Efficient Model Compression for Bayesian Neural Networks