Traditional machine learning methods heavily rely on the independent and
identically distribution assumption, which imposes limitations when the test
distribution deviates from the training distribution. To address this crucial
issue, out-of-distribution (OOD) generalization, which aim