Fix build_laplace_objective behaviour

[st--]

Resolves #109

[willtebbutt]

LGTM subject to patch bump and CI passing (I'm convinced by the test)

[st--]

Hm, I'm not quite sure how to fix the autodiff issues 😞 @willtebbutt any ideas ?

[willtebbutt]

Hmmm a good place to start might be to figure out where the Core.Box is coming from, and why its length is being asked for. It doesn't look obviously like an AD-specific thing to me, although perhaps it is if things that aren't using AD are working properly?

[st--]

Hmmm a good place to start might be to figure out where the Core.Box is coming from, and why its length is being asked for. It doesn't look obviously like an AD-specific thing to me, although perhaps it is if things that aren't using AD are working properly?

That I can tell you: it's coming from the objective closure within build_laplace_objective!, because the externally-scoped f doesn't have a defined type, as it's nothing when first called and then Vector in subsequent calls (once it was first assigned in the line I added). The length comes from the map call in loglik_and_derivs.

[willtebbutt]

Hmm okay. So is a Core.Box making its way into the call to map before it gets assigned then? So it this actually an AD issue, or is it just to do with ensuring that f is initialised by the time it's needed?

[st--]

Hmm okay. So is a Core.Box making its way into the call to map before it gets assigned then? So it this actually an AD issue, or is it just to do with ensuring that f is initialised by the time it's needed?

I'm getting less sure here, but it's definitely not an issue in the forward pass (and all the Boxes seem to get hidden away), but AD fails...

[codecov]

Codecov Report

Merging #115 (0784872) into master (c751c05) will decrease coverage by 0.51%.
The diff coverage is 80.00%.

@@            Coverage Diff             @@
##           master     #115      +/-   ##
==========================================
- Coverage   94.13%   93.61%   -0.52%     
==========================================
  Files           5        5              
  Lines         324      329       +5     
==========================================
+ Hits          305      308       +3     
- Misses         19       21       +2     

Impacted Files	Coverage Δ
src/LaplaceApproximationModule.jl	96.89% <80.00%> (-1.19%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update c751c05...0784872. Read the comment docs.

[st--]

Would there be any significant gains from changing it to two fields, f::Vector and f_initialized::Bool or something like that?

[willtebbutt]

LGTM