GPs with non-Normal likelihoods in PyMC3: FloatingPointError

Thanks to Bill Engels for an example to how to build a GP model with non-conjugate likelihood!
https://bwengals.github.io/non-conjugate%20GPs.html

I tried to run it with the most recent Theano and PyMC3 installed. Unfortunately I get the nasty FloatingPointError when the model is built and NUTS uses ADVI to initialize.

The complete error message can be found here:
https://pastebin.com/BiM78urj

Just a wild guess: you might need to run from the WIP gp branch of @bwengals

Yes, @junpenglao is right, i edited that post to include that important bit of info…

Do you also have an up-to-date version of the FITC-VFE blog post?

That post is updated now

1 Like