1 结构以及模型退化问题

上述为VGGNet的结构图，左边是VGG16D的步骤 16代表16层，有参数的层，其他结构也是如此

对于这6个机构来说，D中的VGG16是最优的，而VGG19相比较而言，没有VGG16优，VGG16有16层（包括13个卷积层和3个全连接层），而VGG19有19层（包括16个卷积层和3个全连接层），对于出现了的模型退化问题，有几个原因：

1 网络层次越深，提取的特征越是高度抽象，这可以帮助模型理解复杂的、高级的模式。然而，如果网络过深，最后几层可能会从数据中提取过于抽象的特征，与具体的输入图像的关联性减弱，导致模型性能降低

2 在非常深的网络中，梯度在反向传播过程中可能会因为多层传递而变得非常小（梯度消失）或非常大（梯度爆炸），这会使得网络难以训练

3 更深的网络通常有更多的参数，这可能会使模型更容易发生过拟合，特别是当训练数据有限时

等等一些原因

2 代码复现（6种结构的复现）

只是简单的复现

import torch.nn as nn
import torchclass VggBlock(nn.Module):def __init__(self, in_channel, out_channel, n, use_11 = False, LRN = False) -> None:super().__init__()layers = []for i in range(n):if use_11 and (i == n-1):kernel_size = (1, 1)padding = 0    else:kernel_size = (3, 3)padding = 1conv = nn.Sequential(nn.Conv2d(in_channel, out_channel, kernel_size=kernel_size, stride=1, padding=padding),nn.ReLU())layers.append(conv)in_channel = out_channelif LRN == True:layers.append(nn.LocalResponseNorm(size=5))layers.append(nn.MaxPool2d(2))self.block = nn.Sequential(*layers)def forward(self, x):return self.block(x)class VGGNet(nn.Module):def __init__(self, num_classes, features, classify_input_channel) -> None:super().__init__()self.num_classes = num_classesself.features = featuresself.pooling = nn.AdaptiveMaxPool2d(7)self.classify = nn.Sequential(nn.Linear(in_features=7 * 7 * classify_input_channel, out_features=4096),nn.ReLU(),nn.Linear(in_features=4096, out_features=4096),nn.ReLU(),nn.Linear(in_features=4096, out_features=self.num_classes),)def forward(self, images):z = self.features(images)z = self.pooling(z)z = z.flatten(1)return self.classify(z) class VGGNet11A(nn.Module):def __init__(self, num_classes) -> None:super().__init__()features = nn.Sequential(VggBlock(3, 64, 1),VggBlock(64, 128, 1),VggBlock(128, 256, 2),VggBlock(256, 512, 2),VggBlock(512, 512, 2))self.vggnet = VGGNet(num_classes, features, classify_input_channel=512)def forward(self, images):return self.vggnet(images)class VGGNet11ALRN(nn.Module):def __init__(self, num_classes) -> None:super().__init__()features = nn.Sequential(VggBlock(3, 64, 1, LRN=True),VggBlock(64, 128, 1),VggBlock(128, 256, 2),VggBlock(256, 512, 2),VggBlock(512, 512, 2))self.vggnet = VGGNet(num_classes, features, classify_input_channel=512)def forward(self, images):return self.vggnet(images)class VGGNet13B(nn.Module):def __init__(self, num_classes) -> None:super().__init__()features = nn.Sequential(VggBlock(3, 64, 2),VggBlock(64, 128, 2),VggBlock(128, 256, 2),VggBlock(256, 512, 2),VggBlock(512, 512, 2))self.vggnet = VGGNet(num_classes, features, classify_input_channel=512)def forward(self, images):return self.vggnet(images)class VGGNet16C(nn.Module):def __init__(self, num_classes) -> None:super().__init__()features = nn.Sequential(VggBlock(3, 64, 2),VggBlock(64, 128, 2),VggBlock(128, 256, 3, use_11=True),VggBlock(256, 512, 3, use_11=True),VggBlock(512, 512, 3, use_11=True))self.vggnet = VGGNet(num_classes, features, classify_input_channel=512)def forward(self, images):return self.vggnet(images)class VGGNet16C(nn.Module):def __init__(self, num_classes) -> None:super().__init__()features = nn.Sequential(VggBlock(3, 64, 2),VggBlock(64, 128, 2),VggBlock(128, 256, 3),VggBlock(256, 512, 3),VggBlock(512, 512, 3))self.vggnet = VGGNet(num_classes, features, classify_input_channel=512)def forward(self, images):return self.vggnet(images)class VGGNet16C(nn.Module):def __init__(self, num_classes) -> None:super().__init__()features = nn.Sequential(VggBlock(3, 64, 2),VggBlock(64, 128, 2),VggBlock(128, 256, 4),VggBlock(256, 512, 4),VggBlock(512, 512, 4))self.vggnet = VGGNet(num_classes, features, classify_input_channel=512)def forward(self, images):return self.vggnet(images)