Comparing Differentiable Logics for Learning with Logical Constraints

An experimental comparison of differentiable logics for machine learning with logical constraints.

Repository Structure

.
├── README.md                    - this file
├── base
│   ├── backends.py              - implementations of logic operators in Numpy, SymPy, and PyTorch
│   ├── dl2.py                   - implementation of DL2
│   ├── fuzzy_logics.py          - implementations of various fuzzy logics
├── experiments
│   ├── constraints.py           - the constraints used in the experiments
│   ├── main.py                  - the entry point containing training and test loop
│   ├── models.py                - the neural networks used in the experiments
│   ├── run.sh                   - script to run the experiments and replicate results from the paper
│   └── util.py                  - implementations of PGD and GradNorm
├── extras
│   ├── Makefile                 - creates the consistency and derivatives tables
│   ├── consistency.py           - evaluates the consistency of fuzzy logics
│   ├── derivatives.py           - creates LaTeX tables of derivatives of differentiable logic operators
│   ├── tautologies.py           - the tautologies used in the consistency calculation
├── reports                      - folder containing experimental results (.csv files)
│   ├── Makefile                 - creates plots and tables from the experimental results
└── requirements.txt             - pip requirements

Reproducing the results from the paper

The results from the paper can be replicated by running cd experiments && sh run.sh.

Assuming you have a reasonably up-to-date LaTeX distribution installed, the plots and tables from the publication can then be generated with cd reports && make.

Requirements

The experiments were run on Python 3.12.3 (but should probably work with newer versions as well). The provided requirements.txt can be used to install the required packages using pip install -r requirements.txt.