peal.adaptors.clarc¶

Class Artifact Compensation (ClArC) adaptors for removing known shortcuts.

ClArC (Anders et al.) corrects a classifier that relies on an artifact (a spurious “confounder” feature) by estimating a Concept Activation Vector (CAV) for the artifact from group-annotated activations and then either

projecting the artifact direction out of the activations at a chosen layer (PClArC, optionally followed by fine-tuning the downstream head), or
fine-tuning the downstream head with a right-for-the-right-reasons penalty that pushes the input gradient orthogonal to the CAV (RRClArC).

Both are PEAL adaptors driven by a ClArCConfig: they load a trained predictor, sweep layer_index x correction_strength, evaluate group accuracies on the validation (and optionally an unpoisoned test) split, write one csv per evaluation dataset plus best_model_result.txt into base_dir and save every corrected model as *.cpl. Module-level helpers split a model into feature extractor / head and compute pattern or SVM CAVs.

Module attributes

CavCache

Memo of the CAV estimated for one layer: layer (int), cav, annotations and activations (tensors from ClArC.get_annotations_and_activations()).

Functions

`calculate_cav`	Dispatch to `calculate_pcav()` or `calculate_svm_cav()`.
`calculate_pcav`	Pattern CAV: covariance of activations with the +-1 artifact label.
`calculate_svm_cav`	CAV from the normal of a balanced `LinearSVC` (C=1, up to 10000 iters).
`extract_all_children`	Flatten nested `nn.Sequential` containers into a list of leaf children.
`get_layer_name`	Return the dotted name of the child at 1-based `layer_index`.
`split_model`	Split a model into `Sequential` feature extractor and downstream head.

Classes

`CavCache`	Memo of the CAV estimated for one layer: `layer` (int), `cav`, `annotations` and `activations` (tensors from `ClArC.get_annotations_and_activations()`).
`CavProjection`	Layer that projects the CAV direction out of its input.
`ClArC`	Base class of the ClArC adaptors: data setup, sweep loop and evaluation.
`ClArCConfig`	Configuration shared by the ClArC adaptors.
`PClArC`	Projective ClArC: remove the CAV direction from activations at a layer.
`PClArCConfig`	Config of `PClArC`.
`RRClArC`	Right-for-the-right-reasons ClArC.
`RRClArCConfig`	Config of `RRClArC`.
`SimpleProjection`	Layer that subtracts the mean artifact minus non-artifact activation.