diff --git a/src/common/types.cpp b/src/common/types.cpp
index 8c8de7765161e61dc75036a87a34fc6abd2df43e..b90fb70f2a81b365f049632cc7281a69ec58e18d 100644
--- a/src/common/types.cpp
+++ b/src/common/types.cpp
@@ -40,9 +40,11 @@ const char *G_OP_TYPE_POOL2D = "pool2d";
 const char *G_OP_TYPE_PRIOR_BOX = "prior_box";
 const char *G_OP_TYPE_RELU = "relu";
 const char *G_OP_TYPE_RESHAPE = "reshape";
+const char *G_OP_TYPE_RESHAPE2 = "reshape2";
 const char *G_OP_TYPE_SIGMOID = "sigmoid";
 const char *G_OP_TYPE_SOFTMAX = "softmax";
 const char *G_OP_TYPE_TRANSPOSE = "transpose";
+const char *G_OP_TYPE_TRANSPOSE2 = "transpose2";
 const char *G_OP_TYPE_SPLIT = "split";
 const char *G_OP_TYPE_FEED = "feed";
 const char *G_OP_TYPE_FETCH = "fetch";
@@ -90,6 +92,7 @@ std::unordered_map<
         {G_OP_TYPE_FEED, {{"X"}, {"Out"}}},
         {G_OP_TYPE_FETCH, {{"X"}, {"Out"}}},
         {G_OP_TYPE_TRANSPOSE, {{"X"}, {"Out"}}},
+        {G_OP_TYPE_TRANSPOSE2, {{"X"}, {"Out", "XShape"}}},
         {G_OP_TYPE_BOX_CODER,
          {{"PriorBox", "PriorBoxVar", "TargetBox"}, {"OutputBox"}}},
         {G_OP_TYPE_FUSION_CONV_ADD_BN_RELU, {{"Input"}, {"Out"}}},
@@ -99,6 +102,7 @@ std::unordered_map<
         {G_OP_TYPE_POLYGON_BOX_TRANSFORM, {{"Input"}, {"Output"}}},
         {G_OP_TYPE_FC, {{"X", "Y", "Z"}, {"Out"}}},
         {G_OP_TYPE_RESHAPE, {{"X"}, {"Out"}}},
+        {G_OP_TYPE_RESHAPE2, {{"X"}, {"Out", "XShape"}}},
         {G_OP_TYPE_DEPTHWISE_CONV, {{"Input"}, {"Output"}}},
         {G_OP_TYPE_FILL_CONSTANT, {{}, {"Out"}}},
         {G_OP_TYPE_FUSION_CONV_ADD_RELU, {{"Input"}, {"Out"}}},
diff --git a/src/framework/load_ops.h b/src/framework/load_ops.h
index 2b76b0158fe06e8678208f6f98fcdb71f8d91e51..982f1c0f3525afde8475866c0121343fafc9d5a0 100644
--- a/src/framework/load_ops.h
+++ b/src/framework/load_ops.h
@@ -109,9 +109,15 @@ LOAD_FUSION_MATCHER(fusion_conv_add_bn_relu);
 #ifdef RESHAPE_OP
 LOAD_OP2(reshape, CPU, MALI_GPU);
 #endif
+#ifdef RESHAPE2_OP
+LOAD_OP2(reshape2, CPU, MALI_GPU);
+#endif
 #ifdef TRANSPOSE_OP
 LOAD_OP1(transpose, CPU);
 #endif
+#ifdef TRANSPOSE2_OP
+LOAD_OP1(transpose2, CPU);
+#endif
 #ifdef PRIORBOX_OP
 LOAD_OP1(prior_box, CPU);
 #endif
@@ -221,5 +227,9 @@ LOAD_FUSION_MATCHER(fusion_conv_bn);
 #ifdef ELEMENTWISESUB_OP
 LOAD_OP1(elementwise_sub, CPU)
 #endif
+#ifdef QUANT_OP
 LOAD_OP1(quantize, CPU);
+#endif
+#ifdef DEQUANT_OP
 LOAD_OP1(dequantize, CPU);
+#endif
diff --git a/src/operators/kernel/arm/quantize_kernel.cpp b/src/operators/kernel/arm/quantize_kernel.cpp
index e7552d2602b31f9a5c10e3d81122babae8fcf1a8..11a1f0a53d4886e1a07d258b76b3827671471dca 100644
--- a/src/operators/kernel/arm/quantize_kernel.cpp
+++ b/src/operators/kernel/arm/quantize_kernel.cpp
@@ -135,11 +135,15 @@ static void quantize_round_to_even(const Tensor *input, const float scale,
 #if defined(__ARM_NEON__) || defined(__ARM_NEON)
   size_t loop = size >> 4;
   size_t remain = size & 0xF;
+
+  #pragma omp parallel for
   for (size_t i = 0; i < loop; ++i) {
-    float32x4_t r0 = vld1q_f32(x);
-    float32x4_t r1 = vld1q_f32(x + 4);
-    float32x4_t r2 = vld1q_f32(x + 8);
-    float32x4_t r3 = vld1q_f32(x + 12);
+    const float *local_x = x + (i << 4);
+    int8_t *local_y = y + (i << 4);
+    float32x4_t r0 = vld1q_f32(local_x);
+    float32x4_t r1 = vld1q_f32(local_x + 4);
+    float32x4_t r2 = vld1q_f32(local_x + 8);
+    float32x4_t r3 = vld1q_f32(local_x + 12);
     r0 = vmulq_n_f32(r0, scale);
     r1 = vmulq_n_f32(r1, scale);
     r2 = vmulq_n_f32(r2, scale);
@@ -156,12 +160,12 @@ static void quantize_round_to_even(const Tensor *input, const float scale,
     int16x8_t q6 = vcombine_s16(d2, d3);
     int8x8_t d5 = vmovn_s16(q5);
     int8x8_t d6 = vmovn_s16(q6);
-    vst1_s8(y, d5);
-    vst1_s8(y + 8, d6);
-    x += 16;
-    y += 16;
+    vst1_s8(local_y, d5);
+    vst1_s8(local_y + 8, d6);
   }
   size = remain;
+  x += (loop << 4);
+  y += (loop << 4);
 #endif
   for (size_t i = 0; i < size; ++i) {
     float value = x[i] * scale;
@@ -187,11 +191,15 @@ static void quantize_round_to_zero(const Tensor *input, const float scale,
 #ifdef defined(__ARM_NEON__) || defined(__ARM_NEON)
   size_t loop = size >> 4;
   size_t remain = size & 0xF;
+
+  #pragma omp parallel for
   for (size_t i = 0; i < loop; ++i) {
-    float32x4_t r0 = vld1q_f32(x);
-    float32x4_t r1 = vld1q_f32(x + 4);
-    float32x4_t r2 = vld1q_f32(x + 8);
-    float32x4_t r3 = vld1q_f32(x + 12);
+    const float *local_x = x + (i << 4);
+    int8_t *local_y = y + (i << 4);
+    float32x4_t r0 = vld1q_f32(local_x);
+    float32x4_t r1 = vld1q_f32(local_x + 4);
+    float32x4_t r2 = vld1q_f32(local_x + 8);
+    float32x4_t r3 = vld1q_f32(local_x + 12);
     r0 = vmulq_n_f32(r0, scale);
     r1 = vmulq_n_f32(r1, scale);
     r2 = vmulq_n_f32(r2, scale);
@@ -208,12 +216,12 @@ static void quantize_round_to_zero(const Tensor *input, const float scale,
     int16x8_t q6 = vcombine_s16(d2, d3);
     int8x8_t d5 = vmovn_s16(q5);
     int8x8_t d6 = vmovn_s16(q6);
-    vst1_s8(y, d5);
-    vst1_s8(y + 8, d6);
-    x += 16;
-    y += 16;
+    vst1_s8(local_y, d5);
+    vst1_s8(local_y + 8, d6);
   }
   size = remain;
+  x += (loop << 4);
+  y += (loop << 4);
 #endif
   for (size_t i = 0; i < size; ++i) {
     y[i] = trunc(x[i] * scale);
@@ -228,11 +236,15 @@ static void quantize_round_to_nearest(const Tensor *input, const float scale,
 #if defined(__ARM_NEON__) || defined(__ARM_NEON)
   size_t loop = size >> 4;
   size_t remain = size & 0xF;
+
+  #pragma omp parallel for
   for (size_t i = 0; i < loop; ++i) {
-    float32x4_t r0 = vld1q_f32(x);
-    float32x4_t r1 = vld1q_f32(x + 4);
-    float32x4_t r2 = vld1q_f32(x + 8);
-    float32x4_t r3 = vld1q_f32(x + 12);
+    const float *local_x = x + (i << 4);
+    int8_t *local_y = y + (i << 4);
+    float32x4_t r0 = vld1q_f32(local_x);
+    float32x4_t r1 = vld1q_f32(local_x + 4);
+    float32x4_t r2 = vld1q_f32(local_x + 8);
+    float32x4_t r3 = vld1q_f32(local_x + 12);
     r0 = vmulq_n_f32(r0, scale);
     r1 = vmulq_n_f32(r1, scale);
     r2 = vmulq_n_f32(r2, scale);
@@ -249,12 +261,12 @@ static void quantize_round_to_nearest(const Tensor *input, const float scale,
     int16x8_t q6 = vcombine_s16(d2, d3);
     int8x8_t d5 = vmovn_s16(q5);
     int8x8_t d6 = vmovn_s16(q6);
-    vst1_s8(y, d5);
-    vst1_s8(y + 8, d6);
-    x += 16;
-    y += 16;
+    vst1_s8(local_y, d5);
+    vst1_s8(local_y + 8, d6);
   }
   size = remain;
+  x += (loop << 4);
+  y += (loop << 4);
 #endif
   for (size_t i = 0; i < size; ++i) {
     y[i] = round(x[i] * scale);
diff --git a/src/operators/kernel/arm/reshape2_kernel.cpp b/src/operators/kernel/arm/reshape2_kernel.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..83bbf112abb8b5e290126d6909a0fe77291f8fac
--- /dev/null
+++ b/src/operators/kernel/arm/reshape2_kernel.cpp
@@ -0,0 +1,37 @@
+/* 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. */
+
+#ifdef RESHAPE2_OP
+
+#include "operators/kernel/reshape2_kernel.h"
+#include "operators/kernel/central-arm-func/reshape2_arm_func.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <>
+bool Reshape2Kernel<CPU, float>::Init(Reshape2Param<CPU> *param) {
+  return true;
+}
+
+template <>
+void Reshape2Kernel<CPU, float>::Compute(
+    const Reshape2Param<CPU> &param) const {
+  Reshape2Compute<float>(param);
+}
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif
diff --git a/src/operators/kernel/arm/transpose2_kernel.cpp b/src/operators/kernel/arm/transpose2_kernel.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..656d2768840a52f50c42d3797018aa9aec037783
--- /dev/null
+++ b/src/operators/kernel/arm/transpose2_kernel.cpp
@@ -0,0 +1,36 @@
+/* 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. */
+#ifdef TRANSPOSE2_OP
+
+#include "operators/kernel/transpose2_kernel.h"
+#include "operators/kernel/central-arm-func/transpose2_arm_func.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <>
+bool Transpose2Kernel<CPU, float>::Init(Transpose2Param<CPU> *param) {
+  return true;
+}
+
+template <>
+void Transpose2Kernel<CPU, float>::Compute(
+    const Transpose2Param<CPU> &param) const {
+  Transpose2Compute<float>(param);
+}
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif
diff --git a/src/operators/kernel/central-arm-func/elementwise_add_arm_func.h b/src/operators/kernel/central-arm-func/elementwise_add_arm_func.h
index 0c01ef0072444479d2d2e2f7676b842d89e432ec..b6288380a04c71b3d6467f7f6648db046ae9acc9 100644
--- a/src/operators/kernel/central-arm-func/elementwise_add_arm_func.h
+++ b/src/operators/kernel/central-arm-func/elementwise_add_arm_func.h
@@ -58,6 +58,7 @@ void ElementwiseAddCompute(const ElementwiseAddParam<CPU> &param) {
   const float *input_data = input_x->data<float>();
   float *output_data = Out->mutable_data<float>();
   for (int i = 0; i < batch; ++i) {
+    #pragma omp parallel for
     for (int j = 0; j < channels; ++j) {
       size_t offset = (i * channels + j) * elementwise_num;
       const float *input = input_data + offset;
diff --git a/src/operators/kernel/central-arm-func/reshape2_arm_func.h b/src/operators/kernel/central-arm-func/reshape2_arm_func.h
new file mode 100644
index 0000000000000000000000000000000000000000..c22cf120313b039944932fb4e6cc52aa59a68fd4
--- /dev/null
+++ b/src/operators/kernel/central-arm-func/reshape2_arm_func.h
@@ -0,0 +1,59 @@
+/* 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. */
+
+#ifdef RESHAPE2_OP
+#pragma once
+
+#include <vector>
+#include "operators/kernel/reshape_kernel.h"
+#include "operators/op_param.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <typename P>
+void Reshape2Compute(const Reshape2Param<CPU> &param) {
+  const auto *input_x = param.InputX();
+  const auto &input_x_dims = input_x->dims();
+  auto *out = param.Out();
+  framework::DDim out_dims = out->dims();
+  const auto *input_shape = param.InputShape();
+
+  if (input_shape) {
+    auto *shape_data = input_shape->data<int>();
+    framework::Tensor cpu_shape_tensor;
+    auto shape =
+        std::vector<int>(shape_data, shape_data + input_shape->numel());
+    out_dims = ValidateShape(shape, input_x->dims());
+  } else {
+    auto &shape = param.Shape();
+    out_dims = ValidateShape(shape, input_x_dims);
+  }
+
+  bool inplace = param.Inplace();
+  out->Resize(out_dims);
+  if (!inplace) {
+    out->mutable_data<float>();
+    framework::TensorCopy(*input_x, out);
+    out->Resize(out_dims);
+  } else {
+    out->ShareDataWith(*input_x);
+    out->Resize(out_dims);
+  }
+}
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif
diff --git a/src/operators/kernel/central-arm-func/transpose2_arm_func.h b/src/operators/kernel/central-arm-func/transpose2_arm_func.h
new file mode 100644
index 0000000000000000000000000000000000000000..dea90e863b20f19820d60d9cce67b6849d3c467b
--- /dev/null
+++ b/src/operators/kernel/central-arm-func/transpose2_arm_func.h
@@ -0,0 +1,70 @@
+/* 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. */
+
+#ifdef TRANSPOSE2_OP
+#pragma once
+
+#include <vector>
+#include "operators/op_param.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <typename P>
+void Transpose2Compute(const Transpose2Param<CPU>& param) {
+  const auto* input_x = param.InputX();
+  const auto input_x_dims = input_x->dims();
+  auto* out = param.Out();
+  const auto axis = param.Axis();
+  const auto* input_x_data = input_x->data<float>();
+  auto* out_data = out->mutable_data<float>();
+
+  size_t ndim = axis.size();
+  std::vector<int> xdim(ndim);
+  std::vector<int> xstride(ndim);
+  std::vector<int> xout(ndim);
+  for (int i = 0; i < ndim; i++) {
+    int j = ndim - 1 - i;
+    xdim[j] = input_x_dims[axis[i]];
+    xstride[j] = 1;
+    for (int k = axis[i] + 1; k < ndim; k++) {
+      xstride[j] *= input_x_dims[k];
+    }
+    xout[j] = xstride[j] * xdim[j];
+  }
+
+  auto numel = input_x->numel();
+  size_t pind = 0;
+  std::vector<int> ind(ndim);
+  for (int i = 0; i < numel; i++) {
+    out_data[i] = input_x_data[pind];
+    ind[0]++;
+    pind += xstride[0];
+    for (int j = 0; j < ndim - 1; j++) {
+      if (ind[j] == xdim[j]) {
+        ind[j + 1]++;
+        ind[j] = 0;
+        pind += xstride[j + 1];
+        pind -= xout[j];
+      } else {
+        break;
+      }
+    }
+  }
+}
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif
diff --git a/src/operators/kernel/reshape2_kernel.h b/src/operators/kernel/reshape2_kernel.h
new file mode 100644
index 0000000000000000000000000000000000000000..8d15a619d314e3f5d3085a34cff503e286b5ee37
--- /dev/null
+++ b/src/operators/kernel/reshape2_kernel.h
@@ -0,0 +1,36 @@
+/* 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. */
+
+#ifdef RESHAPE2_OP
+
+#pragma once
+
+#include <vector>
+#include "framework/operator.h"
+#include "operators/op_param.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <typename DeviceType, typename T>
+class Reshape2Kernel
+    : public framework::OpKernelBase<DeviceType, Reshape2Param<DeviceType>> {
+ public:
+  void Compute(const Reshape2Param<DeviceType>& param) const;
+  bool Init(Reshape2Param<DeviceType>* param);
+};
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif
diff --git a/src/operators/kernel/transpose2_kernel.h b/src/operators/kernel/transpose2_kernel.h
new file mode 100644
index 0000000000000000000000000000000000000000..8ae75ea483ddb887d9c53b32228ff72b41c76097
--- /dev/null
+++ b/src/operators/kernel/transpose2_kernel.h
@@ -0,0 +1,37 @@
+/* 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. */
+
+#ifdef TRANSPOSE2_OP
+
+#pragma once
+
+#include <vector>
+
+#include "framework/operator.h"
+#include "operators/op_param.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <typename DeviceType, typename T>
+class Transpose2Kernel
+    : public framework::OpKernelBase<DeviceType, Transpose2Param<DeviceType>> {
+ public:
+  void Compute(const Transpose2Param<DeviceType>& param) const;
+  bool Init(Transpose2Param<DeviceType>* param);
+};
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif
diff --git a/src/operators/op_param.h b/src/operators/op_param.h
index 568cf77b8e4e81732cd9a783c1a9ea64d347102b..c60014094b582036ef2038b04edf7be3313e571e 100644
--- a/src/operators/op_param.h
+++ b/src/operators/op_param.h
@@ -243,6 +243,12 @@ class OpParam {
     return GetVarValue<T>("Y", outputs, scope);
   }
 
+  template <typename T>
+  static T *OutputXShapeFrom(const VariableNameMap &outputs,
+                             const Scope &scope) {
+    return GetVarValue<T>("XShape", outputs, scope);
+  }
+
   template <typename T>
   static T *OutputBoxesFrom(const VariableNameMap &outputs,
                             const Scope &scope) {
@@ -1126,6 +1132,37 @@ class TransposeParam : public OpParam {
 };
 #endif
 
+#ifdef TRANSPOSE2_OP
+template <typename Dtype>
+class Transpose2Param : public OpParam {
+  typedef typename DtypeTensorTrait<Dtype>::gtype GType;
+  typedef typename DtypeTensorTrait<Dtype>::rtype RType;
+
+ public:
+  Transpose2Param(const VariableNameMap &inputs, const VariableNameMap &outputs,
+                  const AttributeMap &attrs, const Scope &scope) {
+    input_x_ = InputXFrom<GType>(inputs, scope);
+    out_ = OutFrom<GType>(outputs, scope);
+    output_xshape_ = OutputXShapeFrom<GType>(outputs, scope);
+    axis_ = GetAttr<vector<int>>("axis", attrs);
+  }
+
+  const RType *InputX() const { return input_x_; }
+
+  RType *Out() const { return out_; }
+
+  RType *OutputXShape() const { return output_xshape_; }
+
+  const vector<int> &Axis() const { return axis_; }
+
+ private:
+  RType *input_x_;
+  RType *out_;
+  RType *output_xshape_;
+  vector<int> axis_;
+};
+#endif
+
 #ifdef LOOKUP_OP
 template <typename Dtype>
 class LookupParam : public OpParam {
@@ -1233,6 +1270,49 @@ class ReshapeParam : public OpParam {
 };
 #endif
 
+#ifdef RESHAPE2_OP
+template <typename Dtype>
+class Reshape2Param : public OpParam {
+  typedef typename DtypeTensorTrait<Dtype>::gtype GType;
+  typedef typename DtypeTensorTrait<Dtype>::rtype RType;
+
+ public:
+  Reshape2Param(const VariableNameMap &inputs, const VariableNameMap &outputs,
+                const AttributeMap &attrs, const Scope &scope) {
+    input_x_ = InputXFrom<GType>(inputs, scope);
+    input_shape_ = InputShapeFrom<GType>(inputs, scope);
+    out_ = OutFrom<GType>(outputs, scope);
+    output_xshape_ = OutputXShapeFrom<GType>(outputs, scope);
+    shape_ = GetAttr<vector<int>>("shape", attrs);
+    if (HasAttr("inplace", attrs)) {
+      inplace_ = GetAttr<bool>("inplace", attrs);
+    } else {
+      inplace_ = false;
+    }
+  }
+
+  const RType *InputX() const { return input_x_; }
+
+  const RType *InputShape() const { return input_shape_; }
+
+  RType *Out() const { return out_; }
+
+  RType *OutputXShape() const { return output_xshape_; }
+
+  const vector<int> &Shape() const { return shape_; }
+
+  const bool &Inplace() const { return inplace_; }
+
+ private:
+  RType *input_x_;
+  RType *input_shape_;
+  RType *out_;
+  RType *output_xshape_;
+  vector<int> shape_;
+  bool inplace_;
+};
+#endif
+
 #ifdef SCALE_OP
 template <typename Dtype>
 class ScaleParam : public OpParam {
diff --git a/src/operators/reshape2_op.cpp b/src/operators/reshape2_op.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..d1623076570d466fc53f885374060c5e744365ed
--- /dev/null
+++ b/src/operators/reshape2_op.cpp
@@ -0,0 +1,47 @@
+/* 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. */
+
+#ifdef RESHAPE2_OP
+
+#include "operators/reshape2_op.h"
+#include <vector>
+#include "operators/kernel/reshape_kernel.h"
+namespace paddle_mobile {
+namespace operators {
+
+template <typename Dtype, typename T>
+void Reshape2Op<Dtype, T>::InferShape() const {
+  auto &shape = this->param_.Shape();
+  auto input_x_dims = this->param_.InputX()->dims();
+  auto out_dims = ValidateShape(shape, input_x_dims);
+  this->param_.Out()->Resize(out_dims);
+  std::vector<int64_t> xshape_dims(input_x_dims.size() + 1, 0);
+  for (int i = 0; i < input_x_dims.size(); ++i) {
+    xshape_dims[i + 1] = input_x_dims[i];
+  }
+  this->param_.OutputXShape()->Resize(framework::make_ddim(xshape_dims));
+}
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+namespace ops = paddle_mobile::operators;
+#ifdef PADDLE_MOBILE_CPU
+REGISTER_OPERATOR_CPU(reshape2, ops::Reshape2Op);
+#endif
+#ifdef PADDLE_MOBILE_MALI_GPU
+REGISTER_OPERATOR_MALI_GPU(reshape2, ops::Reshape2Op);
+#endif
+
+#endif
diff --git a/src/operators/reshape2_op.h b/src/operators/reshape2_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..3a06c2b9b90233b6ad0bacb6176f4cc274ff1cc0
--- /dev/null
+++ b/src/operators/reshape2_op.h
@@ -0,0 +1,54 @@
+/* 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. */
+
+#ifdef RESHAPE2_OP
+
+#pragma once
+
+#include <string>
+
+#include "framework/operator.h"
+#include "operators/kernel/reshape2_kernel.h"
+#include "operators/op_param.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+using paddle_mobile::framework::Tensor;
+
+template <typename DeviceType, typename T>
+class Reshape2Op : public framework::OperatorWithKernel<
+                       DeviceType, Reshape2Param<DeviceType>,
+                       operators::Reshape2Kernel<DeviceType, T>> {
+ public:
+  Reshape2Op(const std::string &type, const VariableNameMap &inputs,
+             const VariableNameMap &outputs,
+             const framework::AttributeMap &attrs,
+             std::shared_ptr<framework::Scope> scope)
+      : framework::OperatorWithKernel<DeviceType, Reshape2Param<DeviceType>,
+                                      operators::Reshape2Kernel<DeviceType, T>>(
+            type, inputs, outputs, attrs, scope) {}
+
+  using framework::OperatorWithKernel<
+      DeviceType, Reshape2Param<DeviceType>,
+      operators::Reshape2Kernel<DeviceType, T>>::OperatorWithKernel;
+  void InferShape() const override;
+
+ protected:
+};
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif
diff --git a/src/operators/transpose2_op.cpp b/src/operators/transpose2_op.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..64d07991f60b4057e3d2841afa1bfe6483f31a88
--- /dev/null
+++ b/src/operators/transpose2_op.cpp
@@ -0,0 +1,64 @@
+/* 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. */
+
+#ifdef TRANSPOSE2_OP
+
+#include <vector>
+
+#include "common/enforce.h"
+#include "operators/transpose2_op.h"
+namespace paddle_mobile {
+namespace operators {
+
+template <typename Dtype, typename T>
+void Transpose2Op<Dtype, T>::InferShape() const {
+  auto input_x_dims = this->param_.InputX()->dims();
+  auto axis = this->param_.Axis();
+
+  size_t x_dims_size = input_x_dims.size();
+  size_t axis_size = axis.size();
+
+  PADDLE_MOBILE_ENFORCE((x_dims_size == axis_size),
+                        "input_dims must "
+                        "be equal to the axis_size. ")
+
+  std::vector<int> count(axis_size, 0);
+  for (size_t i = 0; i < axis_size; i++) {
+    PADDLE_MOBILE_ENFORCE(
+        axis[i] < static_cast<int>(axis_size) && ++count[axis[i]] == 1,
+        "Each element of Attribute axis should be a unique value "
+        "range from 0 to (dims - 1), "
+        "where the dims is the axis's size");
+  }
+  framework::DDim out_dims(input_x_dims);
+  for (size_t i = 0; i < axis_size; i++) {
+    out_dims[i] = input_x_dims[axis[i]];
+  }
+  this->param_.Out()->Resize(out_dims);
+  std::vector<int64_t> xshape_dims(input_x_dims.size() + 1, 0);
+  for (int i = 0; i < input_x_dims.size(); ++i) {
+    xshape_dims[i + 1] = input_x_dims[i];
+  }
+  this->param_.OutputXShape()->Resize(framework::make_ddim(xshape_dims));
+}
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+namespace ops = paddle_mobile::operators;
+#ifdef PADDLE_MOBILE_CPU
+REGISTER_OPERATOR_CPU(transpose2, ops::Transpose2Op);
+#endif
+
+#endif  // TRANSPOSE_OP
diff --git a/src/operators/transpose2_op.h b/src/operators/transpose2_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..f1339cc59e0c71a232eddd5dcef47f62994b80da
--- /dev/null
+++ b/src/operators/transpose2_op.h
@@ -0,0 +1,53 @@
+/* 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. */
+
+#ifdef TRANSPOSE2_OP
+
+#pragma once
+
+#include <string>
+
+#include "framework/operator.h"
+#include "operators/kernel/transpose2_kernel.h"
+#include "operators/op_param.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+using paddle_mobile::framework::Tensor;
+
+template <typename DeviceType, typename T>
+class Transpose2Op : public framework::OperatorWithKernel<
+                         DeviceType, Transpose2Param<DeviceType>,
+                         operators::Transpose2Kernel<DeviceType, T>> {
+ public:
+  Transpose2Op(const std::string &type, const VariableNameMap &inputs,
+               const VariableNameMap &outputs,
+               const framework::AttributeMap &attrs,
+               std::shared_ptr<framework::Scope> scope)
+      : framework::OperatorWithKernel<
+            DeviceType, Transpose2Param<DeviceType>,
+            operators::Transpose2Kernel<DeviceType, T>>(type, inputs, outputs,
+                                                        attrs, scope) {}
+
+  using framework::OperatorWithKernel<
+      DeviceType, Transpose2Param<DeviceType>,
+      operators::Transpose2Kernel<DeviceType, T>>::OperatorWithKernel;
+  void InferShape() const override;
+};
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index a10088f9b417b628418404b8df3d340b851af383..2bd7169533f637add2a752feaceca8df132cb262 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -184,6 +184,10 @@ if (NOT FOUND_MATCH)
     ADD_EXECUTABLE(test-transpose-op operators/test_transpose_op.cpp test_helper.h test_include.h)
     target_link_libraries(test-transpose-op paddle-mobile)
 
+    # gen test
+    ADD_EXECUTABLE(test-transpose2-op operators/test_transpose2_op.cpp test_helper.h test_include.h)
+    target_link_libraries(test-transpose2-op paddle-mobile)
+
     # gen test
     ADD_EXECUTABLE(test-multiclassnms-op operators/test_multiclass_nms_op.cpp test_helper.h test_include.h)
     target_link_libraries(test-multiclassnms-op paddle-mobile)
@@ -200,6 +204,10 @@ if (NOT FOUND_MATCH)
     ADD_EXECUTABLE(test-reshape-op operators/test_reshape_op.cpp test_helper.h test_include.h)
     target_link_libraries(test-reshape-op paddle-mobile)
 
+    # gen test
+    ADD_EXECUTABLE(test-reshape2-op operators/test_reshape2_op.cpp test_helper.h test_include.h)
+    target_link_libraries(test-reshape2-op paddle-mobile)
+
     # gen test
     ADD_EXECUTABLE(test-relu-op operators/test_relu_op.cpp test_helper.h test_include.h)
     target_link_libraries(test-relu-op paddle-mobile)
diff --git a/test/net/test_googlenet.cpp b/test/net/test_googlenet.cpp
index c88a78974c330ec270fbcb3f5c28e368ef16440e..f7d29942224b51734cf62988ba8f271f1fa05bc3 100644
--- a/test/net/test_googlenet.cpp
+++ b/test/net/test_googlenet.cpp
@@ -25,8 +25,8 @@ int main() {
   paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
 #endif
 
-  paddle_mobile.SetThreadNum(1);
-  bool optimize = false;
+  paddle_mobile.SetThreadNum(4);
+  bool optimize = true;
   auto time1 = time();
   if (paddle_mobile.Load(g_googlenet, optimize)) {
     auto time2 = time();
@@ -35,10 +35,10 @@ int main() {
     std::vector<float> output;
     std::vector<int64_t> dims{1, 3, 224, 224};
     GetInput<float>(g_test_image_1x3x224x224, &input, dims);
-    //    // 预热十次
-    //    for (int i = 0; i < 10; ++i) {
-    //      output = paddle_mobile.Predict(input, dims);
-    //    }
+    // 预热十次
+    for (int i = 0; i < 10; ++i) {
+      output = paddle_mobile.Predict(input, dims);
+    }
     auto time3 = time();
     for (int i = 0; i < 10; ++i) {
       output = paddle_mobile.Predict(input, dims);
@@ -47,9 +47,6 @@ int main() {
 
     std::cout << "predict cost :" << time_diff(time3, time4) / 10 << "ms"
               << std::endl;
-    for (int i = 0; i < output.size(); ++i) {
-      DLOG << "result[" << i << "] = " << output[i];
-    }
   }
   return 0;
 }
diff --git a/test/operators/test_batchnorm_op.cpp b/test/operators/test_batchnorm_op.cpp
index 4ccad8c1512036c2400a09575b3775e75b26acce..5f064d27f3f3f9cca5428467557c9412f76735c7 100644
--- a/test/operators/test_batchnorm_op.cpp
+++ b/test/operators/test_batchnorm_op.cpp
@@ -12,8 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#pragma once
-
 #include "../test_helper.h"
 #include "../test_include.h"
 #include "operators/batchnorm_op.h"
diff --git a/test/operators/test_box_coder_op.cpp b/test/operators/test_box_coder_op.cpp
index 92cba3995c866c67c00491ad5cc38fb094594ad3..aeef10be9650623767af4d2de8913ce53b1d2c59 100644
--- a/test/operators/test_box_coder_op.cpp
+++ b/test/operators/test_box_coder_op.cpp
@@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#pragma once
 #include "../test_include.h"
 #include "operators/box_coder_op.h"
 
diff --git a/test/operators/test_elementwise_sub_op.cpp b/test/operators/test_elementwise_sub_op.cpp
index cfac83eff7a012d52d47f96e088bd8519603cadc..e27361b21c3146675ea856d02d70878e73e8912f 100644
--- a/test/operators/test_elementwise_sub_op.cpp
+++ b/test/operators/test_elementwise_sub_op.cpp
@@ -12,8 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#pragma once
-
 #include "../test_helper.h"
 #include "../test_include.h"
 #include "operators/elementwise_sub_op.h"
diff --git a/test/operators/test_fill_constant_op.cpp b/test/operators/test_fill_constant_op.cpp
index b099217d1641eb221b3d0d86d780fb6ecfa929bd..99c65ed821c0a90691070b661a6967a11d4694f7 100644
--- a/test/operators/test_fill_constant_op.cpp
+++ b/test/operators/test_fill_constant_op.cpp
@@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#pragma once
 #include "../test_include.h"
 #include "operators/fill_constant_op.h"
 
diff --git a/test/operators/test_fusion_fc_op.cpp b/test/operators/test_fusion_fc_op.cpp
index a23bde45cb74f0f75e655821b15e66b1cef4c081..aaa2d7b578dbda4c6919210eb4a2fb42ba243e53 100644
--- a/test/operators/test_fusion_fc_op.cpp
+++ b/test/operators/test_fusion_fc_op.cpp
@@ -12,8 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#pragma once
-
 #include <framework/program/program-optimize/program_optimize.h>
 #include "../test_include.h"
 #include "operators/fusion_fc_op.h"
diff --git a/test/operators/test_im2sequence_op.cpp b/test/operators/test_im2sequence_op.cpp
index b45e437e12f95cd9f7050247fc03a152246d8122..6c69d1cc9d94ffd958251ee4ed783d6b5531c455 100644
--- a/test/operators/test_im2sequence_op.cpp
+++ b/test/operators/test_im2sequence_op.cpp
@@ -12,8 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#pragma once
-
 #include "../test_helper.h"
 #include "../test_include.h"
 #include "operators/im2sequence_op.h"
diff --git a/test/operators/test_multiclass_nms_op.cpp b/test/operators/test_multiclass_nms_op.cpp
index d1b98d4965fd182ab1adc480279f38cea53974be..3447bbdd10b64d2c2f497bdb4d5af15958a9a95b 100644
--- a/test/operators/test_multiclass_nms_op.cpp
+++ b/test/operators/test_multiclass_nms_op.cpp
@@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#pragma once
 #include "../test_include.h"
 #include "operators/multiclass_nms_op.h"
 
@@ -31,14 +30,12 @@ class TestMultiClassNMSOp {
 
     const std::vector<std::shared_ptr<BlockDesc>> blocks =
         to_predict_program_->Blocks();
-    //  DLOG << " **block size " << blocks.size();
     for (auto block_desc : blocks) {
       std::vector<std::shared_ptr<OpDesc>> ops = block_desc->Ops();
-      //    DLOG << " ops " << ops.size();
       for (auto op : ops) {
         if (op->Type() == "multiclass_nms" &&
             op->Input("BBoxes")[0] == "box_coder_0.tmp_0") {
-          DLOG << " mul attr size: " << op->GetAttrMap().size();
+          DLOG << " attr size: " << op->GetAttrMap().size();
           DLOG << " inputs size: " << op->GetInputs().size();
           DLOG << " outputs size: " << op->GetOutputs().size();
           DLOG << " BBoxes is : " << op->Input("BBoxes")[0];
@@ -55,14 +52,6 @@ class TestMultiClassNMSOp {
                << op->GetAttrMap().at("nms_top_k").Get<int>();
           DLOG << " score_threshold : "
                << op->GetAttrMap().at("score_threshold").Get<float>();
-          //                            DLOG << " variances : " <<
-          //                            op->GetAttrMap().at("variances").Get<std::vector<float>>();
-          //                            DLOG << " aspect_ratios : " <<
-          //                            op->GetAttrMap().at("aspect_ratios").Get<std::vector<float>>();
-          //                            DLOG << " min_sizes : " <<
-          //                            op->GetAttrMap().at("min_sizes").Get<std::vector<float>>();
-          //                            DLOG << " max_sizes : " <<
-          //                            op->GetAttrMap().at("max_sizes").Get<std::vector<float>>();
           std::shared_ptr<operators::MultiClassNMSOp<Dtype, float>> priorbox =
               std::make_shared<operators::MultiClassNMSOp<Dtype, float>>(
                   op->Type(), op->GetInputs(), op->GetOutputs(),
@@ -88,16 +77,12 @@ class TestMultiClassNMSOp {
     auto *output_tensor = output->GetMutable<LoDTensor>();
     output_tensor->mutable_data<float>({1917, 6});
 
-    //  DLOG << typeid(output_tensor).name();
-    //  DLOG << "output_tensor dims: " << output_tensor->dims();
-
     std::shared_ptr<Tensor> out_tensor = std::make_shared<LoDTensor>();
     out_tensor.reset(output_tensor);
 
     predict(t1, t2, 0);
 
     return out_tensor;
-    // return outvars_tensor;
   }
 
  private:
diff --git a/test/operators/test_polygon_box_transform_op.cpp b/test/operators/test_polygon_box_transform_op.cpp
index a71177ddbd8e4d8b0f204fd6ec9c948882499cbd..5b30ce1ebfd59db972953e16e4506fa2595b8f04 100644
--- a/test/operators/test_polygon_box_transform_op.cpp
+++ b/test/operators/test_polygon_box_transform_op.cpp
@@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#pragma once
 #include "../test_include.h"
 #include "operators/polygon_box_transform_op.h"
 
diff --git a/test/operators/test_prior_box_op.cpp b/test/operators/test_prior_box_op.cpp
index 8c697a9a7982f05b71caa5bb5f4d12e50dc9d418..2c75d01df297030b4633829ac4b29f7592aaf5c4 100644
--- a/test/operators/test_prior_box_op.cpp
+++ b/test/operators/test_prior_box_op.cpp
@@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#pragma once
 #include "../test_include.h"
 #include "operators/prior_box_op.h"
 
diff --git a/test/operators/test_reshape2_op.cpp b/test/operators/test_reshape2_op.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..42c348a6274592eb23332620131faa0784a71d28
--- /dev/null
+++ b/test/operators/test_reshape2_op.cpp
@@ -0,0 +1,142 @@
+/* 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 "../test_include.h"
+#include "operators/reshape2_op.h"
+
+namespace paddle_mobile {
+namespace framework {
+
+template <typename Dtype>
+class TestReshape2Op {
+ public:
+  explicit TestReshape2Op(const Program<Dtype> p) : program_(p) {
+    if (use_optimize_) {
+      to_predict_program_ = program_.optimizeProgram;
+    } else {
+      to_predict_program_ = program_.originProgram;
+    }
+    const std::vector<std::shared_ptr<BlockDesc>> blocks =
+        to_predict_program_->Blocks();
+    for (auto block_desc : blocks) {
+      std::vector<std::shared_ptr<OpDesc>> ops = block_desc->Ops();
+      for (auto op : ops) {
+        if (op->Type() == "reshape2") {
+          DLOG << " attr size: " << op->GetAttrMap().size();
+          std::unordered_map<std::string, Attribute> attrs = op->GetAttrMap();
+          for (std::unordered_map<std::string, Attribute>::iterator it =
+                   attrs.begin();
+               it != attrs.end(); ++it) {
+            DLOG << "  " << it->first << " " << it->second;
+          }
+
+          DLOG << " inputs size: " << op->GetInputs().size();
+          VariableNameMap inputs = op->GetInputs();
+          for (VariableNameMap::iterator it = inputs.begin();
+               it != inputs.end(); ++it) {
+            DLOG << "  " << it->first << " " << it->second;
+          }
+
+          DLOG << " outputs size: " << op->GetOutputs().size();
+          VariableNameMap outputs = op->GetOutputs();
+          for (VariableNameMap::iterator it = outputs.begin();
+               it != outputs.end(); ++it) {
+            DLOG << "  " << it->first << " " << it->second;
+          }
+
+          input_var_name = op->Input("X")[0];
+          output_var_name = op->Output("Out")[0];
+          std::shared_ptr<operators::Reshape2Op<Dtype, float>> op_ptr =
+              std::make_shared<operators::Reshape2Op<Dtype, float>>(
+                  op->Type(), op->GetInputs(), op->GetOutputs(),
+                  op->GetAttrMap(), program_.scope);
+          ops_of_block_[*block_desc.get()].push_back(op_ptr);
+          return;
+        }
+      }
+    }
+  }
+
+  std::shared_ptr<Tensor> predict(const Tensor &t) {
+    auto scope = program_.scope;
+    Variable *input_feed_value = scope->Var(input_var_name);
+    auto tensor_input = input_feed_value->GetMutable<LoDTensor>();
+    tensor_input->ShareDataWith(t);
+
+    Variable *output = scope->Var(output_var_name);
+    auto *output_tensor = output->GetMutable<LoDTensor>();
+
+    std::shared_ptr<Tensor> out_tensor = std::make_shared<LoDTensor>();
+    out_tensor.reset(output_tensor);
+
+    predict(t, 0);
+
+    return out_tensor;
+  }
+
+ private:
+  const framework::Program<Dtype> program_;
+  std::shared_ptr<ProgramDesc> to_predict_program_;
+  std::map<framework::BlockDesc,
+           std::vector<std::shared_ptr<OperatorBase<Dtype>>>>
+      ops_of_block_;
+  bool use_optimize_ = false;
+  string input_var_name;
+  string output_var_name;
+
+  void predict(const Tensor &t, int block_id) {
+    std::shared_ptr<BlockDesc> to_predict_block =
+        to_predict_program_->Block(block_id);
+    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();
+    }
+  }
+};
+
+template class TestReshape2Op<CPU>;
+}  // namespace framework
+}  // namespace paddle_mobile
+
+int main() {
+  DLOG << "----------**********----------";
+  DLOG << "begin to run Reshape2 Test";
+  paddle_mobile::Loader<paddle_mobile::CPU> loader;
+  auto program = loader.Load(std::string(g_ocr) + "/model",
+                             std::string(g_ocr) + "/params");
+
+  paddle_mobile::framework::Tensor input;
+  SetupTensor<float>(&input, {1, 4, 4}, static_cast<float>(0),
+                     static_cast<float>(1));
+  auto *input_ptr = input.data<float>();
+  for (int i = 0; i < 16; ++i) {
+    *(input_ptr + i) = i;
+  }
+  DLOG << "input : ";
+  for (int i = 0; i < input.numel(); ++i) {
+    DLOG << " index " << i << " : " << input_ptr[i];
+  }
+
+  paddle_mobile::framework::TestReshape2Op<paddle_mobile::CPU> testReshape2Op(
+      program);
+
+  auto output = testReshape2Op.predict(input);
+  auto *output_ptr = output->data<float>();
+
+  DLOG << "output : ";
+  for (int i = 0; i < output->numel(); ++i) {
+    DLOG << " index " << i << " : " << output_ptr[i];
+  }
+  return 0;
+}
diff --git a/test/operators/test_sum_op.cpp b/test/operators/test_sum_op.cpp
index e51d1cff5e99c5d9c444db046e78eee6a03f9243..467529d8d3877fcb9ac5527daf5f037aea6d18fc 100644
--- a/test/operators/test_sum_op.cpp
+++ b/test/operators/test_sum_op.cpp
@@ -12,8 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#pragma once
-
 #include "../test_helper.h"
 #include "../test_include.h"
 #include "operators/sum_op.h"
diff --git a/test/operators/test_transpose2_op.cpp b/test/operators/test_transpose2_op.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..b75a957cd5c1cd08dc09895e9e2448761e822274
--- /dev/null
+++ b/test/operators/test_transpose2_op.cpp
@@ -0,0 +1,143 @@
+/* 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 "../test_include.h"
+#include "operators/transpose2_op.h"
+
+namespace paddle_mobile {
+namespace framework {
+
+template <typename Dtype>
+class TestTranspose2Op {
+ public:
+  explicit TestTranspose2Op(const Program<Dtype> p) : program_(p) {
+    if (use_optimize_) {
+      to_predict_program_ = program_.optimizeProgram;
+    } else {
+      to_predict_program_ = program_.originProgram;
+    }
+    const std::vector<std::shared_ptr<BlockDesc>> blocks =
+        to_predict_program_->Blocks();
+    for (auto block_desc : blocks) {
+      std::vector<std::shared_ptr<OpDesc>> ops = block_desc->Ops();
+      for (auto op : ops) {
+        if (op->Type() == "transpose2") {
+          DLOG << " attr size: " << op->GetAttrMap().size();
+          std::unordered_map<std::string, Attribute> attrs = op->GetAttrMap();
+          for (std::unordered_map<std::string, Attribute>::iterator it =
+                   attrs.begin();
+               it != attrs.end(); ++it) {
+            DLOG << "  " << it->first << " " << it->second;
+          }
+
+          DLOG << " inputs size: " << op->GetInputs().size();
+          VariableNameMap inputs = op->GetInputs();
+          for (VariableNameMap::iterator it = inputs.begin();
+               it != inputs.end(); ++it) {
+            DLOG << "  " << it->first << " " << it->second;
+          }
+
+          DLOG << " outputs size: " << op->GetOutputs().size();
+          VariableNameMap outputs = op->GetOutputs();
+          for (VariableNameMap::iterator it = outputs.begin();
+               it != outputs.end(); ++it) {
+            DLOG << "  " << it->first << " " << it->second;
+          }
+
+          input_var_name = op->Input("X")[0];
+          output_var_name = op->Output("Out")[0];
+          std::shared_ptr<operators::Transpose2Op<Dtype, float>> op_ptr =
+              std::make_shared<operators::Transpose2Op<Dtype, float>>(
+                  op->Type(), op->GetInputs(), op->GetOutputs(),
+                  op->GetAttrMap(), program_.scope);
+          ops_of_block_[*block_desc.get()].push_back(op_ptr);
+          return;
+        }
+      }
+    }
+  }
+
+  std::shared_ptr<Tensor> predict(const Tensor &t) {
+    auto scope = program_.scope;
+    Variable *input_feed_value = scope->Var(input_var_name);
+    auto tensor_input = input_feed_value->GetMutable<LoDTensor>();
+    tensor_input->ShareDataWith(t);
+
+    Variable *output = scope->Var(output_var_name);
+    auto *output_tensor = output->GetMutable<LoDTensor>();
+    output_tensor->mutable_data<float>({1, 2, 8});
+
+    std::shared_ptr<Tensor> out_tensor = std::make_shared<LoDTensor>();
+    out_tensor.reset(output_tensor);
+
+    predict(t, 0);
+
+    return out_tensor;
+  }
+
+ private:
+  const framework::Program<Dtype> program_;
+  std::shared_ptr<ProgramDesc> to_predict_program_;
+  std::map<framework::BlockDesc,
+           std::vector<std::shared_ptr<OperatorBase<Dtype>>>>
+      ops_of_block_;
+  bool use_optimize_ = false;
+  string input_var_name;
+  string output_var_name;
+
+  void predict(const Tensor &t, int block_id) {
+    std::shared_ptr<BlockDesc> to_predict_block =
+        to_predict_program_->Block(block_id);
+    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();
+    }
+  }
+};
+
+template class TestTranspose2Op<CPU>;
+}  // namespace framework
+}  // namespace paddle_mobile
+
+int main() {
+  DLOG << "----------**********----------";
+  DLOG << "begin to run Transpose2 Test";
+  paddle_mobile::Loader<paddle_mobile::CPU> loader;
+  auto program = loader.Load(std::string(g_ocr) + "/model",
+                             std::string(g_ocr) + "/params");
+
+  paddle_mobile::framework::Tensor input;
+  SetupTensor<float>(&input, {1, 8, 2}, static_cast<float>(0),
+                     static_cast<float>(1));
+  auto *input_ptr = input.data<float>();
+  for (int i = 0; i < 16; ++i) {
+    *(input_ptr + i) = i;
+  }
+  DLOG << "input : ";
+  for (int i = 0; i < input.numel(); ++i) {
+    DLOG << " index " << i << " : " << input_ptr[i];
+  }
+
+  paddle_mobile::framework::TestTranspose2Op<paddle_mobile::CPU>
+      testTranspose2Op(program);
+
+  auto output = testTranspose2Op.predict(input);
+  auto *output_ptr = output->data<float>();
+
+  DLOG << "output : ";
+  for (int i = 0; i < output->numel(); ++i) {
+    DLOG << " index " << i << " : " << output_ptr[i];
+  }
+  return 0;
+}
diff --git a/tools/op.cmake b/tools/op.cmake
index f7a6ed4b134f78ddb23487cd3a861f244e6a86db..2e1e311a2c96bac02257cfdce2d2fbebcd962dfb 100644
--- a/tools/op.cmake
+++ b/tools/op.cmake
@@ -201,9 +201,11 @@ if(NOT FOUND_MATCH)
   set(PRIORBOX_OP ON)
   set(RELU_OP ON)
   set(RESHAPE_OP ON)
+  set(RESHAPE2_OP ON)
   set(SIGMOID_OP ON)
   set(SOFTMAX_OP ON)
   set(TRANSPOSE_OP ON)
+  set(TRANSPOSE2_OP ON)
   set(FUSION_CONVADDBNRELU_OP ON)
   set(FUSION_CONVADDADDPRELU_OP ON)
   set(FUSION_DWCONVBNRELU_OP ON)
@@ -246,9 +248,11 @@ endif()
   # option(PRIORBOX_OP "" ON)
   # option(RELU_OP "" ON)
   # option(RESHAPE_OP "" ON)
+  # option(RESHAPE2_OP "" ON)
   # option(SIGMOID_OP "" ON)
   # option(SOFTMAX_OP "" ON)
   # option(TRANSPOSE_OP "" ON)
+  # option(TRANSPOSE2_OP "" ON)
 # endif ()
 
 if (BATCHNORM_OP)
@@ -314,6 +318,9 @@ endif()
 if (RESHAPE_OP)
   add_definitions(-DRESHAPE_OP)
 endif()
+if (RESHAPE2_OP)
+  add_definitions(-DRESHAPE2_OP)
+endif()
 if (SIGMOID_OP)
   add_definitions(-DSIGMOID_OP)
 endif()
@@ -323,6 +330,9 @@ endif()
 if (TRANSPOSE_OP)
   add_definitions(-DTRANSPOSE_OP)
 endif()
+if (TRANSPOSE2_OP)
+  add_definitions(-DTRANSPOSE2_OP)
+endif()
 if (FUSION_CONVADDBNRELU_OP)
   add_definitions(-DFUSION_CONVADDBNRELU_OP)
 endif()