This repository contains the code of our paper UMIX: Improving Importance Weighting for Subpopulation Shift via Uncertainty-Aware Mixup. The code is implemented on the code provided by WILDs. If you have any questions, please contact me via the following email zongbo at tju.edu.cn.
- Python 3
- torch 1.7.0
- torch-scatter 2.0.5
- torch-geometric 1.6.2
- torchvision 0.8.1+cu101
- torch-cluster 1.5.8
- torch-sparse 0.6.8
- numpy 1.18.5
- pandas 1.1.5
- pillow 7.2.0
- scikit-learn 0.23.2
- scipy 1.5.2
- transformers 4.15.0
- waterbirds
- CelebA
- civilcomments
- camelyon17
You can run our algorithm from the shell file in the script directory. Specifically, taking the Waterbirds dataset as an example,
- you should first run the files in the UMIX_trajectory folder to get the uncertainty.
- Then run the files in the UMIX folder to get the model results.
- You can then re-search for hyperparameters through the files under the UMIX_nni folder.
We also provide our saved checkpoints in this link.
The code is built on WILDS codebase v1.2.2 (https://github.com/p-lambda/wilds/releases/tag/v1.2.2). We thank the builders of the original repository. Because our method requires some adjustments to the original codebase, we directly upload our modified files for efficiency.
This is not an official Tencent product.
This tool is developed in Tencent AI Lab.
The copyright holder for this project is Tencent AI Lab.
All rights reserved.