Checkpointing-Abstraction

[LeanderK]

I am thinking about writing an ML-Flow logging hook, but to use it's full potential need to commit the final checkpoint to ML-Flow. Both pytorch and tensorflows checkpointing-class does essentially the same and I think a simple abstraction would allow me to interact with the hook and retrieve the last checkpoint. Would this change be welcome?

[jhaux]

I would totally love it! Should you add the new functionality, could you then also add a small motivation in the documentation, why ML-Flow and pushing checkpoints to it is a good idea?

[pesser]

+1 for checkpoint abstraction!