| name | tide-overfit-protocol |
| description | Runs tide single-song overfit smoke tests across conv1d, mel, and MERT frontends via WSL GPU. Use when the user mentions tide overfit, overfit suite, frontend ablation, smoke test on tide, or comparing preprocessing frontends. |
Tide overfit protocol
When to use
Single-song memorization check on tide before scaling to multi-song val. Default 50 epochs; 100 ep only when explicitly testing epoch sensitivity.
Preconditions
- Tide audio/chart paths in
configs/overfit_tide/*.json
- MERT:
data/v2/test/tide.mert.npy (or path in config mert_features_dir)
- Do not set
pipeline_check_shortcuts=true unless explicitly debugging pipeline wiring
Run (WSL GPU)
Use the WSL command template with:
scripts/run_overfit_tide_suite.py --epochs=50
Subset frontends: --frontends=conv1d,mel
Scripts auto-dispatch from Windows when WSL is available.
Configs and artifacts
| Item | Path |
|---|
| Configs | configs/overfit_tide/{conv1d,mel,mert}.json |
| Models | models_wsl/overfit_tide/<frontend>/ |
| Summary | models_wsl/overfit_tide/suite_summary.json |
| Callbacks | callbacks/overfit_tide/<frontend>/ |
After the run
- Read
suite_summary.json for per-frontend F1 / precision / recall
- Log
EXP-YYYYMMDD-NN in EXPERIMENT_LOG.md with full timestamp (index + entry)
- If F1=0 or conv1d collapse → follow onset-event-eval-matching before retraining
- If plateau persists → tide-ablations or formulation change per EXPERIMENT_LOG.md § Recommended next step
Related
