espnet2.main_funcs.average_nbest_models.average_nbest_models

source

espnet2.main_funcs.average_nbest_models.average_nbest_models(output_dir: Path, reporter: Reporter, best_model_criterion: Sequence[Sequence[str]], nbest: Collection[int] | int, suffix: str | None = None) → None

Generate averaged model from n-best models.

This function averages the parameters of the n-best models based on specified criteria and saves the averaged model to the output directory. It creates symlinks for easy access to the best and averaged models.

• Parameters:
  • output_dir (Path) – The directory contains the model file for each epoch.
  • reporter (Reporter) – Reporter instance for tracking and managing model performance.
  • best_model_criterion (Sequence *[*Sequence *[*str ] ]) – Criteria to decide the best model, e.g. [(“valid”, “loss”, “min”), (“train”, “acc”, “max”)].
  • nbest (Union *[*Collection *[*int ] , int ]) – Number of best model files to be averaged. This can be a single integer or a collection of integers.
  • suffix (Optional *[*str ]) – A suffix added to the averaged model file name.
• Raises:UserWarning – If no nbest values are provided, a warning will be issued and defaults to 1.

Examples

>>> from pathlib import Path
>>> from espnet2.train.reporter import Reporter
>>> reporter = Reporter()
>>> average_nbest_models(
...     output_dir=Path("/path/to/models"),
...     reporter=reporter,
...     best_model_criterion=[("valid", "loss", "min")],
...     nbest=3,
...     suffix="final"
... )

NOTE

Ensure that the model files exist in the specified output directory before calling this function.