Gpytorch: [Docs]

[dialuser]

📚 Documentation/Examples

Hi,
I have a problem where the output samples are 2D images of sizes N*N. Different images represents outputs of a PDE solver from different times. There's strong spatial correlation among the grid cells.
I have looked into MultitaskGPModel and GridGPRegressionModel.
The former maybe used if I treat the grid cells as different tasks, the latter seems to be similar. But I wonder if this is the best way to cast the problem using gpytorch. I'm mainly concerned with scalability (N^2). Any advice? Thanks ahead.

* Is there documentation missing? *

* Is documentation wrong? *

* Is there a feature that needs some example code? *

* Think you know how to fix the docs? * (If so, we'd love a pull request from you!)

• Link to GPyTorch documentation
• Link to GPyTorch examples

[jacobrgardner]

You might try the variational multitask models, which scale reasonably well with the number of output GPs: https://github.com/cornellius-gp/gpytorch/blob/master/examples/04_Variational_and_Approximate_GPs/SVGP_Multitask_GP_Regression.ipynb

[Balandat]

@wjmaddox has implemented the HOGP model from http://proceedings.mlr.press/v89/zhe19a.html that works quite well in these settings. Should be able to diff put this up as a PR once it's cleaned up and tested some more.

[dialuser]

@Balandat, I'm looking forward to using HOGP. Thanks.

[Balandat]

It might be a little bit until this is all polished, in particular the current version does not support efficiently computing predictive variances, but we should get there.

[wjmaddox]

Just bumping this thread now that the implementation is in botorch here. Let us know how it works for your use case.