Implementing new Models (Distributions)


I need to implement some distributions (such as a gaussian walk, or in general an AR1 distribution) that aren’t available in edward (or tensorflow). In general, these are probability distributions over variable length 1-d tensors (strictly speaking they are stochastic processes, but programmatically it’s useful to reason about them as distributions like any other).

Is the recommended course of action to subclass tf.contrib.distributions.Distribution in my own code, then to use a similar method as the edward source to auto generate a RandomVariable subclass?


Apologies, missed this page in the docs