masked_loss and different GIoU coefficient

[hazby2002]

I want to know why the model calls masked_loss when calculating box loss and cycle loss. The last bounding box of a batch is almost ignored because of a low coefficient when calculating loss as a result.

https://github.com/NExT-ChatV/NExT-Chat/blob/main/mllm/models/nextchat/nextchat_base.py#L360

---

Besides, I also wonder why the coefficient of $GIoU(b, b_{gt})$ in the code is 0.2, which is different from the 0.8 in the paper.

https://github.com/NExT-ChatV/NExT-Chat/blob/main/mllm/models/nextchat/nextchat_base.py#L358