Error in dual task loss?

[wangq95]

Hi, @tovacinni Thanks for your work, I have a question about the implementation of dual task loss in loss.py. I think the dual task loss should take the estimation from classical stream N, K, H, W and the edge prediction from shape stream N, 1, H, W as inputs, but in loss.py, the dual task loss actually takes segmentation estimation and segmentation lable as inputs: losses['dual_loss'] = self.dual_weight * self.dual_task(segin, segmask) Is this a bug or what?

[WaterKnight1998]

Hi, @tovacinni Thanks for your work, I have a question about the implementation of dual task loss in loss.py. I think the dual task loss should take the estimation from classical stream N, K, H, W and the edge prediction from shape stream N, 1, H, W as inputs, but in loss.py, the dual task loss actually takes segmentation estimation and segmentation lable as inputs: losses['dual_loss'] = self.dual_weight * self.dual_task(segin, segmask) Is this a bug or what?

The code is not working for me I tried to solved the code inside dualtaskloss as follows:

N, C, H, W = input_logits.shape
        th = 1e-8  # 1e-10
        eps = 1e-10
        ignore_mask = (gts == ignore_pixel).detach()
        input_logits = torch.where(ignore_mask.view(N, 1, H, W).expand(N, C, H, W),
                                   torch.zeros(N,C,H,W).cuda(),
                                   input_logits)
        gt_semantic_masks = gts.detach()
        gt_semantic_masks = torch.where(ignore_mask, torch.zeros(N,H,W).long().cuda(), gt_semantic_masks)
        gt_semantic_masks = _one_hot_embedding(gt_semantic_masks, 19).detach()

        g = _gumbel_softmax_sample(input_logits.view(N, C, -1), tau=0.5)
        g = g.reshape((N, C, H, W))
        g = compute_grad_mag(g, cuda=self._cuda)
 
        g_hat = compute_grad_mag(gt_semantic_masks, cuda=self._cuda)

        g = g.view(N, -1)
        g_hat = g_hat.reshape(N, -1)
        loss_ewise = F.l1_loss(g, g_hat, reduction='none', reduce=False)

        p_plus_g_mask = (g >= th).detach().float()
        loss_p_plus_g = torch.sum(loss_ewise * p_plus_g_mask) / (torch.sum(p_plus_g_mask) + eps)

        p_plus_g_hat_mask = (g_hat >= th).detach().float()
        loss_p_plus_g_hat = torch.sum(loss_ewise * p_plus_g_hat_mask) / (torch.sum(p_plus_g_hat_mask) + eps)

        total_loss = 0.5 * loss_p_plus_g + 0.5 * loss_p_plus_g_hat

However, loss_ewise = F.l1_loss(g, g_hat, reduction='none', reduce=False) this line is not working either. What are the suggestion you make to solve it?