MEDI

Hi, this is the core code of our rencent work "Meta Discovery: Learning to Discover Novel Classes given Very Limited Data" (ICLR 2022, https://openreview.net/pdf?id=MEpKGLsY8f). This work is done by

• Haoang Chi (NUDT), haoangchi618@gmail.com
• Dr. Feng Liu (UTS), feng.liu@uts.edu.au
• Dr. Bo Han (HKBU), bhanml@comp.hkbu.edu.hk
• Dr. Wenjing Yang (NUDT), wenjing.yang@nudt.edu.cn
• Dr. Long Lan (NUDT), long.lan@nudt.edu.cn
• Dr. Tongliang Liu (USYD), tongliang.liu@sydney.edu.au
• Dr. Gang Niu (RIKEN AIP), gang.niu.ml@gmail.com
• Dr. Mingyuan Zhou (UT), mingyuan.zhou@mccombs.utexas.edu
• Prof. Masashi Sugiyama (RIKEN AIP), sugi@k.u-tokyo.ac.jp

Software

Torch version is 1.7.1. Python version is 3.7.6. CUDA version is 11.0.

These python files, of cause, require some basic scientific computing python packages, e.g., numpy. I recommend users to install python via Anaconda (python 3.7.6), which can be downloaded from https://www.anaconda.com/distribution/#download-section . If you have installed Anaconda, then you do not need to worry about these basic packages.

After you install anaconda and pytorch (gpu), you can run codes successfully.

MEDI

Please feel free to test the MEDI method by running main.py.

Specifically, please run

CUDA_VISIBLE_DEVICES=0 python main.py

in your terminal (using the first GPU device).

The pretrained checkpoints can be downloaded from https://github.com/google-research/simclr?tab=readme-ov-file and converted to Pytorch format with https://github.com/tonylins/simclr-converter.

Citation

If you are using this code for your own researching, please consider citing

@inproceedings{chi2022meta,
  title={Meta discovery: Learning to discover novel classes given very limited data},
  author={Chi, Haoang and Liu, Feng and Han, Bo and Yang, Wenjing and Lan, Long and Liu, Tongliang and Niu, Gang and Zhou, Mingyuan and Sugiyama, Masashi},
  booktitle={International Conference on Learning Representations},
  year={2022}
}

Acknowledgment

This work was partially supported by the National Natural Science Foundation of China (No. 91948303-1, No. 61803375, No. 12002380, No. 62106278, No. 62101575, No. 61906210) and the National Grand R&D Plan (Grant No. 2020AAA0103501). BH was supported by NSFC Young Scientists Fund No. 62006202 and RGC Early Career Scheme No. 22200720. TLL was supported by Australian Research Council Projects DE-190101473 and DP-220102121. MS was supported by JST CREST Grant Number JPMJCR18A2.