pip install -r requirements.txt

The code uses Hydra for configuration.

You will have to download the CUB200 and GTSRB dataset and extract them to a folder called data/.

Each dataset code lives in its own subfolder with a train.py that:

• Trains neural networks (serving as interpretations of predicates and functions in the constraints).

• Saves:

  • Model checkpoints
  • Network predictions on in-distribution (ID) and out-of-distribution (OOD) datasets
  • Network feature representations for ID and OOD data

Run test.py to:

1. Train a Markov Logic Network (MLN)
2. Evaluate OOD detectors using the pre-extracted predictions and features

Use mining.py to automatically mine logical constraints from the data.

• compiler.py Compiles a domain-specific FOL notation into Python code.
• detectors.py Wraps detectors from the pytorch-ood library for easy use.
• mln.py Implements the Markov Logic Network.
• shared.py Utilities shared across datasets (e.g., MLN training loops, evaluation helpers).