updated pytorch, fixed tensorboard, removed clip by open ai

PDillis / stylegan3-fun

Modifications of the official PyTorch implementation of StyleGAN3. Let's easily generate images and videos with StyleGAN2/2-ADA/3!

Other

230 stars 36 forks source link

updated pytorch, fixed tensorboard, removed clip by open ai #8

Closed ZibbeZabbe closed 2 years ago

ZibbeZabbe commented 2 years ago

Had issues training. Would get different issues. Specifying pytorch 1.10.2 cuda 11.3 ensures cuda version of pytorch is downloaded. Removed clip by open AI as that forced torch 1.7.1 due to compability Added setuptools specification to fix an issue that occasionaly occured with tensorboard when attempting to train.

I do believe this should work without any tweaking. Possibly useful to specify versions of pyspng, psutil and tensorboard to avoid future compability issues

PDillis commented 2 years ago

I have merged the PR from the main branch, this lets you use PyTorch 1.11. So in the environment.yml, setting - pytorch>=1.9.1 should work. Right now, I've managed to generate images with this modified environment, as well as training a small network (currently in it). Let me know if this works for you! I'll go ahead and test it in Ubuntu and if everything works, I'll accept your PR as you have other changes I want to add.