lmj.rbm is a small Python library that contains code for using and training Restricted Boltzmann Machines (RBMs), the basic building blocks for many types of deep belief networks. Variations available include the "standard" RBM (with optional sparsity-based hidden layer learning); the temporal net introduced by Taylor, Hinton & Roweis; and convolutional nets with probabilistic max-pooling described by Lee, Grosse, Ranganath & Ng.
Mostly the code is being used for research in our lab. Hopefully others will find it instructive, and maybe even useful !
Installation
Just install using the included setup script :
python setup.py install
Or you can install the package from the internets using pip :
pip install lmj.rbm
Testing
To try things out, download the MNIST digits dataset :
curl http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz | gunzip -c > train-images.ubyte curl http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz | gunzip -c > train-labels.ubyte
Then install glumpy :
pip install glumpy
Then run the test :
python test/mnist.py -i train-images.ubyte -l train-labels.ubyte
If you're feeling overconfident, go ahead and try out the gaussian visible units :
python test/mnist.py
--images train-images.ubyte --labels train-labels.ubyte --batch-size 257 --l2 0.0001 --learning-rate 0.2 --momentum 0.5 --sparsity 0.01 --gaussian
The learning parameters are squirrely, but if things go right you should see a number of images show up on your screen that represent the "basis functions" that the network has learned when trying to auto-encode the images you are feeding it.
Product's homepage
Requirements:
· Python
Find us on Google+
