In the "simple demonstration of coregionalisation" example in GPflow a VGP is used with Gaussian likelihoods for each GP. Is this approach equivalent to an exact regression model? I was trying to reimplement the example using GPR, but was unsuccessful so far, so I am not sure how to verify otherwise.
Edit:
I was following this example here, however in my case I have variable length input/output, i.e. each Xi-Yi pair have different length and own mean functions, by using mea = gpflow.functions.SwitchedMeanFunction(mean_list). kern is a kernel modelled as in the given example by kern = k1 * coreg and coreg = gpflow.kernels.Coregion(output_dim=output_dim, rank=rank, active_dims=[1]).
Using:
m = gpflow.models.GPR([X_augmented, Y_augmented], kernel=kern, mean_function=mea)
gives me one reasonable result regressing through the data points and a second one which seems very random and consists of parallel lines not regressing through any data points.
Thanks in advance, appreciate any hint.