Abstract: Existing domain adaptation (DA) methods try to handle various DA scenarios subject to imbalanced label sets or multiple source/target domains, e.g., Closed set, Open set, Multi-Source, Partial and Multi-Target DA. Though Universal Domain Adaptation (UniDA) and Versatile Domain Adaptation (VDA) have been proposed to address these scenarios simultaneously, the related proposed methods still suffer from two issues: i) UniDA and VDA can hardly cover all of existing DA scenarios, e.g., UniDA cannot handle Multi-Source and Multi-Target DA scenarios, and VDA does not include Open set DA. ii) The proposed UniDA and VDA methods mainly focus on the versatility, and they ignore the essential DA problem, i.e., domain mismatch. This paper introduces Grand Unified Domain Adaptation (GUDA) scenario, which needs no prior knowledge about the number of source/target domains or the overlap. GUDA can cover more existing DA scenarios. Towards tackling GUDA, we formulate a grand unified adaptation network called Graph Contrastive Adaptation Network (GCAN), which can handle above mentioned DA scenarios and further reduces the domain mismatch without any modification. GCAN includes a graph contrastive adaptation objective at the node level, and a transferability rule to gain the common category identification loss. The results illustrate that GCAN works stably on different GUDA settings and shows comparable performance against recent DA methods on five benchmarks.