A bug with a nan value when model,img half

[YoungjaeDev]

I try to save_txt to make a mot15 det file, but there is a problem that model can't detect any of it
I tested it with the command below

python classy_track.py --source MOT15/train/ADL-Rundle-6/img1/000001.jpg --img-size 640 --save-img

The pred output is as follows:

My pip list

torch                  1.8.2+cu111
torchaudio             0.8.2
torchvision            0.9.2+cu111

[YoungjaeDev]

@tensorturtle
Oh, the supplementary explanation is missing, the first result is forced to half=false in the code, and the second result is the result of turning the original code