google-research/tuning_playbook

The Deep Learning Tuning Playbook is a written guide from researchers at Google Brain on how to systematically get better results when training deep learning models. Deep learning models (the kind of AI systems used for image recognition, language processing, and similar tasks) have many settings called hyperparameters, things like learning rate (how aggressively the model updates itself during training), batch size (how many examples it processes at once), and the number of training steps. Getting these settings right is often the difference between a model that works well and one that does not, but the process has historically been more art than science. The playbook documents a principled approach to this tuning process. It covers how to choose an initial configuration when starting a new project, how to run experiments in a way that actually tells you something useful (rather than just randomly trying different settings), how to decide which changes to keep, and how to determine when to stop training. It also addresses common failure modes and gives specific guidance on topics like batch normalization (a technique for stabilizing training) and setting up experiment tracking. You would use this if you are an engineer or researcher training deep learning models and want a structured method for improving their performance. It is written documentation rather than runnable code, designed as a reference to consult throughout a project. No specific programming language or framework is required to follow it.