In pymc3, how can I make predictions by averaging the particles (models) if I use Stein variational gradient descent for Bayesian neural nets?
In the tutorial, I found the code for GMM:
approx = pm.fit(method=pm.SVGD(n_particles=200, jitter=1.))
trace = approx.sample(10000)
And I also wonder what distribution the second line is sampling. I am confused because unlike ADVI, “approx” in SVGD should return a set of particles instead of a distribution.