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..
Data-free code for pruning mobilenets- Data-driven code for pruning resnets
pruned models for resnet-50 in the data-driven regime.
Available Code:
- Code for data-free experiments is available in the
Data-Free
folder. - 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)
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}
}
We are grateful for the code made available by pytorch imagenet example, Regularization-Pruning, and pytorch-summary.