Share Email Print
cover

Proceedings Paper • new

A deep learning model compression algorithm based on optimal clustering
Author(s): Qirui Wu; Wenzhen Li; Xu Lu; Hailong Zhang; Hanwu Luo; Cheng Lei
Format Member Price Non-Member Price
PDF $14.40 $18.00
cover GOOD NEWS! Your organization subscribes to the SPIE Digital Library. You may be able to download this paper for free. Check Access

Paper Abstract

Nowadays, the models of deep learning are increasingly used in various industrial applications. However, the storage space of the models is too large, which makes it quite difficult to apply to mobile devices. In order to solve this problem, a model compression algorithm based on optimal clustering is proposed in this paper. Firstly, the model parameters of fully connected layer in deep convolutional neural network are clustered according to the best clustering method. Then the cluster center of the parameters is selected as the representative of the original parameter matrix. At last, the parameters of the cluster center are transformed differently in the forward calculation of the model to achieve the effect of compressing the parameters of the model and ensured the accuracy of the model. The compression algorithm proposed here is compared with other model compression algorithms in several common deep learning models such as Alexnet, VGG16 and so on. The results show that the algorithm proposed in this paper can compress the memory of the model greatly and improve the accuracy.

Paper Details

Date Published: 6 May 2019
PDF: 10 pages
Proc. SPIE 11069, Tenth International Conference on Graphics and Image Processing (ICGIP 2018), 1106904 (6 May 2019); doi: 10.1117/12.2524450
Show Author Affiliations
Qirui Wu, Wuhan Nari Ltd. Liability Co. (China)
Wenzhen Li, State Grid East Inner Mongolia Electric Power Supply Co., Ltd. (China)
Xu Lu, State Grid East Inner Mongolia Electric Power Supply Co., Ltd. (China)
Hailong Zhang, Wuhan Nari Ltd. Liability Co. (China)
Hanwu Luo, State Grid East Inner Mongolia Electric Power Supply Co., Ltd. (China)
Cheng Lei, Wuhan Sanjiang China Electronics Technology Co., Ltd. (China)


Published in SPIE Proceedings Vol. 11069:
Tenth International Conference on Graphics and Image Processing (ICGIP 2018)
Chunming Li; Hui Yu; Zhigeng Pan; Yifei Pu, Editor(s)

© SPIE. Terms of Use
Back to Top