ermongroup / NDA

Negative Data Augmentation

Official Code for the paper Negative Data Augmentation accepted at ICLR 2020

Paper Link.

To train with Jigsaw NDA for unconditional Cifar-10, run the following -

bash train_C10.sh

To evaluate the trained model, run -

bash eval_C10.sh jigsaw_C10

For conditional Cifar-10, run the following -

bash train_C10_cond.sh

To evaluate the trained model, run -

bash eval_C10_cond.sh jigsaw_C10_cond

Evaluating pre-trained model

To evaluate pretrained model for unconditional Cifar-10, run the following -

bash eval_C10.sh jigsaw_seed2_C10_alpha_0.25_beta_0.75

For conditional Cifar-10, run the following -

bash eval_C10_cond.sh jigsaw_C10_conditional_seed2_alpha_0.25_beta_0.75

Using other NDA

Lines 242-246 in train_fns_aug.py contain other NDA augmentations, uncomment the corresponding line to use that NDA. Change the experiment_name argument in train_C10.sh or train_C10_cond.sh to generate a seperate model for that NDA

If you use this code for your research, Please cite using

@article{sinha2021negative,
  title={Negative data augmentation},
  author={Sinha, Abhishek and Ayush, Kumar and Song, Jiaming and Uzkent, Burak and Jin, Hongxia and Ermon, Stefano},
  journal={arXiv preprint arXiv:2102.05113},
  year={2021}
}