Revive Re-weighting in Imbalanced Learning by Density Ratio Estimation

by Jiaan Luo*, Feng Hong*, Jiangchao Yao, Bo Han, Ya Zhang, Yanfeng Wang at SJTU, Shanghai AI Lab, and HKBU.

Citation

@inproceedings{
luo2024revive,
title={Revive Re-weighting in Imbalanced Learning by Density Ratio Estimation},
author={Luo Jiaan and Hong Feng and Yao Jiangchao and Han Bo and Zhang Ya and Wang Yanfeng},
booktitle={NeurIPS},
year={2024}
}

Environment

The project has been tested under the following environment settings:

• OS: Ubuntu 18.04.5 LTS
• GPU: NVIDIA GeForce RTX 3090
• CUDA: 11.7
• Cudatoolkit: 11.0.221
• Python: 3.8.18
• PyTorch: 1.13.1

Content

Structure

• ./utils: data augmentation, model management and evaluation
• ./models: backbone models
• ./data: datasets (automatically downloaded)
• main.py: main function script
• imbalance_cifar.py: custom datasets script

Run

python main.py --gpu 0 --imb_type exp --dataset cifar10 --imb_factor 0.01

Evaluate

python main.py --gpu 0 --evaluate --resume path/to/saved/model --dataset cifar10 --imb_factor 0.01

Implement Other Datasets

• Add raw data to ./[_data_name].
• Create a imbalanced version of the dataset in ./data. (Optional)
• Create a dataloader in ./utils.

Contact

If you have any problem with this code, please feel free to contact luojiaan@sjtu.edu.cn.