Class-Conditional Sharpness Aware Minimization for Deep Long-Tailed Recognition

This is the implementation of our CVPR'23 paper "Class-Conditional Sharpness-Aware Minimization for Deep Long-Tailed Recognition". In this paper, we propose a new variant of SAM to better adapt to the scenario of deep long-tailed recognition.

Installation

Requirements

• python 3.6
• PyTorch 1.4.0
• torchvision 0.5.0
• opencv 4.2
• numpy 1.19.5

Training

CIFAR-10-LT

Specify the data path ("data_root") in configs/Cifar10.json. Then running the following commend:

$ python3 train_cifar.py --config ./configs/Cifar10.json

CIFAR-100-LT

Specify the data path ("data_root") in configs/Cifar100.json. Then running the following commend:

$ python3 train_cifar.py --config ./configs/Cifar100.json

Stronger Augmentation Usage

Uncomment the CIFAR10Policy and Cutout in datasets/Cifar.py, and change the "mixup" to false in Cifar10.json/Cifar100.json:

self.transform = transforms.Compose([
                transforms.RandomCrop(32, padding=4),
                transforms.RandomHorizontalFlip(),
                CIFAR10Policy(),    # add AutoAug
                transforms.ToTensor(),
                Cutout(n_holes=1, length=16),
                transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
            ])

Citation

@inproceedings{zhou2023ccsam,
author = {Zhou, Zhipeng and Li, Lanqing and Zhao, Peilin and Heng, Pheng-Ann and Gong, Wei},
title = {Class-Conditional Sharpness Aware Minimization for Deep Long-Tailed Learning},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year = {2023}
}

Correspondence

If you have any further questions, please feel free to contact Zhipeng Zhou by zzp1994@mail.ustc.edu.cn