4 ms·
PPL was one of the problems in scaling up StyleGAN2. It contributes to the over-regularization that made diverse images impossible to learn. There might be some
by gwern 6y ago
PPL was one of the problems in scaling up StyleGAN2. It contributes to the over-regularization that made diverse images impossible to learn. There might be some very weak value of PPL which is beneficial (probably the optimal value of the PPL regularization is not exactly 0) but when the runs take a month or two, it's not easy to experiment with them!
The dataset is actually ~3x bigger than ImageNet, FWIW.
- godelski 6y ago> The dataset is actually ~3x bigger than ImageNet, FWIW. Oh that is very interesting. Thanks. Do you know anything else about the dataset? I was looking at the link for some quick stats and didn't see them. I'm curious if there are metrics to support or disprove my hypothesis about body features. If so ADA would still help because there isn't an even distribution of the features.
- gwern 6y agoI included the basic stats in https://www.gwern.net/Faces#extended-stylegan2-danbooru2019-aydao https://www.gwern.net/Faces#extended-stylegan2-danbooru2019-... and links to the writeups for the individual datasets.