rodsveiga / sgd.stability

Analyze the dynamic stability of SGD

How SGD Selects the Global Minima in Over-parameterized Learning: A Dynamical Stability Perspective

by Lei Wu, Chao Ma, Weinan E

Training

python train.py --dataset fashionmnist --training_size 1000 --model_file net.pkl

Computing sharpness and non-uniformity

python diagnose.py --dataset fashionmnist --training_size 1000 --model_file net.pkl

Dependencies

• pytorch >= 0.4

Citation

@inproceedings{leiwu2018,
	title={How SGD Selects Global Minima in Over-parameterized Learning: A Dynamical Stability Perspetive},
	author={Wu, Lei and Ma, Chao and E, Weinan},
	booktitle={Advances in Neural Information Processing Systems},
	year={2018}
}