Difference in Yolov5 stock config and FairMOT config.

[skyimager]

Hi, I am trying to train FairMOT with yolov5s as backbone. I am using my custom yolov5s model trained from the repo: https://github.com/ultralytics/yolov5

When starting the training I get the below message:

Creating model.....
/src/lib/models/networks/config/yolov5s.yaml
Transferred 283/334 items from ../models/yolov5s.pt
Starting training...
loaded ../models/fairmot_lite.pth, epoch 30

When it is trying to load the yolov5 model, only certain layer weights are transferred (283/334).

Also, I see difference in configuration between the stock yolo and yolo provided in FairMOT repo: Stock yolov5s config: https://github.com/ultralytics/yolov5/blob/master/models/yolov5s.yaml

FairMOT yolov5s config: https://github.com/ifzhang/FairMOT/blob/15d2694f2737b869aaebc259bdea09ef73e5dd27/src/lib/models/networks/config/yolov5s.yaml

Is it suppose to be this way or am I missing something here?

[zengjie617789]

because the least down-ratio of yolov5 is 8 while the fairmot is 4, so it should modified the structures to adapt .