@noelwelsh Can you recommend a good AD library for Scala?
-
-
Replying to @darrenjw
I don't think there is a good standalone package. You can find it embedded in various deep learning packages (e.g. Deep learning 4j)
1 reply 0 retweets 1 like -
Replying to @noelwelsh
Thanks. It's a shame, though. If we had a good stand-alone package, we could easily use it for both DL and HMC, for example, and other things...
2 replies 0 retweets 0 likes -
Replying to @darrenjw @noelwelsh
I think a major challenge is finding a good typesafe design for tensors in Scala. There's https://github.com/ctongfei/nexus or my similar https://github.com/avibryant/nda but these are heavyweight and opinionated. For HMC I cop out and just use scalars; *much* simpler API but wouldn't fly for DL.
1 reply 0 retweets 3 likes -
-
Replying to @noelwelsh @avibryant
Breeze covers a reasonable proportion of numpy, but it doesn't yet do tensors - there's currently discussion on github regarding the "best" way to do them...
1 reply 0 retweets 0 likes
It seems very difficult to retrofit tensors + AD to a system that wasn't designed for them ground-up, in the same way that TF had to reimplement NumPy rather than building on top of it. Especially in a Scala context where I would expect a static compute graph vs tape-based AD.
-
-
agreed - tensors + AD need to be designed together (with very low overhead for scalars)
0 replies 0 retweets 1 likeThanks. Twitter will use this to make your timeline better. UndoUndo
-
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.