Learning Policies for Contextual Submodular Prediction