Merge branch 'develop' into add_reorg_op

CMakeLists.txt

@@ -62,7 +62,6 @@ option(WITH_DISTRIBUTE  "Compile with distributed support"              OFF)
 option(USE_EIGEN_FOR_BLAS   "Use matrix multiplication in Eigen"        OFF)
 option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen"            OFF)
 option(WITH_ARM_FP16    "Use half precision support on armv8.2-a cpu"   OFF)
-option(WITH_FAST_BUNDLE_TEST    "Bundle tests that can be run in a single process together to reduce launch overhead"   OFF)
 option(WITH_CONTRIB     "Compile the third-party contributation"        OFF)
 option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
 option(WITH_ANAKIN      "Compile with Anakin library"                   OFF)

paddle/fluid/API.spec

@@ -175,6 +175,7 @@ paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dim
 paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.space_to_depth ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None))
 paddle.fluid.layers.hash ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None))

paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc

@@ -15,6 +15,7 @@
 #include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h"
 
 #include <gtest/gtest.h>
+#include "paddle/fluid/framework/op_proto_maker.h"
 
 namespace paddle {
 namespace framework {
@@ -36,6 +37,8 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
     op->SetInput("X", inputs);
   }
   op->SetOutput("Out", outputs);
+  op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
+              static_cast<int>(OpRole::kForward));
 }
 
 // a->OP0->b

paddle/fluid/framework/ir/fc_fuse_pass_tester.cc

@@ -15,6 +15,7 @@
 #include "paddle/fluid/framework/ir/fc_fuse_pass.h"
 
 #include <gtest/gtest.h>
+#include "paddle/fluid/framework/op_proto_maker.h"
 
 namespace paddle {
 namespace framework {
@@ -32,6 +33,8 @@ void SetOp(ProgramDesc* prog, const std::string& type,
     op->SetInput("X", inputs);
   }
   op->SetOutput("Out", outputs);
+  op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
+              static_cast<int>(OpRole::kForward));
 }
 
 // a->OP0->b

paddle/fluid/framework/ir/graph.cc

@@ -23,8 +23,62 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 namespace ir {
+namespace {
+
+void CheckProgram(const ProgramDesc &program) {
+  std::map<int, bool> visit;
+#define _INT(role) static_cast<int>(role)
+
+  for (size_t i = 0; i < program.Size(); ++i) {
+    for (OpDesc *op : program.Block(i).AllOps()) {
+      // For backward compatibility, some program doesn't have role added.
+      if (!op->HasAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) continue;
+      int role_id = boost::get<int>(
+          op->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName()));
+      visit[role_id] = true;
+      switch (role_id) {
+        case _INT(OpRole::kForward):
+          PADDLE_ENFORCE(
+              visit.find(_INT(OpRole::kBackward)) == visit.end(),
+              "Cannot add forward operator before backward operator.");
+          break;
+        case _INT(OpRole::kBackward):
+        case _INT(OpRole::kBackward) | _INT(OpRole::kLoss):
+          PADDLE_ENFORCE(
+              visit.find(_INT(OpRole::kOptimize)) == visit.end(),
+              "Cannot add backward operator before optimize operator.");
+          break;
+        case _INT(OpRole::kForward) | _INT(OpRole::kLoss):
+          PADDLE_ENFORCE(visit.find(_INT(OpRole::kBackward) |
+                                    _INT(OpRole::kLoss)) == visit.end(),
+                         "Cannot add backward|loss operator before "
+                         "forward|loss operator.");
+          PADDLE_ENFORCE(
+              visit.find(_INT(OpRole::kOptimize)) == visit.end(),
+              "Cannot add backward operator before optimize operator.");
+          break;
+        case _INT(OpRole::kOptimize):
+        case _INT(OpRole::kOptimize) | _INT(OpRole::kLRSched):
+          PADDLE_ENFORCE(visit.find(_INT(OpRole::kBackward)) != visit.end(),
+                         "Optimize operators must follow backward operator.");
+          break;
+        case _INT(OpRole::kLRSched):
+        case _INT(OpRole::kDist):
+        case _INT(OpRole::kRPC):
+        case _INT(OpRole::kNotSpecified):
+          break;
+        default:
+          LOG(FATAL) << "Unknown operator role. Don't add new role because "
+                        "you don't know what you are doing.";
+      }
+    }
+  }
+#undef _INT
+}
+}  // namespace
 
 Graph::Graph(const ProgramDesc &program) : program_(program) {
+  CheckProgram(program_);
   // Make the nodes id start from 0.
   Node::ResetId();
   auto var_nodes = InitFromProgram(program_);

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -259,6 +259,15 @@ GraphPatternDetector::DetectPatterns() {
   return result;
 }
 
+bool GraphItemCMP(const std::pair<PDNode *, Node *> &a,
+                  const std::pair<PDNode *, Node *> &b) {
+  if (a.first != b.first) {
+    return a.first < b.first;
+  } else {
+    return a.second < b.second;
+  }
+}
+
 // TODO(Superjomn) enhance the function as it marks unique unique as duplicates
 // see https://github.com/PaddlePaddle/Paddle/issues/13550
 void GraphPatternDetector::UniquePatterns(
@@ -267,12 +276,16 @@ void GraphPatternDetector::UniquePatterns(
   std::vector<GraphPatternDetector::subgraph_t> result;
 
   std::unordered_set<size_t> set;
+  std::hash<std::string> hasher;
   for (auto &g : *subgraphs) {
-    size_t key = 0;
-    for (auto &item : g) {
-      key ^= std::hash<void *>{}(item.first);
-      key ^= std::hash<void *>{}(item.second);
-    }
+    // Sort the items in the sub-graph, and transform to a string key.
+    std::vector<std::pair<PDNode *, Node *>> sorted_keys(g.begin(), g.end());
+    std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemCMP);
+    std::stringstream ss;
+    for (auto &item : sorted_keys) {
+      ss << item.first << ":" << item.second;
+    }
+    auto key = hasher(ss.str());
     if (!set.count(key)) {
       result.emplace_back(g);
       set.insert(key);

paddle/fluid/framework/lod_tensor.cc

@@ -418,7 +418,7 @@ void LoDTensor::MergeLoDTensor(
     PADDLE_ENFORCE_EQ(new_lod.size(), lod.size());
     for (size_t j = 0; j < lod.size(); ++j) {
       auto &sub_lod = new_lod[j];
-      auto &offset = sub_lod.back();
+      size_t offset = sub_lod.back();
       for (size_t k = 1; k < lod[j].size(); ++k) {
         sub_lod.push_back(lod[j][k] + offset);
       }

paddle/fluid/framework/operator.cc

@@ -354,18 +354,18 @@ void OperatorBase::GenerateTemporaryNames() {
   }
 }
 
-static bool VarIsTensor(const Variable* var) {
-  return var->IsType<LoDTensor>() || var->IsType<SelectedRows>();
+static bool VarIsTensor(const Variable& var) {
+  return var.IsType<LoDTensor>() || var.IsType<SelectedRows>();
 }
 
-const Tensor* GetTensorFromVar(Variable* var) {
-  if (var->IsType<LoDTensor>()) {
-    return var->GetMutable<LoDTensor>();
-  } else if (var->IsType<SelectedRows>()) {
-    return var->GetMutable<SelectedRows>()->mutable_value();
+const Tensor* GetTensorFromVar(const Variable& var) {
+  if (var.IsType<LoDTensor>()) {
+    return static_cast<const Tensor*>(&(var.Get<LoDTensor>()));
+  } else if (var.IsType<SelectedRows>()) {
+    return &(var.Get<SelectedRows>().value());
   } else {
     PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
-                 var->Type().name());
+                 var.Type().name());
   }
 }
 
@@ -415,8 +415,7 @@ bool ExecutionContext::HasOutput(const std::string& name) const {
 template <>
 const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
   auto* var = InputVar(name);
-  return var == nullptr ? nullptr
-                        : GetTensorFromVar(const_cast<Variable*>(var));
+  return var == nullptr ? nullptr : GetTensorFromVar(*var);
 }
 
 template <>
@@ -428,7 +427,7 @@ const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
   std::transform(names.begin(), names.end(), std::back_inserter(res),
                  [&](const std::string& sub_name) {
                    auto var = scope_.FindVar(sub_name);
-                   return var == nullptr ? nullptr : GetTensorFromVar(var);
+                   return var == nullptr ? nullptr : GetTensorFromVar(*var);
                  });
   return res;
 }
@@ -770,8 +769,10 @@ void OperatorWithKernel::TransferInplaceVarsBack(
   for (auto& var_name : inplace_vars) {
     VLOG(3) << "share inplace var " + var_name + " back to it's original scope";
     auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name));
-    auto* transformed_tensor =
-        GetTensorFromVar(transfer_scope.FindVar(var_name));
+    auto* var = transfer_scope.FindVar(var_name);
+    PADDLE_ENFORCE(var != nullptr, "The var[%s] should not be nullptr",
+                   var_name);
+    auto* transformed_tensor = GetTensorFromVar(*var);
     original_tensor->ShareDataWith(*transformed_tensor);
   }
 }
@@ -784,11 +785,11 @@ Scope* OperatorWithKernel::TryTransferData(
     for (auto& var_name : var_name_item.second) {
       auto* var = scope.FindVar(var_name);
       // Only tensor can be tranfer to another device.
-      if (var == nullptr || !VarIsTensor(var)) {
+      if (var == nullptr || !VarIsTensor(*var)) {
         continue;
       }
 
-      auto* tensor_in = GetTensorFromVar(var);
+      auto* tensor_in = GetTensorFromVar(*var);
       if (!tensor_in->IsInitialized()) {
         continue;
       }

paddle/fluid/framework/operator.h

@@ -63,7 +63,7 @@ inline std::string GradVarName(const std::string& var_name) {
 }
 
 proto::VarType::Type GetDataTypeOfVar(const Variable* var);
-const Tensor* GetTensorFromVar(Variable* var);
+const Tensor* GetTensorFromVar(const Variable& var);
 
 class OperatorBase;
 class ExecutionContext;

paddle/fluid/framework/tensor_test.cc

@@ -75,6 +75,19 @@ TEST(Tensor, MutableData) {
                                         platform::CPUPlace());
     EXPECT_EQ(p1, p2);
   }
+  // Not sure if it's desired, but currently, Tensor type can be changed.
+  {
+    framework::Tensor src_tensor;
+    int8_t* p1 = src_tensor.mutable_data<int8_t>(framework::make_ddim({1}),
+                                                 platform::CPUPlace());
+    EXPECT_NE(p1, nullptr);
+    *p1 = 1;
+
+    uint8_t* p2 = src_tensor.mutable_data<uint8_t>(framework::make_ddim({1}),
+                                                   platform::CPUPlace());
+    EXPECT_NE(p2, nullptr);
+    EXPECT_EQ(static_cast<int>(p2[0]), 1);
+  }
 
 #ifdef PADDLE_WITH_CUDA
   {

paddle/fluid/framework/tensor_util.cc

@@ -153,6 +153,12 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
     auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
     auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
     memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);
+  } else if (platform::is_cuda_pinned_place(src_place) &&
+             platform::is_gpu_place(dst_place)) {
+    auto src_pinned_place = boost::get<platform::CUDAPinnedPlace>(src_place);
+    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
+    memory::Copy(dst_gpu_place, dst_ptr, src_pinned_place, src_ptr, size,
+                 nullptr);
   }
 #endif
 }

paddle/fluid/inference/CMakeLists.txt

+if(WITH_TESTING)
+  include(test.cmake) # some generic cmake funtion for inference
+endif()
 # analysis and tensorrt must be added before creating static library,
 # otherwise, there would be undefined reference to them in static library.
 add_subdirectory(analysis)

paddle/fluid/inference/analysis/CMakeLists.txt

@@ -20,22 +20,17 @@ cc_test(test_node SRCS node_tester.cc DEPS analysis)
 cc_test(test_dot SRCS dot_tester.cc DEPS analysis)
 cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis paddle_fluid)
 
-function (inference_analysis_test TARGET)
+function(inference_analysis_test TARGET)
   if(WITH_TESTING)
      set(options "")
      set(oneValueArgs "")
      set(multiValueArgs SRCS ARGS EXTRA_DEPS)
      cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-        set(mem_opt "")
-        if(WITH_GPU)
-            set(mem_opt "--fraction_of_gpu_memory_to_use=0.5")
-        endif()
-        cc_test(${TARGET}
-                SRCS "${analysis_test_SRCS}"
+     inference_base_test(${TARGET}
+             SRCS ${analysis_test_SRCS}
              DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS}
-                ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model ${mem_opt} ${analysis_test_ARGS})
-        set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec)
-    endif(WITH_TESTING)
+             ARGS --inference_model_dir=${WORD2VEC_MODEL_DIR} ${analysis_test_ARGS})
+  endif()
 endfunction(inference_analysis_test)
 
 inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS paddle_inference_api)

paddle/fluid/inference/analysis/data_flow_graph_tester.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/inference/analysis/data_flow_graph.h"
+#include "paddle/fluid/framework/op_proto_maker.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/inference/analysis/ut_helper.h"
 
@@ -130,6 +131,8 @@ void SetOp(framework::ProgramDesc* prog, const std::string& type,
   op->SetType(type);
   op->SetInput("Xs", inputs);
   op->SetOutput("Xs", outputs);
+  op->SetAttr(framework::OpProtoAndCheckerMaker::OpRoleAttrName(),
+              static_cast<int>(framework::OpRole::kForward));
 }
 
 TEST(DataFlowGraph, Build_IR_Graph) {

paddle/fluid/inference/api/CMakeLists.txt

@@ -17,39 +17,12 @@ if(APPLE)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move")
 endif(APPLE)
 
-
-set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager naive_executor ${GLOB_PASS_LIB}
-        )
+set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager naive_executor ${GLOB_PASS_LIB})
 
 if(WITH_GPU AND TENSORRT_FOUND)
     set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine analysis_predictor)
 endif()
 
-function(inference_api_test TARGET_NAME)
-    if (WITH_TESTING)
-        set(options "")
-        set(oneValueArgs SRC)
-        set(multiValueArgs ARGS)
-        cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-
-	if (WITH_GPU)
-		cc_test(${TARGET_NAME}
-			SRCS ${inference_test_SRC}
-			DEPS "${inference_deps}"
-			ARGS --dirname=${PYTHON_TESTS_DIR}/book/ --fraction_of_gpu_memory_to_use=0.15)
-        else()
-		cc_test(${TARGET_NAME}
-			SRCS ${inference_test_SRC}
-			DEPS "${inference_deps}"
-			ARGS --dirname=${PYTHON_TESTS_DIR}/book/)
-	endif()
-        if(inference_test_ARGS)
-            set_tests_properties(${TARGET_NAME}
-                    PROPERTIES DEPENDS "${inference_test_ARGS}")
-        endif()
-    endif(WITH_TESTING)
-endfunction(inference_api_test)
-
 cc_library(reset_tensor_array SRCS details/reset_tensor_array.cc DEPS lod_tensor scope)
 cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS reset_tensor_array lod_tensor scope)
 cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api analysis naive_executor zero_copy_tensor)
@@ -59,8 +32,11 @@ cc_test(test_paddle_inference_api
         SRCS api_tester.cc
         DEPS paddle_inference_api)
 
-inference_api_test(test_api_impl SRC api_impl_tester.cc
-                    ARGS test_word2vec test_image_classification)
+if(WITH_TESTING)
+  inference_base_test(test_api_impl SRCS api_impl_tester.cc DEPS ${inference_deps}
+                      ARGS --word2vec_dirname=${WORD2VEC_MODEL_DIR} --book_dirname=${PYTHON_TESTS_DIR}/book)
+  set_tests_properties(test_api_impl PROPERTIES DEPENDS test_image_classification)
+endif()
 cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_predictor ${inference_deps} paddle_inference_api
         ARGS --dirname=${PYTHON_TESTS_DIR}/book)
 
@@ -68,8 +44,10 @@ if(WITH_GPU AND TENSORRT_FOUND)
 cc_library(paddle_inference_tensorrt_subgraph_engine
         SRCS api_tensorrt_subgraph_engine.cc
         DEPS paddle_inference_api analysis tensorrt_engine paddle_inference_api paddle_fluid_api tensorrt_converter zero_copy_tensor_dummy)
-
-inference_api_test(test_api_tensorrt_subgraph_engine SRC api_tensorrt_subgraph_engine_tester.cc ARGS test_word2vec)
+  if(WITH_TESTING)
+    inference_base_test(test_api_tensorrt_subgraph_engine SRCS api_tensorrt_subgraph_engine_tester.cc DEPS ${inference_deps}
+                      ARGS --dirname=${WORD2VEC_MODEL_DIR})
+  endif()
 endif()
 
 if (WITH_ANAKIN AND WITH_MKL) # only needed in CI

paddle/fluid/inference/api/api_impl_tester.cc

@@ -22,12 +22,14 @@ limitations under the License. */
 #include "paddle/fluid/inference/tests/test_helper.h"
 
 #ifdef __clang__
-#define ACC_DIFF 4e-2
+#define ACC_DIFF 4e-3
 #else
-#define ACC_DIFF 1e-2
+#define ACC_DIFF 1e-3
 #endif
 
-DEFINE_string(dirname, "", "Directory of the inference model.");
+DEFINE_string(word2vec_dirname, "",
+              "Directory of the word2vec inference model.");
+DEFINE_string(book_dirname, "", "Directory of the book inference model.");
 
 namespace paddle {
 
@@ -49,7 +51,7 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) {
 
 NativeConfig GetConfig() {
   NativeConfig config;
-  config.model_dir = FLAGS_dirname + "/word2vec.inference.model";
+  config.model_dir = FLAGS_word2vec_dirname;
   LOG(INFO) << "dirname  " << config.model_dir;
   config.fraction_of_gpu_memory = 0.15;
 #ifdef PADDLE_WITH_CUDA
@@ -116,7 +118,7 @@ void MainImageClassification(bool use_gpu) {
   NativeConfig config = GetConfig();
   config.use_gpu = use_gpu;
   config.model_dir =
-      FLAGS_dirname + "/image_classification_resnet.inference.model";
+      FLAGS_book_dirname + "/image_classification_resnet.inference.model";
 
   const bool is_combined = false;
   std::vector<std::vector<int64_t>> feed_target_shapes =
@@ -220,7 +222,7 @@ void MainThreadsImageClassification(bool use_gpu) {
   NativeConfig config = GetConfig();
   config.use_gpu = use_gpu;
   config.model_dir =
-      FLAGS_dirname + "/image_classification_resnet.inference.model";
+      FLAGS_book_dirname + "/image_classification_resnet.inference.model";
 
   auto main_predictor = CreatePaddlePredictor<NativeConfig>(config);
   std::vector<framework::LoDTensor> jobs(num_jobs);

paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc

@@ -29,13 +29,13 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) {
 
   //# 1. Create PaddlePredictor with a config.
   NativeConfig config0;
-  config0.model_dir = FLAGS_dirname + "word2vec.inference.model";
+  config0.model_dir = FLAGS_dirname;
   config0.use_gpu = true;
   config0.fraction_of_gpu_memory = 0.3;
   config0.device = 0;
 
   MixedRTConfig config1;
-  config1.model_dir = FLAGS_dirname + "word2vec.inference.model";
+  config1.model_dir = FLAGS_dirname;
   config1.use_gpu = true;
   config1.fraction_of_gpu_memory = 0.3;
   config1.device = 0;

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

@@ -62,7 +62,7 @@ for WITH_STATIC_LIB in ON OFF; do
     -DWITH_GPU=$TEST_GPU_CPU \
     -DWITH_STATIC_LIB=$WITH_STATIC_LIB
   make -j
-  word2vec_model=${PADDLE_ROOT}'/build/python/paddle/fluid/tests/book/word2vec.inference.model'
+  word2vec_model=$DATA_DIR'/word2vec/word2vec.inference.model'
   if [ -d $word2vec_model ]; then
     for use_gpu in $use_gpu_list; do
       ./simple_on_word2vec \

paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc

@@ -70,12 +70,8 @@ void Main(bool use_gpu) {
     // The outputs' buffers are in CPU memory.
     for (size_t i = 0; i < std::min(static_cast<size_t>(5), num_elements);
          i++) {
-      // Here will result random fail, for that the model is trained by CI, the
-      // train phase is not stable, so the result will be random.
-      // TODO(Superjomn) will restore after the model is upload.
-      // CHECK_NEAR(static_cast<float*>(outputs.front().data.data())[i],
-      // result[i],
-      // 0.001);
+      CHECK_NEAR(static_cast<float*>(outputs.front().data.data())[i], result[i],
+                 0.001);
     }
   }
 }

paddle/fluid/inference/test.cmake

+set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com" CACHE STRING "inference download url")
+set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING
+    "A path setting inference demo download directories.")
+function (inference_download install_dir url filename)
+    message(STATUS "Download inference test stuff from ${url}/${filename}")
+    execute_process(COMMAND bash -c "mkdir -p ${install_dir}")
+    execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}/${filename}")
+    message(STATUS "finish downloading ${filename}")
+endfunction()
+
+function (inference_download_and_uncompress install_dir url filename)
+    inference_download(${install_dir} ${url} ${filename})
+    execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}")
+endfunction()
+
+set(WORD2VEC_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/word2vec")
+if (NOT EXISTS ${WORD2VEC_INSTALL_DIR})
+    inference_download_and_uncompress(${WORD2VEC_INSTALL_DIR} ${INFERENCE_URL} "word2vec.inference.model.tar.gz")
+endif()
+set(WORD2VEC_MODEL_DIR "${WORD2VEC_INSTALL_DIR}/word2vec.inference.model")
+
+function (inference_base_test TARGET)
+   set(options "")
+   set(oneValueArgs "")
+   set(multiValueArgs SRCS ARGS DEPS)
+   cmake_parse_arguments(base_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+   if(WITH_GPU)
+       set(mem_opt "--fraction_of_gpu_memory_to_use=0.5")
+   endif()
+   cc_test(${TARGET} SRCS ${base_test_SRCS} DEPS ${base_test_DEPS} ARGS ${mem_opt} ${base_test_ARGS})
+endfunction()

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

-set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com")
-set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING
-    "A path setting inference demo download directories.")
 set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor)
-function (inference_download install_dir url filename)
-    message(STATUS "Download inference test stuff from ${url}/${filename}")
-    execute_process(COMMAND bash -c "mkdir -p ${install_dir}")
-    execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}/${filename}")
-    message(STATUS "finish downloading ${filename}")
-endfunction()
-
-function (inference_download_and_uncompress install_dir url filename)
-    inference_download(${install_dir} ${url} ${filename})
-    execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}")
-endfunction()
 
 function(download_model_and_data install_dir model_name data_name)
     if (NOT EXISTS ${install_dir})

paddle/fluid/operators/affine_grid_cudnn_op.cu.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/op_registry.h"
+#include "paddle/fluid/platform/cudnn_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using ScopedSpatialTransformerDescriptor =
+    platform::ScopedSpatialTransformerDescriptor;
+
+template <typename T>
+class CUDNNAffineGridOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use CUDAPlace.");
+    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    auto handle = dev_ctx.cudnn_handle();
+    auto* theta = ctx.Input<Tensor>("Theta");
+    auto* output = ctx.Output<Tensor>("Output");
+    const T* theta_data = theta->data<T>();
+
+    int n = theta->dims()[0];
+    auto size_attr = ctx.Attr<std::vector<int>>("output_shape");
+    Tensor h_sizes;
+    int* h_size_data;
+    if (size_attr.size() == 0) {
+      auto* output_shape = ctx.Input<Tensor>("OutputShape");
+      framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes);
+      h_size_data = h_sizes.data<int>();
+    } else {
+      h_size_data = h_sizes.mutable_data<int>({4}, platform::CPUPlace());
+      h_size_data[0] = n;
+      h_size_data[1] = size_attr[1];
+      h_size_data[2] = size_attr[2];
+      h_size_data[3] = size_attr[3];
+    }
+
+    T* output_data = output->mutable_data<T>(
+        {n, h_size_data[2], h_size_data[3], 2}, ctx.GetPlace());
+    ScopedSpatialTransformerDescriptor st_desc;
+    cudnnSpatialTransformerDescriptor_t cudnn_st_desc =
+        st_desc.descriptor<T>(4, h_size_data);
+
+    PADDLE_ENFORCE(platform::dynload::cudnnSpatialTfGridGeneratorForward(
+        handle, cudnn_st_desc, theta_data, output_data));
+  }
+};
+
+template <typename T>
+class CUDNNAffineGridGradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use CUDAPlace.");
+    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    auto handle = dev_ctx.cudnn_handle();
+    auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
+    auto theta_grad = ctx.Output<Tensor>(framework::GradVarName("Theta"));
+
+    int n = output_grad->dims()[0];
+    auto size_attr = ctx.Attr<std::vector<int>>("output_shape");
+    Tensor h_sizes;
+    int* h_size_data;
+    if (size_attr.size() == 0) {
+      auto* output_shape = ctx.Input<Tensor>("OutputShape");
+      framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes);
+      h_size_data = h_sizes.data<int>();
+    } else {
+      h_size_data = h_sizes.mutable_data<int>({4}, platform::CPUPlace());
+      h_size_data[0] = n;
+      h_size_data[1] = size_attr[1];
+      h_size_data[2] = size_attr[2];
+      h_size_data[3] = size_attr[3];
+    }
+
+    ScopedSpatialTransformerDescriptor st_desc;
+    cudnnSpatialTransformerDescriptor_t cudnn_st_desc =
+        st_desc.descriptor<T>(4, h_size_data);
+
+    const T* output_grad_data = output_grad->data<T>();
+    T* theta_grad_data = theta_grad->mutable_data<T>(ctx.GetPlace());
+
+    PADDLE_ENFORCE(platform::dynload::cudnnSpatialTfGridGeneratorBackward(
+        handle, cudnn_st_desc, output_grad_data, theta_grad_data));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace plat = paddle::platform;
+REGISTER_OP_KERNEL(affine_grid, CUDNN, plat::CUDAPlace,
+                   paddle::operators::CUDNNAffineGridOpKernel<float>,
+                   paddle::operators::CUDNNAffineGridOpKernel<double>);
+REGISTER_OP_KERNEL(affine_grid_grad, CUDNN, plat::CUDAPlace,
+                   paddle::operators::CUDNNAffineGridGradOpKernel<float>,
+                   paddle::operators::CUDNNAffineGridGradOpKernel<double>);

paddle/fluid/operators/affine_grid_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/affine_grid_op.h"
+#include <string>
+#include "paddle/fluid/framework/op_registry.h"
+#ifdef PADDLE_WITH_CUDA
+#include "paddle/fluid/platform/cudnn_helper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T>
+struct Linspace<paddle::platform::CPUDeviceContext, T> {
+  framework::Tensor operator()(T start, T end, int count,
+                               const framework::ExecutionContext& ctx) {
+    Tensor numbers;
+    T* number_data = numbers.mutable_data<T>({count}, platform::CPUPlace());
+    T slice = (end - start) / (T)(count - 1);
+    for (int i = 0; i < count; ++i) {
+      number_data[i] = start + (T)i * slice;
+    }
+    return numbers;
+  }
+};
+
+class AffineGridOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Theta"),
+                   "Input(Theta) of AffineGridOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Output"),
+                   "Output(Output) of AffineGridOp should not be null.");
+    auto theta_dims = ctx->GetInputDim("Theta");
+    PADDLE_ENFORCE(theta_dims.size() == 3,
+                   "AffineGrid's Input(Theta) should be 3-D tensor.");
+
+    auto output_shape = ctx->Attrs().Get<std::vector<int>>("output_shape");
+    if (output_shape.size() == 0) {
+      PADDLE_ENFORCE(ctx->HasInput("OutputShape"),
+                     "Input(OutputShape) of AffineGridOp should not be null if "
+                     "attr(output_shape) is not configured.");
+      auto output_shape_dims = ctx->GetInputDim("OutputShape");
+      PADDLE_ENFORCE(output_shape_dims.size() == 1,
+                     "AffineGrid's Input(OutputShape) should be 1-D tensor.");
+    } else {
+      PADDLE_ENFORCE(output_shape.size() == 4,
+                     "The size of attr(output_shape) should be 4.");
+    }
+
+    PADDLE_ENFORCE(theta_dims[1] == 2, "Input(theta) dims[1] should be 2.");
+    PADDLE_ENFORCE(theta_dims[2] == 3, "Input(theta) dims[2] should be 3.");
+    // N * H * W * 2
+    ctx->SetOutputDim("Output",
+                      framework::make_ddim({theta_dims[0], -1, -1, 2}));
+    ctx->ShareLoD("Theta", "Output");
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    framework::LibraryType library{framework::LibraryType::kPlain};
+#ifdef PADDLE_WITH_CUDA
+    if (platform::CanCUDNNBeUsed(ctx)) {
+      library = framework::LibraryType::kCUDNN;
+    }
+#endif
+    auto data_type = framework::ToDataType(ctx.Input<Tensor>("Theta")->type());
+    return framework::OpKernelType(data_type, ctx.GetPlace(),
+                                   framework::DataLayout::kAnyLayout, library);
+  }
+};
+
+class AffineGridOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput(
+        "Theta",
+        "(Tensor) A batch of affine transform parameters with shape [N, 2, 3]. "
+        "It is used to transform coordinate (x_0, y_0) to coordinate (x_1, "
+        "y_1).");
+    AddInput("OutputShape",
+             "(Tensor) The shape of target image with format [N, C, H, W].")
+        .AsDispensable();
+    AddOutput("Output", "(Tensor) Output Tensor with shape [N, H, W, 2].");
+    AddAttr<bool>(
+        "use_cudnn",
+        "(bool, default false) Only used in cudnn kernel, need install cudnn")
+        .SetDefault(true);
+    AddAttr<std::vector<int>>(
+        "output_shape",
+        "The target output image shape with format [N, C, H, W].")
+        .SetDefault(std::vector<int>());
+
+    AddComment(R"DOC(
+    It generates a grid of (x,y) coordinates using the parameters of the
+    affine transformation that correspond to a set of points where the input
+    feature map should be sampled to produce the transformed output feature map.
+
+    Given:
+        Theta = [[[x_11, x_12, x_13]
+                  [x_14, x_15, x_16]]
+                 [[x_21, x_22, x_23]
+                  [x_24, x_25, x_26]]]
+    
+        OutputShape = [2, 3, 5, 5]
+
+    Step 1:
+
+        Generate relative coordinates according to OutputShape.
+        The values of relative coordinates are in the interval between -1 and 1.
+        The shape of the relative coordinates is [2, H, W] as below:
+    
+        C = [[[-1.  -1.  -1.  -1.  -1. ]
+              [-0.5 -0.5 -0.5 -0.5 -0.5]
+              [ 0.   0.   0.   0.   0. ]
+              [ 0.5  0.5  0.5  0.5  0.5]
+              [ 1.   1.   1.   1.   1. ]] 
+             [[-1.  -0.5  0.   0.5  1. ]
+              [-1.  -0.5  0.   0.5  1. ]
+              [-1.  -0.5  0.   0.5  1. ]
+              [-1.  -0.5  0.   0.5  1. ]
+              [-1.  -0.5  0.   0.5  1. ]]]
+        C[0] is the coordinates in height axis and  C[1] is the coordinates in width axis.
+    
+    Step2:
+        Tanspose and reshape C to shape [H * W, 2] and append ones to last dimension. The we get:
+        C_ = [[-1.  -1.   1. ]
+              [-0.5 -1.   1. ]
+              [ 0.  -1.   1. ]
+              [ 0.5 -1.   1. ]
+              [ 1.  -1.   1. ]
+              [-1.  -0.5  1. ]
+              [-0.5 -0.5  1. ]
+              [ 0.  -0.5  1. ]
+              [ 0.5 -0.5  1. ]
+              [ 1.  -0.5  1. ]
+              [-1.   0.   1. ]
+              [-0.5  0.   1. ]
+              [ 0.   0.   1. ]
+              [ 0.5  0.   1. ]
+              [ 1.   0.   1. ]
+              [-1.   0.5  1. ]
+              [-0.5  0.5  1. ]
+              [ 0.   0.5  1. ]
+              [ 0.5  0.5  1. ]
+              [ 1.   0.5  1. ]
+              [-1.   1.   1. ]
+              [-0.5  1.   1. ]
+              [ 0.   1.   1. ]
+              [ 0.5  1.   1. ]
+              [ 1.   1.   1. ]]
+    Step3:
+        Compute output by equation $$Output[i] = C_ * Theta[i]^T$$
+    )DOC");
+  }
+};
+
+class AffineGridOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    auto theta_dims = ctx->GetInputDim("Theta");
+    if (ctx->HasOutput(framework::GradVarName("Theta"))) {
+      ctx->SetOutputDim(framework::GradVarName("Theta"), theta_dims);
+    }
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    framework::LibraryType library_{framework::LibraryType::kPlain};
+#ifdef PADDLE_WITH_CUDA
+    if (platform::CanCUDNNBeUsed(ctx)) {
+      library_ = framework::LibraryType::kCUDNN;
+    }
+#endif
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("Theta")->type()),
+        ctx.GetPlace(), framework::DataLayout::kAnyLayout, library_);
+  }
+};
+
+class AffineGridGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto* op = new framework::OpDesc();
+    op->SetType("affine_grid_grad");
+    op->SetInput("Theta", Input("Theta"));
+    op->SetInput("OutputShape", Input("OutputShape"));
+    op->SetInput(framework::GradVarName("Output"), OutputGrad("Output"));
+
+    op->SetAttrMap(Attrs());
+
+    op->SetOutput(framework::GradVarName("Theta"), InputGrad("Theta"));
+    return std::unique_ptr<framework::OpDesc>(op);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(affine_grid, ops::AffineGridOp, ops::AffineGridOpMaker,
+                  ops::AffineGridGradMaker);
+REGISTER_OPERATOR(affine_grid_grad, ops::AffineGridOpGrad);
+
+REGISTER_OP_CPU_KERNEL(
+    affine_grid,
+    ops::AffineGridOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::AffineGridOpKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    affine_grid_grad,
+    ops::AffineGridGradOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::AffineGridGradOpKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/affine_grid_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 <vector>
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+template <typename T, size_t D, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
+
+using Array1 = Eigen::DSizes<int64_t, 1>;
+using Array2 = Eigen::DSizes<int64_t, 2>;
+using Array3 = Eigen::DSizes<int64_t, 3>;
+using Array4 = Eigen::DSizes<int64_t, 4>;
+
+/**
+ *Return a tensor with evenly spaced numbers over a specified interval.
+ */
+template <typename DeviceContext, typename T>
+struct Linspace {
+  framework::Tensor operator()(T start, T end, int count,
+                               const framework::ExecutionContext& ctx);
+};
+
+template <typename DeviceContext, typename T>
+class AffineGridOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
+    auto* theta = ctx.Input<Tensor>("Theta");
+    int n = theta->dims()[0];
+
+    auto size_attr = ctx.Attr<std::vector<int>>("output_shape");
+    int h = 0;
+    int w = 0;
+    if (size_attr.size() == 0) {
+      auto* output_shape = ctx.Input<Tensor>("OutputShape");
+      Tensor h_sizes;
+      framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes);
+      const int* h_size_data = h_sizes.data<int>();
+      h = h_size_data[2];
+      w = h_size_data[3];
+    } else {
+      h = size_attr[2];
+      w = size_attr[3];
+    }
+
+    auto* output = ctx.Output<Tensor>("Output");
+    output->mutable_data<T>({n, h, w, 2}, ctx.GetPlace());
+
+    math::SetConstant<DeviceContext, T>()(
+        ctx.template device_context<DeviceContext>(), output,
+        static_cast<T>(0));
+
+    Linspace<DeviceContext, T> linspace;
+    // Get indexes of height with shape [height, width, 1]
+    auto h_idx = linspace((T)-1, (T)1, h, ctx);
+    auto h_idx_t = EigenTensor<T, 1>::From(h_idx);
+    // Get indexes of width with shape [height, width, 1]
+    auto w_idx = linspace((T)-1, (T)1, w, ctx);
+    auto w_idx_t = EigenTensor<T, 1>::From(w_idx);
+    // Get constant ones tensor with shape [height, width, 1]
+    Tensor ones;
+    ones.mutable_data<T>({h, w, 1}, ctx.GetPlace());
+    auto ones_t = EigenTensor<T, 3>::From(ones).setConstant((T)1);
+    // Get grid tensor with shape [n, h, w, 3] by concatenating h_idx, w_idx and
+    // ones
+    Tensor grid;
+    grid.mutable_data<T>({n, h, w, 3}, ctx.GetPlace());
+    auto grid_t = EigenTensor<T, 4>::From(grid);
+
+    grid_t.device(place) = w_idx_t.reshape(Array2(1, w))
+                               .broadcast(Array2(h, 1))
+                               .reshape(Array3(h, w, 1))
+                               .concatenate(h_idx_t.reshape(Array2(1, h))
+                                                .broadcast(Array2(w, 1))
+                                                .shuffle(Array2(1, 0))
+                                                .reshape(Array3(h, w, 1)),
+                                            2)
+                               .eval()
+                               .concatenate(ones_t, 2)
+                               .reshape(Array4(1, h, w, 3))
+                               .broadcast(Array4(n, 1, 1, 1));
+
+    // output = grid * theta.T
+    // TODO(wanghaoshuang): Refine batched matrix multiply
+    auto blas = math::GetBlas<DeviceContext, T>(ctx);
+    for (int i = 0; i < n; ++i) {
+      Tensor sliced_grid = grid.Slice(i, i + 1).Resize({h * w, 3});
+      Tensor sliced_theta = theta->Slice(i, i + 1).Resize({2, 3});
+      Tensor sliced_out = output->Slice(i, i + 1).Resize({h * w, 2});
+      blas.MatMul(sliced_grid, false, sliced_theta, true, T(1), &sliced_out,
+                  T(0));
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+class AffineGridGradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
+    auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
+    auto theta_grad = ctx.Output<Tensor>(framework::GradVarName("Theta"));
+
+    int n = output_grad->dims()[0];
+    auto size_attr = ctx.Attr<std::vector<int>>("output_shape");
+    int h = 0;
+    int w = 0;
+    if (size_attr.size() == 0) {
+      auto* output_shape = ctx.Input<Tensor>("OutputShape");
+      Tensor h_sizes;
+      framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes);
+      const int* h_size_data = h_sizes.data<int>();
+      h = h_size_data[2];
+      w = h_size_data[3];
+    } else {
+      h = size_attr[2];
+      w = size_attr[3];
+    }
+
+    theta_grad->mutable_data<T>({n, 2, 3}, ctx.GetPlace());
+
+    math::SetConstant<DeviceContext, T>()(
+        ctx.template device_context<DeviceContext>(), theta_grad,
+        static_cast<T>(0));
+
+    Linspace<DeviceContext, T> linspace;
+
+    // Get indexes of height with shape [height, width, 1]
+    auto h_idx = linspace((T)-1, (T)1, h, ctx);
+    auto h_idx_t = EigenTensor<T, 1>::From(h_idx);
+    // Get indexes of width with shape [height, width, 1]
+    auto w_idx = linspace((T)-1, (T)1, w, ctx);
+    auto w_idx_t = EigenTensor<T, 1>::From(w_idx);
+    // Get constant ones tensor with shape [height, width, 1]
+    Tensor ones;
+    ones.mutable_data<T>({h, w, 1}, ctx.GetPlace());
+    auto ones_t = EigenTensor<T, 3>::From(ones).setConstant((T)1);
+    // Get grid tensor with shape [n, h, w, 3] by concatenating h_idx, w_idx and
+    // ones
+    Tensor grid;
+    grid.mutable_data<T>({n, h, w, 3}, ctx.GetPlace());
+    auto grid_t = EigenTensor<T, 4>::From(grid);
+    grid_t.device(place) = w_idx_t.reshape(Array2(1, w))
+                               .broadcast(Array2(h, 1))
+                               .reshape(Array3(h, w, 1))
+                               .concatenate(h_idx_t.reshape(Array2(1, h))
+                                                .broadcast(Array2(w, 1))
+                                                .shuffle(Array2(1, 0))
+                                                .reshape(Array3(h, w, 1)),
+                                            2)
+                               .eval()
+                               .concatenate(ones_t, 2)
+                               .reshape(Array4(1, h, w, 3))
+                               .broadcast(Array4(n, 1, 1, 1));
+    // output = grid * theta.T
+    // TODO(wanghaoshuang): Refine batched matrix multiply
+    auto blas = math::GetBlas<DeviceContext, T>(ctx);
+    for (int i = 0; i < n; ++i) {
+      Tensor sliced_grid = grid.Slice(i, i + 1).Resize({h * w, 3});
+      Tensor sliced_out_grad = output_grad->Slice(i, i + 1).Resize({h * w, 2});
+      Tensor sliced_theta_grad = theta_grad->Slice(i, i + 1).Resize({2, 3});
+      blas.MatMul(sliced_out_grad, true, sliced_grid, false, T(1),
+                  &sliced_theta_grad, T(0));
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/delete_var_op.cc

@@ -32,6 +32,11 @@ class DeleteVarOp : public framework::OperatorBase {
   }
 };
 
+class DeleteVarOpShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *ctx) const override {}
+};
+
 class DeleteVarOpInfoMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
@@ -48,4 +53,5 @@ It should not be configured by users directly.
 
 REGISTER_OPERATOR(delete_var, paddle::operators::DeleteVarOp,
                   paddle::framework::EmptyGradOpMaker,
-                  paddle::operators::DeleteVarOpInfoMaker);
+                  paddle::operators::DeleteVarOpInfoMaker,
+                  paddle::operators::DeleteVarOpShapeInference);

paddle/fluid/operators/sum_op.cc

@@ -67,6 +67,7 @@ class SumOp : public framework::OperatorWithKernel {
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     auto x_vars = ctx.MultiInputVar("X");
+    auto x_vars_name = ctx.Inputs("X");
 
     framework::LibraryType library{framework::LibraryType::kPlain};
     framework::DataLayout layout{framework::DataLayout::kAnyLayout};
@@ -81,10 +82,11 @@ class SumOp : public framework::OperatorWithKernel {
 
     if (x_vars[0]->IsType<framework::LoDTensor>()) {
       int dtype = -1;
-      for (auto& x_var : x_vars) {
+      for (size_t idx = 0; idx < x_vars.size(); ++idx) {
+        PADDLE_ENFORCE(x_vars[idx] != nullptr,
+                       "Input var[%s] should not be nullptr", x_vars_name[idx]);
         // FIXME(zcd): The input x_var may be SelectedRows or LoDTensor.
-        auto tensor = framework::GetTensorFromVar(
-            const_cast<framework::Variable*>(x_var));
+        auto tensor = framework::GetTensorFromVar(*x_vars[idx]);
         if (tensor->numel() == 0) {
           continue;
         }

paddle/fluid/platform/cudnn_helper.h

@@ -341,6 +341,28 @@ class ScopedPoolingDescriptor {
   DISABLE_COPY_AND_ASSIGN(ScopedPoolingDescriptor);
 };
 
+class ScopedSpatialTransformerDescriptor {
+ public:
+  ScopedSpatialTransformerDescriptor() {
+    PADDLE_ENFORCE(dynload::cudnnCreateSpatialTransformerDescriptor(&desc_));
+  }
+  ~ScopedSpatialTransformerDescriptor() {
+    PADDLE_ENFORCE(dynload::cudnnDestroySpatialTransformerDescriptor(desc_));
+  }
+
+  template <typename T>
+  inline cudnnSpatialTransformerDescriptor_t descriptor(const int nbDims,
+                                                        const int dimA[]) {
+    PADDLE_ENFORCE(dynload::cudnnSetSpatialTransformerNdDescriptor(
+        desc_, CUDNN_SAMPLER_BILINEAR, CudnnDataType<T>::type, nbDims, dimA));
+    return desc_;
+  }
+
+ private:
+  cudnnSpatialTransformerDescriptor_t desc_;
+  DISABLE_COPY_AND_ASSIGN(ScopedSpatialTransformerDescriptor);
+};
+
 inline bool CanCUDNNBeUsed(const framework::ExecutionContext& ctx) {
   bool use_cudnn = ctx.Attr<bool>("use_cudnn");
   use_cudnn &= paddle::platform::is_gpu_place(ctx.GetPlace());

paddle/fluid/platform/dynload/cudnn.h

@@ -90,6 +90,13 @@ extern void EnforceCUDNNLoaded(const char* fn_name);
   __macro(cudnnSetConvolutionNdDescriptor);          \
   __macro(cudnnGetConvolutionNdDescriptor);          \
   __macro(cudnnDeriveBNTensorDescriptor);            \
+  __macro(cudnnCreateSpatialTransformerDescriptor);  \
+  __macro(cudnnSetSpatialTransformerNdDescriptor);   \
+  __macro(cudnnDestroySpatialTransformerDescriptor); \
+  __macro(cudnnSpatialTfGridGeneratorForward);       \
+  __macro(cudnnSpatialTfGridGeneratorBackward);      \
+  __macro(cudnnSpatialTfSamplerForward);             \
+  __macro(cudnnSpatialTfSamplerBackward);            \
   __macro(cudnnCreate);                              \
   __macro(cudnnDestroy);                             \
   __macro(cudnnSetStream);                           \

paddle/scripts/paddle_build.sh

@@ -147,7 +147,6 @@ function cmake_gen() {
         -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON}
         -DCUDNN_ROOT=/usr/
         -DWITH_TESTING=${WITH_TESTING:-ON}
-        -DWITH_FAST_BUNDLE_TEST=ON
         -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake
         -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
         -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF}
@@ -180,7 +179,6 @@ EOF
         -DWITH_PYTHON=${WITH_PYTHON:-ON} \
         -DCUDNN_ROOT=/usr/ \
         -DWITH_TESTING=${WITH_TESTING:-ON} \
-        -DWITH_FAST_BUNDLE_TEST=ON \
         -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake \
         -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \
         -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \

python/paddle/fluid/io.py

@@ -884,12 +884,13 @@ def _load_slice_up_vars(executor, dirname, slice_vars_and_attrs):
 
     load_prog = Program()
     load_block = load_prog.global_block()
+    need_delete_vars = []
 
     for var_tuple in slice_vars_and_attrs:
         orig_var = var_tuple[0]
         start = var_tuple[1]
         slice_var = var_tuple[2]
-        end = start + reduce(lambda x, y: x * y, slice_var.shape)
+        end = start + slice_var.shape[0]
 
         clone_orig_var = load_block.create_var(
             name=orig_var.name,
@@ -917,5 +918,8 @@ def _load_slice_up_vars(executor, dirname, slice_vars_and_attrs):
             attrs={'axes': [0],
                    'starts': [start],
                    'ends': [end]})
-
+        need_delete_vars.append(clone_orig_var)
+    load_block.append_op(
+        type='delete_var',
+        inputs={'X': need_delete_vars}, )
     executor.run(load_prog)

python/paddle/fluid/layers/nn.py

@@ -155,6 +155,7 @@ __all__ = [
     'sigmoid_cross_entropy_with_logits',
     'maxout',
     'space_to_depth',
+    'affine_grid',
     'sequence_reverse',
     'affine_channel',
     'hash',
@@ -711,8 +712,18 @@ def dynamic_gru(input,
               The first part are weights of the update gate and reset gate with
               shape :math:`(D \\times 2D)`, and the second part are weights for
               candidate hidden state with shape :math:`(D \\times D)`.
-        bias_attr(ParamAttr): The parameter attribute for learnable the
-            hidden-hidden bias.
+
+            If it is set to None or one attribute of ParamAttr, dynamic_gru will
+            create ParamAttr as param_attr. If the Initializer of the param_attr
+            is not set, the parameter is initialized with Xavier. Default: None.
+        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias
+            of GRU. Note that the bias with :math:`(1 \\times 3D)` concatenates 
+            the bias in the update gate, reset gate and candidate calculations.
+            If it is set to False, no bias will be applied to the update gate, 
+            reset gate and candidate calculations. If it is set to None or one 
+            attribute of ParamAttr, dynamic_gru will create ParamAttr as 
+            bias_attr. If the Initializer of the bias_attr is not set, the bias
+            is initialized zero. Default: None.
         is_reverse(bool): Whether to compute reversed GRU, default
             :attr:`False`.
         gate_activation(str): The activation for update gate and reset gate.
@@ -811,10 +822,29 @@ def gru_unit(input,
 
     Args:
         input (Variable): The fc transformed input value of current step.
-        hidden (Variable): The hidden value of lstm unit from previous step.
+        hidden (Variable): The hidden value of gru unit from previous step.
         size (integer): The input dimension value.
-        param_attr (ParamAttr): The weight parameters for gru unit. Default: None
-        bias_attr (ParamAttr): The bias parameters for gru unit. Default: None
+        param_attr(ParamAttr|None): The parameter attribute for the learnable
+            hidden-hidden weight matrix. Note:
+
+            - The shape of the weight matrix is :math:`(T \\times 3D)`, where
+              :math:`D` is the hidden size.
+            - All elements in the weight matrix can be divided into two parts.
+              The first part are weights of the update gate and reset gate with
+              shape :math:`(D \\times 2D)`, and the second part are weights for
+              candidate hidden state with shape :math:`(D \\times D)`.
+
+            If it is set to None or one attribute of ParamAttr, gru_unit will
+            create ParamAttr as param_attr. If the Initializer of the param_attr
+            is not set, the parameter is initialized with Xavier. Default: None.
+        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias
+            of GRU. Note that the bias with :math:`(1 \\times 3D)` concatenates 
+            the bias in the update gate, reset gate and candidate calculations.
+            If it is set to False, no bias will be applied to the update gate, 
+            reset gate and candidate calculations. If it is set to None or one 
+            attribute of ParamAttr, gru_unit will create ParamAttr as 
+            bias_attr. If the Initializer of the bias_attr is not set, the bias
+            is initialized zero. Default: None.
         activation (string): The activation type for cell (actNode).
                              Default: 'tanh'
         gate_activation (string): The activation type for gates (actGate).
@@ -4444,7 +4474,10 @@ def transpose(x, perm, name=None):
     Examples:
         .. code-block:: python
 
-            x = fluid.layers.data(name='x', shape=[5, 10, 15], dtype='float32')
+            # use append_batch_size=False to avoid prepending extra 
+            # batch size in shape
+            x = fluid.layers.data(name='x', shape=[5, 10, 15], 
+                            dtype='float32', append_batch_size=False)
             x_transposed = layers.transpose(x, perm=[1, 0, 2])
     """
 
@@ -6109,6 +6142,124 @@ def crop(x, shape=None, offsets=None, name=None):
     return out
 
 
+def affine_grid(theta, out_shape, name=None):
+    """
+    It generates a grid of (x,y) coordinates using the parameters of
+    the affine transformation that correspond to a set of points where
+    the input feature map should be sampled to produce the transformed
+    output feature map.
+
+    .. code-block:: text
+
+        * Case 1:
+
+          Given:
+
+              theta = [[[x_11, x_12, x_13]
+                        [x_14, x_15, x_16]]
+                       [[x_21, x_22, x_23]
+                        [x_24, x_25, x_26]]]
+      
+              out_shape = [2, 3, 5, 5]
+      
+          Step 1:
+      
+              Generate normalized coordinates according to out_shape.
+              The values of the normalized coordinates are in the interval between -1 and 1.
+              The shape of the normalized coordinates is [2, H, W] as below:
+      
+              C = [[[-1.  -1.  -1.  -1.  -1. ]
+                    [-0.5 -0.5 -0.5 -0.5 -0.5]
+                    [ 0.   0.   0.   0.   0. ]
+                    [ 0.5  0.5  0.5  0.5  0.5]
+                    [ 1.   1.   1.   1.   1. ]]
+                   [[-1.  -0.5  0.   0.5  1. ]
+                    [-1.  -0.5  0.   0.5  1. ]
+                    [-1.  -0.5  0.   0.5  1. ]
+                    [-1.  -0.5  0.   0.5  1. ]
+                    [-1.  -0.5  0.   0.5  1. ]]]
+              C[0] is the coordinates in height axis and  C[1] is the coordinates in width axis.
+
+          Step2:
+
+              Tanspose and reshape C to shape [H * W, 2] and append ones to last dimension. The we get:
+              C_ = [[-1.  -1.   1. ]
+                    [-0.5 -1.   1. ]
+                    [ 0.  -1.   1. ]
+                    [ 0.5 -1.   1. ]
+                    [ 1.  -1.   1. ]
+                    [-1.  -0.5  1. ]
+                    [-0.5 -0.5  1. ]
+                    [ 0.  -0.5  1. ]
+                    [ 0.5 -0.5  1. ]
+                    [ 1.  -0.5  1. ]
+                    [-1.   0.   1. ]
+                    [-0.5  0.   1. ]
+                    [ 0.   0.   1. ]
+                    [ 0.5  0.   1. ]
+                    [ 1.   0.   1. ]
+                    [-1.   0.5  1. ]
+                    [-0.5  0.5  1. ]
+                    [ 0.   0.5  1. ]
+                    [ 0.5  0.5  1. ]
+                    [ 1.   0.5  1. ]
+                    [-1.   1.   1. ]
+                    [-0.5  1.   1. ]
+                    [ 0.   1.   1. ]
+                    [ 0.5  1.   1. ]
+                    [ 1.   1.   1. ]]
+          Step3:
+              Compute output by equation $$Output[i] = C_ * Theta[i]^T$$
+
+    Args:
+        theta (Variable): A batch of affine transform parameters with shape [N, 2, 3].
+        out_shape (Variable | list | tuple): The shape of target output with format [N, C, H, W].
+        out_shape can be a Variable or a list or tuple.
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+
+    Returns:
+        Variable: The output with shape [N, H, W, 2].
+
+    Raises:
+        ValueError: If the type of arguments is not supported.
+
+    Examples:
+
+        .. code-block:: python
+            theta = fluid.layers.data(name="x", shape=[2, 3], dtype="float32")
+            out_shape = fluid.layers.data(name="y", shape=[-1], dtype="float32")
+            data = fluid.layers.affine_grid(theta, out_shape)
+
+            # or
+            data = fluid.layers.affine_grid(theta, [5, 3, 28, 28])
+
+    """
+    helper = LayerHelper('affine_grid')
+
+    if not (isinstance(out_shape, list) or isinstance(out_shape, tuple) or \
+        isinstance(out_shape, Variable)):
+        raise ValueError("The out_shape should be a list, tuple or Variable.")
+
+    if not isinstance(theta, Variable):
+        raise ValueError("The theta should be a Variable.")
+
+    out = helper.create_variable_for_type_inference(theta.dtype)
+    ipts = {'Theta': theta}
+    attrs = {}
+    if isinstance(out_shape, Variable):
+        ipts['OutputShape'] = out_shape
+    else:
+        attrs['output_shape'] = out_shape
+
+    helper.append_op(
+        type='affine_grid',
+        inputs=ipts,
+        outputs={'Output': out},
+        attrs=None if len(attrs) == 0 else attrs)
+    return out
+
+
 def rank_loss(label, left, right, name=None):
     """
     **Rank loss layer for RankNet**

python/paddle/fluid/tests/unittests/dist_save_load.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 os
+import sys
+import signal
+import subprocess
+import argparse
+import time
+import math
+import random
+from multiprocessing import Process
+from functools import reduce
+
+import numpy as np
+import unittest
+import six
+
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid import core
+from paddle.fluid import io
+
+from test_dist_base import TestDistRunnerBase, runtime_main, RUN_STEP
+from dist_simnet_bow import TestDistSimnetBow2x2, DATA_URL, DATA_MD5
+
+
+class TestDistSaveLoad2x2(TestDistSimnetBow2x2):
+    def _load_persistable_vars(self, executor, dirname, program):
+        def _is_checkpoint_var(var):
+            """
+            the checkpoint will not save or load all the variables.
+            var type is FEED_MINIBATCH/FETCH_LIST/RAW or var name ends with @GRAD are discarded.
+
+            : param var(Variable)
+            """
+            if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \
+                    var.desc.type() == core.VarDesc.VarType.FETCH_LIST or \
+                    var.desc.type() == core.VarDesc.VarType.RAW:
+                return False
+            # @GRAD are named for gradient variables, checkpoint will not save it.
+            if "@GRAD" in var.name:
+                return False
+            # .trainer_ are named for distribute train variables, checkpoint will not save it.
+            if ".trainer_" in var.name:
+                return False
+
+            # .block is named for distribute train variables, checkpoint will not save it.
+            if ".block" in var.name:
+                return False
+
+            if "tmp_" in var.name:
+                return False
+
+            return var.persistable
+
+        io.load_vars(
+            executor,
+            dirname=dirname,
+            main_program=program,
+            predicate=_is_checkpoint_var,
+            filename=None)
+
+    def run_pserver(self, args):
+        self.get_model(batch_size=2)
+        # NOTE: pserver should not call memory optimize
+        t = self.get_transpiler(args.trainer_id,
+                                fluid.default_main_program(), args.endpoints,
+                                args.trainers, args.sync_mode)
+        pserver_prog = t.get_pserver_program(args.current_endpoint)
+        startup_prog = t.get_startup_program(args.current_endpoint,
+                                             pserver_prog)
+
+        need_load = bool(int(os.getenv("LOAD", "0")))
+        model_dir = os.getenv("MODEL_DIR", "")
+
+        place = fluid.CPUPlace()
+        exe = fluid.Executor(place)
+        exe.run(startup_prog)
+
+        if need_load and model_dir:
+            self._load_persistable_vars(exe, model_dir, startup_prog)
+        exe.run(pserver_prog)
+
+    def run_trainer(self, args):
+        test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \
+            self.get_model(batch_size=2)
+
+        if args.mem_opt:
+            fluid.memory_optimize(fluid.default_main_program(), skip_grads=True)
+        if args.is_dist:
+            t = self.get_transpiler(args.trainer_id,
+                                    fluid.default_main_program(),
+                                    args.endpoints, args.trainers,
+                                    args.sync_mode)
+
+            trainer_prog = t.get_trainer_program()
+        else:
+            trainer_prog = fluid.default_main_program()
+
+        if args.use_cuda:
+            place = fluid.CUDAPlace(0)
+        else:
+            place = fluid.CPUPlace()
+
+        startup_exe = fluid.Executor(place)
+        startup_exe.run(fluid.default_startup_program())
+
+        strategy = fluid.ExecutionStrategy()
+        strategy.num_threads = 1
+        strategy.allow_op_delay = False
+
+        build_stra = fluid.BuildStrategy()
+
+        if args.use_reduce:
+            build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
+        else:
+            build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce
+
+        exe = fluid.ParallelExecutor(
+            args.use_cuda,
+            loss_name=avg_cost.name,
+            exec_strategy=strategy,
+            build_strategy=build_stra)
+
+        feed_var_list = [
+            var for var in trainer_prog.global_block().vars.values()
+            if var.is_data
+        ]
+
+        feeder = fluid.DataFeeder(feed_var_list, place)
+        reader_generator = train_reader()
+
+        def get_data():
+            origin_batch = next(reader_generator)
+            if args.is_dist and args.use_reader_alloc:
+                new_batch = []
+                for offset, item in enumerate(origin_batch):
+                    if offset % 2 == args.trainer_id:
+                        new_batch.append(item)
+                return new_batch
+            else:
+                return origin_batch
+
+        need_save = bool(int(os.getenv("SAVE", "0")))
+        model_dir = os.getenv("MODEL_DIR", "")
+
+        if need_save:
+            for _ in six.moves.xrange(RUN_STEP):
+                loss, = exe.run(fetch_list=[avg_cost.name],
+                                feed=feeder.feed(get_data()))
+            if need_save and model_dir:
+                io.save_persistables(startup_exe, model_dir, trainer_prog)
+
+        var = np.array(fluid.global_scope().find_var('__fc_b__').get_tensor())
+        print(np.ravel(var).tolist())
+
+
+if __name__ == "__main__":
+    paddle.dataset.common.download(DATA_URL, 'simnet', DATA_MD5, "train")
+    runtime_main(TestDistSaveLoad2x2)

python/paddle/fluid/tests/unittests/test_affine_grid_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 unittest
+import numpy as np
+from op_test import OpTest
+
+
+def AffineGrid(theta, size):
+    n = size[0]
+    w = size[3]
+    h = size[2]
+    h_idx = np.repeat(
+        np.linspace(-1, 1, h)[np.newaxis, :], w, axis=0).T[:, :, np.newaxis]
+    w_idx = np.repeat(
+        np.linspace(-1, 1, w)[np.newaxis, :], h, axis=0)[:, :, np.newaxis]
+    grid = np.concatenate(
+        [w_idx, h_idx, np.ones([h, w, 1])], axis=2)  # h * w * 3
+    grid = np.repeat(grid[np.newaxis, :], size[0], axis=0)  # n * h * w *3
+
+    ret = np.zeros([n, h * w, 2])
+    theta = theta.transpose([0, 2, 1])
+    for i in range(len(theta)):
+        ret[i] = np.dot(grid[i].reshape([h * w, 3]), theta[i])
+
+#    print ret.reshape([h * w, 2]).astype("float32")    
+    return ret.reshape([n, h, w, 2]).astype("float32")
+
+
+class TestAffineGridOp(OpTest):
+    def setUp(self):
+        self.initTestCase()
+        self.op_type = "affine_grid"
+        theta = np.random.randint(1, 3, self.theta_shape).astype("float32")
+        theta = np.ones(self.theta_shape).astype("float32")
+        self.inputs = {'Theta': theta}
+        self.attrs = {"use_cudnn": True}
+        if self.dynamic_shape:
+            self.inputs['OutputShape'] = self.output_shape
+        else:
+            self.attrs['output_shape'] = self.output_shape
+        self.outputs = {'Output': AffineGrid(theta, self.output_shape)}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(
+            ['Theta'],
+            'Output',
+            no_grad_set=['OutputShape'],
+            max_relative_error=0.006)
+
+    def initTestCase(self):
+        self.theta_shape = (3, 2, 3)
+        self.output_shape = np.array([3, 2, 5, 7]).astype("int32")
+        self.dynamic_shape = False
+
+
+class TestAffineGridOpCase1(TestAffineGridOp):
+    def initTestCase(self):
+        self.theta_shape = (3, 2, 3)
+        self.output_shape = np.array([3, 2, 5, 7]).astype("int32")
+        self.dynamic_shape = True
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_dist_save_load.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 os
+import shutil
+import unittest
+import tempfile
+
+import numpy as np
+
+from test_dist_base import TestDistBase, RUN_STEP
+
+
+class TestDistSaveLoadDense2x2(TestDistBase):
+    def _setup_config(self):
+        self._sync_mode = True
+        self._enforce_place = "CPU"
+
+    def check_with_place(self,
+                         model_file,
+                         delta=1e-3,
+                         check_error_log=False,
+                         need_envs={}):
+
+        required_envs = {
+            "PATH": os.getenv("PATH", ""),
+            "PYTHONPATH": os.getenv("PYTHONPATH", ""),
+            "LD_LIBRARY_PATH": os.getenv("LD_LIBRARY_PATH", ""),
+            "http_proxy": ""
+        }
+
+        required_envs.update(need_envs)
+
+        if check_error_log:
+            required_envs["GLOG_v"] = "7"
+            required_envs["GLOG_logtostderr"] = "1"
+
+        model_dir = tempfile.mkdtemp()
+
+        local_env = {}
+        local_env["SAVE"] = "1"
+        local_env["MODEL_DIR"] = model_dir
+        local_env.update(required_envs)
+
+        cluster_env = {}
+        cluster_env["LOAD"] = "1"
+        cluster_env["MODEL_DIR"] = model_dir
+        cluster_env.update(required_envs)
+
+        local_var = self._run_local(model_file, local_env, check_error_log)
+        tr0_var, tr1_var = self._run_cluster(model_file, cluster_env,
+                                             check_error_log)
+
+        shutil.rmtree(model_dir)
+
+        local_np = np.array(eval(local_var[0]))
+        train0_np = np.array(eval(tr0_var[0]))
+        train1_np = np.array(eval(tr1_var[0]))
+        self.assertAlmostEqual(local_np.all(), train0_np.all(), delta=delta)
+        self.assertAlmostEqual(local_np.all(), train1_np.all(), delta=delta)
+        self.assertAlmostEqual(train0_np.all(), train1_np.all(), delta=delta)
+
+    def test_dist(self):
+        need_envs = {
+            "IS_DISTRIBUTED": '0',
+            "IS_SPARSE": '0',
+            'IS_SELF_CONTAINED_LR': '1'
+        }
+        self.check_with_place(
+            "dist_save_load.py",
+            delta=0,
+            check_error_log=False,
+            need_envs=need_envs)
+
+
+if __name__ == "__main__":
+    unittest.main()

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

@@ -283,6 +283,25 @@ class TestDecayedAdagrad(TranspilerTest):
         trainer, _ = self.get_trainer()
 
 
+class TestFtrl(TranspilerTest):
+    def net_conf(self):
+        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
+        y_predict = fluid.layers.fc(input=x,
+                                    size=1000,
+                                    act=None,
+                                    param_attr=fluid.ParamAttr(name='fc_w'),
+                                    bias_attr=fluid.ParamAttr(name='fc_b'))
+        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+        avg_cost = fluid.layers.mean(cost)
+        opt = fluid.optimizer.Ftrl(learning_rate=0.1)
+        opt.minimize(avg_cost)
+
+    def transpiler_test_impl(self):
+        pserver, startup = self.get_pserver(self.pserver1_ep)
+        trainer, _ = self.get_trainer()
+
+
 class TestLRDecayConditional(TranspilerTest):
     def net_conf(self):
         x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
@@ -405,18 +424,43 @@ class TestL2DecayWithPiecewise(TranspilerTest):
             ["sum", "scale", "scale", "elementwise_add", "momentum"])
 
 
+class TestEmptyPserverOptimizeBlocks(TranspilerTest):
+    def net_conf(self):
+        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
+        # only one parameter
+        y_predict = fluid.layers.fc(input=x,
+                                    size=1000,
+                                    act=None,
+                                    param_attr=fluid.ParamAttr(name='fc_w'),
+                                    bias_attr=False)
+        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+        avg_cost = fluid.layers.mean(cost)
+        sgd_optimizer = fluid.optimizer.SGD(learning_rate=1.0)
+        sgd_optimizer.minimize(avg_cost)
+
+    def transpiler_test_impl(self):
+        config = fluid.DistributeTranspilerConfig()
+        config.slice_var_up = False
+
+        pserver, startup = self.get_pserver(ep=self.pserver2_ep, config=config)
+
+        self.assertEqual(len(pserver.blocks), 2)
+        self.assertEqual(len(pserver.blocks[1].ops), 0)
+
+
 class TestDistLookupTableBase(TranspilerTest):
     def network_with_table(self, is_sparse, is_distributed):
         self.table_size = 1000
         self.emb_size = 64
         self.lookup_table_name = 'shared_w'
 
-        def emb_pool(ids):
+        def emb_pool(ids, table_name, is_distributed):
             emb = fluid.layers.embedding(
                 input=ids,
                 size=[self.table_size, self.emb_size],
                 dtype='float32',
-                param_attr=self.lookup_table_name,  # share parameter
+                param_attr=table_name,
                 is_sparse=is_sparse,
                 is_distributed=is_distributed)
             pool = fluid.layers.sequence_pool(input=emb, pool_type='average')
@@ -426,9 +470,13 @@ class TestDistLookupTableBase(TranspilerTest):
             name='title_ids', shape=[1], dtype='int64', lod_level=1)
         brand_ids = fluid.layers.data(
             name='brand_ids', shape=[1], dtype='int64', lod_level=1)
-        title_emb = emb_pool(title_ids)
-        brand_emb = emb_pool(brand_ids)
-        fc0 = fluid.layers.concat(input=[title_emb, brand_emb], axis=1)
+        profile_ids = fluid.layers.data(
+            name='brand_ids', shape=[1], dtype='int64', lod_level=1)
+        title_emb = emb_pool(title_ids, self.lookup_table_name, is_distributed)
+        brand_emb = emb_pool(brand_ids, self.lookup_table_name, is_distributed)
+        profile_emb = emb_pool(profile_ids, "profile_emb", False)
+        fc0 = fluid.layers.concat(
+            input=[title_emb, brand_emb, profile_emb], axis=1)
         predict = fluid.layers.fc(input=fc0,
                                   size=2,
                                   act=None,
@@ -449,7 +497,7 @@ class TestLocalLookupTable(TestDistLookupTableBase):
     def transpiler_test_impl(self):
         pserver1, startup1 = self.get_pserver(self.pserver1_ep)
 
-        self.assertEqual(len(pserver1.blocks), 3)
+        self.assertEqual(len(pserver1.blocks), 4)
         # 0 listen_and_serv
         # 1 optimize for fc_w or fc_b adam
         self.assertEqual([op.type for op in pserver1.blocks[1].ops],
@@ -459,16 +507,23 @@ class TestLocalLookupTable(TestDistLookupTableBase):
         self.assertEqual([op.type for op in pserver1.blocks[2].ops],
                          ["sum", "scale", "adam", "scale", "scale"])
 
+        # 3 optimize for table 2 adam
+        # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num
+        self.assertEqual([op.type for op in pserver1.blocks[3].ops],
+                         ["sum", "scale", "adam", "scale", "scale"])
+
         trainer, _ = self.get_trainer()
         self.assertEqual(len(trainer.blocks), 1)
         ops = [
             'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool',
-            'concat', 'mul', 'elementwise_add', 'cross_entropy', 'mean',
-            'fill_constant', 'mean_grad', 'cross_entropy_grad',
-            'elementwise_add_grad', 'send', 'mul_grad', 'send', 'concat_grad',
-            'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad',
-            'lookup_table_grad', 'sum', 'split_selected_rows', 'send',
-            'send_barrier', 'recv', 'recv', 'recv', 'fetch_barrier', 'concat'
+            'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add',
+            'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
+            'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
+            'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
+            'split_selected_rows', 'send', 'sequence_pool_grad',
+            'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad',
+            'sum', 'split_selected_rows', 'send', 'send_barrier', 'recv',
+            'recv', 'recv', 'recv', 'fetch_barrier', 'concat', 'concat'
         ]
         self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
 
@@ -480,39 +535,45 @@ class TestDistLookupTable(TestDistLookupTableBase):
     def transpiler_test_impl(self):
         pserver1, startup1 = self.get_pserver(self.pserver1_ep)
 
-        self.assertEqual(len(pserver1.blocks), 5)
+        self.assertEqual(len(pserver1.blocks), 6)
         # 0 listen_and_serv
         # 1 optimize for fc_w or fc_b adam
         self.assertEqual([op.type for op in pserver1.blocks[1].ops],
                          ["sum", "scale", "adam", "scale", "scale"])
-        # 2 optimize for table sgd
+        # 4 prefetch -> lookup_sparse_table for data0
         self.assertEqual([op.type for op in pserver1.blocks[2].ops],
+                         ["sum", "scale", "adam", "scale", "scale"])
+        # 2 optimize for table sgd
+        self.assertEqual([op.type for op in pserver1.blocks[3].ops],
                          ["sum", "sgd"])
         # 3 prefetch -> lookup_sparse_table for data0
-        self.assertEqual([op.type for op in pserver1.blocks[3].ops],
+        self.assertEqual([op.type for op in pserver1.blocks[4].ops],
                          ["lookup_sparse_table"])
-        # 4 save table
-        self.assertEqual([op.type for op in pserver1.blocks[4].ops], ["save"])
+        # 5 save table
+        self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"])
 
         trainer, trainer_startup = self.get_trainer()
         self.assertEqual(len(trainer.blocks), 1)
         ops = [
             'split_ids', 'prefetch', 'merge_ids', 'sequence_pool',
-            'sequence_pool', 'concat', 'mul', 'elementwise_add',
-            'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
-            'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
-            'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
-            'sequence_pool_grad', 'lookup_table_grad', 'sum', 'split_ids',
-            'send', 'send_barrier', 'recv', 'recv', 'fetch_barrier'
+            'sequence_pool', 'lookup_table', 'sequence_pool', 'concat', 'mul',
+            'elementwise_add', 'cross_entropy', 'mean', 'fill_constant',
+            'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send',
+            'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad',
+            'lookup_table_grad', 'split_selected_rows', 'send',
+            'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad',
+            'lookup_table_grad', 'sum', 'split_ids', 'send', 'send_barrier',
+            'recv', 'recv', 'recv', 'fetch_barrier', 'concat'
         ]
         self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
-
         startup_ops = [
             'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
             'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
             'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
-            'fill_constant', 'fill_constant', 'uniform_random', 'recv', 'recv',
-            'fetch_barrier', 'fake_init'
+            'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
+            'fill_constant', 'fill_constant', 'uniform_random',
+            'uniform_random', 'recv', 'recv', 'recv', 'fetch_barrier', 'concat',
+            'fake_init'
         ]
         self.assertEqual([op.type for op in trainer_startup.blocks[0].ops],
                          startup_ops)
@@ -526,7 +587,7 @@ class TestAsyncLocalLookupTable(TestDistLookupTableBase):
         config = fluid.DistributeTranspilerConfig()
         pserver1, startup1 = self.get_pserver(self.pserver1_ep, config, False)
 
-        self.assertEqual(len(pserver1.blocks), 3)
+        self.assertEqual(len(pserver1.blocks), 4)
         # 0 listen_and_serv
         # 1 optimize for fc_w or fc_b adam
         self.assertEqual([op.type for op in pserver1.blocks[1].ops],
@@ -535,17 +596,23 @@ class TestAsyncLocalLookupTable(TestDistLookupTableBase):
         # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num
         self.assertEqual([op.type for op in pserver1.blocks[2].ops],
                          ["adam", "scale", "scale"])
+        # 3 optimize for table adam
+        # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num
+        self.assertEqual([op.type for op in pserver1.blocks[3].ops],
+                         ["adam", "scale", "scale"])
 
         trainer, _ = self.get_trainer(config)
         self.assertEqual(len(trainer.blocks), 1)
         ops = [
             'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool',
-            'concat', 'mul', 'elementwise_add', 'cross_entropy', 'mean',
-            'fill_constant', 'mean_grad', 'cross_entropy_grad',
-            'elementwise_add_grad', 'send', 'mul_grad', 'send', 'concat_grad',
-            'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad',
-            'lookup_table_grad', 'sum', 'split_selected_rows', 'send', 'recv',
-            'recv', 'recv', 'concat'
+            'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add',
+            'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
+            'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
+            'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
+            'split_selected_rows', 'send', 'sequence_pool_grad',
+            'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad',
+            'sum', 'split_selected_rows', 'send', 'recv', 'recv', 'recv',
+            'recv', 'concat', 'concat'
         ]
         self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
 
@@ -559,29 +626,34 @@ class TestAsyncDistLookupTable(TestDistLookupTableBase):
 
         pserver1, startup1 = self.get_pserver(self.pserver1_ep, config, False)
 
-        self.assertEqual(len(pserver1.blocks), 5)
+        self.assertEqual(len(pserver1.blocks), 6)
         # 0 listen_and_serv
         # 1 optimize for fc_w or fc_b adam
         self.assertEqual([op.type for op in pserver1.blocks[1].ops],
                          ["adam", "scale", "scale"])
-        # 2 optimize for table sgd
-        self.assertEqual([op.type for op in pserver1.blocks[2].ops], ["sgd"])
-        # 3 prefetch -> lookup_sparse_table for data0
-        self.assertEqual([op.type for op in pserver1.blocks[3].ops],
+        # 2 optimize for table adam
+        self.assertEqual([op.type for op in pserver1.blocks[2].ops],
+                         ["adam", "scale", "scale"])
+        # 3 optimize for table sgd
+        self.assertEqual([op.type for op in pserver1.blocks[3].ops], ["sgd"])
+        # 4 prefetch -> lookup_sparse_table for data0
+        self.assertEqual([op.type for op in pserver1.blocks[4].ops],
                          ["lookup_sparse_table"])
-        # 4 save table
-        self.assertEqual([op.type for op in pserver1.blocks[4].ops], ["save"])
+        # 5 save table
+        self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"])
 
         trainer, _ = self.get_trainer(config)
         self.assertEqual(len(trainer.blocks), 1)
         ops = [
             'split_ids', 'prefetch', 'merge_ids', 'sequence_pool',
-            'sequence_pool', 'concat', 'mul', 'elementwise_add',
-            'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
-            'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
-            'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
-            'sequence_pool_grad', 'lookup_table_grad', 'sum', 'split_ids',
-            'send', 'recv', 'recv'
+            'sequence_pool', 'lookup_table', 'sequence_pool', 'concat', 'mul',
+            'elementwise_add', 'cross_entropy', 'mean', 'fill_constant',
+            'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send',
+            'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad',
+            'lookup_table_grad', 'split_selected_rows', 'send',
+            'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad',
+            'lookup_table_grad', 'sum', 'split_ids', 'send', 'recv', 'recv',
+            'recv', 'concat'
         ]
         self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
 

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

@@ -876,6 +876,22 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(out)
         print(str(program))
 
+    def test_affine_grid(self):
+        program = Program()
+        with program_guard(program):
+            data = layers.data(name='data', shape=[2, 3, 3], dtype="float32")
+            out, ids = layers.argsort(input=data, axis=1)
+
+            theta = layers.data(name="theta", shape=[2, 3], dtype="float32")
+            out_shape = layers.data(
+                name="out_shape", shape=[-1], dtype="float32")
+            data_0 = layers.affine_grid(theta, out_shape)
+            data_1 = layers.affine_grid(theta, [5, 3, 28, 28])
+
+            self.assertIsNotNone(data_0)
+            self.assertIsNotNone(data_1)
+        print(str(program))
+
 
 if __name__ == '__main__':
     unittest.main()

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

@@ -55,6 +55,46 @@ def run_pserver(use_cuda, sync_mode, ip, port, trainers, trainer_id):
     exe.run(pserver_prog)
 
 
+def run_pserver_with_empty_block(use_cuda, sync_mode, ip, port, trainers,
+                                 trainer_id):
+    x = fluid.layers.data(name='x', shape=[1], dtype='float32')
+    y_predict = fluid.layers.fc(input=x, size=1, act=None, bias_attr=False)
+    y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+
+    # loss function
+    cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+    avg_cost = fluid.layers.mean(cost)
+
+    # optimizer
+    sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
+    sgd_optimizer.minimize(avg_cost)
+
+    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
+    exe = fluid.Executor(place)
+
+    ps1 = ip + ":" + str(int(port) + 1)
+    ps2 = ip + ":" + port
+    pserver_endpoints = ps1 + "," + ps2
+
+    config = fluid.DistributeTranspilerConfig()
+    config.slice_var_up = False
+    t = fluid.DistributeTranspiler(config=config)
+    t.transpile(
+        trainer_id,
+        pservers=pserver_endpoints,
+        trainers=trainers,
+        sync_mode=sync_mode)
+    pserver_prog = t.get_pserver_program(ps2)
+
+    # pserver2 have no parameter
+    assert (len(pserver_prog.blocks) == 2)
+    assert (len(pserver_prog.blocks[1].ops) == 0)
+
+    pserver_startup = t.get_startup_program(ps2, pserver_prog)
+    exe.run(pserver_startup)
+    exe.run(pserver_prog)
+
+
 class TestListenAndServOp(OpTest):
     def setUp(self):
         self.ps_timeout = 5
@@ -63,9 +103,9 @@ class TestListenAndServOp(OpTest):
         self.trainers = 1
         self.trainer_id = 0
 
-    def _start_pserver(self, use_cuda, sync_mode):
+    def _start_pserver(self, use_cuda, sync_mode, pserver_func):
         p = Process(
-            target=run_pserver,
+            target=pserver_func,
             args=(use_cuda, sync_mode, self.ip, self.port, self.trainers,
                   self.trainer_id))
         p.daemon = True
@@ -92,7 +132,24 @@ class TestListenAndServOp(OpTest):
 
     def test_handle_signal_in_serv_op(self):
         # run pserver on CPU in sync mode
-        p1 = self._start_pserver(False, True)
+        p1 = self._start_pserver(False, True, run_pserver)
+        self._wait_ps_ready(p1.pid)
+
+        # raise SIGTERM to pserver
+        os.kill(p1.pid, signal.SIGINT)
+        p1.join()
+
+        # run pserver on CPU in async mode
+        p2 = self._start_pserver(False, False, run_pserver)
+        self._wait_ps_ready(p2.pid)
+
+        # raise SIGTERM to pserver
+        os.kill(p2.pid, signal.SIGTERM)
+        p2.join()
+
+    def test_list_and_serv_run_empty_optimize_block(self):
+        # run pserver on CPU in sync mode
+        p1 = self._start_pserver(False, True, run_pserver_with_empty_block)
         self._wait_ps_ready(p1.pid)
 
         # raise SIGTERM to pserver
@@ -100,7 +157,7 @@ class TestListenAndServOp(OpTest):
         p1.join()
 
         # run pserver on CPU in async mode
-        p2 = self._start_pserver(False, False)
+        p2 = self._start_pserver(False, False, run_pserver_with_empty_block)
         self._wait_ps_ready(p2.pid)
 
         # raise SIGTERM to pserver

python/paddle/fluid/tests/unittests/test_py_reader_pin_memory.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 paddle
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+import numpy as np
+from threading import Thread
+
+
+def user_reader(inputs):
+    def _reader():
+        for d in inputs:
+            yield d
+
+    return _reader
+
+
+def batch_feeder(batch_reader, pin_memory=False, img_dtype="float32"):
+    def _feeder():
+        for batch_data in batch_reader():
+            sample_batch = []
+            label_batch = []
+            for sample, label in batch_data:
+                sample_batch.append(sample)
+                label_batch.append([label])
+            tensor = core.LoDTensor()
+            label = core.LoDTensor()
+            place = core.CUDAPinnedPlace() if pin_memory else core.CPUPlace()
+            tensor.set(np.array(sample_batch, dtype=img_dtype), place)
+            label.set(np.array(label_batch, dtype="int64"), place)
+            yield [tensor, label]
+
+    return _feeder
+
+
+class TestPyReader(unittest.TestCase):
+    def setUp(self):
+        self.capacity = 10
+        self.shapes = [(-1, 3, 2, 1), (-1, 1)]
+        self.lod_levels = [0, 0]
+        self.dtypes = ['float32', 'int64']
+
+    def test_pin_memory_pyreader(self):
+        with fluid.program_guard(fluid.Program(), fluid.Program()):
+            place = fluid.CUDAPlace(0) if fluid.core.is_compiled_with_cuda(
+            ) else fluid.CPUPlace()
+            executor = fluid.Executor(place)
+
+            data_file = fluid.layers.py_reader(
+                capacity=self.capacity,
+                dtypes=self.dtypes,
+                lod_levels=self.lod_levels,
+                shapes=self.shapes)
+            # feed_queue = data_file.queue
+            read_out_data = fluid.layers.read_file(data_file)
+
+            self.inputs = []
+            for _ in range(10):
+                sample = np.random.uniform(
+                    low=0, high=1, size=[3, 2, 1]).astype("float32")
+                label = np.random.uniform(
+                    low=0, high=10, size=[1]).astype("int64")
+                self.inputs.append((sample, label))
+
+            self.input_tensors = []
+            for d, l in batch_feeder(
+                    paddle.batch(
+                        user_reader(self.inputs), batch_size=2),
+                    pin_memory=True
+                    if fluid.core.is_compiled_with_cuda() else False)():
+                ta = fluid.LoDTensorArray()
+                ta.append(d)
+                ta.append(l)
+                self.input_tensors.append(ta)
+
+            self.batched_inputs = []
+            for batch in paddle.batch(user_reader(self.inputs), batch_size=2)():
+                feed_d = []
+                feed_l = []
+                for d, l in batch:
+                    feed_d.append(d)
+                    feed_l.append([l])
+                self.batched_inputs.append([feed_d, feed_l])
+
+            data_file.decorate_tensor_provider(
+                batch_feeder(
+                    paddle.batch(
+                        user_reader(self.inputs), batch_size=2),
+                    pin_memory=True
+                    if fluid.core.is_compiled_with_cuda() else False))
+
+            executor.run(fluid.default_startup_program())
+            self.outputs = []
+
+            data_file.start()
+            for _ in self.input_tensors:
+                self.outputs.append(
+                    executor.run(fetch_list=list(read_out_data)))
+            data_file.reset()
+            self.validate()
+
+    def validate(self):
+        self.assertEqual(len(self.batched_inputs), len(self.outputs))
+        for in_data_list, out_data_list in zip(self.batched_inputs,
+                                               self.outputs):
+            self.assertEqual(len(in_data_list), len(out_data_list))
+            in_data_list_np = [
+                np.array(in_lod_tensor) for in_lod_tensor in in_data_list
+            ]
+            for in_data, out_data in zip(in_data_list_np, out_data_list):
+                self.assertTrue((in_data == out_data).all())
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -35,6 +35,7 @@ import sys
 import numpy as np
 import collections
 import six
+import logging
 
 from .ps_dispatcher import RoundRobin, HashName, PSDispatcher
 from .. import core, framework
@@ -767,6 +768,15 @@ in a single call.")
             prefetch_var_name_to_block_id.extend(
                 lookup_table_var_name_to_block_id)
 
+        if len(optimize_blocks) == 0:
+            logging.warn("pserver [" + str(endpoint) +
+                         "] has no optimize block!!")
+            pre_block_idx = pserver_program.num_blocks - 1
+            empty_block = pserver_program._create_block(pre_block_idx)
+            optimize_blocks.append(empty_block)
+
+        # In some case, some parameter server will have no parameter to optimize
+        # So we give an empty optimize block to parameter server.
         attrs = {
             "optimize_blocks": optimize_blocks,
             "endpoint": endpoint,
@@ -910,11 +920,11 @@ to transpile() call.")
             block_idx = int(block_name.split(block_suffix)[1])
             orig_var = self.origin_program.global_block().vars[orig_var_name]
 
-            skip_numel = 0
+            skip_dim0 = 0
             slice_vars = self.param_var_mapping[orig_var_name]
             for slice_var in slice_vars[:block_idx]:
-                skip_numel += reduce(lambda x, y: x * y, slice_var.shape)
-            slice_vars_and_attrs.append([orig_var, skip_numel, param])
+                skip_dim0 += slice_var.shape[0]
+            slice_vars_and_attrs.append([orig_var, skip_dim0, param])
 
         return slice_vars_and_attrs
 
@@ -1065,7 +1075,12 @@ to transpile() call.")
             continue_search_lookup_table_op = False
             all_ops = program.global_block().ops
             for op in all_ops:
-                if op.type == LOOKUP_TABLE_TYPE:
+                if op.type == LOOKUP_TABLE_TYPE and self.table_name == op.input(
+                        "W")[0]:
+                    if not op.attr('is_distributed'):
+                        raise RuntimeError(
+                            "lookup_table_op that lookup an distributed embedding table"
+                            "should set is_distributed to true")
                     continue_search_lookup_table_op = True
 
                     lookup_table_op_index = lookup_table_op_index if lookup_table_op_index != -1 else list(
@@ -1275,7 +1290,6 @@ to transpile() call.")
         }
         outputs = {"ParamOut": [param_var]}
         # only support sgd now
-        import logging
         logging.warn(
             "distribute lookup table only support sgd optimizer, change it's optimizer to sgd instead of "
             + table_opt_op.type)
@@ -1442,6 +1456,9 @@ to transpile() call.")
         elif op_type == "decayed_adagrad":
             if varkey == "Moment":
                 return param_shape
+        elif op_type == "ftrl":
+            if varkey in ["SquaredAccumulator", "LinearAccumulator"]:
+                return param_shape
         elif op_type == "sgd":
             pass
         else:

python/paddle/fluid/transpiler/inference_transpiler.py

@@ -61,6 +61,9 @@ class InferenceTranspiler(object):
             raise TypeError("scope should be as Scope type or None")
         use_mkldnn = bool(os.getenv("FLAGS_use_mkldnn", False))
 
+        if use_mkldnn:
+            self._depthwise_conv_mkldnn(program)
+
         self._fuse_batch_norm(program, place, scope)
         if use_mkldnn:
             self._fuse_conv_bias_mkldnn(program)
@@ -70,6 +73,31 @@ class InferenceTranspiler(object):
                 program)  # ResNet residual block merging
             self._fuse_bn_relu_mkldnn(program)
 
+    def _depthwise_conv_mkldnn(self, program):
+        '''
+        Transpile the program by replacing depthwise_conv2d to conv2d for MKLDNN program.
+        The result is:
+            - before:
+                - any_other_op->depthwise_conv->any_other_op
+            - after:
+                - any_other_op->conv->any_other_op
+        :param program: program to transpile
+        :type program: Program
+        '''
+        self.block = program.block(0)
+
+        i = 0
+        while i < len(self.block.ops):
+            current_op = self.block.ops[i]
+            if current_op.type == 'depthwise_conv2d':
+                current_op.desc.set_type("conv2d")
+            i = i + 1
+
+        # TODO(luotao): use clone() method to flush the program.desc in force,
+        # since some large program.desc will not be flushed immediately.
+        # And a better solution will be considered later.
+        program = program.clone()
+
     def _fuse_conv_eltwise_mkldnn(self, program):
         '''
         Transpile the program fusing elementwise_add into conv for MKLDNN