yeah this is a known issue, matplotlib is producing different results dependent on both version and platform.

ok so the tests should pass for those two 🤞

hmm this label_categorical color was wrong and is still wrong. Given that it was broken I am ok with this PR, but we need to fix it. What is indicated as being C is actually background label. @timtreis I vaguely remember you workin on a fix for this. Am I correct?

I'm afraid the tests failing are the ones of the figures I have added, I wonder if the reason is precisely this version differing behaviour. Increase tolerance ? or someone has an ubuntu machine to recreate figures?

Increasing tolerance will not help as this does not account for large difference in colors which happens with different color for the labels or background.

Yeah, no idea why this is happening. Noticed it in #259 originally but I still have no idea what's causing it. Local to me it looks fine. For this case, I'm just using the image of the runner itself.

Good point, uploaded the new ones from the runner artifacts

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 83.76%. Comparing base (d43d3e7) to head (7935a44).

@@           Coverage Diff           @@
##             main     #287   +/-   ##
=======================================
  Coverage   83.76%   83.76%           
=======================================
  Files           8        8           
  Lines        1694     1694           
=======================================
  Hits         1419     1419           
  Misses        275      275           

wtf how can 3.9 and 3.10 not be consistent?

Because matplotlib. Had this before, in these cases for now if this happens we accept the PR

mmh but this is a bit suspicious, like the results seems to put labels with different orders, despite everything being generated by RNG. In the last commit, I took artefacts from 3.9 and copied it to 3.10 and still now both fails

@timtreis are you still planning to pick this up? Otherwise I can.

I think I tightened the GH actions enough so that 3.9 and 3.10 are now consistent. At least I didn't run into inconsistencies in the few last PRs. We still get the black bg on some of the raccoons, only on the runner though.

Can this be closed?

I checked the code and the one from Giovanni is correct. The max of the labels is 5, but the labels are non-contiguous, so using len(get_element_instances()) is the way to go.

As pointed out by @melonora, the plots from this commit 42a4ee6 were still wrong as the background should have been black and not colored.

The reason was this line here:

I have corrected this, verified the consistency with napari-spatialdata and regenerated the ground-truth plot. Now it's ready to merge.