This worked.
Thanks for the idea @junpenglao
I can use this for my current work.
But can’t we have a more clean way of accessing the training time (including MCMC and SVGD etc), because that will be very helpful for those who want to experiment on various Bayesian inference techniques that is supported by PyMC3.
Even if one tries to log/dump the time consumed for inferences exposing such measurements to the users will be useful.