LIMEcraft: Handcrafted superpixel selectionand inspection for Visual eXplanations
The LIMEcraft algorithm is an explanatory method based on image perturbations. Its prototype was the LIME algorithm - but not without significant drawbacks, especially in the context of difficult cases and medical imaging. The superiority of the LIMEcraft algorithm is primarily the ability to freely select the areas that we want to analyze, as well as the aforementioned perturbations, thanks to which we can compare with the original image, and thus better understand which features have a significant impact on the prediction of the model.
The interface is displayed in the browser, where the user can upload the image and the superpixel mask from their own computer. They can also manually mark interesting areas in the image with the help of a computer mouse. Then, they choose the number of superpixels the areas selected by them are divided into. The same procedure is also applied to the areas uploaded in the mask and, independently, to the areas outside. Then, they just need to confirm their actions and wait for the result of the LIMEcraft algorithm. The user receives the model prediction results expressed as the percentage for both the class originally predicted and the class after image editing.
- upload image
- upload mask
- manually select mask
- change number of superpixels inside and outside the mask
- show how the prediction changed
- change color - RGB
- change shape - power expansion
- rotate - degrees
- shift - left/right and down/up
- remove object by shifting it
- generate report
git clone https://github.com/MI2DataLab/LIMEcraft.git
cd LIMEcraft
virtualenv venv
source venv/bin/activate
python -m ipykernel install --user --name=venv
pip install -r requirements.txt
In case of problems with library versions, try to install the latest ones
git submodule update --init
type jupyter notebook in the console
go to code/dashboard_LIMEcraft.ipynb
choose venv in Kernel -> Change kernel
run the whole notebook
type http://127.0.0.1:8001/ in the web browser
If you have problems with running code on GPU, put os.environ['CUDA_VISIBLE_DEVICES'] = '-1' after import os in code/dashboard_LIMEcraft.ipynb.
go to web browser and download full_skin_cancer_model.h5 from https://www.kaggle.com/kmader/deep-learning-skin-lesion-classification/data
put the model in the folder code
change selected model in code/dashboard_LIMEcraft.ipynb in section "Choose model"
Paper for this work is avaliable at: https://link.springer.com/article/10.1007/s10994-022-06204-w
If you find our work useful, please cite our paper:
@article{Hryniewska2022LIMEcraft,
title={{LIMEcraft: Handcrafted superpixel selection and inspection for Visual eXplanations}},
author={Weronika Hryniewska and Adrianna Grudzień and Przemysław Biecek},
journal = {Machine Learning},
year={2022},
month = {jul},
eprint={2111.08094},
primaryClass={cs.CV},
publisher = {Springer Science and Business Media {LLC}},
keywords = {Explainable AI, superpixels, LIME, image features, interactive User Interface},
doi = {10.1007/s10994-022-06204-w},
url = {https://doi.org/10.1007%2Fs10994-022-06204-w}
}
