• Laser & Optoelectronics Progress
  • Vol. 59, Issue 24, 2415003 (2022)
Jinghui Chu, Meng Li, and Lü Wei*
Author Affiliations
  • School of Electrical and Information Engineering, Tianjin University, 300072, China
  • show less
    DOI: 10.3788/LOP202259.2415003 Cite this Article Set citation alerts
    Jinghui Chu, Meng Li, Lü Wei. Adaptive Dynamic Filter Pruning Approach Based on Deep Learning[J]. Laser & Optoelectronics Progress, 2022, 59(24): 2415003 Copy Citation Text show less
    Structure of filter pruning approach
    Fig. 1. Structure of filter pruning approach
    Combination of AWGM and basic blocks in various networks
    Fig. 2. Combination of AWGM and basic blocks in various networks
    TrainingTest
    5000010000
    Table 1. CIFAR-10 dataset
    CategoryTrainingTest
    Drive Safe2764922
    Drinking1209403
    Talk Right917306
    Talk Left1020341
    Text Right1480494
    Text Left975326
    Adjust Radio915305
    Hair & Makeup901301
    Reach Behind869290
    Talk to Passenger1927643
    Table 2. AUC driving behavior dataset
    MethodAccuracy /%FLOPs /MBParameters /MB
    VGG-16693.96313.7314.98
    Li2093.40206.005.40
    HRank1893.43145.612.51
    Proposed method(0.75)93.65135.792.11
    HRank1892.34108.612.64
    Proposed method(0.85)93.1494.351.53
    Table 3. Performance comparison between proposed method and other methods on VGG16
    MethodAccuracy /%FLOPs /MBParameters /MB
    ResNet-5693.26125.490.85
    He1190.8062.00
    GAL1290.3649.990.29
    HRank1890.7232.520.27
    Proposed method90.8429.830.18
    Table 4. Performance comparison between proposed method and other methods on ResNet-56
    MethodAccuracy /%FLOPs /MBParameters /MB
    GoogLeNet95.051.526.15
    GAL1293.930.943.12
    Li2094.541.023.51
    HRank1894.530.692.74
    Proposed method94.840.622.94
    Table 5. Performance comparison between proposed method and other methods on GoogLeNet
    MethodAccuracy /%FLOPs /MBParameters /MB
    DenseNet-4094.812821.04
    Zhao2893.161560.42
    GAL1293.53128.110.45
    HRank1893.68110.150.48
    Proposed method93.11000.28
    Table 6. Performance comparison between proposed method and other methods on DenseNet-40
    MethodAccuracy /%FLOPs /MBParameters /MB
    VGG-1695.0115.4140.43
    Proposed method94.874.7214.18
    VGG-1995.0819.5845.74
    Proposed method95.034.6228.10
    ResNet-5693.476.20.85
    Proposed method93.212.360.37
    Table 7. Experimental results on VGG16, VGG-19, and ResNet-56