/* Copyright (c) 2019 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 #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/pten/kernels/funcs/blas/blas.h" #include "paddle/pten/kernels/funcs/math_function.h" namespace paddle { namespace operators { template using EigenTensor = framework::EigenTensor; using Tensor = framework::Tensor; using Array1 = Eigen::DSizes; using Array2 = Eigen::DSizes; using IndexPair = Eigen::IndexPair; template static inline void TransCompute(const int rank, const Tensor& in, Tensor* out, const std::vector& perm, const DeviceContext& dev_ctx) { if (rank <= 1 || rank > 5) { PADDLE_THROW(paddle::platform::errors::Fatal( "Weight rank of SpectralNorm should be in range [2, 5], but got %d.", rank)); } switch (rank) { case 2: pten::funcs::Transpose trans2; trans2(dev_ctx, in, out, perm); break; case 3: pten::funcs::Transpose trans3; trans3(dev_ctx, in, out, perm); break; case 4: pten::funcs::Transpose trans4; trans4(dev_ctx, in, out, perm); break; case 5: pten::funcs::Transpose trans5; trans5(dev_ctx, in, out, perm); break; default: break; } } template static inline void CalcMatrixSigmaAndNormWeight( Tensor* sigma, Tensor* u, Tensor* v, Tensor* weight, const int power_iters, const float eps, const framework::ExecutionContext& ctx) { auto& place = *ctx.template device_context().eigen_device(); auto blas = pten::funcs::GetBlas(ctx); auto sigma_t = EigenTensor::From(*sigma); auto weight_t = EigenTensor::From(*weight); auto u_t = EigenTensor::From(*u); auto v_t = EigenTensor::From(*v); const int h = weight->dims()[0]; const int w = weight->dims()[1]; for (int i = 0; i < power_iters; i++) { // V = W^T * U / ||W^T * U||_2 blas.MatMul(*weight, true, *u, false, T(1), v, T(0)); auto v_t_norm = v_t.square().sum().sqrt().eval().reshape(Array1(1)).broadcast( Array1(w)); v_t.device(place) = v_t / (v_t_norm + v_t_norm.constant(eps)); // U = W^T * V / ||W^T * V||_2 blas.MatMul(*weight, false, *v, false, T(1), u, T(0)); auto u_t_norm = u_t.square().sum().sqrt().eval().reshape(Array1(1)).broadcast( Array1(h)); u_t.device(place) = u_t / (u_t_norm + u_t_norm.constant(eps)); } Tensor weight_v; weight_v.mutable_data({h, 1}, ctx.GetPlace()); blas.MatMul(*weight, false, *v, false, T(1), &weight_v, T(0)); auto weight_v_t = EigenTensor::From(weight_v); sigma_t.device(place) = (u_t * weight_v_t) .sum() .eval() .reshape(Array2(1, 1)) .broadcast(Array2(h, w)); weight_t.device(place) = weight_t / sigma_t; } template class SpectralNormKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto& dev_ctx = ctx.template device_context(); auto weight = ctx.Input("Weight"); auto u = ctx.Input("U"); auto v = ctx.Input("V"); auto out = ctx.Output("Out"); int dim = ctx.Attr("dim"); int power_iters = ctx.Attr("power_iters"); float eps = ctx.Attr("eps"); const int h = u->dims()[0]; const int w = v->dims()[0]; Tensor weight_mat; auto dims = weight->dims(); const int rank = dims.size(); std::vector real_dims; if (dim != 0) { std::vector perm; perm.push_back(dim); real_dims.push_back(dims[dim]); for (int i = 0; i < rank; i++) { if (i != dim) { perm.push_back(i); real_dims.push_back(dims[i]); } } weight_mat.mutable_data(framework::make_ddim(real_dims), ctx.GetPlace()); TransCompute(rank, *weight, &weight_mat, perm, dev_ctx); } else { for (int i = 0; i < rank; i++) { real_dims.push_back(i); } paddle::framework::TensorCopySync(*weight, ctx.GetPlace(), &weight_mat); } weight_mat = weight_mat.Resize({h, w}); Tensor sigma; sigma.mutable_data(weight_mat.dims(), ctx.GetPlace()); Tensor uu, vv; paddle::framework::TensorCopySync(*u, ctx.GetPlace(), &uu); paddle::framework::TensorCopySync(*v, ctx.GetPlace(), &vv); CalcMatrixSigmaAndNormWeight( &sigma, &(uu.Resize({h, 1})), &(vv.Resize({w, 1})), &weight_mat, power_iters, eps, ctx); if (dim != 0) { std::vector perm; for (int i = 0; i < rank; i++) { if (i < dim) { perm.push_back(i + 1); } else if (i == dim) { perm.push_back(0); } else { perm.push_back(i); } } out->mutable_data(dims, ctx.GetPlace()); TransCompute( rank, weight_mat.Resize(framework::make_ddim(real_dims)), out, perm, dev_ctx); } else { paddle::framework::TensorCopySync(weight_mat.Resize(dims), ctx.GetPlace(), out); } } }; template class SpectralNormGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto& place = *ctx.template device_context().eigen_device(); auto& dev_ctx = ctx.template device_context(); auto blas = pten::funcs::GetBlas(ctx); auto weight = ctx.Input("Weight"); auto u = ctx.Input("U"); auto v = ctx.Input("V"); auto out_grad = ctx.Input(framework::GradVarName("Out")); auto weight_grad = ctx.Output(framework::GradVarName("Weight")); int dim = ctx.Attr("dim"); int power_iters = ctx.Attr("power_iters"); float eps = ctx.Attr("eps"); const int h = u->dims()[0]; const int w = v->dims()[0]; Tensor weight_mat, out_grad_mat; auto dims = weight->dims(); const int rank = dims.size(); std::vector real_dims; if (dim != 0) { std::vector perm; perm.push_back(dim); real_dims.push_back(dims[dim]); for (int i = 0; i < rank; i++) { if (i != dim) { perm.push_back(i); real_dims.push_back(dims[i]); } } weight_mat.mutable_data(framework::make_ddim(real_dims), ctx.GetPlace()); out_grad_mat.mutable_data(framework::make_ddim(real_dims), ctx.GetPlace()); TransCompute(rank, *weight, &weight_mat, perm, dev_ctx); TransCompute(rank, *out_grad, &out_grad_mat, perm, dev_ctx); } else { for (int i = 0; i < rank; i++) { real_dims.push_back(i); } paddle::framework::TensorCopySync(*weight, ctx.GetPlace(), &weight_mat); paddle::framework::TensorCopySync(*out_grad, ctx.GetPlace(), &out_grad_mat); } weight_mat = weight_mat.Resize({h, w}); out_grad_mat = out_grad_mat.Resize({h, w}); Tensor sigma; sigma.mutable_data(weight_mat.dims(), ctx.GetPlace()); Tensor uu, vv; paddle::framework::TensorCopySync(*u, ctx.GetPlace(), &uu); paddle::framework::TensorCopySync(*v, ctx.GetPlace(), &vv); CalcMatrixSigmaAndNormWeight( &sigma, &(uu.Resize({h, 1})), &(vv.Resize({w, 1})), &weight_mat, power_iters, eps, ctx); Tensor uv; uv.mutable_data({h, w}, ctx.GetPlace()); blas.MatMul(uu.Resize({h, 1}), false, vv.Resize({w, 1}), false, T(1), &uv, T(0)); Tensor weight_grad_mat; weight_grad_mat.mutable_data({h, w}, ctx.GetPlace()); auto weight_grad_mat_t = EigenTensor::From(weight_grad_mat); auto weight_mat_t = EigenTensor::From(weight_mat); auto out_grad_mat_t = EigenTensor::From(out_grad_mat); auto sigma_t = EigenTensor::From(sigma); auto uv_t = EigenTensor::From(uv); weight_mat_t.device(place) = weight_mat_t.sum().eval().reshape(Array2(1, 1)).broadcast(Array2(h, w)); weight_grad_mat_t.device(place) = out_grad_mat_t * (out_grad_mat_t.constant(1.0) - uv_t * weight_mat_t) / sigma_t; if (dim != 0) { std::vector perm; for (int i = 0; i < rank; i++) { if (i < dim) { perm.push_back(i + 1); } else if (i == dim) { perm.push_back(0); } else { perm.push_back(i); } } weight_grad->mutable_data(dims, ctx.GetPlace()); TransCompute( rank, weight_grad_mat.Resize(framework::make_ddim(real_dims)), weight_grad, perm, dev_ctx); } else { paddle::framework::TensorCopySync(weight_grad_mat.Resize(dims), ctx.GetPlace(), weight_grad); } } }; } // namespace operators } // namespace paddle