track_id bug with fp16

[yanghaibin-cool]

Hi,here.When targets is converted to FP16, the track_id will lose the precision, resulting in wrong labels for reid.
How to separate track_id annotations from variable targets. And set targets to torch.float16as current code, but keep track_id to be torch.float32.I tried to modify it, but it didn't work.
Looking forward to your update on this bug.Thank you.

[HanGuangXin]

I have already made a PR to ByteTrack and it is merged to master branch. Have you check it already?

ifzhang/ByteTrack#184
ifzhang/ByteTrack#182

And yes, the PR I made doesn't include operations in yolo_head.py, where reid loss is computed.

[HanGuangXin]

The changes in yolo_head.py should be 4 lines, more or less. I just have no time to verify the results in current version.

[yanghaibin-cool]

The changes in yolo_head.py should be 4 lines, more or less. I just have no time to verify the results in current version.

Thanks, I will try to modify the id_loss in yolo_head.py myself. Thanks again for open-sourcing this project, the combination of bytetrack_reid and other tracking strategies has been very helpful for my work.

[HanGuangXin]

Just like I mentioned in ifzhang/ByteTrack#184: "Hi, I'm thinking about separate track_id annotations from variable targets. And set targets to torch.float16as current code, but keep track_id to be torch.float32."

So the only thing you should do is eparate track_id annotations from variable targets, then pass it to yolo head. I think it is just a matter of some function interfaces.

If you finish it, a PR is welcome.

[HanGuangXin]

Or I can update it when I got time.

[yanghaibin-cool]

Or I can update it when I got time.

Could I change it like this directly? It can be trained like this, but I don't know if it's right or not.

[HanGuangXin]

Hi, I don't think cat 2 tensors with different dtype is a correct operation.