【学习记录：Loss不变】复现LeNet时，Loss在训练过程中不发生改变

class LeNet(nn.Module):

    def __init__(self, input_channels, output_channels, kernel_size, strides):
        super(LeNet, self).__init__()

        self.conv = nn.Sequential(
            # 3x32x32 -> 6x28x28 k = 5, s = 1
            nn.Conv2d(in_channels=input_channels[0], out_channels=output_channels[0],
                      kernel_size=kernel_size[0], stride=strides[0]),
            nn.ReLU(),
            # 6x28x28 -> 6x14x14 k = 2, s = 2
            nn.MaxPool2d(kernel_size=kernel_size[1], stride=strides[1]),
            # 6x14x14 -> 16x10x10 k = 5, s = 1
            nn.Conv2d(in_channels=input_channels[1], out_channels=output_channels[1],
                      kernel_size=kernel_size[2], stride=strides[2]),
            nn.ReLU(),
            # 16x10x10 -> 16x5x5 k = 2, s = 2
            nn.MaxPool2d(kernel_size=kernel_size[3], stride=strides[3]),
            nn.ReLU()
        )

        self.fc = nn.Sequential(
            # 16x5x5 -> 120x1x1
            nn.Linear(in_features=input_channels[2], out_features=output_channels[2], bias=True),
            nn.ReLU(),
            # 120x1x1 -> 84x1x1
            nn.Linear(in_features=input_channels[3], out_features=output_channels[3], bias=True),
            nn.Softmax(dim=1),
            # 84x1x1 -> 10x1x1
            nn.Linear(in_features=input_channels[4], out_features=output_channels[4], bias=True),
            nn.Softmax(dim=1)
        )

    def forward(self, x):
        x = self.conv(x)
        x = x.reshape(x.size(0), -1)
        x = self.fc(x)
        return x

原因是全连接层的每一层都使用了SoftMax激活函数，将中间的全连接层的激活函数用ReLU函数替代（如上图），只有最后一层的全连接层使用SoftMax激活函数。