When we change the inducing points before re-training SVGPs and Deep GPs, we need to update q_mu and q_sqrt such that the posterior distribution is unchanged at the new inducing points.
The current equations only work when the mean function is zero. We need to add a correction term at those places;
When we change the inducing points before re-training SVGPs and Deep GPs, we need to update q_mu and q_sqrt such that the posterior distribution is unchanged at the new inducing points.
The current equations only work when the mean function is zero. We need to add a correction term at those places;
(new_q_mu =new_q_mu - model mean prediction at Z) https://github.com/secondmind-labs/trieste/blob/develop/trieste/models/gpflow/models.py#L1014
(f_mu = f_mu - model mean prediction at Z) https://github.com/secondmind-labs/trieste/blob/a4b25f8cde453c2a19399e66c3c440d51cfbac84/trieste/models/gpflow/utils.py#L330
DeepGPs currently do not have this problem because we don't have a scheme to update inducing points.
It is unclear whereas we have a test for whiten_points, but we should definitely have one with a non-zero mean function.