Code for models in the ICML 2015 paper: "Variational Generative Stochastic Networks with Collaborative Shaping"

I'm in the process of working back from my more recent work -- which involved substantial changes to the basic model implementations, optimization methods, etc. -- to match this repo's state to what I was doing when I wrote the ICML paper. The newer models in the Sequential-Generation repo are more powerful, FYI. The main problem with a "unimodal" reconstruction distribution is that forcing more overlap in the corruption distributions q(z|x1) and q(z|x2) -- via KL regularization -- becomes unreasonable when the required reconstruction distribution p(x|z) contains multiple significant modes. Using a sequential construction for the corruption process and reconstruction distribution allows truncation of the corruption process at varying degrees of overlap in q(z|x1)/q(z|x2). It also readily permits multi-modal reconstruction via a chain of composed unimodal reconstruction distributions.

The file "MnistWalkoutTests.py" shows the basic incantations for initializing and training the models described in the paper. The general VAE implementation is in "OneStageModel.py" and the collaboratively guided Markov chain stuff is in "VCGLoop.py". To perform more unrolling steps, the VCGLoop code would need to be modified to use Theano's scan op, rather than manual loop unrolling.