[实现]ResNet Implementation in PyTorch • Liu Yang's Blog

思想

原始的方法需要学习一个直接的映射， $f(x)$ ，但是直接学习的难度比较大，学习对x的修改，更简单。就像做什么任务，照着模板改会更容易。

所以对于正常块只学习 $f(x)-x$ ，减少了网络的工作量

主要代码

class ShortcutProject(nn.Module):
    def __init__(self, in_channels=1, out_channels=1, kernel_size=3, stride=1) -> None:
        super().__init__()
        self.conv = nn.Conv2d(
            in_channels=in_channels,
            out_channels=out_channels,
            stride=stride,
            kernel_size=kernel_size,
            padding=1,
            bias=None,
        )
        self.bn = nn.BatchNorm2d(out_channels)

    def forward(self, x):
        return self.bn(self.conv(x))


class ResNetBlock(nn.Module):
    def __init__(self, in_channels=1, out_channels=1, stride=1, kernel_size=3) -> None:
        super().__init__()
        self.conv1 = nn.Conv2d(
            in_channels=in_channels,
            out_channels=out_channels,
            stride=stride,
            kernel_size=kernel_size,
            padding=1,
            bias=None,
        )
        self.bn1 = nn.BatchNorm2d(out_channels)
        self.relu1 = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv2d(
            in_channels=out_channels,
            out_channels=out_channels,
            stride=1,
            kernel_size=3,
            padding=1,
            bias=None,
        )
        self.bn2 = nn.BatchNorm2d(out_channels)
        if stride != 1 or in_channels != out_channels:
            # 对输入X
            self.shortcut = ShortcutProject(in_channels, out_channels, stride=stride)
        else:
            self.shortcut = nn.Identity()
        self.relu2 = nn.ReLU(inplace=True)

    def forward(self, x):
        shortcut = self.shortcut(x)
        x1 = self.relu1(self.bn1(self.conv1(x)))
        x1 = self.bn2(self.conv2(x1))
        return self.relu2(x1 + shortcut)

定义模型

class ResNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.res1 = ResNetBlock(in_channels=1, out_channels=128, kernel_size=3)
        self.res2 = ResNetBlock(in_channels=128, out_channels=128, kernel_size=3)
        self.res3 = ResNetBlock(in_channels=128, out_channels=128, kernel_size=3)
        # mlp替换为1x1卷积
        self.con1x1 = nn.Conv2d(in_channels=128, out_channels=10, kernel_size=1)
        self.linear = nn.Linear(28 * 28 * 10, 10)

    def forward(self, x):
        x = self.res1(x)
        x = self.res2(x)
        x = self.res3(x)
        x = self.con1x1(x)
        x = x.view(x.shape[0], -1)
        return self.linear(x)

解释

有padding不需要扩充，1x1卷积控制通道数，最后一层MLP用于分类
先Conv -> Bn -> ReLU，ShortcutProject不能有ReLU，只能有前两项，后面加了之后再过ReLU层
ReLU中的inplace参数表示原地操作，省时省空间，类似于tensor.zero_()这种，原地变0
x只考虑单个样本即可，无需考虑整个batch