Lifted @abstractmethod on instance segmentation model superclass; implementations were actually calling infer twice so obvious performance hit