Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Inquiry Regarding Loss Weights in Group DETR Implementation #38

Open
Anchor1566 opened this issue Dec 19, 2024 · 0 comments
Open

Inquiry Regarding Loss Weights in Group DETR Implementation #38

Anchor1566 opened this issue Dec 19, 2024 · 0 comments

Comments

@Anchor1566
Copy link

First of all, I would like to extend my sincere gratitude and respect for your implementation of Group DETR as well as the provision of additional useful code. Your contributions have been invaluable for our learning and research, and we truly appreciate the effort you've put into this project.

I am currently reviewing the Group DETR implementation, which is quite elegant and insightful. However, I have a question regarding the loss function used within your implementation. Specifically, I noticed that loss_giou, loss_bbox, and loss_class are utilized but it appears there is no explicit setting of weights for these three losses; all weights seem to be set to 1 by default. Is this correct?

In some other models, such as RTDETR, different weights are assigned to different types of losses (for example, loss_vfl: 1, loss_bbox: 5, loss_giou: 2). I am curious if there is any particular reason or consideration behind the decision not to assign specific weights to the losses in Group DETR? Are there any notable points I should be aware of when considering the balance between these losses?

Your guidance on this matter would be greatly appreciated, as it could provide important insights into how to properly adjust and understand the loss components in Group DETR.

Looking forward to your reply.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant