Submodel docs

[penelopeysm]

• The two 'outputs' of a model: return value + random variables and how to extract them
• Basic submodels, prefixing
• Sharp edges with this (i.e. likelihoods in submodels - can't condition on a ~ to_submodel(...))
• How to actually condition values in submodels
• Look to the future Unification of submodels and distributions TuringLang/Turing.jl#2485 and technical work in [wip] submodel explorations TuringLang/DynamicPPL.jl#815

[penelopeysm]

@lunafazio You found this repo! 😄

If you were interested in the docs, I've written them up here: https://turinglang.org/docs/pr-previews/627/usage/submodels/ It's currently in a PR and not merged to main yet (someone will need to review).

[lunafazio]

Thank you, that's very helpful! I'm currently implementing a shrinkage prior that I would like to make easily available for potential users as a package. Using submodels seemed like the right approach, but given the interface has already changed once, I wanted to look into it more deeply this time around to know what to expect in the future.

Personally I really liked the proposal you show in #2485. I feel it would reduce the cognitive overhead of my potential users (applied researchers) since I could tell them that they can just "plug and play" with our prior as if it was any other prior distribution. It's also just prettier to look at IMO. 🙂

[penelopeysm]

It's, unfortunately, a bit of a contentious topic. I still like my original proposal a ~ to_submodel(model) => b the best, and I think it's a lot clearer (the only difference from the original issue is that we still need to have to_submodel() as that allows us / the user to control prefixing), but Hong really likes NamedTuples.

[penelopeysm]

(I don't think that actually implementing it is a super difficult thing; I pretty much know how to do it. It's just that getting consensus on the interface is tricky 😅 )

[lunafazio]

That was indeed my impression after reading through the thread, which is why I figured that as an uninformed newcomer I'd just abstain from adding to the pile of opinions. 🙃

In any case, thank you for your work! I really appreciate the clarity with which you continuously convey all the considerations going into these decisions as it is very helpful for getting a sense of where things are headed.