Weights and Biases Sweeps

Introduction

Weights and Biases (W&B) provides automated hyperparameter search utilities with interactive visualizations. For more details, visit the wandb Sweeps documentation.

Usage

W&B Sweeps require two primary files:

• config.yaml: a standard configuration file for the model.

• sweep.yaml: a file that defines the sweep configuration.

When W&B Sweeps call the config.yaml, they override parameters in the config.yaml in a manner consistent with Hydra’s override syntax. For clarity, it may be convenient to make sweep variables top-level using variable interpolation in the config.yaml.

An example of a sweep configuration file for the tutorial config (configs/tutorial.yaml) is provided in misc/wandb_sweep/sweep.yaml, where paths are relative to the root of the nequip repository. The tutorial.yaml is an example config.yaml that uses variable interpolation for W&B Sweep parameters used in the associated sweep.yaml.

Once a config.yaml and sweep.yaml have been written, a sweep agent can be created:

wandb sweep sweep.yaml

and then the sweep agent can be started:

wandb agent <sweep-ID>