wehub-resource-sync
111 lines
4.3 KiB
C++
111 lines
4.3 KiB
C++
// Copyright (c) 2024 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 <math.h>
|
|
|
|
#include "paddle/phi/kernels/funcs/eigen/common.h"
|
|
#include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
|
|
#include "paddle/phi/kernels/radam_kernel.h"
|
|
|
|
namespace phi {
|
|
|
|
template <typename T, typename Context>
|
|
void RAdamKernel(const Context& dev_ctx,
|
|
const DenseTensor& param,
|
|
const DenseTensor& grad,
|
|
const DenseTensor& learning_rate,
|
|
const DenseTensor& beta1_pow,
|
|
const DenseTensor& beta2_pow,
|
|
const DenseTensor& rho,
|
|
const DenseTensor& moment1,
|
|
const DenseTensor& moment2,
|
|
const optional<DenseTensor>& master_param UNUSED,
|
|
float beta1,
|
|
float beta2,
|
|
float epsilon,
|
|
bool multi_precision UNUSED,
|
|
DenseTensor* param_out,
|
|
DenseTensor* beta1_pow_out,
|
|
DenseTensor* beta2_pow_out,
|
|
DenseTensor* rho_out,
|
|
DenseTensor* moment1_out,
|
|
DenseTensor* moment2_out,
|
|
DenseTensor* master_param_out UNUSED) {
|
|
dev_ctx.template Alloc<T>(param_out);
|
|
dev_ctx.template Alloc<T>(beta1_pow_out);
|
|
dev_ctx.template Alloc<T>(beta2_pow_out);
|
|
dev_ctx.template Alloc<T>(rho_out);
|
|
dev_ctx.template Alloc<T>(moment1_out);
|
|
dev_ctx.template Alloc<T>(moment2_out);
|
|
|
|
T beta1_ = static_cast<T>(beta1);
|
|
T beta2_ = static_cast<T>(beta2);
|
|
T epsilon_ = static_cast<T>(epsilon);
|
|
|
|
auto eigen_param = EigenVector<T>::Flatten(param);
|
|
auto eigen_grad = EigenVector<T>::Flatten(grad);
|
|
auto eigen_lr = EigenVector<T>::Flatten(learning_rate);
|
|
auto eigen_beta1_pow = EigenVector<T>::Flatten(beta1_pow);
|
|
auto eigen_beta2_pow = EigenVector<T>::Flatten(beta2_pow);
|
|
auto eigen_rho = EigenVector<T>::Flatten(rho);
|
|
auto eigen_moment1 = EigenVector<T>::Flatten(moment1);
|
|
auto eigen_moment2 = EigenVector<T>::Flatten(moment2);
|
|
|
|
auto eigen_param_out = EigenVector<T>::Flatten(*param_out);
|
|
auto eigen_beta1_pow_out = EigenVector<T>::Flatten(*beta1_pow_out);
|
|
auto eigen_beta2_pow_out = EigenVector<T>::Flatten(*beta2_pow_out);
|
|
auto eigen_rho_out = EigenVector<T>::Flatten(*rho_out);
|
|
auto eigen_moment1_out = EigenVector<T>::Flatten(*moment1_out);
|
|
auto eigen_moment2_out = EigenVector<T>::Flatten(*moment2_out);
|
|
|
|
T rho_inf =
|
|
static_cast<T>(2) / (static_cast<T>(1) - beta2_) - static_cast<T>(1);
|
|
|
|
eigen_beta1_pow_out = eigen_beta1_pow * beta1_;
|
|
eigen_beta2_pow_out = eigen_beta2_pow * beta2_;
|
|
eigen_rho_out =
|
|
(eigen_rho * (beta2_ - eigen_beta2_pow_out) + eigen_beta2_pow_out) /
|
|
(static_cast<T>(1) - eigen_beta2_pow_out);
|
|
|
|
eigen_moment1_out =
|
|
beta1_ * eigen_moment1 + (static_cast<T>(1) - beta1_) * eigen_grad;
|
|
eigen_moment2_out = beta2_ * eigen_moment2 +
|
|
(static_cast<T>(1) - beta2_) * eigen_grad * eigen_grad;
|
|
|
|
Eigen::DSizes<int, 1> p_dsize(param_out->numel());
|
|
auto eigen_moment1_hat =
|
|
eigen_moment1_out / (static_cast<T>(1) - eigen_beta1_pow_out);
|
|
|
|
T rho_t = rho_inf - static_cast<T>(2) * eigen_rho_out.data()[0];
|
|
|
|
if (rho_t > static_cast<T>(5)) {
|
|
auto l_t = (static_cast<T>(1) - eigen_beta2_pow_out).sqrt() /
|
|
(eigen_moment2_out.sqrt() + epsilon_);
|
|
auto r_t = std::sqrt(
|
|
((rho_t - static_cast<T>(4)) * (rho_t - static_cast<T>(2)) * rho_inf) /
|
|
((rho_inf - static_cast<T>(4)) * (rho_inf - static_cast<T>(2)) *
|
|
rho_t));
|
|
|
|
eigen_param_out = eigen_param - eigen_lr.broadcast(p_dsize) *
|
|
eigen_moment1_hat * r_t * l_t;
|
|
|
|
} else {
|
|
eigen_param_out =
|
|
eigen_param - eigen_lr.broadcast(p_dsize) * eigen_moment1_hat;
|
|
}
|
|
}
|
|
} // namespace phi
|