altairmn / Diss-CD

Investigation of dissimilar contrastive divergence method.

Instructions

The dataset(s) must be present in the same folder as the program. They should have the .pkl.gz extension.

To download MNIST and CIFAR10 (processed for this project), use the following links:

MNIST: http://bit.ly/1Tvjfdu
CIFAR10: http://bit.ly/24wzcez

Description

rbm.py

This file contains the core program. To view help on how to run, use the --help flag.

Example:

python rbm.py --help

Example:

python rbm.py --dataset=mnist.pkl.gz --neg-dataset=cifar_disscd.pkl.gz --epochs=1 --lr=0.1 --k=1 --output-folder=verification --n-hidden=200 --batch-size=40 --pcd

mnist_digits_pcd.py

This file trains an RBM for each digit separately using PCD.

Example:

python mnist_digits_pcd.py

mnist_digits_disscd.py

This file trains an RBM for each digit separately and untrains on the rest of the digits.

Example:

python mnist_digits_disscd.py