Aie Confiance...
https://www.youtube.com/watch?v=2gflQEF1WQU
KAA is a Python framework allowing to apply several methods and metrics from several dedicated XAI libraries on AI model to verify them. It is built around a kernel offering a TUI Text-based User Interface and hosting two sets of plugins:
- the plugins of the explainability libraries called pluginsXAI
- the plugins of the use cases called pluginsUCXAI
KAA must be considered as a comprehensibility tool. It is part of these mediation tools allowing to implement explainability methods in a guided way, to facilitate the adjustment of their parameters by adding knowledge of their influence. It then offers appropriate and relevant interfaces and visualizations of results to select the methods and metrics that are most suited to the user’s problem.
In this version of KAA, the AIX360 (v0.3.0), Alibi (v0.9.6), Captum (v0.7.0), PAIR saliency (v0.2.1), Shap (v0.44.1) and Xplique (v1.4.0) plugins are provided.
Full documentation can be found in ./documentation/Kaa_User_and_Integration_Manual.pdf
KAA requires a version of python equal to 3.8 and several libraries including Tensorflow and Numpy. Installation can be done:
- using Pypi:
pip install KAA-XAI- via downlading the v25.12 tag on our github:
Option 1: Using Git
git clone https://github.com/IRT-Saint-Exupery/KAA.git
cd KAA
git checkout tags/v25.12Option 2: Download the ZIP archive
wget https://github.com/IRT-Saint-Exupery/KAA/archive/refs/tags/v25.12.zip
unzip v25.12.zip
cd KAA-25.12Install the requirements:
pip install -r requirements.txt
Then install the XAI libraries you want:
pip install -r requirementsXAIlibs.txtNow that KAA is installed, you can launch the TUI KAA .
python3 kaa/kaa.py -u example/pluginsUCXAI/ -b example/UseCase/ -r ~/KAA_exampleor execute a command file of the example:
python3 kaa/kaa.py -u example/pluginsUCXAI/ -b example/UseCase/ -r ~/KAA_example -c example/example.cmdThe results are available in the -r path:
~/KAA_example
+-- fullReports
| +-- Resnet
| +-- UCRenaultWeldingResnet_fullReport__expl-data.pdf
+-- Xplique
+-- Occlusion__bz64_pz35_ps3_ov1
| +-- contactSheet.pdf
+-- Rise__bz64_ns2000_gz7_pp0-5
+-- contactSheet.pdf