What does LIME really see in images?

Python code for the paper What does LIME really see in images? No installation required, the main requirements are LIME (tested with version 0.2.0.0) and Tensorflow (>=2.1.0).

General organization

The script shape_detector.py produces the plots for Figure 3, linear_model.py produces Figure 4. Other scripts have to run in a certain order (explained below). Consider running qualitative_results.py'' to produce figures for specific images. Depending on your hardware, some experiments may take time (especially computing empirical LIME explanations). You can modify the value of n_imagesaccordingly. All auxilliary functions are stored in theutils'' folder, while the models used for the CIFAR-10 experiments are saved as h5 files in the ``models'' folder.

Experiments with CIFAR-10 images

• run train_model.py to train the models, or directly use the provided h5 files
• run compute_empirical.py to get empirical LIME explanations
• run compute_approx.py to get approximated explanations
• run compare_exp.py to compare explanations

Experiments with ILSVRC2017 images

• download the archive http://image-net.org/image/ILSVRC2017/ILSVRC2017_DET_test_new.tar.gz
• extract the archive and modify the data_path variable in the scripts accordingly
• run compute_empirical.py to get empirical LIME explanations
• run compute_approx.py to get approximated explanations
• run compare_exp.py to compare explanations

Citing this work

If you use this code please cite

@InProceedings{garreau2021what,
  title = 	 {What does LIME really see in images?},
  author =       {Garreau, Damien and Mardaoui, Dina},
  booktitle = 	 {Proceedings of the 38th International Conference on Machine Learning},
  pages = 	 {3620--3629},
  year = 	 {2021},
  editor = 	 {Meila, Marina and Zhang, Tong},
  volume = 	 {139},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {18--24 Jul},
  publisher =    {PMLR}
}