1/ In developing DeOldify, I've had a very helpful, yet seemingly obvious, rule of thumb to help guide figuring out improvements in training: What can be done to close the gap between training and inference? Getting critical about this can yield some seriously low hanging fruit.
-
Show this thread
-
3/ This is also what lead to reshaping all images at inference as squares. Remember that training in batches is done with squares. Close the gap with inference by simulating the stretching done to "squarify" the images!
1 reply 0 retweets 12 likesShow this thread -
Replying to @citnaj
Are you also deteriorating images on the edges to simulate worn out photos and perform "light restoration"?
2 replies 0 retweets 1 like
Replying to @aviopene
Colorization has always for DeOldify been targeted strictly to bringing back color and no other component, so I make sure it's trained to be the otherwise same image in and out. This is deliberate to keep the task modular, focused, and predictable.
9:26 AM - 27 Aug 2020
0 replies
0 retweets
1 like
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.