Customize the summary of metrics in wandb table

[GillianGrayson]

Hi!

Is there an easy way to customize the summary of metrics in wandb table?

By default, the summary is set to the last value you logged for each key.
But the last value is not always optimal, because there is a non-zero patience parameter in the EarlyStopping callback.
In the result table are the last (not optimal values).
How to deal with it in the simplest way?

Thanks!

P.S.: The best template repository for PyTorch!

[GillianGrayson]

One possible way :

import wandb

...

    def on_fit_start(self) -> None:
        wandb.define_metric(f"train/loss", summary='min')
        wandb.define_metric(f"val/loss", summary='min')
        wandb.define_metric(f"test/loss", summary='min')
        wandb.define_metric(f"train/acc", summary='max')
        wandb.define_metric(f"val/acc", summary='max')
        wandb.define_metric(f"test/acc", summary='max')

[ashleve]

Thanks! Didn't know about that