Inconsistency in KMeans clustering result

[gobear6212]

Hi, I was trying to retrain PGP but I run into an issue with scikit-learn's KMeans implementation. Sometimes when the model tries to compute the Ward distances, it throws a broadcast exception for dists = wts * centroid_dists + np.diag(np.inf * np.ones(len(cluster_counts))) because the shapes of wts and centroid_dists are different.

The root cause seems to be that cluster_lbls and cluster_ctrs are inconsistent, so performing np.unique() for the cluster labels returns the wrong cluster_cnts. In scikit-learn's documentation, I notice the following

cluster_centers_ndarray of shape (n_clusters, n_features) Coordinates of cluster centers. If the algorithm stops before fully converging (see tol and max_iter), these will not be consistent with labels_.

May I ask how should I handle this exception?

[nachiket92]

It looks like K-means returned an empty cluster. This is very strange and has not happened during any of my training runs. Can you consistently reproduce the error? Were any model parameters changed?

[gobear6212]

I didn't change the model parameters, but I tried to introduce additional edges (e.g. on the left/right of the lane instead of only the proximal ones). This exception only occurs once/twice as far as I recall, so I can't reproduce it. But I suspected that it's related to bad initialization of the clusters, so I removed init='random' from KMeans and let it uses the default k-means++ strategy, which seems to work for now. However, I'm not sure if the same exception will occur again.

[qihuihu20]

Have you ever meet the same error after that? I removed init='random' as you said but not effect, May I ask how should I handle this exception? @gobear6212