Strange results on QAT YOLOv5

Hi, I quantized YOLOv5 by adding QuantStub and DeQuantStubs to all the ConvBnReLU2d in the network, exported to ONNX and folded the QDQ nodes into QLinearConv.
https://drive.google.com/file/d/11-7Mnss4g6_RJ2CizITIWFo9bGPDNkO_/view?usp=sharing

However, during inference, this is the result for the whole image:
image