This repository contains code for the paper A Signal Propagation Perspective for Pruning Neural Networks at Initialization (ICLR 2020).

• tensorflow >= 1.14
• python >= 3.6
• packages in requirements.txt

Put the following datasets in your preferred location (e.g., ./data).

• MNIST
• Fashion-MNIST
• CIFAR-10
• Tiny-ImageNet

To run the code (MLP-7-linear on MNIST by default):

$ python main.py --path_data=./data

To enforce approximate dynamical isometry while checking signal propagation on network:

$ python main.py --path_data=./data --check_jsv --enforce_isometry

See main.py to run with other options.

If you use this code for your work, please cite the following:

@inproceedings{lee2020signal,
  title={A signal propagation perspective for pruning neural networks at initialization},
  author={Lee, Namhoon and Ajanthan, Thalaiyasingam and Gould, Stephen and Torr, Philip HS},
  booktitle={ICLR},
  year={2020},
}

This project is licensed under the MIT License. See the LICENSE file for details.