Skip to contents

CmdStan Fitter (user-supplied Stan programs)

Source: R/cmdstan-fitter.R
CmdStanFitter.Rd

Run user-supplied Stan programs via cmdstanr without brms. Compilation is cached by cmdstanr (by file hash); each daemon compiles-or-cache-hits on first use, so there is no model-bank integration (D2).

The Stan program may declare generated-quantities blocks for log_lik (a vector of pointwise log-likelihoods) and optionally epred (the expectation mu). Declare their names via the log_lik / epred arguments.

Newdata prediction is out of scope (raw Stan has no newdata semantics): supports_predictions is always FALSE. When epred is supplied, predict_epred returns the in-sample GQ matrix for LOO-based metrics, while predict_fit remains unsupported. Test-set metrics require brms or a custom fitter.

Usage

CmdStanFitter(
  stan_file = NULL,
  stan_code = NULL,
  stan_data,
  log_lik = NULL,
  epred = NULL,
  chains = 4L,
  iter_warmup = 1000L,
  iter_sampling = 1000L,
  adapt_delta = NULL,
  max_treedepth = NULL,
  parallel_chains = 1L,
  init = NULL,
  ...
)

Arguments

stan_file

Path to a .stan file, or NULL if stan_code is supplied.

stan_code

Character string of Stan code (used when stan_file is NULL). Either stan_file or stan_code is required.

stan_data

Function function(data_bundle, fit_spec) -> list mapping a bayesim data bundle + fit spec to the Stan data list. Required.

log_lik

Name of the generated-quantities log-lik vector in the Stan program, or NULL if the program has no log_lik GQ. When NULL, supports_log_lik/supports_loo are FALSE and elpd metrics degrade.

epred

Optional name of an epred/mu GQ matrix or vector. When supplied, predict_epred returns the in-sample GQ matrix (S x N) for LOO-based metrics; it does not enable newdata predictions.

chains

Integer number of MCMC chains.

iter_warmup

Integer warmup iterations per chain.

iter_sampling

Integer sampling iterations per chain.

adapt_delta

Numeric adapt_delta control parameter, or NULL for Stan default.

max_treedepth

Integer max_treedepth control parameter, or NULL.

parallel_chains

Integer number of chains to run in parallel.

init

Passed through to $sample() (NULL, "random", "0", or a list).

...

Additional named arguments passed through to cmdstanr's $sample().

Value

An S7 CmdStanFitter object.

See also

Fitter, BrmsFitter, LinearRegressionFitter

Examples

if (FALSE) { # \dontrun{
fitter <- CmdStanFitter(
  stan_file = "model.stan",
  stan_data = function(bundle, spec) list(N = nrow(bundle$train), y = bundle$train$y),
  log_lik = "log_lik"
)
} # }