espnet3.components.trainers.trainer.ESPnet3LightningTrainer

source

ESPnetLightningModule = None, exp_dir: str = None, config: DictConfig | Namespace | Dict[str, Any] = None, best_model_criterion=None)

Bases: object

A wrapper around Lightning’s Trainer to provide ESPnet3-specific integration.

This trainer ensures compatibility with ESPnet’s dataloader, callbacks, and configuration system. It initializes the model, handles weight initialization, sets up the training strategy, logger, plugins, and integrates with ESPnet-specific callbacks and samplers.

config

Training configuration.

• Type: Union[DictConfig, Namespace, Dict[str, Any]]

model

ESPnet3 LightningModule instance.

• Type:ESPnetLightningModule

trainer

Underlying PyTorch Lightning trainer object.

• Type: lightning.Trainer

Initialize the trainer with model, configuration, and training setup.

Sets up weight initialization, accelerator/strategy/logger/profiler/plugins, applies ESPnet-specific dataloader constraints, prepares callbacks, and finally constructs the underlying Lightning Trainer.

• Parameters:
  • model (ESPnetLightningModule , optional) – LightningModule to train.
  • exp_dir (str , optional) – Experiment directory for logs and checkpoints.
  • config (DictConfig | Namespace | Dict *[*str , Any ] , optional) – Training config.
  • best_model_criterion (ListConfig , optional) – Criteria for selecting ckpt.

collect_stats(*args, **kwargs)

Collect dataset statistics with the espnet-3’s parallel package.

• Parameters:
  • *args – Positional arguments passed to model.collect_stats().
  • **kwargs – Keyword arguments passed to model.collect_stats().

fit(*args, **kwargs)

Start the training loop using Lightning’s fit method.

• Parameters:
  • *args – Positional arguments passed to trainer.fit().
  • **kwargs – Keyword arguments passed to trainer.fit().

NOTE

Always uses the internally stored model (self.model) when calling fit.

validate(*args, **kwargs)

Run validation using Lightning’s validate method.

• Parameters:
  • *args – Positional arguments passed to trainer.validate().
  • **kwargs – Keyword arguments passed to trainer.validate().
• Returns: Validation results.
• Return type: List[Dict[str, Any]]