espnet3.components.callbacks.get_default_callbacks

source

espnet3.components.callbacks.get_default_callbacks(expdir: str = './exp', log_interval: int = 500, best_model_criterion: List[Tuple[str, int, str]] | List[List] = [('valid/loss', 3, 'min')]) → List[Callback]

Returns a list of standard PyTorch Lightning callbacks tailored for most training workflows.

Includes: : - ModelCheckpoint for saving the last model checkpoint (save_last)

• One or more <br/>

  `

<br/> ModelCheckpoint`s for saving the top-K checkpoints according to : specific metrics

• AverageCheckpointsCallback to compute and save the average model from top-K : checkpoints
• LearningRateMonitor to track and log learning rates during training
• TQDMProgressBar to show a rich progress bar during training

• Parameters:

  • expdir (str) – Directory to store checkpoints and logs.

  • log_interval (int) – Frequency (in training steps) to refresh the progress bar.

  • best_model_criterion (List *[*Tuple *[*str , int , str ] ]) –

    A list of criteria for saving top-K checkpoints. Each item is a tuple: (metric_name, top_k, mode), where:

    • metric_name (str): The name of the validation metric to monitor

    (e.g., “val/loss”).

    • top_k (int): Number of best models to keep.
    • mode (str): “min” to keep models with lowest metric, “max” for highest.
• Returns: A list of callbacks to be passed to the PyTorch Lightning : Trainer.
• Return type: List[Callback]

Example

>>> from callbacks import get_default_callbacks
>>> callbacks = get_default_callbacks(
...     expdir="./exp",
...     log_interval=100,
...     best_model_criterion=[("val/loss", 5, "min"), ("val/acc", 3, "max")]
... )
>>> trainer = Trainer(callbacks=callbacks, ...)