Select variables from the draw target

[yizhang-yiz]

In cmdstanr, for large models it's more efficient to do

CmdStanMCMC_object$draws(variables=c("v1","v2"))

than dumping all the variables into the memory. Is there a way to do that in x_draws_y target? From the code

stantargets/R/tar_stan_mcmc.R

Line 502 in 7c2e6f4

fit$draws() # Do not specify variables or inc_warmup.

it looks like the target doesn't take that into consideration.

For example, in the following scenario:

fit$draws(c("v1", "v2")) %>% mcmc_dens_overlay()

would be more efficient than

tar_read(x_draws_y) %>% posterior::as_draws_df() %>% mcmc_dens_overlay(pars=c("v1", "v2"))

[wlandau]

The variables argument of tar_stan_mcmc() does allow you to set specific variables for x_draws_y. That line you cited applies to the x_mcmc_y target, which saves the whole model object so you can change variables without needing to rerun the whole MCMC from scratch. In the short term for you, if you want to avoid saving the whole object, you can instead run tar_stan_mcmc_rep_draws(variables = "...") with 1 batch and 1 rep.

Where applicable, I am open to adding entirely new arguments to tar_stan_mcmc(), tar_stan_gq(), tar_stan_vb(), and tar_stan_mle() for variables and inc_warmup to control the big cmdstanr object. However, I'm a bit short on time for this. Care to send a pull request?

[yizhang-yiz]

Thanks @wlandau. Wouldn't changing variables argument in tar_stan_mcmc lead to targets reconstruction hence rerun the MCMC? During model exploration I'd like to fit model once, then examine different variables by retrieving them from draws, and generally I wouldn't know before hand which variables to look at. Having to refitting the model is exactly what the package tries to avoid.

[wlandau]

The last few commits fix both issues: add new variables_fit and inc_warmup_fit for the CmdStanFit target separately from the variables and inc_warmup of the draws target.

[yizhang-yiz]

Thanks a lot!