Learning by Self-Explaining

Repository for Learning by Self-Explaining paper.

This repo contains the code for both the CNN-LSX and NeSy-LSX experiments.

Concerning the VLM-LSX instantiation we refer to the following repo and corresponding paper: ILLUME.

Disclaimer: this repository is an agglomerate of multiple repositories and from several projects. It is therefore missing a global structure and might be buggy in terms of requirements etc (please don't hate me).

---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:

cd docker/
docker build -t pytorch-lsx -f Dockerfile .
docker run -it -v /pathtofolder:/workspace/repositories/learning_by_self-explaining -v /pathtodata:/workspace/datasets/ --name pytorch-lsx --entrypoint='/bin/bash' --runtime nvidia pytorch-lsx

---Parameter settings---

In all subdirectories we provide the folder parameters/ which contains argument files for different experimental settings. We thus refer to these for hyperparameter details as well as the existing example bash scripts.

---Running code---

We provide information on how to run the individual experiments in the README's of each subdirectory.