pytorch 学习率调整

def adjust_learning_rate(optimizer, lr, gamma, step):
    lr = lr * (gamma ** (step))
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr
    return lr

调用

optimizer = optim.Adam(net.parameters(),lr=lr)
for epoch in range(Start_iter,max_epoch):
	adjust_learning_rate(optimizer,lr,0.9,epoch)
	lr = adjust_learning_rate(optimizer,lr,0.9,epoch)
	print("lr:",lr)

2

def adjust_learning_rate(optimizer, cur_epoch, base_lr, lr_schedule):
    """
    Sets the learning rate
    """
    lr = 0
    for i, e in enumerate(lr_schedule):
        if cur_epoch < e:
            lr = base_lr * (0.1 ** i)
            break
    if lr == 0:
        lr = base_lr
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

    return lr

调用

# update lr
learning_rate = 0.001
lr_schedule = [5, 30, 80, 130] # 最后一个数必须大于max_epoch
lr = adjust_learning_rate(optimizer, epoch, learning_rate, lr_schedule)

3.按步数

def adjust_learning_rate(optimizer, gamma, lr, step):
    """Sets the learning rate to the initial LR decayed by 10 at every
        specified step
    # Adapted from PyTorch Imagenet example:
    # https://github.com/pytorch/examples/blob/master/imagenet/main.py
    """
    lr = lr * (gamma ** (step))
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

调用

epoch_size = len(train_dataset) // train_batch_size
max_iter = max_epoch * epoch_size
# max_iter 可自己设定

lr_steps = [8000,100000,150000]
step_index = 0
for iteration in range(start_iter, max_iter):
    if iteration in cfg['lr_steps']:
        step_index += 1
        adjust_learning_rate(optimizer, 0.1, lr,step_index)
	
	# 加载数据