动手学深度学习之权重衰退

使用均方范数作为硬性限制

• 通过限制参数值的选择范围来控制模型容量$min l(w,b) \quad subject to \quad \left |w \right |^2 \leq \theta $
• 通常不限制偏移b(限不限制都差不多)
• 小的$\theta$意味着更强的正则项

使用均方范数作为柔性限制

• 对每个$\theta$，都可以找到$\lambda$使得之前的目标函数等价于下面$minl(w,b)+\frac{\lambda }{2}{\Vert w\Vert }^2$
  • 可以通过拉格朗日乘子来证明
• 超参数$\lambda$控制了正则项的重要程度
  • $\lambda =0$：无作用
  • $\lambda \to \infty ,w^{\ast }\to 0$

参数更新法则

• 计算梯度：$\frac{\partial (l(w,b)+\frac{\lambda }{2}{\Vert w\Vert }^2)}{\partial w}=\frac{\partial l(w,b)}{\partial w}+\lambda w$
• 时间t更新参数$w_{t+1}=(1-\eta \lambda )w_t-\eta \frac{\partial l(w_t,b_t)}{\partial w_t}$
  • 通常$\eta \lambda \le 1$，在深度学习中通常叫做权重衰退

总结

• 权重衰退通过L2正则项使得模型参数不会过大，从而控制模型复杂度。
• 正则项权重是控制模型复杂度的超参数

权重衰退

权重衰退是最广泛使用的正则化的技术之一

import torch
from torch import nn
from d2l import torch as d2l

像以前一样生成一些数据
$y=0.05+{\Sigma }_{i=1}^{d}0.01x_i+ϵwhereϵ\sim N(0,0.01^2)$

n_train, n_test, num_inputs, batch_size = 20, 100, 200, 5
true_w, true_b = torch.ones((num_inputs, 1)) * 0.01, 0.05
train_data = d2l.synthetic_data(true_w, true_b, n_train)
train_iter = d2l.load_array(train_data, batch_size)
test_data = d2l.synthetic_data(true_w, true_b, n_test)
test_iter = d2l.load_array(test_data, batch_size, is_train=False)

初始化模型参数

def init_params():
    w = torch.normal(0, 1, size=(num_inputs, 1), requires_grad=True)
    b = torch.zeros(1, requires_grad=True)
    return [w, b]

定义$L_2$范数惩罚

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

定义训练函数

def train(lambd):
    w, b = init_params()
    net, loss = lambda X: d2l.linreg(X, w, b), d2l.squared_loss
    num_epochs, lr = 100, 0.003
    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:
        
            # 增加了L2范数惩罚项，广播机制使l2_penalty(w)成为一个长度为`batch_size`的向量。
            l = loss(net(X), y) + lambd * l2_penalty(w)
            l.sum().backward()
            d2l.sgd([w, b], lr, batch_size)
        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范数是：', torch.norm(w).item())

train(lambd=0)

w的L2范数是： 13.561446189880371

train(lambd=3)

w的L2范数是： 0.3731406033039093

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
    # 偏置参数没有衰减。
    trainer = torch.optim.SGD([{
        "params": net[0].weight,
        '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())

train_concise(0)

w的L2范数： 13.962190628051758

train_concise(3)

w的L2范数： 0.41519707441329956