现代卷积网络实战系列3：PyTorch从零构建AlexNet训练MNIST数据集

1、AlexNet

AlexNet提出了一下5点改进：

1. 使用了Dropout，防止过拟合
2. 使用Relu作为激活函数，极大提高了特征提取效果
3. 使用MaxPooling池化进行特征降维，极大提高了特征提取效果
4. 首次使用GPU进行训练
5. 使用了LRN局部响应归一化（对局部神经元的活动创建竞争机制，使得其中响应比较大的值变得相对更大，并抑制其他反馈较小的神经元，增强了模型的泛化能力）

2、AlexNet网络结构

AlexNet(
 (feature): Sequential(
  (0): Conv2d(1, 32, kernel_size=(5, 5), stride=(1, 1), padding=(1, 1))
  (1): ReLU(inplace=True)
  (2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (3): ReLU(inplace=True)
  (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  (5): Conv2d(64, 96, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (6): ReLU(inplace=True)
  (7): Conv2d(96, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (8): ReLU(inplace=True)
  (9): Conv2d(64, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (10): ReLU(inplace=True)
  (11): MaxPool2d(kernel_size=2, stride=1, padding=0, dilation=1, ceil_mode=False)
  )
 (classifier): Sequential(
  (0): Dropout(p=0.5, inplace=False)
  (1): Linear(in_features=4608, out_features=2048, bias=True)
  (2): ReLU(inplace=True)
  (3): Dropout(p=0.5, inplace=False)
  (4): Linear(in_features=2048, out_features=1024, bias=True)
  (5): ReLU(inplace=True)
  (6): Linear(in_features=1024, out_features=10, bias=True)
  )
)

3、PyTorch构建AlexNet

class AlexNet(nn.Module):
    def __init__(self, num=10):
        super(AlexNet, self).__init__()
        self.feature = nn.Sequential(
            nn.Conv2d(1, 32, kernel_size=5, stride=1, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(32, 64, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Conv2d(64, 96, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(96, 64, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(64, 32, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=2, stride=1),
        )
        self.classifier = nn.Sequential(
            nn.Dropout(),
            nn.Linear(32 * 12 * 12, 2048),
            nn.ReLU(inplace=True),
            nn.Dropout(),
            nn.Linear(2048, 1024),
            nn.ReLU(inplace=True),
            nn.Linear(1024, num),
        )

    def forward(self, x):
        x = self.feature(x)
        x = x.view(-1, 32 * 12 * 12)
        x = self.classifier(x)
        return x

10个epoch训练过程的打印：

D:\conda\envs\pytorch\python.exe A:\0_MNIST\train.py

Reading data…
train_data: (60000, 28, 28) train_label (60000,)
test_data: (10000, 28, 28) test_label (10000,)

Initialize neural network
test loss: 2302.56
test accuracy: 10.1 %

epoch step: 1
training loss: 167.49
test loss: 46.66
test accuracy: 98.73 %

epoch step: 2
training loss: 59.43
test loss: 36.14
test accuracy: 98.95 %

epoch step: 3
training loss: 49.94
test loss: 24.93
test accuracy: 99.22 %

epoch step: 4
training loss: 38.7
test loss: 20.42
test accuracy: 99.45 %

epoch step: 5
training loss: 35.07
test loss: 26.18
test accuracy: 99.17 %

epoch step: 6
training loss: 30.65
test loss: 22.65
test accuracy: 99.34 %

epoch step: 7
training loss: 26.34
test loss: 20.5
test accuracy: 99.31 %

epoch step: 8
training loss: 26.24
test loss: 27.69
test accuracy: 99.11 %

epoch step: 9
training loss: 23.14
test loss: 22.55
test accuracy: 99.39 %

epoch step: 10
training loss: 20.22
test loss: 28.51
test accuracy: 99.24 %

Training finished

进程已结束，退出代码为 0

效果已经非常好了