PyTorch-ProbGraph is a library based on amazing PyTorch (https://pytorch.org) to easily use and adapt directed and undirected Hierarchical Probabilistic Graphical Models. These include Restricted Boltzmann Machines, Deep Belief Networks, Deep Boltzmann Machines and Helmholtz Machines (Sigmoid Belief Networks).
Models can be set up in a modular fashion, using UnitLayers, layers of Random Units and Interactions between these UnitLayers. Currently, only Gaussian, Categorical and Bernoulli units are available, but an extension can be made to allow all kinds of distributions from the Exponential family. (see https://en.wikipedia.org/wiki/Exponential_family)
The Interactions are usually only linear for undirected models, but can be built from arbitrary PyTorch torch.nn.Modules (using forward and the backward gradient).
There is a pre-implemented fully-connected InteractionLinear, one for using existing torch.nn.Modules and some custom Interactions / Mappings to enable Probabilistic Max-Pooling. Interactions can also be connected without intermediate Random UnitLayers with InteractionSequential.
This library was built by Korbinian Poeppel and Hendrik Elvers during a Practical Course "Beyond Deep Learning - Uncertainty Aware Models" at TU Munich. Disclaimer: It is built as an extension to PyTorch and not directly affiliated.
A more detailed documentation is included, using the Sphinx framework. Go inside directory 'docs' and run 'make html' (having Sphinx installed). The documentation can then be found inside the _build sub-directory.
There are some example models, as well as an evaluation script using the EMNIST dataset in the examples
folder.
This library is distributed in a BSD 3-clause license.
The library is accessible via the PyPi repository and can be install by: pip install pytorch_probgraph
Ian Goodfellow and Yoshua Bengio and Aaron Courville, http://www.deeplearningbook.org
Jörg Bornschein, Yoshua Bengio Reweighted Wake-Sleep https://arxiv.org/abs/1406.2751
Geoffrey Hinton, A Practical Guide to Training Restricted Boltzmann Machines https://www.cs.toronto.edu/~hinton/absps/guideTR.pdf
Ruslan Salakhutdinov, Learning Deep Generative Models https://tspace.library.utoronto.ca/handle/1807/19226
Honglak Lee et al., Convolutional Deep Belief Networks for Scalable Unsupervised Learning of Hierarchical Representations, ICML09
G.Hinton, S. Osindero A fast learning algorithm for deep belief nets