Prompt Pre-Training with Twenty-Thousand Classes for Open-Vocabulary Visual Recognition