GitHub - google-research/tuning_playbook: A playbook for systematically maximizing the performance of deep learning models.

Abstraction: Systematic hyperparameter tuning playbook from Google Brain researchers

Key points:

• Authored by five Google Research / Brain team researchers (plus Harvard); frames tuning as a scientific process of incrementally building understanding rather than greedy validation-error chasing
• Classifies hyperparameters into three roles per experiment: scientific (what we're measuring), nuisance (must be optimized away for fair comparison), and fixed (held constant to save budget, introduces caveats)
• Recommends quasi-random search during the exploration phase for unbiased coverage of the search space; switches to Bayesian optimization (e.g., Open-Source Vizier) only after search spaces are refined
• Batch size guidance: use the largest hardware-supported batch size; it governs training speed not final performance — batch size is not a tunable hyperparameter for validation quality
• Training curve inspection checklist: look for problematic overfitting (validation error rising), high step-to-step variance, and whether the model is still improving at run end (compute-bound vs. not)
• Hyperparameter transfer across training lengths: warmup and initialization transfer reliably; learning rate schedules transfer poorly — schedule must be re-tuned when training steps change significantly

Connections: Google Research · Hyperparameter Tuning · Deep Learning · Model Training

Source: https://github.com/google-research/tuning_playbook