AL-GradNorm

Pytorch implementation of our paper: Boosting Active Learning via Improving Test Performance, AAAI 2022. arXiv version is here

Introduction

• With the aid of the influence function, we derive that unlabeled data of higher gradient norm should be selected for annotation in active learning.
• This work explains why some data can benefit test performance whereas some data cannot.

Requirements

numpy

pytorch 1.7+

torchvision 0.10+

Data Prep

Download the dataset (e.g. Cifar-10) and unzip it to your preferred folder.

Specify the path of your folder in line 43-45 of the main.py file.

Run the Code

python main.py

Notes

The config.py file includes all the hyper-parameters. Set SCHEME=0 for the expected-gradnorm scheme and SCHEME=1 for the entropy-gradnorm scheme. Set NUM_CLASS accordingly, for example, 10 for Cifar10.

Currently, the TRIALS is set to 3 in config.py because the reported results are averaged over 3 runs. One can change it to 1 if using 3 GPUs to run the program concurrently.

Citation

@inproceedings{wang2022boosting,
  title={Boosting Active Learning via Improving Test Performance}, 
  author={Wang, Tianyang and Li, Xingjian and Yang, Pengkun and Hu, Guosheng and Zeng, Xiangrui and Huang, Siyu and Xu, Cheng-Zhong and Xu, Min},  
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)},  
  year={2022}
 }

Contact

Tianyang Wang

toseattle@siu.edu