DNR: A Tunable Robust Pruning FrameworkThrough Dynamic Network Rewiring of DNNs

This repo contains the test codes and saved models of our ASP-DAC 2021 paper: DNR: A Tunable Robust Pruning FrameworkThrough Dynamic Network Rewiring of DNNs

Authors:

1. Souvik Kundu (souvikku@usc.edu)
2. Mahdi Nazemi (nazemi@usc.edu)
3. Peter A. Beerel (pabeerel@usc.edu)
4. Massoud Pedram (pedram@usc.edu)

Robust RESNET saved modes:

1. ResNet18 Channel pruned on CIFAR-10
2. ResNet18 Irregular pruned on CIFAR-10

Robust VGG saved modes:

1. VGG16 Channel pruned on CIFAR-10
2. VGG16 Irregular pruned on CIFAR-10

To test adversarial accuracy of a saved model, please:

1. Copy and bring the model to same location as the *.py files.
2. Open the run_test.py file to change model_type ['resnet18' or 'vgg16'] and provide the dataset='cifar10'.
3. Provide --adv_eval to enable adversarial evaluation.
4. Run python file: 'python run_test.py'

Arxiv pre-print version:

arxiv_version

Cite this work:

  @inproceedings{kundu2021dnr,
  title={DNR: A Tunable Robust Pruning Framework Through Dynamic Network Rewiring of DNNs},
  author={Kundu, Souvik and Nazemi, Mahdi and Beerel, Peter A and Pedram, Massoud},
  booktitle={Proceedings of the 26th Asia and South Pacific Design Automation Conference},
  pages={344--350},
  year={2021}
  }