search

AmazaspShumik / sklearn-bayes

Python package for Bayesian Machine Learning with scikit-learn API
MIT License
513 stars 118 forks source link

VBGMMARD vs VBGMM in scikit-learn #11

Closed BayesianHuman closed 8 years ago

BayesianHuman commented 8 years ago

Is there any between VBGMMARD class in sklearn_bayes and VBGMM in sklearn ?

AmazaspShumik commented 8 years ago

In scikit-learn implementation it is assumed that weight vector is random variable, while in my implementation it is assumed to be parameter. At each iteration of algorithm some weights are driven closer and closer to zero, and after they become close enough components corresponding to those weights are explicitly removed from the model (note in scikit-learn there is no removal of components). So VBGMMARD automatically selects number of components through relevance determination technique described above (the same idea is used in Relevance Vector Machine ( Tipping 2001) ) while VBGMM in sklearn does not do that.

AmazaspShumik commented 8 years ago

If you want to automatically select number of components sklearn has DPGMM class. DPGMM uses Bayesian nonparametric model to infer number of components