David Widmann
David Widmann
Another advantage of handling it in the `propose` step is that you don't have to construct the `MvNormal` proposal, which probably is not very efficient.
Slightly unrelated thought: > being able to track acceptances/rejections might be useful I think probably one would not want to keep track of these statistics in a separate `trackstep` function...
> Any way we could reuse the Welford implementation from AdvancedHMC? Only if it's moved to a separate package I assume.
Intuitively I would have thought `adaptor` should be part of an `AdaptiveProposal` but not a general property of `MetropolisHastings`. That seems a bit cleaner but maybe you noticed some advantages...
IMO one should just not return a `Transition` object from `propose`. When the `Transition` is constructed, the log density is already evaluated which seems weird. This came up in a...
@willtebbutt Is that the test you had in mind? Generally, I'm not sure how interesting/relevant tests with repeated columns are in the two-argument case as this will just duplicate entries...
> > Pair{ > Seems no? See minimal example below. It did for me. The main point of my suggestion was to use `Pair{
> EDIT: Ignore me. Basically it has to be A(args::Tuple{\u003C:Real,Int}...) instead of the signature above. Yes, that's what I tried. But I did not realize that it makes a difference.
It's not used in the tests, so it seems some dependency holds it back.