Create a new Python 3.9 environment and install the packages in the requirements.txt file.
To fit a ComFe head to the Flowers-102 dataset using frozen DINOv2 embeddings, run
python main.py +run=base.yaml +R=comfe +model/dataset=torchvision_flowers +model/networks=dinov2_vits_14 seed=1
This should fit pretty quickly on a local GPU, and to visualize the results use
python main.py +run=base.yaml +R=comfe-view +model/dataset=torchvision_flowers +model/networks=dinov2_vits_14
Which will save the explanations for the test images in ../output/local/comfe-view_0/tensorboard/version_0/explanations.