This work is highly inspired by van Bergen et al. (2015) and reimplments that model and the underlying generative models using TensorFlow. This allows for easy calculation of the gradients of the parameters wrt the likelihood.
This leads to model fits that are orders of magnitude faster than implementations that use standard numpy or matlab, even when "only" CPUs are used.