Improving the Reconstruction of Disentangled Representation Learners via Multi-Stage Modelling