CPE-LTSSL

Official code for "Complementary Experts for Long-tailed Semi-Supervised Learning" (AAAI 2024)

Chengcheng Ma^1,2, Ismail Elezi³, Jiankang Deng³, Weiming Dong¹, Changsheng Xu¹.

¹ Chinese Academy of Sciences Institute of Automation (CASIA)
² University of the Chinese Academy of Sciences (UCAS)
³ Huawei Technologies Co Ltd

Introduction

This repo is based on the public and widely-used codebase USB.

What I've done is just adding our CPE algorithm in semilearn/imb_algorithms/cpe, plus some baseline algorithms which are missing in USB, such as RDA and ACR.

Also, I've made corresponding modifications to semilearn/nets/ and several __init__.py.

How to run

For example, on CIFAR-10-LT with $\gamma_l=\gamma_u=100$

CUDA_VISIBLE_DEVICES=0 python train.py --c config/classic_cv_imb/fixmatch_cpe/fixmatch_cpe_cifar10_lb1500_100_ulb3000_100_0_2_4_0.yaml

(Note: I know that USB supports multi-GPUs, but I still recommend you to run on single GPU, as some weird problems may occur.)

The model will be automatically evaluated every 1024 iterations during training. After training, the last two lines in saved_models/classic_cv_imb/fixmatch_cpe_cifar10_lb1500_100_ulb3000_100_0_2_4_0/log.txt will tell you the best accuracy.

For example,

[2023-08-07 15:46:32,882 INFO] model saved: ./saved_models/classic_cv_imb/fixmatch_cpe_cifar10_lb1500_100_ulb3000_100_0_2_4_0/latest_model.pth
[2023-08-07 15:46:32,884 INFO] Model result - eval/best_acc : 0.8479
[2023-08-07 15:46:32,884 INFO] Model result - eval/best_it : 177151

Results

I've put all the log.txt on the CIFAR-10-LT dataset in saved_models. Please take a look. I also put them on Google Drive (link).

Note that a few training process were killed halfway by me. But you can still check the BEST_EVAL_ACC in their log.txt. I promise they are real.

The reported accuracies in Table 2 and 3 in our AAAI paper are the average over three different runs (random seeds are 0/1/2). I wrote a tiny script to realize quick observation and computation.

Citation

If you find this work useful, please consider citing our paper. We provide a BibTeX entry of our paper below:

@misc{ma2023heads,
      title={Three Heads Are Better Than One: Complementary Experts for Long-Tailed Semi-supervised Learning}, 
      author={Chengcheng Ma and Ismail Elezi and Jiankang Deng and Weiming Dong and Changsheng Xu},
      year={2023},
      eprint={2312.15702},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Contact

Feel free to contact me via machengcheng2016@gmail.com if you have any problems about our paper or codes.