The code shows it loads the visual encoder from a CLIP model (clip-vit-b16.pth). I did not find anything mentioned where it comes from. I tried to load clip-vitb16 from OpenAI huggingface, but it has unmatched keys when loading. Is OpenAI's CLIP the required or you have your own trained CLIP?