A python implementation of training Deep Boltzmann Machine and generating Gibbs samples.

To train a DBM:

sh run.sh

After training, obtain Gibbs samples from the trained model:

sh run_gibbs.sh

To train a DBM with H1 units in the first hidden layer, and H2 units the second hidden layer, using initial learning rate r, batch size N, and train the model for T epochs with K persistent chains, run:

python trainDBM.py --trainData <path/to/train/data> --valData <path/to/val/data> \
  --hiddenSizes <H1> <H2> --nepoch <T> --rate <r> --batchSize <N> --numChains <K> \
  --outfile <output_filename> --outdir <output_directory>

To obtain $M$ Gibbs samples, each with $K$ steps, run the following command, where you need to specify the path to the .pickle file (the trained model given by trainDBM.py), and the path to save the Gibbs samples in a .csv file:

python gibbsDBM.py --steps <K> --chains <M> \
  --model <path/to/input/pickle/file> --outfile <path/to/output/csv/file>