Pretrained weights for lucid.models.clip_vit_base_16.
The same tower shape as ViT-B/32 with 16-pixel patches, so it reads 196 patch tokens where the other reads 49 — four times the attention for the same 151M-scale budget, and the paper's better-scoring base model because of it.
Examples
>>> from lucid.models.multimodal.clip._weights import CLIPViTBase16Weights
>>> list(CLIPViTBase16Weights.__members__)
['OPENAI_WIT400M', 'DEFAULT']
DEFAULT is an alias rather than a fourth entry, so a bare
pretrained=True and the tag it resolves to cannot drift apart.
>>> CLIPViTBase16Weights.DEFAULT is CLIPViTBase16Weights.OPENAI_WIT400M
True
>>> CLIPViTBase16Weights.OPENAI_WIT400M.num_classes
0