espnet3.utils.run_utils.validate_experiment_context

source

espnet3.utils.run_utils.validate_experiment_context(, training_config: DictConfig | None, inference_config: DictConfig | None, metrics_config: DictConfig | None, stages_to_run: Sequence[str]) → None

Validate that runtime configs have enough experiment identity.

Supported runner modes.

1. Training-backed mode, where training_config is present and provides experiment identity for inference and metrics.
2. Standalone inference/metrics mode, where the runtime config must define its own exp_tag or concrete exp_dir.

• Parameters:
  • training_config (DictConfig | None) – Training config selected for the current run.
  • inference_config (DictConfig | None) – Inference config for the infer stage.
  • metrics_config (DictConfig | None) – Metrics config for the measure stage.
  • stages_to_run (Sequence *[*str ]) – Resolved stage names requested by the runner.
• Returns: Validation succeeds silently.
• Return type: None
• Raises:ValueError – If infer or measure is requested without training_config and the corresponding runtime config does not define exp_tag or a concrete exp_dir.

Notes

Validation is stage-aware. For example, a missing metrics identity is ignored when measure is not in stages_to_run.

Examples

Allow standalone inference when the config already defines its experiment identity:

inference_config = OmegaConf.create(
    {"exp_tag": "whisper_eval", "exp_dir": "./exp/whisper_eval"}
)
validate_experiment_context(
    training_config=None,
    inference_config=inference_config,
    metrics_config=None,
    stages_to_run=["infer"],
)
# Succeeds because standalone inference can derive exp_dir.

Reject standalone inference when experiment identity is missing.

inference_config = OmegaConf.create(
    {"inference_dir": "${exp_dir}/inference"}
)
validate_experiment_context(
    training_config=None,
    inference_config=inference_config,
    metrics_config=None,
    stages_to_run=["infer"],
)
# Raises:
#   ValueError: infer stage requires --training_config or a
#   standalone inference_config with exp_tag/exp_dir.

Accept training-backed inference even when inference_config does not define exp_tag:

training_config = OmegaConf.create(
    {"exp_tag": "train_asr_rnn", "exp_dir": "./exp/train_asr_rnn"}
)
inference_config = OmegaConf.create(
    {"inference_dir": "${exp_dir}/inference"}
)
validate_experiment_context(
    training_config=training_config,
    inference_config=inference_config,
    metrics_config=None,
    stages_to_run=["infer"],
)
# Succeeds because training-backed mode is active.