search

lukemelas / unsupervised-image-segmentation

[ICLR 2022] Finding an Unsupervised Image Segmenter in each of your Deep Generative Models
77 stars 11 forks source link

Help producing results with the StyleGAN models #4

Closed james-oldfield closed 1 year ago

james-oldfield commented 1 year ago

Hi there!

I'm having trouble producing meaningful results on StyleGAN2 on AFHQ. I've been using the default setup and hyperparameters. After 50 iterations (with the default batch size of 32) I get visualisations that look initially promising: (https://i.imgur.com/eR79Wyd.png). But as training progresses, and indeed when it reaches 300 iterations, these are the visualisation results: https://i.imgur.com/36zhBzT.png.

I've tried playing with the learning rate, and the number of iterations with no success yet. Did you have tips here or ideas as to what might be going wrong here?

Thanks! James.

james-oldfield commented 1 year ago

I've just seen from the ICLR version of the paper's sup. material that a 0.2 radius is useful for StyleGAN, which indeed helps a little bit (100 iterations is not too bad: https://i.imgur.com/kzPgzuJ.png). If you have any other tips here to produce more faithful results please do let me know!