What is the monitor of lr_scheduler?

[Co1lin]

Discussed in #8541

^{Originally posted by Co1lin July 24, 2021}
Hi! I’m wondering what is metric_to_track for the parameter monitor of a learning rate scheduler? I see the example that uses "val_loss" as its value, but I don’t know is val_loss a name of a variable (where should I declare this variable?) or a keyword that can be recognized by pytorch_lightning such that it will monitor the loss of the validation_step?

I get a warning saying:

packages/pytorch_lightning/callbacks/model_checkpoint.py:610: LightningDeprecationWarning: Relying on `self.log('val_loss', ...)` to set the ModelCheckpoint monitor is deprecated in v1.2 and will be removed in v1.4. Please, create your own `mc = ModelCheckpoint(monitor='your_monitor')` and use it as `Trainer(callbacks=[mc])`.

I'm not sure whether this usage will be deprecated for scheduler's monitor too.

[ethanwharris]

Hi @Co1lin, you can configure the monitor for LR scheduling in the configure_optimizers hook of your LightningModule as detailed here: https://pytorch-lightning.readthedocs.io/en/stable/common/optimizers.html?highlight=lr%20scheduler#learning-rate-scheduling

metric_to_track is the name of the logged metric to use as the monitor as a string (similarly for "val_loss"). So when you then call self.log("metric_to_track", my_metric_value) that is the metric that will be used. The deprecation for ModelCheckpoint relates to the fact that we used to just take whichever value was logged as the val_loss as the quantity to monitor. So the usage for LR scheduling hasn't changed, we just give a warning for users who rely on the behaviour for checkpointing.

Hope that helps 😃