SEAM关键代码解析

Self-supervised Equivariant Attention Mechanism for Weakly Supervised Semantic Segmentation 关键代码解析

 def PCM(self, cam, f):
        n,c,h,w = f.size()
        cam = F.interpolate(cam, (h,w), mode='bilinear', align_corners=True).view(n,-1,h*w)
        # 对 CAM 进行双线性插值，使其尺寸与特征图 f 一致，然后将其展平
        f = self.f9(f)
        f = f.view(n,-1,h*w) 
        f = f/(torch.norm(f,dim=1,keepdim=True)+1e-5) # 将特征图 f 展平，然后归一化
		
        aff = F.relu(torch.matmul(f.transpose(1,2), f),inplace=True)
        aff = aff/(torch.sum(aff,dim=1,keepdim=True)+1e-5)
        # 计算特征之间的关联矩阵（亲和性矩阵），并进行归一化
        cam_rv = torch.matmul(cam, aff).view(n,-1,h,w)
        # 将 CAM 与关联矩阵相乘，得到经过协同显著图增强的 CAM
        return cam_rv