4.5.权重衰减

1
def l2_penalty(w):
2
    return torch.sum(w.pow(2)) / 2
3

4
def train(lambd):
5
    w, b = init_params()
6
    net, loss = lambda X: d2l.linreg(X, w, b), d2l.squared_loss
7
    num_epochs, lr = 100, 0.003
8
    animator = d2l.Animator(xlabel='epochs', ylabel='loss', yscale='log',
9
                            xlim=[5, num_epochs], legend=['train', 'test'])
10
    for epoch in range(num_epochs):
11
        for X, y in train_iter:
12
            # 增加了L2范数惩罚项，
13
            # 广播机制使l2_penalty(w)成为一个长度为batch_size的向量
14
            l = loss(net(X), y) + lambd * l2_penalty(w)
15
            l.sum().backward()
5 collapsed lines
16
            d2l.sgd([w, b], lr, batch_size)
17
        if (epoch + 1) % 5 == 0:
18
            animator.add(epoch + 1, (d2l.evaluate_loss(net, train_iter, loss),
19
                                     d2l.evaluate_loss(net, test_iter, loss)))
20
    print('w的L2范数是：', torch.norm(w).item())

1
def train_concise(wd):
2
    net = nn.Sequential(nn.Linear(num_inputs, 1))
3
    for param in net.parameters():
4
        param.data.normal_()
5
    loss = nn.MSELoss(reduction='none')
6
    num_epochs, lr = 100, 0.003
7
    # 偏置参数没有衰减
8
    trainer = torch.optim.SGD([
9
        {"params":net[0].weight,'weight_decay': wd},
10
        {"params":net[0].bias}], lr=lr)
11
    animator = d2l.Animator(xlabel='epochs', ylabel='loss', yscale='log',
12
                            xlim=[5, num_epochs], legend=['train', 'test'])
13
    for epoch in range(num_epochs):
14
        for X, y in train_iter:
15
            trainer.zero_grad()
9 collapsed lines
16
            # loss 上面 loss = nn.MSELoss
17
            l = loss(net(X), y)
18
            l.mean().backward()
19
            trainer.step()
20
        if (epoch + 1) % 5 == 0:
21
            animator.add(epoch + 1,
22
                         (d2l.evaluate_loss(net, train_iter, loss),
23
                          d2l.evaluate_loss(net, test_iter, loss)))
24
    print('w的L2范数：', net[0].weight.norm().item())