nirs4all.pipeline.dagml.training_controls module

Historical sklearn training controls, applied inside each native fit scope.

These are estimator set_params overrides, not arbitrary fit keyword arguments. The historical controller ignored unknown keys; the general DAG host diagnoses them instead. Explicit SGD CV-weight requests use a separate run-scoped transfer owner; setting an estimator’s warm_start parameter alone cannot transfer weights.

nirs4all.pipeline.dagml.training_controls.apply_model_training_controls(model: Any, metadata: Mapping[str, Any], phase: str) → dict[str, Any]

Apply recognized estimator overrides after candidate selection, before fit.

nirs4all.pipeline.dagml.training_controls.apply_pipeline_folds_to_model(model: Any, metadata: Mapping[str, Any], phase: str, fit_ids: list[str]) → bool

Apply the materialized outer FoldSet to an AOM model’s current fit scope.

nirs4all.pipeline.dagml.training_controls.cv_weight_transfer_request(metadata: Mapping[str, Any]) → WarmStartRequest | None

Decode the explicit REFIT request without treating it as a fit keyword.

nirs4all.pipeline.dagml.training_controls.effective_training_controls(metadata: Mapping[str, Any], phase: str) → dict[str, Any]

Resolve refit-over-train precedence without mutating graph metadata.

nirs4all.pipeline.dagml.training_controls.report_model_training_controls(evidence: Mapping[str, Any], model: Any, sample_count: int) → None

Honor controller verbosity without forcing it into estimator parameters.