Initialization of parameters of Bayesian Neural Network

What is the best practice to initialize parameters of the Bayesian Neural Network? I am stuck in a situation where I don’t initialize the parameters and every time I run the variational inference, I get different training errors (variaies between 2-7%) and parameters. I am looking for stable results, basically the same result every time I run the complete variational inference. I was hoping to at least get the same training error when I run the model. Any help is greatly appreciated.