Adapt code to include learnable parameters in interpolation matrix?

[finnBsch]

Hi,
I am trying to work off your paper + code to include learnable hyperparameters in the interpolation matrix W. However, it results in a GPyTorch error "Trying to backward through the graph a second time". I guess this is due to some issues with the caching where the computation graph for the W matrix is lost as it is not intended to be used with hyperparameters. Could you give me any advice on how to fix that?

[wjmaddox]

in general, thats going to break the online nature of the algorithm so that's most likely why it's throwing errors. I suppose one could do a meta-learning like strategy if you're not going to follow our approach in the paper of a somewhat stale head, which are implemented as stems in our model classes here

[finnBsch]

I understand. I was attempting to extend your work to non-stationary fields by warping the inputs using parameters that are learnable. Consequently, this will lead to a changing warp function. This will however lead to a time-varying matrix W as it is computed from interpolations from grid to warped inputs. If I understand correctly, this is not going to work, right?

[wjmaddox]

that's correct. however, if you fix the W matrix, you can still move the inputs around to some extent