Describe the bug

when i qat resnet50 and yolov7 ,i find it evals quantized_model at the same time evals original_model,and it aims to reduce the gap of above of two; after the 10 epochs training, i got a low precision quantized_model and original_model.
here is the train log of resnet50

here is v7

so why does it happen? could someone illustrate the pipline of qat training? why does it training the twos at the same time?
thanks a lot
[here]

Post related information

here is my resnet50 config

here is yolov7

i train resnet use mmcls 1.0.0rc6 i train yolov7 use mmdet 3.0.0