A Method for Recovering on Unsupervised Domain Adaptation Models Compression

Unsupervised domain adaptation (UDA) aims to transfer learned knowledge from a source domain with labeled data, to a target domain with unlabeled data. For practical uses, UDA models are usually compressed to fit into smaller devices. However, previous work on UDA compression cannot obtain satisfactory accuracy after the compression and tine tuning Our study shows one of the major reason is the growth of domain discrepancy during the tine tuning process. To recover the model accuracy for the target domain data, we propose a two-step tine -tune strategy: the first step is like ordinary tuning process whose goal is to reduce the classification error; the second step is designed to suppress the growth of domain discrepancy. We propose a sampling technique to estimate domain discrepancy of the pruned model, and use it in the loss function of the second step. We validated the performance of our method on ImageCLEF-DA, Office -31 and Office -home datasets. The results show that our method reaches higher average accuracy than previous work, especially for unbalanced dataset, such as Office 31.