Einsum Networks -- Fast and Scalable Learning of Tractable Probabilistic Circuits

This is a personal fork of the official EinsumNetworks implementation.

Additional features of this fork include:

• Add PoissonArray to model poisson distributed random variables
• Add BernoulliArray for bernoulli distributed random variables
• Add optional random seed argument to EinsumNetwork.sample(...) which sets the RNG seed for this sample(...) call
• Add MultiDistArray to allow for different random variables to be modeled by different distributions

---

PyTorch implementation of Einsum Netwrks, proposed in

R. Peharz, S. Lang, A. Vergari, K. Stelzner, A. Molina, M. Trapp, G. Van den Broeck, K. Kersting, Z. Ghahramani, Einsum Networks: Fast and Scalable Learning of Tractable Probabilistic Circuits, ICML 2020.

We are still about to clean the code and add some experiments, but the implementation is already fully there and ready to play.

Setup

This will clone the repo, install a python virtual env (requires pythn 3.6), the required packages, and will download some datasets.

git clone https://github.com/cambridge-mlg/EinsumNetworks
cd EinsumNetworks
./setup.sh

Demos

We have add some quick run demos, to illustrate the usage of the code.

source ./venv/bin/activate
cd src
python demo_mnist.py
python demo_debd.py

Train Mixture of EiNets on SVHN

source ./venv/bin/activate
cd src
python train_svhn_mixture.py
python eval_svhn_mixture.py