David Widmann

The docstring of `sample` (which BTW is not Turing-specific or owned by Turing) explains how it is used: https://github.com/TuringLang/AbstractMCMC.jl/blob/master/src/sample.jl#L71 Maybe we could just add a link to `sample` in the...

Yes, probably it would be useful to add the signatures to the docstrings. We could use DocStringExtensions.jl to ensure that they are correct and simplify updates. Would you like to...

> Essentially, we generate a Module and use that as a pseudo-function struct This seems a very surprising and non-standard use of `Module`s :eyes: Maybe we do something similar to...

Isn't this only a question of naming? Also with something like `initialize_state` implementations should still be the same, shouldn't they? > Also, is there a reason why `init_params` became optional?...

> but `AdvancedHMC` ends up doing two transitions while `EllipticalSliceSampling` does just one. IMO this is a bug in AdvancedHMC. I think the expected behaviour is as in EllipticalSliceSampling and...

AdvancedMH just wraps the initial sample in a state, it does not perform any "transition".

Isn't that a more general issue/question that is not specific for AdvancedMH? Also with other samplers you have to save the RNG if you want to continue with exactly the...

I think that use case is also related to https://github.com/TuringLang/AbstractMCMC.jl/issues/109. The annoying part about handling it in `step` is that every sampler package has to adjust for it. I wonder...

Hmm since it is already possible to use arbitrary callbacks and one can always combine them manually, I'd say it's not necessary to support tuples or arrays of callbacks. I...

> Not even adding a `MultiCallback` struct which is like 5 lines of code? I opened this issue because this is what I did myself and I was like "well,...