Weights & Biases logging

AutoCast contains built-in support for logging to Weights & Biases (W&B), a popular tool for tracking machine learning experiments. This allows you to monitor your training progress and visualise metrics as your experiment is running (or after it has completed).

To enable W&B logging you need to first log in with the wandb command-line app:

uv run wandb login

Follow the instructions there to log in to your W&B account.

Then, apply the ++logging.wandb.enabled=false override when training a model. You can additionally also supply

• ++logging.wandb.project=<project_name> to specify the W&B project name (defaults to autocast)

• ++logging.wandb.name=<run_name> to specify the run name

By default the run name is inferred from the --run-id command-line option, so you can also set that if you prefer:

uv run autocast ... --run-id=my_id ...

Since --run-id also controls the output directory for AutoCast (see the Output paths page for more information), this allows you to use a single flag to set both the output directory as well as the W&B run name.