ESPnet3 Train Dataset

Masao SomekiAbout 1 min

ESPnet3 Train Dataset

This page focuses on how training configs describe datasets today.

The full dataset resolution and builder story is documented in:

Dataset references and builders

Current pattern

Training configs use DataOrganizer plus dataset reference entries.

dataset:
  _target_: espnet3.components.data.data_organizer.DataOrganizer
  recipe_dir: ${recipe_dir}

  train:
    - data_src: mini_an4/asr
      data_src_args:
        split: train

  valid:
    - data_src: mini_an4/asr
      data_src_args:
        split: valid

  test:
    - name: test
      data_src: mini_an4/asr
      data_src_args:
        split: test

Each item can specify the dataset in three ways.

1. Omit `data_src` and use the local recipe dataset

If data_src is omitted, ESPnet3 loads:

${recipe_dir}/dataset/__init__.py

This is what mini_an4 and librispeech_100 do in their training configs.

train:
  - data_src_args:
      split: train
valid:
  - data_src_args:
      split: valid

2. Use a dataset tag

train:
  - data_src: mini_an4/asr
    data_src_args:
      split: train

This resolves to:

egs3.mini_an4.asr.dataset

3. Use an explicit module path

train:
  - data_src: egs3.mini_an4.asr.dataset
    data_src_args:
      split: train

What is forwarded to the dataset constructor

Only data_src_args is passed to the exported Dataset class:

Dataset(**data_src_args)

So a config like:

- name: train-clean
  data_src: librispeech_100/asr
  data_src_args:
    split: train-clean-100
    recipe_dir: ${recipe_dir}

becomes:

Dataset(split="train-clean-100", recipe_dir=recipe_dir)

name and transform are handled by DataOrganizer, not by Dataset.

Current recipe examples

`mini_an4`

mini_an4 exports:

dataset/__init__.py
dataset/dataset.py
dataset/builder.py

Its local dataset mode is:

dataset:
  recipe_dir: ${recipe_dir}
  train:
    - data_src: mini_an4/asr
      data_src_args:
        split: train
  valid:
    - data_src: mini_an4/asr
      data_src_args:
        split: valid

`librispeech_100`

librispeech_100 uses the same local-mode pattern, but the dataset reads the raw LibriSpeech directory directly instead of manifest TSVs.

Train and valid requirements

Current DataOrganizer requires:

both train and valid, or
neither

If training is the goal, define both.

Test entries

test entries are optional for the train stage itself, but adding them in the same config is useful because:

inference can reuse the same dataset definition
measurement can reuse test-set names

Each test entry should define name, because that becomes the test-set name used in inference_dir/<test_name>/.