chunked/nltk_chunking/model_doc_lxmert/chunk_16.txt

Further, while the cross-modality encoder
  contains self-attention for each respective modality and cross-attention, only the cross attention is returned and
  both self attention outputs are disregarded.