ESPnet3 Experiment Naming Examples

Below are some examples of how experiment names are determined based on the YAML configs.

Note

${self_name:} is a resolver: at config-load time it is rewritten to the stem of the config file currently being loaded (e.g. training.yaml → training, training_e_branchformer.yaml → training_e_branchformer). It is not tied to exp_tag — it is just the filename. Source: espnet3/utils/config_utils.py.

---

Example 1: training-driven naming

The shipped TEMPLATE/asr/conf/training.yaml uses the resolver by default:

exp_tag: ${self_name:}
exp_dir: ${recipe_dir}/exp/${exp_tag}

So running --training_config conf/training.yaml resolves exp_tag to training and writes outputs under:

exp/training/

If the same run.py invocation also passes --inference_config, apply_training_experiment_context() copies exp_tag and exp_dir from training_config into inference_config (overwriting any conflicting value there, with a warning). TEMPLATE/asr/conf/inference.yaml then resolves:

inference_dir: ${exp_dir}/${self_name:}

Here ${self_name:} is the stem of inference.yaml itself (inference), so the result is:

exp/training/inference

---

Example 2: standalone inference naming

If inference runs on its own (no --training_config in the same invocation), inference.yaml must carry its own identity — its shipped default exp_tag: is blank:

exp_tag: whisper_eval
exp_dir: ${recipe_dir}/exp/${exp_tag}
inference_dir: ${exp_dir}/${self_name:}

That produces:

exp/whisper_eval/inference

---

Example 3: naming a tuning variant

A common convention is one training config per variant under conf/tuning/, each relying on the ${self_name:} default instead of hand-typing a tag:

conf/
  tuning/
    training_e_branchformer.yaml   # exp_tag: ${self_name:}  (inherited from training.yaml)
  inference_beam5.yaml             # exp_tag: ${self_name:}

Because ${self_name:} resolves independently for each file, running each config on its own produces:

exp/training_e_branchformer/
exp/inference_beam5/inference_beam5/

(inference_beam5/inference_beam5 because inference_dir: ${exp_dir}/${self_name:} resolves self_name a second time, against the inference_beam5.yaml filename.)

If inference_beam5.yaml is instead run together with training_e_branchformer.yaml via --training_config/--inference_config in the same run.py call, the training config's exp_tag/exp_dir win (see Example 1), so the decoding outputs move under the training experiment directory: exp/training_e_branchformer/inference_beam5/.