Export Gumbi model to PyMC3 and example workflow for cross_validate

[lilianschuster]

Thanks a lot for creating Gumbi. I was playing around a bit with it and I hope it will evolve further!
What I was trying to do is actually explained here: https://discourse.pymc.io/t/use-exact-gaussian-process-model-from-gpytorch-as-emulator-in-pymc3/8680. Do you think that I can use Gumbi to do sth. similar as done in GPyTorch ( https://docs.gpytorch.ai/en/stable/examples/01_Exact_GPs/Simple_GP_Regression.html) ?
If yes, is there a simple method to export the fitted gumbi model in order that the gumbi.predict can be used inside of "pure" pymc3 again (aesara compatible that it can run with the NUTs sampler) ?

Apart from that, I have another question: I was able to fit a GP of my data with Gumbi, however I could not really check its performance? Do you have an example code where you use the cross_validate method? I did not really manage to get it working:
I did:
gp.cross_validate(['melt_f', 'prcp_fac', 'temp_bias'], n_train = 200)

-> but then I got a TypeError: init() missing 1 required positional argument: 'outputs'. When I do gp.outputs, however, I get the right output name ?!

Thanks a lot in advance!

[JohnGoertz]

Hi Lilian, thanks for giving Gumbi a try!

With regards to cross_validate, this is indeed a bug, I'm working on fixing it now. Thanks for pointing it out!

For your other question, I'm not really certain. The underlying model is accessible via gp.model, so NUTS sampling should be as simple as

with gp.model:
    pm.sample()

That being said, using GP predictions within a larger pymc model can be bit tricky. Often Latent GPs are better for that (although slower), but I haven't decided on the best way to implement those in Gumbi yet. I'm not exactly sure what you want to do, could you provide some code/pseudocode to illustrate how you would want this to work?

[lilianschuster]

Hi John,
Thanks for your fast answer. I added some code in the discourse forum (https://discourse.pymc.io/t/use-exact-gaussian-process-model-from-gpytorch-as-emulator-in-pymc3/8680/4) . I have tried to do the same with Gumbi but the problem of how I can use the GP model for the actual calibration remains (see my additional comment in PyMC discourse). I guess this is what you meant that GP predictions can be tricky in larger PyMC models? Do you know about an example workflow using Latent GP that could help me in my case?

[JohnGoertz]

Okay, I made significant changes to cross_validate that should fix the issue (hopefully without creating new ones). Can you give it a try and see if it works on your data? Note that I've changed around some of the inputs and the output format, take a look at its new signature in the docs.