You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
1/ "ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks" is so packed with great insights. Of particular note: They pretrained their generator on L1 loss, and report that it actually improves quality. https://arxiv.org/abs/1809.00219
2/ I've been experimenting with pretraining both generator and critic with non-gan losses for DeOldify, it turns out, because I suspected it would lead to not only faster training but better results as well. It's still early but I'll just say it looks promising!
3/ The core concept definitely works functionally, I can tell you that much. Both for colorzation in DeOldify, as well as de-artifacting/super-res. This was part of lesson 7 in http://fast.ai V3 part 1, which will be released soon.
4/ The way I think of pretraining a GAN is this: A lot of time is wasted in going back and forth with the generator and critic not knowing what they're doing. Why not teach them directly with faster/simpler loss first to get most of the way there, -then- let them go GAN?
5/ Not only is it (much) faster to do it this way, but I've suspected it could lead to a better outcome because of less "blind leading the blind" going on and more productive training. You're benefiting from the strengths of both more straightforward training and GANs.
6/ I think I may be able to demonstrate that with DeOldify soon, but that's what they're reporting for ESRGAN. Exciting!
7/ I mentioned this is in the fastai course coming up but I want to make sure to credit is given where it's due. @jeremyphoward with the idea of actually using a simpler loss (mse, etc) for pretraining the generator, and doing binary classification for pretraining the critic.
One particular part of ESRGAN work that confuses me is that they use upconv instead of pixelshuffle unlike preceding works like RCAN, EDSR or RDN. The only explanation find is that the paper is mainly about enhancing SRGAN and it uses upconv. Any thoughts?
That is weird! Didn’t catch that... I’ll have to take a look again.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
