Trait ModelFactory 

pub trait ModelFactory: Send + Sync {
    // Required methods
    fn config_space(&self) -> SearchSpace;
    fn create(
        &self,
        params: &ParamMap,
    ) -> Result<Box<dyn StreamingLearner>, FactoryError>;
    fn name(&self) -> &str;

    // Provided methods
    fn warmup_hint(&self) -> usize { ... }
    fn complexity_hint(&self) -> usize { ... }
    fn n_features_hint(&self) -> usize { ... }
    fn supports_auto_builder(&self) -> bool { ... }
}

Factory for creating streaming learner instances from hyperparameter configurations.

Implementations define the hyperparameter search space (a typed SearchSpace) and how to construct a model from a sampled ParamMap.

Migration from positional HyperConfig

Pre-v10 factories returned a ConfigSpace of positional HyperParam entries and consumed a HyperConfig (a Vec<f64> indexed by position). That API is deprecated in favor of typed, named-access SearchSpace / ParamMap. The legacy types remain for one release cycle behind #[deprecated] to give downstream crates time to migrate.

Required Methods

fn config_space(&self) -> SearchSpace

The hyperparameter search space for this model type.

fn create( &self, params: &ParamMap, ) -> Result<Box<dyn StreamingLearner>, FactoryError>

Create a new model instance from a sampled parameter map.

The params are values drawn from Self::config_space. Factories access them by name via ParamMap::float / ParamMap::int / ParamMap::category. Conditional parameters whose gate did not fire are absent from the map; factories must use the _optional variants for those reads.

Returns Err(FactoryError) when the sampled hyperparameter combination is structurally invalid. The AutoML racing layer catches this error, logs a warning, and skips the offending arm rather than panicking.

fn name(&self) -> &str

Human-readable name for this model type (e.g., “SGBT”, “ESN”).

Provided Methods

fn warmup_hint(&self) -> usize

Minimum samples a new model needs before its metrics are meaningful.

Candidates that have seen fewer than warmup_hint() samples are protected from elimination during tournament rounds. This prevents neural architectures with warmup phases (ESN, Mamba, SpikeNet) from being prematurely killed by models that start predicting immediately.

The default is 0 (no warmup protection).

fn complexity_hint(&self) -> usize

Approximate model complexity (effective parameter count).

Used for complexity-adjusted elimination: models with higher complexity are penalized more when evaluation data is scarce. This naturally favors simpler models on sparse data and lets complex models prove themselves when data is abundant.

The default is 100 (moderate complexity).

fn n_features_hint(&self) -> usize

Number of input features this factory expects.

Used by the auto-builder to initialize the FeasibleRegion with correct dimensionality, ensuring config bounds (especially grace period and lambda) are properly calibrated.

The default is 1 (conservative estimate).

fn supports_auto_builder(&self) -> bool

Return true if the SPSA auto-builder (FeasibleRegion + DiagnosticLearner) is meaningful for this factory.

The auto-builder is designed for the SGBT family. Non-SGBT factories should return false (the default) so that the AutoTuner can log a warning and skip activating the adaptor rather than silently no-oping.

Dyn Compatibility

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors

impl ModelFactory for Factory