使用块的网络VGG
1.AlexNet比LeNet更深更大来得到精度,能不能更深和更大?
选项:
①更多的全连接层(太贵)
②更多的卷积层
③将卷积层组合成块
2.VGG块的选择
①深或者宽?
5*5卷积
3*3卷积
选择深且窄的效果更好
②VGG块
3*3卷积(填充1):学习到更多的特征且参数更少(n层,m通道)
2*2最大池化层(步幅2)
3.VGG架构
①多个VGG块后连接成全连接层
②不同次数的重复块得到的不同架构VGG-16,VGG-19
4.VGG进度
①LeNet(1995)
2卷积+池化层
2全连接层
②AlexNet
更大更深
ReLu,DropOut数据增强
③VGG
更大更深的AlexNet(重复的VGG块)
【总结】
①VGG使用可重复使用的卷积块来构建深度卷积神经网络
②不同的卷积块个数和超参数可以得到不同复杂度的变种。
【代码实现】
import torch
from torch import nn
from d2l import torch as d2l
# 函数由三个参数 卷积层数量num_convs,输入通道in_channels,输出通道out_channels
# VGG块
def vgg_block(num_convs, in_channels, out_channels):
layers = []
for _ in range(num_convs):
layers.append(nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1))
layers.append(nn.ReLU())
in_channels = out_channels # 上一次的输出是下一次的输入
layers.append(nn.MaxPool2d(kernel_size=2, stride=2))
return nn.Sequential(*layers)
# VGG网络 超参数变量conv_arch:每个VGG块里卷积层个数和输出通道数
conv_arch = ((1, 64), (1, 128), (2, 256), (2, 512), (2, 512))
# 定义VGG-11
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)
# 模型训练
ratio = 4
small_conv_arch = [(pair[0], pair[1] // ratio) for pair in conv_arch]
net = vgg(small_conv_arch)
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())
d2l.plt.show()