We provide two Jupyter notebooks, namely
./notebooks/lrp-nfnet.ipynbdemonstrates our LRP implementation for NFNet-F0. It reproduces Fig. D.2 in Supplementary Note D../notebooks/demo.ipynbdemonstrates our disentangled explanation framework. More specifically, the notebook shows how to obtain disentanged explanations from PRCA and DRSA for classbasketballusing activation and LRP context vectors fromVGG16-TVatConv4_3. The demonstration reproduces Fig. 2 in the main paper.
Remark: Make sure that PYTHONPATH includes $(pwd)/cxai when starting a Jupyter instance. Or, start the instance using PYTHONPATH=$(pwd)/cxai jupyter notebook.
We use Python version 3.8.6. Necessary dependencies can be installed via
pip install
58E0
-r requirements.txt
Please run the unit test command below to check that necessary functionalities work.
# testing important functions (approximately 3 minutes on CPUs)
make fast-test
# test all functions (approximately 6 minutes on Tesla V100)
make test