BIRCH fails with "'_CFSubcluster' object has no attribute 'centroid_'" #6172
Description
This issue occurs to me during an evaluation with growing feature set sizes. In under a second after the call to fit_predict on the BIRCH instance.
Strangely it works for smaller sets of features and does not run out of memory.
For my evaluation I employ 183719 datasets and construct feature sets of growing size. It works with 6 features, but returns this error with 12. I use the version 1.7.
Here is the full stack trace:
Traceback (most recent call last):
File "clusterer.py", line 27, in score
labels = self.clusterer.fit_predict(x)
File "/usr/lib/python2.7/dist-packages/sklearn/base.py", line 371, in fit_predict
self.fit(X)
File "/usr/lib/python2.7/dist-packages/sklearn/cluster/birch.py", line 424, in fit
return self._fit(X)
File "/usr/lib/python2.7/dist-packages/sklearn/cluster/birch.py", line 458, in _fit
split = self.root_.insert_cf_subcluster(subcluster)
File "/usr/lib/python2.7/dist-packages/sklearn/cluster/birch.py", line 193, in insert_cf_subcluster
subcluster)
File "/usr/lib/python2.7/dist-packages/sklearn/cluster/birch.py", line 212, in insert_cf_subcluster
closest_subcluster, new_subcluster1, new_subcluster2)
File "/usr/lib/python2.7/dist-packages/sklearn/cluster/birch.py", line 170, in update_split_subclusters
self.init_centroids_[ind] = new_subcluster1.centroid_
AttributeError: '_CFSubcluster' object has no attribute 'centroid_'
Sadly I'm not able to reproduce this with an minimal example