Pytorch-lightning: Early stopping + checkpoint key

[williamFalcon]

Consider updating how we condition early stopping or checkpoint

return {'early_stop_on': mse_loss, 'checkpoint_on': other_metric}

Instead of:

# only if val_loss is present
return {'val_loss': val_loss}

[jeremyjordan]

in my opinion all of the configuration for early stopping or model checkpointing should occur in the initialization of the callback object. if we condition based on a special key in the step return, users have to change their model to modify the behavior of a callback.

[williamFalcon]

as a shortcut, can the trainer flags already support the key?

early_stop_callback=‘val_loss’

for example

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.