-
Notifications
You must be signed in to change notification settings - Fork 103
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add save and sampling #23
Conversation
This imports the same save and sampling method as used in your denoising diffusion repo.
gigagan_pytorch/gigagan_pytorch.py
Outdated
|
||
G_kwargs = dict(batch_size=batches) | ||
all_images_list = list(map(lambda n: self.G(batch_size = n), batches)) | ||
print(all_images_list) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
stray print
@CerebralSeed nice! do you want to try doing the sampling for the upsampler too? perhaps the sampled images can have 2 columns, with original -> upsampled. original can be taken from the training dataloader for now, and extra credit would be to support a different validation dataloader specifically for eval |
have you tested it? i think you should be able to see something with a toy dataset |
Did, but I see I forgot to delete a debugging print statement. |
I'll give it a try later this week. |
@lucidrains Was running some tests, and, like most GANs, there are stability issues. The Generator loss keeps increasing. I noticed in their paper, they were using AdamW(see Table A2 on page 19 where they specify their settings). Or maybe Wasserstein loss might be worth trying? Also, was checking on the raw model sizes, as it seems the Discriminator is doing a little too well, relatively speaking. Sure enough, the Discriminator is almost twice the size of the Generator, while in their paper, it's about half the size or less.
Not sure where that's coming from, yet. It looks like you might have gotten the stability issues sorted here #17 So I will update and try again to see if that improves. |
@CerebralSeed nice to hear you are running tests! yup, it should be a lot more stable in the latest version |
I moved the logic for building kwargs for the generator into its own method, so it can be called in the save_and_sample_every. This is tested as working, but the original image colors seem a bit off. Maybe you can figure out what's causing the color issue.
@lucidrains Disregard that latest commit. I copied the whole file and seem to have overwritten some updates you were making. I will push a new commit. |
Closing this Pull request and opening a new one. |
This imports the same save and sampling method as used in your denoising diffusion repo.