Vibashan/irg-sfda

loss_bbox_reg = 0.0

Tajamul21 opened this issue · 5 comments

Tajamul21 commented

Hii,

=I am getting loss_bbox_reg - 0 for all, anything can help me in this regard. Thanks

image

Vibashan commented

Hi @Tajamul21 ,

I think there is some issue with dataloading. Maybe can you look into it?

EST09 commented

Hi,

@Tajamul21 - I get the same thing - did you ever manage to solve it?

Thank you!
Best wishes,
Emily

xcx121 commented

@Vibashan I also get the same thing -
I trained on a custom dataset,Could you please provide me with some solutions?Thank you!
Best wishes

Vibashan commented

Hi @Tajamul21 ,

Thank you for your interest in my work; sorry for the delayed response.

Could you please check the following:

  1. Ensure you're running the default COCO experiment and verify if you still encounter loss_bbox_reg = 0.
  2. Check the function _dense_box_regression_loss to ensure the values are being passed correctly. You can refer to this link for the relevant code:
    detectron2/box_regression.py (line 310)
Vibashan commented

Hi @xcx121

To train a model on a new dataset with Detectron2:

Ensure your dataset is in COCO format. Refer to the Detectron2 Dataset Guide

Configure and train your model. Full details in the Detectron2 Training Tutorial