Standardize logp, logweights, and "info" per-sample metadata

[cscherrer]

As I understand, there's currently no standard for passing values like the log-density or log-weight (when that exists) for each sample. Algorithms that require these will each have to come up with their own standards, which will then need to be managed separately.

I have some description of how I'm handling these one instance of SampleChains here.

It would make things much easier if things like this can be standardized, to the extent they're common across samplers.

[cpfiffer]

We had this but removed it. Our view is that this should be standardized within the various ecosystems that touch AbstractMCMC, for exactly the reason you've highlighted -- some samplers and model types need vastly different internals, and it's basically impossible (I think) to build a robust internal interface that standardizes everything.

[cscherrer]

"Everything" yes, sure. I'm suggesting standardizing optional per-sample named fields:

• samples, for samples (EDIT: This name only really makes sense for a collection of samples)
• logp or something similar for log-density the samples were drawn from
• logweights for those
• info for miscellaneous, which will vary from one sampler to the next

[cpfiffer]

Yeah, that's about what we had. We kind of decided that it not worth standardizing because cross-talk between samplers or whatever is already tricky enough. I think if this were to be standardized, it would be more at the AbstractPPL level and not at the AbstractMCMC level.