Discriminative Feature Attributions: A Bridge between Post Hoc Explainability and Inherent Interpretability

The script for the full pipeline is at script.sh

To set up datasets:

Hard MNIST: Clone https://github.com/jayaneetha/colorized-MNIST, then run python hard_mnist.py

Chest X-ray: Download from https://www.kaggle.com/datasets/paulti/chest-xray-images

CelebA: Download from https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html or https://www.kaggle.com/datasets/jessicali9530/celeba-dataset

To train baseline models (f_b):

python mnist_dataset.py
python xray_dataset.py
python celeba_dataset.py

DiET:

mkdir mnist_ups4_outdir/
python distillation.py -dataset mnist -out mnist_ups4_outdir/ -ups 4 --model_path trained_models/hard_mnist_rn34.pth -lr 2000
python inference_distill.py -dataset mnist -out mnist_ups4_outdir/ -ups 4 --model_path mnist_ups4_outdir/fs_1.pth -lr 2000

Evaluation:

Pixel Perturbation:

python pixel_perturbation.py -dataset mnist --model_path trained_models/hard_mnist_rn34.pth -ups 4 --mask_path mnist_ups4_outdir/ --mask_num 1 -p 10 20 50 100
python pixel_perturbation_ours.py -dataset mnist --model_path mnist_ups4_outdir/fs_1.pth -ups 4 --mask_path mnist_ups4_outdir/ --mask_num 1 -p 10 20 50 100

IOU:

python iou.py --model_path trained_models/hard_mnist_rn34.pth -ups 4 --mask_path mnist_ups4_outdir/ --mask_num 1

Model Faithfulness:

Results reported during distillation (in distillation.py)

Packages required:

• torch
• torchvision
• numpy
• pillow
• pandas
• https://github.com/idiap/fullgrad-saliency