I found during the process of reading the code that when inputting images into the network and obtaining semantic and semantic features, the network was not set to evaluation mode（model.eval()）. May I ask why this is the case?