Gaussian processes fit using plates?

Is it possible to fit Gaussian process models in Pyro using plates? Since the GP interface appears to be higher-level than standard Pyro models (e.g. no model or guide need to be specified), it is unclear how to apply plate context managers to this class of model. The reason I ask is that I want to fit sets of GPs that share common lengthscales, so having each GP (and associated data) assigned to a plate seemed like a logical approach. Can this be done?

1 Like

Hi @fonnesbeck, batching GP is currently not supported, it is still an open issue. I think that we just need to make sure kernels and conditional works with batched inputs, then given a 3D input (batch_size x num_data x num_features), batching will automatically be activated. I will try to support this feature next week. Hopefully, it is not so involved. :slight_smile: For now, you can consider using GPyTorch, which supports batching out of the box.

1 Like