4 ms·

> "The computation for the 512x512px training set was run for 20 days on an IBM AC922 POWER9 server running RHEL7, utilising all four Nvidia Tesla V100 SXM2 16G

by hwers 5y ago

> "The computation for the 512x512px training set was run for 20 days on an IBM AC922 POWER9 server running RHEL7, utilising all four Nvidia Tesla V100 SXM2 16GB GPUs." That's super overkill. If you just transfer learn from a good checkpoint you'll get good results in a day or two on a V100 (probably even a P100), especially at just 512x512. (I've done this lots of times and a colab pro account for $10 is more than plenty.)

edent 5y ago
Oh, that's interesting. Do you have a tutorial on how to do that?
- hwers 5y ago
  I learned everything I know about stylegan (prior to experimenting myself) from this channel: https://www.youtube.com/c/ArtificialImages/videos https://www.youtube.com/c/ArtificialImages/videos For stylegan this playlist was useful: https://www.youtube.com/playlist?list=PLWuCzxqIpJs8ViuBIUtAk-dsAtdrApYoy https://www.youtube.com/playlist?list=PLWuCzxqIpJs8ViuBIUtAk... (in particular this video https://www.youtube.com/watch?v=KxXqCsl-z2k&list=PLWuCzxqIpJs8ViuBIUtAk-dsAtdrApYoy https://www.youtube.com/watch?v=KxXqCsl-z2k&list=PLWuCzxqIpJ...). Or just go straight to this colab: https://colab.research.google.com/github/dvschultz/stylegan2-ada-pytorch/blob/main/SG2_ADA_PyTorch.ipynb https://colab.research.google.com/github/dvschultz/stylegan2... I'd recommend sticking with stylegan2 for the moment though since it's much easier (cheaper time/computation wise) to transfer from (and the result is about the same - in fact better at the same time and computation budget).
  - edent 5y ago
    That's amazing - thanks. Gives me something to play with this weekend :-)
siliconpotato 5y ago
what checkpoint would you use though? AFAIK there are no existing datasets trained on benches.
- hwers 5y ago
  You could just use ffhq/faces, it's fine. You'd think it gets stuck in a local minima but it really doesn't. Transfer learning from any 'finished' checkpoint is much much much faster than doing it from scratch.
- eliaspro 5y ago
  Mapillary had object verification tasks for benches in the past, but I'm not sure they're part of their public dataset.

by hwers 5y ago

edent 5y ago

Oh, that's interesting. Do you have a tutorial on how to do that?

hwers 5y ago
I learned everything I know about stylegan (prior to experimenting myself) from this channel: https://www.youtube.com/c/ArtificialImages/videos https://www.youtube.com/c/ArtificialImages/videos For stylegan this playlist was useful: https://www.youtube.com/playlist?list=PLWuCzxqIpJs8ViuBIUtAk-dsAtdrApYoy https://www.youtube.com/playlist?list=PLWuCzxqIpJs8ViuBIUtAk... (in particular this video https://www.youtube.com/watch?v=KxXqCsl-z2k&list=PLWuCzxqIpJs8ViuBIUtAk-dsAtdrApYoy https://www.youtube.com/watch?v=KxXqCsl-z2k&list=PLWuCzxqIpJ...). Or just go straight to this colab: https://colab.research.google.com/github/dvschultz/stylegan2-ada-pytorch/blob/main/SG2_ADA_PyTorch.ipynb https://colab.research.google.com/github/dvschultz/stylegan2... I'd recommend sticking with stylegan2 for the moment though since it's much easier (cheaper time/computation wise) to transfer from (and the result is about the same - in fact better at the same time and computation budget).
- edent 5y ago
  That's amazing - thanks. Gives me something to play with this weekend :-)

siliconpotato 5y ago

what checkpoint would you use though? AFAIK there are no existing datasets trained on benches.

hwers 5y ago
You could just use ffhq/faces, it's fine. You'd think it gets stuck in a local minima but it really doesn't. Transfer learning from any 'finished' checkpoint is much much much faster than doing it from scratch.
eliaspro 5y ago
Mapillary had object verification tasks for benches in the past, but I'm not sure they're part of their public dataset.