search

TuringLang / Bijectors.jl

Implementation of normalising flows and constrained random variable transformations
https://turinglang.org/Bijectors.jl/
MIT License
199 stars 32 forks source link

Remove ADBackend and ADBijector #242

Closed torfjelde closed 1 year ago

torfjelde commented 1 year ago

ADBijector is, AFAIK, never used. In addition, it's a very trivial thing to implement on your own.

Because of this, I feel like it's unnecessary to add all the additional complexity that is ADBackend just for the sake of maybe one or two use-cases of ADBijector. Hence I think we should just drop it as part of the breaking release.

yebai commented 1 year ago

@torfjelde can you resolve the merge conflicts introduced by #214?

torfjelde commented 1 year ago

Aight, tests should now be passing.

BUT shouldn't this also technically be another breaking change? I.e. we should bump to 0.12? It's a bit annoying given that we just made a breaking release, but seems like the right course of action IMO. @devmotion @yebai

devmotion commented 1 year ago

Yes, it removes exported and documented functionality, so IMO it deserves another breaking release. I don't think it's too bad as the number of direct dependents is fairly low: https://juliahub.com/ui/Packages/Bijectors/39uFz/0.10.6?page=2

torfjelde commented 1 year ago

Version bumped. Should be ready to go once tests pass.

yebai commented 1 year ago

thanks @torfjelde and @devmotion!