Imagine by Reasoning: A Reasoning-Based Implicit Semantic Data Augmentation for Long-Tailed Classification (AAAI 2022)

Prerequisite

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

Dataset

• Imbalanced CIFAR. The original data will be downloaded and converted by imbalancec_cifar.py
• Imbalanced ImageNet
• The paper also reports results on iNaturalist 2018(https://github.com/visipedia/inat_comp).

CIFAR100

In the code, we calculate the accuracy, which is different from that in the paper.

CIFAR-LT-100,long-tailed imabalance ratio of 200
python RISDA.py --gpu 3 --lr 0.1 --alpha 0.5 --beta 1 --imb_factor 0.005 --dataset cifar100 --num_classes 100 --save_name simple --idx cifar_im200

CIFAR-LT-100,long-tailed imabalance ratio of 100
python RISDA.py --gpu 3 --lr 0.1 --alpha 0.5 --beta 0.75 --imb_factor 0.01 --dataset cifar100 --num_classes 100 --save_name simple --idx cifar_im100

Image Classification on ImageNet

Run

Train ResNet-50 on ImageNet-LT

CUDA_VISIBLE_DEVICES=1,0 python imagenet_ISDA_train.py  /datapath/ILSVRC2012_LT/ --model resnet50 --batch-size 128 --lr 0.1 --epochs 100 --alpha_0 0 --beta_0 7.5 --workers 1  --world-size 1 --rank 0  --stage1 80 --stage2 90 

Test ResNet-50 on ImageNet-LT

CUDA_VISIBLE_DEVICES=1,0 python imagenet_ISDA_train.py  /datapath/ILSVRC2012_LT/ --model resnet50 --batch-size 128 --lr 0.1 --epochs 100 --alpha_0 0 --beta_0 7.5 --workers 1  --world-size 1 --rank 0  --stage1 80 --stage2 90 --evaluate checkpoint/best.pth.tar

More details will be uploaded soon.

Acknowledgements

Some codes in this project are adapted from MetaSAug and ISDA. We thank them for their excellent projects.

Citation

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

@inproceedings{chen2021imagine,
  title={Imagine by Reasoning: A Reasoning-Based Implicit Semantic Data Augmentation for Long-Tailed Classification},
  author={Chen, Xiaohua and Zhou, Yucan and Wu, Dayan and Zhang, Wanqian and Zhou, Yu and Li, Bo and Wang, Weiping},
  booktitle = {Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI)},
  year={2022}
}