nirs4all.pipeline.dagml.structural_torch module
Closed declarations for native structural early/learned-late Torch HPO.
Only declarations, immutable input buffers and scope identity cross this bridge. DAG-ML owns proposals, grouped nested OOF, REFIT, scoring and selection. The existing real Torch estimator and Ridge host owners perform numerical fitting.
- nirs4all.pipeline.dagml.structural_torch.declared_torch_topologies(steps: list[Any], splitter: Any) list[list[Any]][source]
Validate the existing sequence _or_/branch/merge grammar once.
- nirs4all.pipeline.dagml.structural_torch.is_torch_choice(steps: list[Any]) bool[source]
Recognize Torch declarations without instantiating any network.
- nirs4all.pipeline.dagml.structural_torch.lower_torch_choices(steps: list[Any], splitter: Any) tuple[dict[str, Any], dict[str, Any], list[str]][source]
Transport unique logical model IDs and numeric conditional destinations.
- nirs4all.pipeline.dagml.structural_torch.prepare_torch_structure(steps: list[Any], splitter: Any, dataset: Any, spec: DagMLTuningSpec, run_options: dict[str, Any], native: Any) dict[str, Any][source]
Attest all raw inputs before native catalogue/model/proposal callbacks.