Table 14. CochResNet50 block components.
1. input (x) |
2. 1 × 1 Conv2d(inplanes, planes, stride = 1) |
3. BatchNorm2d(planes) |
4. ReLU |
5. 3 × 3 Conv2d (planes, planes, stride = 1) |
6. BatchNorm2d(planes) |
7. ReLU |
8. 1 × 1 Conv2d (planes, planes * expansion, stride = 1) |
9. BatchNorm2d(planes) |
10. Residual connection on x (if inplanes! = planes * expansion): 1 × 1 Conv2D (inplanes, planes * expansion, stride) |
11. Residual connection on x (if inplanes! = planes * expansion): BatchNorm2d(planes * expansion) |
12. Add output from [9] to output from [11] |
13. (Output) ReLU |
Multiple of these residual blocks (num_blocks) are stacked together to form a single ResNetBlock. The expansion factor was set to 4 for all layers (expansion = 4).