, class_dim=102)
-                loss = fluid.layers.cross_entropy(input=out, label=label)
-                loss = fluid.layers.mean(loss)
-                opt = fluid.optimizer.Momentum(
-                    learning_rate=0.1,
-                    momentum=0.9,
-                    regularization=fluid.regularizer.L2Decay(1e-4))
-
-                opt.minimize(loss)
+    def run_parallel_exe_with_feed(self, compiled_program, exe, use_cuda, data,
+                                   label, loss):
+        def get_data(batch_size=8):
+            np.random.seed(5)
+            while True:
+                train_data = []
+                for _ in range(batch_size):
+                    img = np.random.random(
+                        size=[1, 3, 224, 224]).astype(np.float32)
+                    label = (np.random.random(size=[1, 1]) *
+                             10).astype(np.int64)
+                    train_data.append([img, label])
+                yield train_data
 
         place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
         feeder = fluid.DataFeeder(place=place, feed_list=[data, label])
-        reader = feeder.decorate_reader(
-            paddle.batch(
-                flowers.train(), batch_size=16), multi_devices=True)
-
-        exe = fluid.Executor(place)
-        exe.run(startup)
-
-        pe = fluid.ParallelExecutor(
-            use_cuda=use_cuda, loss_name=loss.name, main_program=main)
+        reader = feeder.decorate_reader(get_data, multi_devices=True)
 
         for batch_id, data in enumerate(reader()):
-            loss_np = pe.run(feed=data, fetch_list=[loss.name])[0]
+            loss_np = exe.run(compiled_program,
+                              feed=data,
+                              fetch_list=[loss.name])[0]
             print(batch_id, loss_np)
             if batch_id == 2:
                 break
 
-    @unittest.skip(reason="CI timeout")
-    def test_feed_op(self):
+    def test_fetch(self):
+        os.environ['CPU_NUM'] = str(4)
+        if core.is_compiled_with_cuda():
+            self.parallel_exe(
+                use_cuda=True,
+                run_parallel_exe=self.run_parallel_exe_with_fetch)
+        self.parallel_exe(
+            use_cuda=False, run_parallel_exe=self.run_parallel_exe_with_fetch)
+
+    def test_feed(self):
         os.environ['CPU_NUM'] = str(4)
         if core.is_compiled_with_cuda():
-            self.parallel_exe(use_cuda=True, seed=1)
-        self.parallel_exe(use_cuda=False, seed=1)
+            self.parallel_exe(
+                use_cuda=True, run_parallel_exe=self.run_parallel_exe_with_feed)
+        self.parallel_exe(
+            use_cuda=False, run_parallel_exe=self.run_parallel_exe_with_feed)
 
 
 if __name__ == '__main__':

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

@@ -74,7 +74,11 @@ class TestMNIST(TestParallelExecutorBase):
         label = np.ones(shape=[32, 1], dtype='int64')
         return img, label
 
-    def _compare_reduce_and_allreduce(self, model, use_cuda):
+    def _compare_reduce_and_allreduce(self,
+                                      model,
+                                      use_cuda,
+                                      delta1=1e-6,
+                                      delta2=1e-4):
         if use_cuda and not core.is_compiled_with_cuda():
             return
 
@@ -95,9 +99,9 @@ class TestMNIST(TestParallelExecutorBase):
             use_reduce=True)
 
         for loss in zip(all_reduce_first_loss, reduce_first_loss):
-            self.assertAlmostEqual(loss[0], loss[1], delta=1e-6)
+            self.assertAlmostEqual(loss[0], loss[1], delta=delta1)
         for loss in zip(all_reduce_last_loss, reduce_last_loss):
-            self.assertAlmostEqual(loss[0], loss[1], delta=1e-4)
+            self.assertAlmostEqual(loss[0], loss[1], delta=delta2)
 
     # simple_fc
     def check_simple_fc_convergence(self, use_cuda, use_reduce=False):
@@ -174,8 +178,9 @@ class TestMNIST(TestParallelExecutorBase):
                 self.check_batchnorm_fc_convergence(use_cuda, use_fast_executor)
 
     def test_batchnorm_fc_with_new_strategy(self):
-        # FIXME(zcd): close this test temporally.
-        # self._compare_reduce_and_allreduce(fc_with_batchnorm, True)
+        # NOTE: the computation result of nccl_reduce is non-deterministic,
+        # related issue: https://github.com/NVIDIA/nccl/issues/157
+        self._compare_reduce_and_allreduce(fc_with_batchnorm, True, 1e-5, 1e-2)
         self._compare_reduce_and_allreduce(fc_with_batchnorm, False)
 
 

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

@@ -16,6 +16,7 @@ from __future__ import print_function
 
 import paddle.fluid as fluid
 import paddle.fluid.core as core
+from paddle.fluid import compiler
 import numpy as np
 import unittest
 import os
@@ -61,22 +62,21 @@ class TestPassBuilder(unittest.TestCase):
             exe.run(startup)
             feed_dict = {'image': image, 'label': label}
 
-            train_exe = fluid.ParallelExecutor(
-                use_cuda=use_cuda,
+            train_cp = compiler.CompiledProgram(main).with_data_parallel(
+                loss_name=loss.name, build_strategy=build_strategy)
+            test_cp = compiler.CompiledProgram(test_program).with_data_parallel(
                 loss_name=loss.name,
-                main_program=main,
-                build_strategy=build_strategy)
-
-            test_exe = fluid.ParallelExecutor(
-                use_cuda=use_cuda,
-                main_program=test_program,
-                share_vars_from=train_exe,
-                build_strategy=build_strategy)
+                build_strategy=build_strategy,
+                share_vars_from=train_cp)
 
             for i in range(5):
-                test_loss, = test_exe.run([loss.name], feed=feed_dict)
-
-                train_loss, = train_exe.run([loss.name], feed=feed_dict)
+                _ = exe.run(train_cp, fetch_list=[loss.name], feed=feed_dict)
+                test_loss, = exe.run(test_cp,
+                                     fetch_list=[loss.name],
+                                     feed=feed_dict)
+                train_loss = exe.run(train_cp,
+                                     fetch_list=[loss.name],
+                                     feed=feed_dict)
 
                 avg_test_loss_val = np.array(test_loss).mean()
                 if math.isnan(float(avg_test_loss_val)):

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

@@ -14,6 +14,7 @@
 
 import os
 import paddle.fluid as fluid
+from paddle.fluid import compiler
 import paddle
 import unittest
 import six
@@ -140,9 +141,10 @@ def test_main(use_cuda, use_py_func_op, use_parallel_executor):
 
             exe = fluid.Executor(place)
             exe.run(fluid.default_startup_program())
+
+            train_cp = compiler.CompiledProgram(fluid.default_main_program())
             if use_parallel_executor:
-                exe = fluid.ParallelExecutor(
-                    use_cuda=use_cuda, loss_name=loss.name)
+                train_cp = train_cp.with_data_parallel(loss_name=loss.name)
                 fetch_list = [loss.name]
             else:
                 fetch_list = [loss]
@@ -150,9 +152,10 @@ def test_main(use_cuda, use_py_func_op, use_parallel_executor):
             ret = []
             for epoch_id in six.moves.range(2):
                 for d in r():
-                    L, = exe.run(feed=feeder.feed(d), fetch_list=fetch_list)
+                    L, = exe.run(train_cp,
+                                 feed=feeder.feed(d),
+                                 fetch_list=fetch_list)
                     ret.append(L)
-
             return np.array(ret)
 
 

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

@@ -16,6 +16,7 @@ from __future__ import print_function
 
 import unittest
 import paddle.fluid as fluid
+from paddle.fluid import compiler
 import paddle.fluid.core as core
 import numpy as np
 import threading
@@ -188,18 +189,18 @@ class TestPyReaderUsingExecutor(unittest.TestCase):
 
             place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 
-            startup_exe = fluid.Executor(place)
-            startup_exe.run(startup_program)
+            exe = fluid.Executor(place)
+            exe.run(startup_program)
 
+            train_cp = compiler.CompiledProgram(main_program)
             if use_parallel_executor:
-                main_exe = fluid.ParallelExecutor(use_cuda, loss_name=loss.name)
+                train_cp = train_cp.with_data_parallel(loss_name=loss.name)
                 if use_cuda:
                     self.batch_size_times = core.get_cuda_device_count()
                 else:
                     self.batch_size_times = int(
                         os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
             else:
-                main_exe = startup_exe
                 self.batch_size_times = 1
 
             reader = self.tensor_reader(use_decorate_paddle_reader)
@@ -214,13 +215,17 @@ class TestPyReaderUsingExecutor(unittest.TestCase):
 
             self.outputs = []
             for _ in range(self.iterations):
-                fetches = main_exe.run(fetch_list=[in_data.name, label.name])
+                fetches = exe.run(train_cp,
+                                  fetch_list=[in_data.name, label.name])
                 fetches = [as_numpy(fetch) for fetch in fetches]
                 self.outputs.append(fetches)
 
             feed_queue.close()
             self.validate()
-            if not use_decorate_paddle_reader:
+            if use_decorate_paddle_reader:
+                py_reader.exited = True
+                py_reader.thread.join()
+            else:
                 thread.join()
 
     def validate(self):

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

@@ -15,6 +15,7 @@
 from __future__ import print_function
 import os
 import paddle.fluid as fluid
+from paddle.fluid import compiler
 import paddle
 import numpy as np
 import unittest
@@ -74,37 +75,21 @@ class TestReaderReset(unittest.TestCase):
         exe = fluid.Executor(place)
         exe.run(startup_prog)
 
-        build_strategy = fluid.BuildStrategy()
-        exec_strategy = fluid.ExecutionStrategy()
-        parallel_exe = fluid.ParallelExecutor(
-            use_cuda=self.use_cuda,
-            main_program=main_prog,
-            build_strategy=build_strategy,
-            exec_strategy=exec_strategy)
-
-        data_appeared = [False] * self.total_ins_num
+        train_cp = compiler.CompiledProgram(main_prog).with_data_parallel()
         pass_count = 0
         while (True):
             try:
-                data_val, label_val = parallel_exe.run(fetch_list,
-                                                       return_numpy=True)
+                data_val, label_val = exe.run(train_cp,
+                                              fetch_list=fetch_list,
+                                              return_numpy=True)
                 ins_num = data_val.shape[0]
                 broadcasted_label = np.ones((ins_num, ) + tuple(
                     self.ins_shape)) * label_val.reshape((ins_num, 1))
                 self.assertEqual(data_val.all(), broadcasted_label.all())
-                for l in label_val:
-                    self.assertFalse(data_appeared[l[0]])
-                    data_appeared[l[0]] = True
 
             except fluid.core.EOFException:
                 pass_count += 1
-                if with_double_buffer:
-                    data_appeared = data_appeared[:-parallel_exe.device_count *
-                                                  self.batch_size]
-                for i in data_appeared:
-                    self.assertTrue(i)
                 if pass_count < self.test_pass_num:
-                    data_appeared = [False] * self.total_ins_num
                     data_reader_handle.reset()
                 else:
                     break

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

@@ -22,6 +22,14 @@ import numpy
 import functools
 
 
+def convert_to_offset(lod):
+    offset = [[0] for i in lod]
+    for i, level in enumerate(lod):
+        for seq_len in level:
+            offset[i].append(offset[i][-1] + seq_len)
+    return offset
+
+
 class TestReorderLoDTensor(unittest.TestCase):
     num_seq = 5
     # [name, shape, lod_level] pair indicating data info of source and target
@@ -91,13 +99,6 @@ class TestReorderLoDTensor(unittest.TestCase):
             self.inputs[desc[0]] = tensor
 
     def reorder(self):
-        def convert_to_offset(lod):
-            offset_lod = [[0] for i in lod]
-            for i, level in enumerate(lod):
-                for seq_len in level:
-                    offset_lod[i].append(offset_lod[i][-1] + seq_len)
-            return offset_lod
-
         level = 0
         # compute the rank_table according to ref_lod
         ref_lod = self.data[self.data_desc[1][0]][1][level]

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

@@ -17,33 +17,43 @@ from __future__ import print_function
 import unittest
 import numpy as np
 from op_test import OpTest
+from test_reorder_lod_tensor import convert_to_offset
 
 
-class TestSeqAvgPool(OpTest):
-    def convert_to_offset(self, lod):
-        offset = [[0] for i in lod]
-        for i, level in enumerate(lod):
-            for seq_len in level:
-                offset[i].append(offset[i][-1] + seq_len)
-        return offset
+def compute_seqpool_sum(x, offset, out):
+    for i in range(len(offset[0]) - 1):
+        sub_x = x[offset[0][i]:offset[0][i + 1], :]
+        out[i] = sub_x.sum(axis=0)
+
+
+def compute_seqpool_avg(x, offset, out):
+    for i in range(len(offset[0]) - 1):
+        sub_x = x[offset[0][i]:offset[0][i + 1], :]
+        out[i] = sub_x.mean(axis=0)
+
 
+def compute_seqpool_sqrt(x, offset, out):
+    for i in range(len(offset[0]) - 1):
+        sub_x = x[offset[0][i]:offset[0][i + 1], :]
+        seq_len = offset[0][i + 1] - offset[0][i]
+        out[i] = sub_x.sum(axis=0) / np.sqrt(seq_len)
+
+
+class TestSeqAvgPool(OpTest):
     def set_data(self):
         self.op_type = 'sequence_pool'
         # one level, batch size is 4
         x = np.random.uniform(0.1, 1, [11, 23]).astype('float32')
         lod = [[11]]
         self.inputs = {'X': (x, lod)}
-        offset = self.convert_to_offset(lod)
-
+        offset = convert_to_offset(lod)
         out = np.zeros((len(lod[0]), 23)).astype('float32')
         self.outputs = {'Out': out}
         return x, offset, out
 
     def compute(self, x, offset, out):
         self.attrs = {'pooltype': "AVERAGE"}
-        for i in range(len(offset[0]) - 1):
-            sub_x = x[offset[0][i]:offset[0][i + 1], :]
-            out[i] = sub_x.mean(axis=0)
+        compute_seqpool_avg(x, offset, out)
 
     def setUp(self):
         x, offset, out = self.set_data()
@@ -62,9 +72,7 @@ class TestSeqAvgPool(OpTest):
 class TestSeqSumPool(TestSeqAvgPool):
     def compute(self, x, offset, out):
         self.attrs = {'pooltype': "SUM"}
-        for i in range(len(offset[0]) - 1):
-            sub_x = x[offset[0][i]:offset[0][i + 1], :]
-            out[i] = sub_x.sum(axis=0)
+        compute_seqpool_sum(x, offset, out)
 
 
 class TestSeqMaxPool(TestSeqAvgPool):
@@ -72,7 +80,7 @@ class TestSeqMaxPool(TestSeqAvgPool):
         self.op_type = 'sequence_pool'
         x = np.random.uniform(0.1, 1, [13, 23]).astype('float32')
         lod = [[13]]
-        offset = self.convert_to_offset(lod)
+        offset = convert_to_offset(lod)
         for i in range(len(offset[0]) - 1):
             l = offset[0][i + 1] - offset[0][i]
             x[offset[0][i] + np.random.randint(l), :] += 2.0
@@ -93,10 +101,7 @@ class TestSeqMaxPool(TestSeqAvgPool):
 class TestSeqSqrtPool(TestSeqAvgPool):
     def compute(self, x, offset, out):
         self.attrs = {'pooltype': "SQRT"}
-        for i in range(len(offset[0]) - 1):
-            sub_x = x[offset[0][i]:offset[0][i + 1], :]
-            seq_len = offset[0][i + 1] - offset[0][i]
-            out[i] = sub_x.sum(axis=0) / np.sqrt(seq_len)
+        compute_seqpool_sqrt(x, offset, out)
 
 
 class TestSeqLastPool(TestSeqAvgPool):
@@ -122,7 +127,7 @@ class TestSeqAvgPool2D(TestSeqAvgPool):
         x = np.random.uniform(0.1, 1, [13, 3, 17]).astype('float32')
         lod = [[4, 1, 3, 5]]
         self.inputs = {'X': (x, lod)}
-        offset = self.convert_to_offset(lod)
+        offset = convert_to_offset(lod)
 
         out = np.zeros((4, 3, 17)).astype('float32')
         self.outputs = {'Out': out}
@@ -167,7 +172,7 @@ class TestSeqMaxPool2D(TestSeqAvgPool2D):
         x = np.random.uniform(0.1, 1, [13, 3, 11]).astype('float32')
         lod = [[4, 1, 3, 5]]
         self.inputs = {'X': (x, lod)}
-        offset = self.convert_to_offset(lod)
+        offset = convert_to_offset(lod)
         for i in range(len(offset[0]) - 1):
             l = offset[0][i + 1] - offset[0][i]
             x[offset[0][i] + np.random.randint(l), :] += 1.0

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

@@ -28,6 +28,7 @@ class TestSoftmaxWithCrossEntropyOp(OpTest):
 
     def initParams(self):
         self.numeric_stable_mode = False
+        self.dtype = np.float64
 
     def setUp(self):
         self.initParams()
@@ -36,19 +37,19 @@ class TestSoftmaxWithCrossEntropyOp(OpTest):
         class_num = 37
 
         logits = np.random.uniform(0.1, 1.0,
-                                   [batch_size, class_num]).astype("float64")
+                                   [batch_size, class_num]).astype(self.dtype)
         softmax = np.apply_along_axis(stable_softmax, 1, logits)
         labels = np.random.randint(0, class_num, [batch_size, 1], dtype="int64")
 
         cross_entropy = np.asmatrix(
             [[-np.log(softmax[i][labels[i][0]])]
              for i in range(softmax.shape[0])],
-            dtype="float64")
+            dtype=self.dtype)
 
         self.inputs = {"Logits": logits, "Label": labels}
         self.outputs = {
-            "Softmax": softmax.astype("float64"),
-            "Loss": cross_entropy.astype("float64")
+            "Softmax": softmax.astype(self.dtype),
+            "Loss": cross_entropy.astype(self.dtype)
         }
         self.attrs = {"numeric_stable_mode": self.numeric_stable_mode}
 
@@ -56,7 +57,7 @@ class TestSoftmaxWithCrossEntropyOp(OpTest):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(["Logits"], "Loss")
+        self.check_grad(["Logits"], "Loss", max_relative_error=0.05)
 
 
 class TestSoftmaxWithCrossEntropyOpNoCudnn(TestSoftmaxWithCrossEntropyOp):
@@ -64,6 +65,55 @@ class TestSoftmaxWithCrossEntropyOpNoCudnn(TestSoftmaxWithCrossEntropyOp):
         self.numeric_stable_mode = True
 
 
+class TestSoftmaxWithCrossEntropyOpFp16(TestSoftmaxWithCrossEntropyOp):
+    def initParams(self):
+        self.numeric_stable_mode = False
+        self.dtype = np.float16
+
+    def setUp(self):
+        self.initParams()
+        self.op_type = "softmax_with_cross_entropy"
+        batch_size = 41
+        class_num = 37
+
+        # NOTE: numpy float16 have very low accuracy, use float32 for numpy check.
+        logits = np.random.uniform(0.1, 1.0,
+                                   [batch_size, class_num]).astype(np.float32)
+        softmax = np.apply_along_axis(stable_softmax, 1, logits)
+        labels = np.random.randint(0, class_num, [batch_size, 1], dtype="int64")
+
+        cross_entropy = np.asmatrix(
+            [[-np.log(softmax[i][labels[i][0]])]
+             for i in range(softmax.shape[0])],
+            dtype=np.float32)
+
+        self.inputs = {
+            "Logits": logits.astype(self.dtype).view(np.uint16),
+            "Label": labels
+        }
+        self.outputs = {
+            "Softmax": softmax.astype(self.dtype),
+            "Loss": cross_entropy.astype(self.dtype)
+        }
+        self.attrs = {"numeric_stable_mode": self.numeric_stable_mode}
+
+    def test_check_output(self):
+        self.check_output(atol=1e-2)
+
+    def test_check_grad(self):
+        self.check_grad(["Logits"], "Loss", max_relative_error=0.1)
+
+
+class TestSoftmaxWithCrossEntropyOpNoCudnnFp16(
+        TestSoftmaxWithCrossEntropyOpFp16):
+    def initParams(self):
+        self.numeric_stable_mode = True
+        self.dtype = np.float16
+
+    def test_check_grad(self):
+        self.check_grad(["Logits"], "Loss", max_relative_error=0.1)
+
+
 class TestSoftmaxWithCrossEntropyOp2(OpTest):
     """
     Test softmax with cross entropy operator with soft labels.

python/paddle/fluid/tests/unittests/test_teacher_student_sigmoid_loss_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.
+
+import numpy as np
+from math import log
+from math import exp
+from op_test import OpTest
+from scipy.special import logit
+from scipy.special import expit
+import unittest
+
+
+class TestTeacherStudentSigmoidLossOp(OpTest):
+    """
+        Test teacher_student_sigmoid_loss with discrete one-hot labels.
+    """
+
+    def setUp(self):
+        self.op_type = "teacher_student_sigmoid_loss"
+        batch_size = 16
+        num_classes = 1
+        self.inputs = {
+            'X': logit(
+                np.random.uniform(0, 1, (batch_size, num_classes))
+                .astype("float32")),
+            'Label': np.random.uniform(0, 2, (batch_size, num_classes))
+            .astype("float32")
+        }
+        outs = []
+        for index, label in enumerate(self.inputs["Label"]):
+            x = self.inputs["X"][index]
+            if label < -1.0:
+                outs.append(max(x, 0.0) + log(1.0 + exp(-abs(x))))
+            elif label < 0.0:
+                outs.append(max(x, 0.0) - x + log(1.0 + exp(-abs(x))))
+            elif label < 1.0:
+                outs.append(max(x, 0.0) + log(1.0 + exp(-abs(x))) + \
+                            max(x, 0.0) - x * label + log(1.0 + exp(-abs(x))))
+            else:
+                outs.append(max(x, 0.0) - x + log(1.0 + exp(-abs(x))) + \
+                            max(x, 0.0) - x * (label - 1.0) + log(1.0 + exp(-abs(x))))
+        self.outputs = {'Y': np.array(outs)}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Y", numeric_grad_delta=0.005)

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

@@ -22,6 +22,7 @@ import paddle
 import paddle.fluid.core as core
 
 import paddle.fluid as fluid
+from paddle.fluid import compiler
 
 
 def get_places():
@@ -111,17 +112,17 @@ class TestWeightDecay(unittest.TestCase):
                 if use_reduce else fluid.BuildStrategy.ReduceStrategy.AllReduce
         build_strategy.memory_optimize = use_ir_memory_optimize
 
-        parallel_exe = fluid.ParallelExecutor(
-            use_cuda,
+        train_cp = compiler.CompiledProgram(fluid.default_main_program(
+        )).with_data_parallel(
             loss_name=loss.name,
             exec_strategy=exec_strategy,
             build_strategy=build_strategy)
 
         loss_set = []
         for data in self.train_data:
-            out = parallel_exe.run(feed=feeder.feed(data),
-                                   fetch_list=[loss.name])
-            print("loss              %s" % (np.average(out)))
+            out = exe.run(train_cp,
+                          feed=feeder.feed(data),
+                          fetch_list=[loss.name])
             loss_set.append(np.average(out))
 
         return loss_set

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -752,12 +752,6 @@ class DistributeTranspiler(object):
             elif op not in lr_ops:
                 self._append_pserver_non_opt_ops(block, op)
 
-        def __op_have_grad_input__(op):
-            for varname in op.input_arg_names:
-                if varname.find("@GRAD") >= 0:
-                    return varname
-            return ""
-
         def __clone_lr_op_sub_block__(op, program, lr_block):
             if not op.has_attr('sub_block'):
                 return
@@ -808,7 +802,7 @@ class DistributeTranspiler(object):
             merged_var = None
             for _, op in enumerate(self.optimize_ops):
                 # find the origin grad var before clipping/L2Decay,
-                # merged_var should be the input var name of L2Decaybuil
+                # merged_var should be the input var name of L2Decay
                 grad_varname_for_block = op.attr(OP_ROLE_VAR_ATTR_NAME)[1]
                 if op.attr(OP_ROLE_VAR_ATTR_NAME)[
                         0] == optimize_target_param_name:
@@ -1684,7 +1678,16 @@ class DistributeTranspiler(object):
                 if self.config.enable_dc_asgd:
                     new_inputs[key] = dc
                 else:
-                    new_inputs[key] = merged_var
+                    # Note!! This is for l2decay on sparse gradient, because it will create a new tensor for
+                    # decayed gradient but not inplace modify the origin one
+                    origin_grad_name = opt_op.input(key)[0]
+                    if core.kNewGradSuffix(
+                    ) in origin_grad_name and pserver_block.has_var(
+                            origin_grad_name):
+                        new_grad = pserver_block.var(origin_grad_name)
+                        new_inputs[key] = new_grad
+                    else:
+                        new_inputs[key] = merged_var
             elif key == "Param":
                 param_block = _get_param_block(opt_op)
                 if not param_block: