Should I implement Yolo2Output3DLayer by referring Yolo2OutputLayer or implement a SameDiff op to do Yolo2 operation for 3D input? Thanks.
@Booker I would use samediff + a samediff layer for that. Otherwise you have to hand derive the gradients for the backward op yourself. Try to use the samediff namespace ops like:
SameDiff sd = SameDiff.create(); sd.nn().conv2d(..) sd.image().
to set everything up. If you see anything missing please do let us know.