Participants discovered interesting techniques to squeeze into time limits and generate plausible pictures with small resources. Check the top solutions, they generate much better results using tuned Big/Style GAN architectures.
Probably this competition is not the most useful in the world, but definitely could help to emerge some advances in GANs training and architectures, I believe. So probably such interesting applications of these networks like GauGAN eventually become less computationally expensive and more accessible for the community of researchers.
Hurrah! Well done! I really enjoyed competing in this challenge too.
The advancements made by competitors in this challenge are really interesting, especially the context of small images, trained on a single GPU for less than 9 hours. A sort of ImageWoof for GANs. The winners all used BigGAN. I hadn’t gone near it since I read it took many days to train. More fool me. I may try to port some of the winning solutions to fastai before I get pulled into another challenge.
Hey, thank you! (BTW, nice work on AI blogging, and a great DL rig).
Oh completely agree! I also easily excluded BigGAN from consideration when reading how many GPU they used to train their model Sure enough, it was a bit too fast conclusion. Also, I pretty quickly discarded conditional GANs as soon as my first attempts showed bad results. Should try more, I believe.
So this idea of adapting complex architectures to simpler tasks is one of the greatest insights taken from this competition, I would say.