Open mohammadmirzaee25 opened 1 month ago
Hi,
Yes, they're randomly sampled with the same seed, so the small differences in the 0th cycles are because of the network optimization (gradient descent). Except for ours where we use the self-supervised loss, and that's why the difference is bigger.
@TheRevanchist Thank you for answering . So your method use Pseudo-labeling in first cycle too? and i dont understand what you mean by optimization . isnt network the same for all other methods(same seed for training)?
@TheRevanchist Hi. could you please clarify why in paper reults(for example in figure 3 left image) first cycle of all method are different? because i think you said in paper that all labels for first cycle is same and i read your code and could not find why there is difference in first cycle for different methods. arent they sampled random for first cycle with same random seed?