/* Copyright (c) 2016 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 #include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/depthwise_conv.h" #include "paddle/fluid/operators/math/im2col.h" #include "paddle/fluid/operators/math/vol2col.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; // Base convolution operator definations for other conv // like operators to reuse the implementation. 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; PADDLE_ENFORCE( output_size > 0, "Due to the settings of padding(%d), filter_size(%d), dilation(%d) and " "stride(%d), the output size is less than 0, please check " "again. Input_size:%d", padding, filter_size, dilation, stride, input_size); return output_size; } inline bool IsExpand(const std::vector& filter_dim, const std::vector& strides, const std::vector& paddings, const std::vector& 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(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); } // Define Op classes in .h file so that other conv // operator implementations can reuse the code. class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() final; protected: virtual void Apply() {} }; class Conv3DOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() final; protected: virtual void Apply() {} }; class ConvOpInferVarType : public framework::PassInDtypeAndVarTypeToOutput { protected: std::unordered_map GetInputOutputWithSameType() const override { return std::unordered_map{ {"Input", /*->*/ "Output"}}; } }; class ConvOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override; protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override; }; class ConvOpGrad : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override; protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override; }; template class GemmConvKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { const Tensor* input = context.Input("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("Filter"); Tensor* output = context.Output("Output"); output->mutable_data(context.GetPlace()); int groups = context.Attr("groups"); std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); std::vector dilations = context.Attr>("dilations"); const int batch_size = static_cast(input->dims()[0]); // filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w} std::vector filter_shape_vec(framework::vectorize(filter.dims())); // output_shape_vec: {o_n, o_c, o_h, o_w} or {o_n, o_c, o_d, o_h, o_w} std::vector output_shape_vec(framework::vectorize(output->dims())); // use col_shape in the im2col calculation // col_shape_vec: {i_c/g, k_h, k_w, o_h, o_w} or {i_c/g, k_d, k_h, k_w, o_d, // o_h, o_w} size_t data_dim = filter_shape_vec.size() - 2; std::vector 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)); // use col_matrix_shape in the gemm calculation // size: (i_c/g * k_h * k_w, o_h * o_w) or (i_c/g * k_d * k_h * k_w, o_d * // o_h * o_w) 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; // col_matrix shares the same piece of data with col, // but will be reshaped into a two-dimensional matrix shape // to call the matrix multiplication interface. Tensor col_matrix; if (is_expand) { col.mutable_data(col_shape, context.GetPlace()); col_matrix.ShareDataWith(col); col_matrix.Resize(col_matrix_shape); } framework::DDim input_shape = framework::slice_ddim( input->dims(), 1, static_cast(input->dims().size())); framework::DDim filter_matrix_shape = {filter.dims()[0], filter.numel() / filter.dims()[0]}; filter.Resize(filter_matrix_shape); 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(input->dims()[1]) / groups; int out_step = static_cast(output->dims()[1]) / groups; math::Vol2ColFunctor vol2col; math::Im2ColFunctor im2col; auto& dev_ctx = context.template device_context(); auto blas = math::GetBlas(dev_ctx); 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); 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(dev_ctx, in_slice, dilations, strides, std::vector{paddings[0], paddings[1], paddings[0], paddings[1]}, &col); } else if (data_dim == 3U) { // vol2col vol2col(dev_ctx, 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); blas.MatMul(filter_slice, false, col_matrix, false, T(1.0), &out_slice, T(0.0)); } } } }; template class GemmConvGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { const Tensor* input = context.Input("Input"); const Tensor* output_grad = context.Input(framework::GradVarName("Output")); Tensor* input_grad = context.Output(framework::GradVarName("Input")); Tensor* filter_grad = context.Output(framework::GradVarName("Filter")); // The filter and filter_grad will be reshaped in the calculations, // so here use an assignment operation, // that avoids modifying the variable in the Scope. Tensor filter = *context.Input("Filter"); if (!input_grad && !filter_grad) return; int groups = context.Attr("groups"); std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); std::vector dilations = context.Attr>("dilations"); const int batch_size = static_cast(input->dims()[0]); // filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w} std::vector filter_shape_vec(framework::vectorize(filter.dims())); // output_shape_vec: {o_n, o_c, o_h, o_w} or {o_n, o_c, o_d, o_h, o_w} std::vector output_shape_vec( framework::vectorize(output_grad->dims())); // use col_shape in the im2col calculation // col_shape_vec: {i_c/g, k_h, k_w, o_h, o_w} or {i_c/g, k_d, k_h, k_w, o_d, // o_h, o_w} size_t data_dim = filter_shape_vec.size() - 2; std::vector 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)); // use col_matrix_shape in the gemm calculation // size: (i_c/g * k_h * k_w, o_h * o_w) // or // (i_c/g * k_d * k_h * k_w, o_d * o_h * o_w) framework::DDim col_matrix_shape = framework::flatten_to_2d(col_shape, data_dim + 1); framework::DDim input_shape = framework::slice_ddim( input->dims(), 1, static_cast(input->dims().size())); framework::DDim filter_matrix_shape = {filter.dims()[0], filter.numel() / filter.dims()[0]}; filter.Resize(filter_matrix_shape); framework::DDim output_matrix_shape = { output_grad->dims()[1], output_grad->numel() / (output_grad->dims()[0] * output_grad->dims()[1])}; // convolution backward input operator: gemm + col2im(or col2vol) // convolution backward weight operator: im2col(or vol2col) + gemm int in_step = static_cast(input->dims()[1]) / groups; int out_step = static_cast(output_grad->dims()[1]) / groups; bool is_expand = IsExpand(filter_shape_vec, strides, paddings, dilations); Tensor col; // col_matrix shares the same piece of data with col, // but will be reshaped into a two-dimensional matrix shape // to call the matrix multiplication interface. Tensor col_matrix; if (is_expand) { col.mutable_data(col_shape, context.GetPlace()); col_matrix.ShareDataWith(col); col_matrix.Resize(col_matrix_shape); } math::SetConstant set_zero; auto& dev_ctx = context.template device_context(); auto blas = math::GetBlas(dev_ctx); if (input_grad) { input_grad->mutable_data(context.GetPlace()); // if is_expand is false, the operation of set_zero is unnecessary, // because math::matmul will reset input_grad. if (is_expand) { set_zero(dev_ctx, input_grad, static_cast(0)); } math::Col2VolFunctor col2vol; math::Col2ImFunctor col2im; for (int i = 0; i < batch_size; i++) { Tensor out_grad_batch = output_grad->Slice(i, i + 1).Resize(output_matrix_shape); Tensor in_grad_batch = input_grad->Slice(i, i + 1).Resize(input_shape); for (int g = 0; g < groups; g++) { // gemm Tensor out_grad_slice = out_grad_batch.Slice(g * out_step, (g + 1) * out_step); Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step); Tensor in_grad_slice = in_grad_batch.Slice(g * in_step, (g + 1) * in_step); if (!is_expand) { col_matrix.ShareDataWith(in_grad_slice); col_matrix.Resize(col_matrix_shape); } blas.MatMul(filter_slice, true, out_grad_slice, false, T(1.0), &col_matrix, T(0.0)); if (is_expand && data_dim == 2U) { col2im(dev_ctx, col, dilations, strides, std::vector{paddings[0], paddings[1], paddings[0], paddings[1]}, &in_grad_slice); } else if (is_expand && data_dim == 3U) { col2vol(dev_ctx, col, dilations, strides, paddings, &in_grad_slice); } } } } if (filter_grad) { filter_grad->mutable_data(context.GetPlace()); Tensor filter_grad_ = *filter_grad; filter_grad_.Resize(filter_matrix_shape); set_zero(dev_ctx, filter_grad, static_cast(0)); math::Im2ColFunctor im2col; math::Vol2ColFunctor vol2col; for (int i = 0; i < batch_size; i++) { Tensor out_grad_batch = output_grad->Slice(i, i + 1).Resize(output_matrix_shape); Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape); for (int g = 0; g < groups; g++) { // im2col Tensor out_grad_slice = out_grad_batch.Slice(g * out_step, (g + 1) * out_step); 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(dev_ctx, in_slice, dilations, strides, std::vector{paddings[0], paddings[1], paddings[0], paddings[1]}, &col); } else if (data_dim == 3U) { vol2col(dev_ctx, in_slice, dilations, strides, paddings, &col); } // gemm Tensor filter_grad_slice = filter_grad_.Slice(g * out_step, (g + 1) * out_step); blas.MatMul(out_grad_slice, false, col_matrix, true, T(1.0), &filter_grad_slice, T(1.0)); } } } } }; template class DepthwiseConvKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { const Tensor* input = context.Input("Input"); Tensor filter = *context.Input("Filter"); Tensor* output = context.Output("Output"); output->mutable_data(context.GetPlace()); PADDLE_ENFORCE_EQ( output->dims()[1] % input->dims()[1], 0, "The output channels must be a multiple of the input channels"); std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); std::vector dilations = context.Attr>("dilations"); math::DepthwiseConvFunctor depthwiseConv; auto& dev_ctx = context.template device_context(); depthwiseConv(dev_ctx, *input, filter, strides, paddings, dilations, output); } }; template class DepthwiseConvGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { const Tensor* input = context.Input("Input"); const Tensor* output_grad = context.Input(framework::GradVarName("Output")); Tensor* input_grad = context.Output(framework::GradVarName("Input")); Tensor* filter_grad = context.Output(framework::GradVarName("Filter")); Tensor filter = *context.Input("Filter"); if (!input_grad && !filter_grad) return; std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); std::vector dilations = context.Attr>("dilations"); math::SetConstant set_zero; auto& dev_ctx = context.template device_context(); math::DepthwiseConvInputGradFunctor depthwiseConvInputGrad; math::DepthwiseConvFilterGradFunctor depthwiseConvFilterGrad; if (input_grad) { input_grad->mutable_data(context.GetPlace()); set_zero(dev_ctx, input_grad, static_cast(0)); depthwiseConvInputGrad(dev_ctx, *input, filter, *output_grad, strides, paddings, dilations, input_grad); } if (filter_grad) { filter_grad->mutable_data(context.GetPlace()); set_zero(dev_ctx, filter_grad, static_cast(0)); depthwiseConvFilterGrad(dev_ctx, *input, *output_grad, strides, paddings, dilations, filter_grad); } } }; } // namespace operators } // namespace paddle