danstowell@0: 
danstowell@0: Spectrogram auto-encoder
danstowell@0: (c) Dan Stowell 2016.
danstowell@0: 
danstowell@0: 
danstowell@0: A simple example of an autoencoder set up for spectrograms, with two convolutional layers - thought of as one "encoding" layer and one "decoding" layer.
danstowell@0: 
danstowell@0: It's meant to be a fairly minimal example of doing this in Theano, using the Lasagne framework to make things easier.
danstowell@0: 
danstowell@0: By default it simply makes a training set from different chunks of the same single spectrogram (from the supplied wave file). This is not a good training set!
danstowell@0: 
danstowell@0: Notable (potentially unusual) things about this implementation:
danstowell@0:  * Data is not pre-whitened, instead we use a custom layer (NormalisationLayer) to normalise the mean-and-variance of the data for us. This is because I want the spectrogram to be normalised when it is input but not normalised when it is output.
danstowell@0:  * It's a convolutional net but only along the time axis; along the frequency axis it's fully-connected.
danstowell@0:  * There's no maxpooling/downsampling.
danstowell@0: 
danstowell@0: 
danstowell@0: SYSTEM REQUIREMENTS
danstowell@0: ===================
danstowell@0: 
danstowell@0: * Python
danstowell@0: * Theano (NOTE: please check the Lasagne page for preferred Theano version)
danstowell@0: * Lasagne https://github.com/Lasagne/Lasagne
danstowell@0: * Matplotlib
danstowell@0: * scikits.audiolab
danstowell@0: 
danstowell@0: Tested on Ubuntu 14.04 with Python 2.7.
danstowell@0: 
danstowell@0: USAGE
danstowell@0: =====
danstowell@0: 
danstowell@0:           python autoencoder-specgram.py
danstowell@0: 
danstowell@0: It creates a "pdf" folder and puts plots in there (multi-page PDFs) as it goes along.
danstowell@0: There's a "progress" pdf which gets repeatedly overwritten - you should see the output quality gradually getting better.
danstowell@0: 
danstowell@0: Look in userconfig.py for configuration options.
danstowell@0: