The edge_loss and seg_loss are increasing accumulatively when I train the OST dataset, and the training sample images are wrong.

[zhangyonle]

The edge_loss and seg_loss are increasing accumulatively when I train the OST dataset, whether it is the code 'weight = (label_anno.view(bs, -1) == (self.n_cate_anno-1)).sum(1) == h * w # for cityscapes datasets segmentation' to cause the error in training OST dataset.

and the training sample images are wrong, as the predicted segmentation map, predicted edge map and the inpainted image are error. And this wrong sample image usually occurs after 20000 training iters.