Data-driven methods, especially those based on deep learning, have been successfully applied to various tasks. However, training and testing data may have different distributions in many real cases, ...