different kernel structure for each latent function in LMCVariationalStrategy

[michaelcao28]

Hi,

Suppose I am creating an LMC multi-task GP with num_task = num_latents = 3. How do I enable each latent function to have a different kernel structure? For example, latent_1 to have RBFKernel, latent_2 has PeriodicKernel, and latent_3 has MaternKernel.

Thanks!

[gpleiss]

Are you doing exact regression, or are you using a variational model? If you're doing exact regression, you can use the LCMKernel: https://docs.gpytorch.ai/en/stable/kernels.html#gpytorch.kernels.LCMKernel

[heikestein]

Hi Geoff,
I'm trying to do something similar, with multiple kernels of the same type that have distinct active input dimensions (and that should have a fixed but distinct mapping on output dimensions).
I have a large dataset and am using a variational model. Is there a way to do this?
Thanks!

[heikestein]

Ok, if anyone's reading along, I think I've solved the issue by passing on a num_latents x num_active_dim shaped tensor of kernel.active_dims indices to the LazyEvaluatedKernelTensor's _get_item() method, and then indexing into different input dimensions of inputs x1 and x2 after expanding them to batch_shape.
Regarding fixed output mappings, this is easily solved by fixing the lmc_coefficients of the LMCVariationalStrategy to the respective values.
For now this seems to work as intended, so I consider this issue solved.

[fjaraavilaa]

Hey, it is still not super clear to me. In the end you used the LMC Variational Strategy and also the LMC kernel? I am going on with a similar issue.