It's interesting to observe how the spread of ideas actually works in practice. I've seen DeOldify cited many places, yet the key insights that to me seem like low hanging fruit don't actually get traction. I suppose not having a paper might be part of it but it's still odd.
-
I think he's talking about the n optimizations he did since he released the open source version. Yep, the problem is that those optimizations are just "hints" spread across many tweets, releasing a full paper with all the list and all the experiments would be a great thing.
2/ There has a been a lot of optimizations in the work since the open source version but also some brand new ideas. Like we're not using a GAN to train anymore, we've overturned some training rules of thumb, and the self attention we've developed is quite different now.
-
-
1/ Best practises from http://fast.ai and self attention make sense, just like the previously used NoGAN. Still struggling to understand how you got it working without actual GANs and yet the results are so sharp and clear.
-
2/ What comes to my mind is feature and style losses. Other than that, I can't think of generally applicable techniques, except for plenty of small improvements. I am constantly blown away by the results of DeOldify. You really achieved something amazing!
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.