jitter init should not give up if first try is nan

[twiecki]

Our default init is quite brittle when it comes to models that easily have nans or -infs for certain parts of the parameter space. STAN retries multiple times if the current value is no good and I think we should too. We could output to the user how many times we had to try and if it's high enough maybe tell them to try and make the model more robust.

[ricardoV94]

I could work on this. Does anyone happen to have a simple "brittle" model that could be used as a testing case?

[twiecki]

You could try with a model like htis:

with pm.Model():
    pm.HalfNormal('x', transform=None)

I haven't tried it so it might not work with this, but maybe can serve as a starting point.

[ricardoV94]

Seems to work after a few tries. Thanks