Awesome

MixupE

Codes for "MixupE: Understanding and Improving Mixup from Directional Derivative Perspective"

UAI 2023 Best Student Paper

Dataset

For CIFAR-10/100 and SVHN, these datasets can be downloaded automatically with torchvision.

Script for downloading Tiny-Imgane-Net.

Environments

Environment details used for our study.

Python: 3.8.5
PyTorch: 1.11.0+cu10.2
Torchvision: 0.12.0+cu10.2
CUDA: 10.2
NumPy: 1.22.3 
scipy 1.5.2

Train with MixupE

python main.py --dataset <Dataset_Name> --data_dir <Dataset_Path> --root_dir <Exp_Path> --train mixupe --mixupe_version 3

Take the CIFAR-100 for example,

python main.py --dataset cifar100 --data_dir ~/data/CIFAR100/ \
    --root_dir ./experiments/ --labels_per_class 5000 \
    --arch preactresnet50  --epochs 800 --schedule 10 300 600 --gammas 0.01 0.1 0.1 \
    --train mixupe --adv_gen none --adv_train none  --mixup_alpha 2.0 \
    --decay 0.0001 --mixupe_version 3 --mixup_eta 0.001 --learning_rate 0.1 --threshold 1.0 

where the hyper-parameters in "--gammas" are "initial learning rate", "learning rate decay factors".

Citation

@inproceedings{zou2023mixupe,
  title={Mixupe: Understanding and improving mixup from directional derivative perspective},
  author={Zou, Yingtian and Verma, Vikas and Mittal, Sarthak and Tang, Wai Hoh and Pham, Hieu and Kannala, Juho and Bengio, Yoshua and Solin, Arno and Kawaguchi, Kenji},
  booktitle={Uncertainty in Artificial Intelligence},
  pages={2597--2607},
  year={2023},
  organization={PMLR}
}