|
| Network | Top 1/% | Params (M) | MAdds (M) | CPU | Advantage |
|
| MobileNet v1 | 70.6 | 4.2 | 575 | 113 ms | Proposed depthwise separable convolution |
| MobileNet v2 | 72.0 | 3.4 | 300 | 75 ms | Proposed inverted residuals and linear bottlenecks |
| ShuffleNet(×2) [36] | 73.7 | 5.4 | 524 | - | Combined grouped convolution and channel shuffle |
| NasNet-A | 74.0 | 5.3 | 564 | 183 ms | Designed NasNet search space |
| Proxyless [37] | 74.6 | 4 | 320 | 156 ms | A new path pruning method was proposed, which reduced memory consumption |
| MobileNet v3 | 75.2 | 5.4 | 219 | 69 ms | Combined complementary search technology and introduced the h-swish activation function |
|
