search

NVlabs / stylegan3

Official PyTorch implementation of StyleGAN3
Other
6.3k stars 1.11k forks source link

How to downscale pretrained model? #182

Open NukeBird opened 2 years ago

NukeBird commented 2 years ago

I had quite a bit fun with pretrained model that generates danbooru-like 512x512 images, even managed to use transer learning to teach it how to generate faces from touhou

That was fun, but...

  1. I have only RTX 3070 with 8GB of VRAM, I'm often getting out of memory aka "oops, I can't allocate 2mb! Sorry!" in the middle of training
  2. Personally, I would love to work with 256x256 or even 128x128 cause I wanna train my net faster

So... Is that possible to downsize my stylegan2 model?

kevinsadi commented 1 year ago

I also have this same issue. I have noticed that even if I create my training data to be 256x256 and try to train with a pretrained 256x256 pkl file, I get a tensor size mismatch error. Did you manage to figure this out?