Learning imbalanced datasets with maximum margin loss

Kang, Haeyong and Vu, Thang and Yoo, Chang D

---

This is the official implementation of MM-LDAM-DRW in the paper Learning Imbalanced Datasets with Maximum Margin Loss in PyTorch.

Our baseline code follows the official implementation of LDAM-DRW in the paper Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss in PyTorch.

Dependency

The code is built with following libraries:

• PyTorch 1.11.0
• TensorboardX
• scikit-learn

Dataset

• Imbalanced CIFAR. The original data will be downloaded and converted by imbalancec_cifar.py.
• The paper also reports results on Tiny ImageNet and iNaturalist 2018. We will update the code for those datasets later.

Training

We provide several training examples with this repo:

• To train the ERM baseline on long-tailed imbalance with ratio of 100

python cifar_train.py --gpu 0 --imb_type exp --imb_factor 0.01 --loss_type CE --train_rule None

• To train the LDAM Loss along with DRW training on long-tailed imbalance with ratio of 100

python cifar_train.py --gpu 0 --imb_type exp --imb_factor 0.01 --loss_type LDAM --train_rule DRW

• To train the Maximum Margin Loss along with DRW training on long-tailed imbalance with ratio of 100 (hyper-params of scale, max_m, and gamma should be set)

python cifar_train.py --dataset cifar100 --gpu 0 --imb_type exp --imb_factor 0.01 --loss_type HMM \
           --train_rule DRW --epochs 200 --scale 19.0 --max_m 1.29 --gamma 1.528 --seed 1 --exp_str logits

Reference

If you find our paper and repo useful, please cite as

@inproceedings{kang2021learning,
  title={Learning imbalanced datasets with maximum margin loss},
  author={Kang, Haeyong and Vu, Thang and Yoo, Chang D},
  booktitle={2021 IEEE International Conference on Image Processing (ICIP)},
  pages={1269--1273},
  year={2021},
  organization={IEEE}
}