Reason: CPU training, does not support half precision number
Fix: change.half() to.float() in the detection module.

 
Reference data: https://blog.csdn.net/hxx123520/article/details/112562930