Identifiability of deep generative models without auxiliary information