diff --git a/paddle/fluid/inference/api/api_impl.cc b/paddle/fluid/inference/api/api_impl.cc
index c57fc64bb6bfeebc7935f19d0e977e8fccd4c9a0..dca4386b21b4a064c21b52218682321258f368c4 100644
--- a/paddle/fluid/inference/api/api_impl.cc
+++ b/paddle/fluid/inference/api/api_impl.cc
@@ -22,6 +22,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/feed_fetch_method.h"
 #include "paddle/fluid/inference/api/api_impl.h"
+#include "paddle/fluid/inference/api/helper.h"
 #include "paddle/fluid/inference/api/timer.h"
 #include "paddle/fluid/platform/profiler.h"
 
@@ -215,57 +216,20 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
 template <typename T>
 void NativePaddlePredictor::GetFetchOne(const framework::LoDTensor &fetch,
                                         PaddleTensor *output) {
-  std::vector<int> shape;
-  auto dims_i = fetch.dims();
-  auto lod = fetch.lod();
-  const T *output_ptr = fetch.data<T>();
-  auto num = fetch.numel();
-  std::vector<T> data;
-  if (0 == lod.size()) {
-    std::copy(output_ptr, output_ptr + num, std::back_inserter(data));
-    for (int j = 0; j < dims_i.size(); ++j) {
-      shape.push_back(dims_i[j]);
-    }
-  } else {
-    // for batch detection
-    // image[0] -> output[0] shape {145, 6}
-    // image[1] -> output[1] shape {176, 6}
-    // then,
-    // the batch output shape {321, 6}
-    // the lod {{0, 145, 321}}
-    // so we should append output[0] to {176, 6}
-    size_t max_dim = 0;
-    for (size_t j = 1; j < lod[0].size(); j++) {
-      max_dim = std::max(max_dim, lod[0][j] - lod[0][j - 1]);
-    }
-    size_t common_dim = lod[0].back() == 0 ? 0 : num / lod[0].back();
-    if (max_dim > 0) {
-      data.resize((lod[0].size() - 1) * max_dim * common_dim, 0);
-    }
-    for (size_t j = 1; j < lod[0].size(); j++) {
-      size_t start = lod[0][j - 1] * common_dim;
-      size_t end = lod[0][j] * common_dim;
-      if (end > start) {
-        std::copy(output_ptr + start, output_ptr + end,
-                  data.begin() + (j - 1) * max_dim * common_dim);
-      }
-    }
-    shape.push_back(lod[0].size() - 1);
-    shape.push_back(max_dim);
-    for (int j = 1; j < dims_i.size(); ++j) {
-      shape.push_back(dims_i[j]);
-    }
-  }
-
-  output->shape = shape;
-  auto &buffer = output->data;
-  if (buffer.empty() || buffer.length() < sizeof(T) * data.size()) {
-    buffer.Resize(sizeof(T) * data.size());
-  }
-  std::memcpy(buffer.data(), data.data(), sizeof(T) * data.size());
-  // copy LoD
-  for (const auto &level : fetch.lod()) {
-    output->lod.emplace_back(level);
+  // set shape.
+  auto shape = framework::vectorize(fetch.dims());
+  output->shape.assign(shape.begin(), shape.end());
+  // set data.
+  const T *data = fetch.data<T>();
+  int num_elems = inference::VecReduceToInt(shape);
+  output->data.Resize(num_elems * sizeof(T));
+  // The fetched tensor output by fetch op, should always in CPU memory, so just
+  // copy.
+  memcpy(output->data.data(), data, num_elems * sizeof(T));
+  // set lod
+  output->lod.clear();
+  for (auto &level : fetch.lod()) {
+    output->lod.emplace_back(level.begin(), level.end());
   }
 }
 
diff --git a/paddle/fluid/inference/api/helper.h b/paddle/fluid/inference/api/helper.h
index 8e359a67738c0df180933421b45f15b39fd0e78c..1fec2f96da0f9d978a3537b2d78e4ce5ef628c81 100644
--- a/paddle/fluid/inference/api/helper.h
+++ b/paddle/fluid/inference/api/helper.h
@@ -74,13 +74,17 @@ template <>
 std::string to_string<std::vector<std::vector<float>>>(
     const std::vector<std::vector<std::vector<float>>> &vec);
 
+template <typename T>
+int VecReduceToInt(const std::vector<T> &v) {
+  return std::accumulate(v.begin(), v.end(), 1, [](T a, T b) { return a * b; });
+}
+
 template <typename T>
 static void TensorAssignData(PaddleTensor *tensor,
                              const std::vector<std::vector<T>> &data) {
   // Assign buffer
-  int dim = std::accumulate(tensor->shape.begin(), tensor->shape.end(), 1,
-                            [](int a, int b) { return a * b; });
-  tensor->data.Resize(sizeof(T) * dim);
+  int num_elems = VecReduceToInt(tensor->shape);
+  tensor->data.Resize(sizeof(T) * num_elems);
   int c = 0;
   for (const auto &f : data) {
     for (T v : f) {
@@ -89,7 +93,7 @@ static void TensorAssignData(PaddleTensor *tensor,
   }
 }
 
-std::string DescribeTensor(const PaddleTensor &tensor) {
+static std::string DescribeTensor(const PaddleTensor &tensor) {
   std::stringstream os;
   os << "Tensor [" << tensor.name << "]\n";
   os << " - type: ";
@@ -113,8 +117,7 @@ std::string DescribeTensor(const PaddleTensor &tensor) {
   os << "\n";
   os << " - data: ";
 
-  int dim = std::accumulate(tensor.shape.begin(), tensor.shape.end(), 1,
-                            [](int a, int b) { return a * b; });
+  int dim = VecReduceToInt(tensor.shape);
   for (int i = 0; i < dim; i++) {
     os << static_cast<float *>(tensor.data.data())[i] << " ";
   }
@@ -122,8 +125,8 @@ std::string DescribeTensor(const PaddleTensor &tensor) {
   return os.str();
 }
 
-void PrintTime(int batch_size, int repeat, int num_threads, int tid,
-               double latency, int epoch = 1) {
+static void PrintTime(int batch_size, int repeat, int num_threads, int tid,
+                      double latency, int epoch = 1) {
   LOG(INFO) << "====== batch_size: " << batch_size << ", repeat: " << repeat
             << ", threads: " << num_threads << ", thread id: " << tid
             << ", latency: " << latency << "ms ======";
diff --git a/paddle/fluid/inference/tests/api/tester_helper.h b/paddle/fluid/inference/tests/api/tester_helper.h
index 7189df775227680726a9d4840386280c5ad44c23..9fcb5129d268a7730c11e5910077ad233050484e 100644
--- a/paddle/fluid/inference/tests/api/tester_helper.h
+++ b/paddle/fluid/inference/tests/api/tester_helper.h
@@ -47,11 +47,8 @@ void CompareResult(const std::vector<PaddleTensor> &outputs,
   for (size_t i = 0; i < outputs.size(); i++) {
     auto &out = outputs[i];
     auto &ref_out = ref_outputs[i];
-    size_t size = std::accumulate(out.shape.begin(), out.shape.end(), 1,
-                                  [](int a, int b) { return a * b; });
-    size_t ref_size =
-        std::accumulate(ref_out.shape.begin(), ref_out.shape.end(), 1,
-                        [](int a, int b) { return a * b; });
+    size_t size = VecReduceToInt(out.shape);
+    size_t ref_size = VecReduceToInt(ref_out.shape);
     EXPECT_GT(size, 0);
     EXPECT_EQ(size, ref_size);
     EXPECT_EQ(out.dtype, ref_out.dtype);
@@ -87,10 +84,7 @@ std::unique_ptr<PaddlePredictor> CreateTestPredictor(
   }
 }
 
-size_t GetSize(const PaddleTensor &out) {
-  return std::accumulate(out.shape.begin(), out.shape.end(), 1,
-                         [](int a, int b) { return a * b; });
-}
+size_t GetSize(const PaddleTensor &out) { return VecReduceToInt(out.shape); }
 
 std::unordered_map<std::string, int> GetFuseStatis(AnalysisConfig config,
                                                    int *num_ops) {