arrow
返回

Multidimensional Dynamic Pruning: Exploring Spatial and Channel Fuzzy Sparsity

delete2024-09-01
delete1
PRE
AI
邵明文 封面图
邵明文 (Mingwen Shao) *
J
Jiandong Kuang
C
Chao Wang *
左
左旺孟 (Wangmeng Zuo)
G
Guoyin Wang
DOI:10.1109/TFUZZ.2024.3363220delete
delete原文链接
delete原文求助
delete分享
delete收藏
摘要

摘要

En 中文
Dynamic pruning is an effective model compression method to reduce the computational cost of networks. However, existing dynamic pruning methods are limited to pruning along a single dimension (channel, spatial, or depth), which cannot maximally excavate the redundancy of the network. Meanwhile, most of the current state-of-the-arts usually implement dynamic pruning via masked-out partial channels and pixels for training while failing to accelerate the inference speed. To tackle these limitations, we propose a novel fuzzy-based multidimensional dynamic pruning paradigm to dynamically compress neural networks along both the channel and spatial dimensions. Specifically, we design a multidimensional fuzzy-mask block to simultaneously learn which spatial positions or channels are redundant and need to be pruned. Then, the Gumbel-Softmax trick combined with a sparsity loss is introduced to train these mask modules in an end-to-end manner. During the testing stage, we convert features and convolution kernels into two matrices, respectively, and then implement sparse convolution through matrix multiplication to accelerate the network inference. Extensive experiments demonstrate that our method outperforms existing methods in terms of accuracy and computational cost. For instance, on the CIFAR-10 dataset, our method prunes 68% FLOPs of ResNet-56 with only a 0.07% Top-1 accuracy drop.
Keyword:
Computational complexity
convolutional neural networks
deep learning
image classification
neural network compression
Computational complexity
convolutional neural networks
deep learning
image classification
neural network compression

期刊

IEEE Transactions on Fuzzy Systems 封面图
IEEE Transactions on Fuzzy Systems
IF:
11.9
论文数:
5.0K
被引数:
2.9W

机构

H
harbin institute of technology
学者数:
8.0W
论文数: 6.6W
被引数: 66
C
chongqing university of posts & telecommunications
学者数:
6.7K
论文数: 5.3K
被引数: 5
U
university of technology sydney
学者数:
1.6W
论文数: 2.0W
被引数: 25
C
china university of petroleum
学者数:
4.1W
论文数: 2.7W
被引数: 30
学者 查看更多机构
引用论文

引用论文

Calcium-modulating cyclophilin ligand regulates membrane trafficking of postsynaptic GABAA receptors
err2008-06-01
err0
errOAAI
errXu Yuan; Jun Yao; David Norris; David D. Tran; Richard J. Bram; Gong Chen; Bernhard Luscher
err分享
err收藏
Isolation of Melanosomes
err2005-04-01
err0
PREAI
errHidenori Watabe; Tsuneto Kushimoto; Julio C. Valencia; Vincent J. Hearing
err分享
err收藏
Major Adverse Limb Events and Mortality in Patients With Peripheral Artery Disease
err2018-05-01
err0
errOAAI
errSonia S. Anand; Francois Caron; John W. Eikelboom; Jackie Bosch; Leanne Dyal; Victor Aboyans; Maria Teresa Abola; Kelley R.H. Branch; Katalin Keltai; Deepak L. Bhatt; Peter Verhamme; Keith A.A. Fox; Nancy Cook-Bruns; Vivian Lanius; Stuart J. Connolly; Salim Yusuf
err分享
err收藏
err分享
err收藏
Dopamine-Induced Apoptosis of Lactotropes Is Mediated by the Short Isoform of D2 Receptor多巴胺诱导的催乳激素凋亡是由D2受体的短同工型介导的
err2011-03-25
err0
errOAAI
errDaniela Betiana Radl; Jimena Ferraris; Valeria Boti; Adriana Seilicovich; Dipak Kumar Sarkar; Daniel Pisera
err分享
err收藏
Targeting the EWS–FLI1 transcription factor in Ewing sarcoma
err2015-03-26
err0
PREAI
errR. Tancredi; A. Zambelli; G. A. DaPrada; V. Fregoni; L. Pavesi; A. Riccardi; S. Burdach; P. J. Grohar; M. D’Incalci
err分享
err收藏
学者 查看更多内容