A Pruning Method with Adaptive Adjustment of Channel Pruning Rate

Convolutional neural networks have different channel pruning sensitivities for different layers, and the existing network pruning methods cannot perform adaptive pruning of channels within each layer according to the sensitivity, resulting in irreversible damage to the network during the pruning pro...

Full description

Saved in:
Bibliographic Details
Published in2023 IEEE 4th International Conference on Pattern Recognition and Machine Learning (PRML) pp. 618 - 623
Main Authors Zhao, Zhongjin, Liu, Haijun, Li, Zhiwei, Zhang, Chenxi, Ma, Tuo, Peng, Jie
Format Conference Proceeding
LanguageEnglish
Published IEEE 04.08.2023
Subjects
adaptive adjustment
Adaptive systems
channel pruning rate
iterative method
Iterative methods
loss constraint
Machine learning
Maintenance engineering
network pruning
Pattern recognition
Sensitivity
Weight measurement
Online AccessGet full text

Cover

More Information
Summary:Convolutional neural networks have different channel pruning sensitivities for different layers, and the existing network pruning methods cannot perform adaptive pruning of channels within each layer according to the sensitivity, resulting in irreversible damage to the network during the pruning process. To solve this problem, a network pruning method with adaptive adjustment of channel pruning rate is proposed in this paper, which disassembles the dynamic adjustment process of channel pruning rate in each layer into an iterative optimization problem and finally obtains a network structure with high compression rate and high accuracy. This method performs iterative pruning of the network under the supervision of loss constraint according to the preset pruning initial step for the benchmark model, and then chooses to reduce the pruning step or skip the pruning of this layer according to the accuracy loss of the network. Finally achieves the pruning of each layer in the network with Appropriate pruning rates after multiple iterations. Experimental results show that this method can maintain 93.64% recognition accuracy in CIFAR10-VGG16 pruning task with 87.45% reduction of parameter amount, which has higher compression rate and less accuracy loss compared with PEEC, FPGM, RFP and other pruning algorithms.
DOI:10.1109/PRML59573.2023.10348367