Image features from a small local region often give strong evidence in the classification task. However, CNN suffers from paying too much attention only on these local areas, thus ignoring other discriminative regions. This paper deals with this issue by performing the attentive feature cutmix in a progressive manner, among the multi-branch classifier trained on the same task. Specifically, we build the several sequential head branches, with the first global branch fed the original features without any constrains, and other following branches given the attentive cutmix features. The grad-CAM is employed to guide input features of them, so that discriminative region blocks in the current branch are intentionally cut and replaced by those from other images, hence preventing the model from relying on only the small regions and forcing it to gradually focus on large areas. Extensive experiments have been carried out on reID datasets such as the Market1501, DukeMTMC and CUHK03, showing that the proposed algorithm can boost the classification performance significantly.