LuaTorch & PyTorch implementation of orthoreg.
- Copy the contents of the backward-hook in your code.
- Add the beta (regularization strength, default 0.01), lambda (regularizer radius of influence, default 10), and epsilon (for numerically stable normalization), to the arguments of your script.
- run
@inproceedings{rodriguez2017regularizing,
title={Regularizing CNNs with Locally Constrained Decorrelations},
author={Rodr{\'{i}}guez, Pau and Gonz{\`{a}}lez, Jordi and Cucurull, Guillem and Gonfaus, Josep M. and Roca, Xavier},
booktitle={Proceedings of the International Conference on Learning Representations (ICLR)},
year={2017}
}