chunked/nltk_chunking/model_doc_owlvit/chunk_18.txt

The authors first train CLIP from scratch and fine-tune it end-to-end with the classification and box heads on standard detection datasets using a bipartite matching loss.