reg_loss always 1.0 on custom dataset

tianzhi0549 / FCOS

FCOS: Fully Convolutional One-Stage Object Detection (ICCV'19)

https://arxiv.org/abs/1904.01355

Other

3.27k stars 629 forks source link

reg_loss always 1.0 on custom dataset #310

Open brucetusec opened 3 years ago

brucetusec commented 3 years ago

I tried training this model on my own dataset,

cls_loss converges fine, but reg_loss always 1.0, and pred_left, pred_right, pred_top, pred_bottom become all zero after 5~10 iteration.

Would you please give me some suggestions?

justustulo commented 3 years ago

@brucetusec I know I'm kind of late, but I'm having the same issue as well with the reg_loss stuck around 1 on VOC. Still don't have a solution, were you able to figure it out?

brucetusec commented 3 years ago

use torch.exp for regression output works in my case. I am using linear IOU loss now.