[代码复现]--Dynamic Multi-scale Filters

动态适应输入的多尺度融合结构，代码如下：

class DCMLayer(nn.Module):
    def __init__(self, k, channel):
        super(DCMLayer, self).__init__()
        self.k = k
        self.channel = channel
        self.conv = nn.Conv2d(channel, channel // 4, 1, padding=0, bias=True)
        self.fuse = nn.Conv2d(channel // 4, channel, 1, padding=0, bias=True)
        self.dw_conv = nn.Conv2d(channel // 4, channel // 4, self.k, padding=(self.k-1) // 2, groups=channel // 4)
        self.pool = nn.AdaptiveAvgPool2d(k)

    def forward(self, x):
        N, C, H, W = x.shape
        # [N * C/4 * H * W]
        f = self.conv(x)
        # [N * C/4 * K * K]
        g = self.conv(self.pool(x))

        f_list = torch.split(f, 1, 0)
        g_list = torch.split(g, 1, 0)

        out = []
        for i in range(N):
            #[1* C/4 * H * W]
            f_one = f_list[i]
            # [C/4 * 1 * K * K]
            g_one = g_list[i].squeeze(0).unsqueeze(1)
            self.dw_conv.weight = nn.Parameter(g_one)

            # [1* C/4 * H * W]
            o = self.dw_conv(f_one)
            out.append(o)

        # [N * C/4 * H * W]
        y = torch.cat(out, dim=0)
        y = self.fuse(y)

        return y

class DCM(nn.Module):
    def __init__(self, channel):
        super(DCM, self).__init__()
        self.DCM1 = DCMLayer(1, channel)
        self.DCM3 = DCMLayer(3, channel)
        self.DCM5 = DCMLayer(5, channel)
        self.conv = nn.Conv2d(channel * 4, channel, 1, padding=0, bias=True)

    def forward(self, x):
        dcm1 = self.DCM1(x)
        dcm3 = self.DCM3(x)
        dcm5 = self.DCM5(x)

        out = torch.cat([x, dcm1, dcm3, dcm5], dim=1)
        out = self.conv(out)
        return out