Thank you!!!
I assembled all the test code and results into this colab.
I took your suggestion about just the initial point failing, and I provided initial points for all the distributions. Not because this is the final solution, but it’s a quick and good test. If I start everything with the right initval’s, then all is good. But if I leave one out…
I get an error I don’t understand.
Initial evaluation of model at starting point failed!
Starting values:
{'lam_log__': array(-0.1364851), 'Intercept': array(3.05363133), 'Slope': array(0.07561529)}
Logp initial evaluation results:
{'lam': -0.74, 'Intercept': -3.93, 'Slope': -3.91, 'y': -inf}
You can call `model.debug()` for more details.
I have specified that lam is a HalfNormal. Why the !#%#% should it be trying to evaluate it (initially) at -0.74??? That is not allowed by the prior! And why doesn’t the Intercept have the initval I specified (4)?
My real model is going to add a logistic regression, so it will be harder to initialize and specify priors. So I want to understand what is going on and what is wrong with this simple model.
Thanks for the advice. Any clues about the initial values?
– Malcolm