add depthwise conv forward

paddle/operators/conv_op.cc

@@ -318,9 +318,16 @@ framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
 namespace ops = paddle::operators;
 REGISTER_OP(conv2d, ops::ConvOp, ops::Conv2DOpMaker, conv2d_grad,
             ops::ConvOpGrad);
+REGISTER_OP(depthwiseConv, ops::ConvOp, ops::Conv2DOpMaker, conv2d_grad,
+            ops::ConvOpGrad);
 REGISTER_OP(conv3d, ops::ConvOp, ops::Conv3DOpMaker, conv3d_grad,
             ops::ConvOpGrad);
 
+REGISTER_OP_CPU_KERNEL(
+    depthwiseConv,
+    ops::DepthwiseConvKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::DepthwiseConvKernel<paddle::platform::CPUDeviceContext, double>);
+
 REGISTER_OP_CPU_KERNEL(
     conv2d, ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
     ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);

paddle/operators/conv_op.cu.cc

@@ -16,6 +16,11 @@ limitations under the License. */
 
 namespace ops = paddle::operators;
 
+REGISTER_OP_CUDA_KERNEL(
+    depthwiseConv,
+    ops::DepthwiseConvKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::DepthwiseConvKernel<paddle::platform::CUDADeviceContext, double>);
+
 REGISTER_OP_CUDA_KERNEL(
     conv2d, ops::GemmConvKernel<paddle::platform::CUDADeviceContext, float>,
     ops::GemmConvKernel<paddle::platform::CUDADeviceContext, double>);

paddle/operators/conv_op.h

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include "paddle/framework/eigen.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/depthwise_conv.h"
 #include "paddle/operators/math/im2col.h"
 #include "paddle/operators/math/math_function.h"
 #include "paddle/operators/math/vol2col.h"
@@ -350,5 +351,34 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
     }
   }
 };
+
+template <typename DeviceContext, typename T>
+class DepthwiseConvKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const Tensor* input = context.Input<Tensor>("Input");
+    // The filter will be reshaped in the calculations,
+    // so here use an assignment operation,
+    // that avoids modifying the variable in the Scope.
+    Tensor filter = *context.Input<Tensor>("Filter");
+    Tensor* output = context.Output<Tensor>("Output");
+    output->mutable_data<T>(context.GetPlace());
+
+    std::vector<int> strides = context.Attr<std::vector<int>>("strides");
+    std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
+    std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
+
+    framework::DDim filter_matrix_shape = {filter.dims()[0],
+                                           filter.numel() / filter.dims()[0]};
+    filter.Resize(filter_matrix_shape);
+
+    math::DepthwiseConvFunctor<DeviceContext, T> depthwiseConv;
+
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    depthwiseConv(dev_ctx, input, filter, filter_shape_vec, strides, paddings,
+                  output);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle

paddle/operators/math/depthwise_conv.h

@@ -0,0 +1,57 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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 "paddle/framework/tensor.h"
+#include "paddle/platform/device_context.h"
+#include "paddle/platform/hostdevice.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+template <typename DeviceContext, typename T>
+class DepthwiseConvFunctor {
+ public:
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  const framework::Tensor& filter, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings,
+                  framework::Tensor* output);
+};
+
+/*
+template <typename DeviceContext, typename T>
+class DepthwiseConvInputGradFunctor {
+public:
+  void operator()(const DeviceContext& context,
+                  const framework::Tensor& filter,
+                  const framework::Tensor& output_grad, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings,
+                  framework::Tensor* input_grad);
+};
+
+template <typename DeviceContext, typename T>
+class DepthwiseConvFilterGradFunctor {
+public:
+  void operator()(const DeviceContext& context,
+                  const framework::Tensor& input,
+                  const framework::Tensor& output_grad, std::vector<int>& ksize,
+                  std::vector<int>& strides, std::vector<int>& paddings,
+                  framework::Tensor* filter_grad);
+};
+*/
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle