/* 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/op_registry.h" #include "paddle/operators/math/context_project.h" #include "paddle/operators/math/math_function.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; template class SequenceConvKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in = context.Input("X"); auto* out = context.Output("Out"); auto filter = *context.Input("Filter"); out->mutable_data(context.GetPlace()); context.ShareLoD("X", "Out"); int context_start = context.Attr("contextStart"); int context_length = context.Attr("contextLength"); int context_stride = context.Attr("contextStride"); bool padding_trainable = context.Attr("paddingTrainable"); PADDLE_ENFORCE_EQ(in->lod().size(), 1UL, "Only support one level sequence now."); const Tensor* padding_data = nullptr; if (padding_trainable) { padding_data = context.Input("PaddingData"); } int up_pad = std::max(0, -context_start); int down_pad = std::max(0, context_start + context_length - 1); int sequence_width = static_cast(in->dims()[1]); framework::DDim col_shape = {in->dims()[0], context_length * sequence_width}; Tensor col; col.mutable_data(col_shape, context.GetPlace()); // Because if padding_trainable is false, padding data should be zeros. math::SetConstant set_zero; auto& dev_ctx = context.template device_context(); set_zero(dev_ctx, &col, static_cast(0)); math::ContextProjectFunctor seq_project_functor; seq_project_functor(dev_ctx, *in, *padding_data, padding_trainable, context_start, context_length, context_stride, up_pad, down_pad, &col); math::matmul(dev_ctx, col, false, filter, false, static_cast(1.0), out, static_cast(0.0)); } }; template class SequenceConvGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* in_g = context.Output(framework::GradVarName("X")); auto* out_g = context.Input(framework::GradVarName("Out")); auto* filter_g = context.Output(framework::GradVarName("Filter")); auto* padding_data_g = context.Output(framework::GradVarName("PaddingData")); auto* in = context.Input("X"); auto* filter = context.Input("Filter"); int context_start = context.Attr("contextStart"); int context_length = context.Attr("contextLength"); int context_stride = context.Attr("contextStride"); bool padding_trainable = context.Attr("paddingTrainable"); PADDLE_ENFORCE_EQ(in->lod().size(), 1UL, "Only support one level sequence now."); auto lod_g_level_0 = in->lod()[0]; int up_pad = std::max(0, -context_start); int down_pad = std::max(0, context_start + context_length - 1); int sequence_width = static_cast(in->dims()[1]); math::SetConstant set_zero; auto& dev_ctx = context.template device_context(); // use col_shape in the im2col calculation framework::DDim col_shape = {in->dims()[0], sequence_width * context_length}; Tensor col; if (in_g || filter_g || (padding_trainable && padding_data_g)) { col.mutable_data(col_shape, context.GetPlace()); // Because if padding_trainable is false, padding data should be zeros. set_zero(dev_ctx, &col, static_cast(0)); math::matmul(dev_ctx, *out_g, false, *filter, true, T(1.0), &col, T(1.0)); } math::ContextProjectFunctor seq_project_functor; math::ContextProjectGradFunctor seq_project_grad_functor; if (in_g) { in_g->mutable_data(context.GetPlace()); in_g->set_lod(in->lod()); set_zero(dev_ctx, in_g, static_cast(0)); seq_project_grad_functor(dev_ctx, *in_g, padding_trainable, context_start, context_length, context_stride, up_pad, down_pad, false, true, padding_data_g, &col); } if (padding_trainable && padding_data_g) { padding_data_g->mutable_data(context.GetPlace()); set_zero(dev_ctx, padding_data_g, static_cast(0)); LoDTensor* input = const_cast(in); seq_project_grad_functor( dev_ctx, *input, padding_trainable, context_start, context_length, context_stride, up_pad, down_pad, true, false, padding_data_g, &col); } if (filter_g) { filter_g->mutable_data(context.GetPlace()); set_zero(dev_ctx, filter_g, static_cast(0)); Tensor filter_grad = *filter_g; LoDTensor out_grad = *out_g; const Tensor* padding_data = nullptr; if (padding_trainable) { padding_data = context.Input("PaddingData"); } seq_project_functor(dev_ctx, *in, *padding_data, padding_trainable, context_start, context_length, context_stride, up_pad, down_pad, &col); math::matmul(dev_ctx, col, true, out_grad, false, T(1.0), &filter_grad, T(1.0)); } } }; } // namespace operators } // namespace paddle