[SOLVED] How can I transfer parameters from one gpflow model to another to gain similar results?

How can I transfer parameters from one gpflow model to another to gain similar results?

Suppose I have a trained model

m = gpflow.models.SVGP(
     likelihood=likelihood, kernel=kernel, inducing_variable=Z, num_data = len(X_train)
)

is it possible to transfer its parameters to another model and achieve similar results? For example

model = gpflow.models.SVGP(kernel=m.kernel,
                           likelihood=m.likelihood,
                           inducing_variable=m.inducing_variable,
                           num_data=m.num_data)

But this example fails since model has poor results. Are there some other parameters, which should be added to the signature, or it is impossible in principle?

Solution

Yes, the SVGP (as well as VGP) model predictions crucially depend on the q(u) distribution parametrised by model.q_mu and model.q_sqrt. You can transfer all parameters (including those two) using

params = gpflow.utilities.parameter_dict(model)
gpflow.utilities.multiple_assign(model, params)

(see this notebook for more context)