Try the Optim.jl branch with potential warning fix

[juliohm]

cc: @pkofod

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 77.76%. Comparing base (8e8c6c0) to head (91ab4de).

Additional details and impacted files

@@           Coverage Diff           @@
##             main      #79   +/-   ##
=======================================
  Coverage   77.76%   77.76%           
=======================================
  Files          47       47           
  Lines        1390     1390           
=======================================
  Hits         1081     1081           
  Misses        309      309           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[juliohm]

I've updated the PR to the latest version of the code. The warnings are still there, and there are test failures with this Optim.jl branch.

[juliohm]

Thank you @pkofod. It seems that the new implementation doesn't satisfy the box restriction [0, 1]. Some parameters are estimated outside the box in one test.

[pkofod]

that should definitely not happen, let me see

[pkofod]

in an error like this https://github.com/JuliaEarth/GeoStatsFunctions.jl/blob/optim-new/src/theoretical/matrices.jl#L30C1-L31C1 I think it would be useful if the error prints the obtained values

[pkofod]

try to rerun on my latest commit

[juliohm]

@pkofod I've restarted the jobs, but they still fail with the same issue.

[pkofod]

Suggested change

	Optim = "2.0.1"
	Optim = "2.1.0"

I suppose it should get it from sources, but for some reason LBFGSB was not available on one runner?

[pkofod]

ah, only CI with Julia > 1.11 will have sources I believe, so its is bound to fail.

[pkofod]

if you can print the values in the error branch we can maybe come a bit wiser here

[juliohm]

Thank you. I will look into it after I am done with some tasks I had planned for the week.

[pkofod]

I released it as v2.2.0

[juliohm]

I will try to take a look into it again today.

[pkofod]

I think it's a corner case where all these

GeoStatsFunctions.jl/src/fitting/transiograms.jl

Line 210 in 797ca68

s = sum(abs, p)

are driven to exactly 0. Then you get 0/0 => NaN. LBFGSB can literally step to the bound.

[pkofod]

you could reparameterize such that the proportions were just unconstrained z and then use a p = softmax(z). I'm not sure how you build the gradients, but if you write them up you need to adjust for the transformation of course. Alternatively, you can add a very small lower bound, but then no proportions can be truly zero. This would rule out maxproportions though, I don't think that can be expressed.

[juliohm]

Closing in favor of #85