train.py make_D_label about advsemiseg HOT 6 CLOSED

It just converts from a single value (0/1) to a matrix which contains spatial information. The value in the matrix is multiplied by the single value.

from advsemiseg.

thanks @John1231983 for explaining. close for now if there is no further issue.

from advsemiseg.

since" ignore_mask_remain = np.zeros(D_out_sigmoid.shape).astype(np.bool)" ,so is the result of make_D_label is determined by gt_label?gt_label is 1 so D_label is a tensor all 1?gt_label is 0 so D_label is a tensor all 0?If it's true,so what's the meaning of ignore_mask_remain?

from advsemiseg.

@hfslyc

from advsemiseg.

yeah it's actually all 0. We experimented with different masking loss for the adversarial loss on the unlabeled data, and in the end, we found that the unmasked version is the best. Sorry for the confusion. Will modify the code in the near future for better readability.

from advsemiseg.

Thanks for your reply.Your work is very interesting~

from advsemiseg.