diff --git a/paddle/contrib/inference/CMakeLists.txt b/paddle/contrib/inference/CMakeLists.txt
index 98c2f68a6c39ed12795bad4a905558917c0275a4..87173fc42a46c8218fbf0beb4ebf7760f69b7c24 100644
--- a/paddle/contrib/inference/CMakeLists.txt
+++ b/paddle/contrib/inference/CMakeLists.txt
@@ -45,6 +45,10 @@ endfunction(inference_api_test)
 cc_library(paddle_inference_api
     SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
     DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
+if(NOT APPLE)
+  set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/paddle_inference_api.sym")
+  set_target_properties(paddle_inference_api PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
+endif()
 
 # Here the shared library doesn't depend on other fluid libraries, or double free will occur.
 cc_library(paddle_inference_api_shared SHARED
@@ -53,8 +57,19 @@ add_dependencies(paddle_inference_api_shared ${FLUID_CORE_MODULES} ${GLOB_OP_LIB
 set_target_properties(paddle_inference_api_shared PROPERTIES OUTPUT_NAME paddle_inference_api)
 
 if(NOT APPLE)
-  set(LINK_FLAGS "-fPIC -fvisibility=hidden")
+  set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/paddle_inference_api.map")
   set_target_properties(paddle_inference_api_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
+  FILE(WRITE ${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake
+    "execute_process(COMMAND bash -c \"${CMAKE_CURRENT_SOURCE_DIR}/check_symbol.sh"
+    " ${CMAKE_CURRENT_BINARY_DIR}/libpaddle_inference_api.so\" RESULT_VARIABLE symbol_res)\n"
+    "if(NOT \"\${symbol_res}\" STREQUAL \"0\")\n"
+    "  message(FATAL_ERROR \"Check symbol failed.\")\n"
+    "endif()\n")
+  add_custom_command(
+    OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol"
+    COMMAND ${CMAKE_COMMAND} -P "${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake"
+    DEPENDS paddle_inference_api_shared)
+  add_custom_target(check_symbol ALL DEPENDS "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol")
 endif()
 
 cc_test(test_paddle_inference_api
diff --git a/paddle/contrib/inference/check_symbol.sh b/paddle/contrib/inference/check_symbol.sh
new file mode 100755
index 0000000000000000000000000000000000000000..6547ca1413649968e8a0be146915e07192a99898
--- /dev/null
+++ b/paddle/contrib/inference/check_symbol.sh
@@ -0,0 +1,12 @@
+#!/bin/bash
+
+lib=$1
+if [ $# -ne 1 ]; then echo "No input library"; exit -1 ; fi
+
+num_paddle_syms=$(nm -D --defined-only ${lib} | grep paddle | wc -l)
+num_google_syms=$(nm -D --defined-only ${lib} | grep google | wc -l)
+
+if [ $num_paddle_syms -le 0 ]; then echo "Have no paddle symbols"; exit -1 ; fi
+if [ $num_google_syms -ge 1 ]; then echo "Have some google symbols"; exit -1 ; fi
+
+exit 0
diff --git a/paddle/contrib/inference/demo/CMakeLists.txt b/paddle/contrib/inference/demo/CMakeLists.txt
index ecece6fe3471ad7b89c84c3e2b67af4ae9eb3c36..2d501bf0085b1bd4c39ee1a6dfaaa9622fd72ce1 100644
--- a/paddle/contrib/inference/demo/CMakeLists.txt
+++ b/paddle/contrib/inference/demo/CMakeLists.txt
@@ -13,8 +13,6 @@
 # limitations under the License.
 #
 
-inference_api_test(simple_on_word2vec ARGS test_word2vec)
-
 option(WITH_INFERENCE_DEMO "Compile with Inference demo" OFF)
 if(NOT WITH_INFERENCE_DEMO)
   return()
diff --git a/paddle/contrib/inference/demo_ci/CMakeLists.txt b/paddle/contrib/inference/demo_ci/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..789bff7f23cd89bfaeba180efa95972cef6fc58c
--- /dev/null
+++ b/paddle/contrib/inference/demo_ci/CMakeLists.txt
@@ -0,0 +1,77 @@
+cmake_minimum_required(VERSION 3.0)
+
+project(cpp_inference_demo CXX C)
+
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
+
+if(NOT DEFINED PADDLE_LIB)
+  message(FATAL_ERROR "please set PADDLE_LIB with -DPADDLE_LIB=/path/paddle/lib")
+endif()
+if(NOT DEFINED DEMO_NAME)
+  message(FATAL_ERROR "please set DEMO_NAME with -DDEMO_NAME=demo_name")
+endif()
+
+option(WITH_MKL        "Compile demo with MKL/OpenBlas support, default use MKL."       ON)
+option(WITH_GPU        "Compile demo with GPU/CPU, default use CPU."                    OFF)
+option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static."   ON)
+
+if(WITH_GPU)
+  set(CUDA_LIB "/usr/local/cuda/lib64/" CACHE STRING "CUDA Library")
+endif()
+
+include_directories("${PADDLE_LIB}")
+include_directories("${PADDLE_LIB}/third_party/install/protobuf/include")
+include_directories("${PADDLE_LIB}/third_party/install/glog/include")
+include_directories("${PADDLE_LIB}/third_party/install/gflags/include")
+include_directories("${PADDLE_LIB}/third_party/install/snappy/include")
+include_directories("${PADDLE_LIB}/third_party/install/snappystream/include")
+include_directories("${PADDLE_LIB}/third_party/install/zlib/include")
+
+include_directories("${PADDLE_LIB}/third_party/boost")
+include_directories("${PADDLE_LIB}/third_party/eigen3")
+
+link_directories("${PADDLE_LIB}/third_party/install/snappy/lib")
+link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib")
+link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib")
+link_directories("${PADDLE_LIB}/third_party/install/glog/lib")
+link_directories("${PADDLE_LIB}/third_party/install/gflags/lib")
+link_directories("${PADDLE_LIB}/third_party/install/zlib/lib")
+
+add_executable(${DEMO_NAME} ${DEMO_NAME}.cc)
+
+if(WITH_MKL)
+  include_directories("${PADDLE_LIB}/third_party/install/mklml/include")
+  set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel.so 
+               ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5.so)
+  set(MKLDNN_PATH "${PADDLE_LIB}/third_party/install/mkldnn")
+  if(EXISTS ${MKLDNN_PATH})
+    include_directories("${MKLDNN_PATH}/include")
+    set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0)
+  endif()
+else()
+  set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas.a)
+endif()
+
+if(WITH_STATIC_LIB)
+  set(DEPS
+      "-Wl,--whole-archive"
+      ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid.a
+      "-Wl,--no-whole-archive"
+      ${PADDLE_LIB}/contrib/inference/libpaddle_inference_api.a)
+else()
+  # Note: libpaddle_inference_api.so must put before libpaddle_fluid.so
+  set(DEPS
+      ${PADDLE_LIB}/contrib/inference/libpaddle_inference_api.so
+      ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid.so)
+endif()
+set(EXTERNAL_LIB "-lrt -ldl -lpthread")
+
+set(DEPS ${DEPS}
+    ${MATH_LIB} ${MKLDNN_LIB}
+    glog gflags protobuf snappystream snappy z
+    ${EXTERNAL_LIB})
+if(WITH_GPU)
+  set(DEPS ${DEPS} ${CUDA_LIB}/libcudart.so)
+endif()
+
+target_link_libraries(${DEMO_NAME} ${DEPS})
diff --git a/paddle/contrib/inference/demo_ci/run.sh b/paddle/contrib/inference/demo_ci/run.sh
new file mode 100755
index 0000000000000000000000000000000000000000..e3a7269af795b05c296423cb2dc92b753397c6b3
--- /dev/null
+++ b/paddle/contrib/inference/demo_ci/run.sh
@@ -0,0 +1,34 @@
+set -x
+PADDLE_ROOT=$1
+WITH_MKL=$2
+WITH_GPU=$3
+if [ $3 == "ON" ]; then
+  use_gpu_list='true false'
+else    
+  use_gpu_list='false'
+fi
+
+mkdir -p build
+cd build
+
+for WITH_STATIC_LIB in false; do
+  rm -rf *
+  cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \
+    -DWITH_MKL=$WITH_MKL \
+    -DDEMO_NAME=simple_on_word2vec \
+    -DWITH_GPU=$WITH_GPU \
+    -DWITH_STATIC_LIB=$WITH_STATIC_LIB
+  make
+  for use_gpu in $use_gpu_list; do
+    ./simple_on_word2vec \
+      --dirname=${PADDLE_ROOT}/build/python/paddle/fluid/tests/book/word2vec.inference.model \
+      --use_gpu=$use_gpu
+  done
+done
+if [ $? -eq 0 ]; then
+  exit 0
+else
+  echo "inference demo runs fail."
+  exit 1
+fi
+set +x
diff --git a/paddle/contrib/inference/demo/simple_on_word2vec.cc b/paddle/contrib/inference/demo_ci/simple_on_word2vec.cc
similarity index 68%
rename from paddle/contrib/inference/demo/simple_on_word2vec.cc
rename to paddle/contrib/inference/demo_ci/simple_on_word2vec.cc
index c253014642f39a042430992548a285cc7078a959..9713837f86d40383da946af1681e1945c84336b0 100644
--- a/paddle/contrib/inference/demo/simple_on_word2vec.cc
+++ b/paddle/contrib/inference/demo_ci/simple_on_word2vec.cc
@@ -16,21 +16,27 @@ limitations under the License. */
  * This file contains a simple demo for how to take a model for inference.
  */
 
+#include <gflags/gflags.h>
 #include <glog/logging.h>
-#include <gtest/gtest.h>
 #include <memory>
 #include <thread>
-#include "paddle/contrib/inference/paddle_inference_api.h"
+#include "contrib/inference/paddle_inference_api.h"
+#include "paddle/fluid/platform/enforce.h"
+
+DEFINE_string(dirname, "", "Directory of the inference model.");
+DEFINE_bool(use_gpu, false, "Whether use gpu.");
 
 namespace paddle {
 namespace demo {
 
-DEFINE_string(dirname, "", "Directory of the inference model.");
-
 void Main(bool use_gpu) {
   //# 1. Create PaddlePredictor with a config.
   NativeConfig config;
-  config.model_dir = FLAGS_dirname + "word2vec.inference.model";
+  if (FLAGS_dirname.empty()) {
+    LOG(INFO) << "Usage: ./simple_on_word2vec --dirname=path/to/your/model";
+    exit(1);
+  }
+  config.model_dir = FLAGS_dirname;
   config.use_gpu = use_gpu;
   config.fraction_of_gpu_memory = 0.15;
   config.device = 0;
@@ -54,12 +60,16 @@ void Main(bool use_gpu) {
     CHECK(predictor->Run(slots, &outputs));
 
     //# 4. Get output.
-    ASSERT_EQ(outputs.size(), 1UL);
-    LOG(INFO) << "output buffer size: " << outputs.front().data.length();
+    PADDLE_ENFORCE(outputs.size(), 1UL);
+    // Check the output buffer size and result of each tid.
+    PADDLE_ENFORCE(outputs.front().data.length(), 33168UL);
+    float result[5] = {
+        0.00129761, 0.00151112, 0.000423564, 0.00108815, 0.000932706};
     const size_t num_elements = outputs.front().data.length() / sizeof(float);
     // The outputs' buffers are in CPU memory.
     for (size_t i = 0; i < std::min(5UL, num_elements); i++) {
-      LOG(INFO) << static_cast<float*>(outputs.front().data.data())[i];
+      PADDLE_ENFORCE(static_cast<float*>(outputs.front().data.data())[i],
+                     result[i]);
     }
   }
 }
@@ -68,7 +78,7 @@ void MainThreads(int num_threads, bool use_gpu) {
   // Multi-threads only support on CPU
   // 0. Create PaddlePredictor with a config.
   NativeConfig config;
-  config.model_dir = FLAGS_dirname + "word2vec.inference.model";
+  config.model_dir = FLAGS_dirname;
   config.use_gpu = use_gpu;
   config.fraction_of_gpu_memory = 0.15;
   config.device = 0;
@@ -94,14 +104,17 @@ void MainThreads(int num_threads, bool use_gpu) {
         CHECK(predictor->Run(inputs, &outputs));
 
         // 4. Get output.
-        ASSERT_EQ(outputs.size(), 1UL);
-        LOG(INFO) << "TID: " << tid << ", "
-                  << "output buffer size: " << outputs.front().data.length();
+        PADDLE_ENFORCE(outputs.size(), 1UL);
+        // Check the output buffer size and result of each tid.
+        PADDLE_ENFORCE(outputs.front().data.length(), 33168UL);
+        float result[5] = {
+            0.00129761, 0.00151112, 0.000423564, 0.00108815, 0.000932706};
         const size_t num_elements =
             outputs.front().data.length() / sizeof(float);
         // The outputs' buffers are in CPU memory.
         for (size_t i = 0; i < std::min(5UL, num_elements); i++) {
-          LOG(INFO) << static_cast<float*>(outputs.front().data.data())[i];
+          PADDLE_ENFORCE(static_cast<float*>(outputs.front().data.data())[i],
+                         result[i]);
         }
       }
     });
@@ -111,15 +124,18 @@ void MainThreads(int num_threads, bool use_gpu) {
   }
 }
 
-TEST(demo, word2vec_cpu) { Main(false /*use_gpu*/); }
-TEST(demo_multi_threads, word2vec_cpu_1) { MainThreads(1, false /*use_gpu*/); }
-TEST(demo_multi_threads, word2vec_cpu_4) { MainThreads(4, false /*use_gpu*/); }
-
-#ifdef PADDLE_WITH_CUDA
-TEST(demo, word2vec_gpu) { Main(true /*use_gpu*/); }
-TEST(demo_multi_threads, word2vec_gpu_1) { MainThreads(1, true /*use_gpu*/); }
-TEST(demo_multi_threads, word2vec_gpu_4) { MainThreads(4, true /*use_gpu*/); }
-#endif
-
 }  // namespace demo
 }  // namespace paddle
+
+int main(int argc, char** argv) {
+  google::ParseCommandLineFlags(&argc, &argv, true);
+  paddle::demo::Main(false /* use_gpu*/);
+  paddle::demo::MainThreads(1, false /* use_gpu*/);
+  paddle::demo::MainThreads(4, false /* use_gpu*/);
+  if (FLAGS_use_gpu) {
+    paddle::demo::Main(true /*use_gpu*/);
+    paddle::demo::MainThreads(1, true /*use_gpu*/);
+    paddle::demo::MainThreads(4, true /*use_gpu*/);
+  }
+  return 0;
+}
diff --git a/paddle/contrib/inference/paddle_inference_api.map b/paddle/contrib/inference/paddle_inference_api.map
new file mode 100644
index 0000000000000000000000000000000000000000..5203784dc1fcb672eb6a26d9dfd3ffbe02e08038
--- /dev/null
+++ b/paddle/contrib/inference/paddle_inference_api.map
@@ -0,0 +1,6 @@
+{
+	global:
+		*paddle*;
+	local:
+		*;
+};
diff --git a/paddle/contrib/inference/paddle_inference_api.sym b/paddle/contrib/inference/paddle_inference_api.sym
new file mode 100644
index 0000000000000000000000000000000000000000..ef2a04d788aa86b7f6a61c4af479d70d1137f374
--- /dev/null
+++ b/paddle/contrib/inference/paddle_inference_api.sym
@@ -0,0 +1 @@
+*paddle*
diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt
index 1895aea7f98cb1ad12b2ce16545339252349ea37..b1c33c3415f49f9b1160655034350087432d0cb0 100644
--- a/paddle/fluid/inference/CMakeLists.txt
+++ b/paddle/fluid/inference/CMakeLists.txt
@@ -13,6 +13,12 @@ endif()
 
 # Create static library
 cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api)
+if(NOT APPLE)
+  # TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac.
+  set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.sym")
+  set_target_properties(paddle_fluid PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
+endif()
+
 # Create shared library
 cc_library(paddle_fluid_shared SHARED
     SRCS io.cc
diff --git a/paddle/fluid/inference/analysis/data_flow_graph.cc b/paddle/fluid/inference/analysis/data_flow_graph.cc
index d09bf3ed161703b0cf273522921e157c7360a0bc..bd24e8a7d9c20b8cd9c4e41a76ffc33a004a9a69 100644
--- a/paddle/fluid/inference/analysis/data_flow_graph.cc
+++ b/paddle/fluid/inference/analysis/data_flow_graph.cc
@@ -90,6 +90,20 @@ std::string DataFlowGraph::DotString() const {
   return dot.Build();
 }
 
+std::string DataFlowGraph::HumanReadableInfo(bool show_values,
+                                             bool show_functions) const {
+  std::stringstream values, functions;
+  for (auto &n : nodes.nodes()) {
+    if (show_values && n->IsValue()) {
+      values << n->repr() << "\n";
+    }
+    if (show_functions && n->IsFunction()) {
+      functions << n->repr() << "\n";
+    }
+  }
+  return "Values:\n" + values.str() + "\n\n" + "Functions:\n" + functions.str();
+}
+
 //
 // NodesBFSIterator
 //
@@ -146,7 +160,7 @@ bool GraphTraits<DataFlowGraph>::NodesBFSIterator::operator==(
   if ((!queue_.empty()) && (!other.queue_.empty())) {
     return queue_.front() == other.queue_.front() &&
            visited_.size() == other.visited_.size();  // here need to check the
-                                                      // equality of queue and
+    // equality of queue and
     // visited. Just a light but week implementation.
   }
   return false;
@@ -208,6 +222,76 @@ Node *GraphTraits<DataFlowGraph>::NodesDFSIterator::operator->() {
   return stack_.top();
 }
 
+GraphTraits<DataFlowGraph>::NodesTSIterator::NodesTSIterator(
+    const std::vector<Node *> &source) {
+  PADDLE_ENFORCE(!source.empty(),
+                 "Start points of topological sorting should not be empty!");
+  std::unordered_set<Node *> visited;
+  std::unordered_set<Node *> to_visit{source.begin(), source.end()};
+
+  std::vector<Node *> inlink_visited;
+  while (!to_visit.empty()) {
+    std::vector<Node *> queue(to_visit.begin(), to_visit.end());
+    for (auto *p : queue) {
+      inlink_visited.clear();
+
+      std::copy_if(p->inlinks.begin(), p->inlinks.end(),
+                   std::back_inserter(inlink_visited),
+                   [&](Node *x) { return visited.count(x); });
+
+      if (inlink_visited.size() == p->inlinks.size()) {
+        sorted_.push_back(p);
+        for (auto *_ : p->outlinks) {
+          if (!visited.count(_)) {
+            to_visit.insert(_);
+          }
+        }
+
+        to_visit.erase(p);
+        visited.insert(p);
+      }
+    }
+  }
+}
+
+GraphTraits<DataFlowGraph>::NodesTSIterator::NodesTSIterator(
+    const paddle::inference::analysis::GraphTraits<
+        DataFlowGraph>::NodesTSIterator &other)
+    : sorted_(other.sorted_), cursor_(other.cursor_) {}
+
+Node &GraphTraits<DataFlowGraph>::NodesTSIterator::operator*() {
+  PADDLE_ENFORCE_LT(cursor_, sorted_.size());
+  return *sorted_[cursor_];
+}
+
+paddle::inference::analysis::GraphTraits<DataFlowGraph>::NodesTSIterator
+    &GraphTraits<DataFlowGraph>::NodesTSIterator::operator++() {
+  if (++cursor_ >= sorted_.size()) {
+    sorted_.clear();
+    cursor_ = 0;
+  }
+  return *this;
+}
+paddle::inference::analysis::GraphTraits<DataFlowGraph>::NodesTSIterator &
+GraphTraits<DataFlowGraph>::NodesTSIterator::operator=(
+    const paddle::inference::analysis::GraphTraits<
+        DataFlowGraph>::NodesTSIterator &other) {
+  cursor_ = other.cursor_;
+  sorted_ = other.sorted_;
+  return *this;
+}
+
+bool GraphTraits<DataFlowGraph>::NodesTSIterator::operator==(
+    const paddle::inference::analysis::GraphTraits<
+        DataFlowGraph>::NodesTSIterator &other) {
+  return sorted_ == other.sorted_ && cursor_ == other.cursor_;
+}
+
+Node *GraphTraits<DataFlowGraph>::NodesTSIterator::operator->() {
+  PADDLE_ENFORCE_LT(cursor_, sorted_.size());
+  return sorted_[cursor_];
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle
diff --git a/paddle/fluid/inference/analysis/data_flow_graph.h b/paddle/fluid/inference/analysis/data_flow_graph.h
index a4fefc83e0c551d52bec87299bcbc966e7a2dbf7..5dd914d1971bfb5bcc0b1db41d73e2b67120bc06 100644
--- a/paddle/fluid/inference/analysis/data_flow_graph.h
+++ b/paddle/fluid/inference/analysis/data_flow_graph.h
@@ -48,6 +48,9 @@ struct DataFlowGraph {
   // Output a DOT graph file for debug.
   std::string DotString() const;
 
+  std::string HumanReadableInfo(bool show_values = true,
+                                bool show_functions = true) const;
+
  private:
   // Remove duplicate edges and so on.
   void Clean();
@@ -107,6 +110,32 @@ struct GraphTraits<DataFlowGraph> {
     std::unordered_set<Node *> visited_;
   };
 
+  // Topological sorting iterator on nodes.
+  struct NodesTSIterator
+      : public std::iterator<std::forward_iterator_tag, Node *> {
+    NodesTSIterator() = default;
+    explicit NodesTSIterator(const std::vector<Node *> &source);
+    NodesTSIterator(NodesTSIterator &&other)
+        : sorted_(std::move(other.sorted_)), cursor_(other.cursor_) {
+      other.cursor_ = 0;
+    }
+    NodesTSIterator(const NodesTSIterator &other);
+
+    Node &operator*();
+    NodesTSIterator &operator++();
+    // TODO(Superjomn) current implementation just compare the first
+    // element, need to compare the graph and all the elements in the queue and
+    // set.
+    NodesTSIterator &operator=(const NodesTSIterator &other);
+    bool operator==(const NodesTSIterator &other);
+    bool operator!=(const NodesTSIterator &other) { return !(*this == other); }
+    Node *operator->();
+
+   private:
+    std::vector<Node *> sorted_;
+    int cursor_{0};
+  };
+
   explicit GraphTraits(DataFlowGraph *graph) : graph_(graph) {}
 
   // default use BFS to visit the nodes.
@@ -119,17 +148,24 @@ struct GraphTraits<DataFlowGraph> {
   iterator_range<NodesDFSIterator> nodes_in_DFS() {
     return iterator_range<NodesDFSIterator>(nodes_dfs_begin(), nodes_dfs_end());
   }
+  iterator_range<NodesTSIterator> nodes_in_TS() {
+    return iterator_range<NodesTSIterator>(nodes_ts_begin(), nodes_ts_end());
+  }
 
  private:
   NodesBFSIterator nodes_bfs_begin() {
     return NodesBFSIterator(graph_->inputs);
   }
   NodesBFSIterator nodes_bfs_end() { return NodesBFSIterator(); }
+
   NodesDFSIterator nodes_dfs_begin() {
     return NodesDFSIterator(graph_->inputs);
   }
   NodesDFSIterator nodes_dfs_end() { return NodesDFSIterator(); }
 
+  NodesTSIterator nodes_ts_begin() { return NodesTSIterator(graph_->inputs); }
+  NodesTSIterator nodes_ts_end() { return NodesTSIterator(); }
+
  private:
   DataFlowGraph *graph_;
 };
diff --git a/paddle/fluid/inference/analysis/data_flow_graph_tester.cc b/paddle/fluid/inference/analysis/data_flow_graph_tester.cc
index 9d7cceeb65888b8ba3fdf39e88fc2877abd82d11..7912f8d7f17ae3c79e8f73f36b7095fd52c9ac86 100644
--- a/paddle/fluid/inference/analysis/data_flow_graph_tester.cc
+++ b/paddle/fluid/inference/analysis/data_flow_graph_tester.cc
@@ -24,11 +24,11 @@ TEST(DataFlowGraph, BFS) {
   auto dfg = ProgramDescToDFG(desc);
   dfg.Build();
 
-  for (auto* in : dfg.inputs) {
+  for (auto *in : dfg.inputs) {
     LOG(INFO) << "inputs: " << in->name() << " "
               << static_cast<int>(in->type());
   }
-  for (auto* out : dfg.outputs) {
+  for (auto *out : dfg.outputs) {
     LOG(INFO) << "outputs: " << out->name() << " "
               << static_cast<int>(out->type());
   }
@@ -57,6 +57,71 @@ TEST(DataFlowGraph, DFS) {
   ASSERT_EQ(count, dfg.nodes.size());
 }
 
+// Topological sorting.
+/*
+ * Graph topology
+ * inputs: 0, 1, 2
+ * 0 -> 4
+ * 0 -> 5
+ * 1 -> 6
+ * 2 -> 7
+ * 4 -> 5
+ * 4 -> 7
+ * 4 -> 3
+ * 7 -> 3
+ */
+TEST(DataFlowGraph, TS) {
+  DataFlowGraph graph;
+
+  for (int i = 0; i < 8; i++) {
+    auto *node = graph.nodes.Create(Node::Type::kValue);
+    node->SetName("node-" + std::to_string(i));
+  }
+
+  auto add_link = [&](int i, int j) {
+    Node *source = graph.nodes.GetMutable(i);
+    Node *target = graph.nodes.GetMutable(j);
+    target->inlinks.push_back(source);
+    source->outlinks.push_back(target);
+  };
+
+  graph.inputs.push_back(graph.nodes.GetMutable(0));
+  graph.inputs.push_back(graph.nodes.GetMutable(1));
+  graph.inputs.push_back(graph.nodes.GetMutable(2));
+
+  add_link(0, 4);
+  add_link(0, 5);
+  add_link(1, 6);
+  add_link(2, 7);
+  add_link(4, 5);
+  add_link(4, 7);
+  add_link(4, 3);
+  add_link(7, 3);
+
+  auto its = GraphTraits<DataFlowGraph>(&graph).nodes_in_TS();
+  std::vector<int> sorted_ids;
+  for (auto it = its.begin(); it != its.end(); ++it) {
+    LOG(INFO) << it->name();
+    sorted_ids.push_back(it->id());
+  }
+
+  // Assert a occurs prior to b in the sorted_ids.
+  auto assert_positive_sequence_pair = [&](int a, int b) {
+    auto a_offset = std::find(sorted_ids.begin(), sorted_ids.end(), a);
+    auto b_offset = std::find(sorted_ids.begin(), sorted_ids.end(), b);
+    ASSERT_LT(a_offset, b_offset);
+  };
+
+  assert_positive_sequence_pair(2, 7);
+  assert_positive_sequence_pair(7, 3);
+  assert_positive_sequence_pair(4, 3);
+  assert_positive_sequence_pair(0, 4);
+  assert_positive_sequence_pair(0, 5);
+  assert_positive_sequence_pair(1, 6);
+  assert_positive_sequence_pair(4, 5);
+  assert_positive_sequence_pair(4, 7);
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle
diff --git a/paddle/fluid/inference/paddle_fluid.sym b/paddle/fluid/inference/paddle_fluid.sym
new file mode 100644
index 0000000000000000000000000000000000000000..ef2a04d788aa86b7f6a61c4af479d70d1137f374
--- /dev/null
+++ b/paddle/fluid/inference/paddle_fluid.sym
@@ -0,0 +1 @@
+*paddle*
diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/conv_mkldnn_op.cc
index 6b06913d1c83f4534238ac3dd22ac4035c0f0fbf..5bfa1aaa696d5cbe8bdcb94d708746259952740f 100644
--- a/paddle/fluid/operators/conv_mkldnn_op.cc
+++ b/paddle/fluid/operators/conv_mkldnn_op.cc
@@ -29,6 +29,79 @@ using mkldnn::stream;
 using platform::to_void_cast;
 using platform::GetMKLDNNFormat;
 
+class ConvMKLDNNHandler : public platform::MKLDNNHandler {
+ public:
+  ConvMKLDNNHandler(
+      std::shared_ptr<mkldnn::convolution_forward::primitive_desc> conv_pd,
+      const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine,
+      const std::string& base_key)
+      : platform::MKLDNNHandler(dev_ctx, engine, base_key) {
+    conv_pd_ = conv_pd;
+  }
+
+  std::shared_ptr<mkldnn::memory> AcquireDstMemoryFromPrimitive(void* ptr) {
+    return this->AcquireMemoryFromPrimitive(conv_pd_->dst_primitive_desc(), ptr,
+                                            "@dst_mem_p");
+  }
+
+  std::shared_ptr<mkldnn::memory> AcquireSrcMemoryFromPrimitive(
+      const std::shared_ptr<mkldnn::memory> user_memory_p,
+      std::vector<mkldnn::primitive>& pipeline) {
+    auto src_pd = conv_pd_->src_primitive_desc();
+    auto user_pd = user_memory_p->get_primitive_desc();
+    return this->AcquireMemory(src_pd, user_pd, user_memory_p, "@src_mem_p",
+                               pipeline);
+  }
+
+  std::shared_ptr<mkldnn::memory> AcquireWeightsMemoryFromPrimitive(
+      const std::shared_ptr<mkldnn::memory> user_weights_memory_p,
+      std::vector<mkldnn::primitive>& pipeline) {
+    auto user_weights_pd = user_weights_memory_p->get_primitive_desc();
+    auto weights_pd = conv_pd_->weights_primitive_desc();
+    return this->AcquireMemory(weights_pd, user_weights_pd,
+                               user_weights_memory_p, "@weights_mem_p",
+                               pipeline);
+  }
+
+  std::shared_ptr<mkldnn::convolution_forward> AcquireConvolution(
+      std::shared_ptr<mkldnn::memory> src_memory_p,
+      std::shared_ptr<mkldnn::memory> weights_memory_p,
+      std::shared_ptr<mkldnn::memory> dst_memory_p) {
+    auto prim_key = key_ + "@conv_p";
+    auto prim_desc_key = key_ + "@conv_pd";
+    auto conv_p = std::static_pointer_cast<mkldnn::convolution_forward>(
+        dev_ctx_.GetBlob(prim_key));
+    PADDLE_ENFORCE((conv_p != nullptr) || (is_reusing_ == false),
+                   "Fail to find convolution primitive in device context");
+    if (conv_p == nullptr) {
+      conv_p = std::make_shared<mkldnn::convolution_forward>(
+          *conv_pd_, *(src_memory_p), *(weights_memory_p.get()),
+          *(dst_memory_p.get()));
+
+      dev_ctx_.SetBlob(prim_key, conv_p);
+    } else {
+      is_reusing_ = true;
+    }
+    return conv_p;
+  }
+
+  // Generate keys for storing/retriving primitives for this operator
+  // TODO(jczaja): Make hashing function more optimial
+  static std::string GetHash(memory::dims& input_dims,
+                             memory::dims& weights_dims,
+                             std::vector<int>& strides,
+                             std::vector<int>& paddings,
+                             std::vector<int>& dilations, int groups,
+                             const std::string& suffix) {
+    return dims2str(input_dims) + dims2str(weights_dims) + dims2str(strides) +
+           dims2str(paddings) + dims2str(dilations) + std::to_string(groups) +
+           suffix;
+  }
+
+ private:
+  std::shared_ptr<mkldnn::convolution_forward::primitive_desc> conv_pd_;
+};
+
 template <typename T>
 class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
  public:
@@ -36,10 +109,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                    "It must use CPUPlace.");
 
-    // Get unique name for index
-    const std::string key = ctx.op().Output("Output");
-    const std::string key_conv_pd = key + "@conv_pd";
-
     auto& dev_ctx =
         ctx.template device_context<paddle::platform::MKLDNNDeviceContext>();
     const auto& mkldnn_engine = dev_ctx.GetEngine();
@@ -80,68 +149,62 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
         paddle::framework::vectorize2int(filter->dims());
     std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
 
-    // create mkldnn memory from input tensors (data/weights)
-    auto user_src_memory = memory(
-        {{{src_tz}, memory::data_type::f32, input->format()}, mkldnn_engine},
-        to_void_cast(input_data));
-    auto user_weights_memory =
-        memory({{{weights_tz}, memory::data_type::f32, filter->format()},
-                mkldnn_engine},
-               to_void_cast(filter_data));
+    // Get unique name for storing MKLDNN primitives
+    const std::string key = ConvMKLDNNHandler::GetHash(
+        src_tz, weights_tz, strides, paddings, dilations, groups,
+        ctx.op().Output("Output"));
+    const std::string key_conv_pd = key + "@conv_pd";
+
+    std::vector<primitive> pipeline;
+
+    auto user_src_md = platform::MKLDNNMemDesc(
+        {src_tz}, platform::MKLDNNGetDataType<T>(), input->format());
+    auto user_weights_md = platform::MKLDNNMemDesc(
+        {weights_tz}, platform::MKLDNNGetDataType<T>(), filter->format());
 
     /* create memory descriptor for convolution without specified format
      * ('any') which lets a primitive (convolution in this case) choose
      * the memory format preferred for best performance
      */
-    auto src_md = platform::MKLDNNMemDesc(src_tz, memory::data_type::f32,
-                                          memory::format::any);
+    auto src_md = platform::MKLDNNMemDesc(
+        src_tz, platform::MKLDNNGetDataType<T>(), memory::format::any);
     auto weights_md = platform::MKLDNNMemDesc(
-        weights_tz, memory::data_type::f32, memory::format::any);
-    auto dst_md = platform::MKLDNNMemDesc(dst_tz, memory::data_type::f32,
-                                          memory::format::any);
+        weights_tz, platform::MKLDNNGetDataType<T>(), memory::format::any);
+    auto dst_md = platform::MKLDNNMemDesc(
+        dst_tz, platform::MKLDNNGetDataType<T>(), memory::format::any);
 
     // create a conv primitive descriptor and save it for usage in backward
     std::shared_ptr<conv_fwd::primitive_desc> conv_pd = ConvFwdPrimitiveDesc(
         src_md, weights_md, dst_md, strides, paddings, mkldnn_engine);
+    // Save conv_pd/src_memory/weights_memory for backward pass
+    dev_ctx.SetBlob(key_conv_pd, conv_pd);
 
-    // create reorder primitive if the input format is not the preferred one
-    auto src_memory = user_src_memory;
-    primitive reorder_src;
-    bool is_src_reordered = false;
-    if (memory::primitive_desc(conv_pd->src_primitive_desc()) !=
-        user_src_memory.get_primitive_desc()) {
-      src_memory = memory(conv_pd->src_primitive_desc());
-      reorder_src = reorder(user_src_memory, src_memory);
-      is_src_reordered = true;
-    }
-    auto weights_memory = user_weights_memory;
-    primitive reorder_weights;
-    bool is_weights_reordered = false;
-    if (memory::primitive_desc(conv_pd->weights_primitive_desc()) !=
-        user_weights_memory.get_primitive_desc()) {
-      weights_memory = memory(conv_pd->weights_primitive_desc());
-      reorder_weights = reorder(user_weights_memory, weights_memory);
-      is_weights_reordered = true;
-    }
+    ConvMKLDNNHandler handler(conv_pd, dev_ctx, mkldnn_engine, key);
 
-    // create memory primitive for conv dst
-    auto dst_memory = memory(conv_pd->dst_primitive_desc(), output_data);
+    // create mkldnn memory from input tensors (data/weights)
+    auto user_src_memory_p =
+        handler.AcquireSrcMemory(user_src_md, to_void_cast<T>(input_data));
+    auto user_weights_memory_p = handler.AcquireWeightsMemory(
+        user_weights_md, to_void_cast<T>(filter_data));
+
+    // create reorder primitive if the input format is not the preferred one
+    auto src_memory_p =
+        handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline);
+    auto weights_memory_p = handler.AcquireWeightsMemoryFromPrimitive(
+        user_weights_memory_p, pipeline);
+    auto dst_memory_p =
+        handler.AcquireDstMemoryFromPrimitive(to_void_cast<T>(output_data));
 
     // create convolution op primitive
-    auto conv_prim = conv_fwd(*conv_pd, src_memory, weights_memory, dst_memory);
+    auto conv_p = handler.AcquireConvolution(src_memory_p, weights_memory_p,
+                                             dst_memory_p);
 
     // push primitive to stream and wait until it's executed
-    std::vector<primitive> pipeline;
-    if (is_src_reordered) pipeline.push_back(reorder_src);
-    if (is_weights_reordered) pipeline.push_back(reorder_weights);
-    pipeline.push_back(conv_prim);
+    pipeline.push_back(*conv_p);
     stream(stream::kind::eager).submit(pipeline).wait();
 
-    // Save conv_pd/src_memory/weights_memory for backward pass
-    dev_ctx.SetBlob(key_conv_pd, conv_pd);
-
     output->set_layout(DataLayout::kMKLDNN);
-    output->set_format(GetMKLDNNFormat(dst_memory));
+    output->set_format(GetMKLDNNFormat(*dst_memory_p));
   }
 
  private:
@@ -197,13 +260,10 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
 
     if (!input_grad && !filter_grad) return;
 
-    // Get an unique name from "argument" name of "Output" variable
-    // This name will be used as key when saving info into device context
-    const std::string key = ctx.op().Input("Output");
-    const std::string key_conv_pd = key + "@conv_pd";
-
     std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
     std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
+    std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
+    int groups = ctx.Attr<int>("groups");
 
     const T* input_data = input->data<T>();
     const T* filter_data = filter->data<T>();
@@ -223,6 +283,14 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
         paddle::framework::vectorize2int(filter->dims());
     std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
 
+    // Get an unique name from "argument" name of "Output" variable
+    // This name will be used as key when saving info into device context
+    const std::string key =
+        ConvMKLDNNHandler::GetHash(src_tz, weights_tz, strides, paddings,
+                                   dilations, groups, ctx.op().Input("Output"));
+
+    const std::string key_conv_pd = key + "@conv_pd";
+
     // create mkldnn memory from input tensors (input/weights/output_grad)
     auto user_src_memory = memory(
         {{{src_tz}, memory::data_type::f32, input->format()}, mkldnn_engine},
diff --git a/paddle/fluid/operators/detection/rpn_target_assign_op.cc b/paddle/fluid/operators/detection/rpn_target_assign_op.cc
index 3b0c9b2886504ee381b2b33e06a4552602725e57..9a1643d5b35c067ba9064286bab32019fb34fbe8 100644
--- a/paddle/fluid/operators/detection/rpn_target_assign_op.cc
+++ b/paddle/fluid/operators/detection/rpn_target_assign_op.cc
@@ -86,8 +86,9 @@ class RpnTargetAssignKernel : public framework::OpKernel<T> {
                          std::minstd_rand engine,
                          std::vector<int>* inds) const {
     std::uniform_real_distribution<float> uniform(0, 1);
-    if (inds->size() > num) {
-      for (int i = num; i < inds->size(); ++i) {
+    const int64_t size = static_cast<int64_t>(inds->size());
+    if (size > num) {
+      for (int64_t i = num; i < size; ++i) {
         int rng_ind = std::floor(uniform(engine) * i);
         if (rng_ind < num)
           std::iter_swap(inds->begin() + rng_ind + offset,
diff --git a/paddle/fluid/operators/im2sequence_op.cc b/paddle/fluid/operators/im2sequence_op.cc
index 0669661d225c664010fce97f0a526b62988b92c5..c8c7f36536a76ea103ef6f5689c0fbdb76102688 100644
--- a/paddle/fluid/operators/im2sequence_op.cc
+++ b/paddle/fluid/operators/im2sequence_op.cc
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/im2sequence_op.h"
+#include <string>
 #include <vector>
 
 namespace paddle {
@@ -28,20 +29,19 @@ class Im2SequenceOp : public framework::OperatorWithKernel {
                    "Input(X) of Im2SequenceOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of Im2SequenceOp op should not be null.");
-
     auto in_dim = ctx->GetInputDim("X");
+
     PADDLE_ENFORCE_EQ(in_dim.size(), 4,
                       "Input(X) format must be 4D tensor, eg., NCHW.");
-
-    auto kernels = ctx->Attrs().Get<std::vector<int>>("kernels");
-    auto strides = ctx->Attrs().Get<std::vector<int>>("strides");
-    auto paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
-
     int batch_size = in_dim[0];
     int img_channels = in_dim[1];
     int img_height = in_dim[2];
     int img_width = in_dim[3];
 
+    auto kernels = ctx->Attrs().Get<std::vector<int>>("kernels");
+    auto strides = ctx->Attrs().Get<std::vector<int>>("strides");
+    auto paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
+
     int output_height = Im2SeqOutputSize(img_height, kernels[0], paddings[0],
                                          paddings[2], strides[0]);
     int output_width = Im2SeqOutputSize(img_width, kernels[1], paddings[1],
@@ -61,6 +61,10 @@ class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker {
              "C: channels"
              "H: height"
              "W: width");
+    AddInput("Y",
+             "(Tensor) The input tensor of image real size(H, W)."
+             "2-D with shape [batchsize, 2]")
+        .AsDispensable();
     AddOutput("Out", "(LodTensor) The output data of im2sequence op,");
     AddAttr<std::vector<int>>("kernels",
                               "(vector<int>), the "
@@ -73,6 +77,13 @@ class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker {
                               "(vector<int> default:{0, 0, 0, 0}), the "
                               "paddings(up_pad, left_pad, down_pad, right_pad)")
         .SetDefault({0, 0, 0, 0});
+    AddAttr<std::vector<int>>("out_stride",
+                              "the attribute is valid only when input(Y)"
+                              "is not NULL.this attribute represents the"
+                              "scaling of the pic through the CNN"
+                              "(vector<int> dedault:{1,1}),the out_stride"
+                              " (out_stride_height, out_stride_width)")
+        .SetDefault({1, 1});
     AddComment(R"DOC(
 This op uses kernels to scan images and converts these images to sequences.
 After expanding, The number of time steps are output_height * output_width
@@ -123,7 +134,7 @@ output.data = [[ 6.  2.  8.  3.  2.  4.  6.  3.]
                [ 7.  1.  7.  9.  2.  1.  3.  5.]
                [ 5.  7.  2.  4.  1.  3.  9.  0.]
                [ 7.  9.  4.  8.  3.  5.  0.  8.]]
-output.dims = {8, 9}
+output.dims = {8, 8}
 output.lod = [[0, 4, 8]]
 
 )DOC");
diff --git a/paddle/fluid/operators/im2sequence_op.h b/paddle/fluid/operators/im2sequence_op.h
index d792c68f784d8ffec0eb303a6ab9b59c9f121fa7..5bfb91db1887909c65de5f2e5321a8e6be6cf5ac 100644
--- a/paddle/fluid/operators/im2sequence_op.h
+++ b/paddle/fluid/operators/im2sequence_op.h
@@ -13,6 +13,7 @@
    limitations under the License. */
 
 #pragma once
+#include <string>
 #include <vector>
 #include "paddle/fluid/framework/data_layout.h"
 #include "paddle/fluid/framework/eigen.h"
@@ -39,50 +40,106 @@ class Im2SequenceKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     const Tensor* in = ctx.Input<Tensor>("X");
     LoDTensor* out = ctx.Output<LoDTensor>("Out");
-    out->mutable_data<T>(ctx.GetPlace());
-    // TODO(wanghaoshuang): Add layout checker after 'set_layout'
-    // being available for python API
-    // PADDLE_ENFORCE_EQ(in->layout(), framework::DataLayout::kNCHW,
-    //                  "Input(X) layout must be NCHW");
     auto in_dim = in->dims();
     int batch_size = in_dim[0];
     int img_channels = in_dim[1];
     int img_height = in_dim[2];
     int img_width = in_dim[3];
-
     auto kernels = ctx.Attr<std::vector<int>>("kernels");
     auto strides = ctx.Attr<std::vector<int>>("strides");
     auto paddings = ctx.Attr<std::vector<int>>("paddings");
-    int output_height = Im2SeqOutputSize(img_height, kernels[0], paddings[0],
-                                         paddings[2], strides[0]);
-    int output_width = Im2SeqOutputSize(img_width, kernels[1], paddings[1],
-                                        paddings[3], strides[1]);
-
-    const std::vector<int> dilations({1, 1});
-
-    auto out_dims = out->dims();
-    out->Resize({batch_size, out->numel() / batch_size});
-    for (int i = 0; i < batch_size; i++) {
-      const Tensor src =
-          in->Slice(i, i + 1).Resize({img_channels, img_height, img_width});
-      Tensor dst = out->Slice(i, i + 1).Resize(
-          {output_height, output_width, img_channels, kernels[0], kernels[1]});
-
-      math::Im2ColFunctor<math::ColFormat::kOCF, DeviceContext, T> f;
-      auto& dev_ctx = ctx.template device_context<DeviceContext>();
-      f(dev_ctx, src, dilations, strides, paddings, &dst);
-    }
-    out->Resize(out_dims);
-
-    // set lod information
-    // TODO(wanghaoshuang): Move this to InferShape
-    framework::LoD lod(1);
-    lod[0].reserve(batch_size + 1);
-    for (int i = 0, offset = 0; i < batch_size + 1; ++i) {
+    if (ctx.HasInput("Y") && batch_size > 1) {
+      const Tensor* imgrealsize = ctx.Input<Tensor>("Y");
+      auto out_stride = ctx.Attr<std::vector<int>>("out_stride");
+      Tensor cpu_shape_tensor;
+      TensorCopySync(*imgrealsize, platform::CPUPlace(), &cpu_shape_tensor);
+      std::vector<int> imgreal_h;
+      std::vector<int> imgreal_w;
+      std::vector<int> output_height;
+      std::vector<int> output_width;
+      int result = 0;
+      for (int i = 0; i < batch_size; i++) {
+        int tmp_real_h = static_cast<int>((cpu_shape_tensor.data<T>())[2 * i]);
+        int tmp_real_w =
+            static_cast<int>((cpu_shape_tensor.data<T>())[2 * i + 1]);
+        if (tmp_real_h % out_stride[0] == 0) {
+          tmp_real_h = tmp_real_h / out_stride[0];
+        } else {
+          tmp_real_h = tmp_real_h / out_stride[0] + 1;
+        }
+        if (tmp_real_w % out_stride[1] == 0) {
+          tmp_real_w = tmp_real_w / out_stride[1];
+        } else {
+          tmp_real_w = tmp_real_w / out_stride[1] + 1;
+        }
+        imgreal_h.push_back(tmp_real_h);
+        imgreal_w.push_back(tmp_real_w);
+        output_height.push_back(Im2SeqOutputSize(
+            imgreal_h[i], kernels[0], paddings[0], paddings[2], strides[0]));
+        output_width.push_back(Im2SeqOutputSize(
+            imgreal_w[i], kernels[1], paddings[1], paddings[3], strides[1]));
+        result += output_height[i] * output_width[i];
+      }
+
+      out->mutable_data<T>({result, img_channels * kernels[0] * kernels[1]},
+                           ctx.GetPlace());
+
+      const std::vector<int> dilations({1, 1});
+      int offset_out = 0;
+      for (int i = 0; i < batch_size; i++) {
+        const Tensor src =
+            in->Slice(i, i + 1).Resize({img_channels, img_height, img_width});
+        Tensor dst = out->Slice(offset_out,
+                                offset_out + output_height[i] * output_width[i])
+                         .Resize({output_height[i], output_width[i],
+                                  img_channels, kernels[0], kernels[1]});
+        offset_out += output_height[i] * output_width[i];
+
+        math::Im2ColFunctor<math::ColFormat::kOCF, DeviceContext, T> f;
+        auto& dev_ctx = ctx.template device_context<DeviceContext>();
+        f(dev_ctx, src, dilations, strides, paddings, &dst);
+      }
+      framework::LoD lod(1);
+      lod[0].reserve(batch_size + 1);
+      int offset = 0;
+      lod[0].push_back(offset);
+      for (int i = 0; i < batch_size; ++i) {
+        offset += output_height[i] * output_width[i];
+        lod[0].push_back(offset);
+      }
+      out->set_lod(lod);
+    } else {
+      out->mutable_data<T>(ctx.GetPlace());
+      int output_height = Im2SeqOutputSize(img_height, kernels[0], paddings[0],
+                                           paddings[2], strides[0]);
+      int output_width = Im2SeqOutputSize(img_width, kernels[1], paddings[1],
+                                          paddings[3], strides[1]);
+
+      const std::vector<int> dilations({1, 1});
+      auto out_dims = out->dims();
+      out->Resize({batch_size, out->numel() / batch_size});
+      for (int i = 0; i < batch_size; i++) {
+        const Tensor src =
+            in->Slice(i, i + 1).Resize({img_channels, img_height, img_width});
+        Tensor dst =
+            out->Slice(i, i + 1).Resize({output_height, output_width,
+                                         img_channels, kernels[0], kernels[1]});
+
+        math::Im2ColFunctor<math::ColFormat::kOCF, DeviceContext, T> f;
+        auto& dev_ctx = ctx.template device_context<DeviceContext>();
+        f(dev_ctx, src, dilations, strides, paddings, &dst);
+      }
+      out->Resize(out_dims);
+      framework::LoD lod(1);
+      lod[0].reserve(batch_size + 1);
+      int offset = 0;
       lod[0].push_back(offset);
-      offset += output_height * output_width;
+      for (int i = 0; i < batch_size; ++i) {
+        offset += output_height * output_width;
+        lod[0].push_back(offset);
+      }
+      out->set_lod(lod);
     }
-    out->set_lod(lod);
   }
 };
 
diff --git a/paddle/fluid/operators/math/im2col.cc b/paddle/fluid/operators/math/im2col.cc
index 336d6febc2ce3a55e82ed613bbc1081101f822f0..a50b9ace39249f4f899a46e171bbdced033b46bc 100644
--- a/paddle/fluid/operators/math/im2col.cc
+++ b/paddle/fluid/operators/math/im2col.cc
@@ -43,21 +43,6 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
     int col_height = col->dims()[3];
     int col_width = col->dims()[4];
 
-    PADDLE_ENFORCE_EQ((im_height + padding[0] + padding[2] -
-                       ((dilation[0] * (filter_height - 1) + 1))) /
-                              stride[0] +
-                          1,
-                      col_height,
-                      "Output_height and padding(padding_up, padding_down) are "
-                      "inconsistent.");
-    PADDLE_ENFORCE_EQ((im_width + padding[1] + padding[3] -
-                       ((dilation[1] * (filter_width - 1) + 1))) /
-                              stride[1] +
-                          1,
-                      col_width,
-                      "Output_height and padding(padding_up, padding_down) are "
-                      "inconsistent.");
-
     int channels_col = im_channels * filter_height * filter_width;
 
     const T* im_data = im.data<T>();
@@ -178,17 +163,6 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
     int col_height = col->dims()[0];
     int col_width = col->dims()[1];
 
-    PADDLE_ENFORCE_EQ(
-        (im_height + padding[0] + padding[2] - filter_height) / stride[0] + 1,
-        col_height,
-        "Output_height and padding(padding_up, padding_down) are "
-        "inconsistent.");
-    PADDLE_ENFORCE_EQ(
-        (im_width + padding[1] + padding[3] - filter_width) / stride[1] + 1,
-        col_width,
-        "col_width and padding(padding_left, padding_right) are "
-        "inconsistent.");
-
     const T* im_data = im.data<T>();
     T* col_data = col->data<T>();
 
diff --git a/paddle/fluid/operators/math/im2col.cu b/paddle/fluid/operators/math/im2col.cu
index eecb233d22cea06da016b2671fd606b70eddf5a5..4897767f4d88d9e079f05c921153923c4eb354b0 100644
--- a/paddle/fluid/operators/math/im2col.cu
+++ b/paddle/fluid/operators/math/im2col.cu
@@ -77,21 +77,6 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
     int col_height = col->dims()[3];
     int col_width = col->dims()[4];
 
-    PADDLE_ENFORCE_EQ((im_height + padding[0] + padding[2] -
-                       (dilation[0] * (filter_height - 1) + 1)) /
-                              stride[0] +
-                          1,
-                      col_height,
-                      "Output_height and padding(padding_up, padding_down) are "
-                      "inconsistent.");
-    PADDLE_ENFORCE_EQ((im_width + padding[1] + padding[3] -
-                       (dilation[1] * (filter_width - 1) + 1)) /
-                              stride[1] +
-                          1,
-                      col_width,
-                      "col_width and padding(padding_left, padding_right) are "
-                      "inconsistent.");
-
     int num_outputs = im_channels * col_height * col_width;
     int blocks = (num_outputs + 1024 - 1) / 1024;
     int block_x = 512;
@@ -274,21 +259,6 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
     int col_height = col->dims()[0];
     int col_width = col->dims()[1];
 
-    PADDLE_ENFORCE_EQ((im_height + padding[0] + padding[2] -
-                       (dilation[0] * (filter_height - 1) + 1)) /
-                              stride[0] +
-                          1,
-                      col_height,
-                      "Output_height and padding(padding_up, padding_down) are "
-                      "inconsistent.");
-    PADDLE_ENFORCE_EQ((im_width + padding[1] + padding[3] -
-                       (dilation[1] * (filter_width - 1) + 1)) /
-                              stride[1] +
-                          1,
-                      col_width,
-                      "col_width and padding(padding_left, padding_right) are "
-                      "inconsistent.");
-
     int block_dim_x = 0;
     int block_dim_y = 0;
     if (filter_height <= 4 && filter_width <= 4) {
diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt
index 20037d0764056c2a093af801c9cc1eb788dd46d6..e0d7937ae2f3ce4bda12f3771727e2992d63cb9b 100644
--- a/paddle/fluid/platform/CMakeLists.txt
+++ b/paddle/fluid/platform/CMakeLists.txt
@@ -46,7 +46,7 @@ ENDIF()
 # memcpy depends on device_context, here add deps individually for
 # avoiding cycle dependencies
 cc_library(device_context SRCS device_context.cc init.cc DEPS malloc
-    place eigen3 stringpiece cpu_helper ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
+    place eigen3 stringpiece cpu_helper framework_proto ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
 nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_info)
 
 cc_test(init_test SRCS init_test.cc DEPS device_context)
diff --git a/paddle/fluid/platform/mkldnn_helper.h b/paddle/fluid/platform/mkldnn_helper.h
index 33fec2c1073819d88d85a8872227adcb9df3e8f4..a8f93e6848a1db1f5aa0ee266a076af2b5d0c964 100644
--- a/paddle/fluid/platform/mkldnn_helper.h
+++ b/paddle/fluid/platform/mkldnn_helper.h
@@ -222,15 +222,16 @@ class MKLDNNHandler {
 
   static std::string GetHash(mkldnn::memory::dims& operand_dims,  // NOLINT
                              const std::string& suffix) {
-    auto dims2str = [](const mkldnn::memory::dims& operand_dims) {
-      std::string dstr = "";
-      for (size_t i = 0; i < operand_dims.size(); ++i) {
-        dstr += std::to_string(operand_dims[i]) + "-";
-      }
-      return dstr;
-    };
-
     return dims2str(operand_dims) + suffix;
+  };
+
+ protected:
+  static std::string dims2str(const mkldnn::memory::dims& operand_dims) {
+    std::string dstr = "";
+    for (size_t i = 0; i < operand_dims.size(); ++i) {
+      dstr += std::to_string(operand_dims[i]) + "-";
+    }
+    return dstr;
   }
 
  protected:
diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh
index d173b41e86f61954954b6a5ea9957d2e172deca0..bf45c11a9de53a109c72ff7a89b807bc80feb7c8 100755
--- a/paddle/scripts/paddle_build.sh
+++ b/paddle/scripts/paddle_build.sh
@@ -510,11 +510,23 @@ function gen_fluid_inference_lib() {
 EOF
         make -j `nproc` inference_lib_dist
         cd ${PADDLE_ROOT}/build
-        mv fluid_install_dir fluid
+        cp -r fluid_install_dir fluid
         tar -cf fluid.tgz fluid
       fi
 }
 
+function test_fluid_inference_lib() {
+    if [ ${WITH_C_API:-OFF} == "OFF" ] ; then
+        cat <<EOF
+    ========================================
+    Testing fluid inference library ...
+    ========================================
+EOF
+        cd ${PADDLE_ROOT}/paddle/contrib/inference/demo_ci
+        sh run.sh ${PADDLE_ROOT} ${WITH_MKL:-ON} ${WITH_GPU:-OFF}
+      fi
+}
+
 function main() {
     set -e
     local CMD=$1
@@ -568,6 +580,7 @@ function main() {
         run_test
         gen_capi_package
         gen_fluid_inference_lib
+        test_fluid_inference_lib
         ;;
       *)
         print_usage
diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py
index 4faa06303170488d0de2fda4c1461cfe2d623d35..e7a065599ec034039442449ea20e1fcd02c44f8d 100644
--- a/python/paddle/fluid/backward.py
+++ b/python/paddle/fluid/backward.py
@@ -123,7 +123,8 @@ def _append_grad_suffix_(name):
 def _addup_repetitive_outputs_(op_descs):
     """
     In backward part, an variable may be the output of more than one ops.
-    In this case, the variable should be the accumulation of all the outputs.
+    And one op may yield its multiple outputs to the same variable.
+    In these cases, the variable should be the accumulation of all the outputs.
     `sum_op`s are added to implement the accumulate.
     """
     pending_sum_ops = []
@@ -136,29 +137,46 @@ def _addup_repetitive_outputs_(op_descs):
                     "sum", {"X": renamed_vars[var_name]}, {"Out": [var_name]},
                     {"use_mkldnn": False}), idx))
                 renamed_vars[var_name] = [var_name]
-        for var_name in op_desc.output_arg_names():
-            if var_name == core.empty_var_name(
-            ) or var_name in op_desc.input_arg_names():
-                # empty variable or inplace op
-                continue
-            if len(renamed_vars[var_name]) == 0:
-                # it's the first time we get the variable
-                renamed_vars[var_name] = [var_name]
-            else:
-                if len(renamed_vars[var_name]) == 1:
+        for param_idx, param_name in enumerate(op_desc.output_names()):
+            arg_names = op_desc.output(param_name)
+            for arg_idx, var_name in enumerate(arg_names):
+                if var_name == core.empty_var_name(
+                ) or var_name in op_desc.input_arg_names():
+                    # empty variable or inplace op
+                    continue
+                if len(renamed_vars[var_name]) == 0:
+                    # it's the first time we get the variable
+                    renamed_vars[var_name] = [var_name]
+                else:
+                    if len(renamed_vars[var_name]) == 1:
+                        new_name = var_name + "@RENAME@" + \
+                            str(var_rename_count[var_name])
+                        var_rename_count[var_name] += 1
+                        # rename original var_name
+                        renamed_vars[var_name][0] = new_name
+                        _rename_arg_(op_descs, var_name, new_name, 0, idx)
+                        _rename_arg_(pending_sum_ops, var_name, new_name)
+
+                        for p in op_desc.output_names()[:param_idx]:
+                            p_arg_names = op_desc.output(p)
+                            if var_name in p_arg_names:
+                                op_desc.set_output(p, [
+                                    new_name if x == var_name else x
+                                    for x in p_arg_names
+                                ])
+
+                        arg_names = [
+                            new_name if x == var_name else x
+                            for x in arg_names[:arg_idx]
+                        ] + arg_names[arg_idx:]
+
                     new_name = var_name + "@RENAME@" + \
                         str(var_rename_count[var_name])
                     var_rename_count[var_name] += 1
-                    # rename original var_name
-                    renamed_vars[var_name][0] = new_name
-                    _rename_arg_(op_descs, var_name, new_name, 0, idx)
-                    _rename_arg_(pending_sum_ops, var_name, new_name)
-
-                new_name = var_name + "@RENAME@" + \
-                    str(var_rename_count[var_name])
-                var_rename_count[var_name] += 1
-                op_desc.rename_output(var_name, new_name)
-                renamed_vars[var_name].append(new_name)
+                    arg_names[arg_idx] = new_name
+                    op_desc.set_output(param_name, arg_names)
+                    renamed_vars[var_name].append(new_name)
+
     for var_name, inputs in renamed_vars.iteritems():
         if len(inputs) > 1:
             pending_sum_ops.append(
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
index bcf520d5a4e3bbe1d949d08f42199dd8c5cdc947..07b806f544497ccabe4dde9a370e90da372e6cba 100644
--- a/python/paddle/fluid/layers/nn.py
+++ b/python/paddle/fluid/layers/nn.py
@@ -1,4 +1,18 @@
-#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# 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.
+
+#   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.
@@ -3900,7 +3914,13 @@ def transpose(x, perm, name=None):
     return out
 
 
-def im2sequence(input, filter_size=1, stride=1, padding=0, name=None):
+def im2sequence(input,
+                filter_size=1,
+                stride=1,
+                padding=0,
+                input_image_size=None,
+                out_stride=1,
+                name=None):
     """
     Extracts image patches from the input tensor to form a tensor of shape
     {input.batch_size * output_height * output_width, filter_size_H *
@@ -3937,6 +3957,15 @@ def im2sequence(input, filter_size=1, stride=1, padding=0, name=None):
             padding_up = padding_down = padding_left = padding_right = padding
             Default: padding = 0.
 
+        input_image_size(Variable): the input contains image real size.It's dim
+            is [batchsize, 2]. It is dispensable.It is just for batch inference.
+
+        out_stride(int|tuple): The scaling of image through CNN. It is
+            dispensable. It is valid only when input_image_size is not null.
+            If out_stride is tuple,  it must contain two intergers,
+            (out_stride_H, out_stride_W). Otherwise,
+            the out_stride_H = out_stride_W = out_stride.
+
         name (int): The name of this layer. It is optional.
 
     Returns:
@@ -3987,7 +4016,7 @@ def im2sequence(input, filter_size=1, stride=1, padding=0, name=None):
                            [ 5.  7.  2.  4.  1.  3.  9.  0.]
                            [ 7.  9.  4.  8.  3.  5.  0.  8.]]
 
-            output.dims = {8, 9}
+            output.dims = {8, 8}
 
             output.lod = [[4, 4]]
 
@@ -4009,18 +4038,17 @@ def im2sequence(input, filter_size=1, stride=1, padding=0, name=None):
     if len(padding) == 2:
         padding.append(padding[0])
         padding.append(padding[1])
-
+    inputs = {"X": input}
+    attrs = {"kernels": filter_size, "strides": stride, "padding": padding}
+    if input_image_size:
+        if isinstance(out_stride, int):
+            out_stride = [out_stride, out_stride]
+        inputs["Y"] = input_image_size
+        attrs["out_stride"] = out_stride
     helper = LayerHelper('im2sequence', **locals())
     out = helper.create_tmp_variable(dtype=helper.input_dtype())
     helper.append_op(
-        type='im2sequence',
-        inputs={'X': input},
-        outputs={'Out': out},
-        attrs={
-            'kernels': filter_size,
-            'strides': stride,
-            'paddings': padding,
-        })
+        type='im2sequence', inputs=inputs, outputs={'Out': out}, attrs=attrs)
     return out
 
 
diff --git a/python/paddle/fluid/tests/unittests/test_im2sequence_op.py b/python/paddle/fluid/tests/unittests/test_im2sequence_op.py
index 4946475f11a4fc0ccaffeec6821d3976ea7c6560..13bc5768740ece00bbe285a0b47d82bb8a42d2c7 100644
--- a/python/paddle/fluid/tests/unittests/test_im2sequence_op.py
+++ b/python/paddle/fluid/tests/unittests/test_im2sequence_op.py
@@ -16,23 +16,48 @@ import numpy as np
 from op_test import OpTest
 
 
-def get_output_shape(attrs, in_shape):
+def get_output_shape(attrs, in_shape, img_real_size):
+    batchsize = in_shape[0]
     img_height = in_shape[2]
     img_width = in_shape[3]
+    paddings = np.array(attrs['paddings']).astype("int32")
+    kernels = np.array(attrs['kernels']).astype("int32")
+    strides = np.array(attrs['strides']).astype("int32")
+    output_height = np.zeros((1, batchsize)).astype("int32")
+    output_width = np.zeros((1, batchsize)).astype("int32")
+    if len(img_real_size):
+        out_stride = np.array(attrs['out_stride']).astype("int32")
+        imgreal_h = 0
+        imgreal_w = 0
+        for index in range(batchsize):
+            if img_real_size[index, 0] % out_stride[0] == 0:
+                imgreal_h = img_real_size[index, 0] / out_stride[0]
+            else:
+                imgreal_h = img_real_size[index, 0] / out_stride[0] + 1
+            if img_real_size[index, 0] % out_stride[1] == 0:
+                imgreal_w = img_real_size[index, 1] / out_stride[1]
+            else:
+                imgreal_w = img_real_size[index, 0] / out_stride[1] + 1
+            output_height[0,index] = \
+              1 +  \
+              (imgreal_h + paddings[0] + paddings[2] - kernels[0] + strides[0] - 1) / \
+                  strides[0]
 
-    paddings = attrs['paddings']
-    kernels = attrs['kernels']
-    strides = attrs['strides']
+            output_width[0,index] = \
+              1 + \
+              (imgreal_w + paddings[1] + paddings[3] - kernels[1] + strides[1] - 1) / \
+                  strides[1]
+    else:
+        for index in range(batchsize):
+            output_height[0,index] = \
+              1 +  \
+              (img_height + paddings[0] + paddings[2] - kernels[0] + strides[0] - 1) / \
+                  strides[0]
 
-    output_height = \
-      1 +  \
-      (img_height + paddings[0] + paddings[2] - kernels[0] + strides[0] - 1) / \
-          strides[0]
-
-    output_width = \
-      1 + \
-      (img_width + paddings[1] + paddings[3] - kernels[1] + strides[1] - 1) / \
-          strides[1]
+            output_width[0,index] = \
+              1 + \
+              (img_width + paddings[1] + paddings[3] - kernels[1] + strides[1] - 1) / \
+                  strides[1]
 
     return output_height, output_width
 
@@ -75,22 +100,25 @@ def im2col(attrs, im, col):
                                     im_row_offset][im_col_offset]
 
 
-def Im2Sequence(inputs, attrs):
-    output_height, output_width = get_output_shape(attrs, inputs.shape)
+def Im2Sequence(inputs, img_real_size, attrs):
+    output_height, output_width = get_output_shape(attrs, inputs.shape,
+                                                   img_real_size)
     img_channels = inputs.shape[1]
     batch_size = inputs.shape[0]
-    out = np.zeros([
-        batch_size, output_height, output_width, img_channels,
-        attrs['kernels'][0], attrs['kernels'][1]
-    ]).astype("float32")
-
-    for i in range(len(inputs)):
-        im2col(attrs, inputs[i], out[i])
-
-    out = out.reshape([
-        batch_size * output_height * output_width,
-        img_channels * attrs['kernels'][0] * attrs['kernels'][1]
-    ])
+    out = []
+    for index in range(batch_size):
+        tmp = np.zeros([
+            output_height[0, index], output_width[0, index], img_channels,
+            attrs['kernels'][0], attrs['kernels'][1]
+        ]).astype("float32")
+        out.append(tmp)
+    for index in range(len(inputs)):
+        im2col(attrs, inputs[index], out[index])
+        out[index] = out[index].reshape([
+            output_height[0, index] * output_width[0, index],
+            img_channels * attrs['kernels'][0] * attrs['kernels'][1]
+        ])
+    out = np.concatenate(out, axis=0)
     return out
 
 
@@ -103,7 +131,7 @@ class TestBlockExpandOp(OpTest):
         self.attrs = {
             'kernels': [2, 2],
             'strides': [1, 1],
-            'paddings': [1, 1, 1, 1]
+            'paddings': [1, 1, 1, 1],
         }
 
     def setUp(self):
@@ -113,7 +141,8 @@ class TestBlockExpandOp(OpTest):
             self.batch_size, self.img_channels, self.img_height, self.img_width
         ]).astype("float32")
 
-        out = Im2Sequence(x, self.attrs)
+        real_size = np.array([]).astype("float32")
+        out = Im2Sequence(x, real_size, self.attrs)
         self.inputs = {'X': x}
         self.outputs = {'Out': out}
 
@@ -133,20 +162,20 @@ class TestBlockExpandOpCase2(TestBlockExpandOp):
         self.attrs = {
             'kernels': [2, 1],
             'strides': [2, 1],
-            'paddings': [2, 1, 2, 1]
+            'paddings': [2, 1, 2, 1],
         }
 
 
 class TestBlockExpandOpCase3(TestBlockExpandOp):
     def config(self):
-        self.batch_size = 3
+        self.batch_size = 2
         self.img_channels = 1
         self.img_height = 4
         self.img_width = 5
         self.attrs = {
             'kernels': [2, 1],
             'strides': [2, 1],
-            'paddings': [2, 0, 2, 0]
+            'paddings': [2, 0, 2, 0],
         }
 
 
@@ -159,9 +188,94 @@ class TestBlockExpandOpCase4(TestBlockExpandOp):
         self.attrs = {
             'kernels': [2, 2],
             'strides': [1, 1],
-            'paddings': [0, 0, 0, 0]
+            'paddings': [0, 0, 0, 0],
+        }
+
+
+class TestBlockExpandOpCase5(OpTest):
+    def config(self):
+        self.batch_size = 1
+        self.img_channels = 3
+        self.img_height = 4
+        self.img_width = 5
+        self.attrs = {
+            'kernels': [2, 1],
+            'strides': [2, 1],
+            'paddings': [2, 1, 2, 1],
+            'out_stride': [2, 2],
+        }
+
+    def setUp(self):
+        self.config()
+        self.op_type = "im2sequence"
+        x = np.random.uniform(0.1, 1, [
+            self.batch_size, self.img_channels, self.img_height, self.img_width
+        ]).astype("float32")
+        real_size = np.array([[8, 10], [5, 8]]).astype("float32")
+        out = np.array(Im2Sequence(x, real_size, self.attrs))
+        self.inputs = {'X': x, 'Y': real_size}  #l ??
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestBlockExpandOpCase6(OpTest):
+    def config(self):
+        self.batch_size = 3
+        self.img_channels = 1
+        self.img_height = 4
+        self.img_width = 5
+        self.attrs = {
+            'kernels': [2, 1],
+            'strides': [1, 1],
+            'paddings': [0, 0, 0, 0],
+            'out_stride': [1, 1],
+        }
+
+    def setUp(self):
+        self.config()
+        self.op_type = "im2sequence"
+        x = np.random.uniform(0.1, 1, [
+            self.batch_size, self.img_channels, self.img_height, self.img_width
+        ]).astype("float32")
+        real_size = np.array([[8, 10], [5, 8], [5, 8]]).astype("float32")
+        out = np.array(Im2Sequence(x, real_size, self.attrs))
+        self.inputs = {'X': x, 'Y': real_size}  #l ??
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestBlockExpandOpCase7(OpTest):
+    def config(self):
+        self.batch_size = 2
+        self.img_channels = 2
+        self.img_height = 3
+        self.img_width = 3
+        self.attrs = {
+            'kernels': [2, 2],
+            'strides': [1, 1],
+            'paddings': [1, 0, 1, 0],
+            'out_stride': [2, 2],
         }
 
+    def setUp(self):
+        self.config()
+        self.op_type = "im2sequence"
+        x = np.random.uniform(0.1, 1, [
+            self.batch_size, self.img_channels, self.img_height, self.img_width
+        ]).astype("float32")
+        real_size = np.array([[6, 6], [4, 4]]).astype("float32")
+        out = np.array(Im2Sequence(x, real_size, self.attrs))
+        self.inputs = {'X': x, 'Y': real_size}
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
 
 if __name__ == '__main__':
     unittest.main()
+#set shiftwidth=4 set expandtab set tabstop=4
diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py
index 842d34c07e94a79e3351347e2528ecc478cc56dc..31ae4e7d8c4107fce7022234d3aa47827647d012 100644
--- a/python/paddle/fluid/tests/unittests/test_layers.py
+++ b/python/paddle/fluid/tests/unittests/test_layers.py
@@ -251,12 +251,16 @@ class TestBook(unittest.TestCase):
         print(str(program))
 
     def test_im2sequence(self):
-        print("test_im2sequence")
         program = Program()
         with program_guard(program):
             x = layers.data(name='x', shape=[3, 128, 128], dtype='float32')
+            y = layers.data(name='y', shape=[], dtype='float32')
             output = layers.im2sequence(
-                input=x, stride=[1, 1], filter_size=[2, 2])
+                input=x,
+                input_image_size=y,
+                stride=[1, 1],
+                filter_size=[2, 2],
+                out_stride=[1, 1])
             self.assertIsNotNone(output)
         print(str(program))
 
diff --git a/python/setup.py.in b/python/setup.py.in
index ee4bbf6febc667f487a83631bbefb5ef11a2177f..a064f36cc19dbc626dd85d76290280a01fa57215 100644
--- a/python/setup.py.in
+++ b/python/setup.py.in
@@ -181,6 +181,14 @@ else:
     command = "patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so"
 if os.system(command) != 0:
     raise Exception("patch core.so failed, command: %s" % command)
+if '${WITH_FLUID_ONLY}'== 'OFF':
+    # change rpath of _swig_paddle.so.
+    if "@APPLE@" == "1":
+        command = "install_name_tool -id \"@loader_path/../paddle/libs/\" ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle.so"
+    else:
+        command = "patchelf --set-rpath '$ORIGIN/../paddle/libs/' ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle.so"
+    if os.system(command) != 0:
+        raise Exception("patch _swig_paddle.so failed, command: %s" % command)
 
 setup(name='${PACKAGE_NAME}',
       version='${PADDLE_VERSION}',