PartitionShap demo: channel-wise explanations of raster-based machine learning models

Exploring PartitionShap on raster datasets of an arbitrary number of bands/channels

These notebooks use our SHAP fork with modifications to make it easier to use and visualize multi-channel explanations. The EuroSAT notebooks use TorchSat, a library for working with PyTorch models whose inputs are rasters with arbitraty number of channels.

Citation

@misc{ChannelWisePartitionShap,
  author = {Krell, E. and Friesen, J. and Judge, J.},
  title = {Partitionshap Multiband Demo},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/conrad-blucher-institute/partitionshap-multiband-demo}},
  commit = {5807c4c6eb452b31b88e61276ec7a56b70a61e35}
}

Notebooks

1. PartitionSHAP: ImageNet (RGB) demo
2. PartitionSHAP: EuroSAT (RGB) demo
   • pretrained weights: https://drive.google.com/file/d/14BYvrjem4dbmkmibmmSD2cDq2ZvStfVG/view?usp=sharing
3. PartitionSHAP: EuroSAT (13-band) demo
   • pretrained weights: https://drive.google.com/file/d/1gYtOpYdMCCMxCkEhRzK1dDyE6tOMIOKa/view?usp=sharing

3D SHAP viewer tool (prototype)

python SHAP_3D_viewer.py --help
Usage: SHAP_3D_viewer.py [options]

Options:
  -h, --help            show this help message and exit
  -f FILE, --file=FILE  Path to 3D SHAP values (.npz)
  -d DATA_NAME, --data_name=DATA_NAME
                        Name of SHAP values in the input SHAP values (.npz)
                        file.
  -e, --show_edges      Show edges of grid elements

Todo

• Evan: Use test instead of validation data for XAI runs in notebooks -> not currently repoducable since will choose a random validation case
• Josh: Add colab notebook with RGB & RGB+NIR model train, test, PartitionShap results (other XAI not is not for this repo!)
• Juliana: Add updated 3D tool, update readme with example for both .npz and .pickle files
• Does it make sense to "publish" the poster by including the PDF in this repo?