Python interface for loading

[Adamits]

Previously, I recall that we had methods like get_trainer and get_model to go along with train.get_trainer_from_argparse_args and train.get_model_from_argparse_args.

These seem to have disappeared. What is the rationale behind that? Was this a mistake?

[kylebgorman]

I sort of put them in there anticipating having tests in place not long after, but never got around to finishing the testing, and the interfaces kept changing in the meantime. I think I should complete #91 and we should complete #60 in general before we are ready to freeze this interface.

[Adamits]

Cool! So

before we are ready to freeze this interface.

The interface being the exact argument names? And then we would just add the other method, which would share that interface?

[kylebgorman]

By interface I mean the helper functions to create the pieces you need for training. Maybe something like:

def get_datamodule(...): ...

def get_model(...): ...

def get_trainer(...): ...

You want different flavors of these in train and predict modules.

Alternatively we could have a sort of "god class" that wraps the trainer object.

[Adamits]

Ahhh got. Makes sense.

Alternatively we could have a sort of "god class" that wraps the trainer object.

And this class would have the helpers as method?

[kylebgorman]

Alternatively we could have a sort of "god class" that wraps the trainer object.

And this class would have the helpers as method?

That's what I had in mind. Though I am not sure if that's a good idea vs. the interface I outlined above.

[Adamits]

Yeah it seems clean, but then, I would also want to advocate for exposing the PTL trainer as much as possible (not totally sure whether either suggestion is better than the other to this end).

[kylebgorman]

Arguably we have a reasonably stable API for this as of #110. Do you have particular suggestions for what else is needed?

[kylebgorman]

Going to close this---I think we have a reasonably good interface now. Feel free to reopen with specific requests.