chunked/nltk_chunking/model_doc_detr/chunk_46.txt

During training, the authors of DETR did find it helpful to use auxiliary losses in the decoder, especially to help
  the model output the correct number of objects of each class.