Efficient Learning of Augmentation Policy Schedules