Hello,
I found an error in your compute_auc function that leads to auc = 0.5 almost the times. The Discriminator should not end with softmax layer because CrossEntropyLoss from pytorch receives input as logits (pure outpur from final linear layer). If you use softmax layer, the output of Discriminator will be scaled, and as a result, it is difficult to learn.
Bests.
Posted by: baondq @ May 22, 2024, 2:49 a.m.Also, we should increase number of epochs to reach convergence (I set it to 100 in my offline experiments)
Posted by: baondq @ May 22, 2024, 2:54 a.m.Hi, thank you for spotting this issue. We have done a quick fix and we hope the current version solves this problem.
Posted by: Jiajie @ May 27, 2024, 6:26 a.m.