I found that when I export model SAST to predict image, that made difference predict bounding box between inference model and train model