zkk995 / Robust-Locally-Linear-Networks

The official implementation for the ICLR 19' paper "Towards Robust, Locally Linear Deep Networks"

Towards Robust, Locally Linear Deep Networks:

This repository is for the paper

• "Towards Robust, Locally Linear Deep Networks" by Guang-He Lee, David Alvarez Melis, and Tommi S. Jaakkola in ICLR 19.
• Project page

Package version:

• PyTorch0.4.1
• python3.6.1

Reproducing the MNIST experiment:

• Please execute the shell files (gamma100_fc.sh) to reproduce the experiment on MNIST dataset with gamma=100. The results will be in the folder fc_log/

• parse_log.py is a utility script. After you run all the models using gamma100_fc.sh. Use the following comment:

  • ls fc_log > fc_log.list
  • cd fc_log
  • python ../parse_log.py --file-list ../fc_log.list

• To inspect the best model in terms of the median of L2 margin given each validation accuracy, please look into the log file to see the testing scores of the model.

Other experiments:

• Unfortunately we don't plan to release the codes for other experiments.

Citation:

If you find this repo useful for your research, please cite the paper

@inproceedings{
lee2018towards,
title={Towards Robust, Locally Linear Deep Networks},
author={Guang-He Lee and David Alvarez-Melis and Tommi S. Jaakkola},
booktitle={International Conference on Learning Representations},
year={2019},
url={https://openreview.net/forum?id=SylCrnCcFX},
}