Since the appearance of generative adversarial networks, adversarial learning has been widely used in various branches of machine learning, which drives new developments. In domain adaptation, adversarial domain adaptation methods use a shared feature extractor to extract the domain invariant representation and a discriminator to
judge. Both reach optimal solution through the iterative update of adversarial learning. In terms of data sources, generative adversarial networks and domain adaptation have two similar domains. In the aspect of objective functions, both try to pursue consistency. We attempt to further enhance the domain adaption performance from a deeper level by taking advantage of the mature framework of generative adversarial networks and analyzing intrinsic similarities between the two based on the theoretical level and logical structure. By analogy, the model employs two discriminators to solve the mode collapse problem existing in previous adversarial domain adaptation algorithms, and utilzies pseudo labels for structural improvement. Finally, experiments on standard domain adaptation tasks confirm the feasibility and effectiveness of the method.