Created by: QiJune

Dropout and batch norm operator do different between training and testing mode. We should handle these operators when users inference data after training.