Domain Gap Estimation for Source Free Unsupervised Domain Adaptation with Many Classifiers

https://arxiv.org/abs/2207.05785

In theory, the success of unsupervised domain adaptation (UDA) largely relies on domain gap estimation. However, for source free UDA, the source domain data can not be accessed during adaptation, which poses great challenge of measuring the domain gap. In this paper, we propose to use many classifiers to learn the source domain decision boundaries, which provides a tighter upper bound of the domain gap, even if both of the domain data can not be simultaneously accessed. The source model is trained to push away each pair of classifiers whilst ensuring the correctness of the decision boundaries. In this sense, our many classifiers model separates the source different categories as far as possible which induces the maximum disagreement of many classifiers in the target domain, thus the transferable source domain knowledge is maximized. For adaptation, the source model is adapted to maximize the agreement among pairs of the classifiers. Thus the target features are pushed away from the decision boundaries. Experiments on several datasets of UDA show that our approach achieves state of the art performance among source free UDA approaches and can even compete to source available UDA methods.

Colab Jupyter notebook for SF UDA demonstration

Please refer to https://sites.google.com/site/hejunzz/da

We provide the pre-trained 12-classifier source models for VisDA-2017. The pre-trained source models were trained by using $\alpha_s=0.3$ and $\alpha_s=0.4$. You can set either src_alpha=0.3 or src_alpha=0.4 to load the model respectively. The pre-trained models can be downloaded from my Google drive as follows.

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