st--

@awav How is this issue different from #1175 - is it not just a duplicate ?

This addresses #831

It's a bug in as much as the documentation isn't clear on saying it doesn't work, while it's natural to expect it to work (e.g. https://stackoverflow.com/questions/61852379/computing-the-log-marginal-likelihood-of-gpr-model-with-a-linearcoregionalizatio). Documentation should be clearer...

> sounds like we need two tickets, one bug ticket to handle the missing docs, and another feature ticket to implement it? You're right - there's already #1175 for the...

@tadejkrivec if you're still interested in contributing, that'd be great!

Ah, I had encountered this issue before and changed `set_default_float` and `set_default_int` to canonize the types to the numpy dtypes which works around these issues (TF accepts NumPy dtypes, but...

That seems like a reasonable change (sorry for spotting it rather belatedly!), I'd be happy to review a PR that makes these changes, if you would be up for contributing...

Just for clarification, GPflow **does** already support (L-)BFGS(-B) and other second-order optimizers (e.g. Conjugate Gradients) via [`gpflow.optimizers.Scipy`](https://gpflow.readthedocs.io/en/master/gpflow/optimizers/index.html#gpflow-optimizers-scipy), which connects TensorFlow's gradient computation with [`scipy.optimize.minimize`](https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.minimize.html) (which are implemented using numpy). This...

We missed a bug (#818) because of this lack of tests. #1583 adds a minimal regression test, but more thorough tests for this case would still be valuable.

> Would it make sense for `model.predictor()` to return a `Predictor` class, which could then implement multiple methods (`predict_f`, `predict_y` etc)? Yes, that would certainly be an option as well!...