mmrotate
mmrotate copied to clipboard
Why rotatedRetinaNet can't use the RotatedIoULoss?
I want to replace L1Loss with RotatedIoULoss during rotatedRetinaNet training, but when I train to the 10th epoch, loss_bbox becomes Nan. As shown in the following figure.

reg_decoded_bbox=True,
should be added before
loss_bbox=dict(
it is usefull,thank you!