（动手学习深度学习）第7章使用块的网络---VGG

VGG

总结

• VGG使用可重复使用的卷积块来构建深度卷积神经网络。
• 不同的卷积块个数和超参数可以得到不同复杂度的变种。

VGG代码实现

1. 导入相关库

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

2. 定义网络模型

# 定义VGG块
def vgg_block(num_convs, in_channels, out_channels):
    """
    :param num_convs: 卷积层个数
    :param in_channels: 输入通道数
    :param out_channels: 输出通道数
    :return: 返回具体的VGG——block的每个层
    """
    layers = []
    for _ in range(num_convs):
        # 每个VGG块中具体使用多少个卷积层和Relu
        layers.append(
            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
        )
        layers.append(nn.ReLU())
        in_channels = out_channels

    # 每个VGG块最后加入一个最大池化层：让图像尺寸减半
    layers.append(
        nn.MaxPool2d(kernel_size=2, stride=2)
    )
    return nn.Sequential(
        *layers
        # *作用在形参上，代表这个位置接收任意多个非关键字参数，转化成元组方式；
        # *作用在实参上，代表的是将输入迭代器拆成一个个元素。
    )

# 定义VGG网络模型
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(),  # [1, channels, h ,w] --->[1, channels*h*w]

        nn.Linear(out_channels * 7 * 7, 4096), nn.ReLU(),  # [1, channels*h*w] -->[1, 4096]
        nn.Dropout(0.5),
        nn.Linear(4096, 4096), nn.ReLU(),  # [1, 4096] --> [1, 4096]
        nn.Dropout(0.5),

        nn.Linear(4096, 10)  # [1, 4096] --> [1, 10]  10分类任务
    )

# # 使用VGG16, 训练精度反而下降
# conv_arch = ((2, 64), (2, 128), (3, 256), (3, 512), (3, 512)) # (卷积层个数， 输出通道数)

#使用VGG11
conv_arch = ((1, 64), (1, 128), (2, 256), (2, 512), (2, 512))

net = vgg(conv_arch)

3. 查看网络模型

X = torch.randn(1, 1, 224, 224)
for blk in net:
    X = blk(X)
    print(blk.__class__.__name__, 'output shape:\t', X.shape)

4. 修改模型

• 由于VGG-16比AlexNet计算量更大，因此构建一个通道较少的网络

ratio = 4  # 吧所有的通道除以4
small_conv_arch = [(pair[0], pair[1]// ratio) for pair in conv_arch]
small_net = vgg(small_conv_arch)

5. 加载数据集

batch_size = 256
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size, resize=224)

6. 训练模型

lr, num_epochs = 0.05, 10
d2l.train_ch6(small_net, train_iter, test_iter,num_epochs, lr, d2l.try_gpu())