search

lilanxiao / Rotated_IoU

Differentiable IoU of rotated bounding boxes using Pytorch
MIT License
412 stars 62 forks source link

loss suddenly increase? #32

Closed jacklin602 closed 2 years ago

jacklin602 commented 2 years ago

Hi @lilanxiao, thanks for your great work! When I run the demo, I found that sometimes loss will suddenly increase and iou will suddenly drop. Is there any ideas about this problem? differentiableIOU_loss_problem

lilanxiao commented 2 years ago

hi, thank you for this issue! Yes, it also happens to me sometimes. I think it's due to my non-optimal network design and hyper-parameter choices (e.g. too large learning rate), which leads to unstable training. I made this demo to show that the IoU loss is differentiable and didn't pay much attention to details. But I don't think it's a big problem. I've used the IoU loss to train some 3D detectors. It worked well and didn't show any instability.

jacklin602 commented 2 years ago

Thanks for your reply! Just want to make sure that it's not a personal problem :) And thanks for your great work again!