diff --git a/doc/fluid/new_docs/user_guides/howto/inference/native_infer.rst b/doc/fluid/new_docs/user_guides/howto/inference/native_infer.rst
index 21a6fe5cf54d0c0c760ade4ba602024ffa29675f..6d6f3035c0b5c985cd39d45df9f1bcce50dcefa0 100644
--- a/doc/fluid/new_docs/user_guides/howto/inference/native_infer.rst
+++ b/doc/fluid/new_docs/user_guides/howto/inference/native_infer.rst
@@ -4,13 +4,12 @@ Paddle 预测 API
 为了更简单方便的预测部署，Fluid 提供了一套高层 API
 用来隐藏底层不同的优化实现。
 
-`预测库相关代码 <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/contrib/inference>`__
+`预测库相关代码 <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/inference/api>`_
 包括
 
 -  头文件 ``paddle_inference_api.h`` 定义了所有的接口
 -  库文件\ ``libpaddle_fluid.so`` 或 ``libpaddle_fluid.a``
--  库文件 ``libpaddle_inference_api.so`` 或
-   ``libpaddle_inference_api.a``
+
 
 编译和依赖可以参考 :ref:`install_or_build_cpp_inference_lib` 。
 
@@ -97,8 +96,7 @@ engine
     CHECK(predictor->Run(slots, &outputs));
     // 获取 outputs ...
 
-编译时，联编 ``libpaddle_fluid.a/.so`` 和
-``libpaddle_inference_api.a/.so`` 便可。
+编译时，联编 ``libpaddle_fluid.a/.so`` 便可。
 
 详细代码参考
 ------------
diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec
index bb5f2894c08b5d8941ad8914f6b83280aa053e37..70e5b97770a6c581c6a9c0145b03c42b83f14471 100644
--- a/paddle/fluid/API.spec
+++ b/paddle/fluid/API.spec
@@ -312,7 +312,7 @@ paddle.fluid.layers.iou_similarity ArgSpec(args=[], varargs='args', keywords='kw
 paddle.fluid.layers.box_coder ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.polygon_box_transform ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None))
-paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk'], varargs=None, keywords=None, defaults=('ROC', 200, 1))
+paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk'], varargs=None, keywords=None, defaults=('ROC', 4095, 1))
 paddle.fluid.layers.exponential_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,))
 paddle.fluid.layers.natural_exp_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,))
 paddle.fluid.layers.inverse_time_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,))
@@ -376,7 +376,7 @@ paddle.fluid.optimizer.DecayedAdagradOptimizer.__init__ ArgSpec(args=['self', 'l
 paddle.fluid.optimizer.DecayedAdagradOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.FtrlOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'l1', 'l2', 'lr_power'], varargs=None, keywords='kwargs', defaults=(0.0, 0.0, -0.5))
 paddle.fluid.optimizer.FtrlOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
-paddle.fluid.optimizer.RMSPropOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'rho', 'epsilon', 'momentum'], varargs=None, keywords='kwargs', defaults=(0.95, 1e-06, 0.0))
+paddle.fluid.optimizer.RMSPropOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'rho', 'epsilon', 'momentum', 'centered'], varargs=None, keywords='kwargs', defaults=(0.95, 1e-06, 0.0, False))
 paddle.fluid.optimizer.RMSPropOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.AdadeltaOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'rho'], varargs=None, keywords='kwargs', defaults=(1e-06, 0.95))
 paddle.fluid.optimizer.AdadeltaOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc
index 0bfff745493d069e948e6d277ec2bbfb0673a70b..7a99169849debcbc57d6f197b36c5045b211f3ef 100644
--- a/paddle/fluid/framework/details/multi_devices_graph_pass.cc
+++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc
@@ -326,7 +326,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
   ir::Graph &result = *graph;
 
   for (auto &node : nodes) {
-    if (node->NodeType() == ir::Node::Type::kVariable && node->Var()) {
+    if (node->IsVar() && node->Var()) {
       all_vars_.emplace(node->Name(), node->Var());
     }
   }
@@ -583,18 +583,6 @@ void MultiDevSSAGraphBuilder::InsertDataBalanceOp(
   }
 }
 
-bool MultiDevSSAGraphBuilder::IsParameterGradientOnce(
-    const std::string &og,
-    std::unordered_set<std::string> *og_has_been_broadcast) const {
-  bool is_pg_once =
-      grad_names_.count(og) != 0 && og_has_been_broadcast->count(og) == 0;
-  if (is_pg_once) {
-    // Insert NCCL AllReduce Op
-    og_has_been_broadcast->insert(og);
-  }
-  return is_pg_once;
-}
-
 int MultiDevSSAGraphBuilder::GetOpDeviceID(const ir::Graph &graph,
                                            ir::Node *node) const {
   if (strategy_.reduce_ != BuildStrategy::ReduceStrategy::kReduce) {
@@ -688,20 +676,6 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result,
   return var;
 }
 
-// Find the first occurence of `prev_op_name` and make current `op` depend
-// on it.
-void MultiDevSSAGraphBuilder::ConnectOp(ir::Graph *result, OpHandleBase *op,
-                                        const std::string &prev_op_name) const {
-  for (auto &prev_op : result->Get<GraphOps>(kGraphOps)) {
-    if (prev_op->Name() == prev_op_name) {
-      auto *dep_var = new DummyVarHandle(result->CreateControlDepVar());
-      prev_op->AddOutput(dep_var);
-      result->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
-      op->AddInput(dep_var);
-    }
-  }
-}
-
 void MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
                                                 ir::Node *node) const {
   int op_dev_id = -1;
diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.h b/paddle/fluid/framework/details/multi_devices_graph_pass.h
index 7a6f238f9cf7af18cb10ea271e453fec1902c833..ac6d9c5a64cfde60f75c76dae0a30cc7d735e996 100644
--- a/paddle/fluid/framework/details/multi_devices_graph_pass.h
+++ b/paddle/fluid/framework/details/multi_devices_graph_pass.h
@@ -69,9 +69,6 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
   std::vector<std::string> FindDistTrainRecvVars(
       const std::vector<ir::Node *> &nodes) const;
 
-  void ConnectOp(ir::Graph *result, OpHandleBase *op,
-                 const std::string &prev_op_name) const;
-
   void CreateComputationalOps(ir::Graph *result, ir::Node *node,
                               size_t num_places) const;
 
@@ -83,10 +80,6 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
   void CreateComputationalOp(ir::Graph *result, ir::Node *node,
                              int dev_id) const;
 
-  bool IsParameterGradientOnce(
-      const std::string &og,
-      std::unordered_set<std::string> *og_has_been_broadcast) const;
-
   int GetOpDeviceID(const ir::Graph &graph, ir::Node *node) const;
 
   void InsertAllReduceOp(ir::Graph *result, const std::string &og) const;
diff --git a/paddle/fluid/framework/ir/fc_lstm_fuse_pass.cc b/paddle/fluid/framework/ir/fc_lstm_fuse_pass.cc
index 00f5e7fad2ef5d42eb0de9703389e910090d93c1..9512fd056e73836cdc34a9e409ab2d7809a6aff6 100644
--- a/paddle/fluid/framework/ir/fc_lstm_fuse_pass.cc
+++ b/paddle/fluid/framework/ir/fc_lstm_fuse_pass.cc
@@ -87,15 +87,24 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope,
       }
       op_desc.SetInput("Bias", {new_bias_var});
     }
-
 #undef GET_NODE
 
+    // Create temp variables.
+    scope->Var(name_scope + "/BatchedInput.new")
+        ->GetMutable<framework::LoDTensor>();
+    scope->Var(name_scope + "/BatchCellPreAct.new")
+        ->GetMutable<framework::LoDTensor>();
+    scope->Var(name_scope + "/BatchedGate.new")
+        ->GetMutable<framework::LoDTensor>();
+
     op_desc.SetInput("H0", {});
     op_desc.SetInput("C0", {});
     op_desc.SetOutput("Hidden", {hidden_n->Name()});
     op_desc.SetOutput("Cell", {cell_n->Name()});
     op_desc.SetOutput("XX", {xx_n->Name()});
-    op_desc.SetOutput("BatchedInput", {"blstm_0.tmp_2"});
+    op_desc.SetOutput("BatchedGate", {name_scope + "/BatchedGate.new"});
+    op_desc.SetOutput("BatchCellPreAct", {name_scope + "/BatchCellPreAct.new"});
+    op_desc.SetOutput("BatchedInput", {name_scope + "/BatchedInput.new"});
     op_desc.SetAttr("is_reverse", lstm_n->Op()->GetAttr("is_reverse"));
     op_desc.SetAttr("use_peepholes", lstm_n->Op()->GetAttr("use_peepholes"));
     // TODO(TJ): get from attr
@@ -131,8 +140,8 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope,
 
   int fusion_count{0};
 
-  auto fc_no_bias_handler = [&](
-      const GraphPatternDetector::subgraph_t& subgraph, Graph* g) {
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
 #define GET_NODE(name__)                                \
   std::string name__##key = name_scope + "/" + #name__; \
   auto* name__##n = pattern->RetrieveNode(name__##key); \
@@ -153,21 +162,24 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope,
 
     if (with_fc_bias) {
       GET_NODE(fc_bias);
+      GET_NODE(elementwise_add);
       lstm_creator(lstm, x, w, Weight, Bias, Hidden, Cell, fc_out, fc_bias);
+      // Remove unneeded nodes.
+      std::unordered_set<const Node*> marked_nodes(
+          {mul_n, lstm_n, elementwise_add_n});
+      GraphSafeRemoveNodes(graph, marked_nodes);
     } else {
       lstm_creator(lstm, x, w, Weight, Bias, Hidden, Cell, fc_out, -1);
+      // Remove unneeded nodes.
+      std::unordered_set<const Node*> marked_nodes({mul_n, lstm_n});
+      GraphSafeRemoveNodes(graph, marked_nodes);
     }
 #undef GET_NODE
 
-    // Remove unneeded nodes.
-    std::unordered_set<const Node*> marked_nodes({mul_n, lstm_n});
-
-    GraphSafeRemoveNodes(graph, marked_nodes);
-
     ++fusion_count;
   };
 
-  gpd(graph, fc_no_bias_handler);
+  gpd(graph, handler);
 
   return fusion_count;
 }
diff --git a/paddle/fluid/framework/ir/fc_lstm_fuse_pass.h b/paddle/fluid/framework/ir/fc_lstm_fuse_pass.h
index 5a6687872eb3ab4a032227fda9ff0e7f5254670b..3ee32c63a46fcc34bdccd1e14d4bbaf9668c49e9 100644
--- a/paddle/fluid/framework/ir/fc_lstm_fuse_pass.h
+++ b/paddle/fluid/framework/ir/fc_lstm_fuse_pass.h
@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#pragma once
+
 #include "paddle/fluid/framework/ir/fuse_pass_base.h"
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/graph_pattern_detector.h"
diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc
index a4da69a0a2ea44806b68a27647213759ebd387b1..434bee4ccee1c199088d09c934fe86435ec7d095 100644
--- a/paddle/fluid/framework/ir/graph_pattern_detector.cc
+++ b/paddle/fluid/framework/ir/graph_pattern_detector.cc
@@ -73,7 +73,6 @@ void PDPattern::AddEdge(PDNode* a, PDNode* b) {
 void GraphPatternDetector::operator()(Graph* graph,
                                       GraphPatternDetector::handle_t handler) {
   if (!MarkPDNodesInGraph(*graph)) {
-    LOG(INFO) << "Mark failed";
     return;
   }
 
diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.h b/paddle/fluid/framework/ir/graph_pattern_detector.h
index 9d67c4a6997dfe19561f37bf3ea76eba8b59ff35..eacea1750f6f1e86a8fe79637c3bd757a7275398 100644
--- a/paddle/fluid/framework/ir/graph_pattern_detector.h
+++ b/paddle/fluid/framework/ir/graph_pattern_detector.h
@@ -19,6 +19,9 @@
 #endif
 
 #include <numeric>
+#include <string>
+#include <utility>
+#include <vector>
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/node.h"
 #include "paddle/fluid/inference/analysis/dot.h"
diff --git a/paddle/fluid/inference/analysis/CMakeLists.txt b/paddle/fluid/inference/analysis/CMakeLists.txt
index dadc8a53706fb9edff884dcf6d49168bfef3aa30..f2e18a461fd221252e4a10262a13bc8e942f5988 100644
--- a/paddle/fluid/inference/analysis/CMakeLists.txt
+++ b/paddle/fluid/inference/analysis/CMakeLists.txt
@@ -58,7 +58,7 @@ endif()
 inference_analysis_test(test_analyzer SRCS analyzer_tester.cc
     EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor
     ARGS --infer_ditu_rnn_model=${DITU_INSTALL_DIR}/model
-        --infer_ditu_rnn_data=${DITU_INSTALL_DIR}/data.txt)
+         --infer_ditu_rnn_data=${DITU_INSTALL_DIR}/data.txt)
 
 inference_analysis_test(test_data_flow_graph SRCS data_flow_graph_tester.cc)
 inference_analysis_test(test_data_flow_graph_to_fluid_pass SRCS data_flow_graph_to_fluid_pass_tester.cc)
@@ -74,7 +74,7 @@ inference_analysis_test(test_model_store_pass SRCS model_store_pass_tester.cc)
 set(CHINESE_NER_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/chinese_ner_model.tar.gz")
 set(CHINESE_NER_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/chinese_ner-data.txt.tar.gz")
 set(CHINESE_NER_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo/chinese_ner" CACHE PATH "Chinese ner model and data root." FORCE)
-if (NOT EXISTS ${CHINESE_NER_INSTALL_DIR} AND WITH_TESTING)
+if (NOT EXISTS ${CHINESE_NER_INSTALL_DIR} AND WITH_TESTING AND WITH_INFERENCE)
   inference_download_and_uncompress(${CHINESE_NER_INSTALL_DIR} ${CHINESE_NER_MODEL_URL} "chinese_ner_model.tar.gz")
   inference_download_and_uncompress(${CHINESE_NER_INSTALL_DIR} ${CHINESE_NER_DATA_URL} "chinese_ner-data.txt.tar.gz")
 endif()
@@ -87,7 +87,7 @@ inference_analysis_test(test_analyzer_ner SRCS analyzer_ner_tester.cc
 set(LAC_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/lac_model.tar.gz")
 set(LAC_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/lac_data.txt.tar.gz")
 set(LAC_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo/lac" CACHE PATH "LAC model and data root." FORCE)
-if (NOT EXISTS ${LAC_INSTALL_DIR} AND WITH_TESTING)
+if (NOT EXISTS ${LAC_INSTALL_DIR} AND WITH_TESTING AND WITH_INFERENCE)
     inference_download_and_uncompress(${LAC_INSTALL_DIR} ${LAC_MODEL_URL} "lac_model.tar.gz")
     inference_download_and_uncompress(${LAC_INSTALL_DIR} ${LAC_DATA_URL} "lac_data.txt.tar.gz")
 endif()
@@ -96,3 +96,15 @@ inference_analysis_test(test_analyzer_lac SRCS analyzer_lac_tester.cc
     EXTRA_DEPS paddle_inference_api paddle_fluid_api
     ARGS --infer_model=${LAC_INSTALL_DIR}/model
         --infer_data=${LAC_INSTALL_DIR}/data.txt)
+
+
+set(TEXT_CLASSIFICATION_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/text-classification-Senta.tar.gz")
+set(TEXT_CLASSIFICATION_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo/text_classification" CACHE PATH "Text Classification model and data root." FORCE)
+
+if (NOT EXISTS ${TEXT_CLASSIFICATION_INSTALL_DIR} AND WITH_TESTING AND WITH_INFERENCE)
+  inference_download_and_uncompress(${TEXT_CLASSIFICATION_INSTALL_DIR} ${TEXT_CLASSIFICATION_MODEL_URL} "text-classification-Senta.tar.gz")
+endif()
+
+inference_analysis_test(test_text_classification SRCS test_text_classification.cc
+    EXTRA_DEPS paddle_inference_api paddle_fluid_api analysis_predictor
+    ARGS --infer_model=${TEXT_CLASSIFICATION_INSTALL_DIR}/text-classification-Senta)
diff --git a/paddle/fluid/inference/analysis/analyzer.cc b/paddle/fluid/inference/analysis/analyzer.cc
index 192ac2daa6a78efec6db19870f71e80593c62da9..ca834406451d53fd44887300561a6327d97cafcd 100644
--- a/paddle/fluid/inference/analysis/analyzer.cc
+++ b/paddle/fluid/inference/analysis/analyzer.cc
@@ -14,6 +14,7 @@
 
 #include "paddle/fluid/inference/analysis/analyzer.h"
 #include <string>
+#include <vector>
 #include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h"
 #include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h"
 #include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
@@ -41,20 +42,16 @@ class DfgPassManagerImpl final : public DfgPassManager {
  public:
   DfgPassManagerImpl() {
     // TODO(Superjomn) set the key with pass reprs.
-    LOG(INFO)
-        << "-----------------------------------------------------------------";
-    if (FLAGS_IA_enable_ir) {
-      AddPass("fluid-to-ir-pass", new FluidToIrPass);
-    } else {
+    if (!FLAGS_IA_enable_ir) {
       AddPass("fluid-to-data-flow-graph", new FluidToDataFlowGraphPass);
+    } else {
+      AddPass("fluid-to-ir-pass", new FluidToIrPass);
     }
     TryAddTensorRtPass();
     AddPass("data-flow-graph-to-fluid", new DataFlowGraphToFluidPass);
     if (!FLAGS_IA_output_storage_path.empty()) {
       AddPass("model-store-pass", new ModelStorePass);
     }
-    LOG(INFO)
-        << "-----------------------------------------------------------------";
   }
 
   std::string repr() const override { return "dfg-pass-manager"; }
@@ -101,19 +98,16 @@ class DfgPassManagerImpl final : public DfgPassManager {
 Analyzer::Analyzer() { Register("manager1", new DfgPassManagerImpl); }
 
 void Analyzer::Run(Argument* argument) {
+  std::vector<std::string> passes;
+  for (auto& pass : all_ir_passes_) {
+    if (!disabled_ir_passes_.count(pass)) {
+      passes.push_back(pass);
+      passes.push_back("graph_viz_pass");  // add graphviz for debug.
+    }
+  }
+  passes.push_back("graph_viz_pass");
   // Ugly support fluid-to-ir-pass
-  argument->Set(kFluidToIrPassesAttr,
-                new std::vector<std::string>({
-                    // Manual update the passes here.
-                    "graph_viz_pass",                              //
-                    "infer_clean_graph_pass", "graph_viz_pass",    //
-                    "attention_lstm_fuse_pass", "graph_viz_pass",  //
-                    "fc_lstm_fuse_pass", "graph_viz_pass",         //
-                    "mul_lstm_fuse_pass", "graph_viz_pass",        //
-                    "seq_concat_fc_fuse_pass", "graph_viz_pass",   //
-                    "fc_fuse_pass", "graph_viz_pass"               //
-
-                }));
+  argument->Set(kFluidToIrPassesAttr, new std::vector<std::string>(passes));
 
   for (auto& x : data_) {
     PADDLE_ENFORCE(x->Initialize(argument));
@@ -122,6 +116,11 @@ void Analyzer::Run(Argument* argument) {
   }
 }
 
+Analyzer& Analyzer::DisableIrPasses(const std::vector<std::string>& passes) {
+  disabled_ir_passes_.insert(passes.begin(), passes.end());
+  return *this;
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle
diff --git a/paddle/fluid/inference/analysis/analyzer.h b/paddle/fluid/inference/analysis/analyzer.h
index 2e107c82dd50d5cf22797f4c82e69d302514f955..3fdd2b9ec7537c891a04efb3ca9a1d45075ffa5e 100644
--- a/paddle/fluid/inference/analysis/analyzer.h
+++ b/paddle/fluid/inference/analysis/analyzer.h
@@ -36,16 +36,10 @@ limitations under the License. */
  */
 
 #include <gflags/gflags.h>
+#include "paddle/fluid/inference/analysis/flags.h"
 #include "paddle/fluid/inference/analysis/pass.h"
 #include "paddle/fluid/inference/analysis/pass_manager.h"
 
-// TODO(Superjomn) add a definition flag like PADDLE_WITH_TENSORRT and hide this
-// flag if not available.
-DECLARE_bool(IA_enable_tensorrt_subgraph_engine);
-DECLARE_string(IA_graphviz_log_root);
-DECLARE_string(IA_output_storage_path);
-DECLARE_bool(IA_enable_ir);
-
 namespace paddle {
 namespace inference {
 namespace analysis {
@@ -57,7 +51,26 @@ class Analyzer : public OrderedRegistry<PassManager> {
 
   void Run(Argument* argument);
 
+  Analyzer& DisableIrPasses(const std::vector<std::string>& passes);
+
   DISABLE_COPY_AND_ASSIGN(Analyzer);
+
+ private:
+  // All avaiable IR passes.
+  // The bigger fuse comes first, so that the small operators prefer to be
+  // merged in a larger fuse op. The small fusion will not break the pattern of
+  // larger fusion.
+  const std::vector<std::string> all_ir_passes_{{
+      // Manual update the passes here.
+      "infer_clean_graph_pass",    //
+      "attention_lstm_fuse_pass",  //
+      "fc_lstm_fuse_pass",         //
+      "mul_lstm_fuse_pass",        //
+      "seq_concat_fc_fuse_pass",   //
+      "fc_fuse_pass",              //
+  }};
+
+  std::unordered_set<std::string> disabled_ir_passes_;
 };
 
 }  // namespace analysis
diff --git a/paddle/fluid/inference/analysis/analyzer_tester.cc b/paddle/fluid/inference/analysis/analyzer_tester.cc
index 59e103e1179240a100d492a2475573c8188bebe7..94be6733f63488634ee1302e015a57c9a8892d84 100644
--- a/paddle/fluid/inference/analysis/analyzer_tester.cc
+++ b/paddle/fluid/inference/analysis/analyzer_tester.cc
@@ -271,17 +271,22 @@ void TestDituRNNPrediction(const std::string &model_path,
                            const std::string &data_path, int batch_size,
                            bool use_analysis, bool activate_ir,
                            int num_times = 1) {
-  NativeConfig config;
+  AnalysisConfig config;
   config.prog_file = FLAGS_infer_ditu_rnn_model + "/__model__";
   config.param_file = FLAGS_infer_ditu_rnn_model + "/param";
   config.use_gpu = false;
   config.device = 0;
   config.specify_input_name = true;
+  config.enable_ir_optim = activate_ir;
+  PADDLE_ENFORCE(config.ir_mode ==
+                 AnalysisConfig::IrPassMode::kExclude);  // default
+  config.ir_passes.clear();  // Do not exclude any pass.
 
   auto base_predictor =
       CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
   auto predictor =
-      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kAnalysis>(config);
+      CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
+          config);
   std::vector<PaddleTensor> input_slots;
   DataRecord data(data_path, batch_size);
   // Prepare inputs.
diff --git a/paddle/fluid/inference/analysis/flags.h b/paddle/fluid/inference/analysis/flags.h
new file mode 100644
index 0000000000000000000000000000000000000000..717e543f01dfa071865a5c14c0b7679e65239daf
--- /dev/null
+++ b/paddle/fluid/inference/analysis/flags.h
@@ -0,0 +1,22 @@
+// 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 <gflags/gflags.h>
+
+// TODO(Superjomn) add a definition flag like PADDLE_WITH_TENSORRT and hide this
+// flag if not available.
+DECLARE_bool(IA_enable_tensorrt_subgraph_engine);
+DECLARE_string(IA_graphviz_log_root);
+DECLARE_string(IA_output_storage_path);
+DECLARE_bool(IA_enable_ir);
diff --git a/paddle/fluid/inference/analysis/fluid_to_ir_pass.h b/paddle/fluid/inference/analysis/fluid_to_ir_pass.h
index 6731b1f759363eec5dd8645783212a72ace67b2f..3086085710d6e850ed27e82d2323690dfdd3ef19 100644
--- a/paddle/fluid/inference/analysis/fluid_to_ir_pass.h
+++ b/paddle/fluid/inference/analysis/fluid_to_ir_pass.h
@@ -15,6 +15,7 @@
 #pragma once
 
 #include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/inference/analysis/flags.h"
 #include "paddle/fluid/inference/analysis/ir_pass_manager.h"
 #include "paddle/fluid/inference/analysis/pass.h"
 
@@ -85,9 +86,11 @@ class FluidToIrPass final : public DataFlowGraphPass {
           new Scope *(&argument_->Get<Scope>(ir::kParamScopeAttr)));
     }
 
-    const auto &ir_passes_to_apply =
-        argument_->Get<std::vector<std::string>>(kFluidToIrPassesAttr);
-    ir_passes.Apply(ir_passes_to_apply);
+    if (FLAGS_IA_enable_ir) {
+      const auto &ir_passes_to_apply =
+          argument_->Get<std::vector<std::string>>(kFluidToIrPassesAttr);
+      ir_passes.Apply(ir_passes_to_apply);
+    }
 
     PADDLE_ENFORCE(argument_->main_dfg.get());
     argument_->main_dfg->Build(ir_passes.graph());
diff --git a/paddle/fluid/inference/analysis/test_text_classification.cc b/paddle/fluid/inference/analysis/test_text_classification.cc
new file mode 100644
index 0000000000000000000000000000000000000000..2913824f62301795aea967c22021b2af11f343c1
--- /dev/null
+++ b/paddle/fluid/inference/analysis/test_text_classification.cc
@@ -0,0 +1,109 @@
+// 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 <gflags/gflags.h>
+#include <glog/logging.h>  // use glog instead of PADDLE_ENFORCE to avoid importing other paddle header files.
+#include <gtest/gtest.h>
+#include "paddle/fluid/framework/ir/pass.h"
+#include "paddle/fluid/inference/analysis/analyzer.h"
+#include "paddle/fluid/inference/analysis/ut_helper.h"
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
+#include "paddle/fluid/inference/api/timer.h"
+
+DEFINE_string(infer_model, "", "Directory of the inference model.");
+DEFINE_string(infer_data, "", "Path of the dataset.");
+DEFINE_int32(batch_size, 1, "batch size.");
+DEFINE_int32(repeat, 1, "How many times to repeat run.");
+
+namespace paddle {
+
+template <typename T>
+std::string to_string(const std::vector<T> &vec) {
+  std::stringstream ss;
+  for (const auto &c : vec) {
+    ss << c << " ";
+  }
+  return ss.str();
+}
+
+void PrintTime(const double latency, const int bs, const int repeat) {
+  LOG(INFO) << "===========profile result===========";
+  LOG(INFO) << "batch_size: " << bs << ", repeat: " << repeat
+            << ", avg latency: " << latency / repeat << "ms";
+  LOG(INFO) << "=====================================";
+}
+
+void Main(int batch_size) {
+  // Three sequence inputs.
+  std::vector<PaddleTensor> input_slots(1);
+  // one batch starts
+  // data --
+  int64_t data0[] = {0, 1, 2};
+  for (auto &input : input_slots) {
+    input.data.Reset(data0, sizeof(data0));
+    input.shape = std::vector<int>({3, 1});
+    // dtype --
+    input.dtype = PaddleDType::INT64;
+    // LoD --
+    input.lod = std::vector<std::vector<size_t>>({{0, 3}});
+  }
+
+  // shape --
+  // Create Predictor --
+  AnalysisConfig config;
+  config.model_dir = FLAGS_infer_model;
+  config.use_gpu = false;
+  config.enable_ir_optim = true;
+  config.ir_passes.push_back("fc_lstm_fuse_pass");
+  auto predictor =
+      CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
+          config);
+
+  inference::Timer timer;
+  double sum = 0;
+  std::vector<PaddleTensor> output_slots;
+  for (int i = 0; i < FLAGS_repeat; i++) {
+    timer.tic();
+    CHECK(predictor->Run(input_slots, &output_slots));
+    sum += timer.toc();
+  }
+  PrintTime(sum, batch_size, FLAGS_repeat);
+
+  // Get output
+  LOG(INFO) << "get outputs " << output_slots.size();
+
+  for (auto &output : output_slots) {
+    LOG(INFO) << "output.shape: " << to_string(output.shape);
+    // no lod ?
+    CHECK_EQ(output.lod.size(), 0UL);
+    LOG(INFO) << "output.dtype: " << output.dtype;
+    std::stringstream ss;
+    for (int i = 0; i < 5; i++) {
+      ss << static_cast<float *>(output.data.data())[i] << " ";
+    }
+    LOG(INFO) << "output.data summary: " << ss.str();
+    // one batch ends
+  }
+}
+
+TEST(text_classification, basic) { Main(FLAGS_batch_size); }
+
+}  // namespace paddle
+
+USE_PASS(fc_fuse_pass);
+USE_PASS(seq_concat_fc_fuse_pass);
+USE_PASS(fc_lstm_fuse_pass);
+USE_PASS(graph_viz_pass);
+USE_PASS(infer_clean_graph_pass);
+USE_PASS(attention_lstm_fuse_pass);
diff --git a/paddle/fluid/inference/api/CMakeLists.txt b/paddle/fluid/inference/api/CMakeLists.txt
index 3a43c72e33b3d5d8910b554021bb1c6a626edd93..ea00bf364951b0a4304b380df492d00e84451136 100644
--- a/paddle/fluid/inference/api/CMakeLists.txt
+++ b/paddle/fluid/inference/api/CMakeLists.txt
@@ -44,7 +44,19 @@ function(inference_api_test TARGET_NAME)
 endfunction(inference_api_test)
 
 cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS lod_tensor)
-cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api)
+cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api
+          analysis
+          ir_pass_manager
+          pass
+          fc_fuse_pass
+          fc_lstm_fuse_pass
+          seq_concat_fc_fuse_pass
+          graph_viz_pass
+          infer_clean_graph_pass
+          graph_pattern_detector
+          infer_clean_graph_pass
+          attention_lstm_fuse_pass
+  )
 
 cc_test(test_paddle_inference_api
         SRCS api_tester.cc
diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
index e87abd2feeff7769eb223f83a7a28f5cb3337cdb..a8fa677202d8429c274a6e3fdfd18ef5d48620c2 100644
--- a/paddle/fluid/inference/api/analysis_predictor.cc
+++ b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -14,6 +14,8 @@
 
 #include "paddle/fluid/inference/api/analysis_predictor.h"
 #include <memory>
+#include <string>
+#include <vector>
 #include "paddle/fluid/framework/ir/fuse_pass_base.h"
 #include "paddle/fluid/framework/ir/pass.h"
 #include "paddle/fluid/framework/scope.h"
@@ -28,6 +30,8 @@ bool AnalysisPredictor::Init(
   VLOG(3) << "Predictor::init()";
   if (config_.use_gpu) {
     place_ = paddle::platform::CUDAPlace(config_.device);
+    LOG(WARNING) << "ir optimize only supports CPU currently";
+    config_.enable_ir_optim = false;
   } else {
     place_ = paddle::platform::CPUPlace();
   }
@@ -73,7 +77,7 @@ bool AnalysisPredictor::Init(
 
 void AnalysisPredictor::OptimizeInferenceProgram() {
   LOG(INFO) << "optimize begin";
-  FLAGS_IA_enable_ir = true;
+  FLAGS_IA_enable_ir = config_.enable_ir_optim;
   FLAGS_IA_enable_tensorrt_subgraph_engine = false;
   FLAGS_IA_output_storage_path = "";  // Don't output the model.
   // Analyze inference_program
@@ -90,24 +94,26 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
   }
   argument_.origin_program_desc.reset(
       new ProgramDesc(*inference_program_->Proto()));
-  Analyzer().Run(&argument_);
+  PADDLE_ENFORCE(config_.ir_mode == AnalysisConfig::IrPassMode::kExclude,
+                 "Only kExclude is supported yet.");
+  Analyzer().DisableIrPasses(config_.ir_passes).Run(&argument_);
+
   CHECK(argument_.transformed_program_desc);
   VLOG(5) << "to prepare executor";
-  // LOG(INFO) << "transformed_parogram_desc " <<
-  // argument.transformed_program_desc->DebugString();
   inference_program_.reset(
       new framework::ProgramDesc(*argument_.transformed_program_desc));
-  PADDLE_ENFORCE(argument_.Has(framework::ir::kParamScopeAttr));
-  // Update scope.
-  scope_.reset(
-      argument_.Release<framework::Scope>(framework::ir::kParamScopeAttr));
-  LOG(INFO) << "optimize end ==";
+  if (argument_.Has(framework::ir::kParamScopeAttr)) {
+    // Update scope.
+    scope_.reset(
+        argument_.Release<framework::Scope>(framework::ir::kParamScopeAttr));
+  }
+  LOG(INFO) << "== optimize end ==";
 }
 
 template <>
 std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
-    NativeConfig, PaddleEngineKind::kAnalysis>(const NativeConfig& config) {
-  VLOG(3) << "create NativePredictor";
+    AnalysisConfig, PaddleEngineKind::kAnalysis>(const AnalysisConfig& config) {
+  VLOG(3) << "create AnalysisConfig";
   if (config.use_gpu) {
     // 1. GPU memeroy
     PADDLE_ENFORCE_GT(
diff --git a/paddle/fluid/inference/api/analysis_predictor.h b/paddle/fluid/inference/api/analysis_predictor.h
index e32b6185f6044ab3577bde0a8f8dcf2391688aa8..e53925366e9214cd60422efe56884751297c15e5 100644
--- a/paddle/fluid/inference/api/analysis_predictor.h
+++ b/paddle/fluid/inference/api/analysis_predictor.h
@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <string>
+#include <vector>
 #include "paddle/fluid/inference/analysis/analyzer.h"
 #include "paddle/fluid/inference/api/api_impl.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
@@ -28,7 +30,7 @@ using framework::proto::ProgramDesc;
  */
 class AnalysisPredictor : public NativePaddlePredictor {
  public:
-  explicit AnalysisPredictor(const NativeConfig& config)
+  explicit AnalysisPredictor(const AnalysisConfig& config)
       : NativePaddlePredictor(config), config_(config) {}
 
   bool Init(const std::shared_ptr<framework::Scope>& parent_scope);
@@ -44,7 +46,7 @@ class AnalysisPredictor : public NativePaddlePredictor {
   Argument& analysis_argument() { return argument_; }
 
  private:
-  NativeConfig config_;
+  AnalysisConfig config_;
   Argument argument_;
 };
 
diff --git a/paddle/fluid/inference/api/api_impl.cc b/paddle/fluid/inference/api/api_impl.cc
index 38b11d9113e4b03f8365b969009f7a385a683a70..bd9b4b1a814f995e3979105f5b9830b95fd8ea7d 100644
--- a/paddle/fluid/inference/api/api_impl.cc
+++ b/paddle/fluid/inference/api/api_impl.cc
@@ -176,7 +176,8 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
                                     framework::Scope *scope) {
   VLOG(3) << "Predictor::set_feed";
   if (inputs.size() != feeds_.size()) {
-    LOG(ERROR) << "wrong feed input size.";
+    LOG(ERROR) << "wrong feed input size, need " << feeds_.size() << " but get "
+               << inputs.size();
     return false;
   }
   for (size_t i = 0; i < inputs.size(); ++i) {
diff --git a/paddle/fluid/inference/api/paddle_inference_api.h b/paddle/fluid/inference/api/paddle_inference_api.h
index 1baa64c249f291ec1bc874be5031abe6d4368274..995da11e4a30eca72a91a53d3293aa8b033b012b 100644
--- a/paddle/fluid/inference/api/paddle_inference_api.h
+++ b/paddle/fluid/inference/api/paddle_inference_api.h
@@ -150,6 +150,21 @@ struct TensorRTConfig : public NativeConfig {
   int workspace_size{1 << 30};
 };
 
+// NOTE WIP, not stable yet.
+struct AnalysisConfig : public NativeConfig {
+  //
+  enum class IrPassMode {
+    kSystem,   // Use system default passes, not customize.
+    kInclude,  // Specify the passes in `ir_passes`.
+    kExclude   // Specify the disabled passes in `ir_passes`.
+  };
+
+  bool enable_ir_optim = true;
+  IrPassMode ir_mode{IrPassMode::kExclude};
+  // attention lstm fuse works only on some specific models, disable as default.
+  std::vector<std::string> ir_passes{"attention_lstm_fuse_pass"};
+};
+
 // A factory to help create different predictors.
 //
 // FOR EXTENSION DEVELOPER:
diff --git a/paddle/fluid/operators/auc_op.cc b/paddle/fluid/operators/auc_op.cc
index 5edecd18e673da326ec119cf9a383f24f8045089..dfaa7456f917c1308984b361afed752f96ea6f59 100644
--- a/paddle/fluid/operators/auc_op.cc
+++ b/paddle/fluid/operators/auc_op.cc
@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/auc_op.h"
-#include <string>
 
 namespace paddle {
 namespace operators {
@@ -36,15 +35,12 @@ class AucOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(predict_height, label_height,
                       "Out and Label should have same height.");
 
-    int num_thres = ctx->Attrs().Get<int>("num_thresholds");
+    int num_pred_buckets = ctx->Attrs().Get<int>("num_thresholds") + 1;
 
     ctx->SetOutputDim("AUC", {1});
-    ctx->SetOutputDim("TPOut", {num_thres});
-    ctx->SetOutputDim("TNOut", {num_thres});
-    ctx->SetOutputDim("FPOut", {num_thres});
-    ctx->SetOutputDim("FNOut", {num_thres});
-
-    ctx->ShareLoD("Predict", /*->*/ "AUC");
+    ctx->SetOutputDim("BatchAUC", {1});
+    ctx->SetOutputDim("StatPosOut", {num_pred_buckets});
+    ctx->SetOutputDim("StatNegOut", {num_pred_buckets});
   }
 
  protected:
@@ -66,25 +62,24 @@ class AucOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("Label",
              "A 2D int tensor indicating the label of the training data. "
              "shape: [batch_size, 1]");
-    AddInput("TP", "True-Positive value.");
-    AddInput("FP", "False-Positive value.");
-    AddInput("TN", "True-Negative value.");
-    AddInput("FN", "False-Negative value.");
     // TODO(typhoonzero): support weight input
+    AddInput("StatPos", "Statistic value when label = 1");
+    AddInput("StatNeg", "Statistic value when label = 0");
+
     AddOutput("AUC",
               "A scalar representing the "
               "current area-under-the-curve.");
-    AddOutput("TPOut", "True-Positive value.");
-    AddOutput("FPOut", "False-Positive value.");
-    AddOutput("TNOut", "True-Negative value.");
-    AddOutput("FNOut", "False-Negative value.");
+    AddOutput("BatchAUC", "The AUC for current batch");
+    AddOutput("StatPosOut", "Statistic value when label = 1");
+    AddOutput("StatNegOut", "Statistic value when label = 0");
 
     AddAttr<std::string>("curve", "Curve type, can be 'ROC' or 'PR'.")
         .SetDefault("ROC");
+
     AddAttr<int>("num_thresholds",
                  "The number of thresholds to use when discretizing the"
                  " roc curve.")
-        .SetDefault(200);
+        .SetDefault((2 << 12) - 1);
 
     AddComment(R"DOC(
 Area Under The Curve (AUC) Operator.
diff --git a/paddle/fluid/operators/auc_op.h b/paddle/fluid/operators/auc_op.h
index 0a18585edb54a76aff5ae72ecc71e0eebb9f9361..fb0517d70635e090f8c5b59ff9d8420fc34c747b 100644
--- a/paddle/fluid/operators/auc_op.h
+++ b/paddle/fluid/operators/auc_op.h
@@ -13,9 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+
 #include <string>
 #include <vector>
-#include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 
 namespace paddle {
@@ -23,106 +23,85 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
-
 template <typename DeviceContext, typename T>
 class AucKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    auto* predict = ctx.Input<Tensor>("Predict");
-    auto* label = ctx.Input<Tensor>("Label");
-    auto* auc = ctx.Output<Tensor>("AUC");
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto *predict = ctx.Input<Tensor>("Predict");
+    auto *label = ctx.Input<Tensor>("Label");
+
+    std::string curve = ctx.Attr<std::string>("curve");
+    int num_thresholds = ctx.Attr<int>("num_thresholds");
+    int num_pred_buckets = num_thresholds + 1;
+
     // Only use output var for now, make sure it's persistable and
     // not cleaned up for each batch.
-    auto* true_positive = ctx.Output<Tensor>("TPOut");
-    auto* false_positive = ctx.Output<Tensor>("FPOut");
-    auto* true_negative = ctx.Output<Tensor>("TNOut");
-    auto* false_negative = ctx.Output<Tensor>("FNOut");
+    auto *auc = ctx.Output<Tensor>("AUC");
+    auto *stat_pos = ctx.Output<Tensor>("StatPosOut");
+    auto *stat_neg = ctx.Output<Tensor>("StatNegOut");
 
-    auto* auc_data = auc->mutable_data<double>(ctx.GetPlace());
+    auto *stat_pos_data = stat_pos->mutable_data<int64_t>(ctx.GetPlace());
+    auto *stat_neg_data = stat_neg->mutable_data<int64_t>(ctx.GetPlace());
+    calcAuc(ctx, label, predict, stat_pos_data, stat_neg_data, num_thresholds,
+            auc);
 
-    std::string curve = ctx.Attr<std::string>("curve");
-    int num_thresholds = ctx.Attr<int>("num_thresholds");
-    std::vector<double> thresholds_list;
-    thresholds_list.reserve(num_thresholds);
-    for (int i = 1; i < num_thresholds - 1; i++) {
-      thresholds_list[i] = static_cast<double>(i) / (num_thresholds - 1);
-    }
-    const double kEpsilon = 1e-7;
-    thresholds_list[0] = 0.0f - kEpsilon;
-    thresholds_list[num_thresholds - 1] = 1.0f + kEpsilon;
+    auto *batch_auc = ctx.Output<Tensor>("BatchAUC");
+    std::vector<int64_t> stat_pos_batch(num_pred_buckets, 0);
+    std::vector<int64_t> stat_neg_batch(num_pred_buckets, 0);
+    calcAuc(ctx, label, predict, stat_pos_batch.data(), stat_neg_batch.data(),
+            num_thresholds, batch_auc);
+  }
 
+ private:
+  inline static double trapezoidArea(double X1, double X2, double Y1,
+                                     double Y2) {
+    return (X1 > X2 ? (X1 - X2) : (X2 - X1)) * (Y1 + Y2) / 2.0;
+  }
+
+  inline static void calcAuc(const framework::ExecutionContext &ctx,
+                             const framework::Tensor *label,
+                             const framework::Tensor *predict,
+                             int64_t *stat_pos, int64_t *stat_neg,
+                             int num_thresholds,
+                             framework::Tensor *auc_tensor) {
     size_t batch_size = predict->dims()[0];
     size_t inference_width = predict->dims()[1];
+    const T *inference_data = predict->data<T>();
+    const auto *label_data = label->data<int64_t>();
+
+    auto *auc = auc_tensor->mutable_data<double>(ctx.GetPlace());
 
-    const T* inference_data = predict->data<T>();
-    const auto* label_data = label->data<int64_t>();
-
-    auto* tp_data = true_positive->mutable_data<int64_t>(ctx.GetPlace());
-    auto* fn_data = false_negative->mutable_data<int64_t>(ctx.GetPlace());
-    auto* tn_data = true_negative->mutable_data<int64_t>(ctx.GetPlace());
-    auto* fp_data = false_positive->mutable_data<int64_t>(ctx.GetPlace());
-
-    for (int idx_thresh = 0; idx_thresh < num_thresholds; idx_thresh++) {
-      // calculate TP, FN, TN, FP for current thresh
-      int64_t tp = 0, fn = 0, tn = 0, fp = 0;
-      for (size_t i = 0; i < batch_size; i++) {
-        // NOTE: label_data used as bool, labels > 0 will be treated as true.
-        if (label_data[i]) {
-          if (inference_data[i * inference_width + 1] >=
-              (thresholds_list[idx_thresh])) {
-            tp++;
-          } else {
-            fn++;
-          }
-        } else {
-          if (inference_data[i * inference_width + 1] >=
-              (thresholds_list[idx_thresh])) {
-            fp++;
-          } else {
-            tn++;
-          }
-        }
+    for (size_t i = 0; i < batch_size; i++) {
+      uint32_t binIdx = static_cast<uint32_t>(
+          inference_data[i * inference_width + 1] * num_thresholds);
+      if (label_data[i]) {
+        stat_pos[binIdx] += 1.0;
+      } else {
+        stat_neg[binIdx] += 1.0;
       }
-      // store rates
-      tp_data[idx_thresh] += tp;
-      fn_data[idx_thresh] += fn;
-      tn_data[idx_thresh] += tn;
-      fp_data[idx_thresh] += fp;
     }
-    // epsilon to avoid divide by zero.
-    double epsilon = 1e-6;
-    // Riemann sum to caculate auc.
-    Tensor tp_rate, fp_rate, rec_rate;
-    tp_rate.Resize({num_thresholds});
-    fp_rate.Resize({num_thresholds});
-    rec_rate.Resize({num_thresholds});
-    auto* tp_rate_data = tp_rate.mutable_data<double>(ctx.GetPlace());
-    auto* fp_rate_data = fp_rate.mutable_data<double>(ctx.GetPlace());
-    auto* rec_rate_data = rec_rate.mutable_data<double>(ctx.GetPlace());
-    for (int i = 0; i < num_thresholds; i++) {
-      tp_rate_data[i] = (static_cast<double>(tp_data[i]) + epsilon) /
-                        (tp_data[i] + fn_data[i] + epsilon);
-      fp_rate_data[i] =
-          static_cast<double>(fp_data[i]) / (fp_data[i] + tn_data[i] + epsilon);
-      rec_rate_data[i] = (static_cast<double>(tp_data[i]) + epsilon) /
-                         (tp_data[i] + fp_data[i] + epsilon);
+
+    *auc = 0.0f;
+
+    double totPos = 0.0;
+    double totNeg = 0.0;
+    double totPosPrev = 0.0;
+    double totNegPrev = 0.0;
+
+    int idx = num_thresholds;
+
+    while (idx >= 0) {
+      totPosPrev = totPos;
+      totNegPrev = totNeg;
+      totPos += stat_pos[idx];
+      totNeg += stat_neg[idx];
+      *auc += trapezoidArea(totNeg, totNegPrev, totPos, totPosPrev);
+
+      --idx;
     }
-    *auc_data = 0.0f;
-    if (curve == "ROC") {
-      for (int i = 0; i < num_thresholds - 1; i++) {
-        auto dx = fp_rate_data[i] - fp_rate_data[i + 1];
-        auto y = (tp_rate_data[i] + tp_rate_data[i + 1]) / 2.0f;
-        *auc_data = *auc_data + dx * y;
-      }
-    } else if (curve == "PR") {
-      for (int i = 1; i < num_thresholds; i++) {
-        auto dx = tp_rate_data[i] - tp_rate_data[i - 1];
-        auto y = (rec_rate_data[i] + rec_rate_data[i - 1]) / 2.0f;
-        *auc_data = *auc_data + dx * y;
-      }
+
+    if (totPos > 0.0 && totNeg > 0.0) {
+      *auc = *auc / totPos / totNeg;
     }
   }
 };
diff --git a/paddle/fluid/operators/lookup_table_op.h b/paddle/fluid/operators/lookup_table_op.h
index f5c10ced8305b64c6386c5051804f8c9a8f71802..58463dc4d6fd7cc3454de766814a947fee161070 100644
--- a/paddle/fluid/operators/lookup_table_op.h
+++ b/paddle/fluid/operators/lookup_table_op.h
@@ -57,7 +57,7 @@ class LookupTableKernel : public framework::OpKernel<T> {
           memset(output + i * row_width, 0, row_width * sizeof(T));
         } else {
           PADDLE_ENFORCE_LT(ids[i], row_number);
-          PADDLE_ENFORCE_GE(ids[i], 0);
+          PADDLE_ENFORCE_GE(ids[i], 0, "ids %d", i);
           memcpy(output + i * row_width, table + ids[i] * row_width,
                  row_width * sizeof(T));
         }
diff --git a/paddle/fluid/operators/rmsprop_op.cc b/paddle/fluid/operators/rmsprop_op.cc
index 919ebe48ca38040274bd2052b95ef96eccff4db6..2f773f222e50a440801b06a4fd997bf237b34772 100644
--- a/paddle/fluid/operators/rmsprop_op.cc
+++ b/paddle/fluid/operators/rmsprop_op.cc
@@ -36,9 +36,13 @@ class RmspropOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
                    "Output(param_out) of RmspropOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("MomentOut"),
-                   "Output(Momentum_out) of RmspropOp should not be null.");
+                   "Output(MomentOut) of RmspropOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("MeanSquareOut"),
                    "Output(MeanSquareOut) of RmspropOp should not be null.");
+    if (ctx->Attrs().Get<bool>("centered")) {
+      PADDLE_ENFORCE(ctx->HasOutput("MeanGradOut"),
+                     "Output(MeanGradOut) of RmspropOp should not be null.");
+    }
 
     auto param_dim = ctx->GetInputDim("Param");
     PADDLE_ENFORCE_EQ(
@@ -58,6 +62,9 @@ class RmspropOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("ParamOut", param_dim);
     ctx->SetOutputDim("MomentOut", param_dim);
     ctx->SetOutputDim("MeanSquareOut", param_dim);
+    if (ctx->Attrs().Get<bool>("centered")) {
+      ctx->SetOutputDim("MeanGradOut", param_dim);
+    }
   }
 };
 
@@ -70,6 +77,10 @@ class RmspropOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("MeanSquare",
              "(Tensor, default Tensor<float>)"
              " The mean square value that gets updated.");
+    AddInput("MeanGrad",
+             "(Tensor, default Tensor<float>)"
+             " The moving average of gradient")
+        .AsDispensable();
     AddInput("LearningRate",
              "(Tensor, default Tensor<float>) "
              "The learning rate should be a tensor of size 1.");
@@ -82,6 +93,8 @@ class RmspropOpMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("ParamOut", "(Tensor) Output updated parameter value.");
     AddOutput("MomentOut", "(Tensor) Output updated moment.");
     AddOutput("MeanSquareOut", "(Tensor) Output Mean squared updated value.");
+    AddOutput("MeanGradOut",
+              "(Tensor) Output moving average of gradient updated value.");
 
     AddAttr<float>("epsilon",
                    "(float, default 1e-10) Constant "
@@ -93,6 +106,8 @@ class RmspropOpMaker : public framework::OpProtoAndCheckerMaker {
         .SetDefault(0.9f);
     AddAttr<float>("momentum", "(float, default 0.0) Constant value.")
         .SetDefault(0.0f);
+    AddAttr<bool>("centered", "(bool, default false) use centered rmsprop.")
+        .SetDefault(false);
     AddComment(R"DOC(
 Rmsprop Optimizer. 
 
@@ -103,6 +118,14 @@ MomentOut = momentum * Moment +
 ParamOut = Param -  MomentOut
 $$
 
+if centered is true:
+
+mean_grad = decay * mean_square{t-1} + (1-decay) * gradient
+mean_square = decay * mean_square{t-1} + (1-decay) * gradient ** 2
+mom = momentum * mom{t-1} + learning_rate * g_t /
+    sqrt(mean_square - mean_grad**2 + epsilon)
+param -= mom
+
 The original slides that proposed Rmsprop: Slide 29 of
 http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf)
 
diff --git a/paddle/fluid/operators/rmsprop_op.h b/paddle/fluid/operators/rmsprop_op.h
index 12836f43bde47ac87eb0af33dea501593b659a5d..25ed32c5ebb2ff5be962ac1e3e38c970623d705c 100644
--- a/paddle/fluid/operators/rmsprop_op.h
+++ b/paddle/fluid/operators/rmsprop_op.h
@@ -41,6 +41,7 @@ class RmspropOpKernel : public framework::OpKernel<T> {
     float epsilon = ctx.Attr<float>("epsilon");
     float rho = ctx.Attr<float>("decay");
     float momentum = ctx.Attr<float>("momentum");
+    bool centered = ctx.Attr<bool>("centered");
 
     auto p = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Param"));
     auto ms = EigenVector<T>::Flatten(*ctx.Input<Tensor>("MeanSquare"));
@@ -53,12 +54,24 @@ class RmspropOpKernel : public framework::OpKernel<T> {
     auto ms_out = EigenVector<T>::Flatten(*mean_square_out);
     auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
 
-    Eigen::DSizes<int, 1> grad_dsize(grad->numel());
+    Eigen::DSizes<int, 1> grad_dsize(static_cast<int>(grad->numel()));
 
     ms_out.device(place) = rho * ms + (1 - rho) * g * g;
-    mom_out.device(place) =
-        momentum * mom +
-        lr.broadcast(grad_dsize) * g / (ms_out + epsilon).sqrt();
+    if (centered) {
+      auto mg = EigenVector<T>::Flatten(*ctx.Input<Tensor>("MeanGrad"));
+      auto* mean_grad_out = ctx.Output<Tensor>("MeanGradOut");
+      mean_grad_out->mutable_data<T>(ctx.GetPlace());
+      auto mg_out = EigenVector<T>::Flatten(*mean_grad_out);
+
+      mg_out.device(place) = rho * mg + (1 - rho) * g;
+      mom_out.device(place) = momentum * mom +
+                              lr.broadcast(grad_dsize) * g /
+                                  (ms_out - mg_out.square() + epsilon).sqrt();
+    } else {
+      mom_out.device(place) =
+          momentum * mom +
+          lr.broadcast(grad_dsize) * g / (ms_out + epsilon).sqrt();
+    }
     p_out.device(place) = p - mom_out;
   }
 };
diff --git a/python/paddle/fluid/layers/metric_op.py b/python/paddle/fluid/layers/metric_op.py
index 0182bbeb637ec7b6a341a4822a1cc5fb5aef077d..b1598bfec210474ae1e17f9f88e8b57aa80b8452 100644
--- a/python/paddle/fluid/layers/metric_op.py
+++ b/python/paddle/fluid/layers/metric_op.py
@@ -78,7 +78,7 @@ def accuracy(input, label, k=1, correct=None, total=None):
     return acc_out
 
 
-def auc(input, label, curve='ROC', num_thresholds=200, topk=1):
+def auc(input, label, curve='ROC', num_thresholds=2**12 - 1, topk=1):
     """
     **Area Under the Curve (AUC) Layer**
 
@@ -118,16 +118,14 @@ def auc(input, label, curve='ROC', num_thresholds=200, topk=1):
     """
     helper = LayerHelper("auc", **locals())
     auc_out = helper.create_tmp_variable(dtype="float64")
+    batch_auc_out = helper.create_tmp_variable(dtype="float64")
     # make tp, tn, fp, fn persistable, so that can accumulate all batches.
-    tp = helper.create_global_variable(
-        persistable=True, dtype='int64', shape=[num_thresholds])
-    tn = helper.create_global_variable(
-        persistable=True, dtype='int64', shape=[num_thresholds])
-    fp = helper.create_global_variable(
-        persistable=True, dtype='int64', shape=[num_thresholds])
-    fn = helper.create_global_variable(
-        persistable=True, dtype='int64', shape=[num_thresholds])
-    for var in [tp, tn, fp, fn]:
+    stat_pos = helper.create_global_variable(
+        persistable=True, dtype='int64', shape=[num_thresholds + 1])
+    stat_neg = helper.create_global_variable(
+        persistable=True, dtype='int64', shape=[num_thresholds + 1])
+
+    for var in [stat_pos, stat_neg]:
         helper.set_variable_initializer(
             var, Constant(
                 value=0.0, force_cpu=True))
@@ -137,18 +135,15 @@ def auc(input, label, curve='ROC', num_thresholds=200, topk=1):
         inputs={
             "Predict": [input],
             "Label": [label],
-            "TP": [tp],
-            "TN": [tn],
-            "FP": [fp],
-            "FN": [fn]
+            "StatPos": [stat_pos],
+            "StatNeg": [stat_neg]
         },
         attrs={"curve": curve,
                "num_thresholds": num_thresholds},
         outputs={
             "AUC": [auc_out],
-            "TPOut": [tp],
-            "TNOut": [tn],
-            "FPOut": [fp],
-            "FNOut": [fn]
+            "BatchAUC": [batch_auc_out],
+            "StatPosOut": [stat_pos],
+            "StatNegOut": [stat_neg]
         })
-    return auc_out, [tp, tn, fp, fn]
+    return auc_out, batch_auc_out, [stat_pos, stat_neg]
diff --git a/python/paddle/fluid/metrics.py b/python/paddle/fluid/metrics.py
index 592cb23eb9319658f8542ed5bc6ab3e95cfdb118..0c2800dcf35ed156b71625babea2724f520575e5 100644
--- a/python/paddle/fluid/metrics.py
+++ b/python/paddle/fluid/metrics.py
@@ -558,8 +558,6 @@ class Auc(MetricBase):
         name: metric name
         curve: Specifies the name of the curve to be computed, 'ROC' [default] or
           'PR' for the Precision-Recall-curve.
-        num_thresholds: The number of thresholds to use when discretizing the roc
-            curve.
 
     "NOTE: only implement the ROC curve type via Python now."
 
@@ -574,15 +572,14 @@ class Auc(MetricBase):
                 numpy_auc = metric.eval()
     """
 
-    def __init__(self, name, curve='ROC', num_thresholds=200):
+    def __init__(self, name, curve='ROC', num_thresholds=4095):
         super(Auc, self).__init__(name=name)
         self._curve = curve
         self._num_thresholds = num_thresholds
-        self._epsilon = 1e-6
-        self.tp_list = np.zeros((num_thresholds, ))
-        self.fn_list = np.zeros((num_thresholds, ))
-        self.tn_list = np.zeros((num_thresholds, ))
-        self.fp_list = np.zeros((num_thresholds, ))
+
+        _num_pred_buckets = num_thresholds + 1
+        self._stat_pos = [0] * _num_pred_buckets
+        self._stat_neg = [0] * _num_pred_buckets
 
     def update(self, preds, labels):
         if not _is_numpy_(labels):
@@ -590,41 +587,32 @@ class Auc(MetricBase):
         if not _is_numpy_(preds):
             raise ValueError("The 'predictions' must be a numpy ndarray.")
 
-        kepsilon = 1e-7  # to account for floating point imprecisions
-        thresholds = [(i + 1) * 1.0 / (self._num_thresholds - 1)
-                      for i in range(self._num_thresholds - 2)]
-        thresholds = [0.0 - kepsilon] + thresholds + [1.0 + kepsilon]
-
-        # calculate TP, FN, TN, FP count
-        for idx_thresh, thresh in enumerate(thresholds):
-            tp, fn, tn, fp = 0, 0, 0, 0
-            for i, lbl in enumerate(labels):
-                if lbl:
-                    if preds[i, 1] >= thresh:
-                        tp += 1
-                    else:
-                        fn += 1
-                else:
-                    if preds[i, 1] >= thresh:
-                        fp += 1
-                    else:
-                        tn += 1
-            self.tp_list[idx_thresh] += tp
-            self.fn_list[idx_thresh] += fn
-            self.tn_list[idx_thresh] += tn
-            self.fp_list[idx_thresh] += fp
+        for i, lbl in enumerate(labels):
+            value = preds[i, 1]
+            bin_idx = int(value * self._num_thresholds)
+            assert bin_idx <= self._num_thresholds
+            if lbl:
+                self._stat_pos[bin_idx] += 1.0
+            else:
+                self._stat_neg[bin_idx] += 1.0
+
+    @staticmethod
+    def trapezoid_area(x1, x2, y1, y2):
+        return abs(x1 - x2) * (y1 + y2) / 2.0
 
     def eval(self):
-        epsilon = self._epsilon
-        num_thresholds = self._num_thresholds
-        tpr = (self.tp_list.astype("float32") + epsilon) / (
-            self.tp_list + self.fn_list + epsilon)
-        fpr = self.fp_list.astype("float32") / (
-            self.fp_list + self.tn_list + epsilon)
-        rec = (self.tp_list.astype("float32") + epsilon) / (
-            self.tp_list + self.fp_list + epsilon)
-
-        x = fpr[:num_thresholds - 1] - fpr[1:]
-        y = (tpr[:num_thresholds - 1] + tpr[1:]) / 2.0
-        auc_value = np.sum(x * y)
-        return auc_value
+        tot_pos = 0.0
+        tot_neg = 0.0
+        auc = 0.0
+
+        idx = self._num_thresholds
+        while idx >= 0:
+            tot_pos_prev = tot_pos
+            tot_neg_prev = tot_neg
+            tot_pos += self._stat_pos[idx]
+            tot_neg += self._stat_neg[idx]
+            auc += self.trapezoid_area(tot_neg, tot_neg_prev, tot_pos,
+                                       tot_pos_prev)
+            idx -= 1
+
+        return auc / tot_pos / tot_neg if tot_pos > 0.0 and tot_neg > 0.0 else 0.0
diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py
index 33d6311b9717c66f0d6782eb6b3e348cd4c02a69..215f0cf2fc5ab4fbd06719ac4790a01dd00080eb 100644
--- a/python/paddle/fluid/optimizer.py
+++ b/python/paddle/fluid/optimizer.py
@@ -897,7 +897,20 @@ class RMSPropOptimizer(Optimizer):
 
         r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
 
-        v(w, t) & = \\beta v(w, t-1) + \\frac{\\eta} {\\sqrt{v(w,t) +
+        v(w, t) & = \\beta v(w, t-1) + \\frac{\\eta} {\\sqrt{r(w,t) +
+            \\epsilon}} \\nabla Q_{i}(w)
+
+        w & = w - v(w, t)
+
+    if centered is True:
+
+    ..  math::
+
+        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
+
+        g(w, t) & = \\rho g(w, t-1) + (1 - \\rho)\\nabla Q_{i}(w)
+
+        v(w, t) & = \\beta v(w, t-1) + \\frac{\\eta} {\\sqrt{r(w,t) - (g(w, t))^2 +
             \\epsilon}} \\nabla Q_{i}(w)
 
         w & = w - v(w, t)
@@ -915,6 +928,10 @@ class RMSPropOptimizer(Optimizer):
             avoid division by zero, set 1e-6 by default.
         momentum(float): :math:`\\beta` in equation is the momentum term,
             set 0.0 by default.
+        centered(bool): If True, gradients are normalized by the estimated variance of
+            the gradient; if False, by the uncentered second moment. Setting this to
+            True may help with training, but is slightly more expensive in terms of
+            computation and memory. Defaults to False.
 
     Raises:
         ValueError: If learning_rate, rho, epsilon, momentum are None.
@@ -928,12 +945,14 @@ class RMSPropOptimizer(Optimizer):
 
     _momentum_acc_str = "momentum"
     _mean_square_acc_str = "mean_square"
+    _mean_grad_acc_str = "mean_grad"
 
     def __init__(self,
                  learning_rate,
                  rho=0.95,
                  epsilon=1.0e-6,
                  momentum=0.0,
+                 centered=False,
                  **kwargs):
         super(RMSPropOptimizer, self).__init__(
             learning_rate=learning_rate, **kwargs)
@@ -950,6 +969,7 @@ class RMSPropOptimizer(Optimizer):
         self._rho = rho
         self._epsilon = epsilon
         self._momentum = momentum
+        self._centered = centered
 
     def _create_accumulators(self, block, parameters):
         if not isinstance(block, framework.Block):
@@ -958,6 +978,7 @@ class RMSPropOptimizer(Optimizer):
         for p in parameters:
             self._add_accumulator(self._momentum_acc_str, p)
             self._add_accumulator(self._mean_square_acc_str, p)
+            self._add_accumulator(self._mean_grad_acc_str, p)
 
     def _append_optimize_op(self, block, param_and_grad):
         if not isinstance(block, framework.Block):
@@ -967,6 +988,8 @@ class RMSPropOptimizer(Optimizer):
                                              param_and_grad[0])
         mean_square_acc = self._get_accumulator(self._mean_square_acc_str,
                                                 param_and_grad[0])
+        mean_grad_acc = self._get_accumulator(self._mean_grad_acc_str,
+                                              param_and_grad[0])
         rmsprop_op = block.append_op(
             type=self.type,
             inputs={
@@ -974,17 +997,20 @@ class RMSPropOptimizer(Optimizer):
                 "Grad": param_and_grad[1],
                 "Moment": momentum_acc,
                 "MeanSquare": mean_square_acc,
+                "MeanGrad": mean_grad_acc,
                 "LearningRate": self._create_param_lr(param_and_grad),
             },
             outputs={
                 "ParamOut": param_and_grad[0],
                 "MomentOut": momentum_acc,
-                "MeanSquareOut": mean_square_acc
+                "MeanSquareOut": mean_square_acc,
+                "MeanGradOut": mean_grad_acc
             },
             attrs={
                 "epsilon": self._epsilon,
                 "decay": self._rho,
-                "momentum": self._momentum
+                "momentum": self._momentum,
+                "centered": self._centered
             })
 
         return rmsprop_op
diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py
index 20f1a37a426e9697048d636bf738c9056213e5f6..868df5d79424302700d9f82083cd7e4320c707ea 100644
--- a/python/paddle/fluid/tests/unittests/op_test.py
+++ b/python/paddle/fluid/tests/unittests/op_test.py
@@ -291,7 +291,7 @@ class OpTest(unittest.TestCase):
                             return_numpy=False)
         return outs, fetch_list
 
-    def check_output_with_place(self, place, atol):
+    def check_output_with_place(self, place, atol, equal_nan=False):
         outs, fetch_list = self._calc_output(place)
         for out_name, out_dup in Operator.get_op_outputs(self.op_type):
             if out_name not in self.outputs:
@@ -321,7 +321,7 @@ class OpTest(unittest.TestCase):
                         if isinstance(expect, tuple) else expect
                     self.assertTrue(
                         np.allclose(
-                            actual_t, expect_t, atol=atol),
+                            actual_t, expect_t, atol=atol, equal_nan=equal_nan),
                         "Output (" + sub_out_name + ") has diff at " +
                         str(place))
                     if isinstance(expect, tuple):
@@ -337,7 +337,7 @@ class OpTest(unittest.TestCase):
                 expect_t = expect[0] if isinstance(expect, tuple) else expect
                 self.assertTrue(
                     np.allclose(
-                        actual_t, expect_t, atol=atol),
+                        actual_t, expect_t, atol=atol, equal_nan=equal_nan),
                     "Output (" + out_name + ") has diff at " + str(place) +
                     "\nExpect " + str(expect_t) + "\n" + "But Got" +
                     str(actual_t))
@@ -360,10 +360,10 @@ class OpTest(unittest.TestCase):
             places.append(core.CUDAPlace(0))
         return places
 
-    def check_output(self, atol=1e-5):
+    def check_output(self, atol=1e-5, equal_nan=False):
         places = self._get_places()
         for place in places:
-            self.check_output_with_place(place, atol)
+            self.check_output_with_place(place, atol, equal_nan)
 
     def check_output_customized(self, checker):
         places = self._get_places()
diff --git a/python/paddle/fluid/tests/unittests/test_auc_op.py b/python/paddle/fluid/tests/unittests/test_auc_op.py
index 5393a17e674a3cad6d705a1ff7a45320e644af94..1de4a9d016a177944253d12094722d3a05614be2 100644
--- a/python/paddle/fluid/tests/unittests/test_auc_op.py
+++ b/python/paddle/fluid/tests/unittests/test_auc_op.py
@@ -26,18 +26,15 @@ class TestAucOp(OpTest):
         pred = np.random.random((128, 2)).astype("float32")
         labels = np.random.randint(0, 2, (128, 1))
         num_thresholds = 200
-        tp = np.zeros((num_thresholds, )).astype("int64")
-        tn = np.zeros((num_thresholds, )).astype("int64")
-        fp = np.zeros((num_thresholds, )).astype("int64")
-        fn = np.zeros((num_thresholds, )).astype("int64")
+
+        stat_pos = np.zeros((num_thresholds + 1, )).astype("int64")
+        stat_neg = np.zeros((num_thresholds + 1, )).astype("int64")
 
         self.inputs = {
             'Predict': pred,
             'Label': labels,
-            'TP': tp,
-            'TN': tn,
-            'FP': fp,
-            'FN': fn
+            "StatPos": stat_pos,
+            "StatNeg": stat_neg
         }
         self.attrs = {'curve': 'ROC', 'num_thresholds': num_thresholds}
 
@@ -47,11 +44,10 @@ class TestAucOp(OpTest):
         python_auc.update(pred, labels)
 
         self.outputs = {
-            'AUC': python_auc.eval(),
-            'TPOut': python_auc.tp_list,
-            'FNOut': python_auc.fn_list,
-            'TNOut': python_auc.tn_list,
-            'FPOut': python_auc.fp_list
+            'AUC': np.array(python_auc.eval()),
+            'BatchAUC': np.array(python_auc.eval()),
+            'StatPosOut': np.array(python_auc._stat_pos),
+            'StatNegOut': np.array(python_auc._stat_neg)
         }
 
     def test_check_output(self):
diff --git a/python/paddle/fluid/tests/unittests/test_rmsprop_op.py b/python/paddle/fluid/tests/unittests/test_rmsprop_op.py
index 3d4623c74d9a307b12ab6d72ad0b4d2dae938720..70848e4e2239e2be160bb0c1a28a5aecd01a87dc 100644
--- a/python/paddle/fluid/tests/unittests/test_rmsprop_op.py
+++ b/python/paddle/fluid/tests/unittests/test_rmsprop_op.py
@@ -15,90 +15,164 @@
 from __future__ import print_function
 
 import unittest
+
 import numpy as np
-from op_test import OpTest
-
-
-class TestRmspropOp1(OpTest):
-    ''' Test RMSProp with explicit inputs
-    '''
-
-    def setUp(self):
-        self.op_type = "rmsprop"
-
-        param = np.random.random((123, 321)).astype("float32")
-        mean_square = np.random.random((123, 321)).astype("float32")
-        learning_rate = np.array([0.01]).astype("float32")
-        grad = np.random.random((123, 321)).astype("float32")
-        moment = np.zeros((123, 321)).astype("float32")
-
-        epsilon = 1e-6
-        decay = 0.9
-        momentum = 0.0
-
-        self.inputs = {
-            'Param': param,
-            'MeanSquare': mean_square,
-            'LearningRate': learning_rate,
-            'Grad': grad,
-            'Moment': moment,
-        }
-
-        self.attrs = {'epsilon': epsilon, 'decay': decay, 'momentum': momentum}
-
-        ms_out = decay * mean_square + (1 - decay) * grad * grad
-        moment_out = momentum * moment + \
-            learning_rate * grad / np.sqrt(ms_out + epsilon)
-        param_out = param - moment_out
-
-        self.outputs = {
-            'ParamOut': param_out,
-            'MomentOut': moment_out,
-            'MeanSquareOut': ms_out
-        }
-
-    def test_check_output(self):
-        self.check_output()
-
-
-class TestRmspropOp2(OpTest):
-    '''Test RMSProp with default values for attributes
-    '''
-
-    def setUp(self):
-        self.op_type = "rmsprop"
-
-        param = np.random.random((123, 321)).astype("float32")
-        mean_square = np.random.random((123, 321)).astype("float32")
-        learning_rate = np.array([0.01]).astype("float32")
-        grad = np.random.random((123, 321)).astype("float32")
-        moment = np.zeros((123, 321)).astype("float32")
-
-        epsilon = 1.0e-10
-        decay = 0.9
-        momentum = 0.0
-
-        self.inputs = {
-            'Param': param,
-            'MeanSquare': mean_square,
-            'LearningRate': learning_rate,
-            'Grad': grad,
-            'Moment': moment,
-        }
-
-        ms_out = decay * mean_square + (1 - decay) * grad * grad
-        moment_out = momentum * moment + \
-            learning_rate * grad / np.sqrt(ms_out + epsilon)
-        param_out = param - moment_out
-
-        self.outputs = {
-            'ParamOut': param_out,
-            'MomentOut': moment_out,
-            'MeanSquareOut': ms_out
-        }
-
-    def test_check_output(self):
-        self.check_output()
+import paddle.fluid.core as core
+from paddle.fluid.op import Operator
+
+
+class TestBase(unittest.TestCase):
+    def setup(self, centered, epsilon=1e-6):
+        np.random.seed(5)  # fix seed
+
+        self.param_name = "param"
+        self.param = np.random.random((123, 321)).astype("float32")
+
+        self.mean_square_name = "mean_square"
+        self.mean_square = np.random.random((123, 321)).astype("float32")
+
+        self.mean_grad_name = "mean_grad"
+        self.mean_grad = np.random.random((123, 321)).astype("float32")
+
+        self.lr_name = "lr"
+        self.learning_rate = np.array([0.01]).astype("float32")
+
+        self.grad_name = "grad"
+        self.grad = np.random.random((123, 321)).astype("float32")
+
+        self.moment_name = "moment"
+        self.moment = np.zeros((123, 321)).astype("float32")
+
+        self.epsilon = epsilon
+        self.decay = 0.9
+        self.momentum = 0.0
+        self.centered = centered
+
+        self.ms_out = self.decay * self.mean_square + (1 - self.decay
+                                                       ) * self.grad * self.grad
+        if centered:
+            self.mg_out = self.decay * self.mean_grad + (1 - self.decay
+                                                         ) * self.grad
+            self.moment_out = self.momentum * self.moment + \
+                              self.learning_rate * self.grad / np.sqrt(self.ms_out - np.square(self.mg_out) + self.epsilon)
+        else:
+            self.moment_out = self.momentum * self.moment + \
+                              self.learning_rate * self.grad / np.sqrt(self.ms_out + self.epsilon)
+
+        self.param_out = self.param - self.moment_out
+
+    def check(self,
+              actual_t,
+              expect_t,
+              place,
+              out_name,
+              atol=1e-5,
+              equal_nan=False):
+        self.assertTrue(
+            np.allclose(
+                actual_t, expect_t, atol=atol, equal_nan=equal_nan),
+            "Output (" + out_name + ") has diff at " + str(place) + "\nExpect "
+            + str(expect_t) + "\n" + "But Got" + str(actual_t))
+
+
+class TestRmspropOp(TestBase):
+    def check_with_place(self, place, centered, epsilon):
+        self.setup(centered, epsilon)
+        scope = core.Scope()
+
+        # create and initialize Param Variable
+        param = scope.var(self.param_name).get_tensor()
+        param.set(self.param, place)
+
+        mean_square = scope.var(self.mean_square_name).get_tensor()
+        mean_square.set(self.mean_square, place)
+
+        lr = scope.var(self.lr_name).get_tensor()
+        lr.set(self.learning_rate, place)
+
+        grad = scope.var(self.grad_name).get_tensor()
+        grad.set(self.grad, place)
+
+        moment = scope.var(self.moment_name).get_tensor()
+        moment.set(self.moment, place)
+
+        # create and run sgd operator
+
+        if self.centered:
+            mean_grad = scope.var(self.mean_grad_name).get_tensor()
+            mean_grad.set(self.mean_grad, place)
+
+            rmsprop_op = Operator(
+                "rmsprop",
+                Param=self.param_name,
+                Grad=self.grad_name,
+                MeanSquare=self.mean_square_name,
+                MeanGrad=self.mean_grad_name,
+                Moment=self.moment_name,
+                LearningRate=self.lr_name,
+                ParamOut=self.param_name,
+                MeanSquareOut=self.mean_square_name,
+                MomentOut=self.moment_name,
+                MeanGradOut=self.mean_grad_name,
+                epsilon=self.epsilon,
+                decay=self.decay,
+                momentum=self.momentum,
+                centered=True)
+        else:
+            rmsprop_op = Operator(
+                "rmsprop",
+                Param=self.param_name,
+                Grad=self.grad_name,
+                MeanSquare=self.mean_square_name,
+                Moment=self.moment_name,
+                LearningRate=self.lr_name,
+                ParamOut=self.param_name,
+                MeanSquareOut=self.mean_square_name,
+                MomentOut=self.moment_name,
+                epsilon=self.epsilon,
+                decay=self.decay,
+                momentum=self.momentum,
+                centered=False)
+
+        rmsprop_op.run(scope, place)
+
+        atol = 1e-5
+        equal_nan = False
+
+        if self.centered:
+            atol = 1e-3
+            equal_nan = True
+
+        self.check(
+            np.array(mean_square), self.ms_out, place, self.mean_square_name)
+        self.check(
+            np.array(moment),
+            self.moment_out,
+            place,
+            self.moment_name,
+            atol=atol,
+            equal_nan=equal_nan)
+        self.check(
+            np.array(param),
+            self.param_out,
+            place,
+            self.param_name,
+            atol=atol,
+            equal_nan=equal_nan)
+
+        if self.centered:
+            self.check(
+                np.array(mean_grad), self.mg_out, place, self.mean_grad_name)
+
+    def test_rmsprop(self):
+        places = [core.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(core.CUDAPlace(0))
+        for place in places:
+            self.check_with_place(place, False, 1e-6)
+            self.check_with_place(place, False, 1e-10)
+            self.check_with_place(place, True, 1e-6)
+            self.check_with_place(place, True, 1e-10)
 
 
 if __name__ == "__main__":