Multinomial regression on iris dataset convergence issues

I’m trying to replicate the multinomial regression on the iris dataset described in this Github issue:
https://github.com/pymc-devs/pymc3/issues/1004
As far as I can tell, the only thing I’ve changed from the example in the thread is setting the number of cores to 1 to work around the Windows parallel sampling bug, but when I run this the sampling crawls and doesn’t converge:


I’ve replicated the same behavior on a Mac with parallel sampling in place. The behavior persists if I fit a subset of the data from only two categories (instead of 3). A binomial regression for a single category (changing the softmax function to sigmoid) runs as expected and converges. Am I missing something or could there be a bug here?

Your model is over parameterized - it is thus unidentifiable and doesnt converge. You can for example have a look at Multinomial hierarchical regression with multiple observations per group ("Bad energy issue")

Thanks for the help! I’ll take a look at that thread.