[ICLR 2021] Heteroskedastic and Imbalanced Deep Learning with Adaptive Regularization

Heteroskedastic and Imbalanced Deep Learning with Adaptive Regularization

Kaidi Cao, Yining Chen, Junwei Lu, Nikos Arechiga, Adrien Gaidon, Tengyu Ma

Dependency

The code is built with following libraries:

• PyTorch 1.8

Training

The whole HAR training pipeline can be done in the following three steps:

• To estimate the statistics through a pretrain step

python cifar_hetero_est.py --mislabel_type hetero --gpu 0 --split 0

• To calculate the weights for regularization

python weight_est.py --statspath ./log/estimate_cifar10_resnet32_hetero_0.5_0_example/stats0.pkl

• Finally train a model from the scratch

python cifar_train.py --dataset cifar10  --rand-number 0 --mislabel_type hetero --imb_type None --gpu 0 --reg_weight 10 --exp_str example --reg_path ./data/cifar10_example_weights.npy

Reference

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

@inproceedings{cao2020heteroskedastic,
  title={Heteroskedastic and imbalanced deep learning with adaptive regularization},
  author={Cao, Kaidi and Chen, Yining and Lu, Junwei and Arechiga, Nikos and Gaidon, Adrien and Ma, Tengyu},
  booktitle={International Conference on Learning Representations}, 
  year={2021} 
}