NeSyXIL (Neuro-Symbolic Explanatory Interactive Learning)

This is the official repository of the article: Right for the Right Concept: Revising Neuro-Symbolic Concepts by Interacting with their Explanations by Wolfgang Stammer, Patrick Schramowski, Kristian Kersting, published at CVPR 2021.

This repository contains all source code required to reproduce the experiments of the paper including the ColorMNIST experiments and running the CNN model.

In case you are only looking for the NeSy Concept Learner please visit the separate repository.

Included is a docker file as well as a modified version of captum (official repo) that we had modified by a single line such that the gradients of the explanations are kept. Simply running the docker file will automatically install the captum_xil version. (ref. line 105 in NeSyXIL/src/docker/captum_xil/captum/attr/_utils/gradient.py)

How to Run:

CLEVR-Hans

Dataset

Please visit the CLEVR-Hans repository for instrucitons on how to download the CLEVR-Hans dataset.

NeSy Concept Learner

We have moved the source code for the NeSy Concept Learner to a separate repository in case someone is only interested in the Concept Learner for their work. In order to reproduce our experiments please clone this repository first:

1. cd src/clevr_hans/nesy_xil/

2. git clone https://github.com/ml-research/NeSyConceptLearner.git

Docker

We have attached a Dockerfile to make reproduction easier. We further recommend to build your own docker-compose file based on the DockerFile. To run without a docker-compose file:

1. cd src/docker/

2. docker build -t nesy-xil -f Dockerfile .

3. docker run -it -v /pathto/NeSyXIL:/workspace/repositories/NeSyXIL -v /pathto/CLEVR-Hans3:/workspace/datasets/CLEVR-Hans3 --name nesy-xil --entrypoint='/bin/bash' --runtime nvidia nesy-xil

NeSy XIL experiments

We have included the trained NeSy Concept Learner files for seeds 0 (or 1 for the global rule). To reproduce the NeSy XIL results please run the desired script e.g. as

cd src/clevr_hans/nesy_xil/

./scripts/clevr-hans-concept-learner-xil_global.sh 0 100 /workspace/datasets/CLEVR-Hans3/

For training with the RRRLoss on CLEVR-Hans3 with GPU 0 and run number 100. This corresponds to the setting of our experiments on the global correction rule.

CNN experiments

We have included the trained CNN model files for seed 0. To reproduce the CNN experiments please runs the desired scripts e.g. as

cd src/clevr_hans/cnn/

./scripts/clevr-hans-resnet-xil.sh 0 0 /workspace/datasets/CLEVR-Hans3/

For training with the HINTLoss on CLEVR-Hans3 with GPU 0 and run number 0.

ColorMNIST

For running the ColorMNIST experiment with the provided dockerfile:

1. cd src/color_mnist/data/

2. python generate_color_mnist.py

3. cd ..

4. ./scripts/color_mnist_train.sh 0 for running the default experiment on GPU 0.

Citation

If you find this code useful in your research, please consider citing:

@InProceedings{Stammer_2021_CVPR, author = {Stammer, Wolfgang and Schramowski, Patrick and Kersting, Kristian}, title = {Right for the Right Concept: Revising Neuro-Symbolic Concepts by Interacting With Their Explanations}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, month = {June}, year = {2021}, pages = {3619-3629} } }