Inducing Neural Collapse in Imbalanced Learning: Do We Really Need a Learnable Classifier at the End of Deep Neural Network?

Code for our paper Inducing Neural Collapse in Imbalanced Learning: Do We Really Need a Learnable Classifier at the End of Deep Neural Network? (NeurIPS 2022)

by Yibo Yang, Shixiang Chen, Xiangtai Li, Liang Xie, Zhouchen Lin, and Dacheng Tao.

If you find our work interesting and helpful to your project, please consider citing our paper.

@article{yang2022we,
  title={Do We Really Need a Learnable Classifier at the End of Deep Neural Network?},
  author={Yang, Yibo and Xie, Liang and Chen, Shixiang and Li, Xiangtai and Lin, Zhouchen and Tao, Dacheng},
  journal={arXiv preprint arXiv:2203.09081},
  year={2022}
}

The long-tailed classification code is based on MiSLAS.

Usage

The following command is for training.

python train.py --cfg {path_to_config}

The configs can be found in ./config/DATASET/FILENAME.yaml. The FILENAME that ends with ETF_DR uses our ETF classifier and the Dot-Regression loss. Otherwise, it is the baseline practice using a learnable classifier with the CE loss.

Note that the experiment on ImageNet-LT is performed on multiple GPUs, while the other experiments (CIFAR-10, CIFAR-100, SVHN, and STL-10) are performed on one GPU.

Contact

For any question, please contact Yibo (ibo@pku.edu.cn).