search

Zzh-tju / DIoU-pytorch-detectron

Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression (AAAI 2020)
GNU General Public License v3.0
88 stars 16 forks source link

Got Nan #19

Open Sebastian-X opened 4 years ago

Sebastian-X commented 4 years ago

Hello, thanks for your code. Refering to your code, I reimplemented CIoU loss in my faster-rcnn framework, but got Nan gradients in the first iteration of training. Tracking the Nan values, I found this problem will occur when the RPN module predicts inappropriate rois, e.g. [200, 599, 300, 599] which has the same y1, y2 values. In this case, the gradients of corresponding prediction boxes will be Nan. Have you ever encountered this kind of problem before?

Sebastian-X commented 4 years ago

I understand that you use a mask to avoid computing IoUs for these inappropriate rois while computing CIoU, but it couldn't prevent the Pytorch autograde function from backpropagating through these rois.

Sebastian-X commented 4 years ago

In fact, it seems that this problem will occur when the IoU of certain bounding box is 0.