Doubts about the YOLOV5-P4 training process

[hhaAndroid]

❔Question

Thank you for open source such excellent code. I read the weights after training and found that the BN parameter is nn.BatchNorm2d(c2, momentum=0.03, eps=0.001), but the code is the default value nn.BatchNorm2d(c2), and the training description is python train.py --data coco.yaml --cfg yolov5s.yaml --weights '' --batch-size 64, He did not adopt evolutionary strategies. so this why?

[glenn-jocher]

@hhaAndroid https://github.com/ultralytics/yolov5/blob/03281f8c7613ad808d7d356f0195152b2d46ab99/utils/torch_utils.py#L152-L162

[hhaAndroid]

@glenn-jocher Oh, It's so magical, it will actually be in the parameter initialization.

[glenn-jocher]

@hhaAndroid yes we call this function when creating a model here: https://github.com/ultralytics/yolov5/blob/374957317a5469742b24291caa52dedfd9d31c99/models/yolo.py#L114-L118

[hhaAndroid]

@glenn-jocher Thank you