I was wondering last week where I could find some info about this, looking forward reading it!
I guess I should be making a bit more noise about this, but DeOldify video works as well as it does because of a new GAN training technique developed in collaboration with @fastdotai called NoGAN. Here's a Reddit post on this!https://www.reddit.com/r/MachineLearning/comments/bq8gji/p_new_in_deoldify_smooth_colorization_of_video/ …
-
-
-
Awesome!
End of conversation
New conversation -
-
-
Wanted to learn more about this too. Thanks for sharing the link!
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I used the same trick on neural-enhance. I remember reading about this technique in super-resolution papers, but their goal was mainly to stabilize training rather than speeding things up. https://github.com/alexjc/neural-enhance#1b-training-super-resolution …
-
I believe this is referred to as "warming up" the discriminator before enabling the full GAN loss. L498:https://github.com/alexjc/neural-enhance/blob/master/enhance.py#L498 …
End of conversation
New conversation -
-
-
It should be called Noggin
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Was the NoGAN also the technique applied to the microscopy videos at Salk Institute? The results are impressive. I’ve already tried the U-Net w perceptual loss to enhance satellite imagery from the Lunar Reconnaissance Orbiter, but I’m thinking about maybe testing this NoGAN.
-
Those microscopy results didn’t have NoGAN applied but I did try it on super-resolution and it worked marvelously. I suggested to
@manorlaboratory that he should try it out as well. It should be a generally useful approach. - Show replies
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.