使用块的网络 (VGG)

1. VGG 网络的设计理念

AlexNet 首次证明了模型能自动学习特征的能力，但这一突破并没有为后续的研究提供用于构建新网络的模板。

2. VGG 网络架构

2.1 VGG 块设计

VGG 块是VGG网络的核心构建单元，具有统一而简洁的设计模式：

VGG 块的实现

import torch
from torch import Tensor, nn

class VGGBlock(nn.Module):
    """
    VGG 块
    
    由两层使用 ReLU 激活函数的 3×3 卷积层，后接一个 2×2 步幅为 2 的最大池化层组成
    """
    
    def __init__(self, in_channels: int, out_channels: int):
        super(VGGBlock, self).__init__()
        self.block = nn.Sequential(
            # 第一个3×3卷积层
            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            
            # 第二个3×3卷积层
            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            
            # 2×2最大池化层
            nn.MaxPool2d(kernel_size=2, stride=2)
        )

    def forward(self, x: Tensor) -> Tensor:
        return self.block(x)

2.2 完整的 VGG11 网络

与 AlexNet、LeNet 类似，VGG 网络同样由卷积层、汇聚层组成的特征提取模块和由全连接层组成的分类模块组成。

VGG11 网络实现

class VGG11(nn.Module):
    """VGG11 网络实现"""
    
    def __init__(self, num_classes: int = 10):
        super(VGG11, self).__init__()

        # 特征提取器 - 由5个VGG块组成
        self.features = nn.Sequential(
            VGGBlock(1, 16),    # 输入: 224×224×1 -> 输出: 112×112×16
            VGGBlock(16, 32),   # 输入: 112×112×16 -> 输出: 56×56×32
            VGGBlock(32, 64),   # 输入: 56×56×32 -> 输出: 28×28×64
            VGGBlock(64, 128),  # 输入: 28×28×64 -> 输出: 14×14×128
            VGGBlock(128, 128), # 输入: 14×14×128 -> 输出: 7×7×128
        )

        # 分类器 - 由全连接层组成
        self.classifier = nn.Sequential(
            nn.Dropout(0.5),
            nn.Linear(7 * 7 * 128, 4096),
            nn.ReLU(inplace=True),
            
            nn.Dropout(0.5),
            nn.Linear(4096, 4096),
            nn.ReLU(inplace=True),
            
            nn.Linear(4096, num_classes)
        )

    def forward(self, x: Tensor) -> Tensor:
        # 特征提取
        x = self.features(x)
        
        # 展平特征图
        x = torch.flatten(x, 1)
        
        # 分类
        x = self.classifier(x)
        
        return x

2.3 网络结构分析

使用 torchinfo 库的 summary 函数来详细分析网络结构：

from torchinfo import summary

# 创建模型并分析
model = VGG11(num_classes=10)
summary(model, input_size=(1, 1, 224, 224))

==========================================================================================
Layer (type:depth-idx)                   Output Shape              Param #
==========================================================================================
VGG11                                    [1, 10]                   --
├─Sequential: 1-1                        [1, 128, 7, 7]            --
│    └─VGGBlock: 2-1                     [1, 16, 112, 112]         --
│    │    └─Sequential: 3-1              [1, 16, 112, 112]         2,480
│    └─VGGBlock: 2-2                     [1, 32, 56, 56]           --
│    │    └─Sequential: 3-2              [1, 32, 56, 56]           13,888
│    └─VGGBlock: 2-3                     [1, 64, 28, 28]           --
│    │    └─Sequential: 3-3              [1, 64, 28, 28]           55,424
│    └─VGGBlock: 2-4                     [1, 128, 14, 14]          --
│    │    └─Sequential: 3-4              [1, 128, 14, 14]          221,440
│    └─VGGBlock: 2-5                     [1, 128, 7, 7]            --
│    │    └─Sequential: 3-5              [1, 128, 7, 7]            295,168
├─Sequential: 1-2                        [1, 10]                   --
│    └─Dropout: 2-6                      [1, 6272]                 --
│    └─Linear: 2-7                       [1, 4096]                 25,694,208
│    └─ReLU: 2-8                         [1, 4096]                 --
│    └─Dropout: 2-9                      [1, 4096]                 --
│    └─Linear: 2-10                      [1, 4096]                 16,781,312
│    └─ReLU: 2-11                        [1, 4096]                 --
│    └─Linear: 2-12                      [1, 10]                   40,970
==========================================================================================
Total params: 43,104,890
Trainable params: 43,104,890
Non-trainable params: 0
Total mult-adds (Units.MEGABYTES): 746.44
==========================================================================================
Input size (MB): 0.20
Forward/backward pass size (MB): 24.55
Params size (MB): 172.42
Estimated Total Size (MB): 197.17
==========================================================================================

3. 模型训练与评估

3.1 训练配置

继续使用 training_tools.py 中的工具来训练和评估VGG11模型：

import torch
from torch import nn, optim
from training_tools import fashionMNIST_loader, Trainer

def train_vgg11():
    """训练VGG11模型"""
    # 超参数设置
    BATCH_SIZE = 128
    EPOCHS = 15  # 减少训练轮数用于演示
    LEARNING_RATE = 0.005  # 降低学习率以获得更稳定的训练
    
    # 创建模型和数据加载器
    model = VGG11(num_classes=10)
    train_loader, test_loader = fashionMNIST_loader(BATCH_SIZE, resize=224)
    
    # 损失函数和优化器
    criterion = nn.CrossEntropyLoss()
    optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE, weight_decay=1e-4)
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    
    print(f"使用设备: {device}")
    print(f"模型参数数量: {sum(p.numel() for p in model.parameters()):,}")
    
    # 开始训练
    with Trainer(model, train_loader, test_loader, criterion, optimizer, device) as trainer:
        trainer.train(EPOCHS)

if __name__ == '__main__':
    train_vgg11()

3.2 训练结果分析

使用设备: cuda
模型参数数量: 43,104,890

Epoch [  1/15] Train Loss: 2.3027, Train Acc:  9.90%, Test Loss: 2.3026, Test Acc: 10.00%
Epoch [  2/15] Train Loss: 2.3028, Train Acc:  9.85%, Test Loss: 2.3027, Test Acc: 10.00%
...
Epoch [ 12/15] Train Loss: 2.3027, Train Acc:  9.88%, Test Loss: 2.3025, Test Acc: 10.00%
Epoch [ 13/15] Train Loss: 2.3025, Train Acc: 10.22%, Test Loss: 2.3021, Test Acc: 10.00%
Epoch [ 14/15] Train Loss: 2.2864, Train Acc: 17.02%, Test Loss: 2.3021, Test Acc: 10.00%
Epoch [ 15/15] Train Loss: 1.6031, Train Acc: 38.25%, Test Loss: 0.6503, Test Acc: 74.66%

# 继续训练会看到快速提升：
Epoch [ 16/30] Train Loss: 0.5281, Train Acc: 80.15%, Test Loss: 0.4621, Test Acc: 83.29%
Epoch [ 20/30] Train Loss: 0.2704, Train Acc: 90.00%, Test Loss: 0.2714, Test Acc: 90.12%
Epoch [ 30/30] Train Loss: 0.1174, Train Acc: 95.60%, Test Loss: 0.2576, Test Acc: 92.31%

训练特点分析