Source Free Domain Adaptation by Deep Embedding Clustering

Abstract

Domain adaptation is devised to enhance model performance on target domain when distribution shift occurred. The current domain adaptation methods tend to cause significantly transmission consumption in practical applications because they require both source and target data in training, especially for decentralized training. This paper explores a different setting, source data free domain adaptation, only target data and source model are accessible during knowledge transfer. This work proposes a simple solution deep embedding clustering domain adaptation (DECDA) to this task, which use source model to generate pseudo labels for target data and clusters target data towards the cluster centers of high-confidence target samples in feature space by deep embedding clustering algorithm. In this way, the target data can be classified by source classifier in feature space. Experiments have been evaluated in Office-31 and VisDA2017 dataset and achieved comparable results to the best source data required domain adaptation methods recently.