Is there something like log_marginal_likelihood() ?

[MJimitater]

Hi folks, thank you for pushing forward the developments with v.1.5.0. 🎉

I was wondering: in sklearn after fitting a GP classifier, one can call the log marginal likelihood simply by gp.log_marginal_likelihood().

Is there something like that in GPyTorch?

Thanks

[gpleiss]

Are you looking for the log marginal likelihood of a GP classifier or a regressor? For a regressor, call:

model.likelihood(model(train_x)).log_prob(train_y)

[MJimitater]

Thanks! Indeed Im looking for the corresponding one of a GP classifier, is there something like it?

[gpleiss]

It's not possible to compute it exactly, since inference is intractable. With a variational model, you can get the variational lower bound via:

mll = gpytorch.mlls.VariationalELBO(likelihood, model)
lower_bound = mll(model(train_x), train_y)