pip package not up to date? about spherecluster HOT 8 OPEN

I changed from sklearn.cluster import _k_means to from sklearn.cluster import _k_means_fast as _k_means but got another exception

conda/pip install scikit-learn==0.20.0 may be a better way if the pip package is not up to date

check this issue for more details #26

from spherecluster.

thank you, it fix my bugs

from spherecluster.

Yeah, similar issue here. The package sklearn.externals.joblib was removed from sklearn in version 0.23 so I have to downgrade to be able to use spherecluster again.

from spherecluster.

@jasonlaska @dominikstrb @cv3d Thanks again for providing this package, but are you interested in additional help in maintaining this project? If additional people could be added to the repo and the pypi account other members of the community can help fix some of these minor incompatibilities

from spherecluster.

Hi @AndrewAnnex I only fixed some imports a couple of months ago, as far as I can remember. If there is anything I can do to help, I am happy to!

from spherecluster.

hey @dominikstrb, I think the most pressing thing would be to fix any remaining issues spherecluster running against newer releases of sklearn and cut a new release for pypi to fix existing issues. I am happy to contribute a conda-forge recipe for spherecluster once that is setup.

from spherecluster.

The biggest problem is that the implementation of k-means was changed in sklearn 0.23.2: scikit-learn/scikit-learn#11950

The internals of the implementation are now very different and it does seem like more than a minor fix in the spherecluster codebase is necessary to accomodate these changes, since spherecluster does not only use the public sklearn api, but the internals of the k-means implementation including functions like _centers_dense, which do not exist anymore since the commit I referenced.

I'm not familiar enough with the current sklearn k-means implementation, but maybe @jeremiedbb, who did that overhaul of k-means can help us.

from spherecluster.

I'm not sure I can help you. Basically we can't guarantee that the private api won't change in the future. I'd advise you to only rely on the public api or reimplement all the private parts of the api you need.

In your case it looks like a huge amount of stuff to reimplement while you're just adding a single line in kmeans_single_lloyd.

As a temporary fix you can pin the scikit-learn version but it's not a good solution in the longterm.

Sorry I can't be more helpful

from spherecluster.