Sparse GP examples and deep GP examples about gpytorch HOT 3 CLOSED

The KISS-GP examples are instances of a sparse GP algorithm. I don't think we have specific plans to implement other techniques, e.g. SGPR/SVGP in the immediate future. SGPR wouldn't necessarily be challenging to implement as a new LazyVariable, but our internal "to do" list is pretty long.

To be honest, I'm not totally sure what would be involved in implementing deep GPs. We do support deep kernel learning (e.g. the MNIST example notebook) for both classification and regression.

from gpytorch.

Thanks for your reply. ^_^
I think there may be some interests on a more general sparse GP. As you say, implementing SGPR/SVPG is not difficult with your library, but I think it's still a good way to help the beginners and show your library's flexibility.
DKL is an outstanding work. But I'm not sure if GP community will return to deep GPs in the future. I'm trying to implement deep GPs now. However I can only build a deep GP model by repeating a single GP (layer?). Maybe a GP module like torch.nn is better? I'm not sure.
Thanks for your reply again.

from gpytorch.

SGPR is now implemented via the InducingPointKernel kernel and has an associated test at https://github.com/cornellius-gp/gpytorch/blob/master/test/examples/test_sgpr_regression.py. We eventually plan on making sure it is compatible with our variational inference, which will round out an implementation of SVGP.

It's worth emphasizing that we would expect SKI/KISS-GP (implemented via GridInterpolationKernel)
to be faster than SGPR in essentially all cases where it is applicable (including DKL).

We are going to close the issue for now, since I don't believe we have any plans to implement deep GPs internally in the immediate future. Pull requests are of course welcome for such a thing!

from gpytorch.