Black Box Variational Inference (using Score Gradient)

Does PyMC3 have Black Box Variational Inference using Score Gradient? Like Algorithm 1 in the paper “Black Box Variational Inference” by Rajesh Ranganath, Sean Gerrish, David M. Blei in 2014 (Artificial Intelligence and Statistics).

I have a Finite Element Code as the model, therefore I cannot evaluate the model derivatives (implies, I cannot have likelihood derivatives). So I was looking for the Score Gradient approach for Variational Inference, where I just need to evaluate my likelihood function. In addition, are there any alternatives to perform inference on such models which are implemented in PyMC3?

Does anyone know about this? @junpenglao?


Hi, score gradients are not yet implemented. I guess you may try implementing custom score gradients that do the job. But this may appear too complicated