DivDis

Code for the paper "Diversify and Disambiguate: Learning From Underspecified Data" (arXiv, project page).

The main repulsion loss for DivDis is implemented in divdis.py in the root directory, and shared by all experiments.

Citation

If you find this code useful, please cite our paper:

@inproceedings{lee2022divdis,
  title     = {Diversify and Disambiguate: Learning From Underspecified Data},
  author    = {Lee, Yoonho and Yao, Huaxiu and Finn, Chelsea},
  year      = {2022}
}

Code references

• MNIST-CIFAR repo (base code for MNIST-CIFAR experiment)
• GroupDRO repo (base code for subpopulation shifts)