Sum-of-Parts (SOP) Models: Faithful Attributions for Groups of Features

[Paper] [Blog]

Official implementation for "Sum-of-Parts Models: Faithful Attributions for Groups of Features".

Authors: Weiqiu You, Helen Qu, Marco Gatti, Bhuvnesh Jain, Eric Wong

Prerequisite

To set up the environment:

conda create -n sop python=3.10
conda activate sop
pip install -r requirements.txt

To do experiments on ImageNet first 10 classes, create a folder data/imagenet_m with subfolders data/imagenet_m/train and data/imagenet_m/val, download data from ImageNet and put the 10 classes of data in subfolders in these folders.

Usage

Training

To train SOP for 10 classes on ImageNet on the Huggingface's Vision Transformer google/vit-base-patch16-224, first download our model for the first 10 classes for ImageNet.

python scripts/run/train_imagenet_m.py

or notebook notebooks/train.ipynb.

Evaluation

To use the trained SOP wrapped model at inference time, checkout notebooks/eval.ipynb.