Distribution Alignment Optimization through Neural Collapse for Long-tailed Classification (ICML, 2024)

November 5, 2025 ยท View on GitHub

by Jintong Gao1, He Zhao2, Dandan Guo1, Hongyuan Zha3,

1Jilin University, 2CSIRO's Data61, 3The Chinese University of Hong Kong, Shenzhen

This is the official implementation of Distribution Alignment Optimization through Neural Collapse for Long-tailed Classification in PyTorch.

Requirements:

All codes are written by Python 3.8 with

PyTorch >=1.5
torchvision >=0.6
TensorboardX 1.9
Numpy 1.17.3

Training

To train the model(s) in the paper, run this command:

CIFAR-LT

CIFAR-10-LT (CE-DRW + DisA):

python cifar_train.py --dataset cifar10 --num_classes 10 --loss_type CE --train_rule DRW --lamda 0.1 --gpu 0

CIFAR-100-LT (CE-DRW + DisA):

python cifar_train.py --dataset cifar100 --num_classes 100 --loss_type CE --train_rule DRW --lamda 0.1 --gpu 0

Evaluation

To evaluate my model, run:

python test.py --dataset cifar10 --num_classes 10 --gpu 0 --resume model_path

Citation

If you find our paper and repo useful, please cite our paper.

@inproceedings{
Gao2024DisA,
title={Distribution Alignment Optimization through Neural Collapse for Long-tailed Classification},
author={Jintong Gao and He Zhao and Dandan Guo and Hongyuan Zha},
booktitle={International Conference on Machine Learning (ICML)},
year={2024}
}

Acknowledgement

ETF-DR

INC

RBL

Contact

If you have any questions when running the code, please feel free to concat us by emailing