nsiccha

I've found that behaviour confusing, as it's also not documented "anywhere" (where it's immediately obvious to the user). Example code: ```julia sm = Stan.SampleModel(...) # This runs with engaged=true Stan.stan_sample(sm;...

More flexibly handle NaNs

9

Would be nice to be able to provide some kind of error/NaN handler, to e.g. keep track of how often and why chains run into problematic regions.

Parallel progress logging is slow with the (by default enabled) VSCode progress frontend.

8

While testing out the code in https://github.com/TuringLang/AbstractMCMC.jl/pull/168, I've noticed (again) that there is significant overhead when using VSCode's progress frontent. Using VSCode's progress frontent, something that takes roughly 1s when...

Lift the "(Transformed) parameters cannot be integers" restriction

5

When working with ragged arrays, it's sometimes useful to me to have transformed parameters of the type ```stan tuple( array[n_groups] int, vector[n_total] ) ``` However, this is not allowed, due...

As per the title. There's really no reason to disallow it, and it just makes usage for users more annoying - not even harder.

`cholesky_factor_corr[n]` reliably fails to initialize (from [-2,+2]) for "moderate" `n`

3

Using the `cholesky_factor_corr[n]` parameter type reliably causes initialization to fail from around `n=40`. This happens because for high enough `n` and if uniformly sampling the unconstrained draws from [-2,+2], some...

@penelopeysm: The other reason (apart from the implementation potentially being subtly wrong) I've never used [NUTS.jl](https://github.com/nsiccha/NUTS.jl) myself is because, even though it introduces less overhead than AdvancedHMC's or DynamicHMC's implementations,...

Opening to keep track of: @mhauru - [ ] [says](https://github.com/TuringLang/AdvancedHMC.jl/pull/473#discussion_r2550630867): "The two-argument version of show should, according to the [docs](https://docs.julialang.org/en/v1/base/io-network/#Base.show-Tuple%7BIO,%20Any%7D)," - [ ] [says](https://github.com/TuringLang/AdvancedHMC.jl/pull/473#discussion_r2550763090): "This looks like something that could...

Make it easier to select the used adaptation

1

Currently, AFAICT, dual averaging is kind of hard-coded into NUTS, and the mass matrix adaptation simply depends on the type of the provided metric. Would be nice if there were...

I think it's pretty clear that the current user facing interface (in e.g. `sample`) to e.g. "sample with NUTS with some user-defined number of warm-up and sampling transitions" can be...