Adaptive Dynamic Filter Pruning Approach Based on Deep Learning

Jinghui Chu; Meng Li; Lü Wei

doi:10.3788/LOP202259.2415003

Laser & Optoelectronics Progress, Volume. 59, Issue 24, 2415003(2022)

Adaptive Dynamic Filter Pruning Approach Based on Deep Learning

Jinghui Chu, Meng Li, and Lü Wei^*

Author Affiliations

School of Electrical and Information Engineering, Tianjin University, 300072, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(9)

Fig. 1. Structure of filter pruning approach

Download full size

Fig. 2. Combination of AWGM and basic blocks in various networks

Download full size

Table 1. CIFAR-10 dataset
View table
Table 1. CIFAR-10 dataset
Training Test
50000 10000

Table 2. AUC driving behavior dataset
View table
Table 2. AUC driving behavior dataset
Category Training Test
Drive Safe 2764 922
Drinking 1209 403
Talk Right 917 306
Talk Left 1020 341
Text Right 1480 494
Text Left 975 326
Adjust Radio 915 305
Hair & Makeup 901 301
Reach Behind 869 290
Talk to Passenger 1927 643

Table 3. Performance comparison between proposed method and other methods on VGG16

View table

Table 3. Performance comparison between proposed method and other methods on VGG16

Method	Accuracy /%	FLOPs /MB	Parameters /MB
VGG-16^［6］	93.96	313.73	14.98
Li^［20］	93.40	206.00	5.40
HRank^［18］	93.43	145.61	2.51
Proposed method（0.75）	93.65	135.79	2.11
HRank^［18］	92.34	108.61	2.64
Proposed method（0.85）	93.14	94.35	1.53

Table 4. Performance comparison between proposed method and other methods on ResNet-56
View table
Table 4. Performance comparison between proposed method and other methods on ResNet-56
Method Accuracy /% FLOPs /MB Parameters /MB
ResNet-56 93.26 125.49 0.85
He^［11］ 90.80 62.00
GAL^［12］ 90.36 49.99 0.29
HRank^［18］ 90.72 32.52 0.27
Proposed method 90.84 29.83 0.18

Table 5. Performance comparison between proposed method and other methods on GoogLeNet
View table
Table 5. Performance comparison between proposed method and other methods on GoogLeNet
Method Accuracy /% FLOPs /MB Parameters /MB
GoogLeNet 95.05 1.52 6.15
GAL^［12］ 93.93 0.94 3.12
Li^［20］ 94.54 1.02 3.51
HRank^［18］ 94.53 0.69 2.74
Proposed method 94.84 0.62 2.94

Table 6. Performance comparison between proposed method and other methods on DenseNet-40
View table
Table 6. Performance comparison between proposed method and other methods on DenseNet-40
Method Accuracy /% FLOPs /MB Parameters /MB
DenseNet-40 94.81 282 1.04
Zhao^［28］ 93.16 156 0.42
GAL^［12］ 93.53 128.11 0.45
HRank^［18］ 93.68 110.15 0.48
Proposed method 93.1 100 0.28

Table 7. Experimental results on VGG16, VGG-19, and ResNet-56
View table
Table 7. Experimental results on VGG16, VGG-19, and ResNet-56
Method Accuracy /% FLOPs /MB Parameters /MB
VGG-16 95.01 15.41 40.43
Proposed method 94.87 4.72 14.18
VGG-19 95.08 19.58 45.74
Proposed method 95.03 4.62 28.10
ResNet-56 93.47 6.2 0.85
Proposed method 93.21 2.36 0.37

Tools

Get Citation

Copy Citation Text

Jinghui Chu, Meng Li, Lü Wei. Adaptive Dynamic Filter Pruning Approach Based on Deep Learning[J]. Laser & Optoelectronics Progress, 2022, 59(24): 2415003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Sep. 6, 2021

Accepted: Oct. 27, 2021

Published Online: Nov. 28, 2022

The Author Email: Wei Lü (luwei@tju.edu.cn)

DOI:10.3788/LOP202259.2415003

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. CIFAR-10 dataset

Table 1. CIFAR-10 dataset

Table 2. AUC driving behavior dataset

Table 2. AUC driving behavior dataset

Table 3. Performance comparison between proposed method and other methods on VGG16

Table 3. Performance comparison between proposed method and other methods on VGG16

Table 4. Performance comparison between proposed method and other methods on ResNet-56

Table 4. Performance comparison between proposed method and other methods on ResNet-56

Table 5. Performance comparison between proposed method and other methods on GoogLeNet

Table 5. Performance comparison between proposed method and other methods on GoogLeNet

Table 6. Performance comparison between proposed method and other methods on DenseNet-40

Table 6. Performance comparison between proposed method and other methods on DenseNet-40

Table 7. Experimental results on VGG16, VGG-19, and ResNet-56

Table 7. Experimental results on VGG16, VGG-19, and ResNet-56