typedefs

typedefs ¶

Source contracts and batch container types.

Bases: Protocol

Task fields required by dataset explanation sources.

Bases: Protocol

Experiment fields required by dataset explanation sources.

Bases: Protocol

Config fields required by dataset explanation sources.

ExplanationBatch(observation, actions, display_observation, metadata, preprocess_observation)

A model-ready observation window plus metadata for saving explanations.

Parameters:

Name	Type	Description	Default
`observation`	`ObservationBatch`	Observation dictionary passed to the policy attribution functions.	required
`actions`	`ActionBatch \| None`	Optional action dictionary from the source sample. Offline dataset mode provides normalized/tokenized action targets when the dataset contains them. Online inference mode sets this to `None`.	required
`display_observation`	`dict[str, Tensor]`	Camera tensors used for visual overlays. These are kept separate because dataset batches may contain extra tokenized observation keys that are not displayable.	required
`metadata`	`dict[str, ExplanationMetadataValue]`	Lightweight source metadata used for output filenames and reports.	required
`preprocess_observation`	`bool`	Whether the explainer should run policy normalization/tokenization before attribution.	required

Bases: Protocol

Consumes explanation batches produced by online inference hooks.

explain_batch(batch)

Generate and persist explanations for one batch.

Source code in src/versatil/explainability/sources/typedefs.py

def explain_batch(self, batch: ExplanationBatch) -> None:
    """Generate and persist explanations for one batch."""
    ...