Store accumulators in `Context` instead of `VarInfo`

[yebai]

Does storing accumulators in contexts (e.g. as a field of AccumulationContext ¹) make sense instead of VarInfo?

Thoughts?

cc @mhauru @penelopeysm?

Footnotes

1. see discussions here for renaming DefaultContext to AccumulationContext ↩

[penelopeysm]

If we adopt the view that

VarInfo = state
Context = something that controls how that state is updated

then accumulators, as a type of state, would belong inside VarInfo. (The distinction between varinfo and context might not be super clearly delineated now - for example we have some contexts that track state, although these are slated to be removed in the latest PRs) - but I think that it's one that makes sense. otherwise I don't see why we have two separate arguments to the model evaluation function, we could just have a single struct that contains both varinfo and context)

[mhauru]

My thoughts are similar to Penny's, VarInfo to me seems like a more natural home for accumulators. Or, possibly, some separate object, like something called __state__ that has as its fields both a VarInfo and an AccumulatorTuple.

[yebai]

The proposed semantics for Contexts and VarInfo sound sensible. Thanks for the clarification!.

One more thing for further discussion: would it make sense (in the longer run, see TuringLang/Turing.jl#2422)

1. replace tilde_assume and tilde_observe (and their bangbang versions) with a unified tilde_obbsume, and use value types in VarInfo to determine between parameters, data, fixed, conditioned, and missing values (see Simplify the context mechanism in DynamicPPL #895 (comment))
2. enforce VarInfo acts on tilde_obbsume function, while Context acts on tilde function?

EDIT: proposed callstack tilde --> tilde_obbsume --> accumulate_obbsume.

[mhauru]

I'd be in favour, long-term, of a single unified tilde. This would simplify code logic because we wouldn't have multiple ways of specifying conditioned variables (function arguments vs set with ConditionContext). Happy to investigate the use of value types, I think we need to test them out for some simple case to understand e.g. performance implications. If they can be used without big performance drawbacks they could be really good for these sorts of things.

On 2. I have no strong opinion straight away. You mean that tilde would dispatch on different context types, and then tilde_obssume would not?

[yebai]

Having a unified tilde for everything is also a good idea.