CUT training stopping criterion

[danivelikova]

Hi! I have a question regarding the training of the CUT network.
How was the final CUT model chosen? Did you save each epoch during training and then plot the FID score on a set of test images for each epoch and picked the epoch with the lowest FID? How did you avoid hallucinations?
I have a set of 500 images for each class and I’m training with a lr=5e-06 for about 70 epochs. However, I experience that after the 20-30th epoch, the model starts hallucinating.