ModelCheckpoint instance to JSON

[mpariente]

🚀 Feature

Have a to_json(path) function that writes the best_k_models to a json file in path.

Motivation

Why having a JSON is good?
Then you don't need to load the checkpoints to know the metrics and things like a posteriori stochastic weight averaging gets easier.

It used to be easier in 0.6, but loss values are now tensors, so it's not directly serializable.
This is currently needed:

best_k = {k: v.item() for k, v in checkpoint.best_k_models.items()}
with open(json_path, "w") as f:
    json.dump(best_k, f, indent=0)

Instead, we could have

checkpoint.to_json(json_path)

I'm ready to make a PR for it. Just tell me if this would be useful for you.

[awaelchli]

should this file be saved with every checkpoint, or one file that gets updated/overwritten?

[mpariente]

I just thought about a single function that can be called by the user at the end of training.
But actually, updating/overwritting at every checkpoint also would be cool, at least the user never has to call it.
Link to the PR #3048