diff --git a/CMakeLists.txt b/CMakeLists.txt
index a9382f9697e5c39c7e7195b94fc5332d39dc18a9..f60846e98aa9ca36bd6bd68cccdda6e3d2ff616a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,6 +1,7 @@
 cmake_minimum_required(VERSION 3.0)
 project(paddle-mobile)
-add_definitions(-DPADDLE_MOBILE_DEBUG="true")
+add_definitions(-DPADDLE_MOBILE_DEBUG)
+add_definitions(-DENABLE_EXCEPTION)
 
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
 set(CMAKE_BUILD_TYPE RelWithDebInfo)
diff --git a/scripts/push2android.sh b/scripts/push2android.sh
new file mode 100644
index 0000000000000000000000000000000000000000..44b0ee32e99ccddf5cc6060882dc37158c149693
--- /dev/null
+++ b/scripts/push2android.sh
@@ -0,0 +1,14 @@
+#!/usr/bin/env sh
+
+push_fn () {
+MODELS_PATH="../test/models/*"
+EXE_FILE="../test/build/*"
+EXE_DIR="data/local/tmp/bin"
+MODELS_DIR="data/local/tmp/models"
+LIB_PATH="../build/release/arm-v7a/build/*"
+adb push ${EXE_FILE} ${EXE_DIR}
+adb push ${LIB_PATH} ${EXE_DIR}
+adb push ${MODELS_PATH} ${MODELS_DIR}
+echo "test files sync completed"
+}
+push_fn
diff --git a/src/common/enforce.h b/src/common/enforce.h
index abd6217fbeb62a09a7d4a9fe9b6d85d0ab1cf2d8..52bda2258a00c7444762fe8297380c1c7752dd42 100644
--- a/src/common/enforce.h
+++ b/src/common/enforce.h
@@ -14,7 +14,7 @@ limitations under the License. */
 
 #pragma once
 
-#ifdef PADDLE_MOBILE_DEBUG
+#ifdef ENABLE_EXCEPTION
 #include <stdio.h>
 #include <exception>
 #include <sstream>
@@ -25,7 +25,7 @@ limitations under the License. */
 
 namespace paddle_mobile {
 
-#ifdef PADDLE_MOBILE_DEBUG
+#ifdef ENABLE_EXCEPTION
 struct PaddleMobileException : public std::exception {
   const std::string exception_prefix = "paddle mobile C++ Exception: \n";
   std::string message;
@@ -64,7 +64,7 @@ struct PaddleMobileException : public std::exception {
   }
 #else
 #define PADDLE_MOBILE_THROW_EXCEPTION(...)
-#define PADDLE_MOBILE_ASSERT(stat, ...)
+#define PADDLE_MOBILE_ENFORCE(stat, ...)
 #endif
 
 }  // namespace paddle_mobile
diff --git a/src/framework/operator.cpp b/src/framework/operator.cpp
index 808002d4c8f3193744ef68c1db881a787d19b133..46feb97cb879bb092e1a8f6955f7f50712f9ea3b 100644
--- a/src/framework/operator.cpp
+++ b/src/framework/operator.cpp
@@ -28,18 +28,6 @@ vector<string> OperatorBase<Dtype>::GetOutKeys() const {
   return it->second.second;
 }
 
-template <typename T>
-static T *GetVarValue(const string &key, const VariableNameMap &var_map,
-                      const Scope &scope) {
-  auto var_vec = var_map.at(key);
-  if (!var_vec.empty()) {
-    auto var = scope.FindVar(var_vec[0]);
-    return var->GetMutable<T>();
-  } else {
-    return nullptr;
-  }
-}
-
 template <typename Dtype>
 OperatorBase<Dtype>::OperatorBase(const std::string &type,
                                   const VariableNameMap &inputs,
@@ -60,7 +48,7 @@ void OperatorBase<Dtype>::CheckAllInputOutputSet() const {}
 template <typename Dtype>
 void OperatorBase<Dtype>::Run() const {
   RunImpl();
-#ifdef PADDLE_MOBILE_DEBUG
+#if (PADDLE_MOBILE_DEBUG)
   vector<string> output_keys = GetOutKeys();
   for (const auto key : output_keys) {
     Tensor *out_ = GetVarValue<framework::LoDTensor>(key, outputs_, *scope_);
diff --git a/src/framework/operator.h b/src/framework/operator.h
index 6194e5dcfff2c0320318d3e30e5c8204cd71a749..2de3a9536702ee17daabae969d81e979fc66b0c8 100644
--- a/src/framework/operator.h
+++ b/src/framework/operator.h
@@ -39,6 +39,18 @@ namespace framework {
 using std::string;
 using std::vector;
 
+template <typename T>
+static T *GetVarValue(const string &key, const VariableNameMap &var_map,
+                      const Scope &scope) {
+  auto var_vec = var_map.at(key);
+  if (!var_vec.empty()) {
+    auto var = scope.FindVar(var_vec[0]);
+    return var->GetMutable<T>();
+  } else {
+    return nullptr;
+  }
+}
+
 template <typename Dtype>
 class OperatorBase : PaddleMobileObject {
  public:
diff --git a/src/io.cpp b/src/io.cpp
index b8350a81118d8187552854f8ed7cab3a13dab8f6..ac89106e4988b35d54a59991009a98739e50d024 100644
--- a/src/io.cpp
+++ b/src/io.cpp
@@ -371,31 +371,47 @@ void Executor<Dtype, P>::InitMemory() {
 }
 
 template <typename Dtype, Precision P>
-void Executor<Dtype, P>::Predict(const framework::Tensor &t, int block_id) {
+std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
+    const framework::Tensor &t) {
   framework::Variable *g_feed_value = program_.scope->Var("feed");
   framework::Tensor *feed_tensor =
       g_feed_value->GetMutable<framework::LoDTensor>();
   feed_tensor->Resize(t.dims());
   feed_tensor->ShareDataWith(t);
   std::shared_ptr<framework::BlockDesc> to_predict_block =
-      to_predict_program_->Block(block_id);
+      to_predict_program_->Block(0);
   for (int j = 0; j < ops_of_block_[*to_predict_block.get()].size(); ++j) {
     auto op = ops_of_block_[*to_predict_block.get()][j];
     op->Run();
   }
+  auto ops = ops_of_block_[*to_predict_program_->Block(0)];
+  auto last_op = ops.rbegin();
+  auto output_map = (*last_op)->Outputs();
+  std::vector<std::string> out_keys = (*last_op)->GetOutKeys();
+  PADDLE_MOBILE_ENFORCE(out_keys.size() > 0, "the last op contains no output");
+  framework::LoDTensor *output_tensor =
+      framework::GetVarValue<framework::LoDTensor>(out_keys[0], output_map,
+                                                   *(program_.scope));
+  return std::shared_ptr<framework::Tensor>(output_tensor);
+}
+template <typename Dtype, Precision P>
+std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
+    const framework::Tensor &t, int block_id) {
+  return Predict(t);
 }
 
 template <typename Dtype, Precision P>
 std::vector<typename Executor<Dtype, P>::Ptype> Executor<Dtype, P>::Predict(
     const std::vector<Ptype> &input, const std::vector<int64_t> &dims) {
   framework::Tensor tensor(input, framework::make_ddim(dims));
-
-  Predict(tensor, 0);
-
-  framework::Variable *g_feed_value = program_.scope->Var("col");
-  auto feed_tensor = g_feed_value->GetMutable<framework::Tensor>();
-
-  return {};
+  std::shared_ptr<framework::Tensor> output_tensor = Predict(tensor, 0);
+  Executor<Dtype, P>::Ptype *output_ptr =
+      output_tensor->data<typename Executor<Dtype, P>::Ptype>();
+  std::vector<typename Executor<Dtype, P>::Ptype> result_vector;
+  for (int j = 0; j < output_tensor->numel(); ++j) {
+    result_vector.push_back(output_ptr[j]);
+  }
+  return result_vector;
 }
 
 template class Executor<CPU, Precision::FP32>;
diff --git a/src/io.h b/src/io.h
index de2d359bf58d1ad328defd2f51e87e2d6bfe6295..ae99197baa97c07d2a883f8721d533b85ab7873a 100644
--- a/src/io.h
+++ b/src/io.h
@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <memory.h>
+#include <map>
 #include <string>
 #include <vector>
 
@@ -44,24 +45,25 @@ class Executor {
  public:
   typedef typename PrecisionTrait<P>::ptype Ptype;
 
-  Executor() = default;
-
   Executor(const framework::Program<Dtype> p, int batch_size = 1,
            bool use_optimize = true);
 
-  //  std::shared_ptr<framework::Tensor> Predict(framework::Tensor &t);
+  std::shared_ptr<framework::Tensor> Predict(const framework::Tensor &t);
 
   std::vector<Ptype> Predict(const std::vector<Ptype> &input,
                              const std::vector<int64_t> &dims);
 
  protected:
+  Executor() = default;
+
   void InitMemory();
   void LoadMemory(const framework::VarDesc var_desc,
                   framework::LoDTensor *tensor, const std::string &file_path);
   framework::Program<Dtype> program_;
   int batch_size_ = 1;
   std::shared_ptr<framework::ProgramDesc> to_predict_program_;
-  void Predict(const framework::Tensor &t, int block_id);
+  std::shared_ptr<framework::Tensor> Predict(const framework::Tensor &t,
+                                             int block_id);
   std::map<framework::BlockDesc,
            std::vector<std::shared_ptr<framework::OperatorBase<Dtype>>>>
       ops_of_block_;
diff --git a/test/net/test_googlenet.cpp b/test/net/test_googlenet.cpp
index 139579e9116651c15764997d962b7d2622532146..0640af890cf9857e1dda7eeaaec593b7c613fe8a 100644
--- a/test/net/test_googlenet.cpp
+++ b/test/net/test_googlenet.cpp
@@ -18,20 +18,17 @@ limitations under the License. */
 
 int main() {
   paddle_mobile::Loader<paddle_mobile::CPU> loader;
-  //  ../../../test/models/googlenet
-  //  ../../../test/models/mobilenet
   auto time1 = time();
   auto program = loader.Load(g_googlenet, false);
   auto time2 = time();
-  DLOG << "load cost :" << time_diff(time1, time1) << "ms";
+  DLOG << "load cost :" << time_diff(time1, time2) << "ms\n";
   paddle_mobile::Executor<paddle_mobile::CPU> executor(program, 1, false);
-
   std::vector<float> input;
   std::vector<int64_t> dims{1, 3, 224, 224};
   GetInput<float>(g_test_image_1x3x224x224, &input, dims);
   auto time3 = time();
   executor.Predict(input, dims);
   auto time4 = time();
-  DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
+  DLOG << "predict cost :" << time_diff(time3, time4) << "ms\n";
   return 0;
 }