[多尺度融合]--Res2Net：A New Multi-scale Backbone Architecture

引言：

其实很早之前就知道这篇文章了，最近因为要用到多尺度融合的结构，所以就想到了这篇文章！在目标检测、语义分割以及其他的很多领域，目标尺度分步不均对于网络模型的处理能力要求就比较高，因此多尺度处理以及多尺度融合的策略就用的比较多。本文是提出一个多尺度的backbone网络，可以有效提升分类检测的性能！

创新点：

上图右边是标准的ResNet结构，右图为本文提出的Res2Net模块，将3x3卷积替换一个包含分组卷积和多尺度融合的模块。同时也可以搭配channel attention一起使用，即SE模块。

代码：

##  Res2Net+SE
class R2CAB(nn.Module):
    def __init__(self, channel, stride=1, scale=4, basewidth=28):
        super(R2CAB, self).__init__()
        width = int(math.floor(16 * (basewidth / 64.0)))
        self.conv1 = nn.Conv2d(channel, width * scale, kernel_size=1, stride=stride, bias=True)
        if scale == 1:
            self.nums = 1
        else:
            self.nums = scale - 1
        convs = []
        for i in range(self.nums):
            convs.append(nn.Conv2d(width, width, kernel_size=3, stride=stride, padding=1, bias=True))
        self.convs = nn.ModuleList(convs)
        self.conv3 = nn.Conv2d(width * scale, channel, kernel_size=1, stride=stride, bias=True)

        self.relu = nn.ReLU(inplace=True)
        self.scale = scale
        self.width = width
        # SELayer
        self.ca = CALayer(channel)

    def forward(self, x):
        residual = x
        # [N, width * scale, H, W]
        out = self.relu(self.conv1(x))

        # scale * [N, width , H, W]
        spx = torch.split(out, self.width, 1)
        for i in range(self.nums):
            if i == 0:
                sp = spx[i]
            else:
                sp = sp + spx[i]
            sp = self.convs[i](sp)
            sp = self.relu(sp)
            if i == 0:
                out = sp
            else:
                out = torch.cat((out, sp), 1)

        out = torch.cat((out, spx[self.nums]), 1)
        out = self.conv3(out)
        out = self.ca(out)

        out += residual
        return out