ESPnet3 Trainer Configuration (Training)

ESPnet3 uses the PyTorch Lightning Trainer for the train stage. Most fields under trainer: in train.yaml map directly to Lightning Trainer arguments.

Upstream reference:

• PyTorch Lightning Trainer

Common settings

Key	Description
accelerator	Backend selection (e.g., auto, gpu, cpu).
devices	Number of devices to use.
num_nodes	Number of nodes for distributed training.
max_epochs	Total training epochs.
accumulate_grad_batches	Gradient accumulation steps.
gradient_clip_val	Gradient clipping value.
log_every_n_steps	Logging frequency (steps).
check_val_every_n_epoch	Validation interval (epochs).
precision	Mixed precision setting (e.g., 32, bf16-mixed).
strategy	Distributed strategy (e.g., ddp, auto).
logger	Logger configuration (TensorBoard, CSV, etc.).
callbacks	Training callbacks (checkpointing, averaging, etc.).

ESPnet3 provides default callbacks. For details and extension points, see:

• doc/vuepress/src/espnet3/core/components/callbacks.md

Example

trainer:
  accelerator: auto
  devices: ${num_device}
  num_nodes: ${num_nodes}
  max_epochs: 10
  log_every_n_steps: 100
  gradient_clip_val: 1.0
  logger:
    - _target_: lightning.pytorch.loggers.TensorBoardLogger
      save_dir: ${exp_dir}/tensorboard
      name: tb_logger

Where this is used in code

The trainer config is passed into ESPnet3's training wrapper, which constructs a lightning.pytorch.Trainer internally.

See:

• espnet3/components/training/trainer.py
• doc/vuepress/src/espnet3/config/train_config.md