Cmdstanr: Implement equivalent of model_pars and par_dims of stanfit from rstan

[rok-cesnovar]

Running:

fit <- cmdstanr_example("logistic")
rstanfit <- rstan::read_stan_csv(fit$output_files())
rstanfit@model_pars
rstanfit@par_dims

prints

> rstanfit@model_pars
[1] "alpha" "beta"  "lp__" 
> rstanfit@par_dims
$alpha
integer(0)

$beta
[1] 3

$lp__
integer(0)

Should we implement the equivalent of this? If yes, any name suggestions?

Or are there any plans to add this to posterior (like posterior::variables()). Retrieving dimensions and names of parameters/variables isn't Stan specific. I can also add this to posterior if it would be better to live there.

[jgabry]

I agree this is useful information. I think some version of this could go in posterior but it's a little bit Stan-specific in that the dimension of alpha is integer(0) instead of 1 (to indicate a scalar instead of a 1-element vector or array). But that's a choice that RStan made to in some way match the Stan program. For posterior my first instinct is that alpha should have dimension 1 (but we should think about that and check with @paul-buerkner).

And maybe we can combine model_pars and par_dims into a single function variable_structure() (or something like that) which takes a draws object as input and just returns the named list?

# assume the draws object has variables
# alpha, beta[1], beta[2], beta[3], Sigma[1,1], Sigma[1,2], Sigma[2,1], Sigma[2,1], lp__

> variable_structure(draws)
$alpha
[1] 1

$beta
[1] 3

$Sigma
[1] 2 2

$lp__
[1] 1

Paul will also probably have suggestions if you open an issue over at posterior.

[jgabry]

Although maybe we should also add the equivalent of model_pars to fit$metadata(). Right now fit$metadata()$model_pars gives all element names, so we could add fit$metadata()$model_pars_named or something that only lists the unique non-element names.

[rok-cesnovar]

Although maybe we should also add the equivalent of model_pars to fit$metadata(). Right now fit$metadata()$model_pars gives all element names, so we could add fit$metadata()$model_pars_named or something that only lists the unique non-element names.

Agree. That also makes read_cmdstan_csv a more complete function standalone.

[paul-buerkner]

I think our current way forward in posterior is to implement those dimensional draws arrays (https://github.com/stan-dev/posterior/issues/13) via rvars (https://github.com/stan-dev/posterior/issues/8). On these decisions it will likely depend on how we represent the dims in posterior best.

[jgabry]

@rok-cesnovar This is currently listed under the beta release milestone, but this could probably wait for 1.0 right? That way we could go forward with the beta release sooner.

[rok-cesnovar]

Agreed, can move to 1.0. I was expecting I would have time to work on this in July, that is why I put the milestone on beta. But then other stuff got in the way.