Revisiting Bayesian Autoencoders with MCMC