学习率余弦衰减

#定义学习器
self.nf_optimizer = torch.optim.Adam(params, lr=self.nf_conf.lr)

#学习率余弦衰减函数
def adjust_learning_rate(c, optimizer, epoch):
    lr = c.lr
    if c.lr_cosine:
        eta_min = lr * (c.lr_decay_rate ** 3) #学习率衰减下限
        lr = eta_min + (lr - eta_min) * (
                1 + math.cos(math.pi * epoch / c.meta_epochs)) / 2
    else:
        steps = np.sum(epoch >= np.asarray(c.lr_decay_epochs))
        if steps > 0:
            lr = lr * (c.lr_decay_rate ** steps)

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

for epoch in range(1, total_epochs+1):
    #调用
    adjust_learning_rate(self.nf_conf, self.nf_optimizer, epoch)