RelationMatch

Official implementation of paper "RelationMatch: Matching In-batch Relationships for Semi-supervised Learning" (https://arxiv.org/abs/2305.10397).

• Achieving 15.21% accuracy improvement over FlexMatch on the STL-10 dataset!

Installation

You can create the ssl environment by just running this command: conda env create -f environment.yaml

Then you may activate ssl environment by conda activate ssl

Runnning experiments

Under ssl environment, run following command:

python BASE_METHOD.py --c=PATH/TO/YOUR/CONFIG

BASE_METHOD can be one of fixmatch, flexmatch, with respect to RelationMatch and RelationMatch(w/ CPL) in paper.

We provide our example config in configs/BASE_METHOD/ folder with relation suffix. You can use diff to find the difference between relation match configs and TorchSSL configs, and it's easy to modify other TorchSSL configs to RelationMatch configs follow our sample configs.

Acknowledgement

We implement our RelationMatch method based on TorchSSL repo. Thanks for their wonderful work!

Citations

Please cite the paper and star this repo if you use RelationMatch and find it interesting/useful, thanks!

@article{zhang2023relationmatch,
  title={RelationMatch: Matching In-batch Relationships for Semi-supervised Learning},
  author={Zhang, Yifan and Yang, Jingqin and Tan, Zhiquan and Yuan, Yang},
  journal={arXiv preprint arXiv:2305.10397},
  year={2023}
}