However, recent works show the attention-based module in transformers can be replaced by spatial MLPs and the resulted models still perform quite well.