auto_insert_model_name behavior when not in filename arg in ModelCheckpoint

[felipegb94]

Hi Everyone,

I am currently using pytorch-lightning=1.6.1 and I am confused by the behavior of ModelCheckpoint auto_insert_model_name flag when filename is not specified.

I have the following checkpoint callback:

	ckpt_callback = pl.callbacks.ModelCheckpoint(
		monitor='val_loss'	
		, auto_insert_metric_name=True
		, save_last=True
		, save_top_k=-1 # Of -1 it saves model at end of epoch

With the above callback, I was expecting that the val_loss value would be added to the checkpoint filename. However, it does not get added.

For the val_loss to get added I need to set filename='{epoch:02d}-{step:02d}-{val_loss:.4f}'. But, if I do this, then there is no need to even set auto_insert_metric_name=True because val_loss will be added to the filename eitherway.

So, I guess my question is, what really is the usage of auto_insert_model_name if it does not automatically update the filename?

Thanks!
Felipe