st--

Hi @antonykamp could you make sure to update RELEASE.md and CONTRIBUTORS.md as part of this PR? (See #1660 and #1661).

Names weren't consistent beforehand, either, and they're not really needed for anything specific. However, giving them unique names (note- names only need to be unique within a given python module)...

@awav would you be satisfied by "simply replace all double-underscores with single-underscores"? @vdutor what's the minimum that'd satisfy you to approve this PR?

Overall this looks very good - I've noticed that it doesn't show the type annotations in the signature anymore, but we can fix that by moving to proper Sphinx docstrings...

Hi @irum, thanks for the report; it's just hard to tell without more details what your issue is and what we may be able to do to fix it. Could...

@joacorapela if @vdutor is happy with it, we can merge it as is, but would be good to hear your thoughts on the same simplification for `predict_f`.: )

@a-mch I've just tried to reproduce this but couldn't (TensorFlow 2.2.0, but on Linux and on the CPU). I had to slightly adjust your minimal example (and updated your issue...

(NB- a function you wrap in `tf.function()` should not take complex objects such as the model as input!)

So looks like this is a GPU-only issue, and it still seems to be present with TF 2.3 and GPflow 2.1/develop. @a-mch what did you use as a "native TF"...

Hi @DolanDack, thanks for spotting this - can you explain how you've identified whether it runs on the GPU or not? I can't think of any GPU-specific code in GPflow,...