https://blog.csdn.net/bc521bc/article/details/85864555

这篇bolg说的很详细了，但是具体在代码中怎么用还是有点模糊。自己试验了一下，顺路记一下，其实很简单，在optimizer后面定义一下，然后在每个epoch中step一下就可以了。一开始出错是因为我把step放到

T_optimizer.step()后面了，导致一个epoch后小到看不出来了.

T_optimizer = SGD(net.parameters(), lr=LR, weight_decay=0.0005, momentum=0.9)
scheduler = lr_scheduler.StepLR(T_optimizer, step_size=30, gamma=0.5)

for epoch in enumerate(range(startepoch, startepoch + EPOCH)):

    for anc, pos, neg in Triplet_data:
        net.zero_grad()
        anc_feat = net(anc.to(device))
        pos_feat = net(pos.to(device))
        neg_feat = net(neg.to(device))
        tri_loss = T_loss(anc_feat, pos_feat, neg_feat)
        tri_loss.backward()
        T_optimizer.step()
    ave_loss = np.mean(loss_np)
    scheduler.step()#LR减小