How to implement physical model as prior for mean function

[huerg]

After looking at the issues #674 and #1669 and using what I thought might help I am still uncertain about how to implement a custom mean function in gpytorch (to which I am new). I decided to start with a simple linear mean function derived from a linear regression as an example. Later it should encode, for example, a physical model.

Even though my question concentrates on the class “LinearRegressionMean” I am attaching a fully functional code to help clarify the whole question. Since I am not familiar with the usual way of doing it, please let me know if a shorter one would be better.

( I want to implement a custom mean function since they are, in my humble opinion, easier to model than kernels when one has a physical model describing approximate relationships but wants to make the model better with the help of data, falling back to the analytical function at the data gaps).

Many thanks!
Luis

custom_mean.zip

[lukasfro]

Hi @huerg, a linear mean function is already provided here. Hope this helps :) Let me know if you have further questions.

[huerg]

Dear @lukasfro , thank you for your help. I was wondering whether my question had been formulated in an unpolite manner since no one looked at it for weeks.
If I understand it right the code snippet below tells me that the regression coefficients are getting added to the graph and updated via gradient descent.

self.register_parameter(name="weights", parameter=torch.nn.Parameter(torch.randn(*batch_shape, input_size, 1)))
if bias:
self.register_parameter(name="bias", parameter=torch.nn.Parameter(torch.randn(*batch_shape, 1)))

Since I want to use an external model I will adapt the code and let you know if I get stuck.
Thank you!