4. 轻量级神经网络算法目录
- 轻量级神经网络算法
4.1 各轻量级神经网络算法总结对比
4.2 SqueezeNet
4.3 DenseNet
4.4 Xception
4.5 MobileNet v1
4.6 IGCV
4.7 NASNet
4.8 CondenseNet
4.9 PNASNet
4.10 SENet
4.11 ShuffleNet v1
4.12 MobileNet v2
4.13 AmoebaNet
4.14 IGCV2
4.15 IGCV3
4.16 ShuffleNet v2
4.17 MnasNet
4.18 MobileNet v3
4.1 各轻量级神经网络算法对比
轻量级神经网络准确率、Params、MAdds、推理时间等对比,对比数据集:ImageNet 2012 classification dataset。
Date | Model | Detail | Top-1 Acc. (%) | Top-5 Acc. (%) | Params(M) | MAdds(M) | Infer-time(ms) |
2016.2 | SqueezeNet | 67.5 | 88.2 | 3.2 | 708 | ||
2016.8 | DenseNet | DenseNet(0.5) | 41.4 | 42 | 25 | ||
DenseNet(1.0) | 44.8 | 142 | 63 | ||||
DenseNet(1.5) | 60.1 | 295 | 103 | ||||
DenseNet(2.0) | 65.4 | 519 | 164 | ||||
2016.1 | Xception | Xception(0.5) | 55.1 | 40 | 19 | ||
Xception(1.0) | 65.9 | 145 | 51 | ||||
Xception(1.5) | 70.6 | 305 | 95 | ||||
Xception(2.0) | 72.4 | 525 | 149 | ||||
2017.4 | MobileNet v1 | MobileNet v1(0.25) | 50.6 | 0.5 | 41 | 27 | |
MobileNet v1(0.5) | 63.7 | 1.3 | 149 | 60 | |||
MobileNet v1(0.75) | 68.4 | 2.6 | 325 | 94 | |||
MobileNet v1(1.0) | 70.6 | 89.5 | 4.2 | 569 | 154 | ||
2017.7.10 | IGCV | ||||||
2017.7.21 | NASNet | NASNet-A | 74 | 91.3 | 5.3 | 564 | 183 |
2017.11 | CondenseNet | CondenseNet(G=C=4) | 71 | 90 | 2.9 | 274 | |
CondenseNet(G=C=8) | 73.8 | 91.7 | 4.8 | 529 | |||
2017.12 | PNASNet | PNASNet | 74.2 | 91.9 | 5.1 | 588 | |
2017.9 | SENet | ||||||
2017.12 | ShuffleNet v1 | ShuffleNet(0.5) | 56.8 | 38 | 18 | ||
ShuffleNet v1(1.0)-g=3 | 67.4 | 140 | 46 | ||||
ShuffleNet v1(1.5)-g=3 | 71.5 | - | 3.4 | 292 | 97 | ||
ShuffleNet v1(x2)-g=3 | 73.7 | - | 5.4 | 524 | 156 | ||
2018.1 | MobileNet v2 | MobileNet v2(0.35) | 60.8 | 1.6 | 59.2 | 16.6/19.6/13.9(Pixel*) | |
MobileNet v2(1.0) | 72 | 91 | 3.4 | 300 | 75(Pixel 1 Phone) | ||
MobileNet v2(1.4) | 74.7 | 92.5 | 6.9 | 585 | 143(Pixel 1 Phone) | ||
2018.2 | AmoebaNet | AmoebaNet-A | 74.5 | 92 | 5.1 | 555 | 190 |
2018.4 | IGCV2 | IGCV2-0.25 | 54.9 | 0.5 | 46 | 32 | |
IGCV2-0.5 | 65.5 | 1.3 | 156 | 65 | |||
IGCV2-1.0 | 70.7 | 4.1 | 564 | 204 | |||
2018.6 | IGCV3 | IGCV3-0.7 | 68.45 | 2.8 | 210 | 85 | |
IGCV3-1.0 | 72.2 | 3.5 | 318 | 159 | |||
IGCV3-1.4 | 74.55 | 7.2 | 610 | 222 | |||
2018.7 | ShuffleNet v2 | ShuffleNet v2(0.5) | 60.3 | 1.4 | 41 | 18 | |
ShuffleNet v2(1.0) | 69.4 | 2.3 | 146 | 41 | |||
ShuffleNet v2(1.5) | 72.6 | 3.5 | 299 | 85 | |||
ShuffleNet v2(x2) | 74.9 | 7.4 | 597 | 149 | |||
ShuffleNet v2(x2)-SE | 75.4 | 597 | 179 | ||||
2019.3 | MnasNet | MnasNet-Small | 64.9 | 1.9 | 65.1 | 20.3/24.2/17.2 | |
MnasNet-A1 | 75.2 | 92.5 | 3.9 | 312 | 78(Pixel 1 Phone) | ||
MnasNet-A2 | 75.6 | 92.7 | 4.8 | 340 | 84(Pixel 1 Phone) | ||
MnasNet-A3 | 76.7 | 93.3 | 5.2 | 403 | 103(Pixel 1 Phone) | ||
2019.5.6 | MobileNet v3 | MobileNet v3-Large(1.0) | 75.2 | 5.4 | 219 | 51/61/44(Pixel*) | |
MobileNet v3-Large(0.75) | 73.3 | 4 | 155 | 39/46/40(Pixel*) | |||
MobileNet v3-Small(1.0) | 67.4 | 2.5 | 56 | 15.8/19.4/14.4(Pixel*) | |||
MobileNet v3-Small(0.75) | 65.4 | 2 | 44 | 12.8/15.6/11.7(Pixel*) |
总结
以上就是关于轻量级神经网络算法的对比结果,点击Model列的算法可以详细了解各个算法。