基于灰色关联分析的卷积神经网络模型裁剪方法 
下载： 913次
Method of Convolutional Neural Network Model Pruning Based on Gray Correlation Analysis
1 江南大学轻工过程先进控制教育部重点实验室, 江苏 无锡 214122
2 无锡信捷电气股份有限公司, 江苏 无锡214122
图 & 表
图 1. 模型裁剪框架图
Fig. 1. Framework of model pruning
下载图片 查看原文
图 2. 卷积神经网络结构图
Fig. 2. Schematic of convolutional neutral network
下载图片 查看原文
图 3. 残差网络裁剪示意图
Fig. 3. Schematic of pruning of residual network
下载图片 查看原文
图 4. 各层裁剪量示意图
Fig. 4. Schematic of pruning quantity of each layer
下载图片 查看原文
图 5. 裁剪过程精度变化图
Fig. 5. Precision change in pruning process
下载图片 查看原文
图 6. 多种方法裁剪精度变化对比。(a) VGG-16各方法对比;(b) ResNet-50各方法对比;(c) AlexNet各方法对比
Fig. 6. Comparison of pruning accuracy changes of various methods. (a) Comparison of various methods on VGG-16; (b) comparison of various methods on ResNet-50; (c) comparison of various methods on AlexNet
下载图片 查看原文
表 1总体实验结果
Table1. Overall experimental results
| Model | Pruning ratio /% | Accuracy /% | Acceleration factoron 1080Ti | Acceleration factoron TX2 | Size /MB |
|---|
| VGG-16 | 0 | 93.1 | 1× | 1× | 43.5 | | VGG-16-pruned-A | 40 | 92.0 | 1.36× | 1.83 × | 17.6 | | VGG-16-pruned-B | 80 | 91.8 | 2.7× | 2.8× | 3.2 | | ResNet-50 | 0 | 88.4 | 1× | 1× | 74.3 | | ResNet-50-pruned-A | 40 | 85.4 | 1.46× | 1.6× | 43.8 | | ResNet-50-pruned-B | 80 | 79.3 | 3.4× | 1.9× | 8.9 | | AlexNet | 0 | 78.2 | 1× | 1× | 14.2 | | AlexNet-pruned-A | 40 | 71.5 | 1.6× | 1.9× | 6.4 | | AlexNet-pruned-B | 80 | 54.2 | 3.1× | 3.4 × | 1.8 |
|
查看原文
表 2AlexNet各点精度与加速效果
Table2. Accuracy and acceleration effect of AlexNet
| Pruningratio /% | Accuracy(Top-5) /% | Variation ofaccuracy /% | Accelerationfactor on 1080Ti | Accelerationfactor on TX2 | Time on 1080Tiand TX2 /ms | Size /MB |
|---|
| 0 | 78.2 | 0 | 1.00 | 1.00 | 32,74 | 14.2 | | 20 | 75.2 | -3.0 | 1.28 | 1.42 | 25,52 | 9.6 | | 40 | 71.5 | -6.7 | 1.60 | 1.90 | 20,39 | 6.4 | | 60 | 68.7 | -9.5 | 2.10 | 2.30 | 15,32 | 2.9 | | 80 | 54.2 | -24.0 | 3.10 | 3.40 | 10,21 | 1.8 |
|
查看原文
表 3ResNet-50各点精度与加速效果
Table3. Accuracy and acceleration effect of ResNet-50
| Pruningratio /% | Accuracy(Top-5) /% | Variation ofaccuracy /% | Accelerationfactor on 1080Ti | Accelerationfactor on TX2 | Time on 1080Tiand TX2 /ms | Size /MB |
|---|
| 0 | 88.4 | 0 | 1.00 | 1.0 | 79,248 | 74.3 | | 20 | 86.5 | -1.9 | 1.17 | 1.4 | 67,178 | 67.6 | | 40 | 85.4 | -3.0 | 1.46 | 1.6 | 54,154 | 43.8 | | 60 | 83.1 | -5.3 | 1.97 | 1.8 | 40,136 | 25.3 | | 80 | 79.3 | -9.1 | 3.40 | 1.9 | 23,128 | 8.9 |
|
查看原文
黄世青, 白瑞林, 覃高鄂. 基于灰色关联分析的卷积神经网络模型裁剪方法[J]. 激光与光电子学进展, 2020, 57(4): 041011. Shiqing Huang, Ruilin Bai, Gaoe Qin. Method of Convolutional Neural Network Model Pruning Based on Gray Correlation Analysis[J]. Laser & Optoelectronics Progress, 2020, 57(4): 041011.
PDF全文
