ESPnet3 Publish-related Stages

ESPnet3 provides publish-related stages to package model artifacts and optionally upload them:

• pack_model: build a model package directory (README/meta + artifacts)
• upload_model: upload the packed directory (e.g., to Hugging Face Hub)

Quick usage

Run

python run.py \
  --stages pack_model upload_model \
  --train_config conf/train.yaml \
  --publish_config conf/publish.yaml

Configure (in publish.yaml)

Keep the core settings in publish.yaml. For the full list, see Publish configuration.

Config section	Description
pack_model	Packaging strategy, output directory, and files to include.
upload_model	Hugging Face repo settings for uploading the package.

Outputs

Typical output is a package directory (default: <exp_dir>/model_pack) containing a README.md, meta.yaml, and copied artifacts.

Developer Notes

pack_model details

pack_model gathers files from the experiment directory and builds a package directory (default: <exp_dir>/model_pack). It also generates metadata and an optional README.

Key behaviors:

Key	Meaning
strategy	auto selects espnet2 when train_config.task is set, otherwise espnet3.
out_dir	Output directory for the package.
infer_dir	Where scores.json (or measures.json) is searched for README metrics (if present). If unset, it falls back to infer_config.infer_dir.
files / yaml_files	Used to generate meta.yaml for espnet2-style from_pretrained.
include / exclude	Extra file globs to include or skip.
readme_template	Optional template for README generation.

Example output tree:

<pack_dir>/
├── README.md
├── meta.yaml
└── exp/
    └── ...

upload_model details

upload_model uploads the packed directory to Hugging Face.

Key behaviors:

Key	Meaning
upload_model.hf_repo	Required (e.g., yourname/your-model-repo).
huggingface-cli upload	Used under the hood; make sure you are logged in or set HF_TOKEN.
out_dir	The directory uploaded as a model repo.