songweige/HEX

Example implementation for the paper: Learning Robust Representations by Projecting Superficial Statistics Out

Learning Robust Representations by Projecting Superficial Statistics Out

Example implementation of the paper:

• H. Wang, Z. He, Z. C. Lipton, and E. P. Xing, Learning Robust Representations by Projecting Superficial Statistics Out, Proceedings of Seventh International Conference on Learning Representations (ICLR 2019).

Code structure:

• scripts/
  • __init__.py
  • datagenerator.py Helper function for to load ImageNet data, not part of the contribution of this work.
  • model.py implementation of NGLCM and HEX plugged into AlexNet
    • AlexNet (original implementation): Line 20-170
    • AlexNet (with NGLCM and HEX): Line 5-114, Line 175-280
      • NGLCM: Line 5-17, Line 187-199
      • Equation 3: Line 238-244
      • Expanding the final layer: Line 246-253
      • Equation 4: Line 266-278
      • Normalization is recommended: Line 233-236
  • run.py training and testing the model in classification
    • Prepare the data for NGLCM: Line 23-46

Replication

For the codes that are used to replicate the experiments in the paper, please visit HaohanWang/HEX_experiments

Bibtex

@inproceedings{
wang2018learning,
title={Learning Robust Representations by Projecting Superficial Statistics Out},
author={Haohan Wang and Zexue He and Zachary L. Lipton and Eric P. Xing},
booktitle={International Conference on Learning Representations},
year={2019},
url={https://openreview.net/forum?id=rJEjjoR9K7},
}

Contact

• Haohan Wang
• @HaohanWang