python
import torch
from torch import nn
from d2l import torch as d2l7.2.1 VGG 块
AlexNet 没有提供一个通用的模板来指导后续的研究人员设计新的网络，如今研究人员转向了块的角度思考问题。通过使用循环和子程序，可以很容易地在任何现代深度学习框架的代码中实现这些重复的架构。
经典的卷积神经网络的基本组成部分如下：
- 带填充以保持分辨率的卷积层
- 非线性激活层
- 汇聚层VGG 块与之类似，由一系列卷积层组成，再加上用于空间降采样的汇聚层。
python
def vgg_block(num_convs, in_channels, out_channels): # 卷积核数量，输入通道数，输出通道数
layers = []
for _ in range(num_convs): # 加入卷积层
layers.append(nn.Conv2d(in_channels, out_channels,
kernel_size=3, padding=1))
layers.append(nn.ReLU())
in_channels = out_channels
layers.append(nn.MaxPool2d(kernel_size=2,stride=2)) # 加入池化层
return nn.Sequential(*layers)为什么 VGG 块选择堆 3 × 3 3\times 3 3×3 而不是 5 × 5 5\times 5 5×5？
主要是因为在同样的算力下，堆层数多但是核小的 3 × 3 3\times 3 3×3 要比堆核大但是层数少的 5 × 5 5\times 5 5×5 效果要好。简言之，卷积层深度的影响要大于卷积核大小的影响。
7.2.2 VGG 网络
VGG 网络可以分为两部分：
- 第一部分由卷积层和汇聚层构成
- 第二部分由全连接层组成
python
conv_arch = ((1, 64), (1, 128), (2, 256), (2, 512), (2, 512))
def vgg(conv_arch):
conv_blks = []
in_channels = 1
# 卷积层部分
for (num_convs, out_channels) in conv_arch:
conv_blks.append(vgg_block(num_convs, in_channels, out_channels))
in_channels = out_channels
return nn.Sequential(
*conv_blks, nn.Flatten(),
# 全连接层部分
nn.Linear(out_channels * 7 * 7, 4096), nn.ReLU(), nn.Dropout(0.5),
nn.Linear(4096, 4096), nn.ReLU(), nn.Dropout(0.5),
nn.Linear(4096, 10))
net = vgg(conv_arch)
python
X = torch.randn(size=(1, 1, 224, 224))
for blk in net:
X = blk(X)
print(blk.__class__.__name__,'output shape:\t',X.shape)Sequential output shape: torch.Size([1, 64, 112, 112])
Sequential output shape: torch.Size([1, 128, 56, 56])
Sequential output shape: torch.Size([1, 256, 28, 28])
Sequential output shape: torch.Size([1, 512, 14, 14])
Sequential output shape: torch.Size([1, 512, 7, 7])
Flatten output shape: torch.Size([1, 25088])
Linear output shape: torch.Size([1, 4096])
ReLU output shape: torch.Size([1, 4096])
Dropout output shape: torch.Size([1, 4096])
Linear output shape: torch.Size([1, 4096])
ReLU output shape: torch.Size([1, 4096])
Dropout output shape: torch.Size([1, 4096])
Linear output shape: torch.Size([1, 10])7.2.3 训练模型
python
ratio = 4
small_conv_arch = [(pair[0], pair[1] // ratio) for pair in conv_arch] # 构建一个通道较少的模型以减少运算量
net = vgg(small_conv_arch)
python
lr, num_epochs, batch_size = 0.05, 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()) # 大约需要三十五分钟，慎跑loss 0.188, train acc 0.929, test acc 0.908
327.1 examples/sec on cuda:0
练习
（1）打印层的尺寸时，我们只看到 8 个结果，而不是 11 个结果。剩余的 3 层信息去哪了？
因为 3 到 8 层是以两两一组的块的形式显示的。
（2）与 AlexNet 相比，VGG 的计算要慢得多，而且还需要更多的显存。分析出现这种情况的原因。
层数更多，提取的特征也多。
（3）尝试将Fashion-MNIST数据集图像的高度和宽度从224改为96。这对实验有什么影响？
python
conv_arch96 = ((1, 64), (1, 128), (2, 256), (2, 512), (2, 512))
def vgg96(conv_arch):
conv_blks = []
in_channels = 1
for (num_convs, out_channels) in conv_arch:
conv_blks.append(vgg_block(num_convs, in_channels, out_channels))
in_channels = out_channels
return nn.Sequential(
*conv_blks, nn.Flatten(),
nn.Linear(out_channels * 3 * 3, 4096), nn.ReLU(), nn.Dropout(0.5),
nn.Linear(4096, 4096), nn.ReLU(), nn.Dropout(0.5),
nn.Linear(4096, 10))
ratio = 4
small_conv_arch = [(pair[0], pair[1] // ratio) for pair in conv_arch96]
net96 = vgg96(small_conv_arch)
lr, num_epochs, batch_size = 0.05, 10, 128
train_iter96, test_iter96 = d2l.load_data_fashion_mnist(batch_size, resize=96)
d2l.train_ch6(net96, train_iter96, test_iter96, num_epochs, lr, d2l.try_gpu()) # 快多了loss 0.218, train acc 0.919, test acc 0.904
1496.5 examples/sec on cuda:0
训练速度大幅加快，模型精度略有损失。
（4）请参考VGG论文 (Simonyan and Zisserman, 2014)中的表1构建其他常见模型，如VGG-16或VGG-19。
跑不了一点，累了，略。