Repository files navigation

Official pytorch implementation of "Uncertainty Modeling for Out-of-Distribution Generalization" in International Conference on Learning Representations (ICLR) 2022.

By Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, Ling-Yu Duan.

In this paper, we improve the network generalization ability by modeling the uncertaintyof domain shifts with synthesized feature statistics during training. Specifically,we hypothesize that the feature statistic, after considering the potential uncertainties, follows a multivariate Gaussian distribution. Hence, each feature statistic isno longer a deterministic value, but a probabilistic point with diverse distributionpossibilities. With the uncertain feature statistics, the models can be trained toalleviate the domain perturbations and achieve better robustness against potentialdomain shifts.

We use the following versions of OS and softwares:

• OS: Ubuntu 16.04
• Python: 3.6.8
• PyTorch: 1.5.0
• CUDA: 10.1

Following the instructions of individual repositories to install the required environments.

The experiments include:instance retrieval (person re-identification on DukeMTMC and Market1501),multi-domain generalization (PACS and Office-Home), andsemantic segmentation (from GTA5 to CityScapes). The core code of our method can be found indsu.py

• The detailed instructions to run experiments on instance retrieval can be found atinstance-retrieval-readme.md.

• The detailed instructions to run experiments on multi-domain generalization can be found atmulti-domain-generalization-readme.md.

• The detailed instructions to run experiments on semantic segmentation can be found atsemantic-segmentation-readme.md.

If you find our work is useful for your research, please kindly cite our paper.

@inproceedings{li2022uncertainty,title={Uncertainty Modeling for Out-of-Distribution Generalization},author={Xiaotong Li and Yongxing Dai and Yixiao Ge and Jun Liu and Ying Shan and LINGYU DUAN},booktitle={International Conference on Learning Representations},year={2022},url={https://openreview.net/forum?id=6HN7LHyzGgC}}

If you have any questions, you can contact me from the email:lixiaotong@stu.pku.edu.cn