This repository contains the code for DeDUCE: Generating Counterfactual Explanations Efficiently by Benedikt Höltgen, Lisa Schut, Jan M. Brauner, Yarin Gal.
run_training.py trains ResNets, drawing on files in DDU.
Sanity checks for the DDU models are performed in nb_DDU_FashionMNIST.ipynb.
run_DeDUCE.py, run_JSMA.py, and run_REVISE.py generate counterfactuals, drawing on files in CE.
nb_AnoGAN_eval.ipynb is used for tuning the AnoGAN metric as well as computing scores, drawing on files in metrics.
Sanity checks are performed in nb_metrics_EMNIST.ipynb.
nb_tune_DeDUCE.ipynb and nb_tune_REVISE.ipynb are used for tuning the respective algorithm on the validation set examples in valset_batch.
nb_eval_testset.ipynb performs the testset evaluation, with files in _testset_results.
nb_eval_testset2-5.ipynb performs further evaluations, on additional testsets, with files in _testset_results*.