A Three-Stage Model Fusion Method for Out-of-Distribution Generalization

Jiahao Wang, Hao Wang, Zhuojun Dong, Hua Yang, Yuting Yang, Qianyue Bao, Fang Liu, Licheng Jiao

Abstract: Training a model from scratch in a data-deficient environment is a challenging task. In this challenge, multiple differentiated backbones are used to train, and a number of tricks are used to assist in model training, such as initializing weights, mixup, and cutmix. Finally, we propose a three-stage model fusion to improve our accuracy. Our final accuracy of Top-1 on the public test set is 84.62421%.