Probabilistic Integral Circuits - PICs

This repository is the official implementation of Probabilistic Integral Circuits.

@InProceedings{gala24pic,
    title={Probabilistic Integral Circuits},
    author={Gala, Gennaro and de Campos, Cassio and Peharz, Robert and Vergari, Antonio and Quaeghebeur, Erik},
    booktitle={Proceedings of The 27th International Conference on Artificial Intelligence and Statistics},
    year={2024}
}

Datasets

• MNIST-famility datasets: all MNIST-famility datasets are available in torchvision, and automatically downloaded if needed.

• PTB288: download it here, and place it in data/ptbchar_288

• Binary datasets (DEBD): download them here, and place them in data/debd

• UCI datasets: download non pre-processed datasets here or pre-processed datasets here, and place them in data/UCI

Training PICs

python train_pic.py -ds mnist                   -bs 256 -nip 128 -int trapezoidal 
python train_pic.py -ds fashion_mnist           -bs 256 -nip 128 -int trapezoidal 
python train_pic.py -ds emnist -split mnist     -bs 256 -nip 128 -int trapezoidal 
python train_pic.py -ds emnist -split letters   -bs 256 -nip 128 -int trapezoidal 
python train_pic.py -ds emnist -split balanced  -bs 256 -nip 128 -int trapezoidal 
python train_pic.py -ds emnist -split byclass   -bs 256 -nip 128 -int trapezoidal

Training HCLTs

python train_hclt.py -ds mnist                  -bs 256 -hd 128
python train_hclt.py -ds fashion_mnist          -bs 256 -hd 128
python train_hclt.py -ds emnist -split mnist    -bs 256 -hd 128
python train_hclt.py -ds emnist -split letters  -bs 256 -hd 128
python train_hclt.py -ds emnist -split balanced -bs 256 -hd 128
python train_hclt.py -ds emnist -split byclass  -bs 256 -hd 128