权重衰退简洁实现

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

def train_concise(wd):     net = nn.Sequential(nn.Linear(num_inputs, 1))           for param in net.parameters():         param.data.normal_()           loss = nn.MSELoss()     num_epochs, lr = 100, 0.003           # 随机梯度下降     trainer = torch.optim.SGD([{         "params": net[0].weight,                   # weight_decay：就是那个超参数         'weight_decay': wd}, {             "params": net[0].bias}], lr=lr)           animator = d2l.Animator(xlabel='epochs', ylabel='loss', yscale='log',                             xlim=[5, num_epochs], legend=['train', 'test'])           for epoch in range(num_epochs):         for X, y in train_iter:             with torch.enable_grad():                 trainer.zero_grad()                 l = loss(net(X), y)             l.backward()             trainer.step()         if (epoch + 1) % 5 == 0:             animator.add(epoch + 1, (d2l.evaluate_loss(net, train_iter, loss),                                      d2l.evaluate_loss(net, test_iter, loss)))     print('w的L2范数：', net[0].weight.norm().item())