WARNING: This is just a hack for now and completely untested.
There is only support for a DBN where the first layer is a Gaussian-Bernoulli RBM and the subsequent layers are Bernoulli-Bernoulli RBMs. The goal is implement methods for training Deep Belief Networks in Julia. We are using the Restricted Boltzmann Machine implementation from Boltzmann.jl and a HDF5 data source based on an implementation in Mocha.jl. We may either replace the latter with a simpler HDF5 source implementation or implement a training algorithm in a way that's more compatible with Mocha.
over 2 years ago