Evolving Transferable Neural Pruning Functions [Paper] [ArXiv]

Paper in GECCO2022

@inproceedings{liu2022evolving,
  title     = {Evolving transferable neural pruning functions},
  author    = {Liu, Yuchen and Kung, Sun-Yuan and Wentzlaff, David},
  booktitle = {Proceedings of the Genetic and Evolutionary Computation Conference},
  pages     = {385--394},
  year      = {2022}
}

This repository contains the codes for Evolving Transferable Neural Pruning Functions, which is published in GECCO.

Overview

We propose to learn transferable channel pruning functions via an evolution strategy, genetic programming. The evolved functions show generalizability when applied to datasets and networks different from that in evolution.

Usage

Working Environment

We have tested our codes with the following setups:

python==3.6.5
tensorflow==1.14.0
numpy==1.16.0

Evolution

We evolve the pool of functions on two pruning tasks: LeNet-5 on MNIST and VGG-16 on CIFAR-10. The execution codes are under the folder of Evolution.

Transfer Pruning

The evolved functions can be used to conduct transfer pruning on datasets and networks that are different from the evolution. For instance, you can prune ResNet on CIFAR-100 with the evolved function (Equation 3 in the paper) under Transfer_Pruning.