Manually calculate GP hyperparameter gradients

Misc.
1

Due to the slow gradient calculation of gaussian process hyperparameters using autograd, I want to look into implementing my own version of the gradients (for a very limited set of covariance functions).

To do this I looked around and think that guides are what I am looking for.

How would I now define a guide and pass it to the training e.g. using the standard GP regression example? (Gaussian Processes — Pyro Tutorials 1.8.4 documentation)
Just for the sake of the example, let’s say I would just multiply each hyperparameter by 1.1 (i.e. add 10%), how should the guide look like?
Any directions and tips appreciated!

2

guides are unrelated to custom gradients. custom gradients are implemented using torch machinery and as such having nothing to do with pyro per se, e.g. see this example in gpytorch. i don’t know your use case but it’s unlikely that your custom gradients will win you gigantic speed gains. maybe a factor of 2 but not a factor of 5 or 10.

3

Thank you for clarifying my misunderstanding!
I will look into it.