During my experiment, I modified the network structure of the bbox head, but the features extracted by the backbone contained NaN values, leading to errors in the assigner. This issue didn't occur if I didn't make modifications.
I have two areas of confusion:
Why does modifying the bbox head affect the extraction of backbone features, and why does it only affect certain images?
I noticed in the source code of the rotated DETR model that the author debugged this issue in the forward_train function (outputting the image name if NaN is present). Have you encountered the same problem? Finally, how did you resolve it?
Thank you.
During my experiment, I modified the network structure of the bbox head, but the features extracted by the backbone contained NaN values, leading to errors in the assigner. This issue didn't occur if I didn't make modifications. I have two areas of confusion: