606 B
606 B
!
btw, i suspect that the coordconv at the first layer might not be necessary (coord0) at the first layer the CNN just does low level feature dedection (detecting edges and so on) the coordconv are only helpful for higher level analysis deeper in the model at least that's my theory also, do i see correctly that after the first layer, there's 1 maxpool that makes it 4x smaller? from 2048x1024 to 512x256? or does the first layer use stride 2, and the maxpool only makes it 2x smaller? overall the architecture looks good 👍
- Comments Floris