Monitored quantity in ModelCheckpoint does not exist

#1

Hi,

I’m using PL V0.9, and I’m using pl.callbacks.ModelCheckpoint to monitor a quantity named val_loss, to save only the best models.

However, after updating my code base, there is no quantity named val_loss in my validation loop, because I changed it’s name, but my callback still save checkpoints!

I want to know, in this case where the monitored quantity does not exist, which quantity is used by the callback to make the decision of saving?

Thank you in advance!

#2

You did not include how you are creating the ModelCheckpoint, so assuming it is set to monitor=None, it will try to save every epoch.

You can set the monitor argument to your quantity name in order to track it instead.

#3

Hi @carmocca
Thank you for your response.
Actually, I create the ModelCheckpoint in the following way:

checkpoint_callback = pl.callbacks.ModelCheckpoint(
        monitor="val_loss",
        mode="min",
        save_last=True,
        save_top_k=5,
        verbose=False,
    )

However, there is no metric called val_loss, but ModelCheckpoint still save the models, and only 5 at the time + the last one.

#4

5 at the time + last is correct considering you have save_top_k=5, save_last=True.

There is this check to stop the run if the monitor is not found:


So you must be actually monitoring something.

Can you share your training and validation step?