List view
Proposed changes: - Make `GPModel` `ExactGPModel`. This class would then only be used when performing exact inference. All Likelihoods would only be used for variational inference. - Refactor variational inference a bit. Rather than strapping on `_variational_strategy` to the outputs (which can be really hacky), I think we should have a `register_variational_parameters` approach. - Create a `gpytorch.optim.SVI` class for doing SVI (similar to Pyro). This would just be a thin wrapper around an existing pytorch optimizer, and would make sure to add the appropriate KL terms (and possibly hyperprior terms). - Similarly, it might not hurt to also have a `gpytorch.optim.ExactMLL` optimizer as well, which would be a really thin wrapper (this would mostly be for symmetry). - A slight refactor to the `.size()` method for lazy variables. - Refactor `KroneckerProductLazyVariable`
Due by February 20, 2018•12/12 issues closed