Struct FitOptions 

pub struct FitOptions {

    pub latent_cloglog: Option<LatentCLogLogState>,
    pub mixture_link: Option<MixtureLinkSpec>,
    pub optimize_mixture: bool,
    pub sas_link: Option<SasLinkSpec>,
    pub optimize_sas: bool,
    pub compute_inference: bool,
    pub skip_rho_posterior_inference: bool,
    pub max_iter: usize,
    pub tol: f64,
    pub nullspace_dims: Vec<usize>,
    pub linear_constraints: Option<LinearInequalityConstraints>,
    pub firth_bias_reduction: bool,
    pub adaptive_regularization: Option<AdaptiveRegularizationOptions>,
    pub penalty_shrinkage_floor: Option<f64>,
    pub rho_prior: RhoPrior,
    pub kronecker_penalty_system: Option<KroneckerPenaltySystem>,
    pub kronecker_factored: Option<KroneckerFactoredBasis>,
    pub persist_warm_start_disk: bool,
}

Fields

latent_cloglog: Option<LatentCLogLogState>

mixture_link: Option<MixtureLinkSpec>

optimize_mixture: bool

sas_link: Option<SasLinkSpec>

optimize_sas: bool

compute_inference: bool

skip_rho_posterior_inference: bool

Internal lifecycle knob for fits whose result will be immediately superseded. Keeps ordinary inference work but skips the live-objective rho posterior certificate/escalation until the returned model is known.

max_iter: usize

tol: f64

nullspace_dims: Vec<usize>

linear_constraints: Option<LinearInequalityConstraints>

firth_bias_reduction: bool

Use Jeffreys/Firth bias reduction for supported likelihoods.

Model-fitting paths must pass this explicitly through every objective evaluator so baseline fits, spatial hyperparameter evaluations, outer line searches, final refits, and inference all optimize the same target.

adaptive_regularization: Option<AdaptiveRegularizationOptions>

penalty_shrinkage_floor: Option<f64>

Relative shrinkage floor for penalized block eigenvalues.

When Some(epsilon), a rho-independent ridge of magnitude epsilon * max_balanced_eigenvalue is added to each eigenvalue of the combined penalty on the penalized block. This acts as a weak proper complexity prior that prevents barely-penalized directions from causing pathological non-Gaussianity in the posterior (e.g., extreme skewness under logit link with high-dimensional spatial smooths).

The ridge is rho-independent, so LAML gradients remain correct without modification (d(epsilon*I)/d(rho_k) = 0).

Typical value: Some(1e-6). Set to None or Some(0.0) to disable. Default: Some(1e-6).

rho_prior: RhoPrior

Fixed prior on smoothing parameters for explicit joint HMC sampling flows.

This prior is part of the sampled target itself, unlike rho_mode, which is only used to initialize chains near the REML solution.

kronecker_penalty_system: Option<KroneckerPenaltySystem>

Kronecker-factored penalty system for tensor-product smooth terms. When set, the REML evaluator uses O(∏q_j) logdet and KroneckerMarginal penalty coordinates instead of O(p³) eigendecomposition.

kronecker_factored: Option<KroneckerFactoredBasis>

Full Kronecker factored basis for P-IRLS factored reparameterization.

persist_warm_start_disk: bool

Engage the cross-process ON-DISK persistent warm-start layer.

Default false: only the always-on in-memory warm start runs, so a single fit and throwaway/replicate/CI-coverage loops pay zero disk I/O (#1082). Set true (threaded from FitConfig::persist_warm_start_disk) to engage cross-process / repeat-fit resume; the standard RemlState then calls enable_persistent_warm_start_disk().

Trait Implementations

impl Clone for FitOptions

fn clone(&self) -> FitOptions

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for FitOptions

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl Default for FitOptions

fn default() -> FitOptions

Returns the “default value” for a type.