Master Thesis -- Lilli Joppien
Concept-Based Methods and Their Fidelity in Presence of Spurious Features Under a Causal Lense.
Advisors: Oana-Iuliana Popescu, Simon Bing
First Examiner: Prof. Dr. Jakob Runge
Second Examiner: Prof. Dr. Grégoire Montavon
Next to the conda environment (env.yml
) some other packages need to be installed to run this experiment
conda env create -n ENVNAME --file env.yml
zennit-crp
tigramite
from isntructions at https://github.com/jakobrunge/tigramite/If one desires to rerun the whole experiment, follow the instructions below,
else all data necessary for the compare_measures
notebook should be here.
If one wants to look at further visualizations of what the explanation heatmaps look like etc.
The trained models can be requested by making an issue here, or otherwise recomputed as seen below.
Training Models if access to SLURM based cluster with gpu nodes:
experiments
folder to clusterEXPERIMENT
is selected in python3 run_iterations.py
python3 script_parallel_iterations.py
python3 extract_infos.py
else ask for pretrained models Computing Explanations and Measures
python3 compute_normal_measures.py [1,2]
(choose 1 for watermark experiment or 2 for pattern)
This computes the relevance and attribution maps measures as well as the region-specific measures (RMA, RRA, PG)
note that if this has not previously been done, it will take about 20 minutes to compute all explanationspython3 compute_relmax.py recompute=True
python3 compute_latent_factors_gt.py
, this will also take a while when done for the first time.
This computes relevances and model ground truth effect for the other latent factors of our causal model: shape, rotation, scale, posX, posYExplore Measures and Plot Visualizations
compare_measures.ipynb
further_visualizations.ipynb