search

soniajoseph / ViT-Prisma

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

Fix the ImageNet checkpoints so they properly load from Huggingface #60

Closed soniajoseph closed 7 months ago

soniajoseph commented 8 months ago

Right now, the toy ImageNet checkpoints need ImagenetConfig in the same folder to load. This is not expected behavior.

One possible fix is loading all the checkpoints again, saving them again, ensuring that they can load without ImagenetConfig in he folder, and then reuploading them to Huggingface.

@YashVadi, someone else also welcome to pick this up