Context- regarding the new Bit-M models. I'm personally excited because I know the Facebook WSL models make a big difference practically. There's an easy to miss variation that looks more practical: 2x wide 152 layer. Competitive benchmarks, way less memory needed.https://twitter.com/wightmanr/status/1263615215108870151 …
-
Show this thread
-
To clarify- the Facebook wsl resnets do a great job on image to image applications (like DeOldify). I just can't use them for anything commercial. BiT-M models are Apache 2.0 :)
2 replies 0 retweets 3 likesShow this thread -
Replying to @citnaj
Do you think it would be worth trying danbooru on with this model?
3 replies 0 retweets 1 like
Replying to @CpnTaters
2/ as I train these is that they're natively trained at higher resolution (480px vs 224px) and it seems that what results from that is that you get lower performance at 64px if you start training there. Pretty sure it'll get better past that point (haven't gotten there yet).
12:51 PM - 25 May 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.