5a2f76ede044b4904af9461e18253f2929cfc5a4,fairseq/criterions/nat_loss.py,LabelSmoothedDualImitationCriterion,_compute_loss,#LabelSmoothedDualImitationCriterion#Any#Any#Any#Any#Any#Any#,29

Before Change

        else:  // soft-labels
            losses = F.kl_div(logits, targets, reduction="none")
            losses = losses.float().sum(-1).type_as(losses)

        nll_loss = mean_ds(losses)
        if label_smoothing > 0:

After Change

        if masks is not None:
            outputs, targets = outputs[masks], targets[masks]

        if not masks.any():
            nll_loss = torch.tensor(0)
            loss = nll_loss
        else:
            logits = F.log_softmax(outputs, dim=-1)
            if targets.dim() == 1:
                losses = F.nll_loss(logits, targets.to(logits.device), reduction="none")

            else:  // soft-labels
                losses = F.kl_div(logits, targets.to(logits.device), reduction="none")
                losses = losses.sum(-1)

            nll_loss = mean_ds(losses)
            if label_smoothing > 0:
                loss = nll_loss * (
                    1 - label_smoothing) - mean_ds(logits) * label_smoothing
            else:
                loss = nll_loss

        loss = loss * factor
        return {"name": name, "loss": loss, "nll_loss": nll_loss, "factor": factor}

    def _custom_loss(self, loss, name="loss"):

Instances