Towards a direct Bayesian adaptation framework for deep models