search

soniajoseph / ViT-Prisma

ViT Prisma is a mechanistic interpretability library for Vision Transformers (ViTs).
Other
165 stars 18 forks source link

Folding fix #73

Closed themachinefan closed 7 months ago

themachinefan commented 7 months ago

-Fix folding to include the final layer (needed for https://github.com/soniajoseph/ViT-Prisma/issues/72).

-Also fixed import errors from renaming

Aside: "head" is a confusing name for the final layer since it conflicts with the attention heads. We should probably rename it.

soniajoseph commented 7 months ago

Great catch, thank you.