Roadmap for depreciating `VarInfo` in favour of `SimpleVarInfo`

[yebai]

This is meant as an umbrella issue for documenting the steps required to depreciate VarInfo in favour of SimpleVarInfo-equivalwent.

Some known items:

• [x] Lineralisation for SimpleVarInfo: solved by VarNamedVector
• [x] Support for Gibbs sampling: e.g. GID stuff in VarInfo: solved by condition-based Gibbs sampler
• [x] Support for dynamic model sizes: solved by VarNamedVector
• [ ] Support for particle MCMC samplers:

cc @torfjelde

[devmotion]

Support for Gibbs sampling: e.g. GID stuff in VarInfo

IIRC we don't use GID etc in the Gibbs sampler anymore.

[torfjelde]

IIRC we don't use GID etc in the Gibbs sampler anymore.

Are you referring to Gibbs or GibbsConditional? Because in the former I'm pretty sure it's used to determine which variables are related to which sampler.

[devmotion]

I think that's done solely by the sub-samplers and their parameter space (eg https://github.com/TuringLang/Turing.jl/blob/9f8a9c4c476095d45246b4924d5cd542f3f8d506/src/inference/gibbs.jl#L243)? The logic was changed quite a bit in https://github.com/TuringLang/Turing.jl/pull/1500.

[torfjelde]

I think that's done solely by the sub-samplers and their parameter space

Exactly what I was referring to yeah. And the current "selector" stuff for samplers IMO should go away, and we should instead keep track of which varnames correspond to which sampler in the Gibbs sampler or something.

[torfjelde]

@yebai we might want to move this issue to a discussion instead. In a discussion we can make sub-threads, which makes it much easier to keep track of the ongoing conversations regarding particular points rather than having a flat thread like in an issue.