BUG: NeptuneCallback fails for losses other than BaseLoss

[slawekslex]

Describe the bug

The line here:

neptune-fastai/neptune_fastai/impl/__init__.py

Line 214 in b8e0f3c

return repr(self.loss_func.func)

assumes that Learner.loss_func is an object with a func attribute. However loss_func can be any callable, like torch.nn.MSELoss or just a plain python function.

This causes is to crash with AttributeError

Reproduction

Create a Learner by passing loss_func=nn.MSELoss() or any other custom function. Together with NeptuneCallback

Expected behavior

Model trains without crashing

[Blaizzy]

Hi @slawekslex,

Prince Canuma here, from Neptune.ai,

Thank you for submitting this feedback, I have managed to replicate it with your description and submit a couple of possible fixes to the engineering team, I will let you know once it is fixed.

For now, if this is blocking you. You can edit your local ./neptune_fastai/impl/init.py file line 214 to:

@property
def _optimizer_criterion(self) -> str:
    return repr(self.loss_func)

Let me know this helps you!

[slawekslex]

Thanks for looking into it @Blaizzy. I was able to work around it by setting a dummy attribute on my loss function.
The problem was more in figuring out what's wrong and the general experience of trying to use a library and needing to deal with error after error.

[Blaizzy]

I understand, but thanks to your feedback we can improve and fix those errors.

Are there any other errors you are facing besides this one(#22) and vocab attr(#21) ?? I would be glad to hear all and dig into them so I provide feedback to engineering.

[slawekslex]

I'm going to open issues here if I run into anything else. Thanks again for the responsiveness.

[Blaizzy]

Most welcome! 👍

Have a great weekend 💯