F&F - Fast and Functional Structured Data Generators

Code for the paper "Fast and Functional Structured Data Generators Rooted in Out-of-Equilibrium Physics" by Alessandra Carbone, Aurélien Decelle, Lorenzo Rosset and Beatriz Seoane.

Installation

All the dependencies needed to run the code in this repository can be installed using pip though

pip install -r requirements.txt

or, using conda,

conda install --yes --file requirements.txt

To set up the environment variables, add the following lines to the .bashrc file:

export PATH=${PATH}:/installation_path/RBM-Proteins
export RBMHOME=/installation_path/RBM-Proteins

Example data

To download an example of an input file for training the RBM using the MNIST dataset, use

wget --no-check-certificate 'https://drive.google.com/uc?export=download&id=1XiP_KPKuGZmxoqQz6tnVqUFlxf44S5kX' -O 'data/MNIST.h5'

Training RBMs

Dataset format

Datasets are handled through the .h5 format. The following fields are needed:

Mandatory:

• train: training set data;
• train_labels: labels for the training data in string format. If no label is available for specific data points, use the dummy label '-1';

Optional:

• train_names: train set names for the tree leaves;
• test: test set data;
• test_labels: labels for the test data in string format. If no label is available for specific data points, use the dummy label '-1';
• test_names: test set names for the tree leaves.

Launching a training

The basic syntax for starting the training of an RBM model is

./rbm-train.sh -m <model_type> -d <path_to_data>

The list of available models can be printed through

./rbm-train.sh -h

and it consists of all the classes in the repository src/RBMs without the suffix RBM.py.

In order to know all the available training options for a particular type of RBM model, use

./rbm-train.sh -m <model_type> -h

The Python files that handle the training of the models are contained in src/train.