about self-training and Pseudo-label

[L0310]

When I was studying your code implementation, I noticed that the pseudo labels generated during training were not saved and used for the self-training stage. Instead, only the LAT loss + p2_cons + p2_reg was used in the latter part of the training process. So, your work is an end-to-end one.

        if process <= division:
            if mask is None:
                # pseudo = minMaxNorm(uphw(attn_pos, size=size)).gt(0.5).float()
                pseudo = self.crf(uphw(minMaxNorm(x), size=size), minMaxNorm(uphw(attn_pos.detach(), size=size)), iters=self.cfg.crf_round).gt(0.5).float()
            else:
                loss = {}  ## clear
                pseudo = uphw(mask, size=size)
            bce_pr_wrt_ps = F.binary_cross_entropy_with_logits(pred, pseudo, reduction="none").reshape(n, -1).mean(dim=-1)
            loss["p1_bce"] = (bce_pr_wrt_ps * torch.softmax(self.cfg.rew_kai / (bce_pr_wrt_ps + 1.0), dim=-1)).sum()
            loss["p1_cons"] = F.l1_loss(torch.sigmoid(pred[0:n]), torch.sigmoid(pred[n::]))
        if process > division:
            loss["p2_lwt"] = self.lwt(torch.sigmoid(pred), minMaxNorm(x), margin=0.5)
            loss["p2_cons"] = F.l1_loss(torch.sigmoid(pred[0:n]), torch.sigmoid(pred[n::]))
            loss["p2_reg"] = 0.5 - torch.abs(torch.sigmoid(pred) - 0.5).mean()

[guanhuankang]

As mentioned in "Training Strategy" section in our paper, we adopt selfmask model as our detector for self-training stage. In the above code snippet, "mask" is alway None, i.e., "pseudo = self.crf(uphw(minMaxNorm(x), size=size), minMaxNorm(uphw(attn_pos.detach(), size=size)), iters=self.cfg.crf_round).gt(0.5).float()" This is the output of our CSE Module, whose results is suggested as $Our_{attn}$ in table 1. It is used to supervise the outputs of our FRC Decoder, whose results are the initial results as suggested in $Our_{init}$ in table 1.

[L0310]

Is the self-training stage the code below? Does this command "python train.py --name training_csnet" include the self-training stage? Or does the self-training stage need to start training all over again? Thank you very much！