Discriminator loss converges to zero early in training about magvit2-pytorch HOT 9 OPEN

Thanks @lucidrains. I'll try again with those parameters. I saw it in the taming implementation here:
https://github.com/CompVis/taming-transformers/blob/3ba01b241669f5ade541ce990f7650a3b8f65318/taming/modules/losses/vqperceptual.py#L51

from magvit2-pytorch.

@lucidrains Nice. Let me try it out again. No, I don't have any relatives in Massachusetts. Did you meet someone with the last name Pfeil?

from magvit2-pytorch.

That's amazing. It's not a common name. Sorry to hear about your friend.

from magvit2-pytorch.

@jpfeil can you screenshot the paper section where they propose delaying the discriminator training? (and link the paper too)

from magvit2-pytorch.

@jpfeil do you have adversarial_loss_weight greater than 0.? also try another run where your perceptual_loss_weight is 0.1

from magvit2-pytorch.

@jpfeil welp.. whatever Robin and Patrick does goes; they are the best in the world.

let me add that

from magvit2-pytorch.

@jpfeil ok, added that same functionality here. try removing the learning rate schedule in your next run too, shouldn't need it for something this easy

from magvit2-pytorch.

@jpfeil you don't happen to have relatives in Massachusetts, do you?

from magvit2-pytorch.

yea, I knew someone back in high school with the Pfeil family name. Tragedy struck and they moved away though. You are the second Pfeil I've met!

from magvit2-pytorch.