But then again. The vanilla Transformer model -- it does not have a sense of hierarchical representations at least for vision modality.