Struct MultinomialLogisticRegression 

pub struct MultinomialLogisticRegression { /* private fields */ }

Multinomial (softmax) logistic regression over an ordered K >= 2 label set.

See the module documentation for the objective, the sklearn C relation, and the deliberate omissions.

Implementations

impl MultinomialLogisticRegression

pub fn new(n_classes: usize) -> MultinomialLogisticRegression

Creates an unfitted head for n_classes classes.

n_classes is validated at fit time, not here, so that every rejection travels through the same typed channel.

pub fn from_stored_coefficients( ordered_labels: Vec<String>, n_features: usize, weights: Vec<f32>, intercepts: Vec<f32>, ) -> Result<MultinomialLogisticRegression, HeadFitError>

Rebuild a head from stored coefficients — the RELOAD door (plan 03-08).

This is not a fit, and it does not claim to be one

Self::report stays None: no optimizer ran in this process, and a synthesized report would assert a convergence status nobody observed. The provenance of the coefficients belongs to whatever artifact carried them, and it is that artifact’s job to record the fit that produced them.

Every input is validated before the head exists, through the SAME label-set rule the fit path uses ([validate_label_set]) plus the coefficient-arity and finiteness checks a fit cannot get wrong. A head returned from here therefore satisfies predict_proba’s preconditions by construction, exactly as a fitted one does.

Errors

HeadFitError::InvalidInput carrying the offending HeadInputError: too few classes, a mis-sized label map, an empty or duplicated label, a zero feature dimension, a weight or intercept array of the wrong length, or a non-finite coefficient.

pub fn with_max_iter(self, max_iter: usize) -> MultinomialLogisticRegression

Sets the maximum L-BFGS iteration budget (default DEFAULT_MAX_ITER).

pub fn with_tol(self, tol: f64) -> MultinomialLogisticRegression

Sets the gradient-norm convergence tolerance (default DEFAULT_TOL).

pub fn with_history_size( self, history_size: usize, ) -> MultinomialLogisticRegression

Sets the L-BFGS correction-pair history size (default DEFAULT_HISTORY_SIZE).

pub fn n_classes(&self) -> usize

Number of classes this head is configured for.

pub fn n_features(&self) -> Option<usize>

Fitted feature dimension, or None before a successful fit.

pub fn weights(&self) -> &[f32]

Fitted weights as K * d values in row-major order, or an empty slice.

pub fn intercepts(&self) -> &[f32]

Fitted intercepts (K values), or an empty slice.

pub fn labels(&self) -> &[String]

Ordered labels; index == weight-matrix row.

pub fn report(&self) -> Option<&HeadFitReport>

Report from the last successful fit.

pub fn fit( &mut self, features: &[Vec<f32>], class_indices: &[usize], ordered_labels: &[String], regularization: Regularization, ) -> Result<HeadFitReport, HeadFitError>

Fits the head.

Returns the HeadFitReport on convergence, or a typed HeadFitError otherwise. Non-convergence is an error, not a warning.

A failed fit leaves the head UNFITTED

The fitted state is discarded on entry, before anything can fail. Otherwise a head that fitted once and was then re-fitted with data the gate rejects would keep the PREVIOUS fit’s weights, n_features, labels and report — so predict would answer with a model the caller believes it failed to build, and report() would describe a converged run that the last call did not perform. NotFitted is the honest answer after a rejected fit, and it is only reachable if the state is cleared here.

pub fn predict_logits( &self, features: &[Vec<f32>], ) -> Result<Vec<Vec<f64>>, HeadFitError>

Per-row class logits, accumulated in f64 from the f32 store.

THE single logit implementation: Self::predict_proba is this plus a softmax, so a caller that needs both — VerifiedSetFitModel::classify, which reports logits alongside probabilities — cannot obtain a pair that disagrees with itself. Extracting it also removed the standing temptation to write a second copy of the accumulation loop next to the caller that wanted logits.

The accumulation ORDER is unchanged from the original predict_proba: intercept first, then j ascending. That order is what the artifact writer recorded its probe logits in, so it is load-bearing rather than incidental.

pub fn predict_proba( &self, features: &[Vec<f32>], ) -> Result<Vec<Vec<f64>>, HeadFitError>

Per-row class probabilities, computed with f64 logit accumulation.

Each returned row is finite and sums to 1 within 1e-6.

pub fn predict_indices( &self, features: &[Vec<f32>], ) -> Result<Vec<usize>, HeadFitError>

Per-row predicted class indices, breaking exact ties to the lowest index.

pub fn predict( &self, features: &[Vec<f32>], ) -> Result<Vec<String>, HeadFitError>

Per-row predicted labels, breaking exact ties to the lowest label index.

Trait Implementations

impl Clone for MultinomialLogisticRegression

fn clone(&self) -> MultinomialLogisticRegression

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for MultinomialLogisticRegression

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

Formats the value using the given formatter.