85. Learning Transferable Visual Models From Natural Language Supervision(2)