Native Julia implementations of NLE, NPE, etc.

[rsenne]

Just curious if you have any interest in implementing native julia versions these inference methods. I love the python sbi package but I think it would be a cool project to bring these methods to the language using Flux etc.

[bgroenks96]

Hi @rsenne , yes absolutely. I'm a bit unsure if it would really yield any practical benefits (e.g. in terms of performance), but avoiding the need to install python dependencies would already be a big win.

Just FYI though, I need to do some (significant) refactoring on this package which I hope to get done in the next month or so.

Also, regarding the choice of NN library, I would probably favor Lux.jl unless there is a very compelling reason to choose Flux.

[bgroenks96]

Though actually InvertibleNetworks.jl would perhaps be a good argument in favor of Flux.

[rsenne]

Ah great. Realistically, I'm not looking for any huge performance benefits as I imagine they will be nearly identical in speed, or at least not unreasonably different. I mostly just like the idea of this being all Julia. Would you want to ping me when refactoring is done, or is there anything you would like help on? I'd be interested in working this. Also, I haven't used Lux admittedly., but I'd be happy to learn the package for this purpose.

[bgroenks96]

Sure no problem, I'll drop a comment here when the refactoring is done. It will also probably correspond to a v0.2 release.

[bgroenks96]

@rsenne See #14 and let me know if you have any questions, concerns, or wishes for version 0.2, especially any suggested breaking changes to the public-facing APIs.

[rsenne]

Hey @bgroenks96 I think after looking at the current API--a lot of my ideas should be purely additive so I don't think i will break any current user facing API structure. If you want i can give you a quick set of my ideas if that would help you in any way

[bgroenks96]

Any input would be greatly appreciated!

[rsenne]

Are you able to request me as a reviewer? Otherwise I can give feedback here

[bgroenks96]

Unfortunately not. I guess I would opt for just discussing it here for now? Or on the PR if specific to that.

[rsenne]

I left a few comments over on the PR!

[bgroenks96]

Hi @rsenne, have you decided whether or not you want to proceed with this?

[rsenne]

Hi @bgroenks96 yes this is still in my radar. Hoping to start soon!

[rsenne]

Hey @bgroenks96 I was planning on getting this going this week--but i just stumbled upon this. Seems like I would be just re-inventing the wheel now. Maybe it would be worth joining forces or inter-operating with NeuralEstimators.jl?

[bgroenks96]

Ah, thanks, I hadn't seen that! Yes, absolutely, I think we could even just make this a direct dependency of SimulationBasedInference.jl.