丢弃法/暂退法（Dropout）

1.实现dropout_layer函数，以dropout概率丢弃张量输入x元素

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

def dropout_layer(X, dropout):
    assert 0 <= dropout <= 1
    if dropout == 1:
        return torch.zeros_like(X)
    if dropout == 0:
        return X
    mask = (torch.rand(X.shape) > dropout).float()
return mask * X / (1.0 - dropout)

2.测试dropout_layer函数

X= torch.arange(16, dtype = torch.float32).reshape((2, 8))
print(X)
print(dropout_layer(X, 0.))
print(dropout_layer(X, 0.5))
print(dropout_layer(X, 1.))

3.定义具有两个隐藏层的多层感知机，每个隐藏层包含256个单元

num_inputs, num_outputs, num_hiddens1, num_hiddens2 = 784, 10, 256, 256

dropout1, dropout2 = 0.2, 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, self.num_inputs))))
        # 只有在训练模型时才使用dropout
        if self.training == True:
            # 在第一个全连接层之后添加一个dropout层
            H1 = dropout_layer(H1, dropout1)
        H2 = self.relu(self.lin2(H1))
        if self.training == True:
            # 在第二个全连接层之后添加一个dropout层
            H2 = dropout_layer(H2, dropout2)
        out = self.lin3(H2)
        return out

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

4.简洁实现

net = nn.Sequential(nn.Flatten(),
        nn.Linear(784, 256),
        nn.ReLU(),
        # 在第一个全连接层之后添加一个dropout层
        nn.Dropout(dropout1),
        nn.Linear(256, 256),
        nn.ReLU(),
        # 在第二个全连接层之后添加一个dropout层
        nn.Dropout(dropout2),
        nn.Linear(256, 10))

def init_weights(m):
    if type(m) == nn.Linear:
        nn.init.normal_(m.weight, std=0.01)

net.apply(init_weights);