bayesian neural networks are an active area of research. generally speaking, variational approaches don’t work very well/reliably. this is especially true if the neural network has a large number of parameters (like a convnet). and things get even worse if you do things more or less naively (as would be the case if you use pyro.random_module). so i’d probably advise against trying to build bayesian neural networks in pyro (or in general) unless you’re an expert on the many difficult aspects of the problem.