MetaSAug: Meta Semantic Augmentation for Long-Tailed Visual Recognition (CVPR 2021) arXiv

Prerequisite

• PyTorch >= 1.2.0
• Python3
• torchvision
• PIL
• argparse
• numpy

Evaluation

For faster evaluation, we provide several pre-trained models of MetaSAug. We can use MetaSAug_test_CE.sh & MetaSAug_test_LDAM.sh to test MetaSAug with cross-entropy loss and LDAM loss, respectively. The models are stored in checkpoints/ours.

Evaluation examples:

• sh MetaSAug_test_CE.sh
• sh MetaSAug_test_LDAM.sh

Training example

CIFAR-LT-100, MetaSAug with LDAM loss
python3.6 MetaSAug_LDAM_train.py --gpu 0 --lr 0.1 --lam 0.75 --imb_factor 0.05 --dataset cifar100 --num_classes 100 --save_name MetaSAug_cifar100_LDAM_imb0.05 --idx 1

Acknowledgements

Some codes in this project are adapted from Meta-class-weight. We thank them for their excellent projects.

Citation

If you find this code useful for your research, please cite our paper:

@inproceedings{li2021MetaSAug,
author = {Li, Shuang and Gong, Kaixiong and Liu, Chi Harold and Wang, Yulin and Qiao, Feng and Cheng, Xinjing},
title = {MetaSAug: Meta Semantic Augmentation for Long-Tailed Visual Recognition},
year = {2021},
booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
}