This repository contains all the experiments presented in the paper: "XAI-TRIS: Non-linear benchmarks to quantify ML explanation performance".
We use Pipfiles to create Python environments. Our main experiments use Captum, however inclusion of some methods (PatternNet, PatternAttribution, DTD) from the innvestigate library requires particular dependencies, such as Python=3.7, tensorflow==1.13.1/1.14, and keras==2.2.4. If you do not wish to include these methods, you are able to remove tensorflow and keras from the Pipfile and change any specific package requirements to * along with the Python version of choice (we have also tested this with Python 3.10).
To setup pipenv, run:
pipenv installand then
pipenv shellIn four steps we can reproduce the results: (i) generate tetromino data, (ii) train models (iii) apply XAI methods, (iv) run the evaluation steps and generate plots.
If one wishes to use the data from the ImageNet-1k, we sourced it from HuggingFace by making an account and agreeing to the license terms. Please respect these license terms when using this data. Download and extract images into a folder imagenet_images/ in the top level of this repository. We used the N=50,000 validation set due to the appropriate volume size for our datasets and benchmarks.
Once ImageNet data is downloaded and extracted above (if using it), configuration is set by the data/data_config.json file, with fields explained in data/README.md. Generate data via:
python -m data.generate_data One can also generate the original 8x8 data (without ImageNet backgrounds), which is quicker and less computationally demanding to run, and also produces interesting results. Configuration here is set by the data/data_config_8by8.json file. Generate data via:
python -m data.generate_data data_config_8by8Update the data_path parameter of the training/training_config.json with the path to the folder of freshly generated pickle files containing data, of the form artifacts/tetris/data/YYYY-MM-DD-HH-mm.
To train models for a particular scenario, background, and experiment number:
python -m training.train_models SCENARIO BACKGROUND EXP_NUM OUT_DIR where SCENARIO=[linear|multiplicative|translations_rotations|xor], BACKGROUND=[uncorrelated|correlated|imagenet], EXP_NUM=[0,...,N], and OUT_DIR=[str]. For example python -m training.train_models linear uncorrelated 0 neurips_training will train models for the linear scenario with uncorrelated (WHITE) background for the 0th experiment and output models and logs to the artifacts/tetris/training/neurips_training directory.
python -m training.train_models_keras SCENARIO BACKGROUND EXP_NUM OUT_DIR where input args are the same as specified above.
As you may have noticed, you will have to run this individually for all scenario-background pairs. You can also input regex into the SCENARIO and BACKGROUND input args, so running python -m training.train_models * * EXP_NUM OUT_DIR will train all scenario and background types for PyTorch.
One can use a somewhat approach like ./train_all.sh which simply runs all parameterizations one after another. We can also provide a script combining all parameterizations for each frameworks if needed.
Update the data_path, training_output, and num_experiments parameters of the xai/xai_config.json file with the same data_path as used in the training step, and the corresponding training output path of the form artifacts/tetris/training/OUT_DIR.
To calculate explanations for a particular scenario and background and across all experiments:
python -m xai.calculate_explanations SCENARIO BACKGROUND OUT_DIR where SCENARIO=[linear|multiplicative|translations_rotations|xor], BACKGROUND=[uncorrelated|correlated|imagenet], and OUT_DIR=[str] as in the training step. Note: as of 14/06/23, you can't use regex for SCENARIO and BACKGROUND as in the training step, as it is incompatible with the input to the next step. If/when changed, this README will be updated.
python -m xai.calculate_explanations_keras SCENARIO BACKGROUND OUT_DIR where input args are the same as specified above.
Update the parameters data_path, training_path, xai_path, and num_experiments of eval/eval_config.json, and specify the desired output directory here too, in the form ./artifacts/tetris/eval/OUT_DIR.
python -m eval.quantitative_metrics SCENARIO BACKGROUND where SCENARIO=[linear|multiplicative|translations_rotations|xor] and BACKGROUND=[uncorrelated|correlated|imagenet]. Note: as of 14/06/23, you can't use regex for SCENARIO and BACKGROUND as in the training step, as it is incompatible with the input to the next step. If/when changed, this README will be updated.
python -m eval.quantitative_metrics_keras SCENARIO BACKGROUND OUT_DIR where input args are the same as specified above.
With the same parameters of eval/eval_config.json as before, simply run:
python -m eval.plot_qualitativepython -m eval.plot_quantitativeIf you have any questions and/or issues with the above, please do not hesitate to contact the authors!