Domain Gap Estimation for Source Free Unsupervised Domain Adaptation with Many Classifiers
Proposed to use many classifiers as a strong domain discriminator for source free unsupervised domain adaptation. Not only the optimal number of classifiers is provided, but also the trade-off between domain gap estimation and limited source domain training data is rigorously discussed.
Theoretically answered why using many classifiers can measure the domain gap more accurately and stably than bi-classifier in source free setting.
Proposed a framework of source free unsupervised domain adaptation via many classifiers (DAMC). By exploiting many classifiers in source domain training, the pre-trained source model approaches the tightest domain gap which makes the downstream target adaptation better performance.
Figure 1: Comparison of disagreement ratio between bi-classifier model and 3-classifier model for 3-category task. 3-classifier model gives a lower disagreement ratio than bi-classifier model, which induces a tighter domain gap estimation
Ziyang Zong, Jun He, Lei Zhang, Hai Huan. Domain Gap Estimation for Source Free Unsupervised Domain Adaptation with Many Classifiers. https://arxiv.org/abs/2207.05785 (under review)
Update :
10/12/2022: DAMC source code is released!
Please refer the code for better understanding our paper: https://github.com/hejunzz/damc
We also provide our pre-trained source models and Colab notebook for SF UDA demonstration. See below!
damc.visda.train.cls12.smo0.0.3.10.pth: https://drive.google.com/file/d/1lzMr8aab5hOvIpDTC9gLn_g8YCU3Dyy_/view?usp=sharing
damc.visda.train.cls12.smo0.0.4.10.pth: https://drive.google.com/file/d/1LX59Nq9gV-xqtBrV7szojDzFYubqI_nX/view?usp=sharing