We recommend to create a conda environment and install the following:
- Python 3.7.11
- Pytorch 1.8.0 (GPU-enabled version)
In some cases, the package request might be missing in the environment. To install request,
python -m pip install requests
Simple run
python open_lth.py
In response, you will see the following message.
==================================================================================
OpenLTH: A Framework for Research on Lottery Tickets and Beyond
----------------------------------------------------------------------------------
Choose a command to run:
* open_lth.py train [...] => Train a model.
* open_lth.py lottery [...] => Run a lottery ticket hypothesis experiment.
* open_lth.py lottery_branch [...] => Run a lottery branch.
==================================================================================
This repo includes codes for the experiment implementation of the paper On the Neural Tangent Kernel Analysis of Randomly Pruned Neural Networks, by Hongru Yang and Zhangyang Wang.
The implementation is heavily based on Jonathan Frankle's implemenation for experiments on the lottery ticket hypothesis, and further developed in Xiaohan Chen's work.
To run random pruning, use the following command:
python open_lth.py lottery --default_hparams=cifar_resnet_50 --levels=16 --pruning_strategy=random_global --replicate=1
If you find this useful, please cite the following paper:
@InProceedings{pmlr-v206-yang23b,
title = {On the Neural Tangent Kernel Analysis of Randomly Pruned Neural Networks},
author = {Yang, Hongru and Wang, Zhangyang},
booktitle = {Proceedings of The 26th International Conference on Artificial Intelligence and Statistics},
pages = {1513--1553},
year = {2023},
editor = {Ruiz, Francisco and Dy, Jennifer and van de Meent, Jan-Willem},
volume = {206},
series = {Proceedings of Machine Learning Research},
month = {25--27 Apr},
publisher = {PMLR},
pdf = {https://proceedings.mlr.press/v206/yang23b/yang23b.pdf},
url = {https://proceedings.mlr.press/v206/yang23b.html},
}
This repository is licensed under the MIT license, as found in the LICENSE file.