diff --git a/paddle/framework/data_transform.cc b/paddle/framework/data_transform.cc
index 9d6a8424426a68ae66cf93b803c35e33e30226f2..ac6e40a3ae8f6655eecc8279543d6afe6bbb1399 100644
--- a/paddle/framework/data_transform.cc
+++ b/paddle/framework/data_transform.cc
@@ -11,6 +11,7 @@ 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 <functional>
 
 #include "paddle/framework/data_transform.h"
 #include "paddle/framework/lod_tensor.h"
@@ -74,26 +75,28 @@ void TransDataType(const platform::DeviceContext* ctx,
   }
 }
 
-void TransDataLayout(const platform::DeviceContext* ctx,
+void TransDataLayout(const std::vector<int>& axis,
+                     const platform::DeviceContext* ctx,
                      const KernelTypePair& kernel_pair, const Variable& in,
                      Variable* out) {
-  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only Support Tensor transform!.");
+  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only support Tensor transform!.");
   PADDLE_ENFORCE(
       platform::places_are_same_class(kernel_pair.first.place_,
                                       kernel_pair.second.place_),
-      "TransDataType Only Support DataType transform on same place!");
+      "TransDataLayout only support DataLayout transform on same place!");
+  PADDLE_ENFORCE(kernel_pair.first.data_type_ == kernel_pair.second.data_type_,
+                 "TransDataLayout only support Datatype are same!");
 
   auto src = in.Get<Tensor>();
   auto* dst = out->GetMutable<Tensor>();
   PADDLE_ENFORCE(arity(src.dims()) == 4, "Input Arity Only Suppport 4!");
 
-  auto src_dim = src.dims();
-  dst->Resize(src_dim);
   auto place = kernel_pair.second.place_;
   CopyFrom(src, place, *ctx, dst);
-  const std::vector<int> axis = {0, 2, 3, 1};
 
+  auto src_dim = src.dims();
   std::vector<int64_t> dst_dim;
+
   dst_dim.resize(axis.size());
   for (size_t i = 0; i < axis.size(); i++) {
     dst_dim[i] = src_dim[axis[i]];
@@ -102,7 +105,7 @@ void TransDataLayout(const platform::DeviceContext* ctx,
   dst->Resize(make_ddim(dst_dim));
 
   auto src_type = kernel_pair.first.data_type_;
-  framework::VisitDataType(src_type, CastDataLayout(src, dst, ctx, axis));
+  framework::VisitDataType(src_type, CastDataLayout(ctx, axis, src, dst));
 
   dst->set_layout(kernel_pair.second.data_layout_);
 }
@@ -111,5 +114,22 @@ void TransDataLayout(const platform::DeviceContext* ctx,
 }  // namespace paddle
 
 namespace f = paddle::framework;
+
+namespace {
+std::vector<int> NHWC2NCHW = {0, 3, 1, 2};
+std::vector<int> NCHW2NHWC = {0, 2, 3, 1};
+}
+
 REGISTER_DATA_TRANSFORM_FN(f::KernelFP32, f::KernelFP64, f::TransDataType);
-REGISTER_DATA_TRANSFORM_FN(f::KernelNHWC, f::KernelNCHW, f::TransDataLayout);
+REGISTER_DATA_TRANSFORM_FN(f::KernelNHWC, f::KernelNCHW,
+                           std::bind(f::TransDataLayout, NHWC2NCHW,
+                                     std::placeholders::_1,
+                                     std::placeholders::_2,
+                                     std::placeholders::_3,
+                                     std::placeholders::_4));
+REGISTER_DATA_TRANSFORM_FN(f::KernelNCHW, f::KernelNHWC,
+                           std::bind(f::TransDataLayout, NCHW2NHWC,
+                                     std::placeholders::_1,
+                                     std::placeholders::_2,
+                                     std::placeholders::_3,
+                                     std::placeholders::_4));
diff --git a/paddle/framework/data_transform.h b/paddle/framework/data_transform.h
index 9abb3c99bf30fcf9deab59dc7ee9c02e7c7c775b..56ebc80f4386958608213f30e745f2d9528e9e5e 100644
--- a/paddle/framework/data_transform.h
+++ b/paddle/framework/data_transform.h
@@ -73,6 +73,7 @@ struct CastDataType {
     auto numel = in_.numel();
     auto* in_end = in_begin + numel;
     auto* out_begin = out_->mutable_data<OutType>(place);
+
     if (platform::is_cpu_place(place)) {
       platform::Transform<platform::CPUDeviceContext> trans;
       auto* context = static_cast<const platform::CPUDeviceContext*>(ctx_);
@@ -86,9 +87,9 @@ struct CastDataType {
 };
 
 struct CastDataLayout {
-  CastDataLayout(const framework::Tensor& in, framework::Tensor* out,
-                 const platform::DeviceContext* ctx,
-                 const std::vector<int>& axis)
+  CastDataLayout(const platform::DeviceContext* ctx,
+                 const std::vector<int>& axis, const framework::Tensor& in,
+                 framework::Tensor* out)
       : in_(in), out_(out), ctx_(ctx), axis_(axis) {}
   const framework::Tensor in_;
   framework::Tensor* out_;
@@ -98,6 +99,7 @@ struct CastDataLayout {
   template <typename T>
   void operator()() {
     auto place = ctx_->GetPlace();
+
     if (platform::is_cpu_place(place)) {
       operators::math::Transpose<platform::CPUDeviceContext, T, 4> trans4;
       auto* context = static_cast<const platform::CPUDeviceContext*>(ctx_);
diff --git a/paddle/framework/data_transform_test.cc b/paddle/framework/data_transform_test.cc
index 8665b6248faa2d218230449c45a10f022f3fbf4f..edd305fd17ae202926b83fbec10089719baa2e16 100644
--- a/paddle/framework/data_transform_test.cc
+++ b/paddle/framework/data_transform_test.cc
@@ -106,7 +106,7 @@ TEST(DataTransform, Register) {
   ASSERT_EQ(test_value, 2);
 }
 
-TEST(DataTransform, Layout) {
+TEST(DataTransform, DataLayout) {
   using namespace paddle::framework;
   using namespace paddle::platform;
 
@@ -127,7 +127,19 @@ TEST(DataTransform, Layout) {
   }
 
   Tensor dst = out.Get<Tensor>();
-  EXPECT_TRUE(dst.layout() != src->layout());
+
+  EXPECT_TRUE(dst.layout() == DataLayout::kNCHW);
+  EXPECT_TRUE(dst.dims() == make_ddim({2, 2, 3, 1}));
+
+  {
+    auto kernel1 = GenFromBit({1, 0, 1, 0});
+    auto kernel2 = GenFromBit({1, 0, 0, 0});
+    auto pair0 = std::make_pair(kernel1, kernel2);
+    instance.Get(pair0)(ctx, pair0, out, &in);
+  }
+
+  EXPECT_TRUE(src->layout() == DataLayout::kNHWC);
+  EXPECT_TRUE(src->dims() == make_ddim({2, 3, 1, 2}));
 }
 
 TEST(DataTransform, DataType) {