Very insightful work! I would like to ask, during training, the ADE dataset is trained on a single category, what about during inference?