• Python, Numpy, Scipy
• Theano 0.6 (Bleeding edge version)
• Pylearn2 0.1

Easily reproduce the results of:
"BinaryConnect: Training Deep Neural Networks with binary weights during propagations".

python mnist_mlp.py

This python script will train a MLP on MNIST with the stochastic version of BinaryConnect. It should run for about an hour on an old GPU (Tesla M2050). The final test error should be around 1.2%.

mnist_mlp.py contains all the relevant hyperparameters. It is very straightforward to modify it. layer.py contains the binarization function (binarize_weights).