Domain adaptive person re-identification (re-ID) aims to re-identify persons across domains with distinct distributions. The key to this task lies in how to effectively mitigate the domain gap between source and target domain. We observe that the attention of a network, which is crucial for identifying a person, may shift from source to target. Previous works don’t explicitly model and mitigate the shift of attention mechanism, largely constraining the re-ID performance. To address this issue, we propose to align the attention mechanism across domains to reduce the domain gap and facilitate the person re-ID. Specifically, we assume that the discriminative parts of a person should be consistent across domains with different styles. We firstly adopt CycleGAN to acquire paired images with different domain styles. Then we minimize the distance of attention maps across domains to rectify the attention shift. Extensive experiments demonstrate that our method performs favorably against previous state-of-the-arts.