nirs4all.pipeline.dagml.operator_routing module

Operator-routing registry: dag-ml graph node → real nirs4all/sklearn operator.

Phase 2a lowers a live pipeline to compat DSL; the compiler turns each step into a graph node carrying an operator ref + params. This module inverts that for execution: given a node it instantiates the concrete operator a host controller fits/transforms/predicts with. Mechanism-independent — both the CLI process-adapter and an in-process callback route the same way.

The operator ref shape differs by node kind (verified against the compiled vertical slice):

  • transform — {"class": "<FQN>"}; params live on the node.

  • y_transform — {"class": "<FQN>", "params": {...}}; params live on the operator.

  • model — a bare short name string ("PLSRegression"); params on the node.

Transforms/y_transforms carry a fully-qualified class resolved by import; models carry a short name resolved through an explicit allow-table (short names are ambiguous across modules, so silent fuzzy resolution is refused). The transform-vs-target distinction is the node kind, never the class — a bare scaler is an X transform, a y_processing scaler is a target transform.

nirs4all.pipeline.dagml.operator_routing.route_operator(operator_kind: str, operator_ref: str, params: dict[str, Any] | None = None, *, variant_overrides: dict[str, Any] | None = None) → object[source]

Instantiate the operator for one node.

Parameters:

  • operator_kind – transform | y_transform | model.

  • operator_ref – a fully-qualified class (transform/y_transform) or a short model name (model).

  • params – operator constructor params.

  • variant_overrides – per-variant params that win over params (generator sweeps).