Fix CUML HDBSCAN predictions by using correct method.

[Bougeant]

This fix is intended to fix the bug found here: #1463

[MaartenGr]

Thanks for the PR! Just to be sure I understand correctly, why is this specific change a solution to the problem? Also, which version of cuML are you using? I believe there were some changes to recent versions that might have fixed this.

[Bougeant]

So I think when func == "membership_vector", we should calculate probabilities using membership_vector(model, embeddings), not using approximate_predict(model, embeddings) which merely calculates the most relevant cluster and the associated probability.

membership_vector returns a 2D of size (n_sample, n_topics), while approximate_predict returns a tuple of two 1D arrays of size (n_sample) each. This explains why probabilities is a tuple, not an array in #1463 (AttributeError: 'tuple' object has no attribute 'shape'). I stumbled upon the same issue and was able to get past it with this fix.

I believe that this was just a typo (you probably meant membership_vector but ended up writing approximate_predict)

[MaartenGr]

@Bougeant I believe this relates to #1324 where approximate_predict was used for earlier versions of cuML compared to newer versions which do use membership_vector. So this was not a typo but merely the functionality that was in v23.03 and lower I believe. If I'm not mistaken, membership_vector was introduced in v23.04.

Perhaps that related PR should be updated and merged instead?

[Bougeant]

Oh right, makes sense. In any case, the current implementation does not work since we expect a 2D array but are getting a 2-tuple. #1324 adds some unit tests as well, so that's probably the way to go.