ESPnet3 Train Stage

Masao SomekiAbout 1 min

ESPnet3 Train Stage

The train stage runs model training using:

espnet3.systems.base.training.train
espnet3.components.modeling.lightning_module.ESPnetLightningModule
espnet3.components.trainers.trainer.ESPnet3LightningTrainer

Quick usage

Run

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

Configure (in `training.yaml`)

Keep the core settings in training.yaml.

Section	Description
`task`	task entrypoint for ESPnet2-style models
`model`	model definition and normalization settings
`dataset`	train and valid splits
`dataloader`	collate and iterator settings
`trainer`	Lightning trainer configuration
`optimizer`, `scheduler`	single-optimizer training path
`optimizers`, `schedulers`	named multi-optimizer path
`exp_dir`	training output directory
`stats_dir`	stats output directory used by `collect_stats`

Main config sections

Training uses training.yaml.

Typical sections are:

task or model
dataset
dataloader
optimizer / scheduler or optimizers / schedulers
trainer
exp_dir
stats_dir

Outputs

Training writes under exp_dir, including:

checkpoints
logs
TensorBoard output if configured

collect_stats writes under stats_dir.

Typical outputs are written under:

exp_dir: checkpoints, logs, TensorBoard files, saved configs
stats_dir: feature shapes and normalization stats from collect_stats

Key ideas

Dataset

Current dataset definitions are based on DataOrganizer plus dataset reference entries using data_src and data_src_args.

Example:

dataset:
  _target_: espnet3.components.data.data_organizer.DataOrganizer
  recipe_dir: ${recipe_dir}
  train:
    - name: train
      data_src: mini_an4/asr
      data_src_args:
        split: train
        data_path: ${dataset_dir}
  valid:
    - name: valid
      data_src: mini_an4/asr
      data_src_args:
        split: valid
        data_path: ${dataset_dir}

Details:

Dataloader

Topic	Summary	Details
`dataloader`	supports ESPnet iterator mode and plain PyTorch `DataLoader` mode	Dataloader + Collate
`trainer`	uses the ESPnet3 Lightning trainer wrapper, not raw Lightning directly	Trainer
`optimizer`, `scheduler`, `optimizers`, `schedulers`	supports both single-optimizer and named multi-optimizer training	Optimizer + Scheduler, Multiple optimizers and schedulers
`model`	supports both task-backed ESPnet2 model construction and direct Hydra instantiation	Model
`callbacks`	handles logging, checkpointing, and metric reporting such as `MetricsLogger`	Callbacks

Details by topic

The train stage is intentionally thin: most customization happens in one of those component configs rather than in a stage-specific CLI path.