search

nickgkan / butd_detr

Code for the ECCV22 paper "Bottom Up Top Down Detection Transformers for Language Grounding in Images and Point Clouds"
Other
74 stars 11 forks source link

Question about loss #48

Closed WeitaiKang closed 6 months ago

WeitaiKang commented 6 months ago

Hi authors, I found that in your loss function here, you sum up all the number of ground truth bboxes among all the gpus, without averaging this number by the number of gpus. It seems that it would lower down the loss value if we use more gpus to train right? Is it a right way to do this?

WeitaiKang commented 6 months ago

I make a mistake. It is the issue about the code (https://github.com/yanmin-wu/EDA/blob/master/models/losses.py), which is based on this code, but they remove this line