espnet3.systems.base.inference_runner.InferenceRunner

source

EnvironmentProvider, , idx_key: str = 'idx', hyp_key: str | Sequence[str] = 'hyp', ref_key: str | Sequence[str] = 'ref', **kwargs)

Bases: BaseRunner

Inference runner with strict output-format validation.

This runner implements forward to call a recipe-provided output function. The key names are configurable via idx_key and hyp_key/ref_key. hyp_key and ref_key may be a single string or a list of strings to support multiple hypothesis/reference fields.

Output format requirements: : - The result is a dict with exactly the configured keys.

• idx_key value is a scalar (list/tuple is not allowed).
• hyp_key and ref_key values may be scalars or lists/tuples. If lists are returned, each entry is written to its own SCP file (e.g., hyp0.scp, hyp1.scp).

Initialize the inference runner with output key settings.

static forward(idx, dataset=None, model=None, **kwargs)

Run inference for one dataset item and return output dict.

• Parameters:
  • idx – Integer index into the dataset.
  • dataset – Dataset providing inference entries.
  • model – Inference model callable on the configured input.
  • **kwargs – Expects input_key and output_fn_path.
• Returns: Dict containing idx and output fields for a single item, or a list of dicts for batched inputs (as returned by output_fn).
• Raises:
  • RuntimeError – If required I/O settings are missing.
  • KeyError – If required input keys are missing from the dataset item(s).
  • RuntimeError – If batched inference fails; includes guidance to disable batching when unsupported.

Notes

• input_key may be a string or a list/tuple of strings.
• Batched inputs are passed to the model as lists per key; padding is the model’s responsibility.

Examples

>>> # Single-item inference
>>> out = InferenceRunner.forward(
...     0, dataset=dataset, model=model,
...     input_key="speech", output_fn_path="m.mod.out_fn"
... )
>>> # Batched inference
>>> out = InferenceRunner.forward(
...     [0, 1], dataset=dataset, model=model,
...     input_key=["speech", "text"], output_fn_path="m.mod.out_fn"
... )