Add best_model_monitor to ModelCheckpoint

[carmocca]

Motivation

If I have a directory with many checkpoints, possibly belonging to runs with different monitors, I don't know which metric was tracked by each checkpoint.

Pitch

ckpt = torch.load("saved_model.ckpt")
assert hasattr(ckpt, "checkpoint_callback_best_model_monitor"

Alternatives

Guess If I included the monitor in the checkpoint name, I could parse it to know. But seems much more convenient to add the property.

[rohitgr7]

another alternative: you can also add that to checkpoint itself in on_save_checkpoint.

[carmocca]

Do you mean by overriding it in my LightningModule?

It is an ugly solution because it does not know about any ModelCheckpoints used. Only the trainer knows about them.

So I would need to instantiate my LightningModule with an added checkpoint_callback_monitor parameter in order to do:

def on_save_checkpoint(self, checkpoint):
    checkpoint["checkpoint_callback_best_model_monitor"] = self.checkpoint_callback_monitor

[ananthsub]

@carmocca here

[carmocca]

If you guys agree with this issue, I can work on the PR

[ananthsub]

I think it makes sense to add monitor especially since best_model_score is already included. Lightning should track what that score actually represents 😃

[rohitgr7]

Do you mean by overriding it in my LightningModule?

na.. not the LightningModule but ModelCheckpoint, I know it's ugly haha..
IMO I think it's good to have the monitor as well. But need to check whether it is required to load it back or not in on_load_checkpoint.