hbyido / LpSS

Training sparse neural network with Lp normalized weight

Lp-spherical Sparse Learning

Controlling Sparsity of Neural Network with Weight Constrained on Unit Lp-Sphere.
For more information, please contact author by: [e-mail]

The corresponding paper is under review.

System Requirements

Operating systems: Windows or Linux
Dependencies: CUDA, CuDNN, Python (version >= 3.5) Pytorch (version >= 1.2) and corresponding torchvision

Demo

After installed all of the requirements, please run demo.py

Results

• Procedures of LpSS

• Weight distribution of a neuron with respect of p

• Summarization of test accuracy

Dataset		nw	Epochs	s=0.5	s=0.8	s=0.9
UCI machine learning	DNA	936195	40	0.9533	0.9233	0.9755
	Mushroom	90562	20	1.0000	0.9890	0.7172
	Climate	76738	20	0.9714	0.9643	0.9214
MNIST		170298	40	0.9957	0.9953	0.9946
Fashion-MNIST		267514	50	0.9443	0.9391	0.9329
CIFAR-10		1154074	250	0.9423	0.9383	0.9310
Tiny ImageNet		4644120	250	0.6377	0.6295	0.6042

License

Apache-2.0