UnMixMatch

Official implementation of our AAAI 2024 paper:

Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data
Shuvendu Roy, Ali Etemad
In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI-24)

Unconstrained Setting

OpenSet Setting

Data

• Labelled data: CIFAR-10, CIFAR-100, SVHN, STL-10 will be downloaded automatically.
• Download and organize imagenet100 following the instructions at https://github.com/danielchyeh/ImageNet-100-Pytorch

Run the experiments

1. Modify the config file in config/cifar10_40_0.yaml as you need. Include your data directory for imagenet100 in the config file.
2. Run python unmixmatch.py --c config/cifar10_40_0.yaml

This settings will run UnMixMatch on CIFAR-10 with 40 labels per class and get and accuracy of 47.91±1.1.

Acknowledgement

We thank the authors of the following repositories for releasing their code. The implementation of UnMixMatch is built over the implementation of ReMixMatch from this repository: https://github.com/TorchSSL/TorchSSL

Citing UnMixMatch

If you think this toolkit or the results are helpful to you and your research, please cite our paper:

@inproceedings{UnMixMatch,
  title={Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data},
  author={Roy, Shuvendu and Etemad, Ali},
  booktitle={AAAI Conference on Artificial Intelligence},
  year={2024}
}