Tickling-Vision-Models

activation maximized feature visualizations for top feature directions
Donor-Target influence (top5 logits)	ablation effectiveness per sample

core functionalities

collect activations from a selected set of layers passing images from a subset of classes

python -m scrips.collect_activaions

train decomposers on pixel / patch actviations and checkpoint weights

python -m scripts.train_decomposers

top direction feature visualization using sparse representation activation maximization

python -m scripts.paint_atlas

test causal impact feature patching with donor activations on model predictions

python -m scripts.patch_features

test logits shift on adversarial examples and directional adversarial residual ablation for recovery

python -m scripts.adversarial_attack_ablation

run all and make plots with:

./run.sh

summer fever headache (some may call it "Grad-CAM")