Home

Awesome

Adaptive Sharpness-Aware Minimization (ASAM)

This repository contains Adaptive Sharpness-Aware Minimization (ASAM) for training rectifier neural networks. This is an official repository for ASAM: Adaptive Sharpness-Aware Minimization for Scale-Invariant Learning of Deep Neural Networks which is accepted to International Conference on Machine Learning (ICML) 2021.

<p align="center"> <img src="img/thumbnail.png" alt="Trajectories of SAM and ASAM" width="512"/> </p>

Abstract

Recently, learning algorithms motivated from sharpness of loss surface as an effective measure of generalization gap have shown state-of-the-art performances. Nevertheless, sharpness defined in a rigid region with a fixed radius, has a drawback in sensitivity to parameter re-scaling which leaves the loss unaffected, leading to weakening of the connection between sharpness and generalization gap. In this paper, we introduce the concept of adaptive sharpness which is scale-invariant and propose the corresponding generalization bound. We suggest a novel learning method, adaptive sharpness-aware minimization (ASAM), utilizing the proposed generalization bound. Experimental results in various benchmark datasets show that ASAM contributes to significant improvement of model generalization performance.

Getting Started

Requirements

Train Examples (CIFAR)

CIFAR-10 dataset:

python example_cifar.py --dataset CIFAR10 --minimizer ASAM --rho 0.5

CIFAR-100 dataset:

python example_cifar.py --dataset CIFAR100 --minimizer ASAM --rho 1.0

We can also run SAM optimizer for CIFAR-10 or CIFAR-100 dataset:

python example_cifar.py --dataset CIFAR10 --minimizer SAM --rho 0.05
python example_cifar.py --dataset CIFAR100 --minimizer SAM --rho 0.10

Citation

If you found this code useful please cite our paper

@article{kwon2021asam,
  title={ASAM: Adaptive Sharpness-Aware Minimization for Scale-Invariant Learning of Deep Neural Networks},
  author={Kwon, Jungmin and Kim, Jeongseop and Park, Hyunseo and Choi, In Kwon},
  journal={arXiv preprint arXiv:2102.11600},
  year={2021}
}

Contributors

Jungmin Kwon (jungmin.kwon@samsung.com)
Jeongseop Kim (jisean.kim@samsung.com)
Hyunseo Park (hyunseo.park@samsung.com)
In Kwon Choi (ik21.choi@samsung.com)