ESPnet3 Train Stage

The train stage runs model training using a PyTorch Lightning trainer based on the dataset and hyperparameters defined in training.yaml and saves model checkpoints and logs.

1. Run

python run.py --stages train --training_config conf/training.yaml

2. Configuration

Training is configured in training.yaml using the sections shown in the table below. For a detailed list of options, see Training Configuration and the links in the table.

Section	Description	Details
task	task entrypoint for ESPnet2-style models
model	model definition and normalization settings
dataset	train and valid splits	Data Organizer
dataloader	collate and iterator settings	Dataloader + Collate
trainer	Lightning trainer configuration	Trainer
optimizer, scheduler	single-optimizer training path	Optimizer + Scheduler
optimizers, schedulers	named multi-optimizer path	Multiple Optimizers and Schedulers
exp_dir	training output directory

3. Outputs

Training outputs are written under exp_dir, including:

• Checkpoints
• Logs
• (If configured) TensorBoard output

Related pages

Training configuration

See all options for cofiguring the train stage.

Inference stage

Information on the inference stage.

Trainer

Information about the trainer component