【一起深度学习——NIN】

原理图：

代码实现：

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

def nin_block(in_channels, out_channels, kernel_size, strides, padding):
    return nn.Sequential(
        #模块化设计思想，方便重复使用。
        #经过一个卷积层，再加两个1*1的卷积层。起到全连接层的作用。
        nn.Conv2d(in_channels,out_channels,kernel_size,strides,padding),
        nn.ReLU(),
        # 两个1*1的卷积层，并不会改变形状，不改变通道数
        nn.Conv2d(out_channels, out_channels, kernel_size=1, stride=1),
        nn.ReLU(),
        nn.Conv2d(out_channels,out_channels,kernel_size=1,stride=1),
        nn.ReLU(),
    )

#nn.Sequential 容器
net = nn.Sequential(
    nin_block(1,96,kernel_size=11,strides=4,padding=0),
    nn.MaxPool2d(kernel_size=3,stride=2),
    nin_block(96,256,kernel_size=5,strides=1,padding=2),
    nn.MaxPool2d(kernel_size=3,stride=2),
    nin_block(256,384,kernel_size=3,strides=1,padding=1),
    nn.MaxPool2d(kernel_size=3,stride=2),
    nn.Dropout(0.5),
    # [1,10,5,5]
    nin_block(384, 10, kernel_size=3, strides=1, padding=1),
    #自适应平均池化。与普通的池化层不同，允许指定输出特征图的大小而不是池化层的大小。
    #[1,10,5,5] => [1,10,1,1]
    nn.AdaptiveAvgPool2d(output_size=(1,1)),
    # [1,10,1,1] =>[1,10*1*1] =>[1,10]
    nn.Flatten()
)

#开始训练
lr, num_epochs, batch_size = 0.1, 10, 128
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size, resize=224)
d2l.train_ch6(net, train_iter, test_iter, num_epochs, lr, d2l.try_gpu())

输出结果：