Fix CI

[simsurace]

This PR is to try to fix #114. I will start by re-testing on Julia 1.8, because the last green CI run was before 1.9 was released.

[simsurace]

1.6 is expected to fail. But interesting that the old manifest + Julia 1.8 are needed to pass all tests.

[simsurace]

Ok, adding some rrules was sufficient. They are probably not as fast as they could be and should be upstreamed to ChainRules.jl, but it should do for now.
Let's revert the debugging changes now.

[coveralls]

Pull Request Test Coverage Report for Build 6403558960

• 11 of 11 (100.0%) changed or added relevant lines in 2 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+20.1%) to 88.008%

---

Totals
Change from base Build 6289961279:	20.1%
Covered Lines:	1365
Relevant Lines:	1551

---

💛 - Coveralls

[simsurace]

Managed to fix it. There was an API problem with dtc from AbstractGPs 0.5.16->0.5.17 . I'm now using the new API, I think TemporalGPs should probably not be extending the old one (#116).
Will do some cleanup work later to bump compat and get rid of the deprecation warnings before merging, but I ask for a review to see if the changes are ok.

[simsurace]

Unbelieveable, how did it break again?

[simsurace]

Seems like I repeated the same "mistake" as Theo in this commit. Seems like fixing those deprecation warnings is not so easy. I clearly don't know how to fix all this AD stuff :)

[simsurace]

The examples seem to take prohibitively long and hitting the timeout.

[willtebbutt]

@simsurace modulo the tests timing out, this looks great.

Before I review properly, I'd like to know whether you would you be willing to upstream the kron rules to ChainRules.jl? This would avoid type-piracy here, and would ensure that everyone gets the benefit of them.

[simsurace]

I meant to do that, but I haven't yet managed to write down rules that pass all tests.

[willtebbutt]

Would it be helpful if I took a look at testing the code currently on this branch?

[simsurace]

Thanks, I'll open a PR draft and link it here.

[simsurace]

I think getting the ChainRules PR merged will take a while. To not leave TemporalGPs in a broken state, and to be able to release these fixes, it would probably be good to invent an internal method _kron and define rules on that in order to not commit type piracy.

[willtebbutt]

Agreed. Looking at the Zygote implementation of kron I'm a bit confused why it is not working correctly here though. Do you have any intuition why it isn't?

Sorry to slow things down with additional questions, I just want to make sure that we've not missed anything.

[willtebbutt]

I think it's just https://github.com/JuliaGaussianProcesses/TemporalGPs.jl/blob/master/bench/single_output_gps.jl ? In any case, it's really computing the logpdf and its gradient as a function of the kernel parameters and observations for the three Matern kernels that we've got implemented, and a variety of different numbers of observations (definitely uses RegularSpacing rather than arbitrary spacing of inputs).

[simsurace]

Two-argument map strikes back. I think they can be replaced by broadcasts or comprehensions...

[simsurace]

Taken together, these optimizations amount to a 2-3 orders of magnitude improvement in approx_space_time_learning for T = 1000.

[simsurace]

Finally all green 🎉. I will now restore the original problem sizes for the space-time examples.

@willtebbutt, are you happy to merge this now?

[simsurace]

Fixes #118

[simsurace]

I removed the "needs version bump" label because the version that is currently on master hasn't been released yet.

[simsurace]

The SIGTERM in the last example is new.. I will investigate

[simsurace]

Ok, it seems that the augmented_inference example is a bit flaky in CI. Locally it has not failed me. I don't know why it occasionally gets killed. @willtebbutt can we merge?

[willtebbutt]

Thanks for sorting this. Yup, please feel to merge!