Merge pull request #1210 from lijiancheng0614/transpose2-dev

Transpose2 dev

src/common/types.cpp

@@ -44,6 +44,7 @@ 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";
@@ -91,6 +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"}}},

src/framework/load_ops.h

@@ -115,6 +115,9 @@ LOAD_OP2(reshape2, CPU, MALI_GPU);
 #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

src/operators/kernel/arm/transpose2_kernel.cpp

+/* 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

src/operators/kernel/central-arm-func/transpose2_arm_func.h

+/* 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

src/operators/kernel/transpose2_kernel.h

+/* 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

src/operators/op_param.h

@@ -1132,6 +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 {

src/operators/transpose2_op.cpp

+/* 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

src/operators/transpose2_op.h

+/* 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

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)

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"

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"
 

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"

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"
 

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"

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"

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:

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"
 

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"
 

test/operators/test_reshape2_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/reshape2_op.h"
 

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"

test/operators/test_transpose2_op.cpp

+/* 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;
+}

tools/op.cmake

@@ -205,6 +205,7 @@ if(NOT FOUND_MATCH)
   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)
@@ -251,6 +252,7 @@ endif()
   # option(SIGMOID_OP "" ON)
   # option(SOFTMAX_OP "" ON)
   # option(TRANSPOSE_OP "" ON)
+  # option(TRANSPOSE2_OP "" ON)
 # endif ()
 
 if (BATCHNORM_OP)
@@ -328,6 +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()