Suggested Features

[ParadaCarleton]

I'm leaving these here as examples of things I wanted while building ParetoSmooth.jl:

1. Some kind of AbstractModel type that I can dispatch on.
2. A log_pdf function.
3. A way to instantiate a model on some counterfactual dataset -- either by calling some AbstractModel object on a dataset or by having condition/decondition functions.

[devmotion]

1. There exists AbstractPPL.AbstractProbabilistivProgram <: AbstractMCMC.AbstractModel. DynamicPPL.Model (ie Turing models) are a subtype of it.
2. There exists AbstractPPL.logdensity and loglikelihood, logprior, and logprior for Turing models.
3. There exists AbstractPPL.condition and AbstractPPL.decondition, and it's also implemented for Turing models (with some restriction for backward compatibility, as explained in the docstring).

[phipsgabler]

Yeah, this is basically all described in the interface draft. (The abstract supertype is not, I think, but already exists.)

Note though that we are still in the process of refining this. condition/decondition work in DynamicPPL, but only with named tuples. As described in the draft, there needs to come some more flexible syntax. And for counterfactuals -- I'm not sure what exactly you mean, but I mention the possibility of having a do-operator in the prob-syntax at some place below.

Feel free to reach out to me with question, I'm always keen on refining my understanding of what downstream packages need from the interface (are you in the Turing Slack?)