Option for loo

[mattocci27]

Help

• I understand and agree to https://books.ropensci.org/targets/help.html.

Description

The addition of an option like return_loo = TRUE, similar to return_summary = TRUE might be a good idea because it is hard to compute loo::loo when tar_stan_mcmc is run inside tar_map.

For example, we can use summary inside this tar_map like:

mapped <- tar_map(
  values = list(data = 1:5),
  tar_stan_mcmc(
      fit,
      "model.stan",
      ...),
  tar_targets(
     summary_modified,
     some_functions(fit_summary)
  ))

However, to use loo::loo() for each output, we need another tar_map, which can be something like:

loo_mapped <- tar_map(
  values = list(mcmc = rlang::syms(paste0("fit_mcmc_model_, 1:5))),
  tar_target(
    loo,
    loo_function(mcmc)
  )
)

Considering that tar_map is useful for creating different parameter settings, it makes sense that stantargets should be able to handle loo computations for model comparisons. What do you think?

[wlandau]

Definitely a neat idea. However, I hesitate to directly build in features this specific, since the Stan ecosystem changes rapidly and there are many different ways to invoke loo.