Is the backbone frozen in yolov3 training?
Created by: jcdubron
I download the pre-trained models from https://github.com/PaddlePaddle/models/blob/v1.4/PaddleCV/yolov3/weights/download.sh.
The backbone part in yolov3.tar.gz has almost the same weights in darknet53.tar.gz. So I wonder whether the backbone part is frozen during the yolov3 training process. If so, is it helpful to better converge or achieve higher accuracy?