New ModelCheckpoint Code Update Breaking Model Saving Functionality : v1.2.5->v.1.2.6

[valayDave]

I am using Latest PT-Lightning and your upgrade to ModelCheckpoint broke my functionality to my codebase.

V1.2.6 specifically coz I started facing these issues yesterday. Downgraded to V1.2.5. Please check once. The dict based metrics returned from the train_step or validation_step are not captured in the ModelCheckpointCallback

 monitor_candidates = self._monitor_candidates(trainer)

monitor_candidates doesn't have any values returned from the trainer. And I am not using the Result class. I am returning a pure dictionary.

Didn't have time to create replicate the bug exactly in a notebook because I am on a tight deadline. But please do check.

As I have more time, I will update more on the issue.

[awaelchli]

You need to use self.log("your_metric", metric)
And then set the monitor argument in ModelCheckpoint(monitor="your_metric")

Let me know if that helps.

[carmocca]

Exactly. Setting monitor candidates through dictionary returns is legacy behaviour and no longer supported.