Code to accompany Efficient training of energy-based models via spin-glass control

Alejandro Pozas-Kerstjens, Gorka Muñoz-Gil, Miguel Angel García-March, Antonio Acín, Maciej Lewenstein, and Przemyslaw R. Grzybowski

This is a repository containing the code for models with Restricted Axons and training via Pattern-InDuced correlations (RAPID), developed in the article "Efficient training of energy-based models via spin-glass control. Alejandro Pozas-Kerstjens, Gorka Muñoz-Gil, Eloy Piñol, Miguel Ángel García-March, Antonio Acín, Maciej Lewenstein, and Przemysław R. Grzybowski. arXiv:1910.01592."

All code is written in Python.

Libraries required:

• ebm-torch for energy-based models
• gc for garbage collection
• itertools for combinatorial operations
• math and numpy for math operations
• matplotlib for plot generation
• pytorch >= 0.4.0 as ML framework
• tqdm for custom progress bar

Files:

• comparison: example of use. Trains restricted Boltzmann machines using commonplace samplers and via RAPID, and computes various indicators of quality of training.

• MNIST training: example of use. Trains an RA-RBM in the MNIST dataset.

• rapid: contains the relevant classes.

• utils: additional functions relevant for the example.