This repo contains the code to my masterthesis at TU Darmstadt (Quantifying and Explaining Latent Uncertainty: Probabilistic Circuits for Robust Deep Learning).
As backbone, efficientnet-v2 is used, while the probabilistic circuit employed is simple-einet.
Please note that the code in this repo is research-code and not very usable at the moment without modifications.
The dependencies can be found in the Dockerfile.
All models can be found in Models.py. This includes DDeXU (called 'EfficientNetSPN' there) and the baseline methods that we compare to.
To run an experiment load the corresponding run-method from the experiment, e.g. start_cifar10_calib_run(...) from experiments/cifar10_calib_experiment.py, or modify experiment_runner.py.