Add improved goodness of fit implementation

[stes]

This adds a better goodness of fit measure. Instead of the old variant which simply matched the InfoNCE and depends on the batch size, the proposed measure

• is at 0 for "chance level" (vs. log batch size)
• does not need an adjustment for single session vs. multi-session solvers
• increases as the model gets better, which might be more intuitive

The conversion is quite simply done via

GoF(model) = log (batch_size_per_session * num_sessions) - InfoNCE(model)

This measure is also used in DeWolf et al., 2024, Eq. (43)

Application example (GoF improves from 0 to a larger value during training):

---

OLD example, for comparison

---

Close https://github.com/AdaptiveMotorControlLab/CEBRA-dev/pull/669

[stes]

TODO: Fix case where batch_size is None

[CeliaBenquet]

@stes about what I implemented in #202 that I do see here.

I think it would be good to have a really basic function where you provide the loss and the batch size, so that it is easily usable in the pytorch implementation as well.

Also, it would be nice to test for the default CEBRA.batch_size = None, not sure it is handled here.

[stes]

Unrelated build issue due to upstream change in sklearn (#204 ), attempted fix in #205

[stes]

The build issue is fixed, and once #205 is merged tests should pass here as well.

[CeliaBenquet]

Thank you @stes! This looks nice!! Some minor suggestions on the docstrings and maybe add some tests for the different corner cases based on the arguments provided in infonce_to_goodness_of_fit.

[stes]

one review comment to resolve (= add more relevant tests for one batchsize None), then ready!

[MMathisLab]

lgtm! thanks