/* 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/eigen.h" #include "paddle/framework/op_registry.h" namespace paddle { namespace operators { template class ActivationKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* X = context.Input("X"); auto* Y = context.Output("Y"); Y->mutable_data(context.GetPlace()); auto x = framework::EigenVector::Flatten(*X); auto y = framework::EigenVector::Flatten(*Y); auto place = context.GetEigenDevice(); Functor functor; functor(place, x, y); } }; template class ActivationGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* X = context.Input("X"); auto* Y = context.Input("Y"); auto* dY = context.Input(framework::GradVarName("Y")); auto* dX = context.Output(framework::GradVarName("X")); dX->mutable_data(context.GetPlace()); auto dy = framework::EigenVector::Flatten(*dY); auto x = framework::EigenVector::Flatten(*X); auto y = framework::EigenVector::Flatten(*Y); auto dx = framework::EigenVector::Flatten(*dX); auto place = context.GetEigenDevice(); Functor functor; functor(place, x, y, dy, dx); } }; struct SigmoidFunctor { template void operator()(Device d, X x, Y y) { y.device(d) = 1. / (1. + (-x).exp()); } }; struct SigmoidGradFunctor { template void operator()(Device d, X x, Y y, dY dy, dX dx) { dx.device(d) = dy * y * (1. - y); } }; struct ExpFunctor { template void operator()(Device d, X x, Y y) { y.device(d) = x.exp(); } }; struct ExpGradFunctor { template void operator()(Device d, X x, Y y, dY dy, dX dx) { dx.device(d) = y; } }; template struct ReluFunctor { template void operator()(Device d, X x, Y y) { y.device(d) = x.cwiseMax(static_cast(0)); } }; template struct ReluGradFunctor { template void operator()(Device d, X x, Y y, dY dy, dX dx) { dx.device(d) = dy * (x > static_cast(0)).template cast(); } }; struct TanhFunctor { template void operator()(Device d, X x, Y y) { y.device(d) = x.tanh(); } }; template struct TanhGradFunctor { template void operator()(Device d, X x, Y y, dY dy, dX dx) { dx.device(d) = dy * (T(1) - y * y); } }; struct SqrtFunctor { template void operator()(Device d, X x, Y y) { y.device(d) = x.sqrt(); } }; template struct SqrtGradFunctor { template void operator()(Device d, X x, Y y, dY dy, dX dx) { const T y_conj = Eigen::numext::conj(y); dx.device(d) = static_cast(0.5) * dy / y_conj; } }; } // namespace operators } // namespace paddle