Abstract Fitter Class

Abstract base class for Bayesian model fitters in bayesim.

The Fitter class defines the interface that all model fitters must implement. It provides a consistent API for fitting Bayesian models, extracting posterior draws, generating predictions, computing log-likelihoods, and performing model diagnostics.

Usage

Fitter(
  name = character(0),
  supports_predictions = FALSE,
  supports_log_lik = FALSE,
  supports_loo = FALSE,
  supports_epred = FALSE
)

Arguments

name

Character string identifying the fitter (e.g., "stan", "brms")

supports_predictions

Logical indicating if the fitter supports predictions

supports_log_lik

Logical indicating if the fitter supports log-likelihood computation

supports_loo

Logical indicating if the fitter supports LOO-CV

supports_epred

Logical indicating if the fitter supports posterior expectation predictions (predict_epred(); required by the r2_loo / rmse_loo LOO metrics)

Value

An S7 class object representing the abstract Fitter

Methods

The following S7 generics form the fitter interface. A minimal custom fitter only needs to implement fit_model() and extract_draws(); diagnostics default to an empty list and unsupported optional capabilities default to NULL.

fit_model(fitter, data_bundle, fit_spec, seed, task_ctx)

Main fitting method

extract_draws(fitter, fit_result, variables = NULL)

Extract posterior draws

predict_fit(fitter, fit_result, newdata = NULL, seed = NULL)

Generate predictions

log_lik_matrix(fitter, fit_result, newdata = NULL)

Pointwise log-likelihood

loo_fit(fitter, fit_result)

LOO-CV computation

fit_diagnostics(fitter, fit_result)

Extract fit diagnostics

Creating Custom Fitters

To create a custom fitter, extend this class and implement methods for the core S7 generics: fit_model() and extract_draws(). Implement optional predict_fit(), log_lik_matrix(), and loo_fit() methods only when the matching supports_* property is TRUE. All matrices follow the draws-by-observations (S x N) orientation; see vignette("custom-fitters") for the full contract.

See also

LinearRegressionFitter, BrmsFitter, and CmdStanFitter for the built-in implementations, and validate_fitter() to check a custom fitter against the contract