YOLOX-Reslayer

# ResLayer的中文意思是残差层，它可以解决深度网络的梯度消失问题
class ResLayer(nn.Module):
    "Residual layer with `in_channels` inputs."                         # 有in_channels个输入的残差层

    def __init__(self, in_channels: int):                               # 输入通道数
        super().__init__()                                              # 初始化父类
        mid_channels = in_channels // 2                                 # 中间通道数
        self.layer1 = BaseConv(
            in_channels, mid_channels, ksize=1, stride=1, act="lrelu"   # 1x1卷积，将通道数降低一半
        )
        self.layer2 = BaseConv(
            mid_channels, in_channels, ksize=3, stride=1, act="lrelu"   # 3x3卷积，将通道数恢复到原来的大小
        )

    def forward(self, x):
        out = self.layer2(self.layer1(x))                               # 1x1卷积，3x3卷积，通道数恢复到原来的大小
        return x + out                                                  # shortcut  y=y+x
# ResLayer的数据流图如下：
# x -> layer1(1x1) -> layer2(3x3) -> y
# |______________________________|
# 其中layer1是1x1卷积，layer2是3x3卷积