---

前引

---

这个课程真的很好 通过自己不停的去写代码 复现的过程中
会发现如果自己去复现 必须要自己完整理解核心过程 才能实现出来 从中又能学到很多东西

我觉得一句话特别好 纸上得来终觉浅 绝知此事要躬行
就是这样的

---

从零开始的推荐系统学习之路（十一）---- 动手学深度学习系列 从零实现L2正则化 & DropOut

---

1、L2正则化

---

1、理论部分

---

---

2、从零实现L2正则化（裸代码）

---

通常放在loss里面 如果手动实现 或者 放在trainer也可以。。

%matplotlib inline
import torch
import numpy
from torch import nn
from d2l import torch as d2l


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)

# train_data, test_data


def InitWeight():
    w, b = torch.normal(0, 0.1, true_w.shape, requires_grad=True), torch.zeros(1, requires_grad=True)
    return w, b


def L2(w):
    return 0.5 * torch.sum(w ** 2)


def MSELoss(y_hat, y):
    return torch.mean(((y_hat - y) ** 2), )

def train(lambd):
    w, b = InitWeight()
    trainer = torch.optim.SGD(params=[w, b], lr=0.01)
    net, loss = lambda X: d2l.linreg(X, w, b), MSELoss
    num_epochs = 100
    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(w)
            trainer.zero_grad()
            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范数是：', torch.norm(w).item())


train(lambd=0)


train(lambd=3)

---

---

3、从零实现L2正则化（pytorch代码）

---

# next(iter(train_iter))[1].shape
net = nn.Sequential(nn.Linear(200, 1))
loss = nn.functional.mse_loss


def train(lambd):
    w, b = InitWeight()
    trainer = torch.optim.SGD(params=net.parameters(), lr=0.01, weight_decay=lambd)
    num_epochs = 100
    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:
            l = loss(net(X), y)
            trainer.zero_grad()
            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范数是：{}'.format(torch.norm(net[0].weight.data).item()))


train(lambd=0)

train(lambd=3)

---

2、DropOut

---

1、理论部分

---

---

2、从零实现DropOut（裸代码）

---

这里需要用 之前写的 fashion那部分代码

def dropout_layer(X, dropout):
    assert dropout >= 0 and dropout <= 1
    if dropout == 1:
        return torch.zeros_like(X)
    elif dropout == 0:
        return X
    else :
        mask = (torch.rand_like(X) > dropout).float()
        return mask * X / (1 - dropout)

dropout_layer(torch.rand(1, 10).reshape(-1, 5), 0)



num_inputs, num_outputs, num_hiddens1, num_hiddens2 = 784, 10, 256, 256
dropout1, dropout2 = 0.1, 0.5

class Net(nn.Module):
    def __init__(self, num_inputs, num_outputs, num_hiddens1, num_hiddens2, is_training = True):
        super(Net, self).__init__()
        self.num_inputs = num_inputs
        self.training = is_training
        
        self.lin1 = nn.Linear(num_inputs, num_hiddens1)
        self.lin2 = nn.Linear(num_hiddens1, num_hiddens2)
        self.lin3 = nn.Linear(num_hiddens2, num_outputs)
        self.relu = nn.ReLU()

    def forward(self, X):
        h1 = self.relu(self.lin1(X.reshape(-1, num_inputs)))        
        if self.training:
            h1 = dropout_layer(h1, dropout1)
        h2 = self.relu(self.lin2(h1))
        if self.training:
            h2 = dropout_layer(h2, dropout2)
        output = self.lin3(h2)

        return output


net = Net(num_inputs, num_outputs, num_hiddens1, num_hiddens2)

num_epochs, lr, batch_size = 10, 0.5, 256
loss = nn.CrossEntropyLoss(reduction='none')
train_iter, test_iter = load_data_fashion_mnist(batch_size)
trainer = torch.optim.SGD(net.parameters(), lr=lr)
train_ch3(net, train_iter, test_iter, loss, num_epochs, trainer)

---

---

3、从零实现DropOut（pytorch代码）

---

net = nn.Sequential(nn.Flatten(),
                    nn.Linear(num_inputs, num_hiddens1),
                    nn.ReLU(),
                    nn.Dropout(dropout1),
                    nn.Linear(num_hiddens1, num_hiddens2),
                    nn.ReLU(),
                    nn.Dropout(dropout2),
                    nn.Linear(num_hiddens2, num_outputs))
trainer = torch.optim.SGD(net.parameters(), lr=lr)
loss = nn.CrossEntropyLoss(reduction='none')
train_ch3(net, train_iter, test_iter, loss, num_epochs, trainer)