A multi-scale loss formulation for learning a probabilistic model with proper score optimisation