image-clustering | means clustering of media using CNN embeddings | Machine Learning library

code based on this article works as expected when I define only 3 clusters. But when I change the number of clusters, I do not get the equal number of clusters back.

I just finished the Bayesian Analysis in Python book by Osvaldo Martin (great book to understand bayesian concepts and some fancy numpy indexing).

I really want to extend my understanding to bayesian mixture models for unsupervised clustering of samples. All of my google searches have led me to Austin Rochford's tutorial which is really informative. I understand what is happening but I am unclear in how this can be adapted to clustering (especially using multiple attributes for the cluster assignments but that is a different topic).

I understand how to assign the priors for the Dirichlet distribution but I can't figure out how to get the clusters in PyMC3. It looks like the majority of the mus converge to the centroids (i.e. the means of the distributions I sampled from) but they are still separate components. I thought about making a cutoff for the weights (w in the model) but that doesn't seem to work the way I imagined since multiple components have slightly different mean parameters mus that are converging.

How can I extract the clusters (centroids) from this PyMC3 model? I gave it a maximum of 15 components that I want to converge to 3. The mus seem to be at the right location but the weights are messed up b/c they are being distributed between the other clusters so I can't use a weight threshold (unless I merge them but I don't think that's the way it is normally done).