HigherOrderGP Posterior eval mode

[ChristianMichelsen]

In HigherOrderGP, once you call .posterior it automatically sets it in eval mode, but does not set it back to train mode once it's finished. Is there are reason why it does not do that?

I've had a couple of times where I initialize a HOGP and just quickly check predictions at first before optimizing the MLL to be able to see the progress but without any progress due to the initial call to .posterior. This had left me spending more time on debugging more than I am proud of 😅

[saitcakmak]

Hi @ChristianMichelsen. We generally do not put models in .train outside of model training. Calling train() on a GPyTorch Module will trigger a self._clear_cache() call, which will destroy important caches that make repeated model inference efficient.

[ChristianMichelsen]

Would it then make sense to put the gp into .train() mode in e.g. fit_gpytorch_mll_scipy?

Also, regarding the caches, where can I read more about them, that sounds interesting!

[saitcakmak]

Would it then make sense to put the gp into .train() mode in e.g. fit_gpytorch_mll_scipy?

We do this in fit_gpytorch_mll, the main entry point for model fitting, which accepts fit_gpytorch_mll_scipy/torch as the optimizer input. I don't think it'd hurt to do this in fit_gpytorch_mll_scipy as well, but I already see a note about this in the docstring:

The model and likelihood in mll must already be in train mode.

Also, regarding the caches, where can I read more about them, that sounds interesting!

I don't know if there is a good place to read about them. I'm mostly familiar with them through working with the code but you can check GPyTorch docs https://docs.gpytorch.ai/.