Remove entries from the progress bar later in training.

[ChristophReich1996]

Hi everyone,

I have the following "issue" I have multiple training phases with different loss functions. However, I would only like to show the current loss values in my progress bar. Is there any elegant way to discard old loss entries (eg. loss_bootstrap_1) from the progress bar?

Thanks for the help! :)

Cheers
Christoph

[akihironitta]

Hi @ChristophReich1996! Not sure if it's the best (or if this is what you're looking for), but turning it off with

self.log("loss_bootstrap_1", random_value, prog_bar=False)

once at a point your training phase changes might probably work.

[ChristophReich1996]

Hi @akihironitta, thanks for your help, unfortunately, if I set prog_bar to False, after the respective training phase, Lightning throws a MisconfigurationException...

pytorch_lightning.utilities.exceptions.MisconfigurationException: You called `self.log(loss_bs_2, ...)` twice in `training_step` with different arguments. This is not allowed

I'm using pytorch-lightning==1.6.4.

[akihironitta]

@ChristophReich1996 Ah right... I don't think it's feasible with the current progress bar. A simple workaround is to give up using a progress bar 😆 and use one of the loggers supported by PL and see metrics of your interest in the browser.