Turn off Automatic Mixed Precision but maps are wrong in model summary

[P42ca1]

I turned off the Automatic Mixed Precision in train.py by setting "amp = False", since using FP16 in my model&dataset make loss becomes nan easily, but when the training process is finished, model summary in the output is wrong(map50=0.904 at 500th epoch, but shows map50=0.611 in the summary),does someone know how to fix this?