Remove abc

[metodj]

• removed all ABC stuff (abstract methods etc.)
• refactored baselaplace.py -> BaseLaplace now has two subclasses, ParametricLaplace and FunctionalLaplace (this will greatly help with bringing in the GP inference)

[aleximmer]

It looks good to me and since tests pass I don’t see a problem. However, I would suggest to make the raised errors a bit more verbose. For example, if an interface does not implement gradients, this could be explicitly stated as the cause of the problem with a message like “class XY does not provide gradients.” Or something similar. What do you think?

[metodj]

Sure, will make raised NotImplementedError more verbose. Another thing I just remembered is that the documentation will also have to be updated, will add this to the next commit as well. The change in the documentation will be minor, though (mostly renaming of the classes).

As an aside, the proposed refactor is backward-compatible in a sense that ParametricLaplace supports everything that BaseLaplace supported before. What is different now is that BaseLaplace now is more generic so that it will be easier to bring in FunctionalLaplace.

[metodj]

I created a dummy class in curvature.py in order to test how are the raised NotImplementedError displayed:

class TestInterface(CurvatureInterface):

    def __init__(self, model, likelihood, last_layer=False):
        super().__init__(model, likelihood, last_layer)

When using simply

    def gradients(self, x, y):
        raise NotImplementedError

the following is raised,

while using

    def gradients(self, x, y):
        raise NotImplementedError(f"{self.__class__.__name__} does not provide gradients.")

yields

If you prefer the second option, I can change all raise NotImplentedErrors accordingly. In my opinion, though, this might not be necessary as the error thrown is quite informative as it is and also by making it more verbose we would be slightly deviating from the Python convention.

[metodj]

Also, in the laplace.py I changed it so that only subclasses of ParametricLaplace are exposed through Laplace for the time being. Once FunctionalLaplace is ready, we can switch back to BaseLaplace. Let me know if that's fine.

[aleximmer]

Okay, we can leave the plain NotImplementedErrors then. That's actually verbose enough. I think you only need to recompile the documentation (see the readme for this) and then push the changed documents. Let me know if that doesn't work for you.
@runame Are you okay with merging this in?

[runame]

Sorry for the late reply, I was pretty busy this week. Looks good from my side, nice work!

One nitpick: the string style is inconsistent in some places, i.e. " " instead of our convention ' '. I'll comment on some lines where I have seen it (haven't checked everything carefully though).

[metodj]

Thanks for pointing out the convention for the string style, wasn't aware of it before

[runame]

Nice, thanks!

[aleximmer]

Thanks both!