Closed MyeongJin-Kim closed 5 years ago
The weight for the discriminator is randomly initialized.
Thanks for the comment!
One more question!
If discriminator is randomly initialized, when train again with previous segmentation weight, mIoU drop at early stage. Did you get the same pattern?
I don't quite understand your question. what do you mean by "train again"? If you mean training a new model with pseudo labels, I don't have overfitting problem.
In inner loop for SSL, if N=2, segmentation network is trained 3 times in a row. When training segmentation network, how did you set the discriminator's weight? Initialize discriminator every time or use previous discriminator's weight??
Also same question for outer loop's weight.