4 ms·

I would love to replicate this. It looks like the dataset is open source. https://www.gwern.net/Danbooru2018 https://www.gwern.net/Danbooru2018 I don't have a

I would love to replicate this. It looks like the dataset is open source. https://www.gwern.net/Danbooru2018 https://www.gwern.net/Danbooru2018 I don't have a sense for hardware requirements though. Does anyone have a good idea of how much time and money it would take to train such a model?

gwern 7y ago
I'm glad you asked: https://www.gwern.net/Faces#compute https://www.gwern.net/Faces#compute Figure a few GPU-weeks and a few hundred dollars if you want to go from scratch.
- b_tterc_p 7y ago
  Thanks! I was just in the process of digging through your stuff.
- weeb_throwaway 7y ago
  I saw the section about not training from scratch (via transfer learning) in https://www.gwern.net/Faces#transfer-learning https://www.gwern.net/Faces#transfer-learning. The Holo example is really impressive! How expensive is it in terms of labelled data and compute? Do you know if anyone tried this for just ahegao faces?
  - gwern 7y ago
    > How expensive is it in terms of labelled data and compute? All the stuff you see in that page (except the BigGAN ones) is unconditional, no labels. You just dump the images in and it figures it out. StyleGAN does support labels via one-hot embedding as I understand it, but I don't know how to use it so none of my experiments use it. A few people have mentioned or used it, but there's no good documentation about how to make it work, so... For unconditional samples, it depends on how many you have and how different they are. You can see in the various examples transfer learning with a few hundred to a few thousand (with and without data augmentation). > Do you know if anyone tried this for just ahegao faces? It's funny you ask that because I was corresponding with an anonymous who was using it for just that (and ball gags). He'd run into some issues with the encoder/editing functionality and wanted advice, but the regular transfer learning worked fine. He'd compiled a small dataset of a few hundred to a few thousand examples on his own, and it worked disturbingly well: sufficiently so I didn't want to write it up. (I try to keep my site SFW.)

gwern 7y ago

I'm glad you asked: https://www.gwern.net/Faces#compute https://www.gwern.net/Faces#compute Figure a few GPU-weeks and a few hundred dollars if you want to go from scratch.

b_tterc_p 7y ago
Thanks! I was just in the process of digging through your stuff.
weeb_throwaway 7y ago
I saw the section about not training from scratch (via transfer learning) in https://www.gwern.net/Faces#transfer-learning https://www.gwern.net/Faces#transfer-learning. The Holo example is really impressive! How expensive is it in terms of labelled data and compute? Do you know if anyone tried this for just ahegao faces?
- gwern 7y ago
  > How expensive is it in terms of labelled data and compute? All the stuff you see in that page (except the BigGAN ones) is unconditional, no labels. You just dump the images in and it figures it out. StyleGAN does support labels via one-hot embedding as I understand it, but I don't know how to use it so none of my experiments use it. A few people have mentioned or used it, but there's no good documentation about how to make it work, so... For unconditional samples, it depends on how many you have and how different they are. You can see in the various examples transfer learning with a few hundred to a few thousand (with and without data augmentation). > Do you know if anyone tried this for just ahegao faces? It's funny you ask that because I was corresponding with an anonymous who was using it for just that (and ball gags). He'd run into some issues with the encoder/editing functionality and wanted advice, but the regular transfer learning worked fine. He'd compiled a small dataset of a few hundred to a few thousand examples on his own, and it worked disturbingly well: sufficiently so I didn't want to write it up. (I try to keep my site SFW.)