[ICLR 2023] DFPC: Data flow driven pruning of coupled channels without data.

August 25, 2023 ยท View on GitHub

This repository is for the new deep neural network pruning method introduced in the following ICLR 2023 paper:

DFPC: Data flow driven pruning of coupled channels without data. [Camera Ready]
Tanay Narshana, Chaitanya Murti, and Chiranjib Bhattacharyya
Indian Institute of Science, Bengaluru, India.

TLDR: This paper introduces a novel method for pruning of networks containing coupled connections without using data, DFPC.

We posit "coupled connections" as a bottleneck to obtain lower inference latencies when pruning networks. We then provide a formalization to abstract "coupled connections" and use it to derive a data-free way to measure the importance of coupled neurons in a network. Our experimental results display the merit in pruning "coupled connections" for they obtain pruned models with a better latency-vs-accuracy.

We're working up to clean up our code and provide our models in a clean way. Everything should be up by mid June Coming up..

  1. Data-free code for pruning mobilenets
  2. Data-driven code for pruning resnets
  3. pruned models for resnet-50 in the data-driven regime.

Available Code:

  1. Code for data-free experiments is available in the Data-Free folder.
  2. Pruned models for the data-driven experiment for ResNet-50 on the ImageNet dataset is available in the Pruned-Models folder.

Feel free to contact us at tanay.narshana@gmail.com. (Email is more recommended if you'd like quicker reply)

Reference

Please cite this in your publication if our work helps your research:

@inproceedings{narshana2023dfpc,
title={{DFPC}: Data flow driven pruning of coupled channels without data.},
author={Tanay Narshana and Chaitanya Murti and Chiranjib Bhattacharyya},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=mhnHqRqcjYU}
}

Acknowledgments

We are grateful for the code made available by pytorch imagenet example, Regularization-Pruning, and pytorch-summary.

Visitors