大核注意力Large Kernel Attention(LKA)

这种分解方式也被叫做大核注意力(Large Kernel Attention)，即LKA。如上图所示，一个很大kernel size的卷积被分解成一个Depth-wise卷积+一个Depth-wise空洞卷积+一个1× \times× 1卷积。这样，就可以大大减少FLOPs和参数量。很有效地解决了小核卷积的local性。

class AttentionModule(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.conv0 = nn.Conv2d(dim, dim, 5, padding=2, groups=dim)  #depth-wise conv
        self.conv_spatial = nn.Conv2d(dim, dim, 7, stride=1, padding=9, groups=dim, dilation=3)   #conv_spatial
        self.conv1 = nn.Conv2d(dim, dim, 1)  # 1x1 conv


    def forward(self, x):
        u = x.clone()        
        attn = self.conv0(x)
        attn = self.conv_spatial(attn)
        attn = self.conv1(attn)

        return u * attn