split_rhat() and bulk_ess() return NA for a run they cannot score: too
few iterations, one chain, or a coordinate that never moved. Every parameter
can come back NA at once, and max(na.rm = TRUE) over all-NA is -Inf
with a warning, which the print method used to show as if it were a value.
A partially scored run is reported from the parameters that did get a
number, with a count of the ones that did not.
Arguments
- d
The data frame from
mcmc_diagnostics().
Details
When d carries an n_evals attribute – slice_sample_surrogate() sets
one, since it is the cost of the run the adapted slice width is trying to
keep down (see nsbi_mcmc) – it is appended to the summary. A "stan"
sampler's diagnostics have no such concept and print without it.