INNER CODE UNIT · Python

edge_attention

nv-tlabs/GSCNN · loss.py:94

    def edge_attention(self, input, target, edge):
        n, c, h, w = input.size()
        filler = torch.ones_like(target) * 255
        return self.seg_loss(input, 
                             torch.where(edge.max(1)[0] > 0.8, target, filler))

    def forward(self, inputs, targets):
        segin, edgein = inputs
        segmask, edgemask = targets

        losses = {}

        losses['seg_loss'] = self.seg_weight * self.seg_loss(segin, segmask)
        losses['edge_loss'] = self.edge_weight * 20 * self.bce2d(edgein, edgemask)
        losses['att_loss'] = self.att_weight * self.edge_attention(segin, segmask, edgein)
        losses['dual_loss'] = self.dual_weight * self.dual_task(segin, segmask)
              
        return losses

View source record →