peal.adaptors.projection_adaptor¶
Adaptor that repairs a classifier by projecting concept directions out of its head.
The ProjectionAdaptor is a training-free repair: it loads a trained
predictor, takes the component directions of a sparse dictionary (SAE, MSAE,
SVD, …) fitted on the penultimate features, and registers a forward pre-hook
on the final nn.Linear layer that removes the subspace spanned by the
selected components from its input. Accuracy and group statistics on the chosen
data partition are printed before and after the projection so that the effect
of removing a (spurious) concept can be read off directly.
Functions
Modifies a torch.nn.Linear layer to project out specific components from the input before the standard forward pass. |
Classes
Training-free repair: project dictionary components out of the classifier head. |
|
Config of |