Skip to content

SHI-Labs/Boosted-Dynamic-Networks

Repository files navigation

Boosted-Dynamic-Networks

Haichao Yu, Haoxiang Li, Gang Hua, Gao Huang, Humphrey Shi

This repository is the official implementation for our paper Boosted Dynamic Neural Networks. In the paper, we propose a new early-exiting dynamic neural network (EDNN) architecture, where we formulate an EDNN as an additive model inspired by gradient boosting, and propose multiple training techniques to optimize the model effectively. Our experiments show it achieves superior performance on CIFAR100 and ImageNet datasets in both anytime and budgeted-batch prediction modes.

Framework

Results in Anytime Prediction Mode

MSDNet on CIFAR100 MSDNet on ImageNet RANet on CIFAR100 RANet on ImageNet

Results in Budgeted-batch Prediction Mode

MSDNet on CIFAR100 MSDNet on ImageNet RANet on CIFAR100 RANet on ImageNet

Usage

Please use the scripts in msdnet_scripts/ and ranet_scripts/ for model training and evaluation. For ImageNet experiments, please first download the dataset and put it into the proper folder.

Citation

@article{yu2022boostdnn,
	title        = {Boosted Dynamic Neural Networks},
	author       = {Yu, Haichao and Li, Haoxiang and Hua, Gang and Huang, Gao and Shi, Humphrey},
	year         = 2022,
	url          = {https://arxiv.org/abs/2211.16726},
	eprint       = {2211.16726},
	archiveprefix = {arXiv},
	primaryclass = {cs.LG}
}

Acknowledgments

This repository is built based on previous open-sourced efforts:

About

Boosted Dynamic Neural Networks, AAAI 2023

Resources

License

MIT, MIT licenses found

Licenses found

MIT
LICENSE
MIT
LICENSE_MSDNet_PyTorch

Stars

Watchers

Forks

Releases

No releases published

Packages