余弦学习率实现

xinfeng2005

已于 2023-02-22 12:18:13 修改

阅读量653

点赞数

文章标签：学习 python 机器学习

于 2023-02-17 21:06:51 首次发布

本文链接：https://blog.csdn.net/xinfeng2005/article/details/129093847

版权

from math import cos, pi

def adjust_learning_rate(optimizer, epoch, iteration, num_iter):
    lr = optimizer.param_groups[0]['lr']

    warmup_epoch = 5 if args.warmup else 0
    warmup_iter = warmup_epoch * num_iter
    current_iter = iteration + epoch * num_iter
    max_iter = args.epochs * num_iter

    if args.lr_decay == 'step':
        lr = args.lr * (args.gamma ** ((current_iter - warmup_iter) / (max_iter - warmup_iter)))
    elif args.lr_decay == 'cos':
        lr = args.lr * (1 + cos(pi * (current_iter - warmup_iter) / (max_iter - warmup_iter))) / 2
    elif args.lr_decay == 'linear':
        lr = args.lr * (1 - (current_iter - warmup_iter) / (max_iter - warmup_iter))
    elif args.lr_decay == 'schedule':
        count = sum([1 for s in args.schedule if s <= epoch])
        lr = args.lr * pow(args.gamma, count)
    else:
        raise ValueError('Unknown lr mode {}'.format(args.lr_decay))

    if epoch < warmup_epoch:
        lr = args.lr * current_iter / warmup_iter

    for param_group in optimizer.param_groups:
        param_group['lr'] = lr