@themachinefan adapted the video vision transformer VivitForVideoClassification to Prisma from transformers (thank you!), but it's not yet clear if the standard mech interp techniques apply or break.
Try running direct logit attribution, etc on the video ViT. Put your results in a jupyter notebook as research code.
@themachinefan adapted the video vision transformer
VivitForVideoClassification
to Prisma from transformers (thank you!), but it's not yet clear if the standard mech interp techniques apply or break.Try running direct logit attribution, etc on the video ViT. Put your results in a jupyter notebook as research code.