Convolutional neural network optimization via channel reassessment attention module
The architecture of CRA module.
| model | top-1 error | top-5 error | FLOPs | params |
|---|---|---|---|---|
| ResNet-50 | 24.20 | 7.15 | 4.11G | 25.56M |
| CRA-ResNet-50 | 22.77 | 6.47 | 4.11G | 26.31M |
| ResNet-101 | 23.12 | 6.67 | 7.84G | 44.55M |
| CRA-ResNet-101 | 21.60 | 5.93 | 7.84G | 46.17M |
| ResNeXt-101 | 21.27 | 5.79 | 8.01G | 44.18M |
| CRA-ResNeXt-101 | 20.71 | 5.47 | 8.02G | 45.80M |