Module contrib.semseg.manipulation

Manipulate representations by increasing or decreasing the presence of a feature in a ViT activation, then use the linear probe for inference.

Record class-specific scores before and after manipulation to see that you can directly manipulate abilities to complete downstream tasks.

Functions

def main(cfg: Manipulation)

def manipulate(cfg: Manipulation,
sae: SparseAutoencoder,
acts_BWHD: jaxtyping.Float[Tensor, 'batch width height d_vit']) ‑> tuple[jaxtyping.Float[Tensor, 'batch width height d_vit'], jaxtyping.Float[Tensor, 'batch width height d_vit']]