Shengyu Zhao
Shengyu Zhao
I've experienced this issue occasionally. I'm not sure what is the root cause, but it usually becomes normal if you resume training from the checkpoint just before the sudden spike.
I think that this is a normal phase in training when the generator's initial progress cannot be reflected by the FID score.
How are the black patches? It can also be that the discriminator does not learn well at the beginning of training, so you can try training only the discriminator for...
Hard to say the exact reason... But I feel that the short spike won't affect the performance as I think it understandable that training can be pretty random at the...
How does your dataset look like? It seems to me a more severe discriminator overfitting issue.
Looks like a challenging dataset... I think the model will not learn well if there are only hundreds of such images. DiffAugment can reduce discriminator overfitting by some degree, while...
Maybe you need to run the training script with `--schedule=evaluate`, referring to [compare_gan's README](https://github.com/google/compare_gan#training-and-evaluation).
This is borrowed from recent GAN implementations like StyleGAN2, but I think maybe they do not make much difference.
We always use the first 1k images in the FFHQ tfrecord file.
To our purpose, i.e. to reduce discriminator over fitting, augmenting only the discriminator would be enough. I think it's also totally fine to augment the encoder, but serving as a...