使用块的网络 - VGG。

使用多个$3\times 3$的要比使用少个$5\times 5$的效果要好。


VGG全称是Visual Geometry Group，因为是由Oxford的Visual Geometry Group提出的。AlexNet问世之后，很多学者通过改进AlexNet的网络结构来提高自己的准确率，主要有两个方向：小卷积核和多尺度。而VGG的作者们则选择了另外一个方向，即加深网络深度。

网络架构

卷积网络的输入是224 * 224的RGB图像，整个网络的组成是非常格式化的，基本上都用的是3 * 3的卷积核以及 2 * 2的max pooling，少部分网络加入了1 * 1的卷积核。因为想要体现出“上下左右中”的概念，3*3的卷积核已经是最小的尺寸了。

VGG16相比之前网络的改进是3个33卷积核来代替7x7卷积核，2个33卷积核来代替5*5卷积核，这样做的主要目的是在保证具有相同感知野的条件下，减少参数，提升了网络的深度。

多个VGG块后接全连接层。

不同次数的重复块得到不同的架构，如VGG-16，VGG-19等。

VGG：更大更深的AlexNet。

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

代码实现

使用数据集CIFAR

model.py

import torch
from torch import nn

class Vgg16(nn.Module):

    def __init__(self, *args, **kwargs) -> None:
        super().__init__(*args, **kwargs)
        self.model = nn.Sequential(
            nn.Conv2d(3,64,kernel_size=3,padding=1),
            nn.ReLU(),
            nn.Conv2d(64,64,kernel_size=3,padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2,2),
            nn.Conv2d(64,128,kernel_size=3,padding=1),
            nn.ReLU(),
            nn.Conv2d(128,128,kernel_size=3,padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2,2),
            nn.Conv2d(128,256,kernel_size=3,padding=1),
            nn.ReLU(),
            nn.Conv2d(256,256,kernel_size=3,padding=1),
            nn.ReLU(),
            nn.Conv2d(256,256,kernel_size=3,padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2,2),
            nn.Conv2d(256,512,kernel_size=3,padding=1),
            nn.ReLU(),
            nn.Conv2d(512,512,kernel_size=3,padding=1),
            nn.ReLU(),
            nn.Conv2d(512,512,kernel_size=3,padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2,2),
            nn.Conv2d(512, 512, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv2d(512, 512, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv2d(512, 512, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2,2),
            nn.Flatten(),
            nn.Linear(7*7*512,4096),
            nn.Dropout(0.5),
            nn.Linear(4096,4096),
            nn.Dropout(0.5),
            nn.Linear(4096,10)
        )

    def forward(self,x):
        return self.model(x)

# 验证模型正确性
if __name__ == '__main__':
    net = Vgg16()
    x = torch.ones((64,3,244,244))
    output = net(x)
    print(output)

train.py

import torch
from torch import nn
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from torchvision import datasets
from torchvision.transforms import transforms
from model import Vgg16

# 扫描数据次数
epochs = 3
# 分组大小
batch = 64
# 学习率
learning_rate = 0.01
# 训练次数
train_step = 0
# 测试次数
test_step = 0


# 定义图像转换
transform = transforms.Compose([
    transforms.Resize(224),
    transforms.ToTensor()
])
# 读取数据
train_dataset = datasets.CIFAR10(root="./dataset",train=True,transform=transform,download=True)
test_dataset = datasets.CIFAR10(root="./dataset",train=False,transform=transform,download=True)
# 加载数据
train_dataloader = DataLoader(train_dataset,batch_size=batch,shuffle=True,num_workers=0)
test_dataloader = DataLoader(test_dataset,batch_size=batch,shuffle=True,num_workers=0)
# 数据大小
train_size = len(train_dataset)
test_size = len(test_dataset)
print("训练集大小：{}".format(train_size))
print("验证集大小：{}".format(test_size))

# GPU
device = torch.device("mps" if torch.backends.mps.is_available() else "cpu")
print(device)
# 创建网络
net = Vgg16()
net = net.to(device)
# 定义损失函数
loss = nn.CrossEntropyLoss()
loss = loss.to(device)
# 定义优化器
optimizer = torch.optim.SGD(net.parameters(),lr=learning_rate)

writer = SummaryWriter("logs")
# 训练
for epoch in range(epochs):
    print("-------------------第 {} 轮训练开始-------------------".format(epoch))
    net.train()
    for data in train_dataloader:
        train_step = train_step + 1
        images,targets = data
        images = images.to(device)
        targets = targets.to(device)
        outputs = net(images)
        loss_out = loss(outputs,targets)
        optimizer.zero_grad()
        loss_out.backward()
        optimizer.step()

        if train_step%100==0:
            writer.add_scalar("Train Loss",scalar_value=loss_out.item(),global_step=train_step)
            print("训练次数：{}，Loss：{}".format(train_step,loss_out.item()))

    # 测试
    net.eval()
    total_loss = 0
    total_accuracy = 0
    with torch.no_grad():
        for data in test_dataloader:
            test_step = test_step + 1
            images, targets = data
            images = images.to(device)
            targets = targets.to(device)
            outputs = net(images)
            loss_out = loss(outputs, targets)
            total_loss = total_loss + loss_out
            accuracy = (targets == torch.argmax(outputs,dim=1)).sum()
            total_accuracy = total_accuracy + accuracy
        # 计算精确率
        print(total_accuracy)
        accuracy_rate = total_accuracy / test_size

        print("第 {} 轮，验证集总损失为：{}".format(epoch+1,total_loss))
        print("第 {} 轮，精确率为：{}".format(epoch+1,accuracy_rate))
        writer.add_scalar("Test Total Loss",scalar_value=total_loss,global_step=epoch+1)
        writer.add_scalar("Accuracy Rate",scalar_value=accuracy_rate,global_step=epoch+1)
    torch.save(net,"./model/net_{}.pth".format(epoch+1))
    print("模型net_{}.pth已保存".format(epoch+1))