Merge pull request #321 from smilejames/develop

submit depthwise_conv_op and test

src/operators/conv_op.cpp

@@ -21,13 +21,6 @@ limitations under the License. */
 namespace paddle_mobile {
 namespace operators {
 
-int ConvOutputSize(int input_size, int filter_size, int dilation, int padding,
-                   int stride) {
-  const int dkernel = dilation * (filter_size - 1) + 1;
-  int output_size = (input_size + 2 * padding - dkernel) / stride + 1;
-  return output_size;
-}
-
 template <typename Dtype, typename T>
 void ConvOp<Dtype, T>::InferShape() const {
   //  std::cout << " begin get dims: " << std::endl;

src/operators/conv_op.h

@@ -44,5 +44,12 @@ class ConvOp : public framework::OperatorWithKernel<DeviceType> {
   ConvParam param_;
 };
 
+inline int ConvOutputSize(int input_size, int filter_size, int dilation,
+                          int padding, int stride) {
+  const int dkernel = dilation * (filter_size - 1) + 1;
+  int output_size = (input_size + 2 * padding - dkernel) / stride + 1;
+  return output_size;
+}
+
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/depthwise_conv_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 "operators/depthwise_conv_op.h"
+#include <vector>
+#include "framework/data_type.h"
+#include "framework/op_proto_maker.h"
+#include "framework/op_registry.h"
+#include "operators/conv_op.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <typename Dtype, typename T>
+void DepthwiseConvOp<Dtype, T>::InferShape() const {
+  auto in_dims = param_.Input()->dims();
+  auto filter_dims = param_.Filter()->dims();
+  const std::vector<int> &strides = param_.Strides();
+  std::vector<int> paddings = param_.Paddings();
+  int groups = param_.Groups();
+  std::vector<int> dilations = param_.Dilations();
+
+  PADDLE_MOBILE_ENFORCE((in_dims.size() == filter_dims.size() &&
+                         dilations.size() == paddings.size() &&
+                         paddings.size() == strides.size()),
+                        "ConvParam is not suitable");
+
+  std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
+  for (size_t i = 0; i < strides.size(); ++i) {
+    output_shape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
+                                          dilations[i], paddings[i],
+                                          strides[i]));
+  }
+
+  framework::DDim ddim = framework::make_ddim(output_shape);
+  param_.Output()->Resize(ddim);
+}
+
+template class DepthwiseConvOp<CPU, float>;
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+namespace ops = paddle_mobile::operators;
+USE_OP(depthwise_conv2d);
+REGISTER_OPERATOR(depthwise_conv2d, ops::DepthwiseConvOp);

src/operators/depthwise_conv_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. */
+
+#pragma once
+
+#include <string>
+#include "framework/operator.h"
+#include "operators/kernel/depthwise_conv_kernel.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <typename DeviceType, typename T>
+class DepthwiseConvOp : public framework::OperatorWithKernel<DeviceType> {
+ public:
+  DepthwiseConvOp(const std::string &type, const VariableNameMap &inputs,
+                  const VariableNameMap &outputs,
+                  const framework::AttributeMap &attrs,
+                  std::shared_ptr<framework::Scope> scope)
+      : framework::OperatorWithKernel<DeviceType>(type, inputs, outputs, attrs,
+                                                  scope),
+        param_(inputs, outputs, attrs, *scope) {}
+
+  using framework::OperatorWithKernel<DeviceType>::OperatorWithKernel;
+  void InferShape() const override;
+
+  void RunImpl() const {
+    operators::DepthwiseConvKernel<DeviceType, T> kernel;
+    kernel.Compute(param_);
+    this->ClearVariables({"Filter", "Input"});
+  }
+
+ private:
+  ConvParam param_;
+};
+
+}  // namespace operators
+}  // namespace paddle_mobile

src/operators/kernel/arm/conv_kernel.cpp

@@ -17,19 +17,6 @@ limitations under the License. */
 namespace paddle_mobile {
 namespace operators {
 
-bool IsExpand(const std::vector<int64_t> &filter_dim,
-              const std::vector<int> &strides, const std::vector<int> &paddings,
-              const std::vector<int> &dilations) {
-  bool filter_1 = true, strides_1 = true, padding_0 = true, dilation_1 = true;
-  for (size_t j = 0; j < strides.size(); ++j) {
-    filter_1 = filter_1 && (static_cast<int>(filter_dim[j + 2]) == 1);
-    strides_1 = strides_1 && (strides[j] == 1);
-    padding_0 = padding_0 && (paddings[j] == 0);
-    dilation_1 = dilation_1 && (dilations[j] == 1);
-  }
-  return !(filter_1 && strides_1 && padding_0 && dilation_1);
-}
-
 template <>
 void ConvKernel<CPU, float>::Compute(const ConvParam &param) const {
   LOG(kLOG_DEBUG) << param;

src/operators/kernel/arm/depthwise_conv_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. */
+
+#include "operators/kernel/depthwise_conv_kernel.h"
+#include "operators/kernel/conv_kernel.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <>
+void DepthwiseConvKernel<CPU, float>::Compute(const ConvParam &param) const {
+  LOG(kLOG_DEBUG) << param;
+
+  const Tensor *input = param.Input();
+  Tensor filter = *param.Filter();
+  Tensor *output = param.Output();
+  output->mutable_data<float>();
+
+  int groups = param.Groups();
+  std::vector<int> strides = param.Strides();
+  std::vector<int> paddings = param.Paddings();
+  std::vector<int> dilations = param.Dilations();
+
+  DLOG << " compute end get Attrs " << strides[0];
+
+  const int batch_size = static_cast<int>(input->dims()[0]);
+
+  std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
+  std::vector<int64_t> output_shape_vec(framework::vectorize(output->dims()));
+
+  size_t data_dim = filter_shape_vec.size() - 2;
+  std::vector<int64_t> col_shape_vec(1 + 2 * data_dim);
+  col_shape_vec[0] = input->dims()[1] / groups;
+  for (size_t j = 0; j < data_dim; ++j) {
+    col_shape_vec[j + 1] = filter_shape_vec[j + 2];
+    col_shape_vec[j + 1 + data_dim] = output_shape_vec[j + 2];
+  }
+  framework::DDim col_shape(framework::make_ddim(col_shape_vec));
+
+  framework::DDim col_matrix_shape =
+      framework::flatten_to_2d(col_shape, data_dim + 1);
+
+  bool is_expand = IsExpand(filter_shape_vec, strides, paddings, dilations);
+  Tensor col;
+  Tensor col_matrix;
+  if (is_expand) {
+    col.mutable_data<float>(col_shape);
+    col_matrix.ShareDataWith(col);
+    col_matrix.Resize(col_matrix_shape);
+  }
+  DLOG << " col_shape = " << col_shape;
+  DLOG << " col_matrix_shape = " << col_matrix_shape;
+
+  framework::DDim input_shape = framework::slice_ddim(
+      input->dims(), 1, static_cast<int>(input->dims().size()));
+  DLOG << " input_shape = " << input_shape;
+
+  framework::DDim filter_matrix_shape = {filter.dims()[0],
+                                         filter.numel() / filter.dims()[0]};
+  filter.Resize(filter_matrix_shape);
+  DLOG << " filter.dims() = " << filter.dims();
+
+  framework::DDim output_matrix_shape = {
+      output->dims()[1],
+      output->numel() / (output->dims()[0] * output->dims()[1])};
+
+  // convolution operator: im2col(or vol2col) + gemm
+  int in_step = static_cast<int>(input->dims()[1]) / groups;
+  int out_step = static_cast<int>(output->dims()[1]) / groups;
+
+  math::Vol2ColFunctor<CPU, float> vol2col;
+  math::Im2ColFunctor<math::ColFormat::kCFO, CPU, float> im2col;
+
+  for (int i = 0; i < batch_size; i++) {
+    Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
+    Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
+    DLOG << " in_batch.dims() = " << in_batch.dims();
+    DLOG << " out_batch.dims() = " << out_batch.dims();
+
+    for (int g = 0; g < groups; g++) {
+      Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
+
+      if (!is_expand) {
+        col.ShareDataWith(in_slice);
+        col_matrix.ShareDataWith(col);
+        col_matrix.Resize(col_matrix_shape);
+      } else if (data_dim == 2U) {
+        // im2col
+        im2col(in_slice, dilations, strides,
+               std::vector<int>{paddings[0], paddings[1], paddings[0],
+                                paddings[1]},
+               &col);
+      } else if (data_dim == 3U) {
+        // vol2col
+        vol2col(in_slice, dilations, strides, paddings, &col);
+      }
+
+      // gemm
+      Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
+      Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
+      DLOG << " out_slice " << out_slice.dims();
+      DLOG << " filter_slice " << filter_slice.dims();
+      DLOG << " col_matrix " << col_matrix.dims();
+      math::matmul<float>(filter_slice, false, col_matrix, false,
+                          static_cast<float>(1), &out_slice,
+                          static_cast<float>(0));
+      auto filter_ptr = filter_slice.data<float>();
+    }
+  }
+}
+
+template class DepthwiseConvKernel<CPU, float>;
+
+}  // namespace operators
+}  // namespace paddle_mobile

src/operators/kernel/conv_kernel.h

@@ -12,6 +12,7 @@ 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 <vector>
 #include "framework/operator.h"
 #include "operators/math/im2col.h"
 #include "operators/math/math_function.h"
@@ -23,12 +24,28 @@ limitations under the License. */
 namespace paddle_mobile {
 namespace operators {
 
-using namespace framework;
+using framework::OpKernelBase;
 
 template <typename DeviceType, typename T>
-class ConvKernel : public framework::OpKernelBase<DeviceType, ConvParam> {
+class ConvKernel : public OpKernelBase<DeviceType, ConvParam> {
  public:
   void Compute(const ConvParam &param) const;
 };
+
+inline bool IsExpand(const std::vector<int64_t> &filter_dim,
+                     const std::vector<int> &strides,
+                     const std::vector<int> &paddings,
+                     const std::vector<int> &dilations) {
+  bool filter_1 = true, strides_1 = true, padding_0 = true, dilation_1 = true;
+  for (size_t j = 0; j < strides.size(); ++j) {
+    filter_1 = filter_1 && (static_cast<int>(filter_dim[j + 2]) == 1);
+    strides_1 = strides_1 && (strides[j] == 1);
+    padding_0 = padding_0 && (paddings[j] == 0);
+    dilation_1 = dilation_1 && (dilations[j] == 1);
+  }
+
+  return !(filter_1 && strides_1 && padding_0 && dilation_1);
+}
+
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/depthwise_conv_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. */
+
+#include "framework/operator.h"
+#include "operators/math/im2col.h"
+#include "operators/math/math_function.h"
+#include "operators/math/vol2col.h"
+#include "operators/op_param.h"
+
+#pragma once;
+
+namespace paddle_mobile {
+namespace operators {
+
+using framework::OpKernelBase;
+
+template <typename DeviceType, typename T>
+class DepthwiseConvKernel : public OpKernelBase<DeviceType, ConvParam> {
+ public:
+  void Compute(const ConvParam &param) const;
+};
+}  // namespace operators
+}  // namespace paddle_mobile

test/CMakeLists.txt

@@ -91,3 +91,7 @@ target_link_libraries(test-googlenet paddle-mobile)
 # gen test
 ADD_EXECUTABLE(test-sigmoid operators/test_sigmoid_op.cpp  test_include.h)
 target_link_libraries(test-sigmoid paddle-mobile)
+
+# gen test
+ADD_EXECUTABLE(test-depthwise-conv-op operators/test_depthwise_conv_op.cpp test_helper.h test_include.h executor_for_test.h)
+target_link_libraries(test-depthwise-conv-op paddle-mobile)

test/operators/test_depthwise_conv_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 "../executor_for_test.h"
+#include "../test_include.h"
+#include "operators/depthwise_conv_op.h"
+
+int main() {
+  paddle_mobile::Loader<paddle_mobile::CPU> loader;
+  //  ../models/image_classification_resnet.inference.model
+  auto program = loader.Load(g_mobilenet_ssd);
+
+  PADDLE_MOBILE_ENFORCE(program.originProgram != nullptr,
+                        "program file read fail");
+
+  Executor4Test<paddle_mobile::CPU, paddle_mobile::operators::DepthwiseConvOp<
+                                        paddle_mobile::CPU, float>>
+      executor(program, "depthwise_conv2d");
+
+  paddle_mobile::framework::LoDTensor input;
+  // GetInput<float>(g_test_image_1x3x224x224, &input, {1, 3, 224, 224});
+  // use SetupTensor if not has local input image .
+  SetupTensor<float>(&input, {1, 32, 150, 150}, static_cast<float>(0),
+                     static_cast<float>(1));
+  auto input_ptr = input.data<float>();
+  auto out_ddim = paddle_mobile::framework::make_ddim({1, 32, 150, 150});
+  auto output = executor.Predict(input, "batch_norm_0.tmp_3",
+                                 "depthwise_conv2d_0.tmp_0", out_ddim);
+
+  auto output_ptr = output->data<float>();
+  for (int j = 0; j < output->numel(); ++j) {
+    DLOG << " value of output: " << output_ptr[j];
+  }
+  return 0;
+}