Using vision/references/detection/train.py with augmentation training mAP is abnormal

The offical vision/references/detection/train.py using COCO api to calculate mAP.
When it not using data augmentation that train mAP is normal,
Training one epoch like that:
train mAP is 0.25383283202173673
validation mAP is 0.30275321883612916

However, I turned RandomHorizontalFlip on and probability setting as 0.5.
Training one epoch like that:
train mAP is 0.03392322767009024
validation mAP is 0.18803215716046826

Futhermore, I turned RandomHorizontalFlip on and probability setting as 1.0.
Training one epoch like that:
train mAP is 0.24607017764763847
validation mAP is 0.24640787952490323

I thought RandomHorizontalFlip with Annotation transfer is nothing wrong. What others problem do I meet?